content
stringlengths 1
15.9M
|
|---|
\section{Introduction}\label{intro}
The Hardy space $H^2(\Pi^+)$ and the weighted Bergman spaces $\mathcal{A}_{\alpha}^2 (\Pi^+ )$ of the upper half-plane $\Pi^+ $ are reasonably well understood Hilbert spaces of analytic functions (the spaces are defined in Section \ref{prelim}). \textit{Composition operators} $ C_{\tau} : f\mapsto f\circ \tau $, where $ \tau : \Pi^+ \longrightarrow \Pi^+$ is analytic, acting on these spaces are, however, much less studied when compared to their counterparts in the unit disc setting. Matache \cite{Ma1} found a condition for boundedness of $C_{\tau}$ on $H^2(\Pi^+) $ in terms of Carleson measures and showed later in \cite{Ma2} that there are no compact composition operators on $H^2(\Pi^+)$. In \cite{SS} Shapiro and Smith extended the non-compactness result to $\mathcal{A}_{\alpha}^2 (\Pi^+ )$. Boundedness of a composition operator $C_{\tau}$ on the Hardy or the weighted Bergman spaces of the half-plane has been proved to be equivalent with the angular derivative of the inducing map at infinity, denoted by $\tau' (\infty)$, being finite and positive (see \cite{Ma3, EW}). In addition, Elliott et al. \cite{EW, EJ} have shown that whenever $C_{\tau}$ is bounded, the operator norm, the essential operator norm and the spectral radius are all equal and determined by the quantity $\tau'(\infty )$. The above properties show that the spaces $H^2(\Pi^+)$ and $\mathcal{A}_{\alpha}^2 (\Pi^+ )$ differ significantly from their unit disc analogues $H^2 (\mathbb{D})$ and $\mathcal{A}_{\alpha}^2 (\mathbb{D})$ regarding the composition operators acting on them. We refer e.g. to \cite{Sh, CM} for expositions of the rich theory of composition operators on spaces defined on the unit disc.
It is natural to ask what the spectral properties of composition operators acting on the half-plane are. In fact, in the unit disc setting the spectral picture of composition operators has been completely determined when the inducing maps are \textit{linear fractional transformations} and the operators act on weighted Dirichlet spaces, including $H^2 (\mathbb{D})$ and $\mathcal{A}_{\alpha}^2 (\mathbb{D})$ for all $\alpha >-1$ (see, for instance, \cite{No, CM, Hu, Hi, Po, GS}). The spectra in the corresponding setting on the half-plane are largely unknown; in the (unweighted) Dirichlet space of $\Pi^+$ the spectra are known but besides that only the spectra (and the essential spectra) for invertible or self-adjoint parabolic and invertible hyperbolic composition operators acting on the Hardy space $H^2(\Pi^+)$ have been computed (see \cite{Ma4}). In contrast to the unit disc case, not all linear fractional transformations $\tau$ induce bounded composition operators on $H^2(\Pi^+)$ or $\mathcal{A}_{\alpha}^2 (\Pi^+ )$. Indeed, $C_{\tau}$ is bounded only when $\tau$ is a parabolic or a hyperbolic self-map of $ \Pi^+$ fixing infinity. In this paper we compute the spectra and the essential spectra of these composition operators.
In the parabolic case (see Theorem \ref{sppara} in Section \ref{secpara}) we obtain the following result:
\vspace{5pt}
\textbf{Theorem A.} Let $\tau$ be a parabolic self-map of $\Pi^+$, that is, $\tau (w)=w+w_0$, where $ \textnormal{Im}\, w_0 \geq 0$ and $w_0 \neq 0$. Then the spectrum of $C_{\tau}$ acting on the Hardy or the weighted Bergman spaces of the upper half-plane equals
$$ \overline{ \{ e^{iw_0t} : t\in [0,\infty ) \} } = \left\{ \begin{array}{ll}
\mathbb{T}, & \textnormal{ when } w_0\in \mathbb{R},\\
\{ e^{iw_0t} : t\in [0,\infty ) \} \cup \{ 0\}, & \textnormal{ when } w_0\in \Pi^+.
\end{array}
\right.
$$
Moreover, the essential spectrum coincides with the spectrum in both cases.
\vspace{5pt}
For a moment, write formally $H^2(\Pi^+) = \mathcal{A}_{-1}^2 (\Pi^+ ) $. We can summarize the results in the hyperbolic case (see Theorems \ref{sphypauto}, \ref{hyptauyks}, \ref{hyptaukaks} and Corollary \ref{sphypautoinverse} in Section \ref{hypsec}) as follows:
\vspace{5pt}
\textbf{Theorem B.} Let $\tau$ be a hyperbolic self-map of $\Pi^+$, that is, $\tau (w)=\mu w+w_0$, where $\mu\in (0,1)\cup (1,\infty)$ and $\textnormal{Im}\, w_0 \geq 0$. Then, for all $\alpha \geq -1$, the spectrum of $C_{\tau}$ acting on $\mathcal{A}_{\alpha}^2 (\Pi^+ )$ is
\vspace{5pt}
\begin{itemize}
\item[i)] $ \big\{ \lambda \in \mathbb{C} : \vert \lambda \vert = \mu^{ -(\alpha +2 )/2}\big\} $, when $w_0 \in \mathbb{R}$,
\vspace{5pt}
\item[ii)] $ \big\{ \lambda \in \mathbb{C} : \vert \lambda \vert\leq \mu^{ -(\alpha +2 )/2}\big\}$, when $w_0\in \Pi^+$.
\end{itemize}
Moreover, the essential spectrum coincides with the spectrum in both cases.
\vspace{5pt}
From these results we are also able to deduce the spectra of the parabolic and the hyperbolic composition operators on weighted Dirichlet spaces $\mathcal{D}_{\alpha}^2 (\Pi^+ ) $ for $\alpha >-1$ (see Theorems \ref{dirpara} and \ref{dirhyper} in Section \ref{secdir}).
There are similarities as well as differences when we consider the composition operators $C_{\tau}$ on the half-plane and $C_{\varphi}$ on the unit disc, where the inducing maps satisfy $\tau =h \circ \varphi \circ h^{-1}$ for some conformal map $h: \mathbb{D} \longrightarrow \Pi^+$. It is worth noting that the composition map $ C_h $ which gives a unitary equivalence in the (unweighted) Dirichlet space setting between $ C_{\tau}: \mathcal{D}_0^2 (\Pi^+)\longrightarrow \mathcal{D}_0^2 (\Pi^+)$ and $C_{\varphi}: \mathcal{D}^2 (\mathbb{D})/\mathbb{C} \longrightarrow \mathcal{D}^2 (\mathbb{D})/\mathbb{C} $ does not give even similarity in the Hardy or the weighted Bergman spaces since $C_h$ is not bounded below. Therefore, in general, we cannot deduce the spectra of $C_{\tau}$ on the half-plane from the spectra of $C_{\varphi}$ on the unit disc. Nevertheless, in the parabolic case the spectra are the same. The hyperbolic composition operators, on the other hand, seem to have a life of their own when comparing the spectral results between $\mathbb{D}$ and $\Pi^+$.
\section{Preliminaries and notation}\label{prelim}
Throughout the paper we will use the following notations: $\overline{\mathbb{C}}:=\mathbb{C}\cup \{\infty\}$ denotes the extended complex plane, $\Pi^+ := \{ w\in \mathbb{C}: \textnormal{Im}\, w >0\}$ the upper half-plane, $\mathbb{D}:=\{z\in \mathbb{C} : \vert z\vert <1\}$ the unit disc and $\mathbb{T}:= \{z\in \mathbb{C} : \vert z\vert =1\} $ the unit circle.
\subsection*{The Hardy and weighted Bergman spaces of the upper half-plane}
The \textit{Hardy space} $H^2 (\Pi^{+})$ of the upper half-plane consists of the analytic functions $F:\Pi^{+} \longrightarrow \mathbb{C} $ such that
$$ \Vert F\Vert_{H^2 (\Pi^{+})} = \sup_{y>0} \Big(\int_{-\infty}^{\infty} \vert F(x+iy)\vert^2 \, dx \Big)^{1/2} < \infty .$$
The space $H^2 (\Pi^{+})$ is isometrically embedded into $ L^2 (\mathbb{R})$ via the mapping $F \longmapsto F^{*}$, where $F^{*}(x)= \lim_{y \longrightarrow 0^{+}} F(x+iy)$ (for more information on Hardy spaces of the half-plane, see \cite[Chapter 11]{Du}, for instance).
For $\alpha > -1$, the \textit{weighted Bergman space} $\mathcal{A}_{\alpha}^2 (\Pi^{+})$ of the upper half-plane consists of the functions $F$ analytic on $\Pi^{+}$ satisfying
$$ \Vert F\Vert_{\mathcal{A}_{\alpha}^2 (\Pi^{+})} = \Big( \int_{\Pi^{+}} \vert F(x+iy)\vert^2 \, y^{\alpha} \, dx \, dy \Big)^{1/2} < \infty .$$
The Hardy space $H^2 (\Pi^{+})$ can often be formally interpreted as the ``limit case'' of the weighted Bergman spaces as $\alpha \longrightarrow -1$, that is $ H^2 (\Pi^{+}) = \mathcal{A}_{-1}^2 (\Pi^{+})$. In the sequel, we will use this formal convention. The spaces $\mathcal{A}_{\alpha}^2 (\Pi^{+})$, for $\alpha \geq -1$, are Hilbert spaces and we use the notation $\langle \cdot , \cdot \rangle_{\mathcal{A}_{\alpha}^2 (\Pi^{+}) }$ for the inner product.
\vspace{5pt}
Often we will find it useful to change the perspective from $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ to the corresponding space in the unit disc, namely to the classical Hardy space
$$H^2 (\mathbb{D})= \big\{f: \mathbb{D} \longrightarrow \mathbb{C} \textnormal{ analytic}, \Vert f \Vert^2= \sup_{0 <r <1 }\int_{0}^{2\pi} \vert f(re^{i\theta})\vert^2 \, \frac{d\theta}{2\pi} <\infty \big\},$$
or the weighted Bergman spaces defined for all $\alpha > -1$ by (for $z=u+iv$)
$$\mathcal{A}_{\alpha}^2 (\mathbb{D})= \{f: \mathbb{D} \longrightarrow \mathbb{C} \textnormal{ analytic}, \Vert f \Vert_{\alpha}^2= \int_{\mathbb{D}} \vert f(z)\vert^2 (1-\vert z\vert^2)^{\alpha}\, du \, dv <\infty \big\} .$$
We will write also here $H^2 (\mathbb{D}) = \mathcal{A}_{-1}^2 (\mathbb{D})$ for convenience. Recall that the function $h$, where
\begin{equation}\label{hoo}
h(z)=i \frac{1+z}{1-z},
\end{equation}
is a conformal map from $\mathbb{D}$ onto $\Pi^+$ with the inverse $h^{-1}(w)=\frac{w-i}{w+i}$. For any $\alpha \geq -1$ there is an isometric isomorphism (see e.g. \cite[pp. 128-131]{Ho} for the Hardy space and \cite{DGM} for the weighted Bergman spaces)
\begin{equation}\label{jii}
J: \mathcal{A}_{\alpha}^2 (\mathbb{D}) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+), \quad (Jf)(w)= f \Big( \frac{w-i}{w+i}\Big)\frac{c_{\alpha}}{(w+i)^{\alpha +2}},
\end{equation}
where
$$c_{\alpha}=\left\{ \begin{array}{ll}
1/\sqrt{\pi}, & \textnormal{ when } \alpha =-1, \\
2^{\alpha +1}, & \textnormal{ when } \alpha > -1 .
\end{array} \right.
$$
The inverse of $J$ is given by
$$ J^{-1}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 (\mathbb{D}) , \quad (J^{-1}F)(z)= F \Big( i\frac{1+z}{1-z}\Big)\frac{d_{\alpha}}{(1-z)^{\alpha +2}},
$$
where
$$d_{\alpha}=\left\{ \begin{array}{ll}
2\sqrt{\pi}i, & \textnormal{ when } \alpha =-1, \\
2i, & \textnormal{ when } \alpha > -1 .
\end{array} \right.
$$
\vspace{5pt}
Recall that the \textit{reproducing kernels} of $\mathcal{A}_{\alpha}^2 ( \Pi^+)$, for $\alpha \geq -1$, are functions $K_{w_0}^{\alpha} \in \mathcal{A}_{\alpha}^2 ( \Pi^+)$ satisfying
$$ \langle F, K_{w_0}^{\alpha}\rangle_{\mathcal{A}_{\alpha}^2 ( \Pi^+)} = F(w_0) \textnormal{ for any } F \in \mathcal{A}_{\alpha}^2 ( \Pi^+) \textnormal{ and } w_0 \in \Pi^+ .$$
We will next compute the explicit form of $ K_{w_0}^{\alpha}$ since we could not find a suitable reference for the general case.
\begin{lemma}
The reproducing kernels $ K_{w_0}^{\alpha} \in \mathcal{A}_{\alpha}^2 ( \Pi^+) $ for $\alpha \geq -1$ are of the form
\begin{equation}\label{reprkernela} K_{w_0}^{\alpha} (w)= \frac{k_{\alpha}}{ (w-\overline{w_0})^{\alpha +2}},
\end{equation}
where
$$k_{\alpha}=\left\{ \begin{array}{ll}
i/2\pi, & \textnormal{ when } \alpha =-1, \\
i(\alpha +1) 2^{\alpha }, & \textnormal{ when } \alpha > -1.
\end{array} \right.
$$
\end{lemma}
\begin{proof}
Recall first that for $\alpha \geq -1$, the reproducing kernels $g_{z_0}^{\alpha}$ of $\mathcal{A}_{\alpha}^2 (\mathbb{D})$ at a point $z_0\in \mathbb{D}$ are of the form (see \cite[Chapter 2]{CM}, for instance)
$$g_{z_0}^{\alpha} (z)= \frac{ \nu_{\alpha}}{(1-\overline{z_0}z)^{\alpha +2}}, \textnormal{ where } \nu_{-1}=1 \textnormal{ and }\nu_{\alpha }= (\alpha +1) \textnormal{ otherwise} .$$
But for each $F\in \mathcal{A}_{\alpha}^2 ( \Pi^+)$ and $ w_0 = i \frac{1+z_0}{1-z_0}\in \Pi^+$ it holds that
$$F(w_0)= \frac{(1-z_0)^{\alpha +2}}{d_{\alpha}} (J^{-1}F)(z_0) = \frac{(1-z_0)^{\alpha +2}}{d_{\alpha}} \langle J^{-1}F, g_{z_0}^{\alpha}\rangle_{\mathcal{A}_{\alpha}^2 (\mathbb{D})} =\frac{(1-z_0)^{\alpha +2}}{d_{\alpha}} \langle F, Jg_{z_0}^{\alpha}\rangle_{\mathcal{A}_{\alpha}^2 ( \Pi^+)} ,$$
where $J$ is the isometric isomorphism in \eqref{jii}. On the other hand, $F(w_0)=\langle F, K_{w_0}^{\alpha} \rangle_{ \mathcal{A}_{\alpha}^2 ( \Pi^+)} $, so that
\begin{equation*}
\begin{split}
K_{w_0}^{\alpha} (w)& =\frac{(1-\overline{z_0})^{\alpha +2}}{\overline{d_{\alpha}}} \big(Jg_{z_0}^{\alpha}\big)(w) = \frac{(1-\overline{z_0})^{\alpha +2}}{\overline{d_{\alpha}}}\frac{\nu_{\alpha}}{\Big(1-\overline{z_0}\big( \frac{w-i}{w+i}\big)\Big)^{\alpha +2}} \frac{c_{\alpha}}{(w+i)^{\alpha +2}} \\
&= \frac{\nu_{\alpha} c_{\alpha}}{\overline{d_{\alpha}}} \frac{1}{ \Big(w- \overline{ i \frac{1+z_0}{1-z_0}}\Big)^{\alpha +2}} = \frac{k_{\alpha}}{ (w-\overline{w_0})^{\alpha +2}}.
\end{split}
\end{equation*}
\end{proof}
\vspace{5pt}
For any $\beta \geq 0$, let us denote by $ L_{\beta}^2$ the space consisting of the measurable functions $f:\mathbb{R}_+ \longrightarrow \mathbb{C} $ with finite norm
$$ \Vert f\Vert_{L_{\beta}^2} = \left\{ \begin{array}{ll}
\Big( 2\pi\int_{0}^{\infty} \vert f(t)\vert^2 \, dt \Big)^{1/2} , & \textnormal{ when } \beta = 0,\\
\Big(\frac{2\pi \Gamma (\beta)}{2^{\beta}}\int_{0}^{\infty} \vert f(t)\vert^2 t^{-\beta} \, dt \Big)^{1/2} , & \textnormal{ when } \beta > 0.
\end{array} \right.
$$
For future reference we will write explicitly the classical Paley-Wiener theorem (\cite{PW}, see also \cite[Thm. 11.9]{Du} or \cite[Thm. 19.2]{R}) and its generalization to the weighted Bergman spaces (a detailed proof for all $\alpha >-1$ can be found in \cite[Thm. 1]{DGM}) by using the above notation: For any $\alpha \geq -1$, the space $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ is isometrically isomorphic to $L_{\alpha +1}^2$ under the Fourier transform $\mathcal{F}$. Indeed, $F \in \mathcal{A}_{\alpha}^2 ( \Pi^+)$ if and only if there exists a function $f\in L_{\alpha +1}^2$ such that
$$ F(w)= (\mathcal{F}^{-1} f)(w) = \int_0^{\infty} f(t) e^{iwt} \, dt , \quad w\in \Pi^+ .$$
Moreover,
\begin{equation}\label{pwnormeq}
\Vert F\Vert_{\mathcal{A}_{\alpha}^2 ( \Pi^+)}^2 = \Vert f\Vert_{L_{\alpha +1}^2}^2 = b_{\alpha} \int_0^{\infty} \vert f(t) \vert^2 t^{-(\alpha +1)} \, dt,
\end{equation}
where
$$b_{\alpha}=\left\{ \begin{array}{ll}
2\pi, & \textnormal{ when } \alpha =-1, \\
2^{-\alpha}\pi \Gamma (\alpha +1), & \textnormal{ when } \alpha > -1.
\end{array} \right.
$$
\subsection*{Composition operators $C_{\tau}$ on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$}
Let $\tau : \Pi^+\longrightarrow \Pi^+$ be an analytic map. Denote by $C_{\tau}$ the composition map induced by $\tau$,
$$C_{\tau} F = F \circ \tau , \quad F:\Pi^+ \longrightarrow\mathbb{C} \textnormal{ analytic} .$$
(The definition can be generalized by replacing $\Pi^+$ by any open subset of $\mathbb{C}$.)
\vspace{5pt}
In what follows we will consider the composition operator $C_{\tau}$ acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$, where $\tau$ is a linear fractional self-map of $\Pi^+ $ such that $\tau (w)\not\equiv w$. Recall that, in general, the linear fractional transformations (abbreviated by LFT) $\psi$ are meromorphic bijections $ \overline{\mathbb{C}} \longrightarrow \overline{\mathbb{C}} $ and they can be written in the form
$$\psi (z)= \frac{az+b}{cz+d}, \quad \textnormal{where } a,b,c,d \in \mathbb{C}, \, ad-bc\neq 0 .$$
The non-identity LFTs have exactly two \textit{fixed points}, i.e. points $z \in \overline{\mathbb{C}}$ such that $\psi (z)=z$, counting multiplicities. According to their behaviour at the fixed points, the LFTs can be classified into \textit{parabolic, hyperbolic, elliptic} or \textit{loxodromic} maps. For more information on LFTs, see \cite[Chapter 0]{Sh}, for instance.
\vspace{5pt}
A necessary and sufficient condition for the boundedness of $C_{\tau}$ (for any analytic $\tau : \Pi^+ \longrightarrow \Pi^+$) on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $, for $\alpha \geq -1$, is that $\tau$ has a finite positive \textit{angular derivative} at infinity (see \cite[Thm. 15]{Ma3} or \cite[Thm. 3.1]{EJ} for the Hardy space and \cite[Thm. 3.4]{EW} for the weighted Bergman spaces). That is, $\tau$ must fix $\infty$ and the non-tangential limit is
$$\lim_{w \longrightarrow\infty} \frac{w}{\tau (w)} =: \tau' (\infty) \in (0,\infty).$$
It is a standard exercise to show that the linear fractional self-maps of $\Pi^+$ that fix $\infty$ are precisely of the form
\begin{equation}\label{tau} \tau (w)= \mu w + w_0, \textnormal{ where } \mu >0 \textnormal{ and Im}\, w_0\geq 0.
\end{equation}
Depending on the value of $\mu$ in Equation \eqref{tau}, we have two different types of mappings that induce bounded composition operators on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$:
\begin{itemize}
\item[1.] $\tau$ is \textit{parabolic} when $\mu =1$ (and necessarily $w_0 \neq 0$). For a parabolic mapping it always holds that the derivative at its only fixed point is $1$, that is $\tau' (\infty)=1$.
\item[2.] $\tau$ is \textit{hyperbolic} when $\mu \in (0,1)\cup (1,\infty)$ and $\textnormal{Im}\, w_0 \geq 0 $. In this case, by definition $ \tau' (\infty)=1/\mu$. It holds that if $p$ and $q$ are, respectively, the attractive and repulsive fixed points of a hyperbolic mapping $\tau$, then
\begin{equation}\label{attrepder}
\tau' (p)= \frac{1}{\tau' (q)}.
\end{equation}
\end{itemize}
Note that in both cases $\tau$ is an automorphism of $ \Pi^+$ if and only if $\textnormal{Im}\, w_0 =0$.
\vspace{5pt}
The condition that $\tau$ must fix $\infty$ for $C_{\tau}$ to be bounded on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$ excludes the elliptic, loxodromic and certain hyperbolic mappings from inducing composition operators on the half-plane. The counterparts of these mappings in the unit disc setting induce diagonal (the elliptic case) or compact (the loxodromic and hyperbolic non-automorphisms having fixed points in $\mathbb{D}$ and $\overline{\mathbb{C}}\setminus \overline{\mathbb{D}}$) composition operators, for instance, on $ \mathcal{A}_{\alpha}^2 ( \mathbb{D})$ for $\alpha \geq -1$.
\subsection*{Spectra of linear operators}
We recall some basic definitions and facts from general spectral theory that are used in the sequel. A suitable reference for the spectral theory of linear operators on Hilbert spaces is \cite{Mu}, for instance. Let $T$ be a bounded linear operator acting on a Hilbert space $\mathcal{H}$. The \textit{spectrum} of $T: \mathcal{H} \longrightarrow \mathcal{H}$ is defined by
$$\sigma \big(T; \mathcal{H}\big)= \{ \lambda \in \mathbb{C} : T-\lambda \textnormal{ is not invertible on } \mathcal{H}\}.$$
We will also use the notation $\sigma (T)$ for the spectrum if the space is obvious from the context. Recall that $\sigma (T)\subset \mathbb{C}$ is always a non-empty compact set. The \textit{point spectrum}, i.e. the set of eigenvalues of $T$, is denoted by $\sigma_p (T)$. The \textit{approximative point spectrum} of $T$ is defined by
$$ \sigma_{a} (T)=\{ \lambda \in \mathbb{C}: \exists \textnormal{ a sequence } (h_n) \subset \mathcal{H} \textnormal{ of unit vectors such that } \Vert (T-\lambda)h_n\Vert \longrightarrow 0\}.$$
Obviously, $\sigma_p (T)\subseteq \sigma_{a} (T)$. The following inclusions always hold
\begin{equation}\label{spincl}
\partial \sigma (T) \subseteq \sigma_{a} (T)\subseteq\sigma (T),
\end{equation}
where $\partial \sigma (T) $ denotes the boundary of $ \sigma (T)$.
If the operators $T: \mathcal{H}_1 \longrightarrow \mathcal{H}_1$ an $V: \mathcal{H}_2 \longrightarrow \mathcal{H}_2$ are \textit{similar}, i.e. there exists an invertible operator $U: \mathcal{H}_1 \longrightarrow \mathcal{H}_2$ such that $T= U^{-1}VU$, then their spectra, point spectra and approximative point spectra coincide. This is very useful for our purposes since if we can write $\tilde{\tau}=g\circ \tau \circ g^{-1}$, where $g$ is an automorphism of $\Pi^+$ fixing infinity (so that $C_g$ is bounded and invertible on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$ and $ C_g^{-1}=C_{g^{-1}}$), then
$$ C_{\tilde{\tau}}= C_g^{-1} C_{\tau} C_{g}.$$
This means that $C_{\tilde{\tau}}$ and $C_{\tau}$ both acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$ are similar and therefore $\sigma (C_{\tilde{\tau}} ) =\sigma (C_{\tau})$. Note that the operators $C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+)$ and $C_{\varphi}: \mathcal{A}_{\alpha}^2 (\mathbb{D}) \longrightarrow \mathcal{A}_{\alpha}^2 (\mathbb{D})$, where $\alpha \geq -1$ and $\varphi = g^{-1}\circ \tau \circ g $ for an analytic bijection $g: \mathbb{D} \longrightarrow \Pi^+$, are not similar in general. This is due to the fact that the composition operator $ C_{g}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 (\mathbb{D}) $ is not bounded below. However, $C_{\varphi}$ on $\mathcal{A}_{\alpha}^2 (\mathbb{D}) $ is similar to a weighted composition operator $MC_{\tau}$ on $ \mathcal{A}_{\alpha}^2 (\Pi^+)$, where the multiplication operator $M$ is in some cases simple enough to allow one to deduce the spectrum of $C_{\tau}$ from that of $C_{\varphi}$.
We are also able to determine the \textit{essential spectrum} of $C_{\tau}$ in all cases considered. The essential spectrum of $T$ is the closed set defined by
$$\sigma_{ess} (T; \mathcal{H}) = \{\lambda \in \mathbb{C}: T-\lambda \textnormal{ is not a Fredholm operator on }\mathcal{H} \} .$$
Recall that, by Atkinson's theorem, the Fredholmness of $T$ on $\mathcal{H}$ is equivalent to the invertibility of $ (T-\lambda ) + \mathcal{K}(\mathcal{H})$ in the Calkin algebra $ \mathcal{L}(\mathcal{H})/ \mathcal{K}(\mathcal{H})$. Here and in the sequel $ \mathcal{L}(\mathcal{H})$ denotes the set of bounded linear operators on $\mathcal{H}$ and $\mathcal{K}(\mathcal{H})$ the set of compact operators on $\mathcal{H}$.
We will use the fact that the eigenvalues of infinite multiplicity are contained in the essential spectrum and that the essential spectra of similar operators coincide.
We denote by $\rho (T)$ the spectral radius of $T$, which is defined by $\rho (T):= \max \{\vert\lambda\vert : \lambda \in \sigma (T) \}$ and satisfies the spectral radius formula
\begin{equation}\label{spradfor}
\rho (T)= \lim_{n\longrightarrow \infty} \Vert T^n\Vert^{1/n}.
\end{equation}
\vspace{5pt}
Concerning the bounded composition operators acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$, the spectral radius is obtained from the angular derivative of the inducing map: The results by Elliott, Jury and Wynn (see \cite[Thm. 3.4]{EJ} and \cite[Thm. 3.5]{EW}) guarantee that whenever $C_{\tau}$ is bounded on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$, the spectral radius as well as the operator norm and the essential operator norm of $C_{\tau}$ satisfy
\begin{equation}\label{sprad}
\rho (C_{\tau}) = \Vert C_{\tau}\Vert =\Vert C_{\tau}\Vert_{ess} =\big( \tau' (\infty)\big)^{(\alpha + 2)/2} .
\end{equation}
Here, the essential operator norm is defined for $T:\mathcal{H} \longrightarrow \mathcal{H}$ by $\Vert T \Vert_{ess} = \inf \{ \Vert T -K\Vert : K \textnormal{ is a compact operator on }\mathcal{H} \}$. Moreover, replacing $\rho (T)$ by the essential spectral radius $ \rho_{ess} (T):= \max \{\vert\lambda\vert : \lambda \in \sigma_{ess} (T) \}$ and $ \Vert \cdot\Vert$ by $ \Vert \cdot\Vert_{ess}$ in Equation \eqref{spradfor} we have the essential spectral radius formula.
\section{The spectrum of $C_{\tau}$ when $\tau$ is a parabolic self-map of $\Pi^+$}\label{secpara}
In this section we consider the composition operator $C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$, induced by a parabolic self-map of $\Pi^{+}$, that is
$$\tau (w)= w+w_0, \quad \textnormal{where } \textnormal{Im}\, w_0 \geq 0, \, w_0\neq 0.$$
Recall that the only fixed point of $\tau$ is $\infty$ and $\tau' (\infty)=1$. Moreover, $\tau$ is an automorphism of $ \Pi^+$ if and only if $\textnormal{Im}\, w_0 =0$ ($w_0\neq 0$), in which case $C_{\tau}$ is invertible.
\vspace{5pt}
Even though we cannot find a composition operator $ C_{g}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 (\mathbb{D}) $ (see Remark \ref{notsimi}) that would give a similarity between the operators $C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+) $ and $C_{\varphi} : \mathcal{A}_{\alpha}^2 (\mathbb{D}) \longrightarrow \mathcal{A}_{\alpha}^2 (\mathbb{D})$, where $\varphi = g^{-1}\circ \tau \circ g $, it turns out that their spectra are the same. (For the spectra of parabolic composition operators acting on $\mathcal{A}_{\alpha}^2 ( \mathbb{D})$ for $\alpha \geq -1$, see \cite[Thm. 7.5 and Cor. 7.42]{CM}.) The (essential) spectra of invertible ($\textnormal{Im}\, w_0 =0$) and self-adjoint (corresponding the case $\textnormal{Re}\, w_0 =0$ in the upper half-plane) parabolic composition operators on the Hardy space of the half-plane have been computed in \cite[Thm. 2.7 and Thm. 3.1]{Ma4}. Our proof is very different and works on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $, for all $\alpha >-1$, as well. We are also able to show that the essential spectrum coincides with the spectrum in all cases.
\vspace{5pt}
\begin{thm}\label{sppara}
Let $\tau : \Pi^+ \longrightarrow \Pi^+$ be a parabolic map of the form $\tau (w)= w+w_0$, where $\textnormal{Im}\, w_0 \geq 0$ and $w_0\neq 0$. Then the spectrum of $C_{\tau}$ acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $, for all $\alpha \geq -1$, is
\vspace{5pt}
\begin{itemize}
\item[i)] $ \sigma \big( C_{\tau}\big)= \mathbb{T}$, when $w_0\in \mathbb{R}$,
\vspace{5pt}
\item[ii)] $\sigma \big( C_{\tau}\big) = \{e^{iw_0t} : t\in [0,\infty ) \} \cup \{0\}$, when $w_0\in \Pi^+$.
\end{itemize}
\vspace{5pt}
\noindent Moreover, the essential spectrum of $C_{\tau}$ coincides with its spectrum in both cases.
\end{thm}
\begin{proof}
Denote
$$ \mathcal{S}:= \overline{ \{ e^{iw_0t} : t\in [0,\infty ) \} } = \left\{ \begin{array}{ll}
\mathbb{T}, & \textnormal{ when } w_0\in \mathbb{R}, \\
\{ e^{iw_0t} : t\in [0,\infty ) \} \cup \{ 0\}, & \textnormal{ when } \textnormal{Im}\, w_0 >0.
\end{array}
\right.
$$
We will split the proof into two parts. In \textit{Step 1} we will prove that $ \sigma \big( C_{\tau}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) \subseteq \mathcal{S}$ for all $\alpha \geq -1$ and in \textit{Step 2} we show that each point in $\mathcal{S} $ belongs to the essential spectrum of $ C_{\tau}$ on $ \mathcal{A}_{\alpha}^2 ( \Pi^+)$ for all $\alpha \geq -1$.
\vspace{5pt}
\textit{Step 1.} Note first that $ \big( \mathcal{F}^{-1}( C_{\tau} G) \big)(t) = e^{iw_0t} (\mathcal{F}^{-1}G) (t)$ for any $G\in \mathcal{A}_{\alpha}^2 ( \Pi^+) $ and $t>0$. Therefore, by the Paley-Wiener theorem (see p. \pageref{pwnormeq}), the Fourier transform gives us a similarity between the operators $C_{\tau} : \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+)$ and $\widehat{C}_{\tau} : L_{\alpha +1}^2 \longrightarrow L_{\alpha +1}^2$, where
\begin{equation}\label{ceetausim}\widehat{C}_{\tau}f(t)=e^{iw_0t}f(t).
\end{equation}
Since $\widehat{C}_{\tau} $ is a multiplication operator on $L_{\alpha +1}^2$, we can easily find points $\lambda \in \mathbb{C}$ for which $ \widehat{C}_{\tau} -\lambda$ \textit{is} invertible. Let $\lambda \in \mathbb{C}\setminus \mathcal{S}$. Since $\mathbb{C}\setminus \mathcal{S}$ is an open set, $\textnormal{dist}\, (\lambda, \mathcal{S})=:\delta >0$. For all $t> 0$, define a function $k$ by $k(t)=\frac{1}{e^{iw_0t}-\lambda}$. Since $ \vert k(t)\vert \leq 1/\delta$, the multiplication operator $M_k$, where $M_kf =kf$, is bounded on $ L_{\alpha +1}^2 $. Now,
$$M_k(\widehat{C}_{\tau} - \lambda)f =k (e^{iw_0t}-\lambda )f = f= (e^{iw_0t}-\lambda )kf = (\widehat{C}_{\tau} - \lambda )M_k f,$$
which means that $M_k$ is the inverse of $\widehat{C}_{\tau} - \lambda $ and therefore $\lambda \in \mathbb{C}\setminus \sigma \big( \widehat{C}_{\tau}\big)$. It follows that $ \sigma \big( C_{\tau}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) = \sigma \big( \widehat{C}_{\tau} ; L_{\alpha +1}^2\big)\subseteq \mathcal{S} $.
\vspace{5pt}
\textit{Step 2.} In order to show that $\mathcal{S}\subseteq \sigma_{ess} \big( C_{\tau}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big)$ for all $\alpha \geq -1$ we need the following fact (see \cite[Thm. XI.2.3]{Con}): \textit{Let $T:\mathcal{H} \longrightarrow \mathcal{H} $ be a bounded linear operator on a Hilbert space $\mathcal{H}$. If there exists an orthonormal sequence $(f_n)\subset \mathcal{H}$ such that}
$$
\Vert (T-\lambda) f_n\Vert \longrightarrow 0, \textnormal{ as } n\longrightarrow \infty ,
$$
\textit{then} $\lambda \in \sigma_{ess} \big(T; \mathcal{H}\big)$.
\vspace{5pt}
As in \textit{Step 1} we will use the similarity provided by the Paley-Wiener theorem. Now our aim is to find an orthogonal sequence of functions $(g_n)$ in $L_{\alpha +1}^2$ for any $\alpha \geq -1$ and $\lambda \in \mathcal{S}$,
$$\frac{\Vert (\widehat{C}_{\tau} -\lambda)g_n \Vert_{L_{\alpha +1}^2}}{\Vert g_n \Vert_{L_{\alpha +1}^2 }} \longrightarrow 0,\textnormal{ as } n \longrightarrow \infty.$$
Fix $t_0 \geq 0$ and $w_0\in \Pi^+ \cup \mathbb{R}$. Let $\lambda =e^{iw_0t_0}\in\mathcal{S}$ and $(\epsilon_n)\subset \mathbb{R}$ be any sequence satisfying $\epsilon_n >\epsilon_{n+1} $ and $\epsilon_n \longrightarrow 0$, as $n\longrightarrow \infty$ (for instance, set $\epsilon_n =1/n$). Consider the orthogonal sequence $(g_n)\subset L_{\alpha +1}^2$ of characteristic functions
$$ g_n= \chi_{\big[t_0 + \epsilon_{n+1} , t_0 + \epsilon_{n} ]} $$
for which it holds that
$$(\widehat{C}_{\tau} -\lambda ) g_n= (e^{iw_0t} -e^{iw_0t_0}) g_n .$$
Now we have that
\begin{equation*}
\frac{\Vert (\widehat{C}_{\tau} -\lambda ) g_n \Vert_{L_{\alpha +1}^2}^2}{\Vert g_n \Vert_{L_{\alpha +1}^2}^2}
\leq \sup_{t_0 +\epsilon_{n+1} \leq t \leq t_0 +\epsilon_{n}} \vert e^{iw_0t} - e^{iw_0t_0}\vert^2 \longrightarrow 0, \textnormal{ as } n \longrightarrow \infty.
\end{equation*}
It follows that $\mathcal{S} \subseteq \sigma_{ess} \big(\widehat{C}_{\tau}; L_{\alpha +1}^2\big)= \sigma_{ess} \big( C_{\tau}; A_{\alpha }^2 (\Pi^+)\big)$ for all $\alpha \geq -1$.
\end{proof}
\section{The spectrum of $C_{\tau}$ when $\tau$ is a hyperbolic self-map of $\Pi^+$}\label{hypsec}
We will first consider the case where $\tau$ is a hyperbolic automorphism of $\Pi^+$, that is, $\tau$ is of the form $\tau (w)=\mu w + w_0$, where $\mu\in (0,1)\cup (1,\infty)$ and $w_0\in \mathbb{R}$. Any mapping of this form can be written as
$$\tau = g \circ \tilde{\tau} \circ g^{-1},$$
where $ \tilde{\tau} (w)=\mu w$ and $g(w)= w+ \frac{w_0}{1-\mu}$. Since $g$ is an automorphism of $\Pi^+$ fixing $\infty$, it induces an invertible composition operator $C_g$ on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $ so that $ C_{\tau}$ and $C_{ \tilde{\tau}} $ acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$, for $\alpha \geq -1$, are similar. Therefore, in order to compute the spectrum for $C_{\tau}$ it is enough to compute the spectrum for the composition operator induced by the mapping
$$w \mapsto \mu w,\quad \mu\in (0,1)\cup (1,\infty).$$
We will first compute the spectrum and the essential spectrum in the case $\mu \in (0,1)$ (Theorem \ref{sphypauto}) and from this we will also get the case $\mu \in (1,\infty)$ (Corollary \ref{sphypautoinverse}). The spectrum and the essential spectrum of the invertible hyperbolic composition operators in the Hardy space of the (right) half-plane $\mathbb{C}_+$ have been computed by Matache in \cite[Thm. 2.12]{Ma4}. Matache's proof for the spectrum is based on the similarity between $C_{\tau}$ on $H^2 (\mathbb{C}_+)$ and a certain weighted composition operator on $H^2 (\mathbb{D})$. The spectrum is then obtained from the results in \cite[Thm. 4.8]{HLNS}. It it possible that this idea could also be used in the weighted Bergman spaces. Our approach is different and more direct revealing simultaneously also the essential spectrum on all of the spaces.
\vspace{5pt}
It is worth noting that the result in this case differs essentially from the unit disc setting, where, in the classical Hardy and the weighted Bergman spaces, the spectrum of an invertible hyperbolic composition operator is always an annulus centred at the origin (see \cite[Thm. 7.4]{CM}). In the half-plane setting we get that the spectrum is always a circle centred at the origin with radius depending on the space.
\vspace{5pt}
\begin{thm}\label{sphypauto}
Let $\tau$ be a hyperbolic automorphism of $\Pi^+$ of the form $\tau (w)=\mu w $, where $ \mu \in (0,1)$. Then the spectrum and the essential spectrum of $C_{\tau}$ acting on $ \mathcal{A}_{\alpha}^2 ( \Pi^+)$, for $\alpha \geq -1$, are
$$ \sigma \big( C_{\tau} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big) = \sigma_{ess} \big( C_{\tau} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big)= \big\{ \lambda \in \mathbb{C} : \vert \lambda \vert = \mu^{-(\alpha +2)/2}\big\} .$$
\end{thm}
\begin{proof}
The fixed points of $\tau$ are $0$ (attractive) and $\infty$. Moreover, by \eqref{attrepder}
$$\tau' (\infty) = 1/\tau' (0)=\mu^{-1},$$
so that the spectral radius of $ C_{\tau}$ acting on $ \mathcal{A}_{\alpha}^2 ( \Pi^+)$ is $\mu^{-(\alpha +2)/2} $ (see \eqref{sprad}). We will split the proof into two parts: in \textit{Step 1} we show that $\sigma \big( C_{\tau} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big) \subseteq \big\{ \lambda \in \mathbb{C} : \vert \lambda \vert = \mu^{-(\alpha +2)/2}\big\} =: \mathcal{Y}_{\alpha}$ and in \textit{Step 2} we prove that each point in $\mathcal{Y}_{\alpha} $ belongs to the essential spectrum of $C_{\tau} $ on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$.
\vspace{5pt}
\textit{Step 1.} Let $F \in \mathcal{A}_{\alpha}^2 ( \Pi^+)$, $\alpha >-1$, and $\lambda \in \mathbb{C}$. Now, by the change of variable $ w \longrightarrow \mu^{-1}w$, we get that (for $w=x+iy$)
\begin{equation*}
\begin{split}
\Vert C_{\tau} F\Vert_{\mathcal{A}_{\alpha}^2 ( \Pi^+)}^2 &= \int_{\Pi^+} \vert F(\mu w)\vert^2 y^{\alpha} \, dx \, dy = \int_{\Pi^+} \vert F( w)\vert^2 y^{\alpha} \mu^{-\alpha -2}\, dx \, dy \\
&= \mu^{-\alpha -2} \Vert F\Vert_{\mathcal{A}_{\alpha}^2 ( \Pi^+)}^2.
\end{split}
\end{equation*}
For $H^2 (\Pi^+)$ ($\alpha =-1$), we get similarly that
$$\Vert C_{\tau} F\Vert_{H^2 (\Pi^+)}^2 =\sup_{\mu y>0} \int_{-\infty}^{\infty} \vert F(\mu x+i\mu y)\vert^2 \, dx =\sup_{y>0} \int_{-\infty}^{\infty} \vert F(x+iy)\vert^2 \mu^{-1} \, dx = \mu^{-1} \Vert F\Vert_{H^2 (\Pi^+)}^2.$$
Since $\Vert C_{\tau} F\Vert_{\mathcal{A}_{\alpha}^2 ( \Pi^+)} =\mu^{-(\alpha +2)/2} \Vert F\Vert_{\mathcal{A}_{\alpha}^2 ( \Pi^+)}$ for all $\alpha \geq -1$ and the spectrum of an isometric operator is contained in the unit circle, we are able to deduce that
$$\sigma \big( C_{\tau} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big) \subseteq \mathcal{Y}_{\alpha}.$$
\vspace{5pt}
\textit{Step 2.} Here we use a similar technique as in \textit{Step 2} on p. \pageref{sppara}. Since $ \mathcal{F}^{-1}\big(G (\mu x)\big)= \frac{1}{\mu}( \mathcal{F}^{-1}G )(x/\mu) $ for any $G\in \mathcal{A}_{\alpha}^2 ( \Pi^+)$ and $x>0$, we have, by the (versions of) Paley-Wiener theorem, that the operator $ \widehat{C}_{\tau} : L_{\alpha +1}^2\longrightarrow L_{\alpha +1}^2$ defined by
$$ \widehat{C}_{\tau} f (x) = \frac{1}{\mu} f(x/\mu) , \quad x>0,$$
is similar to $C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+)\longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+)$, for all $\alpha \geq -1$. For each $\lambda\in \mathcal{Y}_{\alpha}$, we find an orthogonal sequence $(g_n)\subset L_{\alpha +1}^2$ such that
$$ \frac{\Vert ( \widehat{C}_{\tau} - \lambda )g_n \Vert_{L_{\alpha +1}^2}}{\Vert g_n\Vert_{L_{\alpha +1}^2}} \longrightarrow 0, \textnormal{ as }n\longrightarrow\infty .$$
Fix $\lambda_{\alpha} = \mu^{-(\alpha +2)/2}e^{it}$, where $t\in \mathbb{R}$, and define (cf. \cite[Thm. 3.2]{Hi} in the Dirichlet space case)
$$g(x)= x^{\alpha / 2} \mu^{-(\alpha +2)/2 }e^{-it \log_{\mu} x} , \quad x>0.$$
By inspection,
\begin{equation}\label{geemyy}
\frac{1}{\mu }g(x/ \mu) = \mu^{-(\alpha +2)/2}e^{it} g(x) = \lambda_{\alpha} g(x).
\end{equation}
Let $a_0=1$ and $a_n = e^{-n}a_{n-1}=e^{-1-2- \cdots -n}$ for all $n\geq 1$. Note that $ a_n < a_{n-1}$ for all $n\in \mathbb{N}$ and that $a_n \longrightarrow 0$, as $n\longrightarrow\infty$. Define a sequence $(g_n)\subset L_{\alpha +1}^2$ by setting
$$ g_n = g\cdot \chi_{[a_n, a_{n-1}]}. $$
Since the sequence $(a_n)$ is decreasing, $ \langle g_n ,g_m \rangle_{L_{\alpha +1}^2} =0$ for all $n \neq m$. Moreover,
\begin{equation}\label{normg}
\begin{split}
\Vert g_n \Vert_{L_{\alpha +1}^2}^2 &= b_{\alpha}\int_{a_n}^{a_{n-1}} \big\vert x^{\alpha / 2} \mu^{-(\alpha +2)/2 }e^{-it \log_{\mu} x} \big\vert^2 x^{-(\alpha +1)}\, dx = b_{\alpha} \mu^{-(\alpha +2)} \int_{a_n}^{a_{n-1}} x^{-1}\, dx \\
&= b_{\alpha} \mu^{-(\alpha +2)} n.
\end{split}
\end{equation}
On the other hand, by using \eqref{geemyy}, we get (for $n$ big enough)
\begin{equation}\label{normctau}
\begin{split}
\Vert ( \widehat{C}_{\tau} - \lambda_{\alpha})g_n \Vert_{L_{\alpha +1}^2}^2 &= b_{\alpha} \int_0^{\infty} \big\vert \lambda_{\alpha} g(x)\chi_{[\mu a_n ,\mu a_{n-1} ]} -\lambda_{\alpha} g(x)\chi_{ [a_n ,a_{n-1} ]} \big\vert^2 x^{-(\alpha +1)} \, dx \\
&\stackrel{(*)}{=} b_{\alpha} \vert \lambda_{\alpha} \vert^2 \Big( \int_{\mu a_n}^{a_n} \vert g(x)\vert^2 x^{-(\alpha +1)}\, dx + \int_{\mu a_{n-1}}^{a_{n-1}} \vert g(x)\vert^2 x^{-(\alpha +1)}\, dx\Big) \\
&= b_{\alpha}\vert \lambda_{\alpha} \vert^2 \mu^{-(\alpha +2)} \Big( \int_{\mu a_n}^{a_n} x^{-1}\, dx + \int_{\mu a_{n-1}}^{a_{n-1}} x^{-1}\, dx\Big) \\
&= -2b_{\alpha}\vert \lambda_{\alpha} \vert^2 \mu^{-(\alpha +2)} \ln \mu .
\end{split}
\end{equation}
In $(*)$ we have used the fact that $a_n <\mu a_{n-1}$ for all $n > \ln (1/ \mu )$ and so the function $ g(x) \chi_{[\mu a_n ,\mu a_{n-1}]}(x) - g(x)\chi_{ [a_n ,a_{n-1}]}$ vanishes on $[0,\mu a_n]\cup [ a_n, \mu a_{n-1}]$.
By Equations \eqref{normg} and \eqref{normctau} we have that
$$ \frac{\Vert ( \widehat{C}_{\tau} -\lambda_{\alpha} ) g_n\Vert_{L_{\alpha +1}^2}^2}{\Vert g_n \Vert_{L_{\alpha +1}^2}^2} = \frac{ -2\vert \lambda_{\alpha} \vert^2 \ln \mu}{n} \longrightarrow 0, \textnormal{ as } n\longrightarrow \infty .$$
Now \cite[Thm. XI.2.3]{Con} guarantees that $\lambda_{\alpha} \in \sigma_{ess} \big( \widehat{C}_{\tau} ; L_{\alpha +1}^2\big)$. Furthermore, by similarity, this yields
$$\mathcal{Y}_{\alpha}:= \{\lambda \in \mathbb{C} : \vert \lambda\vert = \mu^{-(\alpha +2)/2}\}\subseteq \sigma_{ess} \big( C_{\tau}; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big).$$
\end{proof}
From the (essential) spectrum of $C_{\tau}$ we will obtain the (essential) spectrum of the inverse of $C_{\tau}$ by using general spectral theory.
\begin{cor}\label{sphypautoinverse}
Let $\tau^{-1} (w)= \mu^{-1} w$, where $\mu\in (0,1)$. The spectrum of $C_{\tau^{-1}}$ acting on $ \mathcal{A}_{\alpha}^2 ( \Pi^+)$, $\alpha \geq -1$, is
$$ \sigma \big( C_{\tau^{-1}} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big) =\sigma_{ess} \big( C_{\tau^{-1}} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big)= \big\{ \lambda \in \mathbb{C} : \vert \lambda \vert = \mu^{(\alpha +2)/2}\big\} .$$
\end{cor}
\begin{proof}
Since $C_{\tau^{-1}}= C_{\tau}^{-1}$, we have that $\lambda \in \sigma \big( C_{\tau^{-1}} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big)$ if and only if $ \lambda^{-1}\in \sigma \big( C_{\tau} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big)$, where $\tau (w)=\mu w$. The same is true also for the essential spectrum and so the result follows from Theorem \ref{sphypauto}.
\end{proof}
\vspace{5pt}
Let us consider then the cases where $\tau$ is a hyperbolic self-map of $\Pi^+$ and $\tau (\Pi^+)\subsetneq \Pi^+$. Let $\mu \in (0,1)$ so that $ \mu^{-1}\in (1,\infty)$, and consider first the mapping
$$\tau (w)= \frac{w}{\mu} +w_0 \textnormal{, where } \textnormal{Im}\, w_0 >0.$$ The attractive fixed point of $\tau$ is $\infty$ and the repulsive fixed point will necessarily have negative imaginary part (just solve $\tau (w)=w$ to see this). Moreover, we can write
$$ \tau_1 = g_1^{-1}\circ \tau \circ g_1, $$
where
$$\tau_1 (w)= \frac{w}{\mu} + \frac{i(1-\mu )}{\mu} $$
and $g_1(w) = \mu (1-\mu )^{-1} \big( (\textnormal{Im}\, w_0 )w - \textnormal{Re}\, w_0 \big)$. Since $\frac{\mu \textnormal{Im}\, w_0}{1-\mu} >0$ and $\frac{ \mu \textnormal{Re}\, w_0}{\mu -1}\in \mathbb{R}$, $g_1$ is an automorphism of $ \Pi^+$ fixing $\infty$ and, therefore, induces a bounded invertible composition operator $ C_{g_1}$ on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for all $\alpha \geq -1$. It follows that $ C_{\tau}$ and $C_{\tau_1}$ acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha \geq -1$ are similar operators and in order to compute the spectrum for $C_{\tau}$ it is enough to compute the spectrum for $C_{\tau_1}$. In the proof below, recall our convention $\mathcal{A}_{-1}^2 (\mathbb{D})$ for $H^2 (\mathbb{D})$.
\vspace{5pt}
\begin{thm}\label{hyptauyks}
Let $\mu \in (0,1)$ and $\tau_1$ be a hyperbolic self-map of $\Pi^+$ of the form $\tau_1 (w)= \frac{w}{\mu} + \frac{i(1-\mu )}{\mu} $. Then, for all $\alpha \geq -1$, the spectrum and the essential spectrum are
$$ \sigma \big(C_{\tau_1} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big) = \sigma_{ess} \big(C_{\tau_1} ; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big) =\{\lambda \in \mathbb{C} : \vert \lambda\vert \leq \mu^{(\alpha +2)/2}\}.$$
\end{thm}
\begin{proof}
The fixed points of $\tau_1$ are $-i$ and $\infty$ (the attractive one). By \eqref{sprad} and \eqref{attrepder}, for $\alpha \geq -1$,
$$\rho \big( C_{\tau_1}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) = \big(\tau_1' (\infty)\big)^{(\alpha +2)/ 2}=\big(\tau_1' (-i)\big)^{-(\alpha +2)/ 2} = \mu^{(\alpha +2)/ 2},$$
so that
$$\sigma \big(C_{\tau_1}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big)\subseteq \{\lambda \in \mathbb{C}: \vert \lambda\vert \leq \mu^{(\alpha +2)/2}\}.$$
We will show that each point in the open disc $\{\lambda \in \mathbb{C} : \vert \lambda\vert < \mu^{(\alpha +2)/2}\}$ is an eigenvalue of infinite multiplicity for $C_{\tau_1}$ on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $ for $\alpha \geq -1 $. Since the eigenvalues of infinite multiplicity belong to the essential spectrum, which is always a closed set, the claim will then follow.
We will take advantage of the known spectrum in the corresponding case of the unit disc. Namely, by Hurst \cite[Theorem 8]{Hu}, we know that the spectrum of $C_{\varphi_1}$, where $\varphi_1 (z)=\mu z +1 -\mu$, acting on $\mathcal{A}_{\alpha}^2 (\mathbb{D})$ for $\alpha \geq -1$, is the closed disc with radius $\varphi_1' (1)^{-(\alpha +2)/2} = \mu^{-(\alpha +2)/2}$. Moreover, each point $\lambda $ satisfying $ \vert \lambda \vert < \mu^{-(\alpha +2)/2}$ is an eigenvalue of infinite multiplicity for $C_{\varphi_1} : \mathcal{A}_{\alpha}^2 (\mathbb{D})\longrightarrow \mathcal{A}_{\alpha}^2 (\mathbb{D})$.
By a simple computation, we see that $\varphi_1 = h^{-1} \circ \tau_1 \circ h$, where $h: \mathbb{D} \longrightarrow \Pi^+$ is the bijection defined in \eqref{hoo}. Using the isometric isomorphism $J: \mathcal{A}_{\alpha}^2 (\mathbb{D})\longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+)$ (see \eqref{jii}) we get that
\begin{equation*}
\begin{split}
J(f\circ \varphi_1)(w)& = (f\circ \varphi_1 \circ h^{-1})(w)\frac{2^{\alpha +1}}{(w+i)^{\alpha +2}} = (f\circ h^{-1} \circ \tau_1 )(w)\frac{2^{\alpha +1}}{(\tau_1 (w)+i)^{\alpha +2}}\Big(\frac{\tau_1 (w)+i}{w+i}\Big)^{\alpha +2} \\
&= \Big(\frac{\tau_1 (w)+i}{w+i}\Big)^{\alpha +2} \big(C_{\tau_1} (Jf)\big) (w) .
\end{split}
\end{equation*}
On the other hand, observe that
$$ \frac{\tau_1 (w)+i}{w+i} = \frac{\frac{w}{\mu} + \frac{i(1-\mu )}{\mu}+i}{w+i}= \mu^{-1}\frac{w+ i -i\mu +i\mu}{w+i}= \mu^{-1},$$
so that
$$\big( J(C_{ \varphi_1}f)\big) (w)= \mu^{-(\alpha +2)} \big(C_{\tau_1} (Jf)\big) (w).$$
That is, $C_{\varphi_1}$ on $ \mathcal{A}_{\alpha}^2 (\mathbb{D})$ is similar to the operator $\mu^{-(\alpha +2)}C_{\tau_1}$ acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $.
\vspace{5pt}
Recall that, by \cite[Theorem 8]{Hu}, for any $\lambda$ with $ \vert\lambda\vert < \mu^{-(\alpha +2)/2} $ there exists $f_{\lambda}\in \mathcal{A}_{\alpha}^2 (\mathbb{D})$ (in fact, an $\infty$-dimensional subspace of them) such that $C_{\varphi_1} f_{\lambda} = \lambda f_{\lambda}$. But now, writing $F_{\lambda} =Jf_{\lambda}$, we have that
$$ \mu^{-(\alpha +2)} C_{\tau_1}F_{\lambda}= \lambda F_{\lambda} $$
which is equivalent to
$$ C_{\tau_1}F_{\lambda} = \mu^{\alpha +2} \lambda F_{\lambda} .$$
Since $ \vert\lambda\vert < \mu^{-(\alpha +2)/2} $, we obtain that $C_{\tau_1}$ has an eigenfunction for each point in the open disc of radius $\mu^{(\alpha +2)/ 2} $. Moreover, the eigenvalues are of infinite multiplicity since $J$ preserves the linear independence of functions.
\vspace{5pt}
In fact, it is possible to compute the eigenfunctions in $ \mathcal{A}_{\alpha}^2 ( \Pi^+)$ explicitly from the corresponding functions in the unit disc setting: In the unit disc setting (see the proof of Theorem $8$ in \cite{Hu}), for any $p > -(\alpha +2)/2 $ and $q\in \mathbb{R}$, the function $f_{p,q} \in \mathcal{A}_{\alpha}^2 (\mathbb{D})$,
$$f_{p,q} (z)= \exp \{ (p+iq) \log (1-z)\} = (1-z)^{p+iq},$$
where $\log (\cdot )$ is the principal branch of the natural logarithm, is an eigenvector for $C_{\varphi_1}$ corresponding to the eigenvalue $\lambda = \mu^{p+iq}$. Using the isometric isomorphism we get that
$$(Jf_{p,q})(w) = \Big(1-\frac{w-i}{w+i}\Big)^{p+iq} \frac{c_{\alpha}}{ (w+i)^{\alpha +2}}= \frac{(2i)^{p+iq}c_{\alpha}}{(w+i)^{\alpha +2 + p+iq }}
$$
is an eigenvector for $C_{\tau_1}$ acting on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ corresponding to the eigenvalue $\mu^{\alpha + 2 + p+iq}$.
\end{proof}
\vspace{5pt}
Consider next the mapping $\tau (w)= \mu w +w_0$, where $\mu \in (0,1)$ and $\textnormal{Im}\, w_0 >0$. Here, the repulsive fixed point of $\tau$ is $\infty$ and the attractive fixed point belongs to $\Pi^+$. In this case, to compute the spectrum for $C_{\tau}$, it is enough to consider the mapping $ \tau_2 : \Pi^+ \longrightarrow \Pi^+$ of the form
$$ \tau_2 (w)= \mu w+ i (1-\mu),$$
having $i$ as its attractive fixed point. This is due to the fact that we can write
$$\tau_2 = g_2^{-1}\circ \tau \circ g_2 , $$
where $g_2(w)= (1-\mu)^{-1} \big((\textnormal{Im}\, w_0)w + \textnormal{Re}\, w_0 \big) $ is an automorphism of $\Pi^+$ fixing $\infty$, so that $C_{g_2}$ is bounded and invertible on $ \mathcal{A}_{\alpha}^2 ( \Pi^+)$ for all $\alpha \geq -1$. From this it follows that
$$C_{\tau_2} = C_{g_2}C_{\tau}C_{g_2}^{-1}.$$
Therefore, for any $\alpha \geq -1$, we have that $\sigma \big(C_{\tau_2} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) =\sigma \big(C_{\tau} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big)$ and that the essential spectra of $C_{\tau_2}$ and $C_{\tau}$ on $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ coincide. The spectrum and the essential spectrum for $C_{\tau_2} $ are determined in the following result.
\vspace{5pt}
\begin{thm}\label{hyptaukaks}
Let $\mu \in (0,1)$ and $\tau_2$ be a hyperbolic self-map of $\Pi^+$ of the form $\tau_2 (w)=\mu w+ i (1-\mu) $. Then, for all $\alpha \geq -1$, the spectrum and the essential spectrum of $C_{\tau_2}$ on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $ are
$$ \sigma \big(C_{\tau_2} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) = \sigma_{ess} \big(C_{\tau_2} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) =\{\lambda \in \mathbb{C} : \vert \lambda\vert \leq \mu^{-(\alpha +2)/2}\}.$$
\end{thm}
\begin{proof}
The mapping $\tau_2$ has $i$ (attractive) and $\infty$ as its fixed points. Again, by \eqref{sprad}, we know that the spectral radius of $C_{\tau_2}$ is $\big(\tau_2' (\infty)\big)^{(\alpha +2)/2}= \big(\tau_2' (i)\big)^{-(\alpha +2)/2} =\mu^{- (\alpha +2)/2}$ so that
$$\sigma \big(C_{\tau_2}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big)\subseteq \{\lambda \in \mathbb{C} : \vert \lambda\vert \leq \mu^{-(\alpha +2)/2}\}.$$
To prove the reverse containment (and that the essential spectrum coincides with the spectrum) it is enough to show that each point in the disc $\{\lambda \in \mathbb{C} : \vert \lambda\vert < \mu^{-(\alpha +2)/2}\}$ is an eigenvalue (of infinite multiplicity) for the adjoint $C_{\tau_2}^{*}$. From this it follows that
$$\{\lambda \in \mathbb{C} : \vert \lambda\vert < \mu^{-(\alpha +2)/2}\} \subseteq \sigma \big(C_{\tau_2}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) $$
since $\lambda \in \sigma (C_{\tau_2})$ if and only if $\overline{\lambda } \in \sigma ( C_{\tau_2}^{*}) $. Since the eigenvalues of infinite multiplicity belong to the essential spectrum and an operator is a Fredholm operator if and only if its adjoint is (see \cite[Thm. 16.4.]{Mu}, for instance), we will also get that
$$ \sigma_{ess} \big( C_{\tau_2} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) = \sigma_{ess}\big( C_{\tau_2}^{*} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) = \sigma \big( C_{\tau_2}^{*} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) = \sigma \big( C_{\tau_2} ; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big).
$$
Our proof relies on a useful observation, namely, that in this case, $C_{\tau_2}^{*} = \mu^{-(\alpha +2)} C_{\tau_1}$, where $\tau_1(w)= \mu^{-1} w + i\mu^{-1}(1-\mu )$ as in Theorem \ref{hyptauyks}. The claim follows then from Theorem \ref{hyptauyks}.
Recall first that the reproducing kernel $K_z^{\alpha}$ of $\mathcal{A}_{\alpha}^2 ( \Pi^+)$ at the point $z\in \Pi^{+}$ is of the form $K_z^{\alpha} (w)= \frac{k_{\alpha}}{(w-\bar{z})^{\alpha +2}}$, where the constant $k_{\alpha}$ depends only on $\alpha$. Let $F\in \mathcal{A}_{\alpha}^2 ( \Pi^+)$. Now
$$ C_{\tau_2}^{*} F (z) = \langle C_{\tau_2}^{*} F , K_z^{\alpha} \rangle_{\mathcal{A}_{\alpha}^2 ( \Pi^+)} = \langle F , C_{\tau_2} K_z^{\alpha} \rangle_{\mathcal{A}_{\alpha}^2 ( \Pi^+)} ,$$
where
\begin{equation*}
\begin{split}
C_{\tau_2} K_z^{\alpha} (w) &= \frac{k_{\alpha}}{(\tau_2 (w)-\bar{z})^{\alpha +2}}
= \frac{k_{\alpha}}{(\mu w + i(1-\mu )-\bar{z})^{\alpha +2}} \\
& = \mu^{-(\alpha +2)} \frac{k_{\alpha}}{\Big( w - \overline{\big( \frac{z}{\mu} +\frac{i(1-\mu )}{\mu}\big)}\Big)^{\alpha +2}} = \mu^{-(\alpha +2)} K_{\tau_1 (z)}^{\alpha} (w).
\end{split}
\end{equation*}
Here, $\tau_1 $ is as in Theorem \ref{hyptauyks}. It follows that for all $z\in \Pi^{+} $
\begin{equation*}
\begin{split}
C_{\tau_2}^{*} F (z) &= \langle F , C_{\tau_2} K_z^{\alpha} \rangle_{\mathcal{A}_{\alpha}^2 ( \Pi^+)} = \mu^{-(\alpha +2)} \langle F , K_{\tau_1 (z)}^{\alpha} \rangle_{\mathcal{A}_{\alpha}^2 ( \Pi^+)} \\
& = \mu^{-(\alpha +2)} F\big(\tau_1 (z) \big) = \mu^{-(\alpha +2)} C_{\tau_1}F(z).
\end{split}
\end{equation*}
\end{proof}
\vspace{5pt}
\begin{rem}\label{notsimi} Note that in the hyperbolic automorphism case $\sigma \big(C_{\tau}; \mathcal{A}_{\alpha}^2 ( \Pi^+)\big) \neq \sigma \big(C_{\varphi}; \mathcal{A}_{\alpha}^2 ( \mathbb{D})\big)$ for $\alpha \geq -1$, where $ \varphi = g^{-1}\circ \tau \circ g$ for any analytic bijection $g: \mathbb{D} \longrightarrow \Pi^+$. From this we can deduce that there does not exist a composition operator $C_g$ that would give an isomorphism between the spaces $\mathcal{A}_{\alpha}^2 ( \Pi^+) $ and $\mathcal{A}_{\alpha}^2 (\mathbb{D})$. The fact that in the non-automorphism case the operators $C_{\tau_1}$ and $C_{\tau_2}$ are (up to a multiplication by a scalar) adjoints of each other is a phenomenon that does not occur in the unit disc setting.
\end{rem}
\section{The spectrum of $C_{\tau}$ on $\mathcal{D}_{\alpha}^2 (\Pi^+) $ for $\alpha > -1$}\label{secdir}
From the spectral results for the parabolic and the hyperbolic composition operators in the Hardy space and the weighted Bergman spaces of the upper half-plane we will get the (essential) spectra in the weighted Dirichlet spaces of the upper half-plane in a straightforward manner.
\vspace{5pt}
We will use the definition given in \cite[Section 6]{DGM} for the weighted Dirichlet spaces on $\Pi^+$: For all $\alpha >-1$, the space $\mathcal{D}_{\alpha}^2 (\Pi^+) $ consists of the analytic functions $ F: \Pi^+ \longrightarrow \mathbb{C}$ such that $F' \in \mathcal{A}_{\alpha}^2 ( \Pi^+) $, i.e.
\begin{equation}\label{intdir} \int_{\Pi^+} \vert F' (x+iy)\vert^2 y^{\alpha} \, dx \, dy < \infty ,
\end{equation}
together with the condition that for all $x\in \mathbb{R}$
$$ F(x+iy) \longrightarrow 0, \textnormal{ when } y \longrightarrow \infty .$$
The spaces $ \mathcal{D}_{\alpha}^2 (\Pi^+) $ for $\alpha >-1$ defined in this way actually consists of the unique members of each equivalence class when identifying the functions satisfying \eqref{intdir} that differ by a constant. This being the case, we put $ \Vert F\Vert_{\mathcal{D}_{\alpha}^2 (\Pi^+) } = \Big(\int_{\Pi^+} \vert F' (x+iy)\vert^2 y^{\alpha} \, dx \, dy \Big)^{1/2}$.
\vspace{5pt}
Recall that in the classical Dirichlet space setting ($\alpha =0$) the spectra of \textit{all} linear fractional composition operators $C_{\varphi}$ and $C_{\tau}$ coincide whenever $\tau = g \circ \varphi \circ g^{-1}$ for some conformal map $g: \mathbb{D} \longrightarrow \Pi^+$. (For the spectra of linear fractional composition operators acting on the Dirichlet space $\mathcal{D}^2 (\mathbb{D})$, see \cite{CM, Hu, Hi}.) This, and the boundedness of any linear fractional composition operator acting on the Dirichlet space of the upper half-plane $\mathcal{D}_{0}^2 (\Pi^+)$, follows from the fact that the composition operator $C_g : \mathcal{D}_0^2 (\Pi^+) \longrightarrow \mathcal{D}^2 (\mathbb{D})/ \mathbb{C}$ is unitary.
\vspace{5pt}
There is also a version of the Paley-Wiener theorem for weighted Dirichlet spaces (see \cite[Thm. 3]{DGM}): Let $\alpha >-1$. Denote by $L^2_{\alpha -1} $ the space consisting of the measurable functions on $\mathbb{R}_{+}$ with finite norm
$$ \Vert f\Vert_{ L^2_{\alpha -1} }=\Big(\frac{\Gamma (\alpha +1)}{2^{\alpha}} \int_0^{\infty} \vert f (t)\vert^2 t^{1-\alpha} \, dt \Big)^{1/2}.$$
The Paley-Wiener theorem states that $F\in\mathcal{D}_{\alpha}^2$ if and only if $F(w)=\int_0^{\infty} f(t)e^{iwt} \, dt$ for all $w\in \Pi^+$, where $f \in L^2_{\alpha -1}$. Moreover, $ \Vert F\Vert_{\mathcal{D}_{\alpha}^2 }= \Vert f\Vert_{L^2_{\alpha -1}} $.
\vspace{5pt}
By noting that $ \alpha -1 =( \alpha -2) +1$, the above version of Paley-Wiener theorem gives a justification for writing $ \mathcal{D}_{\alpha}^2 (\Pi^+) = \mathcal{A}_{\alpha -2}^2 (\Pi^+) $. This point of view is also supported by the spectral results below. Note that we consider only those linear fractional composition operators which are bounded also on $ \mathcal{A}_{\alpha}^2 ( \Pi^+)$ for $\alpha > -1$. In fact, as far as we know, it is not known whether there are also other linear fractional transformations that induce bounded composition operators on $ \mathcal{D}_{\alpha}^2 (\Pi^+)$ for $ \alpha \in (-1,0) $ or $\alpha \in (0, 1) $ as there are in the unweighted Dirichlet space.
\begin{lemma}\label{lemmadirsim} Let $\tau : \Pi^+ \longrightarrow \Pi^+$ be a hyperbolic or parabolic map of the form $\tau (w)=\mu w +w_0$, where $\mu >0$ and $\textnormal{Im}\, w_0\geq 0$ (and if $\mu =1$, then $w_0\neq 0$). Then, for all $\alpha >-1$, the operators $ C_{\tau}: \mathcal{D}_{\alpha}^2 (\Pi^+) \longrightarrow \mathcal{D}_{\alpha}^2 (\Pi^+) $ and $ \mu C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow\mathcal{A}_{\alpha}^2 ( \Pi^+) $ are similar.
\end{lemma}
\begin{proof}
It is clear from the definition of the weighted Dirichlet spaces that the differentiation operator $D: F \longmapsto F'$ is an isomorphism $\mathcal{D}_{\alpha}^2 (\Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+) $. Recall that $C_{\tau} $ is bounded on $\mathcal{A}_{\alpha}^2 ( \Pi^+) $ for all $\alpha >-1$. Since, moreover, $ (F\circ \tau)' = \tau' (F' \circ \tau)$ and $\tau'(w)=\mu$ for all $w\in\Pi^+$, the composition operator $C_{\tau}$ is bounded on $\mathcal{D}_{\alpha}^2 (\Pi^+) $ and similar to the operator
$$ \mu C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+) $$
for any $\alpha >-1$.
\end{proof}
\vspace{5pt}
The following theorems are obvious consequences of Lemma \ref{lemmadirsim} and the spectral results in Sections $3$ and $4$.
\begin{thm}\label{dirpara}
Let $\tau$ be a parabolic self-map of $\Pi^+ $, that is, $\tau (w)=w+w_0$, where $\textnormal{Im}\, w_0 \geq 0$ ($w_0 \neq 0$). Then, for all $\alpha >-1$
\vspace{5pt}
\begin{itemize}
\item[i)] $ \sigma \big( C_{\tau} ; \mathcal{D}_{\alpha}^2 (\Pi^+) \big)=\sigma_{ess} \big( C_{\tau} ; \mathcal{D}_{\alpha}^2 (\Pi^+) \big) = \mathbb{T}$, when $w_0\in \mathbb{R}$,
\vspace{5pt}
\item[ii)] $\sigma \big( C_{\tau} ; \mathcal{D}_{\alpha}^2 (\Pi^+) \big) = \{e^{iw_0t} : t\in [0,\infty ) \} \cup \{0\}$, when $w_0\in \Pi^+$.
\end{itemize}
Moreover, the essential spectrum coincides with the spectrum in both cases.
\end{thm}
\begin{proof}
By Lemma \ref{lemmadirsim}, the operators $C_{\tau} : \mathcal{D}_{\alpha}^2 (\Pi^+) \longrightarrow \mathcal{D}_{\alpha}^2 (\Pi^+) $ and $ C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+) $ are similar and hence the result follows from Theorem \ref{sppara}.
\end{proof}
\begin{thm}\label{dirhyper}
Let $\tau$ be a hyperbolic self-map of $\Pi^+ $, that is, $\tau (w)=\mu w + w_0$, where $\mu\in (0,1)\cup (1,\infty)$ and $\textnormal{Im}\, w_0 \geq 0$. Then, for all $\alpha >-1$
\vspace{5pt}
\begin{itemize}
\item[i)] $\sigma \big( C_{\tau}; \mathcal{D}_{\alpha}^2 (\Pi^+) \big) = \big\{ \lambda \in \mathbb{C} : \vert \lambda \vert = \mu^{ -((\alpha -2) +2 )/2}\big\} =\big\{ \lambda \in \mathbb{C} : \vert \lambda \vert = \mu^{ -\alpha /2}\big\}$, when $w_0\in \mathbb{R}$.
\vspace{5pt}
\item[ii)] $\sigma \big( C_{\tau}; \mathcal{D}_{\alpha}^2 (\Pi^+) \big) = \big\{ \lambda \in \mathbb{C} : \vert \lambda \vert\leq \mu^{ -((\alpha -2) +2 )/2}\big\} =\big\{ \lambda \in \mathbb{C} : \vert \lambda \vert \leq \mu^{ -\alpha /2}\big\} $, when $w_0\in \Pi^+$.
\end{itemize}
Moreover, the essential spectrum coincides with the spectrum in both cases.
\end{thm}
\begin{proof}
Since the operators $C_{\tau} : \mathcal{D}_{\alpha}^2 (\Pi^+) \longrightarrow \mathcal{D}_{\alpha}^2 (\Pi^+) $ and $ \mu C_{\tau}: \mathcal{A}_{\alpha}^2 ( \Pi^+) \longrightarrow \mathcal{A}_{\alpha}^2 ( \Pi^+) $ are similar by Lemma \ref{lemmadirsim}, we have that
$$\sigma \big( C_{\tau}; \mathcal{D}_{\alpha}^2 (\Pi^+) \big) = \sigma \big( \mu C_{\tau}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big) =\big\{\lambda \in \mathbb{C} : \mu^{-1}\lambda \in \sigma \big( C_{\tau}; \mathcal{A}_{\alpha}^2 ( \Pi^+) \big)\big\}. $$
The result follows from Theorem B in Section \ref{intro} (for the proof, see Theorems \ref{sphypauto}, \ref{hyptauyks}, \ref{hyptaukaks} and Corollary \ref{sphypautoinverse} in Section \ref{hypsec}).
\end{proof}
\section*{Acknowledgements}
This article is part of the author's PhD Thesis and she would like to thank her advisors Hans-Olav Tylli and Pekka Nieminen (University of Helsinki) for helpful discussions and comments. The author is also grateful to Professor Eva Gallardo-Guti\'{e}rrez (Universidad Complutense de Madrid) for inspiring conversations.
\bibliographystyle{plain}
|
\section{Introduction}
\label{sect:intro}
W UMa-type binaries consist of two cool stars (F, G, K spectral
type) in contact with each other, surrounded by a common
convective envelope lying between the inner and outer critical
Roche surfaces. In result their components possess almost
identical surface brightness, i.e. temperature
(\citealt{Lucy+1968}, \citealt{Lucy+1976}).
\newpage
The periods of W UMa binaries are in the range 0.22--0.70 days.
They present numerous family: around of 1/500--1/130 MS stars in
the solar neighborhood (\citealt{Rucinski+2002}). There are many
studies on them (\citealt{Liu+etal+2011},
\citealt{Qian+etal+2013}, \citealt{Liao+etal+2014}, etc.) but a
complete theory of their origin, structure, evolution and future
fate still lacks.
The most present theoretical models explain the formation of
(short-period) contact systems by the systematic angular momentum
loss (AML) in initially detached binaries with orbital periods of
a couple of days, due to the magnetized stellar winds and tidal
coupling (\citealt{Vilhu+1981}, \citealt{Rahunen+1982},
\citealt{Stepien+1995}). But according to
\cite{Pribulla+Rucinski+2006} a third (distant) companion is
necessary for formation of systems with a period under 1 day.
There are two models of the evolution during the contact phase
itself. The thermal relaxation oscillation (TRO) model assumes
that each component of the binary is out of thermal equilibrium
and its size oscillates around the inner Roche lobe
(\citealt{Lucy+1976}, \citealt{Flannery+1976},
\citealt{Webbink+1977}, \citealt{Yakut+Eggleton+2005}). The binary
spends a part of its present life in contact (when both stars fill
their Roche lobes and mass flows from the secondary to the
primary) and the rest as a semi-detached binary (when only the
primary fills its Roche lobe and mass flows from the primary to
the secondary), slowly evolving towards an extreme mass ratio
system. TRO model explains well the geometry of the W UMa-type
stars: the primary component is an ordinary MS star and the
secondary is also a MS star but swollen to its Roche lobe by
energy transfer. The main problem of the TRO model is the
mechanism of the energy transfer. The alternative model
(\citealt{Stepien+2004}, \citealt{Stepien+2006},
\citealt{Stepien+2009}, \citealt{Stepien+2011}) assumes that mass
transfer occurs with the mass ratio reversal, similarly as in
Algol-type binaries, following the Roche lobe overflow (RLOF) by
the massive component. The contact configuration is formed
immediately after that or after some additional AML. Each
component is in thermal equilibrium and the large size of the
currently less massive component results from its advanced
evolutionary stage (its core is hydrogen depleted).
The final products of the W UMa-type evolution are also debatable
(\citealt{Li+etal+2007}, \citealt{Eker+etal+2008}). It is supposed
that they may become: single blue stragglers (by merging of the W
UMa components as a result of high rate of angular momentum loss);
two brown dwarfs (\citealt{Li+etal+2007}) or two stars with very
low mass (if mass-loss rate is very high).
The W-phenomenon is another unresolved problem and interesting
peculiarity of the W UMa stars appearing by the lower apparent
surface brightness of the more massive components of the W-type
systems (\citealt{Binnendijk+1970}). They are recognized by the
primary minima which are occultations (indicating that the small
components are the hotter ones). It was suspected that this effect
is due to a large coverage of the primary with cool, dark spots
reducing significantly its apparent luminosity
(\citealt{Eaton+etal+1980}, \citealt{Stepien+1980},
\citealt{Hendry+etal+1992}), but this explanation was not entirely
convincing.
\cite{Gazeas+Niarchos+2006} suggested that subtype A systems have
higher total angular momentum (AM) and can evolve into subtype W
which is the opposite of the earlier conclusion.
\newpage
Besides their key role for understanding of the stellar evolution,
the contact binaries are natural laboratories to study important
astrophysical processes: interaction of stellar winds; magnetic
activity; mass, energy and angular momentum transfer and loss;
phenomenon ''mass ratio reversal''; merging or fusion of the stars
(\citealt{Martin+etal+2011}). The period-color-luminosity relation
of the contact binary stars are an useful tool for distance
determination (\citealt{Rucinski+1994}, \citealt{Rucinski+1996};
\citealt{Rucinski+Duerbeck+1997};
\citealt{Klagyivik+Csizmadia+2004}; \citealt{Eker+etal+2008}).
Hence, the study of the properties of the W UMa stars and their
variety is important for the modern astrophysics. But the
statistics of the most interesting W UMa stars, those with short
periods, is still quite poor (\citealt{Terrell+etal+2012}) mainly
due to their faintness (they are late stars).
In this paper we present photometric observations and light curve
solutions of four short-period W UMa stars: NSVS 2244206, NSVS 908513, CSS
J004004.7+385531 $\equiv$ 2MASS J00400476+3855318 $\equiv$ GSC
02797-00705 $\equiv$ UCAC4-645-002474; VSX J062624.4+570907
$\equiv$ 2MASS J06262444+5709075 $\equiv$ CSS J062624.5+570907
$\equiv$ GSC 03772-01134. Table~1 presents their coordinates and
available (preliminary) information for their light variability.
\begin{table*}[tp]
\begin{center}
\caption[]{Previous information for our targets \label{Tab1}}
\footnotesize
\begin{tabular}{ccccccccc}
\hline\hline
\noalign{\smallskip}
Name & RA & Dec & Period & Epoch & V & Ampl & Type & Ref \\
& & & [d] & [d] & [mag]& [mag]& & \\
\hline
\noalign{\smallskip}
NSVS 2244206 & 06 06 20.21 & +65 07 21.0 & 0.280727 & - & 11.997& 0.32 & EB/EW & 1 \\
NSVS 908513 & 12 30 39.36 & +83 23 07.8 & 0.399592 & - & 11.772& 0.53 & EB/EW & 1 \\
CSS J004004.7+385531 & 00 40 04.73 & +38 55 31.9 & 0.251206 & - & 13.95 & 0.72 & EW & 2 \\
VSX J004004.4+385513 & 00 40 04.40 & +38 55 13.6 & 0.251206 & 2451359.756& 13.42 & 0.34 & EW & 3 \\
VSX J062624.4+570907 & 06 26 24.43 & +57 09 07.4 & 0.280628 & 2455162.795& 12.72 & 0.68 & EW & 3 \\
\hline\hline
\end{tabular}
\end{center}
References: 1 -- \cite{Gettel+etal+2006}; 2 -- \cite{Drake+etal+2014}; 3 --
\cite{Wozniak+etal+2004};
\end{table*}
\begin{table*}[tp]\footnotesize
\begin{center}
\caption[]{Journal of the Rozhen photometric observations
\label{t2}}
\begin{tabular}{ccccc}
\hline\hline
Target& Date & Exposure ($g',i'$) & Number ($g',i'$) & Error ($g',i'$) \\
& & [sec] & & [mag] \\
\hline
NSVS 2244206 & 2015 Jan 8 & 60, 90 & 64, 64 & 0.004, 0.004\\
& 2015 Jan 11 & 60, 90 & 78, 78 & 0.003, 0.004\\
& 2015 Jan 12 & 60, 90 & 91, 84 & 0.004, 0.004\\
& 2015 Jan 15 & 60, 90 & 147, 171 & 0.003, 0.003\\
\hline
NSVS 908513 & 2015 Mar 31 & 60, 90 & 55, 55 & 0.003, 0.004\\
& 2015 Apr 11 & 60, 90 & 106, 144 & 0.002, 0.003\\
& 2015 Apr 16 & 60, 90 & 138, 137 & 0.004, 0.005\\
\hline
CSS J004004.7+385531 & 2014 Nov 10 & 120, 120 & 110, 112 & 0.014, 0.014\\
& 2014 Nov 20 & 120, 120 & 16, 16 & 0.011, 0.015\\
& 2014 Nov 22 & 120, 120 & 44, 60 & 0.012, 0.014\\
& 2014 Nov 26 & 120, 120 & 74, 66 & 0.008, 0.010\\
\hline
VSX J062624.4+570907 & 2014 Dec 24 & 150, 150 & 118, 117 & 0.003, 0.006\\
\hline\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[tp]
\begin{center}
\caption[]{Coordinates and magnitudes of the standard and check stars \label{t3}}
\footnotesize
\begin{tabular}{ccccccccc}
\hline
\noalign{\smallskip}
Label & Star ID & Other designations& RA & Dec & \emph{g'} & \emph{ i' }\\
\hline\noalign{\smallskip}
Target 1& NSVS 2244206 & UCAC4 776-023507 & 06 06 20.21 & +65 07 21.0 & 11.977 & 11.144 \\
Chk & UCAC4 776-023451 & GSC 04103-01161 & 06 05 04.84 & +65 07 17.21 & 11.736 & 11.367 \\
C1 & UCAC4 777-021920 & GSC 04103-00028 & 06 06 58.03 & +65 18 15.55 & 13.905 & 12.452 \\
C2 & UCAC4 777-021910 & GSC 04103-00908 & 06 06 39.95 & +65 17 56.15 & 13.270 & 12.842 \\
C3 & UCAC4 777-021896 & GSC 04103-00086 & 06 06 31.01 & +65 15 33.41 & 12.338 & 11.895 \\
C4 & UCAC4 776-023478 & GSC 04103-01089 & 06 05 43.92 & +65 08 32.90 & 13.219 & 12.628 \\
C5 & UCAC4 776-023468 & GSC 04103-01386 & 06 05 27.41 & +65 03 00.31 & 12.335 & 11.802 \\
C6 & UCAC4 776-023541 & GSC 04103-00274 & 06 06 57.18 & +65 00 11.24 & 13.121 & 12.600 \\
C7 & UCAC4 775-024696 & GSC 04103-00798 & 06 07 27.43 & +64 59 18.52 & 12.602 & 11.433 \\
C8 & UCAC4 776-023570 & GSC 04103-00108 & 06 07 39.47 & +65 04 53.91 & 13.765 & 13.233 \\
Target 2 & NSVS 908513 & UCAC4 867-006050 & 12 30 39.36 & +83 23 07.8 & 11.772 & 11.105 \\
Chk & UCAC4 867-006038 & GSC 04633-01779 & 12 29 15.53 & +83 20 58.94 & 13.285 & 12.466 \\
C1 & UCAC4 867-006016 & GSC 04633-01264 & 12 25 15.86 & +83 16 24.90 & 11.335 & 10.733 \\
C2 & UCAC4 867-006026 & GSC 04633-01365 & 12 26 44.73 & +83 15 13.74 & 11.088 & 10.747 \\
C3 & UCAC4 867-006048 & GSC 04633-01496 & 12 30 02.21 & +83 15 12.10 & 13.035 & 11.580 \\
C4 & UCAC4 867-006056 & GSC 04633-01284 & 12 31 17.68 & +83 22 11.16 & 13.786 & 12.918 \\
C5 & UCAC4 867-006057 & GSC 04633-01516 & 12 31 39.72 & +83 21 50.46 & 11.334 & 10.880 \\
C6 & UCAC4 868-005899 & GSC 04633-01339 & 12 24 15.66 & +83 27 36.10 & 12.305 & 11.610 \\
C7 & UCAC4 868-005922 & GSC 04633-01435 & 12 28 54.40 & +83 26 09.13 & 12.605 & 11.971 \\
C8 & UCAC4 868-005969 & GSC 04633-01360 & 12 35 27.00 & +83 26 58.19 & 12.049 & 10.779 \\
C9 & UCAC4 868-005942 & GSC 04633-01599 & 12 31 56.99 & +83 29 43.30 & 12.548 & 11.966 \\
C10 & UCAC4 868-005965 & GSC 04633-01703 & 12 34 40.00 & +83 32 27.62 & 12.985 & 12.260 \\
C11 & UCAC4 868-005971 & GSC 04633-01424 & 12 35 37.57 & +83 33 21.72 & 13.667 & 12.479 \\
C12 & UCAC4 868-005980 & GSC 04633-01616 & 12 37 02.02 & +83 32 47.47 & 11.259 & 10.850 \\
Target 3 & CSS J004004.7+385531 & UCAC4 645-002474 & 00 40 04.73 & +38 55 31.9 & 14.531 & 13.314 \\
Chk & UCAC4 645-002460 & GSC 02784-01714 & 00 39 51.35 & +38 55 14.15 & 14.129 & 13.642 \\
C1 & UCAC4 645-002459 & GSC 02784-00660 & 00 39 51.08 & +38 52 47.94 & 13.984 & 13.560 \\
C2 & UCAC4 645-002487 & GSC 02797-00707 & 00 40 16.35 & +38 51 36.13 & 14.213 & 13.208 \\
C3 & UCAC4 645-002499 & GSC 02797-00841 & 00 40 26.60 & +38 55 17.97 & 14.117 & 13.528 \\
C4 & UCAC4 645-002488 & GSC 02797-00759 & 00 40 16.49 & +38 57 06.96 & 14.186 & 13.674 \\
C5 & UCAC4 645-002434 & GSC 02784-00090 & 00 39 22.45 & +38 55 07.13 & 13.935 & 13.808 \\
C6 & UCAC4 645-002509 & GSC 02797-00433 & 00 40 41.68 & +38 55 19.64 & 14.343 & 13.750 \\
C7 & UCAC4 644-002468 & GSC 02784-01325 & 00 39 34.96 & +38 42 54.15 & 13.725 & 13.214 \\
C8 & UCAC4 644-002496 & GSC 02784-01757 & 00 39 59.82 & +38 43 47.26 & 14.275 & 13.392 \\
Target 4 & VSX J062624.4+570907 & UCAC4 736-046853 & 06 26 24.43 & +57 09 07.4 & 13.413 & 12.469 \\
Chk & UCAC4 737-044764 & GSC 03773-00028 & 06 27 28.22 & +57 14 51.82 & 13.912 & 12.355 \\
C1 & UCAC4 737-044774 & GSC 03773-00266 & 06 27 41.41 & +57 16 07.39 & 14.768 & 14.029 \\
C2 & UCAC4 737-044766 & GSC 03773-00279 & 06 27 30.58 & +57 16 07.12 & 14.517 & 13.872 \\
C3 & UCAC4 737-044778 & GSC 03773-00030 & 06 27 48.83 & +57 18 39.37 & 13.565 & 13.001 \\
C4 & UCAC4 736-046930 & GSC 03773-00302 & 06 27 53.71 & +57 11 14.10 & 13.344 & 12.349 \\
C5 & UCAC4 736-046874 & GSC 03773-00004 & 06 26 46.16 & +57 11 30.66 & 13.964 & 13.318 \\
C6 & UCAC4 736-046859 & GSC 03772-01339 & 06 26 33.14 & +57 04 49.70 & 13.514 & 12.656 \\
C7 & UCAC4 736-046863 & GSC 03772-00447 & 06 26 36.42 & +57 02 51.35 & 14.111 & 13.185 \\
C8 & UCAC4 736-046818 & GSC 03772-01556 & 06 25 48.21 & +57 08 54.91 & 13.787 & 12.977 \\
\hline\hline
\end{tabular}
\end{center}
\end{table*}
\section{Observations}
Our CCD photometric observations of the targets in Sloan \emph{g',
i'} bands were carried out at Rozhen Observatory with the 30-cm
Ritchey Chretien Astrograph (located into the \emph{IRIDA South}
dome) using CCD camera ATIK 4000M (2048 $\times$ 2048 pixels, 7.4
$\mu$m/pixel, field of view 35 x 35 arcmin). Information for our
observations is presented in Table~2.
The photometric data were reduced by {\textsc{AIP4WIN2.0}
(\citealt{Berry+Burnell+2005}). We performed aperture ensemble
photometry with the software \textsc{VPHOT} using more than six
standard stars in the observed field of each target. The
coordinates and magnitudes of the standard and check stars (Table
3) were taken from the catalogue UCAC4
(\citealt{Zacharias+etal+2012}).
We established that there are two close objects,
CSS~J004004.7+385531 and VSX~J004004.4+385513, with the same
periods and types of variability in the VSX database (Table 1).
Our observations revealed that the true variable is CSS
J004004.7+385531 while VSX~J004004.4+385513 is a stationary star
(Fig. 1).
\begin{figure}
\centering
\includegraphics[width=8 cm, angle=0]{MS2618fig1.eps}
\caption{The field of CSS J004004.7+385531(V1) with the close
star VSX J004004.4+385513 (V2) wrongly considered as a variable}
\label{Fig1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.55\columnwidth]{MS2618fig2.eps}
\caption{Top: the folded light curves of NSVS 2244206 and their fits; Bottom: the corresponding residuals
(shifted vertically by different number to save space).
Color version of this figure is available in the online journal.}
\label{Fig2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.55\columnwidth]{MS2618fig3.eps}
\caption{Same as Fig. 2 for NSVS 908513}
\label{Fig3}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.55\columnwidth]{MS2618fig4.eps}
\caption{Same as Fig. 2 for CSS J004004.7+385531}
\label{Fig4}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.55\columnwidth]{MS2618fig5.eps}
\caption{Same as Fig. 2 for VSX J062624.4+570907}
\label{Fig5}
\end{figure}
We determined the times of the individual minima (Table 4)
by the method of \cite{Kwee+Woerden+1956}.
\begin{table}
\begin{center}
\caption[]{Times of minima of our targets \label{Tab1}}
\footnotesize
\begin{tabular}{ccc}
\hline\hline
\noalign{\smallskip}
Target & MinI & MinII \\
\hline
\noalign{\smallskip}
NSVS 2244206 & - & 2457033.51298\\
& - & 2457034.35629 \\
& - & 2457038.28497 \\
& 2457038.42614 & - \\
\hline
NSVS 908513 & - & 2457113.31730 \\
& 2457124.30454 & 2457124.50608 \\
& 2457129.50051 & 2457129.30045 \\
\hline
CSS J004004.7+385531 & 2456972.34036 & 2456972.46714 \\
& 2456972.59113 & 2456982.26682 \\
& 2456984.39854 & 2456988.29364 \\
& 2456988.41840 & - \\
\hline
VSX J062624.4+570907 & 2457016.33657 & 2457016.47792 \\
& 2457016.61853 & - \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\section{Light curve solutions}
We carried out modeling of the photometric data by the code
\textsc{PHOEBE} (\citealt{Prsa+Zwitter+2005}). It is based on the
Wilson--Devinney (WD) code (\citealt{Wilson+Devinney+1971},
\citealt{Wilson+1979}). \textsc{PHOEBE} incorporates all the
functionality of the WD code but also provides a graphical user
interface alongside other improvements, including updated filters
as Sloan ones used in our observations. We apply the traditional
convention the MinI (phase 0.0) to be the deeper light minimum and
the star that is eclipsed at MinI to be the primary (hotter)
component.
\newpage
We determined in advance the mean temperatures $T_{m}$ of the
binaries (Table~5) by their infrared color indices \emph{(J-K)}
from the 2MASS catalog and the calibration color-temperature of
\cite{Tokunaga+2000}. In fact, the determination of stellar
temperatures from the infrared flux is a method first developed by
\cite {Blackwell+Shallis+1977}.
Our procedure of the light curve solutions was carried out in
several stages.
At the first stage we fixed $T_{1}^0$ = $T_{m}$ and searched for fit
varying the secondary temperature $T_{2}$, orbital inclination
$i$, mass ratio $q=m_2/m_1$ and potentials $\Omega_{1,2}$ (and
thus relative radii $r_{1,2}$ and fillout factor \emph{f}). The
fit quality was estimated by the value of $\chi^2$.
\newpage
Coefficients of gravity brightening and reflection effect
appropriate for stars with convective envelopes were adopted.
Initially we used linear limb-darkening law with limb-darkening
coefficients corresponding to the stellar temperatures and
Sloan photometric system (\citealt{Claret+Bloemen+2011}).
In order to reproduce the light curve distortions of the targets
we added cool spots on the stellar surfaces and varied spot
parameters: longitude $\lambda$, latitude $\beta$, angular size
$\alpha$ and temperature factor $\kappa=T_{sp}/T_{st}$.
As a result of the first stage of the light curve solution
we obtained the values $T_{2}^{0}$, $i^{0}$, $\Omega_{1,2}^{0}$, $q$
as well as the spot parameters for each target.
After reaching the best fit we adjusted $T_{1}$ and $T_{2}$ around
the value $T_m$ by the formulae (\citealt{Kjurkchieva+etal+2015})
\begin{equation}
T_1^c=T_{\rm {m}} + \frac{c \Delta T}{c+1}
\end{equation}
\begin{equation}
T_2^c=T_1^c -\Delta T
\end{equation}
where $c=l_2/l_1$ and $\Delta T=T_{1}^{0}-T_{2}^{0}$ are determined from the \textsc{PHOEBE} solution.
Finally, we varied slightly $T_{1}$, $T_{2}$, $i$ and
$\Omega_{1,2}$ around their values $T_{1}^{c}$, $T_{2}^{c}$, $i^{0}$ and
$\Omega_{1,2}^{0}$ and obtained the final \textsc{PHOEBE}
solution.
The first part of Table~5 contains the parameters of our light
curve solutions: mass ratio $q$; orbital inclination $i$;
potentials $\Omega_{1,2}$; fillout factor \emph{f}; stellar
temperatures $T_{1, 2}$; relative radii $r_{1, 2}$; ratio of
relative luminosities $l_{2}/l_{1}$. The errors of these
parameters are the formal \textsc{PHOEBE} errors. Table~6 gives
the obtained spot parameters. The synthetic curves corresponding
to our light curve solutions are shown in Figs. 2-5.
\begin{figure}
\begin{center}
\includegraphics[width=3cm]{MS2618fig6a.eps}
\includegraphics[width=3cm]{MS2618fig6b.eps}
\includegraphics[width=3cm]{MS2618fig6c.eps}
\includegraphics[width=3cm]{MS2618fig6b.eps}
\caption{From the left to the right 3D configurations of: NSVS 2244206, NSVS 908513, CSS J004004.7+385531, VSX J062624.4+570907}\label{fig4}
\end{center}
\end{figure}
Due to the lack of radial velocity measurements we had not a
possibility to determine reliable values of the global parameters
of the target components. We were able to obtain some estimations
of these quantities by the following procedure.
The primary luminosity $L_1$ was determined by the empirical
relation luminosity-temperature for MS stars. The secondary
luminosity $L_2$ was calculated by the relation $L_2=c L_1$ where
$c=l_2/l_1$ is the luminosity ratio from our light curve solution.
We obtained the orbital separation $a$ (in solar radii) from the
equation
\begin{equation}
\log a = 0.5 \log L_i - \log r_{i} - 2 \log T_{i} + 2 \log
T_{\odot}.
\end{equation}
and then calculated the absolute stellar radii by $R_{i}=ar_{i}$.
\newpage
The total target mass $M$ (in solar units) was calculated from the
third Kepler law
\begin{equation}
M=\frac{0.0134 a^3}{P^2}
\end{equation}
where \emph{P} is in days while \emph{a} is in solar radii. Then
the individual masses $M_i$ were determined by the formulae $ M_1
= M /(1+q)$ and $ M_2= M- M_1$.
The global parameters of the target components obtained by the
foregoing procedure are given in the second part of Table~5. Their
errors are calculated from the corresponding formulae by the
errors of the quantities of the light curve solutions or
observations.
\begin{table*}[tp]\footnotesize
\begin{center}
\caption{Parameters of the best light curve solutions \label{t4}
(top) and global parameters (bottom) of the targets}
\begin{tabular}{ccccc}
\hline\hline
Star name & NSVS 2244206 & NSVS 908513 & CSS J004004.7+385531 & VSX J062624.4+570907 \\
\hline
\emph{q} & 0.735$\pm$0.003 & 0.709 $\pm$ 0.002 & 0.548 $\pm$ 0.004 & 0.777 $\pm$ 0.002 \\
\emph{i}, ($^{\circ}$) & 76.42$\pm$0.07 & 75.15 $\pm$ 0.03 & 89.77 $\pm$ 0.01 & 78.88 $\pm$ 0.11 \\
$\Omega_1=\Omega_2$ & 3.1961$\pm$.004 & 3.2 $\pm$ 0.002 & 2.9 $\pm$ 0.009 & 3.307 $\pm$ 0.004 \\
\emph{f} & 0.260$\pm$0.002 & 0.146 $\pm$ 0.004 & 0.206 $\pm$ 0.009 & 0.162 $\pm$ 0.004 \\
$T_m$, (K) & 5000 & 5810 & 4560 & 5230 \\
$T_1$, (K) & 5157 $\pm$ 36 & 5923 $\pm$ 25 & 4560 $\pm$ 37 & 5350 $\pm$ 20 \\
$T_2$, (K) & 4702 $\pm$ 32 & 5615 $\pm$ 23 & 4560 $\pm$ 38 & 5044 $\pm$ 7 \\
$r_1$ & 0.429$\pm$0.001 & 0.422 $\pm$ 0.001 & 0.449 $\pm$ 0.003 & 0.416 $\pm$ 0.002 \\
$r_2$ & 0.376$\pm$0.001 & 0.363 $\pm$ 0.001 & 0.344 $\pm$ 0.004 & 0.372 $\pm$ 0.002 \\
$l_2/l_1$ & 0.5301 & 0.5946 & 0.5883 & 0.6314 \\
\hline
$L_{1}^{bol}$ & 0.519$\pm$0.023 & 1.031 $\pm$ 0.023 & 0.254 $\pm$ 0.04 & 0.604 $\pm$ 0.073 \\
$L_{2}^{bol}$ & 0.338$\pm$0.045 & 0.646 $\pm$ 0.046 & 0.149 $\pm$ 0.062 & 0.421 $\pm$ 0.096 \\
\emph{a}, (R$_{\odot}$) & 2.214$\pm$0.06 & 2.312 $\pm$ 0.066 & 1.798 $\pm$ 0.055 & 2.23 $\pm$ 0.067 \\
$R_1$, (R$_{\odot}$) & 0.951$\pm$0.028 & 0.976 $\pm$ 0.03 & 0.808 $\pm$ 0.03 & 0.927 $\pm$ 0.032 \\
$R_2$, (R$_{\odot}$) & 0.833$\pm$0.032 & 0.839 $\pm$ 0.033 & 0.619 $\pm$ 0.026 & 0.829 $\pm$ 0.027 \\
$M_1$, (M$_{\odot}$) & 1.064$\pm$0.084 & 0.607 $\pm$ 0.051 & 0.798 $\pm$ 0.07 & 1.061 $\pm$ 0.094 \\
$M_2$, (M$_{\odot}$) & 0.782$\pm$0.067 & 0.430 $\pm$ 0.038 & 0.437 $\pm$ 0.043 & 0.825 $\pm$ 0.077 \\
\hline\hline
\end{tabular}
\end{center}
\end{table*}
\section{Analysis of the results}
The analysis of the light curve solutions of our short-period W
UMa stars led to several important results.
(1) CSS J004004.7+385531 reveals total eclipses while the remained
three targets undergo partial eclipses.
(2) The temperatures of the stellar components of the targets
correspond to G-K spectral type (Table~5). The temperature
differences of their components do not exceed 450 K while the
components of CSS J004004.7+385531 are in precise thermal contact.
(3) The targets have overcontact configurations which fillout
factors \emph{f} are in the range 0.15--0.26 (Table 5). It should
be pointed out that the preliminary classification of NSVS 2244206
and NSVS 908513 was EB/EW (Table 1) but our observations and light
curve solutions led to the conclusion that their configurations
are overcontact (Fig. 6).
\newpage
(4) The target components are relatively close in size and
luminosity: the size ratios $r_2/r_1$ are within 0.75--0.90; the
luminosity ratios $l_2/l_1$ are within 0.53--0.63.
(5) We observed slightly different levels of the two
quadratures (O'Connell effect) of three our targets. They were
reproduced by small cool spots (Table 6) on their primaries. We
obtained solutions with the same fit quality for combinations of
slightly different spot sizes (within 1$^{o}$) and latitudes ($\pm
25^{o}$ around the stellar equator). Table~6 presents the
parameters of the equatorial spots whose angular sizes have
minimum values.
(6) The residuals of CSS J004004.7+385531 are bigger than those of
the other three targets that is expected taking into account that
this totally-eclipsed binary is the faintest member of our sample
(Table 1). But the residuals are biggest at its primary eclipse
because the synthetic eclipse is narrower than the observed one
(Fig. 4). The reason for this discrepancy is that the observed
primary eclipse of CSS J004004.7+385531 turns out slightly wider
than the secondary one. This may due to some additional structure
in the system which presence cannot be taken into account from the
software for light curve synthesis: equatorial bulge around the
less massive component (accretor) formed as a result of the
transferred mass; disk-like feature; clouds at some Lagrangian
points (result of previous nonconservative mass transfer,
\citealt{Stepien+Kiraga+2013}).
(7) We managed to reproduce the almost flat bottom of the
primary eclipse of CSS J004004.7+385531, especially in \emph{i'}
band, only by very small limb-darkening coefficient of 0.18, a
value considerably smaller than that corresponding to its
temperature. This formally means faint limb-darkening effect, i.e.
almost homogeneous stellar disk of the primary component. There
are two possible reasons for this effect. The first one is that
the theory and corresponding codes for light curve synthesis
cannot take into account precisely the limb-darkening effect for
overcontact binaries with strongly distorted components whose
photospheres are deeply inside the envelopes. The second reason
for the flat bottom of the primary eclipse might be light
contribution of optically thick region around L1 that is covered
by the secondary component at the primary eclipse. Such an
argument refers just for CSS J004004.7+385531 that undergoes
almost central eclipse.
(8) Quite often photometric solutions of W UMa light curves
appear to be ambiguous since both A and W configurations can fit well
the observations (\citealt{vanHamme+1982a}; \citealt{Lapasset+Claria+1986}).
The mass ratio of the W UMa binaries is important
parameter for their W/A subclassification. But the rapid rotation
of their components does not allow to obtain precise spectral mass
ratio from measurement of their highly broadened and blended
spectral lines (\citealt{Bilir+etal+2005};
\citealt{Dall+Schmidtobreick+2005}). As a result the W/A
subclassification of the W UMa binaries
is made mainly on the widely-accepted empirical relation
''spectral type -- mass'' (\citealt{vanHamme+1982a}; \citealt{Lapasset+Claria+1986}):
the G-K binaries are of W subtype while A and earlier F
binaries are of A subtype.
\newpage
Particularly, our targets are faints objects and we have
obtained only their photometric mass ratios \emph{q} by
varying this parameter within the range 0.1-2.5.
Thus, we obtained a pair solutions for each target with
close quality: (a) A-subtype solution with parameters $q^A < 1$,
$r_1^A$, $r_2^A < r_1^A$; (b) W-subtype solution with parameters
$q^W \approx 1/q^A$, $r_1^W \approx r_2^A$, $r_2^W \approx r_1^A$.
To choose one of them we introduced the relative difference
$\Delta Q$ (in percentage) of the solution quality \emph{Q(W)}
corresponding to the W-subtype configuration and the solution
quality \emph{Q(A)} corresponding to the A-subtype configuration
\begin{equation}
\Delta Q = \frac{Q(W)-Q(A)}{Q(A)} .
\end{equation}
For all our targets we established $\Delta Q \geq 3 \%$,
i.e. the A solutions are better than the W solutions. This allowed
us to assume that A subtype is more probable subclassification of
our targets than the W subtype. Additional considerations for this
choice were: (i) the W solutions did not reproduce so good the
observed eclipse depths; (ii) the big fillout factors of our
targets are inherent to A subtype binaries
(\citealt{vanHamme+1982a}, \citealt{vanHamme+1982b});
(iii) the A solutions are quite sensible to the mass ratio (Fig. 7).
Hence, our targets could be assigned to the exceptions
from the statistical relation ''spectral type -- mass''. The
possible reason for this discrepancy may be that this relation is
derived for W UMa binaries with periods $>$ 0.3 d (\citealt{vanHamme+1982a}, \citealt{vanHamme+1982b}).
\begin{table}\footnotesize
\begin{center}
\caption[]{Parameters of the cool spots on the targets
\label{t5}}
\begin{tabular}{ccccc}
\hline\hline
Target & $\beta$ & $\lambda$ & $\alpha$ & $\kappa$ \\
\hline
NSVS 908513 & 90 & 330 & 13 & 0.90 \\
CSS J004004.7+385531 & 90 & 70 & 7 & 0.90 \\
VSX J062624.4+570907 & 90 & 270 & 10 & 0.90 \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\centering
\includegraphics[width=4.5cm, angle=0]{MS2618fig7.eps}
\caption{Sensibility of our light curve solution of NSVS 2244206 (measured by $\chi^2$) to the mass ratio
(the rest parameters last fixed at their final values)}
\label{Fig7}
\end{figure}
\newpage
\section{Conclusion}
We obtained light curve solutions of four short-period W UMa
binaries which main results are as follows.
(1) The temperatures of the stellar components of the targets
correspond to G-K spectral type and they are almost in thermal
contacts.
(2) All targets are overcontact configurations with
fillout factor within 0.15--0.26.
(3) The target components are relatively close in size and
luminosity: the size ratios $r_2/r_1$ are within 0.75--0.90; the
luminosity ratios $l_2/l_1$ are within 0.53--0.63.
(4) The results of the light curve solution of
CSS~J004004.7+385531 imply weak limb-darkening effect of its
primary component and possible presence of additional absorbing
feature in the system.
This study adds new four systems with estimated parameters to the
family of short-period binaries. They could help to improve the
statistical relations between the stellar parameters of the
low-massive stars and to better understanding the evolution of
close binaries.
\normalem
\begin{acknowledgements}
The research was supported partly by funds of project RD 08-244 of
Scientific Foundation od Shumen University. It used the SIMBAD
database and NASA Astrophysics Data System Abstract Service. This
research was made possible through the use of the AAVSO
Photometric All-Sky Survey (APASS), funded by the Robert Martin
Ayers Sciences Fund. The authors are grateful to the
anonymous referee for the valuable notes and recommendations.
\end{acknowledgements}
\bibliographystyle{raa}
|
\section{Policy Synthesis Algorithm}
\label{sec:algorithm}
In this section, we define our policy synthesis algorithm based on SMT solving, called ${\cal S}_{\sf smt}$. The
algorithm takes as input a resource structure~$S$, a set~$R$ of
requirements, and a set $C$ of configurations. The set $C$ is encoded
symbolically, as we describe shortly. The algorithm outputs a
configuration~$c$ such that $S, c\Vdash R$, if there is such a
configuration in $C$; otherwise, it returns $\sf unsat$. To synthesize
a configuration $c$, the algorithm encodes the question $\exists c\in
C.\ S, c\Vdash R$ in a decidable logic supported by standard SMT
solvers.
Due to its technical nature, we relegate a detailed description of the
encoding to the end of this section.
Our algorithm takes as input a set of configurations,
and we refer to the symbolic encoding of this set as a {\em
configuration template}.
The configuration template enables us to restrict the search space:
the algorithm confines its search to the configurations described by
the template.
Our algorithm ${\cal S}_{\sf smt}$ is sound, independent of the provided
configuration template. Its completeness, however, depends on
the template.
We show that one can construct a template for which ${\cal S}_{\sf smt}$ is
complete, but the resulting template would, in practice, encode so many configurations that the resulting SMT problem would be infeasible to solve.
We therefore strike a balance between the algorithm's completeness and
its efficiency: since real-world local policies often have small
syntactic representations, as demonstrated by our experiments in
Section~\ref{sec:eval}, our policy synthesis tool starts with a
template that defines configurations with succinct local policies, and iteratively
executes ${\cal S}_{\sf smt}$, increasing the template's size in each iteration.
It turns out that in our case studies
a small number of iterations is
sufficient to synthesize all local policies.
Below, we describe the algorithm ${\cal S}_{\sf smt}$'s components.
\subsection{Configuration Templates}
A {\em configuration template} assigns to each edge of the resource
structure a symbolic encoding of a set of local policies.
To illustrate this encoding, consider the set of local policies
$\{{\sf true}, {\sf role} = {\sf employee}, {\sf role} \neq {\sf
visitor}\}$. We symbolically encode this set for an edge, say $({\sf cor}, {\sf bur})$,
as a constraint over subject and contextual attributes, as well as
a {\em control} variable~$z_{({\sf cor}, {\sf bur})}$:
\begin{equation}\tag{T1}
\label{ex:template}
\begin{array}{llcl}
\hspace{-10pt}C(({\sf cor}, {\sf bur})) =\hspace{-6pt} & (z_{({\sf cor}, {\sf bur})} = 1 & \hspace{-6pt}\Rightarrow\hspace{-6pt} & {\sf true})\ \wedge\\
& (z_{({\sf cor}, {\sf bur})} = 2 & \hspace{-6pt}\Rightarrow\hspace{-6pt} & {\sf role} = {\sf employee})\ \wedge\\
& (z_{({\sf cor}, {\sf bur})} = 3 & \hspace{-6pt}\Rightarrow\hspace{-6pt} & {\sf role} \neq {\sf visitor})~.
\end{array}
\end{equation}
The control variable $z_{({\sf cor}, {\sf bur})}$ encodes the choice of one of three local policies for the edge~$({\sf cor}, {\sf bur})$.
Hence, for this example, the set of configurations defined by the
configuration template contains~$3^{|E|}$ elements, where~$E$ is the
set of edges in the resource structure.
Note that for a set of local policies of size~$n$ (here $n=3$),~$\lceil \log n\rceil$ propositional variables are sufficient for
representing each edge's control variables. To avoid clutter, we will write $C_{r_0, r_1}$ for $C((r_0, r_1))$.
We remark that configuration templates can be used to restrict the search space of configurations to those that satisfy {\em attribute availability} constraints, which restrict the set of attributes that PEPs can retrieve.
Suppose that only the side-entrance door of our running example is equipped with a keypad.
To account for this constraint, we will restrict the configurations in the template to those that use the $\operatorname{\sf correct-pin}$ attribute only in the local policy of side entrance's lock.
\subsection{Algorithm}
The main steps of~${\cal S}_{\sf smt}$ are given in Algorithm~\ref{alg:smt}. We
describe the algorithm with an example: the input to the algorithm
consists of the resource structure and the requirements $\mathbf{R2}$
and $\mathbf{R5}$ of our running example, along with the above
configuration template~$C$, which maps edges to the set of local
policies $\{{\sf true}, {\sf role} = {\sf employee}, {\sf role} \neq
{\sf visitor}\}$.
The algorithm starts by creating for each requirement a constraint
that asserts the satisfaction of the requirement in the resource
structure, given the template. This constraint is called~$\psi$ in the
algorithm, and is expressed in the logic of an SMT solver.
This step is implemented by the subroutine~\textsc{Encode},
defined in Figure~\ref{fig:encode}. To encode the satisfaction of access constraints, we follow the standard model-checking algorithm for CTL based on labeling~\cite{Huth:2004:LCS:975331}; we explain this encoding at the end of this section.
As an example, the result of $\textsc{Encode}(S, \textbf{R2}, C)$,
after straightforward simplifications, is the following constraint:
\begin{align*}
\psi_{R2} := {\sf role} = {\sf visitor} \Rightarrow (\neg C_{\sf out, cor}\vee \neg C_{\sf cor, mr}) \, .
\end{align*}
Here ${\sf role}$ is an \emph{attribute variable}, originating from
\textbf{R2}'s target, and $C_{\sf out, cor}$ and $C_{\sf cor, mr}$
are the symbolic encodings of the local policies for the edges~$({\sf out, cor})$ and $({\sf cor, mr})$, respectively.
This constraint states that if the
requirement's target ${\sf role} = {\sf visitor}$ is satisfied, then
one of the PEPs along the path that starts at the entry resource and
reaches the meeting room directly through the corridor must deny
access.
Similarly, $\textsc{Encode}(S, \textbf{R5}, C)$ returns the
constraint:
\begin{align*}
\psi_{R5} :=\ & {\sf role} \neq {\sf employee} \Rightarrow \\
& ((\neg C_{\sf out, cor}\vee \neg C_{\sf cor, bur}) \\
& \wedge (\neg C_{\sf out, lob}\vee \neg C_{\sf lob, cor} \vee \neg C_{\sf cor, bur})) \, .
\end{align*}
This states that any access request that maps the attribute
$\sf role$ to a value other than $\sf employee$ must be denied by
at least one PEP along the path to the bureau that goes directly through the corridor, and moreover it must be denied by
at least one PEP along the path that passes through the lobby.
\begin{algorithm}[t]
\DontPrintSemicolon
\KwIn{A resource structure $S = ({\cal R}, E, r, L)$,\hspace{45pt}
a set~$\{R_1, \cdots, R_n\}$ of requirements,\hspace{55pt}
a configuration template~$C$}
\KwOut{A configuration~$c$ or ${\sf unsat}$}
\Begin{
$\phi \gets {\sf true}$\;
\For {$R \in \{R_1, \cdots, R_n\}$} {
$\psi\gets \textsc{Encode}(S, R, C)$\; \label{line:smt-encode}
$\phi \gets \phi \wedge \psi$\;
}
\If {$(\exists{\vec{z}}. \forall{\vec{a}}.\ \phi)\ \mathrm{is}\
{\sf sat}$} { \label{line:smt-solve}
${\cal M}\gets \textsc{Model}(\exists{\vec{z}}. \forall{\vec{a}}.\ \phi)$\;
\For {$e \in E$} {
$c(e) \gets \textsc{Derive}(C(e), {\cal M})$\;
}
\Return{$c$}\;
} \Else {
\Return{${\sf unsat}$}
}
}
\caption{The algorithm ${\cal S}_{\sf smt}$ for synthesizing policies
using SMT solving.}
\label{alg:smt}
\end{algorithm}
\input{translation-rules}
The conjunction of the constraints created for all the requirements is
called~$\phi$ in Algorithm~\ref{alg:smt}. To check whether there is a
configuration in~$C$ that satisfies the requirements, the
algorithm calls an SMT solver to find a model for the formula $\exists \vec{z}. \forall \vec{a}.\ \phi$. Here this is
\[ \exists \vec{z}. \forall \vec{a}.\ (\psi_{R2}\wedge \psi_{R5})~, \]
where $\vec{z}$ and $\vec{a}$ consist, respectively, of all the
control and attribute variables.
If~$\phi$ is unsatisfiable, then no configuration in $C$ satisfies the
requirements. In this case, the algorithm returns~$\mathsf{unsat}$.
If however the formula is satisfiable, then
the SMT solver returns a model of the formula, which instantiates all
the control variables (but not the attribute variables since they are
universally quantified).
We refer to the SMT solver's procedure that returns such a model as
{\sc Model} in Algorithm~\ref{alg:smt}.
The model~$\mathcal{M}$ generated by the SMT solver in effect
identifies the local policy for each edge~$e$: by instantiating the
control variables in~$C(e)$, we obtain~$e$'s local policy; see
template~\ref{ex:template}. This procedure is called $\textsc{Derive}(C(e), {\cal M})$ in the algorithm.
For our example, a model $\cal M$ that satisfies $\exists \vec{z}. \forall \vec{a}.\ (\psi_{R2}\wedge \psi_{R5})$ maps $z_{(\sf cor, bur)}$ to $2$, $z_{(\sf out, cor)}$ to $3$, and all other control variables to $1$.
It is then evident from template~\ref{ex:template} that, e.g., the local policy for the edge $({\sf cor, bur})$ is $({\sf role} = {\sf employee})$.
\para{Complexity}
Let $S$ be a resource structure, $R$ be a set of requirements, and $C$
be configuration template.
The running time of the ${\cal S}_{\sf smt}$ algorithm is determined by
the size of the generated formula $\phi$ and
the complexity of finding a model of~$\phi$.
The size of the formula~$\phi$ is in~${\cal O}(d\cdot |R|\cdot |{\cal R}|)$, where~$d$ is the size of the largest access constraint that appears in the requirements, $R$ is the set of requirements, and $\cal R$ is the set of resources in~$S$.
The formula $\phi$ is defined over Boolean control variables~$\vec{z}$
and attribute variables~$\vec{a}$. The number of control and attribute variables is
$\lceil{\it log}(|C|)\rceil$ and
$|{\cal A}|$, respectively. In the worst case, one must check all possible models of
the formula $\phi$, so finding a model of $\phi$ is in
${\cal O}(2^{\lceil{\it log}(|C|)\rceil +
k\cdot |A|})$, where $k$ is the largest domain
that appears in the constraints.
Note that such domains are always finite. For example, ${\sf time}
\geq 10$ is a shorthand for $\neg ({\sf time} \in \{0, \ldots, 9\})$.
We conclude that the overall running time of the algorithm ${\cal S}_{\sf smt}$ is in
${\cal O}(2^{\lceil{\it log}(|C|)\rceil +
k\cdot |A|} + d\cdot |R|\cdot |{\cal R}|)$.
\subsection{Soundness and Completeness}
The algorithm ${\cal S}_{\sf smt}$ is sound.
\begin{theorem}
Let $S$ be resource structure, $R$ a set of requirements, and $C$ a configuration template. If ${\cal S}_{\sf smt}(S, R, C) = c$ then $S, c\Vdash R$. If ${\cal S}_{\sf smt}(S, R, C) = {\sf unsat}$, then there is no configuration $c$ in $C$ such that $ S, c\Vdash R$.
\label{thm:soundness}
\end{theorem}
${\cal S}_{\sf smt}$'s completeness depends on the template~$C$ provided as input to the algorithm. We show that one can construct a template for which the algorithm is complete.
A template $C$ is {\em complete} for a given resource structure~$S$
and set of requirements~$R$ if ${\cal S}_{\sf smt}(S, R, C)$ returns a configuration
whenever there is a configuration that satisfies the requirements.
For the algorithm's completeness, it is in fact sufficient to start the
algorithm with a template~$C_{S,R}$ that contains all the
configurations that the algorithm based on controller synthesis,
described in Section~\ref{sec:decidability}, may output.
The following theorem formalizes this observation.
\begin{theorem}
Given a resource structure~$S$ and a set~$R$ of requirements, the
configuration template~$C_{S,R}$ is complete for $S$ and~$R$.
\label{thm:completeness}
\end{theorem}
The number of configurations in $C_{S,R}$ is exponential in $|E|$ and
$|R|$ (which we prove in~\cite{tech-report}). Hence this template, although
complete, is not useful in practice as it would overwhelm SMT solvers,
rendering~${\cal S}_{\sf smt}$ ineffective. In \secref{sec:implementation}, where we explain our implementation in detail, we describe a configuration template that works
well for synthesizing configurations for practically-relevant
examples.
We conclude this discussion by pointing out that our synthesis
algorithm can be readily used to verify whether a candidate
configuration~$c$ satisfies a set~$R$ of
global access-control requirements in a resource
structure~$S$. Namely, if the configuration template input to ${\cal S}_{\sf smt}$
consists only of the configuration~$c$, then ${\cal S}_{\sf smt}$ returns~$c$
if~$S,c\models R$; otherwise, the algorithm returns ${\sf unsat}$,
which means that the configuration~$c$ does not satisfy~$R$.
\subsection{Encoding into SMT}
We now explain Algorithm~\ref{alg:smt}'s procedure~{\sc Encode}, which translates a resource structure $S$, a requirement $R = (T\Rightarrow \varphi)$, and a configuration template~$C$, into an SMT constraint $T\Rightarrow \tau(\varphi, r_e)$. The generated constraint encodes that whenever the requirement $T\Rightarrow \varphi$ is applicable to an access request~$q$, i.e. $q\vdash T$, then $\varphi$ must be satisfied for the entry resource $r_e$ in the structure~$S_{c,q}$. Here,~$c$ is the configuration selected from the template~$C$.
The constraint $\tau(\varphi, r_e)$ is generated using the rewrite rules $\tau$ as defined in Figure~\ref{fig:encode}.
Given an access constraint $\varphi$ and a resource~$r_0$, the rewrite rules $\tau$ produce an SMT constraint $\tau(\varphi, r_0)$ that encodes $S, r_0\models \varphi$; see Figure~\ref{fig:semantics}.
The rewrite rules for access constraints of the form $\sf true$, $a\in D$, $\neg \varphi$, and $\varphi_1\wedge \varphi_2$ are as expected.
The rewrite rule for access constraints of the form ${\sf EX}\varphi$
encodes that the access constraint~$\varphi$ is satisfied at~$r_0$ if there is an edge from~$r_0$ to some node~$r_1$ such that $C_{r_0, r_1}$ holds and $S, r_1\models \varphi$. In this rule, the constraint $C_{r_0, r_1}$ returns the symbolic encoding of the local policies for the edge $(r_0,r_1)$, and $\tau(\varphi, r_1)$ returns the encoding of $S, r_1\models \varphi$ as an SMT constraint.
In contrast to~${\sf EX}$, the rewrite rule for ${\sf AX}\varphi$ access constraints states that for any resource~$r_1$, such that~$(r_0,r_1)\in E$, if $C_{r_0,r_1}$ is true then
the constraint~$\tau(\varphi, r_1)$ is satisfied.
To encode the semantics of the connectives ${\sf EU}$ (${\sf AU}$), we use the \emph{until}
rewrite rules $\tau_{\sf U}$, which reduce an until construct ${\sf E}[\varphi_1 {\sf U}\varphi_2]$ (${\sf A}[\varphi_1 {\sf U}\varphi_2]$), a resource $r_0\in {\cal R}$, and a set of resources $X\subseteq {\cal R}$ to an SMT constraint. We use the set of resources $X$ to record for which resources the satisfaction of the until access constraint has already been encoded. This is necessary to guarantee the reduction system's termination.
The rule for access constraints of the form ${\sf E}[\varphi_1 {\sf U}\varphi_2]$ encodes that either $S,r_0\models \varphi_2$, or $S,r_0\models \varphi_1$ and there is an edge from~$r_0$ to some node~$r_1$ such that $C_{r_0, r_1}$ holds and $S, r_1\models {\sf E}[\varphi_1 {\sf U}\varphi_2]$. Here $\tau_{\sf U}(E[\varphi_1 {\sf U}\varphi_2], r_1, X\cup \{r_0\})$ returns the encoding of $S, r_1\models {\sf E}[\varphi_1 {\sf U}\varphi_2]$. Note that we add
$r_0$ to $X$ to ensure that no resource is revisited during~${\sf EU}$-rewriting.
Similarly, the rule for access constraints ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ encodes that either $S,r_0\models \varphi_2$, or $S,r_0\models \varphi_1$ holds, for any outgoing edge to a node~$r_1$ we have $S, r_1\models {\sf A}[\varphi_1 {\sf U}\varphi_2]$
and it has no outgoing edges to nodes in $X$.
\ifext
We illustrate the encoding with examples in Appendix~\ref{sec:soundness}. There, we also prove that this rewrite system always terminates, and the generated SMT encoding of access constraints is correct.
\else
We illustrate the encoding with examples in~\cite{tech-report}. There, we also prove that this rewrite system always terminates, and the generated SMT encoding of access constraints is correct.
\fi
\section{Conclusion}
\label{sec:conc}
We have presented a framework for synthesizing locally enforceable
policies from global, system-wide access-control requirements for
physical spaces. Its key components are (1)~a~declarative language
along with patterns for writing global requirements, (2) a model of
the physical space describing how subjects access resources, and (3)
an efficient policy synthesis algorithm for generating policies
compliant with the requirements and the spatial constraints.
Using real-world case studies, we have demonstrated that our synthesis framework is practical and scales to systems with complex requirements and numerous policy enforcement points.
As future work, we plan to extend our policy synthesis framework with architectural constraints. Examples include \emph{optimality} constraints, which can be used to synthesize local policies that avoid re-checking attributes that have been checked by other enforcement points.
Handling such constraints is important for large-scale access-control systems in practice.
We also plan to apply our framework to synthesize locally enforceable policies in other access-control domains, such as access control in networks, which may require tailored synthesis heuristics.
\section{Implementation and Evaluation}
\label{sec:eval}
We report on an implementation of our policy synthesis algorithm, the case studies we conducted to evaluate its efficiency and scalability, and our empirical results.
\subsection{Implementation}
\label{sec:implementation}
We have implemented a synthesizer that encodes policy synthesis instances into the QF\_LIA and QF\_UA logics of SMT-LIB v2~\cite{Barrett10c} and uses the Z3 SMT solver~\cite{DeMoura:2008:ZES:1792734.1792766}.
Our synthesizer is configured with configuration templates of different sizes.
The local policies defined by these configuration templates are in disjunctive normal form.
Namely, the local policies are defined as a disjunction of clauses, each clause consisting of a conjunction of terms, where each term is either an equality constraint for non-numerical
attributes (e.g. ${\sf role} = {\sf employee}$) or an interval constraint for numeric attributes (e.g. $t_1 \le{\sf time}\le t_2$).
We denote by~$C_k$ the configuration template that defines local policies with $k$ clauses, each consisting of $k$ terms.
Note that the local policies defined in the template $C_k$ may refer to at most $k^2$ attributes.
Our synthesizer implements the following procedure:
it iteratively executes ${\cal S}_{\sf smt}(S, R, C_1)$, ${\cal S}_{\sf smt}(S, R, C_2)$, ${\cal S}_{\sf smt}(S, R, C_3), \ldots$, stopping with the first call to ${\cal S}_{\sf smt}$ that returns a satisfying configuration, and returning this configuration.
By iterating over templates increasing in size, our synthesizer generates small local policies, which is desirable for avoiding redundant attribute checks.
For the running example, for instance, our synthesizer's output includes the constraint $\operatorname{\sf correct-pin}$ only for the entrance gates' local policies, and does not include this check, e.g., for the office room's policy.
A satisfying solution for each case study can be found in the configuration template~$C_3$. This indicates that real-world local policies have concise representations.
Note that our synthesizer may not terminate in a reasonable amount of time if no configuration satisfies the global requirements for the given resource structure.
In our case studies, we used a simple iterative method to pinpoint such unsatisfiable requirements: we start with a singleton set of requirements, consisting of one satisfiable requirement, and iteratively extend this set by one requirement. This helped us identify a minimal set of conflicting requirements and revise problematic ones.
\subsection{Case Studies} \label{case-studies-sec}
To investigate $\mathcal{S}_\mathsf{smt}$'s efficiency and scalability, we have conducted case studies in collaboration with KABA~AG. We used real-world requirements and resource structures, and used our tool to synthesize policy configurations for a university building, a corporate building, and an airport terminal.
Our synthesizer and all data are publicly available\footnote{\url{http://www.infsec.ethz.ch/research/software/spctl.html}}. Below, we briefly explain
the three case studies; relevant complexity metrics are summarized in Table~\ref{tab:metrics}.
\para{University Building}
We modeled the main floor of ETH~Zurich's computer science building. This floor consists of $66$ subspaces including labs, offices, meeting rooms, and shared areas.
The subspaces are labeled with four attributes that indicate: the research group to which a physical space is assigned, the physical space type (e.g., office, teaching room, or server room), the room number, and whether the physical spaces belongs to a secretary or a faculty member. Example requirements stipulate that a research group's PhD~students can access all offices assigned to the group except those assigned to the faculty members and secretaries. The policies are defined over eight attributes.
\para{Corporate Building}
We modeled an office space that consists of $20$ subspaces, including
a lobby, meeting rooms, offices, and restricted
areas such as a server room, a mail room, and an HR office. The rooms
are connected by three corridors,
and they are labeled with attributes to mark public areas and employee-only zones.
Access to these spaces is controlled by locks that are equipped with smartcard readers and PIN keypads. These locks are connected to a time server.
Example requirements are that only the postman and HR employees can access the mail room, and that between noon and 1PM employees can access their offices without entering their PIN. The policies are defined over four attributes.
\para{Airport Terminal}
We modeled the main terminal of a major international airport. The part of the terminal that we modeled includes subspaces such as the boarding pass control, security, and shopping areas.
We have used the actual plan of the terminal, and considered $15$
requirements, all currently enforced by the airport's access-control
system. The area is divided into $13$ subspaces, each
labeled with zone identifiers (such as check-in and
passport control). Example requirements stipulate that no passenger can
access departure areas before passing through security, passengers with
economy boarding passes cannot pass through the business/first-class
ticket-control gates, and that only airport staff can access certain
elevators.
\begin{table}
\centering
\normalsize
\tabcolsep2pt
\begin{tabular}{l|lrrr}
\toprule
\multicolumn{2}{l}{} & \multicolumn{1}{l}{University}& \multicolumn{1}{l}{Corporate} & \multicolumn{1}{l}{Airport}\\
\multicolumn{2}{l}{} & \multicolumn{1}{l}{building}& \multicolumn{1}{l}{building}& \multicolumn{1}{l}{terminal}\\
\midrule
Complexity & Requirements& 14 & 10 & 15\\
metrics & PEPs & 127 & 41 & 32\\
& Subspaces & 66 & 20 & 13\\
\midrule
Performance & Synthesis time & 10.32 & 25.30 & 1.92\\
& Std. dev. & 0.04 & 0.15 & 0.01\\
\bottomrule
\end{tabular}
\caption{Complexity metrics and policy synthesis times (in seconds) for the three cases studies}
\label{tab:metrics}
\end{table}
\subsection{Empirical Results}
We ran all experiments on a Linux machine with a quad-code
$\operatorname{i7-4770}$ CPU, $32$GB of RAM, running Z3 SMT~v$4.4.0$.
We present two sets of results: (1) the
synthesizer's performance when used to synthesize the local policies
for the three case studies, and
(2) the synthesizer's scalability.
\para{Performance}
We used our tool to synthesize the local policies for the three
case studies, measuring the time taken for policy
synthesis. We report the average synthesis time, measured over $10$
runs of the synthesizer, in the bottom two rows of Table~\ref{tab:metrics}. The reported
synthesis time is the sum of the time taken for encoding the policy
synthesis instance into SMT constraints, the time for
solving the generated SMT constraints, and the time for iterating over
the smaller templates for which the synthesizer returns $\sf unsat$.
In all three case studies, our tool synthesizes the local
policies in less than $30$ seconds.
The standard deviation is under $0.2$ seconds.
This indicates that synthesizing local policies is practical, and can
be used for real-world systems.
\begin{figure}
\centering
\begin{tikzpicture}
\tikzstyle{every node}=[font=\small]
\begin{axis}
[
width=0.45\textwidth,
height=0.25\textwidth,
grid=major,
grid style={dashed,gray!30},
xlabel={Number of PEPs},
legend cell align=left,
ylabel={Synthesis time (min)},
x label style={at={(axis description cs:0.5,0.05)},anchor=north},
y label style={at={(axis description cs:0.1,0.5)},anchor=south},
draw=gray!50,
line width=1pt,
xmin=0,
ymin=0,
xmax=650,
scaled y ticks = false,
scaled x ticks = false,
legend style={at={(0,1)},anchor=north west},
xticklabel = \pgfmathprintnumber{\tick},
]
\addplot+[only marks, mark=x, mark options={fill=black},color=black,solid,line width=1pt] table[x=peps,y=time] {data/scalability.txt};
\addplot+[smooth, mark=none,color=black,solid,line width=1pt] table[x=peps,y=trend] {data/scalability.txt};
\end{axis}
\end{tikzpicture}
\caption{Scaling the number of PEPs}
\label{fig:scale-res}
\end{figure}
\para{Scalability Experiments}
To investigate the scalability of our synthesis tool, we synthetically generated larger problem instances based on the corporate building case study. Although the case study originally consisted of a single floor, we increased the number of the floors in the building.
We kept the same labeling for the newly added subspaces, so the original requirements also pertain to the newly added floors.
Based on this method, we scaled the number of PEPs up to $650$.
The time needed to synthesize local policies for different numbers of PEPs is given in Figure~\ref{fig:scale-res}. The results show that our tool can synthesize a large number of local policies in a reasonable amount of time. For example, synthesizing up to $500$ local policies takes less than five hours.
The tool's performance can be further improved using domain-specific heuristics for solving the resulting SMT constraints.
Nevertheless, the tool already scales to most real-world scenarios: protected physical spaces usually have less than $500$ PEPs.
\subsection{Generic Requirements}
\label{generic-requirements-formalization}
In this section we illustrate the formalization of two generic requirements: deadlock-freeness and deny-by-default.
\para{Deadlock-freeness}
Given a resource structure, we can formalize a global requirement stipulating that there are no deadlock resources.
Formally, let ~$S = ({\cal R}, E, r, L)$ be a resource structure. The deadlock-freeness requirement for $S$ is:
\[ {\sf true}\Rightarrow {\sf AG}\ {\sf EX}\ {\sf true}~. \]
This requirement applies to all access requests. The access constraint ${\sf AG}\ {\sf EX}\ {\sf true}$ states that for any resource a subject can access, there is a resource that the subject can access next.
A resource structure~$S$ and a configuration $c$ satisfy this requirement iff for any access requests $q\in {\cal Q}$, $S_{c,q}$ has no deadlock resources.
\para{Deny-by-default}
We first define positive and negative requirements. Let $S$ be a resource structure.
We write $C_S$ for the set of all configurations for~$S$.
The set~$C_S$ is partially ordered under the relation $\sqsubseteq_S$, defined as:~$c\sqsubseteq_S c'$ iff $c(e)\subseteq c'(e)$, for any edge~$e$ of~$S$ and any two configurations~$c$ and~$c'$.
Let $c$ and $c'$ be two configurations for a resource structure~$S$.
A requirement $T\Rightarrow \varphi$ is \emph{positive} iff $S,c\Vdash (T\Rightarrow \varphi)$ and $c \sqsubseteq c'$ imply $S,c'\Vdash (T\Rightarrow \varphi)$.
A requirement $T\Rightarrow \varphi$ is \emph{negative} iff $S,c\Vdash (T\Rightarrow \varphi)$ and $c' \sqsubseteq c$ imply $S,c'\Vdash (T\Rightarrow \varphi)$.
Intuitively, if a configuration satisfies a positive (negative) requirement, then any more (less) permissive configuration also satisfies the requirement. Although not all requirements are positive or negative, most real-world requirements are, including the requirements specified using our requirement patterns.
The deny-by-default requirement is defined relative to a set of positive requirements. Let $S = ({\cal R}, E, r, L)$ be a resource structure, $R$ a set of requirements for $S$ that contains only
positive and negative requirements, and let $\{T_1\Rightarrow \varphi), \cdots, T_n\Rightarrow \varphi_n\}$ be the set of all positive requirements contained in $R$.
The deny-by-default requirement for $R$ is
\[ \neg (T_1\vee \cdots \vee T_n)\Rightarrow \textsc{Deny}\ ({\sf id} \neq r)~. \]
Here we assume that $L(r)({\sf id}) = r$, i.e. the entry resource $r$ is labeled with its identifier $r$. Note that we use the deny pattern to formalize deny-by-default.
This requirement states that if none of the positive requirements are applicable to a request, then
only the entry space is accessible to the subject who makes such a request.
Adding deny-by-default to our running example's requirements would ensure that an intern cannot access, e.g., the meeting room.
\section{Introduction}
\label{sec:intro}
Physical access control is used to restrict access to physical spaces. For example, it controls who can access which parts of an office building or how personnel can move within critical spaces such as airports or military facilities. As physical spaces are usually comprised of subspaces, such as rooms connected by doors, policies are enforced by multiple policy enforcement points (PEPs). Each PEP is associated to a control point, like a door, and enforces a \emph{local} policy.
Consider, for example, an office building. An electronic door lock might control access to an office by enforcing a policy that states that only an \emph{employee} may enter the office. This policy is local in the sense that its scope is limited to an individual enforcement point, here the office's door. The policy therefore does not guarantee that non-employees cannot enter the office, since the office may have other doors. Neither does it guarantee that employees can actually access the office. If employees cannot enter the corridor leading to the office's door, then the local policy is useless.
In contrast to the local policies for enforcement points,
access-control requirements for physical
spaces are typically \emph{global}. They express
constraints on the access paths through the entire space.
In the example above, a requirement might be
that employees should be able to access the office
from the lobby.
This requirement is global in that no single PEP alone
can guarantee its satisfaction. A standard electronic
lock, which enforces only local policies such as \emph{grant
access to employees}, is oblivious to the physical constraints of the
office building and what policies the other PEPs enforce. It therefore
cannot address this requirement.
\para{Problem Statement}
The discrepancy between global requirements and local policies creates an abstraction gap that must be bridged when configuring access-control mechanisms.
We consider the problem of automatically synthesizing a set of PEP policies that together enforce global access-control requirements in a given physical space.
This problem is nontrivial.
A given physical space
usually constrains the ways
subjects may access its subspaces. These constraints must be accounted for
when configuring the individual PEPs.
Moreover, global access-control requirements may have interdependencies
and hence their individual solutions may not contribute to an overall
solution.
To illustrate this lack of compositionality, suppose in addition to the requirement that
employees can access an office room from the lobby, we require that
they must not enter the area where auditing documents are stored.
Giving employees access to their office through \emph{any} path
satisfies the first requirement, but it would violate the second one if
the path goes through the audit area.
In practice, constructing local policies for a physical space
is a manual task where a security engineer writes individual policies, one per PEP,
that collectively enforce the space's global requirements.
This manual process results in errors, such as
granting access to unauthorized subjects or denying access to
authorized ones; the literature contains numerous examples
of such
problems~\cite{Fitzgerald_anomalyanalysis,vmcai11,prediction:codaspy12}.
Moreover, engineers must manually revise their policies whenever
requirements are changed, or when the physical space changes,
e.g.\ due to construction work.
In short, writing local policies manually is error-prone and
scales poorly.
Our thesis is that it is also unnecessary: the automatic synthesis of
local policies with system-wide security guarantees is a viable
alternative.
\input{intro-figure}
\para{Approach and Contributions} We propose a formal framework for automatically
synthesizing local policies that run on distributed PEPs
from a set of global access-control
requirements for a given physical space.
The framework's main ingredients are depicted in
Figure~\ref{fig:intro}.
The key component is a \emph{synthesizer}, which takes as input a
model of the physical space and a set of global requirements.
The synthesizer's output is the set of local policies that the PEPs enforce.
If the global requirements are satisfiable, then the synthesizer is guaranteed to output a correct set of local policies; otherwise, it returns ${\sf unsat}$ to indicate that the requirements cannot be satisfied.
Hence, using our framework, engineers can generate local policies from global requirements simply by formalizing the global requirements and modeling the physical space.
Below, we briefly describe the framework's components, depicted in
Figure~\ref{fig:intro}. We use directed graphs to model physical
spaces: a node represents an enclosed space, such as an office or a
corridor, and an edge represents a PEP, for example installed on a
door or turnstile. The nodes are labeled to denote their
attributes. These attributes may include the assets the node contains
(audit documents), its physical attributes (international terminal),
and its clearance level (high security zone). These attributes may be
used when specifying policies.
Formally, our model of a physical space is a Kripke structure.
We give a declarative language, called \lang, for specifying global requirements.
Our language is built on the computation tree logic
(CTL)~\cite{Emerson1982241} and supports subject attributes (e.g., an organizational
role), time constraints (e.g., business-hour requirements), as well
as quantification over paths and branches in physical spaces.
To demonstrate its expressiveness, we show how
common physical access-control requirements can be directly written in \lang.
Moreover, to simplify the task of formalizing such requirements, we
develop requirement patterns and illustrate their use through
examples.
Our synthesis algorithm outputs
attribute-based policies,
expressed as constraints over subject attributes and contextual
conditions, such as organizational roles and the current time. This
covers a wide range of practical setups and scenarios, including
attribute-based and
role-based access control.
We strike a balance between the requirement language's expressiveness
and the complexity of synthesizing local policies.
The synthesis problem we consider is NP-hard.
However, we show that
for practically-relevant requirements, it can
be efficiently solved using existing SMT solvers.
This is intuitively because physical spaces, in practice,
induce directed graphs that have
short simple-paths. We illustrate our framework's effectiveness
using three case studies where we synthesize
access-control policies for a university building, a corporate
building, and an airport terminal. Synthesizing local
policies in each case takes less than $30$ seconds.
The last two case studies are based on real-world examples developed
together with KABA~AG, a leading physical access control company.
To the best of our knowledge, ours is the first framework for
synthesizing policies from system-wide access-control requirements.
We thereby solve a fundamental problem in access control for physical spaces.
An immediate practical consequence is that security engineers can
focus on
system-wide requirements, and delegate to our synthesizer the task of
constructing the local policies with correctness guarantees.
We remark that although this work is focused on access control for
physical spaces, the ideas presented are general and can be extended
to other domains, such as computer networks partitioned into subnetworks
by distributed firewalls.
\para{Organization}
We give an overview of our access-control synthesis framework
in~\secref{sec:overview}. In~\secref{sec:phys-spec}, we describe and
formalize our system model. In~\secref{sec:reqs}, we define our
\lang language for specifying global requirements, and
present requirement patterns.
In \secref{sec:synthesis}, we
define the policy synthesis problem and prove its decidability. In \secref{sec:algorithm}, we define an efficient policy synthesis algorithm. In \secref{sec:eval}, we describe our implementation and
report on our experiments. We review related work in \secref{sec:rw},
and we draw conclusions and discuss future work in~\secref{sec:conc}.
\ifext
The appendices contain all proofs.
\else
Proofs can be found in the extended version of this paper~\cite{tech-report}.
\fi
\section{Overview} \label{sec:overview}
We start with a simple example that illustrates the challenges
of constructing local policies that cumulatively enforce global
access-control requirements. We also explain how our framework is
used, that is, we describe its inputs and outputs.
\input{synthesis-overview}
\subsection{Running Example} \label{sec:example} Consider a small
office space consisting of a lobby, a bureau, a meeting room, and a
corridor. The office layout is given in
Figure~\ref{fig:synth-overview}(a).
Access within this physical space is secured using electronic locks.
Each door has a lock and a card reader. The lock stores a policy
that defines who can open the door from the card reader's side. The
door can be opened by anyone from the opposite side. We annotate
locks with arrows
in Figure~\ref{fig:synth-overview}(a) to indicate the direction that the locks
restrict access. For example, the lock at the main entrance restricts who can
access the lobby, and it allows anyone to exit the office space from
the lobby.
To open a door from the card reader's side, a subject presents a
smartcard that stores the holder's credentials. The lock can access
additional information, such as the current time, needed to evaluate the
policy. The lock opens whenever the policy evaluates to grant.
The global requirements for this physical space are given in
Figure~\ref{fig:synth-overview}(b). The requirements \textbf{R1},
\textbf{R3}, and {\bf R4} define permissions, while \textbf{R2} and
\textbf{R5} define prohibitions. To meet these requirements, the
electronic locks must be configured with appropriate local policies.
As previously observed, this is challenging because
one must account for both spatial constraints
and all global access-control requirements.
We illustrate these points below.
\para{Spatial Constraints} The layout of the physical space
prevents subjects from
freely requesting access to any resource.
For example, the requirement \textbf{R1} is not met just
because the meeting room's lock grants access to visitors;
the visitor must also be able to enter the corridor from the outside.
Such constraints must be accounted for when defining
the local policies.
To satisfy \textbf{R1}, we may for instance choose a path from the main
entrance to the meeting room and configure all the locks along that
path to grant access to visitors.
\para{Global Requirements}
Each global requirement typically has
multiple sets of local policies that satisfy it. The
local policies must however be constructed to ensure that
\emph{all} requirements are satisfied \emph{simultaneously}. For example, the
requirement \textbf{R1} is satisfied if the side-entrance lock and the
meeting room lock both grant access to visitors between $8$AM and $8$PM. It can also be
satisfied by ensuring that the main entrance, the lobby, and the
meeting room locks all grant access to visitors between $8$AM and $8$PM. Granting visitors
access through the side entrance however violates the requirement
\textbf{R2}, which requires that visitors pass through the lobby.
Hence, to meet both requirements, the locks along the path through
the lobby must grant access to visitors between $8$AM and $8$PM, while the side-entrance lock
must always deny access to visitors.
\subsection{Synthesis Framework}
Figure~\ref{fig:synth-overview}(c-e) depicts our framework's input and output for our running example. The input is a model of the physical space and a specification of its global requirements. The output produced by our synthesizer is a set of local policies.
A physical space is modeled as a rooted directed graph
called a \emph{resource structure}. We have depicted
the root node in gray. In our example, this
corresponds to the public space that surrounds the office space, e.g.\
public streets. The remaining (non-root) nodes are the spaces
inside the building. The locks control access along the edges.
A subject can traverse a solid edge of the resource structure
only if the lock's policy evaluates to
grant, whereas any subject can follow the dashed edges. Hence, the locks
effectively enforce the grant-all policy along the dashed edges.
We use two attributes to label the physical spaces: the
attribute~$\mathsf{id}$ represents room identifiers, and $\operatorname{\sf sec-zone}$ formalizes that a space is inside the security zone.
Global requirements are specified using a declarative language, called
\lang. In Figure~\ref{fig:synth-overview}(d) we show the formalization
of our running example's requirements in \lang.
For instance,~{\bf R1}, which states that visitors can access the
meeting room between $8$AM and $8$PM, is formalized as $\big(\mathsf{(role
= visitor)} \wedge \mathsf{(8 \le time\le 20)}\big)\Rightarrow
\textsc{Grant}\mathsf{(id = mr)}$.
This formalization instantiates \lang's permission pattern
$\textsc{Grant}$ to state that there is a path from outside to
the meeting room such that every lock on the path grants access to any
visitor between $8$AM and $8$PM.
We define \lang's syntax and semantics and present several patterns in~\secref{sec:reqs}.
Given these inputs, the synthesizer automatically constructs a local policy for each lock.
The synthesized policies are attribute-based policies that
collectively enforce the global requirements. The synthesized policies
for our running example are given in
Figure~\ref{fig:synth-overview}(e).
We write, for example, ${\sf cor}\ \pgfuseimage{policy}\ {\sf bur}$
for the synthesized policy deployed at the bureau's lock. This
policy $\sf (role = employee)$
grants access to subjects with the role $\sf employee$. We define the synthesis problem, and the syntax
and semantics of attribute-based local policies
in~\secref{sec:synthesis}.
\section{Physical Access Control}
\label{sec:phys-spec}
\subsection{Basic notions}
In this section, we formalize our system model for physical access control. Our terminology is based, in part, on the XACML reference architecture \cite{xacml3}.
Each physical space is partitioned into finitely many enclosed spaces
and one open (public) space. We call the enclosed spaces
\emph{resources}. Two spaces may be directly connected with a
\emph{gate}, controlled by a \emph{policy enforcement point} (PEP).
Examples of gates include doors, turnstiles, and security checkpoints.
Each PEP has its own \emph{policy decision point} (PDP), which stores
a \emph{local policy} mapping access requests to access decisions.
Each {\em access request} consists of subject credentials, which the
PDP receives from the PEP, as well as contextual attributes, if needed,
obtained from \emph{policy information points} (PIPs). The PIPs are
distributed information sources that provide contextual attributes
required by the PDP for access decisions. Examples of PIPs
include revocation list servers and secure time servers.
To enter a space, a subject provides his credentials to the PEP that
controls the gate. The PEP forwards the subject's credentials
to its PDP. The PDP, in turn, queries the PIP if needed, evaluates the
policy, and then forwards the access decision --- either grant or deny ---
to the PEP. The PEP then enforces the PDP's
decision. See Figure~\ref{fig:sys-model}.
We assume that access requests contain all relevant information for making access decisions.
PDPs can thus make their decisions independent of past access requests. Hence it has no bearing on our model whether the PDPs are actually distributed or are realized through a centralized system.
This is desirable from a practical standpoint since PDPs and PIPs need not be equipped with logging mechanisms. Moreover, different PDPs and PIPs need not synchronize their local views on the request history. In this sense they are autonomous entities.
The access requests and local policies we consider are attribute
based and may reference three kinds of attributes.
A \emph{subject attribute} contains information about a subject. For
example, Alice's \emph{organizational role} and \emph{clearance level}
are her subject attributes. Subjects can provide PEPs with their
attributes in the form of credentials.
A {\em contextual attribute} represents information about the security
context provided by a PIP, such as the list of revoked credentials and the current time.
We also introduce {\em resource attributes}, which represent
information about resources. For example, the attributes \emph{floor}
and \emph{department} may represent the floor of an office space and
the department it belongs to. We use resource attributes to specify
global requirements. They are however not needed for expressing access
requests or local policies in our model.
This is because the PDPs associated to any resource can be hardwired
with all the attributes of that resource. In this sense, each PDP
``knows'' the space under its control.
Our system model targets electronic PEP/PDPs
that can
enforce attribute-based policies and read digital credentials,
e.g.\ stored on smart cards and mobile phones.
Manufacturers often refer to these as smart locks~\cite{goji,augustus,lockitron}.
In large physical access-control systems, smart locks are
rapidly replacing mechanical locks and keys, which can only enforce
simple, crude policies.
\input{sys-model}
\subsection{Formalization}
We now formalize the above notions.
\para{Attributes}
Fix a finite set~$\cal A$ of \emph{attributes}
and a set~$\cal V$ of attribute {\em values}.
The {\em domain} function ${\sf dom}\colon {\cal A}\to {\cal P}({\cal
V})$ associates each attribute with the set of values it admits.
For instance, the current time attribute is associated with the set of
natural numbers, and the clearance level attribute is associated with
a fixed finite set of levels.
We assume that any attribute can take the designated
value~$\bot$, representing the situation where the
attribute's value is unknown.
We partition the set of attributes into subject attributes~${\cal
A}_S$, contextual attributes~${\cal A}_C$, and resource
attributes~${\cal A}_R$.
\para{Access Requests} We represent an access request as a total function that
maps subject and contextual attributes to values from their respective
domains.
This function is computed by PDPs after receiving a subject's
credentials and querying PIPs.
For instance, the PDP maps the attribute~$\mathsf{role}$
to~$\mathsf{visitor}$ when the subject's credentials indicate
this. It maps the attribute~$\operatorname{\sf correct-pin}$ to
$\mathsf{true}$ when the PIN entered through the keypad
attached to the PDP is correct. Finally, it maps the contextual
attribute~$\mathsf{time}$ to~$8$ after querying a time server at $8$AM.
We denote the set of all access requests by~${\cal Q}$.
A remark on set-valued attributes is due here. In some settings, attributes take a finite set of values, as opposed to a single value.
For example, in role-based access control, a subject may activate multiple roles.
The attribute $\sf role$ must then be assigned with the set of all the activated roles.
We account for such set-valued attributes simply by defining a Boolean attribute for each value; for example, we define $\sf role\_employee$ and $\sf role\_manager$. An access request~$q$
assigns true to both Boolean attributes whenever a subject has activated both the employee and the manager roles.
\para{Local Policies}
Local policies map access requests to grant or deny. We extensionally define local policies as subsets of~$\mathcal{Q}$: a local policy is defined as the set of requests that it grants.
The structure $({\cal P}({\cal Q}), \subseteq, \cap, \cup, \emptyset, {\cal Q})$ is a complete lattice that orders local policies by their permissiveness. The least permissive policy, namely~$\emptyset$, denies all access requests, and the most permissive one, i.e.~$\mathcal{Q}$, grants them all.
In section~\ref{sec:problem}, we will intensionally define local policies as constraints over subject and contextual attributes.
The local policies shown in Figure~\ref{fig:synth-overview}, for example, are defined by such constraints.
\para{Resource Structures}
We now give a formal model of physical spaces.
A \emph{resource structure} is a tuple~$S = ({\cal R}, E, r_e, L)$, where ${\cal R}$ is a set of {\em resources}, $E\subseteq {\cal R}\times {\cal R}$ is an irreflexive edge relation, $r_e\in {\cal R}$ is the \emph{entry resource}, and $L: {\cal R}\to ({\cal A}_R\to {\cal V})$ is a total function mapping resources to resource attribute valuations.
We assume that every resource~$r\in \mathcal{R}$ is reachable from the entry resource~$r_e$, that is,~$(r_e,r)\in E^*$, where~$E^*$ is the reflexive-transitive closure of~$E$.
The edges in a resource structure model PEPs. The irreflexivity of
$E$ captures the condition that once a subject enters a physical
space, he cannot re-enter the space before first leaving it.
We assume that resource structures do not contain \emph{deadlocks}. A
resource~$r_0$ in~$S$ is a deadlock if there does not exists an~$r_1$
such that~$(r_0,r_1)\in E$. This assumption is valid in physical-space
access control: a deadlock resource corresponds to a ``black hole''
that no one can leave. Note that dead-end corridors are not deadlocks,
provided one can backtrack.
The entry resource~$r_e$ represents the public space and the remaining
resources denote enclosed spaces.
A resource structure describes how subjects can access resources.
A subject accesses a resource along a path, which is a sequence
of resources connected by edges, starting from the entry resource.
For example, before entering a room in a hotel, a subject enters the
hotel's lobby from the street, and then goes through the corridor.
Figure~\ref{fig:synth-overview}(c) gives an example of a resource
structure.
\para{Configurations}
Each edge of a resource structure represents a gate controlled
by a local policy installed on the gate's PDP. We therefore
define a \emph{configuration} for a resource structure~$S$ as a
function that assigns to each edge of~$S$ a local policy.
We write $C_S$ for the set of all configurations for~$S$.
The set~$C_S$ is partially ordered under the relation $\sqsubseteq_S$, defined as:~$c\sqsubseteq_S~\!\!c'$ if for any edge~$e$ of~$S$ we have $c(e)\subseteq c'(e)$.
Namely, a configuration is less permissive than another configuration if for any edge the former assigns a less permissive local policy than the latter.
\input{restricted-example}
We can now define which resources are accessible given an access
request and a configuration.
For a resource structure~$S$, a configuration~$c$ for~$S$, and an
access request~$q$, we define~$S_{c,q}$ as the resource structure
obtained by removing all the edges from~$S$ whose policies deny~$q$,
and then removing all nodes that are not reachable from the entry
resource.
The structure~$S_{c,q}$'s entry resource is the same as~$S$'s.
To illustrate, consider the resource structure $S$ and the configuration
$c$ given in Figure~\ref{fig:synth-overview}, and the access request
$q = \{{\sf role} \mapsto {\sf visitor}, {\sf time}\mapsto 10,
\operatorname{\sf
correct-pin}\mapsto \bot\}$. The
side-entrance PEP and the bureau PEP deny $q$ and therefore these two
edges are removed from~$S$. The node that represents the bureau is not reachable from the entry resource and it is thus also removed. In Figure~\ref{fig:restricted-example} we
depict the removed edges and nodes.
We remark that the structure~$S_{c,q}$ is defined for a fixed access request~$q$. Access requests, which assign values to subject and contextual attributes, can however change, for instance when a subject's role is revoked or as time progresses.
We abstract away such changes in~$S_{c,q}$'s definition. In our running example, this amounts to assuming that a subject's role does not change during this time, and the time needed to move through the office building is negligible compared to the time needed for a subject's access rights to change; for example, the requirements {\bf R1-5} stipulate that subject's access rights may change only twice per day --- at $8$AM and at $8$PM.
This abstraction corresponds to taking a snapshot of all the attributes, and then computing~$S_{c,q}$ based on the snapshot.
We refer to these snapshots as \emph{sessions}.
Henceforth we interpret global requirements and local policies in the context of such sessions.
Interpreting requirements and polices in the context of a session is justified for many practical scenarios. This is because, in most practical settings, changes in subject and contextual attributes are addressed through out-of-band mechanisms.
To illustrate, consider a subject who has the role visitor and enters the meeting room of our running example at $3$PM as permitted by the system's requirements. Now, suppose that the subject's visitor role is revoked at $4$PM, or that the subject remains in the meeting room until $10$PM. No access-control system can force the subject to leave. In practice, out-of-band mechanisms, such as security guards, address such concerns.
In the following sections, we confine our attention to configurations that do not introduce deadlocks. That is, we consider those configurations~$c$ where for any~$q\in\mathcal{Q}$, the structure~$S_{c,q}$ is deadlock-free. In Section~\ref{sec:gen-req}, we describe how this provision can be encoded as a global requirement.
\section{Policy Synthesis Problem}
\label{sec:problem}
We first describe our system model, and then formalize it. Afterward,
we formally define the policy synthesis problem.
\subsection{Formal Setting}
Below we formalize our system model, including attributes, resource
structures, policies, and requirements. In our formal model,
we do not distinguish between physical spaces and protected resources.
This modeling choice avoids unnecessary cluttering, without limiting
the reach and scope of our results.
\para{Attributes}
We fix a finite set~$\cal A$ of \emph{attributes} and an infinite
set~$\cal V$ of attribute {\em values}. Each attribute $a\in {\cal A}$
is associated with a (possibly infinite) set of values. We partition
the set of attributes into subject attributes~${\cal A}_S$, contextual
attributes~${\cal A}_C$, and resource attributes~${\cal A}_R$. An
{\em access request} is a total function mapping subject and
contextual attributes to values, and we denote the set of all access
requests by~${\cal Q}$. A {\em resource valuation} is a total function
that maps resource attributes to their values.
We refer to the pair~$(\mathcal{A},\mathcal{V})$ as the attribute
signature.
Two remarks are due here. First, in contrast to, e.g.,
XACML~\cite{xacml3}, an
access request in our setting does not refer to the resource
attributes.
This is because in our systems model, each PDP controls access to one
physical space. Since, as mentioned above, a PDP ``knows'' the space
under its control, there is no need to include resource attributes in
access requests.
Second, we assume that subject attributes and contextual attributes do
not change in the course of making a policy decision. We relegate
discussing such attribute changes to~\secref{}.
\para{Resource Structures}
A \emph{resource structure} is a tuple~$S = ({\cal R}, E, I, L)$,
where ${\cal R}$ is a set of {\em resources}, $E\subseteq {\cal
R}\times {\cal R}$ is a reflexive relation, $I\subseteq {\cal R}$ is a
set of \emph{entry resources}, and $L: {\cal R}\to ({\cal A}_R\to
{\cal V})$ is a total function mapping resources to resource
valuations.
An example of a resource structure is given in
Figure~\ref{fig:synth-overview}. For simplicity we omit the self
edges, which connect a node to itself, when depicting resource
structures.
Any edge represents an enforcement point.
A resource structure constrains how subjects access resources: A
subject can directly access only the entry resources, while all the
remaining resources are accessed along paths starting from an entry
resource.
After a subject has accessed a resource, then the subject may access
the resource indefinitely. For example, a subject who accesses a room
can remain there indefinitely. This motivates the self edges, which we
assume permit all requests.
\subsection{Correctness and Complexity of the Algorithm~${\cal S}_{\sf cs}$}
\label{sec:decidability}
\para{Correctness}
We now prove the correctness of~${\cal S}_{\sf cs}$.
\begin{theorem}
Let $S$ be a resource structure and $R$ a set of requirements.
If ${\cal S}_{\sf cs}(S, R) = c$ then $S,c\Vdash R$.
If ${\cal S}_{\sf cs}(S, R) = {\sf unsat}$ then there is no configuration $c$ such that $S, c\Vdash R$.
\label{thm:decompose}
\end{theorem}
\begin{proof}
We prove the two implications by contradiction.
Assume that ${\cal S}_{\sf cs}(S, R)$ returns a configuration~$c$.
Suppose for the sake of contradiction that $S,c\not\Vdash R$.
Then, by definition of $\Vdash$, there is an access request $q$ and a requirement $T\Rightarrow \varphi$ such that $q\vdash T$ and $S_{c,q}\not\models \varphi$.
Given a subset $R'\subseteq R$ of the requirements, let $T_{R'} = T_1\wedge ...\wedge T_i \wedge \neg T_{i+1} \wedge \cdots \wedge \neg T_n$, where $\{T_1\Rightarrow \varphi_1, \ldots, T_i\Rightarrow \varphi_i\} = R'$ and $\{T_{i+1}\Rightarrow \varphi_{i+1}, \ldots, T_n\Rightarrow \varphi_n\} = R\setminus R'$.
The constraint $T_{R'}$ corresponds to the target computed at line~\ref{alg:cs:target} of Algorithm~\ref{alg:cs}.
Let $R_q = \{(T\Rightarrow \varphi)\in R\mid q\vdash T\}$ be the set of all requirements in $R$ that are applicable to~$q$.
We have $q\vdash T_{R_q}$ (1). Furthermore, for any $R'\subseteq R$ where $R'\neq R_q$, we have $q\not\vdash T_{R'}$ (2).
By definition of $S_{c,q}$, $S_{c,q}$ contains an edge $e$ if $e$ is an edge of $S$ and $q\vdash c(e)$. Algorithm~\ref{alg:cs} constructs the configuration $c$ by conjoining targets $T_{R'}$, where $R'\subseteq {\cal R}$, to the local policies $c(e)$; see line~\ref{alg:cs:conf}. From (1) and (2) we conclude the following:
First, adding $\neg T_{R_q}$ to a local policy $c(e)$ removes the edge $e$ in $S_{c,q}$ because $q\not\vdash c(e)\wedge (\neg T_{R_q})$. Second, adding $\neg T_{R'}$ to a local policy $c(e)$, where $R'\neq R_q$, does not remove the edge $e$ in $S_{c,q}$ because $q\vdash c(e) \wedge (\neg T_{R'})$ iff $q\vdash c(e)$.
It is immediate that $S_{c,q}$ contains those edges of $S$ for which the target $\neg T_{R_q}$ is not conjoined to the local policy $c(e)$. We conclude that $S_{c,q}$ contains the edges $E' = {\sf cs}(S, \varphi_{R_q})$ (see line~\ref{alg:cs:edges} of Algorithm~\ref{alg:cs}), where $\varphi_{R_q}$ conjoins the access constraints of all requirements in $R_q$.
By definition of controller synthesis, we have $(S, E', r, L)\models \varphi_{R_q}$. Since $S_{c,q} = ({\cal R}, E', r, L)$, $S_{c,q}\models \varphi_{R_q}$.
We can now deduce that $S_{c,q}\models \varphi$ because $(T\Rightarrow\varphi)\in R_q$. But previously we deduced that $S_{c,q}\not\models \varphi$. Thus we have a contradiction, and there is no access request $q$ and requirement $T\Rightarrow \varphi$ such that $q\vdash T$ and $S_{c,q}\not\models \varphi$. Therefore, $S, c\Vdash R$.
Assume that ${\cal S}_{\sf cs}(S, R) = {\sf unsat}$. Suppose for the sake of contradiction that there is a configuration $c$ such that $S, c\Vdash R$.
From ${\cal S}_{\sf cs}(S, R) = {\sf unsat}$, by definition of Algorithm~\ref{alg:cs}, it follows that there is a subset $R' = \{T_1\Rightarrow \varphi_1, \ldots, T_k\Rightarrow \varphi_k\}\subseteq R$ of the requirements and an access request $q$, such that $q\vdash T_1\wedge \cdots\wedge T_k$ (1) and ${\sf cs}(S, \varphi_1\wedge \cdots\wedge \varphi_k) = {\sf unsat}$ (2).
From (1), we know that all requirements in $R'$ are applicable to $q$ .
Furthermore, since $S, c\Vdash R$, it must be that $S_{c,q}\models \varphi_i$, for $1\leq i\leq k$. We get $S_{c,q}\models \varphi_1\wedge \cdots\wedge \varphi_k$.
From (2), by definition of controller synthesis, there is no resource structure $S' = ({\cal R}, E', r, L)$, with $E'\subseteq E$, such that $S'\models \varphi_1\wedge \cdots\wedge \varphi_k$. Thus we have a contradiction, and we conclude that there is no configuration $c$ such that $S,c\Vdash R$.
This concludes our proof.
\end{proof}
\para{Complexity}
The running time of algorithm ${\cal S}_{\sf cs}$ is determined by the
number of iterations of the loops, the complexity of checking
the satisfiability of the conjunction of targets (line~\ref{line:cs-sat}),
and the complexity of solving each controller synthesis instance
(line~\ref{line:cs-synth}).
The first loop is executed $|E|$ times, where $|E|$ is the number of
edges, and the second loop is executed $2^{|R|}$ times. The second
loop checks one satisfiability instance and one controller synthesis
instance. The complexity of checking satisfiability is in ${\cal
O}(2^{k\cdot |A|})$ where $k = |D_{\sf max}|$ for the largest set
$D_{\sf max}$ of values that appears in the constraint $T$, and $|A|$ is the
number of attributes.
Solving a controller synthesis instance requires checking $({\cal R}, E', r_e, L) \models \varphi$ at most $2^{|E|}$ times, where $E'\subseteq E$ and $\varphi$ is a conjunction of access constraints.
The problem $({\cal R}, E', r_e, L) \models \varphi$ can be decided using the model checking algorithm for CTL based on labeling, which is in ${\cal O}( |\varphi| \cdot (|{\cal R}|+ |E'|))$, where $|\varphi|$ is the size of the access constraint $\varphi$~\cite{huth11}. The size of the largest access constraint given as input to $\sf cs$ is in ${\cal O}(|d|\cdot |R|)$, where $d$ is the largest access constraint that appears in the requirements $R$.
The running time of ${\cal S}_{\sf cs}$ is therefore ${\cal O}(2^{|R|}\cdot (2^{k\cdot |A|} + 2^{|E|}\cdot |R|\cdot d \cdot (|{\cal R}|+ |E|)))$.
\subsection{SMT Encoding}
\label{sec:soundness}
\para{Example}
We illustrate the SMT encoding of an exists-until and an always-until access constraint in Figure~\ref{fig:until-constraints}. The SMT encoding of the access constraint ${\sf E}[(\neg {\sf sec\_zone}){\sf U}({\sf id}={\sf bur})]$ for the resource~$\sf out$ formalizes that the two PEPs $({\sf out, cor})$ and $({\sf cor, bur})$ grant access or the PEP $({\sf out, bur})$ grants access. This guarantees the existence of a path that satisfies the access constraint.
The SMT encoding of ${\sf A}[(\neg {\sf sec\_zone}){\sf U}({\sf id}={\sf bur})]$ for the resource $\sf out$ formalizes that the always-until constraint is satisfied along any path that starts from the resource $\sf out$. Since any path that start with $\sf (out, bur, \ldots)$ satisfies the access constraint, the SMT constraint imposes no constrains on the PEP $(\sf out, bur)$. However, not all paths that start with $(\sf out, cor, \ldots)$ satisfy the access constraint. Concretely, the infinite path $(\sf out, cor, out, cor, \ldots)$ violates the access constraint. The SMT constraint therefore formalizes that if there are paths starting with $(\sf out, cor, \ldots)$, i.e. the PEP $(\sf out, cor)$ grants access, then the PEP $\sf (cor, out)$ denies access. This guarantees that the path violating the access constraint is not present in the resulting resource structure. Note that, since we consider only deadlock-free resource structures, the absence of the edge $\sf (cor, out)$ guarantees that the resulting resource structure has the edge $\sf (cor, bur)$, and therefore all paths starting with $(\sf out, cor, \ldots)$ continue along resource $\sf bur$.
\input{smt-encoding-example}
\para{Termination}
We first prove that the rewrite rules given in Figure~\ref{fig:encode} terminate.
\begin{theorem}
Let $S = ({\cal R}, E, r, L)$ be a resource structure.
For any resource~$r_0\in {\cal R}$ and access constraint $\varphi$, the rewrite function $\tau(\varphi, r_0)$ terminates.
\label{lemma:termination}
\end{theorem}
\begin{proof}
The proof proceeds by induction on the length of the access constrain~$\varphi$. Formally,
we define the {\em length} of an access constraint $\varphi$, denoted by $l(\varphi)$, as
\[
\begin{array}{rcl}
l({\sf true}) & = & 1\\
l(a\in D) & = & 1\\
l(\neg \varphi) & = & 1 + l(\varphi)\\
l({\sf EX}\varphi) & = & 1 + l(\varphi)\\
l({\sf AX}\varphi) & = & 1 + l(\varphi)\\
l(\varphi_1\wedge \varphi_2) & = & 1 + {\sf max}(l(\varphi_1), l(\varphi_2))\\
l({\sf E}[\varphi_1 {\sf U}\varphi_2]) & = & 1 + {\sf max}(l(\varphi_1), l(\varphi_2))\\
l({\sf A}[\varphi_1 {\sf U}\varphi_2]) & = & 1 + {\sf max}(l(\varphi_1), l(\varphi_2))
\end{array}
\]
where ${\sf max}(n_1, n_2)$ returns $n_1$ if $n_1\geq n_2$, otherwise it returns~$n_2$. Note that $l(\varphi)\geq 1$ for any access constraint~$\varphi$.
\para{Base Case}
For the base case, $l(\varphi) = 1$, the access constraint is of the form $\sf true$ or $a\in D$. The rewrite function $\tau$ terminates in one step.
\para{Inductive Step}
Assume that $\tau(\varphi, r_0)$ terminates for any access constraint $\varphi$ of length $l(\varphi)\leq k$~(H1). We prove that $\tau(\varphi, r_0)$ terminates for any access constraint of length $l(\varphi) = k+1$.
\begin{compactitem}
\item For the cases where the access constraint~$\varphi$ is of the form $\neg \varphi_1$, ${\sf EX}\varphi_1$, ${\sf AX}\varphi_1$, the rewrite function $\tau(\varphi, r_0)$ calls $\tau(\varphi_1, r_0)$. By induction, $\tau(\varphi_1, r_0)$ terminates because $l(\varphi_1) = k$.
\item The case where $\varphi = \varphi_1\wedge \varphi_2$ also terminates because $l(\varphi_1)\leq k$ and $l(\varphi_2)\leq k$.
\item For the cases where $\varphi$ is of the form ${\sf E}[\varphi_1 {\sf U} \varphi_2]$ or ${\sf A}[\varphi_1 {\sf U} \varphi_2]$, we need to show that $\tau_{\sf U}(\varphi, r_0, \emptyset)$ terminates. We prove that $\tau_{\sf U}(\varphi, r_0, X)$ terminates for any set $X\subseteq {\cal R}$ by descending induction on the size of the set $X$.
For the base case, we have $|X| = |{\cal R}|$. Then, $\tau_{\sf U}(\varphi, r_0, {\cal R})$
calls $\tau(\varphi_1, r_0)$ and $\tau(\varphi_2, r_0)$. By our inductive hypothesis~(H1), both $\tau(\varphi_1, r_0)$ and $\tau(\varphi_2, r_0)$ terminate since $l(\varphi_1)\leq k$ and $l(\varphi_2)\leq k$.
For the inductive step, assume that $\tau_{\sf U}({\sf E}[\varphi_1 {\sf U} \varphi_2]), r_0, X)$ terminates for any $X\subseteq {\cal R}$ of size $k\leq |X|\leq |{\cal R}|$~(H2). Consider a set $X'\subseteq {\cal R}$ of size $|X'| = k-1$. Then, $\tau_{\sf U}(\varphi, r_0, X)$ calls the rewrite functions $\tau(\varphi_1, r_0)$, $\tau(\varphi_2, r_0)$, and $\tau_{\sf U}(\varphi, r_1, X'\cup\{r_0\})$, for $r_1\in E(r_0)\setminus X$. The rewrite function $\tau(\varphi_1, r_0)$, $\tau(\varphi_2, r_0)$ terminate by the inductive hypothesis (H1). By the inductive hypothesis~(H2), the rewrite function $\tau_{\sf U}(\varphi, r_1, X'\cup\{r_0\})$ terminates because $|X\cup \{r_0\}| = k$.
\end{compactitem}
This completes our proof.
\end{proof}
\para{Correctness}
We now prove that the correctness of our SMT-based policy synthesis algorithm.
We start with several definitions. Our definitions are similar to those used to describe the decision procedure for CTL satisfiability given in~\cite{Emerson:1991:TML:114891.114907}.
Let $\varphi_1$ and $\varphi_2$ be two access constraints and $S = ({\cal R}, E, r, L)$ be a resource structure.
We assume that $S$ does not contain deadlock resources, i.e.\ for any resource $r_0\in {\cal R}$, the set $E(r_0) = \{r_1\in {\cal R}\mid (r_0, r_1)\in E\}$ is nonempty.
We call access constraints of the form ${\sf A}[\varphi_1{\sf U}\varphi_2]$ and ${\sf E}[\varphi_1{\sf U}\varphi_2]$ eventuality constraints.
We first define the derivation of a rooted directed graph from $S$ for a given access constraint~$\varphi_2$ and root node~$r_0\in {\cal R}$. We call this graph an {\em eventuality graph}. We then give two conditions
over such eventuality graphs. The first condition is satisfied iff $S, r_0\models {\sf A}[\varphi_1{\sf U}\varphi_2]$, while the second one is satisfied iff $S, r_0\models {\sf E}[\varphi_1{\sf U}\varphi_2]$.
We define the eventuality graph $G(S, r_0, \varphi_2)$ as the rooted directed graph obtained by taking the node $r_0$ and all nodes and edges along all paths emanating from $r_0$ up to and including the first node $r_1$ such that $S, r_1\models \varphi_2$; if there is no such node $r_1$ along a path, then all nodes and edges along the path are included in $G(S, r_0, \varphi_2)$.
We call a node of $G(S, r_0, \varphi_2)$ an {\em interior node} if it has successors; otherwise, we call it a {\em frontier node}.
We now define the two conditions.
We say that an eventuality graph $G(S, r_0, \varphi_2)$ {\em fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$} if
\begin{compactenum}
\item the graph is acyclic,
\item for any of its interior nodes $r_1$ we have $S, r_1\models \varphi_1$, and
\item for any of its frontier nodes $r_2$ we have $S, r_2\models \varphi_2$.
\end{compactenum}
Note that for resource structures without deadlock resources, (1) implies (3).
We say that an eventuality graph $G(S, r_0, \varphi_2)$ {\em fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$} if
\begin{compactenum}
\item the graph contains a frontier node $r_2$ such that $S, r_2\models \varphi_2$, and
\item there is a path from $r_0$ to this frontier node $r_2$ such that for any interior node $r_1$ along the path we have $S, r_1\models\varphi_1$.
\end{compactenum}
From the CTL satisfiability decision procedure of~\cite{Emerson:1991:TML:114891.114907}, it follows that $S, r_0\models {\sf A}[\varphi_1 {\sf U}\varphi_2]$ iff $G(S, r_0, \varphi_2)$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$, and $S, r_0\models {\sf E}[\varphi_1 {\sf U}\varphi_2]$ iff $G(S, r_0, \varphi_2)$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$.
To prove the correctness of our SMT-based synthesis algorithm, we first prove that $\tau$ correctly encodes access constraint into SMT constraints. Towards this end, Theorem~\ref{lemma:smt-encoding} establishes that the SMT encoding is correct for any access constraint and any singleton configuration template~$C=\{c\}$, i.e. a template consisting of one configuration.
To prove this theorem, we give two lemmas (Lemma~\ref{lemma:au} and Lemma~\ref{lemma:eu}), which show that the rewrite function~$\tau_{\sf U}$ correctly encodes eventuality access constraints.
Afterwards, with Lemma~\ref{lemma:smt-correct-reqs} we lift the correctness of the access constraints' encoding to requirements.
Finally, we restate and prove Theorem~\ref{thm:soundness}.
\begin{theorem}
Let $S = ({\cal R}, E, r, S)$ be a resource structure. For any configuration~$c$ for~$S$, resource $r_0\in {\cal R}$, access request $q\in{\cal Q}$, and access constraint $\varphi$,
we have
\[ S_{c,q}, r_0\models \varphi\ \text{iff}\ q\vdash \tau(\varphi, r_0).\]
\label{lemma:smt-encoding}
\end{theorem}
\begin{proof}
The proof proceeds by induction on the derivation of $\tau(\varphi, r_0)$.
\begin{itemize}
\item For the case $\varphi = {\sf true}$, we have $\tau(\varphi, r_0) = {\sf true}$. We get $S_{c,q}, r_0\models \varphi$ and $q\vdash \tau(\varphi, r_0)$.
\item For the case $\varphi = (a\in D)$, we have $\tau(\varphi, r_0) = {\sf true}$ if $L(r_0)(a)\in D$, and $\tau(\varphi, r_0) = {\sf false}$ if $L(r_0)(a)\not\in D$. Recall that $S_{c,q}, r_0\models (a\in D)$ iff $L(r_0)(a)\in D$. It is immediate that $S_{c,q}, r_0\models \varphi$ iff $q\vdash \tau(\varphi, r_0)$.
\item For the case $\varphi = \neg \varphi'$, we have $\tau(\varphi, r_0) = \neg \tau(\varphi', r_0)$.
\begin{itemize}
\item[$\Rightarrow$:] Assume $S_{c,q}, r_0\models \varphi$. We get $S_{c,q}, r_0\not\models \varphi'$. By induction, $q\not\vdash \tau(\varphi', r_0)$. Therefore $q\vdash \tau(\varphi, r_0)$.
\item[$\Leftarrow$:] Assume $q\vdash \tau(\varphi, r_0)$. We get $q\not\vdash \tau(\varphi', r_0)$. By induction, $S_{c,q}\not\models \varphi'$. Therefore $S_{c,q}\models \varphi$.
\end{itemize}
\item For the case $\varphi = \varphi_1\wedge \varphi_2$, we have $\tau(\varphi, r_0) = \tau(\varphi_1, r_0)\wedge \tau(\varphi_2, r_0)$.
\begin{itemize}
\item[$\Rightarrow$:] Assume $S_{c,q}, r_0\models \varphi$. Therefore $S_{c,q}, r_0\models \varphi_1$ and $S_{c,q}, r_0\models\varphi_2$. By induction, $q\vdash \tau(\varphi_1, r_0)$ and $q\vdash \tau(\varphi_2, r_0)$, and therefore $q\vdash \tau(\varphi, r_0)$.
\item[$\Leftarrow$:] Assume $q\vdash \tau(\varphi, r_0)$. Then $q\vdash \tau(\varphi_1, r_0)$ and $q\vdash \tau(\varphi_2, r_0)$. By induction, $S_{c,q}, r_0\models \varphi_1$ and $S_{c,q}, r_0\models \varphi_2$, and therefore $S_{c,q}, r_0\models \varphi$.
\end{itemize}
\item For the case $\varphi = {\sf EX}\varphi'$, we have $\tau(\varphi, r_0) = \exists{r_1\in E(r_0)}.\ (C_{r_0, r_1} \wedge \tau(\varphi', r_1))$.
\begin{itemize}
\item[$\Rightarrow$:] Assume $S_{c,q}, r_0\models \varphi$. By definition of $S_{c,q}$, there is an edge $(r_0, r_1)$ in $E$ such that $q\vdash c((r_0, r_1))$ (1) and $S_{c,q}, r_1\models \varphi'$ (2).
Since $C = \{c\}$, we have $C_{r_0, r_1} = c((r_0, r_1))$. From (1), we thus get $q\vdash C_{r_0, r_1}$.
From (2), by induction, we get $q\vdash \tau(\varphi', r_1)$. It follows that $q\vdash \tau(\varphi, r_0)$.
\item[$\Leftarrow$:] Assume $q\vdash \tau(\varphi, r_0)$. There is an edge $r_1\in E(r_0)$ such that $q\vdash C_{r_0, r_1}$ (1) and $q\vdash \tau(\varphi', r_1)$ (2). From (1), we get $q\vdash c((r_0, r_1))$, and thus there is an edge $(r_0, r_1)$ also in $S_{c,q}$. From (2), by induction, we get $S, r_1\models \varphi'$. Therefore, $S_{c,q}, r_0\models \varphi$.
\end{itemize}
\item For the case $\varphi = {\sf AX}\varphi'$, we have $\tau(\varphi, r_0) = \forall{r_1\in E(r_0)}.\big( C_{r_0, r_1}\Rightarrow \tau(\varphi', r_1)\big)$.
\begin{itemize}
\item[$\Rightarrow$:] Assume $S_{c,q}, r_0\models \varphi$. Then, for any edge $(r_0, r_1)$ of $S_{c,q}$ we have $S_{c,q}, r_1\models \varphi'$.
Consider an edge $(r_0, r_1)\in E$ such that $q\vdash C_{r_0, r_1}$. From $q\vdash C_{r_0, r_1}$, we know that $(r_0, r_1)$ is also an edge in $S_{c,q}$. Therefore, $S_{c,q}, r_1\models \varphi'$. By induction, $q\vdash \tau(\varphi', r_1)$. We get $q\vdash \tau(\varphi, r_0)$.
\item[$\Leftarrow$:] Assume $q\vdash \tau(\varphi, r_0)$. Then for any edge $(r_0, r_1)\in E$, $q\vdash C_{r_0, r_1}$ implies $q\vdash \tau(\varphi', r_1)$. Consider an edge $(r_0, r_1)$ of $S_{c,q}$. We know that $q\vdash C_{r_0, r_1}$, and thus $q\vdash \tau(\varphi', r_1)$. By induction, $S_{c,q}, r_1\models \varphi'$. Therefore $S_{c,q}, r_0\models \varphi$.
\end{itemize}
\item For the case $\varphi = {\sf E}[\varphi_1 {\sf U}\varphi_2]$, we have $\tau(\varphi, r_0) = \tau_{\sf U}(\varphi, r_0, \emptyset)$.
By induction, for any resource $r_1\in {\cal R}$ and any access request $q\in {\cal Q}$ we have
\[
\bigwedge_{i\in \{1,2\}} S_{c,q}, r_1\models \varphi_i\ \text{iff}\ q\vdash \tau(\varphi_i, r_1).
\]
By Lemma~\ref{lemma:eu}, we get $S_{c,q}, r_0\models {\sf E}[\varphi_1 {\sf U}\varphi_2]$ iff $q\vdash \tau_{\sf U}(\varphi, r_0, \emptyset)$.
\item For case $\varphi = {\sf A}[\varphi_1 {\sf U}\varphi_2]$, have $\tau(\varphi, r_0) = \tau_{\sf U}(\varphi, r_0, \emptyset)$.
By induction, for any resource $r_1\in {\cal R}$ and any access request $q\in {\cal Q}$ we have
\[
\bigwedge_{i\in \{1,2\}} S_{c,q}, r_1\models \varphi_i\ \text{iff}\ q\vdash \tau(\varphi_i, r_1).
\]
By Lemma~\ref{lemma:au}, we get $S_{c,q}, r_0\models {\sf A}[\varphi_1 {\sf U}\varphi_2]$ iff $q\vdash \tau_{\sf U}(\varphi, r_0, \emptyset)$.
\end{itemize}
This concludes our proof.
\end{proof}
\begin{lemma}
Let $S = ({\cal R}, E, r, S)$ be a resource structure, $\varphi_1$ and $\varphi_2$ be two access constraints, and $C = \{c\}$ be a configuration template.
If for any resource $r_1\in {\cal R}$ and any access request $q\in {\cal Q}$ we have
\begin{equation}\tag{A1}
\bigwedge_{i\in \{1,2\}} S_{c,q}, r_1\models \varphi_i\ \text{iff}\ q\vdash \tau(\varphi_i, r_1), \label{tau-assumption}
\end{equation}
then for any resource $r_0\in {\cal R}$ we have
\begin{align*}
S_{c,q}, r_0\models {\sf A}[\varphi_1 {\sf U}\varphi_2]\ \text{iff}\ q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, \emptyset).
\end{align*}
\label{lemma:au}
\end{lemma}
\begin{proof}
Assume (A1).
Given a set $X\subseteq {\cal R}$ of resources, we say that $G(S_{c,q}, r_0, \varphi_2)$ is {\em $X\text{-disjoint}$} if no node of $G(S_{c,q}, r_0, \varphi_2)$ is contained in $X$. To avoid clutter, we will write $G[r_0]$ for $G(S_{c,q}, r_0, \varphi_2)$. We prove that for any set~$X\subseteq {\cal R}\setminus\{r_0\}$ of resources,
\[
\begin{array}{ccc}
G[r_0]\ \text{fulfills}\ {\sf A}[\varphi_1 {\sf U}\varphi_2] & {\it iff} & q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, X).\\
\text{and}\ G[r_0]\ \text{is}\ X\text{-disjoint}
\end{array}
\]
The proof proceeds by descending induction on the size of the set~$X$. Note that for the case $X = \emptyset$ we have $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ iff
$q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, \emptyset)$. This case proves the lemma because $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ iff $S_{c,q}, r_0\models {\sf A}[\varphi_1 {\sf U}\varphi_2]$.
Before we start, we expand $\tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, X)$ to
{\small
\begin{align}
\tau(\varphi_2, r_0)\vee & \label{tau-u-var2} \\
& \Big( \tau(\varphi_1, r_0)\label{tau-u-var1} \\
& \wedge \big( \forall{r_1\in E(r_0)\cap X}.\ \neg C_{r_0, r_1} \big) \label{tau-u-no-loop-back}\\
& \wedge \big( \forall{r_1\in E(r_0)\setminus X}. (C_{r_0, r_1} \Rightarrow \notag \\
& \hspace{60pt} \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_1, X\cup \{r_0\})) \big) \Big), \label{tau-u-succ}
\end{align}}
as defined in Figure~\ref{fig:encode}. To avoid clutter, we write, e.g., {\em (\ref{tau-u-no-loop-back}) is true} for $q\vdash \forall{r_1\in E(r_0)\cap X}.\ \neg C_{r_0, r_1}$.
\para{Base Case}
For the base case we have $X = {\cal R}\setminus\{r_0\}$.
\begin{itemize}
\item[$\Rightarrow$:] Assume $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ and $G[r_0]$ is ${\cal R}\setminus\{r_0\}$-disjoint. From ${\cal R}\setminus\{r_0\}$, $G[r_0]$ consists of a single node, $r_0$.
Furthermore, $r_0$ is a frontier node, and since $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$, we have $S_{c,q}, r_0\models \varphi_2$. From~(\ref{tau-assumption}), we get $q\vdash \tau(\varphi_2, r_0)$. Since (\ref{tau-u-var2}) is true, we get $q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus \{r_0\})$.
\item[$\Leftarrow$:] Assume $q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus \{r_0\})$.
Since the resource structure $S_{c,q}$ is deadlock-free, there is a resource $r_1$ in $E(r_0)\cap ({\cal R}\setminus \{r_0\}$) such that $q\vdash C_{r_0, r_1}$.
It follows that (\ref{tau-u-no-loop-back}) is false. Therefore, it must be that
$q\vdash \tau(\varphi_2, r_0)$. By~(\ref{tau-assumption}), $S_{c,q}, r_0\models\varphi_2$. By definition of the eventuality graph $G[r_0]$, we conclude that it consists of a single node, $r_0$. It is immediate that $G[r_0]$ fulfills $ {\sf A}[\varphi_1 {\sf U}\varphi_2]$ and that it is ${\cal R}\setminus \{r_0\}$-disjoint.
\end{itemize}
\para{Inductive Step}
Assume that for any set $X\subseteq {\cal R}\setminus \{r_0\}$ of size $k\leq |X|< |{\cal R}|$,
$G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ and
it is $X$-disjoint
iff
$\tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, X)$.
We show that this holds for any set $X\subset {\cal R}\setminus\{r_0\}$ with $|X| = k-1$.
\begin{itemize}
\item[$\Rightarrow$:]
Assume $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ and it is $X$-disjoint.
\begin{itemize}
\item[Case 1:]
If $S_{c,q}, r_0\models \varphi_2$, then from (\ref{tau-assumption}) we get $q\vdash \tau(\varphi_2, r_0)$. Since~(\ref{tau-u-var2}) is true, it is immediate that $q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, X)$.
\item[Case 2:] If $S, r_0\not\models \varphi_2$, then by (\ref{tau-assumption}) we have $q\not\vdash \tau(\varphi_2, r_0)$.
Therefore, (\ref{tau-u-var2}) is false, so we need to show that (\ref{tau-u-var1}), (\ref{tau-u-no-loop-back}),
and (\ref{tau-u-succ}) are all true:
\begin{compactitem}
\item
Since $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$, we have
$S, r_0\models \varphi_1$ because $r_0$ is an interior node. By~(\ref{tau-assumption}), we get $q\vdash \tau(\varphi_1, r_0)$, and thus (\ref{tau-u-var1}) is true.
\item If $G[r_0]$ has an edge $(r_0, r_1)$, then it must be that the resource structure $S$ has an edge $(r_0, r_1)$ and $q\vdash c((r_0, r_1))$; otherwise, the edge $(r_0, r_1)$ is removed from $S_{c,q}$. Furthermore, since $C = \{c\}$, $C$ does not contain any control variables, and so $C_{r_0, r_1} = c((r_0, r_1))$.
Now, since $G[r_0]$ is $X$-disjoint, we know that $r_0$ does not have any successors contained in $X$. Therefore, for any successor $r_1$ of $r_0$, we have $q\not\vdash C_{r_0, r_1}$. We conclude that (\ref{tau-u-no-loop-back}) is true.
\item Finally, consider an edge $r_1\in E(r_0)\setminus X$ such that $q\vdash C_{r_0, r_1}$.
Since $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ and $r_1$ is a successor of $r_0$, it follows that
$G(S_{c,q}, r_1, \varphi_2)$ also fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$.
Furthermore, since $G[r_0]$ is $X$-disjoint, $G(S_{c,q}, r_1, \varphi_2)$ must be also $X$-disjoint. Furthermore, $G(S_{c,q}, r_1, \varphi_2)$ does not contain the node $r_0$ because $G[r_0]$ is acyclic. We conclude that $G(S_{c,q}, r_1, \varphi_2)$ is $X\cup \{r_0\}$-disjoint and it fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$. By induction, we get $q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_1, X\cup \{r_0\})$. Therefore, (\ref{tau-u-succ}) is true.
\end{compactitem}
\end{itemize}
\item[$\Leftarrow$:]
Assume $q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_0, X)$.
\begin{itemize}
\item[Case 1:] If~$q\vdash \tau(\varphi_2, r_0)$, then (\ref{tau-u-var2}) is true. By~(\ref{tau-assumption}), $S_{c,q}, r_0\models \varphi_2$. It is immediate that
$G[r_0]$ consists of a single node, namely $r_0$. Therefore, $G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ and it is $X$-disjoint because $X\subset {\cal R}\setminus \{r_0\}$.
\item[Case 2:] If~$q\not\vdash \tau(\varphi_2, r_0)$, then (\ref{tau-u-var2}) is false. Therefore,
(\ref{tau-u-var1}), (\ref{tau-u-no-loop-back}),
and (\ref{tau-u-succ}) must be true.
From (\ref{tau-u-var1}) and~(\ref{tau-assumption}), we have $S_{c,q}, r_0\models \varphi_1$.
Consider any node $r_1\in E(r_0)\setminus X$ such that $q\vdash C_{r_0, r_1}$. Then, $r_1$ is a successor of $r_0$ in the graph $G[r_0]$.
From~(\ref{tau-u-succ}), we get $q\vdash \tau_{\sf U}({\sf A}[\varphi_1 {\sf U}\varphi_2], r_1, X\cup \{r_0\})$. By induction, $G(S_{c,q}, r_1, \varphi_2)$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$ and it is $X\cup \{r_0\}$-disjoint.
Since $r_0$ is an internal node, $S, r_0\models \varphi_1$, and all subgraphs rooted
at $r_0$'s successors fulfill ${\sf A}[\varphi_1 {\sf U}\varphi_2]$, it follows that
$G[r_0]$ fulfills ${\sf A}[\varphi_1 {\sf U}\varphi_2]$.
Furthermore, from~(\ref{tau-u-no-loop-back}) we know that $r_0$ has no successors in $X$.
Since all subgraphs rooted at $r_0$'s successors are $X\cup \{r_0\}$-disjoint, it follows that $G[r_0]$ is $X$-disjoint.
\end{itemize}
\end{itemize}
This concludes our proof.
\end{proof}
\begin{lemma}
Let $S = ({\cal R}, E, r, S)$ be a resource structure, $\varphi_1$ and $\varphi_2$ be two access constraints, and $C = \{c\}$ be a configuration template.
If for any resource $r_1\in {\cal R}$ and any access request $q\in {\cal Q}$ we have
\begin{equation}
\tag{A2}
\bigwedge_{i\in\{1,2\}} S_{c,q}, r_1\models \varphi_i\ \text{iff}\ q\vdash \tau(\varphi_i, r_1), \label{tau_eu_assumption}
\end{equation}
then for any resource $r_0\in {\cal R}$ we have
\begin{align*}
S_{c,q}, r_0 \models {\sf E}[\varphi_1 {\sf U}\varphi_2]\ \text{iff}\ q\vdash \tau({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, \emptyset).
\end{align*}
\label{lemma:eu}
\end{lemma}
\begin{proof}
Given a directed graph $G = ({\cal R}, E)$ and a subset $X\subset {\cal R}$ of resources, we
define {\em the projection of $G$ on $X$} as $G|_X = (X, \{ (r_0, r_1)\in E\mid \{r_0, r_1\}\subseteq X\})$. We will write $G[r_0]$ for $G(S_{c,q}, r_0, \varphi_2)$.
We prove by induction on the size of the set $X$ that for any $\{r_0\}\subseteq X\subseteq {\cal R}$, we have
\[
G[r_0]|_X\ \text{fulfills}\ {\sf E}[\varphi_1 {\sf U}\varphi_2]\ {\it iff}\
q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus X).
\]
Note that since $G|_{\cal R} = G$ and $\tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus {\cal R}) = \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, \emptyset)$, the case for $X = {\cal R}$ proves the lemma.
We first expand $\tau_{\sf U}({\sf E}[\varphi_1 {\sf U} \varphi_2], r_0, {\cal R}\setminus X)$ to
\begin{align}
\tau(\varphi_2, r_0) \vee & \label{tau_eu_var2} \\
& \Big( \tau(\varphi_1, r_0) \label{tau_eu_var1}\\
& \wedge \exists{r_1\in E(r_0)\!\setminus\! ({\cal R}\!\setminus\! X)}.\ \big( C_{r_0, r_1} \notag \\
& \hspace{20pt} \wedge \tau_{\sf U}({\sf E}[\varphi_1 {\sf U} \varphi_2], r_1, ({\cal R}\setminus X) \cup \{r_0\}) \big) \Big) \label{tau_eu_succ}
\end{align}
\para{Base Case}
For the base case, we have $X = \{r_0\}$.
\begin{itemize}
\item[$\Rightarrow$:] Assume that $G(S_{c,q}, r_0, \varphi_2)|_{\{r_0\}}$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$. The graph $G[r_0]|_{\{r_0\}}$ consists of the single node $r_0$. The node $r_0$ is a frontier node, and therefore it must be that $S_{c,q}, r_0\models \varphi_2$. By~(\ref{tau_eu_assumption}), we have $q\vdash \tau(\varphi_2, r_0)$. Then (\ref{tau_eu_var2}) is true and therefore $q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus \{r_0\})$.
\item[$\Leftarrow$:] Assume that $q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus \{r_0\})$. Since here $X = \{r_0\}$ and $S$'s edge relation is irreflexive, we have $E(r_0)\setminus ({\cal R}\setminus \{r_0\}) = \emptyset$. Therefore, (\ref{tau_eu_succ}) is false and it must be that $q\vdash \tau(\varphi_2, r_0)$. By~(\ref{tau_eu_assumption}), we have $S_{c,q}, r_0\models \varphi_2$. It is immediate that the graph $G[r_0]|_{\{r_0\}}$ consists of the single node $r_0$, and that it fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$.
\end{itemize}
\para{Inductive Step}
Assume that $G[r_0]|_X$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$ iff
$q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus X)$ holds for any set $\{r_0\}\subseteq X\subset {\cal R}$ of size $1 \leq |X| \leq k$, for some $k$, $1\leq k < |R|$. We show that this also holds for any set $\{r_0\}\subset X\subseteq {\cal R}$ of size $|X| = k+1$.
\begin{itemize}
\item[$\Rightarrow$:] Assume $G[r_0]|_X$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$.
\begin{itemize}
\item[Case 1:] If $S_{c,q}, r_0\models \varphi_2$, then from (\ref{tau_eu_assumption}) we get $q\vdash \tau(\varphi_2, r_0)$. It is immediate that $q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, {\cal R}\setminus X)$.
\item[Case 2:] If $S_{c,q}, r_0\not\models \varphi_2$, then from (\ref{tau_eu_assumption}) we get $q\not\vdash \tau(\varphi_2, r_0)$. Since $G[r_0]|_X$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$, $r_0$ is an internal node and $S_{c,q}, r_0\models \varphi_1$. By (\ref{tau_eu_assumption}), $q\vdash \tau(\varphi_1, r_0)$, and so (\ref{tau_eu_var1}) is true. Furthermore, $r_0$ has a successor $r_1$ with $q\vdash C_{r_0, r_1}$ such that $r_1$ has a path to a node $r_n$ with $S_{c,q}, r_n\models \varphi_2$.
We conclude that $G(S_{c,q}, r_1, \varphi_2)|_{(X \setminus \{r_0\})}$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$.
By induction, since $|X\setminus \{r_0\}| = k-1$, we have $q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_1, {\cal R}\setminus (X\setminus \{r_0\}))$.
Since $r_0\in X$, from ${\cal R}\setminus (X\setminus \{r_0\}) = ({\cal R}\setminus X)\cup \{r_0\}$ we conclude that
(\ref{tau_eu_succ}) is also true.
We conclude that $q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, X)$.
\end{itemize}
\item[$\Leftarrow$:] Assume $q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_0, X)$.
\begin{itemize}
\item[Case 1:] If $q\vdash \tau(\varphi_2, r_0)$, then from (\ref{tau_eu_assumption}) we get $S_{c,q}, r_0\models \varphi_2$. Therefore the graph $G[r_0]|_X$ consists of the single node $r_0$ with $S_{c,q}, r_0\models \varphi_2$. It is immediate that $G[r_0]|_X$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$.
\item[Case 2:] If $q\not\vdash \tau(\varphi_2, r_0)$, then it must be that (\ref{tau_eu_var1}) and (\ref{tau_eu_succ}) are true. From (\ref{tau_eu_var1}) and (\ref{tau_eu_assumption}), we get $S_{c,q}, r_0\models\varphi_1$. From (\ref{tau_eu_succ}), it follows that $r_0$ has a successor $r_1$ with $q\vdash C_{r_0, r_1}$ such that $q\vdash \tau_{\sf U}({\sf E}[\varphi_1 {\sf U}\varphi_2], r_1, ({\cal R}\setminus X)\cup \{r_0\})$.
Since $r_0\in X$, we have $({\cal R}\setminus X)\cup \{r_0\} = {\cal R}\setminus (X\setminus \{r_0\})$.
By induction, $G(S_{c,q}, r_1, \varphi_2)|_{X\setminus \{r_0\}}$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$, so there is a path from $r_1, \ldots, r_n$ in $G(S_{c,q}, r_1, \varphi_2)|_{X\cup \{r_0\}}$ along nodes in $X\setminus \{r_0\}$ where $S_{c,q}, r_n\models\varphi_2$ and $S_{c,q}, r_i\models \varphi_1$ for $1\leq i< n$. It is immediate that there is a path $r_0, r_1, \ldots, r_n$ in $G[r_0]|_X$ such that $S_{c,q}, r_n\models\varphi_2$ and $S_{c,q}, r_i\models \varphi_1$ for $1\leq i< n$. Therefore $G[r_0]|_X$ fulfills ${\sf E}[\varphi_1 {\sf U}\varphi_2]$.
\end{itemize}
\end{itemize}
This concludes our proof.
\end{proof}
\begin{lemma}
Given a resource structure~$S = ({\cal R}, E, r, L)$, a set $R = \{T_1\Rightarrow \varphi_1, \ldots, T_n\Rightarrow \varphi_n \}$ of requirements, and a configuration template $C = \{c\}$, let $\phi = \textsc{Encode}(S, T_1\Rightarrow \varphi_1, C) \wedge \cdots \wedge \textsc{Encode}(S, T_n\Rightarrow \varphi_n, C)$. The constraint $\forall {a}.\ \phi$ is satisfiable iff $S, c\Vdash R$.
\label{lemma:smt-correct-reqs}
\end{lemma}
\begin{proof}
Note that since $C = \{c\}$, the formula $\phi$ contains no control variables, i.e.\ it contains only attribute variables.
Since there is a one-to-one mapping from a valuation of the attribute variables $\vec{a}$ to an access request~$q$, we have $\forall \vec{a}.\ \phi$ iff $\forall q\in {\cal Q}.\ q\vdash \phi$.
We expand the constraint $\phi$ to $(T_1\Rightarrow \tau(\varphi_1, r)) \wedge \cdots \wedge (T_n\Rightarrow \tau(\varphi_n, r))$. We get that $\forall \vec{a}.\ \phi$ iff for any access request $q\in {\cal Q}$, and for any requirement $T\Rightarrow \varphi$, $q\vdash T$ implies $q\vdash \tau(\varphi, r)$.
By Lemma~\ref{lemma:smt-encoding}, $q\vdash \tau(\varphi, r)$ iff $S_{c,q}, r\models \varphi$. We get $\forall \vec{a}.\ \phi$ iff
for any access request $q\in {\cal Q}$, and for any requirement $T\Rightarrow \varphi$, $q\vdash T$ implies $S_{c,q}, r\models \varphi$. By definition of $\Vdash$, we get $\forall \vec{a}.\ \phi$ iff $S, c\Vdash R$.
\end{proof}
\iffalse
We state a useful lemma pertaining to the procedures {\sc Model} and
{\sc Decode}.
\begin{lemma}
Let $S$ be a resource structure, $R = \{T_1\Rightarrow \varphi_1, \ldots, T_n\Rightarrow \varphi_n\}$ a set of requirements, and $C$ a configuration template. Let $\phi = \textsc{Encode}(S, T_1\Rightarrow \varphi_1, C)\wedge \cdots\wedge \textsc{Encode}(S, T_n\Rightarrow \varphi_n, C)$. If $\exists{\vec{z}}. \forall{\vec{a}}.\ \phi$, then
$c = \lambda e.\ \textsc{Derive}(C(e), {\cal M})$, where ${\cal M} = \textsc{Model}(\exists{\vec{z}}. \forall{\vec{a}}.\ \phi)$, is s
\label{lm:derive} \end{lemma}
This lemma states that procedure {\sc Derive} correctly derives a policy
configuration from a model $\cal M$ of the formula
$\exists{\vec{z}}. \forall{\vec{a}}.\ \phi$
returned by the procedure {\sc Model}.
\fi
We now restate and prove Theorem~\ref{thm:soundness}.
\noindent {\bf Theorem~\ref{thm:soundness}.}
{\em Let $S$ be resource structure, $R$ a set of requirements, and $C$ a configuration template. If ${\cal S}_{\sf smt}(S, R, C) = c$ then $S, c\Vdash R$. If ${\cal S}_{\sf smt}(S, R, C) = {\sf unsat}$, then there is no configuration $c$ in $C$ such that $ S, c\Vdash R$.}
\begin{proof}
Let $C = \{c_1, \ldots, c_n\}$. The formula $\exists{\vec{z}}.
\forall{\vec{a}}.\ \phi$ generated by Algorithm~\ref{alg:smt} is
equivalent to the formula $(\forall{\vec{a}}.\ \phi_{c_1})\vee
\cdots\vee (\forall{\vec{a}}.\ \phi_{c_n})$ where $\phi_{c_i}$ is the
formula obtained by grounding the control variables $\vec{z}$ in
$\phi$ with those values that encode the configuration $c_i$. Note
that each formula $\phi_{c_i}$ is equivalent to the one obtained when
using a configuration template $C_i = \{c_i\}$.
Assume that $\exists c\in C.\ S,c\Vdash R$. By Lemma~\ref{lemma:smt-correct-reqs}, $\forall{\vec{a}}.\ \phi_{c_i}$ is satisfiable for some $c_i$ in $C$.
Therefore ${\cal S}_{\sf smt}(S, R, C)$ returns some configuration $c_i$. Assuming the \textsc{Derive} procedure correctly derives a configuration $c_i$ from a model of $\exists{\vec{z}}\forall{\vec{a}}.\ \phi$, then
${\cal S}_{\sf smt}(S, R, C) = c_i$ for some $c_i$ such that $\forall{\vec{a}}.\ \phi_{c_i}$. By Lemma~\ref{lemma:smt-correct-reqs}, $S, c_i\Vdash R$.
Assume that $\neg\exists c\in C.\ S,c\Vdash R$. By Lemma~\ref{lemma:smt-correct-reqs}, $\forall{\vec{a}}.\ \phi_{c_i}$ is not satisfiable for any $c_i$ in $C$. Therefore ${\cal S}_{\sf smt}(S, R, C)$ returns ${\sf unsat}$.
\end{proof}
\
\section{Specifying Requirements}
\label{sec:reqs}
In this section we define \lang, a simple declarative language for specifying requirements.
We give the language's syntax and semantics in Section~\ref{sec:syn-sem}. To simplify the specification of global requirements, in Section~\ref{sec:patterns} we present four requirement patterns that capture common access-control idioms for physical spaces.
Finally, in Section~\ref{sec:gen-req}, we illustrate the specification of two generic access-control requirements: deny-by-default and deadlock-freeness.
\subsection{Requirement Specification Language}
\label{sec:syn-sem}
The design of \lang has been guided by real-world physical access-control requirements. Virtually all such requirements can be formalized as properties that specify which physical spaces subjects can and cannot access, directly and over paths, based on the security context and on the physical spaces they have accessed.
In our physical access-control model, subjects choose which physical spaces to access, which induces a branching structure over the spaces they access.
We therefore build our requirement specification language \lang upon the computation tree logic
(CTL)~\cite{Emerson:1991:TML:114891.114907}, whose branching semantics is a natural fit for physical spaces.
\para{Syntax}
A requirement specified in \lang is a formula of the form
$T\Rightarrow \varphi$ given by the following BNF:
\[ \begin{array}{rcl}
T & ::= & {\sf true}\mid a_s\in D\mid a_c\in D\mid \neg T \mid T
\wedge T\\ \varphi & ::= & {\sf true}\mid a_r\in D\mid \neg \varphi
\mid \varphi \wedge \varphi\mid {\sf EX}\varphi\mid {\sf AX}\varphi
\\[2pt]
& \mid & {\sf E}[\varphi {\sf U} \varphi]\mid {\sf A}[\varphi {\sf U}
\varphi]~.
\end{array}
\]
Here $a_s\in {\cal A}_S$ is a subject attribute, $a_c\in {\cal A}_C$
is a contextual attribute, $a_r\in {\cal A}_R$ is a resource
attribute, and~$D\subseteq {\cal V}$ is a finite subset of values.
The formula $T$ is a constraint over subject and contextual attributes
that defines the access requests to which the requirement applies. We
call $T$ the \emph{target}. The formula $\varphi$ is a CTL formula
over resource attributes. It defines a path property that must hold
for all access requests to which the requirement is applicable. We
call $\varphi$ an \emph{access constraint}.
Note that additional Boolean and
CTL operators can be defined in the standard
way. For example, we write~$\mathsf{false}$ for~$\neg \mathsf{true}$, and
define the Boolean connectives~$\vee$ and~$\Rightarrow$ in the
standard manner using~$\neg$ and~$\wedge$.
We will later make use of the CTL operators ${\sf EF}\varphi$, ${\sf AG}\varphi$,
and ${\sf A}[\varphi{\sf R}\psi]$, which are defined as ${\sf
E}[{\sf true}~{\sf U}~\varphi]$, $\neg ({\sf EF} \neg \varphi)$, and $\neg ( {\sf E}[\neg \varphi{\sf
U}\neg\psi])$, respectively.
Below we give intuitive explanations of~$\sf EX$, $\sf AX$, $\sf EU$, and $\sf AU$,
which are standard CTL connectives.
\begin{figure}[t]
\centering
\tabcolsep=4pt
\fbox{\begin{tabular}{rcl}
$a = c$ & $:=$ & $a \in \{c\}$\\
$a \neq c$ & $:=$ & $\neg (a = c)$\\
$a_{\sf bool}$ & $:=$ & $a_{\sf bool} = {\sf true}$\\
$a_{\sf num} \leq n$ & $:=$ & $a_{\sf num} \in \{0, \ldots, n\}$\\
$a_{\sf num} \geq n$ & $:=$ & $\neg (a_{\sf num} \leq n - 1)$\\
$n \leq a_{\sf num}\leq n'$ & $:=$ & $(a_{\sf num}\geq n) \wedge (a_{\sf num}\leq n')$\\
\end{tabular}}
\caption{Syntactic shorthands: $a\in {\cal A}$ is an attribute,
$a_{\sf num}\in {\cal A}_{\sf num}$ is a numeric attribute,
$a_{\sf bool}\in {\cal A}_{\sf bool}$ is a boolean attribute, $n, n'\in
\mathbb{N}$ are natural numbers.}
\label{fig:syn-ext}
\end{figure}
The
connectives {\em exists-next} $\sf EX$ and {\em always-next} $\sf AX$
constrain the physical spaces that a subject can access next. In our
running example, suppose that a subject has entered the lobby. The
subject can next enter the corridor or go to the public space: these
are immediately accessible from the lobby.
In the lobby,~${\sf EX} \varphi$ states that the formula~$\varphi$
is true in at least one of these ``next'' spaces. In
contrast,~${\sf AX} \varphi$ states that~$\varphi$ is true both
in the corridor and in the public space.
The operators {\em exists-until} $\sf EU$ and {\em always-until} $\sf
AU$ relate two access constraints $\varphi_1$ and $\varphi_2$ over
paths. The formula~$\sf E[\varphi_1 U\varphi_2]$ states that there
exists a path that reaches a
resource~$r$ that satisfies~$\varphi_2$, and any resource prior to~$r$
on the path satisfies~$\varphi_1$. We use this connective to
formalize, for example, waypointing requirements such as: visitors cannot
access the meeting room until they have accessed the lobby.
The formula~$\sf A[\varphi_1 U\varphi_2]$ states that
every path reaches some resource~$r$
that satisfies~$\varphi_2$, and that any resource prior to~$r$ on the
path satisfies~$\varphi_1$.
To simplify writing attribute constraints in~\lang,
we introduce in Figure~\ref{fig:syn-ext} abbreviations
for numeric and boolean attributes.
Based on the attributes' domains, we partition the set of
attributes~$\cal A$ into {\em numeric attributes}~$A_{\sf num}$,
{\em boolean attributes}~$A_{\sf bool}$, and
{\em enumerated attributes}~$A_{\sf enum}$: An attribute~$a$ is {\em
numeric} if ${\sf dom}(a)= \mathbb{N}\cup \{\bot\}$; it is
{\em boolean} if ${\sf dom}(a) = \{{\sf false}, {\sf true}, \bot\}$; otherwise, it
is enumerated and ${\sf dom}(a)$ is finite.
We may write~$a_{\sf num}$ or $a_{\sf bool}$ to emphasize that an
attribute $a$ is numeric or boolean, respectively.
\begin{figure}[t]
\centering
\[\begin{tabular}{llll}
$S, r_0$ & $\models {\sf true}$\\[2pt]
$S, r_0$ & $\models a\in D$ & \ if~$\quad$ & $L(r_0)(a)\in D$\\[2pt]
$S, r_0$ & $\models \neg \varphi$ & \ if~$\quad$ & $S, r_0 \not\models \varphi$\\[2pt]
$S, r_0$ & $\models \varphi_1 \wedge \varphi_2$ & \ if~$\quad$ & $S, r_0 \models \varphi_1\ \text{and}\ S, r_0\models \varphi_2$\\[2pt]
$S, r_0$ & $\models {\sf EX}\varphi$ & \ if~$\quad$ & $\exists (r_0, r_1, \cdots)\in S(r_0).\ S, r_1\models \varphi$ \\[2pt]
$S, r_0$ & $\models {\sf AX}\varphi$ & \ if~$\quad$ & $\forall (r_0, r_1, \cdots)\in S(r_0).\ S, r_1\models \varphi$ \\[2pt]
$S, r_0$ & $\models {\sf E}[\varphi_1 {\sf U}\varphi_2]$ & \ if~$\quad$ & $\exists (r_0, r_1, \cdots)\in S(r_0).\ \exists i\geq 0.$\\ [2pt]
&&& $S, r_i\models \varphi_2 \wedge\forall j\in [0, i).\ S, r_j\models \varphi_1$\\
$S, r_0$ & $\models {\sf A}[\varphi_1 {\sf U}\varphi_2]$ & \ if~$\quad$ & $\forall(r_0, r_1, \cdots)\in S(r_0).\ \exists i\geq 0.$\\[2pt]
&&& $S, r_i\models \varphi_2 \wedge\forall j\in [0, i).\ S, r_j\models \varphi_1$
\end{tabular}\]
\caption{The relation~$\models$ between a resource structure $S = ({\cal R}, E, r_e, L)$, a resource $r_0\in {\cal R}$, and an access constraints~$\varphi$.}
\label{fig:semantics}
\end{figure}
\para{Semantics}
We first inductively define the
satisfaction relation~$\vdash$ between an access request~$q\in \mathcal{Q}$
and a target:
\[
\begin{array}{rclcl}
q & \vdash & {\sf true}\\
q & \vdash & a\in D & \text{if} & q(a)\in D\\
q & \vdash & \neg T & \text{if} & q\not\vdash T\\
q & \vdash & T_1\wedge T_2 & \text{if} & q\vdash T_1\ \text{and}\ q\vdash T_2~.
\end{array}
\]
A requirement $T\Rightarrow \varphi$ is {\em applicable} to an access
request~$q$ iff $q$ satisfies the target $T$, i.e.\ $q\vdash T$.
For example, the requirement $({\sf role = visitor})\Rightarrow
\varphi$ is applicable to all access requests that assign the value
$\sf visitor$ to the subject attribute $\sf role$.
\input{table-patterns}
Let $S = ({\cal R}, E, r_e, L)$ be a resource structure. A {\em path}
of $S$ is an infinite sequence of resources~$(r_0, r_1, \cdots)$ such
that $\forall i\geq 0.\ (r_i, r_{i+1})\in E$, and we denote the set of
all paths rooted at a resource $r_0$ by $S(r_0)$.
In Figure~\ref{fig:semantics},
we inductively define the satisfaction relation~$\models$ between a
resource structure, a resource, and an access constraint.
A resource structure $S$ with an entry resource $r_e$ {\em satisfies} an
access constraint~$\varphi$, denoted by $S \models \varphi$, iff $S,
r_e\models \varphi$.
\begin{definition}
Let $S$ be a resource structure, $c$ a configuration for $S$, and
$T\Rightarrow \varphi$ a requirement. $S$ configured with $c$ {\em
satisfies} $T\Rightarrow \varphi$, denoted by $S, c\Vdash
(T\Rightarrow \varphi)$, iff $q\vdash T$ implies
$S_{c,q}\models\varphi$, for any access request $q\in\mathcal{Q}$.
\end{definition}
We extend $\Vdash$ to sets of requirements as expected. Given a set of
requirements $R = \{T_1\Rightarrow \varphi_1, \ldots, T_n\Rightarrow
\varphi_n\}$, a resource structure $S$ configured with $c$ satisfies
$R$, denoted by $S, c\Vdash R$, iff $S, c\Vdash (T_i \Rightarrow \varphi_i)$ for
all~$i$, $1\leq i \leq n$.
We remark that resource structures can easily be represented using
standard Kripke structures~\cite{Emerson:1991:TML:114891.114907} by
mapping each resource to a Kripke state and each resource attribute
valuation to sets of atomic propositions. The access constraints can
be similarly mapped to standard CTL formulas by translating attribute
constraints into propositional logic.
Note however that while Kripke structures are often used to represent
changes of, say, a concurrent system's state over time, resource
structures model static physical spaces.
\subsection{Requirement Patterns}
\label{sec:patterns}
\lang can be directly used to specify global requirements. However, to illustrate its use and expressiveness, we present the formalization of common physical access-control idioms.
We have studied the requirements of an airport, a corporate building,
and a university campus to elicit the common structure of physical
access-control requirements. To distill the basic requirement
patterns, we split complex requirements into their atomic parts. Our
analysis revealed four common patterns, which we formalize below. The
first pattern abstracts {\em positive} requirements, which
stipulate that the access-control system must grant certain access
requests. The remaining three patterns capture {\em negative}
requirements, which stipulate that the access-control system must deny
certain access requests.
We use the following terminology when describing requirements.
Given a target $T$, we call an access request $q$ a {\em
$T$-request} if $q\vdash T$, i.e.\ $q$ satisfies the target $T$.
Given a resource structure $S$ and an access constraint $\varphi$, we say
that a subject can access a {\em $\varphi$-space} of $S$ if the
subject can access a physical space $r_0$ of $S$ such that $S, r_0\models
\varphi$, i.e. the space $r_0$ satisfies the access constraint~$\varphi$.
Our patterns are summarized in Figure~\ref{fig:patterns}.
\para{Permission}
The {\em permission} pattern abstracts requirements stating that
$T$-requests can access $\varphi$-spaces from the entry resource.
Permission requirements have the form $T\Rightarrow ({\sf EF}\
\varphi)$. The exists-future operator $\sf EF$ formalizes
that a $\varphi$-space is reachable from
the entry resource.
For example, the requirement {\bf R3} stipulating that employees can access the
bureau between $8$AM and $8$PM is formalized as
\[ \big( ({\sf role} = {\sf employee}) \wedge (8\leq {\sf time}\leq 20)\big) \Rightarrow {\sf EF}\ ({\sf id} = {\sf bur}). \]
The target $({\sf role} = {\sf
employee}) \wedge (8\leq {\sf time}\leq 20)$ formalizes that this
requirement is applicable only to access requests made by visitors at
times between $8$AM and $8$PM. The access constraint ${\sf EF}({\sf id} =
{\sf bur})$ is satisfied iff the resource structure has a path from the
entry resource to the bureau.
The requirements {\bf R1} and {\bf R4} of our running example are also
instances of the permission pattern.
\para{Prohibition}
Dual to the permission pattern, the {\em prohibition} pattern captures
requirements stating that $T$-access requests cannot access a
$\varphi$-space.
Prohibition requirements have the form $T\Rightarrow {\sf AG}(\neg
\varphi)$. The operator $\sf AG$ quantifies over
all paths reachable from the entry resource.
An example taken from our airport requirements is: Passengers cannot
access the departure gate zones
without a boarding pass. Another example is
requirement~{\bf R5}, formalized as
\[ ({\sf role}\neq {\sf employee})\Rightarrow {\sf AG}(\neg
\operatorname{\sf sec-zone})~. \]
The target ${\sf role}\neq {\sf employee}$ is satisfied by access
requests that assign a value other than $\sf employee$ to the
attribute $\sf role$. The access constraint ${\sf AG}(\neg
\operatorname{\sf sec-zone})$ is satisfied if no path leads to a
security zone.
\para{Blocking}
The {\em blocking} pattern captures requirements stating that subjects
cannot access a $\psi$-space after they have accessed a
$\varphi$-space.
Intuitively, accessing a $\varphi$-space \emph{blocks} the subject from
accessing $\psi$-spaces.
At international airports, for example, passengers may not access
departure gate zones after they have accessed the baggage claim.
Blocking requirements have the form $T\Rightarrow {\sf AG} (\varphi
\Rightarrow {\sf AG}(\neg \psi))$.
The airport example is formalized as:
\begin{align*} ({\sf role} = {\sf passenger}) \Rightarrow {\sf AG}
& \big( ({\sf zone} = \operatorname{\sf baggage-claim}) \\ & \Rightarrow {\sf AG}\
\neg (\sf zone = {\sf departure})\big)~. \end{align*}
This requirement instantiates the blocking pattern: the target $T$ is
$({\sf role} = {\sf passenger})$, and the two access constraints~$\psi$ and~$\varphi$ are $(\operatorname{\sf zone} = {\sf departure})$
and $({\sf zone} = \operatorname{\sf baggage-claim})$.
\para{Waypointing}
The {\em waypointing} pattern captures requirements stipulating that
subjects must first access a $\varphi$-space before accessing a
$\psi$-space. For example, passengers cannot access an airport's terminal before
they have passed through a security check. This is a negative
requirement that restricts how passengers can access the terminal.
Waypointing requirements have the form $T\Rightarrow ({\sf A}[\varphi{\sf R}\psi])$.
The globally-release operator $\sf AR$ quantifies over all
paths from the entry resource and formalizes that if
$\psi$ holds at some point,
then $\varphi$ was valid at least once beforehand.
The requirement \textbf{R2} of our running example is an instance of
the waypointing pattern and is formalized as
\[ ({\sf role} = {\sf visitor}) \Rightarrow {\sf A}[({\sf id} = {\sf lob}){\sf
R}({\sf id} = {\sf mr})]~. \]
The target specifies that this requirement applies to all access
requests made by visitors. The access constraint is satisfied if all
paths to the meeting room go through the lobby.
The four idioms just described cover \emph{all} the requirements that arose in the case studies
that we report on in Section~\ref{case-studies-sec}.
We remark though that there are global requirements that are not instances of these four patterns.
For example, in corporate buildings, a subject must be able to access the parking lot if he or she has access to an office.
Although this requirement cannot be expressed using the above patterns, it can be directly formalized in \lang as follows:
\begin{align*}
{\sf true} \Rightarrow \big( ({\sf EF} ({\sf zone} = {\sf office})) \Rightarrow ({\sf EF} ({\sf id}
=\operatorname{\sf parking-lot}))\big)~.
\end{align*}
In general, as \lang supports all CTL operators, it can specify any branching property expressible in CTL.
\subsection{Generic Requirements}
\label{sec:gen-req}
We now describe two commonly-used generic requirements.
\para{Deny-by-default}
The deny-by-default principle stipulates that if an access request can be denied without violating the requirements, then it should be denied; cf.~\cite{saltzer75}.
Security engineers often follow this principle to avoid overly permissive local policies.
To illustrate, consider our running example and imagine that the role $\sf intern$ is contained in the domain of the attribute $\sf role$.
The requirements given in Figure~\ref{fig:synth-overview}(b) do not prohibit an intern from accessing, say, the meeting room. However, denying interns access to the meeting room is also compliant with these requirements.
The following requirement, called {\em deny-by-default}, instantiates the above principle: If no positive requirement is applicable to an access request, then only the entry space is accessible to the subject who makes such a request.
To formalize this requirement, we first define positive and negative requirements.
Let $c$ and $c'$ be two configurations for a given resource structure~$S$.
A requirement $T\Rightarrow \varphi$ is \emph{positive} if $S,c\Vdash (T\Rightarrow \varphi)$ and $c \sqsubseteq_S c'$ imply $S,c'\Vdash (T\Rightarrow \varphi)$.
A requirement $T\Rightarrow \varphi$ is \emph{negative} if $S,c\Vdash (T\Rightarrow \varphi)$ and $c' \sqsubseteq_S c$ imply $S,c'\Vdash (T\Rightarrow \varphi)$.
Intuitively, if a configuration satisfies a positive (negative) requirement, then any more (less) permissive configuration also satisfies the requirement.
We remark that although not all requirements are positive or negative, most real-world requirements are, including all requirements specified in this paper.
Let $R$ be a set of requirements that contains only positive and negative requirements, and let
$\{T_1\Rightarrow \varphi), \cdots, T_n\Rightarrow \varphi_n\}$ be the set of all positive requirements contained in $R$.
The deny-by-default requirement for $R$ is
\[ (\neg T_1)\wedge \cdots \wedge (\neg T_n)\Rightarrow {\sf AX}\ ({\sf id} = {\sf entry})~. \]
Here we assume that $L(r_e)({\sf id}) = {\sf entry}$, i.e. the entry resource $r_e$ is labeled with $\sf entry$.
Adding this requirement to our running example's requirements would ensure that an intern cannot access, e.g., the meeting room.
\para{Deadlock-freeness}
A {\em deadlock-freeness} requirement stipulates that there are no deadlocks in a system, i.e.\ resources that a subject can access and then never leave. For example, the meeting room of our running example would be a deadlock if visitors could enter it, but never leave.
As discussed in our system model, local policies that introduce deadlocks are undesirable.
Formally,
the deadlock-freeness requirement is defined as:
\[ {\sf true}\Rightarrow {\sf AG}\ {\sf EX}\ {\sf true}~. \]
This requirement applies to all access requests. The access constraint ${\sf AG}\ {\sf EX}\ {\sf true}$ states that for any resource a subject can access, there is a resource that the subject can access next.
A resource structure~$S$ and a configuration $c$ satisfy this requirement iff for any access requests $q\in {\cal Q}$, $S_{c,q}$ has no deadlocks.
\section{Related Work}
\label{sec:rw}
\para{Physical Access Control}
The Grey project was an experiment in deploying a physical
access-control system at the campus of Carnegie Mellon
University~\cite{bauer:distprove,bgr07}. As part of this project,
researchers developed formal languages for specifying policies and
credentials, and also developed techniques for detecting policy
misconfigurations~\cite{prediction:codaspy12,prediction-tissec}.
The work on credential management, such as delegation, is orthogonal
to the specification of the locks' local policies. In contrast to their
work on detecting policy misconfigurations,
we have developed a framework to synthesize policies
that are guaranteed to
enforces the global requirements, avoiding misconfigurations.
Several researchers have investigated SAT-based and model-checking
techniques for reasoning about physical access
control~\cite{Fitzgerald_anomalyanalysis,vmcai11}. Similarly to our
work, these approaches model spatial constraints, and formalize global
requirements that physical access-control systems must enforce. The
authors of~\cite{Fitzgerald_anomalyanalysis}, for instance, model
physical spaces using directed graphs and formalize global
requirements in first-order logic. Their goal is to identify undesired
denials due to blocked paths and unintended grants to restricted zones
using SAT solvers. In contrast to these verification approaches, we
develop a synthesis framework for generating correct local policies.
\para{Network Policy Synthesis}
The problems of configuring networks with access-control and routing
policies are related to the problem of constructing local policies
from global requirements.
In the network problem domain, one has an explicit resource structure defined
by the network topology and must enforce global requirements using
local rules deployed at the switches.
Several synthesis algorithms for networks have been studied;
e.g.\ see~\cite{DBLP:conf/sp/Guttman97,firmato99,McClurg:2015:ESN:2737924.2737980,Narain2008,Zhang:2011:SDF:2147671.2147677,RahmanA13,mooly:popl15}.
The authors of \cite{DBLP:conf/sp/Guttman97} and \cite{firmato99}, for
example, propose techniques for synthesizing local firewall rules that
collectively enforce global network requirements in a given network
topology. These approaches are sufficiently expressive for formalizing
simple connectivity constraints, such as which hosts can access which
services in a network.
Similarly to our approach, recent techniques for synthesizing network
configurations, such
as~\cite{McClurg:2015:ESN:2737924.2737980,Narain2008,Zhang:2011:SDF:2147671.2147677,RahmanA13},
also leverage SAT and SMT solvers. In addition to access-control
constraints, these techniques also consider business constraints, such
as deployment cost and usability.
However, none of the above approaches for network synthesis supports
branching properties, which are necessary for specifying requirements
such as those stipulating that a fire-exit is reachable from any
office room, as well as those that instantiate our waypointing and
blocking requirement patterns; see Section~\ref{sec:patterns} for
examples. These requirements, which can be expressed in our
framework, are central to physical access control. Existing
network policy synthesis algorithms, therefore, are not sufficiently
expressive for handling access-control requirements for physical spaces.
Policy verification has also been studied in the context of computer
networks; see e.g.~\cite{infocom1354680}. However, this line of research is not
concerned with synthesis, which is our work's main focus. We remark though that our synthesis algorithm can be readily
used for verifying the conformance of a set of local policies to
global access-control requirements; see Section~\ref{sec:algorithm}.
\para{Program Synthesis}
Program synthesis techniques, such as template-based
synthesis~\cite{armando-phd, Srivastava:2010:PVP:1706299.1706337,
armando-fmcad13,Srivastava:2011:PIS:1993498.1993557}, reactive
program synthesis from temporal
specifications~\cite{Clarke:1981:DSS:648063.747438,jobstmann06,
Morgenstern:2011:PSV:2032692.2032706}, and program repair
techniques~\cite{Jobstmann:2005:PRG:2153230.2153260,
Buccafurri:1999:EMC:319103.319105}, are related to policy synthesis
for physical spaces.
Similarly to our SMT-based algorithm,
most of these synthesis frameworks also supplement the logical
specification with a template, and exploit SMT solvers to
efficiently explore the search space defined by the template.
They cannot however express the relevant access-control requirements
we have considered, such as those pertaining to branching
properties.
Our synthesis framework builds upon these techniques, and extends them
with support for specifications that are needed for physical spaces.
Methods for synthesizing models of logical formulas, such as those in
linear-temporal logic or CTL, have been extensively studied in the
literature~\cite{Clarke:1981:DSS:648063.747438,Pnueli:1989:SRM:75277.75293,Ramadge:1987:SCC:35469.35482,pistore-tr,pistore-mbp,
antoniotti95}.
In Section~\ref{sec:synthesis}, we have described a policy
synthesis algorithm based on CTL controller synthesis. This algorithm however comes
at the expense of an exponential blow-up.
Therefore, existing CTL synthesis tools and algorithms cannot be readily applied to synthesize attribute-based local policies in practice.
Our efficient SMT-based algorithm addresses this practical challenge.
\section{Policy Synthesis Problem}
\label{sec:synthesis}
We now define the policy synthesis problem.
We show that this problem is decidable but NP-hard.
\subsection{Problem}
\label{sec:problem}
\begin{definition}
The {\em policy synthesis problem} is as follows:\\[4pt]
\begin{tabular}{p{0.15\columnwidth}p{0.8\columnwidth}}
{\bf Input.} &
A resource structure $S$ and a set of requirements~$R$.\\
{\bf Output.} & A configuration $c$ such that $S, c\Vdash R$,
if such a configuration exists, and $\sf unsat$ otherwise.
\end{tabular}
\end{definition}
The synthesized configuration defines the local policies to be deployed at
the PEPs.
Recall that a policy is extensionally defined as the set of access
requests for which the PEP grants access. As such a set may, in
general, be infinite, one cannot simply output an extensional definition of the
synthesized configuration.
We therefore define local policies intensionally by constraints over
subject and contextual attributes, expressed in the same language that we specify requirement targets in~\secref{sec:reqs}.
The semantics of an intensional local policy $P$ is then simply
$\lambda q.\ \text{if}\ q\vdash P\ \text{then}\ {\sf grant}\ \text{else}\ {\sf deny}$.
Figure~\ref{fig:synth-overview} illustrates the input and output to the policy synthesis problem for our running example.
An example of a local policy defined over the attributes $\sf role$ and $\sf time$ is
$({\sf role} = {\sf visitor}) \wedge (8 \leq {\sf time}\leq 20)$.
This local policy grants all access requests that assign the value $\sf
visitor$ to the attribute $\sf role$ and a number between $8$ and $20$ to
the attribute $\sf time$. Note that this local policy is also the target of requirement~\req{1}.
\subsection{Decidability}
\label{sec:decidability}
To show that the policy synthesis problem is decidable, we give a synthesis algorithm, called ${\cal S}_{\sf cs}$, that uses controller synthesis as a subroutine.
In the following, we first define the controller synthesis problem.
We then show how the algorithm ${\cal S}_{\sf cs}$ constructs the PEPs' local policies by solving multiple controller synthesis instances.
\para{Controller Synthesis Problem}
Controller synthesis algorithms take as input a description of an uncontrolled system, called a plant, along with a specification, and output a controller that restricts the plant so that it satisfies the given specification. In our setting, the plant is the resource structure and the specification is an access constraint, i.e.\ a CTL formula over resource attributes. The synthesized controller then defines which PEPs must grant or deny the access request so that the access constraint is satisfied.
For simplicity, we do not define the controller synthesis problem in its most general form. For our needs the following simpler definition suffices.
\begin{definition}
The {\em controller synthesis problem} is as follows:\\[4pt]
\begin{tabular}{p{0.15\columnwidth}p{0.8\columnwidth}}
{\bf Input.} & A resource structure $S = ({\cal R}, E, r_e, L)$ and an access constraint $\varphi$.\\
{\bf Output.} & A set $E'\subseteq E$ of edges such that $({\cal R}, E', r_e, L)\models
\varphi$, if such an $E'$ exists, and $\sf unsat$ otherwise.
\end{tabular}
\end{definition}
The controller synthesis problem can be reduced to synthesizing a memoryless controller for a Kripke structure given a CTL specification. Deciding whether a controller synthesis instance has a solution is NP-complete~\cite{Antoniotti:1995}. Systems such as MBP~\cite{Bertoli01b} can be used to synthesize controllers. For a comprehensive overview of controller synthesis see~\cite{Ramadge:1987:SCC:35469.35482}.
\para{Algorithm}
The algorithm ${\cal S}_{\sf cs}$ is based on two insights. First, for a given access request $q$, we can use controller synthesis to identify which PEPs must grant or deny~$q$.
In more detail, we can compute $({\cal R}, E', r_e, L)\models \varphi_q$, where $\varphi_q$ conjoins all access constraints of the requirements that are applicable to~$q$.
The edges in $E'$ represent the PEPs that must grant~$q$ and those in $E\setminus E'$ the PEPs that must deny~$q$.
A configuration can thus be synthesized by solving one controller synthesis instance for each access request. However, there are infinitely many access requests.
Our second insight is that we can construct a configuration by solving finitely many controller synthesis instances.
We partition the set~$\cal Q$ of access requests into~$2^{|R|}$ equivalence classes, where two access requests are equivalent if the same set of requirements are applicable to them.
Solving one controller synthesis instance for one representative access request per equivalence class is sufficient for our purpose.
\begin{algorithm}[t]
\DontPrintSemicolon
\KwIn{Resource stricture $S = ({\cal R}, E, r_e, L)$, \hspace{50pt}a set of requirements~$R$}
\KwOut{A configuration $c$ or ${\sf unsat}$}
\Begin{
\For {$e\in E$} {\label{line:foredges}
$c(e) \gets {\sf true}$\;\label{line:setedgetotrue}
}
\For{$R' \subseteq R$} {\label{line:for-loop}
$T \gets T_1\wedge \cdots \wedge T_i \wedge \neg T_{i+1}\wedge
\cdots \wedge \neg T_n$, where\;\label{line:compute-target}
\hspace{10pt} $\{ T_1\Rightarrow \varphi_1, \ldots, T_i\Rightarrow \varphi_i\} = R'$ and\;
\hspace{10pt} $\{ T_{i+1}\Rightarrow \varphi_{i+1}, \ldots, T_n\Rightarrow \varphi_n\} = R\setminus R'$\;
\label{alg:cs:target}
\If {$\exists q\in {\cal Q}.\ q\vdash T$} { \label{line:cs-sat}
$\varphi \gets \varphi_1\wedge \cdots\wedge \varphi_i$\;\label{alg:cs:varphi}
\If {${\sf cs}(S, \varphi) = {\sf unsat}$} { \label{line:cs-synth}
\Return $\sf unsat$\;
} \Else {
\For {$e\in E\setminus {\sf cs}(S, \varphi)$} { \label{alg:cs:edges}
$c(e) \gets c(e) \wedge (\neg T)$\;\label{alg:cs:conf}
}
}
}
}
\Return $c$\;
}
\caption{The algorithm ${\cal S}_{\sf cs}$ for synthesizing policies
using controller synthesis. The controller synthesis algorithm,
denoted~${\sf cs}(S, \varphi)$, outputs either a subset of~$E$ or $\mathsf{unsat}$.}
\label{alg:cs}
\end{algorithm}
The main steps of the algorithm ${\cal S}_{\sf cs}$ are given in Algorithm~\ref{alg:cs}.
The algorithm iteratively constructs a configuration $c$ as follows.
Initially, it sets all local policies to $\sf true$ (lines~\ref{line:foredges}-\ref{line:setedgetotrue}). The algorithm iterates over all subsets $R' = \{T_1\Rightarrow\varphi_1, \ldots,
T_i\Rightarrow \varphi_i\}$ of the requirements~$R$ (line~\ref{line:for-loop}).
The conjunction $T = T_1\wedge \cdots\wedge T_i \wedge \neg T_{i+1} \wedge \cdots \neg T_n$ constructed at line~\ref{line:compute-target} is satisfied by all access requests to which only the requirements contained in $R'$ are applicable.
The set~$\{q\in {\cal Q}\mid q\vdash T\}$ is an equivalence class of access requests.
If this equivalence class is nonempty, i.e. $\exists q\in {\cal Q}.\ q\vdash T$, then $c$ must grant and deny all access requests contained in it in conformance with the access constraints defined by~$R'$.
Lines~\ref{alg:cs:varphi}-\ref{alg:cs:conf} define how the algorithm~${\cal S}_{\sf cs}$ updates~$c$. First, it constructs the conjunction $\varphi$ of the access constraints defined by the requirement in~$R'$.
It then executes the controller synthesis algorithm, denoted by ${\sf cs}$, with the inputs $S$ and $\varphi$.
If the algorithm $\sf cs$ returns $\sf unsat$, then the requirements are not satisfiable for the given resource structure, and the algorithm~${\cal S}_{\sf cs}$ thus returns $\sf unsat$. Otherwise, the algorithm $\sf cs$ returns a set $E'\subseteq E$ of edges. The algorithm updates the configuration $c$ as follows: for any edge in $E\setminus E'$, the configuration is modified to deny access to all requests in the equivalence class defined by~$R'$.
The algorithm terminates when all subsets of the global requirements have been considered.
\begin{theorem}
Let $S$ be a resource structure and $R$ a set of requirements.
If ${\cal S}_{\sf cs}(S, R) = c$ then $S,c\Vdash R$.
If ${\cal S}_{\sf cs}(S, R) = {\sf unsat}$ then there is no configuration $c$ such that $S, c\Vdash R$.
\label{thm:decompose}
\end{theorem}
\ifext
We prove this theorem and give the complexity of~${\cal S}_{\sf cs}$ in Appendix~\ref{sec:decidability}.
\else
We prove this theorem and give the complexity of~${\cal S}_{\sf cs}$ in~\cite{tech-report}.
\fi
\para{Example} To illustrate~${\cal S}_{\sf cs}$,
consider our running example and the requirements {\bf R2} and {\bf
R5} formalized as follows:
\begin{align*}
\textbf{R2} &:= ({\sf role} = {\sf visitor})\Rightarrow ({\sf A}[
({\sf id} = {\sf lob})\ {\sf R}\ ({\sf id} = {\sf mr})]) \\
\textbf{R5} &:= ({\sf role} \neq {\sf employee})\Rightarrow ({\sf AG}\
\neg \operatorname{\sf sec-zone})~.
\end{align*}
We remark that ${\sf dom}({\sf role}) = \{\bot, {\sf visitor}, {\sf employee} \}$, and therefore the targets ${\sf role} = {\sf visitor}$ and ${\sf role} \neq {\sf employee}$ are not equivalent.
To synthesize a configuration, the algorithm ${\cal S}_{\sf cs}$ executes
the second for-loop four times. Let the
selected subset of requirements in the first iteration be $\{\textbf{R2}, \textbf{R5}\}$.
The conjunction~$T$ of the targets is $({\sf role} = {\sf
visitor})\wedge ({\sf role} \neq {\sf employee})$, which is
equivalent to $({\sf role} = {\sf visitor})$. Hence,~$T$ is
satisfiable. The access constraint $\varphi$ (see
Algorithm~\ref{alg:cs}, line \ref{alg:cs:varphi}) is then $ ({\sf A}[ ({\sf id} = {\sf
lob})\ {\sf R}\ ({\sf id} = {\sf mr})]) \wedge ({\sf AG}\ \neg \operatorname{\sf sec-zone})$. A possible output by the controller synthesis
algorithm ${\sf cs}(S, \varphi)$ is $E\setminus \{({\sf cor}, {\sf
bur}), ({\sf out}, {\sf cor}) \}$. The updated configuration $c$
after the first iteration is therefore
\begin{align*}
c(e)
= \left\{
\begin{array}{ll}
{\sf true}\wedge {\sf role} \neq {\sf visitor} & \text{if}\ e = ({\sf cor}, {\sf bur})\\
{\sf true}\ \wedge {\sf role}\neq {\sf visitor} & \text{if}\ e = ({\sf out}, {\sf cor})\\
{\sf true}& \text{otherwise}~.
\end{array} \right.
\end{align*}
Suppose the outputs to the remaining three controller synthesis instances are
${\sf cs}(S, \varphi_{\{\textbf{R2}\}}) = E\setminus \{ ({\sf out},{\sf cor})\}$,
${\sf cs}(S, \varphi_{\{ \textbf{R5}\}}) = E\setminus \{ ({\sf cor}, {\sf bur})\}$, and
${\sf cs}(S, \varphi_\emptyset) = E$, where $\varphi_X$ denotes
the conjunction of the access constraints of the requirements in $X$.
The simplified configuration $c$ returned by
${\cal S}_{\sf cs}$ is
\begin{align*}
c(e)
= \left\{
\begin{array}{ll}
{\sf role} = {\sf employee} & \text{if}\ e = ({\sf cor}, {\sf bur})\\
{\sf role}\neq {\sf visitor} & \text{if}\ e = ({\sf out}, {\sf cor})\\
{\sf true}& \text{otherwise}\, .
\end{array} \right.
\end{align*}
\para{Limitations}
The main limitation of the algorithm ${\cal S}_{\sf cs}$ is that the running time is exponential in the number of requirements, rendering it impractical for nontrivial instances of policy synthesis.
For example, while the algorithm~$\mathcal{S}_\mathsf{cs}$ takes $2$ seconds to synthesize a configuration for our running example, it does not terminate within an hour for our case studies, reported in Section~\ref{case-studies-sec}.
We give a practical policy synthesis algorithm based on SMT solving in Section~\ref{sec:algorithm}.
\subsection{NP-hardness}
To show NP-hardness, we reduce propositional satisfiability to the policy synthesis problem. It is easy to see that a propositional formula~$\varphi$ can be encoded, in logarithmic space, as a target~$T_\varphi$ over Boolean attributes.
Consider the policy synthesis problem for the inputs $S$ and $\{(T_\varphi\Rightarrow {\sf false})\}$, where $S$ is an arbitrary resource structure.
If the output to this policy synthesis instance is $\sf unsat$ then for some access request $q$, we have $q\vdash T_\varphi$. Hence $\varphi$ is satisfiable. Alternatively, the output to the policy synthesis problem is a configuration $c$. Since for any access request $q$ where $q\vdash T_\varphi$ we have $S_{c,q}\models {\sf false}$, it is immediate that there is no access request $q$ such that $q\vdash T_\varphi$. Therefore, $\varphi$ is unsatisfiable.
|
\section{Definitions}
Two ordinal indices, the Szlenk index \cite{Szlenk} and $w^*$-dentability index, have been used to classify and study Asplund spaces. These indices are distinct, but happen to coincide for a large class of spaces. Indeed, due to a result of H\'{a}jek and Schlumprecht \cite{HajekSchlumprecht}, the Szlenk and $w^*$-dentability indices coincide for those Banach spaces whose Szlenk index lies in the interval $[\omega^\omega, \omega_1]$. Here, $\omega$ denotes the first infinite ordinal and $\omega_1$ is the first uncountable ordinal. Since each index has applications to renorming theory, we seek to better understand the relationship between them. Given a Banach space $X$, a $w^*$-compact subset $K$ of $X^*$, and $\ee>0$, we let $s_\ee(K)$ denote those $x^*\in K$ such that for every $w^*$-neighborhood $V$ of $x^*$, $\text{diam}(V\cap K)>\ee$. We let $d_\ee(K)$ denote those $x^*\in K$ such that for every $w^*$-open slice $S$ containing $x^*$, $\text{diam}(S\cap K)>\ee$. We recall that a $w^*$-open slice in $X^*$ is a subset of $X^*$ of the form $\{y^*\in X^*: \text{Re\ }y^*(x)>a\}$ for some $a\in \rr$ and $x\in X$. Of course, $s_\ee(K)\subset d_\ee(K)$. We define $$s^0_\ee(K)=d_\ee^0(K)=K,$$ $$s^{\xi+1}_\ee(K)=s_\ee(s^\xi_\ee(K)),\hspace{5mm} d^{\xi+1}_\ee(K)=d_\ee(d^\xi_\ee(K)),$$ and if $\xi$ is a limit ordinal, $$s^\xi_\ee(K)=\underset{\zeta<\xi}{\bigcap} s_\ee^\zeta(K), \hspace{5mm} d^\xi_\ee(K)=\underset{\zeta<\xi}{\bigcap} d_\ee^\zeta(K).$$ Note that for every $\ee>0$ and every ordinal $\xi$, $s_\ee^\xi(K), d_\ee^\xi(K)$ are $w^*$-compact, and $s_\ee^\xi(K)\subset d_\ee^\xi(K)$. Moreover, if $K$ is convex, so is $d^\xi_\ee(K)$. We let $Sz_\ee(K)=\min\{\xi: s^\xi_\ee(K)=\varnothing\}$ if this class is non-empty, and we write $Sz_\ee(K)=\infty$ otherwise. We let $Sz(K)=\sup_{\ee>0}Sz_\ee(K)$, with the convention that this supremum is $\infty$ if $Sz_\ee(K)=\infty$ for some $\ee>0$. We define $Dz_\ee(K)$, $Dz(K)$ similarly. Given a Banach space $X$, we let $Sz(X)=Sz(B_{X^*})$ and $Dz(X)=Dz(B_{X^*})$. If $\Phi:X\to Y$ is an operator, we define $Sz(\Phi)=Sz(\Phi^* B_{Y^*})$. The index $Sz(X)$ is the \emph{Szlenk index} of $X$, and $Dz(X)$ is the $w^*$-\emph{dentability index} of $X$. We observe that $Sz(K)\leqslant Dz(K)$ for any $w^*$-compact $K$.
Given a compact, Hausdorff space $K$, we let $K'$ denote the Cantor-Bendixson derivative of $K$. That is, $K'$ consists of those points in $K$ which are not isolated in $K$. Note that $K'$ is closed in $K$, and so is compact, Hausdorff with its relative topology as long as it is non-empty. We define $$K^0=K,$$ $$K^{\xi+1}=(K^\xi)',$$ and if $\xi$ is a limit ordinal, we let $$K^\xi=\underset{\zeta<\xi}{\bigcap} K^\zeta.$$ We let $CB(K)$ denote the Cantor-Bendixson index of $K$, which is the minimum ordinal $\xi$ such that $K^\xi=\varnothing$ if such an ordinal exists, and otherwise we write $CB(K)=\infty$. The space $K$ is said to be \emph{scattered} if $CB(K)$ is an ordinal. A standard compactness argument yields that if $K$ is scattered, $CB(K)$ is a successor ordinal. If $K$ is scattered, we let $\Gamma(K)$ denote the minimum ordinal $\xi$ such that $CB(K)\leqslant \omega^\xi$. Note that the inequality $CB(K)\leqslant \omega^\xi$ is strict except in the trivial case that $CB(K)=1$ (that is, when $K$ is finite), since $CB(K)$ cannot be a limit ordinal, while $\omega^\xi$ is a limit ordinal for any $\xi>0$. If $K$ is not scattered, we let $\Gamma(K)=\infty$. We agree to the convention that $\omega \infty=\infty$.
Our proofs work for both the real and complex scalars. In what follows, $C(K)$ is the Banach space of continuous, scalar-valued functions defined on the compact, Hausdorff space $K$. Given a Banach space $X$ and $1<p<\infty$, $L_p(X)$ denotes the space of (equivalence classes of) Bochner-integrable, $X$-valued functions defined on $[0,1]$ with Lebesgue measure.
\begin{theorem} For any compact, Hausdorff $K$ and any $1<p<\infty$, $$Sz(C(K))=\Gamma(K),$$ $$Dz(C(K))=Sz(L_p(C(K)))=\omega \Gamma(K).$$
\label{main theorem}
\end{theorem}
For any ordinals $\xi, \zeta$ such that $\omega^\zeta\leqslant \xi<\omega^{\zeta+1}$, it is easy to see that $CB([0,\xi])=\zeta+1$. Thus our results recover the known facts that if $\omega^{\omega^\zeta}\leqslant \xi <\omega^{\omega^{\zeta+1}}$, $Sz(C([0, \xi]))=\omega^{\zeta+1}$. This was shown by Samuel \cite{Samuel} in the case that $\xi$ is countable, by Lancien and H\'{a}jek when $\xi<\omega_1\omega$, and by Brooker \cite{BrookerOrdinals} in the general case. We also recover the values of $Dz([0, \xi])$ and $Sz(L_p(C([0, \xi])))$, which was shown for countable $\xi$ in \cite{HajekLancienProchazka}, and in the general case by Brooker \cite{BrookerOrdinals}.
\section{Preliminaries}
We collect a few facts concerning the Szlenk and $w^*$-dentability indices.
\begin{proposition}Let $X$ be a Banach space. \begin{enumerate}[(i)]\item If $X$ is isomorphic to a subspace of $Y$, then $Sz(X)\leqslant Sz(Y)$ and $Dz(X)\leqslant Dz(Y)$. \item $Sz(X)=1$ if and only if $\dim X<\infty$, and otherwise $Sz(X)\geqslant \omega$. \item $Dz(X)=1$ if and only if $X=\{0\}$, and otherwise $Dz(X)\geqslant \omega$. \item $Sz(X)=\infty$ if and only if $Dz(X)=\infty$ if and only if $X$ fails to be Asplund. \item For any Asplund space $X$, there exist ordinals $\xi, \zeta$ such that $Sz(X)=\omega^\xi$ and $Dz(X)=\omega^\zeta$. \item For any $1<p<\infty$, $Dz(X)\leqslant Sz(L_p(X))$. \item For any Banach space $Y$, $Sz(X\oplus Y)=\max\{Sz(X), Sz(Y)\}$. \item If $Y$ is a Banach space and $\Phi:X\to Y$ is an isomorphic embedding, $Sz(\Phi)=Sz(X)$. \item For any $K,L\subset X^*$ $w^*$-compact, $\ee>0$, and any ordinal $\xi$, if $K\subset L+\ee B_{X^*}$, then $$s_{12\ee}^\xi(K)\subset s_{3\ee}^\xi(L)+\ee B_{X^*}.$$ \end{enumerate}
\label{szlenk facts}
\end{proposition}
Items $(i)$-$(vi)$ can be found in the survey paper \cite{LancienSurvey}. Item $(vii)$ was stated in \cite{HajekLancien} in the case that $Y=X$, but the proof yields the version here. Item $(viii)$ is due to Brooker \cite{BrookerAsplund}.
The idea for $(ix)$ is in \cite[Lemma $6.2$]{Lancien}, alhough the statement was slightly different. We give the proof of item $(ix)$ as it is stated here.
\begin{proof}[Proof of $(ix)$] By induction. The $\xi=0$ case is clear. Assume $\xi$ is a limit ordinal and the result holds for every $\zeta<\xi$, and fix $x^*\in s^\xi_{12\ee}(K)$. For every $\zeta<\xi$, since $x^*\in s_{12\ee}^\zeta(K)\subset s_{3\ee}^\zeta(L)+\ee B_{X^*}$, we may fix $y^*_\zeta\in s_{3\ee}^\zeta(L)$ and $z^*_\zeta\in \ee B_{X^*}$ such that $x^*=y^*_\zeta+ z^*_\zeta$. We pass to a subnet $(y^*_\lambda)_{\lambda\in D}$ of $(y^*_\zeta)_{\zeta<\xi}$ with $w^*$-limit $y^*\in s^\xi_{3\ee}(L)$ and note that over the same subnet, $z^*_\lambda \underset{\lambda\in D, w^*}{\to}x^*-y^*\in \ee B_{X^*}$. Therefore $x^*=y^*+z^*\in s_{3\ee}^\xi(L)+\ee B_{X^*}$. Last, assume the result holds for some ordinal $\xi$ and suppose $x^*\in s^{\xi+1}_{12\ee}(K)$. Fix a net $(x^*_\lambda)\subset s^\xi_{12\ee}(K)$ converging $w^*$ to $x^*$ and such that for every $\lambda$, $\|x^*_\lambda- x^*\|>6\ee$. For every $\lambda$, fix $y^*_\lambda\in s_{3\ee}^\xi(L)$ and $z^*_\lambda\in \ee B_{X^*}$ such that $x^*_\lambda=y^*_\lambda+z^*_\lambda$. By passing to a subnet, we may assume $y^*_\lambda\underset{w^*}{\to} y^*\in s_{3\ee}^\xi(L)$ and note that over the same subnet, $z^*_\lambda\underset{w^*}{\to} x^*-y^*\in \ee B_{X^*}$. For every $\lambda$, $$\|y^*_\lambda - y^*\| = \|x^*_\lambda-z^*_\lambda- x^* +x^*-y^*\| \geqslant \|x^*_\lambda - x^*\|-\|z^*_\lambda\|-\|x^*-y^*\| >6\ee-2\ee>3\ee.$$ From this it follows that $y^*\in s_{3\ee}^{\xi+1}(L)$.
\end{proof}
Rudin \cite{Rudin} showed that if $K$ is compact, Hausdorff, scattered, $C(K)^*=\ell_1(K)$, where the canonical $\ell_1(K)$ basis is the set of Dirac functionals $\{\delta_x: x\in K\}$. By a result of Namioka and Phelps \cite{NamiokaPhelps}, if $K$ is scattered, $C(K)$ is Asplund. We note that a Banach space $X$ is an Asplund space if and only if every separable subspace of $X$ has separable dual. This was shown in \cite{DevilleGodefroyZizler} in the real case, and it is explained in \cite{BrookerAsplund} how to deduce the complex case from the real case. Stegall \cite{Stegall} showed that if every separable subspace of $X$ has separable dual, then $X^*$ has the Radon-Nikodym property. It follows from \cite[Page 98]{DiestelUhl} that if $X^*$ has the Radon-Nikodym property, for any $1<p<\infty$, $L_p(X)^*=L_q(X^*)$ via the canonical embedding of $L_q(X^*)$ into $L_p(X)^*$. Here, $1/p+1/q=1$. Therefore if $K$ is scattered and $1<p<\infty$, $L_p(C(K))^*=L_q(\ell_1(K))$.
Given a closed subset $F$ of $K$, we let $C_F(K)$ denote those $f\in C(K)$ such that $f|_F\equiv 0$. If $F=\varnothing$, we let $C_F(K)=C(K)$. If $K$ is scattered, we let $K_\infty=K^{CB(K)-1}$. This is well-defined, since $CB(K)$ is a successor ordinal. We let $C_0(K)=C_{K_\infty}(K)$. Note that $K_\infty$ is finite and non-empty, so $0<\dim C(K)/C_0(K)<\infty$. From this it follows that $L_p(C(K))$ is isomorphic to $L_p(C_0(K))\oplus L_p$. Then if $K$ is infinite, \begin{align*} Sz(L_p(C(K))) & =\max\{Sz(L_p(C_0(K))), Sz(L_p)\}=\max\{Sz(L_p(C_0(K))), \omega\} \\ & =Sz(L_p(C_0(K))).\end{align*} If $K$ is finite, $K_\infty=K$ and $C_0(K)=\{0\}$, so that $$Sz(L_p(C(K)))=Sz(L_p)=\omega,$$ since $L_p$ is asymptotically uniformly smooth. This shows that in the non-trivial case that $K$ is infinite, to compute the Szlenk index of $L_p(C(K))$, it is sufficient to compute the Szlenk index of $L_p(C_0(K))$.
With $F\subset K$ still closed, we let $\approx_F $ denote the equivalence relation on $K$ given by $s\approx_F t$ if $s=t$ or if $s,t\in F$. We let $K/F$ denote the space of equivalence classes $[s]$ of members $s$ of $K$ and endow $K/F$ with the quotient topology coming from the function $\chi_F$ given by $\chi_F(s)\mapsto [s]$. Of course, the equivalence classes in $K/F$ are $F$ and $\{s\}$, $s\in K\setminus F$. Note that $C_F(K)$ is canonically isometrically isomorphic to $C_{\{F\}}(K/F)$ via the operator $T:C_{\{F\}}(K/F)\to C_F(K)$ given by $Tf(t)=f(\chi_F(t))$. If $F=K^\gamma$ for some $\gamma<CB(K)$, it is straightforward to check that for $0\leqslant \zeta\leqslant \gamma$, then $(K/K^\gamma)^\zeta=q_F(K^\zeta)$. In particular, $(K/K^\gamma)^\gamma=\{K^\gamma\}$, and $CB(K/K^\gamma)=\gamma+1$. This means that $(K/K^\gamma)_\infty$, the last non-empty Cantor-Bendixson derivative of $K/K^\gamma$, is the space consisting of the single equivalence class $K^\gamma$. From this and our previous remarks, it follows that $C_{K^\gamma}(K)$ is isometrically isomorphic to $C_0(K/K^\gamma)$.
Note that the restriction map $\rho_F:C(K)\to C(F)$ given by $f\mapsto f|_F$ is a quotient map by the Tietze extension theorem. The adjoint $\rho_F^*:\ell_1(F)\to \ell_1(K)$ is the inclusion, and $\rho^*_F B_{\ell_1(F)} =\{\mu\in B_{\ell_1(K)}: |\mu|(K\setminus F)=0\}$, which we identify with $B_{\ell_1(F)}$ throughout. Note that this identification is a linear, $w^*$-$w^*$-continuous isometry, so that for any ordinal $\xi$ and any $\ee>0$, $s^\xi_\ee(B_{\ell_1(F)})= s^\xi_\ee(\rho^*_F B_{\ell_1(F)})$ and $d^\xi_\ee(B_{\ell_1(F)})= d^\xi_\ee(\rho^*_F B_{\ell_1(F)})$. We will use this fact throughout. Moreover, if $R_F:L_p(C(K))\to L_p(C(F))$ is the restriction given by $R_F f(t)= f(t)|_F$, $R_F^*:L_q(\ell_1(F))\to L_q(\ell_1(K))$ is the inclusion, and we may identify $B_{L_q(\ell_1(F))}$ with its image under $R^*_F$ when computing the Szlenk and $w^*$-derivations.
Let $\varphi_F:C_F(K)\to C(K)$, $\Phi_F:L_p(C_F(K))\to L_p(C(K))$ denote the inclusions. Moreover, with this identification, Note that $L_q(\ell_1(F))$ is canonically included in $L_q(\ell_1(K))=L_p(C(K))^*$, and $$L_p(C_F(K))^\perp = L_q(C_F(K)^\perp)=L_q(\ell_1(F)).$$ From this it follows that $L_q(\ell_1(F))$ is $w^*$-closed in $L_q(\ell_1(K))$ and $$L_p(C_F(K))^*= L_q(\ell_1(K))/L_q(\ell_1(F)).$$ Note that any operator $p:\ell_1(K)\to \ell_1(K)$ extends to a function $P:L_q(\ell_1(K))\to L_q(\ell_1(K))$ given by $(Pf)(t)=p(f(t))$, and $\|P\|=\|p\|$. Let $p_F:\ell_1(K)\to \ell_1(F)\subset \ell_1(K)$ be the canonical projection, and let $q_F$ denote the complementary projection $I_{\ell_1(K)}-p_F$. Note that $\|q_F\|, \|p_F\|\leqslant 1$. Given an ordinal $\gamma<CB(K)$, we let $p_\gamma$, $q_\gamma$ denote the projections $p_{K^\gamma}$, $q_{K^\gamma}$. Let $P_F, Q_F:L_q(\ell_1(K))\to L_q(\ell_1(F))$ be the maps induced by $p_F$, $q_F$, respectively. Let $P_\gamma$, $Q_\gamma$ be the maps induced by $p_\gamma$, $q_\gamma$. Note that the quotient map $\varphi_F^*:\ell_1(K)\to \ell_1(K)/\ell_1(F)$ is given by $$\varphi^*_F(\mu)=q_F(\mu)+\ell_1(F),$$ and $\|\phi^*_F (\mu)\|=\|q_F (\mu)\|$. The quotient map $\Phi_F^*:L_q(\ell_1(K))\to L_q(\ell_1(K))/L_q(\ell_1(F))$ is given by $$\Phi^*_F(f)= Q_F(f)+L_q(\ell_1(F)),$$ and $\|\Phi^*_F(f)\|=\|Q_F(f)\|$.
Note also that for any $f\in L_q(\ell_1(K))$, $$\|P_Ff\|^q+\|Q_F f\|^q\leqslant \|f\|^q.$$ Indeed, for any $t\in [0,1]$, $$\|p_F f(t)\|^q+\|q_F f(t)\|^q \leqslant (\|p_F f(t)\|+\|q_F f(t)\|)^q = \|f(t)\|^q.$$ Integrating over $t$ yields the inequality. In particular, if $\|Q_F f\|^q=\|\Phi^*_F f\|^q \geqslant 1-\ee^q$ and $\|f\|\leqslant 1$, then $\|P_F f\|^q \leqslant 1-(1-\ee^q)=\ee^q$. Thus if $f,g\in B_{L_q(\ell_1(K))}$, $\|f-g\|>3\ee$, and $\|\Phi^*_F f\|^q, \|\Phi^*_F g\|^q>1-\ee^q$, then $$\|\Phi^*_F f- \Phi^*_F g\| = \|Q_F f- Q_Fg\| \geqslant \|f-g\|-\|P_F f\|-\|P_F g\|>\ee.$$ We also note that if $\|\varphi^*_F \mu\|, \|\varphi^*_F \mu'\|>1-\ee$ and $\|\mu-\mu'\|>3\ee$, then $\|\varphi^*_F \mu-\varphi^*_F \mu'\|>\ee$.
We will use these fact to prove the following. Item $(i)$ is based on \cite[Lemma $3.3$]{HajekLancien} and $(ii)$ is based on \cite[Lemma $6$]{HajekLancienProchazka}.
\begin{lemma} Suppose $\ee>0$, $F$ is a closed subset of $K$, and $\beta$ is an ordinal. \begin{enumerate}[(i)]\item If $\mu\in s_{3\ee}^\beta(B_{\ell_1(K)})$ and $\|\varphi^*_F \mu\|>1-\ee$, $\varphi^*_F \mu\in s^\beta_\ee(\varphi^*_F B_{\ell_1(K)})$. \item If $f\in s_{3\ee}^\beta(B_{L_q(\ell_1(K))})$ and $\|\Phi^*_F f\|^q>1-\ee^q$, then $\Phi^*_F f\in s_\ee^\beta(\Phi^*_F B_{L_q(\ell_1(K))})$. \end{enumerate}
\label{lemma1}
\end{lemma}
\begin{proof}$(i)$ We work by induction on $\beta$, with the base and limit ordinal cases clear. Assume $\mu\in s^{\beta+1}_{3\ee}(B_{\ell_1(K)})$ and $\|\varphi^*_F \mu\|>1-\ee$. Then there exist two nets $(\mu_\lambda)$, $(\eta_\lambda)$ contained in $s_{3\ee}^\beta(B_{\ell_1(K)})$ indexed by the same directed set, converging $w^*$ to $\mu$, and such that $\|\mu_\lambda-\eta_\lambda\|>3\ee$ for all $\lambda$. By passing to a subnet and using $w^*$-$w^*$ continuity, we may assume that $\|\varphi^*_F \mu_\lambda\|$, $\|\varphi^*_F \eta_\lambda\|>1-\ee$. From this it follows that $\varphi^*_F \mu_\lambda, \varphi^*_F \eta_\lambda\in s_\ee^\beta(\varphi^*_F B_{\ell_1(K)})$. By our previous remarks, $\|\varphi^*_F \mu_\lambda -\varphi^*_F \eta_\lambda\|>\ee$, and since $\varphi^*_F \mu_\lambda$, $\varphi^*_F \eta_\lambda\underset{w^*}{\to} \varphi^*_F \mu$, we deduce that $\varphi^*_F \mu\in s^{\beta+1}_\ee(\varphi^*_F B_{\ell_1(K)})$.
$(ii)$ This follows from an inessential modification of $(i)$.
\end{proof}
\begin{proposition} Let $F\subset K$ be a closed subset of the compact, Hausdorff space $K$. \begin{enumerate}[(i)]\item If $F$ is a finite set of isolated points, $Sz(\varphi_{K\setminus F})=1$. \item If $F$ is a finite, non-empty set of isolated points, $Sz(\Phi_{K\setminus F})=\omega$. \item For any $\mu\in \ell_1(K)$ and any real number $a>0$, if $\|\varphi^*_1 \mu\|>a$, then there exists a finite set $F$ of isolated points such that $\|\varphi^*_{K\setminus F} \mu\|>a$. \item For any $f\in L_q(\ell_1(K))$ and any real number $a>0$, if $\|\Phi^*_1 f\|>a$, then there exists a finite set $F$ of isolated points such that $\|\Phi^*_{K\setminus F} f\|>a$. \item If $\xi<CB(K)$ is a limit ordinal, $\mu\in \ell_1(K)$, and $\|\varphi^*_\xi \mu\|>a$, then there exists an ordinal $\gamma<\xi$ such that $\|\varphi^*_\gamma \mu\|>a$. \item If $\xi<CB(K)$ is a limit ordinal, $f\in L_q(\ell_1(K))$, and $\|\Phi^*_\xi f\|>a$, then there exists an ordinal $\gamma<\xi$ such that $\|\Phi^*_\gamma f\|>a$. \end{enumerate}
\label{barry}
\end{proposition}
\begin{proof} We use Proposition \ref{szlenk facts} for $(i)$ and $(ii)$.
$(i)$ This follows from the fact that the inclusion operator $\varphi_{K\setminus F}:C_{K\setminus F}(K)\to C(K)$ has the same Szlenk index as $C_{K\setminus F}(K)$, which is $1$, since $C_{K\setminus F}(K)$ has finite dimension.
$(ii)$ This follows from the fact that $\Phi_{K\setminus F}$ has the same Szlenk index as $L_p(C_{K\setminus F}(K))\approx L_p$.
$(iii)$ This follows from regularity and the fact that $K\setminus K'=K\setminus K^1$ is the set of isolated points in $K$. If $\|q_1 \mu\|=|\mu|(K\setminus K')>a$, then there exists a compact, and therefore finite, subset $F$ of $K\setminus K'$ such that $\|q_{K\setminus F} \mu\|=|\mu|(F)>a$.
$(iv)$ One uses $(iii)$ to deduce the result first for simple functions and then extend by density.
$(v)$ This follows from regularity and the fact that any compact subset $F$ of $K\setminus K^\xi$ is contained in $K\setminus K^\gamma$ for some $\gamma<\xi$. Indeed, if such a gamma did not exist, we could choose for every $\gamma<\xi$ some $x_\gamma\in F\cap K^\gamma$. Any convergent subnet of the net $(x_\gamma)_{\gamma<\xi}$ necessarily converges to a member of $K^\xi\subset K\setminus F$, contradicting the compactness of $F$. If $\|q_\xi \mu\|=|\mu|(K\setminus K^\xi)>a$, there exists $\gamma<\xi$ and a compact subset $F$ of $K\setminus K^\gamma$ such that $|\mu|(F)>a$. Then $\|q_\gamma \mu\|=|\mu|(K\setminus K^\gamma)\geqslant |\mu|(F)>a$.
$(vi)$ One uses $(v)$ to deduce the result first for simple functions and then extend by density.
\end{proof}
We conclude this section with a technical fact. The general idea behind this fact is well-known. We recall the Hessenberg sum of two ordinals, the details of which can be found in \cite{Monk}. Given a non-zero ordinal $\xi$, we may write $\xi=\omega^{\gamma_1}n_1+\ldots + \omega^{\gamma_k}n_k$, where $\gamma_1>\ldots >\gamma_k$ and $k, n_i\in \nn$. Given two ordinals $\xi, \zeta$, by representing them in Cantor normal and then including zero terms if necessary, we may assume there exist $k\in \nn$, $m_i, n_i\in \nn\cup\{0\}$, and $\gamma_1>\ldots >\gamma_k$ such that $\xi=\omega^{\gamma_1}m_1+\ldots + \omega^{\gamma_k}m_k$ and $\zeta=\omega^{\gamma_1}n_1+\ldots + \omega^{\gamma_k}n_k$. Then the Hessenberg sum of $\xi, \zeta$ is $$\xi\oplus \zeta=\omega^{\gamma_1}(m_1+n_1)+\ldots + \omega^{\gamma_k}(m_k+n_k).$$ We remark that $(\xi\oplus \zeta)+1=(\xi+1)\oplus \zeta$. We also note that for any ordinal $\xi$, the set $\{(\zeta_1, \zeta_2): \zeta_1\oplus \zeta_2=\xi\}$ is finite. Last, if $\xi$ is any ordinal and $r\in \nn$, $\{(\zeta_1, \zeta_2): \zeta_1\oplus \zeta_2=\omega^\xi r\}=\{(\omega^\xi k, \omega^\xi(r-k): 0\leqslant k\leqslant r\}$.
\begin{lemma} Suppose $B, L, A\subset X^*$ are $w^*$-compact and $B\subset L+A$. \begin{enumerate}[(i)]\item For any $\ee>0$, $s_{2\ee}(B)\subset (s_\ee(L)+A)\cup (L+s_\ee(A))$. \item For any ordinal $\xi$ and any $\ee>0$, $$s_{4\ee}^\xi(B)\subset \underset{\zeta\oplus\eta=\xi}{\bigcup} [s_\ee^\zeta(L)+ s_\ee^\eta(A)].$$ \item If for some $\ee>0$, some ordinal $\xi$, and some $r\in \nn$, $Sz_{3\ee}(L)<\omega^\xi r$, then $$s_{12\ee}^{\omega^\xi r}(B)\subset L+s_{3\ee}^{\omega^\xi}(A).$$ \end{enumerate}
\label{smooth}
\end{lemma}
\begin{proof}$(i)$ If $x^*\in s_{4\ee}(B)$, we may fix a net $(x^*_\lambda)\subset B$ converging $w^*$ to $x^*$ and such that $\|x^*_\lambda- x^*\|>2\ee$ for all $\lambda$. For each $\lambda$, write $x^*_\lambda= y^*_\lambda+z^*_\lambda\in L+A$. We may pass to a subnet twice and assume $y^*_\lambda\underset{w^*}{\to} y^*\in L$, $z^*_\lambda\underset{w^*}{\to} z^*\in A$, and either $\|y^*_\lambda- y^*\|\geqslant \|z^*_\lambda - z^*\|$ for all $\lambda$ or $\|y^*_\lambda - y^*\| \leqslant \|z^*_\lambda-z^*\|$ for all $\lambda$. In the first case, it follows that $y^*\in s_\ee(L)$, and in the second, $z^*\in s_\ee(A)$. Since $y^*+z^*=x^*$, we deduce the first statement.
$(ii)$ We work by induction on $\xi$. The $\xi=0$ case is trivial. Assume $$s^\xi_{4\ee}(B)\subset \underset{\zeta\oplus \eta=\xi}{\bigcup} [s_\ee^\zeta(L)+ s_\ee^\eta(A)]$$ and $x^*\in s_\ee^{\xi+1}(B)$. Fix a net $(x^*_\lambda)\subset s^\xi_{4\ee}(B)$ converging $w^*$ to $x^*$ such that $\|x^*_\lambda - x^*\|>2\ee$ for all $\lambda$. Since $\{(\zeta, \eta): \zeta\oplus \eta=\xi\}$ is finite, we may pass to a subnet and assume that for some $\sigma, \tau$ with $\sigma\oplus \tau=\xi$, $x^*_\lambda\in s^\sigma_\ee(B)+s^\tau_\ee(L)$ for all $\lambda$. Then $x^*\in s_{2\ee}(s^\sigma_\ee(L)+s^\tau_\ee(A))$, and we deduce by $(i)$ that $$x^*\in (s^{\sigma+1}_\ee(L)+s^\tau_\ee(A))\cup (s^\sigma_\ee(L)+s^{\tau+1}_\ee(A))\subset \underset{\zeta\oplus \eta=\xi+1}{\bigcup}[ s^\zeta_\ee(L)+ s^\eta_\ee(A)].$$
Last, assume $\xi$ is a limit ordinal and the result holds for every $\zeta<\xi$. For every $\zeta<\xi$, $x^*\in s^{\zeta+1}_{4\ee}(B)$, and there exist $\alpha_\zeta$, $\beta_\zeta$ such that $\alpha_\zeta\oplus \beta_\zeta=\zeta+1$ and $x^*\in s^{\alpha_\zeta}_\ee(L)+s^{\beta_\zeta}_\ee(A)$. By \cite[Proposition $2.5$]{Causey}, there exist a subset $S$ of $[0, \xi)$ and ordinals $\alpha, \beta$ with $\alpha\oplus\beta=\xi$, and such that either $$\alpha\text{\ is a limit ordinal,}\hspace{5mm}\sup_{\zeta\in S}\alpha_\zeta=\alpha, \hspace{5mm} \min_{\zeta\in S}\beta_\zeta\geqslant \beta,$$ or $$\beta\text{\ is a limit ordinal,}\hspace{5mm} \sup_{\zeta\in S} \beta_\zeta=\beta, \hspace{5mm} \min_{\zeta\in S} \alpha_\zeta\geqslant \alpha.$$ In either case, $$x^*\in s^\alpha_\ee(L)+s^\beta_\ee(A)\subset \underset{\zeta\oplus \zeta=\xi}{\bigcup} [s^\zeta_\ee(L)+s^\eta_\ee(A)].$$ Indeed, in the first case, for every $\zeta\in S$, we may fix $y^*_\zeta\in s^{\alpha_\zeta}_\ee(L)$ and $z^*_\zeta\in S^{\beta_\zeta}_\ee(A)\subset s^\beta_\ee(A)$ such that $x^*=y^*_\zeta+z^*_\zeta$. By passing to a $w^*$-converging subnet $(y^*_\zeta)_{\zeta\in D}$ of $(y^*_\zeta)_{\zeta\in S}$, we note that the $w^*$-limit must lie in $\cap_{\zeta\in S}s^\zeta_\ee(L)=s^\alpha_\ee(L)$, and over the same subnet, $z^*_\zeta\underset{w^*}{\to} x^*-y^*\in s^\beta_\ee(A)$. Thus $x^*=y^*+z^*\in s_\ee^\alpha(L)+s_\ee^\beta(A)$. The other case is identical.
$(iii)$ For the final statement, note that since $s_{3\ee}^{\omega^\xi r}(L)=\varnothing$, $$s^{\omega^\xi r}_{12\ee}(B)\subset \underset{k=0}{\overset{r}{\bigcup}}[s^{\omega^\xi k}_{3\ee}(L)+s^{\omega^\xi(r- k)}_{3\ee}(A)]= \underset{k=0}{\overset{r-1}{\bigcup}}[s^{\omega^\xi k}_{3\ee}(L)+s^{\omega^\xi (r-k)}_{3\ee}(A)] \subset L+s^{\omega^\xi}_{3\ee}(A).$$
\end{proof}
\section{The upper estimates}
Recall that for $\gamma<CB(K)$, $\varphi_\gamma$ denotes the canonical inclusion of $C_{K^\gamma}(K)$ into $C(K)$, $C_{K^\gamma}(K)$ is isomorphic to $C_0(K/K^\gamma)$, and $CB(K/K^\gamma)=\gamma+1$. Moreover, $\Phi_\gamma$ denotes the canonical inclusion of $L_p(C_{K^\gamma}(K))$ into $L_p(C(K))$, and $L_p(C_{K^\gamma}(K))$ is isomorphic to $L_p(C_0(K/K^\gamma))$.
\begin{theorem} Given an ordinal $\xi$, consider the following statements: \begin{enumerate}[(i)]\item $A_\xi$: For any compact, Hausdorff space $K$ such that $CB(K)\leqslant \omega^\xi$, $Sz(C_0(K))\leqslant Sz(C(K))\leqslant \omega^\xi$. \item $B_\xi$: If $\gamma<\omega^\xi$, then for any compact, Hausdorff space $K$ such that $\gamma<CB(K)$, if $\varphi_\gamma:C_{K^\gamma}(K)\to C(K)$ denotes the inclusion, $Sz(\varphi_\gamma)\leqslant \omega^\xi$. \item $C_\xi$: For any compact, Hausdorff $K$ such that $\omega^\xi<CB(K)$, and any $\ee\in (0,1)$, $$\varphi_{\omega^\xi}^*(s^{\omega^\xi}_{3\ee}(B_{\ell_1(K)}))\subset (1-\ee)\varphi^*_{\omega^\xi} B_{\ell_1(K)}.$$ \end{enumerate}
For every ordinal $\xi$, $A_\xi$, $B_\xi$, $C_\xi$ hold.
\label{main}
\end{theorem}
\begin{theorem}
Given an ordinal $\xi$, consider the following statements: \begin{enumerate}[(i)]\item $A_\xi':$ For any compact, Hausdorff space $K$ such that $CB(K)\leqslant \omega^\xi$, $Sz(L_p(C_0(K)))\leqslant Sz(L_p(C(K)))\leqslant \omega^{1+\xi}$. \item $B_\xi':$ If $\gamma<\omega^\xi$, then for any compact, Hausdorff space $K$ such that $\gamma<CB(K)$, if $\Phi_\gamma:L_p(C_{K^\gamma}(K))\to L_p(C(K))$ denotes the inclusion, $Sz(\Phi_\gamma)\leqslant \omega^{1+\xi}$. \item $C_\xi':$ For any compact, Hausdorff $K$ such that $\omega^\xi<CB(K)$, and any $\ee\in (0,1)$, $$\Phi^*_{\omega^\xi}(s^{\omega^{1+\xi}}_{3\ee}(B_{L_q(\ell_1(K))}))\subset (1-\ee^q)^{1/q}\Phi^*_{\omega^\xi} B_{L_q(\ell_1(K))} .$$ \end{enumerate}
For every ordinal $\xi$, $A_\xi'$, $B_\xi'$, $C_\xi'$ hold.
\label{main thing}
\end{theorem}
The proofs of Theorems \ref{main} and \ref{main thing} are nearly identical. We will prove Theorem \ref{main thing}. In order to prove Theorem \ref{main}, one simply runs the same proof replacing Lemma \ref{lemma1}$(ii)$ and Proposition \ref{barry}$(ii)$, $(iv)$, and $(vi)$ with Lemma \ref{lemma1}$(i)$ and Proposition \ref{barry}$(i)$, $(iii)$, and $(v)$, respectively. This step of the proof is where our methods necessarily diverge from those used to prove upper estimates for $Sz(C(K))$ and $Sz(L_p(C(K)))$ when $K$ is countable. The only part of the proof which requires work will be to deduce $A_{\xi+1}$ from $C_\xi$. Given $C_\xi$, it is easy to deduce that if $CB(K)= \omega^\xi+1$, then $Sz(C(K))\leqslant \omega^{\xi+1}$, which is a particular case of $A_{\xi+1}$. It follows from Bessaga and Pe\l czy\'{n}ski's isomorphic classification of $C(K)$ spaces for countable $K$ that for two countable, compact, Hausdorff, infinite spaces $K,L$, $C(K)$ is isomorphic to $C(L)$ if and only if $\Gamma(K)=\Gamma(L)$. Therefore for countable $K$, in order to deduce $A_{\xi+1}$ from $C_\xi$, it is sufficient to only consider the special case that $CB(K)=\omega^\xi+1$. However, this isomorphic classification fails for uncountable compact, Hausdorff spaces, and even for uncountable intervals of ordinals. More specifically, $[0, \omega_1]$ and $[0, \omega_1 2]$ have the same Cantor-Bendixson index, $\omega_1+1$, while the spaces $C([0, \omega_1])$ and $C([0, \omega_1 2])$ are not isomorphic \cite{Semadeni}.
\begin{proof}[Proof of Theorem \ref{main thing}] We first prove that $A_\xi'\Rightarrow B_\xi'\Rightarrow C_\xi'$. Assume $A_\xi'$, $K$ is compact, Hausdorff, $\gamma<\omega^\xi$ and $\gamma<CB(K)$. Then if $\Phi_\gamma:L_p(C_{K^\gamma}(K))\to L_p( C(K))$ is the inclusion, $$Sz(\Phi_\gamma)=Sz(L_p(C_{K^\gamma}(K)))=Sz(L_p(C_0(K/K^\gamma))),$$ and $CB(K/K^\gamma)=\gamma+1\leqslant \omega^\xi$. By $A_\xi'$, $Sz(\Phi_\gamma)=Sz(L_p(C_0(K/K^\gamma)))\leqslant \omega^{1+\xi}$. Thus $A_\xi'\Rightarrow B_\xi'$.
Next, assume $B_\xi'$ holds. Assume $f\in s^{\omega^{1+\xi}}_{3\ee}(B_{L_q(\ell_1(K))})$ and, to obtain a contradiction, assume $\|\Phi^*_{\omega^\xi} f\|^q>1-\ee^q$. If $\xi=0$, then by Proposition \ref{barry}$(iv)$, there exists a finite set $F$ of isolated points such that $\|\Phi^*_{K\setminus F} f\|^q>1-\ee^q$. If $\xi>0$, by Proposition \ref{barry}$(vi)$, there exists $\gamma<\omega^\xi$ such that $\|\Phi^*_\gamma f\|^q>1-\ee^q$. Then by Lemma \ref{lemma1}, $\Phi^*_{K\setminus F} f\in s_\ee^\omega(\Phi^*_{K\setminus F} B_{L_q(\ell_1(K))})$ if $\xi=0$, and $\Phi^*_\gamma f\in s_\ee^{\omega^{1+\xi}}(\Phi^*_\gamma B_{L_q(\ell_1(K))})$ if $\xi>0$. In the case that $\xi=0$, we obtain a contradiction to Proposition \ref{barry}$(ii)$, since in this case $s_\ee^\omega(\Phi^*_{K\setminus F} B_{L_q(\ell_1(K))})=\varnothing$. In the case that $\xi>0$, we obtain a contradiction to $B_\xi'$, since $Sz(\Phi_\gamma)\leqslant \omega^{1+\xi}$, so $s_\ee^{\omega^{1+\xi}}(\Phi^*_\gamma B_{L_q(\ell_1(K))})=\varnothing$. In either case, the contradiction yields that $\|Q_{K^{\omega^\xi}} f\|=\|\Phi^*_{\omega^\xi} f\|\leqslant (1-\ee^q)^{1/q}$. Then $$\Phi^*_{\omega^\xi} f= \Phi^*_{\omega^\xi}Q_{K^{\omega^\xi}}f\in (1-\ee^q)^{1/q} \Phi^*_{\omega^\xi} B_{L_q(\ell_1(K))}.$$ Thus $A_\xi'\Rightarrow B_\xi'\Rightarrow C_\xi'$.
We remark that $C'_\xi$ is equivalent to: For any compact, Hausdorff $K$ such that $\omega^\xi<CB(K)$, any $\ee\in (0,1)$, and any $a\in (0,1]$, $$\Phi^*_{\omega^\xi}(s^{\omega^{1+\xi}}_{3\ee}(a B_{L_q(\ell_1(K))}))\subset a(1-\ee^q)^{1/q}\Phi^*_{\omega^\xi} B_{L_q(\ell_1(K))} .$$ Indeed, by homogeneity, if $X$ is any Banach space, $L\subset X^*$ is $w^*$-compact, $\ee>0$, $a\in (0,1]$, and $\xi$ is any ordinal, $s^\xi_\ee(a L)=a s^\xi_{\ee/a}(L)\subset a s^\xi_\ee(L)$. We apply this with $L=B_{L_q(\ell_1(K))}$ to deduce that $$\Phi^*_{\omega^\xi}(s^{\omega^{1+\xi}}_{3\ee}(a B_{L_q(\ell_1(K))}))= a\Phi^*_{\omega^\xi}(s^{\omega^{1+\xi}}_{3\ee/a}( B_{L_q(\ell_1(K))}))\subset a(1-\ee^q)^{1/q}\Phi^*_{\omega^\xi} B_{L_q(\ell_1(K))} .$$
We turn now to the proof of the theorem. Since $A_\xi'\Rightarrow B_\xi'\Rightarrow C_\xi'$, it is sufficient to prove that for any ordinal $\xi$, $A_\xi'$ holds given $A_\zeta', B'_\zeta$, $C'_\zeta$ for every $\zeta<\xi$. Seeking a contradiction, assume there exists an ordinal $\xi$ such that $A_\xi'$ fails, and assume that $\xi$ is the minimum such ordinal. We note that $\xi>0$, since $A_0'$ is true. Indeed, if $CB(K)\leqslant \omega^0=1$, $K$ is finite, whence $L_p(C(K))\approx L_p$ and $Sz(L_p)=\omega$.
Assume $\xi$ is a limit ordinal. Fix a compact, Hausdorff space $K$ with $CB(K)\leqslant \omega^\xi$. Since $CB(K)$ cannot be a limit ordinal, $CB(K)<\omega^\xi$. Since $\xi$ is a limit ordinal, there exists $\zeta<\xi$ such that $CB(K)<\omega^\zeta$. Since $A_\zeta$ holds, we deduce that $Sz(L_p(C(K)))\leqslant \omega^{1+\zeta}< \omega^{1+\xi}$, a contradiction. Thus $\xi$ cannot be a limit ordinal.
It follows that $\xi$ must be a successor, say $\xi=\zeta+1$. Since $\omega^\xi$ is a limit ordinal, $CB(K)=\omega^\xi$ is impossible for any compact, Hausdorff $K$, so it follows that there exists some compact, Hausdorff $K$ such that $CB(K)<\omega^\xi$ and such that the statement of $A_\xi'$ fails for this $K$. Let $n$ be the minimum natural number such that there exists a compact, Hausdorff space $K$ with $CB(K)\leqslant \omega^\zeta n+1$ and such that the $A_\xi'$ fails for this $K$. First suppose that $n=1$. Then $CB(K)\leqslant \omega^\zeta+1$. In this case it must be that $CB(K)=\omega^\zeta+1$, since if $CB(K)\leqslant \omega^\zeta$, we would deduce that $Sz(L_p(C(K)))\leqslant \omega^{1+\zeta}<\omega^{1+\xi}$ by $A_\zeta'$. Since $K$ is infinite, it must be that $Sz(L_p(C_0(K)))=Sz(L_p(C(K)))$, so we compute $Sz(L_p(C_0(K)))$. Note that since $K_\infty=K^{\omega^\zeta}$, $\Phi_{\omega^\zeta}$ is the inculsion of $L_p(C_0(K))$ into $L_p(C(K))$. By $C_\zeta'$, for any $\ee\in (0,1)$, $$\Phi^*_{\omega^\zeta} s_{3\ee}^{\omega^{1+\zeta}}(B_{L_q(\ell_1(K))})\subset (1-\ee^q)^{1/q}\Phi^*_{\omega^\zeta} B_{L_q(\ell_1(K))}.$$ By \cite[Lemma $2.5$]{BrookerAsplund}, $$s_{6\ee}^{\omega^{1+\zeta}}(\Phi^*_{\omega^\zeta} B_{L_q(\ell_1(K))}) \subset \Phi^*_{\omega^\zeta} s_{3\ee}^{\omega^{1+\zeta}}(B_{L_q(\ell_1(K))}).$$ These inclusions together with a standard homogeneity argument yield that for any $j\in \nn$, and $\ee\in (0,1)$, $$s_{6\ee}^{\omega^{1+\zeta} j }(\Phi^*_{\omega^\zeta} B_{L_q(\ell_1(K))})\subset (1-\ee^q)^{j/q} \Phi^*_{\omega^\zeta} B_{L_q(\ell_1(K))},$$ from which it follows that $Sz_{6\ee}(\Phi^*_{\omega^\zeta} B_{L_q(\ell_1(K))})<\omega^{1+\zeta}\omega=\omega^{1+\xi}$. This shows that $Sz(\Phi_{\omega^\zeta})\leqslant \omega^{1+\xi}$. Since $Sz(\Phi_{\omega^\zeta})=Sz(L_p(C_0(K)))=Sz(L_p(C(K)))$, we reach a contradiction if $n=1$.
Since it must be that $n>1$, we may write $n=m+1$. Let $F=K^{\omega^\zeta}$ and note that $CB(F)\leqslant \omega^\zeta m+1$. Indeed, $$F^{\omega^\zeta m+1} = (K^{\omega^\zeta})^{\omega^\zeta m+1}= K^{\omega^\zeta+\omega^\zeta m+1}= K^{\omega^\zeta n+1}=\varnothing.$$ Let $L=B_{L_p(C(F))^*}=B_{L_q(\ell_1(F))}\subset B_{L_q(\ell_1(K))}$. Recall that viewing $B_{L_q(\ell_1(F))}$ as the unit ball of $L_p(C(F))^*$ as well as the subset of $B_{L_q(\ell_1(K))}$ consisting of measures supported on $F$ is a $w^*$-$w^*$-continuous linear isometry, so that $L\subset B_{L_q(\ell_1(K))}$ is $w^*$-compact and $Sz_{3\ee}(L)=Sz_{3\ee}(B_{L_q(\ell_1(F))})=Sz_{3\ee}(B_{L_p(C(F))^*})\leqslant \omega^{1+\xi}$ by the minimality of $n$ and the fact that $CB(F)\leqslant \omega^\zeta m+1$. But the usual compactness argument yields that if $Sz_{3\ee}(L)\leqslant \omega^{1+\xi}$, then $Sz_{3\ee}(L)<\omega^{1+\xi}=\omega^{1+\zeta}\omega$, and there exists some $r\in \nn$ such that $Sz_{3\ee}(L)<\omega^{1+\zeta} r$. We claim that for $i=0, 1, \ldots$ and $f\in s_{12\ee}^{\omega^{1+\zeta}ri} (B_{L_q(\ell_1(K))})$, $\|\Phi^*_F f\|\leqslant (1-\ee^q)^{i/q}$. The $i=0$ case is obvious. Assume we have the result for some $i$. Let $A=(1-\ee^q)^{i/q}B_{L_q(\ell_1(K))}$. Note that our assumption on $i$ yields that $$s^{\omega^{1+\zeta}ri}_{12\ee}(B_{L_q(\ell_1(K))})=\{P_F f+Q_F f: f\in s^{\omega^{1+\zeta}ri}_{12\ee}(B_{L_q(\ell_1(K))})\} \subset L+A.$$ Here we are using the fact that $\|P_F\|\leqslant 1$, so that $P_F$ maps $B_{L_q(\ell_1(K))}$ into $L$. By Lemma \ref{smooth}, $$s^{\omega^{1+\zeta} r(i+1)}_{12\ee}(B_{L_q(\ell_1(K))}) \subset s^{\omega^{1+\zeta}}_{12\ee}\Bigl(s^{\omega^{1+\zeta}ri}_{12\ee}(B_{L_q(\ell_1(K))})\Bigr)\subset s^{\omega^{1+\zeta}}_{12\ee}(L+A)\subset L+s^{\omega^{1+\zeta}}_{3\ee}(A).$$ Using the equivalent version of $C_\zeta'$ given above with $a=(1-\ee^q)^{i/q}$, \begin{align*} \Phi^*_{\omega^\zeta} s^{\omega^{1+\zeta}}_{3\ee}(A) & = \Phi^*_{\omega^\zeta} s^{\omega^{1+\zeta}}_{3\ee}(a B_{L_q(\ell_1(K))})\subset a(1-\ee^q)^{1/q}\Phi^*_{\omega^\zeta} B_{L_q(\ell_1(K))} \\ & =(1-\ee^q)^{\frac{i+1}{q}}\Phi^*_{\omega^\zeta} B_{L_q(\ell_1(K))}.\end{align*} Fix $f\in s^{\omega^{1+\zeta} r(i+1)}_{12\ee}(B_{L_q(\ell_1(K))})\subset L+s^{\omega^{1+\zeta}}_{3\ee}(A)$ and write $f=g+h$ with $g\in L$ and $h\in s^{\omega^{1+\zeta}}_{3\ee}(A)$. Then since $\Phi^*_F|L\equiv 0$, $\Phi^*_F g=0$. This fact together with our inclusion above yields that $\|\Phi^*_F f\|= \|\Phi^*_F h\| \leqslant (1-\ee^q)^{\frac{i+1}{q}}.$ This yields the inductive claim on $i$.
Fix $j\in \nn$ such that $(1-\ee^q)^{j/q}<\ee$. For any $f\in s^{\omega^{1+\zeta} rj}_{12\ee}(B_{L_q(\ell_1(K))})$, $f=P_{\omega^\zeta} f+Q_{\omega^\zeta} f$, $P_{\omega^\zeta} f\in L$, and $$\|Q_{\omega^\zeta} f\|= \|\Phi^*_{\omega^\zeta} f\|\leqslant (1-\ee^q)^{j/q}<\ee.$$ Therefore $$s^{\omega^{1+\zeta} rj}_{12\ee}(B_{L_q(\ell_1(K))}) \subset L+ \ee B_{L_q(\ell_1(K))}.$$ Using Proposition \ref{szlenk facts}$(vi)$, \begin{align*} s^{\omega^{1+\zeta}r(j+1)}_{12\ee}(B_{L_q(\ell_1(K))}) & \subset s_{12\ee}^{\omega^{1+\zeta}r}\bigl(s^{\omega^{1+\zeta} rj}_{12\ee}(B_{L_q(\ell_1(K))})\bigr)\subset s_{12\ee}^{\omega^{1+\zeta}r}(L+\ee B_{L_q(\ell_1(K))}) \\ & \subset s_{3\ee}^{\omega^{1+\zeta}r}(L)+\ee B_{L_q(\ell_1(K))}=\varnothing.\end{align*} But this shows that $Sz_{12\ee}(B_{L_q(\ell_1(K))})<\omega^{1+\zeta}r(j+1)<\omega^{1+\xi}$. Since $\ee\in (0,1)$ was arbitrary, we deduce that $Sz(B_{L_q(\ell_1(K))})\leqslant \omega^{1+\xi}$, and this contradiction finishes the proof.
\end{proof}
\section{The lower estimates}
\begin{lemma} Let $K$ be compact, Hausdorff and fix $\ee\in (0,1)$. For any ordinal $\xi<CB(K)$, $B_{\ell_1(K^\xi)}\subset s_\ee^\xi(B_{\ell_1(K)})\cap d_\ee^{\omega\xi}(B_{\ell_1(K)})$.
\label{lower est}
\end{lemma}
\begin{proof}We prove the results separately for the Szlenk and $w^*$-dentability indices, each by induction. We first prove that $B_{\ell_1(K^\xi)}\subset s_\ee^\xi(B_{\ell_1(K)})$. The base and limit ordinal cases are clear. Assume $K^{\xi+1}\neq \varnothing$ and $B_{\ell_1(K^\xi)}\subset s_\ee^\xi(B_{\ell_1(K)})$. Fix any finite, non-empty subset $F$ of $K^{\xi+1}$ and scalars $(a_x)_{x\in F}$ such that $\sum_{x\in F}|a_x|\leqslant 1$. Let $\mu=\sum_{x\in F}a_x\delta_x$ and let $V$ be any $w^*$-neighborhood of $\mu$. Fix $f_1, \ldots, f_n\in C(K)$, and $\sigma>0$ such that $$\{\eta\in \ell_1(K): (\forall 1\leqslant i\leqslant n)(|\langle \mu, f_i\rangle- \langle \eta, f_i\rangle|<\sigma)\}\subset V.$$ For each $x\in F$, fix a neighborhood $U_x$ of $x$ such that for each $1\leqslant i\leqslant n$ and each $y\in U_x$, $|f_i(x)-f_i(y)|<\sigma$. We may assume that the sets $(U_x)_{x\in F}$ are pairwise disjoint. Since the points $x\in F$ are not isolated in $K^\xi$, for each $x\in F$ we may fix $y_x\in U_x\cap K^\xi$ such that $y_x\neq x$. Let $E=\{y_x: x\in F\}$. Let $\eta=\sum_{x\in F} a_x \delta_{y_x}$ and note that $\eta\in V\cap B_{\ell_1(K^\xi)}\subset V\cap s_\ee^\xi(B_{\ell_1(K)})$. Since $K^{\xi+1}\neq \varnothing$, $K^\xi$ is infinite, and we may find some $\eta'\in \text{span}\{\delta_t: t\in K^\xi\setminus (E\cup F)\}\cap_{i=1}^n \ker(f_i)$. By scaling, we may assume $\|\eta'\|=1-\|\eta\|$. Then $\eta+\eta'\in V\cap B_{\ell_1(K^\xi)}$ and $$\|\mu-(\eta+\eta')\|=\|\mu\|+\|\eta\|+\|\eta'\|\geqslant 1>\ee.$$ Since $V$ was arbitrary, we deduce that $\mu\in s_\ee^{\xi+1}(B_{\ell_1(K)})$. This shows that the members of $B_{\ell_1(K^\xi)}$ with finite support lie in $s_\ee^{\xi+1}(B_{\ell_1(K)})$. Since such measures are dense in $B_{\ell_1(K^\xi)}$ and since $s_\ee^{\xi+1}(B_{\ell_1(K)})$ is $w^*$-closed, we deduce the successor case.
We next prove the statement concerning the $w^*$-dentability index, also by induction. Again, the base and limit cases are trivia. Assume $K^{\xi+1}\neq \varnothing$ and assume $B_{\ell_1(K^\xi)}\subset d^{\omega \xi}_\ee(B_{\ell_1(K)})$. For each $n=0, 1, \ldots$, let $$A_n=\Bigl\{\frac{1}{2^n}\sum_{t\in F}\ee_t\delta_t: |\ee_t|=1, F\subset K^\xi, |F|=2^n\Bigr\}\subset B_{\ell_1(K^\xi)}\subset d_\ee^{\omega\xi}(B_{\ell_1(K)}).$$ We claim that for every $m=0, 1, \ldots$ and every $n\geqslant m$, $A_n\subset d_\ee^{\omega\xi+m}(B_{\ell_1(K)})$. The inductive hypothesis yields the result for each $n$ when $m=0$. Assume that for some $m$ and every $n\geqslant m$, $A_n\subset d_\ee^{\omega\xi+m}(B_{\ell_1(K)})$. Fix some $n\geqslant m+1$ and some $\mu=\frac{1}{2^n}\sum_{t\in F}\ee_t\delta_t\in A_n$. Let $F_1, F_2$ be a partition of $F$ with $|F_1|=|F_2|=2^{n-1}$. Then $\mu_1:=\frac{1}{2^{n-1}}\sum_{t\in F_1}\ee_t\delta_t$, $\mu_2:=\frac{1}{2^{n-1}}\sum_{t\in F_2}\ee_t\delta_t\in A_{n-1}\subset d_\ee^{\omega\xi+m}(B_{\ell_1(K)})$. Moreover, $\mu=\frac{1}{2}\mu_1+\frac{1}{2}\mu_2$. Let $S$ be any $w^*$-open slice containing $\mu$, and note that this slice must contain either $\mu_1$ or $\mu_2$ by convexity. But $$\|\mu-\mu_1\|=\|\mu-\mu_2\|=\frac{1}{2}\|\mu_1-\mu_2\|=1>\ee,$$ so that $\text{diam}(S\cap d_\ee^{\omega \xi+m}(B_{\ell_1(K)}))>\ee$. From this it follows that $\mu\in d^{\omega\xi +m+1}_\ee(B_{\ell_1(K)})$. This yields the claim concerning $A_n$.
Next, we note that for any $x\in K^{\xi+1}$, any unimodular scalar $a$, and any $n\in \nn$, $a\delta_x\in \overline{A}_n^{w^*}$. Indeed, fix $f_1, \ldots, f_k\in C(K)$ and $\eta>0$. Fix a neighborhood $U$ of $x$ such that for every $y\in U$ and $1\leqslant i\leqslant k$, $|f_i(y)-f_i(x)|<\eta$. Fix any finite subset $F\subset K^\xi$ of $U$ with $|F|=2^n$, as we may, since $x$ is not isolated in $K^\xi$. Then if $\mu=\frac{1}{2^n}\sum_{t\in F}a\delta_t\in A_n$, $|\langle f_i, a\delta_x\rangle-\langle f_i, \mu\rangle|<\eta$ for $i=1, \ldots, k$. This yields that $a\delta_x\in \overline{A}^{w^*}_n\subset \overline{d^{\omega\xi+n}_\ee(B_{\ell_1(K)})}^{w^*}= d^{\omega\xi+n}_\ee(B_{\ell_1(K)})$. From this we deduce that $a\delta_x\in \cap_{n<\omega}\delta_\ee^{\omega\xi+n}(B_{\ell_1(K)})=d^{\omega\xi+\omega}_\ee(B_{\ell_1(K)})=d^{\omega(\xi+1)}_\ee(B_{\ell_1(K)})$. Since $d^{\omega(\xi+1)}_\ee(B_{\ell_1(K)})$ is $w^*$-closed and convex, it follows that $$B_{\ell_1(K^{\xi+1})} = \overline{\text{co}}^{w^*}\{a \delta_x:|a|=1, x\in K^{\xi+1}\}\subset d_\ee^{\omega(\xi+1)}(B_{\ell_1(K)}).$$
\end{proof}
\begin{corollary} For any compact, Hausdorff space $K$ and any $1<p<\infty$, $Sz(C(K))=\Gamma(K)$ and $Dz(C(K))=Sz(L_p(C(K)))=\omega \Gamma(K)$.
\end{corollary}
\begin{proof} We will use Proposition \ref{szlenk facts} throughout. Since $Dz(C(K))\leqslant L_p(C(K))$, it suffices to prove that $\omega\Gamma(K) \leqslant Dz(C(K))$ and $Sz(L_p(C(K)))\leqslant \omega\Gamma(K)$ in order to see that $Dz(C(K))=Sz(L_p(C(K)))=\omega\Gamma(K)$.
By Lemma \ref{lower est}, if $K$ is not scattered, $C(K)$, and therefore $L_p(C(K))$, is not Asplund. From this it follows that $$Sz(C(K))=Dz(C(K))=Sz(L_p(C(K)))= \omega\Gamma(K)=\infty.$$
Assume $K$ is scattered. In this case, $CB(K)$ is an ordinal and there exists an ordinal $\xi$ such that $\Gamma(K)=\omega^\xi$. By the definition of $\Gamma(K)$, $CB(K)\leqslant \omega^\xi$. From Theorems \ref{main} and \ref{main thing}, $Sz(C(K))\leqslant \omega^\xi= \Gamma(K)$ and $Sz(L_p(C(K)))\leqslant \omega^{1+\xi}=\omega\Gamma(K)$.
If $CB(K)=1$, $Sz(C(K))=1$ and $Dz(C(K))=\omega$, since $C(K)$ is finite-dimensional and non-zero. This yields the result in the trivial case that $K$ is finite. Otherwise $CB(K)=\zeta+1$ for some ordinal $\zeta$ and there exists an ordinal $\xi$ such that $\omega^\xi<\zeta+1<\omega^{\xi+1}=\Gamma(K)$. By Lemma \ref{lower est}, $B_{\ell_1(K^\zeta)}\subset s_{1/2}^\zeta(B_{\ell_1(K)})\cap d_{1/2}^{\omega\zeta}(B_{\ell_1(K)})$, and it follows that $Sz_{1/2}(B_{\ell_1(K)})>\zeta>\omega^\xi$ and $Dz_{1/2}(B_{\ell_1(K)})>\omega\zeta >\omega^{1+\xi}$. Therefore we deduce that $\omega^\xi<Sz(C(K))\leqslant \omega^{\xi+1}$, and since $Sz(C(K))=\omega^\gamma$ for some ordinal $\gamma$, $Sz(C(K))=\omega^{\xi+1}$. We deduce from $\omega^{1+\xi}<Dz(C(K))\leqslant \omega^{1+\xi+1}$ that $Dz(C(K))\geqslant \omega^{1+\xi+1}$ similarly.
\end{proof}
|
\section{INTRODUCTION}
Our knowledge of the evolution of galaxies has been almost exclusively acquired by large-scale surveys using ground and space-based telescopes
at optical and adjacent wavelengths \citep[e.g.,][]{York:2000,Colless:2001,Martin:2005,Skrutskie:2006}. The information obtained from these surveys is mainly valuable for tracing the evolution of the stellar content of galaxies. In contrast, little is known about the evolution of the (cold) gaseous component in galaxies. In particular, our knowledge of the neutral atomic hydrogen gas ({\hbox{\rmfamily H\,{\textsc i}}}) content (i.e. the raw material for star formation) of individual galaxies is essentially limited to the very low redshift universe \citep{Meyer:2004,Haynes:2011}.
Understanding the evolution of the atomic gas content of galaxies remains one of the key challenges in the study of galaxy evolution. The sensitivity of the current generation of radio telescopes is insufficient to detect {\hbox{\rmfamily H\,{\textsc i}}} from individual galaxies at cosmologically significant redshifts in reasonable integration times. Indeed, the detection of {\hbox{\rmfamily H\,{\textsc i}}} from individual galaxies at high redshifts was one of the original motivations, and remains one of the key science drivers for the proposed Square Kilometre Array \citep[SKA, e.g.,][]{Blyth:2015,Staveley-Smith:2015,Santos:2015}. It is also one of the key programs for several of the upcoming SKA pathfinder telescopes, for instance the Widefield ASKAP L-band Legacy All-sky Blind surveY \citep[WALLABY,][]{Koribalski:2009}, the Deep Investigation of Neutral Gas Origins \citep[DINGO,][]{Meyer:2009}, and Looking At the Distant
Universe with the MeerKAT Array \citep[LADUMA,][]{Holwerda:2012}. In combination with data at other wavelengths, this knowledge of {\hbox{\rmfamily H\,{\textsc i}}} gas
obtained from the SKA and its pathfinders will allow us to fully understand the complex processes that govern galaxy evolution \citep{Meyer:2015}.
Although it is challenging to detect {\hbox{\rmfamily H\,{\textsc i}}} in individual galaxies at $z \ga 0.2$ with the current generation of radio telescopes, it is possible to make measurements of the average {\hbox{\rmfamily H\,{\textsc i}}} content of a sample of galaxies. The volume of space probed in a single pointing and correlator setting typically contains many hundreds of galaxies. If the positions and redshifts of all of these galaxies are known, one can stack their spectra to determine their average {\hbox{\rmfamily H\,{\textsc i}}} content \citep{Chengalur:2001,Zwaan:2001}. This spectral stacking technique has been applied to interferometers such as the Giant Metrewave Radio Telescope (GMRT) and the Westerbork Radio Telescope (WSRT) as well as single dishes such as the Parkes to measure the {\hbox{\rmfamily H\,{\textsc i}}} content of galaxies at redshifts $z \la 0.4$, resulting in determination of the evolution of the gas content in galaxies (i.e. {\hbox{$\Omega_{\HIsub}$}}) out to redshifts
$z \sim 0.2$ \citep{Lah:2007, Lah:2009,Delhaize:2013,Rhee:2013}.
Here we apply the technique to GMRT observations of the COSMOS field to determine {\hbox{$\Omega_{\HIsub}$}} at a redshift of $z\sim 0.37$. The GMRT observation of
the COSMOS field plays an important role as a precursor of future {\hbox{\rmfamily H\,{\textsc i}}} deep surveys using SKA pathfinders. To check the viability of the surveys and develop suitable observing modes and strategies, a smaller-scale {\hbox{\rmfamily H\,{\textsc i}}} survey, as presented here, is a good testbed. Both expected and unexpected issues relevant to wide and deep {\hbox{\rmfamily H\,{\textsc i}}} surveys can be explored through such a pilot study. The redshift studied ($z \sim$~0.37 ) allows us to understand how the {\hbox{\rmfamily H\,{\textsc i}}} gas in galaxies evolves out to $z \sim$~0.4, and perhaps beyond.
This paper is structured as follows: In Section~2 we detail the optical data of the COSMOS field. Section~3 describes the GMRT observations, data quality and data reduction procedures. We explain the stacking analysis used to measure the average {\hbox{\rmfamily H\,{\textsc i}}} gas mass at $z \sim$~0.37 in Section~4. The main results are presented in Section~5. In Section~6, we discuss the implications of our observations for the cosmic evolution of {\hbox{$\Omega_{\HIsub}$}}. We present the summary and conclusions in Section~7. Throughout the paper we adopt the concordance cosmological parameters: $\Omega_{\Lambda}=$~0.7, $\Omega_{M}=$~0.3 and $H_{0}=$~70~km~s$^{-1}$~Mpc$^{-1}$.
\section{Data}
\subsection{Target Field Selection}
Our radio observations are centred on the COSMOS field. This field has a wealth of multi-wavelength data ranging from X-ray to radio \citep{Scoville:2007}. Spectroscopic redshifts are also available for a very large number of galaxies in the COSMOS field \citep[zCOSMOS,][]{Lilly:2007}. This makes it an excellent target for {\hbox{\rmfamily H\,{\textsc i}}} studies. Indeed it has also been selected as the target for a very deep JVLA\footnote{The Karl G. Jansky Very Large Array} {\hbox{\rmfamily H\,{\textsc i}}} survey \citep[CHILES,][]{Fernandez:2013}.
\begin{table}
\centering
\caption[COSMOS photometry band filters and their zero-point offsets]{Photometric band filters and their zero-point offsets in the COSMOS photometry catalogue taken from \citet{Ilbert:2009}. All magnitudes of our sample in these photometric bands were used as inputs for the SED fitting procedure with {\sc le phare}.}
\begin{tabular}{@{}lcccr}
\hline
Filter & Telescope & $\lambda_{\rm eff}$ & FWHM & offset \\
& & ($\AA$) & ($\AA$) & \\
\hline
FUV & $GALEX$ & 1551.3 & 230.8 & 0.314 \\
NUV & $GALEX$ & 2306.5 & 789.1 & -0.022 \\
$u^{*}$ & CFHT & 3911.0 & 538.0 & 0.054 \\
$B_{J}$ & Subaru & 4439.6 & 806.7 & -0.242 \\
$V_{J}$ & Subaru & 5448.9 & 934.8 & -0.094 \\
$g^{+}$ & Subaru & 4728.3 & 1162.9 & 0.024 \\
$r^{+}$ & Subaru & 6231.8 & 1348.8 & 0.003 \\
$i^{+}$ & Subaru & 7629.1 & 1489.4 & 0.019 \\
$i^{*}$ & CFHT & 7628.9 & 1460.0 & -0.007 \\
$z^{+}$ & Subaru & 9021.6 & 955.3 & -0.037 \\
$J$ & UKIRT & 12444.1 & 1558.0 & 0.124 \\
$K_{s}$ & KPNO/CTIO & 21434.8 & 3115.0 & 0.022 \\
$K$ & CFHT & 21480.2 & 3250.0 & -0.051 \\
$IA$427 & Subaru & 4256.3 & 206.5 & 0.037 \\
$IA$464 & Subaru & 4633.3 & 218.0 & 0.013 \\
$IA$484 & Subaru & 4845.9 & 228.5 & 0.000 \\
$IA$505 & Subaru & 5060.7 & 230.5 & -0.002 \\
$IA$527 & Subaru & 5258.9 & 242.0 & 0.026 \\
$IA$574 & Subaru & 5762.1 & 271.5 & 0.078 \\
$IA$624 & Subaru & 6230.0 & 300.5 & 0.002 \\
$IA$679 & Subaru & 6778.8 & 336.0 & -0.181 \\
$IA$709 & Subaru & 7070.7 & 315.5 & -0.024 \\
$IA$738 & Subaru & 7358.7 & 323.5 & 0.017 \\
$IA$767 & Subaru & 7681.2 & 364.0 & 0.041 \\
$IA$827 & Subaru & 8240.9 & 343.5 & -0.019 \\
$NB$711 & Subaru & 7119.6 & 72.5 & 0.014 \\
$NB$816 & Subaru & 8149.0 & 119.5 & 0.068 \\
\hline
\end{tabular}
\label{tab:cosmos_cat}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=150mm]{z_diagram_ra_new.pdf}
\caption[Redshift cone diagram of zCOSMOS field]{Redshift cone diagram of zCOSMOS 10k-bright sample to $z \sim$~1. Red points denote the 506 galaxies covered by GMRT beam and frequency at $z \sim$~0.37.}
\label{fig:z_diagram}
\end{figure*}
\subsection{Photometric Data}
Optical positions and redshifts are crucial inputs required for {\hbox{\rmfamily H\,{\textsc i}}} stacking analysis. Multi-band photometric data is also very important, because it allows one to determine the cosmic {\hbox{\rmfamily H\,{\textsc i}}} density as well as the dependence of {\hbox{\rmfamily H\,{\textsc i}}} gas content on other galaxy properties such as the morphological or spectral type, etc. Photometry of the galaxies in the COSMOS field is available over a very large range of wavelengths: X-ray with {\it XMM-Newton} \citep{Hasinger:2007}, UV with {\it GALEX} \citep{Zamojski:2007}, optical/NIR with CFHT, CTIO, KPNO, Subaru \citep{Taniguchi:2007} and {\it Hubble Space Telescope} \citep[{\it HST},][]{Koekemoer:2007}, mid-infrared with the {\it Spitzer} space telescope \citep{Sanders:2007}, mm/sub-millimetre with the Caltech Submillimeter Observatory (CSO) and Institut de Radioastronomie Millim\'{e}trique (IRAM) telescope \citep{Bertoldi:2007} and radio continuum with the Very Large Array \citep[VLA,][]{Schinnerer:2004, Schinnerer:2007}.
We cross-matched the zCOSMOS redshift catalogue with the publicly released photometric catalogue of the COSMOS field from the {\it IRSA} website\footnote{NASA/IPAC Infrared Science Archive, {\url {http://irsa.ipac.caltech.edu/data/COSMOS/}}}. This provides observed magnitudes in 27 photometric bands covered by CFHT, Subaru, KPNO/CTIO, and {\it GALEX}. This catalogue is an update on the previous optical/NIR catalogue by \citet{Capak:2007}. The updated photometric catalogue was compiled using the same point-spread function (PSF) from $u^*$ to the $K$ band. The photometry was derived over the same aperture of 3\arcsec diameter centred on the position of $i^{+}$ and $i^{*}$ bands. For other bands such as FUV and NUV, wide aperture total fluxes were measured first and then converted to 3\arcsec aperture flux. This consistency in measuring photometry allows one to determine accurate colours, leading to small uncertainties in the measurement of the spectral energy distribution (SED) fitting for {\it k}-correction, galaxy classification and measuring the stellar mass. A more detailed description of the photometric catalogue can be found in \citet{Capak:2007,Ilbert:2009}.
Table~\ref{tab:cosmos_cat} lists the photometric bands that we used in our analysis. We corrected all magnitudes for the zero-point offset listed in Table~\ref{tab:cosmos_cat} following \citet{Ilbert:2009} and then applied the corrections of the Galactic dust extinction using the dust map of \citet{Schlegel:1998}. {\it k}-corrections were also applied for all magnitudes using values obtained through a $\chi^{2}$ template fitting procedure.
\subsection{Spectroscopic Data}
Our spectroscopic data is taken from the zCOSMOS \citep{Lilly:2007} survey. The zCOSMOS \citep{Lilly:2007} is a large optical redshift survey undertaken in the COSMOS field using the VIMOS spectrograph mounted on the VLT at the European Southern Observatory (ESO), Chile.
The main goal of the survey was to trace the large-scale structure of the Universe up to $z \sim$~1 and to characterise galaxy groups and clusters. The zCOSMOS survey consists of two distinct parts, called zCOSMOS-bright and zCOSMOS-deep. The zCOSMOS-bright is a magnitude-limited survey ($I_{AB} <$~22.5 measured in the $I$-band of {\it HST} ACS), targeting $\sim$20,000 galaxies in the redshift range of 0.1~$< z <$~1.2 (see Fig.~\ref{fig:z_diagram}). This survey was undertaken on the entire 1.7~deg$^{2}$ COSMOS field. The zCOSMOS-deep surveys $\sim$10,000 galaxies, colour-selected to be in 1.4~$< z <$~3.0, in the central 1~deg$^{2}$ of the COSMOS field.
In this paper the spectroscopic data used comes from the zCOSMOS-bright catalogue which has spectra for 10,644 objects, the so-called 10k-bright sample. These contain a statistically complete subset of 10,109 objects. Spectra obtained from the zCOSMOS-bright survey cover a wavelength range of approximately 5550 to 9450~\AA, yielding a spectral resolution of $R \sim$~600 sampled at $\sim$2.5~\AA~pixel$^{-1}$. The velocity uncertainty of the zCOSMOS-bright redshifts is $\sim$110~\hbox{km\,s$^{-1}$}. For more details about data and data reduction, refer to \citet{Lilly:2007,Lilly:2009}. The zCOSMOS 10k-bright catalogue provides redshifts along with confidence classes indicating reliability of its redshift measurements. The confidence classes \citep[see Table~1 in][]{Lilly:2009} vary from class 0 (no redshift obtained) to class 4 (most secure redshift) with additional class 9 for one-line redshifts where the line is believed to be either [OII] or {\hbox{\rmfamily H\,$\alpha$}}. For the analysis in this paper, we restrict the sample to galaxies with the most secure redshift, i.e. those in class of 3 or 4. This selection produces a sample of 506 redshifts with reliable redshifts that lie within the GMRT data cube ($\sim$1~deg$^{2}$ at the observed frequency 1040 MHz) and the redshift range of 0.35~$< z <$~0.39 (see also Fig.~\ref{fig:object_distribution}).
\begin{figure*}
\centering
\includegraphics[width=84mm]{plot_zcosmos_CCD_morph_final.pdf}
\includegraphics[width=84mm]{plot_zcosmos_CCD_phot_final.pdf}
\caption[colour--colour diagrams of subsamples classified by morphological and spectro-photometric approaches]{colour--colour diagrams of subsamples classified by morphological ({\it left} panels) and spectro-photometric ({\it right} panels) approaches. The upper panels are $NUV-r_+$ and $r_{+}-J$ colour--colour diagrams and the lower panels are $u^{*}-V$ and $V-J$ colour--colour diagrams. In each panel on the left, morphological subclasses of early-types, spirals and irregulars are shown in red circles, green diamonds and blue squares, respectively. Four subsamples of the spectro-photometric approach from Type~1 to Type~4 are denoted on the right panels by red circles, green triangles, blue diamonds and squares.}
\label{fig:ccd_diagram}
\end{figure*}
\subsection{Galaxy Classification}
We used two different methods to classify our sample galaxies: one based on galaxy morphology and the other on spectro-photometry. The COSMOS field has the {\it HST} ACS imaging data \citep{Koekemoer:2007} with sufficient depth and resolution to perform morphological analysis. The COSMOS archival database provides a morphological class catalogue, which is based on applying an automatic and objective morphological classification technique to high-quality {\it HST} images \citep{Cassata:2007,Tasca:2009}. The morphological classification divides the galaxies into three morphological classes, i.e., early-types including ellipticals and lenticulars, spirals, and irregulars.
\begin{figure}
\centering
\includegraphics[width=86mm]{plot_compare_morph_phot.pdf}
\caption{The spectro-photometric classes of galaxies morphologically classified. Red ($\slash$), green, and blue ($\backslash$) bars denote morphological classes--early, spiral, and irregular, respectively.}
\label{fig:comp_morph_phot}
\end{figure}
The spectro-photometric classification is based on matching the rest-frame magnitudes and colours to a set of templates. Here, {\sc le phare}\footnote{\url {http://www.cfht.hawaii.edu/~arnouts/lephare.html}} \citep{Arnouts:1999,Ilbert:2006}, a $\chi^{2}$ template fitting code, using 20 photometric magnitudes from the COSMOS photometry catalogue, was applied in combination with spectroscopic redshifts from the zCOSMOS catalogue for each galaxy in the sample. This spectro-photometric classification divides our sample galaxies into four types; early-type E/S0 (Type~1), early spirals Sa/Sb (Type~2), late spirals Sc/Sd (Type~3), and irregular and starburst galaxies (Type~4).
\begin{figure}
\centering
\includegraphics[width=86mm]{Blue_early.pdf}
\includegraphics[width=86mm]{Spiral_Type1.pdf}
\caption{{\it HST} cutout images of blue early-type ({\it upper row}) and red spiral ({\it lower row}) examples. The size of each thumbnail image is 6\arcsec by 6 \arcsec. The images were obtained using G10/COSMOS image cutout tool.}
\label{fig:HST_cutout}
\end{figure}
Following \citet{Williams:2009,Ilbert:2009} we used $u^{*}-V$ vs. $V-J$ and $NUV-r_+$ vs. $r_{+}-J$ colour--colour diagrams to compare these two classification schemes. In Fig.~\ref{fig:ccd_diagram}, the left-hand panels plot the morphological classes of the galaxies, while the right-hand panels show the spectro-photometric classes. As can be seen, the morphologically classified early-types and spirals are not sharply separated in these diagrams. In particular, there are galaxies which are morphologically classified as early type but are blue in colour. Fig.~\ref{fig:ccd_diagram} shows that spectro-photometry provides a much cleaner separation. Fig.~\ref{fig:comp_morph_phot} summarises the leakage of the morphological classes between the spectro-photometric classes.
Visual inspection of the {\it HST} images of the blue early types shows that some of them have an obvious spiral or ring feature as seen in the upper row of Fig.~\ref{fig:HST_cutout}. This indicates that the uncertainty of the morphological classification can affect the {\hbox{\rmfamily H\,{\textsc i}}} stacking analysis for different galaxy types that we do below. Although the spectro-photometry clearly separates the early types, there exist Type~1 galaxies having spiral morphology (called ``red spirals''). These galaxies look highly-inclined or edge-on spirals in their {\it HST} images (see the lower row of Fig.~\ref{fig:HST_cutout}). Their high inclination can cause more internal reddening by dust, resulting in redder colour and mis-matched template. These galaxies have a negligible effect on the {\hbox{\rmfamily H\,{\textsc i}}} stacking analysis because the fraction of such galaxies is below 5~\%. We have adopted the spectro-photometric classification in our further analysis.
\section{GMRT Radio Data}
\subsection{Observation}
\begin{figure*}
\centering
\includegraphics[width=130mm]{plot_zcosmos_fov.pdf}
\caption[Distribution of objects in the zCOSMOS catalogue]{The spatial distribution of objects in the zCOSMOS catalogue. The blue circles denote galaxies covered by the GMRT beam and frequency. The solid and the dashed circles indicate FWHM (32.25~arcmin) and 10~per~cent level (58.8~arcmin) of the GMRT primary beam at 1040~MHz, respectively.}
\label{fig:object_distribution}
\end{figure*}
The zCOSMOS field was observed for a total of 134~hours using the GMRT. The observations were conducted over 20 days spread over the years 2008 and 2009. The total observation time includes 115~hours of on-source time with the remaining time spent on calibrator scans. The central frequency of the GMRT observations was 1040~MHz and the bandwidth was 32 MHz corresponding to an {\hbox{\rmfamily H\,{\textsc i}}} redshift range of 0.345~$< z <$~0.387. The observations were done using the old hardware correlator which divided the 32~MHz bandwidth into two 16~MHz-wide sidebands. Each sideband had two polarisations and 128 spectral channels, giving a channel width of 0.125~MHz ($\sim$36.3 \hbox{km\,s$^{-1}$} at $z =$ 0.37). The pointing centre of the GMRT observations was R.A.~10$^{\rm h}$00$^{\rm m}$10$\rm \fs$01~Dec.~+02$^{\rm d}$19$^{\rm m}$19$\rm \fs$95 (J2000). The primary beam size (FWHM) of the GMRT is approximately 32.25~arcmin at 1040~MHz and the 10$\%$ beam level is 58.8~arcmin which was the limit used for selecting galaxies for the stacking analysis. Observations of 3C~48, 3C~147 and 3C~286 were used to calibrate the flux density scale. 0943-083 served as a phase calibrator.
\begin{figure}
\centering
\includegraphics[width=84mm]{plot_rfi_zcosmos.pdf}
\caption[Fraction of flagged data]{The fraction of flagged data after all data reduction procedures.}
\label{fig:rfi_frac}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=120mm]{contmap_bw_new.pdf}
\caption[The GMRT continuum image of the COSMOS field]{The central 30$\arcmin$ area of the entire GMRT continuum image of the COSMOS field.}
\label{fig:contmap_zcosmos}
\end{figure*}
\subsection{Data Reduction}
The GMRT data reduction of the COSMOS field followed a standard reduction procedure including flagging, calibration and imaging. The GMRT data were first processed using {\sc flagcal} \citep{Prasad:2012, Chengalur:2013}, an automated flagging and calibration software developed for the GMRT data. As seen in Fig.~\ref{fig:rfi_frac}, the frequency range of the COSMOS field covered by the GMRT was affected by radio frequency interference (RFI); there were also several instrumental malfunctions during the observations that caused several antennas to be unusable. The edges of each sideband were completely flagged, which lead to the peak around 1040~MHz in Fig.~\ref{fig:rfi_frac} where 100 per~cent of the data has been flagged.
Data near 1030~MHz appear to be severely contaminated by RFI. It is known that aircrafts often generate RFI at this frequency. In addition, the GMRT front-end system had a limited dynamic range during the time when these observations were done, which can generate intermodulation products from strong RFI bursts. This system has since been significantly improved. About half the data in each sideband has been flagged during iterations
through the data reduction process.
Subsequent processing was done using the Common Astronomy Software Applications ({\sc casa})\footnote{{\url {http://casa.nrao.edu}}} package. If necessary, additional flagging was carried out manually, using {\sc casa} plotting and editting tools. Broad band flux and phase calibration was done using the central 80 channels, and the bandpass was determined using scans of both the flux and phase calibrators. After the calibration solutions were determined and applied, the COSMOS field datasets for each day were split and re-inspected separately as continuum images to check whether there remained any bad data appearing as artefacts in the images. Data for the 20 observing runs were then concatenated separately for each sideband. By design \citep{Scoville:2007} the COSMOS field does not contain any bright radio sources. We found that self-calibration did not significantly improve the quality of the images, and hence the final analysis was done using images without any self-calibration.
\subsection{Continuum Image}
To make the final continuum image of the COSMOS field at $z \sim$~0.37, the central 100 channels from each sideband (lower and upper sideband) were selected and concatenated. In order to avoid bandwidth smearing, each sideband used 100 channels subdivided into 10 channel averages during the imaging. A $\sim$1$\times$1~deg$^2$ continuum image was made with pixel size 0.9\arcsec~pixel$^{-1}$, ``robust'' weighting \citep{D_Briggs:1995} with a robust value of 0, and the $w$-projection \citep{Cornwell:2008} algorithm. The rms noise in the central regions of the final continuum (Fig.~\ref{fig:contmap_zcosmos}) is $\sim$12.3 $\mu$Jy~beam$^{-1}$ with a synthesised beam resolution of 3.5\arcsec $\times$ 2.4\arcsec. The astrometric accuracy of the GMRT imaging was determined by comparing the positions of sources detected in this image with high signal-to-noise ratio ( $\geq$ 5 $\sigma$) against the positions listed in the VLA FIRST survey catalogue \citep{Becker:1995}. The average positional offset found was $\sim$0.68 arcsec which is less than the image pixel size of 0.9 arcsec. The maximum measured astrometric offset was 1.8 arcsec, which is still significantly smaller than the GMRT synthesised beam.
\subsection{Line Data Cube}
The spectral data cubes for the two sidebands were made with the same pixel size, robust weighting and wide-field imaging algorithm as were used to make the continuum image. The synthesised beam of the data cubes is 3.5\arcsec $\times$ 2.4\arcsec, corresponding to $\sim$17.9~$\times$~12.0 kpc$^{2}$ at $z \sim$~0.37. The final spectral data cubes of each sideband were made by subtracting the continuum from these data cubes. For continuum subtraction, clean components of individual continuum sources were subtracted from the $uv$ data using the {\sc casa} task `\texttt{uvsub}'and any residual continuum flux was removed in the `$image$' domain by the {\sc casa} task `\texttt{imcontsub}'. The rms noise levels per frequency channel of each sideband are $\sim$140~$\mu$Jy~beam$^{-1}$ (lower sideband) and $\sim$117~$\mu$Jy~beam$^{-1}$ (upper sideband), respectively.
\section{Stacked {\hbox{\rmfamily H\,{\textsc i}}} Emission and {\hbox{\rmfamily H\,{\textsc i}}} Mass Measurements}
\begin{figure}
\centering
\includegraphics[width=84mm]{plot_redshift_hist_new.pdf}
\caption[The redshift distribution of COSMOS sample observed by the GMRT]{The redshift distribution of the COSMOS sample observed by the GMRT. Blue and red hatch areas are galaxies that lie in the upper and lower sidebands, respectively.}
\label{fig:z_hist}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=120mm]{plot_stacked_spectra.pdf}
\caption[Stacked {\hbox{\rmfamily H\,{\textsc i}}} spectra for each subsample]{Stacked {\hbox{\rmfamily H\,{\textsc i}}} spectra of galaxies in each subgroup ({\it left} panels). The {\it right} panels show the re-binned stacked spectra with a velocity width of 500~\hbox{km\,s$^{-1}$} (the vertical dotted lines). The horizontal dashed line in each {\it left} panel is the 1$\sigma$ error of the stacked spectrum.}
\label{fig:stacked_zcosmos}
\end{figure*}
Before performing the {\hbox{\rmfamily H\,{\textsc i}}} stacking analysis, an automated source-finding software \citep[{\sc duchamp},][]{Whiting:2012} was used to search for any directly detected {\hbox{\rmfamily H\,{\textsc i}}} emitters in the final data cubes. An eyeball inspection was also done. No significant source was found by either method.
As plotted in Fig.~\ref{fig:z_hist}, the lower and upper sideband data cubes contain 269 and 237 sample galaxies, respectively. However, some of the galaxies lie in channels seriously affected by RFI or at the edges of the data cubes. These galaxies were excluded, leaving 474 galaxies for the {\hbox{\rmfamily H\,{\textsc i}}} stacking analysis. The usable 474 galaxies consist of 95 galaxies with spectro-photometric Type~1, 58 with Type~2 and 321 with Type 3 or 4. Using their known positions and redshifts, the spectra for these galaxies were extracted from the data cubes with dimensions of 1$\degr \times$1$\degr \times$128 channels and corrected for the primary beam attenuation. The GMRT primary beam pattern was assumed to have a circular Gaussian profile given by
\begin{equation}
\label{eq:pbcor}
gain = e^{-(2\sqrt{{\rm ln}2}d/\theta)^{2}}, \quad \theta = 26.2 \times \frac{1280~{\rm MHz}}{f_{\rm obs}},
\end{equation}
where $f_{\rm obs}$ is the observing frequency, $\theta$ is the Half Power Beam Width (HPBW) and $d$ is angular separation from the GMRT pointing centre, given in units of arcmin. The HPBW was taken from the measurements provided in the National Centre for Radio Astrophysics (NCRA) website\footnote{{\url {http://www.ncra.tifr.res.in}}}. After this correction the spectra were shifted and aligned to the same rest frame velocity. The stacked spectrum was computed from a weighted-average, using the rms noise of each primary beam corrected spectra. Stacked spectra were computed separately for the different galaxy samples, as well as for the full 474 galaxy sample.
Fig.~\ref{fig:stacked_zcosmos} shows the {\hbox{\rmfamily H\,{\textsc i}}} stacked spectra for each of the subtypes of galaxies. Most of the {\hbox{\rmfamily H\,{\textsc i}}} gas in the COSMOS field observed by the GMRT resides in the Type~3 or 4 (late spiral, irregular and star burst galaxies). As expected there is no statistically significant signal from the early-type galaxies.
To calculate the average {\hbox{\rmfamily H\,{\textsc i}}} mass from the co-added spectra for each subsample, the following equation \citep{Wieringa:1992} was used:
\begin{equation}
\label{eq:H_mass}
\frac{{\hbox{$M_{\HIsub}$}}}{\hbox{${\rm M}_{\odot}$}} = \frac{236}{(1+z)} \left( \frac{D_L}{{\rm Mpc}} \right)^2
\bigg( \frac{\int S_V dV}{{\rm mJy \,km~s^{-1}}} \bigg),
\end{equation}
where $z$ is redshift, $D_{L}$ is the luminosity distance in units of Mpc, and $\int S_V dV$ is the integrated {\hbox{\rmfamily H\,{\textsc i}}} emission flux in units of $\rm {mJy~\hbox{km\,s$^{-1}$}}$. The redshift of 0.37 is the median redshift value of stacked galaxies for this calculation, which is also used for the luminosity distance in Eq.~\ref{eq:H_mass}. In calculating the integrated {\hbox{\rmfamily H\,{\textsc i}}} flux from the above equation, we must specify the width of velocity window within which all the {\hbox{\rmfamily H\,{\textsc i}}} emission flux is contained. This velocity window is estimated from the Tully-Fisher relation \citep{Tully:1977}. For our late-type sample galaxies, the mean $w_{20}$ is 278.8~\hbox{km\,s$^{-1}$} and the maximum is $\sim$571~\hbox{km\,s$^{-1}$}. Taking into account the redshift uncertainty of the zCOSMOS survey, $\pm$110~\hbox{km\,s$^{-1}$}, a velocity width of 500~\hbox{km\,s$^{-1}$} was used to calculate the integrated {\hbox{\rmfamily H\,{\textsc i}}} flux with the stacked {\hbox{\rmfamily H\,{\textsc i}}} spectra.
The average {\hbox{\rmfamily H\,{\textsc i}}} masses for each galaxy type are listed in Table~\ref{tab:HI_measure}. The errors in the average {\hbox{\rmfamily H\,{\textsc i}}} masses were estimated by applying a jackknife resampling method \citep{Efron:1982}. As expected, the late type (Type~3 and Type~4) galaxies have higher average {\hbox{\rmfamily H\,{\textsc i}}} gas content than the early type galaxies. The {\hbox{$M_{\HIsub}$}}$/L_{B}$ ratio of the late type galaxies is similar with the median {\hbox{$M_{\HIsub}$}}$/L_{B}$ of the Sc and Sd type galaxies in the local universe \citep{Roberts:1994}.
The above estimate of the {\hbox{\rmfamily H\,{\textsc i}}} mass assumes that the sample galaxies observed in {\hbox{\rmfamily H\,{\textsc i}}} are unresolved by the GMRT synthesised beam. To check this assumption the sizes of sample galaxies in {\hbox{\rmfamily H\,{\textsc i}}} were estimated from the relationship between optical and {\hbox{\rmfamily H\,{\textsc i}}} properties in \citet{Broeils_Rhee:1997}. These authors provided two relationships between the {\hbox{\rmfamily H\,{\textsc i}}} diameter and the $B$-band absolute magnitude as follows:
\begin{eqnarray}
{\rm log}(D_{\hbox{{\scriptsize H}\,{\tiny I}}}) & = & (-0.1673\pm0.0142) \times M_{B}-1.9545 \label{eq:relation_opt_HI1}\\
{\rm log}(D_{\rm eff}) & = & (-0.1674\pm0.0152) \times M_{B}-2.1689 \label{eq:relation_opt_HI2},
\end{eqnarray}
where $D_{\hbox{{\scriptsize H}\,{\tiny I}}}$ is the {\hbox{\rmfamily H\,{\textsc i}}} diameter (in kpc) at a surface density of 1 \hbox{${\rm M}_{\odot}$}~pc$^{-2}$, and $D_{\rm eff}$ is the diameter containing 50 per~cent of the {\hbox{\rmfamily H\,{\textsc i}}} mass. Fig.~\ref{fig:HI_size_zcosmos} shows the distribution of the estimated diameters from the Eq.~\ref{eq:relation_opt_HI1} and \ref{eq:relation_opt_HI2}. The blue and red histograms indicate the distribution of $D_{\rm eff}$ and $D_{\hbox{{\scriptsize H}\,{\tiny I}}}$, respectively. As can be seen, the galaxies seem to be partially resolved in terms of $D_{\rm eff}$ by the GMRT synthesised beam of ~3.5$\arcsec$, while $D_{\hbox{{\scriptsize H}\,{\tiny I}}}$ of most galaxies is larger than the GMRT synthesised beam. We note however that \citet{Broeils_Rhee:1997} warned that these local correlations might be biased due to the fact that their selection criteria required the galaxies to have large {\hbox{\rmfamily H\,{\textsc i}}} mass and optical size. To see whether the GMRT beam partially resolving the large galaxies causes a significant amount of {\hbox{\rmfamily H\,{\textsc i}}} flux to be lost, we repeated the analysis using several larger synthesised beams. The data cubes were smoothed to four additional synthesised beam sizes--5.9$\arcsec$, 7.8$\arcsec$, 9.8$\arcsec$ and 11.7$\arcsec$ which are equivalent to 30, 40, 50, and 60 kpc at $z \sim$~0.37, respectively. The same stacking procedure was conducted for all smoothed data cubes with different beam sizes, and then the average {\hbox{\rmfamily H\,{\textsc i}}} masses were calculated and compared. We found no statistically significant change in the stacked signal. For the subsequent calculations in the next sections, we hence assume that there is no significant effect of the GMRT beam size on the {\hbox{\rmfamily H\,{\textsc i}}} stacking results.
\begin{table*}
\caption[The measured properties of each galaxy type at $z \sim$~0.37]{The measured properties of each galaxy type at the redshift of $z \sim$~0.37.}
\begin{center}
\label{tab:HI_measure}
\begin{tabular}{@{}lccccccc}
\hline
Sample & $N_{\rm gal}$ & $\langle \hbox{$M_{\HIsub}$} \rangle$ & $\langle L_B \rangle$ & $\langle \hbox{$M_{\HIsub}$} \rangle/\langle L_B \rangle$ & $\rho_{L_{B}}$ & $\hbox{$\rho_{\HIsub}$}$ & {\it f} \\
& & (10$^{9}$~{\hbox{${\rm M}_{\odot}$}}) & (10$^{9}$~{\hbox{${\rm L}_{\odot}$}}) & ({\hbox{${\rm M}_{\odot}$}}/{\hbox{${\rm L}_{\odot}$}}) & (10$^{7}$~{\hbox{${\rm L}_{\odot}$}}~Mpc$^{-3}$) & (10$^{7}$~{\hbox{${\rm M}_{\odot}$}}~Mpc$^{-3})$ & \\
\hline
Type~1 & 95 & 2.20~$\pm$~2.60 & 17.76~$\pm$~0.04 & 0.12~$\pm$~0.15 & 5.11~$\pm$~0.65 & 0.63~$\pm$~0.75 & - \\
Type~2 & 58 & 1.50~$\pm$~2.74 & 21.33~$\pm$~0.05 & 0.07~$\pm$~0.13 & 2.69~$\pm$~0.65 & 0.21~$\pm$~0.38 & 1.10~$\pm$~0.09 \\
Type~3+4 & 321 & 3.83~$\pm$~1.20 & 10.07~$\pm$~0.01 & 0.38~$\pm$~0.12 & 8.12~$\pm$~2.15 & 4.86~$\pm$~1.99 & 1.57~$\pm$~0.02\\
\hline
All & & \multicolumn{5}{c}{{\hbox{$\Omega_{\HIsub}$}}$=$~(0.42~$\pm$~0.16)~$\times$~10$^{-3}$} & \\
\hline
\end{tabular}
\end{center}
Note: $N_{\rm gal}$ is the number of galaxies that are co-added, $\langle \hbox{$M_{\HIsub}$} \rangle$ is the average {\hbox{\rmfamily H\,{\textsc i}}} mass per galaxy, $\langle L_B \rangle$ is the mean $B$-band luminosity, $\rho_{L_{B}}$ is the luminosity density and $\hbox{$\rho_{\HIsub}$}$ is the {\hbox{\rmfamily H\,{\textsc i}}} density that the correction factor in the last column has been applied to. {\it f} is the correction factor for incomplete sampling of the luminosity function.
\end{table*}
\begin{figure}
\centering
\includegraphics[width=84mm]{beam_galaxy_size_zcosmos.pdf}
\caption[{\hbox{\rmfamily H\,{\textsc i}}} size distribution of zCOSMOS sample]{{\hbox{\rmfamily H\,{\textsc i}}} size distribution of zCOSMOS sample derived using the relation between optical magnitude and {\hbox{\rmfamily H\,{\textsc i}}} size based on \citet{Broeils_Rhee:1997}. Two {\hbox{\rmfamily H\,{\textsc i}}} sizes differently defined are used. Blue ($\slash$) and red histograms denote effective {\hbox{\rmfamily H\,{\textsc i}}} diameter and {\hbox{\rmfamily H\,{\textsc i}}} diameter at a surface density of 1 \hbox{${\rm M}_{\odot}$}~pc$^{-2}$, respectively. The GMRT synthesised beam size is shown as the vertical dashed line.}
\label{fig:HI_size_zcosmos}
\end{figure}
\section{Cosmic {\hbox{\rmfamily H\,{\textsc i}}} Mass Density ({\hbox{$\Omega_{\HIsub}$}})}
The measured {\hbox{\rmfamily H\,{\textsc i}}} mass from the stacked {\hbox{\rmfamily H\,{\textsc i}}} spectra can be converted to an {\hbox{\rmfamily H\,{\textsc i}}} density ({\hbox{$\rho_{\HIsub}$}}) and a cosmic {\hbox{\rmfamily H\,{\textsc i}}} density ({\hbox{$\Omega_{\HIsub}$}}) allowing one to examine how the {\hbox{\rmfamily H\,{\textsc i}}} gas content of the Universe evolves over cosmic time. The {\hbox{$\Omega_{\HIsub}$}} measurement of the COSMOS field is important because it is the highest-redshift measurement ever made with an {\hbox{\rmfamily H\,{\textsc i}}} spectral stacking technique. While there has been a previous measurement of the {\hbox{\rmfamily H\,{\textsc i}}} content of galaxies at $z \sim 0.37$ \citep{Lah:2009}, the sample for that study was galaxies in a rich cluster (Abell~370). Hence unlike the current work, the results from \citet{Lah:2009} cannot be used to study the evolution of the gas content of field galaxies.
In order to determine {\hbox{$\Omega_{\HIsub}$}}, the first step is to derive {\hbox{\rmfamily H\,{\textsc i}}} density from the average {\hbox{\rmfamily H\,{\textsc i}}} mass ({\hbox{$M_{\HIsub}$}}) measured above. Simply dividing {\hbox{$M_{\HIsub}$}} by the survey volume does not take into account the incompleteness of our sample to be stacked or the effect of cosmic variance. Following \citet{Rhee:2013} and \citet{Delhaize:2013}, we made use of a volume normalisation method and a correction factor to account for these effects. This approach to derive {\hbox{$\rho_{\HIsub}$}} adopts the optical luminosity ($L$) as a weight. We determined the ratio of {\hbox{$M_{\HIsub}$}} to $L$ which we then multiplied by the optical luminosity density \citep[e.g., see][]{Fall:1993}. This assumes that all of the {\hbox{\rmfamily H\,{\textsc i}}} gas is located in galaxies with optical counterparts \citep{Briggs:1990,Fall:1993,Rao:1993}. Several studies show that this is a reasonable assumption in the local universe. Firstly, the agreement between {\hbox{$\Omega_{\HIsub}$}} measurements based on optically selected galaxies and those based on {\hbox{\rmfamily H\,{\textsc i}}}-selected samples shows that there is little neutral hydrogen gas associated with galaxies faint or no optical counterparts \citep{Fall:1993,Rao:1993,Zwaan:1997}. Furthermore large blind 21-cm surveys have found that optically invisible but gas-rich galaxies below optical detection threshold do not exist in numbers sufficiently large to bias the inventory of {\hbox{$\rho_{\HIsub}$}} \citep{Taylor:2005,Doyle:2005}.
Accurate photometric measurements of luminosity, which are needed to normalise the stacked {\hbox{\rmfamily H\,{\textsc i}}} measurement, are available for the COSMOS field
as discussed in Section 2.2.
Thanks to a wealth of photometric and spectroscopic data available, the luminosity functions and luminosity densities have been quite well measured up to $z \sim 1$ in the COSMOS field \citep{Zucca:2009}. Since our galaxy classification scheme is the same as that used in \citet{Zucca:2009}, the luminosity function and luminosity density derived in that paper can be directly adopted to derive {\hbox{$\rho_{\HIsub}$}}. The equations below are used to derive {\hbox{\rmfamily H\,{\textsc i}}} density for each galaxy type:
\begin{equation}
\sum \hbox{$M_{\HIsub}$} = \langle \hbox{$M_{\HIsub}$} \rangle \times N_{\rm gal}, \qquad {\hbox{$\rho_{\HIsub}$}} = \frac{\sum \hbox{$M_{\HIsub}$}}{\sum L_B} \times \rho_{L_B}(\it z),
\label{eq:rhohi_zcosmos}
\end{equation}
where $\langle \hbox{$M_{\HIsub}$} \rangle$ denotes the average {\hbox{\rmfamily H\,{\textsc i}}} mass measured using the {\hbox{\rmfamily H\,{\textsc i}}} stacking technique, $N_{\rm gal}$ is the number of co-added galaxies, $L_{B}$ and $\rho_{L_{B}}$ are luminosity and luminosity density of the zCOSMOS sample galaxies in the $B$-band, respectively. We separately calculated the {\hbox{\rmfamily H\,{\textsc i}}} gas density for each galaxy type of Type~1, Type~2 and Type~3+4. However, this {\hbox{\rmfamily H\,{\textsc i}}} density is calculated without accounting for galaxies fainter than the optical survey limit. In the nearby universe, the low-luminosity late-types are known to be gas-rich, so we need to make a first order compensation for the fact that these objects are not all represented in our sample. To correct for this incomplete sampling of the luminosity function, correction factors for each type were obtained using the luminosity function parameters such as the faint-end slope ($\alpha$) and characteristic luminosity ($L^*$) of the luminosity function given by \citet{Zucca:2009}. Refer to Appendix~A in \citet{Rhee:2013} for more details about the correction factor calculation. This correction factor was not applied to early-type galaxies because the contribution of faint early-type population to {\hbox{$M_{\HIsub}$}} and {\hbox{$\rho_{\HIsub}$}} is small. Moreover the correction factor for the early-type sample (Type~1) made no difference in calculation of the total {\hbox{$\rho_{\HIsub}$}}. These correction factors as well as the corrected {\hbox{$\rho_{\HIsub}$}} for each type are listed in Table~\ref{tab:HI_measure}.
Consistent with earlier works, we define the cosmic {\hbox{\rmfamily H\,{\textsc i}}} gas density ({\hbox{$\Omega_{\HIsub}$}}) as the ratio of {\hbox{$\rho_{\HIsub}$}} to the critical density ($\rho_{\rm crit}$):
\begin{equation}
{\hbox{$\Omega_{\HIsub}$}} = \frac{\hbox{$\rho_{\HIsub}$}}{\rho_{\rm crit}} = \frac{8\pi G \hbox{$\rho_{\HIsub}$}}{3H_{0}^{2}},
\label{eq:ohi_zcosmos}
\end{equation}
where $H_{0}$ is the Hubble constant and $G$ is the gravitational constant. The critical density at present is $\rho_{\rm crit}$ = 2.78~$\times$~10$^{11} h^2$~\hbox{${\rm M}_{\odot}$}~Mpc$^{-3}$, where $h=H_{0}/$100 \hbox{km\,s$^{-1}$} ($h=$~0.7). Since in principle all galaxy types contribute to the cosmic {\hbox{\rmfamily H\,{\textsc i}}} density, the total {\hbox{$\rho_{\HIsub}$}} is obtained by summing {\hbox{$\rho_{\HIsub}$}} contributions from all types. This gives {\hbox{$\Omega_{\HIsub}$}}~$=$~(0.42~$\pm$~0.16)~$\times$~10$^{-3}$. This $2.6\sigma$ measurement is shown in Fig.~\ref{fig:omega_HI_zcosmos} along with other available measurements taken from the literature (see the caption for more details). As can be seen, our measurement taken in conjunction with earlier measurements at lower redshifts indicates that there has been no significant evolution in {\hbox{$\Omega_{\HIsub}$}} from $z = 0$ to $z \sim 0.4$. The weighted mean average of {\hbox{$\Omega_{\HIsub}$}} from all 21-cm measurements at redshifts $z < 0.4$ gives $\hbox{$\Omega_{\HIsub}$} = (0.35 \pm 0.01) \times 10^{-3}$.
Regarding the cosmic variance, it is known that the zCOSMOS field suffers less than 10 per~cent cosmic variance in the redshift interval of $z =$~0 to 0.5 \citep{Driver:2010}. The zCOSMOS field surveyed for this paper by the GMRT has a limited redshift interval of 0.345~$< z <$~0.387 and a small sky area compared to the original, which would result in increased cosmic variance. However, since in our {\hbox{$\Omega_{\HIsub}$}} calculation above we do a volume normalisation using luminosity density derived from the full sample of the zCOSMOS field, the cosmic variance that we are subject to is the same as that computed for the entire zCOSMOS field. Moreover, we compared the luminosity density of the zCOSMOS field at $z \sim$~0.37 used for volume normalisation to that derived from a large and complete spectroscopic survey such as Galaxy and Mass Assembly \citep[GAMA,][]{Loveday:2012}. The two luminosity densities are in excellent agreement for the overall galaxy population as well as for each galaxy population. This implies that the volume which we used for normalisation can represent the average of the universe at the redshift that we explored.
The stacking analysis using 21-cm emission carried out in this paper is based on the assumption that {\hbox{\rmfamily H\,{\textsc i}}} gas in our galaxies is optically thin. This means that any possible influence of {\hbox{\rmfamily H\,{\textsc i}}} self-absorption, which will lead to an underestimation of the {\hbox{\rmfamily H\,{\textsc i}}} mass, is negligible. Since it is very difficult to assess the effect quantitatively and statistically, even large blind {\hbox{\rmfamily H\,{\textsc i}}} surveys have made only a rough estimate for the effect, i.e., an underestimation of less than 15 per~cent in {\hbox{$\Omega_{\HIsub}$}} \citep{Zwaan:1997, Zwaan:2005}. However based on high-resolution maps of the {\hbox{\rmfamily H\,{\textsc i}}} distribution in M31, M32 and LMC, \citet{Braun:2012} recently suggested that galaxies contain a significant population of {\hbox{\rmfamily H\,{\textsc i}}} clouds with a size of 100~pc and high {\hbox{\rmfamily H\,{\textsc i}}} column density ($>$~10$^{23}$~cm$^{-2}$) which are optically thick. He derived a global opacity correction factor of 1.34$\pm$0.05 and applied this to the local {\hbox{$\Omega_{\HIsub}$}} measurements resulting in $\sim$34 per~cent increased {\hbox{$\Omega_{\HIsub}$}} as seen in Fig~\ref{fig:omega_HI_zcosmos}. Although the sample used to derive the correction factor covers a fairly large range of {\hbox{\rmfamily H\,{\textsc i}}} mass the total number of galaxies in his sample is very small. As such, the validity of applying this correction factor to high redshift measurement like ours seems to be unclear. In any case, were this correction to be uniformly applied to all the {\hbox{\rmfamily H\,{\textsc i}}} emission surveys, it would result only in a shift of all the values upwards, and would not affect the conclusion that there appears to be no evolutionary trend in {\hbox{$\Omega_{\HIsub}$}} at least out to $z \la 0.4$.
\begin{figure*}
\centering
\includegraphics[width=150mm]{plot_Omega_HI_final_tz_fitting_new.pdf}
\caption[The cosmic {\hbox{\rmfamily H\,{\textsc i}}} gas density ({\hbox{$\Omega_{\HIsub}$}}) at $z \sim$~0.4]{This shows the cosmic {\hbox{\rmfamily H\,{\textsc i}}} gas density ({\hbox{$\Omega_{\HIsub}$}}) measurements as a function of redshift ({\it top} axis) and look-back time ({\it bottom} axis). All measurements have been corrected to the same cosmological parameters. Our {\hbox{$\Omega_{\HIsub}$}} measurement of the COSMOS field is presented by the green star. The small black square and triangle at $z \sim$~0 are the HIPASS and ALFALFA 21-cm emission measurements by \citet{Zwaan:2005, Martin:2010}, respectively. The red diamond is not a measured value but the average {\hbox{$\Omega_{\HIsub}$}} of the two measurements at $z = 0$ to which a correction factor for self-opaque effect has been applied \citep{Braun:2012}. The open diamonds are from the Parkes telescope and an {\hbox{\rmfamily H\,{\textsc i}}} stacking technique \citep{Delhaize:2013}. The cyan open circle is the preliminary result from the AUDS \citep{Freudling:2011}. The red right-pointing triangle is the 21-cm direct detection measurement from 60\% data of the total AUDS survey \citep{Hoppmann:2015}. Two yellow stars are measured by \citet{Rhee:2013} using the WSRT and stacking technique. The big purple triangle is measured by \citet{Lah:2007} using the GMRT 21-cm emission stacking. The red star is the {\hbox{$\Omega_{\HIsub}$}} measurement of VVDS~14h field at $z \sim$~0.32 \citep{Rhee:2016} using the GMRT along with stacking technique. The blue diamonds, red big circle, green circles, and orange squares are damped Lyman-$\alpha$ measurements from the {\it HST} and the SDSS by \citet{Rao:2006}, \citet{Prochaska:2005}, \citet{Noterdaeme:2009}, \citet{Noterdaeme:2012}, respectively. The grey downward triangles at high redshift of $z >$~2 are ESO UVES measurements of DLAs and sub-DLAs by \citet{Zafar:2013}. The black lines with grey shade areas are least-square fits and their 95~per~cent confidence interval with all {\hbox{$\Omega_{\HIsub}$}} measurements at lower redshifts and higher redshifts, respectively. The blue dashed and the red solid lines are model predictions of {\hbox{$\Omega_{\HIsub}$}} taken from \citet{Lagos:2014} and \citet{Kim:2015}, respectively.}
\label{fig:omega_HI_zcosmos}
\end{figure*}
\section{{\hbox{\rmfamily H\,{\textsc i}}} Gas Evolution over last 4~Gyr}
In Fig.~\ref{fig:omega_HI_zcosmos}, our {\hbox{$\Omega_{\HIsub}$}} measurement of the COSMOS field is compared with other {\hbox{$\Omega_{\HIsub}$}} values available in the published literature \citep{Zwaan:2005,Martin:2010,Freudling:2011,Delhaize:2013,Hoppmann:2015,Rhee:2013,Lah:2007,Rhee:2016,Rao:2006,Prochaska:2005,Noterdaeme:2009,Noterdaeme:2012,Zafar:2013}. Two main observational techniques have been used to measure {\hbox{$\Omega_{\HIsub}$}}: 21-cm emission observations at low redshifts and damped Lyman alpha absorption (DLA) observations at high redshifts. All measurements using {\hbox{\rmfamily H\,{\textsc i}}} 21-cm emission from both direct detection and stacking are in good agreement. At high redshifts ($z >$~2) all {\hbox{$\Omega_{\HIsub}$}} measurements from DLA observations are consistent with one another, showing increase in {\hbox{$\Omega_{\HIsub}$}} with redshift. Also there seems to be at least 2 times more {\hbox{\rmfamily H\,{\textsc i}}} gas amount than at lower redshifts. We note that the {\hbox{\rmfamily H\,{\textsc i}}} 21-cm measurements correspond to the {\hbox{\rmfamily H\,{\textsc i}}} gas inside galaxies, while the DLA observations measure the total gas, regardless of whether it lies inside or outside galaxies. In principle these could be different quantities, although, as noted above, in the local universe at least, there is no evidence for a large reservoir of {\hbox{\rmfamily H\,{\textsc i}}} that lies in optically dark galaxies.
Many galaxy evolution models have recently attempted to predict {\hbox{\rmfamily H\,{\textsc i}}} gas densities across cosmic time to reproduce observations \citep[e.g.,][]{Power:2010,Lagos:2011,Duffy:2012,Dave:2013,Lagos:2014,Rahmati:2015,Kim:2015}. However, there is a tension between observations and theoretical models. We show in Fig.~\ref{fig:omega_HI_zcosmos} theoretical predictions for the evolution of {\hbox{$\Omega_{\HIsub}$}} from the recent semi-analytic `Lagos12' and `Kim15' models of \citet{Lagos:2014} and \citet{Kim:2015} for comparison between observations and theories. As can be seen, the theoretical models do match the low redshift data, but not the high-redshift DLA based measurements.
In contrast, a hydrodynamical simulation study of the distribution of {\hbox{\rmfamily H\,{\textsc i}}} around high-redshift galaxies at $z >$~ 1 \citep[e.g.,][]{Rahmati:2015} shows that their {\hbox{$\Omega_{\HIsub}$}} predictions are in good agreement with the {\hbox{$\Omega_{\HIsub}$}} evolution at high redshifts while they disagree with lower-redshift {\hbox{$\Omega_{\HIsub}$}}. In some simulations this problem is interpreted as reflecting the difference between what the {\hbox{\rmfamily H\,{\textsc i}}} 21cm measurements and the DLA measurements are sensitive to \citep{Altay:2011,Faucher-Giguere:2011,Fumagalli:2011}. The forthcoming surveys by the SKA pathfinders and the SKA itself will be critically important to resolve this issue.
\section{Summary and Conclusion}
We present the results of an {\hbox{\rmfamily H\,{\textsc i}}} spectral stacking analysis using GMRT observations of the COSMOS field. Our sample is chosen from the zCOSMOS-bright 10k catalogue \citep{Lilly:2009}. The individual {\hbox{\rmfamily H\,{\textsc i}}} 21-cm line spectra obtained from the GMRT are stacked using the known optical positions and redshifts of the galaxies. The {\hbox{\rmfamily H\,{\textsc i}}} spectra are separately stacked for galaxy types classified by SED template fitting and then converted to the average {\hbox{\rmfamily H\,{\textsc i}}} mass per galaxy. We find that the {\hbox{\rmfamily H\,{\textsc i}}} signal comes primarily from late type galaxies, as expected. The inferred {\hbox{$M_{\HIsub}$}}$/L_{B}$ ratio is consistent with that of galaxies in the local ($z=0$) universe. Using the average {\hbox{\rmfamily H\,{\textsc i}}} mass along with the integral optical $B$-band luminosity of the sample galaxies and the luminosity density of the COSMOS field, a volume normalisation is applied to obtain the cosmic {\hbox{\rmfamily H\,{\textsc i}}} density ({\hbox{$\Omega_{\HIsub}$}}). We measure {\hbox{$\Omega_{\HIsub}$}}~$=$~(0.42~$\pm$~0.16)~$\times$~10$^{-3}$ at $z \sim$~0.37. This $2.6 \sigma$ measurement is the highest-redshift measurement of {\hbox{$\Omega_{\HIsub}$}} ever made using {\hbox{\rmfamily H\,{\textsc i}}} spectral stacking. The value of {\hbox{$\Omega_{\HIsub}$}} that we measure is consistent within the error bars with both the {\hbox{$\Omega_{\HIsub}$}} at $z=0$ as measured from large blind 21-cm surveys as well as that measured from other {\hbox{\rmfamily H\,{\textsc i}}} stacking experiments. All the {\hbox{\rmfamily H\,{\textsc i}}} 21-cm emission measurements to date show no evidence for evolution of {\hbox{\rmfamily H\,{\textsc i}}} gas abundance over the last 4 Gyr; the weighted mean of {\hbox{$\Omega_{\HIsub}$}} from all 21-cm measurements at $z < 0.4$ is (0.35~$\pm$~0.01)~$\times$~10$^{-3}$. This value of {\hbox{$\Omega_{\HIsub}$}} is however smaller than that measured at $z \ga 2$ from DLA observations. The next generation of radio telescopes will be sensitive enough to detect the {\hbox{\rmfamily H\,{\textsc i}}} signal to redshifts greater than $z =$~1 and will be crucial in understanding the evolution of {\hbox{$\Omega_{\HIsub}$}} in the redshift range intermediate between $z \sim 0.4$ and the redshifts probed by DLA observations.
\section*{Acknowledgments}
We are grateful to an anonymous referee for helpful comments and suggestions that improved this work. JR would like to thank Lister Staveley-Smith for useful comments. We thank Claudia del P. Lagos and Han-Seek Kim for providing their model predictions used in Fig.~\ref{fig:omega_HI_zcosmos}. We also thank the staff of the GMRT for their assistance. The GMRT is operated by the National Centre for Radio Astrophysics of the Tata Institute of Fundamental Research. This research was funded by an Australian Indian Strategic Research Fund (AISRF) grant. This fund was jointly administered by the Department of Innovation, Industry, Science and Research in Australia and by the Department of Science and Technology in India. The project title was ``Gas in Galaxies in the Distant Past''. Parts of this research were conducted by the Australian Research Council Centre of Excellence for All-sky Astrophysics (CAASTRO), through project number CE110001020. The G10/COSMOS cutout tool uses data acquired as part of the Cosmic Evolution Survey (COSMOS) project and spectra from observations made with ESO Telescopes at the La Silla or Paranal Observatories under programme ID 175.A-0839. The G10/COSMOS cutout tool is hosted and maintained by funding from the Inertnational Centre for Radio Astronomy Research (ICRAR) at the University of Western Australia. Full details of the catalogue can be found in \citep{Davies:2015} or on the G10 website: {\url {http://ict.icrar.org/cutout/G10/}}.
\bibliographystyle{mn2e}
|
\section{Introduction}
Quantum Electrodynamics in $2+1$ dimensions (QED$_3$) has been extensively studied during more than three decades now.
Originally, the interest in QED$_3$ came from its similarities to ($3+1$)-dimensional QCD and the fact that phenomena such as dynamical
chiral symmetry breaking (D$\chi$SB) and mass generation may be studied systematically in such a toy model, see, {\it e.g.},
Refs.~[\onlinecite{Pisarski84,AppelquistNW88,Pennington91+92,Pisarski91,AtkinsonJM90,DagottoKK89+90,Azcoiti93+96,KarthikN16,AppelquistW04,AppelquistCS99,GiombiKT16,DiPietroKSS16,Nash89,Kotikov93+12,BashirR07,BashirRSR09}].
Later, a strong interest in QED$_3$ arose in connexion with planar condensed matter physics systems
having relativistic-like low-energy excitations such as some two-dimensional antiferromagnets~\cite{MarstonA89+IoffeL89}
and graphene;~\cite{Semenoff84+Wallace47} the study of a dynamically generated gap in the fermion spectrum of graphene
has now become an active area of research, see, {\it e.g.}, the reviews Refs.~[\onlinecite{KotovUPGC12,MiranskyS15}].
In all cases, the understanding of the phase structure of QED$_3$ is a crucial pre-requisite to understand
non-perturbative dynamic phenomena in more realistic particle and condensed matter physics models.
Despite the fact that a large number of investigations have been carried out to study D$\chi$SB
in QED$_3$, very different results have been obtained.
Without being extensive, let us indeed recall that, in his seminal paper~\cite{Pisarski84}, Pisarski solved the Schwinger-Dyson (SD) gap
equation using a leading order (LO) $1/N$-expansion and found that a fermion mass is generated for all values of $N$, decreasing exponentially with $N$ and vanishing only in the limit $N \rightarrow \infty$.
Later, he confirmed his finding by a renormalization group analysis.~\cite{Pisarski91}
Support of Pisarski's result was given by Pennington and collaborators~\cite{Pennington91+92} who adopted
a more general non-perturbative approach to solving the SD equations.
On the other hand, in a more refined analysis of the gap equation at LO of the $1/N$-expansion,
Appelquist et al.~\cite{AppelquistNW88} have shown that the theory exhibits a critical
behaviour as the number $N$ of fermion flavours approaches $N_c = 32/ \pi^2$;
that is, a fermion mass is dynamically generated only for $N<N_c$.
Contrary to all previous results, an alternative non-perturbative study by
Atkinson et al.~\cite{AtkinsonJM90} suggested that chiral symmetry is unbroken at
sufficiently large $N$. The theory has also been simulated on the lattice.~\cite{DagottoKK89+90,Azcoiti93+96,KarthikN16}
Remarkably, the conclusions of Ref.~[\onlinecite{DagottoKK89+90}] are in the agreement
with the existence of a critical $N$ as predicted in the analysis of Ref.~[\onlinecite{AppelquistNW88}]
while the second paper~\cite{Azcoiti93+96} finds D$\chi$SB for all $N$ and the recent third one~\cite{KarthikN16} no sign of D$\chi$SB at all.
Even in the case where a finite $N_c$ is found, its value is subject to uncertainty with estimates ranging from $N_c=1$ to $N_c=4$, see
Ref.~[\onlinecite{AppelquistW04}] for a review. Moreover, Ref.~[\onlinecite{AppelquistCS99}] found an upper bound, $N_c <3/2$,
while, more recently, Ref.~[\onlinecite{GiombiKT16}] found that $N_c < 4.4$ and Ref.~[\onlinecite{DiPietroKSS16}] that $N_c<9/4$.
Clearly, all these disagreements reflect our poor understanding of this problem.
The purpose of the present work is to include $1/N$ corrections to the LO
result of Ref.~[\onlinecite{AppelquistNW88}].
Because the critical value $N_c$ is not large, the contribution of such higher
orders in the $1/N$ expansion can be essential and their proper study may lead to a better
understanding of the problem.
This important issue has been rarely addressed in the past. To the best of our knowledge, the main references
are [\onlinecite{Nash89}] and [\onlinecite{Kotikov93+12}] where rather different results
were obtained.
The well-known results of Ref.~[\onlinecite{Nash89}] demonstrated a quite strong stability
of the $1/N$ expansion while the ones of Ref.~[\onlinecite{Kotikov93+12}] showed that a similar property holds only in the Landau gauge.
The strong gauge dependence found in Ref.~[\onlinecite{Kotikov93+12}]
is in agreement with the studies of Ref.~[\onlinecite{BashirR07}] in the so-called rainbow approximation.
In the following, we shall refine the analysis of Ref.~[\onlinecite{Kotikov93+12}] and perform an accurate computation of all $1/N$ corrections in the Landau gauge,
with a special focus on the most complicated ones, in order to extract the value of $N_c$.
The last years witnessed a strong progress in the study of the gauge dependence of D$\chi$SB
in various models, see Ref.~[\onlinecite{AhmadCCR16}] as well as references and
discussions therein. The progress is related to the use of the Landau-Khalatnikov-Fradkin
transformation.~\cite{LandauK55+Fradkin56} In the case of QED$_3$ in the $1/N$-expansion,
the application of this transformation~\cite{BashirRSR09} has revealed the almost complete lack of gauge dependence
for $N_c$. This confirms that we can limit our analysis to the case of the Landau gauge.
\section{Model and Schwinger-Dyson equations}
The Lagrangian of massless QED$_3$ with $N$ flavours of fermions reads
\begin{equation}
L = \overline \Psi ( i \hat \partial - e \hat A ) \Psi - \frac{1}{4} F_{ \mu \nu}^2\, ,
\label{L-QED3}
\end{equation}
where $ \Psi$ is taken to be a four component complex spinor.
In the massless case, which we are considering, the model contains infrared divergences.
The latter soften when the model is analysed in a $1/N$ expansion.~\cite{AppelquistP81,JackiwT81+AppelquistH81}
Since the theory is super-renormalizable, the mass scale is given by the dimensionful coupling constant: $a = Ne^2/8$,
which is kept fixed as $N \rightarrow \infty$.
In the four component case, we can introduce the matrices $\gamma_3$ and $\gamma_5$
which anticommute with $\gamma_0$, $\gamma_1$ and $\gamma_2$.
Then, the massless case is invariant under the transformations:
$\Psi \rightarrow \exp(i \alpha_1 \gamma_3) \Psi$ and
$ \Psi \rightarrow \exp(i \alpha_2 \gamma_5) \Psi$.
Together with the identity matrix and $[\gamma_3, \gamma_5 ]$, we have a $U(2)$ symmetry for each spinor and the
full global ``chiral'' (or rather flavour) symmetry is $U(2N)$. A mass term will break this symmetry
to $U(N) \times U(N)$. It is the dynamical generation of such a mass that we shall consider in the following.
It is also possible to include a parity non-conserving mass, see for
example Ref.~[\onlinecite{AppelquistBKW86}], but we will not consider this possibility here.
Following Ref.~[\onlinecite{AppelquistNW88}], we now study the solution of the SD equation.
The inverse fermion propagator has the form
\begin{equation}
S^{-1}(p) = [1 + A(p)]\,\left( i\hat p + \Sigma (p) \right)\, ,
\label{Sigma}
\end{equation}
where $A(p)$ is the wave-function renormalization and $\Sigma (p)$ is the dynamically generated parity-conserving mass which
is taken to be the same for all the fermions. Notice that in our definition of $\Sigma (p)$, Eq.~(\ref{Sigma}), the choice of the free vertex
corresponds to the so-called central Ball-Chiu vertex~\cite{BallC80} for the ``more standard''
definition $\tilde{\Sigma} (p) =
\Sigma (p)[1 + A(p)]$.
With these conventions, the SD equation for the fermion propagator may be decomposed into scalar and vector components
as follows:
\begin{subequations}
\label{SD-sigma+A}
\begin{eqnarray}
\tilde{\Sigma}(p)
= \frac{2a}{N} \, \mbox{Tr} \int \frac{d^3 k}{(2 \pi )^3}
\frac{\gamma^{\mu} D_{\mu \nu}(p-k) \Sigma (k) \Gamma^{\nu}(p,k)}
{\left[1 + A(k) \right] \left( k^2 + \Sigma^2(k) \right)}\, ,
\label{SD-sigma} \\
A(p) p^2 = -\frac{2a}{N} \, \mbox{Tr} \int \frac{d^3 k}{(2 \pi )^3}
\frac{ D_{\mu \nu}(p-k) \hat p \gamma^{\mu} \hat k \Gamma^{\nu}(p,k)}
{\left[1 + A(k) \right] \left( k^2 + \Sigma^2(k) \right)} \, ,
\label{SD-A}
\end{eqnarray}
\end{subequations}
where $D_{\mu \nu}(p)$ is the photon propagator in the Landau gauge:
\begin{equation}
D_{\mu \nu}(p) = \frac{g_{\mu \nu} - p_{ \mu} p_{ \nu} / p^2}{p^2 \left[1 + \Pi (p) \right]}\, ,
\label{photon}
\end{equation}
$\Pi(p)$ is the polarization operator and $ \Gamma ^{ \nu}(p,k)$ is the vertex function.
In the following, we shall first consider Eqs.~(\ref{SD-sigma+A}) at the LO approximation
and then study Eq.~(\ref{SD-sigma}) at the NLO level.
\begin{figure}[tl]
\includegraphics[width=0.2\textwidth]{fig1}
\caption{\label{fig:diags-LO}
LO diagram to the dynamically generated mass $\Sigma(p)$. The crossed line denotes mass insertion.}
\end{figure}
\section{Leading order}
The LO approximations in the $1/N$ expansion are given by:
\begin{equation}
A(p) = 0, \quad \Pi (p) = a/ |p|, \quad \Gamma^{\nu}(p,k) = \gamma^{\nu}\, ,
\end{equation}
where the fermion mass has been neglected~\footnote{A study of the fermion mass contribution to $\Pi(p)$ can be
found, for example, in Ref.~[\onlinecite{GusyninHR96}].} in the calculation of $\Pi(p)$.
A single diagram contributes to the gap equation (\ref{SD-sigma}) at LO, see Fig.~\ref{fig:diags-LO}, and the latter reads:
\begin{equation}
\Sigma (p) =\frac{16a}{N}
\int \frac{d^3 k}{(2 \pi )^3}
\frac{ \Sigma (k) }{ \left( k^2 + \Sigma^2(k)
\right)
\bigl[ (p-k)^2 + a \,|p-k| \bigr]} \, .
\label{SD-sigma-LO1}
\end{equation}
Performing the angular integration in Eq.~(\ref{SD-sigma-LO1}) yields:
\begin{equation}
\Sigma (p) =\frac{4a}{\pi^2 N |p|}
\int_0^{\infty} \! \! \! d |k|\,
\frac{ |k| \Sigma (|k|) }{k^2 + \Sigma^2(|k|)}
\ln \left(\frac{|k|+|p|+a}{|k-p|+a}\right)
\, .
\label{SD-sigma-LO2}
\end{equation}
The study of Eq.~(\ref{SD-sigma-LO2}) in Ref.~[\onlinecite{AppelquistNW88}] has revealed the existence
of a critical number of fermion flavours $N_c$, such that for $N>N_c$, $\Sigma (p) =0$.
As it was argued in this reference, QED$_3$ is strongly damped for $|p|>a$, {\it i.e.},
all relevant physics occur at $|p|/a <1$. Hence, only the lowest order terms in $|p|/a$ have to be kept on the r.h.s.\ of Eq.~(\ref{SD-sigma-LO2})
with a hard cut-off at $|p|=a$. Moreover, considering $N$
close to $N_c$, the value of $\Sigma (|k|)$ can be made arbitrarily small. Thus,
$k^2 + \Sigma^2(|k|)$ can be replaced by $k^2$ on the r.h.s.\ of Eq.~(\ref{SD-sigma-LO2})
which then further simplifies as:
\begin{equation}
\Sigma (p) =\frac{8}{\pi^2 N}
\int_0^{a} \! \! \! d |k| \,
\frac{\Sigma (|k|) }{ \mbox{Max}({|k|,|p|})}\, .
\label{SD-sigma-LO3}
\end{equation}
Following Ref.~[\onlinecite{AppelquistNW88}], the mass function may then be parametrized as:
\begin{equation}
\Sigma (k) = B \, (k^2)^{ -\alpha} \, ,
\label{sigma-parametrization}
\end{equation}
(with an arbitrary $B$ value) where the index $\alpha$ has to be self-consistently determined. Substituting (\ref{sigma-parametrization}) in Eq.~(\ref{SD-sigma-LO3}), the gap equation reads:
\begin{eqnarray}
1 = \frac{2\beta}{L} \quad \text{where} \quad \beta = \frac{1}{\alpha \left( 1/2 - \alpha \right)} \quad \text{and} \quad L \equiv \pi^2 N\, .
\label{gap-eqn-LO}
\end{eqnarray}
Solving the gap equation, the following values of $\alpha$ are obtained:
\begin{eqnarray}
\alpha_{\pm} = \frac{1}{4}\,\left( 1 \pm \sqrt{1 - \frac{32}{L}} \right) \, ,
\label{al-LO}
\end{eqnarray}
which reproduces the solution given by Appelquist et al.\ in Ref.~[\onlinecite{AppelquistNW88}].
Their analysis yields a critical number of fermions:
$N_c = 32/ \pi^2 \approx 3.24$ ({\it i.e.}, $L_c = 32$),
such that $\Sigma(p) = 0$ for $N>N_c$ and
\begin{equation}
\Sigma(0) \simeq \exp
\bigl[ -2 \pi / (N_c/N - 1)^{1/2} \bigr]\, ,
\label{Sigma(0)}
\end{equation}
for $N<N_c$. Thus, D$\chi$SB occurs when $\alpha$ becomes
complex, that is for $N<N_c$.
As it was shown in Ref.~[\onlinecite{Kotikov93+12}], the same result for $\Sigma(p)$ can be obtained
in another way. Taking the limit of large $a$, the linearized version of Eq.~(\ref{SD-sigma-LO1})
has the following form:
\begin{equation}
\Sigma (p) =\frac{16}{N}
\int \frac{d^3 k}{(2 \pi )^3} \frac{ \Sigma (k) }{k^2 \, |p-k| } \, .
\label{SD-sigma-LO4}
\end{equation}
Interestingly, the large-$N$ limit of the photon propagator in QED$_3$
has precisely the same momentum dependence as the one in the so-called reduced QED, see Ref.~[\onlinecite{GorbarGM01}]
and also Refs.~[\onlinecite{Marino93+DoreyM92+KovnerR90}].
The multi-loop structure of the latter has been recently explored in Refs.~[\onlinecite{Teber12+KotikovT13,KotikovT14}].
With the help of the ansatz (\ref{sigma-parametrization}), one can then see that
the r.h.s.\ of Eq.~(\ref{SD-sigma-LO4}) may be
calculated with the help of the standard rules of perturbation theory for massless Feynman diagrams
as in Ref.~[\onlinecite{Kazakov83}], see also the recent short review Ref.~[\onlinecite{TeberK16}].
Indeed, given these rules, the computation of Eq.~(\ref{SD-sigma-LO4}) is straightforward and reads:
\begin{equation}
\Sigma^{(\text{LO})}(p) = \frac{8B}{N}\,\frac{(p^2)^{-\alpha}}{(4\pi)^{3/2}}\, \frac{2\beta}{\pi^{1/2}}\, .
\label{sigma-LO-res}
\end{equation}
This immediately yields the gap equation (\ref{gap-eqn-LO}) and, hence, the results
of Eq.~(\ref{al-LO}) together with the critical value $N_c = 32/ \pi^2 $ at which the index $\alpha$
becomes complex.
Similarly, such rules allow for a straightforward evaluation of the wave function renormalization. At LO, Eq.~(\ref{SD-A}) simplifies as:
\begin{equation}
A(p) p^2 = -\frac{2a}{N}
\text{Tr} \int \! \! \frac{d^D k}{(2 \pi )^D}
\frac{
( g_{\mu \nu} - \frac{(p-k)_\mu (p-k)_\nu}{(p-k)^2})
\hat p
\gamma^{\mu}
\hat k
\gamma^{\nu}}{
k^2 |p-k|}\, ,
\label{SD-A1}
\end{equation}
where the integral has been dimensionally regularized with $D=3 -2 \varepsilon$.
Taking the trace and computing the integral on the r.h.s.\ yields:
\begin{equation}
A(p) =
\frac{\Gamma(1+\varepsilon)(4\pi)^{\varepsilon} \mu^{2\varepsilon}}{p^{2\varepsilon}} \, C_1 =
\frac{\overline{\mu}^{2\varepsilon}}{p^{2\varepsilon}} \, C_1 \, + {\rm O}(\varepsilon) \, ,
\label{SD-A2}
\end{equation}
where the $\overline{MS}$ parameter $\overline{\mu}$ has the standard form
$\overline{\mu}^2 = 4\pi e^{-\gamma_E} \mu^2$ with the Euler constant $\gamma_E$ and
\begin{equation}
C_1 =
+\frac{4}{3\pi^2N} \left(\frac{1}{\varepsilon} + \frac{7}{3} - 2\ln 2\right)\, .
\end{equation}
The corresponding anomalous scaling dimension of the fermion field then reads:
$\eta = \mu^2 (d/d\mu^2) A(p) = 4 / (3\pi^{2}N)$,
and coincides with the one in Ref.~[\onlinecite{Gracey94}].
\section{Next-to-leading order}
The ease at which the standard rules for computing massless Feynman diagrams allowed us to derive LO results suggests the possibility to extend these computations beyond LO.
We therefore consider the NLO contributions to the dynamically generated mass and parametrize them as:
\begin{equation}
\Sigma^{(\text{NLO})}(p) = \left(\frac{8}{N}\right)^2 B\,\frac{(p^2)^{-\alpha}}{(4\pi)^{3}}\,
\left( \Sigma_A + \Sigma_1 + 2\,\Sigma_2 + \Sigma_3 \right) \, ,
\end{equation}
where each NLO contribution is represented graphically in Fig.~\ref{fig:diags-NLO}.
Because we are dealing with the linearized gap equation, each contribution contains a single mass insertion.
Adding these contributions to the LO result, Eq.~(\ref{sigma-LO-res}),
the gap equation has the following general form:
\begin{eqnarray}
1 = \frac{2\beta}{L} +
\frac{\pi}{L^2}\,\Bigl[\Sigma_A + \Sigma_1 + 2\,\Sigma_2 + \Sigma_3 \Bigr]\, .
\label{gap-eqn-NLO}
\end{eqnarray}
After very tedious and lengthy calculations, all NLO contributions could be evaluated exactly using the rules for computing massless Feynman diagrams.
For the most complicated scalar diagrams, see $I_1(\alpha)$ and $I_2(\alpha)$ below, the Gegenbauer-polynomial technique has been used following the paper [\onlinecite{Kotikov95}].
We now summarize our results (details of the calculations will be published elsewhere).
The contribution $\Sigma_A$, see Fig.~\ref{fig:diags-NLO} A), originates from the LO value of $A(p)$ and is singular.
Using dimensional regularization, it reads:
\begin{equation}
\overline{\Sigma}_A =
+\frac{16}{3}\, \frac{\overline{\mu}^{2\varepsilon}}{p^{2\varepsilon}}\, \beta \left(\frac{1}{\varepsilon} + \Psi_1 + \frac{4}{3} - \frac{\beta}{4} \right) \, + {\rm O}(\varepsilon)\, ,
\label{sigma-NLO-A}
\end{equation}
where $\overline{\Sigma}_i = \pi \Sigma_i$, $(i=1,2,3.A)$ and
\begin{equation}
\Psi_1 = \Psi(\alpha)+ \Psi(1/2-\alpha)-2\Psi(1) + \frac{3}{1/2-\alpha}
-2 \ln 2\, ,
\label{psi1}
\end{equation}
and $\Psi$ is the digamma function.
The contribution of diagram 1) in Fig.~\ref{fig:diags-NLO} is finite and reads:
\begin{equation}
\overline{\Sigma}_1 = -4 \hat{\Pi} \beta, ~~~
\hat{\Pi} = \frac{92}{9}-\pi^2\, ,
\label{sigma-NLO-1}
\end{equation}
where~\cite{GusyninHR01,Teber12+KotikovT13} the contribution of $\hat{\Pi}$ arises from the two-loop polarization operator in dimension $D=3$ which may be graphically represented as:
\begin{equation}
\parbox{8mm}{
\begin{fmfgraph*}(8,7)
\fmfleft{i}
\fmfright{o}
\fmfleft{ve}
\fmfright{vo}
\fmffreeze
\fmfforce{(-0.3w,0.5h)}{i}
\fmfforce{(1.3w,0.5h)}{o}
\fmfforce{(0w,0.5h)}{ve}
\fmfforce{(1.0w,0.5h)}{vo}
\fmffreeze
\fmf{photon}{i,ve}
\fmf{photon}{vo,o}
\fmffreeze
\fmfdot{ve,vo}
\fmf{phantom,tag=1}{ve,vo}
\fmfposition
\fmfipath{p[]}
\fmfiset{p1}{vpath1(__ve,__vo)}
\def\blob#1{%
\fmfiv{decor.shape=circle,decor.filled=shaded,decor.size=1w}{#1}}
\blob{point length(p1)/2 of p1}
\end{fmfgraph*}
} \qquad = \quad 2 \times ~
\parbox{8mm}{
\begin{fmfgraph*}(14,8)
\fmfleft{i}
\fmfright{o}
\fmf{photon}{i,v1}
\fmf{photon}{v2,o}
\fmf{phantom,right,tension=0.1,tag=1}{v1,v2}
\fmf{phantom,right,tension=0.1,tag=2}{v2,v1}
\fmf{phantom,tension=0.1,tag=3}{v1,v2}
\fmfdot{v1,v2}
\fmfposition
\fmfipath{p[]}
\fmfiset{p1}{vpath1(__v1,__v2)}
\fmfiset{p2}{vpath2(__v2,__v1)}
\fmfiset{p3}{vpath3(__v1,__v2)}
\fmfi{plain}{subpath (0,length(p1)) of p1}
\fmfi{plain}{subpath (0,length(p2)/4) of p2}
\fmfi{plain}{subpath (length(p2)/4,3length(p2)/4) of p2}
\fmfi{plain}{subpath (3length(p2)/4,length(p2)) of p2}
\fmfi{photon}{point length(p2)/4 of p2 .. point length(p3)/2 of p3 .. point 3length(p2)/4 of p2}
\def\vert#1{%
\fmfiv{decor.shape=circle,decor.filled=full,decor.size=2thick}{#1}}
\vert{point length(p2)/4 of p2}
\vert{point 3length(p2)/4 of p2}
\end{fmfgraph*}
} \qquad + \quad
\parbox{8mm}{
\begin{fmfgraph*}(14,8)
\fmfleft{i}
\fmfright{o}
\fmf{photon}{i,v1}
\fmf{photon}{v2,o}
\fmf{phantom,right,tension=0.1,tag=1}{v1,v2}
\fmf{phantom,right,tension=0.1,tag=2}{v2,v1}
\fmf{phantom,tension=0.1,tag=3}{v1,v2}
\fmfdot{v1,v2}
\fmfposition
\fmfipath{p[]}
\fmfiset{p1}{vpath1(__v1,__v2)}
\fmfiset{p2}{vpath2(__v2,__v1)}
\fmfi{plain}{subpath (0,length(p1)/2) of p1}
\fmfi{plain}{subpath (length(p1)/2,length(p1)) of p1}
\fmfi{plain}{subpath (0,length(p2)/2) of p2}
\fmfi{plain}{subpath (length(p2)/2,length(p2)) of p2}
\fmfi{photon}{point length(p1)/2 of p1 -- point length(p2)/2 of p2}
\def\vert#1{%
\fmfiv{decor.shape=circle,decor.filled=full,decor.size=2thick}{#1}}
\vert{point length(p1)/2 of p1}
\vert{point length(p2)/2 of p2}
\end{fmfgraph*}
} \qquad .
\label{polar-2loops}
\end{equation}
The contribution of diagram 2) in Fig.~\ref{fig:diags-NLO} is again singular. Dimensionally regularizing it yields:
\begin{equation}
2\,\overline{\Sigma}_2 =
-\frac{16}{3} \,\frac{\overline{\mu}^{2\varepsilon}}{p^{2\varepsilon}} \,\beta \left(\frac{1}{\varepsilon} + \Psi_1 + \frac{7}{3} + \frac{5\beta}{8} \right)
- 2 \hat{\Sigma}_2 + O(\varepsilon) \, ,
\label{sigma-NLO-2}
\end{equation}
where
\begin{eqnarray}
&&\hat{\Sigma}_2 = (1-4\alpha) \beta \Bigl[\Psi'(\alpha) - \Psi'(1/2-\alpha)\Bigl]
\nonumber \\
&&-\frac{\pi}{2\alpha}\,\tilde{I}_1(\alpha) - \frac{\pi}{2(1/2-\alpha)} \tilde{I}_1(\alpha+1) \, ,
\label{sigma-NLO-22}
\end{eqnarray}
and $\Psi'$ is the trigamma function.
Notice that the singularities in $\overline{\Sigma}_A$ and $\overline{\Sigma}_2$ cancel each other and their sum is therefore finite:
\begin{equation}
\overline{\Sigma}_A + 2\,\overline{\Sigma}_2 = -\frac{2}{3} \beta \Bigl(7\beta+8 \Bigr) -2 \hat{\Sigma}_2\, .
\end{equation}
This cancellation corresponds to the one of the logarithms,
$\ln(p/\alpha)$, in Ref.~[\onlinecite{Nash89}]; the importance of such cancellations
was discussed before, in Ref.~[\onlinecite{AppelquistNW88}]. The dimensionless integral $\tilde{I}_1(\alpha)$ appearing in Eq.~(\ref{sigma-NLO-22}) is defined as:
\begin{eqnarray}
&&I_1(\alpha) \equiv \frac{(p^2)^{-\alpha}}{(4\pi)^3}\, \tilde{I}_1(\alpha)
\label{I1-def} \\
&&= \int \frac{d^3k_1}{(2\pi)^3} \frac{d^3k_2}{(2\pi)^3}
\frac{1}{|p-k_1|k_1^{2\alpha} (k_1-k_2)^2 (p-k_2)^2|k_2|}\, ,
\nonumber
\end{eqnarray}
and obeys the following relation (it can be obtained by analogy with the ones in Ref.~[\onlinecite{Kazakov83}]):
\begin{widetext}
\begin{figure}
\includegraphics[width=0.99\textwidth]{fig2}
\caption{\label{fig:diags-NLO}
NLO diagrams to the dynamically generated mass $\Sigma(p)$. The shaded blob is defined in Eq.~(\ref{polar-2loops}).}
\end{figure}
\begin{eqnarray}
\tilde{I}_1(\alpha+1) = \frac{(\alpha-1/2)^2}{\alpha^2}
\tilde{I}_1(\alpha) - \frac{1}{\pi \alpha^2}
\Bigl[ \Psi'(\alpha) - \Psi'(1/2-\alpha) \Bigr]\, .
\label{I1-relation}
\end{eqnarray}
Using the results of Ref.~[\onlinecite{Kotikov95}], the integral $\tilde{I}_1(\alpha)$ can be represented
in the form of a two-fold series
\begin{equation}
\tilde{I}_1(\alpha) = \sum_{n=0}^{\infty} \sum_{l=0}^{\infty}\,\frac{B(l,n,1,1/2)}{(n+1/2)\,\Gamma(1/2)}
\times \Biggl[ \frac{2}{n+1/2} \left( \frac{1}{l+n+\alpha} + \frac{1}{l+n+3/2-\alpha} \right)
+\frac{1}{(l+n+\alpha)^2} + \frac{1}{(l+n+3/2-\alpha)^2} \Biggr], \,
\label{I1-series}
\end{equation}
where
\begin{eqnarray}
B(m,n,\alpha,1/2) = \frac{\Gamma(m+n+\alpha) \Gamma(m+\alpha-1/2)}{m!\Gamma(m+n+3/2)
\Gamma(\alpha)\Gamma(\alpha-1/2)} \, .
\label{B}
\end{eqnarray}
Finally, the contribution of diagram 3) in Fig.~\ref{fig:diags-NLO} is finite and reads:
\begin{eqnarray}
\overline{\Sigma}_3 = \hat{\Sigma}_3 + 3 \beta^2, \qquad
\hat{\Sigma}_3 =
(1/2-\alpha) \pi \tilde{I}_2(1+\alpha) + \frac{\pi}{2} \tilde{I}_2(\alpha) +
(\alpha-2) \pi \tilde{I}_3(\alpha)\, .
\label{sigma-NLO-3}
\end{eqnarray}
The dimensionless integrals in Eq.~(\ref{sigma-NLO-3}) are defined as: $\tilde{I}_2(\alpha)= \tilde{I}(\gamma=1/2,\alpha)$ and
$\tilde{I}_3(\alpha)= \tilde{I}(\gamma=-1/2,1+\alpha)$, where:
\begin{equation}
I(\gamma,\alpha) \equiv \frac{(p^2)^{-\alpha -\gamma+1/2}}{(4\pi)^3}\, \tilde{I}(\gamma,\alpha)
=\int \frac{d^3k_1}{(2\pi)^3} \frac{d^3k_2}{(2\pi)^3}
\frac{1}{(p-k_1)^{2\gamma}k_1^{2} (k_1-k_2)^{2\alpha} (p-k_2)^2|k_2|} \, .
\label{I-def}
\end{equation}
They satisfy the following relations:
\begin{equation}
\tilde{I}_2(\alpha) = \tilde{I}_2(3/2-\alpha),~~~
\tilde{I}_3(\alpha) = \frac{2}{4\alpha -1} \Bigl(\alpha \tilde{I}_2(1+\alpha) - (1/2-\alpha)
\tilde{I}_2(\alpha) \Bigr)
- \frac{\beta^2}{\pi}\, ,
\label{I2-I3-relations}
\end{equation}
and, thus, only one of them is independent.
Using the results of Ref.~[\onlinecite{Kotikov95}], the integral $\tilde{I}_2(\alpha)$ can be represented
in the form of a
three-fold series:
\begin{subequations}
\label{I2}
\begin{eqnarray}
&&\tilde{I}_2(\alpha)
= \sum_{n=0}^{\infty} \sum_{m=0}^{\infty}
B(m,n,\beta,1/2) \sum_{l=0}^{\infty}
B(l,n,1,1/2)
\times C(n,m,l,\alpha) \, ,
\label{I2-series}\\
&&C(n,m,l,\alpha) =
\frac{1}{(m+n+\alpha)(l+n+\alpha)}
+
\frac{1}{(m+n+\alpha)(l+m+n+1)}
+ \frac{1}{(m+n+1/2)(l+m+n+\alpha)}
\nonumber \\
&&+ \frac{1}{(m+n+1/2)(l+n+3/2-\alpha)}
+ \frac{1}{(n+l+\alpha)(l+m+n+\alpha)}
+ \frac{1}{(l+n+3/2-\alpha)(l+n+m+\alpha)}\, .
\label{I2-C}
\end{eqnarray}
\end{subequations}
\end{widetext}
Combining all of the above results, the gap equation (\ref{gap-eqn-NLO}) may be written in an explicit form as:
\begin{equation}
1 = \frac{2\beta}{L} + \frac{1}{L^2}\,
\Bigl[8 S(\alpha) -\frac{5}{3}\beta^2-\frac{16}{3}\beta -4 \hat{\Pi} \beta \Bigr]\, ,
\label{gap-eqn-NLO-explicit}
\end{equation}
where
\begin{equation}
S(\alpha) = (\hat{\Sigma}_3(\alpha)-2 \hat{\Sigma}_2(\alpha))/8 \, .
\label{delta}
\end{equation}
At this point, we consider Eq.~(\ref{gap-eqn-NLO-explicit}) directly at the critical point $\alpha=1/4$, {\it i.e.}, at $\beta=16$.
This yields:
\begin{eqnarray}
L_c^2 -32 L_c - 8\bigl(S-64- 8 \hat{\Pi}) = 0 \, ,
\label{Lc-eqn}
\end{eqnarray}
where $ S= S(\alpha=1/4)$.
Solving Eq.~(\ref{Lc-eqn}), we have two standard solutions:
\begin{eqnarray}
L_{c,\pm} = 16 \pm \sqrt{D},~~~ D= 8(S-32- 8 \hat{\Pi}) \, .
\label{Lc-solutions}
\end{eqnarray}
It turns out that the ``$-$'' solution is unphysical and has to be rejected because $L_{c,-}<0$. So, the physical solution is unique and corresponds to
$L_{c} = L_{c,+}$.
In order to provide a numerical estimate for $N_c$, we have used the series representations in order to evaluate the integrals:
$\pi \tilde{I}_1(\alpha=1/4) \equiv R_1$ and
$\pi \tilde{I}_2(\alpha=1/4 + i\delta) \equiv R_2 - iP_2 \delta + O(\delta^2)$ where $\delta \rightarrow 0$ regulates an artificial singularity in $\pi \tilde{I}_3(\alpha=1/4)=R_2 + P_2/4$.
With 10000 iterations for each series, we obtain the following numerical estimates:
\begin{equation}
R_1=163.7428, \quad R_2=209.175, \quad P_2=1260.720 \, .
\label{Is-numerics}
\end{equation}
From these results, we may then obtain the numerical value of $S=R_1-R_2/8-7P_2 /128$ which, combined with the one of $\hat{\Pi}$, yields
$L_c=32.45$ and therefore $N_c=3.29$. This result shows that the inclusion of the $1/N$ corrections increases the
critical value of $N_c$ by only $1.5\%$ with respect to its LO value.
\section{Conclusion}
We have included $O(1/N^2)$ contributions to the SD equation exactly and found that the critical value $N_c$ increased
by $1.5\%$ with respect to the LO result.
Our analysis is in nice agreement with ~[\onlinecite{Nash89}] and
therefore gives further evidence in favour of the solution found by Appelquist et al.~[\onlinecite{AppelquistNW88}].
Our results are in support of the fact
that the $1/N$ expansion of the kernel of the SD equation describes reliably the
critical behaviour of the theory.
In closing, let us briefly compare our study with the one of Nash~\cite{Nash89} which, to the best of our knowledge,
is the only popular reference which included NLO contributions in the gap equation of QED$_3$.
Our good agreement with Ref.~[\onlinecite{Nash89}] is nice but rather strange because
the two analyses are done in quite different ways. While we have used the Landau gauge (in accordance with recent results~\cite{BashirRSR09}
showing the gauge invariance of $N_c$ in this gauge when using the Ball-Chiu vertex), Nash worked with an arbitrary
gauge fixing parameter, $\xi$. He has resummed the most important NLO terms ($\propto \beta^2$
in our definition) which, together with the LO ones, lead to a gauge invariant result for $N_c$.
This result is larger by a factor $4/3$ than the pure LO one.~\cite{AppelquistNW88} The
rest of the NLO terms ($\propto \beta $) were evaluated (mostly numerically) in the Feynman gauge, which
modifies $N_c$ another time and gives the final result of Nash: $N_c=3.28$.
Finally, we also note that Nash obtained two possible solutions (one was considered as unphysical) while
we obtained a unique one.
For these reasons, and despite the surprising closeness of the final results,
our analysis substantially differs from that of Nash and intermediate expressions are difficult to compare.
\footnote{We have been informed by V.~Gusynin that the situation looks even more tricky; Eq.~(15) in Ref.~[\onlinecite{Nash89}] contains an error:
``341'' should be replaced by ``277'' which then leads to $N_c=3.52$.}
We also note that, very recently, NLO corrections were computed by Gusynin and Pyatkovskiy~\cite{Gusynin:2016som}
using a slightly different approach
than ours; they obtained a gauge-independent value $N_c =2.85$. Their value is remarkably close to the one
recently obtained by Herbut,~\cite{Herbut:2016ide} $N_c = 2.89$ using a completely different method.
In order to clear up the beautiful agreement we have with
Nash's results~\cite{Nash89} as well as the difference with the results of Gusynin and Pyatkovskiy,~\cite{Gusynin:2016som}
we plan to take into account of all $\xi$-dependent terms in our forthcoming publication.
\acknowledgments
We are grateful to Valery Gusynin for discussions.
One of us (A.V.K.) was supported by RFBR grant 16-02-00790-a.
Financial support from Universit\'e Pierre et Marie Curie and CNRS is acknowledged.
|
\section{The Problem}
This report is highly technical, treates a very special topic,
can not even reach the final answer to and is therefore not
considered for publication. Detailed Introductions are found in
\cite{kkn}, \cite{kn}, \cite{lmy}, \cite{fks}, \cite{ich}. Let
us jump over them here. Some are lengthy (e.\,g.\ 40 pages
in \cite{ich}).
According to Karabali, Kim and Nair \cite{kkn} the Hamiltonian can
be prepared to act in the space of wave functionals depending on
only the currents $j^a$. It reads
\be{1.1}
T + V = m \int\!\o_r^a \d_{j_r^a} +
m\int\!\!\int^{\prime}\!\O_{r r'}^{a b}
\, \d_{j_r^a} \d_{j_{r{\lower .3mm\hbox{\tiny '}}}^b} + V \quad , \quad V = {N\0{m\pi}}\int
\( {\ov \6} j_r^a \)\, {\ov \6 } j_r^a \quad . \quad
\ee
Here $a$ is the color index running from $1$ to $n \; {\dop} =} \def\glr{= {\dep} \; N^2-1 \,$,
$m = e^2 N / 2 \pi$ ($e$ the coupling), and $\int$ is shorthand for
$\int\!d^2 r$ running over the space of YM 2+1\,. Also $\int' = \int d^2r'\,$,
$\int'' = \int d^2r''$ etc\,. An Index $r$ on a quantity means that
it is a function of $\vc r = (x,y)\,$. Through $x-iy =z$, $x+iy = \ov z$
we have $r^2 = z\ov z$ and may define $\6 \; {\dop} =} \def\glr{= {\dep} \; d/dz$ and $\ov \6 \; {\dop} =} \def\glr{= {\dep} \;
d/d\ov z\,$. By $\; {\dop} =} \def\glr{= {\dep} \;$ or $\glr$ the object near to the colon is defined.
In strict temporal gauge there are only $2*n$ real gauge filds
$A^a_1\,,\;A^a_2\,$, combined to antihermitean and traceless
$N\times N$ fields by $A_j=-iA_j^aT^a$ or even to $A = (A_1+iA_2)/2\,$.
$T^a$ are the $n$ traceless generators of SU(N)\,: $\lk T^a,
T^b \rk = i f^{abc} T^c \; , \, \hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp} \( T^a T^b \) = {1\02} \d^{ab}
\; , \; T^a T^a = n/(2N)\;$. Matrices $M \in\;$SL(N,C)
parametrize the gauge fields by
\be{1.2}
A = - (\6 M)M^{-1} \;\; , \;\; A = -iT^a A^a \;\; , \;\; A^a =
i\,2 \hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp} \(T^a A \) \;\; , \;\; A^{ab} = -f^{abc} A^c \quad . \quad
\ee
The above third relation rests on $2 \hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp} \( T^a T^b \) = \d^{ab}\,$.
But it is also obtained by using $M^{ab} = 2\hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp}\(T^a M T^b M^{-1}\)\,$
in $\lk (\6 M)M^{-1}\rki^{ab}\,$.
The single-indexed fields $A^a$ specify the functional derivatives
in \eq{2.3} below.
The arguments of the functional derivatives in \eq{1.1} are
the currents
\be{1.3}
j^a = 2 \hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp} \,\( T^a j \) \;\, \hbox{with}\;\,
j = (\6 H)H^{-1} = T^a j^a \;\, , \;\,j^{ab}
= -if^{abc} j^c \;,\; j^a = {i\0N} f^{abc}j^{bc} \;, \quad
\ee
of the WZW model, where $H = M^\dagger M\,$ is a gauge invariant.
The last relation in \eq{1.3} derives via
$ H^{ab} = 2 \hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp} \( T^a H T^b H^{-1} \)\,$.
Again a superscript $r$ stands for $\vc r\,$ but mostly
$H_r$ is considered as depending on $z$ and $\ov z\,$:
$H_r = H_{z \ov z}\,$, $H_r{\lower .3mm\hbox{\tiny '}} = H_{z^\prime \ov z^\prime}$ etc.
Round brackets are used to stop the action of
a differentiation at the right bracket.
To distinguish fundamental from adjoint traces we shall write $\hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp}$ for
the first and $( \; ... ¸\; )^{aa}$ for the latter. Under an adjoint
trace inner color--index--pairs are often omitted as e\,.g.\ in
$\(j k \ell\)^{aa} \; {\dop} =} \def\glr{= {\dep} \; j^{ab} k^{bc} \ell^{ca}\,$.
As long as the regularization parameter $\epsilon$ is kept non--zero
positive the functional Schr\"o\-din\-ger equation $(T+V)\,\psi\!\lk j_r^a \rk
= E\,\psi\!\lk j_r^a \rk$ is fully regularized. There are no
singularities ({\sl ``} \def\grqq{'' no field theory\grqq } $\; - \;$ apart from the
continuum of variables). At first the solution $\psi$ is to be obtained ($\psi$
depending on $\epsilon\,$, of course). The limit $\epsilon \to +0$
is allowed only afterwards. The expressions $\o_r^a$ and $\O_{r r'}^{a b}$
in \eq{1.1} can be booked down in closed form (sction 2)
{\sl and} can be fully expanded as power series in $\epsilon$ with all
coefficients finite (section 3).
Former treatments perfomed $\epsilon \to +0$ too early, so they either
needed arguments or/and normal ordering or/and contact to known limiting
cases. Fine. Note that, thereby, all these authors accepted a certain
break in the consequent working through the Karabali--Nair setup. The
reason clearly is in the overwhelming complications otherwise. At least
Leigh, Minic and Yelnikov \cite{lmy} made an attempt in their appendix A
({\small \sl Regulated Computations}). They could not reach the expected
result. However they raised a conjecture about, what would have to
be done. In the present report we follow it up and can justify their idea.
Admittedly the desired final result will be not reached here as well
$\; - \;$ though by quite different reasons.
Adopting the ansatz $\psi = e^P$ for the wave functional, with $H$ from
\eq{1.1} and with $e^{-P} \d_{j^a} e^P = P^{\prime a} + \d_{j^a}\,$
in mind $\,\( P^{\prime a} \; {\dop} =} \def\glr{= {\dep} \; \d_{j^a} P \)$ the Schr\"odinger equation
turns into
\be{1.4}
m \int\!\o_r^a P_r^{\prime a} + m\int\!\!\int^{\prime}
\!\O_{r r'}^{a b} \,\( \d_{j_r^a} P_{r'}^{\prime a}
+ P_r^{\prime a} P_{r'}^{\prime b} \) = E - V \quad . \quad
\ee
Neither $\o$ nor $\O$ depend on $m\,$ (see section 2). By organizing
$P$ in powers of $1/m^2\,$, i.e. $P = P_0 +P_1 +P_2 +P_3 + \cdots$ with
$P_n \sim 1/m^{2n}\,$, \eq{1.4} decomposes in recursive equations \cite{kkn}
for $P_n\,$. In this note we concentrate on the leading nontrivial
term $P_1$ towards large $N$, i.e. $m \to \infty$ (the constant $P_0$
may be set equal to zero). For this task the term quadratic in $P$
in \eq{1.4} may be deleted since $\sim 1/m^4$. The now linear equation
for $P_1$ can simply be read off from \eq{1.4}. But in booking it down,
let us multiply the $P_1$--equation by $\pi m/N$ and rescale $P_1$, $E$ and $V$
to reach a more convenient form\,:
\be{1.5}
\cl T \,\;\cl P \; = \;\cl E - \cl V \quad , \quad \hbox{where} \quad
\ee
\be{1.6}
\cl T = {T \0 m} = \!\int\!\o_r^a \d_{j_r^a}
+ \!\int\!\!\!\int^{\prime}\!\O_{r r'}^{a b}
\,\d_{j_r^a}\d_{j_{r{\lower .3mm\hbox{\tiny '}}}^b} \quad \hbox{and} \quad
\ee
\be{1.7}
\cl P = {\pi m^2 P_1 \0 N} \quad , \quad \cl E =
{\pi m E \0 N} \quad , \quad \cl V = {\pi m V \0 N }
\!\int (\ov\6 j^a) (\ov\6 j^a) \quad . \quad
\ee
All the experts \cite{kkn} to \cite{fks} agree, that the solution to
\eq{1.5} at $\epsilon \to +0$ is given by
\be{1.8}
\cl P = -{1\02}\,\cl V = -{1\02}\int\! \({\ov \6} j_r^a \)\,{\ov \6 } j_r^a
\quad \hbox{or} \quad P_1 = -{\pi \0 2 N m^2 }\int
\( {\ov \6} J_r^a \)\, {\ov \6 } J_r^a \quad \quad
\ee
where $P_1$ is rewritten in terms of the original currents
$J^a= N j^a/\pi\,$ in \cite{kkn} and \cite{lmy} (but $j^a$ is $-J^a$
in \cite{fks}).
For solving $\,\cl T \,\cl P = \cl E - \cl V\,$ we will first observe
that $\cl T$ can be written as a power series in $\epsilon\,$.
If the unknown $\cl P$ is written as a power series aswell, a
coefficient comparison yields to a set of equations, one for each
given $\epsilon$--power. We emphasize that {\bf \,a\,l\,l\,} positive
powers of $\epsilon$ are included in Secions 2 and 3. Also Section 4
does so $\; - \;$ only the list \eq{4.10}, \eq{4.11} of invariants
remains too poor there.
To learn about the solution step by step, one can break apart
at a given $\epsilon^n$\,: ``} \def\grqq{'' step $n$\grqq\ neglects powers
higher than $n\,$. Each step ends up with a system of equations
for the pefactors of the holomorphic invariants involved. Then
$\epsilon \to 0$ always leads to $\cl P = c_0 \, \cl V\,$. As will be
seen in step 2 the result for $c_0\,$ varies under change
of the highest included power $n\,$. Ultimately $n \to \infty$
is required $\; - \;$ a strong support of the conjecture in \cite{lmy}.
The solutions to step 1 and step 2 are obtained in the sequel.
Also step 3 is attacked but will remain incomplete. This is all we
were able to really work through.
\let\dq=\thq \renewcommand{\theequation}{2.\dq}
\setcounter{equation}{0}
\section{Closed expressions for the kernels \boldmath$\o$ and \boldmath$\O$ in \eq{1.6}}
To derive (rederive in essence) these kernels we go a few steps
back in the Karabali--Kim--Nair analysis \cite{kkn}. Thereby special
attention is paid to maintain all $\epsilon$--dependence.
The regularised kinetic energy $\cl T := T/m$ originally reads
\be{2.1}
\cl T = {\pi\0 N}\int\!\int^\prime\!\int^{\prime\prime} H_r^{ab}
\; \ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}} \; \ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u \;
{\cl G }^{bv}_{r r{\lower .3mm\hbox{\tiny ''}}} \; p_{r{\lower .3mm\hbox{\tiny ''}}}^v \quad . \quad
\ee
Here the regularized Green functions are
\bea{2.2}
\ov {\cl G}^{au}_{r r{\lower .3mm\hbox{\tiny '}}} &=& \ov G_{r r{\lower .3mm\hbox{\tiny '}}} \( \d^{au}
- e^{ - (\vcsm r - \vcsm r^{\prime})^2 / \epsilon}
\( H_{z \ov z'}\; {H_{z' \ov z'}}^{\! -1}
\)^{au} \; \) \quad , \quad \nonumber \\
{\cl G}^{bv}_{r r{\lower .3mm\hbox{\tiny ''}}} &=& G_{r r{\lower .3mm\hbox{\tiny ''}}} \( \d^{bv}
- e^{ - (\vcsm r - \vcsm r^{\prime\prime})^2 / \epsilon}
\( {H_{z'' \ov z}}^{\! -1} \, H_{z'' \ov z''} \)^{bv} \; \)
\quad , \quad
\eea
and the operators $\ov p$, $p$ are linear in functional differentiations
with respect to $A^{a*}$ and $A^a\,$, respectively\,:
\be{2.3}
\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u = - \ov \6'\,
( M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger )^{ud} \d_{A^{d *}(\vcsm r')}
\quad , \quad
\; p_{r{\lower .3mm\hbox{\tiny ''}}}^v = \6^{\prime\prime}
( {M_{r{\lower .3mm\hbox{\tiny ''}}}}^{\!-1} )^{vc} \d_{A^c(\vcsm r^{\prime\prime})}
\quad , \quad
\ee
see \eq{1.3}\,.
The prefactors $\ov G$ and $G$ in \eq{2.2} are the bare Green functions
\be{2.4}
\ov G_{r r{\lower .3mm\hbox{\tiny '}}} = {1\0 \pi} {\ov z -\ov z' \0 (\vcsm r -
{\vcsm r}^\prime)^2 +\varepsilon^2}
\quad , \quad
G_{r r{\lower .3mm\hbox{\tiny ''}}} = {1\0 \pi} {z -z^{\prime\prime} \0 (\vcsm r -
{\vcsm r}^{\prime\prime})^2 +\varepsilon^2} \quad , \quad
\ee
Since the singularity, which is controlled by $\varepsilon$ here, is
compensated by the round brackets in \eq{2.2} ($\epsilon$ positive. Note that
$\epsilon \neq \varepsilon$), one may perform $\varepsilon \to +0$ there.
But otherwise it may happen that the limit $\varepsilon \to +0$ must wait
at least until $\ov \6\,\ov G_{r r{\lower .3mm\hbox{\tiny '}}} = \d (\vc r -\vc r')$ or
$\6\,G_{r r{\lower .3mm\hbox{\tiny ''}}} = \d (\vc r -\vc r^{\prime\prime})$ are
reached.\footnote{Just to illustrate this let us ignore the
above regularization details for a moment. It is a digression
but one that could support confidence in the whole setup.
We replace the round brackets in \eq{2.2} by $\d^{au}$ and
$\d^{bc}$, respectively. Now the Delta functions just mentioned
come into play by partial integrations in \eq{2.1}. $H^{ab}=
M^{\dagger a g} M^{gb}$ combines with the $M$'s in \eq{2.3}
to $\d^{cd}$\, ($M$ does not depend on $A^\ast\,$), and things
turn into familar quantum mechanics\,:
\vskip -4mm \hspace*{6.1cm} $\displaystyle} \def\0{\over } \def\6{\partial
\cl T \, = \, -{\pi\0 N} \int \d_{A^c *(\vcsm r)}
\d_{A^c(\vcsm r)} \, = \, {1\0 m} \; {e^2\0 2} \int {1\0 i}
\,\d_{A_j^a(\vcsm r)} {1\0 i} \d_{A_j^a(\vcsm r)} \;\; . $ \\[-4mm]
End of the digression.}
Since the kinetic energy $\cl T$ operates in the space of $\psi\!\lk j_r^a \rk\,$
it can be further reformulated. When applying \eq{2.1}
\be{2.5}
p_{r{\lower .3mm\hbox{\tiny ''}}}^v \psi = i\,\6^{\prime\prime} H_{r{\lower .3mm\hbox{\tiny ''}}}^{ev}\, \d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e}
\psi \quad
\ee
is needed first. Appendix A starts with a short derivation of \eq{2.5}\,.
Hence over this $\psi$--space \eq{2.1} becomes
\be{2.6}
\cl T = {\pi\0 N}\int\!\int^\prime\!\int^{\prime\prime} H_r^{ab}
\; \ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}} \; \ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u \;
{\cl G }^{bv}_{r r{\lower .3mm\hbox{\tiny ''}}} \;
i \6^{\prime\prime} H_{r{\lower .3mm\hbox{\tiny ''}}}^{ev} \d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e} \quad . \quad
\ee
$\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u$ does {\sl\,n\,o\,t\,} commute with ${\cl G }^{bv}_{r
r{\lower .3mm\hbox{\tiny ''}}}\,$, see \eq{3.8} below. But note that even positive powers of
$\epsilon$ are under study here. Hence by the chain rule the functional
derivative in $\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u\,$, \eq{2.3} splits $\cl T$ into three
terms (acting only on ${\cl G}^{bv}\,$, only on $H^{ev}$ and only at
the right end (i.e. on $ \d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e} \psi\,$), respectively.
The first two terms are linear in $\d$, the third term is quadratic.
Hence, with the notation of \eq{1.6}, there will be {\sl\,t\,w\,o\,}
contributions to $\o_r^a\,$ $\; - \;$ and hence in total three to $\cl T\,$:
\be{2.7}
\cl T = \cl T_1 + \cl T_2 + \cl T_3 \quad , \quad
\cl T_1 = \!\int\! \o_r^{\heartsuit a}\,\d_{j_r^a} \ \ , \ \
\cl T_2 = \!\int\! \o_r^{\diamondsuit a}\,\d_{j_r^a} \ \ , \ \
\cl T_3 = \!\int\!\!\!\int^{\prime}\!\O_{r r'}^{a b}
\,\d_{j_r^a}\d_{j_{r{\lower .3mm\hbox{\tiny '}}}^b} \quad . \quad
\ee
$\cl T_1$ is studied with much detail in Appendix A since this
has possibly never been done. But $\cl T_2$, $\cl T_3$ merely
recapitulate \cite{kkn} in Appendix A. Here we summarize the results\,:
\bea{2.8}
\o_r^{\heartsuit a} &=& -{i\0 N}\,\lk \6 -j_r \rki^{ab} f^{buv} \!\int^\prime\!
{ \,e^{ \,- 2\, (\vcsm r -{\vcsm r}^{\,\prime} )^2 / \epsilon } \0
\pi (\vcsm r -{\vcsm r}^{\,\prime})^2 }
\,\( H_{z \ov z'}\; {H_{z' \ov z'}}^{\!\! -1}
H_{z' \ov z}\; {H_{z \ov z}}^{\!\! -1} \)^{uv} \quad , \quad
\\ \label{2.9}
\o_r^{\diamondsuit a} &=& {i\0 N}\,f^{auv}\,\Theta_{rr}^{uv} \quad , \quad
\\ \label{2.10}
\O_{r r{\lower .3mm\hbox{\tiny '}}}^{ab} &=& {1\0 N}\,\lk\6 -j_r\rki^{ac}\;\Theta_{r r{\lower .3mm\hbox{\tiny '}}}^{cb}
\qquad \hbox{with} \quad
\\ \label{2.11}
& & \Theta_{r r{\lower .3mm\hbox{\tiny '}}}^{ab} \,=\, -\pi \!\int^{\prime\prime} \!
\ov {\cl G }^{ua}_{r{\lower .3mm\hbox{\tiny ''}} r}\; H_{r{\lower .3mm\hbox{\tiny ''}}}^{uv} \;
\( \6^\prime {\cl G }^{vc}_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} \) \( H_{r{\lower .3mm\hbox{\tiny '}}}^{-1}\)^{cb} \quad . \quad
\eea
By $\Theta = \pi \L\,$ the notation of \cite{kkn} is reached. There we are. And now\,?
\let\dq=\thq \renewcommand{\theequation}{3.\dq}
\setcounter{equation}{0}
\section{Expansion of \boldmath$\o^{\heartsuit}$, \boldmath$\o^{\diamondsuit}$
and \boldmath$\O$ in powers of \boldmath$\epsilon$}
\vskip -3mm
\subsection{ \boldmath$\o^{\heartsuit}$}
This is the sometimes omitted term of order $O(\epsilon)\,$. It is tempting
to perform the shift of variables $\vc r' \to \vc r' +\vc r$ under
the primed integral in \eq{2.8} and then writing the shifts by
$\ov z$ or $z$ in the $H$--arguments as Taylor expansions\,:
\be{3.1}
\o_r^{\heartsuit a} = -{i\0 N}\,\lk \6 -j_r \rki^{ab} f^{buv} H_r^{vw}\!\int^\prime\!
{ \,e^{ \,- 2\, r^{\,\prime\,2} / \epsilon } \0
\pi r^{\,\prime \,2 } } \sum_{s,t=0}^\infty {z^{\prime\,s}\,
{\ov z}^{\prime \,t} \0 s!\, t!} \,\hat{\6}^s \,\underline{\ov \6}^t
\,\( \underline{H}_{z \ov z}\; {\hat{\underline{H}}_{z \ov z}}^{\!\! -1}
\hat{H}_{z \ov z}\; \)^{uw} \quad . \quad
\ee
Here $\hat \6$ acts on $\hat H$ only and $\underline{\ov \6}$ only on
$\underline H\,$. The primed integral is nonzero only for $s=t$, and
at $s=t=0$ \eq{3.1} vanishes due to $(H H^{-1} H H^{-1})^{vu} = \d^{vu}\,$.
Now the integration can be done. Using $\int e^{-2 r^2/\epsilon } (r^2)^{s-1}/\pi
= (\epsilon/2)^s (s-1)!$ we get
\be{3.2}
\o_r^{\heartsuit a} = -{i\0 N}\,\lk \6 -j_r \rki^{ab} f^{buv} H_r^{vw}
\sum_{s=1}^\infty { (\epsilon/2)^s \0 s\; s!} \hat{\6}^s \underline{\ov \6}^s
\,\( \underline{H}_{z \ov z}\; {\hat{\underline{H}}_{z \ov z}}^{\!\! -1}
\hat{H}_{z \ov z}\; \)^{uw} \quad . \quad
\ee
$\hat\6$ acts on two of the $H$'s. Correspondingly, $\hat\6^s$ can be decomposed
binomically by
\be{3.3}
\hat\6^s = \sum_{k=0}^s \, { s! \0 k! \, (s-k)! } \;\6_{\hbox{\footnotesize} \def\ou{^{[U]} first}}^{k}
\; \6_{\hbox{\footnotesize} \def\ou{^{[U]} second}}^{s-k} \qquad , \quad \hbox{hence} \qquad
\ee
\vskip -8mm
\bea{3.4}
\sum_{s=1}^\infty \cdots \;\glo{8}\, \sum_{s=1}^\infty \sum_{k=0}^s
{ (\epsilon/2)^s \0 s\; k!\,(s-k)! }
\lk \hbox{\Large $($}\, {\ov\6}^s H ( \6^k H^{-1} )\,\hbox{\Large $)$}\,
\6^{s-k} H \rki^{uw} \quad \\ \label{3.5}
\glu{8}\, \sum_{s=1}^\infty \sum_{k=0}^s
{ (\epsilon/2)^s \0 s\; k!\,(s-k)! } \lk \( \ov \6^s B(k) \) A(s-k) H
\rki^{uw} \quad \quad
\eea
with the definitions
\be{3.6}
B(p) \; {\dop} =} \def\glr{= {\dep} \; H \6^p H^{-1} \quad \hbox{and} \quad A(p) \; {\dop} =} \def\glr{= {\dep} \; \( \6^p H\) H^{-1}
\quad , \quad B(p)^{ab} = A(p)^{ba} \quad , \quad
\ee
since $(H^{-1})^{ab} = H^{ba}$ from $H^{ab}=2\,\hbox{Tr}} \def{\rm \,Tr}{\hbox{Sp} \(T^a H T^b H^{-1}\)\,$.
Inserting \eq{3.5} into \eq{3.2}, redefining $k \to s-k$ and using
$B^{ab} = A^{ba}$ one obtains
\be{3.7}
\o^{\heartsuit a} = {i\0 N}\,\lk \6 -j \rki^{ab} f^{buv}
\sum_{s=1}^\infty \sum_{k=0}^s { (\epsilon/2)^s \0 s\; k!\,(s-k)! }
\( B(k)\, \ov \6^s A(s-k) \)^{uv}
\quad . \quad
\ee
For the final $\o^{\heartsuit a}$--version
\be{3.8}
\o^{\heartsuit a} = {i\0 N}\,\lk \6 -j \rki^{ab} f^{buv} \sum_{s=1}^\infty \sum_{k=0}^s
{ (\epsilon/2)^s \0 s\; k!\,(s-k)! } \; M(k,s,s-k,0)^{uv} \qquad
\ee
the more general definition
\be{3.9}
M(a,b,c,d) \; {\dop} =} \def\glr{= {\dep} \; B(a) \,\ov\6^b A(c) B(d)
\ee
has been introduced. Note that $B(0)=1$. With $j = (\6 H) H^{-1}$ it follows
from \eq{3.6} that $B(p) = \lk \6-j\rki^p\,1\,$ and in particular {\small
\bea{3.10}
B(1)\!\!&=&\!\! -j \;\; , \;\; B(2) = -\6 j + j\,j \;\; , \;\;
B(3) = -\6^2 j +(\6 j)\,j +2 j\,\6j - j\,j\,j
\quad \hbox{and} \qquad \qquad \nonumber \\
B(4)\!\!&=&\!\! -\6^3 j +(\6^2 j)\,j +3\,j\,\6^2 j +3\,(\6 j)\,\6 j
-(\6 j)\,j\,j -2\,j\,(\6 j)\,j -3\,j\,j\,\6 j +j\,j\,j\,j
\;\; . \qquad
\eea
} \noindent
$\!\!$For special $A$'s use $A^{ab}=B^{ba}\,$ ($A(0)=1,\, A(1)=j,\, \cdots$).
\eq{3.8} shows that the contribution $\o^{\heartsuit a}$ to $\o^a$
would have been bypassed by a too rush $\epsilon \to +0\,$.
\subsection{ \boldmath$\Theta$}
\eq{2.9} and \eq{2.10} show that both quantities, $\o^{\diamondsuit}$
and $\O\,$, are traced back to the object $\Theta_{r r{\lower .3mm\hbox{\tiny '}}}^{ab}\,$.
The rhs of \eq{2.11} has not quite the form of {\small\sl something}$^{ab}$.
But the indices on the troublemaker $\ov {\cl G}_{r{\lower .3mm\hbox{\tiny ''}} r}^{ua} \,$
(see \eq{2.2}\,) are easily reversed by
$ \( H_{z^{\prime\prime} \ov z}\; H_r^{-1}\)^{au} =
H_{z^{\prime\prime} \ov z}^{ac}\, H_r^{uc} =
\( H_r\; H_{z^{\prime\prime} \ov z}^{-1}\)^{ua}\,$.
In the sequel let us omit (but keep in mind) the superscipts $^{ab}$ on both sides.
Also let $e_{r{\lower .3mm\hbox{\tiny ''}} r}$ stand for $\exp\(-(r^{\prime\prime} - \vc r)^2/\epsilon\)\,$.
In otherwise full detail \eq{2.11} now reads
\be{3.11}
\Theta_{r r{\lower .3mm\hbox{\tiny '}}} = -\pi \!\int^{\prime\prime} \!
\ov G_{r{\lower .3mm\hbox{\tiny ''}} r} \( 1 - e_{r{\lower .3mm\hbox{\tiny ''}} r} H_r
H_{z^{\prime\prime} \ov z}^{-1} \) H_r{\lower .3mm\hbox{\tiny ''}}
\hbox{\Large $($}\,\6^\prime G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} \( 1 - e_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}
H_{z^\prime \ov z^{\prime\prime}}^{-1} H_r{\lower .3mm\hbox{\tiny '}} \) \,\hbox{\Large $)\,$}
H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \quad . \quad
\ee
Now $\ov G$ and $G$ may be replaced by their bare versions $(\e=0)$.
Note that $\6^\prime $ and $G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}$ can be commuted since
the Delta function would be multiplied by the vanishing round bracket.
Hence
\bea{3.12}
\Theta_{r r{\lower .3mm\hbox{\tiny '}}} &=& {1\0 \pi} \!\int^{\prime\prime} \!
{ e_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} \0 ( z^{\prime\prime} -z)
( \ov z^{\prime\prime} -\ov z^\prime )}
\Bigg( {\ov z^{\prime\prime} -\ov z^\prime \0 \epsilon}
H_r{\lower .3mm\hbox{\tiny ''}} H_{z^\prime \ov z^{\prime\prime}}^{-1}
+ H_r{\lower .3mm\hbox{\tiny ''}} \( \6^\prime H_{z^\prime \ov z^{\prime\prime}}^{-1} H_r{\lower .3mm\hbox{\tiny '}} \)
H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \hspace*{2cm} \nonumber \\
& & \hspace*{1mm}
- e_{r{\lower .3mm\hbox{\tiny ''}} r} {\ov z^{\prime\prime} -\ov z^\prime \0 \epsilon}
H_r H_{z^{\prime\prime} \ov z}^{-1}
H_r{\lower .3mm\hbox{\tiny ''}} H_{z^\prime \ov z^{\prime\prime}}^{-1}
- e_{r{\lower .3mm\hbox{\tiny ''}} r} H_r H_{z^{\prime\prime} \ov z}^{-1}
H_r{\lower .3mm\hbox{\tiny ''}} \( \6^\prime H_{z^\prime \ov z^{\prime\prime}}^{-1} H_r{\lower .3mm\hbox{\tiny '}} \)
H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \Bigg) \quad . \quad
\eea
\eq{3.12} is in perfect agreement with the four terms I to IV in
\cite{kkn} (the $\L$ there is $=\Theta/\pi\,$)\,. In \cite{kkn}
I to IV are eqs. (4.7 b to e). (Some of their subsequent equations
are not free from errors or typos). What follows here is new ground.
Towards evaluation of the integral the shift $\vc r^{\prime\prime} \to
\vc r^{\prime\prime} + \vc r^\prime$ is a first useful step\,:
\bea{3.13}
& & \hspace*{-17mm}
e_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} \to \exp{\( - r^{\prime\prime \,2} / \epsilon \,\)}
\; , \;
e_{r{\lower .3mm\hbox{\tiny ''}} r} \to \exp{\(- r^{\prime\prime \,2} / \epsilon
- \s \ov z^{\prime\prime} / \epsilon
- \ov\s z^{\prime\prime} / \epsilon
- \s \ov\s / \epsilon \,\)} \nonumber \\[1mm]
& & \hspace*{-17mm} \hbox{where}
\quad \s \; {\dop} =} \def\glr{= {\dep} \; z^\prime -z \quad \hbox{and} \quad
\ov \s \; {\dop} =} \def\glr{= {\dep} \; \ov z^\prime -\ov z \quad . \qquad
\eea
\vskip -11mm
\bea{3.14}
\Theta_{r r{\lower .3mm\hbox{\tiny '}}} \!\!&=&\!\! {1\0 \pi} \!\int^{\prime\prime} \! \Bigg( \;
{ e^{- r^{\prime\prime \,2} / \epsilon}
\0 \epsilon ( z^{\prime\prime} + \s ) }
\,H_{z^\prime + z^{\prime\prime}\;\ov z^\prime + \ov z^{\prime\prime}}
\, H_{z^\prime \; \ov z^\prime + \ov z^{\prime\prime}}^{-1}
\hspace*{7cm} \nonumber \\
& & \hspace*{11mm}
+ { e^{- r^{\prime\prime \,2} / \epsilon}
\0 \ov z^{\prime\prime} ( z^{\prime\prime} + \s ) }
\,H_{z^\prime + z^{\prime\prime}\;\ov z^\prime + \ov z^{\prime\prime}}
\( \6^\prime\, H_{z^\prime \; \ov z^\prime + \ov z^{\prime\prime}}^{-1}
H_r{\lower .3mm\hbox{\tiny '}} \) H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \nonumber \\
& & \hspace*{11mm}
- { e^{- 2 \, r^{\prime\prime \,2} / \epsilon}
\0 \epsilon ( z^{\prime\prime} + \s ) }
\,e^{- \s \ov z^{\prime\prime} / \epsilon}
\,e^{- \ov \s z^{\prime\prime} / \epsilon} \,e^{- \s \ov\s / \epsilon}
\,H_r H_{z^\prime + z^{\prime\prime}\; \ov z}^{-1}
\,H_{z^\prime + z^{\prime\prime}\;\ov z^\prime + \ov z^{\prime\prime}}
\, H_{z^\prime \; \ov z^\prime + \ov z^{\prime\prime}}^{-1} \nonumber \\
& & \hspace*{-16mm}
- { e^{- 2 \, r^{\prime\prime \,2} / \epsilon}
\0 \ov z^{\prime\prime} ( z^{\prime\prime} + \s ) }
\,e^{- \s \ov z^{\prime\prime} / \epsilon}
\,e^{- \ov \s z^{\prime\prime} / \epsilon} \,e^{- \s \ov \s / \epsilon}
\,H_r H_{z^\prime + z^{\prime\prime}\; \ov z}^{-1}
\,H_{z^\prime + z^{\prime\prime}\;\ov z^\prime + \ov z^{\prime\prime}}
\( \6^\prime\, H_{z^\prime \; \ov z^\prime + \ov z^{\prime\prime}}^{-1}
H_r{\lower .3mm\hbox{\tiny '}} \) H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \; \Bigg) \qquad
\eea
now shows that all the $z^{\prime\prime}$ and $\ov z^{\prime\prime}$
in the $H$--arguments can be transformed into powers by means of Taylor
expansions. Hence all the integrals to be done will be of the form
\be{3.15}
I_{p,q}(\epsilon,\s ) = {1\0 \pi} \!\int^{\prime\prime}\!
{ e^{- r^{\prime\prime \,2} / \epsilon} \0 z^{\prime\prime} +\s }
\; z^{\prime\prime \, p} \; \ov z^{\prime\prime \; q-1}
\quad . \quad
\ee
The four lines of \eq{3.14} become
\bea{3.16} \!\!
\Theta_{r r{\lower .3mm\hbox{\tiny '}}} \!\!&=&\!\! {1\0 \epsilon} \sum_{p=0}^\infty \sum_{q=1}^\infty
\,{1\0 p!\,(q-1)!}\, I_{p,q}(\epsilon,\s) \; \ov \6^{\prime \; q-1} \6^{\prime \; p}
H_{r{\lower .3mm\hbox{\tiny '}}} H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \nonumber
\\[-6mm] & & \hbox{\tiny
\hspace{5.1cm} $\uparrow$
\hspace{6.1mm} $\bullet$
\hspace{5mm} $\bullet\!\uparrow$
\hspace{2.4mm} $\uparrow$} \nonumber \\[2mm]
& &
+ \sum_{p=0}^\infty \sum_{q=0}^\infty \,
\,{1\0 p!\,q!}\, I_{p,q}(\epsilon,\s ) \; \ov \6^{\prime \; q} \6^{\prime \; p}
H_r{\lower .3mm\hbox{\tiny '}} \( \6^\prime H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} H_r{\lower .3mm\hbox{\tiny '}} \) H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \nonumber
\\[-6mm] & & \hbox{\tiny
\hspace{4.3cm} $\uparrow$
\hspace{2.9mm} $\bullet$
\hspace{5.1mm} $\bullet\!\uparrow$
\hspace{8mm} $\uparrow$} \nonumber \\[2mm]
& & \hspace*{-23mm}
-{1\0 \epsilon} e^{-\s \ov \s /\epsilon} H_r \sum_{p=0}^\infty \sum_{q=1}^\infty
\,{1\0 p!\,(q-1)!}\, I_{p,q}({\epsilon \0 2},\s) \
\!\lk \ov \6^\prime -{\s\0\epsilon}\rki^{q-1} \!\lk \6^\prime
-{\ov \s \0 \epsilon} \rki^p
H_{z^\prime \; \ov z}^{-1} H_r{\lower .3mm\hbox{\tiny '}} H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \nonumber
\\[-6mm] & & \hbox{\tiny
\hspace{5.4cm} $\uparrow$
\hspace{20.3mm} $\bullet$
\hspace{16.6mm} $\bullet$
\hspace{4.5mm} $\bullet\!\uparrow$
\hspace{2.2mm} $\uparrow$} \nonumber \\[2mm]
& & \hspace*{-23mm}
-e^{-\s \ov \s /\epsilon} H_r \sum_{p=0}^\infty \sum_{q=0}^\infty
{1\0 p!\,q!}\, I_{p,q}({\epsilon \0 2},\s)\,
\!\lk \ov \6^\prime -{\s\0\epsilon}\rki^q \!\!\lk \6^\prime
-{\ov \s \0 \epsilon}\rki^p
\!H_{z^\prime \; \ov z}^{-1} H_r{\lower .3mm\hbox{\tiny '}} \( \6^\prime H_{r{\lower .3mm\hbox{\tiny '}}}^{-1}
H_r{\lower .3mm\hbox{\tiny '}}\) H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \; . \\[-6mm] & & \hbox{\tiny
\hspace{4cm} $\uparrow$
\hspace{15.1mm} $\bullet$
\hspace{16.1mm} $\bullet$
\hspace{4.8mm} $\bullet\!\uparrow$
\hspace{8.1mm} $\uparrow$} \nonumber
\eea
The meaning of the subscipts\,: a dotted $\6^\prime$ acts {\sl \,o\,n\,l\,y\,}
on dotted $H$--arguments and an arrowed only on arrowed. In particular,
$\ov \6^\prime$ in the last two lines does {\sl \,n\,o\,t\,} act on
$\ov \s $ in the second square bracket.
The integral $I\,$, \eq{3.15}, is evaluated in Appendix B
to be
\be{3.17}
I_{p,q}(\epsilon,\s ) \; = \; (-\s )^{p-q} \!\int_{\s \ov \s }^\infty\!\! dt\;\,
t^{q-1} e^{-t/\epsilon}\,\;-\,\;\theta_{q>p}\; (q-1)! \,(-\s)^{p-q} \,\epsilon^q
\quad . \;\;
\ee
Here $\theta_{q>p}$ is 1 for $q>p$ and zero otherwise. Surprizingly,
the left half of \eq{3.17} does not contribute to \eq{3.16} at all.
The check to this nice outcome is left to the
reader\footnote{\,Perform each of the four $p$--sums. Note that the
operator $\exp(-\s \6^\prime)$\hspace{-4.4mm}\lower 1.3mm\hbox{\tiny $\bullet$}
\,\,\,\,changes a dotted $H$--index $z^\prime$ into
$z^\prime -(z-z^\prime) = z\,$. Show that the third line of \eq{3.16}
is the negative of the first one. ``} \def\grqq{'' But $I(\epsilon/2,\ldots)$ contains
the wrong $\exp(-2 t/\epsilon)$\,!?\grqq . Well, use the $q$--sum
to produce the missing $\exp(+t/\epsilon)\,$: \\
$ \sum_{q=0}^\infty {1\0q!} \lk -t {\ov\6}^\prime/\s + t/\epsilon \rki^q
= {\rm e}^{\lk -t{\ov\6}^\prime / \s +t/\epsilon \rk} =
\sum_{q=0}^\infty {1\0q!} \lk -t{\ov\6}^\prime/\s \rki^q \,
{\rm e}^{+t/\epsilon}\,$. Show fourth line $= -$\,second.}.
Hence for use in \eq{3.16} the second term of \eq{3.17} suffices.
Due to $\theta_{q>p}$ all $q$--sums in \eq{3.16} become
$\sum_{q=p+1}^\infty\,$. The factor $(q-1)!$ either compensates with
the denominator or leaves a $q$ there. For just one more detail note
that the ends of the second and fourth line may be written as \\[-3mm]
\be{3.18}
\( \6^\prime H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} H_r{\lower .3mm\hbox{\tiny '}} \) H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \;=\;
H_{r{\lower .3mm\hbox{\tiny '}}}^{-1}\,\( -j_r{\lower .3mm\hbox{\tiny '}} + j_r{\lower .3mm\hbox{\tiny '}} \) \quad . \quad
\\ \lower 3mm\hbox{\tiny
\hspace{-6.4cm} $\uparrow$
\hspace{2.85cm} $\uparrow$
\hspace{9.5mm} $\uparrow$ }
\ee
For the further lengthy analysis we denote the four lines of
\eq{3.16} by \fall 1 , \fall 2 , \fall 3 , \fall 4 . One obtains
\bea{3.19}
\fall 1 + \fall 2 = {1\0\s} + {\cl C} j_{r^\prime}
+ \6^\prime {\cl C} &\hbox{with}&
{\cl C} = - \sum_{p=1}^\infty \sum_{q=p+1}^\infty
{(-\s)^{p-q} \0 p!\, q } \e^q \ov \6^{\prime \; q} \6^{\prime \;p}
H_{r{\lower .3mm\hbox{\tiny '}}} {H_{r{\lower .3mm\hbox{\tiny '}}}}^{\!\!-1} \; . \qquad
\\[-6mm] & & \hbox{\tiny
\hspace{4.7cm} $\uparrow$
\hspace{3mm} $\bullet$
\hspace{4mm} $\bullet\!\uparrow$
\hspace{2mm} $\uparrow$} \nonumber
\eea
More laboriously one arrives at
\bea{3.20}
\fall 3 + \fall 4 = - {1\0\s} e^{-\s \ov \s /\epsilon} H_r\,
H_{z^\prime \ov z}^{-1} + {\cl E} j_{r^\prime}
+ \6^\prime {\cl E} &\hbox{with}& \hbox{\hspace*{5.6cm}} \nonumber \\
& & \hspace*{-7cm}
{\cl E} = \,e^{-\s \ov \s /\epsilon} H_r
\sum_{p=1}^\infty \sum_{q=p+1}^\infty
\( {\epsilon\0 2}\)^q
\lk \ov \6^\prime -{\s\0 \epsilon} \rki^q
\lk\6^\prime -{\ov \s \0 \epsilon} \rki^p
H_{z^\prime \ov z}^{-1} \, H_r{\lower .3mm\hbox{\tiny '}} H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \;\, . \quad
\\[-5.2mm] & & \hspace*{-2.4cm}
\hbox{\tiny
\hspace{9mm} $\uparrow$
\hspace{1.8cm} $\bullet$
\hspace{16.5mm} $\bullet$
\hspace{4.1mm} $\bullet\!\uparrow$
\hspace{2.3mm} $\uparrow$} \nonumber
\eea
Through \fall 1 + \fall 2 + \fall 3 + \fall 4 and with
$ {\cl A}:={\cl C} + {\cl E}$ we obtain
\be{3.21}
\Theta_{r r{\lower .3mm\hbox{\tiny '}}} = {1\0 \s } \( 1 - e^{-\s \ov \s /\epsilon} H_r
H_{z^\prime \ov z}^{-1} \) \, + \, \6^\prime
\cl A_{r r{\lower .3mm\hbox{\tiny '}}} + \cl A_{r r{\lower .3mm\hbox{\tiny '}}}\,j_r{\lower .3mm\hbox{\tiny '}} \quad . \quad
\ee
\vskip -2mm
The double sum over $p$ and $q$ is common to ${\cl C}$ and
${\cl E}$, hence also to ${\cl A}\,$.
A rearrangement of this double sum comes into mind by
\be{3.22}
\sum_{p=0}^\infty \sum_{q=p+1}^\infty =
\sum_{p=0}^\infty \( \sum_{q=p+1}^\infty - \sum_{q=1}^\infty \)
+ DS = - \sum_{p=1}^\infty \sum_{q=1}^p + DS
\quad \hbox{with} \quad DS := \sum_{p=0}^\infty \sum_{q=1}^\infty
\; . \quad
\ee
${\cl A}$ can be split into the part with
the finite $q$--sum and the part with $DS$. Let the latter
part be called ${\cl F}$. Then
\bea{3.23}
{\cl F} &=& \sum_{q=1}^\infty {1\0 q} \lk
- \(-{\e\0\s} \ov\6^\prime \)^q + \(-{\e\0 2\s}
\lk \ov\6^\prime - {1\02} \rki^q \) \rk
H_{z \ov z^\prime}\,H_{r^\prime}^{-1} \nonumber
\\[-5.6mm] & & \hspace*{1.8cm}
\hbox{\tiny
\hspace{9mm} $\uparrow$
\hspace{2.6cm} $\uparrow$
\hspace{2.4cm} $\uparrow$} \nonumber \\[2mm]
&=& \lk {\rm ln}} \def\det{{\rm det} \( 1 + {\e\0\s} \ov \6^\prime \)
-{\rm ln}} \def\det{{\rm det} \( 1 + {\e\0 2\s} \ov\6^\prime -{1\02}\) \rk
H_{z \ov z^\prime}\,H_{r^\prime}^{-1}
= {\rm ln}} \def\det{{\rm det} (2)\, H_{z \ov z^\prime}\,H_{r^\prime}^{-1}
\\[-4.2mm] & & \hspace*{1cm}
\hbox{\tiny
\hspace{1.05cm} $\uparrow$
\hspace{2.8cm} $\uparrow$
\hspace{2cm} $\uparrow$} \nonumber
\eea
and this drops out in \eq{3.21} because
$ \6^\prime {\cl F} + {\cl F} j_{r^\prime} =0\,$.
Hence the suitable completion of \eq{3.21} is
\bea{3.24}
\cl A_{r r{\lower .3mm\hbox{\tiny '}}} &=& \sum_{p=1}^\infty \,\sum_{q=1}^p
\,{1\0 p!\,q} (-\s)^{p-q} \,\Bigg( \,\epsilon^q \,\ov \6^{\prime \,q}
\(\6^{\prime \,p} H_r{\lower .3mm\hbox{\tiny '}} \) H_{r{\lower .3mm\hbox{\tiny '}}}^{-1}
\hspace*{5cm} \nonumber
\\[-6mm] & & \hbox{\tiny
\hspace{4.1cm} $\uparrow$
\hspace{4.6mm} $\bullet$
\hspace{4.6mm} $\bullet\!\uparrow$
\hspace{4.8mm} $\uparrow$ } \nonumber \\[2mm]
& & \hspace*{1.8cm}
- \,e^{ -\s \ov \s /\epsilon} H_r \( {\epsilon\0 2}\)^q
\lk \ov \6^\prime -{\s\0 \epsilon} \rki^q \lk\6^\prime -{\ov \s \0 \epsilon} \rki^p
H_{z^\prime \ov z}^{-1} H_r{\lower .3mm\hbox{\tiny '}} H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \Bigg) \quad . \quad
\\[-4.2mm] & & \hbox{\tiny
\hspace{5.2cm} $\uparrow$
\hspace{17.1mm} $\bullet$
\hspace{17.1mm} $\bullet$
\hspace{4.06mm} $\bullet\!\uparrow$
\hspace{2.26mm} $\uparrow$} \nonumber
\eea
Two agreeable properties of \eq{3.19} might be emphasized.
All dependences on $\epsilon$ are contained (all positive and negative
$\epsilon$ powers $\; - \;$ imagine the exponential functions be expanded).
Secondly, there are no negative powers of $\s$ or $\ov \s$ in
$\Theta_{r r{\lower .3mm\hbox{\tiny '}}}\,$. This is a welcome fact towards coincidence limits
$\vc r^\prime \to \vc r$ which make $\s\,,\;\ov \s$ vanish. Such limits
occur when two functional derivatives $\; - \;$ remember
$\cl T_3 = \!\int\!\!\!\int^{\prime}\!\O_{r r'}^{ab} \,\d_{j_r^a}
\d_{j_{r{\lower .3mm\hbox{\tiny '}}}^b}\,$ and $\,\O_{r r{\lower .3mm\hbox{\tiny '}}}^{ab} = {1\0 N}\,\lk\6 -j_r\rki^{ac}
\Theta_{r r{\lower .3mm\hbox{\tiny '}}}^{cb}\,$ $\; - \;$ are applied to holomorphic invariants.
$\,\Theta_{r r{\lower .3mm\hbox{\tiny '}}}$ is further processed in the subsections 4.2 and 4.3\,.
\let\dq=\thq \renewcommand{\theequation}{4.\dq}
\setcounter{equation}{0}
\section{ Coincidence limits}
\nopagebreak
\vskip -3mm
\subsection{ Holomorphic invariants}
The mapping of $A$ to the $M$--space is not unique since
$A_r = - (\6 M_r) M_r^{-1}$ remains unchanged under
\be{4.1}
M\,\to\,M h^{\dagger}(\ov z) \quad
\,\curvearrowright\, \;\; M_r^\dagger\,\to\, h(z) M_r^\dagger\, \;\; \hbox{and} \;\;
H=M^{\dagger}M\,\to\, h(z) H h^{\dagger}(\ov z) \quad . \quad
\ee
Physics must not depend on the choice
of $h\,$. The solution $\cl P$ to $\,\cl T \,\cl P = \cl E - \cl V\,$
is ``} \def\grqq{'' physics\grqq . So $\cl P$ has to be holomorphic invariant.
$\cl P$ will turn out to be a linear combination of the invariants
listed in \eq{4.9} to \eq{4.11} below. By means of \eq{4.1}
one derives
\be{4.2}
j = (\6 H) H^{-1} \,\to\, (\6 h) h^{-1} + h j h^{-1} \qquad \hbox{but}
\qquad \ov\6 j \,\to\, h (\ov\6 j) h^{-1} \quad . \quad
\ee
For convenience we might leave the above $N\times N$--matrix--language
and ask for the holomorphic transformation of $j^a$ and $j^{ab}$.
Let us also shorten the notation a bit by
\be{4.3}
\ov\6 j \glr \ov\jmath \quad , \quad
\ov\6 j^a \glr \ov\jmath^a \quad , \quad
\ov\6 j^{ab} \glr \ov\jmath^{ab} \quad , \quad
\ov\6^2 j \glr \ov{\ov\jmath} \quad , \quad \hbox{etc.} \quad
\ee
The last equation in \eq{4.2} now reads
$\ov\jmath \,\to\, h\, \ov\jmath\, h^{-1}$.
From \eq{1.2}, which now reads $\ov\jmath^a =
2\,{\rm \,Tr} \( T^a \ov\jmath \)\,$, one derives that
\be{4.4}
\ov\jmath^a \, \to\, 2\,{\rm \,Tr} \( T^a\, h\, \ov\jmath\, h^{-1} \)
= 2{\rm \,Tr} \(h^{-1}\, T^a\, h\, T^b\)\,2{\rm \,Tr} \(T^b\, \ov\jmath \)
= h^{ab}\, \ov\jmath^b \quad , \quad
\ee
where $\, h^{ab} = 2\,{\rm \,Tr}\( T^a h T^b h^{-1} \) = 2\,{\rm \,Tr}\( T^b h^{-1} T^a h \)
= (h^{-1})^{ba}\,$.
Hence $ \ov\jmath^a\, \ov\jmath^a$ is invariant, and so is $\cl V =\int
\ov\jmath^a\,\ov\jmath^a\,$. From $j^{ab} = \( (\6 H)H^{-1} \)^{ab}$
and \eq{4.2} one obtains
\bea{4.5}
j^{ab} &\to& \( (\6 h H h^\dagger ) h^{\dagger \, -1} H^{-1}
h^{-1} \)^{ab} \, = \, \( (\6 h) h^{-1} + h j h^{-1} \)^{ab}
\quad \hbox{but} \quad \nonumber \\
\lk\6-j\rki^{ab} &\to& \(\6 - (\6 h) h^{-1} - h j h^{-1} \)^{ab}
\,=\, \( h \, \lk\6-j\rk \,h^{-1} \)^{ab} \quad . \quad
\eea
The above details allow to recognize the following set
\be{4.6}
Q_n = \int \ov\jmath^a \( \lk [\6-j]\,\ov\6 \rki^n \)^{ab}
\ov\jmath^b \quad , \quad n = 0,1,2, \cdots\, \quad , \quad
\ee
as holomorphic invariants. Note that $Q_0 = \cl V\,$.
The invariants are conveniently written as adjoint traces. \eq{4.9}
below shows this trace for $Q_n\,$. It contains the operator
\be{4.7}
\cl D \; {\dop} =} \def\glr{= {\dep} \; \6 - \lk j\; , \,\;\; \rk \;\quad {\rm where} \;\quad
\lk j\; , \,\;\; \rk any \,\; {\dop} =} \def\glr{= {\dep} \;\, j\,any -any\,j \quad . \quad
\ee
Starting from \eq{4.6} the version
\eq{4.9} can be derived.\footnote{\, Consider $Q_n =
\int \ov\jmath^a \( \lk [\6-j]\,\ov\6 \rki^{n-1}\)^{ac}
\,[\6-j]^{cb}\;\ov{\ov\jmath}{\,}^b\,$ and insert
$\,j^{cb} =-i\,j^d f^{dcb}\,$ and $\,\ov{\ov\jmath}{\,}^b
= i\,f^{buv}\;\ov{\ov\jmath}{\,}^{uv}/N \,$ at the right end.
Now the Jacobi identity $\,f^{dcb}f^{buv} = -f^{udb}f^{bcv}
-f^{cub}f^{bdv}\,$ leads to
$\,[ \;\quad\; ]^{cb}\;\ov{\ov\jmath}{\,}^b\, =
\,i\, f^{cbv} \lk \cl D \,\ov \6 \,\ov \jmath \rki^{bv} /N \,$
$\; - \;$ and so on.}
As is seen in section 5 (and latest in section 6) {\sl \,m\,o\,r\,e\,}
holomorphic invariants are to be included. They will all be of the form
\be{4.8}
{1\0 N} \int \big(\, any \,\big)^{aa} \qquad \hbox{with}
\quad any \to h\,any\,h^{-1} \quad . \quad
\ee
Our attempts to solve $\,\cl T \,\cl P = \cl E - \cl V\,$ (up
to some maximal $\epsilon$--power) involve the followoing invariants
\bea{4.9}
Q_n &=& {1\0 N} \int \( \ov\jmath \lk\! \cl D \,\ov\6 \rki^n
\ov\jmath \)^{aa} \quad , \quad \\ \label{4.10}
R_{20} &=& {1\0 N} \int \big( \,\ov\jmath \,\ov\jmath \,\ov\jmath
\,\ov\jmath \,\big)^{aa}
\qquad , \qquad
R_{21} \;=\; {1\0 N} \int \( \ov{\ov\jmath} \,\ov\jmath \,
\cl D \,\ov\jmath \)^{aa}
\quad , \quad \\ \label{4.11}
& & \hspace*{-2.3cm}
R_{31}^{(1)} \,=\, {1\0 N} \!\int\!\( \,\ov{\ov\jmath} \,\ov\jmath
\,\ov\jmath \,\cl D \,\ov\jmath \)^{aa} \, , \,
R_{31}^{(2)} \,=\, {1\0 N} \!\int\!\( \,\ov\jmath \,\ov{\ov\jmath} \,\ov\jmath
\,\cl D \,\ov\jmath \)^{aa} \, , \,
R_{32} \,=\, {1\0 N}\!\int\!\(\,\ov{\ov\jmath}
\,\ov\jmath \,\cl D \,\ov\6 \,\cl D \,\ov\jmath \)^{aa} \; . \quad
\eea
The first index on $R$ refers to the $n$ of that $Q_n$, which under
application of $\cl T$ produces the $R$ in addition. The second index
just denotes the number of $\cl D$'s contained. $R_{n\,\cdots}$ has
the same mass dimension as $Q_n$, namely $m^{2n +2}$. In passing,
$\6,\,\ov\6,\,j$, $\cl D\,$ and $\d_j\,$ have mass dimension $m^1\,$,
$\epsilon$ and $\int$ have $m^{-2}$ and $\cl T$ is dimensionless.
\subsection{ Technicalities in applying \boldmath$\cl T$ to invariants}
If the $\cl D$'s are made explicit, an invariant is a linear
combination of products. To begin with $\cl T_1$ or $\cl T_2\,$,
let the $\d_{j_r^a}$ in there act on one factor of such a product,
for example on $\6 \ov{\ov \jmath}\,$:
$$
\int \o_r^a \d_{j_r^a} {1\0 N}\!\int^{\prime\prime} \! \(
\bullet\bullet\bullet_{r{\lower .3mm\hbox{\tiny ''}}} \;\; \6^{\prime\prime} \, \ov{\ov \jmath}_{r{\lower .3mm\hbox{\tiny ''}}} \;
\circ\circ\circ_{r{\lower .3mm\hbox{\tiny ''}}} \)^{cc} \;\;
\krumm{6}{3}{1.3}{ {-i\0 N} \!\int \o_r^a \,f^{cba} \!\int^{\prime\prime}\!\!
\( \6^{\prime\prime} \ov\6^{\prime\prime \, 2}
\, \d (\vc r - \vc r^{\prime\prime}) \)
\( \circ\circ\circ \; \bullet\bullet\bullet \)_{r{\lower .3mm\hbox{\tiny ''}}}^{bc} }
\;\;\;\; {-i\0 N} \int \o_r^a \d_{j_r^a}\, f^{cbd} \!\int^{\prime\prime}\!\!
\( \6^{\prime\prime} \, \ov{\ov \jmath}_{r{\lower .3mm\hbox{\tiny ''}}}^{\,d} \)
\( \circ\circ\circ \; \bullet\bullet\bullet \)_{r{\lower .3mm\hbox{\tiny ''}}}^{bc} $$ \\[2mm]
\bea{4.12}
& & \hspace*{4mm}
\6^{\prime\prime} \to - \6 \;\,\hbox{(on $\d$)} \;\to +\6 \;
\,\hbox{(by partial int.)\ , and similar with $\ov \6$ \ :}
\nonumber \\[-12mm]
\glu{18}
{-i\0 N}\!\int \( \6 \ov\6^2 \o_r^a \) \,f^{cba}
\( \circ\circ\circ \; \bullet\bullet\bullet \)_r^{bc}
\,=\, {1\0 N}\!\int \( (\6 \ov\6^2 \o )
\circ\circ\circ \; \bullet\bullet\bullet \)_r^{cc} \quad , \quad \hspace*{2.4mm}
\eea
where $ -i f^{cba} \o^a \glr \o^{cb}\,$ was defined. To check \eq{4.12}\,:
if $\o$ were $j\,$, the operator $\int j^a\d_{j^a}$ would just count the
number of $j$'s in the product, indeed. Remember that $\o^a$ is the sum
$\o^a = \o^{\heartsuit a} + \o^{\diamondsuit a}$ with $\o^{\heartsuit a}$
given by \eq{3.8}. But $\o^{\diamondsuit a}$ is rather taken from
\eq{4.18} below. Since $\o^a = \o^{\diamondsuit a} + O(\epsilon)\,$ and
$L^{uv}_{00} = j^{uv} +O(\epsilon)\,$ we have $\o^a=j^a+O(\epsilon)\,$.
To study the action of $\,\cl T_3\,$ on an invariant, two factors in a
product might be made explicit, for example $\ov \jmath$ and $\6 \ov \jmath\,$:
\be{4.13}
\!\int\!\!\!\int^{\prime}\!\O_{r r'}^{a b}
\,\d_{j_r^a}\,\d_{j_{r{\lower .3mm\hbox{\tiny '}}}^b}
{1\0 N}\!\int^{\prime\prime}\!\! \(
\bullet\bullet\bullet_r{\lower .3mm\hbox{\tiny ''}}
\;\; \ov\jmath_{r{\lower .3mm\hbox{\tiny ''}}} \;
\circ\circ\circ_{r{\lower .3mm\hbox{\tiny ''}}} \;\; \6^{\prime\prime} \,\ov\jmath_{r{\lower .3mm\hbox{\tiny ''}}}
\diamond\diamond\diamond_{r{\lower .3mm\hbox{\tiny ''}}} \)^{cc} \quad . \quad
\ee
Here $\O_{r r'}^{a b}$ may be replaced by
$\O_{r r'}^{a b} +\O_{r' r}^{b a} \glr \cl K_{r r'}^{ab}$, if the two
$\d$'s are applied ordered, i.e. interchanging them is forbidden.
Assume $\ov\jmath_{r{\lower .3mm\hbox{\tiny ''}}}$ carries indices$^{uv}$ at its position
in $(\;\;)^{cc}\,$. Then
$\d_{j_r^a}\ov\jmath_{r{\lower .3mm\hbox{\tiny ''}}} =-if^{uva} \ov\6^{\prime\prime}
\d\(\vc r -\vc r^{\prime\prime}\) = i f^{uav} \cdots\,$.
We may relax the notation by $f^{uav} \glr f^a\,$ since the
indices $u,v$ are fixed by the position of $f^a$ in $( \;\; )^{cc}\,$.
Hence
\bea{4.14}
\!\eq{4.13} \!\glo{7}\! {-1\0 N} \!\int\!\!\int^\prime\!\!\!\int^{\prime\prime}\!\!\!
\cl K_{r r'}^{ab} \( \ov\6^{\prime\prime}
\d (\vc r -\vc r^{\prime\prime} ) \)\,\(\6^{\prime\prime}
\ov\6^{\prime\prime} \d (\vc r^\prime -\vc r^{\prime\prime}) \)
\( \bullet\!\bullet\!\bullet f^a \!\circ\!\circ\!\circ f^b
\!\diamond\!\diamond\diamond \)_{r{\lower .3mm\hbox{\tiny ''}}}^{cc} \nonumber \\
\gluo{7}{7} {-1\0 N} \!\int\!\!\int^\prime\! \( \ov\6\,\6^\prime\,\ov\6^\prime
\cl K_{r r'}^{ab} \) \,\d (\vc r -\vc r^{\prime})\,
\( \bullet\bullet\bullet\, f^a \circ\circ\circ\, f^b
\diamond\diamond\diamond \)_r^{cc} \nonumber \\
\glu{7} {-1\0 N} \!\int\! \lb \ov\6\,\6^\prime\,\ov\6^\prime
\cl K_{r r'}^{ab}\rb \( \bullet\bullet\bullet\, f^a
\circ\circ\circ\, f^b
\diamond\diamond\diamond \)_r^{cc} \quad . \quad
\eea
The pair of curly brackets stands for the coincidence limit
$\lb any_{\,rr{\lower .3mm\hbox{\tiny '}}} \rb \; {\dop} =} \def\glr{= {\dep} \; \lim_{{\vcsm r}^\prime \to \vcsm r}
any_{\,rr{\lower .3mm\hbox{\tiny '}}}\,$.
Imagine a definite invariant with its linear--combined products,
each booked down with all combinations of $f^a$ and $f^b$
(the latter to the right of the first). Let each of the various
terms be given the form \eq{4.14}. So far things are done by
hand on paper. But now it is convenient to put the result
(for $\cl T_3\,${\small\sl Invariant}) in a MAPLE--file and continue
with it by keyboard and screen. There the following 5 steps
are done.
{\bf I.}\, Reintroducing $\O$ by $\,\cl K_{r r'}^{ab} = \O_{r r'}^{a b}
+\O_{r' r}^{ba}\,$ means that each $\O_{r r'}^{a b}$ becomes a partner with
interchanged indices and variables (hence $\6$'s become primed and
$\6^\prime$'s lose their primes).
{\bf II.}\, With reason explained in step {\bf V}, convert all
unprimed $\6$'s inside $\lb \,\, \rb$ into primed ones by using
\be{4.15}
\lb \6 \cdots \rb = \6 \lb \cdots \rb - \lb \6^\prime \cdots \rb
\quad \hbox{and} \quad
\lb \ov\6 \cdots \rb
= \ov\6 \lb \cdots \rb - \lb \ov\6^\prime \cdots \rb
\quad . \quad
\ee
To understand \eq{4.15} note that $\int \lb \6 \cdots \rb any_r
= \int\!\!\int^\prime\! (\6 \cdots ) \,\delta(\vc r - \vc r^\prime)\,
any_r\,$. Now partial integrate under $\!\int$, use
$\6\,\d = -\6^\prime \,\d$ and partial integrate with $\6^\prime$
under $\!\int^\prime$. Hence $\!\int \lb \6 \cdots \rb any_r$
$ = - \!\int \lb \6^\prime \cdots \rb any_r - \int
\lb \cdots \rb \6\, any_r \,$, which leads to \eq{4.15} by a
last partial integration.
{\bf III.}\, Partial integrate the unprimed $\6$'s in front of $\lb\;\;\rb\,$.
Thereby some of the $( \quad )^{cc}$'s are differentiated,
but they remain $( \quad )^{cc}$. Note the $1/N$--prefactor
of each invariant. Hence a typical term has the form
\be{4.16} \!
{1\0 N}\!\int\! \(..f^a..f^b..\)^{cc}
\lb \hbox{\footnotesize} \def\ou{^{[U]} primed $\6$'s}\;\; \O \rb^{ab}
= {1\0 N^2}\!\int\! \(..f^a..f^b..\)^{cc}
\lb \hbox{\footnotesize} \def\ou{^{[U]} primed $\6$'s}\;\; (\6 -j ) \,\Theta \rb^{ab}
\,\; ,
\ee
where \eq{2.10} was inserted.
{\bf IV.}\, Use \eq{4.15} for the last $\6$ in \eq{4.16} and partial
integrate if it appears in front of $\lb \;\; \rb$.
{\bf V.}\, It will be shown in the sequel that $\Theta$ can be written as
\be{4.17}
\Theta_{r r{\lower .3mm\hbox{\tiny '}}}^{ab} \;=\; \sum_{s=0}^\infty \,\sum_{t=0}^\infty
\,{\s^s\0 s!} \, {\ov \s^t \0 t!}\;\,L_{st}^{ab} \quad , \quad
\ee
with the coefficients $L$ depending on only the unprimed $z$
and $\ov z$. According to \eq{3.13} primed variables
occur in $ \Theta_{r r{\lower .3mm\hbox{\tiny '}}}$ , \eq{3.14}, only through
$\s =z^\prime -z$ and $ \ov \s =\ov z^\prime -\ov z\,$. This was
the reason, to favour primed differentiations in
point {\bf II.}\, A coincidence limit makes $\s$ and $\ov \s$ vanish.
So, it depends on the $\6^\prime$'s and $\ov\6^\prime$'s in front
of $\Theta$ which coefficient $L$ survives. Replace $\lb
\,{\rm primed}\, \6's \,\;\Theta \rb$
by this coefficient. End of the 5 MAPLE--steps.
An example for the result is
shown in Appendix C. Copy the result for $L$'s to a separate
MAPLE--file. Here $\cl T_1\,${\small\sl Invariant} and
$\cl T_2\,${\small\sl Invariant}\, might be included, and
the ready list (see Appendix D) of $L$'s can be inserted there.
By \eq{4.17} the ``} \def\grqq{'' second $\o$\grqq\ becomes explicit.
\eq{2.10} simply turns into
\be{4.18}
\o_r^{\diamondsuit a} \; = \; {i\0 N}\,f^{auv}\,\Theta_{rr}^{uv}
\; =\; {i\0 N}\,f^{auv}\,L_{00}^{uv}
\ee
\vskip -3mm
\subsection{ Derivation of the coefficients \,\boldmath$L$}
\eq{3.21} sugests to distinguish two parts\,: $\Theta =$ $^1\Theta$ $
+^2\Theta$ with $^1\Theta$ the first term of \eq{3.21} and $^2\Theta$
the remaining two terms in \eq{3.21} containing $\cl A\,$.
Correspondingly $L$ splits into $^{1\!}L$ and $^{2\!}L$. To start with
\bea{4.19}
^1\Theta \glo{7} {1\0 \s } \( 1 - H_r \,\e^{ \s\,(-\ov \s/\epsilon
+ \6\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\, \,)} H_{ z\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\, \ov z}^{-1} \)
= -{1\0 \s } H_r \sum_{s=1}^\infty {1\0 s!}\, \s^s (-\ov \s /\epsilon + \6 \,)^s
H_r^{-1} \qquad \quad \nonumber \\
\gluo{6}{8} - \sum_{s=1}^\infty \sum_{t=0}^s {1\0 t!\,(s-t)!}\,
\s^{s-1} (-\ov \s /\epsilon)^t
H_r \, \6^{s-t} H_r^{-1}
\quad , \quad s\to s+1 \, : \qquad \nonumber \\
\glu{9} \sum_{s=0}^\infty \sum_{t=0}^{s+1} \,{\s^s\0 s!}\,
{\ov \s ^t \0 t!}\;\,
{ s! \0 (s +1 -t)! } { (-1)^{1+t} \0 \epsilon^t } B(s+1-t) \quad . \quad
\eea
Defining $\theta_{t \leqslant s+1}$ to be 1 for $t \leqslant s+1$ and
zero otherwise $^{1\!}L$ is obtained as
\be{4.20}
^{1\!}L_{st} \; = \; \theta_{t \leqslant s+1}\; { s!\, (-1)^{1+t}\,
B(s+1-t) \0 (s +1 -t)!\;\, \epsilon^t } \quad . \quad
\ee
Towards $^{2\!}L$ let us assume that $\cl A$ can be written as
\be{4.21}
\cl A^{ab} = \sum_{s=0}^\infty \,\sum_{t=0}^\infty
\,{\s ^s\0 s!} \, {\ov \s ^t \0 t!}\;\, C_{st}^{ab} \quad , \quad
\ee
in general where the coefficients $C$ do not depend on $z^\prime$,
$\ov z^\prime\,$. If \eq{4.21} can be reached, $^{2\!}L$ can be
traced back to $C$ as follows\,:
\be{4.22}
^2\Theta = \6^\prime \cl A + \cl A \,j_r{\lower .3mm\hbox{\tiny '}} =
\sum_{s=1}^\infty \,\sum_{t=0}^\infty
\,{\s ^{s-1}\0 (s-1)!}\,{\ov \s ^t \0 t!}\;\, C_{st}^{ab}
+ \cl A \sum_{u=0}^\infty {\s^u \0 u!} \sum_{v=0}^\infty
{\ov \s ^v \0 v!}
\, \6^u \ov \6^v j_r \quad . \quad
\ee
Combining \eq{4.22} with \eq{4.21} one has to manipulate
\be{4.23}
\sum_{s=0}^\infty \sum_{u=0}^\infty {\s^{s+u} \0 s!\,u!} \,=\,
\sum_{u=0}^\infty \sum_{s=u}^\infty {\s^s \0 (s-u)!\,u!} \,=\,
\sum_{s=0}^\infty {\s^s \0 s!} \sum_{u=0}^s {s!\0 u! (s-u)! } \quad . \quad
\ee
Doing the same with the sums over $t$ and $v$ one can read off that
\be{4.24}
^{2\!}L_{st} = C_{s+1 \;t} + \sum_{u=0}^s \( \matrix{ s \cr u \cr} \)
\sum_{v=0}^t \( \matrix{ t \cr v \cr} \) C_{s-u \,\; t-v}
\,\; \6^u \ov \6^v j_r \quad . \quad
\ee
It remains to determine $C_{st}\,$ as defined by \eq{4.21}.
Clearly ${\cl A}$ from \eq{3.24} needs a further rearrangement
such that $C_{st}\,$ can be read off. The two lines of \eq{3.24}
have something in common, namely the two sums over $p$ and $q$
followed by $1\0 p!\, q$ and explicit powers $p$ and $q\,$.
In the following formula $\a = - \s
\6^\prime$\lower 2mm \hbox{$\!\!\!\!\!$\tiny$\bullet$\quad} ,
$\b = -{\epsilon\0\s} \ov \6^\prime$\lower 2mm
\hbox{$\!\!$\tiny$\uparrow$\quad}
apply to the first line of \eq{3.24} while
$\a = - \s \6^\prime$\lower 2mm \hbox{$\!\!\!\!$\tiny$\bullet$\quad}
$+\s \ov \s /\epsilon$ and $\b =- {\epsilon \0 2\s} \ov\6^\prime$\lower 2mm
\hbox{$\!\!\!$\tiny$\uparrow$\quad}$+{1\02}$ apply to the second\,:
\bea{4.25}
\sum_{p=1}^\infty {1\0 p!}\,\a^p \sum_{q=1}^p {1\0 q}\,\b^q
&=& \int_0^1 \! d\tau \sum_{p=1}^\infty {1\0 p!} \a^p \,\b\;
{1 - (\tau \b)^p \0 1-\tau\b \;}
= \int_0^1 \! d\tau \; {\b \( e^{(\tau\b -1)\a} -1 \)
\0 \tau\b -1}\,e^\a \nonumber \\
&=& \; \sum_{c=1}^\infty {1\0 c! \; c} \,
\Big( (\a\b-\a)^c - (-\a)^c \Big)\;e^\a \quad . \quad
\eea
The last expression was obtained by expanding the exponential
(sum over $c$) and performing the $\tau$ integration.
The explicit exponential $e^\a$ at the right end of \eq{4.25}
has welcome effects. It removes dot-subscripts by
$e^{-\s \6\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\,^\prime} H_{ z\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\,^\prime \ov z^\prime} =
H_{z \ov z^\prime}\,$ and
$e^{-\s \6\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\,^\prime} H_{ z\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\,^\prime \ov z}^{-1}
H_{ z\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\,^\prime \ov z^\prime}
H_{z^\prime \ov z^\prime}^{-1}\,$ in the first and second line
of \eq{3.24}, respectively. Moreover the second term
in the second-line-$\a$ compensates the prefactor
$e^{-\s \ov\s /\epsilon}$ in \eq{3.24}\,.
The corresponding intermediate result is
\bea{4.26}
\cl A &=& \sum_{c=1}^\infty {1\0 c! \; c} \;\Bigg\{
\( \big[ \s + \epsilon \ov \6^\prime \big]^c -\s^c \)
\6^c H_{z \ov z^\prime} H_r{\lower .3mm\hbox{\tiny '}}^{-1} \hspace*{7cm}\nonumber \\
& & \hspace*{18mm} - H_r {1\0 2^c}
\( \big[ \s + \epsilon \ov \6^\prime \big]^c -(2\s )^c \)
\( \6 -{\ov \s \0 \epsilon} \)^c
H_r^{-1} H_{z \ov z^\prime} H_{z^\prime \ov z^\prime}^{-1}
\Bigg\} \quad . \quad \\[-4.3mm]
& & \hbox{\tiny \hspace*{4.4cm} $\uparrow$ \hspace*{5.3cm}$\uparrow$
\hspace*{6mm}$\uparrow$} \nonumber
\eea
where the arrow--subscript prevents $\ov \6^\prime$ to act on
the explixit $\ov \s\,$.
There is still some cumbersome analysis left. We first separate
powers of $\s$ and $\ov \s$ binomically\,:
\bea{4.27}
\cl A &=&
\sum_{c=1}^\infty \sum_{k=1}^c {\epsilon^k \,\s^{c-k} \0 c\,k!\, (c-k)!} \;\cl B_1
\;\;-\;\;
\sum_{c=1}^\infty \sum_{k=1}^c \sum_{\ell=0}^c {\epsilon^{k-\ell}\,
\s^{c-k} c!\,(-\ov \s )^\ell \0 c\,k!\,(c-k)!\,
\ell!\,(c-\ell)!\,2^c} \;\cl B_2 \qquad \nonumber \\ & &
+\;\sum_{c=1}^\infty \sum_{\ell=0}^c { \s^c (-\ov\s )^\ell
(1 - 2^{-c}) \0 c \;\ell!\, (c -\ell)! \, \epsilon^\ell}
\;\cl B_3 \qquad \hbox{with} \quad
\eea
\be{4.28}
\cl B_1 = \ov\6^{\prime \;k} \6^c H_{z \ov z^\prime} H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \;\; , \;\;
\cl B_2 = H_r \ov\6^{\prime \; k} \6^{c-\ell}
H_{z \ov z}^{-1} H_{z \ov z^\prime} H_{r{\lower .3mm\hbox{\tiny '}}}^{-1} \;\; , \;\;
\cl B_3 = \cl B_2|_{k=0} \;\; . \quad
\ee
Now remember that primed indices $z^\prime$ and $\ov z^\prime$
must not occur in $C_{s\,t}\,$. With view to $z^\prime = z+\s\,$,
$\ov z^\prime = \ov z + \ov \s$ they can be removed by Taylor
expansions\,:
\bea{4.29}
\cl B_1 \glo{7} \ov \6^k
\lower 2.2mm\hbox{$\!\!\!\!\!$\tiny$\uparrow\;\;$}
\big(\6^c H_{z \,\ov z+\ov\s}
\lower 2.3mm\hbox{$\!\!\!\!\!\!\!\!$\tiny$\uparrow\;\quad$}
\;\big) H_{z+\s \, \ov z+\ov\s}
\lower 2.3mm\hbox{$\!\!\!\!\!\!\!\!$\tiny$\uparrow\;$}
\quad =\;
\sum_{s=0}^\infty \sum_{t=0}^\infty {\s^s \0 s!}
{\ov \s ^t \0 t!}\;
\ov \6^{t+k}\big(\6^c H_{z \, \ov z}\big)\; \6^s
H_{z \,\ov z}^{-1} \nonumber \\
\glu{8} \sum_{s=0}^\infty \sum_{t=0}^\infty {\s^s\0 s!}
{\ov \s ^t \0 t!}
\;M(0\,,\,t+k\,,\,c\,,\,s) \quad . \quad
\eea
\vskip -5mm
Remember $M(a,b,c,d)$ from \eq{3.9}. Furthermore
\bea{4.30}
\cl B_2 \glo{9} \sum_{s=0}^\infty \sum_{t=0}^\infty
H_r \,\6\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\,^{c-\ell} H_{ z\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\, \ov z}^{-1}\, {\ov \s ^t \0 t!}
\ov \6^{t+k} H_{ z\hspace{-1.6mm}\lower 1.2mm\hbox{\bf .}\, \ov z}\, {\s^s \0 s!} \,
\6^s H_{z \ov z}^{-1} \;\; , \;\; \nonumber \\
\gluo{6}{8} \sum_{s=0}^\infty \sum_{t=0}^\infty {\s^s\0 s!}
{\ov \s ^t \0 t!}
\sum_{a=0}^{c-\ell} {(c-\ell)! \0 a! \; (c-\ell-a)!}
H \( \6^{c-\ell-a} H^{-1} \) \ov\6^{t+k}\( \6^a H \) \6^s H^{-1}
\nonumber \\
\glu{9} \sum_{s=0}^\infty \sum_{t=0}^\infty {\s^s\0 s!}
{\ov \s ^t \0 t!}
\,\sum_{a=0}^{c-\ell} {(c-\ell)! \0 a! \; (c-\ell-a)!}
\;M(c-\ell-a\,,\,t+k\,,\,a\,,\,s) \quad . \quad
\eea
Inserting \eq{4.29},\eq{4.30} and $\cl B_3 = \cl B_2|_{k=0}$
into \eq{4.27} we are forced to commute a few double sums
to reach the structure \eq{4.21}\,. The final result is
\bea{4.31}
C_{s\,t}
&=& \sum_{c=0}^s {s! \0 (s-c)!} \sum_{k=1}^\infty \;\Bigg(\;
{\;\epsilon^k \; M(0,t+k,c+k,s-c) \0 c!\,k!\,(c+k) }
\hspace{4.7cm} \nonumber \\[1.5mm]
& & \hspace*{-1cm}
-\!\!\!\!\sum_{\ell=0}^{{\rm min} (t,c+k)}
\!\!\!\!\!{t! \0 (t-\ell)!} \sum_{a=0}^{c+k-\ell}
{\;\epsilon^{k-\ell} \; (-1)^\ell (c+k-1)!\,
M(c+k-\ell-a,t+k-\ell,a,s-c) \0
k!\;c!\;\ell!\;a!\;(c+k-\ell-a)!\; 2^{c+k} }
\,\;\Bigg) \nonumber \\[1.5mm]
& & \hspace*{-1.8cm}
+ \,\theta_{s\geqslant 1} \sum_{c=1}^s {s! \0 (s-c)!}
\!\!\sum_{\ell=0}^{{\rm min} (t,c)}
\sum_{a=0}^{c-\ell} {t! \0 (t-\ell)!} \;{(1-2^{-c})
\,(-1)^\ell \,M(c-\ell-a,t-\ell,a,s-c) \0
\epsilon^\ell \;\;c \;\ell!\; a!\; (c-\ell-a)!} \quad . \;\;
\eea
We shall work with \eq{4.31} neither by pen nor on sreen. But it
is an easy task to write \eq{4.31} into a file. MAPLE performs
the sums for index pairs ${s\,t}$ of interest. It also processes
$^{2\!}L_{s\,t}\,$, \eq{4.24}\,, and adds $^{1\!}L_{s\,t}\,$,
\eq{4.18}\,. The coefficients $M(a,b,c,d)$ appearing in $L$ can be
made explicit by hand on screen and read in. Hence all $L_{s\,t}$
of interest are known and expressed by the currents $j$.
A list of some $L$'s is shown in Appendix D.
We are quite sure that every detail of this section 4.3 is correct.
This rests on various successful tests. $\cl T$ applied to a
holomorphic invariant must result in a linear combination of
holomorphic invariants. Any erraneous little detail in e.g.\ \eq{4.31}
(and that had happened) would lead to the desaster of
non--holomorphic remnants in the result (and that had led to
find the error).
The above section is the last one which avoids approximations or
truncations. The kinetic energy operator $\cl T$ $\; - \;$ see \eq{2.8},
\eq{3.8},\eq{4.18}, \eq{4.17} and all about $C$'s and $L$'s $\; - \;$
contains the $\epsilon$--powers to all orders. It remained general.
The sad counterpart is that the whole apparatus developed above
will by far not be exhausted in what follows. \\[3mm]
\hspace*{15mm} \rule[1mm]{19mm}{.1mm} \, {\footnotesize} \def\ou{^{[U]}\sl
end of the general analysis that keeps $\epsilon$ arbitrary}
\, \rule[1mm]{19mm}{.1mm} \\[-8mm]
\let\dq=\thq \renewcommand{\theequation}{5.\dq}
\setcounter{equation}{0}
\section{ Results of {\boldmath$\cl T$}--application}
\vskip -4mm
\bea{5.1}
\cl T \,Q_0 &=& -{\aleph \0 2\,\epsilon^2} +{13\0 8}\,Q_0 + \epsilon\,{5\0 4}\,Q_1
+ \epsilon^2\({19\0 384}\,Q_2 + {1\0 48}\,R_{21} -{5\0 32}\,R_{20} \) \nonumber \\
& & +\,\epsilon^3 \({7\0 576}\,Q_3 + {35\0 288}\,R_{32} -{77\0 144}\,R_{31}^{(1)}
-{49\0 144 }\,R_{31}^{(2)} \) \; + \cl O\(\,\epsilon^4\,\) \quad . \quad
\\[2mm] \label{5.2}
\cl T \,Q_1 &=& +{\aleph \0 2\,\epsilon^3} +{69\0 32}\,Q_1 +\epsilon \(
{149\0 128}\,Q_2 -{11\0 8}\,R_{21} +{3\0 32}\,R_{20} \) \nonumber \\
& & +\,\epsilon^2 \( {33\0 512}\,Q_3 -{27\0 256}\,R_{32}
+{165\0 128}\,R_{31}^{(1)} +{105\0 128}\,R_{31}^{(2)} \)
\; + \cl O\(\,\epsilon^3\,\) \quad . \quad
\\[2mm] \label{5.3}
\cl T \,Q_2 &=& -{3\,\aleph \04\,\epsilon^4} +{1\0 8\,\epsilon^2}\,Q_0
+{3\0 2\,\epsilon}\,Q_1 +{1943\0 768}\,Q_2 -{25\0 96}\,R_{21}
+{7\0 64}\,R_{20} \nonumber \\
& & +\,\epsilon\( {35\0 32}\,Q_3 -{43\0 16}\,R_{32}
-{13\0 8}\,R_{31}^{(1)} -{11\0 8}\,R_{31}^{(2)} \)
\; + \cl O\(\,\epsilon^2\,\) \quad . \quad
\\[2mm] \label{5.4}
\cl T \,Q_3 &=& +{3\,\aleph\0 2\,\epsilon^5} +{3\0 8\,\epsilon^3}\,Q_0 -{103\0 32\,\epsilon^2}\,Q_1
+{1\0 \epsilon}\,\( {259\0 64}\,Q_2 +{43\0 32}\,R_{21} \) \nonumber \\
& & +\,{2749\0 1024}\,Q_3 -{1333\0 1536}\,R_{32} +{1403\0 768}\,R_{31}^{(1)}
+{1111\0 768}\,R_{31}^{(2)} \; + \cl O\(\,\epsilon\,\) \quad . \quad
\\[3mm] \label{5.5}
\cl T \,R_{20} &=& -{5\0 2\,\epsilon^2}\,Q_0 +{13\0 4}\,R_{20}
+\,\epsilon\( -{25 \0 8} R_{31}^{(1)} -{7 \0 4} R_{31}^{(2)} + {9\0 4}\, S \)
\; + \cl O\(\,\epsilon^2\,\) \;\; . \quad
\\[2mm] \label{5.6}
\cl T \,R_{21} &=& +{1\0 8\,\epsilon^2}\,Q_0 +{3\0 16\,\epsilon}\,Q_1 -{3\0 64}\,Q_2
+{45\0 16}\,R_{21} -{13\0 32}\,R_{20} \nonumber \\
& & +\,\epsilon\( {29\0 32}\,R_{31}^{(1)} -{13\0 16}\,\,R_{31}^{(2)} \)
\; + \cl O\(\,\epsilon^2\,\) \quad . \quad
\\[2mm] \label{5.7}
\cl T \,R_{31}^{(1)} &=& -{1\0 2\,\epsilon^3}\,Q_0 +{1\0 2\,\epsilon^2}\,Q_1
-{1\0 8\,\epsilon}\,R_{21} \nonumber \\
& & +\,{1\0 32}\,R_{32} +{133\0 32}\,R_{31}^{(1)}
+{17\0 64}\,\,R_{31}^{(2)} -{3\0 4}\,S \; + \cl O\(\,\epsilon\,\)
\quad . \quad
\\[2mm] \label{5.8}
\cl T \,R_{31}^{(2)} &=& -{1\0 4\,\epsilon^3}\,Q_0 +{1\0 4\,\epsilon^2}\,Q_1
-{1\0 4\,\epsilon}\,R_{21} \nonumber \\
& & -\,{1\0 16}\,R_{32} -{1\0 8}\,R_{31}^{(1)}
+{7\0 2}\,\,R_{31}^{(2)} +{1\0 4}\,S \; + \cl O\(\,\epsilon\,\)
\quad . \quad
\\[2mm] \label{5.9}
\cl T \,R_{32} &=& +{1\0 4\,\epsilon^2}\,Q_1 +{1\0 8\,\epsilon}\,Q_2
+{5\0 8\,\epsilon}\,R_{21} \nonumber \\
& & -\,{7\0 384}\,Q_3 +{637\0 192}\,R_{32} +{125\0 96}\,R_{31}^{(1)}
+{35\0 48}\,R_{31}^{(2)} \; + \cl O\(\,\epsilon\,\) \quad . \quad
\eea
Obviously, on the left hand sides $\cl T$ is applied to all the
invariants named in \eq{4.9} to \eq{4.11}. The necessity to include
$\epsilon$--powers diminishes with increasing index on $Q$ or left on $R\,$.
On the right hand sides there are two unknown quantities, which
have to be invariants too. The most simple holomorphic invariant
of the form \eq{4.8} is obtained by choosing ``} \def\grqq{'' $any$\grqq$=1\,$:
\be{5.10}
\aleph = {1\0 N} \int \big( \; 1 \; \big)^{aa} = {N^2-1\0 N} \,F
\qquad\quad
\ee
with $F$ the divergent area of the plane. The equation
$\cl T \,\cl P = \cl E - \cl V$ to be solved shows, that $\aleph$--terms
can be absorbed in $\cl E\,$. The quantity $S\,$
\be{5.11}
S \; {\dop} =} \def\glr{= {\dep} \; {1\0 N^2} \int \(\; f^a \;\ov\jmath\;\ov\jmath\;
f^a \,\ov{\ov \jmath}\, \cl D \ov\jmath \; \)^{cc} \quad , \quad
\ee
is also a holomorphic invariant (sum over $a$). To check this
invariance, use $h^{ab} = 2{\rm \,Tr} \( T^a h T^b h^{-1}\)\,$,
$f^{eas} T^s = -{\rm i}\lk T^e,T^a\rk$ and $2{\rm \,Tr}( \cdots T^e )
\,2{\rm \,Tr}(T^e \cdots) = 2{\rm \,Tr}(\cdots)$ to get $(h^{-1} f^a h)^{cr} =
f^{cdr} h^{ad}\,$. Hence the integrand $\big( \quad \big)^{cc} \glr
S_{\rm int}$ of $S$ transforms into
\bea{5.12} \!
S_{\rm int} &\!\!\to\!\!& ( h^{-1} f^a h )^{cr}\; (\ov\jmath\;\ov\jmath )^{rs}\;
( h^{-1} f^a h )^{st}\; (\,\ov{\ov \jmath}\,\cl D \ov\jmath \; )^{tc} \qquad \nonumber \\
&\!\!=\!\!& f^{cdr} (h^{-1})^{da} \; (\ov\jmath\;\ov\jmath )^{rs}\;
f^{sgt} h^{ag} (\,\ov{\ov \jmath}\,\cl D \ov\jmath \; )^{tc}
\,=\, \( \; f^g \;\ov\jmath\;\ov\jmath\;
f^g \; \ov{\ov \jmath}\,\cl D \ov\jmath \; \)^{cc} = S_{\rm int}
\;,\,\; \hbox{q.e.d.} \;\;\quad
\eea
All invariants occurring on the right hand sides of \eq{5.1} to \eq{5.9}
migth also appear under $\cl T$--operation on the left. The next
section will explain why. $\cl T \aleph = 0$ is no problem. However
through the evaluation of $\cl T S\,$ some unsurmountable difficulties
will become obvious in Section 6.3.
\let\dq=\thq \renewcommand{\theequation}{6.\dq}
\setcounter{equation}{0}
\section{ The first steps in solving \boldmath$\,\,\cl T \,\;\cl P \;
= \;\cl E - \cl V\,$}
The solution $\cl P$ to this equation is some function of $\epsilon\,$.
Only after this function is obtained the limit $\epsilon \to +0$
can be performed. This requirement is violated from the outset
if we truncate the {\sl ansatz} for $\cl P$ at some highest power
$n$ of $\epsilon$ (``} \def\grqq{'' step $n$\grqq ). Let us nevertheless do so
$\; - \;$ with thoughts on truncated perturbative expansions.
In the most simple $\cl P$--{\sl ansatz} let no positive $\epsilon$-power
be allowed at all\,: {\bf step zero}\,. Remember that $\cl V$ equals
$Q_0\,$ (combine \eq{1.7} with \eq{1.1}\,). With the
\be{6.1}
\hbox{{\sl ansatz }} \qquad \ \cl P = c_0\; Q_0 \qquad . \quad
\ee
The equation reads $c_0 \cl T Q_0 = \cl E - Q_0\,$. Omitting the
positive $\epsilon$ powers in \eq{5.1} one obtains
\be{6.2}
c_0 = -{8\0 13} \qquad \hbox{and} \quad
\cl E = {4\0 13\,\epsilon^2}\,\aleph \qquad . \quad
\ee
The result $\cl P = -{8\0 13}\,Q_0$ differs from \eq{1.8} i.e. from
$-{1\0 2}\,Q_0$ \,, $\; - \;$ the well known misfortune.
\vskip -3mm
\subsection{ Step one\ : \ including \boldmath$\epsilon^1$}
Now $\cl P$ must contain $\epsilon^0$ and $\epsilon^1$. The mass dimension $m^2$
of all terms in $\cl T \,\cl P = \cl E - Q_0$ is dictated by the
inhomogeneity $Q_0$ ($\epsilon$ has $m^{-2}$, $Q_1$ has $m^4\,$). So we are
lead to the
\be{6.3}
\hbox{{\sl ansatz }} \qquad \ \cl P = c_0\;Q_0 + c_1\;\epsilon\,Q_1
\quad \quad
\ee
and expect two equations for the coefficients by comparison.
With \eq{5.1} and \eq{5.2} one obtains
\be{6.4}
\(-c_0 +c_1\)\,{\aleph \0 2\,\epsilon^2} +c_0\,{13\0 8}\,Q_0 +
\epsilon\,\(\,c_0\,{5\0 4} + c_1\,{69\0 32} \)\,Q_1
= \cl E - Q_0 \quad , \quad
\ee
and hence
\be{6.5}
c_0 = -{8\0 13} \quad , \quad c_1 = {320 \0 13*69}
\qquad \hbox{and} \quad
\cl E = {\aleph \0 2\,\epsilon^2}\,\({320\069*13} + {8\0 13}\)
\quad . \quad
\ee
The limit $\epsilon \to +0$ ``} \def\grqq{'' afterwards\grqq\ reduces $\cl P$ to
$ -{8\0 13}\,Q_0$ again. No change.
\vskip -3mm
\subsection{ Step two\ : \ including \boldmath$\epsilon^2$}
$\cl P$ must contain $\epsilon^0\,$, $\epsilon^1\,$ and $\epsilon^2\,$.
Already \eq{5.1} shows that now the invariants $R_{21}$ and $R_{20}$
come into play. It will be seen that their inclusion has influence
on the $Q_0$--prefactor in question. Inspecting \eq{5.5} and \eq{5.6}
one may expect a closed set of equation. However five coefficients
are to be distinguished\,:
\be{6.6}
\hbox{{\sl ansatz }} \qquad \ \cl P = c_0\;Q_0 \,+ c_1\;\epsilon\,Q_1
\,+ c_2\;\epsilon^2\,Q_2 \,+ c_3\;\epsilon^2\,R_{21} \,+ c_4\;\epsilon^2\,R_{20}
\quad . \quad
\ee
Now the $\cl T$--applications \eq{5.1} to \eq{5.3} and \eq{5.5},
\eq{5.6} are in use. We compare the coefficients of $Q_0\,$,
$\epsilon Q_1\,$, $\epsilon^2 Q_2\,$, $\epsilon^2 R_{21}\,$ and $\epsilon^2 R_{20}\,$.
The resulting 5 equations can be read off from the following
MAPLE program\,:
\bea{6.7}
{\rm resu} \!&:=&\! {\rm solve}
\big(\;\big\{\; 13*c_0 + c_2 + c_3 -20*c_4 = -8\,, \hspace{4cm} \nonumber \\
& & \hspace{14mm} 40*c_0 + 69*c_1 + 48*c_2 + 6*c_3 = 0\,, \nonumber \\
& & \hspace{14mm} 38*c_0 + 894*c_1 + 1943*c_2 - 36*c_4 = 0\,, \nonumber \\
& & \hspace{14mm} 25*c_0 - 132*c_1 - 25*c_2 + 270*c_3 = 0\,, \nonumber \\
& & \hspace{-18mm} -10*c_0 + 6*c_1 + 7*c_2 - 26*c_3 + 208*c_4 = 0
\;\big\}\,,\,\big\{\, c_0,c_1,c_2,c_3,c_4 \,\big\}\, \big)\;; \quad . \quad
\eea
Again the limit $\epsilon \to +0$ ``} \def\grqq{'' afterwards\grqq\ reduces $\cl P$ to
the first term, i.e. only $c_0$ remains of interest for $\cl P$\,:
\bea{6.8}
\hbox{request} &:& c_0 = -0.5000 \hspace{5cm} \nonumber \\
\hbox{step 1} &:& c_0 = -0.6154 \,\; =\, -8/13 \nonumber \\
\hbox{step 2} &:& c_0 = -0.6330 \\
\big(\; \hbox{no R's} &:& c_0 = -0.5986 \;\big) \nonumber
\eea
Aside, the energy is $\,\cl E = ( -2\,c_0 +2\,c_1 -3\,c_3)
\;\aleph/(4\,\epsilon^2)\,$. \,Obviously, step 2 has led further away
from the ``} \def\grqq{'' request\grqq . But one should not expect too much from the
present lowest non--trivial order. The important outcome is merely {\bf
that} $c_0$ {\bf does change} under the inclusion of higher $\epsilon$ powers.
In the last line of \eq{6.8} we were curious what happens when
$R_{21}$ and $R_{20}$ were deleted from the {\sl ansatz} and
from \eq{5.1} to \eq{5.3}. Then \eq{6.7} would reduce to the first
three lines with $c_3 = c_4=0\,$. So, the $R$'s {\sl \,d\,o\,} influence
the value of $c_0\,$. Stimulated by the above we might study one more
step at least.
\subsection{ The problems with step three\ : \ including \boldmath$\epsilon^3$}
Now $\cl P$ must include $\epsilon^0\,$ to $\epsilon^3\,$, and (at least)
all $\cl T$--applications \eq{5.1} to \eq{5.9} become relevant.
In addition $\cl T\,S\,$, see \eq{5.11}), has to be evaluated,
but this makes trouble\,:.
\be{6.9}
\cl T\, S = {1\0 2\,\epsilon^3}\,Q_0 -{1\0 2\,\epsilon^2}\, Q_1
+{1\0 16\,\epsilon}\,R_{21} -{1\0 2\,\epsilon}\,S^{(2)} \; + \cl O\(\,\e^0\,\)
\quad . \quad
\ee
In \eq{6.9} one more object appears. $S^{(2)}$ is a holomorphic invariant
too and reads
\be{6.10}
S^{(2)} \; {\dop} =} \def\glr{= {\dep} \; {1\0 N^3} \int \( \; f^a \; \ov{\ov\jmath} \; f^b\;f^a\;
\ov\jmath \; f^b\; \cl D \ov\jmath \;\)^{cc} \qquad
\hbox{(sum over $a$ and $b$)}\footnote{\,$S$ and $S^{(2)}$
can be depicted as
\unitlength .8cm
\hspace*{6mm}
\begin{picture}(0,1)
\put(1.5,0){\oval(2,1)}
\put(2.5,.26){\line(3,1){.5}}
\put(2.5,-.26){\line(3,-1){.5}}
\put(1.5,-.5){\line(0,1){1}}
\put(.5,.26){\line(-3,1){.5}}
\put(.5,-.26){\line(-3,-1){.5}}
\put(3.2,-.55){$\cl D \ov\jmath$}
\put(3.2,.3){$\ov{\ov\jmath}$}
\put(-0.4,-.55){$\ov\jmath$}
\put(-0.4,.3){$\ov\jmath$}
\put(4.1,0){\hbox{and}}
\put(6.3,0){\oval(1,1)}
\qbezier(6.1,-.5)(6.33,-.02)(6.56,.46)
\qbezier(6.04,.46)(6.16,.22)(6.28,-.02)
\qbezier(6.32,-.12)(6.41,-.31)(6.5,-.5)
\put(6.3,.5){\line(0,1){.3}}
\put(6.8,-.1){\line(1,0){.3}}
\put(5.8,-.1){\line(-1,0){.3}}
\put(6.1,.9){$\cl D \ov\jmath$}
\put(5.2,-.2){$\ov\jmath$}
\put(7.2,-.2){$\ov{\ov\jmath}$}
\end{picture} \rule[-6mm]{0pt}{12mm}
\hspace*{5.7cm} \, , respectively\,. } \quad . \quad
\ee
It has mass dimension $m^6\,$ (see end of \S~4.1). $S$ and all terms
in \eq{6.9} have mass dimension $m^8\,$. In \eq{6.9} the term
$\cl O \(\,\epsilon^0\,\)$ is not made explicit (but should be). It is again
an invariant and has the structure of $S^{(2)}$ with one more $\cl D$ in.
Clearly, now $\cl T$ has to be applied to both of these new invariants.
It may be supposed that thereby even more holomorphic invariants
come into play. And so on, and so on $\; - \;$ a desaster. As a warning,
further steps in the desert become quite time consuming.
Is step three at its very end\,? There is a last chance.
We abandon $S$ from the {\sl ansatz}\,:
\be{6.11}
\cl P = c_0 Q_0 +c_1 \epsilon^2 Q_1 +c_2 \epsilon^3 Q_3
+c_4 \epsilon^2 R_{21} +c_5 \epsilon^2 R_{20} +c_6 \epsilon^3 R_{32}
+c_7 \epsilon^3 R_{31}^{(1)} +c_8\epsilon^3 R_{31}^{(2)} \;\; . \;
\ee
Then via \eq{5.5}, \eq{5.8}, \eq{5.9} a term
$\,\Delta \; {\dop} =} \def\glr{= {\dep} \; (9c_5 -3c_7 +c_8)\epsilon^3 S/4\,$ remains in $\cl T \cl P\,$.
May we hope for the miracle that this term vanishes\,? \,MAPLE solves
the 9 equations for $c_0$ to $c_8$ with ease. It produces
$\,c_0 = -0.5562\,$ \ $\; - \;$ \ and $\,\Delta =-0.275\,\epsilon^3 S\,
\neq 0$\,: no miracle.
\let\dq=\thq \renewcommand{\theequation}{7.\dq}
\setcounter{equation}{0}
\section{ Conclusions}
The kinetic energy operator $\cl T$ is worked out as a power
series in the regularizaion parameter $\epsilon\,$. $\cl T$ contains
all powers of $\epsilon$ and their prefactors are made explicit.
$\cl T$ is applied to various holomorphic invariants and leads to
linear combination of such invariants. Admittedly the detailed results
include only the third power of $\epsilon\,$ at highest.
To obtain the ground state $\psi =e^P$ (with $P$ restricted to its
leading nontrivial term towards $m \to \infty$) the Hamilton-Jacobi
type equation $\,\cl T \,\cl P = \cl E -\cl V\,$ needs to be solved
for $\cl P$. But actually only $\epsilon^1$ and $\epsilon^2$ could have been
included in an {\sl ansatz} for $\cl P$.
So far tested, the point splitting regularization works
well and is thus hopefully freed from possible doubts.
Due to the difficulties with breaking apart the $\epsilon$ power
series, one migth think about to sum it up. This could be
possible, but then\,: who has the strength to guess a suitable
{\sl ansatz} for $\cl P\,$?
\newpage
\let\dq=\thq \renewcommand{\theequation}{A.\dq}
\setcounter{equation}{0}
\parag{Appendix A : \,\hbox{\small How \eq{2.5} and
\eq{2.8}\,, \eq{2.9}\,, \eq{2.10} come about}}
\rule[.5mm]{3cm}{.9mm}\, \eq{2.5} : \\[1mm]
When restricting $\cl T$ to act in the space $\psi\!\lk j_r^a\rk$ and since
$p_{r{\lower .3mm\hbox{\tiny ''}}}^v$ is the operator at the right end of \eq{2.1} the first detail to be
studied is $p_{r{\lower .3mm\hbox{\tiny ''}}}^v \psi\!\lk j_r^a\rk\,$. With \eq{2.3} this reads
\be{A1}
p_{r{\lower .3mm\hbox{\tiny ''}}}^v \psi =
\6^{\prime\prime}( M_{r{\lower .3mm\hbox{\tiny ''}}}^{\!-1} )^{vc}
\d_{A^c(\vcsm r^{\prime\prime})} \, \psi
= \6^{\prime\prime}\, M_{r{\lower .3mm\hbox{\tiny ''}}}^{cv} \int^{\prime\prime\prime}\!\!
\( \d_{A^c(\vcsm r^{\prime\prime})} \,j_{r^{\prime\prime\prime}}^e \)
\,\d_{j_{r^{\prime\prime\prime}}^e} \psi \quad . \quad
\ee
For the above round bracket one may combine $j^e = 2{\rm \,Tr} \( T^e (\6 H) H^{-1} \)$
with $H = M^\dagger M$ to get
$j^e = 2{\rm \,Tr} \big( T^e (\6 M^\dagger) M^{\dagger -1}
+ T^e M^\dagger (\6 M) M^{-1} M^{\dagger -1} \big)$.
Now $(\6 M) M^{-1} = i T^b A^b$ (see e.g. p.22 of \cite{ich}) and the
$A$-independence of $M^\dagger$ lead to
\be{A2}
\d_{A^c(\vcsm r^{\prime\prime})} \,j_{r^{\prime\prime\prime}}^e
= i (M_{r{\lower .3mm\hbox{\tiny ''}}}^\dagger)^{ec}
\d (\vc r^{\prime\prime} - \vc r^{\prime\prime\prime})
\ee
Using \eq{A2} in \eq{A1} one arrives at \eq{2.5} in the main text.
\rule[.5mm]{3cm}{.9mm}\, \eq{2.8} : \\[1mm]
As explained below \eq{2.7} the kinetic energy operator $\cl T$ splits
into three parts. The first is $\cl T_1 = \int \o_r^{\heartsuit a}\,
\d_{j_r^a}$ and requires evaluation of
\bea{A3}
\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u \; {\cl G }^{bv}_{r r{\lower .3mm\hbox{\tiny ''}}} \glo{5}
-\ov \6'\,( M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger )^{ud} \d_{A^{d *}(\vcsm r')}
G_{r r{\lower .3mm\hbox{\tiny ''}}} \( \d^{bv} - e_{r r{\lower .3mm\hbox{\tiny ''}}} \( {H_{z'' \ov z}}^{\! -1}
\, H_{z'' \ov z''} \)^{bv} \; \) \quad , \quad \hspace*{2cm} \\[-4mm]
& & \ \ \hbox{\small where} \ e_{r r{\lower .3mm\hbox{\tiny ''}}} \ \hbox{\small is shorthand for} \ \
\exp\( -(\vc r -\vc r^{\prime\prime})^2/\epsilon\) \ , \quad \nonumber \\[-4mm]
\label{A4}
\glu{9} \ov \6'\,( M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger )^{ud} G_{r r{\lower .3mm\hbox{\tiny ''}}} e_{r r{\lower .3mm\hbox{\tiny ''}}}
\lk \( \cl R \d_{A^{d *}(\vcsm r')} H_{r{\lower .3mm\hbox{\tiny ''}}}^{cb} \) H_{r{\lower .3mm\hbox{\tiny ''}}}^{cv}
+ \( \cl R H_{r{\lower .3mm\hbox{\tiny ''}}}^{cb} \) \d_{A^{d *}(\vcsm r')} H_{r{\lower .3mm\hbox{\tiny ''}}}^{cv} \rk
\quad , \qquad \\
& & \ \ \hbox{\small where the operator $\cl R$ replaces
${\ov z}^{\prime\prime}$ by $\ov z$ within its round bracket .}
\quad \nonumber
\eea
To proceed with \eq{A4} we need
\be{A5}
\d_{A^{d *}(\vcsm r')} H_{r{\lower .3mm\hbox{\tiny ''}}}^{cb} =
{\ov G}_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} (M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger)^{gd}\,f^{cgh}\, H_{r{\lower .3mm\hbox{\tiny ''}}}^{hb}
\quad \quad
\ee
and the same with $v$ in place of $b\,$. To obtain \eq{A5}
derive from $M^{\dagger -1} \ov{\6} M^\dagger = -i T^a A^{a*}$
a differential equation for $\d_{A^{d *}(\vcsm r')} M_{r{\lower .3mm\hbox{\tiny ''}}}^\dagger$
and solve it to get
\be{A6}
\d_{A^{d *}(\vcsm r')} M_{r{\lower .3mm\hbox{\tiny ''}}}^\dagger
= -i M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger T^d M_{r{\lower .3mm\hbox{\tiny '}}}^{\dagger -1} \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}
M_{r{\lower .3mm\hbox{\tiny ''}}}^\dagger \ \ \hbox{or} \ \
\d_{A^{d *}(\vcsm r')} H_{r{\lower .3mm\hbox{\tiny ''}}}
= -i \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger T^d M_{r{\lower .3mm\hbox{\tiny '}}}^{\dagger -1}
H_{r{\lower .3mm\hbox{\tiny ''}}} \ \ , \quad
\ee
since in $H=M^\dagger M$ only $M^\dagger$ depends on $A^*\,$.
Now using this and $H_{r{\lower .3mm\hbox{\tiny ''}}}^{cb} = 2{\rm \,Tr}( T^c H_r{\lower .3mm\hbox{\tiny ''}} T^b H_{r{\lower .3mm\hbox{\tiny ''}}}^{-1} )$
on the lhs of \eq{A5} one obtains
\be{A7}
\d_{A^{d *}(\vcsm r')} H_{r{\lower .3mm\hbox{\tiny ''}}}^{cb} = -i \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}
2{\rm \,Tr} (T^c D B -T^c B D ) =
-i \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} 2{\rm \,Tr} (\lk T^h , T^c \rk\!\! D)\,2{\rm \,Tr}( T^h B)
\ \
\ee
with the temporary abbreviations $D = M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger T^d M_{r{\lower .3mm\hbox{\tiny '}}}^{\dagger
-1}$ and $B = H_r{\lower .3mm\hbox{\tiny ''}} T^b H_{r{\lower .3mm\hbox{\tiny ''}}}^{-1}$. From \eq{A7} one derives the
rhs of \eq{A5} indeed.
Using \eq{A5} in \eq{A4} we arrive at
\be{A8}
{\ov p}_{r{\lower .3mm\hbox{\tiny '}}}^u \; {\cl G }^{bv}_{r r{\lower .3mm\hbox{\tiny ''}}} =
G_{r r{\lower .3mm\hbox{\tiny ''}}}\, e_{r r{\lower .3mm\hbox{\tiny ''}}} \, f^{uhc}\,H_{r{\lower .3mm\hbox{\tiny ''}}}^{hv} \,
H_{z^{\prime\prime}\ov z}^{cb} \( {\ov \6}^\prime \,{\ov G}_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}
- \cl R \,{\ov \6}^\prime \, {\ov G}_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} \)\, \quad . \quad
\ee
Note that ${\ov \6}^\prime$ and $\cl R$ do commute. The round
bracket in \eq{A8} restricts their action. $\ov G$ is $\e$--regulated,
see \eq{2.4}. Due $\e \to 0$ the first term may be replaced by
$ - \d\(\vc r^{\prime\prime} -\vc r^\prime\)\,$ but the second
requires more care\,:
\bea{A9}
- \cl R \,{\ov \6}^\prime \,\ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} \,=\,
{\e^2/\pi \0 \lk (z^{\prime\prime} -z^\prime)\,(\ov z - \ov z^\prime)
+ \e^2 \rki^2 } \,=\,
e^{( {\ov z}^{\prime\prime} -\ov z ) {\ov \6}^\prime}
\,{\e^2/\pi \0
\lk (z^{\prime\prime} -z^\prime)\,(\ov z^{\prime\prime} - \ov z^\prime)
+ \e^2 \rki^2 } \;\; , \;
\eea
where the exponential is a translation operator ($\ov z^\prime \to
\ov z^\prime +\ov z^{\prime\prime} -\ov z\,$)\,.
The couple \eq{A8} (including \eq{A9}) is now well prepared for insertion
into \eq{2.6} at the right inner position. Thereby slightly rearranged
the result $\cl T_1$ reads\,
\bea{A10}
\cl T_1 &=& {i\pi\0 N} \!\int\!\!\int^\prime \!\int^{\prime\prime}
G_{r r{\lower .3mm\hbox{\tiny ''}}}\, e_{r r{\lower .3mm\hbox{\tiny ''}}} \, f^{uhc} H_r^{ab} H_{z^{\prime\prime}
\ov z}^{cb}\;\hbox{\Large \boldmath$($}\, -\ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}}\,
\d\(\vc r^{\prime\prime} -\vc r^\prime\) \nonumber \\
&+& \,{\e^2/\pi \0
\lk (z^{\prime\prime} -z^\prime)\,(\ov z^{\prime\prime} - \ov z^\prime)
+ \e^2 \rki^2 } \;
e^{( \ov z -{\ov z}^{\prime\prime}) {\ov \6}^\prime}\,
\ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}}\;\,\hbox{\Large \boldmath$)$} \,\;
H_{r{\lower .3mm\hbox{\tiny ''}}}^{hv} \6^{\prime\prime} H_{r{\lower .3mm\hbox{\tiny ''}}}^{ev} \,\d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e}
\quad . \quad
\eea
Obviously, under way from \eq{A9} to \eq{A10}, the translation operator
has changed its place. This was justified by partial integration under
$\int^\prime\,$. The right end of \eq{A10} that follows the round bracket
can be dealt with as $\lk \6^{\prime\prime} + j_{r{\lower .3mm\hbox{\tiny ''}}} \rki^{eh} \,\d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e}\,$.
We now encounter a surprizing problem in forming the limit $\e \to 0$
(it is the ``} \def\grqq{'' harmless\grqq\ $\e\,$). In \eq{A10} the second term
in the round bracket consists of three factors (counting
$\ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}}$ as two, see \eq{2.2}, and including
only $\ov G_{r r{\lower .3mm\hbox{\tiny '}}}$ next). The product of
the first two reads
\be{A11}
\cl P \,\; {\dop} =} \def\glr{= {\dep} \;\, {\e^2/\pi \0
\lk (z^{\prime\prime} -z^\prime)\,(\ov z^{\prime\prime} - \ov z^\prime)
+ \e^2 \rki^2 } \; \cdot\, {(\ov z^{\prime\prime} -\ov z^\prime )/\pi \0
(z -z^\prime )\,(\ov z^{\prime\prime} -\ov z^\prime ) + \e^2 } \;\; . \;\;
\ee
If (A) the limit is first performed in the first fraction, giving
$\d \(\vc r^{\prime\prime} -\vc r^\prime\,\)$, $\,\cl P = 0\,$ is obtained.
But if (B) the limit is first done in the second fraction one
ends up with $\cl P = \d \(\vc r^{\prime\prime} -\vc r^\prime\) \cdot
{1 \0 \pi (z-z^\prime) }\,$. Who wins $\; - \;$ (A) or (B) or none of\,?
To resolve this puzzle we integrate $\cl P$ over all $\vc r^{\prime\prime}$
and ignore the less important dependences on $\vc r^{\prime\prime}$ outside $\cl
P\,$\,:
\be{A12} \cl Q \,\; {\dop} =} \def\glr{= {\dep} \;\, \int\! d^2 r^{\prime\prime} \; \cl P \,=\,
\int\! d^2 r^{\prime\prime} \; {\e^2/\pi \0
\lk r^{\prime\prime\;2} + \e^2 \rki^2 } \;
\cdot\, {\ov z^{\prime\prime}/\pi \0
(z -z^\prime )\, \ov z^{\prime\prime} + \e^2 } \;\; . \;\;
\ee
With $(z -z^\prime) = \o {\rm e}^{i \b}\,$ ($\o$ real)\,, $\ov z^{\prime\prime}
= \rho \,{\rm e}^{i (\ph - \b)}\,$ (\,hence
$z^{\prime\prime} = \rho \,{\rm e}^{- i (\ph - \b)}\,$) and with
the notation $\schl \e \; {\dop} =} \def\glr{= {\dep} \; \e /(\o \rho)\,$ the integral turns into
\be{A13}
\!\!\!\cl Q = {1 \0 \pi^2 \o {\rm e}^{i \b}} \int_0^\infty\!\!\!d\rho\;
{\rho \0 \lk \rho^2 +1\rki^2}\,\cl J \;\, {\rm with} \;\, \cl J
= \int_{-\pi}^\pi \!\!\! d\ph {1 \0 1 + \schl \e {\rm e}^{-i\ph}} = \ldots =
2\pi \theta\( \o \rho - \e \) \;\,
\ee
and $\theta$ the step function. In the final result for $\cl Q\,$ the limit
$\e \to 0$ is unambiguous\,:
\be{A14}
\cl Q \,=\, {1/\pi \0 z-z^\prime} \int_{\e/\o}^\infty\!d\rho \; {2\,\rho \0
\lk \rho^2 +1\rki^2} \;\; \longrightarrow \;\; {1 \0 \pi \,(z-z^\prime)}
\quad . \quad
\ee
Version (B) wins. We conclude that $\cl P$ can be replaced by
${1 \0 \pi \,(z-z^\prime)} \, \d\(\vc r^{\prime\prime} -\vc r^\prime\)\,$.
The Delta function is now common to both terms in the round bracket of
\eq{A10} and helps performing $\int^{\prime\prime}\,$. The intermediate
result for $\cl T_1$ so far reached is
\bea{A15}
\cl T_1 &=& {i\0 N}\!\int\!\!\int^\prime
{e_{r r{\lower .3mm\hbox{\tiny '}}} \0 \pi (\vc r -\vc r^\prime)^2}\; f^{cuh}\, \cdot \nonumber \\
& & \( \, e_{r r{\lower .3mm\hbox{\tiny '}}} \rule[-.2mm]{0pt}{5mm}
H_r^{ab} H_{z^\prime \ov z}^{cb} H_{z {\ov z}^\prime}^{ad} H_{r{\lower .3mm\hbox{\tiny '}}}^{ud}
-H_r^{ab} H_{z^\prime \ov z}^{cb} H_r^{ad} H_{z^\prime \ov z}^{ud}
\;\) \lk \6^{\prime} + j_{r{\lower .3mm\hbox{\tiny '}}} \rki^{eh} \d_{j_{r{\lower .3mm\hbox{\tiny '}}}^e} \quad . \quad
\eea
Note that the second term in the round bracket reduces to $\d^{cu}\,$.
So it vanishes in contact with $f^{cuh}\,$. By partial integration
the square bracket can be placed as $\lk -\6^{\prime} + j_{r{\lower .3mm\hbox{\tiny '}}} \rki^{eh}$
at the beginning of the integrand. Then of course $-\6^\prime$ does
no more act on $\d_{j_{r{\lower .3mm\hbox{\tiny '}}}^e}\,$. Finally we interchange primed
and unprimed variables. Remembering
$\cl T_1 = \!\int\! \o_r^{\heartsuit a}\,\d_{j_r^a}\,$ the object
$\o_r^{\heartsuit a}$ can now be read off to be \eq{2.8}
in the main text.
\rule[.5mm]{3cm}{.9mm}\, \eq{2.9} : \\[1mm]
To learn about $\cl T_2 = \int \o^{\diamondsuit a} \d_{j_r^a}\,$
the functional derivative $\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u$ in \eq{2.7} has to be
placed in front of $H_{r{\lower .3mm\hbox{\tiny ''}}}^{ev}\,$. Using \eq{2.3} and \eq{A5}
we have
\bea{A16}
\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u H_{r{\lower .3mm\hbox{\tiny ''}}}^{ev}\, &=&
-\ov \6'\,( M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger )^{ud} \d_{A^{d *}(\vcsm r')} H_{r{\lower .3mm\hbox{\tiny ''}}}^{ev}\,
\nonumber \\
&=& -\ov \6'\,( M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger )^{ud}
{\ov G}_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} (M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger)^{gd}\,f^{egh}\, H_{r{\lower .3mm\hbox{\tiny ''}}}^{hv}
= - \d (\vc r^{\prime\prime} - \vc r^\prime) \,f^{euh}\, H_{r{\lower .3mm\hbox{\tiny ''}}}^{hv}
\quad \quad
\eea
and consequently
\bea{A17}
\cl T_2 \glo{7} -{\pi\0 N}\int\!\int^\prime\!\int^{\prime\prime} H_r^{ab}
\; \ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}} {\cl G }^{bv}_{r r{\lower .3mm\hbox{\tiny ''}}} \;
i \6^{\prime\prime} \d (\vc r^{\prime\prime} - \vc r^\prime)
\,f^{euh}\, H_{r{\lower .3mm\hbox{\tiny ''}}}^{hv} \d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e} \nonumber \\
\gluo{5}{6}
{ i \pi\0 N}\int\!\int^\prime\! H_r^{ab}
\; \ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}} \( \6^\prime {\cl G }^{bv}_{r r{\lower .3mm\hbox{\tiny '}}} \)
\,f^{euh}\, H_{r{\lower .3mm\hbox{\tiny '}}}^{hv} \d_{j_{r{\lower .3mm\hbox{\tiny '}}}^e}
\;\, \hbox{\small or by just changing notation} \nonumber \\[-4mm]
& & \hspace*{3cm}
\vc r \to \vc r^{\prime\prime} \,,\, \vc r^\prime \to \vc r \,,\,
v \to c \,,\, h \to v \,,\, a \to d \,,\, e \to a \, : \nonumber \\[-4mm]
\glu{9} \int \lk {i\0 N} f^{auv} (-\pi ) \!\int^{\prime\prime}\!
\ov{\cl G }^{du}_{r{\lower .3mm\hbox{\tiny ''}} r} H_{r{\lower .3mm\hbox{\tiny ''}}}^{db} \( \6 {\cl G }^{bc}_{r{\lower .3mm\hbox{\tiny ''}} r}\)
\, (H_r^{-1})^{cv} \rk \d_{j_r^a} \quad . \quad
\eea
The square bracket is $\o_r^{\diamondsuit a}\,$ indeed, see \eq{2.9}.
\rule[.5mm]{3cm}{.9mm}\, \eq{2.10} : \\[1mm]
According to the text below \eq{2.7} for the third part
$\cl T_3 = \!\int\!\!\!\int^{\prime}\!\O_{r r'}^{a b}
\,\d_{j_r^a}\d_{j_{r{\lower .3mm\hbox{\tiny '}}}^b}$ we have to place $\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u$ in front
of $\d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e} \psi\!\lk j\rk$ ($\glr \chi\!\lk j \rk$ for brevity).
\be{A18}
\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u \chi = -\ov \6'\, ( M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger )^{ud}
\d_{A^{d *}(\vcsm r')} \chi =
-\ov \6'\, ( M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger )^{ud} \int^{\prime\prime}
\( \d_{A^{d *}(\vcsm r')} j_{r{\lower .3mm\hbox{\tiny ''}}}^c \)
\d_{ j_{r{\lower .3mm\hbox{\tiny ''}}}^c} \chi
\ee
The round bracket is evaluated via $j_{r{\lower .3mm\hbox{\tiny ''}}}^c = 2 {\rm \,Tr} \( T^c
(\6^{\prime\prime} H_r{\lower .3mm\hbox{\tiny ''}} ) H_{r{\lower .3mm\hbox{\tiny ''}}}^{-1}\)$ and the right half of
\eq{A6}\,. Again abbreviating $M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger T^d M_{r{\lower .3mm\hbox{\tiny '}}}^{\dagger -1}$
by $D$ it follows that
\bea{A19}
\d_{A^{d *}(\vcsm r')}\,j_{r{\lower .3mm\hbox{\tiny ''}}}^c \glo{6}
-i ( \6^{\prime\prime} \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}) 2 {\rm \,Tr}(T^c D)
-i \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} 2 {\rm \,Tr}( \lk T^c , D\rk j_r{\lower .3mm\hbox{\tiny ''}} ) \qquad \nonumber \\
\gluo{6}{6}
-i (\6^{\prime\prime} \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}) (M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger)^{cd}
-i \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} 2 {\rm \,Tr}( \lk T^c , D\rk T^p )\, 2 {\rm \,Tr} (T^p j_r{\lower .3mm\hbox{\tiny ''}} )
\qquad \nonumber \\
\glu{6}
-i (M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger)^{cd} \6^{\prime\prime} \ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}
+ (M_{r{\lower .3mm\hbox{\tiny '}}}^\dagger)^{qd} \,\ov G_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}} f^{cqp} j_{r{\lower .3mm\hbox{\tiny ''}}}^p
\quad . \quad
\eea
Using \eq{A19} in \eq{A18} we obtain
\bea{A20}
\ov{p}_{r{\lower .3mm\hbox{\tiny '}}}^u \chi &=& \int^{\prime\prime}
\( - i \d^{uc} \6^{\prime\prime} \d (\vc r^{\prime\prime} - \vc r^\prime )
+ \d (\vc r^{\prime\prime} - \vc r^\prime ) f^{cup} j_{r{\lower .3mm\hbox{\tiny ''}}}^p \)
\d_{ j_{r{\lower .3mm\hbox{\tiny ''}}}^c} \chi \qquad \nonumber \\[1mm]
&=&
i \( \,\6^\prime - j_{r{\lower .3mm\hbox{\tiny '}}}\, \)^{uc} \,\d_{ j_{r{\lower .3mm\hbox{\tiny '}}}^c} \chi
\quad \quad
\eea
to be inserted in \eq{2.7}\,:
\bea{A21}
\cl T_3 \glo{7}{\pi\0 N}\int\!\int^\prime\!\int^{\prime\prime} H_r^{ab}
\; \ov{\cl G }^{au}_{r r{\lower .3mm\hbox{\tiny '}}} {\cl G }^{bv}_{r r{\lower .3mm\hbox{\tiny ''}}} \;
i \6^{\prime\prime} H_{r{\lower .3mm\hbox{\tiny ''}}}^{ev}
i \( \,\6^\prime - j_{r{\lower .3mm\hbox{\tiny '}}} \)^{uc}
\,\d_{ j_{r{\lower .3mm\hbox{\tiny '}}}^c} \d_{j_{r{\lower .3mm\hbox{\tiny ''}}}^e} \;\; , \quad \nonumber \\[-4mm]
& & \hspace*{2cm} \hbox{\small partial integrations and
$\vc r , \vc r^\prime , \vc r^{\prime\prime} \,\to\,
\vc r^{\prime\prime} , \vc r , \vc r^\prime\;$:} \nonumber \\[-3mm]
\glu{9}
\int\!\!\int^\prime {1\0 N} \( \,\6 - j_r \)^{cu}
(-\pi)\!\int^{\prime\prime}
\!\ov{\cl G }^{au}_{r{\lower .3mm\hbox{\tiny ''}} r} H_{r{\lower .3mm\hbox{\tiny ''}}}^{ab}
\( \6^{\prime} {\cl G }^{bv}_{r{\lower .3mm\hbox{\tiny ''}} r{\lower .3mm\hbox{\tiny '}}}\)
(H_{r{\lower .3mm\hbox{\tiny '}}}^{-1})^{ve} \,\d_{ j_r^c} \d_{j_{r{\lower .3mm\hbox{\tiny '}}}^e} \;\; . \quad
\eea
Now \eq{2.10} can be read off from \eq{A21}. End of
\cite{kkn}--recapitulations.
\let\dq=\thq \renewcommand{\theequation}{B.\dq}
\setcounter{equation}{0}
\parag{Appendix B : \,\hbox{\small Evaluation of the
integral {\boldmath$I_{p,q}(\epsilon,\s)\,$}, \eq{3.15}}}
At first we drop the double primes in \eq{3.15} and
set $z = r e^{-i\ph}$, $\ov z = r e^{i\ph}\,$:
\be{B1}
I_{p,q}(\epsilon,\s) = {1\0 \pi} \!\int_0^\infty \! dr \; r\, e^{-r^2/\epsilon}\,
r^{p+q-1} \!\int\!\! d\ph \,{ (e^{i\ph})^{q-p-1} \0
r e^{-i\ph} + \s } \quad . \quad
\ee
Next we write $\s = |\s| e^{-i\chi}$ and
\be{B2}
I_{p,q}(\epsilon,\s) = {1\0 \pi} \!\int_0^\infty \!\!\! dr \; r\, e^{-r^2/\epsilon}\,
r^{p+q-1} \s^{p-q} E_{q-p} \qquad \hbox{with} \qquad
\ee
\be{B3}
E_n \,\; {\dop} =} \def\glr{= {\dep} \; \int\!\! d\ph \,{ |\s|^n (e^{-i\chi})^n (e^{i\ph})^{n-1}
\0 r e^{-i\ph} + |\s| e^{-i\chi} } \;=\; \!\int\!\! d\ph \,
{ ( |\s| e^{i\ph} )^n \0 r + |\s| e^{i\ph} } \quad , \quad
\ee
where the last step rests on the shift $\ph \to \ph + \chi\,$.
Writing the numerator as \\
$( |\s| e^{i\ph} )^{n-1} (|\s|e^{i\ph} + r -r)$
one easily derives the recurrence relation
\be{B4}
E_n = \d_{n,1}\, 2\pi - r E_{n-1} \quad , \quad
E_0 = \int\! d\ph {1 \0 r + |\s|e^{i\ph} } = {2\pi \0 r}
\theta(r-|\s|) \quad . \quad
\ee
$\theta$ is the step function, cf.\,\eq{A13}\,. The solution to the problem
\eq{B4} is
\be{B5}
E_n = \lower .8mm\hbox{\LARGE \{ }\!\!\matrix{
\ \,(-r)^{n-1} \,2\pi \,\theta(|\s|-r)\ \ \hbox{for}\ \ n \ge 1 \cr
- (-r)^{n-1} \,2\pi \,\theta(r-|\s|)\ \ \hbox{for}\ \ n \le 0 \cr }
\quad . \quad
\ee
With \eq{B5} at $n=q-p$, substituting $t= r^2$ and using
$|\s|^2=\s \ov \s$ from
\eq{3.13}\,, \eq{B2} turns into
\bea{B6}
I_{q\le p} = (-\s)^{p-q} \int_{\s \ov\s}^\infty \! dt \; t^{q-1} e^{-t/\epsilon}
\quad \hbox{and} \quad
I_{q>p} = -(-\s)^{p-q} \int_0^{\s \ov \s} \! dt \; t^{q-1} e^{-t/\epsilon}
\quad . \quad
\eea
Through $\int_0^{\s \ov \s} = \int_0^\infty -
\int_{\s \ov \s}^\infty$ one arrives at
\eq{3.17} in the main text.
\let\dq=\thq \renewcommand{\theequation}{C.\dq}
\setcounter{equation}{0}
\parag{Appendix C : \,\hbox{\small Example of a MAPLE--file}}
In evaluating $\cl T_3\, Q_1 $ the five steps described
below \eq{4.14} lead to
\hspace*{1mm}
aa9 := \\
+2$*$ab$*$L23 \\
+4$*$abj$*$L13 \\
+6$*$abtj$*$L12 \\
\hspace*{.5mm}
-6$*$abttj$*$L11 \\
\hspace*{.5mm}
-4$*$abtttj$*$L10 \\[-6mm]
+4$*$abdj$*$L03 \\
+6$*$abdtj$*$L02 \\
\hspace*{.5mm}
-6$*$abdttj$*$L01 \\
\hspace*{.5mm}
-4$*$abdtttj$*$L00 \\[-6mm]
+2$*$ab$*$jL13 \\
+4$*$abj$*$jL03 \\
+6$*$abtj$*$jL02 \\
\hspace*{.5mm}
-6$*$abttj$*$jL01 \\
\hspace*{.5mm}
-4$*$abtttj$*$jL00 :
\vskip -8.2cm \hspace*{5cm} \parbox[t]{8cm}{
Of course $a$ stands for $f^a$ and $b$ for $f^b\,$. Read \\
the factor containing them as an adjoint trace, hence especially
ab$ = (ab)^{cc} = f^{cad}f^{dbc}$ \\
$ = -N \d^{ab}$. But the factors containing L carry
indices$^{ab}$. Hence e.g. jL02 = $(j * L_{02} )^{ab}\,$.
\\[3mm]
dttj is $\6 \,\ov\6^2 j\,$, hence \,abdttj$ = f^{cad}f^{dbe}
\,\6 \,\ov\6^2 j^{ec}$. \\
Using $f^{cad}f^{dbe}f^{euc} = -(N/2)\,f^{abu}$ this turns
into $(N/2)\,\6\, \ov{\ov \jmath}{}^{\,ab}$.
\\[3mm]
By the cyclic invariance of trace a can always be moved to the
left end inside $(\;\;\;)^{cc}$. By inversion of the content of
$(\;\;\;)^{cc}$ (the factors are anti--symmetric) also $b$
can be placed near to a. But to have always ab at the left
end is a speciality of the example $\cl T Q_1\,$ choosen. } \\
\let\dq=\thq \renewcommand{\theequation}{D.\dq}
\setcounter{equation}{0}
\parag{Appendix D : \,\hbox{\small List of the \boldmath$L$'s which
occur in Appendix C}}
How $L$--coefficients look like. Here they include only the first
power of $\epsilon\,$ (to limit the length of this paper). Each term
carries indices$^{ab}$. $\,...j\;\!${something} means $\,...j*$something.
\bea{D1}
L00 &=& +j \,\; +\epsilon*\Big( +3/8*dtj +3/8*tjj -3/8*jtj
\;\Big)
\\[3mm]
L01 &=& +1/2*{1\0\epsilon} +1/4*tj \nonumber \\
& & +\epsilon*\Big(\; +7/16*dttj +7/16*ttjj -7/16*jttj +1/8*tjtj
\;\Big)
\\[3mm]
L02 &=& +\epsilon*\Big(\; +15/32*dtttj +15/32*tttjj -15/32*jtttj \nonumber \\
& & \hspace*{14mm} +17/32*ttjtj -7/32*tjttj
\;\Big)
\\[3mm]
L03 &=& -11/48*tttj \,\; +\epsilon*\Big(\; +31/64*dttttj +31/64*ttttjj \nonumber \\
& & +1*tttjtj +3/16*ttjttj -5/8*tjtttj -31/64*jttttj
\;\Big)
\\[3mm]
L10 &=& +1/2*dj -1/2*jj \,\; +\epsilon*\Big(\; +7/24*ddtj +5/24*dtjj -7/24*djtj
\nonumber \\
& & +7/24*tjdj -7/12*jdtj -1/12*tjjj -5/24*jtjj +7/24*jjtj
\;\Big)
\\[3mm]
L11 &=&
-1/2*{j\0 \epsilon} -1/4*tjj \,\; +\epsilon*\Big(\;
+31/96*ddttj
+5/24*dttjj \nonumber \\ & &
-1/12*dtjtj
-31/96*djttj
+31/96*ttjdj
-5/48*tjdtj \nonumber \\ & &
-31/48*jdttj
-11/96*ttjjj
-5/24*jttjj
+31/96*jjttj \nonumber \\ & &
-11/48*tjtjj
+1/48*tjjtj
+1/12*jtjtj
\;\Big)
\\[3mm]
L12 &=& -1/4*{1\0\epsilon^2} -1/4*{tj\0\epsilon} -11/32*dttj \nonumber \\
& & -11/32*ttjj +11/32*jttj -3/16*tjtj \nonumber \\
+\epsilon*\Big( \hspace*{-8mm}
& &
+1/3*ddtttj
+19/96*dtttjj
+1/16*dttjtj
-7/16*dtjttj
-1/3*djtttj \nonumber \\ & &
+1/3*tttjdj
+1/8*ttjdtj
-5/8*tjdttj
-2/3*jdtttj
-13/96*tttjjj \nonumber \\ & &
-19/96*jtttjj
+1/3*jjtttj
-13/32*ttjtjj
-1/16*ttjjtj
-13/32*tjttjj \hspace*{-12mm} \nonumber \\ & &
-1/16*jttjtj
+3/16*tjjttj
+7/16*jtjttj
-1/8*tjtjtj
\;\Big)
\\[3mm]
L13 &=& -1/16*{ttj\0\epsilon} -61/96*dtttj -13/32*tttjj +61/96*jtttj
-17/32*ttjtj \nonumber \\
& & +7/32*tjttj \nonumber \\
+\epsilon*\Big( \hspace*{-8mm}
& &
+43/128*ddttttj
+3/16*dttttjj
+7/32*dtttjtj
-15/32*dttjttj \nonumber \\ & &
-19/16*dtjtttj
-43/64*djttttj
-43/64*jdttttj
-39/32*tjdtttj \nonumber \\ & &
-45/64*ttjdttj
+13/32*tttjdtj
+37/128*ttttjdj
-19/128*ttttjjj \nonumber \\ & &
-3/16*jttttjj
+43/128*jjttttj
-19/32*tttjtjj
-3/16*tttjjtj \nonumber \\ & &
-19/32*tjtttjj
-7/32*jtttjtj
+13/32*tjjtttj
+13/16*jtjtttj \nonumber \\ & &
-57/64*ttjttjj
+15/64*ttjjttj
+15/32*jttjttj
-9/16*ttjtjtj \nonumber \\ & &
-9/16*tjttjtj
+15/32*tjtjttj
\;\Big)
\\[3mm]
L23 &=&
+1/4*{1\0 \epsilon^3}
+3/8*{tj\0 \epsilon^2}
+3/8*{dttj\0 \epsilon}
+1/2*{ttjj\0 \epsilon}
-3/8*{jttj\0 \epsilon}
+3/8*{tjtj\0 \epsilon} \nonumber \\ & &
-79/96*ddtttj
-3/8*dtttjj
-1/4*dttjtj
+25/32*dtjttj
+79/96*djtttj \nonumber \\ & &
+79/48*jdtttj
+35/32*tjdttj
-13/32*ttjdtj
-19/32*tttjdj
+7/32*tttjjj \hspace*{-12mm} \nonumber \\ & &
+3/8*jtttjj
-79/96*jjtttj
+21/32*ttjtjj
+5/32*ttjjtj
+1/4*jttjtj \nonumber \\ & &
+21/32*tjttjj
-5/16*tjjttj
25/32*jtjttj
+5/16*tjtjtj \nonumber \\
+\epsilon*\Big(\; \hspace{-8mm} & &
+129/512*dddttttj
+43/512*ddttttjj
-1/128*ddtttjtj
-139/256*ddttjttj \hspace*{-8mm} \nonumber \\ & &
-275/384*ddtjtttj
-129/512*ddjttttj
-387/512*jddttttj
-199/128*tjddtttj \hspace*{-8mm} \nonumber \\ & &
-309/256*ttjddttj
+9/128*tttjddtj
+129/512*ttttjddj
+139/512*dttttjdj \hspace*{-8mm} \nonumber \\ & &
+3/128*dtttjdtj
-399/256*dttjdttj
-261/128*dtjdtttj
-387/512*djdttttj \hspace*{-8mm} \nonumber \\ & &
-53/512*dttttjjj
-53/128*dtttjtjj
-5/128*dtttjjtj
+121/256*dttjjttj \hspace*{-8mm} \nonumber \\ & &
-159/256*dttjttjj
-15/128*dttjtjtj
-53/128*dtjtttjj
+233/384*dtjjtttj \hspace*{-8mm} \nonumber \\ & &
-15/128*dtjttjtj
+121/128*dtjtjttj
-43/512*djttttjj
+1/128*djtttjtj \hspace*{-8mm} \nonumber \\ & &
+139/256*djttjttj
+275/384*djtjtttj
+129/512*djjttttj
-43/256*jdttttjj \hspace*{-8mm} \nonumber \\ & &
+1/64*jdtttjtj
+139/128*jdttjttj
+275/192*jdtjtttj
+129/256*jdjttttj \hspace*{-8mm} \nonumber \\ & &
-43/64*tjdtttjj
-15/64*tjdttjtj
+85/64*tjdtjttj
+161/192*tjdjtttj \hspace*{-8mm} \nonumber \\ & &
-129/128*ttjdttjj
-15/64*ttjdtjtj
+85/128*ttjdjttj
-43/64*tttjdtjj \hspace*{-8mm} \nonumber \\ & &
-5/64*tttjdjtj
-43/256*ttttjdjj
+387/512*jjdttttj
+137/128*tjjdtttj \hspace*{-8mm} \nonumber \\ & &
+261/128*jtjdtttj
+219/256*ttjjdttj
+219/128*tjtjdttj
+399/256*jttjdttj \hspace*{-8mm} \nonumber \\ & &
-15/128*tttjjdtj
-45/128*ttjtjdtj
-45/128*tjttjdtj
-3/128*jtttjdtj \hspace*{-8mm} \nonumber \\ & &
-119/512*ttttjjdj
-119/128*tttjtjdj
-357/256*ttjttjdj
-119/128*tjtttjdj \hspace*{-8mm} \nonumber \\ & &
-139/512*jttttjdj
+33/512*ttttjjjj
+53/512*jttttjjj
+43/512*jjttttjj \hspace*{-8mm} \nonumber \\ & &
-129/512*jjjttttj
+33/128*tttjtjjj
+33/128*tttjjtjj
+5/128*tttjjjtj \hspace*{-8mm} \nonumber \\ & &
+5/128*jtttjjtj
+53/128*jtttjtjj
+33/128*tjtttjjj
+33/128*tjjtttjj \hspace*{-8mm} \nonumber \\ & &
+53/128*jtjtttjj
-1/128*jjtttjtj
-89/384*tjjjtttj
-233/384*jtjjtttj \hspace*{-8mm} \nonumber \\ & &
-275/384*jjtjtttj
+99/256*ttjttjjj
+99/256*ttjjttjj
-49/256*ttjjjttj \hspace*{-8mm} \nonumber \\ & &
+159/256*jttjttjj
-121/256*jttjjttj
-139/256*jjttjttj
+99/128*ttjtjtjj \hspace*{-8mm} \nonumber \\ & &
+15/128*ttjtjjtj
+15/128*ttjjtjtj
+99/128*tjttjtjj
+15/128*tjttjjtj \hspace*{-8mm} \nonumber \\ & &
+15/128*jttjtjtj
+99/128*tjtjttjj
+15/128*tjjttjtj
+15/128*jtjttjtj \hspace*{-8mm} \nonumber \\ & &
-49/128*tjtjjttj
-49/128*tjjtjttj \nonumber \\ & &
-121/128*jtjtjttj
+15/64*tjtjtjtj
\;\Big)
\eea
|
\section{Supplementary Material}
~ ~
\begin{center}
\begin{Large} \textbf{Supplementary Material} \end{Large}
\end{center}
\medskip
\smallskip
\twocolumngrid
\setcounter{section}{0}
\setcounter{subsection}{0}
\setcounter{figure}{0}
\setcounter{equation}{0}
\setcounter{NAT@ctr}{0}
\renewcommand{\figurename}[1]{FIG. }
\renewcommand{\theequation}{S\arabic{equation}}
\makeatletter
\renewcommand{\bibnumfmt}[1]{[S#1]}
\renewcommand{\citenumfont}[1]{S#1}
\renewcommand{\thefigure}{S\@arabic\c@figure}
\makeatother
\renewcommand{\theequation}{S\arabic{equation}}
\renewcommand\thetable{S\arabic{table}}
\renewcommand{\vec}[1]{{\boldsymbol{#1}}}
\section*{Experimental design}
In our experiment, a BEC of about 10\textsuperscript{6} sodium atoms and few 10\textsuperscript{3} to several 10\textsuperscript{4} lithium atoms are both trapped by the same two beam optical dipole trap (ODT) at 1064 nm. In all the experiments the gases are spin-polarized in the absolute hyperfine ground state. The mean trapping frequency for $^{23}$Na is $\bar{\omega}=2\pi\cdot 150 ~\mathrm{Hz}$. The temperature of the sample is $\approx$ 350 nK. We can choose to work with fermionic \textsuperscript{6}Li as well as bosonic \textsuperscript{7}Li at mean trapping frequencies of $\bar{\omega} = 2\pi\cdot 340 ~\mathrm{Hz}$ for \textsuperscript{6}Li and $\bar{\omega}=2\pi\cdot 310 ~\mathrm{Hz}$ for \textsuperscript{7}Li. An additional optical standing wave close to optical transition for lithium ($\lambda_\mathrm{res, Li} \approx 671 ~\mathrm{nm}$) imposes a very strong confinement in one direction for lithium only. The standing wave allows the implementation of motional Ramsey spectroscopy of external impurity energy levels with a high precision \cite{S_Scelle2013}. Due to the depth of the potential the minima can be treated as independent harmonic oscillators, resulting in multiple realizations of the experiment in a single experimental cycle. The geometry corresponds to 2D gases. For fermions the transverse extension is 10 times larger than the longitudinal one. The background induced effect is isolated by alternately measuring the motional energy difference with and without BEC background and is detected as a phase alteration of the Ramsey readout fringes ($\Delta \phi$). The frequency change is given by $\Delta \omega = \Delta \phi / \Delta t$, where $\Delta t$ is the total evolution time, including the time of the state coupling (pulses). We emphasize that the impurity-impurity interaction is negligible, as fermions at low temperatures do not scatter and the intra-species scattering length for the bosons is 7$~a_0$ \cite{S_Salomon2002}, with $~a_0$ being the Bohr radius.
\subsection*{Species-selective optical lattice}
The species-selective optical lattice is close to D-line transitions for lithium (670 to 672 nm while the depth for Li is kept constant). It consists of two intersecting laser beams, leading to a periodicity of $d_\mathrm{lat} = 1.65 ~\mu\mathrm{m}$ and a typical depth of 24.5 lattice recoil (E$_{\mathrm{rec}}$) for \textsuperscript{6}Li and 33\,E$_{\mathrm{rec}}$ for \textsuperscript{7}Li, corresponding to a frequency of 27.2~kHz.
\subsection*{Coupling of motional states}
All presented measurements are performed using a Ramsey pulse scheme in the two lowest external states of the species-selective optical lattice. For coherent coupling of these states the lattice position is periodically modulated. The resulting Rabi frequency is 1.4~kHz. The finite excitation efficiency is caused by residual atoms in the excited state at the beginning of the experiment and by a weak coupling of the second excited state which is suppressed by the anharmonicity of the potential. By coupling the states for a certain time, an equal superposition of ground and excited state is created ($\pi/2$~pulse). A second $\pi/2$ pulse is used for the readout. The second pulse is shifted in time to record the phase of the oscillation. In order to measure the background induced effect on the frequency, the sodium BEC is removed by a resonant light pulse before the Ramsey sequence starts. This light pulse does not cause any observable heating of the Li sample.
\subsection*{Motional state detection}
The population detection of the external states is done by a band mapping technique \cite{S_Greiner2001}. While adiabatically reducing the lattice depth in 2~ms with a time constant $ \tau= 0.8~\mathrm{ms}$ the longitudinal confinement is turned off, allowing the atoms to expand along the lattice direction. After 10~ms of evolution an absorption image is taken. With this technique a high optical density can be obtained even for low atom numbers, corresponding to $\approx 200$ times more background than impurity atoms.
\section*{Data analysis}
\subsection*{Data taking strategy}
Each Ramsey fringe is sampled at 10 points. One scan consists of at least three runs. For each run the scans with and without background are separately fitted with a sine function and the final value for the phase shift is calculated as a weighted mean of the repetitions. Shots with extreme atom numbers, deviating lattice intensity or at wrong wavelength are discarded. If several scans can be combined, their result is calculated as the weighted mean of the individual values.
\subsection*{Background modulation}
In the Thomas-Fermi approximation the modulation of the BEC density is: $n_\mathrm{mod}=(\mu_\mathrm{B}-V_\mathrm{B})/g_\mathrm{BB}$ with $g_{\mathrm{BB}} = 4\pi\hbar^2 a_{\mathrm{BB}}/m_{\mathrm{B}}$ with $a_\mathrm{BB} = 54.54~a_0$ \cite{S_Knoop2011}. This adds to the trap potential for the impurities: $V_\mathrm{Imod} = g_\mathrm{IB} n_\mathrm{mod}$ with $g_{\mathrm{IB}} = 2\pi\hbar^2 a_{\mathrm{IB}}/m_{\mathrm{r}}$, where $a_{\mathrm{IB}} = -75~a_0$ for fermions and $a_{\mathrm{IB}} = 21~a_0$ for bosons \cite{S_Schuster2012}. The resulting impurity trapping potential is up to the constant offset: $V_\mathrm{Ieff} = V_\mathrm{I} - g_\mathrm{IB} V_\mathrm{B}/ g_\mathrm{BB}$. In harmonic approximation $\omega \propto \sqrt{V}$, therefore $\omega_\mathrm{eff} = \omega_0 \sqrt{1 - g_\mathrm{IB}/g_\mathrm{BB}\cdot V_\mathrm{B}/V_\mathrm{I}}$. The relative shift due to the lattice is then:
\begin{equation}
\delta_\mathrm{latt}=\frac{\omega_\mathrm{eff}-\omega_0}{\omega_0} = -\frac{1}{2}\frac{g_{\mathrm{IB}}}{g_{\mathrm{BB}}}\frac{V_{\mathrm{B}}}{V_{\mathrm{I}}}.
\end{equation}
As $g_{\mathrm{IB}}$ is three times larger for fermionic lithium than for the bosonic isotope, this effect is more pronounced in the fermionic case. Furthermore $g_{\mathrm{IB}}$ has opposite sign for fermionic compared to bosonic lithium, inverting the effect when changing the isotope.
This behaviour can be seen in Fig.~2. The difference of the slope for the two bosonic scenarios arises due to a change of the effective background density $\bar{n}$. Specifically, for the systematic studies shown in Fig.~3 we keep the optical potential fixed and subtract the additional frequency shift due to background modulation.
\subsection*{Calculation of density distribution}
The density distributions are obtained numerically by taking into account the finite temperature as well as the external potentials such as the species-selective optical lattice and the harmonic trapping due to the ODT. For lithium it is necessary to include the additional potential due to the interaction with sodium. In order to capture properly the lithium occupation number in the individual sites we slice the ODT density according to the periodicity of the lattice. These occupation numbers enter the detailed calculation of the density distribution in the total confinement potential, where a 2D description is applied.
\subsection*{Experimental signal}
The band mapping technique yields the spatially averaged phase shift $\Delta \phi$. However, different parts of the lithium cloud are embedded in varying sodium densities, which leads to locally varying phase shifts $\phi({\bf r}) = \omega_0 \cdot \delta_\mathrm{self}({\bf r}) \cdot \Delta t$. These local phase shifts can be calculated within the local density approximation from Eq.~(2) in the main text.
In the bosonic case, the observed total phase shift can then be calculated as a lithium density weighted mean:
\begin{equation}
\Delta \phi_B = \mbox{asin}\left(\frac{\int \mathrm d^3{\bf r}~n_\mathrm{Li, BEC}({\bf r}) \sin(\phi({\bf r})) }{\int \mathrm d^3{\bf r}~n_\mathrm{Li}({\bf r})}\right)
\end{equation}
For bosonic impurities only the condensed part is assumed to experience an energy shift. However, the total signal is given as a weighted sum over shifted and unshifted phase patterns. The weak interaction between the BEC and the bosonic impurites leads to long coherence times ($\tau>50~\mathrm{ms}$), which are much longer than the Ramsey sequence, such that decoherence can be neglected in this case.
For fermionic impurities on the other hand, the interaction with the BEC is sufficiently strong that decoherence effects in the center of the cloud affect the observed signal. They can be taken into account via a position dependent relaxatio†n rate $\Gamma({\bf r})$, which was studied in detail in \cite{S_Scelle2013}. The total signal of the fermions is then given by:
\begin{equation}
\Delta \phi_F = \mbox{asin}\left(\frac{\int \mathrm d^3{\bf r}~n_\mathrm{Li}({\bf r}) e^{-\Gamma({\bf r}) t} \sin(\phi({\bf r})) }{\int \mathrm d^3{\bf r}~n_\mathrm{Li}({\bf r})e^{-\Gamma({\bf r}) t}}\right)
\end{equation}
\section*{Theoretical treatment}
We consider two different geometries. In order to obtain the effective mass $m_\mathrm{I}^*$ we work with an unconfined (free) impurity coupled to a structureless BEC as described by Eq.~(1) in the main text. Computation of $m_\mathrm{I}^*$ can be very conveniently performed following the lines of ~\cite{S_Mahan1990BOOK}, but with different matrix elements
\begin{eqnarray} \label{Vk}
V_{\bf q} = \lambda [(\xi q)^2/ ((\xi q)^2 + 2)]^{1/4} \, ,
\end{eqnarray}
where $\lambda = g_{\mathrm{IB}} \, \sqrt{n_{\mathrm{BEC}}}$, see e.\,g.~\cite{S_Tempere2009, S_Grusdt2014, S_Grusdt2014b, S_Shashi2014}.
For weak interactions one obtains $m_\mathrm{I}^*/m_\mathrm{I} = 1 + \nu \alpha + \dots$, where $\nu\approx 0.364$ and $0.336$ for ${}^{6}$Li and ${}^{7}$Li setups, respectively, and where $\alpha = a_\text{IB}^2/(\xi a_\mathrm{BB})$ is the dimensionless interaction strength.
In the second geometry we consider the impurities as being confined in a parabolic potential (energy parameter $\hbar \omega_0$ and length parameter $a=\sqrt{\hbar/m_\mathrm{I} \omega_0}$) in $x$-direction, and being free in all other spatial dimensions. Their eigenstates then have energies $E_{n, k} = \hbar \omega_0 (n+1/2) + \hbar^2 k^2/2m_\mathrm{I} - \mu$, where ${\bf k}$ is a 2D wave vector and $n$ denotes the respective subband of the confinement potential. The unperturbed fermion/boson Hamiltonian is then
\begin{equation}
H_\mathrm{I} = \sum_{n} \int \frac{\mathrm d^2 {\bf k}}{(2 \pi)^2} E_{n, k} \, \hat{a}^\dag_{n, {\bf k}} \hat{a}_{n, {\bf k}}^{\vphantom{\dagger}} \, ,
\end{equation}
where $\hat{a}^\dag_{n, {\bf k}}$ stands for the impurity creation operator. As $a_{\mathrm{IB}} \neq 0$ the impurities are scattered on the harmonic modes of the BEC -- the phonons. The corresponding interaction terms have been derived in the weak depletion limit in \cite{S_Tempere2009}. Adapting it to the present setup leads to the following interaction term:
\begin{eqnarray}
H_{\mathrm{int}} &=& \int \frac{\mathrm d^2 {\bf k}}{(2 \pi)^2} \int \frac{\mathrm d^3 {\bf q}}{(2 \pi)^3} \sum_{n_1, n_2} V_{\bf q} \, A(n_1,n_2,q_x) \,
\nonumber \\
&&\times
\hat{a}^\dag_{n_1, \bf k+q'} \hat{a}_{n_2, \bf k}^{\vphantom{\dagger}} (\hat{b}_{\bf q}^{\vphantom{\dagger}} + \hat{b}^\dag_{- \bf q}) \,.
\end{eqnarray}
Here $\hat{b}_{\bf q}$ is the annihilation operator for the phonon, ${\bf q'}= (q_y,q_z)$ denotes the transverse component. $A$ is the matrix element for the transition between the harmonic oscillator energy levels:
\begin{equation}
A(n_1,n_2,q_x) = \int \mathrm d x \, \varphi^*_{n_1}(x) \, \varphi_{n_2}^{\vphantom{\dagger}}(x) \, e^{- i q_x x} \, ,
\end{equation}
$\varphi_n(x)$ are the wave functions of the $n$-th eigenstate. Finally, the Hamiltonian of the phonons is
\begin{equation}
H_{\mathrm{ph}} = \int \frac{\mathrm d^3 {\bf q}}{(2 \pi)^3} \, \omega_q \, \hat{b}_{\bf q}^\dag \hat{b}_{\bf q}^{\vphantom{\dagger}} \, ,
\end{equation}
where $\omega_q = c q [1 + (\xi q)^2/2]^{1/2}$ with the sound velocity $c$. Matrix elements for the impurity-phonon scattering are given by Eq.~(\ref{Vk}).
In the present case the relative weakness of $a_{\mathrm{IB}}$ justifies a perturbative treatment of the problem. Our goal is the energy level renormalization for the impurities in the confinement potential. Its computation is best accomplished via lowest order irreducible self-energy correction.
The energy level structure is given by the retarded self-energy, which is computed via analytical continuation of its Matsubara counterpart, see e.\,g.~ \cite{S_Mahan1990BOOK},
\begin{widetext}
\begin{eqnarray} \nonumber \nonumber
\Sigma(n, {\bf k}; i \Omega) = \frac{i \lambda^2}{\beta} \int \frac{\mathrm{d}^3 \bf q}{(2 \pi)^3} \sum_{i \epsilon_j} \sqrt{\frac{(\xi q)^2}{ (\xi q)^2 + 2}} \, \sum_{m=0}^\infty A^*(n,m,q_x) A(m,n,q_x) \,\\
\times G[m, {\bf k-q'}, i \Omega - (m-n) \omega_0 - i \epsilon_j] \, D_0(i\epsilon_j, {\bf q}) \, ,
\end{eqnarray}
where $\Omega$ is a fermionic/bosonic (impurity) and $\epsilon_j = 2 \pi j/\beta$ bosonic (phonons) Matsubara frequencies. $\beta=1/T$ is the inverse temperature. The necessary Green's functions are:
\begin{eqnarray}
D_0(i\epsilon_j, {\bf q}) &=& \frac{2 \omega_{\bf q}}{(i \epsilon_j)^2 - \omega_{\bf q}^2} \\
G[m, {\bf k-q'}, i \Omega - (m-n) \omega_0 - i \epsilon_j] &=& \frac{1}{i \Omega - (m-n) \omega_0 - i \epsilon_j - E_{m,{\bf k - q'}}} \, ,
\end{eqnarray}
\end{widetext}
where the impurity Green's function $G$ has the same shape for both bosonic and fermionic case.
Performing the calculation in the assumption that only the two lowest lying energy levels are populated we can extract the energy difference between them. The result for the bosonic impurities is:
\begin{eqnarray}
\delta_\mathrm{self} &=& \frac{\Delta \omega}{\omega_0} = \alpha_\mathrm{IB} \cdot f(\rho_\mathrm{2D}, a, \xi)\,,\\
f(\rho_\mathrm{2D},a, \xi) &=& \rho_\mathrm{2D} \xi^2 \nonumber \\
&&\times \left[ g_\mathrm{e}(a/\xi) \eta -g_\mathrm{g}(a/\xi) (1-\eta) \right]\,,\\
g_\mathrm{g}(a/\xi) &=& \frac{8 \pi^2}{\sqrt{2}} e^{(a/\xi)^2} \left[1-\erf(a/\xi) \right]\,,\\
g_\mathrm{e}(a/\xi) &=& 4 \pi \frac{a}{\xi} \sqrt{2 \pi} \nonumber \\
&&\times \left [1-\sqrt{\pi} \frac{a}{\xi} e^{(a/\xi)^2} (1-\erf(a/\xi)) \right] \,.
\end{eqnarray}
The coupling strength $\alpha_\mathrm{IB}$ is given by:
\begin{equation}
\alpha_\mathrm{IB} = \frac{\lambda^2 m_\mathrm{B}}{8 \pi^2 \hbar^3 \omega_0 \xi}\, ,
\end{equation}
where $\hbar$ is the reduced Planck constant.
Self-energy evaluation for the case of fermionic impurities is slightly different. Here the numerical prefactors weakly depend on the effective temperature and the ratio $a/\xi$. Assuming $T=0$ and a symmetric superposition we obtain for the energy shift the value:
\begin{equation}
f(a, \xi) = -0.11+\frac{0.94}{a/\xi}\, .
\end{equation}
As the shaded inset in Fig.~2 left panel shows, the expression above is in very good agreement with our measurements.
|
\section{Introduction}
To achieve reliable communication at rates close to the channel capacity, it is well-known that the blocklength must tend to infinity. A more refined question is, ``How fast can the gap to capacity decrease as a function of the blocklength?''. A key result is that, for any rate-$R$ code achieving a block error rate of $\eta<1$ on a non-trivial discrete memoryless channel with capacity $C$, the blocklength $N$ must satisfy $C-R\geq A/\sqrt{N}$ for some $A>0$ that depends only on $\delta$ and the channel~\cite{Strassen-zfw64,Laneman-ita06,Hayashi-it09,Polyanskiy-it10}. Thus, the gap to capacity cannot vanish faster than $O(N^{-1/2})$. Random codes are known to achieve this scaling.
Polar codes are the first codes, with low-complexity encoding and decoding algorithms, that were proven to achieve capacity on binary-input memoryless channels~\cite{Arikan-isit08,Arikan-it09}. Since then, the rate of polarization and the relationship between the blocklength and the error rate has received significant attention~\cite{Arikan-isit09,Hassani-it13,Goldin-it14,Hassani-it14,Fazeli-aller14,Guruswami-arxiv14,Presman-it15,Guruswami-it15,Goldin-arxiv15,Mondelli-arxiv15}. This relationship is typically studied in two distinct regimes by asking two different questions. First, for a fixed rate $R<C$, how fast does the error rate decay with the blocklength? Second, for a fixed probability of decoding failure $\eta\in(0,1)$, how fast can the rate approach the capacity?
The majority of prior work in this area focuses on binary polar codes with $2\times2$ kernels and, for these codes, the gap to capacity cannot decrease faster than $O(N^{-0.276})$~\cite{Hassani-it14,Mondelli-arxiv15}. For $2\times2$ kernels with larger alphabets, the analysis is much more difficult and the provable scaling rates are even smaller~\cite{Guruswami-arxiv14,Goldin-arxiv15}. Recently, $8\times8$ and $16\times16$ binary kernels have been constructed that achieve scaling rates of $O(N^{-0.279})$ and $O(N^{-0.298})$~\cite{Fazeli-aller14}. Until now, no reported results provably established scaling rates faster than $O(N^{-0.30})$.
In the first part of this work, we consider the $q$-ary polar codes introduced by Mori and Tanaka based on $q\times q$ Reed-Solomon (RS) polarization kernels with elements from the Galois field $\mathbb{F}_{q}$~\cite{Mori-itw10,Mori-it14}. Thus, in all statements, $q$ is implicitly assumed to be a prime power. These codes have length $N=q^{n}$, where $n$ is the number of steps in the polarization process. We consider transmission over the $q$-ary erasure channel (QEC) with erasure probability $\epsilon$. Mori and Tanaka have also shown that these polar codes achieve capacity on symmetric $q$-ary channels~\cite{Mori-it14}.
By analyzing the polarization process for the QEC, we show that, for any $\gamma>0$ and $\delta>0$, there is a $q_{0}$ such that, for all $q\geq q_{0}$, the fraction of effective channels with erasure rate at most $N^{-\gamma}$ is at least $1-\epsilon-O(N^{-1/2+\delta})$. Thus, the gap to capacity scales at a nearly-optimal rate. While our proof relies on large alphabet QECs with large polarization kernels, we believe a similar result may also hold for small alphabets (e.g., binary) with large polarization kernels.
Like binary polar codes, the performance of $q$-ary polar codes can be analyzed by tracking the evolution of the effective channels through the polarization process~\cite{Arikan-it09}. At each step, a single effective channel with erasure rate $x$ splits into $q$ new channels. For their codes, Mori and Tanaka showed that the $i$-th new effective channel, for $i\in\mathcal{Q}\triangleq\left\{ 0,1,\ldots,q-1\right\} $, is a $q$-ary erasure channel with erasure probability
\begin{equation}
\psi_{i}(x)=\sum_{j=i+1}^{q}\binom{q}{j}x^{j}(1-x)^{q-j},\label{eq:mori_qec_de}
\end{equation}
Applying this formula recursively, one can compute the erasure rates of the $N=q^{n}$ effective channels after $n$ steps. For a polar code with $k$ information symbols, the next step in the design process consists of choosing the $k$ effective channels with the smallest erasure rates. For $q=2$, these steps are identical to the original polar code construction in~\cite{Arikan-it09} and the resulting codes are closely related to binary Reed-Muller codes. For larger $q$, the resulting codes are closely related to $q$-ary Reed-Muller codes~\cite{Kschischang-isit93}.
In Section~\ref{sec:fixed_alphabet}, a similar analysis is also considered for $q$-ary polar codes with $m\times m$ polarizing matrices. This separates the effect of the alphabet size from the effect of the matrix size. If the polarizing matrix at each stage is drawn independently and uniformly from the set of invertible $m\times m$ matrices, then the linear operator associated with the Lyapunov function analysis can be written in closed form. To prove near-optimal scaling for polar codes with fixed $q$ as $m$ increases, however, two technical obstacles remain. Thus, we conclude by stating two concrete mathematical conjectures that, if proven, would imply near-optimal scaling for fixed $q$.
\section{The Polarization Process}
Let the random variable $X_{n}$ denote the channel erasure probability for a randomly chosen effective channel after $n$ levels of polarization. The sequence $X_{n}$, for $n=0,1,\ldots$, is a homogeneous Markov chain on the compact state space $\mathcal{X}=[0,1]$ with transition probability
\begin{align*}
\mathbb{P}\Big( & X_{n}\,=x_{n}\,\Big|\,(X_{0},\ldots,X_{n-1})=(x_{0},\ldots,x_{n-1})\Big)\\
& =\mathbb{P}\left(X_{n}=x_{n}\midb |X_{n-1}=x_{n-1}\right)\\
& \quad=\frac{1}{q}\left|\left\{ i\in\mathcal{Q}\,|\,x_{n}=\psi_{i}(x_{n-1})\right\} \right|.
\end{align*}
We note that $0$ and $1$ are both absorbing states of this Markov chain and we are interested in the convergence rate to these states~\cite{Arikan-it09}.
Let $C(\mathcal{X})$ denote the set of bounded continuous functions mapping $\mathcal{X}$ to $\mathbb{R}$. One can analyze this Markov chain by focusing on the sequence of functions, $g_{n}(x)\triangleq\mathbb{E}[g_{0}(X_{n})|X_{0}=x]$, generated by $g_{0}\in C(\mathcal{X})$~\cite{Hassani-it14,Mondelli-arxiv15}. Since the Markov chain is homogeneous, this sequence satisfies the recursion
\begin{align*}
g_{n} & (x)\triangleq\mathbb{E}\left[g_{0}(X_{n})\midb |X_{0}=x\right]\\
& =\sum_{i=0}^{q-1}\mathbb{E}\left[g(X_{n})\midb |X_{1}\!=\!\psi_{i}(x)\right]\mathbb{P}\left(X_{1}\!=\!\psi_{i}(x)\midb |X_{0}\!=\!x\right)\\
& =\sum_{i=0}^{q-1}g_{n-1}\left(\psi_{i}(x)\right)\frac{1}{q}.
\end{align*}
The one-step update is given by the linear operator $T_{q}:C(\mathcal{X})\to C(\mathcal{X})$, which is defined by
\begin{equation}
(T_{q}g_{n-1})(x)\triangleq\frac{1}{q}\sum_{i=0}^{q-1}g_{n-1}\left(\psi_{i}(x)\right).\label{eq:mori_qec}
\end{equation}
Since the polarization process preserves the average mutual information, it also preserves average erasure rate. This implies that the function $g_{0}(x)=x$ should be an eigenfunction of $T_{q}$ (with eigenvalue $1$) and, using \eqref{eq:mori_qec}, one can verify that it is. We note that this is a straightforward generalization of the approach used for binary polar codes~\cite{Hassani-it14,Mondelli-arxiv15}.
The rate of polarization is determined by the fraction of channels whose erasure rates are not extremal. The following lemma connects the fraction of non-extremal channels (as a function of $n$) with an easily computable constant associated with $T_{q}$. This can be seen as a standard convergence analysis based on Lyapunov functions and it was first applied to polar codes in~\cite{Hassani-it14}.
\begin{lem}
\label{lem:lyapunov} Suppose there exists a non-negative continuous function $V:\mathcal{X}\to\mathbb{R}_{\geq0}$ and a constant $\lambda\in(0,1)$ such that
\begin{equation}
(T_{q}V)(x)\leq\lambda V(x)\label{eq:exp_drift}
\end{equation}
for all $x\in\mathcal{X}$. Then, for $S(\alpha)\triangleq\left\{ x\in\mathcal{X}\,|\,V(x)\geq\alpha\right\} ,$ it follows that
\[
\mathbb{P}\left(X_{n}\in S(\alpha)\midb |X_{0}=x\right)\leq\frac{\lambda^{n}V(x)}{\alpha}.
\]
Further, if $S(\alpha)$ is a closed interval, then
\[
\mathbb{P}\left(X_{n}\geq\min S(\alpha)\midb |X_{0}=x\right)\leq\frac{\lambda^{n}V(x)}{\alpha}+\frac{x}{\max S(\alpha)}.
\]
\end{lem}
\begin{IEEEproof}
To see this, we choose $g_{0}(x)=V(x)$ and observe that~\eqref{eq:exp_drift} implies $\mathbb{E}\left[V(X_{n})\,|\,X_{0}=x\right]=g_{n}(x)\leq\lambda^{n}V(x)$ for all $x\in\mathcal{X}$. From this, we get
\begin{align*}
\mathbb{P}\left(X_{n}\in S(\alpha)\midb |X_{0}=x\right) & =\mathbb{P}\left(V(X_{n})\geq\alpha\midb |X_{0}=x\right)\\
& \leq\frac{\mathbb{E}\left[V(X_{n})\midb |X_{0}=x\right]}{\alpha}\\
& \leq\frac{\lambda^{n}V(x)}{\alpha}.
\end{align*}
Since the polarization process preserves the average mutual information, we have $\mathbb{E}[X_{n}\,|\,X_{0}=x]=x$. For the second part, we combine this with the Markov inequality to see that
\begin{align*}
\mathbb{P}\left(X_{n}>\max S(\alpha)\midb |X_{0}=x\right) & \leq\frac{\mathbb{E}\left[X_{n}\midb |X_{0}=x\right]}{\max S(\alpha)}\\
& =\frac{x}{\max S(\alpha)}.
\end{align*}
Since $S(\alpha)$ is a closed interval, it follows that
\begin{align*}
\mathbb{P} & \left(X_{n}\geq\min S(\alpha)\midb |X_{0}\!=\!x\right)=\mathbb{P}\left(X_{n}\!\in\!S(\alpha)\midb |X_{0}\!=\!x\right)\\
& \quad+\mathbb{P}\left(X_{n}>\max S(\alpha)\midb |X_{0}=x\right)\\
& \quad\quad\leq\frac{\lambda^{n}V(x)}{\alpha}+\frac{x}{\max S(\alpha)}.
\end{align*}
This completes the proof.
\end{IEEEproof}
\begin{rem}
For the considered problem, this lemma is a slight variation of what is used in~\cite{Hassani-it13,Mondelli-arxiv15}. We use this form to show the close connection to Lyapunov functions. From that perspective, the function $V(x)$ can be seen as a Lyapunov function showing convergence to stationary distributions supported on the set $\left\{ x\in\mathcal{X}\,|\,V(x)=0\right\} $~\cite{Hairer-ssa11}.
\end{rem}
\begin{defn}
\label{def:lambda_q_b} Let $V(x)=(x(1-x))^{\beta}$ for $\beta>0$ and define
\[
\lambda_{q,\beta}\triangleq\sup_{x\in(0,1)}\frac{(T_{q}V)(x)}{V(x)}.
\]
Then, $\lambda_{q,\beta}$ is the largest $\lambda\in\mathbb{R}$ such that $(T_{q}V)(x)\leq\lambda V(x)$ for all $x\in(0,1)$. We also note that $V(x)\leq V(\frac{1}{2})=(\frac{1}{4})^{\beta}$ for $x\in[0,1].$
\end{defn}
\begin{lem}
\label{lem:lambda_q_b_bound} The quantity $\lambda_{q,\beta}$ for $\beta\in(0,\frac{1}{2}]$ satisfies
\[
\lambda_{q,\beta}\leq\frac{6}{\sqrt{q\beta}}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}.
\]
\end{lem}
\begin{IEEEproof}
See Section~\ref{subsec:proof_qary_main}.
\end{IEEEproof}
\begin{cor}
\label{cor:lyapunov} If the conditions of Lemma~\ref{lem:lyapunov} hold for $V(x)=(x(1-x))^{\beta}$ with $\beta>0$, then
\[
\mathbb{P}\left(X_{n}\in[\eta,1-\eta]\midb |X_{0}=x\right)\leq\frac{\lambda^{n}V(x)}{V(\eta)}
\]
for $\eta\in(0,\frac{1}{2})$. This also implies
\[
\mathbb{P}\left(X_{n}\geq\eta\midb |X_{0}=x\right)\leq\frac{\lambda^{n}V(x)}{V(\eta)}+\frac{x}{1-\eta}.
\]
\end{cor}
\begin{IEEEproof}
The first statement follows from applying Lemma~\ref{lem:lyapunov} with $\alpha=V(\eta)$. For the second statement, we observe that $S(\alpha)$ is a closed interval because $V(x)$ is a concave function. Also, $\max S(\alpha)=1-\eta$ because $V(\eta)=V(1-\eta)$ implies that $1-\eta\in S(\alpha)$ and $V(x)<V(\eta)$ for $x>1-\eta$. This completes the proof.
\end{IEEEproof}
The primary purpose of this paper is the statement and proof of the following theorem.
\begin{thm}
\label{thm:main_qary} For the $q$-ary polar codes defined in \cite{Mori-itw10,Mori-it14}, let $X_{n}$ be the erasure rate of a randomly chosen effective channel after $n$ steps of polarization. For any $\gamma>0$, $\beta\in(0,\frac{1}{2}]$, and $N^{-\gamma}\leq\frac{3}{4}$, one finds that
\begin{align*}
\mathbb{P}\Big(X_{n}\in[N^{-\gamma} & ,1-N^{-\gamma}]\,\Big|\,X_{0}=x\Big)\\
& \leq N^{\gamma\beta-\frac{1}{2}+\frac{\ln6-\frac{1}{2}\ln\beta+\left(\beta-\frac{1}{2}\right)\ln4}{\ln q}}
\end{align*}
and
\begin{align*}
\mathbb{P}\Big(X_{n}\geq & N^{-\gamma}\,\Big|\,X_{0}=x\Big)\\
& \leq N^{\gamma\beta-\frac{1}{2}+\frac{\ln6-\frac{1}{2}\ln\beta+\left(\beta-\frac{1}{2}\right)\ln4}{\ln q}}+\frac{x}{1-N^{-\gamma}}.
\end{align*}
\end{thm}
\begin{IEEEproof}
Combining Lemma~\ref{lem:lambda_q_b_bound} and Corollary~\ref{cor:lyapunov} with $\eta=N^{-\gamma}$, one gets the prediction
\begin{align*}
\mathbb{P}\Big(X_{n} & \in[N^{-\gamma},1-N^{-\gamma}]\,\Big|\,X_{0}=x\Big)\\
& \leq\frac{\left(x(1-x)\right)^{\beta}}{\left(N^{-\gamma}(1-N^{-\gamma})\right)^{\beta}}\left(6\sqrt{\frac{1}{q\beta}}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}\right)^{n}\\
& \leq\left(\frac{\frac{1}{4}}{1-N^{-\gamma}}\right)^{\beta}N^{\gamma\beta}q{}^{n\frac{\ln6-\frac{1}{2}\ln q-\frac{1}{2}\ln\beta+\left(\beta-\frac{1}{2}\right)\ln4}{\ln q}}\\
& \leq N^{\gamma\beta-\frac{1}{2}+\frac{\ln6-\frac{1}{2}\ln\beta+\left(\beta-\frac{1}{2}\right)\ln4}{\ln q}},
\end{align*}
for $N^{-\gamma}\leq\frac{3}{4}$. The second statement follows directly from the second part of Corollary~\ref{cor:lyapunov}.
\end{IEEEproof}
\begin{cor}
Consider the $q$-ary polar codes defined in \cite{Mori-itw10,Mori-it14} on a QEC with erasure probability $\epsilon$. For any $\gamma>0$ and $\delta>0$, there is a $\beta\in(0,\frac{1}{2}]$ and a $q_{0}$ such that, for all $q\geq q_{0}$, the fraction of effective channels with erasure rate at most $N^{-\gamma}$ is at least $1-\epsilon-O(N^{-1/2+\delta})$.
\end{cor}
\begin{IEEEproof}
Since the stated condition becomes weaker as $\gamma$ decreases and $\delta$ increases, we assume without loss of generality that $\gamma\geq\frac{1}{2}$ and $\delta\leq\frac{1}{2}$. Using this, we choose $\beta=\frac{\delta}{2\gamma}$ and observe that $\beta\in(0,\frac{1}{2}]$. At error rate $N^{-\gamma}$, the gap to capacity is given by
\begin{align*}
(1-\epsilon) & -\mathbb{P}\left(X_{n}<N^{-\gamma}\,|\,X_{0}=\epsilon\right)\\
& =\mathbb{P}\left(X_{n}\geq N^{-\gamma}\,|\,X_{0}=\epsilon\right)-\epsilon.
\end{align*}
Since $2^{-1/2}\leq\frac{3}{4}$, we can applying Theorem~\ref{thm:main_qary} for $N\geq2$ to see that
\begin{align*}
\mathbb{P} & \left(X_{n}\geq N^{-\gamma}\,\Big|\,X_{0}=\epsilon\right)-\epsilon\\
& \stackrel{(a)}{\leq}N^{\frac{\delta}{2}-\frac{1}{2}+\frac{\ln6-\frac{1}{2}\ln\beta+\left(\beta-\frac{1}{2}\right)\ln4}{\ln q}}+\frac{\epsilon N^{-\gamma}}{1-N^{-\gamma}}\\
& \quad\quad\stackrel{(b)}{\leq}N^{\frac{\delta}{2}-\frac{1}{2}+\frac{\delta}{2}}+4\epsilon N^{-\frac{1}{2}}\\
& \quad\quad\quad\quad\leq5N^{-\frac{1}{2}+\delta},
\end{align*}
where $(a)$ follows from $\frac{\epsilon}{1-N^{-\gamma}}-\epsilon=\frac{\epsilon N^{-\gamma}}{1-N^{-\gamma}}$ and $(b)$ follows from $N^{-\gamma}\leq2^{-1/2}\leq\frac{3}{4}$ and choosing $q\geq q_{0}$ with $\ln q_{0}\triangleq\frac{1}{\delta}(2\beta\ln4-\ln\beta+2\ln6-\ln4)$. This completes the proof.
\end{IEEEproof}
\subsection{Numerical Examples}
In this section, we present some applications of Corollary~\ref{cor:lyapunov} based on numerical computation of $\lambda$.
\begin{figure}[!t]
\begin{center}
\scalebox{1.2}{\includegraphics{qec2x}}
\end{center}\vspace{-4mm}
\caption{\label{fig:lambda_2}Numerical evaluation of the constant $\lambda_{2,0.66}$. }
\end{figure}
\begin{example}
Consider the case of $q=2$ where $T_{q}$ is defined by
\[
(T_{2}g)(x)=\frac{g(x^{2})+g(2x-x^{2})}{2}.
\]
Using $V(x)=\left(x(1-x)\right)^{0.66}$, one can verify numerically that $(T_{2}V)(x)\leq0.832V(x)$ for $x\in[0,1]$. For example, see Figure~\ref{fig:lambda_2}. We note that this calculation was described first in~\cite{Hassani-it14}. Therefore,
\begin{align*}
\mathbb{P}\Big(X_{n}\in[0.01,0.99] & \Big|X_{0}=x\Big)\leq\frac{(1/4)^{0.66}}{(0.0099)^{0.66}}0.832^{n}\\
& =9\cdot2^{n\frac{\ln0.832}{\ln2}}\\
& \leq9N^{-0.265}.
\end{align*}
Let $V_{5}(x)$ be the result of applying $T_{2}$ five times to the function $\left(x(1-x)\right)^{0.66}$ (i.e., $V_{5}(x)=\left(T_{2}^{5}\left(x(1-x)\right)^{0.66}\right)(x)$). Then, one can verify numerically that $(T_{2}V_{5})(x)\leq0.8271V_{5}(x)$ and this gives a decay rate of $O(N^{-0.273})$.
\end{example}
\begin{example}
Consider the case of $q=4$ where $T_{q}$ is defined by
\begin{align*}
(T_{4}g)(x) & =\frac{1}{4}\bigg(g(x^{4})+g(4x^{3}(1-x)+x^{4})+\\
g(1- & 4x(1-x)^{3}-(1-x)^{4})+g(1-(1-x)^{4})\bigg)
\end{align*}
Using $V(x)=\left(x(1-x)\right)^{0.64}$, one can verify numerically that $(T_{4}V)(x)\leq0.657V(x)$ for $x\in[0,1]$. Therefore,
\begin{align*}
\mathbb{P}\Big(X_{n}\in[0.01,0.99] & \Big|X_{0}=x\Big)\leq\frac{(1/4)^{0.64}}{(0.0099)^{0.64}}0.657^{n}\\
& =8\cdot4^{n\frac{\ln0.657}{\ln4}}\\
& \leq8N^{-0.303}.
\end{align*}
\end{example}
\begin{example}
Consider the case of $q=16$ where $T_{q}$ is defined by \eqref{eq:mori_qec}. Using $V(x)=\left(x(1-x)\right)^{0.58}$, one can verify numerically that $(T_{16}V)(x)\leq0.375V(x)$ for $x\in[0,1]$. For example, see Figure~\ref{fig:lambda_16}. Therefore,
\begin{align*}
\mathbb{P}\Big(X_{n}\in[0.01,0.99] & \Big|X_{0}=x\Big)\leq\frac{(1/4)^{0.58}}{(0.0099)^{0.58}}0.375^{n}\\
& =7\cdot16^{n\frac{\ln0.375}{\ln16}}\\
& \leq7N^{-0.353}.
\end{align*}
\end{example}
\begin{figure}[!t]
\vspace{2mm}
\begin{center}
\scalebox{1.2}{\includegraphics{qec16x}}
\end{center}\vspace{-4mm}
\caption{\label{fig:lambda_16}Numerical evaluation of the constant $\lambda_{16,0.58}$.}
\end{figure}
\begin{example}
\label{exa:sqrt_numerical_largeq} Consider the case where $T_{q}$ is defined by \eqref{eq:mori_qec} for $q=2,3,\ldots,1024$. Using $V(x)=\sqrt{x(1-x)}$, one can compute numerically the smallest $\lambda_{q}$ such that $(T_{q}V)(x)\leq\lambda_{q}V(x)$ for $x\in[0,1]$. This computation results in $\lambda_{q}=(T_{q}V)(\frac{1}{2})/V(\frac{1}{2})$ and one observes that $\sqrt{q}\lambda_{q}$ is increasing in $q$ and upper bounded by $1.6142$. Assuming this is true, we observe that
\begin{align*}
\mathbb{P}\Big(X_{n}\!\in\![\eta,1-\eta]\Big| & X_{0}=x\Big)\!\leq\!\frac{\sqrt{1/4}}{\sqrt{\eta(1-\eta)}}\left(\frac{1.6142}{\sqrt{q}}\right)^{\!n}\\
& =\frac{1}{\sqrt{4\eta(1-\eta)}}q{}^{n\frac{\ln1.6142-\frac{1}{2}\ln q}{\ln q}}\\
& \leq\frac{1}{\sqrt{4\eta(1-\eta)}}N^{-\frac{1}{2}(1-\frac{1}{\ln q})}.
\end{align*}
\end{example}
\begin{example}
Let $V(x)=(x(1-x))^{1/12}$ and consider the case where $T_{q}$ is defined by \eqref{eq:mori_qec} for $q=2,3,\ldots,1024$. Again, one can compute numerically the smallest $\lambda_{q}=\lambda_{q,1/12}$ such that $(T_{q}V)(x)\leq\lambda_{q}V(x)$ for $x\in[0,1]$. This computation results in $\lambda_{q}=(T_{q}V)(\frac{1}{2})/V(\frac{1}{2})$ and one observes that $\sqrt{q}\lambda_{q}$ is increasing in $q$ and upper bounded by $4.1218$. Assuming this is true, we observe that, for $N\geq2$, we have
\begin{align*}
\mathbb{P}\Big(X_{n}\in & [N^{-2},1-N^{-2}]\Big|X_{0}=x\Big)\\
& \leq\frac{(1/4)^{1/12}}{\left(N^{-2}(1-N^{-2})\right)^{1/12}}\left(\frac{4.1218}{\sqrt{q}}\right)^{n}\\
& =\left(\frac{1/4}{1-N^{-2}}\right)^{1/12}N^{\frac{1}{6}}q{}^{n\frac{\ln4.1218-\frac{1}{2}\ln q}{\ln q}}\\
& \leq\frac{1}{3}N^{-\frac{1}{3}(1-\frac{4.5}{\ln q})}.
\end{align*}
\end{example}
\section{Large-Alphabet Erasure Channels}
\subsection{Intuitive Approach}
Before delving into the proof of Theorem~\ref{thm:main_qary}, we present an intuitive (but non-rigorous) argument that leads us in the right direction. Consider $q$ random trials with success probability $x$ and let the random variable $\mbox{Bin}(q,x)$ denote number of successes. Then, one finds that
\[
\mathbb{P}\left(\mathrm{Bin}(q,x)=i\right)=\binom{q}{i}x^{i}(1-x)^{n-i}.
\]
The key is to replace the binomial random variable, $\mathrm{Bin}(q,x)$, by a Gaussian random variable with the same mean and variance. While this step is motivated by the central limit theorem, it is not rigorous (even as $q\to\infty$) because the approximation does not hold uniformly for all $x\in[0,1]$. Based on this assumption, we approximate $\mathbb{P}\left(\mathrm{Bin}(q,x)\geq i+1\right)$ by
\[
\psi_{i}(x)\approx Q\left(\frac{i+1-qx}{\sqrt{qx(1-x)}}\right),
\]
where $Q(x)\triangleq(2\pi)^{-1/2}\int_{x}^{\infty}e^{-t^{2}/2}dt$. Let $V(x)=(x(1-x))^{\beta}$ for $\beta\in(0,\frac{1}{2}]$. Using a sequence of approximations, one finds that
\begin{align*}
(T_{q}V)(x) & =\frac{1}{q}\sum_{i=0}^{q-1}V\left(\psi_{i}(x)\right)\\
& \approx\frac{1}{q}\sum_{i=0}^{q-1}V\left(Q\left(\frac{i+1-qx}{\sqrt{qx(1-x)}}\right)\right)\\
& \approx\int_{0}^{1}V\left(Q\left(\frac{q(y-x)}{\sqrt{qx(1-x)}}\right)\right)\mathrm{d} y\\
& =\sqrt{\frac{x(1-x)}{q}}\int_{-x\sqrt{qx/(1-x)}}^{\sqrt{q(1-x)/x}}V\left(Q\left(z\right)\right)\mathrm{d} z\\
& \approx\sqrt{\frac{x(1-x)}{q}}\int_{-\infty}^{\infty}V\left(Q\left(z\right)\right)\mathrm{d} z\\
& =\sqrt{\frac{x(1-x)}{q}}\mathfrak{m}(\beta),
\end{align*}
where $\mathfrak{m}(\beta)\triangleq\int_{-\infty}^{\infty}\left(Q(z)Q(-z)\right)^{\beta}dz$.
For $\beta=\frac{1}{2}$, this implies that $V(x)=\sqrt{x(1-x)}$ is an approximate eigenfunction of $T_{q}$ associated with eigenvalue
\[
\tilde{\lambda}_{q}=\frac{\mathfrak{m}(\frac{1}{2})}{\sqrt{q}},
\]
where $\mathfrak{m}(\frac{1}{2})\approx1.6147<e^{1/2}$. Based on this estimate, one could estimate that rate of polarization scales like
\begin{align*}
\tilde{\lambda}_{q}^{n} & \approx\left(\frac{\mathfrak{m}(\frac{1}{2})}{\sqrt{q}}\right)^{n}\leq\left(\frac{e}{q}\right)^{n/2}\\
& =q^{(n/2)(1-\ln q)/\ln q}=N^{-\frac{1}{2}(1-\frac{1}{\ln q})}.
\end{align*}
In fact, the numerical results in Example~\ref{exa:sqrt_numerical_largeq} support this conclusion and suggest that the true $\lambda_{q,1/2}$ satisfies $\lambda_{q,1/2}\sqrt{q}\nearrow\mathfrak{m}(\frac{1}{2})$. Thus, we believe that this non-rigorous analysis produces an exact and tight characterization as $q\to\infty$.
For $\beta\in(0,\frac{1}{2})$, $V(x)$ is not an approximate eigenfunction but one can still estimate the decay rate
\begin{align*}
\tilde{\lambda}_{q,\beta} & =\max_{x\in[0,1]}\left(x(1-x)\right)^{\frac{1}{2}-\beta}\frac{\mathfrak{m}(\beta)}{\sqrt{q}}\\
& =\frac{\mathfrak{m}(\beta)}{\sqrt{q}}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}.
\end{align*}
Combining this estimate with Corollary~\ref{cor:lyapunov} gives the non-rigorous prediction
\begin{align*}
\mathbb{P}\Big(X_{n}\in & [N^{-\gamma},1-N^{-\gamma}]\,\Big|\,X_{0}=x\Big)\\
& \leq\frac{\left(x(1-x)\right)^{\beta}}{\left(N^{-\gamma}(1-N^{-\gamma})\right)^{\beta}}\left(\frac{\mathfrak{m}(\beta)}{\sqrt{q}}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}\right)^{n}\\
& \leq\left(\frac{\frac{1}{4}}{1-N^{-\gamma}}\right)^{\beta}N^{\gamma\beta}q{}^{n\frac{\ln\left(\mathfrak{m}(\beta)\right)-\frac{1}{2}\ln q+\left(\beta-\frac{1}{2}\right)\ln4}{\ln q}}\\
& \leq N^{\gamma\beta-\frac{1}{2}+\frac{\ln\left(\mathfrak{m}(\beta)\right)+\left(\beta-\frac{1}{2}\right)\ln4}{\ln q}},
\end{align*}
for $N^{-\gamma}\leq\frac{3}{4}$. The key point here is that, for any $\gamma>0$ and $\delta>0$, there is a $\beta\in(0,\frac{1}{2}]$ and a large enough $q$ such that this decay rate is $O(N^{-1/2+\delta})$.
\subsection{Rigorous Approach}
Unfortunately, the intuitive argument does not lead directly to a rigorous statement because the central limit theorem is tight only for small deviations. To make things precise, one must instead use the Chernoff-Hoeffding bound for the binomial tail probability.
We start by establishing some basic properties of the functions under consideration. Looking at~\eqref{eq:mori_qec_de}, one observes that
\begin{align*}
\psi_{i}(x) & =\mathbb{P}\left(\mathrm{Bin}(q,x)\geq i+1\right)\\
& =\mathbb{P}\left(\mathrm{Bin}(q,1-x)\leq q-i-1\right)\\
& =1-\mathbb{P}\left(\mathrm{Bin}(q,1-x)\geq q-i\right)\\
& =1-\psi_{q-i-1}(1-x).
\end{align*}
This implies the following lemma.
\begin{lem}
\label{lem:Tq_g_symmetry} If $g(x)=g(1-x)$, then $(T_{q}g)(x)=(T_{q}g)(1-x)$.
\end{lem}
\begin{IEEEproof}
Working directly, one finds that
\begin{align*}
(T_{q}g)(x) & =\frac{1}{q}\sum_{i=0}^{q-1}g\left(\psi_{i}(x)\right)\\
& =\frac{1}{q}\sum_{i=0}^{q-1}g\left(1-\psi_{i}(x)\right)\\
& =\frac{1}{q}\sum_{i=0}^{q-1}g\left(\psi_{q-i-1}(1-x)\right)\\
& =\frac{1}{q}\sum_{i=0}^{q-1}g\left(\psi_{i}(1-x)\right)\\
& =(T_{q}g)(1-x).
\end{align*}
\end{IEEEproof}
The well-known Chernoff bound for the binomial tail probability implies that, for $i+1\geq qx$, one has
\begin{equation}
\psi_{i}(x)=\mathbb{P}\left(\mathrm{Bin}(q,x)\geq i+1\right)\leq e^{-qD(\frac{i+1}{q}||x)},\label{eq:Binomial_KL1}
\end{equation}
where $D(y||x)\triangleq y\ln\frac{y}{x}+(1-y)\ln\frac{1-y}{1-x}$ is the Kullback-Leibler divergence between two Bernoulli distributions. Similarly, for $i\leq qx$, one has
\begin{equation}
1-\psi_{i}(x)=\mathbb{P}\left(\mathrm{Bin}(q,x)\leq i\right)\leq e^{-qD(\frac{i}{q}||x)}.\label{eq:Binomial_KL2}
\end{equation}
\begin{lem}
\label{lem:Simple_KL_bin_bound} For $x\leq y$, we have
\[
\mathbb{P}\left(\mathrm{Bin}(q,x)\geq qy\right)\leq e^{-qd(y,x)},
\]
where $d(y,x)\triangleq\frac{1}{2}(y-x)^{2}/(x(1-x)+(1-2x)(y-x)/3)$. Similarly, for $x\geq y$, we have $\mathbb{P}\left(\mathrm{Bin}(q,x)\leq qy\right)\leq e^{-qd(y,x)}$.
\end{lem}
\begin{IEEEproof}
It is well known from the Chernoff bound that $\mathbb{P}\left(\mathrm{Bin}(q,x)\geq qy\right)\leq e^{-qD(y||x)}$ for $x\leq y$, where $D(y||x)\triangleq y\ln\frac{y}{x}+(1-y)\ln\frac{1-y}{1-x}$ is the Kullback-Leibler divergence. Thus, the first result holds if $d(y,x)\leq D(y||x)$ for $x\leq y$. Since $D(1-y||1-x)=D(y||x)$ and $d(1-y,1-x)=d(y,x)$, the second result follows from the first by symmetry. Thus, it suffices to prove that $d(y,x)\leq D(y||x)$ for $x\leq y$. To do this, we first observe that
\begin{align*}
\frac{\mathrm{d}}{\mathrm{d} x} & \left(d(y,x)-D(y||x)\right)\\
& =\frac{(1-x(1-x))(y-x)^{3}}{x(1-x)(y-x(x+2y-2))^{2}}\geq0
\end{align*}
for $x\leq y$. Next, we observe that
\begin{align*}
\int_{x}^{y} & \left(\frac{\mathrm{d}}{\mathrm{d} x'}\left(d(y,x')-D(y||x')\right)\right)\mathrm{d} x\\
& =\left(d(y,y)-D(y||y)\right)-\left(d(y,x)-D(y||x)\right)\geq0
\end{align*}
because $x\leq y$ throughout the range of integration. Since $d(y,y)=D(y||y)=0$, this implies $d(y,x)\leq D(y||x)$ for $x\leq y$.
\end{IEEEproof}
\begin{lem}
\label{lem:KL_Poisson} For $x,y\in[0,1]$, we have
\[
D(y||x)\geq(y-x)+(1-y)\ln\frac{1-y}{1-x}
\]
and, for $z\in[0,1]$, we have
\[
1-z+z\ln z\geq\frac{1}{2}(1-z)^{2}.
\]
\end{lem}
\begin{IEEEproof}
The first bound follows from lower bounding the $y\ln\frac{y}{x}$ term in $D(y||x)$ by
\[
y\ln\frac{y}{x}=-y\ln\frac{y-(y-x)}{y}\geq-y\left(-\frac{y-x}{y}\right)=y-x.
\]
Let $f(z)=z+(1-z)\ln(1-z)$ and observe that $f(1-z)=1-z+z\ln z$. Since $f'(0)=f(0)=0$ and $f''(z)=\frac{1}{1-z}\geq1$ for $z\in[0,1]$, it follows that
\[
f(z)=\int_{0}^{z}\int_{0}^{y}f''(x)\mathrm{d} x\,\mathrm{d} y\geq\frac{1}{2}z^{2}.
\]
Thus, $f(1-z)=1-z+z\ln z\geq\frac{1}{2}(1-z)^{2}.$
\end{IEEEproof}
\begin{lem}
\label{lem:g_psi_middle_term} For $\beta\in(0,\frac{1}{2}]$, $V(x)=(x(1-x))^{\beta}$, and $x\in[\frac{1}{2},1]$, we have
\[
\frac{1}{q}V\left(\psi_{\left\lceil qx\right\rceil -1}(x)\right)\leq\frac{\left(2x(1-x)\right)^{\beta}}{\sqrt{2q}}.
\]
\end{lem}
\begin{IEEEproof}
If $x\in[\frac{1}{2},1-\frac{1}{q}]$, then we have
\begin{align*}
\frac{1}{q}V & \left(\psi_{\left\lceil qx\right\rceil -1}(x)\right)\stackrel{(a)}{\leq}\frac{1}{q}\left(\frac{1}{4}\right)^{\beta}\\
& =\frac{\left(2x(1-x)\right)^{\beta}}{\sqrt{2q}}\frac{\sqrt{2q}}{\left(2x(1-x)\right)^{\beta}}\frac{1}{q}\left(\frac{1}{4}\right)^{\beta}\\
& =\frac{\left(2x(1-x)\right)^{\beta}}{\sqrt{2q}}\frac{\sqrt{2}}{\left(2xq(1-x)\right)^{\beta}}\frac{1}{q^{1/2-\beta}}\left(\frac{1}{4}\right)^{\beta}\\
& \stackrel{(b)}{\leq}\frac{\left(2x(1-x)\right)^{\beta}}{\sqrt{2q}}\frac{\sqrt{2}}{q^{1/2-\beta}}\left(\frac{1}{4}\right)^{\beta}\\
& \leq\frac{\left(2x(1-x)\right)^{\beta}}{\sqrt{2q}}\sup_{q\geq2}\frac{q^{\beta}\sqrt{2}}{q^{1/2}4^{\beta}}\\
& \stackrel{(b)}{=}\frac{\left(x(1-x)\right)^{\beta}}{\sqrt{2q}},
\end{align*}
where $(a)$ holds because $V(z)\leq(\frac{1}{4})^{\beta}$, $(b)$ follows from $2x\geq1$ and $q(1-x)\geq1$, and $(c)$ holds because the argument of the supremum is decreasing in $q$. If $x\in(1-\frac{1}{q},1]$, then assume $x=1-\frac{\alpha}{q}$ for $\alpha\in[0,1)$ and observe that
\begin{align*}
\frac{1}{q}V & \left(\psi_{\left\lceil qx\right\rceil -1}(x)\right)=\frac{1}{q}V\left(\psi_{q-1}(x)\right)\\
& =\frac{1}{q}(x^{q}(1-x^{q}))^{\beta}\\
& =\frac{1}{q}\left(\left(1-\frac{\alpha}{q}\right)^{q}\left(1-\left(1-\frac{\alpha}{q}\right)^{q}\right)\right)^{\beta}\\
& \stackrel{(a)}{\leq}\frac{1}{q}e^{-\alpha\beta}\alpha^{\beta}\\
& =\frac{\left(\frac{\alpha}{q}(1-\frac{\alpha}{q})\right)^{\beta}}{\sqrt{q}}\frac{\sqrt{q}}{\left(\frac{\alpha}{q}(1-\frac{\alpha}{q})\right)^{\beta}}\frac{1}{q}e^{-\alpha\beta}\alpha^{\beta}\\
& =\frac{\left(\frac{\alpha}{q}(1-\frac{\alpha}{q})\right)^{\beta}}{\sqrt{q}}\frac{q^{\beta-1/2}}{\left((1-\frac{\alpha}{q})\right)^{\beta}}e^{-\alpha\beta}\\
& \leq\frac{\left(x(1-x)\right)^{\beta}}{\sqrt{q}}\sup_{\alpha\in[0,1)}\sup_{q\geq2}\frac{q^{\beta-1/2}}{\left((1-\frac{\alpha}{q})\right)^{\beta}}e^{-\alpha\beta}\\
& \stackrel{(b)}{\leq}\frac{\left(x(1-x)\right)^{\beta}}{\sqrt{q}}\sup_{\alpha\in[0,1)}\frac{2^{\beta-1/2}}{\left((1-\frac{\alpha}{2})\right)^{\beta}}e^{-\alpha\beta}\\
& \stackrel{(c)}{\leq}\frac{\left(x(1-x)\right)^{\beta}}{\sqrt{q}}2^{\beta-1/2},
\end{align*}
where $(a)$ follows from $1-\alpha\leq\left(1-\frac{\alpha}{q}\right)^{q}\leq e^{-\alpha}$, $(b)$ holds because the argument of the supremum is decreasing in $q$, and $(c)$ holds because the argument of the supremum is decreasing in $\alpha$.
\end{IEEEproof}
\subsection{\label{subsec:proof_qary_main} Proof of Lemma~\ref{lem:lambda_q_b_bound}}
Let $V(x)=(x(1-x))^{\beta}$ with $\beta\in(0,\frac{1}{2}]$. Based on Lemma~\ref{lem:Tq_g_symmetry}, it is sufficient to analyze $(T_{q}V)(x)$ for $x\geq1/2$. To do this, we will use the decomposition
\begin{equation}
\begin{split}\frac{1}{q}\sum_{i=0}^{q-1} & V(\psi_{i}(x))=\frac{1}{q}\Bigg(\sum_{i=0}^{\left\lceil qx\right\rceil -2}V(\psi_{i}(x))\\
& \quad+V(\psi_{\left\lceil qx\right\rceil -1}(x))+\sum_{i=\left\lceil qx\right\rceil }^{q-1}V(\psi_{i}(x))\Bigg).
\end{split}
\label{eq:TV_sum_decomp}
\end{equation}
First, we consider the upper sum in~\eqref{eq:TV_sum_decomp}. Applying~\eqref{eq:Binomial_KL1} to $\psi_{i}(x)$ shows that
\begin{align*}
\psi_{i}(x) & =\mathbb{P}\left(\mathrm{Bin}(q,x)\geq i+1\right)\\
& \leq e^{-qD(\frac{i+1}{q}||x)}
\end{align*}
for $i+1\geq qx$. Thus, for $i\in\{\left\lceil qx\right\rceil ,\ldots,q-1\}$, we have $V(\psi_{i}(x))\leq(\psi_{i}(x))^{\beta}\leq e^{-q\beta D(\frac{i+1}{q}||x)}$ and
\begin{align}
\frac{1}{q}\sum_{i=\left\lceil qx\right\rceil }^{q-1}V\left(\psi_{i}(x)\right) & \leq\frac{1}{q}\sum_{i=\left\lceil qx\right\rceil }^{q-1}e^{-q\beta D(\frac{i+1}{q}||x)}\nonumber \\
& \stackrel{(a)}{\leq}\frac{1}{q}\sum_{i=\left\lceil qx\right\rceil }^{q-1}\int_{0}^{1}e^{-q\beta D(\frac{i+z}{q}||x)}\mathrm{d} z\nonumber \\
& \stackrel{(b)}{=}\int_{\left\lceil qx\right\rceil /q}^{1}e^{-q\beta D(y||x)}\mathrm{d} y,\label{eq:Tq_bound_upper}
\end{align}
where $e^{-q\beta D(\frac{i+1}{q}||x)}\leq\int_{0}^{1}e^{-q\beta D(\frac{i+z}{q}||x)}\mathrm{d} z$ holds in $(a)$ because $e^{-q\beta D(\frac{i+z}{q}||x)}$ is decreasing in $z$ for $i\geq qx$. Also, $(b)$ follows from grouping terms into one integral and changing the variable of integration. Although this bound holds for all $x\in[0,1]$, the sum is empty for $x\in(1-\frac{1}{q},1]$ and trivially equal to zero. For $x\geq\frac{1}{2}$, an upper bound on the integral is given by
\begin{align}
& \int_{\left\lceil qx\right\rceil /q}^{1}e^{-q\beta D(y||x)}\mathrm{d} y\leq\int_{x}^{1}e^{-q\beta D(y||x)}\mathrm{d} y\nonumber \\
& \stackrel{(a)}{\leq}\int_{x}^{1}\exp\left(\frac{-q\beta(y-x)^{2}}{2(x(1-x)+(1-2x)(y-x)/3)}\right)\mathrm{d} y\nonumber \\
& \quad\stackrel{(b)}{\leq}\int_{x}^{\infty}\exp\left(-\frac{q\beta(y-x)^{2}}{2x(1-x)}\right)\mathrm{d} y\nonumber \\
& \quad\quad=\sqrt{\frac{\pi x(1-x)}{2q\beta}},\label{eq:upper_sum_int_ub}
\end{align}
where $(a)$ follows from Lemma~\ref{lem:Simple_KL_bin_bound} and $(b)$ holds because $(1-2x)(y-x)\leq0$ for $y\geq x\geq\frac{1}{2}$.
Now, we consider the lower sum in~\eqref{eq:TV_sum_decomp}. Similarly, for $i\in\{0,\ldots,\left\lceil qx\right\rceil -2\}$, \eqref{eq:Binomial_KL2}~shows that
\[
\psi_{i}(x)=1-\mathbb{P}\left(\mathrm{Bin}(q,x)\leq i\right)\geq1-e^{-qD(\frac{i}{q}||x)}.
\]
For $i\in\{0,1,\left\lceil qx\right\rceil -2\}$, we have $V(1-\psi_{i}(x))\leq(1-\psi_{i}(x))^{\beta}\leq e^{-q\beta D(\frac{i}{q}||x)}$ and thus
\begin{align}
\frac{1}{q}\sum_{i=0}^{\left\lceil qx\right\rceil -2} & V\left(\psi_{i}(x)\right)=\frac{1}{q}\sum_{i=0}^{\left\lceil qx\right\rceil -2}V\left(1-\psi_{i}(x)\right)\nonumber \\
& \leq\frac{1}{q}\sum_{i=0}^{\left\lceil qx\right\rceil -2}e^{-q\beta D(\frac{i}{q}||x)}\nonumber \\
& \stackrel{(a)}{\leq}\frac{1}{q}\sum_{i=0}^{\left\lceil qx\right\rceil -2}\int_{0}^{1}e^{-q\beta D(\frac{i+z}{q}||x)}\mathrm{d} z\nonumber \\
& \stackrel{(b)}{=}\int_{0}^{(\left\lceil qx\right\rceil -1)/q}e^{-q\beta D(y||x)}\mathrm{d} y,\label{eq:Tq_bound_lower}
\end{align}
where $e^{-q\beta D(\frac{i}{q}||x)}\leq\int_{0}^{1}e^{-q\beta D(\frac{i+z}{q}||x)}\mathrm{d} z$ holds in $(a)$ because $e^{-q\beta D(\frac{i+z}{q}||x)}$ is increasing in $z$ for $z\in[0,1]$ and $i+1\leq qx$. Also, $(b)$ follows from grouping terms into one integral and changing the variable of integration.
The expression in~\eqref{eq:Tq_bound_lower} can be upper bounded using the decomposition
\begin{align}
\int_{0}^{(\left\lceil qx\right\rceil -1)/q} & e^{-q\beta D(y||x)}\mathrm{d} y\leq\int_{0}^{x}e^{-q\beta D(y||x)}\mathrm{d} y\nonumber \\
& \leq\int_{0}^{2x-1}e^{-q\beta D(y||x)}\mathrm{d} y+\int_{2x-1}^{x}e^{-q\beta D(y||x)}\mathrm{d} y.\label{eq:Tq_bound_lower_split}
\end{align}
The first term in~\eqref{eq:Tq_bound_lower_split} can be upper bounded with
\begin{align}
& \int_{2x-1}^{x}e^{-q\beta D(y||x)}dy\nonumber \\
& \stackrel{(a)}{\leq}\!\!\int_{2x-1}^{x}\!\!\!\!\!\exp\left(\frac{-q\beta(y-x)^{2}}{2(x(1-x)+(2x-1)(x-y)/3)}\right)\mathrm{d} y\nonumber \\
& \stackrel{(b)}{\leq}\!\!\int_{2x-1}^{x}\!\!\!\exp\left(\frac{-q\beta(y-x)^{2}}{2((1-x)+(1-x)/3)}\right)\mathrm{d} y\nonumber \\
& =\sqrt{\frac{2\pi(1-x)}{3q\beta}}\text{erf}\left(\sqrt{\frac{3\beta q(1-x)}{8}}\right)\nonumber \\
& \stackrel{(c)}{\leq}\sqrt{\frac{4\pi x(1-x)}{3q\beta}},\label{eq:lowersum1_final_ub}
\end{align}
where $(a)$ follows from Lemma~\ref{lem:Simple_KL_bin_bound}, $(b)$ holds because $x(1-x)\leq1-x$ and $(2x-1)(x-y)\leq1-x$ for $y\geq2x-1$ and $x\geq\frac{1}{2}$, and $(c)$ follows from $2x\geq1$ for $x\geq\frac{1}{2}$. The second term in~\eqref{eq:Tq_bound_lower_split} can be upper bounded with
\begin{align}
& \int_{0}^{2x-1}e^{-q\beta D(y||x)}\mathrm{d} y\nonumber \\
& \stackrel{(a)}{\leq}\!\!\int_{0}^{2x-1}\!\!\!\!\!\!\!\!\exp\left(-q\beta\left((y-x)+(1-y)\ln\frac{1-y}{1-x}\right)\right)\mathrm{d} y\nonumber \\
\,\nonumber \\
& \stackrel{(b)}{=}\!\!\int_{\frac{1}{1-x}}^{2}\!\!\!\exp\bigg(-q\beta\bigg((1-z(1-x)-x)\nonumber \\
& \quad\quad\quad\quad\quad\quad+z(1-x)\ln\frac{z(1-x)}{1-x}\bigg)\bigg)(x\!-\!1)\mathrm{d} z\nonumber \\
& =(1\!-\!x)\int_{2}^{\frac{1}{1-x}}\!\!\exp\left(q\beta(x-1)\left((1-z)+z\ln z\right)\right)\mathrm{d} z\nonumber \\
& \stackrel{(c)}{\leq}(1\!-\!x)\int_{2}^{\frac{1}{1-x}}\!\!\exp\left(\frac{1}{2}q\beta(x-1)(z-1)^{2}\right)\mathrm{d} z\nonumber \\
& =(1\!-\!x)\sqrt{\frac{\pi}{2q\beta(1-x)}}\text{erf}\sqrt{\frac{q\beta(1-x)(z-1)^{2}}{2}}\bigg|_{z=2}^{\frac{1}{1-x}}\nonumber \\
& \stackrel{(d)}{\leq}(1-x)\sqrt{\frac{\pi}{2q\beta(1-x)}}\nonumber \\
& \stackrel{(e)}{\leq}\sqrt{\frac{\pi x(1-x)}{q\beta}},\label{eq:lowersum2_final_ub}
\end{align}
where $(a)$ follows from Lemma~\ref{lem:KL_Poisson}, $(b)$ is given by the change of variables $y\mapsto1-z(1-x)$, $(c)$ holds because $1-z+z\ln z\geq\frac{1}{2}(z-1)^{2}$ for $z\in[0,1]$, $(d)$ follows from $\mbox{erf}(b)-\mbox{erf}(a)\leq1$ for $b\geq a\geq0$, and $(e)$ holds because $2x\geq1$ for $x\geq\frac{1}{2}$.
Now, we combine Lemma~\ref{lem:g_psi_middle_term} with~\eqref{eq:upper_sum_int_ub}, \eqref{eq:lowersum1_final_ub}, and \eqref{eq:lowersum2_final_ub} to see that
\begin{align}
& (T_{q}V)(x)=\frac{1}{q}\Bigg(\sum_{i=0}^{\left\lceil qx\right\rceil -2}V\left(\psi_{i}(x)\right)+V\left(\psi_{\left\lceil qx\right\rceil -1}(x)\right)\nonumber \\
& \quad\quad\quad\quad\quad\quad\quad+\sum_{i=\left\lceil qx\right\rceil }^{q-1}V\left(\psi_{i}(x)\right)\Bigg)\nonumber \\
& \leq\int_{0}^{(\left\lceil qx\right\rceil -1)/q}e^{-q\beta D(y||x)}\mathrm{d} y+\frac{1}{q}V\left(\psi_{\left\lceil qx\right\rceil -1}(x)\right)\nonumber \\
& \quad\quad\quad+\int_{\left\lceil qx\right\rceil /q}^{1}e^{-q\beta D(y||x)}\mathrm{d} y\nonumber \\
& \leq\sqrt{\frac{\pi x(1-x)}{q\beta}}+\sqrt{\frac{4\pi x(1-x)}{3q\beta}}\nonumber \\
& \quad\quad\quad+\frac{\left(2x(1-x)\right)^{\beta}}{\sqrt{2q}}+\sqrt{\frac{\pi x(1-x)}{2q\beta}}\nonumber \\
& \leq\frac{\left(2x(1-x)\right)^{\beta}}{\sqrt{2q}}+A\sqrt{\frac{x(1-x)}{q\beta}},\label{eq:TqV_ub1}
\end{align}
where $A=\sqrt{\pi}+\sqrt{\frac{4\pi}{3}}+\sqrt{\frac{\pi}{2}}$. Combining Definition~\ref{def:lambda_q_b} with~\eqref{eq:TqV_ub1}, we see that
\begin{align*}
\lambda_{q,\beta} & \triangleq\sup_{x\in(0,1)}\frac{(T_{q}V)(x)}{V(x)}\\
\leq & \sup_{x\in(0,1)}\left(\frac{2^{\beta}}{\sqrt{2q}}+\frac{A}{\sqrt{q\beta}}\left(x(1-x)\right)^{1/2-\beta}\right)\\
\stackrel{(a)}{\leq} & \frac{2^{\beta}}{\sqrt{2q}}+A\frac{1}{\sqrt{q\beta}}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}\\
\leq & \frac{1}{\sqrt{q\beta}}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}\left(A+\frac{2^{\beta}\sqrt{\beta}}{\sqrt{2}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}}\right)\\
& \stackrel{(b)}{\leq}\frac{6}{\sqrt{q\beta}}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta},
\end{align*}
where $(a)$ holds because the supremum is achieved at $x=\frac{1}{2}$ and $(b)$ holds because
\[
A+\frac{2^{\beta}\sqrt{\beta}}{\sqrt{2}\left(\frac{1}{4}\right)^{\frac{1}{2}-\beta}}\leq6
\]
for $\beta\in(0,\frac{1}{2}]$.
\section{\label{sec:fixed_alphabet} Fixed-Alphabet Erasure Channels with\protect \\
Large Polarizing Matrices}
The results in the previous section show that one can approach the optimal scaling rate for $q$-ary erasure channels with $q\times q$ polarizing matrices. It does not, however, allow one to separate the effect of alphabet size and transform size. In this section, we explicitly consider this distinction by constructing a polar code based on $m\times m$ polarizing matrices over $\mathbb{F}_{q}$ and focus on the scaling as $m$ increases. The polar decoder at each stage is based on APP decoding the $m$ nested subcodes associated with successive cancellation decoding. The rate of polarization in $m$ is investigated numerically and a formula for the general case is conjectured.
Let $GL(m,\mathbb{F}_{q})$ denote the general linear group of invertible $m\times m$ matrices over $\mathbb{F}_{q}$. For a fixed $G'\in GL(m,\mathbb{F}_{q})$, let $\varphi_{i}(x;G')$ be the erasure rate (under APP decoding) of the first input symbol of the $(n,n-i)$ linear encoder defined by the bottom $n-i$ rows of $G'$, assuming the output symbols are erased with probability $x$. Consider the length $N=m^{n}$ polar code defined by the polar transform $G_{0}^{\otimes n}$ for a fixed $G_{0}\in GL(m,\mathbb{F}_{q})$. During each stage of decoding, effective channels with erasure rate $x$ are transformed into $m$ different effective channels with erasure rates $\varphi_{i}(x;G_{0})$ for $i\in\mathcal{M}\triangleq\left\{ 0,1,\ldots,m-1\right\} $. Let the random variable $X_{n}$ denote the erasure probability of a randomly chosen effective channel after $n$ levels of polarization~\cite{Hassani-it13,Mondelli-arxiv15}. Then, the sequence $X_{n}$ is a Markov chain on the compact state space $\mathcal{X}=[0,1]$ with transition kernel
\begin{align*}
\mathbb{P}\Big( & X_{n}\,=x_{n}\,\Big|\,(X_{0},\ldots,X_{n-1})=(x_{0},\ldots,x_{n-1})\Big)\\
& =\mathbb{P}\left(X_{n}=x_{n}\midb |X_{n-1}=x_{n-1}\right)\\
& \quad=\frac{1}{m}\left|\left\{ i\in\mathcal{M}\,|\,x_{n}=\varphi_{i}(x_{n-1};G_{0})\right\} \right|.
\end{align*}
Similar to previous examples, one can analyze this Markov chain by focusing on the cost function $g_{n}(x)\triangleq\mathbb{E}[g_{0}(X_{n})|X_{0}=x]$ generated by $g_{0}\in C(\mathcal{X})$. The one-step update for $g_{n}$ is given by the linear operator $T_{m}:C(\mathcal{X})\to C(\mathcal{X})$ where
\[
g_{n+1}(x)=(T_{m}g_{n})(x)\triangleq\frac{1}{m}\sum_{i=0}^{m-1}g_{n}\left(\varphi_{i}(x;G_{0})\right).
\]
Numerical results show that the length-16 BCH kernel from~\cite{Korada-it10*2} has a better rate of polarization than Ar{\i}kan's binary polar codes. If one chooses $g_{0}(x)=(x(1-x))^{0.6}$ for this kernel, then one finds that
\[
\lambda_{BCH16}=\sup_{x\in[0,1]}\frac{g_{1}(x)}{g_{0}(x)}\approx0.4508.
\]
This is better than the mixing value $\lambda_{2}^{4}\approx(0.827)^{4}\approx0.4677$ given by 4 stages of binary polarization. Indeed, this is quite related to the error exponent of polar codes. Computing this quantity for any fixed $m\times m$ kernel requires summing over all $2^{m}$ subsets. Thus, it is computationally challenging to extend these results much beyond length 16 (e.g., to length 32).
For the sake of analysis, we consider an inhomogeneous polar code where the polarizing matrix (or kernel) at each stage is different. Thus, the overall length-$N=m^{n}$ polar transform is defined by the generator matrix
\[
G=G_{1}\otimes G_{2}\otimes\cdots\otimes G_{n},
\]
where the sequence $G_{1},G_{2},\ldots,G_{n}$ is drawn i.i.d. from the general linear group $GL(m,\mathbb{F}_{q})$ of invertible $m\times m$ matrices over $\mathbb{F}_{q}$. We note that unconstrained i.i.d. random matrices cannot be used in this construction because the preservation of mutual information required by polarization fails if the kernel is not invertible. Let the random variable $X_{n}$ denote the channel erasure probability for a randomly chosen effective channel after $n$ levels of polarization~\cite{Hassani-it13,Mondelli-arxiv15}. Then, the sequence $X_{n}$ is a Markov chain on the compact state space $\mathcal{X}=[0,1]$ with transition kernel
\begin{align*}
\mathbb{P} & \left(X_{n}=x_{n}\midb |X_{n-1}=x_{n-1}\right)=\frac{1}{m\left|GL(m,\mathbb{F}_{q})\right|}\\
& \cdot\left|\left\{ (i,G_{n})\in\mathcal{M}\times GL(m,\mathbb{F}_{q})\,|\,x_{n}=\varphi_{i}(x_{n-1};G_{n})\right\} \right|.
\end{align*}
Since this probability also includes the randomness in the kernel selection, it is worth noting that the conditional fraction of unpolarized channels, $\mathbb{P}\left(X_{n}\in[\eta,1-\eta]\,\big|\,G_{1},\ldots,G_{n}\right)$, is a random variable. Thus, the Markov inequality implies that
\begin{align*}
\mathbb{P}\Big(\Big\{\mathbb{P}\big(X_{n}\in[\eta,1-\eta] & \,\big|\,G_{1},\ldots,G_{n}\big)>\delta\Big\}\Big)\\
& <\frac{1}{\delta}\mathbb{P}\left(X_{n}\in[\eta,1-\eta]\right).
\end{align*}
Hence, choosing $\delta=\mathbb{P}\big(X_{n}\in[\eta,1-\eta]\big)$ shows that there is at least one code with, at most, a fraction $\mathbb{P}\big(X_{n}\in[\eta,1-\eta]\big)$ of unpolarized channels.
Similar to previous examples, one can analyze this Markov chain by focusing on the cost function $g_{n}(x)\triangleq\mathbb{E}[g_{0}(X_{n})|X_{0}=x]$ generated by $g_{0}\in C(\mathcal{X})$. The one-step update for $g_{n}$ is given by the linear operator $T_{m}:C(\mathcal{X})\to C(\mathcal{X})$ where
\begin{align*}
g_{n+1}(x) & =(T_{m}g_{n})(x)\triangleq\frac{1}{\left|GL(m,\mathbb{F}_{q})\right|}\\
& \quad\cdot\sum_{G_{n}\in GL(m,\mathbb{F}_{q})}\frac{1}{m}\sum_{i=0}^{m-1}g_{n}\left(\varphi_{i}(x;G_{n})\right).
\end{align*}
If one chooses $g_{0}(x)$ to be concave, then one also gets the upper bound
\begin{align*}
g_{1}(x) & =\frac{1}{\left|GL(m,\mathbb{F}_{q})\right|}\sum_{G_{1}\in GL(m,\mathbb{F}_{q})}\frac{1}{m}\sum_{i=0}^{m-1}g_{0}\left(\varphi_{i}(x;G_{1})\right)\\
& \leq\frac{1}{m}\sum_{i=0}^{m-1}g_{0}\left(\sum_{G_{1}\in GL(m,\mathbb{F}_{q})}\frac{\varphi_{i}(x;G_{1})}{\left|GL(m,\mathbb{F}_{q})\right|}\right)\\
& \triangleq\overline{g}_{1}(x).
\end{align*}
Continuing recursively, one can define
\begin{align*}
\overline{g}_{n+1}(x) & \triangleq\frac{1}{m}\sum_{i=0}^{m-1}\overline{g}_{n}\left(\sum_{G_{n+1}\in GL(m,\mathbb{F}_{q})}\frac{\varphi_{i}(x;G_{n+1})}{\left|GL(m,\mathbb{F}_{q})\right|}\right)\\
& =\frac{1}{m}\sum_{i=0}^{m-1}\overline{g}_{n}\left(\overline{\varphi}_{i}(x)\right),
\end{align*}
where $\overline{\varphi}_{i}(x)$ is the average of $\varphi_{i}(x;G')$ over all $G'\in GL(m,\mathbb{F}_{q})$. Similarly, if $g_{n}(x)\leq\overline{g}_{n}(x)$ and $\overline{g}_{n}(x)$ is concave, then one can write
\begin{align*}
g_{n+1}(x) & \leq\frac{1}{m}\sum_{i=0}^{m-1}g_{n}\left(\overline{\varphi}_{i}(x)\right)\\
& \leq\frac{1}{m}\sum_{i=0}^{m-1}\overline{g}_{n}\left(\overline{\varphi}_{i}(x)\right)=\overline{g}_{n+1}(x)
\end{align*}
to derive an inductive upper bound. To complete the analysis, we need a formula for $\overline{\varphi}_{i}(x)$ and a concave cost function $g_{0}(x)$ such that $\overline{g}_{n}(x)$ is a sequence of concave functions. A closed-form expression for $\overline{\varphi}_{i}(x)$, which can be evaluated using $O(m^{4})$ real operations, is derived in Appendix~\ref{sec:full_rank}.
For $q=2$ and $m=16,32,64$, one can choose $g_{0}(x)=(x(1-x))^{0.35}$ and observe that $\overline{g}_{1}(x)$ is concave. Moreover, the constants
\[
\lambda_{m}=\sup_{x\in[0,1]}\frac{\overline{g}_{1}(x)}{g_{0}(x)}
\]
are given by $(\lambda_{16},\lambda_{32},\lambda_{64})\approx(0.6729,0.4558,0.2880)$. Comparing with binary polar codes, where the mixing rate is roughly $0.827^{\log_{2}m}$, one finds that $\lambda_{64}\approx0.2880\leq0.3199\approx0.827{}^{6}$. Thus, this predicts that polar codes based on random invertible $64\times64$ binary polarizing kernels will achieve a better scaling rate that Ar{\i}kan's binary polar codes. This statement is not rigorous, however, because we have not shown that $\overline{g}_{n}(x)$ remains concave for all $n$. This leads to the conjecture.
\begin{conjecture}
For any prime-power $q\geq2$, any integer $m\geq2$, and any $\beta\in(0,1)$, let $g_{0}(x)=(x(1-x))^{\beta}$. Then, $\overline{g}_{n}(x)$ is concave on $[0,1]$ for all $n$.
\end{conjecture}
If this conjecture is true, then this sequence of inhomogeneous polar codes can be rigorously characterized in terms of $\lambda_{m}$. Additionally, our next conjecture is sufficient to imply that there is a sequence of inhomogeneous polar codes that achieves near-optimal scaling with fixed $q$ as $m$ increases.
\begin{conjecture}
For any prime-power $q\geq2$ and any $\beta\in(0,\frac{1}{2}]$, let $g_{0}(x)=(x(1-x))^{\beta}$. Then,
\[
\lim_{m\to\infty}\frac{1}{\ln m}\ln\lambda_{m}=\lim_{m\to\infty}\frac{1}{\ln m}\ln\sup_{x\in[0,1]}\frac{\overline{g}_{1}(x)}{g_{0}(x)}=-\frac{1}{2}.
\]
\end{conjecture}
If these conjectures are both true, then Corollary~\ref{cor:lyapunov} implies that, for any $\gamma>0$ and $\beta\in(0,\frac{1}{2}]$, we have
\begin{align*}
\mathbb{P}\left(X_{n}\geq N^{-\gamma}\,|\,X_{0}=x\right) & \leq N^{\gamma\beta}e^{n\ln m\frac{\ln\lambda_{m}}{\ln m}}+\frac{x}{1-N^{-\gamma}}\\
& =N^{\gamma\beta-\frac{1}{2}+o_{m}(1)}+\frac{x}{1-N^{-\gamma}}.
\end{align*}
\section{Conclusion}
In this paper, we first investigate the relationship between the blocklength and the gap to capacity for the $q$-ary Reed-Solomon polar codes introduced by Mori and Tanaka. These codes have length $N=q^{n}$, where $n$ is the number of steps in the polarization process. When one of these codes is transmitted over a $q$-ary erasure channel with erasure probability $\epsilon$, its effective channels are $q$-ary erasure channels and their erasure rate satisfy a closed-form recursion. By analyzing this recursion, we show that , for any $\gamma>0$ and $\delta>0$, there is a $q_{0}$ such that, for all $q\geq q_{0}$, the fraction of effective channels with erasure rate at most $O(N^{-\gamma})$ is at least $1-\epsilon-O(N^{-1/2+\delta})$. Thus, the gap to capacity scales at a rate very close to the optimal rate of $O(N^{-1/2})$.
In the second part of this paper, a similar analysis is also considered for $q$-ary polar codes with $m$ by $m$ polarizing matrices. To prove near-optimal scaling for polar codes with fixed $q$ as $m$ increases, however, two technical obstacles remain. Thus, we conclude by stating two concrete mathematical conjectures that, if proven, would imply near-optimal scaling.
These results naturally suggest two interesting open questions. First, can one prove that $q$-ary inhomogeneous polar codes with random $m\times m$ polarization kernels achieve near-optimal scaling on the $q$-ary erasure channel with fixed $q$ as $m$ increases? Second, can this result be extended to noisy $q$-ary (or even binary) channels?
\appendices
\section{\label{sec:full_rank} The Polar Erasure Rate for\protect \\
Random Full-Rank Matrices}
Let $G$ be an $m\times m$ full-rank matrix and let $G^{(i)}$ be the $(m-i)\times m$ submatrix formed by removing the first $i$ rows from $G$. Then, $G^{(i)}$ has rank $m-i$ (i.e., full rank) because removing a row from a matrix reduces its rank by at most 1. Recall that, during polar decoding, one decodes the sequence of full-rank generator matrix codes defined by $G^{(i)}$. Let $\mathcal{S}$ be the set of indices of correctly received output bits and let $G_{\mathcal{S}}^{(i)}$ be the $(m-i)\times|\mathcal{S}|$ submatrix of $G^{(i)}$ containing only columns whose indices are in $\mathcal{S}$. For any fixed $\mathcal{S}$, the $i$-th input bit can be recovered during the $i$-th decoding step iff the $i$-th input bit can be written as a linear combination of the correctly received bits. In this setup, this occurs iff there is a vector $v$ such that $G_{\mathcal{S}}^{(i)}v=e_{1},$ where $e_{j}$ is the unit column vector with a one in the $j$-th position. In terms of matrix ranks, this is equivalent to the condition
\[
\mathrm{rk}\big(G_{\mathcal{S}}^{(i)}\big)=\mathrm{rk}\big([e_{1}\;G_{\mathcal{S}}^{(i)}]\big).
\]
Since $\mathrm{rk}\big([e_{1}\;G_{\mathcal{S}}^{(i)}]\big)\leq\mathrm{rk}\big(G_{\mathcal{S}}^{(i)}\big)+1$, it follows that
\[
E^{(i)}(\mathcal{S})\triangleq\mathrm{rk}\big([e_{1}\;G_{\mathcal{S}}^{(i)}]\big)-\mathrm{rk}\big(G_{\mathcal{S}}^{(i)}\big)
\]
is the erasure indicator function for the $i$-th step of decoding when $\mathcal{S}$ is the set of correctly received bits.
For a random $G$ matrix, we observe that
\[
\mathbb{E}\left[E^{(i)}(\mathcal{S})\right]=\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)\right)
\]
because the rank is increased by including the column $e_{1}$ iff $e_{1}$ is not in the column space. If $G\in\mathbb{F}_{q}^{m\times m}$ is a uniform random full-rank matrix, then $G^{(i)}\in\mathbb{F}_{q}^{(m-i)\times m}$ is also uniform random full-rank matrix. Since the $G^{(i)}$ ensemble is invariant under column permutations, it follows $\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)\right)$ only depends on the cardinality of $\mathcal{S}$ and not on the individual elements. Hence, we define
\[
\rho(m,i,|\mathcal{S}|,q)\triangleq\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)\right)
\]
for this ensemble.
Now, we assume that each output bit is erased independently with probability $x$ and let $\overline{\varphi}_{i}(x)$ be the erasure probability of the $i$-th input bit during the $i$-th step of polar decoding averaged over the ensemble of full-rank polarizing matrices. In this case, we find that
\begin{align*}
\overline{\varphi}_{i}(x) & \triangleq\mathbb{E}\left[\sum_{\mathcal{S}\subseteq[m]}x^{m-|\mathcal{S}|}(1-x)^{|\mathcal{S}|}E^{(i)}(\mathcal{S})\right]\\
& =\sum_{\mathcal{S}\subseteq[m]}x^{m-|\mathcal{S}|}(1-x)^{|\mathcal{S}|}\mathbb{P}\left(e_{1}\notin\mbox{colspace}\left(G_{\mathcal{S}}^{(i)}\right)\right)\\
& =\sum_{d=0}^{m}\binom{m}{d}x^{m-d}(1-x)^{d}\rho(m,i,d,q).
\end{align*}
A closed-form expression for $\rho(m,i,|\mathcal{S}|,q)$ is derived below.
It is well-known~(e.g., see \cite{Fisher-amm66}) that, for a uniform random matrix $G'\in\mathbb{F}_{q}^{k\times d}$, the rank satisfies
\[
\mathbb{P}\left(\mathrm{rk}(G')=j\right)=\phi(j,d,q){k \brack j}_{q}q^{-kd},
\]
where the number of sets of $j$ linearly independent vectors of length $i$ is given by
\[
\phi(j,i,q)\triangleq\prod_{l=0}^{j-1}(q^{i}-q^{l})
\]
and the number of $j$-dimensional subspaces of $\mathbb{F}_{q}^{k}$ equals the Gaussian binomial coefficient
\[
{k \brack j}_{q}\triangleq\prod_{l=0}^{j-1}\frac{q^{k}-q^{l}}{q^{j}-q^{l}}.
\]
For fixed $\mathcal{S}$, consider the submatrix $G_{\mathcal{S}}'$ of a uniform random matrix $G'\in\mathbb{F}_{q}^{k\times m}$ and define $\mathcal{S}^{c}\triangleq\left\{ 1,2,\ldots,m\right\} \backslash\mathcal{S}$. One can compute the joint rank distribution of $G_{\mathcal{S}}'$ and $G'$ using the fact that
\begin{align*}
\theta & (m,k,r,j,q)\triangleq\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j,\,\mathrm{rk}\big(G'\big)=r\right)\\
& =\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j\right)\mathbb{P}\left(\mathrm{rk}\big([G_{\mathcal{S}}'\;G_{\mathcal{S}^{c}}']\big)=r\big|\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j\right)\\
& =\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j\right)\sum_{l=0}^{r}\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}^{c}}'\big)=l\right)\\
& \quad\quad\cdot\mathbb{P}\left(\mathrm{rk}\big([G_{\mathcal{S}}'\;G_{\mathcal{S}^{c}}']\big)=r\big|\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j,\mathrm{rk}\big(G_{\mathcal{S}^{c}}'\big)=l\right)\\
& =\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j\right)\sum_{l=0}^{r}\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}^{c}}'\big)=l\right)\\
& \quad\quad\cdot\mathbb{P}\left(\dim\left(W\oplus W'\right)=r\,|\,\dim(W)=j,\dim(W')=l\right),
\end{align*}
where $W$ and $W'$ are the independent uniform random column-spaces of $G_{\mathcal{S}}'$ and $G_{\mathcal{S}^{c}}'$. For independent uniform random subspaces $W,W'$ of $\mathbb{F}_{q}^{k}$, it is shown in~\cite{Rathi-iee05} that
\begin{align*}
\mathbb{P} & \left(\dim\left(W\oplus W'\right)=r\,|\,\dim(W)=j,\dim(W')=l\right)\\
& \quad\quad=q^{(r-j)(r-l)}{j \brack r-l}_{q}{k-j \brack k-r}_{q}\bigg/{k \brack l}_{q}.
\end{align*}
Therefore, we can write
\begin{align}
\mathbb{P} & \left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j,\,\mathrm{rk}\big(G'\big)=r\right)=\phi(j,d,q){k \brack j}_{q}q^{-kd}\nonumber \\
& \;\cdot\sum_{l=0}^{r}\phi(l,m-d,q)q^{-k(m-d)}q^{(r-j)(r-l)}{j \brack r-l}_{q}{k-j \brack k-r}_{q}.\label{eq:joint_rank}
\end{align}
Now, we have the tools to complete the main derivation. Let $\mathcal{S}\subseteq[m]$ be an arbitrary subset satisfying $|\mathcal{S}|=d$ and write
\begin{align*}
\rho & (m,i,d,q)=\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)\right)\\
& =\sum_{j=0}^{m-i}\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}^{(i)}\big)=j\right)\\
& \quad\quad\cdot\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)\,\big|\,\mathrm{rk}\big(G_{\mathcal{S}}^{(i)}\big)=j\right)\\
& \stackrel{(a)}{=}\sum_{j=0}^{m-i}\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j\,\big|\,\mathrm{rk}\big(G'\big)=m-i\right)\\
& \quad\quad\cdot\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(G_{\mathcal{S}}'\big)\,\big|\,\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j\right)\\
& \stackrel{(b)}{=}\sum_{j=0}^{m-i}\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j\,\big|\,\mathrm{rk}\big(G'\big)=m-i\right)\frac{q^{m-i}-q^{j}}{q^{m-i}-1}\\
& =\sum_{j=0}^{m-i}\left(\frac{q^{m-i}-q^{j}}{q^{m-i}-1}\right)\frac{\mathbb{P}\left(\mathrm{rk}\big(G_{\mathcal{S}}'\big)=j,\,\mathrm{rk}\big(G'\big)=m-i\right)}{\mathbb{P}\left(\mathrm{rk}\big(G'\big)=m-i\right)}\\
& \stackrel{(c)}{=}\sum_{j=0}^{m-i}\left(\frac{q^{m-i}-q^{j}}{q^{m-i}-1}\right)\frac{\phi(j,d,q){m-i \brack j}_{q}q^{-(m-i)d}}{\phi(m-i,m,q)q^{-(m-i)m}}q^{-(m-i)(m-d)}\\
& \quad\quad\cdot\sum_{l=0}^{m-i}\phi(l,m-d,q)q^{(m-i-j)(m-i-l)}{j \brack m-i-l}_{q}\\
& =\sum_{j=0}^{m-i}\left(\frac{q^{m-i}-q^{j}}{q^{m-i}-1}\right)\frac{\phi(j,d,q){m-i \brack j}_{q}}{\phi(m-i,m,q)}\\
& \quad\quad\cdot\sum_{l=0}^{m-i}\phi(l,m-d,q)q^{(m-i-j)(m-i-l)}{j \brack m-i-l}_{q},
\end{align*}
where $(a)$ follows from the fact that a uniform random matrix $G'\in\mathbb{F}_{q}^{(m-i)\times m}$ has the same distribution as $G^{(i)}$ conditioned on the event that $\mathrm{rk}\big(G_{\mathcal{S}}'\big)=m-i$, $(b)$ follows from the fact that the $G^{(i)}$ ensemble is invariant under left-multiplication by a random invertible matrix and $(c)$ follows from~\eqref{eq:joint_rank}.
Let $G^{(i)}$ be a random full-rank $(m-i)\times m$ matrix. Then, the choice of $G^{(i)}$ can be separated into the choice of a uniform random subspace, $\mathcal{C}\subseteq\mathbb{F}_{q}^{m}$, and a uniform random basis for that subspace. Let $H\in\mathbb{F}_{q}^{i\times m}$ be a uniform random basis for the dual space $\mathcal{C}^{\perp}$. Then, $H$ is a uniform random parity-check matrix for the code generated by $G^{(i)}$ (i.e., $G^{(i)}H^{T}=0_{(m-i)\times i})$, where $0_{a\times b}$ denotes an $a\times b$ all-zero matrix. Since the choice of $H$ can also be separated into the choice of a uniform random subspace, $\mathcal{C}^{\perp}$, and a uniform random basis, it follows that the marginal distribution of $H$ (i.e., averaged over the choice of $G^{(i)}$) is uniform over the set of full-rank $i\times m$ matrices. Similarly, the conditional distribution of $G^{(i)}$ given $H$ is uniform over the set of bases for $\big(\mathcal{C}^{\perp}\big)^{\perp}=\mathcal{C}$.
For any full-rank generator matrix $G^{(i)}\in\mathbb{F}_{q}^{(m-i)\times m}$, one can define the augmented generator matrix $\tilde{G}=[G^{(i)}\;e_{1}]$. For the $(m+1,m-i)$ linear code generated by $\tilde{G}$, the recovery of the first information bit (of either code) is equivalent to the extrinsic recovery of the last code bit of the augmented code. Let $\tilde{H}$ be a $(i+1)\times(m+1)$ parity-check matrix for the augmented code (i.e., $\tilde{G}\tilde{H}^{T}=0_{(m-i)\times(i+1)})$. We construct a (non-uniform) random $\tilde{H}$ matrix on the same probability space by defining
\[
\tilde{H}\triangleq\begin{bmatrix}v^{T} & 1\\
H & 0_{m}
\end{bmatrix},
\]
where $v$ is a uniform random solution to $G^{(i)}v=-e_{1}$. Based on this definition, one can verify that
\[
\tilde{G}\tilde{H}^{T}=[G^{(i)}\;e_{1}]\begin{bmatrix}v & H^{T}\\
1 & 0_{m}^{T}
\end{bmatrix}=\begin{bmatrix}0_{m-i} & 0_{(m-i)\times(i+1)}\end{bmatrix}.
\]
Now, we will show that the matrix
\[
\tilde{H}'\triangleq\begin{bmatrix}v\\
H
\end{bmatrix}
\]
is a uniform random $(i+1)\times m$ full-rank matrix. To do this, we first recall that $H$ is a uniform random $i\times m$ full-rank matrix and, conditioned on $H$, the rows of $G^{(i)}$ form a uniform random basis for the null space of $H$. This means that we can write $G^{(i)}=BG'$, where $B$ is uniform random element of $GL(m-i,\mathbb{F}_{q})$ and $G'$ is any basis for the null space of $H$. Now, the vector $v$ is chosen to be a uniform random solution of the system $G^{(i)}v=-e_{1}$ which can be rewritten as $Gv=-B^{-1}e_{1}$. Since $-B^{-1}e_{1}$ is a uniform random non-zero vector, the vector $v$ is distributed uniformly over the set $A$ of vectors such that $Gv\neq0$. Then, $A=\mathbb{F}_{q}^{m}\backslash\text{rowspace}(H)$ because the null space of $G$ equals the row space of $H$. Moreover, $A$ is exactly equal to the set of row vectors that can be used to extend $H$ to a full-rank $(i+1)\times m$ matrix. Thus, $\tilde{H}'$ is a uniform random $(i+1)\times m$ full-rank matrix.
Now, we observe that there exists a vector $u$ such that $G_{\mathcal{S}}^{(i)}u_{\mathcal{S}}=e_{1}$ and $u_{\mathcal{S}^{c}}=0$ if and only if $e_{1}\in\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)$. If and only if such a $u$ vector exists, then we have the equivalence
\[
\tilde{H}_{\mathcal{S}^{c}}=\begin{bmatrix}v_{\mathcal{S}^{c}}^{T}\\
H_{\mathcal{S}^{c}}
\end{bmatrix}\sim\begin{bmatrix}u_{\mathcal{S}^{c}}^{T}\\
H_{\mathcal{S}^{c}}
\end{bmatrix}=\begin{bmatrix}0\\
H_{\mathcal{S}^{c}}
\end{bmatrix},
\]
where $\sim$ indicates equivalence under elementary row operations. This step holds because the set of $v$'s satisfying $G^{(i)}v=-e_{1}$ is a coset of the dual code, which is generated by $H$. Since
\[
\tilde{H}_{\mathcal{S}^{c}}\sim\begin{bmatrix}0\\
H_{\mathcal{S}^{c}}
\end{bmatrix}
\]
if and only if $e_{1}\notin\mbox{colspace}\big(\tilde{H}_{\mathcal{S}^{c}}\big)$, we find that $e_{1}\in\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)$ if and only if $e_{1}\notin\mbox{colspace}\big(\tilde{H}_{\mathcal{S}^{c}}\big)$.
Based on the above construction, we see that $G_{\mathcal{S}}^{(i)}$ is an $(m-i)\times d$ submatrix of a random full-rank $(m-i)\times m$ matrix and $\tilde{H}_{\mathcal{S}^{c}}$ is an $(i+1)\times(m-d)$ submatrix of a random full-rank $(i+1)\times m$ matrix. Thus, we find that
\begin{align*}
\rho(m,i,d,q) & =\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(G_{\mathcal{S}}^{(i)}\big)\right)\\
& =1-\mathbb{P}\left(e_{1}\notin\mbox{colspace}\big(\tilde{H}_{\mathcal{S}^{c}}\big)\right)\\
& =1-\rho(m,m-i-1,m-d,q).
\end{align*}
Using this, we also observe that
\begin{align*}
1 & -\overline{\varphi}_{i}(x)=\sum_{d=0}^{m}\binom{m}{d}(1-x)^{d}x^{m-d}\left(1-\rho(m,i,d,q)\right)\\
& =\sum_{d=0}^{m}\binom{m}{d}(1-x)^{d}x^{m-d}\rho(m,m-i-1,m-d,q)\\
& =\sum_{d'=0}^{m}\binom{m}{m-d'}(1-x)^{m-d'}x^{d'}\rho(m,m-i-1,d',q)\\
& =\overline{\varphi}_{m-i-1}(1-x).
\end{align*}
It can also be observed, numerically, that
\[
\sum_{i=0}^{m-1}\rho(m,i,d,q)=m-d.
\]
Using this, one can check explicitly that the average polar transform preserves the mutual information
\begin{align*}
\frac{1}{m}\sum_{i=1}^{m} & \overline{\varphi}_{i}(x)=\frac{1}{m}\sum_{i=1}^{m}\sum_{d=0}^{m}\binom{m}{d}x^{m-d}(1-x)^{d}\rho(m,i,d,q)\\
& =\frac{1}{m}\sum_{d=0}^{m}\binom{m}{d}x^{m-d}(1-x)^{d}(m-d)\\
& =x.
\end{align*}
|
\section{\label{sec:AtmosNu}Introduction }
Sterile neutrinos with masses in the range $\Delta m^2 = 0.1$~${\rm eV}^2 - 10~{\rm eV}^2$ have been posited to explain anomalies in accelerator \cite{Athanassopoulos:1996jb,Aguilar:2001ty,Aguilar-Arevalo:2013pmq}, reactor \cite{Mention:2011rk}, and radioactive source~\cite{Bahcall:1994bq}~oscillation experiments. Several null results \cite{Armbruster:2002mp,Abe:2014gda, Adamson:2011ku, Cheng:2012yy, Dydak:1983zq} restrict the available parameter space of the minimal 3+1 model, which assumes mixing of the three active neutrinos with a single sterile neutrino, resulting in three light and one heavier mass state. Global fits to world data \cite{Giunti:2011gz,Kopp:2013vaa,Conrad:2012qt} demonstrate that there remain regions of allowed parameter space around the best fit point of $\Delta m^2 = 1~{\rm eV}^2$ and $\sin^2 2\theta_{24}=0.1$. A consequence of these models is the existence of $\nu_\mu$ ($\bar{\nu}_\mu$) disappearance signatures, which are yet to be observed.
Atmospheric neutrinos produced in cosmic ray air showers throughout the Earth's atmosphere are detected by IceCube \cite{Halzen:2010yj}. To mitigate the large atmospheric muon background, only up-going neutrinos are selected. For these trajectories, the Earth acts as a filter to remove the charged particle background. At high neutrino energies, the Earth also modifies the neutrino flux due to charged current and neutral current interactions \cite{GonzalezGarcia:2005xw}. At E$_\nu~>~$100~GeV, oscillations due to the known neutrino mass splittings have wavelengths larger than the diameter of the Earth and can be neglected.
A previous measurement of the atmospheric flux in the sub-TeV range, performed by the Super-Kamiokande experiment, found no evidence for anomalous neutrino disappearance \cite{Abe:2014gda}.
This paper reports the first searches for $(\nu_\mu+\overline{\nu}_\mu)$ disappearance in the approximate 320~GeV to 20~TeV range, using two independent analyses each based on one-year data samples from the IceCube detector \cite{Aartsen:2015rwa,Aartsen:2013eka}. In this energy regime, sterile neutrinos would produce distinctive energy-dependent distortions of the measured zenith angle distributions \cite{Nunokawa:2003ep}, caused by resonant matter-enhanced oscillations during neutrino propagation through the Earth.
This MSW resonant effect depletes antineutrinos in 3+1 models (or neutrinos in 1+3) \cite{Nunokawa:2003ep,Abazajian:2012ys}. Additional oscillation effects produced by sterile neutrinos include vacuum-like oscillations at low energies for both neutrinos and antineutrinos, and a modification of the Earth opacity at high energies, as sterile neutrinos are unaffected by matter. These effects would lead to detectable distortions of the flux in energy and angle, henceforth called ``shape effects,'' in IceCube for mass splittings in the range 0.01$~{\rm eV}^2\leq \Delta m^2 \leq~10~{\rm eV}^2$ \cite{Choubey:2007ji, Razzaque:2011ab, Barger:2011rc, Esmaili:2012nz, Razzaque:2012tp,Esmaili:2013vza, Esmaili:2013cja, Lindner:2015iaa}.
\begin{figure}[tbp!]
\includegraphics[width=1.0\columnwidth]{neutrino_disappearance_probability-2-eps-converted-to.pdf}\\
\includegraphics[width=1.0\columnwidth]{antineutrino_disappearance_probability-2-eps-converted-to.pdf}\\
\includegraphics[width=1.0\columnwidth]{Signal0_WBar_NoStripe-eps-converted-to.pdf}\\
\caption{Top and center: change in the spectrum due to propagation effects for muon neutrinos and antineutrinos at the 3+1 global best fit point. Bottom: The predicted event rate reduction (in percent) vs. reconstructed muon energy and zenith angle for this model.
\label{fig:Oscillograms}}
\end{figure}
\section{\label{sec:IceCube}Atmospheric Neutrinos in IceCube}
Having crossed the Earth, a small fraction of up-going atmospheric neutrinos undergo charged current interactions in either bedrock or ice, creating muons that traverse the instrumented ice of IceCube. These produce secondary particles that add Cherenkov light, which can be detected by the Digital Optical Modules (DOMs) \cite{Hanson:2006bk,Abbasi:2008aa,Abbasi:2010vc} of the IceCube array. The full detector contains 5160 DOMs on 86 strings arranged with string-to-string spacing of approximately 125 m and typical vertical DOM separation of 17 m.
The analysis detailed in this paper, referred to as IC86, uses data from the full 86-string detector configuration taken during 2011-2012, with up-going neutrinos selected according to the procedure developed in \cite{Weaver:thesis,Aartsen:2015rwa}. The sample contains 20,145 well-reconstructed muons detected over a live time of 343.7 days. A total of 99.9\% of the detected events in the data sample are expected to be neutrino-induced muon events from the decays of atmospheric pions and kaons. The flux contribution from charmed meson decays was found to be negligible \cite{Aartsen:2014muf,Aartsen:2015rwa}, as was the contamination of up-going astrophysical neutrinos with the spectrum and rate measured by IceCube \cite{Aartsen:2015rwa}. A complementary analysis, referred to as IC59 and discussed later, was performed using a sample of 21,857 events observed in 348.1 days of data taken with an earlier 59-string configuration of the detector from 2009-2010 \cite{Aartsen:2013eka}.
Since muon production is very forward at these energies, the muon preserves the original neutrino direction with a median opening angle following 0.7 degrees$\times$(E$_{\nu}$/TeV)$^{-0.7}$ \cite{Ahrens:2003ix}.
The muon zenith angle can be reconstructed geometrically with a resolution of $\sigma_{\mathrm{cos}(\theta_{z})}$ varying between 0.005 and 0.015 depending on the angle. Because of energy sharing in production and radiative losses outside the detector, the detected muon energy is smeared downward from the original neutrino value. Muon energy is reconstructed based on the stochastic light emission profile along the track \cite{Aartsen:2013vja,Aartsen:2015rwa} with a resolution of $\sigma_{\mathrm{log}_{10} (E_\mu / GeV)}\sim0.5$.
To search for shape effects \cite{Esmaili:2013vza,Esmaili:2012nz,Esmaili:2013cja,Barger:2011rc}, including the MSW and parametric resonances, the analyses compare the predicted observable muon spectrum for a given incident neutrino flux and oscillation hypothesis with data. Flavor evolution in the active and sterile neutrino system can be calculated by numerical solution of a master equation \cite{GonzalezGarcia:2005xw,Delgado:2014kpa}. For IC86, this calculation is performed using the $\nu$-\texttt{SQuIDs} software package \cite{squids,nusquids}, while the IC59 analysis approximates the oscillation probability by solving a Schr\"{o}dinger-like equation using the NuCraft package~\cite{Wallraff:2014qka}. This approximation is accurate to better than 10\% below $\Delta m^2 \approx 5~{\rm eV}^2$, where Earth-absorption effects can be neglected. Fig.~\ref{fig:Oscillograms} (top and center) shows the $\nu_\mu$ and $\bar\nu_\mu$ oscillation probability vs. true energy and zenith angle, calculated at the best-fit point from \cite{Conrad:2012qt}. Since IceCube has no sign-selection capability, the reconstructed samples contain both $\mu^+$ and $\mu^-$ events. For illustration, Fig. \ref{fig:Oscillograms} (bottom) shows the predicted depletion of events for the global 3+1 best fit point in the distribution of reconstructed variables from the IC86 analysis; in this case the large depletion is dominated by the parametric resonance.
\section{\label{sec:Systematics}Data Analysis and Systematic Uncertainties}
To search for sterile neutrino oscillations we calculate the negative of a binned Poissonian log-likelihood (LLH) for the data given each sterile neutrino hypothesis on a fine grid in the the $\left[ \mathrm {log}(\Delta m^2), \mathrm{log}(\mathrm{sin}^22\theta_{24}) \right]$ hypothesis space. In the IC86 analysis, the data are histogrammed on a grid with 10 bins in energy ranging from 400~GeV to 20~TeV, and 21 linearly spaced bins starting at $\mathrm{cos}(\theta) = 0.24$ with a spacing of 0.06. The bins were chosen a priori guided by experimental resolution, scale of the disappearance signatures and accumulated MC simulation statistics. The LLH values are compared to the minimum in the space to produce unified confidence intervals \cite{Feldman:1997qc}.
Systematic uncertainties are treated by introducing both continuous and discrete nuisance parameters, which are fitted at each hypothesis point. The list of systematic uncertainties considered is given in Table \ref{tbl:systematics} and discussed below. More information can be found in \cite{BenThesis} and \cite{CarlosThesis}.
\begin{table}[h]
\begin{tabular}{ l c c }
\hline
\multicolumn{3}{c}{\it Atmospheric flux} \\ \hline \hline
$\nu$ flux template & discrete (7) & \\
$\nu$ / $\overline{\nu}$ ratio & continuous & 0.025\\
$\pi$ / K ratio & continuous & 0.1\\
Normalization & continuous & none$^1$ \\
Cosmic ray spectral index & continuous & 0.05 \\
Atmospheric temperature & continuous & model tuned\\
\hline \hline
\multicolumn{3}{c}{\it Detector and ice model} \\ \hline \hline
DOM efficiency & continuous & \\
Ice properties & discrete (4) & \\
Hole ice effect on angular response & discrete (2)\\ \hline
\multicolumn{3}{c}{\it Neutrino propagation and interaction} \\ \hline \hline
DIS cross section & discrete (6) & \\
Earth density & discrete (9) & \\
\hline \hline
\end{tabular}
\caption{List of systematic uncertainties considered in the analysis. The numbers in parentheses show the number of discrete variants used. Full descriptions are given in the text. The third column indicates the gaussian width of a prior if introduced for the parameter in the analysis (see \cite{BenThesis} for details). $^1$A prior of 40\% was applied to the Normalization parameter in the rate+shape analysis described below.}
\label{tbl:systematics}
\end{table}
\subsection{Atmospheric neutrino flux uncertainties}
The atmospheric flux in the energies relevant to this analysis is dominated by the neutrinos that originate from pion and kaon decays in cosmic ray showers. This prompts us to parametrize the atmospheric flux as
\begin{equation}
\phi_{\rm atm}(\cos\theta) = N_0 \mathcal{F}(\delta)\bigg( \phi_\pi + R_{\pi/K} \phi_K \bigg) \left(\frac{E_\nu}{E_0}\right)^{- \Delta \gamma}
\label{eq:atmospheric-parametrization}
\end{equation}
\noindent (and similarly for antineutrinos, with a relative flux normalization uncertainty). The free nuisance parameters are the overall flux normalization $N_0$, the correction to the ratio of kaon- to pion-induced fluxes $R_{K/\pi}$ and the spectral index correction $\Delta \gamma$. The $\phi_\pi$ and $\phi_K$ are the spectrum of atmospheric neutrinos originating from $\pi$ and $K$ decays, respectively. Furthermore, $\Delta \gamma$ allows us to take into account uncertainties in the spectral index of the flux. The term $E_0$ is a pivot point near the median of the energy distribution which renders the $\Delta \gamma$ correction approximately normalization-conserving.
Here, seven $\phi_k$ and $\phi_\pi$ variants are used to encapsulate additional hadronic model uncertainty and the primary cosmic ray model uncertainties. Atmospheric density uncertainties are a subleading effect. We thus parametrize it as a linear function, $\mathcal{F}(\delta)$, which is obtained by fitting fluxes calculated with different atmospheric profiles generated within constraints imposed by temperature data from the AIRS satellite \cite{AIRS}.
The central flux prediction for the analysis is the HKKM model with H3a knee correction \cite{Sanuki:2006yd,Honda:2006qj,Gaisser:2013bla}. Additional flux variants are calculated using the analytic air shower evolution code of \cite{Fedynitch:2015zma,MCeq,Collins2015}. The cosmic spectrum variants considered are the Gaisser-Hillas \cite{Gaisser:2013bla}, Zatsepin-Sokolskaya \cite{Zatsepin:2006ci}, and Poly-gonato models \cite{Hoerandel:2002yg}. The hadronic models considered are QGSJET-II-4 \cite{Ostapchenko:2010vb} and SIBYILL2.3 \cite{Riehn:2015oba}. For each combination of hadronic and primary model, fluxes calculated in various atmospheric density profiles are used to derive the $\mathcal{F}(\delta)$ parameterization.
\subsection{Neutrino propagation and interaction uncertainties}
Two sets of neutrino propagation uncertainties are treated in the search. Neutrino oscillation and absorption effects both depend on the Earth density profile along the neutrino trajectory, which is parameterized by the PREM model \cite{ref:PREM}. Uncertainties in the Earth composition and density are accounted for by creating perturbations of the PREM and re-propagating the neutrino flux. The PREM variants are constructed under the constraints that the Earth mass and moment of inertia are preserved, that the density gradient is always negative in the core and mantle regions, and that the local perturbation is never more than 10\%. The effects of Earth model uncertainty on the final propagated neutrino spectrum are incorporated by minimizing over 9 discrete perturbed models.
A further propagation uncertainty is the neutrino charged-current cross-section that, at these energies, is dominated by deep inelastic scattering (DIS). The uncertainty in the cross-sections arises from parton distribution function (PDF) uncertainties. A parametrization of the cross-section uncertainty uses calculations \cite{CooperSarkar:2011pa} (see also \cite{ref:CarlosXS}) based on three different PDF sets: HERAPDF \cite{Aaron:2009aa}, CT10 \cite{Gao:2013xoa} and NNPDF \cite{,Nocera:2014gqa}. In each case, simulated neutrino interactions are re-weighted using true neutrino energy and inelasticity given calculated doubly-differential cross sections, and the analysis fit is run using the weighted sample.
\subsection{Detector and ice uncertainties}
The absolute optical module photon collection efficiency, $\epsilon$, has been measured in the laboratory \cite{Abbasi:2010vc}. However, shadowing by the DOM cable and unknown local optical conditions after deployment introduce an uncertainty in the optical efficiency {\it in situ}, leading to uncertainty in the detected energy and angular event distribution. Here $\epsilon$ is treated as a continuous nuisance parameter and re-weighting techniques are used to correct Monte Carlo distributions to arbitrary values. We follow the method developed in \cite{Weaver:thesis,Aartsen:2015rwa}, implementing a penalized spline \cite{Whitehorn:2013nh} fitted to Monte Carlo datasets generated at various DOM efficiency values. Variability of the optical efficiency induces changes in the detector energy scale. In practice, the best fit value is tightly constrained by the position of the energy peak in the final sample.
The IceCube ice model applied in this analysis has nearly a thousand free parameters that are minimized in an iterative fit procedure using light-emitting-diode (LED) flasher data \cite{Aartsen:2013rt}. The model implements vertically varying absorption and scattering coefficients across tilted isochronal ice layers. The fit procedure yields a systematic and statistical uncertainty on the optical scattering and absorption coefficients in the ice, as well as a larger uncertainty on the amount of light deposited by the LED flashers. This larger uncertainty was later reduced by introducing azimuthal anisotropy in the scattering length into the ice model, which may result from dust grain shear due to glacial flow \cite{Aartsen:2013ola}. We use the model described in \cite{Aartsen:2013rt} as the central ice model, and then use the model with anisotropy \cite{Aartsen:2013ola} as an alternative to assess the impact of this effect. We also incorporate models with 10\% variations in the optical absorption and scattering coefficients to account for the uncertainty on those parameters. A full Monte Carlo sample is created for each model variation.
The ice column immediately surrounding the DOMs has different optical properties than the bulk ice due to dissolved gases that are trapped during the refreezing process following DOM deployment. This introduces additional scattering near the DOM and has a nontrivial effect on its angular response \cite{Aartsen:2013rt}. To quantify this effect on the final event distribution, a comparison is made between the extreme case of the DOM assumed to have its laboratory-derived angular response vs.\ the nominal hole ice model as discrete ice model variants.
\section{\label{sec:Results}Results}
The analysis detailed here was developed with 90\% of the data sample held blind, and unblinding was a multi-step process. The agreement of Monte Carlo (MC) simulations based on the no-steriles hypothesis (corresponding to more than 360 years of simulated data) with data was evaluated using one-dimensional energy and zenith angle distributions, which would wash out the resonance signature of sterile neutrinos (Fig.~\ref{fig:PreUnblinding}). Good data-MC consistency was observed and no nuisance parameter was found to have a significant pull outside of its prior. Other comparisons, insensitive to the sterile neutrino signature, were made by examining subsets of the data split by reconstructed azimuthal track angle, and by event center-of-gravity. No significant data-Monte Carlo disagreements were observed. The full event distribution in the two-dimensional analysis space, and the pulls-per-bin from the null hypothesis (Fig.~\ref{fig:Pulls}) were then examined. Event-by-event reconstructed data and Monte Carlo can be found in \cite{datarelease}.
\begin{figure}[t]
\includegraphics[width=0.90\columnwidth]{EnergyDist-2.png} \\
\caption{\label{fig:PreUnblinding}Reconstructed energy distribution in data and Monte Carlo for the no-steriles hypothesis in the analysis.}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.99\columnwidth]{PullsPerBin_WBar_NoStripe.png}\\
\caption{\label{fig:Pulls}The statistical-only pulls (shape+rate analysis) per reconstructed energy and zenith angle bin at the best nuisance parameter fit point for the no-sterile hypothesis. The shown empty bins are those that were evaluated in the analysis but had no data events remaining following cuts.}
\end{figure}
The LLH value for the data given each sterile neutrino hypothesis was calculated. No evidence for sterile neutrinos was observed. The best fit of the blind, shape-only analysis is at $\Delta m^2$ = 10 eV$^2$ and $\mathrm{sin}^2 2\theta_{24}$=0.56 with a log likelihood difference from the no-steriles hypothesis of $\Delta \rm{LLH}$=1.91, corresponding to a p-value of 15\%. Since the fit does not constrain flux normalization, LLH minima at $\Delta m^2 \gtrsim 5~{\rm eV^2}$ are highly degenerate with the no-sterile hypothesis. This is because the oscillation effect becomes a fast vacuum-like oscillation smeared out by the energy resolution of the detector, and thus changes the normalization but has no effect on shape.
Post-unblinding tests highlighted two undesirable features of the shape-only analysis, both deriving from the degeneracy between high-$\Delta m^2$, fast oscillation hypotheses and changes in the flux normalization. First, because the high-$\Delta m^2$ space is not penalized by any prior, a log likelihood minimum in this region may not be uniquely defined under extensions of the search space. In some cases, slightly stronger exclusion limits can be found by increasing the search space to higher mass. Second, the degeneracy between normalization and mixing can lead to
unphysical values for the normalization that compensate for the sterile neutrino oscillation effect. To avoid these ambiguities, an extension of the analysis (denoted rate+shape) was developed to constrain the neutrino flux normalization using a prior with 40\% uncertainty in the likelihood function, based on \cite{Fedynitch:2012fs,Honda:2006qj}. This results in a weakened exclusion relative to the blind analysis proposal. However, since it is more robust, we consider it our primary result. For the rate+shape analysis, the best fit is at $\Delta m^2=10$~eV$^2$ and sin$^2 2\theta_{24}$=0.50, with a log likelihood difference from the no-steriles hypothesis of $\Delta$LLH=0.75, corresponding to a p-value of 47\%. This minimum is unique under extension
of the analysis space to higher masses, since the large $\Delta m^2$ region is no longer degenerate with the no-sterile hypothesis. This was checked over an extended parameter space up to $\Delta m^2$=100 ${\rm eV}^2$. The confidence interval for the shape-only and the rate+shape analyses are shown in Fig.~\ref{fig:IceCubeResults}.
A number of checks of the rate+shape analysis result were made (see \cite{BenThesis}). The exclusion is found to be robust under tightening or loosening the nuisance parameter priors by a factor of two. Different strengths of the normalization constraint were tested, and the result was found to be relatively insensitive to values between 30\% and 50\%.The pulls on each continuous nuisance parameter were evaluated at all points in the LLH space and found to behave as expected. The contour was redrawn for each discrete nuisance variant and found to have good stability. The Wilks confidence intervals \cite{Agashe:2014kda} were validated using Feldman-Cousins ensembles along the contour \cite{Feldman:1997qc} and found to be accurate frequentist confidence intervals.
An independent search was conducted using the 59-string IceCube data \cite{Aartsen:2015cwa,MariusThesis}, introduced previously, that also finds no evidence of sterile neutrinos. The IC59 analysis, described in detail in \cite{Aartsen:2013eka}, used different treatments for the systematic uncertainties, for the fitting methods and employed independent Monte Carlo samples that were compared to data using unique weighting methods. In particular, the event selection used for this data set had higher efficiency for low-energy neutrinos, using a threshold at 320 GeV, extending the sensitivity of the analysis to smaller $\Delta m^2$. However, detailed {\it a posteriori} inspections revealed that a background contamination from cosmic ray induced muons, on the level of 0.3\% of the full sample, is largest in this region and could lead to an artificially strong exclusion limit. Furthermore, the energy reconstruction algorithm used in both analyses, which measures the level of bremsstrahlung and other stochastic light emission along the muon track, is vulnerable to subtle detector modeling issues and suffers degraded energy resolution in the low-energy region where most muons are minimum-ionizing tracks and a large fraction either start or stop within the detector. It was therefore decided to exclude these events to avoid biasing the resulting exclusion regions. As a result of this {\it a posteriori} change, the IC59 analysis retains a comparable range of sensitivity in $\Delta m^2$ but the reach in $\rm{sin}^2$$\theta_{24}$ is strongly reduced (see Fig.~\ref{fig:IceCubeResults}). However, we still present this result as it independently confirms the result presented here.
\begin{figure}[h!]
\includegraphics[width=0.99\columnwidth]{AllContours99CL-4-eps-converted-to.pdf}
\caption{\label{fig:IceCubeResults}Results from IceCube sterile neutrino searches (regions to the right of the contours are excluded). The dot-dashed blue line shows the result of the original analysis based on shape alone, while the solid red line shows the final result with a normalization prior included to prevent degeneracies between the no-steriles hypothesis and sterile neutrinos with masses outside the range of sensitivity. The dashed black line is the exclusion range derived from an independent analysis of data from the 59-string IceCube configuration.
}
\end{figure}
\section{\label{sec:Conclusions}Discussion and Conclusion}
Resonant oscillations due to matter effects would produce distinctive signatures of sterile neutrinos in the large set of high energy atmospheric neutrino data recorded by the IceCube Neutrino Observatory. The IceCube collaboration has performed searches for sterile neutrinos with $\Delta m^2$ between $0.1$ and $10 ~{\rm eV^2}$. We have assumed a minimal set of flavor mixing parameters in which only $\theta_{24}$ is non-zero.
A nonzero value for $\theta_{34}$ would change the shape of the MSW resonance while increasing the total size of the disappearance signal \cite{Esmaili:2013vza}. As discussed in \cite{Lindner:2015iaa}, among the allowed values of $\theta_{34}$ \cite{Adamson:2011ku}, the model with $\theta_{34}$=0 presented here leads to the most conservative exclusion in $\theta_{24}$. The angle $\theta_{14}$ is tightly constrained by electron neutrino disappearance measurements \cite{Kopp:2013vaa}, and nonzero values of $\theta_{14}$ within the allowed range do not strongly affect our result.
Figure~\ref{fig:RateShapeExcl} shows the current IceCube results at 90\% and 99\% confidence levels, with predicted sensitivities, compared with 90\% confidence level exclusions from previous disappearance searches \cite{Abe:2014gda, Adamson:2011ku, Cheng:2012yy, Dydak:1983zq}. Our exclusion contour is essentially contained within the expected +/- 95\% range around the projected sensitivity derived from simulated experiments, assuming a no-steriles hypothesis. In any single realization of the experiment, deviations from the mean sensitivity are expected due to statistical fluctuations in the data and, to a considerably lesser extent, in the Monte Carlo data sets. Also shown is the 99\% allowed region from a fit to the short baseline appearance experiments, including LSND and MiniBooNE, from \cite{Kopp:2013vaa, Esmaili:2013vza, Conrad:2012qt}, projected with $|U_{e4}|^2$ fixed to its world best fit value according to global fit analyses \cite{Kopp:2013vaa,Conrad:2012qt,Collin:2016rao}. This region is excluded at approximately the 99\% confidence level, further increasing tension with the short baseline anomalies, and removing much of the remaining parameter space of the 3+1 model.
We note that the methods developed for the IC59 and IC86 analyses are being applied to additional data sets, including several years of data already recorded by IceCube, from which we anticipate improvements in IceCube’s sterile neutrino sensitivity.
\begin{figure}[tbp!]
\includegraphics[width=0.99\columnwidth]{FinalResult_plot_v3-eps-converted-to.pdf}\\
\caption{\label{fig:RateShapeExcl}Results from the IceCube search. (Top) The 90\% (orange solid line) CL contour is shown with bands containing 68\% (green) and 95\% (yellow) of the 90\% contours in simulated pseudo-experiments, respectively. (Bottom) The 99\% (red solid line) CL contour is shown with bands containing 68\% (green) and 95\% (yellow) of the 99\% contours in simulated pseudo-experiments, respectively. The contours and bands are overlaid on 90\% CL exclusions from previous experiments \cite{Abe:2014gda, Adamson:2011ku, Cheng:2012yy, Dydak:1983zq}, and the 99\% CL allowed region from global fits to appearance experiments including MiniBooNE and LSND, assuming $|U_{e4}|^2=0.023$ \cite{Kopp:2013vaa} and $|U_{e4}|^2=0.027$ \cite{Conrad:2012qt} respectively.}
\end{figure}
\input{sterileacknowledgements}
Note added: Recently, an analysis using IceCube public data \cite{weaverdatarelease} was performed \cite{Liao:2016reh}. Though this independent analysis has a limited treatment of systematics, it follows the technique described here and in refs. \cite{BenThesis, CarlosThesis}, and obtains comparable bounds. To allow for better reproduction of the result shown in this paper in the future, we have put forward a data release that incorporates detector systematics \cite{datarelease}.
|
\section*{Introduction}
Active systems are collections of
agents that convert the energy of the environment in systematic movement \cite{Vicsek12,Marchetti13,Cates12}.
Examples include bacterial colonies \cite{Peruani12}, epithelial cell layers \cite{Bi15},
self-propelled colloids \cite{Palacci10}, swimming microorganisms \cite{Berg04}, schools of fish \cite{Hemeltijk08}
and bird flocks \cite{Ballerini08}.
Active particles can form gas, liquid, liquid crystal or glassy phases with structural properties
remarkably similar to those of ordinary materials \cite{Marchetti15,Henkes11,Render13,Berthier14,Berthier13,Tailleur08,Marchetti15,Zhou14}.
Active systems are, however, out-of-equilibrium. Hence their steady state is not described by the Boltzmann distribution
and they can support spontaneous, self-sustained motion, which can in turn
be enhanced, stabilized or suppressed by suitably designed confining geometries \cite{Hol14,Sanchez12,wioland,Galajda14}.
It has been shown that active agents can give rise to ratchet effects, \cite{Galajda07,Wan08,Angelani11,Reichhardt16} power microgears \cite{Angelani09,DiLeonardo10,Sokolov10}, drive spontaneous accumulations of passive colloids over target regions \cite{Koumakis13}, and exhibit
long lived density fluctuations \cite{Narayan07}.
From a theoretical point of view, the effect of confinement
has been used to investigate the concept
of pressure in active systems \cite{Solon15,Solon15b,Yang14,Takatori14}
and the effect of wall curvature on both active particles
\cite{Smallenburg15} and passive tracers \cite{Mallory14}.
Strong confinement can induce hysteretic dynamics \cite{Fily15} or sustained spontaneous density oscillations \cite{Paoluzzi15}.
The role of curved walls on active gas has been investigated in Ref.\cite{Fily14}.
Previous work has focused on confinement by rigid walls.
While recent studies have investigated the effects of active baths on flexible open chains \cite{har2014,kai2014,Shin15},
the case of swimmers confined by deformable boundaries
has recently been analyzed only for case of spherical active Brownian particles by Tian {\itshape et al.} \cite{Tian15}.
An interesting
example of active colloidal cell driven by micro rotators has been theoretically investigated in \cite{Spellings15}.
Here we consider an active vesicle in two dimensions composed
by a flexible one dimensional membrane enclosing active particles representing an active solute.
The corresponding equilibrium system would be a vesicle filled with a suspension and bounded by a flexible membrane that is permeable to the solvent but not to the solute molecules. In this case, the solute concentration would be uniform throughout the vesicle interior and exert a homogeneous pressure on the membrane whose equilibrium configuration would be spherical, or circular in two dimensions.
When the solute molecules are active particles or microswimmers, we find that only for high densities of active particles the membrane shape fluctuates around a circle. When the swimmers packing fraction falls below a characteristic value, depending on particles shape, the vesicle acquires an asymmetric shape characterized by a bimodal distribution of the local curvatures, with a high curvature peak and a near zero curvature component. This effect is driven by a feedback mechanism coupling swimmers density and membrane curvature through local pressure. A local fluctuation of particle density produces a local pressure increase that induces a larger curvature on the flexible membrane. Since active particles tend to accumulate at concave boundaries, this local curvature increase drives further accumulation of swimmers, which in turn raises the local pressure.
The presence of this feedback mechanism is confirmed by a strong correlation between
the local swimmers density (or local pressure on the membrane) and the local curvature of the membrane.
Finally, we examine the center of mass dynamics of the whole vesicle and show that
it performs a persistent random walk with a long time diffusivity that is larger for elongated swimmers due to orientational correlations. Interesting, the resulting migratory behavior shares some similarities
with Eukaryotic directed cell migration \cite{Swaney10,Devreotes03}.
\section*{Methods}
We perform two dimensional simulations of $N_s$ {\it run-and-tumble}
swimmers of width $a$ and length $\ell$ (aspect ratio $\alpha=a/ \ell$)
confined by a deformable membrane.
We specifically consider swimmers of two different aspect ratios, $\alpha=1/2$ (elongated) and $\alpha=1$
(spherical).
\begin{figure}[!t]
\centering
\includegraphics[width=.47\textwidth]{cart}
\caption{ {\bf Pictorial representation of the swimmers.}
Each swimmer consists of $p$ spherically symmetric force centers aligned along the swimming direction $\mathbf{e}_i$, with $p\!=\!1$ describing spherical particles
(panel a) and $p\!=\!2$ elongated ones (panel b).}
\label{fig:mod}
\end{figure}
\subsection*{Swimmers}
We consider $N_s$ run-and-tumble particles in two dimensions.
The model is the same used in \cite{Angelani09,Angelani11c,Angelani10c,Paoluzzi13,Paoluzzi14}
to describe {\itshape E. coli} bacterial suspensions.
Each particle consists of a chain of
$p$ rigidly connected disks of diameter $\ell/2$ aligned along the swimming direction ${\hat{\mathbf e}}_i$.
We denote by $\mathbf{r}_i$ the center of mass of the $i$th
swimmer. The position $\mathbf{r}_i^\beta$, with $\beta=1,...,p$, of the $\beta$-th disk on the $i$-th swimmer is then
\begin{equation}
\mathbf{r}_i^\beta = \mathbf{r}_i + \delta {\mathbf r}^\beta_i \, \, .
\end{equation}
Here we consider $p=1$, corresponding to spherical swimmers with $\delta {\mathbf r}_i^1=\mathbf{0}$, and $p=2$, corresponding to elongated swimmers, with
$\delta {\mathbf r}_i^1=-{\hat{\mathbf e}}_i \ell/4$ and $\delta {\mathbf r}_i^2=+ {\hat{\mathbf e}}_i \ell/4$ (panel (b) in Fig. \ref{fig:mod}).
We assume swimmers interact only through steric repulsion and that the interaction potential is written as the sum of radially symmetric potentials centered at each disk. For this reasons the individual disks that compose our swimmers are also referred to as force centers.
At low Reynolds number, the equations of motion of the $i$-th swimmer are
\begin{eqnarray}\label{eqm}
\mathbf{v}_i &=& \mathbf{M}_i \, \cdot\mathbf{F}_i \;,\\ \nonumber
\boldsymbol{\omega}_i &=&\mathbf{K}_i \,\cdot \mathbf{T}_i \;,
\end{eqnarray}
where $\mathbf{v}_i$ is the center of the mass velocity and $\boldsymbol{\omega}_i$ the angular velocity of the $i-$th swimmer. $\mathbf{M}_i$ and $\mathbf{K}_i$ are the translational and rotational mobility matrices
\begin{eqnarray}
\mathbf{M}_i &=& m_{\parallel} \mathbf{\hat{e}}_i \otimes \mathbf{\hat{e}}_i + m_{\perp} \left( \mathbf{1} - \mathbf{\hat{e}}_i \otimes \mathbf{\hat{e}}_i \right) \\ \nonumber
\mathbf{K}_i &=& k_{\perp} \left( \mathbf{1} - \mathbf{\hat{e}}_i \otimes \mathbf{\hat{e}}_i \right) \, ,
\end{eqnarray}
the symbol $\otimes$ is the dyadic product and $\mathbf{1}$ the identity matrix.
In Eq. (\ref{eqm}), $\mathbf{F}_i$ and $\mathbf{T}_i$ are the total force and the
total torque acting on the of the $i$-th swimmer, given by
\begin{eqnarray}\label{fandt}
\mathbf{F}_i &=& f_0 \mathbf{\hat{e}}_i(1 - \sigma_i) + \sum_{j \neq i, \alpha, \beta} \mathbf{f}(\mathbf{r}_i^\alpha - \mathbf{r}_j^\beta) + \sum_{\alpha} \mathbf{f}_{ext}({\mathbf r}_i^\alpha)\\ \nonumber
\mathbf{T}_i &=& \mathbf{t}_r^i \sigma_i + \sum_{j \neq i, \alpha, \beta} \delta {\mathbf r}^\alpha_i \times \mathbf{f}(\mathbf{r}_i^\alpha - \mathbf{r}_j^\beta) + \sum_{\alpha} \delta {\mathbf r}^\alpha_i \times \mathbf{f}_{ext}({\mathbf r}_i^\alpha) \, .
\end{eqnarray}
The index $j=1,\dots,N_s$ runs over swimmers, the indices $\alpha=1,..,p$ and $\beta=1,...,p$ run over disks, and
$\sigma_i$ is a state variable, with value $0$ for running swimmers and $1$ for tumbling ones.
During the running state each swimmer
is self-propelled along ${\hat{\mathbf e}}_i$ with self-propulsion speed $v=m_{\parallel} f_0$.
In the tumbling state, the random torque $\mathbf{t}_r^i$ rotates
the swimming direction ${\hat{\mathbf e}}_i$
at the tumbling rate, $\lambda$.
Moreover, it takes a finite time $(\lambda 10)^{-1}$ for the swimmers
to reorient the swimming direction. The external force $\sum_{\alpha} \mathbf{f}_{ext}({\mathbf r}_i^\alpha)$ in Eq. (\ref{fandt})
represent the interaction of the swimmers with the flexible confining boundary. The details of this interaction will be specified in the next section.
Finally, the repulsive force $\mathbf{f}(\mathbf{r})$ is conservative and generated by the potential $V(r) = \frac{a f_0}{12} \ \left( \frac{a}{r} \right)^{12}$\cite{Angelani09}.
Below we choose units such that $\ell=m_\parallel=f_0=1$ and use $\lambda=0.1$,
$k_{\perp}=4.8$ and $m_{\perp}=0.87$.
\subsection*{Membrane}
The bounding membrane is modeled as a ring of $N_b$ colloidal beads of diameter $a$
connected by springs.
Denoting with ${\mathbf R}_n$ the position of $n$-th bead,
the equation of motion of the membrane in the low Reynolds number regime is given by
\begin{eqnarray}\label{motion}
\dot{{\mathbf R}}_n &=& \mu_b {\mathbf F}_n\;,\hspace{0.2in} {\mathbf F}_n = -\mbox{\boldmath$\nabla$}_n \varphi(\{{\mathbf R}\} ,\{{\mathbf r}\} )\;,
\end{eqnarray}
where the potential $\varphi(\{{\mathbf R}\} ,\{{\mathbf r}\} )$ consists of harmonic and repulsive parts, $\varphi(\{{\mathbf R}\} ,\{{\mathbf r}\} ) = \varphi(\{{\mathbf R}\})^{harm}+\varphi(\{{\mathbf R}\} ,\{{\mathbf r}\} )^{rep}$, with
\begin{eqnarray}
&&\varphi(\{{\mathbf R}\})^{harm} = \frac{k}{2}\sum_{n=1}^{N_b}\left( | {\mathbf R}_{n+1} - {\mathbf R}_n | - a
\right)^2 \label{phiarm} \, ,\\ \nonumber
&&\varphi(\{{\mathbf R}\} ,\{{\mathbf r}\} )^{rep} = \sum_{n<m} V(|{\mathbf R}_n - {\mathbf R}_m|) + \sum_{n,i,\beta} V(|{\mathbf R}_n -{\mathbf r}_i^\beta|)\, ,
\end{eqnarray}
where ${\mathbf R}_{N_b+1}={\mathbf R}_1$ in the sum in Eq.~(\ref{phiarm}).
We choose $\mu_b\!=\!\mu$, $k=5\cdot10^2$.
The external force in Eq. (\ref{fandt}) is $\mathbf{f}_{ext}(\mathbf{r}_\alpha^i)=-\frac{\partial\varphi(\{{\mathbf R}\},\{{\mathbf r}\})^{rep}}{\partial\mathbf{r}_\alpha^i}$.
The initial configuration of the membrane is a circle of radius $R_0\!=\!a (2\sin(\pi/N_b))^{-1}$ and area $A_{ref}\!=\!\pi R_0^2$.
The swimmers cover a fraction $\phi\!=\!N_s a_{swim} / A_{ref}$ of the initial area of the vescicle, with
$a_{swim}\!=\!p \pi (a/2)^2$ the area of one swimmer.
The entire vesicle moves in a two dimensional box of side $70\ \ell$ with periodic boundary conditions.
We have performed numerical simulations of membranes composed of $N_b=50, 100, 150$ beads enclosing $N_s$
elongated swimmers ($p=2$) with packing fraction from $\phi=0.07$ up to $0.83$
and $N_s$ spherical swimmer ($p=1$) with packing fraction from $\phi=0.05$ to $0.82$. Specifically, in the case of elongated swimmers we have simulated systems with $N_s=12,21,32,37,52,69,80$ for $N_b=50$, $N_s=52,69,80,97,112,137,156,208,225,256,316,384,421,448$ for $N_b=100$, and $N_s=80,112,156,208,256,316$, $384,448,540,616,716,812,973$ for $N_b=150$. For spherical swimmers we have used
$N_s=12,21,32,52,80,112,156,208$ for $N_b=50$, $N_s=52,112,208,316,448,616,812$ for $N_b=100$, and
$N_s=316,448,616,812,1020,1264$ for $N_b=150$.
To quantify the shape of the membrane we measure the gyration tensor ${\mathbf Q}$, given by
\begin{equation}
{\mathbf Q} = \frac{1}{N_b}\sum_{n=1}^{N_b} \left( {\mathbf R}_n - {\mathbf R}_{cm}\right)\otimes \left( {\mathbf R}_{ n}- {\mathbf R}_{ cm}\right)\;,
\end{equation}
with ${\mathbf R}_{cm}$ the center of the mass of membrane beads.
From the average values of the two eigenvalues $\lambda_1$ and $\lambda_2$ of ${\mathbf Q}$ we compute the squared radius of gyration $R_g^2=Tr \, {\mathbf Q}$ that gives a measure of the extension of the cell,
\begin{equation}
R_g^2=\lambda_1+\lambda_2 \;,
\end{equation}
and the asphericity~\cite{Aronovitz86}
\begin{equation}
\Delta = \frac{\left( \lambda_1 - \lambda_2 \right)^2}{\left( \lambda_1 + \lambda_2 \right)^2}\, .
\end{equation}
The value $\Delta=0$ corresponds to a circle and $\Delta=1$ to a rod.
Since the gyration tensor is a dynamical quantity, the observables
$R_g$ and $\Delta$ are computed from the time average of the eigenvalues.
We characterize the local shape of the membrane
by measuring the local curvature, $\kappa$,
defined as \cite{Sokolov01}
\begin{equation}
\kappa = \frac{({\mathbf R}^{\prime} \times {\mathbf R}^{\prime\prime})_z}{|{\mathbf R}^{\prime}|^3} \;,
\end{equation}
where ${\mathbf R}$ is the vector position of a membrane point,
${\mathbf R}^{\prime}$ and ${\mathbf R}^{\prime\prime}$ are the first and second derivatives of ${\mathbf R}$ with respect to the membrane contour length.
Curvature values are evaluated at the beads position along the membrane, using discrete form of the derivatives.
To evaluate the pressure $P$ on the
$n-$th bead, we have computed numerically the total force that swimmers exert along the local normal
${\hat{\mathbf n}}_n$ to the membrane divided by the average length of the segments connecting such a bead to its neighbors.
\section*{Results}
It is well established in the literature that confined active particles
tend to accumulate along the confining walls~\cite{Tailleur08,Wensink08,Yang14}.
In our case the confining walls are flexible
and swimmer accumulation induces strong distortions of the bounding membrane.
These distortions are evident in the
snapshots shown in Fig. \ref{fig:fig1} where elongated swimmers are bounded by a membrane of $N_b=100$ beads.
For low packing fraction (left panel) the membrane
explores different shapes characterized by regions of high curvature.
As the number of swimmers is increased (right panel), the imbalance of particles along the
flexible walls becomes less dramatic and the vesicle assumes more symmetric shapes.
\begin{figure*}[!t]
\centering
\includegraphics[width=1.\textwidth]{fig1}
\caption{
{\bf Shape fluctuations.} The bounding membrane is composed of
$N_b=100$ beads. Left panel: snapshots of vesicle shapes explored by the active vesicle for low packing fraction of elongated swimmers ($\phi=0.16$). Right panel: the vesicle becomes more symmetric as the number of active particles increases,
$\phi$ is $0.16$ (a), $0.31$ (b), $0.51$ (c) and $0.76$ (d).
}
\label{fig:fig1}
\end{figure*}
\subsection*{Pressure and global shape properties}
We first discuss the case of spherical swimmers ($p=1$, aspect ratio $\alpha=1$).
In this case particle reorientations are solely due to tumbles and no aligning interactions exists between swimmers or swimmers and walls.
To quantify the deviations of the active vesicle from circular shape we display in Fig.~\ref{fig:fig2}-a the asphericity $\Delta$ for different values of $N_b$ as a function of the swimmers area fraction.
We find that $\Delta$ rapidly decays to zero with increasing $\phi$ especially for large vesicles ($N_b=150$), indicating
that at high density of swimmers the active vesicle approaches an average circular shape.
In contrast, we observe deviation from a circular shape for small vesicles in the dilute regime.
Now we quantify the membrane stretching for $N_b=150$ (in this case $\Delta\sim0$ in the whole $\phi$ range explored).
We show in Fig.~\ref{fig:fig2}-(b) that the gyration radius,
$R_g$ increases with $N_s$. This is true for all vesicle sizes ($N_b=50,100$ not shown in figure), indicating that the active particles exert a pressure that stretches the bounding membrane.
A simple estimate for the dependence of $R_g$ on swimmer packing fraction can be obtained for a dilute gas of spherical {\itshape run-and-tumble} swimmers. In two dimensions
the pressure of an ideal active gas of $N_s$ spherical swimmers in an area $A$ is the
so-called swim pressure \cite{Takatori14,Yang14}, given by
\begin{equation}\label{swim}
P_{swim} =\frac{N_s}{A} \frac{v^2}{2 \lambda \mu}=\frac{\phi}{a_{swim}} \frac{v^2}{2 \lambda \mu} \, ,
\end{equation}
where we have expressed $P_{swim}$ in terms of the initial packing fraction $\phi$.
In presence of confining structures
the pressure in the bulk is strongly affected
by the finite size effects \cite{Yang14,Solon15b,Marchetti15,Maggi15}.
For example, in the case of one dimensional gas of {\itshape run-and-tumble} particles confined in a box of side $L$ we can write\cite{Maggi15}:
\begin{equation}\label{conf}
P_{ 1d-box}(L)= \frac{P_{swim}}{1+2 v/ \lambda L} \,,
\end{equation}
We assume that the internal pressure is responsible of an isotropic deformation of the vesicle from a circle
of radius $R_0$ to a circle of radius $R_g$.
In the dilute regime, we assume that Eq. (\ref{conf})
can be recast phenomenologically as
\begin{equation}\label{pre}
P(R)= \frac{P_{swim}}{1+\mathcal{L}/R} \;,
\end{equation}
with $\mathcal{L}$ a fitting parameter.
A relation between the internal active pressure and the radius $R_g$ in the deformed configuration can be obtained
as follows.
Since the membrane is composed by elastic springs it will store an elastic energy given by,
\begin{equation}\label{ela}
E(R_g)=\frac{1}{2}\frac{k}{N_b}\left[ 2 \pi (R_g-R_0) \right]^2 \, .
\end{equation}
The membrane tension exerts an inward pressure that has to be balanced by the pressure exerted by the active particles, requiring
\begin{equation} \label{press}
P(R_g)=\frac{1}{2\pi R_g}\frac{\partial}{\partial R_g} E(R_g)= \frac{2 \pi k}{N_b} \frac{R_g - R_0}{R_g} \, ,
\end{equation}
In Fig. \ref{fig:fig2}-(b), the red circles represent the quantity $(R_g-R_0)/R_g$ as a function of the actual area fraction computed as $\phi R_0^2/R_g^2$.
For $N_b=150$ the membrane has a nearly circular shape so that (\ref{press}) holds and $(R_g-R_0)/R_g$ becomes proportional to the average pressure exerted by the swimmers. This is confirmed by plotting on the same graph the swimmers pressure as obtained from simulations and divided by $2\pi k/N_b$ (black squares). In the dilute regime, Eq. (\ref{pre})
holds, as a result
the pressure should scale linearly with the packing fraction, provided the correction term $c(R)$ and thus $R$ does not change significantly with the packing fraction. By fitting the low $\phi$ data in Fig. \ref{fig:fig2}-(b) we obtain $c=0.46$. Deviations of the pressure from the linear regime, due to the excluded volume effects, are visible at high $\phi$ \cite{Yang14,Solon15b,Marchetti15}.
We can repeat the same procedure for membranes of different contour length and obtain $c$ values for different $N_b$.
Assuming $c=(1+\mathcal{L}/R)^{-1}$ we expect that the quantity $c/(1-c)$ should scale linearly with $R$ which is approximately proportional to $N_b$ (see inset of Fig. \ref{fig:fig2}-(b)).
\begin{figure*}[!t]
\centering
\includegraphics[width=.95\textwidth]{fig5}
\caption{{\bf Membrane shape for spherical swimmers.}
(a) Asphericity parameter for $N_b=50$ (blue symbols), $N_b=100$ (red symbols), and $N_b=150$ (green symbols), the lines are a guide to the eye. The membrane approaches a circular shape with increasing $\phi$.
(b) The red circles are $(R_g-R_0)/R_g$ (the red line is a guide to the eye), the black symbols represent $P N_b / 2 \pi k$, and the green line is the fit to Eq. (\ref{press}). The data are plotted as a function of the area fraction computed with respect the circle of radius $R_g$ for $N_b=150$. Inset: the quantity $c/(1-c)$ as a function of $N_b$.
(c) Probability distributions of the local curvatures for $N_b=100$ for increasing $\phi$ from $0.05$ (blue) to $0.80$ (red). (d) Parameter $\beta$ as a function of $\phi$ for $N_b=50,100,150$ (blue, red and green), the black dashed line is $\beta$ for a Gaussian distribution. The grey area represents the estimated $\phi_c$ range. Inset: $\phi_c$ obtained from the decoupling between pressure and deformation (blue symbols) the line is the estimate of $\phi_c$ given by $\phi(N_b=N_s)$.}
\label{fig:fig2}
\end{figure*}
\begin{figure*}[!t]
\centering
\includegraphics[width=.85\textwidth]{fig2}
\caption{ {\bf Membrane shape for elongated swimmers.}
(a) Asphericity parameter for $N_b=50$ (blue symbols), $N_b=100$ (red symbols), and $N_b=150$ (green symbols). The lines are a guide to the eye. With increasing $\phi$ the vesicle approaches an average circular shape ($\Delta \sim 0$).
(b) The red circles are $(R_g-R_0)/R_g$, the red line is a guide to the eye, the black squares represent $P N_b / 2 \pi k$. The data are plotted as a function of the area fraction computed with respect the circle of radius $R_g$ for $N_b=150$. (c) The probability distribution of the local curvatures undergoes a crossover from single to double peacked shape by increasing $\phi$ (in figure from $0.16$ (blue) to $0.83$ (red)). (d) To quantify the bimodal character of the distribution we look at the Sarle's bimodality coefficient $\beta$ as a function of $\phi$ for $N_b=50,100,150$ (blue, red and green), the black dashed line is $\beta$ for a Gaussian distribution. The grey area represents the estimated $\phi_c$ range. Inset: $\phi_c$ obtained from the decoupling between pressure and deformation (blue symbols) the line is the estimate of $\phi_c$ given by $\phi(N_b=N_s)$.
}
\label{fig:fig3}
\end{figure*}
\begin{figure*}[!t]
\centering
\includegraphics[width=1.\textwidth]{fig4}
\caption{{\bf Local curvature.} Joint probability distribution of the local curvature of the membrane $\kappa/\kappa_{ref}$
($\kappa_{ref}$ is the reference curvature of the circular free membrane) and the local pressure $P$
exerted by the swimmers on the membrane. The three panels refer to three different swimmers density,
$\phi=0.16$ (left), $0.51$ (middle) and $0.83$ (right). Data correspond to the case of elongated swimmers and
a membrane of $N_b=100$ beads.
}
\label{fig:fig2b}
\end{figure*}
Now we consider elongated swimmers ($p=2$, aspect ratio $\alpha=1/2$).
In order to evaluate the impact of aligning forces on the membrane shape, we perform numerical
simulations of elongated active particles at almost the same area fractions $\phi$ of the previous case.
Again, to quantify the deviations of the vesicle from circular shape we display in Fig. \ref{fig:fig3}-a the
asphericity $\Delta$. We find that $\Delta$ approaches zero with increasing $\phi$, indicating that
at high density of swimmers the active vesicle approaches an average circular shape (see also the snapshot reported in the right panel of Fig. \ref{fig:fig1}).
On the contrary for small $\phi$ we observe strong deviation from a circle,
as displayed by the four snapshots
shown in Fig. \ref{fig:fig1}, left panel.
The radius of gyration increases with $\phi$, also for elongated swimmers, as one can see in Fig. \ref{fig:fig3}-b.
We observe, however, strong deviations from Eq. (\ref{press}) at low densities where $\Delta \neq 0$.
This is not surprising since the right hand side of Eq. (\ref{press}) only holds when for circular membranes. At high area fractions the vesicle shape becomes more
circular ($\Delta\sim0$) and Eq. (\ref{press}) applies.
We stress that for elongated swimmers, where an aligning torque exists on the boundary, there is not
an ideal active gas equation of state like (\ref{swim}) \cite{Solon15}.
\subsection*{Local shape properties}
A useful characterization of the membrane shape is obtained
by analyzing the distribution of local curvatures, $\mathcal{P}(\kappa/\kappa_{ref})$, with $\kappa_{ref}=R_0^{-1}$ the curvature
of the reference circular configuration, shown in
Fig. \ref{fig:fig2}-c for spherical swimmers and in Fig. \ref{fig:fig3}-c for elongated swimmers.
Low density configurations are generically characterized by pinched spots, where both particle density and curvature are high, separated by straight, low curvature regions that are free from active particles (see the snapshot reported in Fig. \ref{fig:fig1}, left panel).
Let us start by considering $\mathcal{P}(\kappa/\kappa_{ref})$ for elongated swimmers, where the asymmetry at low $\phi$ given by $\Delta(\phi)$ is more pronounced than for spherical swimmers.
The distribution changes from bimodal to unimodal with increasing packing fraction $\phi$.
The bimodal distribution obtained at low $\phi$ corresponds to elongated vesicles. The
two peaks correspond to low curvature regions (where $\kappa\to0$ and the density of active swimmers is very low) and
high curvature regions (where swimmers accumulate and $\kappa > \kappa_{ref}$), respectively.
At large $\phi$ the vesicles are spherical on average ($\Delta\sim0$)
and the distribution of curvature exhibits a single peak. The finite width of the distribution
measures the size of fluctuations about the mean shape
with $\lambda_1 \sim \lambda_2$.
The bimodal character of the distribution can be quantified using the Sarle's bimodality coefficient $\beta=(\gamma^2+1)/k$,
where $\gamma$ is the skewness and $k$ the kurtosis of the distribution.
Fig. \ref{fig:fig3}-c shows the parameter $\beta$ reported as a function of swimmer density and for
three vesicle sizes. Deviations from the $1/3$ value, corresponding to a normal distribution,
observed at low swimmer density signals the appearance of the bimodality and associated elongated vesicle shape.
Particles tend to accumulate in small regions, enhancing the local membrane curvature, and
leaving large parts of the membrane empty. The empty regions are flat and give a peak at a vanishing value of the local curvature.
This results from a positive feedback mechanism by which a local fluctuation of particles density produces a local pressure increase that increases the local curvature on the flexible membrane. Since active particles tend to accumulate at convex boundaries, this local curvature increase drives further accumulation of swimmers.
Fig. \ref{fig:fig2b} shows the joint probability density $\mathcal{P}(\kappa/\kappa_{ref},P)$.
In the low density regime (left panel of Fig. \ref{fig:fig2b}),
flat regions of the membrane -- peak close to (0,0) in the figure -- coexist with
highly curved regions -- lighter regions close to (1.5,1.5) in the figure
(see also the snapshots reported in Fig. \ref{fig:fig1}).
By increasing the number of swimmers inside the vesicle the spot close to the origin
disappears and a single broad peak at high $\kappa/\kappa_{ref}$ survives corresponding to uniform
curvature of the membrane -- see the snapshots of Fig. \ref{fig:fig1}, right panel, corresponding to the high particles density.
Similar results are obtained also for spherical swimmers, where the curvature distribution evolves from double to single peaked with increasing area fraction $\phi$. In this case, however, this transition is sharper and occurs at lower values of $\phi$, and vesicles display a nearly circular shape in a wider range of area fractions.
The crossover from single to double peaked distribution of the membrane curvature
relies on the imbalance of swimmers along the boundaries.
A rough estimate of the packing fraction $\phi_c$ at which the crossover takes place
can be obtained by the following argument.
A membrane composed by $N_b$ beads of diameter $a$ has a length $a N_b$.
The minimum number of swimmers of thickness $a$ and length $\ell$ needed to
uniformly cover the entire length of the membrane is
$N_b$ (we suppose that the swimmers are pushing the membrane and that they are perpendicular to it).
The area fraction of swimmers is defined as $\phi=N_s a_{swim}/A_{ref}$,
where $A_{ref}= (N_b a_{swim})^2 / 4\pi$ is the area of the reference circular configuration of the free membrane.
The critical area fraction of swimmers is then
$\phi_c=\phi(N_{s}=N_b)=p \pi^2/N_b$.
This corresponds to the minimal swimmers density needed to obtain a uniform
distribution of pushing active particles along the membrane.
We obtain values of $\phi_c$ ranging $0.4$ to $0.13$ in the case of elongated swimmers,
and values from $0.2$ to $0.07$ in the case of spherical swimmers,
in agreement with the crossover regions observed in the behavior of $\beta$ (Fig. \ref{fig:fig2}-d and Fig. \ref{fig:fig3}-d
where the grey area represents the $\phi_c$ range).
A numerical estimate of $\phi_c$ is obtained as follows.
When $\Delta \neq 0$, Eq. (\ref{press}) does not hold and the relative displacement $(R_g-R_0)/R_0$
is not proportional to the average pressure exerted by the active particles. We define $\phi_c$ as
the value of $\phi$ where Eq. (\ref{press}) begins to hold.
In the inset of Fig. \ref{fig:fig2}-d the
line is the estimate of $\phi_c$ given by $\phi(N_b=N_s)$ and the symbols are the numerical values (spherical swimmers)
obtained looking at the deviation from Eq. (\ref{press}).
The curve reproduces quite well the data.
Different is the situation for the elongated swimmers (inset in Fig. \ref{fig:fig3}-d), where
the numerical estimate lies above the curve $\phi(N_b=N_s)$, i. e., the steric effect is not enough to justify the rise in $\phi_c$.
\subsection*{ Cell migration}
Flexible vesicles do not just fluctuate in shape but, at the same
time, perform a random walk under the action of the fluctuating
force arising from the combined action of swimmers' propelling
forces.
The case of spherical swimmmers is particularly remarkable
since it can be worked out analytically. Since swimmers
and passive beads have the same size and mobility, the center of mass
velocity $\mathbf V_{cm}$ is given by
\begin{equation}\label{vcdm}
{\mathbf V}_{cm}=\frac{1}{N_b+N_s}\left[\sum_n^{N_b}
{\mathbf V}_n+\sum_j^{N_s} {\mathbf v}_j\right]=\frac{\mu}{N_b+N_s} \sum_j\mathbf f_j \, ,
\end{equation}
where $\mathbf V_n$ and $\mathbf v_j$ are respectively the velocities of a membrane bead and a swimmer.
The sum of all interaction forces has to vanish so that only the sum over propelling forces
$\mathbf f_j$ survives in the last term.
Therefore the center of mass moves as a body of reduced mobility
$\mu/(N_b+N_s)$ under the action of the total propelling force on the
swimmers. The corresponding velocity-velocity correlation function is then given by
\begin{equation}
\langle\dot{\mathbf R}_{cm}(0)\cdot\dot{\mathbf R}_{cm}(t)\rangle=
\frac{\mu^2}{\left(N_s+N_b\right)^2}\sum_{i,j}^{N_s}\langle\mathbf
f_i(0)\cdot\mathbf f_j(t)\rangle\label{cdm} \, .
\end{equation}
For spherical swimmers, propelling forces only reorient due to tumbles
and are therefore uncorrelated so that
\begin{equation}
\sum_{i,j}^{N_s}\langle\mathbf f_i(0)\cdot\mathbf f_j(t)\rangle= N_s
\langle\mathbf f(0)\cdot\mathbf f(t)\rangle=\frac{N_s v^2}{\mu^2}
e^{-\lambda |t|}
\end{equation}
The mean square displacement (MSD) is obtained by a double time
integration of (\ref{cdm}), with the result
\begin{eqnarray}
\label{msd}
\nonumber \langle|\Delta\mathbf R_{cm}(t)|^2\rangle&=& \frac{N_s
v^2}{\left(N_s+N_b\right)^2}\int_0^t dt^\prime\int_0^t
dt^{\prime\prime}e^{-\lambda |t^\prime-t^{\prime\prime}|}\\
&=&\frac{4 D_v}{\lambda}\left( \lambda t - 1 + e^{-\lambda t}\right)
\end{eqnarray}
\begin{figure*}[!t]
\centering
\includegraphics[width=.75\textwidth]{fig6}
\caption{ {\bf Vesicle motion.} (a) Mean square displacement of the vesicle center of mass for $\phi=0.16$ and $N_b=100$.
Data correspond to the cases of elongated swimmers (blue symbols) and spherical swimmers (green symbols).
The red curve is the theoretical prediction given by eq. (\ref{msd}). The red dashed curve is the theoretical mean square displacement of a free run-and-tumble particle. (b) Diffusion coefficient normalized to the reduced value $D_v(N_s)$ for elongated (blue symbols) and spherical (green symbols) swimmers as a function of swimmer density $\phi$ ($N_b=100$). The parameters are obtained by fitting the data with eq. (\ref{msd}). The black dashed line is $D_{fit}/D_v=1$.}
\label{fig:fig5}
\end{figure*}
The MSD of the center of mass of the vesicle is then given by the
MSD of an individual swimmer, with the single swimmer diffusivity $D=v^2/2 \lambda$
replaced by the reduced value $D_v=D N_s/\left(N_s+N_b\right)^2$. The MSD
of a free swimmer \cite{Angelani13} and of a vesicle filled with spherical swimmers
are shown in Fig. \ref{fig:fig5}-a together with the formula
(\ref{msd}).
The case of non spherical swimmers is more complex due
to the rotational couplings between propelling forces induced by
anisotropic interactions. Still the calculated MSD can be fitted with
formula (\ref{msd}) leaving both $D_v$ and $\lambda$ as free fitting
parameters. In this case, however, we expect that due to anisotropic
interactions, correlations between $\mathbf f_j$ will arise whose
relaxation is not solely driven by the tumbling rate $\lambda$ but can
occur on longer time scales. The obtained fitting parameters confirm
those expectation giving $\lambda_{fit}\sim0.3\lambda$.
The fitted diffusion coefficients as a function of particles density are reported in
Fig. (\ref{fig:fig5}-b) for both spherical and elongated swimmers.
As expected, the diffusion coefficient in the spherical case is given by the reduced value $D_v$.
In the case of elongated swimmers the vesicle diffusivity is much larger due to a longer persistence of propelling forces arising from locally aligned configurations of swimmers.
\section*{Discussion}
Understanding the properties of active matter in confined geometries is of great importance
not only for basic science, but also for possible practical applications,
for example in micro bio-mechanics, where synthetic autonomous self-propelled objects could be
used as drug-delivery agent or for mechanical actuation.
Previous studies have focused on the behavior of active particles in the presence of
rigid obstacles or confined by stiff boundaries \cite{Smallenburg15,Mallory14,Fily15,Fily14}.
In this paper we explore the shape changes and spontaneous migration of a flexible vesicle filled with active particles.
We find strong fluctuations of the vesicle's shape, changing from circular to elongated with decreasing
number of enclosed particles.
The transition between these two regimes is associated with the crossover of the distribution of the local
curvatures $\mathcal{P}(\kappa)$
from single-peaked to bimodal.
The observed shape deformation is driven by
the accumulations of active particles in the high
curvature regions, which has been observed also in the case of non interacting
Active Brownian particles under strong confinement \cite{Fily14}. Elongated swimmers enhance shape deformations because alignment tends to increase particle accumulation in high curvature regions.
We have recently become aware of a study similar to ours investigating shape fluctuations in 2D flexible vesicles
filled with \textit{spherical} Active Brownian particles \cite{Tian15}.
Although in this work particles' trajectories are randomized by rotational diffusion while we use run-and-tumble dynamics, both our work and Ref. \cite{Tian15} find similar robust shape fluctuations induced by the active particles. The transition from elongated to circular vesicle shape that we observed by increasing density of enclosed swimmers is found in Ref. \cite{Tian15} upon decreasing the particles' propelling force.
We also show that the filled vesicle effectively behaves like an active object, with
exponentially correlated random motion, whose properties
are strongly dependent on the shape and density of the self-propelled pushing particles inside.
In the case of spherical swimmers we can calculate the diffusion coefficient $D_v$ and the
correlation time $\tau$ of the persistent random walk of the filled vesicle, that can be described in terms of an effective temperature that depends on the number of enclosed swimmers. The migratory properties of
the cell are determined entirely by the motility of the active particles.
We additionally examine the behavior of vesicles filled with elongated particles that was not considered in Ref. \cite{Tian15}.
In this case the diffusion coefficient of the whole vesicle is about one order of magnitude greater than that of the
spherical case and it is a non-monotonic function of the swimmers density,
reaching a maximum value near the critical packing fraction $\phi_c$ controlling the crossover from single to double peaked distribution
of the membrane curvature.
The behavior of vesicles filled with active particles bear some resemblance with the directed migration of Eukaryotic cells, as observed for instance in wound healing assays or in the presence of chemotactic cues. In these situations cells become polarized and perform directed random walks advancing preferentially toward or away from chemical stimuli \cite{Swaney10} or towards regions void of other cells~\cite{Kim13}.
Our work shows (see Fig. \ref{fig:fig5}b) that vesicle migration is most effective when driven by elongated particles that indeed induce a net polarization of the vesicle, as observed in the chemotactic motion of living cells. It would be interesting to study the effect
of chemotaxis on our model by considering a
space-varying tumbling rate $\lambda({\mathbf r})$ which depends on an external chemotactic field $c({\mathbf r})$.
\section*{Acknowledgements}
RDL acknowledges funding from the European Research Council under the European Union's
Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement No. 307940.
MP and MCM were supported by the Simons Foundation Targeted Grant in the Mathematical Modeling of Living Systems Number: 342354 and by the Syracuse
Soft Matter Program.
MCM also acknowledges support
by the National Science Foundation through award DMR-1305184.
|
\section{Supplementary Information to: \\ Absence of Marginal Stability in Self-Generated Coulomb Glasses}
\maketitle
|
\section{Introduction}\label{sec:introduction}
With the prominence of the Internet and the rise of the Internet of Things (IoT), significant amount of data are being generated, stored, and exchanged over the networks. Proper data management has become one of the most important and challenging aspects in system design. Information contained in the data are usually valuable resources that can be harnessed.
However, the extensive use of data incurs some privacy risks especially when sensitive information is involved.
The ultimate goal is to utilize the data to its full extent while still preserving privacy of the sensitive information.
Sankar et al. \cite{srpUPTI13} and du Pin Calmon and Fawaz \cite{dfPASI12} studied the utility-privacy tradeoff from an information theoretic perspective, by relating to the framework of secure lossy source coding \cite{yASCP83,yCTFS94,prOSDS07,gepLCWS08,vpSMSC13,scRDTS14}. Inspired by these works, we consider a problem of secure \emph{remote} source coding where the remote source consists of \emph{public} and \emph{secret} parts (hidden information associated with the data). The legitimate receiver and eavesdropper are assumed to have access to the compact representation of the data as well as separate side information. The goal is to extract public attribute/feature of the data at the legitimate receiver from the compact representation satisfying a distortion criterion, while ensuring a low amount of information leakage of secret attribute/feature of the data to the eavesdropper. Similarly as in \cite{srpUPTI13}, we capture utility of the data by the reconstruction distortion of the public part at the legitimate receiver, and capture the privacy leakage by the normalized mutual information between the secret part and the eavesdropper observation. The compression rate is also considered as a practical constraint on limited storage. We wish to characterize the optimal tradeoff region of the compression rate, incurred distortion, and privacy leakage rate. In this work, inner and outer bounds to the optimal tradeoff are given and shown to be tight for some special cases. The results can be relevant for data sharing scenarios with (external) attacks on sensitive information.
In the same spirit as how the the rate-distortion theorem under logarithmic loss \cite{mfUP98,cwMSCU14,msfmFIBT14} is related to the information bottleneck problem \cite{tpbTIBM00}, in this work, the rate-distortion-privacy leakage tradeoff under logarithmic loss distortion corresponds to the secure version of information bottleneck.
It extends the information bottleneck by including the privacy constraint or alternatively it extends the \emph{dual} privacy funnel problem \cite{msfmFIBT14} by including the compression rate constraint. The secure remote source coding problem under logarithmic loss distortion therefore gives an operational meaning to the secure information bottleneck problem.
\begin{figure}[]
\centering
\psfrag{x}[][][0.9]{\small{$X^{n}$}}
\psfrag{y}[][][0.9]{\small{$Y^n$}}
\psfrag{z}[][[0.9]{\small{$Z^n$}}
\psfrag{yps}[][][0.9]{\small{$(Y_p^n,Y_s^n)$}}
\psfrag{yha}[][][0.9]{\small{$\hat{Y}_p^n$}}
\psfrag{d}[][][0.9]{\small{$E[d(Y_p^n,\hat{Y}_p^n)] \leq D$}}
\psfrag{w}[][][0.9]{\small{$W$, rate $R$}}
\psfrag{px}[][][0.9]{\small{$P_{X,Y,Z|Y_p,Y_s}$}}
\psfrag{enc}[][][0.9]{\small{Encoder}}
\psfrag{dec}[][][0.9]{\small{Decoder}}
\psfrag{eve}[][][0.9]{\small{Eaves.}}
\psfrag{L}[][][0.9]{\small{$\frac{1}{n} I(Y_s^{n};W,Z^{n}) \leq L$}}
\includegraphics[width=0.45\textwidth]{tupic12_3.eps}
\caption{Privacy-constrained remote lossy source coding.}\label{fig:model}
\end{figure}
Our problem is
closely related to works on lossy source coding with a privacy constraint. For instance, Yamamoto considered secure source coding with remote sources to be extracted and protected \cite{yASCP83,yCTFS94}.
Villard and Piantanida \cite{vpSMSC13} considered lossy secure source coding where the source sequence is reconstructed at the decoder satisfying a distortion criterion while limiting the leakage of the source to an eavesdropper below a certain level. Tandon et al. \cite{tspDLSC13} considered privacy of side information at the first decoder against a secondary decoder in the Heegard-Berger setting \cite{hbRDSA85}.
\cite{srpUPTI13} and \cite{msITPI15} considered one-way and interactive data sharing settings where the source consists of public and private parts and characterized the set of all achievable distortion-leakage pairs.
The main differences of our work from \cite{srpUPTI13} are the presence of an eavesdropper with correlated side information and the fact that public and private information are considered as remote sources.
\textit{Notation}: We denote the discrete random variables, their
corresponding realizations or deterministic values, and their
alphabets by the upper case, lower case, and calligraphic letters,
respectively. The term $X_{m}^{n}$ denotes the sequence
$\{X_{m},\ldots,X_{n}\}$ when $m\leq n$, and the empty set
otherwise. Also, we use the shorthand notation $X^{n}$ for
$X_{1}^{n}$. The term $X^{n\setminus i}$ denotes the set
$\{X_{1},\ldots,X_{i-1},X_{i+1},\ldots,X_{n}\}$. Cardinality of the
set $\mathcal{X}$ is denoted by $|\mathcal{X}|$. Notation $[1:K]$ denotes the set $\{1,2,\dots,K\}$. Finally, we use
$X-Y-Z$ to denote that $(X,Y,Z)$ forms a Markov chain.
Other notations follow the standard ones in \cite{ekNIT11}.
\section{Secure Remote Source Coding}
\subsection{Problem Formulation}\label{sec:problem_setting}
Let us consider a secure remote source coding shown in Fig.~\ref{fig:model}. Source and side information alphabets, $\mathcal{X}, \mathcal{Y}_p,\mathcal{Y}_s, \mathcal{Y},\mathcal{Z}$ are finite sets. Let $(X^n,Y_p^n,Y_s^n,Y^n,Z^n)$ be $n$-length sequences which have i.i.d. components distributed according to some fixed distribution $P_{X,Y_p,Y_s,Y,Z}$.
The sequence $X^n$ represents the data to be revealed or shared. Public and secret attributes/features associated with the data (but not accessible/allowed to be processed directly) are represented by $Y_p^n$ and $Y_s^n$, respectively. The rate-limited description $W$ is generated based on $X^n$.
The decoder reconstructs the public attribute of the data based on $W$ and correlated side information $Y^n$. For generality, we consider an eavesdropper which has access to the (publicly) stored description and another correlated side information $Z^n$. The secure remote source coding should ensure the reconstruction quality of the public attribute within a prescribed distortion, and at the same time preserve privacy of the secret part by limiting the amount of information leakage rate at the eavesdropper $\frac{1}{n}I(Y_s^n;W,Z^n)$. We note that if $Z=Y$, the problem reduces to the case where we impose a privacy constraint against the legitimate receiver.
Let $d: \mathbb{R}\times\mathbb{R} \rightarrow [0,\infty)$ be a distortion measure. The distortion between $Y_p^n$ and its reconstruction $\hat{Y}_p^n$ is defined as
\[d^{(n)}(Y_p^n,\hat{Y}_p^n)= \frac{1}{n}\sum_{i=1}^nd(Y_{p,i},\hat{Y}_{p,i}).\]
We are interested in characterizing the optimal tradeoff of the compression rate, incurred average distortion at the legitimate decoder, and information leakage rate at the eavesdropper.
\begin{definition}\label{def:code_mFAP}
A $(|\mathcal{W}^{(n)}|,n)$-code for secure remote source coding consists of
\begin{itemize}
\item an encoder $f^{(n)}: \mathcal{X}^{n} \rightarrow \mathcal{W}^{(n)}$,
\item a decoder $g^{(n)}: \mathcal{W}^{(n)} \times \mathcal{Y}^{n} \rightarrow \mathcal{\hat{Y}}_p^{(n)}$.
\hfill $\lozenge$
\end{itemize}
\end{definition}
\begin{definition} A rate-distortion-leakage tuple $(R,D,L) \in \mathbb{R}^{3}_{+}$ is said to be \emph{achievable} if, for any $\delta>0$ there exists a sequence of $(|\mathcal{W}^{(n)}|,n)$-codes such that, for all sufficiently large $n$,
\begin{align}
\frac{1}{n}\log\big|\mathcal{W}^{(n)}\big| &\leq R+\delta,\label{eq:rate_constraint}\\
E[d^{(n)}(Y_p^n,g^{(n)}(W,Y^n))] &\leq D+\delta,\\
\frac{1}{n}I(Y_s^n;W,Z^n) &\leq L+\delta,\label{eq:leakage_constraint}
\end{align}
where $W = f^{(n)}(X^n)$. The \emph{rate-distortion-leakage} region $\mathcal{R}$ is defined as the closure of the set of all achievable tuples.\hfill $\lozenge$
\end{definition}
\subsection{Results}
\begin{theorem}[Inner bound]\label{theorem:inner}
An inner bound to the rate-distortion-leakage region $\mathcal{R}_{in}$ is given as a set of all tuples $(R,D,L) \in \mathbb{R}_{+}^3$ satisfying
\begin{align}
R &\geq I(X;V|Y) \label{eq:rate}\\
D&\geq E[d(Y_p,g(V,Y))]\label{eq:distortion}\\
L &\geq I(Y_s;V,Y) + I(Z;X,Y_s|U)-I(Y;X,Y_s|U) \nonumber\\
&\qquad -I(X;Z|V,Y_s,Y)+I(X;Y|Y_s,Z),\label{eq:leakage}
\end{align}
for some $P_{X,Y_p,Y_s,Y,Z}$$P_{V|X}P_{U|V}$ and $g: \mathcal{V} \times \mathcal{Y} \rightarrow \hat{\mathcal{Y}}_p$ with $|\mathcal{U}|\leq |\mathcal{X}|+3$ and $|\mathcal{V}|\leq (|\mathcal{X}|+3)(|\mathcal{X}|+2)$.
\end{theorem}
\begin{proof}
The proof is given in Appendix \ref{app:proof_inner} and is based on a random coding argument where the achievable scheme follows the layered (superposition) coding with binning in \cite{vpSMSC13}. The main difference lies in the analysis of achievable leakage rate.
\end{proof}
\begin{remark}
The constraint in \eqref{eq:leakage} can be rewritten as
\begin{align*}
L &\geq I(X,Y_s;U,V,Y) + I(Z;X,Y_s|U)-I(Y;X,Y_s|U) \\
&\qquad -I(X;Z|Y_s)-I(X;V|Y,Y_s,Z),
\end{align*}
in which the terms on the right-hand side may be interpreted as follows. The term $I(X,Y_s;U,V,Y)$ corresponds to the leakage of $(X,Y_s)$ through the description which depends on the remaining uncertainty at the decoder (which in this case can decode $(U,V)$ and knows $Y$). The terms $I(Z;X,Y_s|U)-I(Y;X,Y_s|U) $ is the additional leakage of $(X,Y_s)$ through the difference of side information available at the eavesdropper and decoder given that the codeword $U$ can be decoded at the eavesdropper. Since we are only interested in the leakage of $Y_s$, the remaining terms correspond to the leakage reduction of $X$ that is ``orthogonal" to that of $Y_s$.
The layered coding here provides some degree of freedom to optimize achievable leakage rate for our general setting.
\hfill $\lozenge$
\end{remark}
Next, we provide an outer bound to the rate-distortion-leakage region.
\begin{theorem}[Outer bound]\label{theorem:outer}
An outer bound to the rate-distortion-leakage region $\mathcal{R}_{out}$ is given as a set of all tuples $(R,D,L) \in \mathbb{R}_{+}^3$ satisfying \eqref{eq:rate}, \eqref{eq:distortion}, and
\begin{align}
L &\geq I(Y_s;V,Y) + I(Z;X,Y_s|U)-I(Y;X,Y_s|U) \nonumber\\
&\qquad -I(X;Z|V,Y_s,Y)+I(X;Y|T,Y_s,Z),
\end{align}
for some $P_{X,Y_p,Y_s,Y,Z}$$P_{T,V|X}P_{U|V}$ and $g: \mathcal{V} \times \mathcal{Y} \rightarrow \hat{\mathcal{Y}}_p$.
\end{theorem}
\begin{proof}
The proof is based on standard properties of the entropy function and the Csiszar's sum identity \cite{ekNIT11} and is given in Appendix \ref{app:proof_outer}.
\end{proof}
\begin{remark}
The results in Theorems \ref{theorem:inner} and \ref{theorem:outer} can be extended to a scenario where the sequences $(X^n,Y_p^n,Y_s^n)$ are available directly at the encoder. In this case, we can replace $X$ by $(X,Y_p,Y_s)$ in Theorems \ref{theorem:inner} and \ref{theorem:outer}, and the joint distributions become those of the form $P_{X,Y_p,Y_s,Y,Z}P_{V|X,Y_p,Y_s}P_{U|V}$ and $P_{X,Y_p,Y_s,Y,Z}P_{T,V|X,Y_p,Y_s}P_{U|V}$, respectively.
Theorems \ref{theorem:inner} and \ref{theorem:outer} can also be generalized to the case where $Y_p$ and $Y_s$ are not ``disjoint," i.e., they share some common part. For instance, the decoder may wish to reconstruct some attributes associated with the data that are considered as secret to the eavesdropper. We simply modify the setup by replacing $Y_p$ by $(Y_p,Y_c)$ and $Y_s$ by $(Y_s,Y_c)$, where $Y_c$ acts as a common part that is supposed to be reconstructed at the decoder and protected against the eavesdropper. Theorems~\ref{theorem:inner} and \ref{theorem:outer} continue to hold with $Y_p$ replaced by $(Y_p,Y_c)$ and $Y_s$ replaced by $(Y_s,Y_c)$ and the joint distribution of relevant source and side information is given by $P_{X,Y_p,Y_s,Y_c,Y,Z}$.\hfill $\lozenge$
\end{remark}
We see that inner and outer bounds in Theorems \ref{theorem:inner} and \ref{theorem:outer} do not match in general. In particular, there is a gap between the leakage rate bounds.
The difficulty of proving the tight bound lies in the complex dependency of information available at the eavesdropper and the secret remote source $Y_s^n$.
Nevertheless, there exist some special cases where the bounds are tight.
\begin{corollary}\label{corollary:region1}
For the sources and side information whose joint distributions satisfy $I(X;Y_s,Z|Y)=0$, the \emph{rate-distortion-leakage} region $\mathcal{R}$ is given as a set of all tuples $(R,D,L) \in \mathbb{R}_{+}^3$ satisfying \eqref{eq:rate}, \eqref{eq:distortion}, and $L \geq I(Y_s;Z)$
for some $P_{X,Y_p,Y_s,Y,Z}$$P_{V|X}$ and $g: \mathcal{V} \times \mathcal{Y} \rightarrow \hat{\mathcal{Y}}_p$ with $|\mathcal{V}|\leq |\mathcal{X}|+1$.
Interestingly, in this case, the only leakage of $Y_s^n$ is from correlated side information $Z^n$. There is no additional leakage rate from the source description $W$ sent over the rate-limited link.
This can be explained as follows. Our achievability scheme is based on the binning technique which renders $I(W;Y^n) \leq n\delta_{\epsilon}$. Since we have the Markov chain $W-X^n-Y^n-(Y_s^n,Z^n)$, from the data processing inequality, the additional leakage rate of $Y_s^n$ due to $W$ becomes negligible, i.e., $I(Y_s^n;W|Z^n) \leq I(Y_s^n,Z^n;W)\leq I(Y^n;W)\leq n\delta_{\epsilon}$.
\end{corollary}
\begin{proof}
The proof follows from specializing Theorems \ref{theorem:inner} and \ref{theorem:outer} to the case where $X-Y-(Y_s,Z)$ forms a Markov chain, where in achievability, we choose $U =\emptyset$.
\end{proof}
\begin{corollary}\label{corollary:region}
For the sources and side information whose joint distributions satisfy $I(X;Y|Y_s,Z)=0$, the \emph{rate-distortion-leakage} region $\mathcal{R}$ is given as a set of all tuples $(R,D,L) \in \mathbb{R}_{+}^3$ satisfying \eqref{eq:rate}, \eqref{eq:distortion}, and
\begin{align}
L &\geq I(Y_s;V,Y) + I(Z;X,Y_s|U)-I(Y;X,Y_s|U) \nonumber\\
&\qquad -I(X;Z|V,Y_s,Y),
\end{align}
for some $P_{X,Y_p,Y_s,Y,Z}$$P_{V|X}P_{U|V}$ and $g: \mathcal{V} \times \mathcal{Y} \rightarrow \hat{\mathcal{Y}}_p$ with $|\mathcal{U}|\leq |\mathcal{X}|+3$ and $|\mathcal{V}|\leq (|\mathcal{X}|+3)(|\mathcal{X}|+2)$.
\end{corollary}
\begin{proof}
The achievability proof follows directly from Theorem \ref{theorem:inner} with $I(X;Y|Y_s,Z)=0$, while the converse follows from Theorem \ref{theorem:outer} and the fact that $I(X;Y|T,Y_s,Z) \geq 0$.
\end{proof}
\begin{remark}
Corollaries \ref{corollary:region1} and \ref{corollary:region} hold also for the case of stochastic encoder where the description $W$ is randomly generated according to a conditional PMF $p(w|x^n)$. This follows from the fact that in the proof of Theorem~\ref{theorem:outer}, we do not make any assumption regarding the deterministic encoder.
\hfill $\lozenge$
\end{remark}
\begin{remark}
We see that the rate-distortion-leakage region is known for several classes of sources and side information, e.g., those satisfying $I(X;Y|Y_s,Z)=0$ in Corollary~\ref{corollary:region} which includes also the semi-deterministic mapping with $X$ being a deterministic function of $Y_s$.
Moreover, the result in Corollary~\ref{corollary:region} recovers several existing results in the secure lossy source coding literature, e.g.,
\begin{itemize}
\item when $Y=Z$, we may think of the privacy leakage constraint as one imposed at the legitimate decoder. If $X=(Y_p,Y_s)$, Corollary \ref{corollary:region} recovers the result of the utility-privacy tradeoff with side information in \cite{srpUPTI13}. Furthermore, if $X=(Y_p,Y_s)$ and $Y=Z=\emptyset$,
Corollary~\ref{corollary:region} recovers the result in \cite{yASCP83}.
\item when $X=Y_p=Y_s$, Corollary \ref{corollary:region} recovers the result of secure lossy source coding problem studied in \cite{vpSMSC13}.
\item when $Y=Y_s$, the leakage term becomes $I(Y_s^n;W,Z^n)$ which is of the same type as the side information privacy considered in \cite{tspDLSC13}. The main difference is that in \cite{tspDLSC13} side information privacy is considered at the secondary receiver who observes no additional side information. If the reconstruction constraint at the secondary receiver is neglected, then zero leakage rate is achievable by the Wyner-Ziv coding \cite{wzTRDF76}.
Interestingly, in our case where the eavesdropper has access to the additional side information, the layered random binning scheme turns out to be optimal. An achievable leakage rate in this case is $I(Y_s;U,Z)+I(V;Z|Y_s,U)$ which is larger than $I(Y_s;Z)$. This is due to the fact that conditioned on $Z^n$, the Wyner-Ziv bin indices are still correlated with side information $Y_s^n$, and thus revealing some information about $Y_s^n$ to the eavesdropper. \hfill $\lozenge$
\end{itemize}
\end{remark}
\begin{corollary}\label{corollary:lossless}
When we set $Y_p=X$ and consider a lossless reconstruction of $X^n$ at the decoder, an inner bound to the rate-leakage region is given by the set of all $(R,L)\in \mathbb{R}_{+}^2$ satisfying
\begin{align*}
R &\geq H(X|Y)\\
L &\geq I(Y_s;X,Y) + I(Z;X,Y_s|U) - I(Y;X,Y_s|U)\\&\qquad +I(X;Y|Y_s,Z),
\end{align*}
for some $P_{X,Y_s,Y,Z}P_{U|X}$ with $|\mathcal{U}|\leq |\mathcal{X}|$.
The inner bound above can be proved similarly as in Theorem \ref{theorem:inner}. In fact, it can be obtained from Theorem~\ref{theorem:inner} by setting $Y_p=X=V$.
\end{corollary}
\begin{remark}
We note that the special case of lossless reconstruction above was considered recently in \cite{aalLSSC15} where an inner bound to the rate-equivocation region is provided. In general, the results in Corollary \ref{corollary:lossless} and \cite[Theorem 3]{aalLSSC15} do not match. As an example where $Y=Y_s$ and $Z=\emptyset$, it can be shown that Corollary \ref{corollary:lossless} implies that zero leakage rate is achievable (by choosing $U=\emptyset$). However, the achievable leakage rate according to \cite[Theorem 3]{aalLSSC15} can be strictly positive. \hfill $\lozenge$
\end{remark}
\subsection{Quadratic Gaussian Example}
We consider an example of the tradeoff in Corollary \ref{corollary:region} for Gaussian sources under quadratic distortion.
Assuming that $X \sim \mathcal{N}(0,N_x)$, $Y_s= X + \tilde{N}_s$, where $\tilde{N}_s \sim \mathcal{N}(0,N_s) \perp X$, and $Y_p= Y_s + \tilde{N}_p$, where $\tilde{N}_p \sim \mathcal{N}(0,N_p-N_s) \perp Y_s$. Note that $X-Y_s-Y_p$ forms a Markov chain. Also, we assume that there is no side information, i.e., $Y=Z=\emptyset$. The tradeoff region in Corollary \ref{corollary:region} reduces to the set of all $(R,D,L)$ satisfying
\begin{align}
R &\geq \frac{1}{2}\log\Big(\frac{N_x}{D-N_p}\Big) \label{eq:R_Guassian},\\
L &\geq \frac{1}{2}\log\Big(\frac{N_x+N_s}{D-N_p+N_s}\Big),\label{eq:L_Gaussian}
\end{align}
for $D > N_p$.
While our main results were proven for discrete memoryless sources, the extension to the quadratic Gaussian case is standard
and it follows, e.g., \cite{wTRDF78} and
\cite{ekNIT11}. For achievability, we set $U=\emptyset$, choose $V$ to be jointly Gaussian with $X$, i.e., $V=X+Q, Q \sim \mathcal{N}(0,N_q)$, and choose the reconstruction function $g(\cdot)$ to be an MMSE estimate of $Y_p$ given $V$. By letting $D=E[d(Y_p,g(V))]$ and substituting it into the constraints on $R$ and $L$, we obtain the result above. The converse follows from utilizing the EPI \cite{ekNIT11} together with the fact that $R$ and $L$ are decreasing in $h(X^n|W)$ and $h(Y_s^n|W)$, respectively. For the more detailed proof, please see Appendix~\ref{app:proof_Gaussian_example}.
From \eqref{eq:R_Guassian} and \eqref{eq:L_Gaussian}, we can also write the minimum distortion as a function of $R$ and $L$, i.e., \[D_{\text{min}}(R,L)=\max\{N_p+N_x2^{-2R},N_p-N_s+(N_x+N_s)2^{-2L}\}.\]
For a fixed $R$, the minimum distortion $D_{\text{min}}$ decreases with $L$, illustrating the utility-privacy tradeoff in terms of minimum achievable distortion and information leakage rate.
\section{Logarithmic loss distortion}
Logarithmic loss \cite{mfUP98,cwMSCU14} is a measure of quality of the ``soft" estimate used in several applications
\cite{bPRML06},\cite{aabgEBDM07}.
Under logarithmic loss, we employ a soft estimate of the source in terms of a probability distribution over the source alphabet.
For a sequence $\hat{X}^n \in \hat{\mathcal{X}}^n$, we denote $\hat{X}_i$, $i=1,\ldots,n$, the $i^{th}$ element of $\hat{X}^n$. Then $\hat{X}_i, i=1,\ldots,n$ is a probability distribution on $\mathcal{X}$, i.e., $\hat{X}_i: \mathcal{X} \rightarrow [0,1]$, and $\hat{X}_i(x)$ is a probability distribution on $\mathcal{X}$ evaluated for the outcome $x \in \mathcal{X}$.
\begin{definition}[Logarithmic loss \cite{cwMSCU14}]
The logarithmic loss distortion is defined as $d(x,\hat{x}) = \log(\frac{1}{\hat{x}(x)})= D_{KL}(\mathbf{1}_{\{x\}}||\hat{x})$, where $\mathbf{1}_{\{x\}}: \mathcal{X} \rightarrow \{0,1\}$ is an indicator function such that, for $a \in \mathcal{X}$, $\mathbf{1}_{\{x\}}(a)=1$ if $a=x$, and $\mathbf{1}_{\{x\}}(a)=0$ otherwise.
Using this definition for symbol-wise distortion, it is standard to define the distortion between sequences as $d^{(n)}(x^n,\hat{x}^n)=\frac{1}{n}\sum_{i=1}^n d(x_i,\hat{x}_i)$.
\end{definition}
Under logarithmic loss, the average distortion and conditional entropy (equivocation at decoder) are closely related (see, e.g., Lemma
\ref{lemma:loglossouterbound} below \cite{cwMSCU14}).
This is reminiscent of what is known, e.g., for the Gaussian settings under quadratic distortion.
\begin{lemma}\label{lemma:loglossouterbound}
Let $C=(W,Y^n)$ be the argument of the reconstruction function $g^{(n)}(\cdot)$, i.e., $\hat{X}^n = g^{(n)}(C)$, then under the logarithmic loss distortion measure, we get $E[d^{(n)}(X^n,g^{(n)}(C))] \geq \frac{1}{n}\sum_{i=1}^nH(X_i|C)$.
\end{lemma}
The proof of the lemma follows from definition of logarithmic loss, i.e., $d(x_i,g^{(n)}_i(c)) \triangleq \log(\frac{1}{q(x_i|c)})$ where $q$ is a probability measure on $\mathcal{X}$. Then the expected distortion conditioned on $C=c$,
\begin{align*}
&E[d^{(n)}(X^n,g^{(n)}(c))|C=c]\\
& = E[\frac{1}{n}\sum_{i=1}^n d(X_i,g^{(n)}_i(c))|C=c]\\
&= \sum_{x^n \in \mathcal{X}^n} p(x^n|c) \frac{1}{n}\sum_{i=1}^n d(x_i,g^{(n)}_i(c))\\
&= \frac{1}{n}\sum_{i=1}^n \sum_{x_i \in \mathcal{X}} p(x_i|c) \log\bigg(\frac{1}{q(x_i|c)}\bigg)\\
&= \frac{1}{n}\sum_{i=1}^n \sum_{x_i \in \mathcal{X}} p(x_i|c) \log\bigg(\frac{p(x_i|c)}{q(x_i|c)}\cdot \frac{1}{p(x_i|c)}\bigg)\\
&= \frac{1}{n} \sum_{i=1}^n D_{KL}(p(x_i|c)||q(x_i|c)) + \frac{1}{n} \sum_{i=1}^nH(X_i|C=c)\\
& \geq \frac{1}{n} \sum_{i=1}^n H(X_i|C=c).
\end{align*}
By averaging both sides over all $c \in \mathcal{C}$, from the law of total expectation, the lemma is proved.
We note that Lemma~\ref{lemma:loglossouterbound} holds only for the case of symbol-by-symbol logloss distortion, i.e., $d^{(n)}(x^n,\hat{x}^n)=\frac{1}{n}\sum_{i=1}^n d(x_i,\hat{x}_i)$, which is considered in the problem formulation.
\begin{corollary}\label{corollary:region_logloss}
Consider the privacy-constrained remote source coding in Fig. \ref{fig:model}. When the decoder and eavesdropper observe identical side information, i.e., $Y= Z$ (can be seen as privacy leakage against the legitimate receiver), the \emph{rate-distortion-leakage} region under logarithmic loss is given as a set of all tuples $(R,D,L) \in \mathbb{R}_{+}^3$ satisfying
\begin{align}
R &\geq I(X;V|Y) \\
D&\geq H(Y_p| V,Y)\\
L &\geq I(Y_s;V,Y),
\end{align}
for some $P_{X,Y,Y_p,Y_s}$$P_{V|X}$, with $|\mathcal{V}|\leq |\mathcal{X}|+2$.
\end{corollary}
\begin{proof}
For achievability, we apply Corollary \ref{corollary:region} under logarithmic loss distortion with $Y=Z$. We choose $g(\cdot)$ to be a conditional probability distribution on $\mathcal{Y}_p$, i.e., $g(v,y)=p(y_p|v,y)$ which gives $E[d(Y_p,g(V,Y))]=H(Y_p|V,Y)$.\footnote{The proof of Theorem~\ref{theorem:inner} holds for bounded distortion measures. However,
it can still be extended to logarithmic loss distortion by perturbing the reconstruction probability distribution (see e.g., \cite[Remark 3.4]{ckOSSC13}).}
The converse follows by setting $V_i = (W,Y^{n \setminus i})$ and applying Lemma~\ref{lemma:loglossouterbound} which gives $E[d^{(n)}(Y_{p}^n,g^{(n)}(W,Y^n))]=\frac{1}{n}\sum_{i=1}^n E[d(Y_{p,i},g_i^{(n)}(W,Y^n))]\geq \frac{1}{n}\sum_{i=1}^nH(Y_{p,i}|W,Y^n)$.
\end{proof}
\subsection{Secure Information Bottleneck}
Based on Corollary \ref{corollary:region_logloss}, we can formulate an optimization problem where for a given $P_{X,Y,Y_p,Y_s}$, we wish to minimize the rate $I(X;V|Y)$ over all $P_{V|X}$ subject to the constraints on the distortion and information leakage, i.e.,
\begin{align}
\min_{P_{V|X}} I(X;V|Y)&\\
\text{s.t.} \ I(Y_p;V,Y) &\geq D'\label{eq:info_bottleneck_information}\\
I(Y_s;V,Y)&\leq L,\label{eq:info_bottleneck_leakage}
\end{align}
where in \eqref{eq:info_bottleneck_information}, the distortion constraint is rewritten as the ``information" constraint with $D' \triangleq H(Y_p)-D$.
The minimum above corresponds to the minimum achievable rate for fixed $D'$ and $L$ and may be termed as the rate-information-leakage function $R_{\min}(D',L)$.
We can see that the optimization problem is not convex, e.g., $I(Y_p;V,Y)$ is a convex function in $P_{V|X}$ for a fixed $P_{X,Y,Y_p,Y_s}$. Interestingly, the problem shares some similarity with the information bottleneck problem \cite{tpbTIBM00} where in our case there is an additional constraint on the privacy leakage rate \eqref{eq:info_bottleneck_leakage} and the presence of side information $Y$. Given that some side information $Y$ is known beforehand, the goal here is to represent the data $X$ efficiently by a compact representation $V$, while maximizing the relevance of $V$ on the public attribute $Y_p$ and minimizing the relevance of $V$ on the secret attribute $Y_s$. Due to the additional constraint on the privacy leakage rate, we term this optimization problem as \emph{secure information bottleneck}. When $Y=\emptyset$, it also corresponds to a variant of information bottleneck considered in \cite{ctERSW02}. The secure remote source coding problem
under logarithmic loss distortion therefore gives an operational meaning to the secure information bottleneck.
To solve the secure information bottleneck problem, we may extend the iterative algorithm proposed for the information bottleneck \cite{tpbTIBM00}.
For example, the optimization problem above can be solved by minimizing the function
\begin{align*}
\mathcal{L} &= I(X;V|Y)-\beta[I(Y_p;V,Y)-\gamma I(Y_s;V,Y)]\\
&\qquad -\sum_{x,v}\lambda(x)p(v|x)
\end{align*}
over all $p(v|x)$, where $\lambda(x)$ are functions of $x$, and parameters $\beta$ and $\gamma$ are introduced to capture the tradeoff due to information (distortion) and privacy leakage constraints.
Similarly as in \cite{tpbTIBM00}, \cite{ctERSW02}, it can be shown that the stationary points of $\mathcal{L}$ are given by
\begin{align}
p(v|x) &= \frac{1}{Z(x,\beta,\gamma)}\exp\Big\{\sum_y p(y|x)\Big[\log p(v|y)\nonumber\\ &\qquad -\beta\big[D_{KL}(p(y_p|x,y)||p(y_p|v,y))\nonumber\\
&\qquad -\gamma D_{KL}(p(y_s|x,y)||p(y_s|v,y)) \big]\Big]\Big\}, \label{eq:stationary_point}
\end{align}
where $Z(x,\beta,\gamma)$ is the normalization term satisfying $\sum_v p(v|x)=1$.
Equation \eqref{eq:stationary_point} together with the marginalization constraints for $p(v|y)$, $p(y_p|v,y)$, and $p(y_s|v,y)$ forms a set of equations which, given initial distributions, can be solved iteratively similarly as in \cite{tpbTIBM00},\cite{ctERSW02}. With $\beta$ and $\gamma$ that admit feasible solutions, the algorithm converges to a stationary point which may not be the global optimum.
Alternatively, the problem can also be solved by a heuristic method, e.g., by extending the agglomerative information bottleneck \cite{stAIB99},\cite{msfmFIBT14} to include the privacy leakage constraint. For example, Algorithm $1$ in \cite{msfmFIBT14} can be modified to include a condition that requires the merging indices $i$ and $j$ to satisfy both information and privacy leakage constraints.
\subsection{Example}
The secure information bottleneck may alternatively be formulated as
\begin{align}
\max_{P_{V|X}} I(Y_p;V,Y)& \label{eq:info_bottleneck_dis}\\
\text{s.t.} \ I(X;V|Y)&\leq R\\
I(Y_s;V,Y)&\leq L
\end{align}
for which the maximum in \eqref{eq:info_bottleneck_dis} corresponds to the maximum achievable information for given $R$ and $L$ and may be termed as the information-rate-leakage function $D'_{\max}(R,L)$.
In the following, we consider two simple examples under the assumptions that $Y=\emptyset$ and $X-Y_s-Y_p$ forms a Markov chain, in which we can express $D'_{\max}(R,L)$ in closed form.
\emph{(i) Binary source}:
Let $X \sim \text{Bernoulli}(1/2)$, $Y_s$ be an output of a BSC($p$), $p \in [0,1/2]$, with input $X$, and $Y_p$ be an output of a BSC($q$), $q \in [0,1/2]$, with input $Y_s$. The maximum achievable information for given $R \in [0, H(X)]$ and $L \in [0, H(Y_s)]$ is given by
\begin{align}
D'_{\max}(R,L) &= H(Y_p)-\max\Big\{h(h^{-1}(H(Y_s)-L)*q), \nonumber \\ &\qquad \qquad h(h^{-1}(H(X)-R)*p*q)\Big\},
\end{align}
where $h(\cdot)$ is a binary entropy function with the inverse $h^{-1} : [0, 1] \rightarrow [0, 1/2]$, and $a*b \triangleq a(1-b)+(1-a)b$.
The achievability proof follows by letting $V$ be an output of a BSC with input $X$. Then letting $D'=I(Y_p;V)=H(Y_p)-h(h^{-1}(H(X|V))*p*q) = H(Y_p)-h(h^{-1}(H(Y_s|V))*q)$. We obtain the result above by substituting $D'$ in the constraints on $R$ and $L$ and using the fact that $h(h^{-1}(u)*q)$ is an increasing function in $u$ for $q \in [0,1/2]$. The converse follows from Mrs. Gerber's lemma \cite{ekNIT11}. For the more detailed proof, please see Appendix \ref{app:proof_secureIB_binary_example}.
\emph{(ii) Gaussian source}: Let $(X,Y_s,Y_p)$ be jointly Gaussian, i.e., $X \sim \mathcal{N}(0,N_x)$, $Y_s= X + \tilde{N}_s$, where $\tilde{N}_s \sim \mathcal{N}(0,N_s) \perp X$, and $Y_p= Y_s + \tilde{N}_p$, where $\tilde{N}_p \sim \mathcal{N}(0,N_p-N_s) \perp Y_s$. The minimum achievable distortion for given $R$ and $L$ is given by
\begin{align}\label{eq:distortion-rate-leakage_function-Gaussian}
D'_{\max}(R,L) &= \min\Big\{\frac{1}{2}\log\Big(\frac{N_x+N_p}{N_x2^{-2R}+N_p}\Big), \nonumber \\ &\qquad \frac{1}{2}\log\Big(\frac{N_x+N_p}{(N_x+N_s)2^{-2L}+N_p-N_s}\Big) \Big\}.
\end{align}
The proof follows similarly as in the binary case where in achievability we let $V$ be jointly Gaussian with $X$, i.e., $V=X+Q$, $Q\sim \mathcal{N}(0,N_q)$, and in the converse we use the conditional EPI \cite{ekNIT11}. Note that if the leakage constraint is neglected, e.g., letting $L \rightarrow \infty$, \eqref{eq:distortion-rate-leakage_function-Gaussian} reduces to the optimal information-rate function in \cite{wmRIOG14}.
\appendices
\section{Proof of Theroem \ref{theorem:inner}}\label{app:proof_inner}
The proof is based on the achievable scheme used in \cite{vpSMSC13} which is a layered (superposition) coding scheme with binning.
\begin{figure*}[!t]
\normalsize
\setcounter{MYtempeqncnt}{\value{equation}}
\setcounter{equation}{5}
\begin{align*}
& H(X^n|J,K,Y_s^n,Z^n) \\
&\overset{(a)}{\leq} H(X^n,E|J,K,U^n(J),V^n(J,K),Y_s^n,Z^n)\\
&\leq H(X^n|U^n,V^n,Y_s^n,Z^n,E) + H(E)\\
&= \mathrm{Pr}(E=0) H(X^n|U^n,V^n,Y_s^n,Z^n,E=0) + \mathrm{Pr}(E=1) H(X^n|U^n,V^n,Y_s^n,Z^n,E=1) + H(E)\\
&\overset{(b)}{\leq}H(X^n|U^n,V^n,Y_s^n,Z^n,E=0) +\delta_{\epsilon} H(X^n)+ h(\delta_{\epsilon}) \\
&\leq H(X^n|U^n,V^n,Y_s^n,Z^n,E=0) + n\delta_{\epsilon} \log|\mathcal{X}| + h(\delta_{\epsilon})\\
&= \sum_{(u^n,v^n,y_s^n,z^n) \in \mathcal{T}_{\epsilon}^{(n)}} p(u^n,v^n,y_s^n,z^n|E=0) H(X^n|U^n=u^n,V^n=v^n,Y_s^n=y_s^n,Z^n=z^n,E=0) + n\delta_{\epsilon} \log|\mathcal{X}| + h(\delta_{\epsilon})\\
&\overset{(c)}{\leq} \sum_{(u^n,v^n,y_s^n,z^n) \in \mathcal{T}_{\epsilon}^{(n)}} p(u^n,v^n,y_s^n,z^n|E=0) \log|\mathcal{T}_{\epsilon}^{(n)}(X|u^n,v^n,y_s^n,z^n)| + n\delta_{\epsilon} \log|\mathcal{X}| + h(\delta_{\epsilon})\\&\overset{(d)}{\leq} n(H(X|U,V,Y_s,Z)+\delta_{\epsilon}')
\end{align*}
\setcounter{equation}{\value{MYtempeqncnt}}
\hrulefill
\vspace*{4pt}
\end{figure*}
For random codebook generation, we fix $P_{V|X}P_{U|V}$ and the reconstruction function $g(\cdot)$.
\begin{itemize}
\item Randomly and independently generate codewords $u^n(j)$ for $j \in [1:2^{n(I(X;U)+\delta_{\epsilon})}]$ according to the product distribution $\prod_{i=1}^n P_U(u_i)$. We distribute the codewords uniformly at random into $2^{n(I(X;U|Y)+2\delta_{\epsilon})}$ equal-sized bins $b_U(w_1)$, $w_1 \in [1:2^{n(I(X;U|Y)+2\delta_{\epsilon})}]$. Each bin contains $2^{n(I(U;Y)-\delta_{\epsilon})}$ codewords, each indexed by $w'$. There exists a one-to-one mapping between
index $j$ and the pair of bin/codeword indices $(w_1,w')$ such that, without loss of generality, we can
identify $j = (w_1,w')$.
\item For each $j$, randomly and conditionally independently generate codewords $v^n(j,k)$ where $k\in [1:2^{n(I(X;V|U)+\delta_{\epsilon})}]$ according to the conditional product distribution $\prod_{i=1}^n P_{V|U}(v_i|u_i(j))$, and distribute these codewords uniformly at random into $2^{n(I(X;V|U,Y)+2\delta_{\epsilon})}$ equal-sized bins $b_V(j,w_2)$, $w_2 \in [1:2^{n(I(X;V|U,Y)+2\delta_{\epsilon})}]$. Each bin $b_V(j,w_2)$ contains $2^{n(I(V;Y|U)-\delta_{\epsilon})}$ codewords, each indexed by $w''$. There exists a one-to-one mapping between index $k$ and the pair of bin/codeword indices $(w_2,w'')$ such that, without loss of generality, we can
identify $k = (w_2,w'')$.
\item The codebooks are revealed to all parties.
\end{itemize}
For encoding,
\begin{itemize}
\item Given $X^n=x^n$, the encoder looks for codeword $u^n(j)$ such that $(x^n,u^n(j))$ are jointly typical. From the covering lemma \cite{ekNIT11}, with high probability, there exist such a codeword since there are more than $2^{nI(X;U)}$ codewords $u^n(j)$. If there are more than one, the encoder selects one with the smallest index $j$.
\item Then based on $x^n$ and the chosen $u^n(j)$, the encoder looks for a codeword $v^n(j,k)$ such that $(x^n,u^n(j)),v^n(j,k)$ are jointly typical. From the covering lemma, with high probability, there exists such a codeword since there are more than $2^{nI(X;V|U)}$ codewords $v^n(j,k)$ for each $j$. If there are more than one, the encoder selects one with the smallest index $k$.
\item The encoder sends the bin indices $w_1$ and $w_2$ of the chosen codewords to the decoder. The total rate is thus equal to $I(X;U|Y)+I(X;V|U,Y)+4\delta_{\epsilon}=I(X;V|Y)+4\delta_{\epsilon}$, where the equality follows from the Markov chain $U-V-X-Y$.
\end{itemize}
For decoding,
\begin{itemize}
\item Based on $Y^n=y^n$ and the bin indices $(w_1,w_2)$, the decoder looks for a unique codeword $u^n(j)$ in bin $b_U(w_1)$ such that $(y^n,u^n(j))$ are jointly typical. From the packing lemma \cite{ekNIT11}, there exists such a codeword $u^n(j)$ with high probability since there are less than $2^{nI(U;Y)}$ codewords $u^n(j)$ in each bin $b_U(w_1)$.
\item Then based on the decoded $u^n(j)$, the decoder looks for a unique codeword $v^n(j,k)$ in bin $b_V(j,w_2)$ such that $(y^n,u^n(j),v^n(j,k))$ are jointly typical. From the packing lemma, there exists such a codeword $v^n(j)$ with high probability since there are less than $2^{nI(V;Y|U)}$ codewords $v^n(j,k)$ in each bin $b_V(j,w_2)$.
\item The decoder reconstructs $\hat{y}_p^n$ such that $\hat{y}_{p,i}=g(v_i(j,k),y_i)$ for $i=1,\ldots,n$.
\end{itemize}
Let $J=(W_1,W')$ and $K=(W_2,W'')$ be the indices associated with the chosen codewords $U^n(J)$ and $V^n(J,K)$. From LLN, we have that the sequences $(X^n,Y^n,Y_p^n,Y_s^n,Z^n,U^n(J),V^n(J,K))$ are jointly typical with high probability. Thus, using similar arguments as in \cite{wzTRDF76} for a bounded distortion measure, the distortion constraint is satisfied if $D \geq E[d(Y_p,g(V,Y))]$.
Before proceeding with the analysis of the leakage rate, we give a lemma which provides a bound on the
$n$-letter conditional entropy based on properties of jointly typical sequences.
\begin{lemma} \label{lemma:lower_bound}
Let the index $J$ and the pair of indices $(J,K)$ be the indices specifying codewords $U^n$ and $V^n$, respectively. If $\mathrm{Pr}((X^n,U^n(J),V^n(J,K),Y_s^n,Z^n) \in \mathcal{T}_{\epsilon}^{(n)}) \rightarrow 1$ as $n \rightarrow \infty$, we have that $H(X^n|J,K,Y_s^n,Z^n) \leq n(H(X|U,V,Y_s,Z)+\delta_{\epsilon})$.
\end{lemma}
To prove the lemma, let $E$ be a binary random variable taking value $0$ if $(X^n,U^n(J),V^n(J,K),Y_s^n,Z^n) \in \mathcal{T}_{\epsilon}^{(n)}$, and $1$ otherwise. Since $(X^n,U^n(J),V^n(J,K),Y_s^n,Z^n) \in \mathcal{T}_{\epsilon}^{(n)}$ with high probability, we have $\mathrm{Pr}(E=1) \leq \delta_{\epsilon}$. The proof steps are given on top of the previous page
where step $(a)$ follows from the fact that given the codebook, $U^n$ and $V^n$ are functions of $J$ and $(J,K)$, $(b)$ follows from $\mathrm{Pr}(E=1) \leq \delta_{\epsilon}$ where $h(\cdot)$ is the binary entropy function, and $(c)$ and $(d)$ follow from the property of jointly typical set \cite{ekNIT11} with $\delta_{\epsilon}, \delta_{\epsilon}' \rightarrow 0$ as $\epsilon \rightarrow 0$, and $\epsilon \rightarrow 0$ as $n \rightarrow \infty$.
The privacy leakage average over all randomly chosen codebooks can be bounded as follows.
\begin{align*} \label{eq:bound_step}
&I(Y_s^n;W_1,W_2,Z^n)= H(Y_s^n) -H(Y_s^n|W_1,W_2,Z^n) \\
&=H(Y_s^n)-H(Y_s^n,X^n|W_1,W_2,Z^n) \\ &\qquad +H(X^n|W_1,W_2,Y_s^n,Z^n) \\
&\leq H(Y_s^n)-H(Y_s^n,X^n|J,Z^n)+H(W_2) \\ &\qquad +H(X^n|J,K,Y_s^n,Z^n) \\ &\qquad +I(X^n;W',W''|W_1,W_2,Y_s^n,Z^n) \\
&\leq H(Y_s^n)-H(Y_s^n,X^n,Z^n)+H(J)+H(Z^n|J)+H(W_2) \\&\qquad +H(X^n|J,K,Y_s^n,Z^n) +H(W',W''|W_1,W_2,Y_s^n,Z^n) \\
&\overset{(a)}{\leq} n[-H(X,Z|Y_s)+I(X;U)+H(Z|U)+I(X;V|U,Y) \\& \qquad +H(X|U,V,Y_s,Z)+\delta_{\epsilon}'] \\&\qquad +H(W',W''|W_1,W_2,Y_s^n,Z^n)\\
&\overset{(b)}{\leq} n[P+\delta_{\epsilon}'] + I(W',W'';Y^n|W_1,W_2,Y_s^n,Z^n) + n\epsilon_n \\
&\leq n[P+\delta_{\epsilon}'] + H(Y^n|Y_s^n,Z^n) - H(Y^n|J,K,Y_s^n,Z^n)+ n\epsilon_n \\
&\overset{(c)}{\leq} n[P+H(Y|Y_s,Z)+ \delta_{\epsilon}'-H(Y|U,V,Y_s,Z)+\delta_{\epsilon}]+ n\epsilon_n \\
&\overset{(d)}{=} n[I(Y_s;V,Y) + I(Z;X,Y_s|U)-I(Y;X,Y_s|U) \\
&\qquad -I(X;Z|V,Y_s,Y)+I(X;Y|Y_s,Z) + \delta_{\epsilon}''] \\
&\leq n[L+ \delta_{\epsilon}'']
\end{align*}
if $L \geq I(Y_s;V,Y) + I(Z;X,Y_s|U)-I(Y;X,Y_s|U) -I(X;Z|V,Y_s,Y)+I(X;Y|Y_s,Z)$,
where $(a)$ follows from memoryless property of the sources, from the codebook generation with $J\in[1:2^{n(I(X;U)+\delta_{\epsilon})}]$, $W_2 \in [1:2^{n(I(X;V|U,Y)+2\delta_{\epsilon})}]$, and from the bounds $H(Z^n|J) \leq n(H(Z|U)+\delta_{\epsilon})$ and $H(X^n|J,K,Y_s^n,Z^n) \leq n(H(X|U,V,Y_s,Z)+\delta_{\epsilon})$ which can be shown similarly as in Lemma \ref{lemma:lower_bound}, $(b)$ follows from defining $P = -H(X,Z|Y_s)+I(X;U)+H(Z|U)+I(X;V|U,Y) +H(X|U,V,Y_s,Z)$ and Fano's inequality $H(W',W''|W_1,W_2,Y^n,Y_s^n,Z^n) \leq n\epsilon_n$ which holds since given the codebook and $W_1,W_2,Y^n$, the decoder can decode $(W',W'')$ with high probability, $(c)$ follows from Lemma~\ref{lemma:entropy_bound} below, and $(d)$ follows from the definition of $P$ and the Markov chain $U-V-X-(Y,Y_p,Y_s,Z)$.
\begin{lemma}\label{lemma:entropy_bound}
Given the codebook where $J$ is the codeword index of $U^n$, and $(J,K)$ is the codeword index of $V^n$, if $\mathrm{Pr}((Y^n,U^n(J),V^n(J,K),Y_s^n,Z^n) \in \mathcal{T}_{\epsilon}^{(n)}) \rightarrow 1$ as $n \rightarrow \infty$, we have that $H(Y^n|J,K,Y_s^n,Z^n) \geq n[H(Y|U,V,Y_s,Z)-\delta_{\epsilon}]$
\end{lemma}
\textit{Proof of Lemma \ref{lemma:entropy_bound}} We consider the following bound.
\begin{align*}
&H(Y^n|J,K,Y_s^n,Z^n) \\
&= H(Y^n,J,K,Y_s^n,Z^n)-H(J,K)-H(Y_s^n,Z^n|J,K)\\
&\geq nH(Y,Y_s,Z) + I(X^n;J,K|Y^n,Y_s^n,Z^n)-H(J,K)\\&\qquad -H(Y_s^n,Z^n|J,K)\\
&= nH(X,Y,Y_s,Z) - H(X^n|J,K,Y^n,Y_s^n,Z^n)-H(J,K)\\&\qquad -H(Y_s^n,Z^n|J,K)\\
&\overset{(a)}{\geq} n[H(X,Y,Y_s,Z)-H(X|U,V,Y,Y_s,Z)-I(X;U,V)\\ &\qquad -H(Y_s,Z|U,V)-\delta_{\epsilon}]\\
&\overset{(b)}{=} n[H(Y|U,V,Y_s,Z)-\delta_{\epsilon}],
\end{align*}
where $(a)$ follows from the codebook generation where $J \in [1:2^{n(I(X;U)+ \delta_{\epsilon})}]$ and $K \in [1:2^{n(I(X;V|U)+ \delta_{\epsilon})}]$, and from bounding $H(X^n|J,K,Y^n,Y_s^n,Z^n)$ and $H(Y_s^n,Z^n|J,K)$ similarly as in Lemma \ref{lemma:lower_bound}, and $(b)$ follows from the Markov chain $(U,V)-X-(Y,Y_s,Z)$.
The cardinality bounds can be proved using the support lemma \cite{ckITCT11}.
\section{Proof of Theorem \ref{theorem:outer}}\label{app:proof_outer}
Let $(R,D,L)$ be an achievable tuple. Define $U_i \triangleq (W,Z^{i-1},Y_{i+1}^n)$, $V_i \triangleq (W,Z^{i-1},Y^{n\setminus i})$ and $T_i=(W,Y_{i+1}^n,Y_s^{n\setminus i},Z^{n\setminus i})$ which satisfy $(V_i,T_i)-X_i-(Y_i,Y_{p,i},Y_{s,i},Z_i)$ and $U_i-V_i-(T_i,X_i,Y_i,Y_{p,i},Y_{s,i},Z_i)$ for all $i=1,\ldots,n$. From properties of the entropy function we have that
\begin{align*}
n(R&+\delta_n) \geq H(W)\\
&\geq I(X^n,Z^n;W|Y^n)\\
&= \sum_{i=1}^n H(X_i,Z_i|Y_i) -H(X_i,Z_i|W,X^{i-1},Z^{i-1},Y^n)\\
&\overset{(a)}{\geq} \sum_{i=1}^n H(X_i,Z_i|Y_i) -H(X_i,Z_i|V_i,Y_i)\\
&\geq \sum_{i=1}^n I(X_i;V_i|Y_i),
\end{align*}
where $(a)$ follows from the definition of $V_i$,
and
\begin{align*}
n(D+\delta_n) &\geq nE[d^{(n)}(Y_p^n,g^{(n)}(W,Y^n))]\\
&= \sum_{i=1}^n E[d(Y_{p,i},g^{(n)}_i(W,Y^n))]\\
&\overset{(a)}{=} \sum_{i=1}^n E[d(Y_{p,i},g_i(V_i,Y_i))],
\end{align*}
where $(a)$ follows from the definition of $V_i$ which includes $(W,Y^{n\setminus i})$, implying that there exists a function $g_i(\cdot)$ such that $g_i(V_i,Y_i)=g^{(n)}_i(W,Y^n)$,
and finally
\begin{align*}
n(L&+\delta_n) \geq I(Y_s^n;W,Z^n)\\
&= H(Y_s^n) - H(Y_s^n,X^n|W,Z^n) +H(X^n|W,Y_s^n,Z^n)\\
&= H(Y_s^n) - H(Y_s^n,X^n|W) +I(Y_s^n,X^n;Z^n|W)\\&\qquad +H(X^n|W,Y_s^n,Z^n)\\
&= H(Y_s^n) - H(Y_s^n,X^n|W,Y^n)-I(Y_s^n,X^n;Y^n|W) \\&\qquad +I(Y_s^n,X^n;Z^n|W)+H(X^n|W,Y_s^n,Z^n)\\
&\overset{(a)}{=} H(Y_s^n) - I(X^n;Y_s^n,Z^n|W,Y^n)-H(Y_s^n|X^n,Y^n)\\&\qquad-I(Y_s^n,X^n;Y^n|W) +I(Y_s^n,X^n;Z^n|W)\\
&\qquad +I(X^n;Y^n|W,Y_s^n,Z^n)\\
&\overset{(b)}{=} H(Y_s^n) - I(X^n;Y_s^n,Z^n|W,Y^n)-H(Y_s^n|X^n,Y^n)\\&\qquad-I(Z^n;W)+I(Y^n;W)-I(Y_s^n,X^n;Y^n) \\&\qquad +I(Y_s^n,X^n;Z^n)+I(X^n;Y^n|W,Y_s^n,Z^n),
\end{align*}
where $(a)$ follows from the Markov chain $Y_s^n-(X^n,Y^n)-W$ and $(b)$ follows from the Markov chain $(Y^n,Z^n)-(Y_s^n,X^n)-W$.
Continuing the chain of inequalities, we get
\begin{align*}
&n(L+\delta_n) \\
& \overset{(c)}{\geq} \sum_{i=1}^n H(Y_{s,i})-H(Y_{s,i},Z_i|V_i,Y_i) + H(Z_i|X_i,Y_i,Y_{s,i})\\&\qquad -I(Z_i;W|Z^{i-1})+I(Y_i;W|Y_{i+1}^n)-I(Y_{s,i},X_i;Y_i) \\&\qquad +I(Y_{s,i},X_i;Z_i) + I(X_i;Y_i|W,Y_{i+1}^n,Y_s^n,Z^n)\\
&\overset{(d)}{=} \sum_{i=1}^n I(Y_{s,i};V_i,Y_i)-I(Z_i;X_i|V_i,Y_i,Y_{s,i}) -I(Y_{s,i},X_i;Y_i) \\&\qquad +I(Y_{s,i},X_i;Z_i) -I(Z_i;W,Z^{i-1}) +I(Y_i;W,Y_{i+1}^n)\\
&\qquad +I(X_i;Y_i|T_i,Y_{s,i},Z_i)\\
&\overset{(e)}{=} \sum_{i=1}^n I(Y_{s,i};V_i,Y_i)-I(Z_i;X_i|V_i,Y_i,Y_{s,i}) -I(Y_{s,i},X_i;Y_i) \\&\qquad +I(Y_{s,i},X_i;Z_i) -I(Z_i;W,Y_{i+1}^n,Z^{i-1}) \\&\qquad +I(Y_i;W,Y_{i+1}^n,Z^{i-1})+I(X_i;Y_i|T_i,Y_{s,i},Z_i)\\
&\overset{(f)}{=} \sum_{i=1}^n I(Y_{s,i};V_i,Y_i)-I(Z_i;X_i|V_i,Y_i,Y_{s,i}) \\&\qquad -I(Y_{s,i},X_i;Y_i|U_i) +I(Y_{s,i},X_i;Z_i|U_i)\\&\qquad +I(X_i;Y_i|T_i,Y_{s,i},Z_i),
\end{align*}
where $(c)$ follows from the Markov chain $(Y_s^n,Z^n)-(X^n,Y^n)-W$, the definition of $V_i$, and the Markov chain $Y_i-(W,X_i,Y_{i+1}^n,Y_s^n,Z^n)-X^{n \setminus i}$, $(d)$ follows from the Markov chain $Z_i-(X_i,Y_i,Y_{s,i})-V_i$, the definition of $T_i$, and the facts that $Z_i$ is independent of $Z^{i-1}$ and $Y_i$ is independent of $Y_{i+1}^n$, $(e)$ follows from the Csiszar's sum identity \cite{ekNIT11}, and $(f)$ follows from the definition of $U_i$ and the Markov chain $(Y_i,Z_i)-(X_i,Y_{s,i})-U_i$.
We ends the proof by the standard time-sharing argument and letting $n \rightarrow \infty$ and $\delta_n \rightarrow 0$.
\section{Proof of Gaussian with Quadratic Distortion Example}\label{app:proof_Gaussian_example}
Recalling that $X \sim \mathcal{N}(0,N_x)$, $Y_s= X + \tilde{N}_s$, where $\tilde{N}_s \sim \mathcal{N}(0,N_s) \perp X$, and $Y_p= Y_s + \tilde{N}_p$, where $\tilde{N}_p \sim \mathcal{N}(0,N_p-N_s) \perp Y_s$.
\emph{Proof of Achievability}: Let us choose $V=X+Q, Q\sim \mathcal{N}(0,N_q)$ independent of $X$, and choose $g(V)$ to be an MMSE estimate of $Y_p$ given $V$.
With these choices of $V$ and $g(\cdot)$, it can be shown that
\begin{align*}
I(X;V) &= h(V)-h(V|X)\\
&= \frac{1}{2}\log\bigg(\frac{N_x+N_q}{N_q}\bigg),
\end{align*}
and
\begin{align*}
I(Y_s;V) &= h(Y_s)-h(Y_s|V)\\
&= \frac{1}{2}\log\bigg(\frac{N_x+N_s}{N_s + \frac{N_xN_q}{N_x+N_q}}\bigg),
\end{align*}
and lastly
\begin{align*}
E[d(Y_p,g(V))]&=E[(Y_p-g(V))^2]\\
&=N_p+ \frac{N_xN_q}{N_x+N_q}
\end{align*}
Letting $D=E[d(Y_p,g(V))]=N_p+ \frac{N_xN_q}{N_x+N_q}$ and substituting it into the constraints on $R$ and $L$ complete the achievablity part.
\emph{Proof of Converse}: From the problem formulation, the joint PMF $P_{X^n,Y_p^n,Y_s^n,W,\hat{Y}_p^n}$ is given by
\begin{align*}
P_{X^n,Y_s^n}P_{Y_p^n|Y_s^n} P_{W|X^n}1_{\{\hat{Y}_p^n = g^{(n)}(W)\}}.
\end{align*}
From the EPI, we have that
\begin{equation}
2^{\frac{2}{n}h(Y_s^n|W)} \geq 2^{\frac{2}{n}h(X^n|W)} +2^{\frac{2}{n}h(\tilde{N}_s^n)}\label{eq:EPI1}
\end{equation}
and
\begin{align}
2^{\frac{2}{n}h(Y_p^n|W)} &\geq 2^{\frac{2}{n}h(Y_s^n|W)} +2^{\frac{2}{n}h(\tilde{N}_p^n)}\label{eq:EPI2}\\
&\geq 2^{\frac{2}{n}h(X^n|W)} +2^{\frac{2}{n}h(\tilde{N}_s^n)}+2^{\frac{2}{n}h(\tilde{N}_p^n)},\label{eq:EPI3}
\end{align}
where the last inequality follows from \eqref{eq:EPI1}.
Then it follows that
\begin{align}
n(R + \delta_n) &\geq H(W )\nonumber\\
&\geq I(X^n;W)\nonumber\\
&= h(X^n)-h(X^n|W)\nonumber\\
&\geq h(X^n) \nonumber\\&\qquad - \frac{n}{2}\log(2^{\frac{2}{n}h(Y_p^n|W)} -2^{\frac{2}{n}h(\tilde{N}_s^n)}-2^{\frac{2}{n}h(\tilde{N}_p^n)}) ,\label{eq:R_converse}
\end{align}
where the last inequality follows from \eqref{eq:EPI3}.
\begin{align}
n(L + \delta_n) &\geq I(Y_s^n;W)\nonumber\\
&= h(Y_s^n)-h(Y_s^n|W)\nonumber\\
&\geq h(Y_s^n)- \frac{n}{2}\log(2^{\frac{2}{n}h(Y_p^n|W)} -2^{\frac{2}{n}h(\tilde{N}_p^n)}), \label{eq:L_converse}
\end{align}
where the last inequality follows from \eqref{eq:EPI2}.
\begin{align}
D + \delta_n &\geq E[d^{(n)}(Y_p^n,g^{(n)}(W))]\nonumber \\
&= \frac{1}{n} \sum_{i=1}^n E[(Y_{p,i}-g_i^{(n)}(W))^2]\nonumber\\
&= \frac{1}{2\pi e}2^{\log(\frac{2 \pi e}{n} \sum_{i=1}^n E[(Y_{p,i}-g_i^{(n)}(W))^2])}\nonumber\\
&\overset{(a)}{\geq} \frac{1}{2\pi e}2^{\frac{1}{n}\sum_{i=1}^n\log( 2 \pi eE[(Y_{p,i}-g_i^{(n)}(W))^2])}\nonumber\\
&\overset{(b)}{\geq} \frac{1}{2\pi e}2^{\frac{1}{n}\sum_{i=1}^n\log( 2 \pi e E[\text{var}(Y_{p,i}|W)])}\nonumber\\
&\geq \frac{1}{2\pi e}2^{\frac{2}{n}\sum_{i=1}^n h(Y_{p,i}|W)}\nonumber\\
&\geq \frac{1}{2\pi e} 2^{\frac{2}{n} h(Y_p^n|W)}, \label{eq:D_converse}
\end{align}
where $(a)$ follows from Jensen's inequality \cite{ekNIT11} and the fact that $\log(\cdot)$ is a concave function, and $(b)$ follows from the fact that $E[\text{var}(Y_{p,i}|W)]$ is the MMSE over all possible estimator of $Y_{p,i}$ for each $i=1,\ldots,n$.
Combining \eqref{eq:D_converse} with \eqref{eq:R_converse} and \eqref{eq:L_converse}, and letting $n \rightarrow \infty$ and $\delta_n \rightarrow 0$ complete the converse part.
\section{Proof of Secure Information Bottleneck Example: binary source}\label{app:proof_secureIB_binary_example}
Recalling that $X \sim \text{Bernoulli}(1/2)$, $Y_s$ is an output of a BSC($p$) with input $X$, and $Y_p$ is an output of a BSC($q$) with input $Y_s$.
\emph{Proof of Achievability}: Let $V$ be an output of a BSC with input $X$. We have that
\begin{align}
H(Y_p|V)&=h(h^{-1}(H(X|V))*p*q)\label{eq:secureIB_achive1}\\
&=h(h^{-1}(H(Y_s|V))*q).\label{eq:secureIB_achive2}
\end{align}
We let $D'=H(Y_p)-H(Y_p|V)$. Combining \eqref{eq:secureIB_achive1} with the constraint $R \geq H(X)-H(X|V)$ and noting that $u*q$ is an increasing function in $u \in [0,1/2]$ for some fixed $q \in [0,1/2]$, we get $D' \leq H(Y_p)-h(h^{-1}(H(X)-R)*p*q))$. Similarly, combining \eqref{eq:secureIB_achive2} with $L \geq H(Y_s)-H(Y_s|V)$ gives $D' \leq H(Y_p)-h(h^{-1}(H(Y_s)-L)*q))$.
\emph{Proof of Converse}: From Mrs. Gerber's lemma, we have
\begin{align}
H(Y_s|V) &\geq h(h^{-1}(H(X|V))*p)\label{eq:mrsGerber1}\\
H(Y_p|V) &\geq h(h^{-1}(H(Y_s|V))*q).\label{eq:mrsGerber2}
\end{align}
Since $h(h^{-1}(u)*q)$ is an increasing function in $u$ for $q \in [0,1/2]$, combining \eqref{eq:mrsGerber1} and \eqref{eq:mrsGerber2} gives
\begin{equation}
H(Y_p|V) \geq h(h^{-1}(H(X|V))*p*q). \label{eq:mrsGerber3}
\end{equation}
Then we have that
\begin{align*}
D' &\leq I(Y_p;V)\\
&\overset{(a)}{\leq} H(Y_p)-h(h^{-1}(H(Y_s|V))*q)\\
&\overset{(b)}{\leq} H(Y_p)- h(h^{-1}(H(Y_s)-L)*q),
\end{align*}
where $(a)$ follows from \eqref{eq:mrsGerber2} and $(b)$ follows from the facts that $h(h^{-1}(H(Y_s|V))*q)$ is an increasing function in $H(Y_s|V)$ and that $L \geq H(Y_s)-H(Y_s|V)$.
Similarly, we have
\begin{align*}
D' &\leq I(Y_p;V)\\
&\overset{(a)}{\leq} H(Y_p)- h(h^{-1}(H(X|V))*p*q)\\
&\overset{(b)}{\leq} H(Y_p)- h(h^{-1}(H(X)-R)*p*q),
\end{align*}
where $(a)$ follows from \eqref{eq:mrsGerber3} and $(b)$ follows from the facts that $h(h^{-1}(H(X|V))*p*q)$ is an increasing function in $H(X|V)$ and that $R \geq H(X)-H(X|V)$.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\noindent
Given two binary vectors $\uuu = (u_1,u_2,\ldots,u_n)$ and
$\vvv = (v_1,v_2,\ldots,v_n)$, we define their
\emph{product $\uuu\vvv$} componentwise, namely\vspace{-0.50ex}
\be{product-def}
\uuu\vvv
\ \ \deff \,\
(u_1 v_1, u_2 v_2, \ldots, u_n v_n)
\ee
where $u_1v_1, u_2 v_2, \ldots, u_nv_n$ are computed in $\mathrm{GF}(2)$.
Note that the product operation in \eq{product-def} distri\-butes
over addition in $\F_2^n$. Thus \eq{product-def} turns the vector
space $\F_2^n$ into an algebra $\cA_n$ over $\F_2$.
This~algebra $\cA_n$ is unital, associative, and commutative.
Given a set $X \subseteq \F_2^n$, we define the square of $X$
as the set of products of the elements in $X$.~Explicitly,
$X^2$ is defined as follows:\vspace{-1.00ex}
\be{square-def}
X^2 \,\ \deff\,\
\bigl\{\, \uuu \vvv \,:\, \uuu,\vvv \in X ~\text{and}~ \uuu \ne \vvv\bigr\}
\ee
The following lemmas follow straightforwardly from the
definitions in \eq{product-def} and \eq{square-def}, along
with the fact that
$\cA_n$ is a commutative algebra.
We let $\Span{X}$ denote the linear span over $\F_2$
of a set $X \subseteq \F_2^n$.
\vspace{0.50ex}
\begin{lemma}
\label{L1}
$|X^2| \,\le\, |X|\bigl(|X|-1\bigr)/2$.
\vspace{-0.75ex}
\end{lemma}
\begin{proof}
If $|X| = r$, then $X^2$ consists of the $\binom{r}{2}$ vectors
$\uuu \vvv = \vvv \uuu$ for some $\uuu \ne \vvv$ in $X$.
Some of these vectors may coincide.
\end{proof}
\begin{lemma}
\label{L3}
Let $\uuu,\vvv_1,\vvv_2,\vvv_3 \in \F_2^n$. \,If\,
$\vvv_1\vvv_2 + \vvv_1\vvv_3 + \vvv_2\vvv_3 = \zero$,
then
$$
(\uuu + \vvv_1)(\uuu + \vvv_2)
\,+\,
(\uuu + \vvv_2)(\uuu + \vvv_3)
\,+\,
(\uuu + \vvv_3)(\uuu + \vvv_1)
\ = \
\uuu
$$
\vspace{-4.50ex}
\end{lemma}
\begin{proof}
Follows by straightforward verification using distributivity
and commutativity in $\cA_n$.
\end{proof}
\vspace{1.00ex}
We now show how the foregoing lemmas can be used to establish
a bound on the redundancy of binary $k$-server PIR codes for $k \ge 3$.
These codes are defined in~\cite{FVY15a,FVY15b}
as follows.
\begin{definition}
\label{PIRmatrix}
Let $\eee_i$ denote the binary (column) vector with\/ $1$ in position $i$
and zeros elsewhere. We say that an $s \times n$ binary matrix $G$ has\/
{\dfn property $\!\cP_k$} if for all $i \in \![s]$, there exist $k$ disjoint
sets of columns of $G$ that add up to $\eee_i$. A matrix that has
property $\cP_k$ is also said to be a {\dfn $k$-server PIR matrix}.
A binary linear code\/ $\C$ of length $n$ and dimension $s$ is
called
a {\dfn $k$-server PIR code} if there exists a~generator
matrix $G$ for\, $\C$ with property $\cP_k$.
\end{definition}
For much more on $k$-server PIR codes and their applications
in reducing the storage overhead of private information retrieval,
see~\cite{FVY15a,FVY15b}.
In particular, it is shown in~\cite{FVY15b} that, given
a $k$-server PIR code of length $s+r$ and dimension $s$,
the storage overhead of \emph{any} linear $k$-server PIR protocol
can be reduced from $k$ to $(s+r)/s$. Moreover, for every fixed $k$,
there exist $k$-server PIR codes whose rate (and, hence, storage
overhead) approaches $1$ as their dimension $s$ grows. However,
exactly \emph{how fast} the resulting storage overhead tends to $1$
as $s \to \infty$ was heretofore unknown. For every fixed $k$,
Fazeli, Vardy, and Yaakobi~\cite{FVY15a,FVY15b} construct $k$-server
PIR codes with redundancy $r$ bounded by
$r \le k \sqrt{s}\bigl(1 + o(1)\bigr)$.
But the question of whether codes with even smaller redundancy
exist was left open in~\cite{FVY15a,FVY15b}. The following
theorem shows that the redundancy $O(\sqrt{s})$
of the codes~con\-structed in~\cite{FVY15a,FVY15b} is asymptotically optimal.
\begin{theorem}
\label{main}
Let\, $\C$ be a $3$-server PIR code of length $n$ and dimension $s$.
Let $r = n-s$ denote the redundancy of\, $\C$. Then\,
$r(r\,{-}\,1) \ge 2s$.
\end{theorem}
\begin{proof}
Let $G$ be an $s \times n$ generator matrix for $\C$ with property $\cP_3$,
and let $\xxx_1,\xxx_2,\ldots,\xxx_n$ denote the columns of $G$.
By definition, for each $i \,{\in}\, [s]$, there exist $3$ disjoint subsets
of $\{\xxx_1,\xxx_2,\ldots,\xxx_n\}$ that add up to $\eee_i$. Let
$R_1,R_2,R_3 \subset [n]$ denote the corresponding sets of indices.
Then we can write
\be{thm1-a}
\eee_i
\; = \
{\displaystyle\sum}_{j \in R_1} \xxx_j
\ = \
{\displaystyle\sum}_{j \in R_2} \xxx_j
\ = \
{\displaystyle\sum}_{j \in R_3} \xxx_j
\vspace{0.25ex}
\ee
It is easy to see from \Dref{PIRmatrix} that $G$ has full column rank.
Hence some $s$ columns of $G$ are linearly independent, and we assume
w.l.o.g.\ that these are the first $s$ columns. Consequently, there
exists a nonsingular $s \times s$ matrix $A$ such that
\be{G'}
G'
\ \ \deff \:\
AG
\ = \
\bigl[\,I_s\,|\,P\,\bigr]
\ee
where $I_s$ is the $s \times s$ identity matrix and $P$ is an
$s \times r$ matrix. Let $\xxx'_1,\xxx'_2,\ldots,\xxx'_n$ denote
the columns of $G'$, with $\xxx'_j = \eee_j$ for $j = 1,2,\ldots,s$.
Then it follows from \eq{thm1-a} that
\be{thm1-b}
\aaa_i
\; = \
{\displaystyle\sum}_{j \in R_1} \xxx'_j
\ = \
{\displaystyle\sum}_{j \in R_2} \xxx'_j
\ = \
{\displaystyle\sum}_{j \in R_3} \xxx'_j
\vspace{0.25ex}
\ee
where $\aaa_1,\aaa_2,\ldots,\aaa_s$ are the columns of $A$.
Note that $\dim \Span{\aaa_1,\aaa_2,\ldots,\aaa_s} = s$,
since the matrix $A$ is nonsingular. Let us now further
define\vspace{0.75ex}
\begin{align}
\label{S-def}
S_1 &\,=\, R_1 \cap [s], &
~
S_2 &\,=\, R_2 \cap [s], &
~
S_3 &\,=\, R_3 \cap [s]
\\[0.75ex]
\label{T-def}
T_1 &\,=\, R_1 \cap \bigl([n]{\setminus}[s]\bigr), &
~
T_2 &\,=\, R_2 \cap \bigl([n]{\setminus}[s]\bigr), &
~
T_3 &\,=\, R_3 \cap \bigl([n]{\setminus}[s]\bigr)
\\[0.75ex]
\label{v-def}
\vvv_1 &\,=\, \sum_{j\in S_1}\!\xxx'_j \ = \sum_{j\in S_1}\!\eee_j ~~~~&
\vvv_2 &\,=\, \sum_{j\in S_2}\!\xxx'_j \ = \sum_{j\in S_2}\!\eee_j ~~~~&
\vvv_3 &\,=\, \sum_{j\in S_3}\!\xxx'_j \ = \sum_{j\in S_3}\!\eee_j \hspace*{5.00ex}
\\[-2.50ex]
\nonumber
\end{align}
With this notation, we can rewrite \eq{thm1-b} as follows:\vspace{1.00ex}
\be{thm1-c}
\aaa_i + \vvv_1
\ = \
{\displaystyle\sum}_{j \in T_1} \xxx'_j
\hspace{8.00ex}
\aaa_i + \vvv_2
\ = \
{\displaystyle\sum}_{j \in T_2} \xxx'_j
\hspace{8.00ex}
\aaa_i + \vvv_3
\ = \
{\displaystyle\sum}_{j \in T_3} \xxx'_j
\hspace*{4.00ex}
\ee
Finally, let us define
$X \,\ \deff\ \bigl\{\xxx'_{s+1},\xxx'_{s+2},\ldots,\xxx'_{n}\bigr\}$.
Then it follows from~\eq{thm1-c}~that
$\aaa_i + \vvv_1$, $\aaa_i + \vvv_2$,~and $\aaa_i + \vvv_3$
belong to $\Span{X}$.
We are now
ready to use Lemmas \ref{L1} and \ref{L3} in order
to complete the proof.
Since the sets $S_1,S_2,S_3$ are disjoint, it follows from
\eq{v-def} that the supports of $\vvv_1,\vvv_2,\vvv_3$
are also disjoint. In other words,
$
\vvv_1 \vvv_2 = \vvv_1 \vvv_3 = \vvv_2 \vvv_3 = \zero
$.
Using \Lref{L3}, we conclude that
\begin{eqnarray*}
\aaa_i
& = &
\hspace*{-0.25ex}(\aaa_i + \vvv_1)(\aaa_i + \vvv_2)
\ + \
(\aaa_i + \vvv_2)(\aaa_i + \vvv_3)
\ + \
(\aaa_i + \vvv_3)(\aaa_i + \vvv_1)
\\
& = &\hspace*{-0.75ex
\left(\sum_{j \in T_1} \xxx'_j\right)\!\left(\sum_{j \in T_2} \xxx'_j\right)
\ + \
\left(\sum_{j \in T_2} \xxx'_j\right)\!\left(\sum_{j \in T_3} \xxx'_j\right)
\ + \
\left(\sum_{j \in T_3} \xxx'_j\right)\!\left(\sum_{j \in T_1} \xxx'_j\right)
\\[1.00ex]
& = &\hspace*{-0.50ex
\sum_{j \in T_1}\sum_{k \in T_2} \xxx'_j \xxx'_k\:
\ + \
\sum_{j \in T_2}\sum_{k \in T_3} \xxx'_j \xxx'_k\:
\ + \
\sum_{j \in T_3}\sum_{k \in T_1} \xxx'_j \xxx'_k
\end{eqnarray*}
Since the sets $T_1,T_2,T_3$ are disjoint subsets of
$[n]{\setminus}[s]$, all of the products $\xxx'_j \xxx'_k$
above belong to $X^2$. Consequently, it follows that
\smash{$\aaa_i \in \Span{X^2}$} for all $i$. Hence
$$
\dim {\textstyle\Span{X^2} }
\ \ge \
\dim \Span{\aaa_1,\aaa_2,\ldots,\aaa_s}
\ = \
s
$$
But
$
\dim {\textstyle\Span{X^2}}
\le
|X^2|
\le
r(r-1)/2
$,
where we have used \Lref{L1}.
Thus
$r(r-1)/2 \ge s$,~which completes the
proof of the theorem.
\end{proof}
\vspace{1.00ex}
It is shown in~\cite{FVY15a,FVY15b}
that the redundancy of $k$-server PIR codes is non-decreasing
in $k$. That is,~if $\rho(s,k)$ denotes the lowest possible
redundancy of a $k$-server PIR code of dimension $s$,
then
$$
\rho(s,k+1)
\ \ge \
\rho(s,k)
\hspace{7ex}
\text{for all $s \ge 1$ and all $k \ge 2$}
$$
Consequently, the lower bound of \Tref{main} trivially
extends from $3$-server PIR codes to general $k$-server PIR codes
with $k \ge 3$.
\looseness=-1
The following simple construction achieves the
lower bound of \Tref{main} for $k = 3$. Let $r$ be the smallest
integer such that $\binom{r}{2} \ge s$. Take
$G = \bigl[\,I_s\,|\,P\,\bigr]$,
where $P$ is an $s \times r$ matrix whose rows are distinct
binary vectors of weight $2$. Clearly, the rows of $P$ form
a constant-weight binary code with
distance $2$.
By the results of~\cite{FVY15a,FVY15b}, this implies that
$G$ is a $3$-server PIR matrix, and therefore
\be{rho3}
\rho(s,3)
\ = \
\text{the smallest integer $r$ such that\kern1pt\ $r(r\,{-}\,1) \ge 2s$}
\ = \
\left\lceil \sqrt{2s + \sfourth} ~\,+\: \shalf \right\rceil
\ee
It is also shown in~\cite{FVY15a,FVY15b} that for all even $k$,
we have $\rho(s,k) = \rho(s,k{-}1) + 1$. Consequently, \eq{rho3}
determines the lowest possible redundancy of $4$-server PIR codes
as well.
\vspace{9ex}
|
\section{Introduction}
Throughout this work the ground field is supposed to be the complex number field $\mathbb{C}$.
A \textit{quartic double solid} is a projective variety represented as a
the double cover of $\mathbb{P}^3$ branched along a smooth quartic.
It is known that quartic double solids are unirational but not rational \cite{Beauville1977},
\cite{Tikhomirov1986}, \cite{Voisin1988}, \cite{Clemens1991}.
Moreover, a general quartic double solid is not \textit{stably rational}
\cite{Voisin2015a}.
There are also a lot of results related to rationality problems of \emph{singular} quartic double solids
see e.g. \cite{Artin-Mumford-1972}, \cite{Clemens1983}, \cite{Varley1986}, \cite{Debarre1990},
\cite{Przhiyalkovskij-Cheltsov-Shramov-2015}, \cite{CheltsovPrzyjalkowskiShramov2015b}.
The main result of this note is to give a simple proof of the following
\begin{mtheorem}{\bf Theorem.}\label{theorem-main}
Let $X$ be the quartic double solid branched over the surface
\begin{equation*}
\label{equation-1}
x_{1}^3x_{2}+ x_{2}^3x_{3}+ x_{3}^3x_{4}+ x_{4}^3x_{1}=0.
\end{equation*}
Then the intermediate Jacobian $J(X)$ is not a sum of Jacobians of curves.
As a consequence, $X$ is not rational.
\end{mtheorem}
\begin{mtheorem}{\bf Corollary.}\label{corollary-main}
A general quartic double solid is not rational.
\end{mtheorem}
Our proof uses methods of A. Beauville \cite{Beauville2012}, \cite{Beauville2013} and Yu. Zarhin \cite{Zarhin2009}.
The basic idea is to find a sufficiently symmetric variety in the family.
Then the action of the automorphism group provides a good tool to prove
non-decomposability the intermediate Jacobian into a sum of Jacobians of curves
by using purely \textit{group-theoretic} techniques.
Since the Jacobians and their sums form a
closed subvariety of the moduli space of principally polarized abelian varieties,
this shows that a general quartic double solid is not rational\footnote{Recently
V. Przyjalkowski and C. Shramov used similar method
to prove non-rationality of some double quadrics \cite{PrzyjalkowskiShramov2016}.}.
\section{Preliminaries}
\begin{say}{\bf Notation.}
We use standard group-theoretic notation:
if $G$ is a group, then
${\operatorname{z}}(G)$ denotes its center, $[G,G]$ its derived subgroup,
and ${\operatorname{Syl}}_p(G)$ its (some) Sylow $p$-subgroup.
By $\zeta_m$ we denote a primitive $m$-th root of unity.
The group generated by elements $\upalpha_1,\upalpha_2,\dots$
is denoted by $\langle \upalpha_1,\upalpha_2,\dots\rangle$.
\end{say}
\begin{say}
Let $X$ be a three-dimensional smooth projective variety with $H^3(X,\mathscr{O}_X)=0$
and let $J(X)$ be its intermediate Jacobian regarded as a principally polarized abelian variety
(see \cite{Clemens-Griffiths}). Then $J(X)$ can be written, uniquely up to permutations,
as a direct sum
\begin{equation}
\label{equation-decomposition}
J(X)=A_1 \oplus \dots \oplus A_n,
\end{equation}
where $A_1, \dots, A_p$ are indecomposable
principally polarized abelian varieties (see \cite[Corollary 3.23]{Clemens-Griffiths}).
This decomposition induces a decomposition of tangent spaces
\begin{equation}
\label{equation-decomposition-tangent}
\operatorname{T}_{0,J(X)}= \operatorname{T}_{ 0,A_1} \oplus \dots \oplus \operatorname{T}_{0,A_n}
\end{equation}
Now assume that $X$ is acted on by a finite group $G$.
Then $G$ naturally acts on $J(X)$ and $\operatorname{T}_{0,J(X)}$ preserving decompositions \eqref{equation-decomposition}
and \eqref{equation-decomposition-tangent}.
\end{say}
\begin{mtheorem}{\bf Lemma.}\label{lemma-action-80}
Let $C$ be a curve of genus $g\ge 2$ and let $\Gamma\subset \operatorname{Aut}(C)$ be a subgroup
of order $2^k\cdot 5$ whose Sylow $5$-subgroup ${\operatorname{Syl}}_5(\Gamma)$ is normal
in $\Gamma$. Then the following assertions hold:
\begin{enumerate}
\item\label{lemma-action-80-k=2}
if $k=2$, then $g\ge 3$,
\item\label{lemma-action-80-k=4}
if $k=4$, then $g\ge 6$,
\item\label{lemma-action-80-k=5}\label{faithful-action-on-curve}
if $k=5$, then $g\ge 11$.
\end{enumerate}
\end{mtheorem}
\begin{proof}
Let $C':=C/{\operatorname{Syl}}_5(\Gamma)$ and $g':=g(C')$. Let $P_1,\dots, P_n\in C'$ be
all the branch points. By Hurwitz's formula
\[
g+4=5g'+2n.
\]
The group $\Gamma':=\Gamma/ {\operatorname{Syl}}_5(\Gamma)$ of order $2^k$ faithfully acts on $C'$ and permutes $P_1,\dots, P_n$.
\ref{lemma-action-80-k=2}
Assume that $k=g=2$.
Then $g'=0$, $C'\simeq \mathbb{P}^1$, and $n=3$.
At least one of the points $P_1,P_2,P_3$, say $P_1$, must be fixed by $\Gamma'$.
But then $\Gamma'$ must be cyclic (of order $4$) and it cannot leave the set $\{P_1,P_2,P_3\}\subset \mathbb{P}^1$
invariant. This proves
\ref{lemma-action-80-k=2}.
\ref{lemma-action-80-k=4}
Assume that $k=4$ and $g\le 5$. Then $g'\le 1$.
If $g'=0$, then $n\in \{3,\, 4\}$ and
the group $\Gamma'$ of order $16$ acts on $C'\simeq \mathbb{P}^1$ so that
the set $\{P_1,\dots, P_n\}$ is invariant.
This is impossible. If $g'=1$, then, as above, $\Gamma'$ acts on
an elliptic curve $C'$ leaving
a non-empty set of $n\le 2$ points is invariant. This is again impossible and the contradiction proves
\ref{lemma-action-80-k=4}.
\ref{lemma-action-80-k=5}
Finally, let $k=5$ and $g\le 10$. Then $g'\le 2$ and $n\le 7$.
If $g'\le 1$, then we get a contradiction as above.
Let $g'=2$, let $C'\to \mathbb{P}^1$ the the canonical map, and let
$\Gamma''\subset \operatorname{Aut}(\mathbb{P}^1)$ be the image of $\Gamma'$.
Since $\Gamma''$ is a $2$-subgroup in $\operatorname{Aut}(\mathbb{P}^1)$, it is either cyclic or dihedral.
On the other hand, $\Gamma''$
permutes the branch points $Q_1,\dots,Q_6\in \mathbb{P}^1$ so that
the stabilizer of each $Q_i$ is a subgroup in $\Gamma''$ of index $\le 4$.
Clearly, this is impossible.
\end{proof}
\section{Symmetric quartic double solid}
\begin{say}
Let $X$ be the quartic double solid as in Theorem \ref{theorem-main}.
Then $X$ is isomorphic to a hypersurface given by
\begin{equation}
\label{equation}
y^2+x_{1}^3x_{2}+ x_{2}^3x_{3}+ x_{3}^3x_{4}+ x_{4}^3x_{1}=0,
\end{equation}
in the weighted projective space
$\mathbb{P}:=\mathbb{P}(1^4,2)$, where
$x_1,\dots,x_4,y$ are homogeneous coordinates with $\deg x_i=1$, $\deg y=2$.
Let $\upalpha$ be the automorphism of $X$ induced by the diagonal matrix
\[
{\operatorname{diag}} (1,\, \zeta_{40}^{38},\, \zeta_{40}^{4},\, \zeta_{40}^{26};\, \zeta_{40}^{-1})
\]
and let $\upbeta$ be the cyclic permutation $(1,2,3,4)$ of
coordinates $x_1,x_2,x_3,x_4$.
Since
\[
\upbeta\upalpha\upbeta^{-1}= {\operatorname{diag}} (\zeta_{40}^{26}, 1, \zeta_{40}^{38},\, \zeta_{40}^{4};\, \zeta_{40}^{-1})=
{\operatorname{diag}} (1,\, \zeta_{40}^{14},\, \zeta_{40}^{12},\, \zeta_{40}^{18};\, \zeta_{40}^{27})=
\upalpha^{13},
\]
these automorphisms generate the group
\[
G= \langle \upalpha,\, \upbeta \mid \upalpha^{40}=\upbeta^4=1,\hspace{3pt}
\upbeta\upalpha\upbeta^{-1}=\upalpha^{13}\rangle \subset \operatorname{Aut}(X),
\quad G\simeq \mathbb{Z}/40 \rtimes\mathbb{Z}/ 4.
\]
\end{say}
\begin{mtheorem}{\bf Lemma.}\label{subgroup-index10}
Let $G$ be as above. Then we have
\begin{enumerate}
\item \label{subgroup-index10-1}
${\operatorname{z}}(G)=\langle \upalpha^{10}\rangle$ and
$[G,G]=\langle \upalpha^4\rangle$,
\item \label{subgroup-index10-2}
the Sylow $5$-subgroup ${\operatorname{Syl}}_5(G)$ is normal,
\item \label{subgroup-index10-3}
any subgroup in $G$ of index $10$ contains ${\operatorname{z}}(G)$.
\end{enumerate}
\end{mtheorem}
\begin{proof}
\ref{subgroup-index10-1} can be proved by direct computations and \ref{subgroup-index10-2}
is obvious because ${\operatorname{Syl}}_5(G)\subset \langle\upalpha\rangle$.
To prove \ref{subgroup-index10-3} consider a subgroup
$G'\subset G$ of index $10$.
The intersection $G'\cap \langle\upalpha\rangle$ is of index $\le 4$ in $G'$.
Hence $G'\cap \langle\upalpha\rangle$ is a $2$-group of order $\ge 4$ and so
$\upalpha^{10}\in G'\cap \langle\upalpha\rangle$.
\end{proof}
\begin{mtheorem}{\bf Lemma (cf. \cite[0.1(b)]{Voisin1988}).}\label{Lemma-representation0}
There exists a natural exact sequence
\[
0 \to H^2(X,\Omega_{X}^1 ) \to
H^0(X,-K_X)^\vee {\longrightarrow}\mathbb{C} \to 0.
\]
\end{mtheorem}
\begin{proof}
Since $X$ is contained in the smooth locus of $\mathbb{P}$ and $\mathscr{O}_{\mathbb{P}}(X)=\mathscr{O}_{\mathbb{P}}(4)$, we have
the following exact sequence
\[
0 \longrightarrow \mathscr{O}_X(-4)\longrightarrow \Omega_{\mathbb{P}}^1|_X \longrightarrow \Omega_{X}^1 \longrightarrow 0,
\]
and so
\[
H^2(X, \Omega_{\mathbb{P}}^1|_X) \to H^2(X,\Omega_{X}^1 ) \to
H^0(X,\mathscr{O}_X(2))^\vee \to H^3(X, \Omega_{\mathbb{P}}^1|_X)\to 0
\]
The Euler exact sequence for $\mathbb{P}=\mathbb{P}(1^4,2)$ has the form
\[
0 \longrightarrow \Omega^1_{\mathbb{P}} \longrightarrow\mathscr{O}_{\mathbb{P}}(-2)\oplus \mathscr{O}_{\mathbb{P}}(-1)^{\oplus 4} \longrightarrow \mathscr{O}_{\mathbb{P}}
\longrightarrow 0.
\]
Restricting it to $X$ we obtain
$H^2(X, \Omega_{\mathbb{P}}^1|_X)=0$ and $H^3(X, \Omega_{\mathbb{P}}^1|_X)=\mathbb{C}$.
\end{proof}
\begin{mtheorem}{\bf Lemma.}\label{Lemma-representation}
We have the following decomposition of $G$-modules:
\[
\operatorname{T}_{0,J(X)}= V_4\oplus V_4'\oplus V_2,
\]
where $V_4$, $V_4'$ are irreducible faithful $4$-dimensional representations and
$V_2$ is an irreducible $2$-dimensional representation with kernel $\langle \upalpha^{8},\,\upbeta^2\rangle$.
Moreover, ${\operatorname{z}}(G)$ acts on $V_4$ and $V_4'$ via different characters.
\end{mtheorem}
\begin{proof}
Clearly, $\operatorname{T}_{0,J(X)}\simeq H^0(J(X),\Omega_{J(X)})^\vee \simeq H^2(X,\Omega_{X}^1)$
and by Lemma \ref{Lemma-representation0} we have an injection
$\operatorname{T}_{0,J(X)}\hookrightarrow H^0(X,-K_X)^\vee$.
By the adjunction formula $K_X=(K_{\mathbb{P}}+X)|_X$ and so
\[
H^0(X,-K_X)\simeq H^0(\mathbb{P},\mathscr{O}_{\mathbb{P}}(-K_{\mathbb{P}}-X)).
\]
Consider the affine open subset $U:=\{x_1x_2x_3x_4\neq 0\}$.
Then $v=y/x_1^2$ and $z_i=x_i/x_1$, $i=2,3,4$ are affine coordinates in
$U\subset \{x_1\neq 0\}\simeq \mathbb A^4$.
Let $\upomega$ be the $3$-form
\[
\upomega:= \frac {d z_2\wedge d z_3\wedge d z_4 }{\partial \phi/\partial v} =
\frac {d z_2\wedge d z_3\wedge d z_4 }{2 v},
\]
where $\phi=v^2+z_{2}+ z_{2}^3z_{3}+ z_{3}^3z_{4}+ z_{4}^3$ is the equation of $X$ in $U$.
It is easy to check that for any polynomial $\psi(z_2,z_3,z_4)$ of degree $\le 2$
the element $\psi\cdot \upomega ^{-1}$ extends to a section of $H^0(X, -K_X)$. Thus we have
\[
H^0(X, -K_X)\simeq \{ \psi(z_2,z_3,z_4)\cdot \upomega ^{-1} \mid \deg \psi\le 2\}.
\]
It is easy to check that the forms
\begin{equation}
\label{equation-basis}
\upomega ^{-1}, z_2^2 \upomega ^{-1}, z_3^2 \upomega ^{-1}, z_4^2\upomega ^{-1},
z_2 \upomega ^{-1}, z_2z_3 \upomega ^{-1}, z_3z_4 \upomega ^{-1}, z_4 \upomega ^{-1},
z_3 \upomega ^{-1} , z_2z_4\upomega ^{-1}
\end{equation}
are eigenvectors for $\upalpha$ and $\upbeta$ permutes them.
Moreover, the following subspaces
\begin{itemize}
\item[]
$W_4= \langle \upomega ^{-1},\hspace{5pt}z_2^2 \upomega ^{-1},\hspace{5pt}z_3^2 \upomega ^{-1},\hspace{5pt}z_4^2\upomega ^{-1}\rangle$,
\item[]
$W_4'= \langle z_2 \upomega ^{-1},\hspace{5pt}z_2z_3 \upomega ^{-1},\hspace{5pt}z_3z_4 \upomega ^{-1},\hspace{5pt}z_4 \upomega ^{-1}\rangle$,
\item[]
$W_2= \langle z_3 \upomega ^{-1} ,\hspace{5pt}z_2z_4\upomega ^{-1}\rangle$.
\end{itemize}
are $G$-invariant in $H^0(X, -K_X)$. Moreover, in the basis \eqref{equation-basis}
the element $\upalpha$ acts diagonally:
\begin{equation}
\label{equation-W}
\begin{array}{l}
\upalpha|_{W_4} ={\operatorname{diag}} (\zeta_{40}^{11},\, \zeta_{40}^{7},\, \zeta_{40}^{19},\, \zeta_{40}^{23}),
\\[4pt]
\upalpha|_{W_4'} ={\operatorname{diag}} (\zeta_{40}^{9},\, \zeta_{40}^{13},\, \zeta_{40}^{},\, \zeta_{40}^{37}),
\\[4pt]
\upalpha|_{W_2} ={\operatorname{diag}} (\zeta_{8}^{3},\, \zeta_{8}^{7}),
\end{array}
\end{equation}
and $\upbeta$ acts on each of these subspaces permuting the eigenspaces of $\upalpha$ cyclically.
Thus $\upalpha^{10}$ acts on $W_4$ (resp., $W_4'$)
via scalar multiplication by $\zeta_4^{3}$ (resp., $\zeta_4$).
Put $V_4:=W_4^\vee$, $V_4':=W_4'^\vee$, $V_2:=W_2^\vee$.
\end{proof}
\section{Proof of Theorem \ref{theorem-main}}
\begin{say}
Assume to the contrary to Theorem \ref{theorem-main} that $J(X)$ is a direct sum of Jacobians of curves, i.e.
in the unique decomposition \eqref{equation-decomposition} we have
$A_i\simeq J(C_i)$, where $C_i$ is a curve of genus $\ge 1$
and $J(C_i)$ is its Jacobian regarded as a principally polarized abelian variety.
Let $G_i$ be the stabilizer of $A_i$. There is a natural homomorphism
$\varsigma_i: G_i \to\operatorname{Aut} (C_i)$.
By the
Torelli theorem
$\varsigma_i$ is injective and we have
\begin{equation}
\label{equation-Aut-C}
\operatorname{Aut} (J(C_i))\simeq
\begin{cases}
\operatorname{Aut} (C_i)&\text{if $C_i$ is hyperelliptic,}
\\
\operatorname{Aut} (C_i)\times \{\pm 1\}&\text{otherwise.}
\end{cases}
\end{equation}
Let us analyze the action of $G$ on the set $\{A_1,\dots, A_n\}$.
Up to renumbering we may assume that subvarieties $A_1,\dots, A_m$ form one $G$-orbit
(however, the choice of this orbit is not unique in general).
Clearly, $m\in \{1,2,4,5,8,10\}$.
Denote the stabilizer of $A_i$ by $G_i$.
Consider the possibilities for $m$ case by case.
\end{say}
\begin{say}
\label{lemma-invariant-A}
{\bf Case: $m=1$,}
that is, $A_1\subset J(X)$ is a $G$-invariant subvariety.
Since ${\operatorname{z}}(G)=\langle \upalpha^{10}\rangle$, the only normal subgroup of order 2 in $G$
is $\langle \upalpha^{20}\rangle$. Hence $G$ cannot be decomposed
as a direct product of groups of orders $2$ and $80$ (otherwise the order of $\upalpha$
would be $20$).
If the action of $G$ on $A_1=J(C_1)$ is faithful, then by \eqref{equation-Aut-C} so is the corresponding
action on $C_1$. So, the curve $C_1$ of genus $\le 10$ admits
faithful action of the group $G$ of order $2^{5}\cdot 5$.
This contradicts Lemma \ref{lemma-action-80}\ref{lemma-action-80-k=5}.
Therefore the induced representation on $\operatorname{T}_{0,A_1}$ is not faithful.
By Lemma \ref{Lemma-representation}\ $\operatorname{T}_{0,J(C_1)}= V_2$.
In this case $g(C_1)=2$ and the action of $G$ on $J(C_1)$ induces a faithful action of the group
$\bar G:= G/\langle \upalpha^{8},\,\upbeta^2\rangle$ of order $16$.
Since $C_1$ is hyperelliptic, $\bar G$ is contained in $\operatorname{Aut}(C_1)$.
If $\bar G$ contains the hyperelliptic involution $\tau$, then
$\tau$ generates a
normal subgroup of order 2. In this case $\langle \tau\rangle=[\bar G,\bar G]$
and $\bar G/\langle \tau\rangle$ is an abelian non-cyclic group of order $8$.
But such a group cannot act faithfully on $C_1/\langle \tau\rangle\simeq \mathbb{P}^1$.
Thus $\bar G$ does not contain the hyperelliptic involution.
In this case the image of the induced action of $\bar G$ on canonical sections
$H^0(C_1,\mathscr{O}_{C_1}(K_{C_1}))$ does not contain scalar matrices.
Hence this representation is reducible and so it is trivial on $[\bar G,\bar G]$.
On the other hand, the action of $\operatorname{Aut}(C_1)$ on $H^0(C_1,\mathscr{O}_{C_1}(K_{C_1}))$
must be faithful
a contradiction.
\end{say}
From now on we may assume that the decomposition \eqref{equation-decomposition}
contains no $G$-invariant summands.
\begin{say}{\bf Case: $m=5$.}
The subspace $\operatorname{T}_{0,A_1}\oplus\cdots\oplus \operatorname{T}_{0,A_5}\subset \operatorname{T}_{0,J(X)}$ is a $G$-invariant
of dimension $5$ or $10$.
On the other hand, $\operatorname{T}_{0,J(X)}$ contains no invariant subspaces of
dimension $5$ by Lemma \ref{Lemma-representation}.
Hence, $\operatorname{T}_{0,A_1}\oplus\cdots\oplus \operatorname{T}_{0,A_5}= \operatorname{T}_{0,J(X)}$, $\dim A_i=2$, and
$J(X)= \oplus_{i=1}^{5} A_i$.
The stabilizer $G_i\subset G$ is a Sylow $2$-subgroup that faithfully acts on $C_i$
(because $C_i$ is hyperelliptic, see \eqref{equation-Aut-C}).
Further, $G_i$ permutes the Weierstrass points $P_1,\dots,P_6\in C_i$.
Hence a subgroup $G_i'\subset G_i$ of index 2 fixes one of them.
In this situation, $G_i'$ must be cyclic.
On the other hand, it is easy to see that $G$ does not contain any elements of order $16$,
a contradiction.
\end{say}
\begin{say}{\bf Case: $m=10$.}
Then $A_1,\dots,A_{10}$ are elliptic curves and
$G_i\subset G$ is a subgroup of index $10$.
By Lemma \ref{subgroup-index10} each $G_i$ contains ${\operatorname{z}}(G)$.
Clearly, ${\operatorname{z}}(G)$ acts on $\operatorname{T}_{0,A_i}$ via the same character.
Since the subspaces $\operatorname{T}_{0,A_i}$ generate $\operatorname{T}_{0,J(X)}$, the group ${\operatorname{z}}(G)$
acts on $\operatorname{T}_{0,J(X)}$ via scalar multiplication.
This contradicts Lemma \ref{Lemma-representation}.
\end{say}
\begin{say}
\label{orbit=8}
{\bf Case: $m=8$.}
Then $A_1,\dots,A_8$ are elliptic curves and the stabilizer $G_1\subset G$
is of order $20$. In particular, the Sylow $5$-subgroup ${\operatorname{Syl}}_5(G)$ is contained in $G_1$.
Since ${\operatorname{Syl}}_5(G)$ is normal in $G$, we have ${\operatorname{Syl}}_5(G)\subset G_i$ for $i=1,\dots,8$.
Since the automorphism group of an elliptic curve contains no order $5$ elements, ${\operatorname{Syl}}_5(G)$ acts trivially on $A_i$.
Therefore, ${\operatorname{Syl}}_5(G)$ acts trivially on the $8$-dimensional
$G$-invariant subspace $\operatorname{T}_{0,A_1}\oplus \cdots\oplus \operatorname{T}_{0,A_8}$.
This contradicts Lemma \ref{Lemma-representation}.
\end{say}
\begin{say}\label{orbit=4}{\bf Case: $m=4$.}
The intersection $G_1\cap \langle\upalpha\rangle$ is a subgroup of index $\le 4$
in both $G_1$ and $\langle\upalpha\rangle$. Hence, $G_1\ni \upalpha^{4}$ and so $G_1\supset [G,G]$.
In particular, $G_1$ is normal and $G_1=\cdots=G_4$.
If $\dim A_1=1$, then the element $\upalpha^8$ of order $5$ must
act trivially on elliptic curves $A_i\in 0$, $i=1,\dots,4$.
Therefore, $\upalpha^8$ acts trivially on the $4$-dimensional space $\operatorname{T}_{0,A_1}\oplus\cdots\oplus \operatorname{T}_{0,A_4}$.
This contradicts Lemma \ref{Lemma-representation}.
Thus $\dim A_1=2$. Then $\operatorname{T}_{0,A_1}\oplus\cdots\oplus \operatorname{T}_{0,A_4}=V_4\oplus V_4'$.
An eigenvalue of $\upalpha$ on $\operatorname{T}_{0,A_1}\oplus\cdots\oplus \operatorname{T}_{0,A_4}$
must be a primitive $40$-th root of unity (see \eqref{equation-W}).
Hence the group $G_1\cap \langle\upalpha\rangle$ acts faithfully on $\operatorname{T}_{0,A_1}$
and $C_1$ (see \eqref{equation-Aut-C}).
By Lemma \ref{lemma-action-80}\ref{lemma-action-80-k=2} $G_1\cap \langle\upalpha\rangle$ is of order $10$,
i.e. $G_1\cap \langle\upalpha\rangle=\langle \upalpha^4\rangle$ and
the kernel $N:=\ker (G_1\to \operatorname{Aut}(C_1))$ is of order $4$. Thus $G_1=\langle \upalpha^4\rangle\times N$.
In particular, $G_1$ is abelian.
But then the centralizer $\operatorname{C}(\upalpha^8)$ of $\upalpha^8$ contains $N$ and $\langle\upalpha\rangle$.
Therefore, $\operatorname{C}(\upalpha^8)=G$ and $\upalpha^8\in {\operatorname{z}}(G)$.
This contradicts Lemma \ref{subgroup-index10}\ref{subgroup-index10-1}.
\end{say}
Thus we have excluded the cases $m=1,4,5,8,10$.
The only remaining possibility is that all the orbits of $G$ on $\{ A_i\}$
are of cardinality $2$.
\begin{say}{\bf Case: $m=2$.}
Then $\dim A_1\le 5$ and $G_1$ is a group of order $80$.
By replacing the orbit $\{A_1,A_2\}$ with another one we may assume that
$\operatorname{T}_{0,A_1}\oplus \operatorname{T}_{0,A_2}\not \subset V_2$
and so $\operatorname{T}_{0,A_1}\oplus \operatorname{T}_{0,A_2}$ coincides with either $V_4$, $V_4'$, or $V_4\oplus V_4'$.
In particular, $g(C_1)\ge 2$.
Clearly,
$G_1\cap \langle\upalpha\rangle$ is of order $40$ or $20$. Hence,
$\upalpha^{2}\in G_1$ and so the group $G_1$ cannot be decomposed as a direct product
$G_1=\langle \upalpha^{20}\rangle\times H$. By the Torelli theorem $G_1$ faithfully acts on $C_1$.
This contradicts Lemma \ref{lemma-action-80}\ref{lemma-action-80-k=4}.
\end{say}
Proof of Theorem \ref{theorem-main} is now complete.
\begin{proof}[Proof of Corollary \textup{\ref{corollary-main}}]
The Jacobians and their sums form a
closed subvariety of the moduli space of principally polarized abelian varieties.
By Theorem \ref{theorem-main}, in our case, this subvariety does not contain
the subvariety formed by Jacobians of quartic double solids.
Therefore a general quartic double solid is not rational.
\end{proof}
\subsection*{Acknowledgements.}
The author would like to thank C. Shramov and the referee for useful comments
and corrections.
\input jacobian.bbl
\end{document}
|
\section{From DPD to continuum}
\section{Introduction}
The study of active suspensions is by now a well established field of research in mathematics, physics, and the engineering sciences. Because the literature on active suspensions is vast, an exhaustive review of its contents is infeasible. We therefore include only a few representative citations of prior work \cite{bricard, ibele, koch-sub, kudrolli08, kudrolli10, narayan06, narayan07, schaller}.
In particular, concentrated suspensions of active particles have recently attracted much attention \cite{marchetti-review}. For such media, the derivation of continuum theories from realistic fine scale models constitutes a challenging and still largely open problem. Most available results rely on kinetic theory \cite{baskaran-marchetti2009, saintillan-shelley2013}. Since kinetic closures are typically based on assuming that the suspended particles are dilutely concentrated and interact only weakly \cite{marchetti-review}, it is of interest to develop alternative coarsening procedures which are free of these limitations. Promising alternatives to kinetic theory include, for example, coarsening procedures based on direct calculation of ensemble averages~\cite{Chuang2007} and coarsening procedures based on space-time averages. In this article, we focus on the latter.
The primary objective of this paper is to develop a coarsening method for active suspensions that does not require a kinetic formulation and that can deal with highly concentrated and strongly interacting particle systems. To focus on features induced by self-propulsion, we work with spherical particles (modeled as point particles interacting with appropriate forces) instead of rod-like particles. This makes sense for the following reasons. First, simulations show that suspensions of active point particles manifest a rich variety of semi-ordered and ordered states \cite{HPKF}, including vortical, meso-turbulent, and polar ones. Second, it is important to understand, by means of a consistent bottom-up derivation, which features of the continuum equations are due mainly to self-propulsion and high concentration in contrast to attributes that stem from the presence of orientational degrees of freedom.
Third, the models using spherical particles are simpler than those using elongated-particle models, which typically require additional balance equations for director variables. On this basis, it seems reasonable to develop microstructure-consistent equations for a simpler case (which is done here), and to leave more complicated cases involving, say, orientational degrees of freedom for future research. Finally, the equations developed in this paper can be directly applicable to experimental situations involving spherical swimming bacteria such as a strain of
\emph{Serratia marcescens} studied in \cite{rabani}. More generally, as Dusenbery \cite[page~25]{dusenbery} reports, about
10\% of motile bacterial species are spherical.
Starting from a fine-scale dissipative particle dynamics (DPD) \cite{dpd-first, espanol-warren} model of self-motile point particles, we derive meso-scale continuum equations by applying an averaging technique known as the Irving--Kirkwood--Noll procedure \cite{Kirkwood,Noll}. The most relevant version of this approach is that introduced by Hardy \cite{Hardy} and, later and independently, by Murdoch and Bedeaux \cite{mb, mb96, mb97, murdoch-book}. In this procedure, continuum equations are derived systematically and directly from particle equations.
A collision-based kinetic formulation is thus avoided. In our context, the averages depend on a mesoscopic length scale $\eta$ that is assumed to be much larger than the range $R$ of the DPD forces. The ratio $\eta/R$ embodies a separation of scales. In contrast to results that rely on ensemble averaging, our equations describe single realizations, both in terms of initial conditions and realizations
of fluctuation forces. This has significant practical advantages because stochastic averaging, which is commonly associated with large errors and high computational costs, is no longer required to calculate effective parameters. Explicit constitutive equations are obtained from spatial statistics of fluctuations. In particular, our method results in the commonly encountered cubically nonlinear (internal) body force density associated with self-propulsion. Surprisingly, this arises even though the self-propulsion force in our DPD model does not incorporate a mechanism for aligning neighboring particles or other velocity selection mechanisms.
Our effective continuum equations resemble those arising in the well-known phenomenological model of Toner and Tu \cite{toner-tu95}. Toner--Tu type equations have previously been derived by applying the kinetic theory to systems of self-propelled particles that move with constant speed
in directions that change in response to
a velocity-aligning force \cite{bertin06,ihle11}. It is noteworthy that the same set of continuum equations results from a completely different microscopic model in which short-range interactions dominate but no velocity-aligning force appears.
Our continuum equations combine features of three classical models: the Navier--Stokes equations for a compressible fluid, the Ginzburg--Landau equations, and the Langevin equations. The structure of the momentum balance and a linear constitutive equation for viscous stress are reminiscent of the Navier--Stokes equations for a compressible fluid,
with the hydrostatic contribution to the pressure being determined as a nonlinear function of the effective mass density. The viscosity tensor is also given by a constitutive relation. The cubic nonlinearity of the effective self-propulsion forces is typical of the Ginzburg--Landau equations. A feature of our approach which brings to mind Langevin equations is the presence of an additive stochastic force.
As a consequence of the fluctuation-dissipation relation \cite{espanol-warren}, the physical temperature is incorporated into the constitutive relations via dependence of the DPD forces on the temperature. In addition, the constitutive functions depend on two scalar temperature-like parameters. While one of these characterizes fluctuations of fine scale (DPD) velocities and can be associated with an ``upscaling temperature," the other describes the extent to which relative particle positions fluctuate. For sufficiently dense and slightly compressible systems, position fluctuations tend to remain close to their initial values and, thus, they may be determined without using any fine scale computing. Averaging of random DPD forces yields a coarse-scale fluctuation force which is reminiscent of the stochastic driving term in the Langevin equation. The force is uncorrelated in time and realizes a Gaussian random field at each instant of time. However, in contrast to the Langevin equation, the strength and the variance here are given by constitutive functions of the mass density, the temperature, and the fluctuation strength parameters of the DPD forces. Therefore, the statistics of the average fluctuation force vary with time and position. Importantly, however, the variance decreases with increasing scale separation, so that the probability distribution of the coarse-scale fluctuation force concentrates more and more tightly near the mean. Since the mean is zero, this force vanishes in the limit of infinite scale separation, and the model becomes deterministic. Suggesting that imposing a fine-scale fluctuation-dissipation relation does not in general yield an analogous coarse-scale relation, this construction may be of independent interest.
The remainder of the paper is organized as follows. The DPD model is described in Section \ref{sect:dpd}. Exact averaged equations are provided in Section \ref{sect:exact-av}. Our closure method, which is contingent on the spatial statistics arising for the discretized formulation of the balance equations, is described in Section~\ref{sect:statistics}. The resulting closed-form continuum equations are summarized in Section~\ref{sect:cont-summary}. Closed-form approximations for the average self-propulsion force and the convective stress that models the momentum transfer due to velocity fluctuations are derived in Section~\ref{sect:sp-cs-closure}. The average fluctuation force equation is derived in Section~\ref{sect:rand1} and analyzed in Section~\ref{sect:rand2}. A linear stability analysis is conducted is Section \ref{sect:lin-stab}. Closed-form approximations for the stresses induced by conservative and dissipative forces are derived in Sections~\ref{sect:cons-closure} and Section~\ref{sect:diss-closure}, respectively. A synopsis of our most salient results is provided in Section \ref{sect:conclusions}. Closed-form approximations for the stresses induced by conservative and dissipative forces are derived in the Appendix.
\section{DPD equations of motion}
\label{sect:dpd}
Dissipative particle dynamics have been previously used for modeling of passive colloidal suspensions at low, moderate, and high concentrations \cite{dense1-boek, dense2-bolint, dense3-laurati}. For high concentrations, DPD equations with short-range pairwise forces seem to work reasonably well because long-range hydrodynamics interactions are screened by nearly touching neighboring particles. (See, for example, \cite{panchenko-jfm} for a detailed mathematical treatment of the screening effect.)
It therefore seems natural to develop a suitable DPD model for active colloidal suspensions. To this end, we augment the conventional DPD forces (to be detailed shortly)
with a self-propulsion force
\begin{equation}
\label{dpd-sp-force}
{\boldsymbol f}_i^{\textit{SP}}=A^{\textit{SP}}h(|{\boldsymbol v}_i|){\boldsymbol v}_i,
\end{equation}
where $A^{\textit{SP}}$ is a constant strength parameter with dimensions of force and $h$ is a nonnegative function with dimensions of inverse velocity. A related observation is that the product $A^{SP}h(|{\boldsymbol v}|)$ carries dimensions of mass per unit time or, equivalently, of viscosity per unit length. This product can therefore be viewed as a velocity-dependent viscosity coefficient. Other properties of $h$ will be discussed in the final paragraph of this section.
To simplify the presentation, we assume that all DPD particles have equal mass $m$. The positions ${\boldsymbol q}_i$ of a particles evolve according to the ordinary-differential equations
\begin{equation}
m\ddot{\boldsymbol q}_i
={\boldsymbol f}_i^{\textit{SP}}+\sum_{j\ne i}{\boldsymbol f}^C_{ij}+\sum_{j\ne i}{\boldsymbol f}^D_{ij}+\sum_{j\ne i}{\boldsymbol f}^R_{ij},~~~~~~~i, j=1, 2, \ldots, N,
\label{dpd-odes}
\end{equation}
where ${\boldsymbol f}^C_{ij}$, ${\boldsymbol f}^D_{ij}$, and ${\boldsymbol f}^R_{ij}$ are the conservative, dissipative, and fluctuation forces familiar from conventional DPD \cite{dpd-first, espanol-warren}. Specifically,
\begin{align}
{\boldsymbol f}^C_{ij}&=A w^C(r_{ij}) {\boldsymbol e}_{ij},
\label{dpd-c-force}
\\[4pt]
{\boldsymbol f}^D_{ij}&=-\gamma w^D(r_{ij})({\boldsymbol v}_{ij}\cdot{\boldsymbol e}_{ij}){\boldsymbol e}_{ij},
\label{diss-force}
\\[4pt]
{\boldsymbol f}^R_{ij}&=\alpha \xi_{ij} w^R (r_{ij}){\boldsymbol e}_{ij},
\label{rand-force}
\end{align}
where $A$ is a stiffness coefficient, $\gamma$ is a drag coefficient, $\alpha$ is a strength parameter, $w^C$, $w^D$, and $w^R$ are window functions, $\xi_{ij}$ is a Gaussian random variable with zero mean and unit variance satisfying $\xi_{ij}=\xi_{ji}$, and $r_{ij}$ and ${\boldsymbol e}_{ij}$ are given by
\begin{equation}
r_{ij}=|{\boldsymbol q}_i-{\boldsymbol q}_j|
\qquad\text{and}\qquad
{\boldsymbol e}_{ij}=\frac{{\boldsymbol q}_i-{\boldsymbol q}_j}{|{\boldsymbol q}_i-{\boldsymbol q}_j|}.
\end{equation}
The parameters are related by the fluctuation-dissipation relations \cite{espanol-warren}
\begin{equation}
w^D=\left(w^R \right)^2,\qquad\alpha^2=2\gamma k_B T,
\label{FDrelns}
\end{equation}
where $k_B$ is the Boltzmann constant and $T$ is the absolute temperature. These relations are the consequence of imposing a balance of fluctuation forces and dissipative forces so that the associated Fokker--Planck equation has a steady state solution given by the Gibbs canonical probability density $Z^{-1}e^{-(k_B T)^{-1} H}$, where $H$ is the sum of the kinetic energy and the potential energy of the conservative DPD forces, and $Z$ is a normalizing constant known as the partition function.
From its definition \eqref{dpd-sp-force}, the self-propulsion force ${\boldsymbol f}_i^{\textit{SP}}$ aligns with the direction ${\boldsymbol v}/|{\boldsymbol v}|$ determined by the velocity vector ${\boldsymbol v}_i$.
One possibility is to choose $h$ such that
\begin{equation}
\label{choice-h}
h(\xi)=\frac{1}{\sqrt{\xi^2+\delta^2}}.
\end{equation}
In this case, the magnitude of the self-propulsion force ${\boldsymbol f}_i^{\textit{SP}}$ defined by \eqref{dpd-sp-force} is nearly equal to the constant strength parameter $A^{\textit{SP}}$ for $\delta\ll|{\boldsymbol v}|$, in which case the orientation ${\boldsymbol v}/|{\boldsymbol v}|$ is closely approximated by ${\boldsymbol v}/\sqrt{|{\boldsymbol v}|^2+\delta^2}$. The approximation is illustrated in Figure \ref{Fig:sp-app}.
\begin{figure}[!t]
\label{Fig:sp-app}
\centering
\includegraphics[width=0.416\textwidth]{molification3.eps}
\put(-227,60){\rotatebox{90}{$h(\xi)$}}
\put(-110,-13){$\xi$}
\put(-40,120){\tiny$1/\xi$}
\put(-40,111){\tiny$\delta=0.025$}
\put(-40,102){\tiny$\delta=0.050$}
\put(-40,93){\tiny$\delta=0.100$}
\put(-40,84){\tiny$\delta=0.200$}
\caption{Approximation of $1/\xi$ by $h(\xi)=1/\sqrt{\xi^2+\delta^2}$ for various values of $\delta$.}
\end{figure}
The particular value of $\delta$ should be selected consistent with the requirement that, at any given instant of time, most particles move with the velocities that are considerably larger than $\delta$. Thus,
$\delta$ should be small in comparison to $\sqrt{K}$, where $K$ is the spatio-temporal average kinetic energy per particle and per unit mass. Formally, choosing $\delta>0$ prevents division by zero when ${\boldsymbol v}_i$ vanishes. The choice \eqref{choice-h} can be viewed as a constitutive relation of stick-slip type intended to mimic the observed behavior of bacteria, which swim mostly at constant velocity but occasionally pause.
\section{Exact continuum equations. Stresses corresponding to pair forces}
\label{sect:exact-av}
To derive meso-scale continuum equations from the micro-scale model, we apply spatial averaging to a single realization of DPD equations. (Here, the term ``single realization'' refers to one realization of stochastic forces present in the DPD model.)
The initial conditions for the DPD equations are assumed to be both deterministic and precisely known. In that sense, our strategy is therefore predicated on conditions that differ significantly from those
commonly used to justify
ensemble averaging in statistical mechanics. The expected result of spatial averaging is a system of stochastic continuum equations. The random nature of these equations should be inherited from the underlying stochastic ordinary differential equations.
Spatial averages are defined using a window function $\psi_\eta$. This function depends on the averaging length scale $\eta$ and is normalized to have a unit integral for each $\eta$. Most often, $\psi_\eta$ is non-negative and either compactly supported or rapidly decreasing.
The average mass density $\overline\rho$ and linear momentum $\overline{{\boldsymbol p}}$ are defined according to \cite{Hardy}
\begin{equation}
\label{gen-density}
\overline{\rho}(t, {\boldsymbol x})=\sum_{j=1}^N m_j \psi_\eta ({\boldsymbol x}-{\boldsymbol q}_j(t))
\end{equation}
and
\begin{equation}
\label{gen-momentum}
\overline{{\boldsymbol p}}(t, {\boldsymbol x})=\sum_{j=1}^N m_j {\boldsymbol v}_j \psi_\eta ({\boldsymbol x}-{\boldsymbol q}_j(t)).
\end{equation}
Mimicking conventional derivations \cite{Hardy,mb}, we may differentiate these quantities with respect to time and then use the ordinary-differential equations of the DPD model to eliminate time derivatives of the particle velocities. This yields exact balance equations for mass and linear momentum of the form
\begin{equation}
\label{exact-density}
\partial_t \overline{\rho}+{\rm div}\overline{{\boldsymbol p}}=0
\end{equation}
and
\begin{equation}
\label{exact-momentum}
\partial_t \overline{{\boldsymbol p}}+{\rm div}(\overline{{\boldsymbol p}}\otimes \overline{{\boldsymbol v}})={\rm div}
\left[
-\sum_{j=1}^N m_j ({\boldsymbol v}_j-\overline{{\boldsymbol v}})\otimes ({\boldsymbol v}_j-\overline{{\boldsymbol v}})\psi_\eta({\boldsymbol x}-{\boldsymbol q}_j)\right]
+{\rm div}{\boldsymbol T}+{{\boldsymbol g}}^R+ {{\boldsymbol g}}^{\textit{SP}},
\end{equation}
where
\begin{equation}
\label{average-velocity}
\overline{{\boldsymbol v}}=\frac{\overline{\boldsymbol p}}{\overline\rho}
\end{equation}
is the average velocity, ${\boldsymbol T}$ is the interaction stress,
\begin{equation}
\label{R0}
{{\boldsymbol g}}^R(t, {\boldsymbol x})=\sum_{j, k=1}^N {\boldsymbol f}_{jk}^R \psi_\eta({\boldsymbol x}-{\boldsymbol q}_j(t))
\end{equation}
is the average fluctuating force, and
\begin{equation}
\label{av-sp}
{{\boldsymbol g}}^{\textit{SP}}(t, {\boldsymbol x})=\sum_{j=1}^N {\boldsymbol f}_j^{\textit{SP}}\psi_\eta({\boldsymbol x}-{\boldsymbol q}_j(t))
\end{equation}
is the average self-propulsion force.
The interaction stress ${\boldsymbol T}$ consists of a sum
\begin{equation}
\label{exact-stress}
{\boldsymbol T}={\boldsymbol T}^C+{\boldsymbol T}^D,
\end{equation}
of conservative and dissipative contributions, where ${\boldsymbol T}^C$ and ${\boldsymbol T}^D$ are determined respectively by the DPD pair forces ${\boldsymbol f}^C_{ij}$ and ${\boldsymbol f}^D_{ij}$ defined in \eqref{dpd-c-force} and \eqref{diss-force} through
\begin{equation}
\label{cons-stress}
{\boldsymbol T}^C=\frac 12 \sum_{i=1}^N \sum_{j=1}^N {\boldsymbol f}^C_{ij} \otimes ({\boldsymbol q}_j-{\boldsymbol q}_i) \Psi_\eta({\boldsymbol x}, {\boldsymbol q}_i, {\boldsymbol q}_j)
\end{equation}
and
\begin{equation}
\label{diss-stress}
{\boldsymbol T}^D=\frac 12 \sum_{i=1}^N \sum_{j=1}^N {\boldsymbol f}^D_{ij} \otimes ({\boldsymbol q}_j-{\boldsymbol q}_i) \Psi_\eta({\boldsymbol x}, {\boldsymbol q}_i, {\boldsymbol q}_j),
\end{equation}
with $\Psi_\eta$ being defined by
\begin{equation}
\label{functPsi}
\Psi_\eta({\boldsymbol x}, {\boldsymbol q}_i, {\boldsymbol q}_j)=\int_0^1 \psi_\eta (s({\boldsymbol x}-{\boldsymbol q}_i)+(1-s)({\boldsymbol x}-{\boldsymbol q}_j))\,\text{d}s.
\end{equation}
On using \eqref{dpd-c-force} and \eqref{diss-force} in \eqref{cons-stress} and \eqref{diss-stress}, ${\boldsymbol T}^C$ and ${\boldsymbol T}^D$ are seen to be symmetric. Thus, by \eqref{exact-stress}, the interaction stress obeys
\begin{equation}
{\boldsymbol T}=\transpose{{\boldsymbol T}},
\end{equation}
which is not surprising for a system of point-like particles (regardless of activity).
Although the average fluctuation force ${\boldsymbol g}^R$ can be also written as the divergence of the random interaction stress
\begin{equation}
\label{rand-stress}
{\boldsymbol T}^R=\frac 12 \sum_{i=1}^N \sum_{j=1}^N {\boldsymbol f}^R_{ij} \otimes ({\boldsymbol q}_j-{\boldsymbol q}_i) \Psi_\eta({\boldsymbol x}, {\boldsymbol q}_i, {\boldsymbol q}_j),
\end{equation}
working directly with ${\boldsymbol g}^R$ is more convenient because its statistics are somewhat more easily described and because doing so results in a model that resembles the Langevin equation.
It does not seem possible to apply the simple techniques of Noll \cite{Noll} and Hardy \cite{Hardy} to express the self-propulsion force ${\boldsymbol g}^{\textit{SP}}$ as the divergence of a stress-like object. We have therefore opted to treat this force as a body force in the continuum theory.
The exact equations \eqref{exact-density}--\eqref{exact-momentum} do not constitute a continuum model in the true sense. Indeed, the stress and other terms in the right hand side of \eqref{exact-momentum} cannot be determined without knowledge of the trajectories of all DPD particles. Since a true continuum model should be self-contained (or closed), exact equations should be supplemented by a closure approximation that allows the right-hand side of \eqref{exact-momentum} to be expressed as a function of available averages, namely the mass density, linear momentum, temperature, and so on. Finding a suitable closure is both the most difficult and the most important step in deriving meso-scale equations from a microscopic model.
\section{Average velocity, average deformation and fluctuation-based closure}
\label{sect:statistics}
\subsection{Averaging in the discrete setting}
In conventional continuum theories, averages are defined at each point of space and at each instant of time. However, in most situations of practical interest, the objective is to compute solutions using a discretized version of the governing equations. Discretization reduces the available information because the spatial resolution of any numerical method is inherently finite. In principle, this resolution can be refined indefinitely. In practice, refinement is, however, limited by the available computing power. It is therefore natural to consider the situation where the smallest available resolution is in place and cannot be further reduced, in which case averages are available only at points $\bx_\beta$, $\beta=1,2,\dots,B$, of the computational grid. In principle, time should be also discretized. We nevertheless focus on spatial discretization and assume that the grid values are known at each instant of time.
Discretizing a continuum description in accompanied by a loss of information which turns out to be quite severe. In the continuum setting, the fine-scale velocity can, in principle, be uniquely reconstructed from the knowledge of the average density and momentum \cite{PBG, PT, BP}. A unique reconstruction is possible for each realization of particle dynamics. Once velocities have been determined, integrating with respect to time leads to a unique recovery of positions. Uniqueness implies that fine-scale information is completely transferred to the meso-scale. In the discretized setting, uniqueness is lost and many different particle states can generate the same grid values of, say, the average mass density and momentum. Missing information should be quantified using suitable statistics for fluctuations. The statistics developed here differ from the atomistic ensemble statistics, which originate from indeterminate initial conditions. Since the typical spatial scales in DPD are much larger than the distances between neighboring fluid molecules, it is reasonable to suppose that the DPD initial conditions are given precisely. Averaging of a single realization and measuring the averages on a finite set of points is nevertheless still accompanied by an information deficit.
\subsection{Average mass density and average velocity}
Consider a simple discretized averaging model in dimension $d$, with $d=2$ or $d=3$, assuming that the continuum length scale $\eta$ corresponds to the finest affordable resolution. Consider a cubic computational domain $\Omega$ with $d$-dimensional volume $\mathcal{V}$. Divide $\Omega$ into $B$ non-overlapping cubic averaging cells ${\mathcal C}_\beta$, $\beta=1,2,\dots,B$, each of side length $\eta$ and volume ${\mathcal V}_\beta=\mathcal{V}/B=\eta^d$. Let ${\mathcal C}_\beta$ be centered at $\bx_\beta$. Further, let $I_\beta$ be the time-dependent index set of particles located in ${\mathcal C}_\beta$ and let $n_\beta(t)$ be the number of particles in $I_\beta$ at time $t$.
\begin{figure}[!t]
\label{Fig:ac-cells}
\centering
\includegraphics[width=0.9\textwidth]{fig2_new2}
\caption{Averaging cells within the flow domain $ \Omega$.}
\end{figure}
Define the average mass density $\overline{\rho}_\beta$ of ${\mathcal C}_\beta$ by
\begin{equation}
\label{av-density}
\overline{\rho}_\beta(t)=\overline{\rho}(t, \bx_\beta)
=\frac{1}{{\mathcal V}_\beta}\sum_{j\inI_\beta}m=\frac{m\mskip1mu n_\beta(t)}{{\mathcal V}_\beta}
\end{equation}
where $m$ is the mass of particle $j$ in $I_\beta$. Similarly, define the average velocity $\overline{{\boldsymbol v}}_\beta$ of ${\mathcal C}_\beta$ by
\begin{equation}
\label{v-bar}
\overline{{\boldsymbol v}}_\beta(t)=\overline{{\boldsymbol v}}(t,\bx_\beta)=\frac{1}{n_\beta}\sum_{j\in I_\beta} {\boldsymbol v}_j(t).
\end{equation}
Importantly, for a particle $j$ in ${\mathcal C}_\beta$, the velocity fluctuations
\begin{equation}
\label{vel-fluct}
{\boldsymbol v}_j^\prime={\boldsymbol v}_j-\overline{{\boldsymbol v}}_\beta
\end{equation}
satisfy
\begin{equation}
\label{vel-zero}
\sum_{j\in J_\beta} {\boldsymbol v}_j^\prime=0.
\end{equation}
The averages \eqref{av-density} and \eqref{v-bar} can be obtained, respectively, from \eqref{gen-density} and \eqref{gen-momentum} on choosing the window function $\psi_\eta$ to be of the particular form
\begin{equation}
\label{psi-box}
\psi_\eta({\boldsymbol x})=
\left\{
\begin{array}{cc}
\displaystyle
\frac{1}{\mathcal V}_\beta & {\rm if}\;{\boldsymbol x}\in {\mathcal C}_\beta,
\cr\noalign{\vskip4pt}
0 & {\rm otherwise}.
\end{array}
\right.
\end{equation}
Granted that all particles have equal mass and using Hardy averages \cite{Hardy}, we then have
\begin{eqnarray*}
\overline{{\boldsymbol v}}_\beta(t)
= \frac{\sum_{j=1}^N m{\boldsymbol v}_j \psi_\eta(\bx_\beta-{\boldsymbol q}_j)}{\sum_{j=1}^N m \psi_\eta(\bx_\beta-{\boldsymbol q}_j)}=
\frac{{\mathcal V}_\beta^{-1}\sum_{j\in I_\beta} {\boldsymbol v}_j }{{\mathcal V}_\beta^{-1}\sum_{j\in I_\beta} 1}=
\frac{1}{n_\beta}\sum_{j\in I_\beta} {\boldsymbol v}_j.
\end{eqnarray*}
More generally, we may write
\begin{equation}
\label{average-velocity-gen}
\overline{{\boldsymbol v}}(t, {\boldsymbol x})
=\frac{\sum_{j=1}^N m {\boldsymbol v}_j \psi_\eta({\boldsymbol x}-{\boldsymbol q}_j(t))}{\sum_{j=1}^N m \psi_\eta(\bx_\beta-{\boldsymbol q}_j(t))}=
\frac{1}{n(t, {\boldsymbol x})}\sum_{j\in I(t, {\boldsymbol x})} {\boldsymbol v}_j,
\end{equation}
where $n(t,{\boldsymbol x})$ is the number of particles in the cube ${\mathcal C}_{\boldsymbol x}$ with the volume ${\mathcal V}_\beta$ centered at ${\boldsymbol x}$ at time $t$, and $I(t,{\boldsymbol x})$ is the associated index set of particles located within this box.
\subsection{Average deformation}
By analogy to the classical kinematical connection between the referential and spatial descriptions of velocity, we define the average deformation ${\boldsymbol \chi}$ by
$$
\dot{\boldsymbol \chi}(t, {\boldsymbol X})=\overline{{\boldsymbol v}}(t, {\boldsymbol \chi}(t, {\boldsymbol X})),\;\;\;\;{\boldsymbol \chi}(0,{\boldsymbol X})={\boldsymbol X},
$$
where ${\boldsymbol X}$ denotes a generic point in a fixed reference configuration. Although this quantity is not known a priori, it is useful to represent the relative particle positions
$$
{\boldsymbol q}_{ij}(t)={\boldsymbol q}_i(t)-{\boldsymbol q}_j(t)
$$
in the form
$$
{\boldsymbol q}_{ij}=\overline{{\boldsymbol q}}_{ij}+{\boldsymbol q}_{ij}^\prime,
$$
where $\overline{{\boldsymbol q}}_{ij}=\overline{{\boldsymbol q}}_i-\overline{{\boldsymbol q}}_j$ is an average relative position compatible with ${\boldsymbol \chi}$, and ${\boldsymbol q}_{ij}^\prime$ is the fluctuation. Compatibility is understood as follows. Given that ${\boldsymbol q}_i$ and ${\boldsymbol q}_j$ are in ${\mathcal C}_\beta$, the average positions are prescribed by
\begin{equation}
\label{av-positions}
\overline{{\boldsymbol q}}_{ij}= \nabla {\boldsymbol \chi} (t, {\boldsymbol X}_\beta)({\boldsymbol X}_i-{\boldsymbol X}_j),
\end{equation}
where ${\boldsymbol X}_\beta$ is a point which for present purposes is associated with the cell ${\mathcal C}_\beta$.
A natural choice of ${\boldsymbol X}_\beta$ is
$$
{\boldsymbol X}_\beta={\boldsymbol \chi}^{-1}(t, \bx_\beta)
$$
(the pre-image of the cell center under the inverse average deformation map).
It is important to recognize that, in general, ${\boldsymbol X}_\beta$ may change in time, and that it need not lie within ${\mathcal C}_\beta$.
The points ${\boldsymbol X}_i$ and $ {\boldsymbol X}_j$ lie on a fixed periodic lattice
which may be identified with the undeformed reference lattice covering the whole initial flow domain $\Omega$. The corresponding lattice vectors have equal length determined by placing $N$ particles at ${\boldsymbol X}_i$ in $\Omega$.
The actual initial positions ${\boldsymbol q}_i^0, i=1, 2,\dots,N$, constitute a perturbation of the reference lattice. These perturbations need not be small. We require only that the distribution of the fluctuations ${\boldsymbol q}_i^{\prime, 0}={\boldsymbol q}_i^0-{\boldsymbol X}_i$ be orientation-independent in the sense to be made precise in Sect.~\ref{empiricalstatistics}.
With this choice of ${\boldsymbol X}_i$, the points $\overline{{\boldsymbol q}}_{i}(t)$ within each cell ${\mathcal C}_\beta$ also form a lattice at each instant $t$. The deformations of these lattices relative to the reference lattice are determined by $\nabla{\boldsymbol \chi}(t, \bx_\beta)$. The deformed lattices may consequently differ from cell to cell. In particular, differences in lattice orientation may account for meso-scale vortices.
Once $\overline{{\boldsymbol q}}_i$, $i=1,2,\dots,N$, are chosen, the fluctuations are determined by the actual particle positions, and we require that
\begin{equation}
\label{zero-av-fluct}
\sum_{i, j\in I_\beta} {\boldsymbol q}_{ij}^\prime=\bf0.
\end{equation}
This assumption is consistent with the expectation that particles can be locally ordered on the basis of the average deformation gradient, and, moreover, that the differences between actual and average positions are uncorrelated within a cell.
In the sequel, we will analyze the forces between particles located at nearby lattice points.
These forces depend on the average relative positions $\overline{{\boldsymbol q}}_{ij}$, $i,j=1,2,\dots,N$. Although the number of lattice sites within each cell is large, the assumption that the range $R$ of the DPD forces is much smaller than $\eta={\mathcal V}_\beta^{1/d}$ implies that the number of relevant neighbors of any given site must be small. In any representative cell ${\mathcal C}_\beta$, the vector $\overline{{\boldsymbol q}}_{ij}$ can therefore only assume one of a small number $A$ of values denoted by
\begin{equation}
\label{q-al}
\overline{{\boldsymbol q}}^{\alpha\beta},~~~~~~~\alpha=1, 2, \dots, A,
\end{equation}
where the index $\alpha$ serves to enumerate different possible values within ${\mathcal C}_\beta$.
We refer to $\overline{{\boldsymbol q}}^{\alpha\beta}$, $\alpha=1,2,\dots,A$, $\beta=1,2,\dots,B$, as the \emph{relative lattice vectors}. The number $A$ of these vectors in a given cell $\beta$ is set by the lattice geometry, the range of the DPD forces, and the local density (average interparticle distance). For example, in a two-dimensional triangular lattice with nearest neighbor interactions, $\alpha=6$. If next-to-nearest neighbor interactions are also relevant, then $\alpha=16$. Moreover, in a three-dimensional cubic lattice with only nearest neighbors being relevant, $\alpha= 17$. Relative lattice vectors corresponding to one site of the two-dimensional triangular lattice are shown in Fig. 3.
\begin{figure}[!t]
\label{Fig:rel-lattice}
\centering
\includegraphics[width=0.375\textwidth]{fig3}
\caption{Relative lattice vectors associated with one lattice site. The dashed circle shows the range of the DPD forces.}
\end{figure}
The extent to which local lattice vectors stretch relative to the reference lattice is determined by the parameters $s_{\alpha\beta}$, as defined by
\begin{equation}
\label{q-al1}
\left |\overline{{\boldsymbol q}}^{\alpha\beta}\right|=s_{\alpha\beta} e_{\alpha},
\end{equation}
where $e_{\alpha}$ is the length of the corresponding local lattice vector in the underformed lattice.
When a local deformation gradient is nearly spherical, the corresponding local lattice deformation is close to a uniform expansion (or contraction), as characterized by
\begin{equation}
\label{true-stretch}
\hat s_\beta(t)=\left({\rm det}\nabla{\boldsymbol \chi}(t, \bx_\beta)\right)^{1/d}.
\end{equation}
Given a particular lattice geometry, it is also possible to estimate $\hat s_\beta$ using the average concentration
\begin{equation}
\frac{n_\beta}{{\mathcal V}_\beta}.
\end{equation}
Indeed, the number $n_\beta$ of particles within the cell ${\mathcal C}_\beta$ can be approximated by
$$
n_\beta\approx \frac{{\mathcal V}_\beta}{c_d(\hat s_\beta{\ell})^d},
$$
where $c_d$ is a constant that depends only on the lattice geometry and the dimension $d$ of the underlying point space, and where $\ell$ is the length of the reference lattice vector connecting nearest neighbors. The concentration can be related to the mass density $\overline{\rho}_\beta=\overline{\rho}(\cdot,\bx_\beta)=m\mskip1mun_\beta/{\mathcal V}_\beta$
by writing
$$
\frac{1}{(\hat s_\beta \ell)^d}\approx \frac{c_d\mskip1mun_\beta}{{\mathcal V}_\beta}=\frac{c_d\mskip1mu\overline{\rho}_\beta}{m},
$$
where we recall that $m$ denotes the mass of a single DPD particle. Solving the foregoing relation for $\hat s_\beta$ yields
\begin{equation}
\label{length-density}
\hat s_\beta \approx \frac{1}{\ell}\left(\frac{m}{c_d\mskip1mu\overline{\rho}_\beta}\right)^{1/d}.
\end{equation}
Hereafter, it is convenient to utilize the local lattice length scale $|\overline{{\boldsymbol q}}|_\beta$ defined by
\begin{equation}
\label{local-length}
|\overline{{\boldsymbol q}}|_\beta=\hat s_\beta{\ell}.
\end{equation}
It is worth mentioning that
the affine component $\overline{{\boldsymbol q}}_{ij}^\beta=\nabla{\boldsymbol \chi}_\beta ({\boldsymbol X}_i-{\boldsymbol X}_j)$ of the relative displacement in the cell
${\mathcal C}_\beta$ is obtained by identifying the actual particle deformations with the average deformation ${\boldsymbol \chi}$. It can be thus said that $\overline{{\boldsymbol q}}_{ij}$ complies with the Cauchy--Born rule, while the fluctuating component ${\boldsymbol q}_{ij}^\prime$ describes possible violations of the rule.
While the Cauchy--Born rule is typically associated with deformations of crystalline materials, it should be still relevant for dense fluids and soft matter provided that kinetic energy of velocity fluctuations is sufficiently small compared to the kinetic energy of the average motion.
Moreover, our averaging scheme rests on the introduction of a lattice. This is done to enable efficient calculation of the constitutive equations, since for periodic arrays, the pair contributions in the constitutive equations become highly repetitive. This does not prohibit geometric irregularities in the actual particle placement. Rather, introduction of the local lattice corresponds to the expectation that average deformation gradient is close to piecewise constant (and, thus, that the average deformation itself is nearly piecewise linear) at the chosen meso-scale. Granted that the separation between the averaging scale and the fine scale is sufficiently large and that the initial conditions do not contain strong oscillations at the fine scale, the average deformation should be free of small-scale fluctations, and the above perturbed lattice picture of the local deformation should be reasonable. It is also worth noting that imposing \eqref{zero-av-fluct} is tantamount to stipulating that the Cauchy--Born rule holds on average.
Lattices of many different geometries may be compatible with the same average deformation. It is therefore important to have a method for choosing a specific lattice geometry that best fits the available information, namely the values of the average mass density and velocity and the initial conditions for the DPD model. From the initial conditions we can extract the initial average coordination number in each cell. Combining this information with the knowledge of the initial mass density in this cell (which determines the particle population in each cell) provides a selection method for choosing a unique isotropic lattice at the initial instant of time. The same lattice also serves as the reference lattice. The simplest version of this approach, described above, would yield the same lattice in each cell, provided that the initial density is constant and that the coordination number is the same in all cells. This places restrictions on the initial conditions. A more sophisticated and broadly applicable variant of this approach would be to use the local cell coordination numbers together with the values of the local mass density. Such a strategy might be useful for treating non-uniform initial conditions and might cause the initial (reference) local lattices to be of different geometry---for instance, cubic in one cell and tetrahedral in another cell.
We next provide a criterion for selecting local lattices at subsequent instants of time. Given a cell ${\mathcal C}_\beta$, the simplest option is to assign to it the same lattice geometry chosen for the pre-image ${\boldsymbol \chi}^{-1}({\mathcal C}_\beta)$ at the initial time.
However, doing so is not necessarily optimal because it may result in a local lattice length incompatible with the length changes induced by the average deformation. We therefore select the lattice geometry that minimizes the discrepancy between the value of $\hat s_\beta$ given by \eqref{length-density} and the value given by \eqref{true-stretch}.
This process may result in local lattice geometry which changes in time at a given location. Of course, variations in geometry from one location to another are also possible. Overall, such an approach could be viewed as a relaxation of the standard Cauchy--Born rule. Because of its comparative flexibility, the relaxed version should be applicable to crystalline solids, amorphous solids, soft matter, and dense fluid systems at sufficiently low temperatures.
\subsection{Empirical statistics and fluctuation-based closure}
\label{empiricalstatistics}
We rely on a closure strategy that is simple in the sense that it employs Taylor approximations up to the second order in fluctuations.
The corresponding calculations are straightforward but lengthy. For this reason, most of the details are relegated to the Appendix.
The resulting constitutive relations incorporate the tensorial second moments of fluctuations of both the positions and the velocities of the DPD particles. These fluctuation tensors will be now described in more detail.
The position fluctuation tensor is defined by
\begin{equation}
\label{av-fluct-tensor}
{\boldsymbol Q}^{\prime, {\alpha\beta}}=\sum_{(i, j)\in J_{\alpha\beta}}{\boldsymbol q}^\prime_{ij}\otimes {\boldsymbol q}^\prime_{ij}
\end{equation}
where$J_{\alpha\beta}$ denote the index set of pairs $(i, j)$ such that $\overline{{\boldsymbol q}}_{ij}=\overline{{\boldsymbol q}}^{\alpha\beta}$ for some $\alpha$ within a generic cell ${\mathcal C}_\beta$.
Further, velocity fluctuations are embodied by the second order tensor
\begin{equation}
\label{T0}
\overline{{\boldsymbol v}^\prime\otimes {\boldsymbol v}^\prime}^\beta=\frac{1}{n_\beta}\sum_{j\in I_\beta} {\boldsymbol v}^\prime_j \otimes {\boldsymbol v}^\prime_j
\end{equation}
We assume that the DPD particles in each cell are in local thermodynamic equilibrium, in which case the velocity fluctuation tensor must be nearly spherical and can be characterized by one scalar parameter through a relation of the form
\begin{equation}
\overline{{\boldsymbol v}^\prime\otimes {\boldsymbol v}^\prime}^\beta=\theta{\boldsymbol I},
\label{T}
\end{equation}
where $\theta$ is a temperature-like quantity describing the strength of the velocity fluctuations.
The closure method relies on the following assumptions.
\begin{enumerate}
\item \emph{The moments of all fluctuations of order greater than two are small in comparison to the second moments.}
\item \emph{The second order tensors $\overline{{\boldsymbol v}^\prime\otimes {\boldsymbol v}^\prime}^\beta$ and ${\boldsymbol Q}^{\prime, {\alpha\beta}}$ are nearly uniform in time for all combinations of $\alpha$ and $\beta$.}
\end{enumerate}
The first assumption permits us to truncate Taylor expansions to the second order in fluctuations. Turning to the second assumption, we note that
the portion pertaining to $\overline{{\boldsymbol v}^\prime\otimes {\boldsymbol v}^\prime}$ is reasonable for dense, isothermal flows. A justification of the portion of assumption pertaining to ${\boldsymbol Q}^\prime$ is provided in Appendix \ref{sect:positions-fluct}.
The second assumption also allows us to estimate fluctuations from the DPD initial conditions. Since the DPD model is already an average of a molecular model (associated with a much finer length scale), the initial conditions for DPD may be assumed to be deterministic and known precisely, as is usually done for Langevin-type equations. If the DPD initial conditions are not available, it is instead possible to specify a probability distribution of the initial conditions and to then use ensemble averaging in conjunction with spatiotemporal averaging.
Aside from the foregoing assumptions, several additional assumptions are imposed below. These assumptions lead to significantly simplified constitutive relations. This resulting theory is physically reasonable and provides explicit constitutive relations expressing the pressure and viscosity as functions of the average deformation gradient and the fluctuation tensors entering the second of the above assumptions. If working with more complicated constitutive equations is feasible, it seems possible to relax these assumptions and develop a more accurate closed-form model on their basis.
Finally, we remark that the fluctuation-dependent quantities are reminiscent of a more general notion of ephemeral continua \cite{capriz-fried-seguin,capriz-ephemeral} From that perspective, the special nature of the case under discussion stems from the decision to explore the consequences of having only finitely many ``material points" at the mesoscopic scale. In addition to the loss of information mentioned above, the placement of the relevant points (or, equivalently, the placement of averaging cells) is related to their size, which in the present case is equal to the averaging scale $\eta$. This explicit scale dependence should be considered as one of the fundamental distinctions between a mesosopic model and a classical continuum model. An even more general framework arises when $\eta$ differs from the distance $\xi$ between the centers of two adjacent cells. The resulting constitutive equations would then depend on both length scales. In the event that the ratio of these length scales is fixed, the features of the resulting theory are essentially the same as those of the theory presented here. Significant differences could arise in the case when $\eta$ and $\xi$ are widely disparate, for instance when $\xi/\eta\to 0$ and at the same time
$\xi/R\to \infty$. In this case the averaging cells would significantly overlap, and a deconvolution closure strategy like that described by Panchenko, Barannyk and Gilbert \cite{PBG} could be used in conjunction with the truncated Taylor formula closure described in the subsequent sections. The purpose of the deconvolution closure would be to recover the unknown averages at the smaller scale $\xi$ from the available averages at the larger scale $\eta$. After this is done, closure could be achieved using truncated Taylor expansions and empirical fluctuation statistics.
\section{Summary of closed-form continuum equations}
\label{sect:cont-summary}
The exact equations of balance are approximated by the closed-form continuum equations
\begin{equation}
\label{mass-balance}
\partial_t \overline{\rho}+{\rm div}\overline{{\boldsymbol p}}=0
\end{equation}
and
\begin{equation}
\label{m-balance-closed}
\partial_t \overline{{\boldsymbol p}}+{\rm div}(\mskip1mu\overline{{\boldsymbol p}}\otimes \overline{{\boldsymbol v}}\mskip1mu)=-\nabla
\left(
\theta\mskip1mu\overline\rho
\right)
+{\rm div}{\overline{{\boldsymbol T}}}+ (K_1(\theta)-K_2(\theta)|\overline{\boldsymbol v}|^2) \overline{\boldsymbol v}+\overline{{\boldsymbol g}}^R.
\end{equation}
In \eqref{m-balance-closed}, and in the remainder of the paper, we use the superposed bar notation to emphasize the fundamental difference between an exact quantity such as the stress \eqref{exact-stress} and its closed-form approximation \eqref{total-stress-summary} in the form of a constitutive equation.
The first term on the right-hand side of \eqref{m-balance-closed} is the effective convective stress given by the gradient of the corresponding pressure $\theta\overline{\rho}$. The quantity ${\overline{\boldsymbol T}}$ in the second term on the right-hand side of \eqref{m-balance-closed} is the effective interaction stress and will be discussed below. The third term on the right-hand side of \eqref{m-balance-closed}, which contains ${\boldsymbol v}$, is the effective self-propulsion force. Finally, $\overline{{\boldsymbol g}}^R$ represents a closed-from approximation of the average fluctuation force.
Since the constitutive relation for the effective interaction stress turns out to be conventional, we include the derivation of the constitutive equations for the convective stress, self-propulsion forces, and the average fluctuation force in the main body of the paper and relegate the laborious calculations involved in the derivation of $\overline{{\boldsymbol T}}$ to the Appendix.
Before turning to the derivations, some comments are in order. We first consider the self propulsion force and $\overline{{\boldsymbol g}}^R$. The constants $K_1$ and $K_2$ in the definition of the self-propulsion force obey
\begin{equation}
K_1>0
\qquad\text{and}\qquad
K_2>0.
\end{equation}
We therefore see that our method recovers the typical cubic nonlinearity which appears with only intuitive justification in many phenomenological models of collective behavior (see, for example, Toner and Tu \cite{toner-tu95}, Dunkel et al.~\cite{dunkel2013}, and Marchetti et al.~\cite{marchetti-review}). We arrive at this expression by rigorously upscaling a physically realistic microscale DPD self-propulsion force~\eqref{dpd-sp-force}. Importantly that force bears no resemblance to effective force that is obtained by upscaling. Note that the form of the continuum self-propulsion term mainly depends on how \eqref{dpd-sp-force} and \eqref{choice-h} are chosen, which we decide on the basis of the microscopic physics, as discussed in Section~\ref{sect:dpd}.
The average fluctuation force $\overline{{\boldsymbol g}}^R$ is a Gaussian random field
with vanishing mean and variance $\sigma$. The variance is a time- and position-dependent state variable determined constitutively as a function of $\nabla{\boldsymbol \chi}$, $\theta_q$, and the temperature $T$. When the gradient of the local average deformation is nearly spherical, constitutive dependence of $\sigma$ on $\nabla{\boldsymbol \chi}$ reduces to dependence on the average mass density $\overline\rho$.
The constitutive equation for the variance appears to be a new contribution which may be of broad interest in developing stochastic evolution equations for active continua. In the phenomenological approach, the macro-scale fluctuation force is often linked to the fluid viscosity by a formally postulated fluctuation-dissipation relation. In contrast, bearing in mind that such a relation must hold at the micro-scale (see Eq. \eqref{FDrelns}), the parameters of the fluctuation force at the macro-scale cannot be chosen based on the fluctuation-dissipation relation. Instead, the variance of the fluctuation force is expected to vary in space and time following the evolution of local particle patterns. Put differently, the more concentrated an active suspension is, and the greater the tendency of the system to self-organize, the less likely it is for the average fluctuation force to exhibit a fixed variance. Therefore, one of the more important contributions of this work is the quantification of this expectation in the form of a constitutive relation for $\sigma$. That relation appears to yield a useful refinement of the Toner--Tu equations, which do not posses this feature. Further testing of this finding by simulation and experiments may be warranted.
Another important comment concerning $\overline{{\boldsymbol g}}^R$ is that the variance $\sigma$ depends on the extent to which the scales are separated, as dictated by the ratio of the cell size $\eta$ to the typical range of $R$ of the DPD forces. In Section \ref{sect:rand2}, we prove that $\sigma\to0$ as $\eta/R\to 0$. Consequently, we infer that the model becomes deterministic in the limit of infinite scale separation.
Finally, we comment briefly on the constitutive relation
\begin{equation}
\label{total-stress-summary}
{\overline{{\boldsymbol T}}}=-P(\overline{\rho}, \theta_q) {\boldsymbol I}+{\boldsymbol \mu}(\overline{\rho}, \theta_q) {\boldsymbol e}(\mskip1mu\overline{\boldsymbol v}\mskip1mu)
\end{equation}
for the interaction stress. In \eqref{total-stress-summary}, ${\boldsymbol e}(\mskip1mu\overline{\boldsymbol v}\mskip1mu)$ denotes the symmetric component of the gradient of the average velocity. The overall structure of this constitutive relation is therefore reminiscent of that underlying the Navier--Stokes equations. That said, we emphasize that the pressure $P$ is determined by an unconventional equation of state in terms of the average mass density $\overline{\rho}$ and the fluctuation strength $\theta_q$ of relative positions. Importantly, $\theta_q$ is generally distinct from the fluctuation strength $\theta$ of relative velocities. Like $P$, the viscosity tensor ${\boldsymbol \mu}$ generally varies with both $\overline\rho$ and $\theta_q$.
Derivations of the conservative and viscous contributions to the interaction stress $\overline{{\boldsymbol T}}$
are provided in Sections~\ref{sect:cons-closure} and \ref{sect:diss-closure} of the Appendix.
\section{Averaging the self-propulsion force and convective stress}
\label{sect:sp-cs-closure}
\subsection{Self-propulsion force}
For $\psi_\eta$ of the form \eqref{psi-box}, the average self-propulsion force density (which is a nonlinear volume average) is given by
$$
{{\boldsymbol g}}_\beta^{\textit{SP}}=\frac{1}{n_\beta} \sum_{j\in I_\beta} g(|{\boldsymbol v}_j|^2) {\boldsymbol v}_j=\sum_{j\in I_\beta} h(|{\boldsymbol v}_j|) {\boldsymbol v}_j.
$$
We now approximate ${\boldsymbol g}_\beta^{\textit{SP}}$ by a function of the average velocity $\overline{{\boldsymbol v}}_\beta$.
Toward this, we write
$$
{\boldsymbol v}_j=\overline{{\boldsymbol v}}_\beta+{\boldsymbol v}^\prime_j,
$$
and use Taylor's theorem to expand $g(|\overline{{\boldsymbol v}}_\beta+{\boldsymbol v}^\prime_j|^2)(\overline{{\boldsymbol v}}_\beta+{\boldsymbol v}^\prime_j)$ about $\overline{{\boldsymbol v}}_\beta$. Keeping only terms up to the second order in the velocity fluctuation ${\boldsymbol v}^\prime_j$, we find that
\begin{multline}
g(|\overline{{\boldsymbol v}}_\beta+{\boldsymbol v}^\prime_j|^2)(\overline{{\boldsymbol v}}_\beta+{\boldsymbol v}^\prime_j)
=
g(|\overline{{\boldsymbol v}}_\beta|^2)\overline{{\boldsymbol v}}_\beta+2g^\prime(|\overline{{\boldsymbol v}}_\beta|^2) \overline{{\boldsymbol v}}_\beta \cdot {\boldsymbol v}^\prime_j \overline{{\boldsymbol v}}_\beta +g(|\overline{{\boldsymbol v}}_\beta|^2) {\boldsymbol v}^\prime_j
\\
+
2 g^{\prime\prime}(|\overline{{\boldsymbol v}}_\beta|^2)\overline{{\boldsymbol v}}_\beta
(\overline{{\boldsymbol v}}_{\beta}\otimes \overline{{\boldsymbol v}}_\beta ) :
({{\boldsymbol v}}^\prime_j \otimes {{\boldsymbol v}}^\prime_j) \overline{{\boldsymbol v}}_\beta\\
+
g^\prime (|\overline{{\boldsymbol v}}_\beta|^2) |{\boldsymbol v}^\prime_j|^2 \left({{\boldsymbol v}}^\prime_j \otimes {{\boldsymbol v}}^\prime_j \right) \overline{{\boldsymbol v}}_\beta
+
2 g^\prime (|\overline{{\boldsymbol v}}_\beta|^2) \overline{{\boldsymbol v}}_\beta \cdot {\boldsymbol v}^\prime_j {\boldsymbol v}^\prime_j+ \cdots.
\label{gex1}
\end{multline}
Averaging both sides of \eqref{gex1} while taking into consideration the identity $n_\beta^{-1} \sum_{j\in I_\beta} {\boldsymbol v}^\prime_j=0$ together with \eqref{T}, we obtain
\begin{equation}
\label{v-av}
\frac{1}{n_\beta} \sum_{j\in I_\beta}
g(|\overline{{\boldsymbol v}}_\beta+{\boldsymbol v}^\prime_j|^2)(\overline{{\boldsymbol v}}_\beta+{\boldsymbol v}^\prime_j) \approx
(g(|\overline{{\boldsymbol v}}_\beta|^2)+ (2+d)\theta g^\prime(|\overline{{\boldsymbol v}}_\beta|^2)+
2\theta g^{\prime\prime}(|\overline{{\boldsymbol v}}_\beta|^2)|\overline{{\boldsymbol v}}_\beta|^2)
\overline{{\boldsymbol v}}_\beta,
\end{equation}
where, as before, $d$ is the spatial dimension.
For small $|\overline{{\boldsymbol v}}_\beta|$, \eqref{v-av} simplifies further to
$$
g(|\overline{{\boldsymbol v}}_\beta|^2)+ (2+d)\theta g^\prime(|\overline{{\boldsymbol v}}_\beta|^2)+
2\theta g^{\prime\prime}(|\overline{{\boldsymbol v}}_\beta|^2)|\overline{{\boldsymbol v}}_\beta|^2 \approx
K_1 - K_2 |\overline{{\boldsymbol v}}_\beta|^2,
$$
where $K_1$ and $K_2$ are defined by
\begin{equation}
K_1= g(0)+ (2+d)\theta g^\prime(0)
\qquad\text{and}\qquad
K_2 = -(4+d)\theta g^{\prime\prime}(0).
\label{K12gen}
\end{equation}
For the particular choice $g(\xi)=(\xi^2+\delta^2)^{-1/2}$, \eqref{K12gen} specializes to yield
\begin{equation}
\label{Ks}
K_1=g(0)=\delta^{-1}
\qquad\text{and}\qquad
K_2=(4+d)\theta \delta^{-3},
\end{equation}
which results in the constitutive equation
\begin{equation}
\label{self-prop-const}
{{\boldsymbol g}}^{\textit{SP}}\approx \overline{{\boldsymbol g}}^{\textit{SP}}=
(\delta^{-1}-(4+d)\theta \delta^{-3} |\overline{{\boldsymbol v}}_\beta|^2) \overline{{\boldsymbol v}}_\beta.
\end{equation}
\subsection{Convective stress}
With reference to \eqref{exact-momentum}, the convective stress is given by
\begin{equation}
m\sum_{i=1}^N{\boldsymbol v}^\prime_i\otimes {\boldsymbol v}^\prime_i \psi_\eta({\boldsymbol x}-{\boldsymbol q}_i).
\label{convstress}
\end{equation}
If the window function $\psi_\eta$ has the particular form \eqref{psi-box} and the velocity fluctuations obey the assumption \eqref{T}, then \eqref{convstress} specializes to
\begin{equation}
\label{conv-closure}
\frac{m}{{\mathcal V}_\beta} \sum_{i\in I_\beta} {\boldsymbol v}^\prime_i\otimes {\boldsymbol v}^\prime_i =
\frac{m}{{\mathcal V}_\beta} n_\beta \theta {\boldsymbol I}=\overline{\rho}_\beta\theta {\boldsymbol I}.
\end{equation}
\section{Constitutive equation for the fluctuation force}
\label{sect:rand1}
For the particular choice \eqref{psi-box} of the weight function $\psi_\eta$, the average \eqref{R0} of the fluctuation forces ${\boldsymbol f}^{R}_{ij}$ defined in \eqref{rand-force} yields
\begin{equation}
\label{R1}
{\boldsymbol g}^R(t, \bx_\beta)=\sum_{i, j=1}^N {\boldsymbol f}^R_{ij} \psi_\eta({\boldsymbol x}-{\boldsymbol q}_i)=
\frac{2 \gamma k_B T}{\mathcal{V}_\beta}\sum_{i\in I_\beta}\sum_{j=1}^N\xi_{ij}
\sqrt{w^D(r_{ij})}\,{\boldsymbol e}_{ij},
\end{equation}
where $i$ is an element of $I_\beta$ if and only if particle $i$ is located in cell ${\mathcal C}_\beta$ and where we have invoked the fluctuation-dissipation relations \eqref{FDrelns}. Nontrivial contributions to the double sum may arise under a variety of circumstances. Consider two particles labeled $i$ and $j$. Then, the corresponding contribution to \eqref{R1} is potentially nontrivial if particles $i$ and $j$ both belong to ${\mathcal C}_\beta$ or if particle $i$ is in ${\mathcal C}_\beta$ and particle $j$ is outside of ${\mathcal C}_\beta$ but within the range of the force, in which case $w^D(r_{ij})\ne 0$. However, because ${\boldsymbol f}^R_{ij}=-{\boldsymbol f}^R_{ji}$, the contribution to \eqref{R1} vanishes if particles $i$ and $j$ both belong to ${\mathcal C}_\beta$. Thus, \eqref{R1} reduces to
\begin{equation}
\label{R2}
{\boldsymbol g}^R(t, \bx_\beta)=\frac{2 \gamma k_B T}{\mathcal{V}_\beta}\sum_{i\in I_\beta}\sum_{j\notin I_\beta}
\xi_{ij}\sqrt{w^D(r_{ij})}\,{\boldsymbol e}_{ij}.
\end{equation}
Since ${\boldsymbol q}_i$ and ${\boldsymbol q}_j$ are random variables that depend on the history of the motion, it is quite difficult to describe the probability distribution of ${\boldsymbol g}^R$. However, a reasonable approximation can be developed by assuming that the dynamics are discrete in time. Calculating positions and velocities at a generic time step then proceeds by (i) inserting $r_{ij}$ and ${\boldsymbol e}_{ij}$ obtained at the previous time step into the equations \eqref{rand-force}, (ii) multiplying by $\xi_{ij}$, and (iii) updating positions and velocities. The central point is that $\xi_{ij}$ are independent, identically distributed normal random variables with zero mean and unit variance,
and $\xi_{ij}$ are produced using, for example, a suitable random number generator, and without taking into account any information about $r_{ij}$ and ${\boldsymbol e}_{ij}$. Thus $\xi_{ij}$, $i,j,=1,2,\dots,N$ may be assumed to be statistically
independent of $r_{ij}$ and ${\boldsymbol e}_{ij}$, $i,j,=1,2,\dots,N$.
Therefore, at each time step, ${\boldsymbol g}^R_\beta$ can be treated as a linear combination $\sum_{i \in I_\beta} \sum_{j\notin I_\beta}{\boldsymbol a}_{ij} \xi_{ij}$ of normal random variables $\xi_{ij}$ with the coefficients
\begin{equation}
{\boldsymbol a}_{ij}=\frac{2 \gamma k_B T}{\mathcal{V}_\beta}\sqrt{w^D(r_{ij})}\,{\boldsymbol e}_{ij}.
\end{equation}
Standard results from probability theory lead to the conclusion that each of the $d$ components of ${\boldsymbol g}^R(\cdot,{\boldsymbol x}_\beta)$ is a normal random variable with zero mean and variance
\begin{equation}
\label{variance}
\sigma^{(k)}_\beta=\frac{2\gamma k_B T}{\mathcal{V}_\beta}
\sqrt{\sum_{i \in I_\beta}
\sum_{j\notin I_\beta}
w^D(r_{ij})e_{ij}^{(k)}},
\qquad
k=1,\dots,d,
\end{equation}
where $e_{ij}^{(k)}$ is the component of ${\boldsymbol e}_{ij}$ in the direction of the $k$-th basis element. Assuming that all components of ${\boldsymbol g}^R(\cdot,{\boldsymbol x}_\beta)$ are equally distributed yields
\begin{equation}
\label{iso-var}
\sigma^{(k)}_\beta=\sqrt{\frac{1}{d}\sum\limits_{l=1}^d(\sigma^{(l))}_\beta)^2}=
\sigma_\beta=\frac{2\gamma k_B T}{\mathcal{V}_\beta}\sqrt{\frac{1}{d}\sum_{i \in I_\beta}\sum_{j\notin I_\beta}w^D(r_{ij})}\,.
\end{equation}
The variance therefore becomes another state variable that requires closure. Expanding $w^D$ to the second order in ${\boldsymbol q}^\prime$, we find that
\begin{multline}
\label{single-term}
w^D(r_{ij})=w^D(|\overline{{\boldsymbol q}}_{ij}|)+\left(w^D\right)^\prime(|\overline{{\boldsymbol q}}_{ij}|)
\frac{\overline{{\boldsymbol q}}_{ij}}{|\overline{{\boldsymbol q}}_{ij}|}
\cdot {\boldsymbol q}^\prime_{ij}
\\[4pt]+
\frac 12 \left[
\left(w^D\right)^{\prime\prime}(|\overline{{\boldsymbol q}}_{ij}|)\frac{\overline{{\boldsymbol q}}_{ij}\otimes \overline{{\boldsymbol q}}_{ij}}
{|\overline{{\boldsymbol q}}_{ij}|^2}
- \left(w^D\right)^{\prime}(|\overline{{\boldsymbol q}}_{ij}|)\frac{|\overline{{\boldsymbol q}}_{ij}|^2 {\boldsymbol I}-\overline{{\boldsymbol q}}_{ij}\otimes \overline{{\boldsymbol q}}_{ij}}
{|\overline{{\boldsymbol q}}_{ij}|^3}
\right]: ({\boldsymbol q}^\prime_{ij}\otimes {\boldsymbol q}^\prime_{ij})+ \cdots,
\end{multline}
where $\left(w^D\right)^\prime$ and denotes the derivative of $w^D$ with respect to its argument, and similarly for $\left(w^D\right)^{\prime\prime}$.
Inserting \eqref{single-term} in \eqref{R2} and, as before, first summing the terms with $\overline{{\boldsymbol q}}_{ij}=\overline{{\boldsymbol q}}^{\alpha\beta}$ with a fixed relative lattice vector $\overline{{\boldsymbol q}}^{\alpha\beta}$ and then summing over $\alpha$, we find that
\begin{multline}
\label{rand-cl1}
\sigma_\beta^2 \approx \frac{4 \gamma^2 (k_B T)^2}{d (\mathcal{V}_\beta)^2}
\left(
\sum_{\alpha} w^D(|\overline{{\boldsymbol q}}^{\alpha\beta}|)
|S_\alpha|
+
\sum_{\alpha} \left(w^D\right)^\prime (|\overline{{\boldsymbol q}}^{\alpha\beta}|)
\frac{\overline{{\boldsymbol q}}^{\alpha\beta}}{|\overline{{\boldsymbol q}}^{\alpha\beta}|}
\cdot \sum_{(i,j)\in S_\alpha} {\boldsymbol q}^\prime_{ij}
\right)
\\[6pt]
+ \frac{2 \gamma^2 (k_B T)^2}{d (\mathcal{V}_\beta)^2}
\sum_{\alpha}
\left[
\left(w^D\right)^{\prime\prime}(|\overline{{\boldsymbol q}}^{\alpha\beta}|)\frac{\overline{{\boldsymbol q}}^{\alpha\beta}\otimes \overline{{\boldsymbol q}}^{\alpha\beta}}
{|\overline{{\boldsymbol q}}^{\alpha\beta}|^2}\right.
- \left.\left(w^D\right)^{\prime}(|\overline{{\boldsymbol q}}^{\alpha\beta}|)\frac{|\overline{{\boldsymbol q}}^{\alpha\beta}|^2 {\boldsymbol I}-\overline{{\boldsymbol q}}^{\alpha\beta}\otimes \overline{{\boldsymbol q}}^{\alpha\beta}}
{|\overline{{\boldsymbol q}}^{\alpha\beta}|^3}
\right]:\sum_{(i,j)\in S_\alpha} {\boldsymbol q}^\prime_{ij}\otimes {\boldsymbol q}^\prime_{ij},
\end{multline}
where $S_\alpha$ is the index set defined by
\begin{equation}
S_\alpha=
\{
(i, j): i\in I_\beta, j \notin I_\beta, \overline{{\boldsymbol q}}_{ij}=\overline{{\boldsymbol q}}^{\alpha\beta}
\}
\end{equation}
and $|S_\alpha|$ denotes the number of elements in $S_\alpha$.
Since $|S_\alpha|$ is expected to be large, the simplest reasonable closure assumptions are
\begin{equation}
\label{rand-cl2}
\sum_{(i,j)\in S_\alpha} {\boldsymbol q}^\prime_{ij}={\bf 0}
\qquad{\rm and}\qquad
\sum_{(i,j)\in S_\alpha}{{\boldsymbol q}}_{ij}^\prime \otimes {{\boldsymbol q}}^\prime_{ij}=\theta_q |\overline{{\boldsymbol q}}^{\alpha\beta}|^2{\boldsymbol I},
\end{equation}
which, when inserted into \eqref{rand-cl1}, gives
\begin{equation}
\label{rand-cl3}
\sigma_\beta^2 \approx \frac{4 \gamma^2 (k_B T)^2}{d (\mathcal{V}_\beta)^2}
\sum_{\alpha} w^D(|\overline{{\boldsymbol q}}^{\alpha\beta}|)
|S_\alpha|
+
\theta_q \frac{2\gamma^2 (k_B T)^2}{d (\mathcal{V}_\beta)^2}
\sum_{\alpha}
\left[
\left(w^D\right)^{\prime\prime}(|\overline{{\boldsymbol q}}^{\alpha\beta}|)|\overline{{\boldsymbol q}}^{\alpha\beta} |^2
- (d-1)\left(w^D\right)^{\prime}(|\overline{{\boldsymbol q}}^{\alpha\beta}|)
{|\overline{{\boldsymbol q}}^{\alpha\beta}|}
\right],
\end{equation}
where the trivial identity ${\rm tr}(\overline{{\boldsymbol q}}^{\alpha\beta}\otimes \overline{{\boldsymbol q}}^{\alpha\beta})=|\overline{{\boldsymbol q}}^{\alpha\beta}|^2$ has been used.
Since $w^D(|\overline{{\boldsymbol q}}^{\alpha\beta}|)> 0$, the right-hand side of \eqref{rand-cl3} is guaranteed to be positive if $\theta_q$ is sufficiently small. More detailed analyses are possible for particular choices of $w^D$.
Since $\overline{{\boldsymbol q}}^{\alpha\beta}=\nabla {\boldsymbol \chi}(\bx_\beta){\boldsymbol e}^{\alpha}$, where ${\boldsymbol e}^{\alpha}$ is a relative lattice vector of the undeformed reference lattice, (\ref{rand-cl3}) provides a constitutive equation for the variance, given as a function of the average deformation gradient and other material parameters such as $\theta_q$, $T$, $\gamma$, and $w^D$.
If the local deformation is close to a uniform expansion or contraction, then $|\overline{{\boldsymbol q}}^{\alpha\beta}|\approx
|\overline{{\boldsymbol q}}|^\beta l_\alpha$, where $|\overline{{\boldsymbol q}}|^\beta$ is the length scale
of a uniformly deformed local lattice vector, as defined in \eqref{local-length}, and $l_\alpha$ is a non-dimensional parameter independent of the deformation. Since, consistent with \eqref{length-density}, $|\overline{{\boldsymbol q}}^\beta|\sim (\overline{\rho})^{-1/d}$, the right-hand side of \eqref{rand-cl3} becomes a function of mass density but also depends on the lattice geometry, ${\mathcal V}_\beta$, $d$, $\theta_q$, $\gamma$, and $T$. Thus, for nearly sperhical local deformations,
\begin{equation}
\label{sigma-iso}
\sigma_\beta^2 \approx \frac{4 \gamma^2 (k_B T)^2}{d (\mathcal{V}_\beta)^2}\left(F_1(\overline{\rho}_\beta)+
\frac 12 \theta_q F_2(\overline{\rho}_\beta)
\right),
\end{equation}
where $F_1$ and $F_2$ are determined by the following sums:
\begin{equation}
\left.
\begin{split}
\label{sigma-iso1}
F_1(\overline{\rho}_\beta) &=
\sum_{\alpha} w^D
\left(
\left(
\frac{m}{c_d \overline{\rho}_\beta}
\right)^{1/d}
\right)
|S_\alpha|,
\\[6pt]
F_2(\overline{\rho}_\beta) &=
\sum_{\alpha}
\left[
\left(w^D\right)^{\prime\prime}
\left(
(\frac{m}{c_d \overline{\rho}_\beta}
)^{1/d}
\right)
\left(\frac{m}{c_d \overline{\rho}_\beta}
\right)
^{2/d}
- (d-1)\left(w^D\right)^{\prime}
\left(
(\frac{m}{c_d \overline{\rho}_\beta}
)^{1/d}
\right)
\left(
\frac{m}{c_d \overline{\rho}_\beta}
\right)^{1/d}
\right].
\end{split}
\,\right\}
\end{equation}
In summary, the constitutive approximation $\overline{{\boldsymbol g}}^R$ of the exact fluctuation force ${\boldsymbol g}^R$ is obtained by choosing, at each instant of time,
a Gaussian random field with mean zero and variance given by \eqref{sigma-iso}. The time-correlation properties of $\overline{{\boldsymbol q}}^R$ are identical to those of ${\boldsymbol g}^R$.\
\section{Vanishing of the variance with increasing scale separation}
\label{sect:rand2}
Since the variance $\sigma_\beta$ is influenced by the local state of deformation, the distribution of the average fluctuation force may differ from location to location and may also evolve in time. It is therefore useful to provide as much insight as possible regarding the behavior of this distribution. With this objective in mind, we consider the impact of increasing the mesoscopic length scale $\eta$ with the range $R$ of the DPD forces held fixed. Since this corresponds to increasing the number of particles that contribute to the variance, intuition suggests that the ``randomness'' of the average fluctuation force
${\boldsymbol g}^R$ should decrease as $\eta$ increases. In support of this heuristic expectation, we next show that
\begin{equation}
\sigma_\beta\to0
\qquad\text{as}\qquad
\frac{\eta}{R}\to\infty.
\label{siglim}
\end{equation}
In taking the foregoing limit, we identify $R$ with the range of ${\boldsymbol f}_{ij}^R$ defined, with reference to the fluctuation-dissipation relations \eqref{FDrelns}, by the support of the window function $w^D$ associated with the dissipative force ${\boldsymbol f}^D_{ij}$. This limiting process also requires a condition on the mass density, namely that there exists a positive number $M$, with dimensions of mass per unit volume, independent of $\eta$ and $\beta$ such that
\begin{equation}
\label{rho-upper}
\overline{\rho}_\beta\leq M.
\end{equation}
Our estimate of the variance involves two steps
\begin{enumerate}
\item \emph{Estimate $|S_\alpha|$.} We begin with the observation that the nonvanishing contributions to $\sigma_\beta$ are comprised only of pairs $(i,j)$ such that particle $i$ lies in
${\mathcal C}_\beta$ and particle $j$ lies outside of ${\mathcal C}_\beta$. In addition, ${\boldsymbol f}_{ij}^R$ should not vanish identically, that is, the distance $r_{ij}$ between particles $i$ and $j$ should be less than the range $R$ of the DPD forces. All such particle pairs should be located in the rectangular shell
\begin{equation}
\label{S-bet}
S_\beta=\{ {\boldsymbol x}\in {\bf R}^d: {\rm dist}({\boldsymbol x},\partial{\mathcal C}_\beta)< R\}
\end{equation}
containing all points with Euclidean distance to the boundary $\partial{\mathcal C}_\beta$ of ${\mathcal C}_\beta$ less than $R$. Bearing in mind that $S_\beta$ has volume
\begin{equation}
\label{volume-S}
{\mathcal V}_{S_\beta}=(\eta+R)^d-(\eta-R)^d=2R\sum_{k=0}^{d-1}
(\eta+R)^{d-1-k}(\eta-R)^k
=2R d \eta^{d-1}+O(\eta^{d-2}),
\end{equation}
the number of particles inside $S_\beta$ can be estimated by the number density
\begin {equation}
\label{part-number}
n_{S_\beta}\approx \frac{\overline{\rho}_\beta{\mathcal V}_{S_\beta}}{m},
\end{equation}
where $m$ is the mass of one particle. Increasing $M$, if necessary, we find that
$$
n_{S_\beta}\leq \frac{2M}{m}R d \eta^{d-1}
$$
for all sufficiently large values of $\eta/R$. Thus, fixing $\alpha$ and noting that for each particle $i$ in $S_\beta$ there is at most one $\overline{{\boldsymbol q}}_{ij}$ with
$\overline{{\boldsymbol q}}_{ij}=\overline{{\boldsymbol q}}^{\alpha\beta}$, we arrive at the intermediate estimate
\begin{equation}
\label{e1}
|S_\alpha| \leq \frac{2MRd\eta^{d-1}}{m}.
\end{equation}
\item \emph{Estimate the remaining terms in $\sigma_\eta$}. Since $|\overline{{\boldsymbol q}}^{\alpha\beta}|\leq R$ and since $w^D$ is bounded along with its first and second derivatives, the sums appearing in \eqref{rand-cl3} can be estimated by
\begin{equation}
\label{e2}
\sum_\alpha (w^D)^\prime (|\overline{{\boldsymbol q}}^{\alpha\beta}|)|\overline{{\boldsymbol q}}^{\alpha\beta}|\leq
N_{\alpha\beta} \left(\sup \left|(w^D)^\prime\right|\right) R,
\end{equation}
where $N_{\alpha\beta}$ is the number of relevant $\overline{{\boldsymbol q}}^{\alpha\beta}$. Importantly, this number depends on $R$ and $\overline{\rho}_\beta$. For larger values of $\overline{\rho}_\beta$, the interparticle distance decreases, whereby $N_{\alpha\beta}$ tends to increase. However, if $\overline{\rho}_\beta$ satisfies \eqref{rho-upper}, it follows that there exists a positive number
$
N_{\max}
$
depending only on $R$ and $M$ such that
$$
N_{\alpha\beta}\leq N_{\max}.
$$
Thus,
\begin{equation}
\label{e2-1}
\sum_\alpha (w^D)^\prime (|\overline{{\boldsymbol q}}^{\alpha\beta}|)|\overline{{\boldsymbol q}}^{\alpha\beta}|\leq
N_{\max} \left(\sup \left|(w^D)^\prime\right|\right) R
\end{equation}
and the remaining sums in \eqref{rand-cl3} can be estimated similarly. Finally, using \eqref{e1} and \eqref{e2-1} in \eqref{rand-cl3} and recalling that ${\mathcal V}_\beta=\eta^d$, we obtain
\begin{multline}
\label{e3}
\sigma_\beta^2\leq \frac{8M\gamma^2 (k_B T)^2}{m}
(\sup w^D)
N_{\max} R \eta^{-d-1}
\\
+
\theta_q \frac{2\gamma^2 (k_B T)^2}{d}
N_{\max}
\left[
\sup\left|\left(w^D\right)^{\prime\prime}\right|
R^2\eta^{-2d}
+(d-1)
\sup\left|\left(w^D\right)^{\prime}\right|
R\eta^{-2d}
\right]
\\[4pt]
=
C_1(M, \gamma, T, R)
\left(
\frac{R}{\eta}
\right)^{d+1}+
C_2(M, \gamma, T, R, \theta_q)
\left(
\frac{R}{\eta}
\right)^{2d}
+
C_3(M, \gamma, T, R, \theta_q)
\left(
\frac{R}{\eta}
\right)^{2d},
\end{multline}
from which we conclude that $\sigma_\beta$ obeys the limit \eqref{siglim}.
\end{enumerate}
\section{Linear stability}
\label{sect:lin-stab}
We now study linear stability of constant solutions assuming infinite scale separation, so that, consistent with \eqref{siglim}, ${\boldsymbol g}^R=\bf0$. For simplicity, we restrict attention to two spatial dimensions, and assume that all relevant effective material parameters such the elasticity and viscosity tensors are constant and isotropic.
\subsection{Stability of uniform polar solution}
We first investigate the linear stability of the uniform polar solution $\rho(t, {\boldsymbol x})=\rho_0$, ${\boldsymbol v}(t, {\boldsymbol x})={\boldsymbol v}_0$, with $\rho_0$ and ${\boldsymbol v}_0$ being constants and with $|{\boldsymbol v}_0|=\sqrt{K_1/K_2}$.
Assuming that the mass density and velocity admit expansions of the form $\rho=\rho_0+\delta\rho$ and ${\boldsymbol v} ={\boldsymbol v}_0+\delta\bv} %{{\boldsymbol \epsilon}$, with $\delta\rho\ll\rho_0$ and $|\delta\bv} %{{\boldsymbol \epsilon}|\ll|{\boldsymbol v}_0|=\sqrt{K_1/K_2}$, we first formally linearize the mass balance \eqref{mass-balance} to yield
\begin{equation}
\label{lin-density}
\partial_t \delta\rho+\rho_0 {\rm div}\delta\bv} %{{\boldsymbol \epsilon}+\nabla(\delta\rho)\cdot {\boldsymbol v}_0=0.
\end{equation}
Since ${\boldsymbol v}_0$ is constant, the characteristic streamlines of the hyperbolic equation \eqref{lin-density} are easily determined. With this information, we find that $\delta\rho$ is given in terms of $\delta\bv} %{{\boldsymbol \epsilon}$ by
\begin{equation}
\label{deltarho}
\delta\rho(t, {\boldsymbol x})=-\rho_0 \int_0^t {\rm div}\delta\bv} %{{\boldsymbol \epsilon}(\tau,{\boldsymbol x}+(\tau-t){\boldsymbol v}_0)\,\text{d}\tau.
\end{equation}
Next, we formally linearize the momentum balance \eqref{m-balance-closed} to yield
\begin{equation*}
\rho_0 (\partial_t +{\boldsymbol v}_0\cdot\nabla) \delta\bv} %{{\boldsymbol \epsilon}=-(P^\prime(\rho_0)+\theta)\nabla(\delta\rho)- 2K_2 (\delta\bv} %{{\boldsymbol \epsilon}\cdot{\boldsymbol v}_0){\boldsymbol v}_0+\mu(\rho_0) \Delta \delta\bv} %{{\boldsymbol \epsilon}.
\end{equation*}
Taking time-derivative of both sides, using (\ref{deltarho}) to express $\partial_t \delta\rho$, and neglecting in that expression the term containing the third derivatives of $\delta\bv} %{{\boldsymbol \epsilon}$, we find that
\begin{equation}
\label{linmom2}
\rho_0 \partial^2_{tt}\delta\bv} %{{\boldsymbol \epsilon} +\rho_0{\boldsymbol v}_0\cdot\nabla\partial_t \delta\bv} %{{\boldsymbol \epsilon}=-L\nabla{\rm div}\delta\bv} %{{\boldsymbol \epsilon}-2 K_2(\partial_t \delta\bv} %{{\boldsymbol \epsilon}\cdot{\boldsymbol v}_0){\boldsymbol v}_0+\mu \Delta \partial_t \delta\bv} %{{\boldsymbol \epsilon},
\end{equation}
where we have introduced $L=\rho_0 (P^\prime(\rho_0)+\theta)$. Inserting
\begin{equation}
\delta\bv} %{{\boldsymbol \epsilon}={\boldsymbol A} e^{\sigma t-i{\boldsymbol k}\cdot{\boldsymbol x}}
\end{equation}
into \eqref{linmom2} yields
\begin{equation}
\label{linmom3}
\lambda {\boldsymbol A}=M{\boldsymbol A}
\end{equation}
where, on introducing a positively oriented Cartesian basis $\{{\boldsymbol \imath}_1,{\boldsymbol \imath}_2\}$ and writing $k_r={\boldsymbol k}\cdot{\boldsymbol \imath}_r$ and $v_r={\boldsymbol v}_0\cdot{\boldsymbol \imath}_r$, $\lambda$ and $M$ are given by
\begin{equation}
\lambda=\rho_0 \sigma^2 +\rho_0\sigma (-i{\boldsymbol k} )\cdot {\boldsymbol v}_0+\sigma \mu |{\boldsymbol k}|^2
\end{equation}
and
\begin{equation}
M=
\left(
\begin{array}{cc}
L k_1^2-2K_2\sigma v_1^2 & Lk_1k_2-2K_2\sigma v_1 v_2\\
Lk_1k_2- 2K_2\sigma v_1 v_2 & L k_2^2-2K_2\sigma v_2^2\\
\end{array}
\right).
\end{equation}
The characteristic equation for the matrix $M$ is
\begin{equation}
\lambda^2-({\rm tr}M)\lambda +{\rm det}M=0.
\end{equation}
Since only long wave lengths (small ${\boldsymbol k}$) are of interest at the meso-scale, we observe that, as ${\boldsymbol k}\to\bf0$, ${\rm tr} M=O(1)$ and ${\rm det}M=O(|k|^2)$. Thus, making the approximation
\begin{equation}
\label{quad-app}
\sqrt{({\rm tr}M)^2-4{\rm det}M}\approx {\rm tr}M- 2\frac{{\rm det}M}{{\rm tr}M},
\end{equation}
we find two solutions,
\begin{equation}
\lambda_1={\rm tr}M- \frac{{\rm det}M}{{\rm tr}M}
\qquad\text{and}\qquad
\lambda_2=\frac{{\rm det}M}{{\rm tr}M}.
\end{equation}
These solutions depend on both $\sigma$ and ${\boldsymbol k}$. The possible dispersion relations $\sigma=\sigma({\boldsymbol k})$ should satisfy
\begin{equation}
\label{quad1}
\rho_0 \sigma^2 +\rho_0 \sigma (-i{\boldsymbol k} )\cdot {\boldsymbol v}_0+\sigma \mu |{\boldsymbol k}|^2=\lambda (\sigma, {\boldsymbol k}),
\end{equation}
with right-hand side being either $\lambda_1$ or $\lambda_2$. Although it is possible to solve \eqref{quad1} in closed form without approximation, the result of doing so is difficult to interpret. For this reason, we approximate $\lambda_1$ and $\lambda_2$ using Taylor's formula and keeping terms up to order $O(|{\boldsymbol k}|^2)$. This yields
\begin{equation}
\label{two-lam}
\lambda_1\approx -2K_2 \sigma |{\boldsymbol v}_0|^2+L|{\boldsymbol k}|^2-\frac{L}{|{\boldsymbol v}_0|^2} E({\boldsymbol k})
\qquad\text{and}\qquad
\lambda_2\approx \frac{L}{|{\boldsymbol v}_0|^2} E({\boldsymbol k}),
\end{equation}
where $E$ is given by
\begin{equation}
E({\boldsymbol k})=k_1^2 v_2^2 +k_2^2 v_1^2-2k_1k_2 v_1v_2.
\end{equation}
Substituting $\lambda_1$ from (\ref{two-lam}) into (\ref{quad1}) and approximating the square root in the quadratic formula as before we obtain relations
\begin{equation}
\sigma_{1}({\boldsymbol k})=i{\boldsymbol k}\cdot{\boldsymbol v}_0-\frac{2K_2|{\boldsymbol v}_0|^2}{\rho_0}-\frac{\mu}{\rho_0}|{\boldsymbol k}|^2-\frac{L}{2K_2 |{\boldsymbol v}_0|^2}
\left(|{\boldsymbol k}|^2-\frac{E({\boldsymbol k})}{|{\boldsymbol v}_0|^2}\right)
\end{equation}
and
\begin{equation}
\sigma_{2}({\boldsymbol k})=\frac{L}{2K_2|{\boldsymbol v}_0|^2}
\left(|{\boldsymbol k}|^2-\frac{E({\boldsymbol k})}{|{\boldsymbol v}_0|^2}\right).
\end{equation}
To analyze the stability of the modes corresponding to $\sigma_1$ and $\sigma_2$, we can assume (without loss of generality) that ${\boldsymbol v}_0=v_1{\boldsymbol \imath}_1$. Then
\begin{equation}
\label{F}
|{\boldsymbol k}|^2-\frac{F({\boldsymbol k})}{|{\boldsymbol v}_0|^2}=|{\boldsymbol k}|^2-k_2^2=k_1^2\geq 0.
\end{equation}
This shows that $\sigma_1$ corresponds to a stable mode and $\sigma_2$-mode is unstable. It is interesting to note that in the incompressible case with $\mu>0$ there is no unstable mode \cite{dunkel2013}. In contrast to incompressible models, solutions of compressible equations may therefore exhibit various vortical patterns while approaching the uniform polar (flocking) state.
Finally, the same calculations with $\lambda_2$ in place of $\lambda_1$ yield
\begin{equation}
\sigma_3({\boldsymbol k})=i{\boldsymbol k}\cdot{\boldsymbol v}_0-\frac{\mu|{\boldsymbol k}|^2}{\rho_0}
\qquad\text{and}\qquad
\sigma_4({\boldsymbol k})=0.
\end{equation}
The mode corresponding to $\sigma_3$ is stable and the mode corresponding to $\sigma_4$ is neutrally stable.
\subsection{Stability of the trivial solution}
Formal linearization of the equations enforcing mass and momentum balance about a state in which $\rho_0=0$ and ${\boldsymbol v}_0=\bf0$ yields the velocity perturbation equation
\begin{equation}
\label{linmom3bis}
\rho_0 \partial^2_{tt}\delta\bv} %{{\boldsymbol \epsilon} =-L\nabla{\rm div}\delta\bv} %{{\boldsymbol \epsilon}+\mu \Delta \partial_t \delta\bv} %{{\boldsymbol \epsilon}.
\end{equation}
Inserting $\delta\bv} %{{\boldsymbol \epsilon}={\boldsymbol A} e^{\sigma t-i{\boldsymbol k}\cdot {\boldsymbol x}}$ in \eqref{linmom3bis}, we find that
\begin{equation}
\label{linmom4}
(\rho_0\sigma^2 +\mu|{\boldsymbol k}|^2\sigma) {\boldsymbol A}=L({\boldsymbol k}\cdot{\boldsymbol A}){\boldsymbol A}=\tilde M {\boldsymbol A},
\end{equation}
where
\begin{equation}
\tilde M=-L{\boldsymbol k}\otimes {\boldsymbol k}.
\end{equation}
Since ${\rm det}\tilde M=0$ and ${\rm tr}\tilde M=L|{\boldsymbol k}|^2$, the eigenvalues of $\tilde M$ are
\begin{equation}
\tilde\lambda_1= L|{\boldsymbol k}|^2
\qquad\text{and}\qquad
\tilde\lambda_2=0.
\end{equation}
The corresponding values of $\sigma$ are
\begin{equation}
\left.
\begin{split}
\displaystyle
\sigma_1({\boldsymbol k})&=\frac{1}{2\rho_0}
\left(
-\mu|{\boldsymbol k}|^2+ \sqrt{ \mu^2 |{\boldsymbol k}|^4+4\rho_0 L|{\boldsymbol k}|^2}
\right),
\\[4pt]
\sigma_2({\boldsymbol k})&=\frac{1}{2\rho_0}
\left(
-\mu|{\boldsymbol k}|^2- \sqrt{ \mu^2 |{\boldsymbol k}|^4+4\rho_0 L|{\boldsymbol k}|^2}\right),
\\[4pt]
\sigma_3({\boldsymbol k})&=-\frac{\mu}{\rho_0}|{\boldsymbol k}|^2,
\\[4pt]
\sigma_4({\boldsymbol k})&=0.
\end{split}\,
\right\}
\end{equation}
For $L>0$, which should be considered generic, the modes corresponding to $\sigma_1$ and $\sigma_2$ are unstable, the mode corresponding to $\sigma_3$ is stable, and the mode corresponding to $\sigma_4$ is neutrally stable.
\section{Conclusions}
\label{sect:conclusions}
In this work, the Irving--Kirkwood--Noll procedure is applied to derive the effective meso-scale continuum equations of an active suspension of point particles. The derivations make direct use of the particle equations of motion. A kinetic formulation, often associated with restrictive assumptions of small concentrations and weak interactions, is consequently avoided. The spatially averaged equations enforcing mass and momentum balance are therefore valid for highly concentrated and strongly interacting particle systems. Importantly, we use a realistic model of the self-propulsion force in which the force acting on a particle depends only on the velocity of that particle. This means that in contrast to other agent-based approaches to active suspensions, such as the classical Viscek model, our model does not include a dedicated velocity-aligning mechanism.
Our approach also involves a novel closure strategy in which the average mass density and velocity are measured not at every point of space-time but rather only at a discrete subset of points. Compared to the standard case of continuum fields defined at each point, upscaling in the discretized setting is associated with an additional loss of information. This makes it necessary to impose certain statistical assumptions about fluctuation tensors of velocities and relative positions. The simplest such assumption is that these tensors are nearly spherical
and thus can be characterized by scalar parameters reminiscent of the physical temperature. The resulting constitutive theory involves three parameters: the physical temperature and two fluctuation strength parameters. In addition, constitutive equations depend on the mass density and velocity gradient.
In contrast to previous derivations of continuum models with an ensemble averaging approach \cite{Chuang2007}, the present model includes conservative and dissipative stress tensors, both of which are given by constitutive equations, and the effects due to fluctuations are taken into account.
Our coarse-scale evolution equations are similar to the well-known equations of Toner and Tu \cite{toner-tu95}, the main difference being that our equations involve a constitutive relation for the coarse-scale fluctuation force. According to this relation, the strength and the variance of the fluctuation force depends on time and space through the mass density, temperature, and fluctuation strength parameters.
Previously, kinetic theory has been used to derive Toner--Tu type equations from Vicsek's \cite{vicsek95} model. Since the assumptions underlying classical kinetic theory do not apply to concentrated suspensions, whether the Toner--Tu equations can be reliably applied to such systems was previously unclear. However, the results of the present work justify the use of these equations for modeling dense active suspensions, at least in the case of nearly spherical particles.
\begin{acknowledgments}
E.F.\ gratefully acknowledges support from the Okinawa Institute of Science and Technology Graduate University with subsidy funding from the Cabinet Office, Government of Japan.
\end{acknowledgments}
|
\section{Introduction}
Graphs considered in this paper are undirected, simple and finite (unless otherwise noted). Given a graph $G=(V,E)$ with vertex set $V(G)=V$ and edge set $E(G)=E$, for convenience, we often identify a triangle in $G$ with its edge set. A subset of $E$ is called a {\em triangle cover} if {it intersects} each triangle of $G$. Let $\tau_t(G)$ denote the minimum cardinality of a triangle cover of $G$, referred to as the {\em triangle covering number} of $G$. A set of pairwise edge-disjoint triangles {in $G$} is called a {\em triangle packing} of $G$. Let $\nu_t(G)$ denote the maximum cardinality of a triangle packing of $G$, referred to as the {\em triangle packing number} of $G$. It is clear that $1\leq\tau_t(G)/\nu_t(G)\le3$ holds for every graph $G$. Our research is motivated by the following conjecture {raised by Tuza \cite{tuza1981} in 1981}.
\begin{conjecture}[Tuza's Conjecture \cite{tuza1981}] \label{coj:tuza}
$\tau_t(G)/\nu_t(G)\le2$ holds for every graph $G$.
\end{conjecture}
To the best of our knowledge, the conjecture is still {unsolved in general}. If {it is true}, then the upper bound 2 is sharp as shown by $K_4$ and $K_5$ -- the complete graphs of orders $4$ and $5$.
\paragraph{Related work.} The only known universal upper bound smaller than 3 was given by Haxell \cite{haxell19}, who shown that $\tau_t(G)/\nu_t(G)\le 66/{23} =2.8695...$ holds for all graphs $G$. {Haxell's proof \cite{haxell19} implies a polynomial-time algorithm for finding a triangle cover of cardinality at most 66/23 times that of a maximal triangle packing}. Other partial results on {Tuza's} conjecture concern with special classes of graphs.
Tuza \cite{tuza1990} proved {his} conjecture holds for planar graphs, $K_{5}$-free chordal graphs and graphs with $n$ vertices and at least $7n^{2}/16$ edges. {The proof for planar graphs \cite{tuza1990} gives an elegant polynomial-time algorithm for finding a triangle cover in planar graphs with cardinality at most twice that of a maximal triangle packing. The validity of Tuza's conjecture on the} class of planar graphs was later {generalized} by Krivelevich \cite{Krivelevich1995} to the class of graphs without $K_{3,3}$-subdivision. Haxell and Kohayakawa \cite{HK1998} showed that $\tau_t(G)/\nu_t(G)\le2-\epsilon$ for tripartite graphs $G$, where $\epsilon > 0.044$. Haxell, Kostochka and Thomasse \cite{HKT2012} proved that every $K_{4}$-free planar graph $G$ satisfies $\tau_t(G)/\nu_t(G)\le1.5$.
Regarding the tightness of the {conjectured upper bound 2, Tuza \cite{tuza1990} noticed that infinitely many graphs $G$ attain the conjectured upper bound $\tau_t(G)/\nu_t(G)=2$.} Cui, Haxell and Ma \cite{CHM2009} characterized planar graphs $G$ satisfying $\tau_t(G)/\nu_t(G)=2${; these graphs are edge-disjoint unions of $K_4$'s} plus possibly some vertices and edges that are not in triangles. Baron and Kahn \cite{BK2014} proved that Tuza's conjecture is asymptotically tight for dense graphs.
Fractional and weighted variants of Conjecture \ref{coj:tuza} were studied in literature. Krivelevich \cite{Krivelevich1995} {proved} two fractional versions of the conjecture: $\tau_t(G)\leq2\nu^{\ast}_t(G)$ and $\tau^{\ast}_t(G)\leq2\nu_t(G)${, where $\tau^{\ast}_t(G)$ and $\nu^{\ast}_t(G)$ are the values of an optimal fractional triangle cover and an optimal fractional triangle packing of $G$, respectively. The result was generalized by Chapuy et al. \cite{CDDMS2015} to the weighted version}, which amounts to packing and covering triangles in multigraphs {$G_w$ (obtained from $G$ by adding multiple edges). The authors \cite{CDDMS2015} showed that $\tau(G_w)\leq2\nu^{\ast}(G_w)-\sqrt{\nu^{\ast}(G_w)/6}+1$ and $\tau^{\ast}(G_w)\leq2\nu(G_w)$; the arguments imply an LP-based 2-approximation algorithm for finding a minimum weighted triangle cover.}
\paragraph{Our contributions.} Along a different line, we establish new sufficient conditions for validity of Tuza's conjecture by comparing the {triangle} packing number, the number of triangles and the number of edges. {Given a graph $G$, we use $\mathscr T_G=\{E(T):T$ is a triangle in $G\}$ to denote the set consisting of the (edge sets of) triangles in $G$.} Without loss of generality, we focus on the graphs in which every edge is contained in some triangle. These graphs are called {\em irreducible}.
\begin{theorem}\label{th:condition}
Let $G=(V,E)$ be an irreducible graph. Then a triangle cover of $G$ with cardinality at most $2\nu_t(G)$ can be found in polynomial time, which implies $\tau_t(G)\le2\nu_t(G)$, if one of the following conditions is satisfied:
(i) $\nu_t(G)/|\mathscr T_G|\ge\frac13$, (ii) $\nu_t(G)/|E|\ge\frac14$, (iii) $|E|/|\mathscr T_G|\ge2$.
\end{theorem}
The primary idea behind the theorem is simple: any one of conditions (i) -- (iii) allows us to remove at most $\nu_t(G)$ edges from $G$ to make the resulting graph $G'$ satisfy $\tau_t(G')=\nu_t(G')$; the removed edges and the edges in a minimum triangle cover of $G'$ form a triangle cover of $G$ with size at most $\nu_t(G)+\nu_t(G')\le2\nu_t(G)$. The idea is realized by establishing new results on linear 3-uniform hypergraphs (see Section \ref{sec:hypergraph}); the most important one states that such a hypergraphs could be made acyclic by removing {a number of vertices that is no more than a third of the number of its edges}. A key observation here is that hypergraph $(E,\mathscr T_G)$ is linear and 3-uniform.
To show the qualities of conditions (i) -- (iii) in Theorem \ref{th:condition}, we obtain the following result which complements to the constants $\frac13$, $\frac14$ and 2 in these conditions with $\frac14$, $\frac15$ and $\frac32$, respectively.
\begin{theorem}\label{th:howgood}
Tuza's conjecture holds for every graph if there exists some real $\delta>0$ such that Tuza's conjecture holds for every irreducible graph $G$ satisfying one of the following properties: (i') $\nu_t(G)/|\mathscr T_G|\ge \frac14-\delta$, (ii') $\nu_t(G)/|E|\ge\frac15-\delta$, (iii') $|E|/|\mathscr T_G|\ge\frac32-\delta $.
\end{theorem}
We also investigate Tuza's conjecture on classical Erd\H{o}s-R\'{e}nyi random graph $\mathcal G(n,p)$, and prove that $\text{\bf Pr}[\tau_t(G)/\nu_t(G)\le2]=1-o(1)$ provided $G\in\mathcal G(n,p)$ and $p> \sqrt{3}/2$.
\redcomment{It is worthwhile pointing out that strengthening Theorem \ref{th:condition}, our arguments actually establish stronger results for linear 3-uniform hypergraphs (see Theorem \ref{th:condition'})}.
\medskip
{The rest of paper is organized as follows. Section \ref{sec:hypergraph} proves theoretical and algorithmic results on linear 3-uniform hypergraphs concerning feedback sets, which are main technical tools for establishing new sufficient conditions for Tuza's conjecture in Section \ref{sec:tuza}. Section \ref{sec:conclude} concludes the paper with extensions and future research directions.}
\section{Hypergraphs}\label{sec:hypergraph}
{This section develops hypergraph tools for studying Tuza's conjecture. The theoretical and algorithmic results are of interest in their own right.}
Let $\mathcal H=(\mathcal V,\mathcal E)$ be a hypergraph with vertex set $\mathcal V$ and edge set $\mathcal E$. For convenience, we use $|\!|\mathcal H|\!|$ to denote the number $|\mathcal E|$ of edges in $\mathcal H$. If hypergraph $\mathcal H'=(\mathcal V',\mathcal E')$ satisfies $\mathcal V' \subseteq \mathcal V$ and $ \mathcal E' \subseteq \mathcal E$, we call $\mathcal H'$ a {\em sub-hypergraph} of $\mathcal H$, and write $\mathcal H'\subseteq \mathcal H$. For each $v\in \mathcal V$, the {\em degree} $d_{\mathcal H}(v)$ is the number of edges in $\mathcal E$ that {contain} $v$. We say $v$ is an {\em isolated vertex} of $\mathcal H$ if $d_{\mathcal H}(v)=0$. Let $k\in\mathbb N$ be a positive integer, hypergraph $\mathcal H$ is called {\em $k$-regular} if $d_{\mathcal H}(u)=k$ for each $u\in \mathcal V$, and {\em $k$-uniform} if $|e|=k$ for each $e\in \mathcal E$. Hypergraph $\mathcal H$ is {\em linear} if $|e\cap f|\le1$ for any pair of distinct edges $e,f\in \mathcal E$.
A vertex-edge alternating sequence $ v_{1}e_{1}v_{2}...v_{k}e_{k}v_{k+1}$ of $\mathcal H$ is called a {\em path} (of {\em length} $k$) between $v_{1}$ and $v_{k+1}$ if $v_{1}, v_{2},..., v_{k+1}\in\mathcal V$ are distinct, $ e_{1}, e_{2},..., e_{k}\in\mathcal E$ are distinct, and $\{v_{i},v_{i+1}\}\subseteq e_{i}$ for each $i\in [k]=\{1,\ldots,k\}$. We consider each vertex of $\mathcal H$ as a path of length 0. Hypergraph $\mathcal H$ is said to be {\em connected} if there is a path between any pair of distinct vertices in $\mathcal H$. A maximal connected sub-hypergraph of $\mathcal H$ is called a {\em component} of $\mathcal H$. Obviously, $\mathcal H$ is connected if and only if it has only one component.
A vertex-edge alternating sequence $\mathcal C= v_{1}e_{1}v_{2}e_{2}...v_{k}e_{k}v_{1}$, where $k\ge2$, is called a {\em cycle} (of length $k$) if $v_{1}, v_{2},..., v_{k}\in\mathcal V$ are distinct, $ e_{1}, e_{2},..., e_{k}\in \mathcal E$ are distinct, and $\{v_{i},v_{i+1}\}\subseteq e_{i}$ for each $i\in [k]$, where $v_{k+1}=v_{1}$. We consider the cycle $\mathcal C$ as a sub-hypergraph of $\mathcal H$ with vertex set $\cup_{i\in[k]}e_i$ and edge set {$\{e_{i}: i\in [k]\}$}. For any $\mathcal S\subset\mathcal V$ (resp. $S\subset\mathcal E$), we write $\mathcal H\setminus \mathcal S$ for the sub-hypergraph of $\mathcal H$ obtained from $\mathcal H$ by deleting all vertices in $\mathcal S$ and all edges incident with some vertices in $\mathcal S$ (resp. deleting all edges in $\mathcal E$ and keeping vertices). If $\mathcal S$ is a singleton set $\{s\}$, we write $\mathcal H\setminus s$ instead of $\mathcal H\setminus \{s\}$. For any $\mathcal S\subseteq2^{\mathcal V}$, the hypergraph $(\mathcal V,\mathcal E\cup \mathcal S)$ is often written as $\mathcal H\uplus\mathcal E$, and as $\mathcal H\oplus\mathcal S$ if $\mathcal S\cap\mathcal E=\emptyset$.
A vertex (resp. edge) subset of $\mathcal H$ is called a {\em feedback vertex set} or FVS (resp. {\em feedback edge set} or FES) of $\mathcal H$ if it intersects the vertex (resp. edge) set {of} every cycle of $\mathcal H$. A vertex subset of $\mathcal H$ is called a {\em transversal} of $\mathcal H$ if it intersects every edge of $\mathcal H$. Let $\tau^{{}_{\mathcal V}}_c(\mathcal H)$, $\tau_c^{{}_{\mathcal E}}(\mathcal H)$ and $\tau(\mathcal H)$ denote, respectively, the minimum cardinalities of a FVS, a FES, and a transversal of $\mathcal H$. A {\em matching} of $\mathcal H$ is an nonempty set of pairwise disjoint edges of $\mathcal H$. Let $\nu(\mathcal H)$
denote the maximum cardinality of a matching of $\mathcal H$. It is easy to see that $\tau^{{}_{\mathcal V}}_c(\mathcal H)\leq \tau_c^{{}_{\mathcal E}}(\mathcal H)$, $\tau^{{}_{\mathcal V}}_c(\mathcal H)\leq \tau(\mathcal H)$ and $ \nu(\mathcal H)\le\tau(\mathcal H)$. {Our discussion will frequently use the trivial observation that if no cycle of $\mathcal H$ contains any element of some subset $\mathcal S$ of $\mathcal V\cup\mathcal E$, then $\mathcal H$ and $\mathcal H\setminus S$ have the same set of FVS's, and $\tau_c^{{}_{\mathcal V}}(\mathcal H)= \tau_c^{{}_{\mathcal V}}(\mathcal H\setminus \mathcal S)$. The following theorem is one of main contributions of this paper.}
\begin{theorem}\label{1/3}
Let $\mathcal H$ be a linear $3$-uniform hypergraph. Then $\tau_c^{{}_{\mathcal V}}(\mathcal H)\le |\!|\mathcal H|\!|/3$.
\end{theorem}
\begin{proof} Suppose that the theorem failed. {We} take a counterexample $\mathcal H=(\mathcal V,\mathcal E)$ with $\tau_c^{{}_{\mathcal V}}(\mathcal H)> |\mathcal E|/3$ such that $|\!|\mathcal H|\!|=|\mathcal E|$ is as small as possible.
Obviously $|\mathcal E|\geq 3$. Without loss of generality, we can assume that $\mathcal H$ has no isolated vertices. Since $\mathcal H$ is linear, any cycle in $\mathcal H$ is of length at least 3.
If there exists $e\in \mathcal E$ which does not belong to any cycle of $\mathcal H$, then $\tau_c^{{}_{\mathcal V}}(\mathcal H)= \tau_c^{{}_{\mathcal V}}(\mathcal H\setminus e)$. The minimality of $\mathcal H=(\mathcal V,\mathcal E)$ implies $\tau_c^{{}_{\mathcal V}}(\mathcal H\setminus e)\leq (|\mathcal E|-1)/3$, giving $\tau_c^{{}_{\mathcal V}}(\mathcal H)< |\mathcal E|/3$, a contradiction. So we have
\begin{itemize}
\item[(1)] Every edge in $\mathcal E$ is contained in some cycle of $\mathcal H$.
\end{itemize}
If there exists $v\in\mathcal V$ with $d_{\mathcal H}(v)\geq3$, then $\tau_c^{{}_{\mathcal V}}(\mathcal H\setminus v)\leq (|\mathcal E|-d_{\mathcal H}(v))/3\leq (|\mathcal E|-3)/3$, where the first inequality is due to the minimality of $\mathcal H$. Given a minimum FVS $\mathcal S$ of $\mathcal H\setminus v$, it is clear that $\mathcal S\cup \{v\}$ is a FVS of $\mathcal H$ with size $ |\mathcal S|+1=\tau_c^{{}_{\mathcal V}}(\mathcal H\setminus v)+1\leq |\mathcal E|/3$, a contradiction to $\tau_c^{{}_{\mathcal V}}(\mathcal H)> |\mathcal E|/3$. So we have
\begin{itemize}
\item[(2)] $ d_{\mathcal H}(v)\leq 2$ for all $v\in \mathcal V$.
\end{itemize}
Suppose that there exists $v\in\mathcal V$ with $d_{\mathcal H}(v)= 1$. Let $e_1\in\mathcal E$ be the unique edge that contains $v$. Recall from (1) that $e_1$ is contained in a cycle $\mathcal C= v_{1}e_{1}v_{2}e_{2}v_{3}\cdots e_{k}v_{1}$, where $k\ge3$. By (2), we have $d_{\mathcal H}(v_i)=2$ for all $i\in[k]$. In particular $d_{\mathcal H}(v_1)=d_{\mathcal H}(v_2)=2>d_{\mathcal H}(v)$ implies $v\not\in\{v_1,v_2\}$, and in turn $ v_1,v_2,v \in e_1$ enforces $e_{1}= \{v_{1},v,v_{2}\}$. Let $\mathcal S$ be a minimum FVS of $\mathcal H'=\mathcal H\setminus\{e_1,e_2,e_3\}$. It follows from (2) {that
\[\mathcal H\setminus v_3\subseteq \mathcal H\setminus\{e_2,e_3\}= \mathcal H'\oplus e_1,\]
and} in $\mathcal H'\oplus e_1$, {edge} $e_1$ intersects at most one other edge, and therefore is not contained by any cycle.
Thus {$\mathcal S$ is a FVS of $ \mathcal H'\oplus e_1$, and hence a FVS of $\mathcal H\setminus v_3$, implying} that $\{v_3\}\cup\mathcal S$
is a FVS of $\mathcal H$. We deduce that $|\mathcal E|/3<\tau_c^{{}_{\mathcal V}}(\mathcal H)\le|\{v_3\}\cup \mathcal S|\le 1+|\mathcal S|$. Therefore $\tau_c^{{}_{\mathcal V}}(\mathcal H')=|\mathcal S|>(|\mathcal E|-3)/3=|\!|\mathcal H'|\!|/3$ shows a contradiction to the minimality of $\mathcal H$. Hence the vertices of $\mathcal H$ all have degree at least 2, which together with (2) gives
\begin{itemize}
\item[(3)] $\mathcal H$ is $2$-regular.
\end{itemize}
Let $\mathcal C=(\mathcal V_c,\mathcal E_c)=v_1e_1v_2e_2\ldots v_ke_kv_1$ be a shortest cycle in $\mathcal H$, where $k\ge3$. For each $i\in[k]$, suppose that $e_i=\{v_i,u_i,v_{i+1}\}$, where $v_{k+1}=v_1$.
Because $\mathcal C$ is a shortest cycle, for each pair of distinct indices $i,j\in [k]$, we have $e_{i}\cap e_{j}=\emptyset$ if and only if $e_i$ and $e_j$ are not adjacent in $\mathcal C$, i.e., $|i-j|\not\in\{1,k-1\}$. This fact along with the linearity of $\mathcal H$ says that $v_1,v_2,\ldots,v_k,u_1,u_2,\ldots,u_k$ are distinct. By (3), each $u_i$ is contained in a unique edge $f_i\in\mathcal E\setminus\mathcal E_c$, $i\in[k]$. We distinguish among three cases depending on the values of $k\pmod 3$. In each case, we construct a proper sub-hypergraph $\mathcal H'$ of $\mathcal H$ with $|\!|\mathcal H'|\!|<|\!|\mathcal H|\!|$ and $\tau_c^{{}_{\mathcal V}}(\mathcal H')>|\!|\mathcal H'|\!|/3$ which shows a contradiction to the minimality of $\mathcal H$.
\paragraph{\sc Case 1. $k\equiv0\pmod3$:} Let $\mathcal S$ be a minimum FVS of $\mathcal H'=\mathcal H\setminus\mathcal E_c$. Setting {$\mathcal V_*=\{v_i:i\equiv0\pmod3, i\in [k]\}$ and $\mathcal E_*=\{e_i:i\equiv1\pmod3, i\in [k]\}$}, it follows from (3) that
\[\mathcal H\setminus \mathcal V_*\subseteq (\mathcal H\setminus\mathcal E_c)\oplus \mathcal E_*= \mathcal H'\oplus \mathcal E_*,\]
and in $\mathcal H'\oplus \mathcal E_*$, each edge in $\mathcal E_*$ intersects exactly one other edge, and therefore is not contained by any cycle.
Thus $(\mathcal H'\oplus \mathcal E_*)\setminus \mathcal S$ is also acyclic, so is $(\mathcal H\setminus\mathcal V_*)\setminus\mathcal S$, saying that $\mathcal V_*\cup\mathcal S$
is a FVS of $\mathcal H$. We deduce that $|\mathcal E|/3<\tau_c^{{}_{\mathcal V}}(\mathcal H)\le|\mathcal V_*\cup \mathcal S|\le k/3+|\mathcal S|$. Therefore $\tau_c^{{}_{\mathcal V}}(\mathcal H')=|\mathcal S|>(|\mathcal E|-k)/3=|\!|\mathcal H'|\!|/3$ shows a contradiction.
\paragraph{\sc Case 2. $k\equiv1\pmod3$:} Consider the case where $f_1\ne f_3$ or $f_2\ne f_4$. Relabeling the vertices and edges if necessary, we may assume
without loss of generality that $f_1\ne f_3$. Let $\mathcal S$ be a minimum FVS of $\mathcal H'=\mathcal H\setminus(\mathcal E_c\cup \{f_1,f_3\})$. Set $\mathcal V_*=\emptyset$, $\mathcal E_*=\emptyset$ if $k=4$ and {$\mathcal V_*=\{v_i:i\equiv0\pmod3, i\in [k]-[3]\}$, $\mathcal E_*=\{e_i:i\equiv1\pmod3, i\in [k]-[6]\}$} otherwise. In any case we have $|\mathcal V_*|=(k-4)/3$ and
\[\mathcal H\setminus(\{u_1,u_3\}\cup\mathcal V_*)\subseteq (\mathcal H\setminus(\mathcal E_c\cup \{f_1,f_3\}))\oplus( \{e_2,e_4\}\cup\mathcal E_*)= \mathcal H'\oplus (\{e_2,e_4\}\cup\mathcal E_*).\]
Note from (3) that in $\mathcal H'\oplus (\{e_2,e_4\}\cup\mathcal E_*)$, each edge in $ \{e_2,e_4\}\cup\mathcal E_*$ can intersect at most one other edge, and therefore is not contained by any cycle.
Thus $( \mathcal H'\oplus (\{e_2,e_4\}\cup\mathcal E_*))\setminus \mathcal S$ is also acyclic, so is $(\mathcal H\setminus (\{u_1,u_3\}\cup\mathcal V_*))\setminus\mathcal S$. Thus $\{u_1,u_3\}\cup\mathcal V_*\cup\mathcal S$ is a FVS of $\mathcal H$, and $|\mathcal E|/3<\tau_c^{{}_{\mathcal V}}(\mathcal H)\le|\{u_1,u_3 \}\cup\mathcal V_*\cup\mathcal S|\le2+|\mathcal V_*|+|\mathcal S|=(k+2)/3+|\mathcal S|$. This gives $\tau_c^{{}_{\mathcal V}}(\mathcal H')=|\mathcal S|>(|\mathcal E|-k-2)/3=|\mathcal H'|/3$, a contradiction.
Consider the case where $f_1= f_3$ and $f_2=f_4$. As $u_1,u_2,u_3,u_4$ are distinct and $|f_1|=|f_2|=3$, we have $f_1\ne f_2$. Observe that $u_1e_1v_2e_2v_3e_3u_3f_3u_1$ is a cycle in $\mathcal H$ of length 4. The minimality of $k$ enforces $k=4$. Therefore $\mathcal E_c\cup\{f_1,f_2\}$ consist of 6 distinct edges. Let $\mathcal S$ be a minimum FVS of $\mathcal H'=\mathcal H\setminus(\mathcal E_c\cup\{f_1,f_2\})$. It follows from (3) that \[\mathcal H\setminus \{u_2,u_4\}\subseteq (\mathcal H\setminus (\mathcal E_c\cup\{f_1,f_2\}))\oplus\{e_1,e_3,f_1\}= \mathcal H'\oplus\{e_1,e_3,f_1\}.\]
In $\mathcal H'\oplus\{e_1,e_3,f_1\}$, both $e_{1}$ and $e_{3}$ intersect only one other edge, which is $f_1$, and any cycle through {$f_1$} must contain $e_1$ or $e_3$. It follows that none of $e_1,e_3,f_1$ is contained by a cycle of $\mathcal H'\oplus\{e_1,e_3,f_1\}$. Thus $(\mathcal H'\oplus \{e_1,e_3,f_1\})\setminus \mathcal S$ is acyclic, so is $(\mathcal H \setminus \{u_2,u_4\})\setminus \mathcal S$, saying that $ \{u_2,u_4\}\cup\mathcal S$ is a FVS of $\mathcal H$. Hence $|\mathcal E|/3<\tau_c^{{}_{\mathcal V}}(\mathcal H)\le|\{u_2,u_4 \}\cup\mathcal S|\leq2+ |\mathcal S|$. In turn $\tau_c^{{}_{\mathcal V}}(\mathcal H')=|\mathcal S|>(|\mathcal E|-6)/3=|\!|\mathcal H'|\!|/3$ shows a contradiction.
\paragraph{\sc Case 3. $k\equiv2\pmod3$:} Let $\mathcal S$ be a minimum FVS of $\mathcal H'=\mathcal H\setminus (\mathcal E_c\cup\{f_1\})$. Setting {$\mathcal V_*=\{v_i:i\equiv1\pmod3, i\in [k]-[3]\}$ and $\mathcal E_*=\{e_i:i\equiv2\pmod3, i\in [k]\}$}, we have $|\mathcal V_*|=(k-2)/3$ and
\[\mathcal H\setminus(\{u_1\}\cup\mathcal V_*)\subseteq (\mathcal H\setminus(\mathcal E_c\cup \{f_1\}))\oplus \mathcal E_*= \mathcal H'\oplus\mathcal E_*\]
In $\mathcal H'\oplus\mathcal E_*$, each edge in $\mathcal E_*$ intersects at most one other edge, and therefore is not contained by any cycle. Thus $(\mathcal H'\oplus\mathcal E_*)\setminus \mathcal S$ is acyclic, so is $(\mathcal H\setminus ( \{u_1\}\cup\mathcal V_*))\setminus\mathcal S$. Hence $\{u_1\}\cup\mathcal V_*\cup\mathcal S$ is a FVS of $\mathcal H$, yielding $|\mathcal E|/3<\tau_c^{{}_{\mathcal V}}(\mathcal H)\le| \{u_1\}\cup\mathcal V_*\cup\mathcal S|\le1+(k-2)/3+|\mathcal S|$ and a contradiction $\tau_c^{{}_{\mathcal V}}(\mathcal H')=|\mathcal S|>(|\mathcal E|-k-1)/3=|\!|\mathcal H'|\!|/3$.
\medskip The combination of the above three cases complete the proof.
\end{proof}
{We remark that the upper bound $|\!|\mathcal H|\!|/3$ in Theorem \ref{1/3} is best possible. See Figure~\ref{sharp} for illustrations of five 3-uniform linear hypergraphs attaining the upper bound. It is easy to prove that the maximum degree of every extremal hypergraph (those $\mathcal H$ with $\tau_c^{{}_{\mathcal V}}(\mathcal H)= |\!|\mathcal H|\!|/3$) is at most three. It would be interesting to characterize all extremal hypergraphs for Theorem \ref{1/3}.}
\begin{figure}[sharp]
\begin{center}
\includegraphics[scale=0.65]{sharp.eps}
\caption{\label{sharp}{Extremal linear 3-uniform hypergraphs $\mathcal H$ with $\tau_c^{{}_{\mathcal V}}(\mathcal H)= |\!|\mathcal H|\!|/3$.}}
\end{center}
\end{figure}
\medskip
The proof of Theorem \ref{1/3} actually gives a recursive combinatorial algorithm for finding in polynomial time a FVS of size at most $|\!|\mathcal H|\!|/3$ on a linear 3-uniform hypergraph $\mathcal H$.
\begin{algorithm}
\KwIn{Linear 3-uniform hypergraph $\mathcal H=(\mathcal V,\mathcal E)$.}
\KwOut{$\textsc{Alg1}(\mathcal H)$, which is {a FVS of $\mathcal H$ with cardinality at most $|\!|\mathcal H|\!|/3$}.}
\begin{mylabel}
\vspace{1.5mm}\item[1.] \textbf{If} $|\mathcal E|\le2$ \textbf{Then} $\textsc{Alg1}(\mathcal H)\leftarrow\emptyset$
\vspace{0.mm}\item[2.] \quad\textbf{Else} \textbf{If} {$\exists$ $s\in \mathcal V\cup\mathcal E$ such that $s$ is not contained in any cycle of $\mathcal H$}
\vspace{0.mm}\item[3.] \quad\quad\quad\hspace{3mm} \textbf{Then} $\textsc{Alg1}(\mathcal H)\leftarrow \textsc{Alg1}(\mathcal H\setminus s)$
\vspace{0.5mm}\item[4.] \quad\quad\quad\hspace{.7mm} \textbf{If} $\exists$ $s\in \mathcal V$ {such that} $d_{\mathcal H}(s)\ge3$
\vspace{0.mm}\item[5.] \quad\quad\quad\hspace{3mm} \textbf{Then} $\textsc{Alg1}(\mathcal H)\leftarrow\{s\}\cup\textsc{Alg1}(\mathcal H\setminus s)$
\vspace{0.5mm}\item[6.] \quad\quad\quad\hspace{.7mm} \textbf{If} $\exists$ $v\in\mathcal V$ such that $d_{\mathcal H}(v)=1$
\vspace{0.mm}\item[7.] \quad\quad\quad\hspace{3mm} \textbf{Then} Let $v_1e_1v_2e_2v_3\cdots e_kv_1$ be a cycle of $\mathcal H$ such that $e_1=\{v_1,v_2,v\}$
\vspace{0.mm}\item[8.] \quad\quad\quad\quad\quad\hspace{5mm} {$\textsc{Alg1}(\mathcal H)\leftarrow\{v_3\}\cup\textsc{Alg1}(\mathcal H\setminus \{e_1,e_2,e_3\})$}
\vspace{0.mm}\item[9.] \quad\quad\quad\hspace{0.5mm} Let $(\mathcal V_c,\mathcal E_c)=v_1e_1v_2e_2\ldots v_ke_kv_1$ be a shortest cycle in $\mathcal H$
\vspace{0.mm}\item[10.] \quad\quad\quad\hspace{0.5mm} For each $i\in[k]$, let $u_i\in\mathcal V_c$, $f_i\in\mathcal E\setminus\mathcal E_c$ be such that $\{u_i,v_i,v_{i+1}\}=e_i$, $u_i\in f_i
\vspace{0.mm}\item[11.] \quad\quad\quad\hspace{.7mm} \textbf{If} $k\equiv0\pmod3$ \textbf{Then} {$\textsc{Alg1}(\mathcal H)\leftarrow\{v_i: i\equiv0\pmod3, i\in [k]\}\cup\textsc{Alg1}(\mathcal H\setminus \mathcal E_c)$}
\vspace{0.mm}\item[12.] \quad\quad\quad\hspace{.7mm} \textbf{If} $k\equiv1\pmod3$
\vspace{0.mm}\item[13.] \quad\quad\quad\hspace{3mm} \textbf{Then} \textbf{If} $f_1\ne f_3$ or $f_2\ne f_4$
\vspace{0.mm}\item[14.] \quad\quad\quad\quad\quad\hspace{10mm}\textbf{Then} Relabel vertices and edges if necessary to make $f_1\ne f_3$
\vspace{0.mm}\item[15.] \hspace{37mm} {$\mathcal V_*\leftarrow \{v_i:i\equiv0\pmod3, i\in [k]-[3]\}$}
\vspace{0.mm}\item[16.] \hspace{37mm} $\textsc{Alg1}(\mathcal H)\leftarrow\{u_1,u_3\}\cup \mathcal V_*\cup\textsc{Alg1}(\mathcal H\setminus (\mathcal E_c\cup\{f_1,f_3\}))$
\vspace{0.mm}\item[17.] \quad\quad\quad\quad\quad\hspace{9.5mm} \textbf{Else} $\textsc{Alg1}(\mathcal H)\leftarrow\{u_2,u_4\}\cup \textsc{Alg1}(\mathcal H\setminus (\mathcal E_c\cup\{f_1,f_2\}))$
\vspace{0.mm}\item[18.] \quad\quad\quad\hspace{.7mm} \textbf{If} $k\equiv2\pmod3$
\vspace{0.mm}\item[19.] \quad\quad\quad\hspace{3mm} {\textbf{Then} $\textsc{Alg1}(\mathcal H)\leftarrow\{u_1\}\cup \{v_i:i\equiv1\pmod3, i\in [k]-[3]\}\cup\textsc{Alg1}(\mathcal H\setminus (\mathcal E_c\cup\{f_1\}))$}
\vspace{0mm}\item[20.] Output $\textsc{Alg1}(\mathcal H)$
\vspace{-3mm}
\end{mylabel}
\caption{{Feedback Vertex Sets of Linear 3-Uniform Hypergraphs}} \label{alg1}
\end{algorithm}
Note that Algorithm \ref{alg1} never visits isolated vertices {(it only scans along the edges of the current hypergraph)}. The number of iterations performed by the algorithm is upper bounded by $|\mathcal E|$. {Since $\mathcal H$ is 3-uniform, the} condition in any step is checkable in $O(|\mathcal E|^2)$ time. Any cycle in {Step 7 or Step 9} can be found in $O(|\mathcal E|^2)$ time.\footnote{{The shortest path between any pair of vertices can be find in $O(|\mathcal E|)$ time using breadth first search. A shortest cycle can be find by checking all $O(|\mathcal E|)$ possibilities.}} Thus Algorithm \ref{alg1} runs in {$O(|\mathcal E|^3)$} time.
\begin{corollary}\label{cor:1/3}
Given any linear 3-uniform hypergraph $\mathcal H$, Algorithm \ref{alg1} finds in {$O(|\!|\mathcal H|\!|^3)$} time a FVS of $\mathcal H$ with size at most $|\!|\mathcal H|\!|/3$.\qed
\end{corollary}
\begin{lemma}\label{acycle}
If $\mathcal H=(\mathcal V,\mathcal E)$ is a connected linear $3$-uniform hypergraph without cycles, then $|\mathcal V|=2|\mathcal E|+1$.
\end{lemma}
\begin{proof}
We prove by induction on $|\mathcal E|$. The base case where $|\mathcal E|=0$ is trivial. Inductively, we assume that {$|\mathcal E|\ge1$} and the
lemma holds for all connected acyclic linear 3-uniform hypergraph of edges fewer than $\mathcal H$. Take arbitrary $e\in \mathcal E$. Since $\mathcal H$ is connected, acyclic and 3-uniform, $\mathcal H\setminus e$ contains exactly three components $\mathcal H_{i}=(\mathcal V_{i},\mathcal E_{i})$, $i=1,2,3$. Note that for each $i\in[3]$, hypergraph $\mathcal H_i$ with $|\mathcal E_i|<|\mathcal E|$ is connected, linear, 3-uniform and acyclic. By the induction hypothesis, we have $|\mathcal V_{i}|=2|\mathcal E_{i}|+1$ for $i=1,2,3$. It follows that $|\mathcal V|=\sum_{i=1}^3|\mathcal V_{i}|= 2\sum_{i=1}^3|\mathcal E_{i}| +3= 2|\mathcal E|+1$.
\end{proof}
Given any hypergraph $\mathcal H=(\mathcal V,\mathcal E)$, we can easily find a minimal (not necessarily minimum) FES in $O(|\mathcal E|^2)$ time: Go through the edges of the trivial FES $\mathcal E$ in any order, and remove the edge from the FES immediately if the edge is redundant. The redundancy test can be implemented using Depth First Search.
\begin{lemma}\label{cyclecover}
Let $\mathcal H=(\mathcal V,\mathcal E)$ be a linear $3$-uniform hypergraph with $p$ components. If $\mathcal F$ is a minimal FES of $\mathcal H$, then $|\mathcal F|\leq 2|\mathcal E|-|\mathcal V|+p$. In particular, $\tau_c^{{}_{\mathcal E}}(\mathcal H)\le 2|\mathcal E|-|\mathcal V|+p$.
\end{lemma}
\begin{proof} Suppose that $\mathcal H\setminus \mathcal F$ contains exactly $k$ components $\mathcal H_{i}=(\mathcal V_{i},\mathcal E_{i})$, $i=1,\ldots,k$. It follows from Lemma~\ref{acycle} that $|\mathcal V_{i}|=2|\mathcal E_{i}|+1$ for each $i\in [k]$. Thus $|\mathcal V|=\sum_{i\in [k]}|\mathcal V_{i}|= 2\sum_{i\in [k]}|\mathcal E_{i}|+k= 2(|\mathcal E|- |\mathcal F|)+k$, which means $2|\mathcal F|= 2|\mathcal E|-|\mathcal V|+k$. To establish the lemma, it suffices to prove $k\leq |\mathcal F|+ p$.
In case of $|\mathcal F|=0$, we have $\mathcal F=\emptyset$ and $k=p=|\mathcal F|+p$. In case of $|\mathcal F| \ge1$, suppose that $\mathcal F=\{e_{1},...,e_{|\mathcal F|}\}$. Because $\mathcal F$ is a minimal FES of $\mathcal H$, for each $i\in [|\mathcal F|]$, there is a cycle $\mathcal C_{i}$ in $\mathcal H\setminus (\mathcal F\setminus \{e_{i}\})$ such that $e_{i}\in \mathcal C_{i}$, and $\mathcal C_i\setminus e_i$ is a path in $\mathcal H\setminus \mathcal F$ connecting two of the three vertices in $e_i$. Considering $\mathcal H\setminus\mathcal F$ being obtained from $\mathcal H$ be removing $e_1,e_2,\ldots,e_{|\mathcal F|}$ sequentially, for $i=1,\ldots,|\mathcal F|$, since $|e_i|=3$, the presence of path $\mathcal C_i\setminus e_i$ implies that the removal of $e_i$ can create at most one more component. Therefore we have $k\le p+|\mathcal F|$ as desired.
\end{proof}
Given a {hypergraph} $\mathcal H=(\mathcal V,\mathcal E)$ with $n$ vertices and $m$ edges, {let $M_{\mathcal H}$ be the $\mathcal V\times\mathcal E$ incidence matrix}. From $M_{\mathcal H}$, we may construct a bipartite graph $G_{\mathcal H}$ with bipartition $\mathcal V,\mathcal E$ such that there is an edge of $G_{\mathcal H}$ between $v\in\mathcal V$ and $e\in\mathcal E$ if and only if $v\in e$ in $\mathcal H$.
Suppose that $\mathcal H$ is acyclic. It is easy to see that $G_{\mathcal H}$ is acyclic. Thus $M=M_{\mathcal H}$ falls within the class of {\em restricted totally unimodular} (RTUM) matrices defined by Yannakakis \cite{yannakakis1985}. As the name indicates, RTUM matrices are all totally unimodular. Hence the total unimodularity and LP duality give the well-known result \cite{berge1989} that {$\tau(\mathcal H)=\min\{\mathbf 1^T\mathbf x:M^T\mathbf x\ge\mathbf 1,x\geq 0\}=\max\{\mathbf 1^T\mathbf y:M\mathbf y\le\mathbf 1,y\geq 0\}=\nu(\mathcal H)$}. Moreover, since $M$ is RTUM, both a minimum transversal and a maximum matching of $\mathcal H$ can be found in $O(n(m+n\log n)\log n)$ time using Yanakakis's combinatorial algorithm \cite{yannakakis1985} based on the current best combinatorial algorithms for the $b$-matching problem and the maximum weighted independent set problem on a bipartite mulitgraph with $n$ vertices and $m$ edges, where the bipartite $b$-matching problem can be solved with the minimum-cost flow algorithm in $O(n\log n(m+n\log n))$ time (see Section 21.5 and Page 356 of \cite{schrijver2003}) and the maximum weighted independent set problem can be solved with maximum flow algorithm in $O(nm\log n)$ time (See Pages 300-301 of \cite{yannakakis1985}).
\begin{theorem}[\cite{berge1989,{yannakakis1985}}]\label{cyclefree}
Let $\mathcal H$ be a hypergraph with $n$ {non-isolated} vertices and $m$ edges. If $\mathcal H$ has no cycle, then $\tau(\mathcal H)=\nu(\mathcal H)$, and a minimum transversal and a maximum matching of $\mathcal H$ can be found in $O(n(m+n\log n)\log n)$ time.\qed
\end{theorem}
\section{Triangle packing and covering}\label{sec:tuza}
{This section establish several new sufficient conditions for Conjecture \ref{coj:tuza} as well as their algorithmic implications on finding minimum triangle covers. Section \ref{sec:high} deals with graphs of high triangle packing numbers. Section \ref{sec:irr} investigates irreducible graphs with many edges. Section \ref{sec:er} discusses Erd\H{o}s-R\'{e}nyi graphs with high densities.}
To each graph $G=(V,E)$, we associate a hypergraph $\mathcal H_G=(E,\mathscr T_G)$, referred to as {\em triangle hypergraph} of $G$, such that the vertices and edges of $\mathcal H_G$ are the edges and triangles of $G$, respectively. Since $G$ is simple, it is easy to see that $\mathcal H_G $ is $3$-uniform and linear, $ \nu(\mathcal H_G)=\nu_{t}(G)$ and $\tau(\mathcal H_G)=\tau_{t}(G)$. Note that $|\!|\mathcal H_G|\!|=|\mathscr T_G|<\min\{|V|^3,|E|^3\}$, and {$|E|\le3|\mathscr T_G|$} if $G$ is irreducible{, i.e., $\cup_{T\in\mathscr T_G}E(T)=E$. Note that the number of non-isolated vertices of $\mathcal H_G$ is upper bounded by $3|\!|\mathcal H_G|\!|=3|\mathscr T_G|$}.
\subsection{Graphs with many edge-disjoint triangles}\label{sec:high}
We investigate Tuza's conjecture for graphs with large packing numbers, which are firstly compared with the number of triangles, and then with the number of edges.
\begin{theorem}\label{c}
If graph $G$ and real number $c\in (0,1]$ satisfy $\nu_{t}(G)/|\mathscr T_G|\ge c$, then a triangle cover of $G$ with size at most $ \frac{3c+1}{3c}\nu_t(G)$ can be found in {$O(|\mathscr T_G|^3)$} time, which implies $\tau_t(G)/\nu_t(G)\le\frac{3c+1}{3c}$.
\end{theorem}
\begin{proof}
{We} consider the triangle hypergraph $\mathcal H_G=(E,\mathscr T_G)$ of $G$ {which} is 3-uniform and linear. By Corollary~\ref{cor:1/3}, we can {find} in $O(|\mathscr T_G|^3)$ time a FVS $\mathcal S$ of $\mathcal H_G$ with $|\mathcal S|\le |\mathscr T_G|/3$.
Since $\nu(\mathcal H_G)=\nu_{t}(G)\ge c|\mathscr T_G|$, it follows that $|\mathcal S|\le \nu(\mathcal H_G)/(3c)$. As $\mathcal H_G\setminus S$ is acyclic, Theorem~\ref{cyclefree} enables us to find in $O(|\mathscr T_G|^2\log^2|\mathscr T_G|)$ time a minimum transversal $\mathcal R$ of $\mathcal H_G\setminus S$ such that $|\mathcal R|=\tau(\mathcal H_G\setminus S)=\nu(\mathcal H_G\setminus \mathcal S)$. We observe that $\mathcal S\cup\mathcal R\subseteq E$ and $G\setminus (\mathcal S\cup\mathcal R)$ is triangle-free. Hence $\mathcal S\cup\mathcal R$ is a triangle cover of $G$ with size
\[|\mathcal S\cup\mathcal R|
\leq \frac{\nu(\mathcal H_G)}{3c}+\nu(\mathcal H_G\setminus\mathcal S) \leq\frac{3c+1}{3c}\nu(\mathcal H_G)=\frac{3c+1}{3c}\nu_t(G),\] which proves the theorem.
\end{proof}
The special case of $c=1/3$ in the above theorem {gives the following result providing a new sufficient condition for Tuza's conjecture.}
\begin{corollary}\label{th:1/3}If graph $G$ satisfies $\nu_{t}(G)/|\mathscr T_G|\ge1/3$, then $\tau_t(G)/\nu_t(G)\le2$.\qed
\end{corollary}
The condition $\nu_{t}(G)\ge|\mathscr T_G|/3$ in Corollary~\ref{th:1/3} applies, in some sense, only to the class of large scale sparse graphs (which, e.g., does not include complete graphs on four or more vertices). The mapping from the real number $c$ in the condition $\nu_{t}(G)\ge c|\mathscr T_G|$ to the coefficient $ \frac{3c+1}{3c}$ in the conclusion $\tau_t(G)\le \frac{3c+1}{3c}\nu_t(G)$ of Theorem \ref{c} shows the trade-off between conditions and conclusions. As in {Corollary}~\ref{th:1/3}, $c = \frac13$ maps to $ \frac{3c+1}{3c} = 2$ hitting the boundary of Tuza's conjecture. It remains to study graphs $G$ with $\nu_t(G)/|\mathscr T_G|<\frac13$. The next theorem (Theorem \ref{th:a-fourth}) tells us that actually we only need to take care of graphs $G$ with $\nu_t(G)/|\mathscr T_G|\in(\frac14-\epsilon,\frac13)$, where $\epsilon$ can be any arbitrarily small positive number. So, in some sense, to solve Tuza's conjecture, we only have a gap of $\frac13-\frac14=\frac1{12}$ to be bridged. Interestingly, for $c = \frac14$, we have {$\frac{3c+1}{3c} = \frac73 = 2.333...$}, which is much better than the best known general bound 2.87 due to Haxell \cite{haxell19}. Only when $c \leq \frac16$ does $\frac{3c+1}{3c}$ state a trivial bound equal to or greater than 3.
\begin{theorem}\label{th:a-fourth}
If there exists some real $\delta>0$ such that Conjecture \ref{coj:tuza} holds for every graph $G$ with $\nu_t(G)/|\mathscr T_G|\ge1/4-\delta$, then Conjecture \ref{coj:tuza} holds for every graph.
\end{theorem}
\begin{proof} If $\delta\geq \frac14$, the theorem is trivial. We consider $0 < \delta < \frac14$. As the set of rational numbers is dense, we may assume $\delta\in\mathbb Q$ and $ 1/4-\delta=i/j$ for some $i,j\in\mathbb N$. Therefore $i/j<1/4$ gives {$4i+1\le j$}, i.e., $4+1/i\le j/i$. It remains to prove that for any graph $G$ with $\nu_t(G)<(i/j)|\mathscr T_G|$ there holds $\tau_t(G)\le2\nu_t(G)$.
Write $k$ for the positive integer $i|\mathscr T_G|-j\cdot\nu_t(G)$. Let $G'$ be the disjoint union of $G$ and $k$ copies of {$K_4$}. Clearly, $|\mathscr T_{G'}|=|\mathscr T_G|+k|\mathscr T_{K_4}|=|\mathscr T_G|+4k$, $\tau_t(G')=\tau_t(G)+k\cdot\tau_t(K_4)=\tau_t(G)+2k$ and $\nu_t(G')=\nu_t(G)+k\cdot\nu_t(K_4)=\nu_t(G)+k$.
It follows that
\begin{eqnarray*}
(i/j)|\mathscr T_{G'}|&=&(i/j)(|\mathscr T_G|+4k)\\
&=&(i/j)((k+j\cdot\nu_t(G))/i+4k)\\
&=&(i/j)(j\cdot\nu_t(G)/i+(4+1/i)c)\\
&\le& \nu_t(G)+k\\
&=& \nu_t(G')
\end{eqnarray*}
where the inequality is guaranteed by $4+1/i\le j/i$. So $\nu_t(G')\ge(1/4-\delta)|\mathscr T_{G'}|$ together with the hypothesis of the theorem implies $\tau_t(G')\le2\nu_t(G')$, i.e., $\tau_t(G)+2k\le2(\nu_t(G)+k)$, giving $\tau_t(G)\le2\nu_t(G)$ as desired.
\end{proof}
\redcomment{In the proof of the above theorem, the property of $K_4$ that $\nu_t(K_4)/|\mathscr T_{K_4}|=1/4$ and $\tau_t(K_4)/\nu_t(K_4)=2$ plays an important role. It helps to reduce the general Tuza's conjecture to the special case where $\nu_t(G)\ge(1/4-\delta)|\mathscr T_G|$.}
The sufficient condition that compares the triangle packing number with the number of edges is based on the fact that every simple graph $G=(V,E)$ has a bipartite subgraph of at least $|E|/2$ edges, {which can be found in polynomial time}. Since this subgraph does not {contain} any triangle, we deduce that $\tau_t(G)\le|E|/2$, which implies the following result.
\begin{corollary}\label{cor:1/4} If $G=(V,E)$ is a graph such that $\nu_{t}(G)/|E|\ge c$ for some $c>0$, then $\tau_t(G)/\nu_t(G)\le 1/(2c)$. In particular, if $\nu_{t}(G)/|E|\ge1/4$, then $\tau_t(G)/\nu_t(G)\le 2$.\qed
\end{corollary}
{Thus if $\nu_{t}(G)/|E|\ge c$ for some $c>0$, then a triangle cover of $G$ with size at most $\nu_t(G)/(2c)$ can be found in polynomial time.} Complementary to Corollary \ref{th:1/3} whose condition mainly takes care of sparse graphs, the second statement of Corollary~\ref{cor:1/4} applies to many dense graphs, including complete graphs on $ 25$ or more vertices.
Similar to Corollary \ref{th:1/3} and Theorem \ref{th:a-fourth}, by which our future investigation space on Tuza's conjecture shrinks to interval $(\frac14-\epsilon,\frac13)$ w.r.t. $\nu_t(G)/|\mathscr T_G|$, Corollary \ref{cor:1/4} and the following Theorem \ref{th:1/5} narrow the interval w.r.t. $\nu_t(G)/|E|$ to $(\frac15-\epsilon,\frac14)$. Moreover, when taking $c = \frac15$ in Corollary \ref{cor:1/4}. we obtain $ \frac{1}{2c} = 2.5$, still better than Haxell's general bound $2.87$ \cite{haxell19}.
\begin{theorem}\label{th:1/5}
If there exists some real $\delta>0$ such that Conjecture \ref{coj:tuza} holds for every graph $G$ with $\nu_t(G)/|E|\ge1/5-\delta $, then Conjecture \ref{coj:tuza} holds for every graph.
\end{theorem}
\begin{proof} We use the similar trick to that in proving Theorem \ref{th:a-fourth}; we add a number of complete graphs on five (instead of four) vertices. We may assume $\delta\in(0,\frac{1}{5})\cap \mathbb Q$ and $ 1/5-\delta=i/j$ for some $i,j\in\mathbb N$. Therefore
$i/j<1/5$ and the integrality of $i,j$ imply $5+1/i\le j/i$. To prove Tuza's conjecture for each graph $G$ with $\nu_t(G)<(i/j)|E|$, we write $k=i|E|-j\cdot\nu_t(G)\in\mathbb N$. Let $G'=(V',E')$ be the disjoint union of $G$ and $k$ copies of $K_5$'s. Then $|E'|=|E|+10k$, $\tau_t(G')=\tau_t(G)+k\cdot\tau_t(K_5)=\tau_t(G)+4k$, $\nu_t(G')=\nu_t(G)+k\cdot\nu_t(K_5)=\nu_t(G)+2k$, and
\[(i/j)|E'|=(i/j)(|E|+10k)
=(i/j)(j\cdot\nu_t(G)/i+(10+1/i)k)\le \nu_t(G)+2k= \nu_t(G')\]
where the inequality is guaranteed by $10+1/i\le 2j/i$. So $\nu_t(G')\ge(1/5-\delta)|E'|$ together with the hypothesis the theorem implies $\tau_t(G')\le2\nu_t(G')$, i.e., $\tau_t(G)+4k\le2(\nu_t(G)+2k)$, giving $\tau_t(G)\le2\nu_t(G)$ as desired.
\end{proof}
\subsection{Graphs with many edges on triangles}\label{sec:irr}
Each graph has a unique maximum irreducible subgraph. Tuza's conjecture is valid for a graph if and only the conjecture is valid for its maximum irreducible subgraph. In this section, we study sufficient conditions for Tuza's conjecture on irreducible graphs that bound the number of edges {below in terms of} the number of triangles.
\begin{theorem}\label{th:2}
If $G=(V,E)$ is an irreducible graph such that $|E|/|\mathscr T_G|\ge2$, then a triangle cover of $G$ with cardinality at most $ 2\nu_t(G)$ can be found in $O(|\mathscr T_G|^2\log^2|\mathscr T_G|)$ time, which implies $\tau_t(G)/\nu_t(G)\le2$.
\end{theorem}
\begin{proof}Suppose that the linear 3-uniform hypergraph $\mathcal H=(E,\mathscr T_G)$ associated to $G$ has exactly $p$ components. By Lemma~\ref{cyclecover}, we can find in $O(|\mathscr T_G|^2)$ time a minimal FES $\mathcal F$ of $\mathcal H$ such that $|\mathcal F|\leq 2|\mathscr T_G|-|E|+p\le p$. Since $G $ is irreducible, we see that $\mathcal H$ has no isolated vertices, i.e., every component of $\mathcal H$ has at least one edge. Thus $\nu(\mathcal H)\geq p\geq |\mathcal F|$. For the acyclic hypergraph $ \mathcal H\setminus\mathcal F$, By Lemma~\ref{cyclefree} we may found in $O(|\mathscr T_G|^2\log^2|\mathscr T_G|)$ time a minimum transversal $\mathcal R$ of $\mathcal H\setminus \mathcal F$ such that \[|\mathcal R|=\tau(\mathcal H\setminus\mathcal F)=\nu(\mathcal H\setminus\mathcal F).\] Observe that $\mathcal R\subseteq E$ and $\mathcal F\subseteq\mathscr T_G$. If $\mathcal F=\emptyset$, set $\mathcal S=\emptyset$, else for each $F\in\mathcal F$, take $e_F\in E$ with $e_F\in F$, and set $\mathcal S=\{e_F:F\in\mathcal F\}$. It is clear that $\mathcal R\cup\mathcal S$ is a transversal of $\mathcal H$ (i.e., a triangle cover of $G$) with cardinality $|\mathcal R\cup\mathcal S|\leq\nu( \mathcal H\setminus\mathcal F)+ |\mathcal F| \leq 2\nu(\mathcal H)=2\nu_t(G)$, establishing the theorem.
\end{proof}
We observe that the graphs $G$ which consist of a number of triangles sharing a common edge satisfy $|E(G)|\ge 2|\mathscr T_G|$, but {$\nu_{t}(G)< |\mathscr T_{G}|/3$} when $|\mathscr T_{G}|\geq 4$. So in some sense, Theorem~\ref{th:2} works a supplement of Corollary \ref{th:1/3} for sparse graphs.
A multigraph is {\em series-parallel} if and only if it can be constructed
from a single edge by iteratively performing the {\em D-Operation} of doubling an edge and/or the {\em S-Operation} of subdividing an edge.
{A graph is a {\em $2$-tree} if and only if it can be constructed from a single edge by iteratively performing the {\em DS-Operation} of doubling an edge and subdivide the new edge with a new vertex.}
A subgraph of a 2-tree is called a {\em partial $2$-tree}. It is well-known that a {(simple)} graph is a partial 2-tree if and only if all of its maximal 2-connected subgraphs are series-parallel \cite{bodlaender1998}. Thus, a series-parallel {(simple)}
graph is a partial 2-tree. {In the following we show that every partial 2-tree $G$ satisfies $|E(G)|\ge 2|\mathscr T_G|$.}
\begin{corollary}
If $G=(V,E)$ is a partial $2$-tree, then a triangle cover of $G$ with cardinality at most $ 2\nu_t(G)$ can be found in $O(|E|^2\log^2|E|)$ time.
\end{corollary}
\begin{proof} In $O(|E|^2)$ time, we may remove from $G$ all edges that are not contained in any triangles. The resulting graph is still a partial 2-tree. So we may assume without loss of generality that $G$ is irreducible. Since each triangle of $G$ is contained a unique maximal 2-connected subgraph of $G$, we may further assume that $G$ is 2-connected. It follows that $G$ is series-parallel. Since $G$ is simple, it can be constructed from a single edge by iteratively performing the S-Operation and/or the DS-Operation. The S-Operation increases the number of edges and dose not change the number of triangles, while the DS-Operation increases the number of edges by 2 and the number of triangles by 1. Therefore, we have $|E|\ge 2|\mathscr T_G|$. The conclusion follows from Theorem \ref{th:2}.
\end{proof}
{Note that} partial $2$-trees are $K_4$-free planar graphs. The validity of Tuza's conjecture on partial $2$-trees has been verified in \cite{tuza1990,HKT2012}. {The 2-approximation algorithm for finding a minimum triangle cover in planar graphs implied by Tuza's proof \cite{tuza1990} runs in $O(|E|)$ time.}
Along the same line as in the previous subsection, regarding Tuza's conjecture on graph $G$, Theorem \ref{th:2} and the following Theorem \ref{th:3/2} jointly narrow the interval w.r.t. $|E(G)|/|\mathscr T_G|$ to $(1.5-\epsilon,2)$ for future study.
\begin{theorem} \label{th:3/2}
If there exists some real $\delta>0$ such that Conjecture \ref{coj:tuza} holds for every irreducible graph $G=(V,E)$ with $|E|/|\mathscr T_G|\ge3/2-\delta$, then Conjecture \ref{coj:tuza} holds for every irreducible graph (and therefore every graph).
\end{theorem}
\begin{proof} Again we apply the trick of adding copies of $K_4$. We may assume $\delta\in(0,3/2)\cap\mathbb Q$ and $ 3/2-\delta=i/j$ for some $i,j\in\mathbb N$. Therefore $2i+1\le 3j$, implying $(i/j)(4+1/i)\le 6$.
For any irreducible graph $G$ with $|E|<(i/j)|\mathscr T_G|$, we write $k=i|\mathscr T_G|-j|E|\in\mathbb N$. Let $G'$ be the disjoint union of $G$ and $k$ copies of $K_4$.
Then $G'$ is irreducible, and
\[
(i/j)|\mathscr T_{G'}|=(i/j)(|\mathscr T_G|+4k)=(i/j)(j|E|/i+(4+1/i)k)\le |E|+6k= |E'|.\]
It follows from the hypothesis of the theorem that $\tau_t(G')\le2\nu_t(G')$, i.e., $\tau_t(G)+2k\le2(\nu_t(G)+k)$, giving $\tau_t(G)\le2\nu_t(G)$ as desired.
\end{proof}
\subsection{Erd\H{o}s-R\'{e}nyi graphs with high densities}\label{sec:er}
Let $n$ be a positive integer, and let $p\in[0,1]$. The Erd\H{o}s-R\'{e}nyi random graph model \cite{Alon2008} is a probability space over the set $\mathcal G(n,p)$ of graphs $G=(V,E)$ on the vertex set $V=\{1,...,n\}$, where an edge between vertices $i$ and $j$ is included in $E$ with probability $p$ independent from every other possible edge, i.e., \[\text{\bf Pr}[ij\in E]=p\text{ for each pair of distinct }i,j\in V.\] The $\mathcal G(n,p)$ model is often used in the probabilistic method for tackling problems in various areas such as graph theory and combinatorial optimization.
The following result on the triangle packing numbers of complete graphs \cite{Brualdi2009} is useful in deriving a good estimation for the triangle packing numbers of graphs in $\mathcal G(n,p)$.
\begin{theorem}[\cite{Brualdi2009}]\label{steiner}
$\nu_t(K_{n})=|E(K_{n})|/3$ if and only if $n\equiv1,3\pmod6$.\qed
\end{theorem}
\begin{theorem}\label{random}
Suppose that $p> \sqrt{3}/2$ and $G=(V,E)\in\mathcal G(n,p)$. Then $\text{\bf Pr}[\nu_t(G)\geq |E|/4]=1-o(1)$ and $\text{\bf Pr}(\tau_t(G)\le2\nu_t(G))=1-o(1)$.
\end{theorem}
\begin{proof} Let $K_n$ denote the complete graph on $V$. For each edge $e\in K_{n}$, let $X_{e}$ be the indicator variable satisfying: $X_{e}=1$ if $e\in E$ and $X_{e}=0$ otherwise. Thus $\text{\bf E}[X_{e}]= p$, $X=\sum_{e\in K_{n}}X_{e}=|E|$, $ \text{\bf E}[X]=n(n-1)p/2$. Since $X_{e}, e\in K_{n}$, are independent 0-1 variables, by Chernoff Bounds, for each $\epsilon \in (0,1]$, $ \text{\bf Pr}[X> (1+\epsilon)\text{\bf E}[X]]\leq exp(-\epsilon^{2}\text{\bf E}[X]/3)=exp(-\epsilon^{2}n(n-1)p/6)=o(1)$. So \[\text{\bf Pr}[X\leq (1+\epsilon)\text{\bf E}[X]]= \text{\bf Pr}(X\leq (1+\epsilon)n(n-1)p/2)
=1-o(1).\]
On the other hand, by Theorem~\ref{steiner}, we can make $K_n$ have an edge-disjoint triangle decomposition by deleting at most three vertices, which implies that $\nu_t(K_{n})$ is lower bounded by $k=\lceil(n-3)(n-4)/6\rceil$. Thus we can take $k$ edge-disjoint triangles $T_{1}, \ldots, T_k$ from $K_{n}$. For each $i\in [k]$, let $Y_{i}$ be the indicator variable satisfying: $Y_{i}=1$ if $T_{i}\subseteq G $ and $Y_{i}=0$ otherwise. Note that $\text{\bf E}[Y_{i}]= p^{3}$ for each $i\in[k]$, $ \nu_t(G)\geq Y = \sum_{i=1}^kY_{i}$ and $ \text{\bf E}[Y]=kp^{3}$. Because $T_1,\ldots,T_k$ are edge-disjoint, $Y_1,\ldots,Y_k$ are independent 0-1 variables. By Chernoff Bounds, for each $\epsilon \in (0,1)$, $\text{\bf Pr}[Y< (1-\epsilon)\text{\bf E}[Y]]\leq exp(-\epsilon^{2}\text{\bf E}[Y]/2)\leq exp(-\epsilon^{2}(n-3)(n-4)p^{3}/12)=o(1)$.Thus
\[\text{\bf Pr}[\nu_t(G)\geq (1-\epsilon)(n-3)(n-4)p^{3}/6]\ge\text{\bf Pr}[\nu_t(G)\geq (1-\epsilon)kp^3]\ge\text{\bf Pr}[Y\geq (1-\epsilon)\text{\bf E}[Y]]=1-o(1).\]
Recall that $p> \sqrt{3}/2$. We can take $\epsilon \in (0,1)$ such that $\lim_{n\rightarrow \infty}\frac{(1-\epsilon)(n-3)(n-4)p^{3}/6}{(1+\epsilon)n(n-1)p/8}= \frac{4p^{2}(1-\epsilon)}{3(1+\epsilon)}> 1$. So for sufficient large $n$, we always have $(1-\epsilon)(n-3)(n-4)p^{3}/6 > (1+\epsilon)n(n-1)p/8$. Since we have $\nu_t(G)\geq (1-\epsilon)(n-3)(n-4)p^{3}/6$ with probability $1-o(1)$ and have $|E|=X\leq (1+\epsilon)n(n-1)p/2$ with probability $1-o(1)$, we obtain $\nu_t(G)\geq |E|/4$ with probability $1-o(1)$. It follows from Corollary \ref{cor:1/4} that $\text{\bf Pr}(\tau_t(G)\le2\nu_t(G))=1-o(1)$.
\end{proof}
\section{Conclusion}\label{sec:conclude}
Using tools from hypergraphs, we design polynomial-time algorithms for finding {a} small triangle covers {in} graphs, which particularly imply several sufficient conditions for Tuza's conjecture {(Conjecture \ref{coj:tuza})}.
\paragraph{Triangle packing and covering.} In this paper, we have established new sufficient conditions $\nu_t(G)/|\mathscr T_G|\ge1/3$ and $|E|/|\mathscr T_G|\ge2$ for Tuza's conjecture on packing and covering triangles in graphs $G$. We prove the sufficiency by designing {polynomial-time} combinatorial algorithms for finding a triangle cover of $G$ whose cardinality is upper bounded by $2\nu_t(G)$. The high level {idea} of these algorithms is to remove {\em some edges} from $G$ so that the triangle hypergraph of the remaining graph is {\em acyclic} (see the proofs of Theorems \ref{1/3} and \ref{th:2}), which guarantees that the remaining graph has equal triangle covering number and triangle packing number, and a minimum triangle cover of the remaining graph is computable in polynomial time (see Theorem \ref{cyclefree}). It is well-known that the acyclic condition in Theorem \ref{cyclefree} could be weakened to odd-cycle-freeness \cite{yannakakis1985}. So the lower bound $1/3$ and $2$ in the sufficient conditions could be (significantly) improved if we can remove (much) {\em fewer edges} from $G$ such that the triangle hypergraph of the remaining graph is {\em odd-cycle free}.
In view of Theorems~\ref{th:a-fourth}, \ref{th:1/5} and \ref{th:3/2}, the study on the graphs $G$ satisfying $\nu_t(G) / |\mathscr T_G|\ge1/4 $ or $\nu_t(G)/ |E|\ge1/5$ or $|E|/|\mathscr T_G|\ge3/2$ might suggest more insight and foresight for resolving Tuza's conjecture. These graphs are critical in the sense that they are standing on the border of the resolution.
\redcomment{Let us paying more attention to {\em extremal graphs} $G$ which satisfy Tuza's conjecture with tight ratio $\tau_t(G)/\nu_t(G)=2$. Actually, from Theorem~\ref{c}, Corollary~\ref{cor:1/4} and Theorem~\ref{th:2}, we can get a nice observation: for every irreducible extremal graph $G=(V,E)$, the following three inequalities hold on: $\nu_t(G)/ |\mathscr T_G|\leq 1/3$, $\nu_t(G)/ |E|\leq 1/4$, and $|E|/|\mathscr T_G|\leq 2$. Gregory J. Puleo first notices this observation.}
Another intermediate step towards resolving Tuza's conjecture is investigating its validity for the classical Erd\H{o}s-R\'{e}nyi random graph model $\mathcal G(n,p)$. In this paper, we have shown that Tuza's conjecture holds with high probability for graphs in $\mathcal G(n,p)$ when $p> \sqrt{3}/2$. It would be nice to prove the same result for {$p\in(0,\sqrt3/2]$}.
\paragraph{The generalization to {linear} 3-uniform hypergraphs.}
Our work has shown very close relations between triangle packing and covering in graphs and edge (resp. cycle) packing and covering in linear 3-uniform hypergraphs. The theoretical and algorithmic results on linear 3-uniform hypergraphs (Corollary \ref{cor:1/3} and Lemma \ref{cyclecover}) are crucial for us to establish sufficient conditions for Tuza's conjecture, and to find in strongly polynomial time a ``small'' triangle cover under the conditions (see Corollary \ref{th:1/3} and Theorem \ref{th:2}). Recall that, for any graph $G$, its triangle hypergraph $\mathcal H_G$ is linear 3-uniform, and Tuza's conjecture is equivalent to $\tau(\mathcal H_G)\leq 2\nu(\mathcal H_G)$. \redcomment{As a natural generalization, one may ask: Does $\tau(\mathcal H)\leq 2\nu(\mathcal H)$ hold for all linear 3-uniform hypergraphs $\mathcal H$? It is easy to see that $\{\mathcal H_G:G$ is a graph\} is {\em properly} contained in the set of linear 3-uniform hypergraphs. Unfortunately, Zbigniew Lonc pointed out there is a simple negative example: The Fano projective plane is an example of a linear 3-uniform hypergraph whose matching number is 1 and transversal number is 3(See Figure~\ref{Fano}).
Last but not the least, the arguments in the paper have actually proved the following stronger result.}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.48]{Fano.eps}
\caption{\label{Fano}The Fano projective plane}
\end{center}
\end{figure}
\begin{theorem}\label{th:condition'}
Let $\mathcal H=(\mathcal V,\mathcal E)$ be a linear $3$-uniform hypergraph without isolated vertices. Then a transversal of $\mathcal H$ with cardinality at most $2\nu(\mathcal H)$ can be found in polynomial time, which implies $\tau(\mathcal H)\le2\nu(\mathcal H)$, if one of the following conditions is satisfied:
(i) $\nu(\mathcal H)/|\mathcal E|\ge\frac13$, (ii) $|\mathcal V|/|\mathcal E|\ge2$.\qed
\end{theorem}
\redcomment{Comparing the above result on linear 3-uniform hypergraphs $\mathcal H$ with its counterpart on graphs presented in Theorem \ref{th:condition}, one might notice that the condition on the lower bound of $\nu(\mathcal H)/|\mathcal V|$ is missing. This reason is that we do not have a nontrivial constant upper bound on $\tau(\mathcal H)/|\mathcal V|$.}
\paragraph{Acknowledgements:} \redcomment{The authors are indebted to Gregory J. Puleo and Zbigniew Lonc for their invaluable comments and
suggestions which have greatly improved the presentation of this paper.}
|
\section{Introduction}
\label{sec:introduction}
Clusters represent the largest gravitationally bound
and densest structures in the universe. As such,
they provide a unique context for studying the formation
of both galaxies and large-scale structure. In order to understand
how galaxies form and evolve in dense environments, and
how these environments themselves evolve, it is important
to trace the origins of galaxy clusters. Protoclusters
are identified as regions containing significant overdensities
of galaxies, which may not yet be virialized but
will later evolve into rich galaxy clusters.
These cluster seeds provide insight into the origins of the environmental
trends observed among galaxies \citep[e.g.,][]{Dressler1980}, as well as the assembly of
clusters themselves. At this point, tens of protoclusters
have been spectroscopically confirmed at $z> 2$ \citep{Chiang2013},
with an order of magnitude more high-redshift protocluster candidates
photometrically identified \citep[e.g.,][]{Planck2015a,Chiang2014}.
The protocluster at $z\approx3.09$ in the SSA22 field is one
of the best-studied cluster progenitors in the literature.
This structure was first identified by \citet{Steidel1998}
in the course of a large spectroscopic survey for $z\sim 3$ Lyman
Break Galaxies \citep[LBGs;][]{Steidel2003}, and contains an LBG redshift-space
overdensity of $\delta_{LBG}=5.0\pm 1.2$ \citep{Steidel2000}.
Based on analytic calculations, \citet{Steidel1998} argued
that the SSA22 protocluster will evolve by $z=0$ into a rich Coma-like
cluster with $M\sim 10^{15} M_{\odot}$.
Following the initial protocluster discovery, an extensive suite of multi-wavelength observations has
been assembled in the SSA22 field, spanning from radio through X-ray
wavelengths and including high-resolution {\it Hubble Space Telescope}
imaging. Specifically, mid- and far-infrared, submillimeter, and
millimeter observations have been used to study obscured
star formation and AGN activity in the SSA22 protocluster
\citep[e.g.,][]{Geach2005,Webb2009,Tamura2009,Umehata2015},
while deep JHK imaging has enabled the analysis of the most
massive protocluster galaxies \citep{Uchimoto2012,Kubo2013},
and deep {\it Chandra} observations have unveiled the X-ray AGN population
\citep{Lehmer2009a,Lehmer2009b}. The above observations suggest enhancements
in the star-formation rate density,
and the abundance of both massive galaxies and AGNs in the SSA22 protocluster, relative to
the ``field" at $z\sim 3$.
Quite remarkably, $\sim 1500$ narrowband-selected Ly$\alpha$
emitting galaxies (LAEs) at $z\sim 3.09$ have been identified in the original
$9'\times9'$ SSA22 pointing and the surrounding 1.38 deg$^2$ area
\citep{Steidel2000,Matsuda2004,Hayashino2004,Yamada2012a},
mapping the SSA22 protocluster and its surroundings on scales $>100$ comoving Mpc. The sample of SSA22 LAEs includes 12 giant
Ly$\alpha$ ``blobs" (LABs), i.e., regions of Ly$\alpha$ emission
extending over $100-200$ proper kpc in diameter
\citep{Matsuda2011}, which may be specifically associated with galaxy overdensities
\citep{Prescott2008,Steidel2000}. Follow-up spectroscopy of LAEs in SSA22 has been used to
trace out the three-dimensional structure of this forming protocluster, in which
three 30 Mpc $\times$ 10 Mpc filaments of LAEs may intersect
in the region of highest galaxy overdensity \citep{Matsuda2005}.
However, with a sample of 56 spectroscopically-confirmed LAEs spread out over
170 arcmin$^2$, the sampling in redshift and position is not particularly dense.
We have obtained spatially dense spectroscopic sampling of LBGs
and LAEs, including 146 galaxies at $z\sim 3.09$ within the original $9'\times9'$ pointing of \citet{Steidel1998}.
This region also contains the highest overdensity of LAEs in the
SSA22 structure \citep{Yamada2012a}. Our large spectroscopic sample reveals
previously unknown substructure within the SSA22 overdensity, both in redshift
space and on the sky. Characterizing the substructure in SSA22 is crucial for identifying
analogous protoclusters in cosmological simulations and understanding the origin
and fate of the SSA22 overdensity.
In \S\ref{sec:obsAndMethods}, we describe our observations and
redshift measurements for SSA22 LBGs and LAEs.
In \S\ref{sec:results} we discuss the substructure observed within the SSA22
protocluster both in redshift space and on the sky.
Finally, in \S\ref{sec:discussion} we compare with other work and discuss
the implications of our results for understanding the SSA22 protocluster.
For this work, we adopt cosmological parameters of
$H_0=67.8 \rm \ km \ s^{-1} \ Mpc^{-1}$, $\Omega_M = 0.3089$, and
$\Omega_{\Lambda}=0.6911$ \citep{Planck2015b}.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{f1.pdf}
\caption{{\bf Left:} Redshift histogram of objects within the $z\sim 3.09$ overdensity in the SSA22
field. This histogram exhibits a double-peaked structure, with one peak centered at $z=3.069$
and the other at $z=3.095$. The histogram contains a total of 146 galaxies comprising 106 LAEs and 40 LBGs,
plus one additional LBG at $z=3.04$ to show the larger-scale environment. {\bf Top right:}
Distributions of LAEs (blue histogram) and LBGs (green histogram),
both of which appear to trace two distinct concentrations in redshift space. {\bf Bottom right:}
Distributions of brighter $(R_{AB} < 26.26)$ (red histogram) and fainter $(R_{AB}\ge 26.26)$ (blue
histogram) galaxies. We find no significant difference in the redshift distributions of the brighter and fainter
galaxies.}
\label{fig:histogram-all}
\end{figure*}
\section{Observations \& Methods}
\label{sec:obsAndMethods}
\subsection{Data}
\label{sec:data}
In the analysis presented here, we consider the redshifts and sky
positions of both LAEs and LBGs in the SSA22 field.
The LAEs were selected to lie at $3.05 \leq z \leq 3.12$ using deep broadband BV imaging from
Subaru/Suprime-cam and narrowband $4980\rm \AA$ imaging from
Keck/LRIS and Subaru/Suprime-cam. The LBGs were drawn from the large survey of
\citet{Steidel2003}, in which LBGs were identified using deep $U_nG{\cal R}$ imaging and
photometric selection criteria tuned to find star-forming galaxies at $z\sim3$.
Our key results are based on dense spectroscopic sampling of galaxies over a $5.5' \times 7.6'$
region within the original ``SSA22a" pointing of \citet{Steidel1998}, centered on R.A. = 22:17:34,
decl.=00:15:04 (J2000). Keck/LRIS spectroscopy for the LBGs in our sample was obtained over the course of
several observing campaigns utilizing different instrumental setups.
Within the region of dense spectroscopic sampling, redshifts have been measured for 85\% of LBG photometric
candidates and 82\% of LAE narrowband-selected candidates.
A full description of these imaging and spectroscopic
observations, including the methods of data reduction, can be found in
\cite{Steidel2003} and \citet{Nestor2011, Nestor2013}.
\subsection{Redshift Measurements}
\label{sec:zMeasurement}
We measured redshifts for objects in our sample based on the observed
wavelengths of Ly$\alpha$ emission, and, if present, interstellar metal absorption lines.
For measuring redshifts, we developed code to systematically fit profiles
to either emission or absorption lines.
Ly$\alpha$ emission profiles were initially fit with single Gaussian functions.
In the case of Ly$\alpha$ emission exhibiting a double-peaked morphology,
however, a single Gaussian did not provide an adequate description of the profile,
and an additional Gaussian was required to obtain an acceptable fit.
Galaxies with double-peaked Ly$\alpha$ emission in our sample typically show one of two different
morphologies, with either two peaks of comparable amplitude or else a significantly stronger red peak
and weaker blue peak. We used the same method to measure the redshift for both types of
double-peaked Ly$\alpha$ emission, based on the wavelength of the center of the trough between
the two peaks. To estimate the interstellar absorption redshift in LBG spectra,
we averaged the redshift measurements from individual absorption lines.
In our sample, some objects had repeat observations. In such cases, we adopted
the redshift measured from the observation with the highest signal to noise ratio (SNR).
In total, we measured redshifts for 202 galaxies, including 116 LAEs and 86 LBGs.
Due to our interest in the $z\sim3.09$ protocluster, we now focus only on galaxies
at $3.05 \le z \le 3.12$.
This redshift cut reduces our sample to 146 galaxies,
comprising 106 LAEs and 40 LBGs.
Ly$\alpha$ emission and interstellar absorption features typically trace gas that is outflowing
from galaxies, which perturbs their measured redshifts from the systemic value.
In order to investigate the large-scale structure traced by LBGs and LAEs in SSA22,
we need to estimate the systemic redshift of each galaxy. We obtain
systemic redshifts for objects in our sample by shifting the measured Ly$\alpha$
and interstellar absorption redshifts to the rest frame of the galaxies.
The shift required to estimate the systemic redshift has been measured to be
different for LAEs and LBGs, and depends on which spectral features are
observed. Based on the results of \citet{Trainor2015}, we shifted the redshifts of objects
classified as LAEs with only Ly$\alpha$ emission by $\delta v=-200 \rm \ km \ s^{-1}$, and the
redshifts of LAEs showing both Ly$\alpha$ emission and interstellar absorption
lines by $\delta v = (0.114\times \Delta v_{abs,em} - 230 \rm \ km \ s^{-1})$,
where $\Delta v_{abs,em}$ is the velocity difference between the Ly$\alpha$
and interstellar absorption redshifts. Based on equations presented in \citet{Adelberger2003},
we shifted the redshifts of objects classified as LBGs showing Ly$\alpha$
emission by $\delta v = -310 \rm \ km \ s^{-1}$,
and finally, the redshifts of LBGs showing only interstellar absorption
lines by $\delta v = 150 \rm \ km \ s^{-1}$. To apply
the systemic redshift correction, we converted the shifts from velocity to
redshift space using $\delta v/c = \delta z/(1+z)$. Following these
corrections, we obtained a final list of LAE and LBG systemic redshifts.
\section{Results}
\label{sec:results}
\subsection{The SSA22 Redshift Distribution}
\label{sec:histogram}
We constructed a histogram from the finalized list of 146 systemic redshifts in the
SSA22 field (Figure~\ref{fig:histogram-all}). This histogram shows that
the spike discovered in \citet{Steidel1998} at $z\approx3.09$ contains
two distinct peaks, one with a central redshift of $z_{peak,b}=3.069$ (blue peak), and
another centered on $z_{peak,r}=3.095$ (red peak).
The peaks are separated by $\Delta v\sim 1900 \mbox{ km s}^{-1}$ and $\sim 24$ comoving
Mpc along the line of sight. We fit a Gaussian to each of the two peaks, defining the boundary between the two
peaks as the bottom of the trough of the distribution ($z=3.0788$). Accordingly,
the blue and red peaks contain, respectively, 43 and 103 galaxies.
In velocity space, the standard deviations of the blue and red peaks are, respectively:
$\sigma_{v,b}= 350 \rm \ km \ s^{-1}$ and $\sigma_{v,r} = 540\rm \ km \ s^{-1}$. Taking
into account the uncertainties associated with estimating systemic redshifts
\citep[$\sim 170\mbox{ km s}^{-1}$;][]{Adelberger2003,Trainor2015},
we find the intrinsic widths to be 310 and 520 $\rm \ km \ s^{-1}$, respectively, for the blue and red peaks.
We note that the velocity width and centroid of the red peak are similar to values found by
\citet{kubo2015} for a sample of $K$-selected galaxies in the SSA22 overdensity with
rest-frame optical spectroscopic measurements.
The properties of the two peaks are summarized in Table~\ref{table:hist-parameters}.
\begin{deluxetable*}{ccccc}
\tabletypesize{\footnotesize}
\tablecolumns{5}
\tablewidth{0pt}
\tablecaption{Redshift Histogram Fit Parameters}
\tablehead{
\colhead{Peak} & \colhead{$N_{gal}$} & \colhead{$z_{peak}$} & \colhead{$\sigma_z$\tablenotemark{a}} & \colhead{$\sigma_v$ (km s$^{-1}$)\tablenotemark{a}} }
\startdata
Blue & 43 & 3.069 $\pm$ 0.001 & (4.74 [4.25] $\pm$ 0.72) $\times \ 10^{-3}$ & 350 [310] $\pm$ 53 \\
Red & 103 &3.095 $\pm$ 0.001 & (7.37 [7.12] $\pm$ 0.54) $\times \ 10^{-3}$ & 540 [520] $\pm$ 40
\enddata
\tablenotetext{a}{Values in brackets result after deconvolving the uncertainties in systemic redshifts
\citep{Adelberger2003,Trainor2015} from the observed sample standard deviations in redshift and velocity space.}
\label{table:hist-parameters}
\end{deluxetable*}
We investigated the relative distributions of LAEs and LBGs in the two peaks.
As shown in Figure~~\ref{fig:histogram-all}, both the
LAE and LBG distributions show evidence of double peaked structure.
LBGs [LAEs] comprise 28\% [72\%] of the blue peak, and 27\% [73\%] of the red peak.
A Kolmogorov-Smirnov (K-S) test indicates a probability of 98\% that the LAE
and LBG redshift distributions are drawn from the same parent distribution.
We also considered how brighter and fainter galaxies are distributed within the SSA22 stucture.
All LAEs and LBGs are covered by a deep Subaru/Suprime-Cam R-band image \citep{Hu2004}, ranging in magnitude
from $R_{AB}=21.61$ (a QSO) to limits fainter than $R_{AB}=27.0$ (roughly half of the LAEs).
We classified each galaxy as either brighter or fainter,
depending on its magnitude relative to the
sample median of $R_{AB}=26.26$. As shown in Figure~\ref{fig:histogram-all}, the brighter
and fainter histograms are very similar. A K-S test indicates a probability
of 87\% that the brighter and fainter redshift distributions are drawn from the same parent
distribution. Galaxy correlation functions of $z\sim 3$ LBGs suggest that more UV-luminous galaxies are more
strongly clustered \citep{Ouchi2004, Adelberger2005}. In future work, we will compare the
relative distributions of brighter and fainter galaxies (i.e., more and less massive dark matter halos)
in cosmological simulations of protoclusters \citep[e.g.,][]{Klypin2011} with those observed in the SSA22 overdensity.
\citet{Lehmer2009b} and \citet{Alexander2016} identify
eight X-ray AGNs in the SSA22 field with spectroscopic
redshifts falling within the protocluster range. We used the X-ray catalog of \citet{Lehmer2009a} to
find two additional matches with spectroscopically-confirmed LAEs in our catalogs
\citep[LAE017 at $z=3.105$ and LAE076 at $z=3.066$;][]{Nestor2013},
for a total sample of 10 X-ray AGN at $z=3.06-3.11$. Nine out of these 10 AGNs fall within the red peak of the SSA22 redshift
histogram (and five out of 6 within the field of view of our dense spectroscopic sampling).
Although there are more LBGs and LAEs identified in the higher-redshift peak (103, vs. 43 in the lower-redshift peak),
the AGN distribution is even more skewed towards higher redshift. \citet{Lehmer2009b} previously
noted an enhanced AGN fraction in the SSA22 protocluster, but now we can specifically associate this enhancement
with the red peak at $z=3.095$. We also note that all seven giant LABs from \citet{Matsuda2011} with
spectroscopic redshifts are located in the red peak.
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{f2.pdf}
\caption{Sky and redshift positions of our spectroscopic sample. Sky coordinates are listed relative to the
southeast corner of the field. Blue triangles [red circles] represent galaxies contained in the redshift peak
centered at $z=3.069$ [$z=3.095$]. The shade of each point corresponds
to redshift of the galaxy being plotted (see colorbar). White on the colorbar corresponds to the bottom of the trough between the two
peaks in the redshift histogram. {\bf Bottom left:} Galaxies plotted as decl. vs. R.A.
The field is separated into four quadrants by the grey dashed lines.
The numbers of galaxies in each quadrant in the blue and red peaks are listed, respectively, with blue and red numbers.
The region of dense spectroscopic sampling extends over $5.5' \times 7.6'$.
{\bf Top:} Galaxies plotted as $z$ vs. R.A.. Here the absence of galaxies between the blue and red peaks is
clearly visible, as is the segregation of objects in the blue peak towards the western half of the
field. {\bf Bottom right:}
Galaxies plotted as decl. vs. $z$. There is no significant difference in the distributions of blue and red galaxies in decl.}
\label{fig:scatter}
\end{figure}
\subsection{Sky Positions}
\label{sec:scatter}
We use the sky positions of SSA22 LAEs and LBGs to investigate
the spatial distribution of galaxies within the observed redshift
structure. At $z=3.09$, for the assumed cosmology, one arcsecond corresponds
to $7.81$~kpc (proper). Our observations span $\approx 15$~comoving Mpc
on a side, covering the densest region but not the full extent of the structure
mapped out by \citet{Matsuda2005} and \citet{Yamada2012a} (see \S\ref{sec:discussion}).
The redshift interval $z=3.05-3.12$ corresponds to a Hubble-flow distance of $\sim 66$ comoving
Mpc, but the translation between redshift and distance is affected by peculiar velocities.
Quite strikingly, we find differences in the spatial distributions of galaxies located
in the two redshift peaks. Figure~\ref{fig:scatter} displays the sky positions of objects
in our spectroscopic sample, indicating members of blue and red redshift peaks
with, respectively, blue and red symbols. In order to search for a spatial separation of
galaxies in different redshift peaks, we split the positions of objects in our sample into
quadrants, represented by grey dashed lines in Figure~\ref{fig:scatter} (bottom left), and
count the number of galaxies in each quadrant that are within either blue or red redshift peaks.
Most significantly, we find that galaxies in the blue peak concentrate towards the western
half of the field, with the northeast quadrant showing the strongest deficiency in
lower-redshift galaxies. This result suggests that galaxies in blue and red redshift peaks
do not span the same physical volume, and that we are observing the edge of the
structure containing the galaxies in the blue peak.
\section{Discussion}
\label{sec:discussion}
We have measured redshifts for $146$ LAEs and LBGs at $z=3.05-3.12$ in the SSA22 field,
which clearly delineate two separate peaks in redshift space and also show segregation on the sky.
Collectively, these observations suggest that the the two redshift peaks correspond
to physically distinct structures.
As stated previously, our spectroscopic observations
do not cover the full area of the overdensity in LAEs traced by \citet{Matsuda2005}
and \citet{Yamada2012a}. Therefore, with our data alone, we cannot establish the spatial extent of the double-peaked
redshift histogram. However, we can start to address this question with the observations of \citet{Yamada2012b},
who survey $4980 \rm \ arcmin^2$ in the SSA22 field with Subaru/FOCAS spectroscopy and find 91
LAEs at a redshift consistent with the $z\sim3.09$ structure.
There are 19 LAEs in common with our sample, resulting in 72 additional objects over a significantly
larger area on the sky (though with sparser sampling). As shown in Figure~~\ref{fig:yamData},
in both the \citeauthor{Yamada2012b} sample alone, and the total sample combined with our data, we detect a double-peaked structure
in redshift. This result suggests that the double-peaked structure is not localized
to the field of view of our observations. We also observe a relative lack of blue-peak galaxies towards
the eastern portion of the field in the combined sample, consistent with our results over a smaller field.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.95\textwidth]{f3.pdf}
\caption{{\bf Left:} Redshift histogram of our sample, combined with the sample of \citet{Yamada2012b}. The double-peaked structure
is present in the combined distribution, as well as in the sample of \citet{Yamada2012b} alone.
{\bf Right:} Sky positions of the combined SSA22 sample.
Symbols are as in Figure~\ref{fig:scatter} for galaxies in our sample, with
square points for galaxies from \citet{Yamada2012b}, and color-coding as in Figure~\ref{fig:scatter}.
The original SSA22a pointing from \citet{Steidel1998} is indicated in black.
The relative lack of lower-redshift galaxies (blue points) toward the east side of the
field is now established over a wider area.}
\label{fig:yamData}
\end{figure*}
The large-scale structure of the SSA22 overdensity has been considered by \citet{Matsuda2005}, who identify
three filaments of length $\sim 30$ comoving Mpc intersecting within the highest-density region of LAEs on the sky,
and at $z=3.094$ in redshift space. One of these apparent filaments is traced by a concentration of galaxies at $z=3.074$,
which corresponds to the blue peak of our measured histogram.\footnote{Given that the majority of objects
in our histogram are LAEs, whose redshifts were shifted by $\Delta z\sim -0.003$ to the systemic frame,
our measured $z_{peak,b}=3.069$ would appear at a redshift of $3.072$ in the absence of Ly$\alpha$
velocity corrections, which were not applied by \citeauthor{Matsuda2005}).} With much sparser spectroscopic
sampling, \citeauthor{Matsuda2005} describe this lower-redshift concentration as connecting smoothly
with the large belt-like structure of LAEs at $z=3.088-3.108$ that extends $\sim 40$ comoving
Mpc across the sky in the northwest -- southeast direction, and corresponds to the red peak of our redshift histogram.
In contrast, our well-sampled redshift distribution indicates that the two peaks are distinct structures that do not join
smoothly in redshift space or fully overlap on the sky, leading to a qualitatively different description
of the large-scale structure in the SSA22 overdensity at $z<3.08$.
\citet{Matsuda2005} also present a gradient in redshift from northwest (higher redshift) to southeast (lower redshift)
within the extended band of LAEs at $z=3.088-3.108$. As shown in Figure~\ref{fig:scatterR},
within our smaller field of view, we find a redshift gradient in the red peak ($z=3.08-3.11$) from
southwest (higher redshift) to northeast (lower redshift), i.e., rotated by 90 degrees with respect to the gradient
presented in \citet{Matsuda2005}. Our dense spectroscopic sampling covers only $\sim 30$\% of the linear dimension
of the band of LAEs, so we cannot establish the extent of the redshift gradient we observe. However, based on higher
spatial resolution, our sampling of redshifts as a function of position within the red peak reveals a more complex
picture than the linear gradient presented by \citet{Matsuda2005}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{f4.pdf}
\caption{Sky positions of objects in the red peak of the SSA22 redshift histogram ($z=3.08-3.11$). Points are color coded
by redshift as indicated in the color bar, and follow a gradient in redshift space from the southwest (lower right, higher redshift)
to northeast (upper left, lower redshift).}
\label{fig:scatterR}
\end{figure}
In order to understand the full three-dimensional architecture, dynamics, origin, and fate of the substructure
we have uncovered with our spectroscopic survey of the SSA22 overdensity, we must obtain comparably
dense spectroscopy over a significantly larger area -- at the very least within the $27' \times 34'$ Subaru/Suprime-Cam
pointing \citep[SSA22-Sb1 from][]{Yamada2012a} containing the initial SSA22a field of \citet{Steidel1998}.
The spectroscopic samples of \cite{Matsuda2005} and \citet{Yamada2012b}
begin to address this need, but are both incomplete in terms of areal coverage of the field,
and too sparsely sampled within the areas covered. Detailed comparisons with cosmological simulations of large-scale
structure formation \citep[e.g.,][]{Springel2005,Klypin2011} are also required to understand whether
the double-peaked substructure in SSA22 will coalesce by $z=0$ or remain as two distinct clusters.
We will perform such comparisons in future work.
\section*{Acknowledgements}
We thank Abhimat Gautam and Tommaso Treu for useful comments.
CCS acknowledges support from NSF grants AST-0908805 and AST-1313472.
AES acknowledges support from the David \& Lucile Packard Foundation.
We wish to extend special thanks to those of Hawaiian ancestry on
whose sacred mountain we are privileged to be guests. Without their generous hospitality, most
of the observations presented herein would not have been possible.
|
\section{Introduction}
\label{sec:intro}
At present day the concordance model $\Lambda$CDM is the most famous cosmological model, which together with the paradigm of inflation predicts the hierarchical structure formation, constituted of approximately $4\%$ of luminous matter, $26\%$ of dark matter (DM) and about $70\%$ is a exotic component known as dark energy (DE) which it is traditionally presented as the main responsible for the late accelerated expansion of the Universe. In $\Lambda$CDM model, the DM is made up of colisionless non baryonic particles and the DE is provided by the cosmological constant $\Lambda$ with an equation of state (EoS) $w=-1$. This model is in excellent agreement with the cosmological observations of the anisotropies of the cosmic microwave background (CMB) radiation, baryonic acoustic oscillations (BAO), Supernovaes Ia (SNIa), etc. Nevertheless, the standard model has some fundamental problems related to the nature of the dark matter and dark energy \cite{2008ARA&A..46..385F,2006astro.ph..9591A}. In the context of DE, there are several theoretical arguments against a cosmological constant. One is the coincidence problem, that is, why today the value of dark energy density is of the same order of magnitude than dark matter density. Other important fundamental issue is the fine tuning of the present value of $\Lambda$ which is completely in disagreement with the predictions of particle physics \cite{weinberg89,cop06}. Therefore, several dark energy models with a dynamical EoS have been proposed to alleviate the problems of the cosmological constant \cite{2008ARA&A..46..385F}. For instance, the EoS can be parametrizated in different ways, being one of the most popular the Chevallier-Polarski-Linder (CPL) \cite{Chevallier:2000qy, Linder:2003nc}. In addition, there are several models where a scalar field may play the role of the dark energy of the Universe, for example, the quintessence \cite{Caldwell:1998, Ratra:1988}, phantom \cite{cal02, chi00,par99}, quintom \cite{Guo:2004} and k-essence fields \cite{armendariza,armendarizb, chi00}. On the other hand, a great number of theoretical models of dark energy consider interactions with dark matter (solve the problem of coincidence), e.g. Interacting Dark Energy \cite{CalderaCabral:2008bx}, the Holographic dark energy (are free from the cosmological constant problem.)\cite{coh99,suss95,hooft93, hooft01,zhang10}, modified gravity \cite{Dvali:2000hr} and Braneworld models \cite{garcia}. These models can be classified as phantom if EoS $\omega < -1$ , or as quintessence if $\omega > -1$, where in the first case a fluid multicomponent is required with at least one phantom constituent, which has been shown to suffer from serious theoretical problems and the second case the general relativity needs to be extended to a more general theory on cosmological scales \cite{Nesseris:2006er}. Among all these models we need to discriminate which one is the most favored by the current observations. The most popular method of discrimination is through the Akaike Information Criterion and Bayesian \cite{Shi:2012ma,Akaike:74,Schwarz:78}, which penalizes so much the number of data points as the number of parameters of each model. The most common tests comparing models with observations are SNIa, CMB, BAO,.., etc, which are considered as \textit{geometric test} and determine H(z) independent of the validity of Einstein's equations, directly through the redshift dependence with cosmological distances (e.g. the angular diameter distance $d_A (z)$, the scale of the sound horizon $D_V (z)$ and mass gas fraction $f_{gas}$, among others). Other important observational methods to determine H(z) use \textit{dynamics test}, which measure the evolution of the density of matter-energy (background or perturbations) and they are connected with geometry through a theory of gravity. An example of a dynamical test of geometry (H(z)) is the measured linear growth factor of the matter density perturbations D(a), whose measure can be obtained by different methods as redshift distortion factor in redshift surveys ($A(z)=f(z)\sigma_{8}(z)$), number counts of galaxy clusters ($dN(M,z)/dMdz$), large scale structure power spectrum (P(k)) and Integrated Sachs-Wolfe (ISW) effect. In this paper we include this dynamic test through data from $ A_obs (z) $, which is important to know the effects of dark energy on the growth of structure (baryons and dark matter).\\
Our analysis begin with the cosmological models $\oint 2$, then we introduce the observational tests $\oint 3$, the statistical analysis of data $\oint 4$ and then we finish with the study of the history of expansion through the deceleration parameter $\oint 5$. In section $\oint 6$ we present the summary and discussion of results.
\section{Cosmological models}
\label{sec2}
In order to put constraints on DE cosmological models we need to calculate the theoretical angular diameter distance of the model and then compare it with the observations. The angular diameter distance for a Friedmann-Lemaitre-Robertson-Walker (FLRW) model universe,
for a source at redshift $z$ is
\begin{equation}
d_{A} (z,\Theta_i^m) = \frac{3000h^{-1}}{(1+z)} \frac{1}{\sqrt{\mid \Omega_{k} \mid}}
\sin \varsigma \left( \int_{0}^{z} \frac{\sqrt{\mid \Omega_{k} \mid}}{E(z,\Omega_i)}dz\right),
\label{eq2:2.1}
\end{equation}
where $h=H_{0}/100 \mathrm{kms}^{-1}\mathrm{Mpc}^{-1}$ is dimensionless Hubble parameter and
the function $\sin \varsigma(x)$ is defined such that it can be $\sinh(x)$ when $\Omega_{k} >0$,
$\sin(x)$ when $\Omega_{k} <0$ and $x$ when $\Omega_{k} =0$ \cite{Hogg:1999ad}. At present, all the signs of dark energy come from the measurements of expansion rate of the universe $H(z)$, which gives us its evolution and thus the history of the expansion. In standard FLRW cosmology, the expansion rate as a function of the scale factor $H(a)$ is given by the Friedmann equation as:
\begin{equation}
E^2(a,\Omega_i) = \Omega_{r}a^{-4} + \Omega_{m}a^{-3} + \Omega_{k}a^{-2} + \Omega_{X} e^{3\int_a^1 \frac{da'}{a'} \left( 1+w(a')\right) }
\label{eq2:2.2}
\end{equation}
where $H(a)/H_0=E(a,\Omega_i)$, $H_0$ is the curent value of the expansion rate and the scale factor is related to redshift as $1+z=a^{-1}$. In the equation (\ref{eq2:2.2}) $\Omega_i$ is the current energy density divided by today's critical density $\rho_{cri}= 3 H_0^2 / 8\pi G$, In the form of the \textit{i-th} component of the fluid density of: radiation ($\Omega_{r}$), matter ($\Omega_{m}$), curvature ($\Omega_{k}$) and dark energy ($\Omega_{X}$). The ratio of the pressure to the energy density $\omega (a)=p(a)/\rho (a)$ is the EoS of dark energy, the which divide our models into two cases: in one the energy density of the fluid is constant and in the other the energy density of the fluid is dynamic. In all cosmological models, the density parameter of curvature is free $\Omega_{k}$ and on which constraints are obtained. For each model of dark energy corresponds a vector of parameters $\Theta_{i}^{model}= \left\lbrace \theta_i, \Omega_i \right\rbrace $, where $\theta_i = \left\lbrace h, \sigma_8 \right\rbrace$ and $\Omega_i = \left\lbrace \Omega_r , \Omega_m , \Omega_k , \Omega_x \right\rbrace$ is multicomponent fluid for the analysis of the present work.
\subsection{$\Lambda$CDM model}
We begin our analysis with the standard cosmological model.
In this paradigm, the DE is provided by the cosmological constant $\Lambda$, with an EoS, such that $w=-1$. In this model the dimensionless Hubble parameter $E^{2}(z,\Theta)$ is given by
\begin{equation}
E^{2}(z,\Theta) = \Omega_{r}(1+z)^{4} + \Omega_{m}(1+z)^{3} + \Omega_{k}(1+z)^{2} + \Omega_{X},
\label{eq2:2.3}
\end{equation}
where $\Omega_{m}$ and $\Omega_{X} = \Omega_{\Lambda}=1-\Omega_{m}-\Omega_{k}-\Omega_{r}$ are the density parameters for matter and dark energy respectively and
$\Omega_{r}$ is the radiation parameter. The parameter vector is $\Theta_i^{\Lambda CDM}=\left\lbrace h,\sigma_8, \Omega_{m},\Omega_{k}\right\rbrace$ and the best fit is shown in Table \ref{tab:LCDM}.\\
\begin{table}[htb]
\centering
\begin{tabular}{ll}
\hline
$\Lambda$ \textit{Cold Dark Matter model} \\
\hline
\hline
\textit{h}=$0.6961\pm 0.0021$ & $\sigma_8$=$0.761\pm 0.019$ \\
$\Omega_m$=$0.2867\pm 0.0032$ & $\Omega_k$=$0.00024\pm 0.0011$ \\
\hline
\end{tabular}
\caption{Best fit parameters $\Theta_i$ with all data set to $\Lambda$CDM model.}
\label{tab:LCDM}
\end{table}
Note that the parameter $\sigma_8$ enters into our analysis as a free parameter and from now in all cosmological models. This parameter represents the amplitude of the fluctuations of galaxies and is obtained from the Growth Parameter of structure $A_{obs}(z)$. Using a chi square estimate we find that $\sigma_8=0.761\pm0.019$ (see Figure \ref{fig:Sigma3}, top-right), this result is compatible with the obtained from \cite{planckXVI}, \cite{2015arXiv150201589P}, however PLANCK15 reports a lower uncertainty. Aditionally our best fit value for the dark energy density divided by the critical density today with 68\% confidence is $\Omega_\Lambda=0.7095\pm0.0054$, that is according with the limits reports by \cite{2015arXiv150201589P}. The value of cosmological constant in this work is positive and different from zero ($\Lambda = 1.5277 \pm 0.0092 \times 10^{-35}s^{-2}$), This value of $\Lambda$ is consistent with measurements obtained by the High-Z Supernova Team and the Supernova Cosmology Project \cite{1999ApJ...517..565P} \cite{2001AIPC..586..316C}. Some derived parameters for this model are showed in Table \ref{tab:par2}.\\
To see the evolution of cosmic acceleration we plot the deceleration parameter q(z) in the Figure \ref{fig:qz}. As expected, q(z) has negative values at late times ($q_0=-0.67$), and positive values at an earlier epochs. The transition from one decelerated expansion phase to an accelerated occurs at $z\sim0.92$ using data from Baryon Acoustic Oscillations (BAO), in accordance with previous works \cite{2012PhLB..710..251C}.
\subsection{$w$CDM model}
An extension of the standard model where $\Lambda$ has $w=-1$ is to consider that the EoS still is constant but its value can deviate of $-1$. In this case the dimensionless Hubble parameter $E^{2}(z,\Theta)$ for a FLRW universe with curvature reads as
\begin{eqnarray}
E^{2}(z,\Theta) = \Omega_{r}(1+z)^{4} + \Omega_{m}(1+z)^{3} + \Omega_{k}(1+z)^{2} +
\Omega_{X} (1+z)^{3(1+w)},
\label{eq2:2.4}
\end{eqnarray}
where $\Omega_{X}=(1-\Omega_{m}-\Omega_{k}-\Omega_{r})$. In this model the set of free parameters are \linebreak $\Theta_i^{\omega CDM}=\left\lbrace h,\sigma_8, \Omega_{k}, \Omega_{m},\omega\right\rbrace$. The best fit values are showed in the Table \ref{tab:wCDM}.\\
\begin{table}[htb]
\centering
\begin{tabular}{ll}
\hline
$\omega$ \textit{Cold Dark Matter model} \\
\hline
\hline
\textit{h}=$0.6945\pm 0.0021$ & $\sigma_8$=$0.742\pm 0.020$ \\
$\Omega_m$=$0.2893\pm 0.0033$ & $\Omega_k$=$-0.0038\pm 0.0020$ \\
$\omega$=$-1.091\pm 0.031$ & \\
\hline
\end{tabular}
\caption{Best fit parameters $\Theta_i$ with all data set to $\omega CDM$ model.}
\label{tab:wCDM}
\end{table}
The Figure \ref{fig:Sigma3} shows the diagram of statistical confidence in the space
of parameters ($\Omega_m$ - $\omega$) (top-left) and (h - $\sigma_8$) (top-center) to $\omega CDM$ model, where we can see that $\Lambda CDM$ model ($\omega=-1$) is discarded to $3\sigma$ combining all data sets. The result obtained in \cite{Shi:2012ma} ($w=-0.990 \pm 0.041$) is outside of our restriction to $3\sigma$, however, the constraints obtained by \cite{planckXVI} to $\omega CDM$ ($w=-1.13^{+0.13}_{-0.10}$) are in good agreement with our work to $1\sigma$. Recently in \cite{2015arXiv150201589P} the result obtained for the equation of state of the $\omega CDM$ model is $w=-1.006 \pm 0.045$, whose lower limit is consistent with our restriction to $2\sigma$.\\
Figure \ref{fig:qz} shows the evolution of the deceleration parameter q(z) as a function of redshift. A transition from one decelerated expansion phase to an accelerated occurs at $z \sim 0.86$ using only data from Baryon Acoustic Oscillations (BAO). Therefore $\omega CDM$ model predicts an accelerated expansion of the universe consistent with the cosmological observations, which is consistent with works previously reported \cite{Shi:2012ma}.\\
\subsection{Chevalier-Polarski-Linder model}
Other extension to the standard scenario is allowing that equation of state of
dark energy varyies with redshift via some parametrization.
One of the most popular is the Chevallier-Polarski-Linder (CPL) parametrization
\cite{Chevallier:2000qy} \cite{Linder:2003nc} given by
\begin{equation}
w(z) = w_{0} + w_{1} \frac{z}{1+z},
\label{eq2:2.5}
\end{equation}
where $w_{0}$ y $w_{1}$ are constants to be fitted.
The dimensionless Hubble parameter $E(z)$ for CPL parametrization can be expressed as
\begin{equation}
E^{2}(z,\Theta) = \Omega_{r}(1+z)^{4} + \Omega_{k} (1+z)^2+\Omega_{m}(1+z)^{3} +
\Omega_{X} X(z),
\label{eq2:2.6}
\end{equation}
where $\Omega_{X} = \left( 1 - \Omega_{k} - \Omega_{m} -\Omega_{r} \right)$ and
\begin{equation}
X(z)=(1+z)^{3(1 + w_{0} + w_{1})} \exp \left[ - \frac{3w_{1}z}{1+z}\right].
\label{eq2:2.7}
\end{equation}
The free parameters to be constrained are $\Theta_i^{CPL}=\left\lbrace h,\sigma_8,\Omega_{k},\Omega_{m}, w_{0}, w_{1} \right\rbrace$. The best fit values using all the observational tests are showed in the Table \ref{tab:CPL}.\\
\begin{table}[htb]
\centering
\begin{tabular}{ll}
\hline
\textit{Chevalier-Polarski-Linder model} \\
\hline
\hline
\textit{h}=$0.6952\pm 0.0022$ & $\sigma_8$=$0.737\pm 0.020$ \\
$\Omega_m$=$0.2897\pm 0.0038$ & $\Omega_k$=$-0.0080\pm 0.0024$ \\
$\omega_0$=$-0.90\pm 0.13$ & $\omega_a$=$-1.13\pm 0.85$ \\
\hline
\end{tabular}
\caption{Best fit parameters $\Theta_i$ with all data set to \textit{CPL} model.}
\label{tab:CPL}
\end{table}
The Figure \ref{fig:Sigma3} (down-left) shows the diagram of statistical confidence to 1, 2 and 3 sigma in the space of parameters ($\omega_a - \omega_0$) for the CPL model. Note that the limits on \textit{h}, $\Omega_m$ and $\sigma_8$ using all cosmological data are compatible with those obtained in \cite{Shi:2012ma} and \cite{planckXVI}. However, there is tension for the constraints on $\Omega_k$, $\omega_0$ and $\omega_a$, due to the degeneracy between the curvature and the state equation. Recently \citep{planckXVI} combine Planck+WP+BAO gets $\omega_0 = -1.04^{+0.72}_{-0.69}$, which is in good agreement with our estimates. The CPL cosmological model is reduced to $\Lambda CDM$ when $\omega_0 = -1$ and $\omega_a = 0$, however, we can see that the cosmological constant is outside the contour to $3\sigma$ in our analysis (Figure \ref{fig:Sigma3}, down-left). The main physical parameters derived for this model are found in Table \ref{tab:par2}, which are very close to the reference model $\Lambda CDM$. \\
The point of transition from one decelerated phase to an accelerated phase occurs at $z \sim 0.85$ using only BAO data, which is showed in the Figure \ref{fig:qz}. Furthermore, there is a point at $z \sim 0.11$, where the expansion reaches its maximum acceleration, and then begin to slow down the acceleration at the current epoch ($q_0 = -0.48$). This slowing-down acceleration phase was found by \cite{Shafieloo:2009} and \cite{victor_2012} using SNIa data and recently by C\'ardenas, Bernal \& Bonilla \cite{victor_fgas} using only $f_{gas}$. This feature is presented as a anomalous behavior at low redshift \cite{victor_fgas} \cite{2014JCAP...10..017M}. \\
\subsection{Interacting Dark Energy model}
In interacting dark energy (EDE) models there is a relation between the energy density of dark energy $\rho_{x}$ and the energy density of dark matter $\rho_{m}$ that could alleviate the cosmic coincidence problem. For this, the general approach is common to introduce an interacting term in the right side of continuity equations in the following way \cite{Amendola:1999er,CalderaCabral:2008bx,Cai:2004dk,Dalal:2001dt,Guo:2007zk}
\begin{eqnarray}
\dot{\rho_{m}} &+& 3H\rho_{m} =\delta H \rho_{m},\nonumber\\
\dot{\rho_{x}} &+& 3H\left(1 + w_{x}\right)\rho_{x}= -\delta H \rho_{m},
\label{eq2:2.8}
\end{eqnarray}
where $w_{x}$ is the equation of state of dark energy and $\delta$ is a constant
interacting term to be fitted with the observations. Thus, the dimensionless Hubble parameter for
this interacting model is as follows
\begin{eqnarray}
E^{2}(z,\Theta) = \Omega_{r}(1+z)^{4} + \Omega_{k}(1+z)^{2} + \Omega_{m}\Psi (z) +
\Omega_{X} (1+z)^{3(1 + w_{x})},
\label{eq2:2.9}
\end{eqnarray}
where $\Omega_{X} = (1 - \Omega_{m} - \Omega_{k}-\Omega_{r})$ and
\begin{equation}
\Psi (z) = \frac{\left( \delta(1+z)^{3(1 + w_{x})} + 3 w_{x}(1+z)^{3 - \delta} \right)}{\delta + 3 w_{x}}
\label{eq2:2.10}
\end{equation}
This model is characterized by six parameters $\Theta_i^{IDE}=\left\lbrace h,\sigma_8, \Omega_{k},\Omega_{m}, w_{x},\delta \right\rbrace$.
We show the best fit values of these parameters in Table \ref{tab:IDE}.\\
\begin{table}[htb]
\centering
\begin{tabular}{ll}
\hline
\textit{Interacting Dark Energy model} \\
\hline
\hline
\textit{h}=$0.6873\pm 0.0028$ & $\sigma_8$=$0.715\pm 0.019$ \\
$\Omega_m$=$0.3275\pm 0.0039$ & $\Omega_k$=$0.0030\pm 0.0022$ \\
$\omega_x$=$-1.059\pm 0.037$ & $\delta$=$-0.0063\pm 0.0024$ \\
\hline
\end{tabular}
\caption{Best fit parameters $\Theta_i$ with all data set to \textit{IDE} model.}
\label{tab:IDE}
\end{table}
In this case the Figure \ref{fig:Sigma3} (down-center) shows the parameter space for $\delta$ y $\omega _x$ at $1\sigma$, $2\sigma$ and $3\sigma$ of statistical confidence. The $\Lambda CDM$ is recovered when $\omega _x = -1$ and $\delta = 0$. however, our results show that the model $\Lambda CDM$ is discarded to $3\sigma$ level of statistical confidence. On the other hand, If the term coupling in equation (\ref{eq2:2.8}) takes on a negative value ($\delta < 0$) this corresponds to a transfer from dark matter to dark energy, whereas a positive coupling term ($\delta > 0$) implies the opposite and we can see that the transfer of dark matter to dark energy is favored in this work. The present analysis of the EoS of dark energy shows a phantom behavior ($ \ omega _x <$ -1), result consistent with those obtained by \cite{Shi:2012ma}.\\
The transition state decelerated-accelerated occurs approximately at redshift $z \sim 1.06$ as can be observed in Figure \ref{fig:qz}, using only BAO data. The analysis of the deceleration parameter q(z) shows an anomalous behavior (slowdown of acceleration) at very low redshift (almost currently), which it is evident from the fast change of the slope of q (z) from $z \sim 0.5$ to $z \sim 0.0$.
\subsection{Early Dark Energy model}
In the early dark energy (EDE) scenarios it is proposed that the energy density of dark energy can be significant at high redshifts. This can be if the dark energy tracks the dynamics of the background fluid density \cite{SteinhardtEDE, Wetterich:1988}. These models could ameliorate the coincedence problem of the cosmological constant. Here, we generalize the EDE model proposed by \cite{2006JCAP...06..026D} adding a curvature term. The dimensionless Hubble parameter for this EDE model is
\begin{equation}
E^{2}(z,\Theta) = \frac{\Omega_{r}(1+z)^{4} +\Omega_{m}(1+z)^{3} + \Omega_{k}(1+z)^{2}}{1 - \Omega_{X}},
\label{eq2:2.11}
\end{equation}
where $\Omega_{X}$ is given by
\begin{eqnarray}
\Omega_{X} = \frac{\Omega_{X_0} - \Omega_{e}\left[ 1 - (1+z)^{3 w_{0}} \right] }{\Omega_{X_0} + f(z)} + \Omega_{e} \left[ 1 - (1+z)^{3 w_{0}} \right]\
\label{eq2:2.12}
\end{eqnarray}
and
\begin{equation}
f(z) = \Omega_{m}(1+z)^{-3w_{0}} + \Omega_{r}(1+z)^{-3w_{0} +1 } + \Omega_{k}(1+z)^{-3w_{0} -1 }
\label{eq2:2.13}
\end{equation}
such that $\Omega_{X_0} = 1 - \Omega_{m} - \Omega_{k}- \Omega_{r}$ is the current dark energy density. Thus, we have five free parameters $\Theta_i^{EDE}=\left\lbrace h,\sigma_8,\Omega_{k},\Omega_{m}, \Omega_{e},\omega_{0}\right\rbrace$. The best fit of these parameters is shown in Table \ref{tab:EDE}. \\
\begin{table}[htb]
\centering
\begin{tabular}{ll}
\hline
\textit{Early Dark Energy model} \\
\hline
\hline
\textit{h}= $0.6850\pm 0.0029$ & $\sigma_8$=$0.715\pm 0.019$ \\
$\Omega_m$=$0.3257\pm 0.0044$ & $\Omega_k$=$0.0040\pm 0.0024$ \\
$\omega_0$=$-1.101\pm 0.036$ & $\Omega_e$=$0.014\pm 0.010$ \\
\hline
\end{tabular}
\caption{Best fit parameters $\Theta_i$ with all data set to \textit{EDE} model.}
\label{tab:EDE}
\end{table}
Using the bayesian analysis we found $\Omega_e=0.014\pm0.010$ (see Figure \ref{fig:Sigma3}, down-right), which is in accordance with \cite{2006JCAP...06..026D} reporting a $\Omega_e<0.04$ by using a combination data from WMAP+VSA+CBI+ BOOMERANG+SDSS+SNIa and is also in agreement with \cite{2013PhRvD..87h3009P}, who reports a constraint $\Omega_e<0.015$ at 95\% confidence level. In Figure \ref{fig:Sigma3} (down-rigth) we can see that the model $\Lambda CDM$ ($\Omega_e=0$, $\omega_0 =-1$) is discarded even at $3\sigma$.\\
The deceleration parameter $q(z)$ has a transition phase acelerated-decelerated to $z\sim0.73$, moreover, q(z) shows the same anomalous behavior of slowdown of aceleration that CPL and IDE models for low redshift using only data from BAO, in this case at $z_{low}\sim 0.05$.
\section{Cosmological test}
Fluctuations in the matter and radiation density include dark matter $\Omega_m$, baryon matter $\Omega_b$ and photons $\Omega_\gamma$, which interact with each other through gravitational potential created by themselves \cite{2008CamUnivPres,1989KolbTurner}. These fluctuations grow through gravitational instability as the universe expands, then decouple to form on the one hand, the Cosmic Microwave Background (CMB) and on the other hand, the Large Scale Structure (LSS) of the Universe. In the early universe the radiation was hot enough to ionize hydrogen and due to the Thomson scattering it is coupled to the baryons, forming one photon-baryon fluid. The radiation pressure of photons opposes the gravitational compression of the fluid producing a harmonic motion, whose amplitude perturbations does not grow but slowly decays, establishing thus the so-called Baryon Acoustic Oscillations (BAO). The gravitational driving force is due to the Newtonian potential, where overdensities of photon-baryon fluid produce oscillations, which determines the characteristic patterns of fluctuations in the power spectrum of matter and radiation \cite{2005APh....24..334W}. Subsequently photons are diluted with the expansion and stream out of potential wells. Although effectively without pressure, the baryons still contribute to inertia and gravitational mass of fluid, which produces a change in the balance of pressure and gravity, then the baryons drag photons in potential wells. Previous to $z_{drag}$, perturbations in the photon-baryon fluid propagate as acoustic waves with speed of sound $c_s$, which at the epoch of drag define a sound comoving horizon $r_s$. At recombination epoch ($z_{cmb}$), photons decoupled from matter, where baryons can now constitute neutral elements and radiation is scattered for last time, forming the so-called CMB \cite{2008CamUnivPres}. The resultant fluctuations in CMB, which are observed in radiation maps with anisotropy of the order of $\Delta T/T \sim 10^{-5}$, are better studied in its power spectrum.\\
In the following we present the details of the main observational sample: CMB by using Shift parameter $R$; BAO by means of Distance Ratio Scale $D_v (z)/r_s$ and growth of LSS through Redshift-Space Distortions $A(z)=f(z)\sigma_{8}(z)$, that we adopt in order to constrain the free parameters of the different cosmological models presented in the previous section, Including the calculation of some derived parameters which are presented in Table \ref{tab:par1}, with their corresponding definition and/or physical meaning.\\
\begin{table*}[htb]
\centering
\begin{tabular}{ll}
\hline
\hline
Parameter & Physical meaning and/or definition \\
\hline
\textit{h} & Dimensionless Hubble parameter \\
$\Omega_m$ & Dimensionless dark matter density parameter\\
$\Omega_\Lambda$ & Dimensionless dark energy density parameter to $\Lambda CDM$\\
$\Omega_k$ & Dimensionless curvature density parameter \\
$\sigma_8$ & RMS matter fluctuations in linear theory \\
$\omega$ & Constant EoS to $\omega CDM$ \\
$\omega (a) = \omega_0 + (1+a)\omega_1$ & EoS for \textit{CPL} parametrization \\
$\omega_x$, $\delta$ & EoS and dimensionless coupling term for \textit{IDE} \\
$\omega_0$, $\Omega_e$ & EoS and asymptotic dark energy density term for \textit{EDE}\\
\hline
$H_0 = 100 h$ & Current expansion rate in $Km.s^{-1} Mpc^{-1}$\\
$t_0=H_0^{-1}$ & Age of the Universe today (in Gyr) \\
$\Omega_b=0.045$ & Dimensionless baryon density parameter \\
$\Omega_r =\Omega_\gamma (1 + 0.2271 N_{eff})$ & Dimensionless radiation density parameter \\
$\Omega_\gamma =2.469 \times 10^{-5} h^{-2}$ & Dimensionless photon density parameter\\
$\Omega_\nu$ & Dimensionless neutrino density parameter ($\Omega_r =\Omega_\gamma + \Omega_\nu$)\\
$N_{eff}=3.04$ & Effective number of relativistic neutrino degrees of freedom\\
$\omega_m=\Omega_m h^2$ & Physical dark matter density\\
$\omega_b=\Omega_b h^2$ & Physical baryon density\\
$\rho_{cri}= 3 H_0^2 / 8\pi G$ & Critical density ($1.88 \times10^{â29} h^2g/cm^3$) \\
$\Omega_X$ & Dimensionless dark energy density parameter\\
$\rho_X= \rho_{cri} \Omega_X$ & Physical dark energy density \\
$\Lambda = 8\pi G \rho_{\Lambda}$ & Cosmological constant where $\rho_{\Lambda} = \rho_{cri} 3 H_0^2$ \\
$c_s$ & Sound speed \\
$r_s$ & Comoving size of sound horizon \\
$z_{drag}$ & Redshift at which baryon-drag optical depth equals unity \\
$r_{drag} = r_s (z_{drag})$ & Comoving size of the sound horizon at $z_{drag}$\\
$r_{s}/D_v (z)$ & BAO distance ratio scale \\
$z_{cmb}$ & Redshift at decoupled photon-baryon \\
$R(z_{cmb})$ & Scaled distance at recombination ($z_{cmb}$) \\
$l_A (z_{cmb})$ & Angular scale of sound horizon at recombination ($z_{cmb}$) \\
\hline
\end{tabular}
\caption{Cosmological parameters used in the analysis. For each, we give the symbol and summary definition. The top block contains free parameters used for each cosmological model. The lower blocks define various derived parameters.}
\label{tab:par1}
\end{table*}
\subsection{CMB}
A particular cosmological test to probe dark energy is the angular scale of sound horizon ($r_s$) at time of decoupling ($z_{cmb}\sim 1090$), the which is encrypted in the location $l^{TT}_{1}$ of the first peak of the CMB power spectrum. We include CMB information of WMAP 9-yr and Planck 13 data \cite{Hinshaw:2012,planckXVI} to probe the expansion history up to the last scattering surface. The $\chi^2$ for the CMB data is constructed as
\begin{equation}
\chi^2_{cmb} = X^TC_{cmb}^{-1}X,
\label{eq3:3.1}
\end{equation}
where
\begin{equation}
X _{wmap9}=\left(
\begin{array}{c}
l_A - 302.40 \\
R - 1.7246 \\
z_{cmb} - 1090.88
\end{array}\right).
\label{eq3:3.2}
\end{equation}
Here $l_A$ is the ``acoustic scale'' defined as
\begin{equation}
l_A = \frac{\pi d_L(z_{cmb})}{(1+z)r_s(z_{cmb})},
\label{eq3:3.3}
\end{equation}
where $d_L(z)=D_L(z)/H_0$ and the redshift of decoupling $z_{cmb}$ is
given by \cite{husugi},
\begin{equation}
z_{cmb} = 1048[1+0.00124(\Omega_b h^2)^{-0.738}]
[1+g_1(\Omega_{m}h^2)^{g_2}],
\label{eq3:3.4}
\end{equation}
\begin{equation}
g_1 = \frac{0.0783(\Omega_b h^2)^{-0.238}}{1+39.5(\Omega_b
h^2)^{0.763}},
g_2 = \frac{0.560}{1+21.1(\Omega_b h^2)^{1.81}},
\label{eq3:3.5}
\end{equation}
The ``shift parameter'' $R$ defined as \cite{BET97}
\begin{equation}
R = \frac{\sqrt{\Omega_{m}}}{c(1+z_{cmb})} D_L(z).
\label{eq3:3.6}
\end{equation}
$C_{cmb}^{-1}$ in Eq. (\ref{eq3:3.1}) is the inverse covariance
matrix, the wich is to WMAP9 data
\begin{equation}
C_{cmb^{wmap9}}^{-1} = \left(
\begin{array}{ccc}
3.182 & 18.253 & -1.429\\
18.253 & 11887.879 & -193.808\\
-1.429 & -193.808 & 4.556
\end{array}\right).
\label{eq3:3.7}
\end{equation}
To Planck 13 data we obtained:
\begin{equation}
X _{Planck13}=\left(
\begin{array}{c}
l_A - 301.57 \\
R - 1.7407 \\
\omega_b - 0.02228
\end{array}\right),
\label{eq3:3.8}
\end{equation}
where $\omega_b = \Omega_b h^2$ and $\Omega_b = 0.045$ for this work. The inverse covariance matrix for ($l_A, R, \omega_b$) is how follow:
\begin{equation}
C_{cmb^{Planck13}}^{-1} = \sigma_{i} \sigma_{j} C_{NorCov_{i,j}},
\label{eq3:3.9}
\end{equation}
where $\sigma_{i} = \left( 0.18, 0.0094, 0.00030 \right)$ and normalized covariance matrix is
\begin{equation}
C_{NorCov_{i,j}} = \left(
\begin{array}{ccc}
1.0000 & 0.5250 & -0.4235\\
0.5250 & 1.0000 & -0.6925\\
-0.4235 & -0.6925 & 1.0000
\end{array}\right).
\label{eq3:3.10}
\end{equation}
\subsection{BAO}
The large scale correlation function measured from 2dF Galaxy Redshift Survey and SDSS redshift survey, displays a peak which was identified with the expanding spherical wave of baryonic perturbations from acoustic oscillations at recombination and comoving scale of about $150 h^{-1}Mpc$ \cite{2005MNRAS.362..505C,2005ApJ...633..560E}. Baryon Acoustic Oscillations (BAO) are periodic fluctuations in the density of radiant matter (baryonic), which were printed in the primordial plasma (photon-baryon) before decoupling and as standard rules provide a powerful tool to measure the properties of dark energy \cite{2010deot.book..246B,2005APh....24..334W}. We use the recent results presented by Beutler et al. in \cite{2016MNRAS.455.3230B}. The $\chi^2$ for BOSS-WiggleZ (BW) BAO data is given by:
\begin{equation}
\chi^2_{\scriptscriptstyle BW} =
(\bar{D}_{obs}-\bar{D}_{th})C_{\scriptscriptstyle
BW}^{-1}(\bar{D}_{obs}-\bar{D}_{th})^T,
\label{eq3:3.11}
\end{equation}
where $\bar{D}_{obs} = (2056,2132,2100,2516)Mpc=(CMASS-DR11,cc-BW,WiggleZ-BW,WiggleZ-hz)$ is data vector at $z=(0.57,0.57,0.57,0.73)$ and $\bar{D}_{obs}$ is given by
\begin{equation}
\bar{D}_{obs}=D_v(z) \frac{r_{s}^{fid}}{r_{s}(z_{d})}
\label{eq3:3.12}
\end{equation}
where $r_{s}^{fid}=(149.28,150.18,150,18,148,6)Mpc$ \cite{2016MNRAS.455.3230B} and $r_s(z)$ is the comoving sound horizon
\begin{equation}
r_s(z) = c \int_z^\infty \frac{c_s(z')}{H(z')}dz',
\label{eq3:3.13}
\end{equation}
where the sound speed
\begin{equation}
c_s(z) = \frac{1}{\sqrt{3(1+\bar{R_b}/(1+z)}},
\label{eq3:3.14}
\end{equation}
with $\bar{R_b} = 31500 \Omega_{b}h^2(T_{CMB}/2.7\rm{K})^{-4}$ and $T_{CMB}$ = 2.726K. The redshift $z_{drag}$ at the baryon drag epoch is fitted with the formula proposed by \cite{1998ApJ...496..605E},
\begin{equation}
z_{drag} =
\frac{1291(\Omega_{m}h^2)^{0.251}}{1+0.659(\Omega_{m}h^2)^{0.828}}[1+b_1(\Omega_b
h^2)^{b_2}],
\label{eq3:3.15}
\end{equation}
where $b_1 = 0.313(\Omega_{m}h^2)^{-0.419}[1+0.607(\Omega_{m}h^2)^{0.674}]$ and $b_2 = 0.238(\Omega_{m}h^2)^{0.223}$. The matrix in equation (\ref{eq3:3.11}) is
\begin{equation}
C_{\scriptscriptstyle BW}^{-1} = \left(
\begin{array}{cccc}
250.47 & -3.48 & 0.09 & 0.004\\
-3.48 & 36.32 & -5.67 & -4.79\\
0.09 & -5.67 & 4.27 & -3.26\\
0.004 & -4.79 & -3.26 & 18.90
\end{array}\right) \times 10^{-5}
\label{eq3:3.16}
\end{equation}
that it is the inverse of the covariance matrix
\begin{equation}
C_{\scriptscriptstyle BW}^{cov} = \left(
\begin{array}{cccc}
400 & 56 & 88 & 29 \\
56 & 4225 & 7410 & 2348\\
88 & 7410 & 40000 & 8772\\
29 & 2348 & 8772 & 7396
\end{array}\right)
\label{eq3:3.17}
\end{equation}
which is given by $C_{\scriptscriptstyle BW}^{cov} = \sigma_{i} \sigma_{j} R_{ij}$, where $\sigma_{j} = (20,65,200,86)Mpc$ is the vector of variance and $R_{ij}$ is the normalized correlation matrix given by:
\begin{equation}
R_{ij}= \left(
\begin{array}{cccc}
1 & 0.043 & 0.022 & 0.017 \\
0.043 & 1 & 0.57 & 0.42\\
0.022 & 0.57 & 1 & 0.51\\
0.017 & 0.42 & 0.51 & 1
\end{array}\right)
\label{eq3:3.18}
\end{equation}
which takes into account the BAO measurements from the cross-correlation between CMASS (BOSS) and the WiggleZ samples, including high redshift WiggleZ data point \cite{2016MNRAS.455.3230B}. Similarly, for the SDSS DR7 BAO distance measurements, the $\chi^2$ can be expressed as \cite{2010MNRAS.401.2148P}
\begin{equation}
\chi^2_{\scriptscriptstyle SDSS} =
(\bar{d}_{obs}-\bar{d}_{th})C_{\scriptscriptstyle
SDSS}^{-1}(\bar{d}_{obs}-\bar{d}_{th})^T,
\label{eq3:3.19}
\end{equation}
where $\bar{d}_{obs} = (0.1905,0.1097)$ is the datapoints at $z=0.2$
and $0.35$. $\bar{d}_{th}$ denotes the distance ratio
\begin{equation}
d_z = \frac{r_s(z_d)}{D_V(z)}.
\label{eq3:3.20}
\end{equation}
Here, $r_s(z)$ is the comoving sound horizon as above. $C_{\scriptscriptstyle SDSS}^{-1}$ in Eq. (12) is the inverse
covariance matrix for the SDSS data set given by
\begin{equation}
C_{\scriptscriptstyle SDSS}^{-1} = \left(
\begin{array}{cc}
30124 & -17227\\
-17227 & 86977
\end{array}\right).
\label{eq3:3.21}
\end{equation}
For the 6dFGS BAO data \cite{2011MNRAS.416.3017B}, there is only
one data point at $z=0.106$, the $\chi^2$ is easy to compute:
\begin{equation}
\chi^2_{\scriptscriptstyle 6dFGS} =
\left(\frac{d_z-0.336}{0.015}\right)^2.
\label{eq3:3.22}
\end{equation}
The total $\chi^2$ for all the BAO data sets thus can be written as
\begin{equation}
\chi^2_{BAO} = \chi^2_{\scriptscriptstyle BW} +
\chi^2_{\scriptscriptstyle SDSS} + \chi^2_{\scriptscriptstyle
6dFGS}.
\label{eq3:3.23}
\end{equation}
\begin{figure}[htb]
\centering
\begin{center}
\includegraphics[height=7.5cm]{Dv_z}
\caption{Plot of the distance-redshift relation from various BAO measurements $D_v (z)/r_s$. The gray area is the error propagation to $ 1- \sigma $ for the $\omega CDM$ model.}
\label{fig:Dvz}
\end{center}
\end{figure}
Figure \ref{fig:Dvz} shows the Hubble diagram corresponding to the data points for BAO and the model that best fits to the data ($\omega CDM$) with error propagation to $1- \sigma$.
\subsection{LSS Growth Rate}
The Large Scale Structure (LSS) of the universe is made up of huge empty, sheets, filaments, clusters of galaxies and superclusters, also known as cosmic web. These structures evolved from a homogeneous and isotropic universe and can in principle be treated theoretically as huge deviations from the mean density in the perturbation theory. We can define the linear matter density contrast as: $\delta (r,t)\equiv \delta \rho (r,t) / \rho (r,t)$, where $\rho (r,t)$ is the mean density of matter (background) and $\delta \rho (r,t)$ is the first order perturbation. The dynamics of the cosmic Hubble expansion H(t) is driven by the gravitational field of the mean matter density $\rho (r,t)$, while the density fluctuations $\delta \rho (r,t)$ produce an additional gravitational field. We can consider overdensed ($\delta \rho (r,t) > 0$) and underdensed regions ($\delta \rho (r,t) < 0$), e.g. super galaxy clusters and voids. In overdensed regions the gravitational field is stronger than in the cosmic average and therefore, due to this excess of self gravity, the overdensed region will expand more slowly than the average Hubble expansion. In an underdensed region the gravitational field is weaker than in the cosmic mean and therefore, the expansion is slow down less than in the cosmic mean. Overdense regions increase their density contrast over the course of time, while underdense regions decrease their density contrast. In both situations, $\vert \delta \vert$ increased with time. Then, the characterization of growth of the density perturbations can be made assuming the following relationship $\delta (r,t) = D(t) \delta_0 (r)$, where $D(t)$ is the linear Growth Factor Structure and $\delta_0 (r)$ is an arbitrary function of the spatial coordinate. Under the assumption that general relativity is the correct theory of gravity ($G_{eff}(a)=1$) \cite{2011JCAP...07..037N,2006astro.ph..5313U,2006PhRvD..74h4007S}, we use $D(a)$ in order to characterize the growth of structure, which is obtained numerically of the following differential equation:
\begin{equation}
\ddot{D}(a) + \left( \frac{3}{a} + \frac{\dot{H}(a)}{H(a)} \right) \dot{D}(a) - \frac{3}{2} \frac{\Omega_m}{a^5 H(a)^2} G_{eff}(a)D(a) = 0
\label{eq3:3.24}
\end{equation}
where dots denote differentiation with respect to the scale factor and we assume the initial conditions $D(0) = 0$ and $\dot{D}(0) = 1$ for the growing mode. the solution $\delta (r,t) = D(t) \delta_0 (r)$ indicates that in linear perturbation theory the spatial shape of the density fluctuations is frozen in comoving coordinates and only their amplitude increases.
\begin{figure}[htb]
\centering
\begin{center}
\includegraphics[height=7.5cm]{A_z}
\caption{Growth parameter A(z) using the best fit parameters for different cosmological models. The region in gray shows the propagation of error $1\sigma$ of $\sigma _8$ for $\omega CDM model$. }
\label{fig:Az}
\end{center}
\end{figure}
The linear Growth Factor Structure $D(t)$ of the amplitude follows a simple differential equation (\ref{eq3:3.24}). Through of the Growth Rate Data $f(a)\equiv a\dot{D}(a)/D(a)$ we get an observational stimate. We use the Growth Parameter $A(z)=f(z)\sigma_{8}(z)$ to constrain cosmological models through:
\begin{equation}
\chi^{2}_{GF} = \sum_{i=1}^n \frac{(A(z) - A_{obs}(z_i))^2}{\sigma_i^2}
\label{eq3:3.25}
\end{equation}
where $A_{obs}(z_i)$ is the observed structure factor and $\sigma_{8}(z)$ is rms mass fluctuations in $8h^{-1}$ Mpc spheres. The growth structure data used in this paper are obtained from the following projects: PSCz, 2dF, VVDS, SDSS, 6dF, 2MASS, BOSS and WiggleZ galaxy surveys (Table \ref{tab:Az}, appendix \ref{appendix-sec2}). As $\sigma_{8}(z)=\sigma_{8}^0D(z)/D(0)$ then we can use $\sigma_{8}^0$ as a free parameter. The observable Growth Parameter $A_{obs}(z)$ also accounts the Alcock-Paczynski effect in redshift-space distortions. Figure \ref{fig:Az} shows the results of analysis, where the gray area is the error propagation to $ 1- \sigma $ for $\omega CDM$ model using the constrain on $\sigma _8$. In this paper an independent constraint on $\sigma _8$ was performed for each cosmological model and whose best fit was the one for the $\Lambda CDM$ model, with a value of $\sigma _8 =0.742\pm 0.019$ (See Figure \ref{fig:Sigma3}). As complementary tests we use Mass Gas Fracction ($f_{gas}$), GRBs, Strong Gravitational Lensing and SNIa (Union2.1) (See appendix \ref{appendix-sec1}).
\section{Method and data analysis}
\subsection{Maximum likelihood $\mathcal{L}_{max}$ and confidence level $\sigma_i$.}
Maximum likelihood $\mathcal{L}_{max}$, is the procedure of finding the value of one or more parameters for a given statistic which makes the known likelihood distribution a maximum. The maximum likelihood estimate for the best fit parameters $\Theta_{i}^{m}$ is
\begin{equation}
\mathcal{L}_{max}(\Theta_{i}^{m})= exp \left[ -\frac{1}{2}{\chi_{min}^{2}(\Theta_{i}^{m})} \right]
\label{eq4:4.1}
\end{equation}
If $\mathcal{L}_{max}(\Theta_{i}^{m})$ has a Gaussian errors distribution, then
\begin{equation}
\chi_{min}^{2}(\Theta_{i}^{m})=-2 \ln \mathcal{L}_{max}(\Theta_{i}^{m}),
\label{eq4:4.2}
\end{equation}
which is our case \cite{Andrae:2010gh}. So, for our analysis:
\begin{equation}
\chi_{min}^{2}= \chi^{2}_{CMB} +\chi^{2}_{BAO}+\chi^{2}_{GF}+\chi^{2}_{SNIa}+\chi^{2}_{GRBs}+\chi^{2}_{GC}+\chi^{2}_{SGL}
\label{eq4:4.3}
\end{equation}
Table \ref{tab:par2} presents the main results of cosmological parameters derived from the free parameters constrained in this study with all observational data.
\begin{table*}[htb]
\centering
\begin{tiny}
\begin{tabular}{lccccc}
\hline
\hline
Par & $\Lambda CDM$ & $\omega CDM$ & \textit{CPL} & \textit{IDE} & \textit{EDE} \\
\hline
$H_0$ & $69.58 \pm 0.22$ & $ 69.46 \pm 0.21$ & $69.53 \pm 0.22$ & $68.73\pm 0.28$ & $68.51\pm 0.29$ \\
$t_0$ & $14.064 \pm0.044$ & $14.087 \pm 0.042$ & $14.073 \pm 0.045$ & $14.236 \pm 0.059$ & $14.283 \pm 0.0614$ \\
$\Omega_{r_0} $ & $8.613 \pm 0.054\times 10^{-5}$ & $8.651\pm 0.053\times 10^{-5}$ & $8.634\pm 0.056\times 10^{-5}$ & $8.835 \pm 0.073\times 10^{-5}$ & $8.893 \pm 0.076\times 10^{-5}$ \\
$\Omega_{\gamma _0} $ & $5.100 \pm 0.032\times 10^{-5}$ & $5.118 \pm 0.031\times 10^{-5}$ & $5.108 \pm 0.33\times 10^{-5}$ & $5.226 \pm 0.43\times 10^{-5}$ & $5.261 \pm 0.045\times 10^{-5}$ \\
$\Omega_{\nu _0}$ & $3.521 \pm0.022\times 10^{-5}$ & $3.533\pm 0.021\times 10^{-5}$ & $3.526\pm 0.022\times 10^{-5}$ & $3.608 \pm 0.029\times 10^{-5}$ & $3.632 \pm 0.031\times 10^{-5}$ \\
$\omega_{m_0}$ & $0.14007 \pm 0.00088$ & $0.14047 \pm 0.00085$ & $0.14005 \pm 0.00091$ & $0.1547 \pm 0.0013$ & $0.1529 \pm 0.0013$ \\
$\omega_{b_0}$ & $0.02178 \pm 0.00014$ & $0.02171 \pm 0.00013$ & $0.02175 \pm 0.00014$ & $0.02129 \pm 0.00017$ & $0.02112 \pm 0.00018$ \\
$\Omega_{X_0} $ & $0.7095 \pm 0.0054$ & $0.7126 \pm 0.0053$ & $0.7182 \pm 0.0063$ & $0.6693 \pm 0.0061$ & $0.6701 \pm 0.0068$ \\
$\rho_{cri_0}$ & $9.101 \pm 0.057\times 10^{-30}$ & $9.070\pm 0.055\times 10^{-30}$ & $9.088\pm 0.058\times 10^{-30}$ & $8.881\pm 0.073\times 10^{-30}$ & $8.823\pm 0.075\times 10^{-30}$\\
$\rho_{X _0}$ & $6.457 \pm 0.041\times 10^{-30}$ & $6.463\pm 0.039\times 10^{-30}$ & $6.527\pm 0.042\times 10^{-30}$ & $5.944\pm 0.049\times 10^{-30}$ & $5.912\pm 0.051\times 10^{-30}$ \\
$c_s$ & $0.44965 \pm 0.00056$ & $0.44994 \pm 0.00053$ & $0.44977 \pm 0.00057$ & $0.45179 \pm 0.00073$ & $0.45233 \pm 0.00075$ \\
$z_{drag}$ & $1019.15 \pm 0.39$ & $1019.02 \pm 0.37$ & $1019.08 \pm 0.39$ & $1019.11 \pm 0.49$ & $1018.66 \pm 0.51$ \\
$r_{drag}$ & $152.09 \pm 0.42$ & $152.04\pm 0.38$ & $152.13\pm 0.41$ & $146.78\pm 0.88$ & $148.14\pm 0.97$ \\
$z_{cmb}$ & $1092.52 \pm 0.13$ & $1092.66 \pm 0.12$ & $1092.56 \pm 0.13$ & $1094.61 \pm 0.17$ & $1094.68 \pm 0.18$ \\
\hline
\end{tabular}
\end{tiny}
\caption{Derived parameters for different cosmological dark energy models. We assume $\Omega_{b_0}=0.045$ \cite{Kirkman:2003uv} and $N_{eff}=3.04$ \cite{planckXVI} for all cosmological models. }
\label{tab:par2}
\end{table*}
\subsection{Fisher matrix, error propagation and uncertainties.}
The Fisher matrix is widely used in the analysis of the constraint of cosmological parameters from different observational data sets \cite{Albrecht:2009ct,2012JCAP...09..009W}. They encode the Gaussian uncertainties of the several parameters $\Theta_{i}^{m}$. Given the best fit $\chi_{min}^{2}$, the Fisher matrix can be calculeted as:
\begin{equation}
F_{ij}=\frac{1}{2}\frac{\partial ^2 \chi_{min}^{2}}{\partial p_i \partial p_j},
\label{eq4:4.4}
\end{equation}
where $p_i$ and $p_j$ are the set of free parameters in each cosmological model. In the extended form we have:
\begin{equation}
\left[ F \right] = \frac{1}{2}\left[
\begin{array}{cccccc}
\frac{\partial^2}{\partial p_1^2} & \frac{\partial^2}{\partial p_1 \partial p_2} & . & . & . & \frac{\partial^2}{\partial p_1 \partial p_n}\\
\frac{\partial^2}{\partial p_2 \partial p_1} & \frac{\partial^2}{\partial p_2^2} & . & . & . & \frac{\partial^2}{\partial p_2 \partial p_n}\\
. & . & . & & & .\\
. & . & & . & & .\\
. & . & & & . & .\\
\frac{\partial^2}{\partial p_n \partial p_1} & \frac{\partial^2}{\partial p_n \partial p_2} & . & . & . & \frac{\partial^2}{\partial p_n^2} \\
\end{array}\right] \chi_{min}^{2}(p_1, p_2,...,p_n ).
\label{eq4:4.5}
\end{equation}
where $\chi_{min}^{2}(p_1,p_2,...,p_n)=\chi_{min}^{2}(\Theta_{1}^{m},\Theta_{2}^{m},...,\Theta_{n}^{m})$, for the present analysis. The inverse of the Fisher matrix is the covariance matrix:
\begin{equation}
\left[ C_{cov} \right] = \left[ F \right]^{-1} = \left[
\begin{array}{cccccc}
\sigma^2_1 & \sigma_{12} & . & . & . & \sigma_{1n}\\
\sigma_{21} & \sigma^2_2 & . & . & . & \sigma_{2n}\\
. & . & . & & & .\\
. & . & & . & & .\\
. & . & & & . & .\\
\sigma_{n1} & \sigma_{n2} & . & . & . & \sigma^2_{n}\\
\end{array}\right],
\label{eq4:4.6}
\end{equation}
where $\sigma_{i}$ and $\sigma_{j}$ are the uncertainties associated with each cosmological parameters $p_i$ and $p_j$, to a level of statistical confidence of $1\sigma$. The uncertainties are obtained as $\sigma_i = \sqrt{Diag \left[ C_{cov} \right]_{ij} }$ . \\
\begin{figure}[htb]
\centering
\begin{center}
\includegraphics[height=4.5cm]{wCDM} \includegraphics[height=4.5cm]{wCDM_2} \includegraphics[height=4.5cm]{LCDM_2}\\
\includegraphics[height=4.5cm]{CPL} \includegraphics[height=4.5cm]{IDE} \includegraphics[height=4.5cm]{EDE}\\
\caption{Diagrams of statistical confidence, marginalizing different cosmological parameters at $1\sigma$, $2\sigma$ and $3\sigma$, with all observational tests (CMB+BAO+SNIa+GRBs+GF+GC+SGL)}
\label{fig:Sigma3}
\end{center}
\end{figure}
Table \ref{fig:Sigma3} presents statistical confidence diagrams obtained during this study, for each of the cosmological models of dark energy dynamics. In this paper we adopt the Figure of Merit Science Working Group (FoMSWG), proposed by the Joint Dark Energy Mission (JDEM), which it is a mission planned by NASA and Departamente of Energy (DOE) in USA for launch this year (2016)\footnote{\url{http://science.nasa.gov/missions/jdem/}} (It is still in the study phase). The main goal of this mission is to test the properties of dark energy following the guidelines of the Dark Energy Task Force (DEFT) \cite{2006astro.ph..9591A}, which was established by the Astronomy and Astrophysics Advisory Committee (AAAC) and the High Energy Physics Advisory Panel (HEPAP) as a joint sub-committee to advise the NASA, DOE and National Science Foundation (NSF) on future dark energy research \cite{2006astro.ph..9591A}. Within the main tasks proposed for this observational mission include: 1) Determine whether the accelerating expansion of the universe is consistent with a cosmological constant. 2) Measure as well as possible any sign of evolution with time of the density of dark energy. 3) Search for possible failure of general relativity comparing the effects of dark energy on the expansion with the effect of the same on the growth of structures at different scales. This work maintains the same spirit of the DEFT, especially in points a) and b), as shown by the results of the following sections.
\subsection{Statistical discrimination models.}
As mentioned before, we must give tight constraints on the
parameters of the several models to discern among them which one is the most favored by the observations.
To do that is not enough to calculate the best fit, we need to compare the models with each other
using an appropriate statistical criterion. In this paper we use the Akaike information criterion
(AIC; \cite{Akaike:74}) and the Bayesian information criterion (BIC; \cite{Schwarz:78}),
which allow to compare cosmological models with different degrees of freedom,
with respect to the observational evidence and the set of parameters \cite{Liddle:04}.
The AIC and BIC can be calculated as
\begin{equation}
AIC=-2 \ln \mathcal{L}_{max} +2k,
\label{eq4:4.7}
\end{equation}
\begin{equation}
BIC=-2 \ln \mathcal{L}_{max} + k \ln N,
\label{eq4:4.8}
\end{equation}
where $\mathcal{L}_{max}$ is the maximum likelihood of the model under consideration,
$k$ is the number of parameters. It is worthy to note that BIC
considers the number $N$ of data points used in the fit. Thus, it imposes
a strict penalty against extra parameters for any set of data $\ln N > 2$.
The prefered model is that which minimizes AIC and BIC.
However, the absolute values of them are not of interest,
only the relative values between the different models. Therefore
the ``strength of evidence'' can be characterized in the form $\Delta AIC=AIC_{i}-AIC_{min}$
$\Delta BIC=BIC_{i}-BIC_{min}$, where the subindex $i$ refers to value of $AIC$ ($BIC$)
for model $i$ and $AIC_{min}$ ($BIC_{min}$) is the minimum value of $AIC$ ($BIC$) among all the
models \cite{Burnham:03,Robert:95}. We give the judgements for both criterion
in Tables \ref{tab:AIC} and \ref{tab:BIC}
\begin{table}[!h]
\begin{center}
\begin{tabular}{cc}
\hline
$\Delta AIC$ & Level of Empirical Support For Model $i$\\
\hline
$0-2$ & Substantial \\
$4-7$ & Considerably Less \\
$>10$ & Essentially None \\
\hline
\end{tabular}
\end{center}
\caption[$\Delta AIC$ criterion]{$\Delta AIC$ criterion.\label{tab:AIC}}
\end{table}
\begin{table}[!h]
\begin{center}
\begin{tabular}{cccc}
\hline
$\Delta BIC$ & Evidence Against Model $i$\\
\hline
$0-2$ & Not Worth More Than A Bare Mention \\
$2-6$ & Positive \\
$6-10$ & Strong \\
$>10$ & Very Strong \\
\hline
\end{tabular}
\end{center}
\caption[$\Delta BIC$ criterion]{$\Delta BIC$ criterion.\label{tab:BIC}}
\end{table}
Thus, if we have a set of models of dark energy first we should estimate
the best fit $\chi^{2}$. Then we can apply the $AIC$ and $BIC$ to identify
which model is the prefered one by the observations hence
to the nature of dark energy. We also use the $\chi_{red}^{2}=\chi_{min}^{2}/\nu$
as criterion of good-fitting, for which the degrees of freedom are given by $\nu = N - k$,
where the best model is that one whose value of $\chi_{red}^{2}$ is closet to one.\\
\begin{table}[htb]
\centering
\begin{tabular}{lccccccc}
\hline
Model & k & $\chi_{min}^{2}$ & $\chi_{red}^{2}$ & $AIC$ & $\Delta AIC$ & $BIC$ & $\Delta BIC$ \\
\hline
\hline
$\omega$ CDM & 5 & 809.860 & 1.106 & 819.860 & 0.00 & 842.872 & 0.00 \\
$\Lambda$ CDM & 4 & 821.812 & 1.121 & 829.812 & 9.95 & 848.222 & 5.35 \\
CPL & 6 & 810.738 & 1.109 & 822.738 & 2.88 & 850.353 & 7.48 \\
IDE & 6 & 888.975 & 1.216 & 900.975 & 81.12 & 928.591 & 85.72 \\
EDE & 6 & 893.022 & 1.221 & 905.022 & 85.16 & 932.637 & 89.801 \\
\hline
\end{tabular}
\caption{Comparison of the different cosmological models with the
$\Delta AIC$ y $\Delta BIC$ criterions using all complementary tests ($d_{A,c}+f_{gas}+BAO+CMB+SNIa+SL$).
The $\omega$CDM model is the most favored model by these criterios.
\label{tab:AIC/BIC}}
\end{table}
Table \ref{tab:AIC/BIC} shows the values of $\chi^{2}_{min}$, AIC and BIC for the DE models from all cosmological tests. Notice that the $\omega CDM$ model gives the lowest value of AIC and BIC, therefore, we concluded that is the cosmological model most favored by observational data, as it can also be analyzed in Table \ref{tab:par2}. The $\Delta AIC$ and $\Delta BIC$ values for the other models are measured with respect to $\omega CDM$. Following \cite{Shi:2012ma}, we can classify the DE models in four groups: group 1, the model that shows a substantial level of empirical support and not worth more than a bare mention of evidence against to $\omega CDM$; group 2, considerably less level of empirical support and some positive evidence against to $\Lambda CDM$; group 3, substantial level of empirical support and some strong evidence against to CPL model; group 4, essentially none substantial level of empirical support and very strong evidence against to IDE and EDE models.
\section{Deceleration parameter}
In this section we introduce of the most popular cosmological tests, which can help us to uncover some signs of the nature of dynamic dark energy. \\
It is natural to describe the kinematics of the cosmic expansion through the Hubble parameter $H(t)$ an and its dependence on time, i.e. the deceleration parameter $q(t)$. Following to \cite{2012PhyU...55A...2B} the scale factor $a(t)$ expands into a Taylor series around the current time ($t_0$) as:
\begin{equation}
\frac{a(t)}{a(t_0)} = 1 + \frac{H_0}{1!} \left[ t - t_0 \right] - \frac{q_0}{2!} H^2_0 \left[ t - t_0 \right]^2 + \frac{j_0}{3!} H^3_0 \left[ t - t_0 \right]^3 + \frac{s_0}{4!} H^4_0 \left[ t - t_0 \right]^4 + ... ,
\label{eq5:5.1}
\end{equation}
where in general we can have a kinematic description of the cosmic expasion through the set of parameters:
\begin{equation}
H(t)\equiv\frac{1}{a}\dfrac{da}{dt}; \quad q(t)\equiv-\frac{1}{a}\frac{d^2 a}{dt^2} H(t)^{-2}; \quad j(t)\equiv\frac{1}{a}\frac{d^3 a}{dt^3} H(t)^{-3}; \quad s(t)\equiv\frac{1}{a}\frac{d^4 a}{dt^4} H(t)^{-4}.
\label{eq5:5.2}
\end{equation}
To characterize whether the universe is currently accelerated or decelerated, the history of expansion is fit through deceleration parameter $q(z)\equiv -\ddot{a}(z)/a(z)H(z)^2$. If $q(z)>0$, $\ddot{a}(z)<0$; then the expansion decelerate, as expected due to gravity (i.e. dark matter, baryonic matter, radiation). The discovery that the universe is currently accelerating, already has about one decade and a half old \cite{1999ApJ...517..565P} \cite{1998AJ....116.1009R}. The simplest explanation for the accelerating universe is the cosmological constant $\Lambda$, however, there is still no compelling theoretical explanation based on physical foreground and not only phenomenological. To take account information about the dynamics of the expansion we to use (\ref{eq2:2.2}) and (\ref{eq5:5.2}), then we obtain
\begin{equation}
q(z) = \frac{1+z}{E(z)} \frac{dE(z)}{dz},
\label{eq5:5.3}
\end{equation}
which directly depends on the cosmological model and its matter-energy content. In general, if $\Omega_X \neq 0$ is sufficiently large (i.e. $\Omega_X >\Omega_m $), then $q(z)<0$ and $\ddot{a}(z)>0$, which corresponds to an expansion accelerated the universe as shown by observational data at present, which also indicates a cosmological constant different from zero (see Table \ref{tab:par2}). If the acceleration of the universe is driven by a new fluid, then it is important to identify signs to determine if the energy density of the fluid is constant or dynamic.\\
\begin{figure}[htb]
\centering
\begin{center}
\includegraphics[height=7.5cm]{q_z_2}
\caption{Desaceleration parameter vs redshift using only BAO data, where the gray region corresponds to the best fit to $1\sigma$ for $\omega CDM$ model, with error propagation only on $\Omega_m =0.32 \pm 0.16$. It is shown the transition decelerated-accelerated ($q(z_t)=0$) and the current value of ($q_0$) ($\Lambda CDM$ ($z_t\sim 0.92$, $q_0=-0.67$), $\omega CDM$ ($z_t\sim 0.86$, $q_0=-0.70$), CPL ($z_t\sim 0.85$, $q_0=-0.48$), IDE ($z_t\sim 1.06$, $q_0=-0.72$), EDE ($z_t\sim 0.73$, $q_0=-0.68$)). Note the strange behavior of the deceleration parameter to later times for models of dynamical dark energy (CPL, IDE, EDE).}
\label{fig:qz}
\end{center}
\end{figure}
\begin{table}[htb]
\centering
\begin{tabular}{lcl}
\hline
Model & $\chi_{min}^{2}$ & Parameters \\
\hline
\hline
$\Lambda$ CDM & 2.69 & h=0.5822, $\Omega_m$=0.2232, $\Omega_k$=-0.0107 \\
$\omega$ CDM & 2.63 & h=0.9197, $\Omega_m$=0.3203, $\Omega_k$=-0.3347, $\omega$=-0.9048\\
CPL & 2.68 & h=0.6742, $\Omega_m$=0.2839, $\Omega_k$=-0.0425, $\omega_a$=-1.0254, $\omega_0$=-0.8935\\
IDE & 2.59 & h=0.5931, $\Omega_m$=0.4626, $\Omega_k$=-1.1601, $\omega_x$=-0.7085, $\delta$=0.1069 \\
EDE & 2.61 & h=0.8321, $\Omega_m$=0.3807, $\Omega_k$=-0.2711, $\omega_0$=-0.9875, $\Omega_e$= -0.1383\\
\hline
\end{tabular}
\caption{The best fit values for the free parameters using the only BAO data set.
\label{tab:BAOqz}}
\end{table}
Figure \ref{fig:qz} shows the plot of the deceleration parameter $q(z)$ using only data from BAO (See Table \ref{tab:BAOqz}). As expected, the models studied give $q(z)<0$ at late times and $q(z)>0$ at earlier epoch, which means that the history of the expansion is slowed down in the past and speeded up at present. All cosmological models presents a redshift of transition ($z_t$) between the two periods (see Figure \ref{fig:qz}), however, all models of dynamical dark energy present an interesting behavior of slowing down of acceleration at low redshift (late times) using only data from BAO, which can be characterized through the change of sign of the parameter $j(z)$ (CPL: $j(z_{low}) \rightarrow 0$, when $z_{low} \sim 0.11$; IDE: $j(z_{low})\rightarrow 0$, when $z_{low} \sim 0.00$; EDE: $j(z_{low}) \rightarrow 0$, when $z_{low} \sim 0.05$). We can interpret j(z) as the slope at each point of q(z), which indicates a change in acceleration. This result is consistent with the one presented by J. Barrow, R. Bean and J. Magueijo \cite{2000MNRAS.316L..41B}, who raises the possibility of a scenario consistent with the current accelerating universe and does not involve an eternal accelerated expansion. In \cite{2012PhyU...55A...2B} extensive analysis it is made of this possibility. This can be also a clear behavior of dynamical dark energy at low redshift in this models with variation of the density of dark energy over time.
\section{Summary and discussion}
In this paper we perform the study of the history of the of cosmic expansion through the H(z), q(z) and j(z) parameters with data from SDSS, WiggleZ, BOSS and CMASS, using BAO distance ratio scale $r_{s}/D_v (z)$. We find new evidence presented in previous work, showing peculiar behavior of the deceleration parameter q(z) in later times ($z_{low}< 0.5$), which indicates through the change of sign of the parameter $j(z)$ (+$\rightarrow$ -), that the universe could decelerate in the near future (Figure \ref{fig:qz}). This phenomenon raises the possibility that an accelerated expansion does not imply the eternal accelerated expansion, even in the presence of dark energy. This anomalous behavior is present only in models with dark energy density varying with the time. This behavior is possibly due to the dynamics of the dark energy density in this class of cosmological models, which in principle can be a sign of evolution with time of the density of dark energy and could distinguish from a constant density of dark energy. However, Figure \ref{fig:qz} shows that $\Lambda$ still can not be ruled out in this analysis, since all models are in the region of error propagation to $ 1\sigma $ of $\omega CDM$ and would have to make a deeper analysis , for example by including other observational evidence for better restriction, which will be presented in next paper.\\
In the study of the comparing different cosmological models of dark energy we use the most recent observational data. This data includes statistical combination of CMB (Plank 13 + WMAP 9), BAO (SDSS, WiggleZ, BOSS and CMASS), A(z) (PSCz, 2dF, VVDS, SDSS, 6dF, 2MASS, BOSS and WiggleZ), SNIa (Union 2.1), GRBs, SGL and $f_{gas}$. By using these data sets, we obtained the best-fit parameters for different cosmological models. We use the information criteria including the $\Delta AIC$ and the $\Delta BIC$, to compare different models and to see which is the most favored by current observational data. Our analysis shows that $\omega CDM$ dark energy model is preferred by Bayesian and Akaike criterion. By first time we report that observational data are in favor of the cosmological model $\omega CDM$ and very close to being in favor of models of dynamical dark energy as Chevalier-Polarski-Linder model, which is better than $\Lambda CDM$ for $\Delta AIC$ (See Table \ref{tab:AIC/BIC}). In Table \ref{tab:wCDM} and Figure \ref{fig:Sigma3} we can see that $\Lambda CDM$ model ($\omega=-1$) is excluded to a level of statistical confidence of $3\sigma$ ($\omega =-1.091^{+0.051}_{-0.059}$) combining all observational data for the present analysis and also from all other cosmological models. We include the parameter information that are commonly used in the characterization of models of the universe, with error propagation $1\ sigma$ from the constraint of free parameters for each cosmological models of dark energy presented in Table \ref{tab:par2}. \\
Assuming that general relativity is the correct theory of gravity, we used measures of the growth parameter $A(z)$ (See Table Appendix \ref{appendix-sec2}) to put independent constraints on variance in mass flucuaciones $\sigma _8 =0.742\pm 0.019$. We are able to break the degeneracy between $\Omega _m$ and $\sigma _8$, through the use of the equation (\ref{eq3:3.24}), to thereby achieve a good independent constraint. Constrain as best as possible this parameter is very important in cosmology because through its the normalization of the power spectrum of the matter is done.
\acknowledgments
A. Bonilla wish to express their gratitude to Universidad Distrital FJDC for the academic support and funding. A. Bonilla and J. Garcia wishes to thank to the guest lecturers for his comments and suggestions to this work within the framework of \textit{I Workshop on Current Challenges in Cosmology: Inflation and the Origin of the CMB Anomalies} \footnote{{http://cosmology.univalle.edu.co}}, \textit{I Simposio Andino de Astrof\'isica Relativista} \footnote{https://eventos.redclara.net/indico/event/503/} and the professors Yehuda Hoffman and Stefan Gottloeber within the framework of academic event \textit{Escuela Andina de Cosmolog\'ia} \footnote{http://forero.github.io/AndeanCosmologySchool/}. A. Bonilla and J. Garcia also wish give special thanks to \textit{Dr. Florian Beutler} for their appropriate comments about the most recent BAO data and \textit{Dr. Santiago Vargas} (OAN, UNAL) for their cordial review of "systematic errors" in the preparation of this work.
|
\section{Introduction}
The classical theory of General relativity (GR) admits spacetime singularities, see~\cite{Hawking:1973uf,Wald:1984rg} and \cite{Penrose:1964wq}. Such a singularity manifests itself, at short distances, as a blackhole and, at small time scales, as the cosmological Big Bang singularity. In particular, the latter has a detrimental impact on the construction of an ultraviolet (UV) complete theory of primordial inflation~\cite{Borde:1993xh}.
A major shortcoming of GR is that it may never evade the {\it shadow} of the cosmological singularity problem, unless one violates a relevant energy condition. The appropriate energy condition depends on the nature of the geodesic congruences involved. For null rays, one requires a violation of the \emph{Null Energy Condition} (NEC), i.e. $T_{\mu\nu}k^{\mu}k^{\nu}\geq 0$, where $\mu,~\nu$ run from $0,~1,~2,~3$ in 4-dimensions, $k^{\mu}$ is a null tangent vector, and $T_{\mu\nu}$ is the energy momentum tensor.
For timelike geodesics, one has to violate the \emph{Weak Energy Conditions} (WEC), i.e. $T_{\mu\nu}\xi^{\mu}\xi^{\nu}\geq 0$, where $\xi^{\mu}$ is
timelike. The famous Hawking-Penrose Singularity Theorems~\cite{Penrose:1964wq} were derived in the context of null and timelike versions of the Raychaudhuri equation~\cite{Raychaudhuri:1953yv}~\footnote{The definition of a singularity which forms the basis of the Hawking-Penrose theorems concerns geodesic completeness and it is this notion of a spacetime containing causal (timelike or null) geodesic congruences which focus, (or indeed \emph{defocus}) that interests us here. If a spacetime is causally incomplete, then a freely falling photon passing along this geodesic will cease to exist in a finite `time' (affine parameter). As such, we can reasonably call this a \emph{singular} spacetime, see \cite{Wald:1984rg}, \cite{Geroch:1968ut}, and \cite{Albareti:2014dxa}.}
At time scales close to the Planck scale, i.e. $M_P\sim 2.4\times 10^{18}$~GeV, one would expect the Einstein-Hilbert action to be modified by higher curvature corrections, i.e. higher derivative modifications made up of the Ricci scalar, Ricci tensor and the Riemann/Weyl terms. However, higher derivative theories are beset by classical and quantum instabilities. For instance, quadratic curvature gravity, which is renormalizable, succeeds in improving the UV behaviour, but at the expense of introducing
{\it ghosts} in the spin-2 component of the graviton propagator~\cite{Stelle}. The presence of ghosts is not a welcome sign as it renders the vacuum unstable at both classical
and at quantum level.
Recently, the issue of ghosts in the spin-2 component of quadratic curvature gravity has been addressed in ~\cite{Biswas:2011ar}, where it was found that one requires an infinite number of covariant derivatives acting on the curvature to maintain general covariance and avoid the addition of ghosts at a perturbative level. Although the graviton propagator is modified by these infinite derivatives, it is still possible to retain the original massless graviton degrees of freedom, if one requires that \emph{no new poles} are introduced in either the spin-0 or spin-2 component of the graviton propagator.
This can be achieved if the propagator is modified by an {\it exponent of an entire function} of the d'Alembertian (denoted as $\Box= g^{\mu\nu}\nabla_{\mu}\nabla_{\nu}$), which contains no roots, by construction~\cite{Biswas:2011ar,Tomboulis,Tseytlin:1995uq,Siegel,Biswas:2005qr,Modesto}. For a further {\it covariant generalisation}, see~\cite{Edholm:2016hbt}.
Typically, such an exponential function in the propagator weakens the classical and quantum effects of gravity in the UV. For instance, it was found, at the linearised limit, that there are no blackhole-like singularities in a static spacetime~\cite{Biswas:2011ar}, nor a time-dependent collapse of matter~\cite{Frolov:2015bia}. It was also observed that the gravitational entropy~\cite{Wald-1},
of a static and axisymmetric metric receives zero contribution from the infinite derivative sector of the action, when no additional scalar propagating modes are introduced. In this case, the gravitational entropy is strictly given by the {\it area-law},
arising solely from the contribution of the Einstein-Hilbert action~\cite{Conroy:2015wfa}. In the quantum context, the interplay between an
{\it exponentially enhanced} vertex operator and an {\it exponentially suppressed } graviton propagator ensures that, beyond 1-loop, the theory becomes finite~\cite{Tomboulis,Modesto,Talaganis:2014ida}.
Furthermore, it was conjectured that the high energy scattering of gravitons in infinite derivative theories lead to a finite amplitude, which does not become arbitrarily large for given external momenta~\cite{Talaganis:2016ovm}.
A different and rather intriguing motivation to investigate infinite derivative theories of gravity
arises from string theory (ST)~\cite{Polchinski:1998rr} and string field theory (SFT) models, which appear in the context of
noncommutative geometry \& SFT~\cite{Witten:1985cc}, for a review, see \cite{Siegel:1988yz}, and various {\it toy model}s of SFT such as $p$-adic strings~\cite{Freund:1987kt,Freund:1987ck,Brekke:1988dg,Frampton:1988kr}, zeta strings~\cite{Dragovich:2007wb}, and strings quantized on a random lattice~\cite{Douglas:1989ve,Gross:1989ni,Brezin:1990rb,Ghoshal:2006te,Biswas:2004qu}. A key feature of these models is the presence of an {\it infinite series of higher-derivative} series in $\alpha'$, the inverse of string tension, incorporating non-locality in the form of an {\it exponential kinetic} correction. Similar infinite-derivative modifications have also been argued to arise in the asymptotic safety approach to quantum gravity~\cite{Krasnov:2006du,Krasnov:2007uu}, see for a review~\cite{Ambjorn:2012jv}.
Note that in GR, even in an
{\it asymptotically-flat Minkowski spacetime}, if we perturb matter and curvature while satisfying the NEC, the null congruences will {\it always} converge, in a finite time. The same holds true for timelike geodesics congruences. However, as null rays more readily converge than their timelike counterparts in a geometrically-flat spacetime, it is preferable for our aims to analyse the defocusing of null geodesic congruences rather than timelike. See \cite{Albareti:2014dxa} and \cite{Geroch:1968ut} for details. The prime question is then: \emph{Can null rays ``defocus'' in an infinite derivative theory of gravity (IDG), which is also ghost-free?}
The aim of this paper is to study the time-dependent singularity problem within the ghost-free, infinite derivative theory of gravity proposed in
Refs.~\cite{Biswas:2005qr,Biswas:2011ar}, where a non-singular bouncing solution for an infinite derivative equation of motion was obtained through an {\it Ansatz}-led approach. In this respect our current analysis is quite different from earlier investigations in the context of {\it IDG} theories~\cite{Biswas:2005qr}, and \cite{Conroy:2014dja}. Indeed, our main focus here is to go beyond a specific Ansatz-dependent background solution, and investigate under what
generic conditions an {\it IDG} theory can potentially lead to geodesic completeness, for time-dependent scenarios.
In Section \ref{RC}, we discuss the Raychaudhuri equation in GR, before briefly introducing IDG and its relevant properties, such as dynamical degrees of freedom, see Section \ref{IDG}. In Section \ref{DNC}, we discuss the defocusing of null congruences for ghost-free, infinite derivative theories of gravity, while comparing this behaviour with a well-known finite, quadratic curvature model of inflation. In the final section, we conclude our main results.
\section{Raychaudhuri's equation in General Relativity}\label{RC}
Let us begin our discussion within GR, where the dynamics of null rays can be understood in a {\it model independent} way by studying
the Raychaudhuri Equation (RE) for null geodesic congruences, such that $k^{\mu}k_{\mu}=0$, where $k_{\mu}$ is a four vector tangential to the null geodesic congruence, the metric signature is mostly positive, i.e. $(-,+,+,+)$, and $\theta$ is the expansion parameter, which is defined by $\theta= \nabla_{\mu}k^{\mu}$.
In the simplest case, if we consider the congruence of null rays to be orthogonal to the hypersurface, then the \emph{twist} tensors vanish, and
we may also neglect the shear tensor, which is
purely spatial, thus making a positive contribution to the r.h.s. of the RE. Thus, we have:
\begin{equation}\label{eq1}
\frac{d\theta}{d\lambda}+\frac{1}{2}\theta^{2}\le-R_{\mu\nu}k^{\mu}k^{\nu},
\end{equation}
%
where $\lambda$ is the affine parameter, and $R_{\mu\nu}$
is the Ricci tensor, see~\cite{Wald:1984rg}. In GR, the Einstein equation yields, $$G_{\mu\nu}=\kappa T_{\mu\nu},$$ where $\kappa =8\pi G=M_{P}^{-2}$, which implies $R_{\mu\nu}=\kappa( T_{\mu\nu}-\frac{1}{2} g_{\mu\nu} T)$.
By contracting with the vector fields $k^{\mu}k^{\nu}$ yields $R_{\mu\nu}k^{\mu}k^{\nu}=\kappa T_{\mu\nu}k^{\mu}k^{\nu}$, and demanding that the NEC is always satisfied, $T_{\mu\nu}k^{\mu}k^{\nu}\geq 0$,
we obtain the \emph{null convergence condition} (NCC):
\begin{equation}\label{eq2}
R_{\mu\nu}k^{\mu}k^{\nu}\geq 0,\qquad
\frac{d\theta}{d\lambda}+\frac{1}{2}\theta^{2}\le0
\end{equation}
The above equations suggest that the converging null geodesics cannot start to diverge before meeting the
{\it origin of coordinates}, or in other words converging null rays must meet the spacetime singularity
in a finite time within GR, when the NEC is satisfied.
In a geometrically-flat spacetime, the obeyence of the convergence condition \eqref{eq2} results in the formation of a \emph{closed trapped surface} and subsequently, a singular spacetime. A closed trapped surface is formed when the ingoing and outgoing expansions of these null rays are negative, i.e. $\nabla_\mu k^\mu<0$. However, in the case of a closed Universe in the $k=+1$ frame, the existence of closed trapped surfaces does not necessarily imply a singularity. See, \cite{Ellis:2003mb} for a full exposition.
Our aim in this paper is to use the power of the RE to show that {\it IDG}
can indeed yield the defocusing of null geodesic congruences, without violating the NEC, nor introducing ghosts or tachyons, in a geometrically-flat spacetime.
\section{ Infinite derivative gravity}\label{IDG}
We shall now analyse how the RE is modified in the framework of quadratic curvature {\it IDG}.
First, let us briefly recall the {\it IDG} action we are going to investigate, see~\cite{Biswas:2011ar}
\begin{eqnarray}
\label{action}
S=\frac{1}{2}\int d^{4}x\sqrt{-g}\bigl(M_{P}^{2}R+
R{\cal F}_{1}(\bar \Box)R+R^{\mu\nu}{\cal F}_{2}(\bar\Box)R_{\mu\nu}+ R^{\mu\nu\lambda\sigma}{\cal F}_{3}(\bar\Box)R_{\mu\nu\lambda\sigma}\bigr).
\end{eqnarray}
Note that $\bar\Box = \Box/{\cal M}^2$ where ${\cal M} < M_P$ is the scale of modification of gravity in the UV. Current short distance tests of gravity, where
there is no departure from Newtonian $1/r$ fall of gravity up to $5\times 10^{-6}$~m~\cite{Kapner}, place a rather mild limit on this scale, ${\cal M}\geq 10^{-2}$~eV~\cite{Edholm:2016hbt}. The non-local functions are defined by ${\cal F}_{i}({\bar \Box})\equiv \sum_{n=0}^{\infty}f_{i_n}(\Box/{\cal M}^2)^n$, where the coefficients $f_{i_n}$ are fixed by the {\it ghost-free} condition, see ~\cite{Biswas:2005qr,Biswas:2013cha}~\footnote{Ref.~\cite{Biswas:2013kla} demonstrates how to recover various limits of action Eq.~(\ref{action}), at the level of graviton propagator,
such as Weyl gravity, Gauss-Bonnet Gravity, Massive gravity, and scalar tensor theory. This paper also classifies which theory can be made
ghost free and which cannot.}.
In this paper, we will concentrate on linear perturbations around the Minkowski spacetime such that $$g_{\mu\nu}=\eta_{\mu\nu}+h_{\mu\nu},$$ where $\eta_{\mu\nu}$ is the Minkowski metric and $h_{\mu\nu}\equiv \delta g_{\mu\nu}$ is the metric tensor variation.
The full equations of motion can be found in~\cite{Biswas:2013cha}. At the linearised limit,
the linearised equations of motion can be written in the following concise form
\[
\kappa T_{\mu\nu} = a(\bar\Box)R^{(L)}_{\mu\nu}-\frac{1}{2}\eta_{\mu\nu}c(\bar\Box)R^{(L)}-\frac{f(\bar\Box)}{2}\partial_{\mu}\partial_{\nu}R^{(L)}\,.
\label{eomred}
\]
where $\Box=\eta^{\mu\nu}\partial_\mu \partial_\nu$, $L$ stands for linearised quantities, and
\begin{eqnarray}
&a(\bar \Box)\equiv1+M_{P}^{-2}\left({\cal F}_{2}(\bar\Box)+2{\cal F}_{3}(\bar\Box)\right)\Box
\nonumber
\nonumber\\&
c(\bar\Box)\equiv1+M_{P}^{-2}\left(-4{\cal F}_{1}(\bar\Box)-{\cal F}_{2}(\bar\Box)+\frac{2}{3}{\cal F}_{3}(\bar\Box)\right)\Box
\nonumber\\&
f(\bar\Box)\equiv M_{P}^{-2}\left(4{\cal F}_{1}(\bar\Box)+2{\cal F}_{2}(\bar\Box)+\frac{4}{3}{\cal F}_{3}(\bar\Box)\right),
\label{abc}
\end{eqnarray}
satisfying the following relations~\cite{Biswas:2011ar}
\begin{eqnarray}
-a(\bar\Box)+c(\bar\Box)+f(\bar\Box)\Box=0\,.
\end{eqnarray}
Taking the trace of the above equation Eq.~(\ref{eomred}) yields:
\[\label{trace}
\kappa T=\frac{1}{2}\bigl(a(\bar\Box)-3c(\bar\Box)\bigr)R^{(L)}
.\]
We see here that the trace equation describes the scalar propagating mode, or precisely the spin-0 component of the graviton propagator~\footnote{For detailed computation of finding the spin projection operators and the graviton propagator for an {\it IDG}, see ~\cite{Biswas:2011ar,Biswas:2013kla}.}. This can be seen by studying the graviton propagator for the action Eq.\eqref{action}, by decomposing it in terms of true dynamical degrees of freedom around Minkowski spacetime:
\[
\label{MinkProp4}
\Pi(-k^{2})=\frac{\mathcal{P}^{2}}{k^{2}a(-k^{2})}+\frac{\mathcal{P}_{s}^{0}}{k^{2}\left(a(-k^{2})-3c(-k^{2})\right)}
,\]
where we have suppressed the spacetime indices for clarity. To ensure freedom from ghosts, both $a$ and $c$ must be {\it exponents of entire functions}. In particular, $a$ must not contain any roots in order to avoid the Weyl ghost~\cite{Biswas:2013kla}, while $c$ may contain at most one root, which would yield a scalar tensor theory of gravity. The addition of further poles to the function $c$ will necessarily be ghost-like or tachyonic, see \cite{Biswas:2005qr}, for further details. Moreover, the GR propagator must be recovered in the infrared limit as $k\rightarrow 0$, with $a(-k^2)\rightarrow1$ and $c(-k^2)\rightarrow 1$. For the choice $a=c$, we have
\begin{equation}
\label{MinkPropAC}
\Pi(-k^{2})=\frac{1}{a(-k^{2})}\biggl(\frac{\mathcal{P}^{2}}{k^2}-\frac{\mathcal{P}_{s}^{0}}{2k^2}\biggr).
\end{equation}
which simply modulates the physical graviton propagator by an overall factor of $\sim 1/a(-k^2)$.
\section{Defocusing conditions of Null Congruences for Infinite Derivative Gravity}\label{DNC}
Note that our action Eq.\eqref{action} will modify the RE, since $\kappa T_{\mu\nu}$ is modified, see Eq.\eqref{eomred}, and it is more insightful to first study the linearised limit. By contracting the energy momentum tensor
with the congruences of null geodesics, $k^\mu$, we obtain
\[
\label{defocusgen}
R^{(L)}_{\mu\nu}k^{\mu}k^{\nu}=\frac{1}{a(\bar\Box)}\biggl[\kappa T_{\mu\nu}k^{\mu}k^{\nu}+\frac{1}{2}k^{\mu}k^{\nu}f(\bar\Box)\partial_{\mu}\partial_{\nu}R^{(L)}\biggr]
, \]
where the NCC requires $R^{(L)}_{\mu\nu}k^{\mu}k^{\nu}\geq0$.
Now let us study the simplest case, when the curvature is solely dependent on the cosmic time $t$, so that the D'Alembertian operator is given simply by $\Box=-\partial_t^2$. The contribution of gravity to the RE is then:
\[
\label{defocus2}
R^{(L)}_{\mu\nu}k^{\mu}k^{\nu}=\frac{1}{a(\bar\Box)}\biggl[\kappa T_{\mu\nu}k^{\mu}k^{\nu}-\frac{(k^{0})^{2}}{2}f(\bar\Box)\Box R^{(L)}\biggr]
. \]
To preserve the NEC, we must have $T_{\mu\nu}k^\mu k^\nu \geq 0$, whereas for the null rays to diverge, we require $R_{\mu\nu}k^\mu k^\nu<0$. Furthermore, from the propagator Eq.\eqref{propR2}, we require $a({\bar \Box})$ acting upon the curvature to be positive so as to avoid negative residues in the spin-2 component - a phenomenon known as the \emph{Weyl ghost}, i.e. $a({\bar \Box})R^{(L)}>0$.\\
\noindent
Taking all this into account, we obtain the \emph{minimal defocusing condition} for an {\it IDG} theory, which is independent of a background solution, and would yield a geodesically past-complete trajectory for a spatially-flat, homogeneous and isotropic background:
\[
\label{defocus3}
\frac{f(\bar\Box)\Box}{a(\bar\Box)}R^{(L)}> 0\Rightarrow \frac{a(\bar\Box)-c(\bar\Box)}{a(\bar\Box)}R^{(L)}> 0
,\]
with $T_{\mu\nu}k^\mu k^\nu=0$. From this defocusing condition, we may draw $3$ important conclusions:\\
\begin{enumerate}
\item{$a(\bar \Box)= c(\bar\Box)$: From Eq.\eqref{defocusgen}, defocusing of null geodesics can happen {\it only} when $a(\bar\Box)$ acting upon curvature is negative, which comes at the expense of the Weyl ghost~\cite{Biswas:2013kla}.}\\
\item{$a(\bar \Box)\neq c(\bar\Box)$: Since the curvature is always positive, we need a departure from the purely massless mode of the graviton propagator
in Eq.\eqref{MinkPropAC}. This condition tells us that in order for the null rays to defocus - a minimum requirement of a singularity-free theory of gravity - one requires an additional root in the spin-0 component of the graviton propagator. As $a(\bar\Box)$ does not introduce any new poles, the spin-2 component of the graviton
propagator remains massless. As such, one additional scalar degree of freedom must propagate in the spacetime besides the massless graviton, if we wish to satisfy the defocusing condition.}\\
\item{By inspection of Eq.~(\ref{abc}), it is possible to ``switch off'' any one of the form factors ${\cal F}_1,{\cal F}_2$ or ${\cal F}_3$ that make up the function $a({\bar \Box})$ and $c({\bar \Box})$, without altering the unitarity of the theory, see \cite{Biswas:2016etb,Biswas:2016egy}. The simplest choice is actually
${\cal F}_2=0$. In a conformally flat background such as in Minkowski space, de Sitter or indeed Freedman-Robertson-Walker metric, the Weyl tensor vanishes
in the background. Therefore, in this case, the action non-linear Eq.\eqref{action} reduces to the following form:
\begin{eqnarray}
\label{action1}
S=\frac{1}{2}\int d^{4}x\sqrt{-g}\bigl[M_{P}^{2}R+ R{\cal F}_{1}(\bar \Box)R\bigr]\,.
\end{eqnarray}
}
\end{enumerate}
Indeed, it is reassuring to note that from the most general covariant {\it IDG} action, Eq.\eqref{action}, we can deduce the simplest action of gravity
which has the potential to resolve the cosmological singularity problem, as given by Eq.~(\ref{action1}). Such a reduced action has been studied in~\cite{Biswas:2005qr}, where the authors found an {\it exact} non-singular bouncing solution for a homogeneous and istrotropic background for a spatially-flat geometry. However, one must emphasise that the solution was indeed found based on a given {\it Ansatz}, which was invoked in order to solve the full non-linear equations of motion~\footnote{Such an {\it Ansatz}, taking the form $\Box R = c_1R+c_2$, where $c_{1,2}$ are constants and
$R$ is the Ricci-scalar, was further verified in \cite{Barnaby:2007ve,Biswas:2006bs}. The sub-and super Hubble perturbations were also studied around the given bounce solution, and no instabilities were found~\cite{Biswas:2010zk}. }.
However, in the present work, our conclusion is deduced purely from geometric considerations of geodesics of null rays by analysing the RE.
Now, the relevant \emph{ghost-free condition} of the theory can be derived from the trace equation Eq.\eqref{trace}, where the functions $a(\bar\Box)$ and $c(\bar\Box)$ are recast in terms of
another {\it exponent of an entire function} $\widetilde a(\bar\Box)$, containing no roots, by
\[
\label{GF1}
c(\bar\Box)=\frac{a(\bar\Box)}{3}\left[1+2\left(1-\alpha M_{P}^{-2}\Box\right){\widetilde a}(\bar\Box)\right]
.\]
Here, $\alpha$ is a constant, which a Taylor expansion of the trace equation Eq.\eqref{trace}, in conjunction with Eq.\eqref{abc}, reveals to be
\begin{equation}
\alpha=6f_{1_0}+2f_{2_0}-\frac{M_P^2}{{\cal M}^2}\,.
\end{equation}
The
new graviton propagator can now be deduced by substituting Eq.\eqref{GF1} into Eq.\eqref{MinkProp4}, and subsequently decomposing the scalar propagating mode into partial fractions.
\[
\label{Minkpropdec}
\Pi(-k^2)=\frac{1}{a(-k^2)}\biggl[\frac{{\cal P}^2}{k^2}-\frac{1}{2{\widetilde a}(-k^2)}\biggl(\frac{{\cal P}_s^0}{k^2}-\frac{{\cal P}_s^0}{k^2+m^2}\biggr)\biggr]
,\]
where we have defined the spin-0 particle with a mass,
\begin{equation}\label{Staro-mass}
m^2\equiv \frac{M_P^2}{\alpha} > 0\,,
\end{equation}
which must be positive to ensure that the mass is non-tachyonic, and non-zero to retain the essential new pole, as discussed previously.
Armed with this, we are now in a position to describe the suitable defocusing condition which precludes the existence of ghosts. Substitution of Eq.\eqref{GF1} into Eq.\eqref{defocus3} leads to the linear condition;
\[
\label{defocus2}
(1-\Box/m^{2})\widetilde{a}({\bar \Box})R^{(L)}<R^{(L)}
.\]
Before we conclude, we briefly extend our results to inhomogenous spacetimes. We may expand the general defocusing condition Eq.\eqref{defocusgen}, to include solutions, with spatial as well as temporal dependencies. Without loss of generality, we consider the perturbations along the $x$-direction, where $r=\sqrt{x^2+y^2+z^2}$.
As before, we require $T_{\mu\nu}k^\mu k^\nu \geq 0$ so as not to violate the NEC. We then read off the minimum requirement for the associated null rays to defocus
\[
\frac{f(\bar\Box)}{a(\bar\Box)}\left(\partial_{t}^{2}+\partial_r^2\right)R^{(L)}<0
,\]
where $\partial_r^2=\partial_i\partial^i$ is the Laplace operator.
As an additional note, a closer look at the above action Eq.\eqref{action1} suggests that if we were to consider ${\cal M}\rightarrow \infty$, the action would reduce to that of Starobinsky's model of
inflation~\cite{Starobinsky}:
\[
\label{actionR2}
S_{R^2}=\frac{1}{2}\int d^4 x \bigl(M_P^2 R+f_{1_0}R^2\bigr).
\]
Indeed, a curious question to ask is, could Starobinsky's action avoid the cosmological singularity?
At the limit ${\cal M}\rightarrow \infty$, the propagator Eq.\eqref{Minkpropdec} can be expressed as
\[
\label{propR2}
\Pi_{R^{2}}=\Pi_{GR}+\frac{1}{2}\frac{\mathcal{P}_{s}^{0}}{k^{2}+m^{2}} ,
\]
where $m$ is given by Eq.~(\ref{Staro-mass}), with $\alpha=6f_{1_0}\geq 0$, and to avoid tachyonic mass, $m^2>0$.
However, the fundamental difference can be seen by comparing the propagator for $R^2$-gravity with an IDG propagator, see Eq.\eqref{Minkpropdec}. In the local limit, $a={\widetilde a}\rightarrow 1$. Furthermore, as we are making comparisons with the propagator in momentum space, the D'Alembertian takes the form $\Box\rightarrow -k^2$.
Taking these limits on the defocusing condition Eq. \eqref{defocus2} reveals the analogous condition for the Starobinsky model
\[
(k^2/m^2) R^{(L)}<0
.\]
In this scenario, in order to avoid convergence, we have two options: either 1.) the spin-0 particle has imaginary mass, i.e. $m^2<0$, or 2.) the curvature is negative.
\begin{enumerate}
\item A particle/field with imaginary mass is, by definition, tachyonic. As $m^2=M_P^2/6f_{1_0}$, where $f_{1_0}$ is the coefficient attached to the $R^2$ term in the action Eq.~\eqref{actionR2}, this corresponds to the $S\sim R-R^2$ form of the Starobinsky model.
\item If one dictates that this coefficient is necessarily positive, thus avoiding tachyons, the defocusing condition may be satisfied if one allows for negative curvature. However, negative curvature would contradict the requirement of accelerated expansion of the Universe, which is vital to realise primordial inflation. This corresponds to the $S\sim R+R^2$ form of the Starobinsky model.
\end{enumerate}
As such, the Starobinsky model cannot pair inflation with resolving the Big Bang Singularity.
\section{Conclusion}
From very generic considerations of a covariant, higher derivative theory of gravity, one can conclude that
null congruences can be made complete, or can be defocused, provided two essential criteria are satisfied
at the microscopic level, i.e. in terms of the graviton propagator: 1)
The graviton propagator must contain a scalar propagating mode which contains one additional root,
besides the massless spin-2, i.e. two helicity states; and 2) the infinite derivative theory of gravity must be, at least, ghost and tachyon free.
This conclusion verifies earlier analyses, which were dependent on an {\it Ansatz}-led, time-dependent solution to the equations of motion for an {\it IDG} theory, which described a non-singular cosmology in an homogeneous, isotropic and geometrically-flat framework.
This entails that an \emph{infinite} covariant derivative theory of gravity which can avoid ghosts or tachyons in the graviton propagator,
can yield a non-singular bouncing cosmology, and a possible UV completion of original Starobinsky inflation~\cite{Biswas:2013dry,Craps:2014wga,Koshelev:2016xqb}. In future, we shall be able to generalise our current discussion to de Sitter backgrounds, based on our recent
analysis of propagator in de Sitter spacetimes~\cite{Biswas:2016etb,Biswas:2016egy}. We can also attempt to analyse the defocusing of null congruences for a generic inhomogeneous and anisotropic backgrounds, however, one requires to first analyse the true dynamical propagating degrees of
freedom in such backgrounds, and the graviton propagator. Finding the latter would be a subject of discussion by itself, and we leave these discussions for future publication.
\section*{Acknowledgments}
We would like to thank Tirthabir Biswas, Valery Frolov, Terry Tomboulis, Alexei Starobinsky for helpful discussions. AC is funded
by STFC grant no ST/K50208X/1. AK is supported by the FCT Portugal fellowship SFRH/BPD/105212/2014 and in part by RFBR grant 14-01-00707. The work of A.M. is supported in part by the Lancaster-Manchester-Sheffield Consortium for Fundamental Physics under STFC grant ST/L000520/1.
|
\section{Introduction}
\label{sec:intro}
It was first noted by Hubble that the distribution of galaxies in angular cells on the celestial sphere is well approximated by a lognormal \citep{Hubble1934}. This has been confirmed observationally (\citealt*{Coles1991}, \citealt{Wild2005}) as well as in N-body simulations (\citealt*{Bernardeau_Kofman_1995}, \citealt{Bernardeau1994}, \citealt*{Kayo2001}), which have shown that the underlying mass density field is expected to be lognormal.
Since the weak lensing convergence field along the line of sight is a weighted projection of the mass density contrast field, one might suspect that the lognormal distribution is a reasonable, if not exact, model of this too. This has been tested on numerical simulations and a lognormal PDF shown to be a reasonable model (\citealt*{Taruya2002}, \citealt*{Hilbert2011a}). Even better fits to the convergence PDF, particularly in the tails of the distribution, have been obtained by generalisations of a lognormal PDF (\citealt*{Das2006}, \citealt{Takahashi2011}, \citealt*{Joachimi2011a}).
The Dark Energy Survey (DES) \citep{DES2005,DES2015,DES2016} presents an excellent opportunity to study both of these fields. DES was specifically conceived to produce cutting edge science from four different cosmological probes - large-scale structure, weak gravitational lensing, galaxy clusters and supernovae - using the same instrument. The full survey involves five years of observations, currently in progress. In this paper we focus on data produced during the pre-survey Science Verification (SV) series of observations.
This early data from DES allowed for the construction of two types of density fields. One is from luminous matter, i.e. galaxies of various types, $\delta_g$, which are biased tracers of the underlying dark matter field, $\delta_m$. The other uses the weak lensing of galaxy shapes to construct a convergence, or $\kappa$ map \citep{Vikram2015, Chang2015} that is directly sensitive to the integrated dark matter field out to the lensed galaxies.
Both maps trace the underlying density distribution in the Universe. Galaxies are biased tracers of matter density, preferentially clustering in overdense regions. Galaxy density contrast can then be considered a biased local tracer of the density field.
Weak lensing convergence on the other hand responds directly to the underlying density field and is therefore unbiased. However gravitational deflection of light is a cumulative effect, sensitive to the integrated matter density along the line of sight from source galaxy to observer. The convergence field for a given galaxy source distribution therefore gives us information about the cumulative density field between observer and source, with the exact contribution of matter at different distances along the line of sight governed by the lensing efficiency function.
The purpose of the present study is to analyse the galaxy and mass maps from DES SV simultaneously, testing the two maps separately for log-normality, as well as analysing the joint distribution. To our knowledge this is the first time that the log-normality of the weak lensing convergence field alone has been tested using data rather than numerical simulation \citep{Taruya2002}, and the first time the joint distribution has been tested for log-normality.
The \emph{Counts in Cells} (CiC) method (e.g. \citealt{Hubble1934, White1979, Gaztanaga1992, Szapudi1997,Bernardeau2002}) is a natural way to measure the individual and joint PDFs. The CiC technique splits up a particular data set into spatial cells, in two or three dimensions, and takes an aggregate of the available information inside each cell.
Statistical variation between cells then provides information on the properties of that cosmological field.
DES observations are ideally suited to this sort of analysis. The fact that DES provides a joint galaxy survey and convergence map data set produced from the same observations makes it easier to ensure consistency between data and to control for systematics. The SV data we use were taken before the start of the full five year survey, covers 139 deg$^2$ to full survey depth and forms a test-bed for the kind of analyses planned on the final DES data. All of the analyses in this work are done first on galaxy and convergence maps from MICE simulations in order to test our methodology.
The outline of the paper is as follows. In section \ref{sec:LN_theory} we review the theory and formalism used. We describe the galaxy and weak lensing convergence maps from MICE simulations and DES used in section \ref{sec:data}, and summarise our CiC method in section \ref{sec:method}. In section \ref{sec:MICE} we validate our CiC method on MICE Grand Challenge N-body simulations, checking that we see the expected lognormality in MICE $\delta_g$ and the noise-free convergence. In section \ref{sec:LN_DES} we present lognormal fits to the individual DES galaxy and convergence field distributions as well as their joint distribution. We check the validity of the log-normal model by measuring the variance of the fields and comparing this to the variance derived under the assumption of log-normality. We discuss the results in \ref{sec:conc}, and in the Appendices we give the formalism used to calculate moments from CiC, test the impact of systematic effects, and confirm that assumptions we make in our method do not significantly affect our results.
\section{Lognormal Modelling of cosmic fields}
\label{sec:LN_theory}
Lognormal distributions are very common in nature, from the sizes of clouds, pebbles on a beach, or crystals in icecream; the length of sentences or words in a conversation; to populations of bacteria (see \citealt*{Limpert2001}, \citealt*{Gaskell2004} and references therein). Many of these examples involve multiplicative processes, of either merging or fragmentation. Any process that can be written as a product of terms will, if there are many terms, tend to a lognormal distribution. This is because if a process $X$ can be written as a product of independent terms, then $\rm{ln}(X)$ will be a sum of independent terms, and via the central limit theorem these will be normally distributed. So $\rm{ln}(X)$ is normally distributed, or, $X$ is lognormally distributed.
There are many examples of the hierarchical merging or fragmenting of structure leading to lognormal distributions, such as: the initial mass function of field stars, explained in terms of cloud fragmentation \citep{Zinnecker1984}; X-ray flux variations, suggesting lognormal distribution of emitting regions \citep{Gaskell2004}; luminosity functions of central galaxies, explained in terms of BCGs being formed by steps of mass adding/stripping (e.g. \citealt*{Taghizadeh-Popp2012}); and the angular momentum of disc galaxies \citep{Marr2015a}.
In this paper we test the lognormality of the distribution of matter in the Universe via the galaxy density contrast field, $\delta_g$, and via the weak lensing convergence field, $\kappa_{\rm{WL}}$. Each approach has particular observational and astrophysical noise associated with it, which we discuss and propose models for in this section.
\subsection{Galaxies}
In the standard picture of gravitational instability, the primordial density fluctuations that are the precursor of all structure in the universe are assumed to be a random Gaussian field. Once they enter the non-linear regime, with finite rms fluctuations, their PDF must deviate from a Gaussian to avoid non-zero probabilities being attached to regions with negative densities \citep{Fry1984}. The exact form of the PDFs in this regime is not known but there are various phenomenological models that are fully specified statistically and satisfy the common sense requirement that the matter density, $\rho \geqslant 0$ (e.g. \citealt*{Saslaw1984}, \citealt*{Suto1990}, \citealt{Lahav1993}, \citealt{Gaztanaga1993a}; \citealt{Suto1993}, \citealt*{Ueda1996}).
One such model that is widely used is the lognormal. As well as being completely specified statistically and always having $\rho \geqslant 0$, it becomes arbitrarily close to Gaussian statistics at early times and has the advantage that it can be handled analytically. The merits of this model in a cosmological context are discussed extensively in \citet{Coles1991}. They explain possible motivations for using the lognormal model as: empirical; kinematic; an application of the central limit theorem (as described above); and importantly, simplicity. It is one of the few random fields for which interesting quantities such as its moments can be calculated analytically.
It should be noted that despite these compelling motivations to use a lognormal in the statistical treatment of density perturbations, it does have shortcomings. In particular, it is not uniquely specified by its moments; many distributions can lead to the same set of moments. It must then be the case that information is lost in going from a lognormal field to its moments, an effect quantified in \citet{Carron2011}. However, it remains a popular and useful tool in analysis of the mass density contrast field.
Galaxies are biased tracers of the mass density contrast field. The 1D log-normal distribution of galaxy density contrast $\delta_g$ = $(\rho - \bar\rho)/\bar\rho$ is given by:
\begin{equation}
f(\delta_g) \mathrm{d}\delta_g = \frac{1}{w \sqrt{2 \pi}}exp\left( \frac{-x^2}{2w^2}\right) \mathrm{d}x
\label{eqn:LN}
\end{equation}
where $x = \mathrm{ln} (1+\delta_g) + w^2/2$ and $w^2$ is the variance of the corresponding normal distribution $f[ln(1+\delta_g)]$. The offset $w^2/2$ ensures that $\langle \delta_g \rangle = 0$. The width $w$ is then the single free parameter of the 1D lognormal distribution.
If the lognormal distribution correctly describes the data, the variance of the overdensities will be related to the variance, $w$, of the underlying Gaussian distribution by
\begin{equation}\label{eq:gg}
\langle\delta_g\delta_g\rangle = \mathrm{e}^{w^2}-1.
\end{equation}
Due to the discrete nature of galaxies, shot noise is present. Assuming Poisson sampling of galaxies,
the shot noise in the measurement of the distribution of galaxy overdensities can be accounted for by convolving the log-normal model with a Poisson distribution. The probability distribution function of the galaxy counts $N$ in a cell of given size is then given by:
\begin{equation}\label{eq:PxLN}
P(N) = \int^\infty_{-1} \frac{\bar{N}^N (1+\delta_g)^N}{N!} e^ {-\bar{N}(1+\delta_g)} f(\delta_g) d\delta_g
\end{equation}
This Poisson sampled lognormal distribution has been shown to be a good fit to different galaxy populations in \citet{Coles1991}, \citet{Blanton2000a} and \citet{Wild2005}. In this work the smallest number of DES galaxies in a cell considered is around 300, so the shot noise term is important.
\subsection{Weak Lensing Convergence}\label{sec:LN_theory:WL}
Various expressions for the convergence PDF have been proposed \citep{Munshi2000a,Valageas2000,Kainulainen2011a}. The lognormal model has the advantage - as in the case of the matter density contrast - of mathematical convenience, while offering the chance to extract more information than assuming a purely Gaussian model for the convergence field. Following a study that showed that a lognormal transformation of the matter density contrast increases the signal to noise \citep{Neyrinck2009}, \citet{Seo2011} performed an analogous study of the weak lensing convergence. They found that such a transform, when applied to the positively offset convergence, decorrelated angular frequencies and increased the signal-to-noise in the transformed power spectrum.
The convergence field along a line-of-sight can be expressed as a weighted projection of the mass density contrast field:
\begin{equation}
\kappa(\theta) = \int^{\chi_{\rm hor}}_{0} d\chi w(\chi)\delta[r(\chi)\theta,\chi],
\label{eqn:kappa_from_delta}
\end{equation}
where $\chi$ is the comoving distance, $\chi_{\rm hor}$ is the angular diameter distance to the horizon and $\delta[r(\chi)\theta,\chi]$ is the underlying matter density contrast field. $w(\chi)$ is a geometrical weight function that depends on the relative separations of sources, lens and the observer (see e.g. \citealt{Mellier1999}; \citealt{Bartelmann2001} for reviews). It takes the form
\begin{equation}
w(\chi) = \frac{3}{2}\left(\frac{H_{0}}{c}\right)^{2} \frac{\chi\Omega_{0}}{a(\chi)} \int_{\chi}^{\chi_{\rm hor}} d\chi' n(\chi') \frac{\chi'-\chi}{\chi'} ,
\label{eqn:lensing_weight_fn}
\end{equation}
where $n(\chi)$ is the source galaxy distribution.
The distribution of $\kappa$ is not expected to be exactly lognormal, even if $\delta$ is, since $\kappa$ is a weighted projection of the mass density contrast field along line of sight. However, simulations have shown \citep{Taruya2002} that the convergence field is well approximated by a lognormal outside the regime of extremely high $\kappa$. Hence we choose in this work to model the noise free $\kappa$ field distribution with a shifted lognormal
\begin{equation}\label{eq:WL_LN}
P(\kappa) = \begin{cases}
\cfrac{\exp\Biggl[-
\cfrac{(\textnormal{ln}(\kappa_0+\kappa)-\mu)^2}{2\sigma^2}\Biggr] }
{\textnormal{ln}(\kappa_0+\kappa) \sqrt{2 \pi}\sigma }
& \text{for } \kappa > - \kappa_0,\\
0 & \text{otherwise},
\end{cases}
\end{equation}
where the shift $\kappa_0 = -\kappa_{min}$ and is called the `minimum convergence parameter' \citep{Hilbert2011a}. The lowest possible value of $\kappa$ is given by $-\kappa_0$ since the lognormal is only defined for a positive range. The mean is given by
\begin{equation}\label{eq:mu}
\mu = \textnormal{ln}(\kappa_0+\langle\kappa\rangle) - \sigma^2 / 2
\end{equation}
and the second moment
\begin{equation}\label{eq:kk}
\langle\kappa\kappa\rangle = e^{(2\mu+\sigma^2)}(e^{\sigma^2}-1).
\end{equation}
The value assigned to $\kappa_0$ is a modelling choice that can be approached in several ways. The minimum measured value of $\kappa$ could be used, but this is a noisy quantity and should not be used unless one has access to many realisation of $\kappa$. Or, treating $\kappa$ analogously to $\delta_g$, we could consider a theoretical minimum corresponding to the convergence we would see, for a given source distribution, if the mass distribution was a pure void along the entire line of sight. For the MICE source distribution used in this work this value is $-0.050$, and for the DES source distribution it is $-0.053$. However simulations show that there are no empty lines of sight in a $\Lambda$CDM universe \citep{Taruya2002, Vale2003, Hilbert2011a}. So we choose, where possible, to treat $\kappa_0$ as a free parameter and fit it jointly with the lognormal width.
\begin{figure}
\centering
\includegraphics[scale=0.4]{DES_g_0p2z0p5_K_0p3z1p3_r20_cov100_100_khat_rand_GassianFit.png}
\caption{\label{fig:100_Krand} Demonstration of the Gaussianity of the noise in DES weak lensing convergence, $\kappa$, at a smoothing scale of 20\arcmin. Probability density distributions of the 100 realisations of the $\kappa$ map, in which shears were randomised, are shown in blue, with jackknife errors. A Gaussian PDF was fitted to each, with the mean of the best fitting PDFs shown in magenta. A Gaussian model is an excellent fit to the noise, with average goodness of fit $\chi^2/$DOF$ = 0.95 \pm 0.73$. }
\end{figure}
As for galaxies, we need to modify the lognormal to account for noise. The DES $\kappa$ map is constructed (see section \ref{data:DES_K}) from measurements of shear, which is the change in the ellipticity of galaxies resulting from weak gravitational lensing. Since galaxies are intrinsically elliptical (i.e. in the absence of lensing), the observed shear is the sum of this intrinsic ellipticity and the shear caused by lensing. The variance of the intrinsic ellipticity, called shape noise, is the dominant source of noise in shear measurements, typically by a factor of more than an order of magnitude. An estimate of the shape noise in the DES $\kappa$ map is provided by the 100 noise realisations described in section \ref{data:DES_K}.
To analyse the shape of the noise distribution we construct PDFs via CiC (as described in section \ref{sec:method}) on each of the 100 maps. The resulting distributions appear Gaussian, as shown in figure \ref{fig:100_Krand}, where the thick blue curve is made up of 100 superimposed noise distributions with jackknife error bars, and the magenta line shows the mean bestfit Gaussian PDF. A Gaussian model provides an excellent fit, with average goodness of fit over the 100 maps $\chi^2/dof = 0.95 \pm 0.73$.
We therefore propose that the 1D probability distribution for the weak lensing convergence field is then given by a convolution of a lognormal distribution with a noise contribution modelled as Gaussian:
\begin{equation}\label{eq:GxLN}
f(\kappa) = \frac{1}{\sqrt{2 \pi} \sigma_n} \int^\infty_{-\kappa_0} \exp\Big[-\frac{(\kappa'-\kappa)^2}{2\sigma_n^2}\Big] P(\kappa') \mathrm{d}\kappa'
\end{equation}
where $ P(\kappa)$ is the noise free log-normal distribution given in equation \ref{eq:WL_LN}, and $\sigma_n$ is the Gaussian width of the shape noise. In this work, $\sigma_n$ is determined from the 100 noise realisations.
\subsection{Joint Galaxy and Weak Lensing Convergence Field}
We can try to model the joint distribution of galaxy density contrast $\delta_{g}$ and weak lensing convergence $\kappa_{\rm{WL}}$ as a bivariate lognormal with PDF $f(\delta_{g},\kappa_{\rm{WL}})$. Following the notation used in \citet{Coles1991,Wild2005}, this is given by
\begin{equation}
f(\delta_g,\kappa) = \frac{|V|^{-1/2}}{2\pi} \exp \left[ -\frac{(\tilde{g}_{\delta}^{2} + \tilde{g}_{\kappa}^{2} - 2r_{\text{LN}}\tilde{g}_{\delta}\tilde{g}_{\kappa})}{2(1-r^{2}_{\text{LN}})} \right],
\label{eqn:2D_LN}
\end{equation}
where $g_x = \ln(x) - \left<\ln(x)\right>$, with $x = (1+\delta_g)$ and $x = (1+\kappa/\kappa_0)$ for the galaxy and convergence fields respectively, and $\tilde{g_x} = g_x/\omega_x$ where $\omega_x$ is the variance of the underlying Gaussian field $\ln(x)$.
The lognormal correlation coefficient $r_{\text{LN}}$ is given by
\begin{equation}
r_{\text{LN}} = \frac{\left< g_{\delta}g_{\kappa} \right>}{\omega_{\delta}\omega_{\kappa}} \equiv \frac{\omega^{2}_{\delta\kappa}}{\omega_{\delta}\omega_{\kappa}}
\end{equation}
and $|V|$ is the determinant of the covariance matrix
\begin{equation}
V = \left( \begin{array}{cc}
\omega^{2}_{\delta} & \omega^{2}_{\delta\kappa} \\
\omega^{2}_{\delta\kappa} & \omega^{2}_{\kappa} \end{array} \right).
\end{equation}
Note that $r_{\text{LN}}$ and $V$ are defined in log-density space, and so $r_{\rm{LN}}$ is not the same as the (linear) Pearson correlation coefficient $\rho$.
The conditional probability
\begin{eqnarray}\label{eq:cond}
f(\delta_{g}|\kappa) &=& \frac{f(\delta_{g},\kappa)}{f(\delta_{g})}\\
&=& \frac{w_{\delta}}{(2 \pi |V|)^{1/2}}
exp\Biggl[-\frac{(\tilde{g}_{\delta} - r_{\rm{LN}}\tilde{g}_{\kappa})^2}
{2(1-r^{2}_{\rm{LN}})}\Biggr].
\end{eqnarray}
Since $f(\delta_g,\kappa) = f(\kappa)f(\delta_g|\kappa)$ we can combine equations \ref{eq:GxLN} and \ref{eq:cond} to give the joint probability distribution function $f(\delta_g,\kappa)$. Including the convolutions with Poisson shot noise and Gaussian shape noise then gives
\begin{multline}\label{eq:biv}
P(N,\kappa) = \int^\infty_{-1} \int^\infty_{-\kappa_0}
\frac{1}{\sqrt{2 \pi} \sigma_n} \exp\Big[-\frac{(\kappa'-\kappa)^2}{2\sigma_n^2}\Big] f(\kappa) \\
\times \frac{\bar{N}^N (1+\delta_g)^N}{N!} e^ {-\bar{N}(1+\delta_g)} f(\delta_g|\kappa)\mathrm{d}\delta_g \mathrm{d}\kappa'
\end{multline}
\section{The Data}
\label{sec:data}
\begin{figure}
\centering
\includegraphics[scale=0.2]{fig_z_kernels_skynet_nocmb_250416_alt.png}
\caption{\label{fig:z_kernels} Redshift kernels of the observables considered in this paper: the galaxy redshift distribution of the DES Benchmark galaxy sample using the best-fit Skynet photo-z estimation (black line), and the lensing efficiency function of the sources used to make the DES $\kappa$ map (red line). Also shown is the redshift distribution of the source galaxies (shaded region). Each is shown with an arbitrary normalisation to make comparison easier.}
\end{figure}
This paper uses the DES Science Verfication (SV) galaxy and shape catalogues. The SV data were gathered between November 2012 and February 2013, shortly after DECam \citep{Flaugher2015} commissioning, and before the beginning of the (five year) DES survey proper in August 2013. The operation of the camera, survey planning, data analysis and reduction were all tested in preparation for starting year one of DES itself. The SV goal was to reproduce the properties of the full five-year DES survey over a much smaller sky area.
Five optical filters ($grizY$) are used, with exposure times of 90 seconds for $griz$ and 45 seconds for $Y$. The final median depth in our region of interest, per band, was $g \sim 24.0$, $r \sim 23.9$, $i \sim 23.0$ and $z \sim 22.3$.
In total the SV data covered $\sim 250$ deg$^2$ at close to the nominal depth of the full DES survey. The observing footprint was divided into regions to maximise overlap with other surveys and with several small fields used for SNe searches.
In this paper we concentrate on a large contiguous region of $\sim$139 deg$^2$ called the SPT-E field due to its overlap with the South Pole Telescope CMB survey. This amount of contiguous data makes the SV SPT-E field a powerful data set in its own right, particularly for weak gravitational lensing where it rivals the full CFHTLenS \citep{Erben2013} and is only slightly shallower.
\subsection{DES Galaxy Sample}
\label{data:DES:DES_gals}
We use a particular subset of the DES SV galaxy catalogue known as the ``Benchmark'' sample \citep{Crocce2015}. First a catalogue of galaxies suitable for LSS analysis was constructed from the SV data and dubbed the ``Gold'' sample \citep{Rykoff2016}. Objects were included if detected in all five of the DES photometric bands. This covered $\sim$244 deg$^2$, restricted to $\textit{dec} > -61$ to avoid the Large Magellenic Cloud and R Doradus regions. In addition the Gold catalogue included masking of satellite trails and other artifacts, removal of regions where colors are severely affected by stray light and the application of additional stellar locus correction \citep{Kelly2014}.
From this Gold sample, the Benchmark sample was selected for cosmological analysis by imposing the additional conditions:
\begin{itemize}
\item $18.0 < i < 22.5$
\item $0 < g-r < 3$, $0 < r-i < 2$ and $0 < i-z < 3$.
\item \texttt{wavg\_spread\_model} $>$ 0.003 (star-galaxy separation)
\item $60 < ra < 95$ and $-62 < dec < -40$ (SPT-E),
\end{itemize}
where $i$ refers to \texttt{SExtractor}'s \texttt{MAG_AUTO} quantity. The cuts on position restrict our analysis to the SPT-E region. The redshifts used in this paper come from the Skynet photo-z pipeline (\citealt{Bonnett2015a}, \citealt{Graff2013}).
The galaxy redshift distribution is shown in figure \ref{fig:z_kernels}. The redshift range we use throughout this paper is $0.1<z<1.5$, chosen as in this region the galaxy redshift distribution overlaps with the lensing efficiency function used to make the DES $\kappa_{WL}$ map (see next section).
\subsection{DES $\kappa$ Map}
\label{data:DES_K}
Shear measurement on DES SV galaxy images was performed with two independent pipelines: \textsc{im3shape}\footnote{The open source code can be downloaded at: \texttt{https://bitbucket.org/joezuntz/im3shape/}} \citep{Zuntz2013} and \textsc{ngmix}\footnote{The open source code can be downloaded at: \texttt{https://github.com/esheldon/ngmix}} \citep{Sheldon2014}.
Extensive testing of both codes was carried out by the DES collaboration (see \citealt{Jarvis2015} for details) and both pipelines passed all requirement tests for measurement of cosmic shear with the SV data set. A number of cuts were applied to both catalogues to remove stars, spurious detections, poor measurements and other effects that could bias shear measurement; these are also described in \cite{Jarvis2015}.
Shear measurements for a given galaxy are headless vectors and the cosmic shear field is therefore a spin-2 quantity. To allow us to perform our CiC analysis on a scalar quantity we work with maps of weak lensing convergence, $\kappa$, a spin-0 field. This $\kappa$-reconstruction was performed using the Kaiser-Squires method \citep{Kaiser1993}, and the production and initial analysis of these $\kappa$ maps is described in detail in \citet{Vikram2015}.
The Kaiser-Squires reconstruction method uses the relation of the Fourier transform of the observed shear, $\hat{\gamma}$, to that of the convergence, $\hat{\kappa}$,
\begin{equation}
\hat{\kappa}_{\ell} = D^*_{\ell}\hat{\gamma}_\ell,
\label{eqn:FT}
\end{equation}
\begin{equation}
D_{\ell} = \frac{\ell_{1}^{2}-\ell_{2}^{2}+2i\ell_{1}\ell_{2}}{|\ell|^{2}},
\end{equation}
where $\ell_i$ are the Fourier counterparts of the angular coordinates, $\theta_i, i=1,2$. The inverse Fourier transform of equation \ref{eqn:FT} gives the convergence for the observed field in real space. In the absence of noise, systematics and masking, the convergence will be a real (spin-0) quantity. In reality these effects produce a non-zero imaginary component. It is most convenient to express the real part of the convergence map as a map of curl free E-modes, and the imaginary part as divergence free B-modes. The $\kappa$ maps have pixels of size 2\arcmin. For use in this analysis the original flat sky $\kappa$ maps are transformed into \texttt{HEALPix} \citep{Gorski2005} maps at resolution $N_{\text{side}}$=4096. This is done by dividing each pixel of the flat sky maps into 25 sub-pixels, and creating a \texttt{HEALPix} map by combining these sub-pixels. This procedure reduces inaccuracies in changing from one mapping system to another, and in tests gives the same angular power spectrum measurements as the flat sky map to well within the errors.
The source galaxy selection used to construct the $\kappa$ map used in this paper took galaxies with redshifts in the range $0.6 < z < 1.3$. The resulting redshift efficiency function is shown in figure \ref{fig:z_kernels}. The lensing efficiency function peaks at {$z \sim 0.3$}, and our selection of galaxies at $0.1 < z < 0.5$ overlaps significantly with the range of redshifts to which the $\kappa$ map is sensitive.
In addition to the E-mode $\kappa$ map we make use of a number of other products made in the course of the DES mass-mapping analysis. A B-mode map was constructed by rotating the measured galaxy ellipticities by 45 degrees. The physical process of weak gravitational lensing does not induce B-modes in the convergence field so the B-mode map is a test of systematic effects in our observations, shear measurement and $\kappa$-reconstruction; it should be consistent with zero within our reconstruction noise. We will refer to the B-mode reconstructed map as $\kappa_B$.
In addition to the E- and B-mode maps we also make use of a series of noise-only realisations, made by taking the galaxy shape catalogue and rotating the measured shape of each galaxy by some random angle. $\kappa$ maps were then constructed from each randomised catalogue in the usual way. This has the effect of destroying all cosmological information in the resulting maps, while retaining the same noise properties as the data (because the distribution of galaxies on the sky and in redshift remains the same, as does the overall ellipticity distribution across the sample). 100 of these noise realisations were made and we use them to estimate the noise contribution in our measurement, as described in more detail in section \ref{sec:LN_theory:WL}.
\begin{figure}
\centering
\includegraphics[scale=0.3]
{DES_KE_KB_rands_stdev_with_JKerrs_cap_alpha0p3_biggercircles.png}
\caption{\label{fig:Bmode_SN} Standard deviations of the $\kappa_E$, $\kappa_B$ signal (magenta) and 100 realisations in which the shears have been randomised (blue) at a cell size of 20\arcmin. The random realisations of $\kappa_E$ give an estimate of the shape noise contribution to $\kappa_E$; this accounts for 80\% of the $\kappa_E$ signal. The $\kappa_B$ signal is also a good estimate of the shape noise, with the standard deviation of the $\kappa_B$ signal agreeing with the rms standard deviation of $\kappa_E$ random realisations within 2\%. These standard deviations are calculated via CiC and errors are from jackknife sampling.}
\end{figure}
Fig. \ref{fig:Bmode_SN} shows the standard deviations of the $\kappa_E$, $\kappa_B$ signal (magenta) and 100 noise realisations (blue) for a cell size of 20\arcmin. This shows that the shape noise (given by the random realisations of $\kappa_E$) accounts for 80\% of the $\kappa_E$ signal, underlining the importance of accounting for shape noise in our modelling (as described in section \ref{data:DES_K}). The shape noise dominates the signal most at small scales, accounting for 89\% of the signal at 10\arcmin\ and dropping to 64\% at 40\arcmin. We can see that the $\kappa_B$ signal is also a good estimate of the shape noise, with the standard deviation of the $\kappa_B$ signal agreeing with the rms standard deviation of $\kappa_E$ random realisations within 2\%.
These standard deviations are calculated via CiC (see section \ref{sec:consistency_second_moments} for a prescription for calculating moments from CiC) and errors are from jackknife sampling (see section \ref{sec:method}).
\subsection{MICE Simulations}
\label{sec:data_MICE}
We validate our measurement of CiC from DES SV data using a special set of mock catalogues produced from N-body simulations for the DES collaboration. These come from the Marenostrum Institut de Ci\`{e}ncies de l'Espai Grand Challenges (MICE-GC hereafter) lightcone N-body simulation and associated halo catalogue.
These simulations have been used to produce mock galaxy catalogues for $\sim$200 million galaxies over 5000 deg$^2$ up to a redshift of $z=1.4$. There are also shear estimates for each galaxy made by ray-tracing through the N-body simulations. Every galaxy has a $\kappa_{\rm{WL}}$ value assigned from the integrated dark matter field.
The simulations are made with $4096^3$ particles of mass $2927 M_{\odot} h^{-1}$ in a box of side 3072 $h^{-1}$Mpc. The MICE-GC has an assumed flat $\Lambda$CDM cosmology: $\Omega_{\rm m} = 0.25$, $\Omega_{\rm b} = 0.044$, $\Omega_\Lambda = 0.75$, $\sigma_8 = 0.8$, $h=0.7$, $n_s = 0.95$. The MICE-GC DES mocks approximately reproduce the magnitude limits of the DES survey and are complete down to apparent magnitude $i < 22.0$ at $z=0.5$
For use in this paper we have reduced the effective number densities in the mock galaxy and shear catalogues to reflect the statistics of the DES SV samples as well as normalising the redshift to reflect the distribution shown in Fig. \ref{fig:z_kernels}. Each mock catalogue is projected onto a \texttt{HEALPix} map of $N_{side}=8192$, which is then degraded to match the resolution of our data maps where appropriate.
\begin{figure}
\centering
\includegraphics[scale=0.4]{MICE_kappa_DESified_with_shapenoise_unmasked_NgNz_only.png}
\caption{\label{fig:shape_noise_MICE} Distribution of MICE $\kappa_{\rm{WL}}$ at an angular scale of 10\arcmin when DES-like shape noise is added. An estimate of the width of the shape noise distribution is obtained by fitting a Gaussian to the 100 random realisations of DES $\kappa_{\rm{WL}}$. A noise contribution drawn from a Gaussian of this width is added to MICE $\kappa_{\rm{WL}}$ at the level of the cells used to construct the CiC distribution. The darker, narrow histogram is that of the shape noise free MICE $\kappa_{\rm{WL}}$; the lighter histogram shows the distribution once the Gaussian shape noise is added; the black dashed line shows the observed distribution of DES $\kappa_{\rm{WL}}$. The Gaussian width of the DES shape noise estimate, $\sigma_n$ is 0.0099 at this scale, which is 89\% of the width of the resulting noisy MICE $\kappa_{\rm{WL}}$ distribution. }
\end{figure}
In order to be able to compare the distribution of DES $\kappa_{\rm{WL}}$ (which we know has a significant shape noise contribution) with simualtions, we create a second MICE $\kappa_{\rm{WL}}$ sample that has DES-like shape noise added. An estimate of the width of the shape noise distribution is obtained by fitting a Gaussian to the 100 random realisations of DES $\kappa_{\rm{WL}}$. A noise contribution drawn from a Gaussian of this width is added to MICE $\kappa_{\rm{WL}}$ at the level of the cells used to construct the CiC distribution.
Figure \ref{fig:shape_noise_MICE} shows the effect of adding shape noise to MICE $\kappa_{\rm{WL}}$ in this way at an angular scale of 10\arcmin. The darker, narrow histogram is that of the shape noise free MICE $\kappa_{\rm{WL}}$; the lighter histogram shows the distribution once the Gaussian shape noise is added; and the black dashed line shows the distribution of DES $\kappa_{\rm{WL}}$. At a smoothing scale of 10 arcmin the Gaussian width of the DES shape noise estimate is 0.0099, which is 89\% of the width of the resulting noisy MICE $\kappa_{\rm{WL}}$ distribution; this falls to 63\% at 40\arcmin.
\section{Method}
\label{sec:method}
\subsection{Constructing PDFs via Counts-in-Cells}
The CiC approach is a relatively simple way to measure the distribution of galaxies in a survey, but it is a surprisingly powerful tool. A general CiC distribution for galaxies can be denoted by $f(N,V)$, the probability of finding $N$ galaxies in a volume of space $V$. This can be a 3D volume or, as is the case in this paper, a 2D area on the sky where we count over a population projected along the line of sight. Repeating this procedure with cells of varying radii, $r$, gives us the distribution $f_{r}(N)$, where the moments of $f_{r}(N)$ are related to the volume integrals of the correlation functions of our underlying observable \citep{Peebles1980,Fry1985,Saslaw2000,Fry1994}.
We perform our CiC analysis on the galaxy density contrast and weak lensing covergence maps with \texttt{HEALPix} pixelisation of resolution $N_{\text{side}}=4096$, which corresponds to an average pixel size of 0.9\arcmin. For galaxies, to construct the PDF we sum the galaxy counts, $N$, inside 2D circular cells of fixed radius $r$ in the range 10--40\arcmin. At the median redshift, $z=0.3$, of the sources considered this corresponds to physical scales of 3--10 Mpc. The smallest cells used are 10 times larger than the \texttt{HEALPix} pixels in order to minimise edge effects, and this also avoids any difference in counts across our survey area due to the changing geometry of the \texttt{HEALPix} pixels (see Appendix \ref{sec:Ap_sampling} for a discussion of this assumption). We chose to use randomly positioned circular cells rather than using the \texttt{HEALPix} pixels themselves as this allows us to repeat the analysis straightforwardly at any smoothing scale, rather than using only the fixed scales of \texttt{HEALPix} pixels. The criterion for accepting a cell is that 80\% of its area should fall in unmasked regions (again see Appendix \ref{sec:Ap_sampling} for discussion of this choice). We want to use enough cells that all pixels in the map are covered at least once, and find that this is achieved when the total area of the cells is 20 times that of the survey. We use a coverage of 100 times the total area.
Histograms of the counts give us the distribution $f(N)$, and this procedure is repeated with cells of different radii to obtain the distribution $f_{r}(N)$. Double counting of pixels is accounted for by jackknife errors on the height of each bin in the resulting histogram of counts. We divide the survey area into 152 approximately equal area (1 deg$^2$) jackknife patches. For a fixed set of randomly generated cells, and removing one patch at a time, we re-make the galaxy and convergence PDFs and re-calculate the statistics of interest in order to produce covariances.
We repeat our CiC analysis on the DES reconstructed $\kappa$ maps. The `count' in each cell is now the average of the weak lensing convergence $\kappa$ in pixels contained in that cell.
In Appendix \ref{sec:Ap_syst} we test the impact of spatially varying systematic effects on the DES $\delta_g$ and $\kappa_{\rm{WL}}$ CiC distributions.
It is straightforward to generalise our CiC method to more than one observable. We simply throw the same circles onto each map (using the same mask for each), allowing us to compare counts at the same position for different observables.
\subsection{Fitting the PDFs}
We fit the lognormal models described in section \ref{sec:LN_theory} to these distributions. For the MICE and DES galaxy density contrast distributions we fit a Poisson sampled lognormal using equation \ref{eq:PxLN}. For MICE $\kappa_{\rm{WL}}$, which has no shape noise, we fit a plain lognormal model (equation \ref{eq:WL_LN}). For the $\kappa_{\rm{WL}}$ distributions which include shape noise (i.e. DES $\kappa_{\rm{WL}}$ and the MICE $\kappa_{\rm{WL}}$ to which we add shape noise), we use equation \ref{eq:GxLN}.
\begin{figure}
\centering
\includegraphics[scale=0.33]{covmat_norm_kappa_r20.png}
\caption{\label{fig:covmat} Correlation matrix of bin heights for a histogram of DES $\kappa_{\rm{WL}}$, at a smoothing scale of 10\arcmin. Derived from jackknife sampling of the DES $\kappa_{\rm{WL}}$ map. }
\end{figure}
The histogram bins in $\delta_g$ or $\kappa_{\rm{WL}}$ are correlated. This is demonstrated in figure \ref{fig:covmat}, which shows the correlation matrix of bin heights of DES $\kappa_{\rm{WL}}$ at a smoothing scale of 10\arcmin. In fitting the lognormal model we take into account these correlations by minimising
\begin{equation}
\chi^2 = (\vec{f} - \vec{d})\mathcal{C}^{-1}(\vec{f} -\vec{d})'.
\end{equation}
Here $f$ is the data vector of the lognormal fit at the bin centres, $d$ is the data vector of bin heights, and $\mathcal{C}$ is the covariance matrix. We remove weak eigenvectors of the covariance matrix via singular value decomposition.
\section{Validating Methods on MICE}
\label{sec:MICE}
In this section we verify the methods used to test the lognormality of DES $\delta_{g}$ and $\kappa_{\rm{WL}}$ fields. After checking that the MICE $\delta_{g}$ field is lognormal as we would expect, we see if this is true of the noise-less convergence field.
To enable easier comparison with the DES $\kappa_{\rm{WL}}$ results we also look at the distribution of the simulation $\kappa_{\rm{WL}}$ for the MICE sample with number of galaxies and $n(z)$ matched to our DES sample, and with DES-like shape noise added. We then look at the joint distribution of $\delta_g$ and $\kappa_{\rm{WL}}$, for the cases with and without shape noise.
As an additional check of the validity of the lognormal model, we compare the second moments of the distributions as calculated via CiC with those derived under the assumption of lognormality.
\subsection{Testing Lognormality of MICE Density and Convergence Fields}
\label{sec:LN_MICE}
\subsubsection{One-dimensional PDFs and log-normal fits}
\label{sec:1D_MICE}
\begin{figure*}
\centering
\includegraphics[scale=0.4] {MICE_gal_G_LN_r_10_15_30_log_yaxis.png}
\includegraphics[scale=0.4]
{MICE_kappa_G_LN_cov_r_10_15_30_log_yaxis_svcutoff_0p01.png}
\includegraphics[scale=0.4]
{MICE_Knoisy_unmasked_G_conv_cov_r_10_15_30_fit_k0.png}
\caption{\label{fig:gal_K_1Dfit_r5_MICE} \textbf{UPPER ROW:} measured 1D PDF of MICE galaxies at a smoothing scales of 10, 15 and 30\arcmin. The Poisson sampled lognormal fit (black) provides a better fit to the galaxy CiC distribution than the Gaussian (magenta) at a scale of 10\arcmin. The distribution becomes increasingly Gaussian at larger scales. \textbf{MIDDLE ROW:} same as above but for the MICE $\kappa_{\rm{WL}}$ PDF. Again the lognormal provides a good fit at the smallest scale, with the $\kappa_{\rm{WL}}$ distribution becoming more Gaussian at larger scales. \textbf{BOTTOM ROW:} Fits to $\kappa_{\rm{WL}}$ using the sub-sample of MICE with DES-like galaxy density and $n(z)$, and to which DES-like shape noise has been added. This shape noise makes the distribution of $\kappa_{\rm{WL}}$ more Gaussian at all scales. All $\chi^2$ are per degree of freedom.}
\end{figure*}
We first construct a simple histogram of $\delta_{g}$ from the CiC to estimate the 1D PDF of $\delta_{g}$. The histogram uses 50 bins and we calculate jackknife errors on the bin heights as described in the previous section. The result for cells of radius of 10, 15 and 30\arcmin\ is shown in the upper panel of Fig. \ref{fig:gal_K_1Dfit_r5_MICE}.
We fit a Poisson sampled lognormal distribution as described in eqn. \ref{eqn:LN} with $w$ as the single free parameter. The best-fit lognormal, which minimises $\chi^2$, is shown as a solid black line and the best fit Gaussian (magenta) is shown for comparison. At a cell size of 10\arcmin\ (corresponding to about 3 Mpc at the median redshift $z=0.3$) it is clear that the lognormal model fits the data better, reflecting the non-linear clustering at this scale. The counting of information inside a cell can be thought of as a form of smoothing where the cells form a top-hat filter with a fixed size. As the size of our cells increases we average information on increasingly large scales and lose sensitivity to the effects of non-linear clustering on small scales.
The lognormal distribution is designed to capture some of the information present as a result of non-linear evolution, so we would expect it to become less pronounced as the effective smoothing scale increases. This is indeed the case: at a cell radius of 10\arcmin\ the lognormal model is highly favoured, with a $\chi^{2}/DOF = 1.13$, compared to 9.66 for the Gaussian. At a cell size 30\arcmin\ (corresponding to a physical scale of 8Mpc at the median redshift) the distribution has become much more Gaussian with best-fit $\chi^{2}$/DOF for the Gaussian model now 1.50. The lognormal model is still favoured at this scale, with best-fit $\chi^{2}$/DOF $=0.95$.
The result for the MICE $\kappa_{\rm{WL}}$ PDF is shown in the middle panel of fig. \ref{fig:gal_K_1Dfit_r5_MICE}. Since there is no shape noise in the simulation we fit a plain lognormal, shown by the black line. As discussed in section \ref{sec:LN_theory:WL}, in order to fit a lognormal model to $\kappa_{\rm{WL}}$ one must assign a value to $\kappa_0$, the minimum convergence parameter in equation \ref{eq:WL_LN}. At 10\arcmin\ we jointly fit $\kappa_{0}$ and the lognormal width in equation \ref{eq:WL_LN}, finding best-fit $\kappa_{0}=0.021$. For larger scales we find that it is not possible to jointly constrain $\kappa_0$ and the width of the lognormal as they are degenerate. We therefore use the theoretically derived $\kappa_0=0.050$, described in section \ref{sec:LN_theory:WL}.
Since the convergence is the weighted sum of the mass fluctuations along the line of sight we expect it to be only approximately lognormal. At a smoothing scale of 10\arcmin\ the lognormal is a good fit, with $\chi^2$/DOF$ = 1.19$, and it is significantly preferred to the Gaussian model, which has a best-fit $\chi^2$/DOF$ = 14.43$. This lognormality of $\kappa_{\rm{WL}}$ at small scales is in line with \cite{Taruya2002} who found that a lognormal model was a good fit to simulated $\kappa_{\rm{WL}}$ at angular scales of 2 - 4\arcmin. Increasing the cell radius above 10\arcmin\ removes the clear preference for the lognormal, and the lognormal and Gaussian models fit the data equally well at cell radii of 30\arcmin. The fixed, theoretically derived $\kappa_0 = 0.050$ allows the lognormal model with a single free parameter to fit the distribution well at 15\arcmin, but at larger cell radii this model does very slightly worse than the Guassian model. This suggests that this value of $k_0$ may not be a good estimate for the minimum $\kappa$ in the CiC PDF for larger cells. This makes sense as this $\kappa_0$ corresponds a pure void along the line of sight, which is a decreasingly likely observation as the cell radius increases.
The final row of figure \ref{fig:gal_K_1Dfit_r5_MICE} shows the distribution of $\kappa$ using the sub-sample of MICE with DES-like galaxy density and $n(z)$, and to which DES-like shape noise has been added, as described in section \ref{sec:data_MICE}. The shape noise dominates the resulting distribution, particularly at smaller scales. The width of the distribution of shape noise is 74\% of the width of the noisy $\kappa$ distribution at 40\arcmin, and at 10\arcmin\ it accounts for 89\%. We model the noisy $\kappa$ distribution with a lognormal convolved with Gaussian noise as described in section \ref{sec:LN_theory:WL}, using equation \ref{eq:GxLN}. Again we find that it is not possible to jointly constrain $\kappa_0$ and the width of the lognormal at scales above 10\arcmin\ as they are degenerate. We therefore use the theoretically derived $\kappa_0=0.049$. It can be seen from figure \ref{fig:gal_K_1Dfit_r5_MICE} that at all scales the shape noise makes the noisy $\kappa$ distribution much more Gaussian.
Despite the low signal to noise, at 10\arcmin\ the lognormal convolved with Gaussian noise provides a better fit than the simple Gaussian, with $\chi^2$/DOF$ = 1.06$ and 1.56 respectively. At scales larger than this the Gaussian model performs as well as the lognormal. As with the noise free convergence distribution, the theoretically derived $\kappa_0$ seems to be a less suitable choice at larger scales as the Gaussian model provides a better fit for scales above 30\arcmin.
\subsubsection{Joint galaxy-convergence distribution}
\label{sec:Joint_MICE}
In this sub-section we study the joint distribution of galaxy overdensities and weak lensing convergence and determine to what extent it can be described as a bivariate lognormal distribution. We look at joint distributions using both the full MICE sample, and the subsample with DES-like galaxy density and $n(z)$ and the addition of DES-like shape noise. As in the 1D case, the full sample with higher galaxy density allows us to better capture any lognormal behaviour, and the DES-like sample allows us to compare the results for DES data given in the next section with simulations.
We can make a simple quantitative estimate of the relative correlation of $\delta_{g}$ and $\kappa_{\rm{WL}}$ by calculating the Pearson product-moment correlation coefficient, $r$, for the joint PDF, where
\begin{equation}
\rho_{X,Y} = \frac{\textrm{Cov}(X,Y)}{\sigma_{X}\sigma_{Y}} = \frac{\left< (X-\bar{X}) (Y-\bar{Y})\right>}{\sigma_{X}\sigma_{Y}} .
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.37]{MICE_2Dfit_g_0p2z0p5_K_0p6z1p3_r15_cond_prob.png}
\includegraphics[scale=0.37]{MICE_2Dfit_g_0p1z0p5_K_0p6z1p2_Ng_unmasked_r15_kappa_conv_with_shapenoise.png}
\caption{\textbf{UPPER PANEL:} Joint CiC distribution of weak lensing convergence and galaxy density contrast for MICE simulation at a smoothing scale of 15\arcmin. The top right plot shows the bivariate lognormal fit to MICE simulations. Contours for the simulation are given by the solid blue lines, with dashed magenta contours for the fit. Also shown are the 1D PDFs for $1+\delta_g$ and $1+\hat{\kappa}_{WL}$ individually. PDFs are calculated via the CiC method with cells of radius 15\arcmin. As in the rest of this paper, galaxies are selected over the redshift range $0.1 < z < 0.5$ and WL sources are restricted to the range $0.6 < z < 1.3$. This joint distribution has a Pearson correlation coefficient of $r = 0.83$. \textbf{LOWER PANEL:} Same but with DES-like shape noise added to $\kappa_{\rm{WL}}$. The Pearson correlation coefficient drops to 0.45 with the addition of this shape noise. }
\label{fig:2Dscatter_MICE}
\end{figure}
We begin with the joint distribution of $\delta_{g}$ and $\kappa_{WL}$ with no shape noise, which is shown in the upper panel of Fig. \ref{fig:2Dscatter_MICE} for a smoothing scale of 15\arcmin.
The blue contours in the top right section of this plot show the joint PDF, and the dashed magenta contours show the bivariate lognormal fit.
Since there is no shape noise in this case the bivariate fit is given by equation \ref{eq:biv} but omitting the Gaussian convolution.
We expect the correlation coefficient $\rho$ to be high (close to one) since the galaxies considered are responsible for the lensing. This is indeed what we see: the Pearson correlation coefficient is 0.81 at a smoothing scale 10\arcmin\ and 0.89 at 40\arcmin. We do not see full correlation because the relevant window functions - the lensing efficiency function of the source sample and the galaxy redshift distribution of the galaxy sample - do not overlap precisely.
The lower panel of this figure shows the case where MICE $\kappa_{\rm{WL}}$ has had shape noise added. This noise reduces the correlation of the $\kappa_{\rm{WL}}$ with $\delta_g$, smearing out the joint distribution (shown on the top right of the figure) versus the shape noise free case. The Pearson correlation coefficient is reduced to 0.45.
\subsection{Comparison of Moments}
\label{sec:Joint_MICE}
\begin{figure}
\includegraphics[scale=0.35]
{MICE_mom2_G_LN.png}
\caption{\textbf{UPPER PANEL:} comparison of second moments of MICE galaxy density contrast as a function of smoothing scale, directly measured via CiC and from lognormal and Gaussian fits to the CiC PDF. Blue data points show the ratio of the variance $\left< \delta_g^{2} \right>$ from our fits to the 1D lognormal distribution to that calculated directly from the CiC PDF; black data points show the same but for the Gaussian fit. Data points are offset slightly in scale for clarity. \textbf{LOWER PANEL:} same but for shape noise free MICE weak lensing convergence.}
\label{fig:mom2MICE}
\end{figure}
We can use the second moments to check the validity of the lognormal modelling by comparing the moments derived directly from the CiC with those derived by fitting a lognormal model to the CiC PDF. The second moments of the MICE galaxy and convergence fields $\langle\delta_g^2\rangle$ and $\langle\kappa^2\rangle$ can be calculated from the CiC (as described in Appendix \ref{sec:Ap_mom}). The moments derived under lognormal modelling are given by equations \ref{eq:gg} and \ref{eq:kk}.
First we calculate the variance of the MICE galaxy PDF, shown in the first panel of Fig. \ref{fig:mom2MICE}. Blue data points show the ratio of the variance $\left< \delta_g^{2} \right>$ from fitting a lognormal to the CiC PDF and that calculated directly from the CiC. Errors on $\left< \delta_g^{2} \right>$ directly from CiC are produced by jackknife sampling; errors on the $\left< \delta_g^{2} \right>$ derived from the lognormal fit are from the 1$\sigma$ width of the likelihood of the lognormal width.
The lognormal model gives a good estimate of the variance of the MICE galaxy density contrast distribution (with Poisson shot noise accounted for) at all scales. It gives a better estimate of the variance than a Gaussian model at all scales, and particularly at 10\arcmin.
The lognormal model also gives a good estimate of the variance of the weak lensing convergence distribution at scales up to 20\arcmin. The poorer estimates at 30 and 40\arcmin\ are due to the fact that we fix $\kappa_{min}$ to the theory value at these scales.
These results suggest that within the ranges of scales discussed, the lognormal model can be used to estimate the two point statistics of both the galaxy density contrast and weak lensing convergence distributions to reasonable accuracy in these simulations.
\section{Testing lognormality of DES density and convergence fields}
\label{sec:LN_DES}
Here we repeat the analysis of the previous section with DES galaxy and convergence maps, looking first distributions individually and then at their joint distribution.
\subsection{One-dimensional PDFs and log-normal fits}
\label{sec:1D_DES}
\begin{figure*}
\centering
\includegraphics[scale=0.38]{DES_gal_G_LN_r_10_15_30_log_yaxis_svd.png}
\includegraphics[scale=0.38]
{DES_kappa_G_LNconv_r_10_15_30_fit_w_k0.png}
\caption{\label{fig:smoothing} \textbf{UPPER ROW}: measured 1D PDF of DES galaxies at a smoothing scales of 10, 15 and 30\arcmin. At 10\arcmin\ the Poisson sampled log-normal fit (black) provides a much better fit than the Gaussian (magenta), demonstrating the log-normality of the galaxy CiC distribution at this scale. At larger scales the distribution becomes more Gaussian. \textbf{LOWER ROW}: same but for $\kappa_{\rm{WL}}$. Here the lognormal model includes Gaussian shape noise, which provides a good fit at all scales. Error bars on the counts PDFs are jackknife errors. All $\chi^2$ are per degree of freedom.}
\end{figure*}
Fig. \ref{fig:smoothing} shows 1D CiC PDFs for the DES galaxy density contrast (top row) and $\kappa_{\rm{WL}}$ (second row) fields for different cell radii. The lognormal fit is again shown in black, and for comparison a Gaussian fit is shown in magenta.
For the $\delta_g$ PDF at 10\arcmin\ the lognormal model is clearly favoured, with $\chi^2$/DOF$ = 1.28$ compared to 6.55 for the Gaussian model. This confirms the expected lognormal behaviour at non-linear scales, indicating that our CiC procedure is capturing non-linear clustering information beyond the Gaussian assumption at smaller radii. As in the simulations the $\delta_g$ PDFs clearly appear more Gaussian at larger cell radii, although the lognormal model still provides a better fit than the Gaussian even at 30\arcmin, with $\chi^2$/DOF of 0.97 and 1.82 respectively.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$r$, arcmin & $\sigma$ & $\chi^2_{G}$ & $\chi^2_{LN}$ & DOF & $\langle\delta_g\delta_g\rangle \times 10^{-2}$ \\
\hline
10 & 0.184 & 72.05 & 14.08 &11& $3.44 \pm 0.30$ \\
15 & 0.156 & 27.61 & 15.18 &11& $2.46 \pm 0.28$ \\
20 & 0.146 & 21.48 & 12.36 &12& $2.15 \pm 0.29$ \\
30 & 0.126 & 30.94 & 16.49 &17& $1.60 \pm 0.16$ \\
40 & 0.122 & 22.88 & 18.48 &22& $1.26 \pm 0.17$ \\
\hline
\end{tabular}
\caption{Best-fit parameters and derived statistics from lognormal fits to CiC PDFs of DES galaxy density contrast, for varying cell radii. First first column gives the cell radius, and the second column is the width of the best fitting Poisson-sampled lognormal. The following columns are the minimum $\chi^2$ for Gaussian and Poisson sampled lognormal fits, and the number of degrees of freedom. The final column is the second moment of the best-fitting lognormal PDF, derived from the lognormal width, with $1\sigma$ errors given by the likelihood of the lognormal width.}
\label{table:LN_g}
\end{table}
The best-fit values of the free parameters of the lognormal fits to the DES galaxy density contrast distribution, the $\chi^2$, the number of degrees of freedom (DOF) and the second moment of the best-fit lognormal PDF are given in table \ref{table:LN_g}, for smoothing scales of 10 - 40\arcmin.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$r$, arcmin & $\kappa_{0}$ & $\sigma$ & $\chi^2_{G}$ & $\chi^2_{LN}$ & DOF & $\langle\kappa\kappa\rangle \times 10^{-5}$ \\
\hline
10 & 0.021 & 0.235 & 18.41 & 11.10 &10& $2.44 \pm 0.45 $ \\
15 & 0.017 & 0.248 & 19.17 & 9.09 &9& $1.59 \pm 0.34 $ \\
20 & 0.016 & 0.238 & 10.92 & 4.63 &10& $1.46 \pm 0.29 $ \\
30 & 0.009 & 0.314 & 11.66 & 6.27 &11& $0.84 \pm 0.21 $ \\
40 & 0.008 & 0.300 & 14.82 & 8.58 &13& $0.71 \pm 0.19 $ \\
\hline
\end{tabular}
\caption{Same as Table \ref{table:LN_g} but for DES weak lensing convergence. The lognormal fit accounts for shape noise, so the statistics quoted are for the de-noised $\kappa_{WL}$ distribution. The additional information given vs. Table \ref{table:LN_g} , in the second column, is the best-fit minimum convergence parameter $\kappa_0 = -\kappa_{min}$.}
\label{table:LN_k}
\end{table}
The second row of Fig. \ref{fig:smoothing} shows the DES $\kappa_{\rm{WL}}$ distribution. We find that it is possible to jointly constrain $\kappa_0$, the minimum convergence parameter in equation \ref{eq:WL_LN}, and the width of the lognormal at all smoothing scales. The best-fit values of $\kappa_0$ as well as the lognormal width $\sigma$, best-fit $\chi^2$ and the second moment of the best-fit lognormal PDF are given in table \ref{table:LN_k}. The best-fit $\kappa_0 = 0.021$ and $\sigma = 0.235$ at cell radius 10\arcmin\ are in good agreement with the results from the MICE simulation at this scale, which are 0.023 and 0.226 respectively. Note that for larger scales we fix $\kappa_0$ at the theory value of 0.05 in the simulations, so would not expect close agreement of the best-fit lognormal width with that of the data at the these scales.
The $\kappa_{\rm{WL}}$ distribution appears quite Gaussian at all scales due to the Gaussian shape noise, the distribution of which has a width of 70-90\% of the width of the $\kappa_{\rm{WL}}$ distribution. Despite this low signal to noise, as in the case of simulated $\kappa_{\rm{WL}}$, we find that the lognormal model with Gaussian shape noise (black line) provides a better fit than the simple Gaussian model (magenta line) at small scales. At 10\arcmin\ the lognormal model has $\chi^2 = 1.11$ and the Gaussian 1.84, corresponding to p-values of 0.35 for the lognormal model (i.e. within one $\sigma$) and 0.07 for the Gaussian model.
At 15\arcmin\ the advantage of the lognormal model over the Gaussian is clear, with best-fit $\chi^2$/DOF of 1.01 and 2.13 respectively. At scales larger than this the Gaussian model provides a good fit with best-fit $\chi^2$/DOF of 1.09, 1.06 and 1.14 at 20, 30 and 40\arcmin. The lognormal model is over-fitting the data at these scales, with $\chi^2$/DOF of 0.46, 0.57 and 0.66 at the same scales, so the Gaussian model is sufficient in this regime.
\subsection{Joint galaxy-convergence distribution}
\label{sec:Joint_DES}
\begin{figure}
\centering
\includegraphics[scale=0.36]{DES_2Dfit_g_0p1z0p5_K_0p6z1p3_r15_kappa_conv_Skynet.png}
\caption{Joint distribution of weak lensing convergence and galaxy density contrast for DES at smoothing scale 15\arcmin. Upper right panel: Fit of bivariate lognormal to DES SV data. Contours for the data are given by the solid blue lines, with dashed magenta contours for the fit. Also shown are the individual 1D PDFs for $1+\delta_g$ and $1+\hat{\kappa}_{WL}$. DES Benchmark galaxies are used, selecting the redshift range $0.1 < z < 0.5$ and WL sources from the imshape catalogue are used over the range $0.6 < z < 1.3$. All redshifts are best-fits from the Skynet pipeline. PDFs are calculated via the CiC method with cells of radius 15\arcmin. This joint distribution has a Pearson correlation coefficient of $r = 0.45$.
}
\label{fig:2Dscatter_DES}
\end{figure}
The joint distribution of DES galaxy density contrast and weak lensing convergence data at an angular scale of 15\arcmin\ is shown in the top right panel of fig. \ref{fig:2Dscatter_DES}. The data are shown by the blue contours, and the bivariate fit is shown by the dashed magenta contours. The individual 1D PDFs for $1+\delta_g$ and $1+{\kappa/\kappa_0}$ are also shown.
Before we account for shot noise in the galaxies and shape noise in the convergence, the galaxy counts and $\kappa_{\rm{WL}}$ have a Pearson correlation coefficient of 0.45. This is in line with what we see in the MICE simulations once DES-like shape noise is added (bottom row of Fig. \ref{fig:2Dscatter_MICE}).
Once we account for these sources of noise, the correlation coefficient is 0.82, again in line with the noise-free MICE simulations, where the Pearson correlation coefficient was 0.83 (top row of Fig. \ref{fig:2Dscatter_MICE}).
\subsection{Comparison of Second Moments}
\label{sec:consistency_second_moments}
\begin{figure}
\centering
\includegraphics[scale=0.35]{DES_Skynet_v7_pix2_mom2_G_LN.png}
\caption{Same as figure \ref{fig:mom2MICE}, but for DES galaxies (upper panel) and convergence (lower panel).
\label{fig:mom2DES}
\end{figure}
\begin{table*}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$r$, arcmin & $\langle \delta_g\delta_g\rangle$ & $\langle \kappa\kappa\rangle$ & $\langle\ \kappa\kappa\rangle_{SN}$ \\
\hline
10 & $(3.70 \pm 0.22)\times 10^{-2}$ & $(2.52 \pm 0.41)\times 10^{-5}$ & $(1.00 \pm 0.03)\times 10^{-4}$ \\
15 & $(2.76 \pm 0.20)\times 10^{-2}$ & $(1.69 \pm 0.32)\times 10^{-5}$ & $(4.68 \pm 0.18)\times 10^{-5}$ \\
20 & $(2.26 \pm 0.18)\times 10^{-2}$ & $(1.39 \pm 0.25)\times 10^{-5}$ & $(2.60 \pm 0.16)\times 10^{-5}$ \\
30 & $(1.65 \pm 0.14)\times 10^{-2}$ & $(9.84 \pm 1.66)\times 10^{-6}$ & $(1.18 \pm 0.06)\times 10^{-5}$ \\
40 & $(1.38 \pm 0.16)\times 10^{-2}$ & $(8.40 \pm 1.30)\times 10^{-6}$ & $(6.75 \pm 0.44)\times 10^{-6}$ \\
\hline
\end{tabular}
\caption{Second moments of DES galaxy density contrast and weak lensing convergence, as calculated by CiC, for different cell radii. Shot and shape noise have been accounted for, and these are the de-noised moments. The final column gives our estimate the shape noise of the weak lensing convergence. This is derived from the 100 realisations of the $\kappa_{\rm{WL}}$ map with randomised shears, which we find to agree with the second moment of the $\kappa_{\rm{WL}}$ B-mode within 2\%.}
\label{table:ellcuts}
\end{table*}
In this section we check the validity of the lognormal model by comparing second moments derived from the log-normal assumption with those measured directly from the data.
The variance of the DES galaxy PDF is shown in the first panel of Fig. \ref{fig:mom2DES}. Blue data points show the ratio of the variance $\left< \delta_g^{2} \right>$ from our fits to the 1D lognormal distribution to that calculated directly from the CiC PDF. Errors on $\left< \kappa^{2} \right>$ directly from CiC are produced by jackknife sampling; errors on $\left< \kappa^{2} \right>$ derived from the lognormal fit are from the 1$\sigma$ width of the likelihood of the lognormal width. The second panels shows the same for DES $\kappa_{\rm{WL}}$.
The lognormal model with appropriate noise contribution gives an estimate of the variance that is consistent with that calculated directly from the CiC, for both galaxies and $\kappa_{\rm{WL}}$, at all scales from 10 - 40\arcmin.
For the galaxy density contrast distribution, the Gaussian model provides a less accurate estimate of the variance calculated directly from the CiC at all scales.
For the convergence distribution the Gaussian model again gives variance estimates less accurate than the lognormal model at all scales.
For both galaxies and weak lensing convergence, the Gaussian and lognormal approaches underestimate the variance as compared to measuring it directly from the CiC. This is because in constructing the CiC PDF to which we fit the lognormal model, we bin the cell counts. We account for noise via singular value decomposition, and one of the things this removes is contributions to the fit from the outermost bins, which have very few cell counts. This makes the effective distribution narrower, with lower second moment, than if these noisy data points were included. In calculating the variance directly from the CiC (as described in Appendix \ref{sec:Ap_mom}) this binning is not necessary and all cells, including those with the most extreme values of $\delta_g$ or $\kappa_{\rm{WL}}$, are included in the calculation, resulting in a larger variance in $\delta_g$ or $\kappa_{\rm{WL}}$. This effect is less stark in the MICE simulations where there are a greater number of galaxies than in the DES data, so fewer bins are discarded due to low counts of cells. This underestimation of the variance, however, is not significant within the errors.
\section{Discussion}
\label{sec:conc}
We have tested the lognormality of the DES galaxy density contrast and weak lensing convergence PDFs at angular scales of 10 - 40\arcmin\ (corresponding to physical scales of 3 - 10 Mpc at median redshift $z=0.3$). In the context of this work, estimating the CiC PDF is a way of quantifying the non-linear growth of mass and galaxy fluctuations, as well as the visual impression of comparing the $\kappa_{\rm{WL}}$ mass maps with the galaxy distribution on the same patch of the sky. It is also a test of systematics.
\vspace{5mm}
Our main findings are as follows:
\begin{itemize}
\item In agreement with many earlier papers we find that the 1D DES galaxy PDF is well fitted by a lognormal model, taking into account Poisson shot noise, with best-fit $\chi^2$/DOF$ = 1.28$ vs. 6.55 for a Gaussian model at a scale of 10\arcmin.
\item In modelling the weak lensing convergence distribution it is important to account for shape noise since the width of this noise is a significant fraction (70-90\%) of the width of the $\kappa_{\rm{WL}}$ signal. We find that the shape noise estimate derived from the 100 realisations of DES $\kappa_{\rm{WL}}$ in which the shears have been randomised agrees with that of the $\kappa_B$ mode within 2\% at all scales from 10 to 40\arcmin, and that the distribution of the shape noise can be well modeled by a Gaussian PDF. This allows us to model the $\kappa_{\rm{WL}}$ distribution with a lognormal convolved with Gaussian PDF. In future work it would be interesting to investigate the spatial correlation of this noise.
\item The convergence field is not expected to be exactly lognormal even if the mass density contrast field is, as it is a weighted projection of the mass density field along the line of sight. We find however, in agreement with previous work on simulations, that the 1D $\kappa_{\rm{WL}}$ PDF is well fitted by a lognormal model, taking into account shape noise. This is the first such measurement on data. The best-fit $\chi^2$/DOF for the lognormal model is 1.11, compared to 1.84 for a Gaussian model, corresponding to p-values of 0.35 (i.e. within one $\sigma$) and 0.07 respectively. At scales above 15\arcmin\ the Gaussian model is a sufficient approximation.
\item The bivariate ($\kappa_{\rm{WL}}$, $\delta_g$) PDF is also well fitted by a bivariate lognormal.
\item De-noised second moments derived via the lognormal fit are consistent with variances derived directly from the data up to scales of 40\arcmin, for both the DES galaxy density contrast and weak lensing convergence distributions.
\end{itemize}
This pilot study could be extended to much larger areas with weak lensing surveys such as the full DES (5000 deg$^2$) survey, LSST (20,000 deg$^2$) and Euclid (15,000 deg$^2$). In this work we have tested the lognormality of the $\kappa_{\rm{WL}}$ PDF; with the higher signal/noise that future surveys will provide it might be possible to deduce from the observed $\kappa_{\rm{WL}}$ PDF whether or not the underlying matter density field is lognormal - essentially inverting equation 4.
In this work we have used the CiC to probe lognormality, but there is a wealth of information contained within it that could be exploited in future work. The CiC contains the full PDF so as well as the variance, higher order moments such as skewness and kurtosis can also be extracted.
The method used in this work, of constructing PDFs via CiC and cross correlating them, could be used to extract information on galaxy bias and to derive cosmological parameters. It could also be interesting to repeat this analysis using manipulations of the shear field than other $\kappa_{\rm{WL}}$ that avoid the reconstruction noise due to the Kaiser Squires method.
Quantifying $P(\kappa_{\rm{WL}})$ will be important for the emerging field of mass reconstruction using $\kappa_{\rm{WL}}$, since it is required as a prior input for this process. We have demonstrated that a lognormal model is a better choice than a Gaussian model at scales of 10 - 20\arcmin. As well as the improved ability to capture non-linear behaviour versus a Gaussian model, the lognormal model still allows fast production of, for example, simulated realisations of the convergence field for testing, and covariance matrices.
|
\section{Introduction}
Professor Nambu made great achievements in so much inexhaustible depth and wideness, and thus it may be something like the picture of ``The Blind Men and the Elephant'' to talk
about only a single aspect of his physics. But the subject I am going to talk about is not just one of them, but probably his most influential one. In fact the 2018 Nobel Prize announcement is
``for the discovery of the mechanism of spontaneous broken symmetry in subatomic physics'' \cite{2008Nobel}.
His spontaneous symmetry breaking (SSB) \cite{Nambu:1961tp} was the theory for the origin of mass of the nucleon, then an elementary particle, generated dynamically from nothing (the vacuum) through
the nucleon-antinucleon pair condensate. Although the nucleon is no longer an elementary particle, the essence of his mechanism to generate mass of composite particles, hadrons including the nucleon, as well as the near masslessness of the
composite pion,
is now realized through the quark-antiquark
condensate in the underlying theory QCD. This mass constitutes 99 \% of mass of the nucleon, namely of the ordinary matter made out of the atoms, and thus the Nambu's theory already accounted for the
origin of the dominant part of the mass of the visible world.
The problem of the origin of mass in the modern particle physics is only for the rest 1\% of the mass of the matter, the elementary particles
of the Standard Model (SM), which is attributed to
the Higgs boson whose origin is still mysterious. I will discuss that this 1\% may also be explained by the dynamical symmetry breaking in some underlying theory, similarly to the Nambu's theory.
Origin of Mass of all the SM particles is the Higgs VEV $v=\sqrt{\frac{-\mu_0^2}{\lambda}}=246\,{\rm GeV}$ or the Higgs mass $M_\phi^2 =2\lambda v^2=-2\mu_0^2$ read from the SM Higgs Lagrangian:
\begin{eqnarray}
{\cal L}_{\rm Higgs}&=&
|\partial_\mu h|^2 -\mu_0^2 |h|^2 -\lambda|h|^4
\label{Higgs}\\
&=& \frac{1}{2} \left[
\left(\partial_\mu {\hat \sigma}\right)^2 +\left(\partial_\mu {\hat \pi_a}\right)^2
\right]
-\frac{1}{2}
\mu_0^2 \left[
{\hat \sigma}^2+{\hat \pi_a}^2
\right]-\frac{\lambda}{4} \left[
{\hat \sigma}^2+{\hat \pi_a}^2
\right]^2
\label{sigma}\\
&=& \frac{1}{2} {\rm tr} \left( \partial_\mu M\partial^\mu M^\dagger \right)
- \left[\frac{\mu_0^2}{2} {\rm tr}\left(M M^\dagger\right)+\frac{\lambda}{4} \left({\rm tr}\left(M M^\dagger\right)\right)^2\right] \,,
\label{sigmaMatrix}
\end{eqnarray}
where we have rewritten the conventional form in Eq.(\ref{Higgs}) into the Gell-Mann-Levy (GL) $SU(2)_L \times SU(2)_R$ linear sigma model \cite{GellMann:1960np} in Eq.(\ref{sigma}) through
\begin{equation}
h=\left(\begin{array}{c}
\phi^+\\
\phi^0\end{array} \right)=\frac{1}{\sqrt{2}} \left(\begin{array}{c}
i{\hat \pi}_1+{\hat \pi}_2\\
\hat{\sigma}-i {\hat \pi}_3\end{array}\right)\,,
\end{equation}
and further into Eq.(\ref{sigmaMatrix}) with the $2\times 2$ matrix $M$
\begin{equation}
M=(i \tau_2 h^*, h) = \frac{1}{\sqrt{2}}\left({\hat \sigma}\cdot 1_{2\times 2} +2i {\hat \pi}\right)\, \quad \left({\hat \pi} \equiv {\hat \pi}_a \frac{\tau_a}{2}\right)\,,
\end{equation}
which transforms under $G=SU(2)_L\times SU(2)_R$ as:
\begin{equation}
M \rightarrow g_L \, M\, g_R^\dagger \,,\quad \left(g_{R,L} \in SU(2)_{R,L}\right)\,.
\end{equation}
Then the origin of mass
is attributed to the mysterious input mass parameter of the {\it tachyons}, $\hat \pi$ and $\hat \sigma$ (not physical particles), with the mass $\mu_0$ such that
\begin{equation}
\mu_0^2<0
\end{equation}
{\it as a free parameter}.
But why the tachyon? How is the tachyon mass determined?
SM cannot answer to these questions, even though the Higgs boson has been discovered with the mass near 125 GeV.
Historically, the GL linear sigma model in the form of Eq.(\ref{sigma}) as the prototype of the Higgs Lagrangian Eq.(\ref{Higgs}) was proposed for phenomenologically describing the pion (as well as the nucleon)
{\it without concept of the SSB}, while the Nambu-Jona-Lasinio (NJL) model \cite{Nambu:1961tp} explained the same property at the deeper level in terms of the dynamical symmetry breaking
due to the vacuum property.
Here we should
recall that
{\it the SSB was born as a dynamical symmetry breaking (DSB)}, where the tachyons are in fact generated as composites of the dynamical consequence of the strong dynamics, but not ad hoc inputs as in the GL theory. Actually, the GL theory is now regarded as an effective theory (macroscopic theory) for the NJL model as a microscopic theory, as is the Ginzburg-Landau (GL) theory
for the Bardeen-Cooper-Schrieffer (BCS) theory \cite{BCS} for the superconductor. So history may repeat itself.
I first discuss that although his idea was motivated by the BCS theory, it was not just a copy of it but essentially new in the most important
aspect, namely it created a new dynamics, although based on the same kind of four-fermion interaction: the NJL dynamics is the {\it strong coupling theory having non-zero critical coupling} to separate the SSB phase with $\mu_0^2<0$ (above the critical coupling) from the non-SSB phase with $\mu_0^2>0$ (below the critical coupling). It is in sharp contrast to the BCS theory which is a {\it weak coupling theory having the zero critical coupling}, always in the SSB phase $\mu_0^2<0$ even for infinitesimal (attractive) coupling,
due to the Fermi surface. The Fermi surface reduces the effective dimensions by 2 so as to make the theory in effectively $1+1$ dimensions like the Thirring model and/or Gross-Neveu model.
The non-zero critical coupling is also hidden in the asymptotically free gauge theories including the QCD: it reveals itself in the chiral symmetry restoration
when the system is in the extreme condition such as the
high temperature, high density, and large number of light fermions, where the coupling cannot grow above the hidden critical coupling in the infrared region as strong enough to form the fermion-antifermion
condensate.
The existence of the non-zero critical coupling in the gauge theory was first recognized by Maskawa-Nakajima \cite{Maskawa:1974vs} in the ladder Schwinger-Dyson (SD) equation, a gauge theory analogue of the NJL gap equation. The solution of the ladder SD equation in the weak coupling region (existing even for the infinitesimal coupling) \cite{Johnson:1964da} disappears at zero fermion bare mass at finite cutoff,
which is actually the explicit chiral symmetry breaking solution vanishing when the cutoff is removed. \footnote{
At TV interview after announcement of the Nobel prize together with Professor Nambu in 2008, Toshihide Maskawa confessed that the paper he most studied was the Nambu's paper on SSB,
``I exhausted it''. He
in fact discovered the non-zero critical coupling for SSB in the gauge theory \cite{Maskawa:1974vs} not just in the NJL four-fermion model.
At that time I was a graduate student at Kyoto University to hear it first hand and have been influenced by this work, strong coupling gauge theory (SCGT) with non-zero criticality, ever since. See Nagoya SCGT workshops, http://www.kmi.nagoya-u.ac.jp/workshop/SCGT15/ }
Although the asymptotically-free gauge theory like QCD has no explicit critical coupling to divide the SSB phase from non-SSB phase (having only a single phase of SSB), the running coupling always becomes strong in the infrared region where the coupling exceeds a hidden critical coupling to trigger the SSB having the condensate of order of the scale of this mass region \cite{Higashijima:1983gx}.
The main purpose of this article is to describe the expanding horizon of such a strong coupling dynamics characterized by the non-zero critical coupling initiated by Professor Nambu in view of the
modern version of the
origin of mass,
namely the composite Higgs models having large anomalous dimension. First, the (weakly gauged) strong coupling four-fermion models like the top quark condensate model \cite{Miransky:1988xi,Nambu:1989jt,Bardeen:1989ds}, where only
the top quark has the strong coupling above the criticality (anomalous dimension $\gamma_m \simeq 2$ \cite{Miransky:1988gk}) so as to be responsible for the electroweak symmetry breaking \cite{Miransky:1988xi}. Second,
the gauge theories such as the walking technicolor based on the
near conformal gauge theory just above the criticality having anomalous dimension $\gamma_m\simeq1$ and a
composite dilaton (technidilaton) as the composite Higgs \cite{Yamawaki:1985zg,Bando:1986bg}.
The technidilaton in the walking technicolor has been shown to be consistent with the 125 GeV Higgs at the present LHC experimental data \cite{Matsuzaki:2012mk,Matsuzaki:2012xx, Matsuzaki:2015sya}.
Before discussing possible underlying theory for SM, I show that the SM Higgs Lagrangian itself already has some hints for the
theory beyond the SM. It was shown \cite{Fukano:2015zua} that the SM Higgs Lagrangian itself possesses nonlinearly realized ``hidden'' symmetries (scale symmetry and Hidden Local Symmetry (HLS) \cite{Bando:1984ej,Bando:1987br, Harada:2003jx}, both spontaneously broken), in addition to the well-known symmetry, nonlinearly realized global $SU(2)_L\times SU(2)_R$ chiral symmetry (also spontaneously broken)
to be gauged by the electroweak symmetry. It is in fact straightforward to show \cite{Fukano:2015zua} that the SM Higgs Lagrangian is cast into the scale-invariant nonlinear chiral Lagrangian \cite{Matsuzaki:2012mk},
and then
further shown to be
gauge equivalent to the scale-invariant HLS (s-HLS) Lagrangian \cite{Kurachi:2014qma}. The SM Higgs is nothing but a (pseudo-) dilaton! \cite{Fukano:2015zua}
{\it This is the very nature of the SM Higgs Lagrangian, quite independent of details of the
possible underlying theory as the UV completion.}
Also the HLS can
naturally accommodate the vector bosons, analogues of the rho mesons in the QCD, into the SM (``SM rho meson'') \cite{MY16}:
It would be the simplest extension of the SM to account for the 2 TeV diboson events at LHC \cite{Aad:2015owa}.
Then I elaborate \cite{Yamawaki:2015tmu} on the well-known fact that the NJL model
can be regarded as the microscopic theory (underlying theory or ultraviolet (UV) completion) for the SM Higgs Lagrangian,
or the GL linear sigma model,
as the macroscopic theory (effective theory) at composite level. With the coupling larger than the non-zero critical coupling, the NJL model equivalent to the SM Higgs has also the nonlinearly realized hidden (approximate) scale symmetry for the SM Higgs as a composite pseudo-dilaton (``NJL dilaton''), together with the HLS for the dormant composite spin 1 boson (``NJL rho meson'') as a possible
candidate for the LHC diboson events \cite{Aad:2015owa}.
Although both are trivial theories having no interaction in the infinite cutoff limit (Gaussian fixed point), I will discuss possible way out, one \cite{Kondo:1991yk} being the gauged NJL model
in combination with the walking gauge theory, another \cite{Yamawaki:2015tmu} the recently suggested different way of the continuum limit where the composite Higgs becomes massless
(up to the trace anomaly) as the pseudo-dilaton in the same sense as the SM Higgs.
The simplest possibility for such a composite model would be the top quark condensate model (``top-mode SM'') \cite{Miransky:1988xi,Nambu:1989jt,Bardeen:1989ds}, where crucial is the
non-zero criticality \cite{Miransky:1988xi}: only top (may also bottom) has the coupling larger than the non-zero critical coupling to acquire the dynamical mass due to SSB.
Near the scale-invariant limit, the top-mode dilaton may be the 125 GeV Higgs, and the HLS gauge boson (``top-mode rho meson'') may be identified with the recent 2 TeV diboson excess (and the top-mode axion, $\bar b b$ bound state, may be identified with the 750 GeV diphoton excess at LHC \cite{750} which was reported after this symposium).
We then discuss the walking technicolor proposed based on the SSB solution of the ladder SD equation to have a large anomalous dimension $\gamma_m=1$ and technidilaton as a pseudo-Nambu-Goldstone (NG) boson of the approximate scale symmetry \cite{Yamawaki:1985zg,Bando:1986bg}. Such a scale-symmetric walking gauge theory may be realized when flavor number $N_F$ of massless fermions
is large in the asymptotically free gauge theory (``large $N_F$ QCD'') \cite{Appelquist:1996dq,Miransky:1996pd}, with $N_F (\gg 2)$ slightly smaller than that having an infrared fixed point (conformal window) where the coupling in the infrared region is almost constant and below the critical coupling
so that the SSB does not take place. The effective theory of the walking technicolor is the s-HLS Lagrangian with a larger chiral symmetry $SU(N_F)_L \times SU(N_F)_R$, with typically $N_F=8$ (one-family model), precisely the same type of the s-HLS as in the case for the SM Higgs Lagrangian with $SU(2)_L \times SU(2)_R$. The technidilaton as a composite Higgs has been shown \cite{Matsuzaki:2012mk,Matsuzaki:2012xx, Matsuzaki:2015sya} to be
consistent with the
present LHC 125 GeV Higgs, and the HLS vector mesons (walking technirhos) have also been shown \cite{Fukano:2015hga} to be consistent with the LHC diboson events \cite{Aad:2015owa}.
(We also showed \cite{Matsuzaki:2015che}
that one of the technipions can be identified consistently with the 750 GeV diphoton events at LHC \cite{750} reported after the symposium).
Several theoretical issues are discussed such as the recent lattice studies of the walking theories, as well as the ladder, and the renormalizability of the gauged-NJL model and the conformal phase transition, etc.
\section{NJL the Strong Dynamics vs. BCS the Weak Dynamics}
It is widely believed that the NJL model is a copy of the BCS. Here I emphasize that they are essentially different dynamics,
NJL as the strong coupling with critical coupling no-zero, while the BCS as a weak coupling with the critical coupling zero.
The difference comes from the Fermi surface in BCS which reduces the effective phase space from 3+1 to 1+1, while the NJL case is in the free space
of full 3+1 dimensions. The attractive forces are more efficient in smaller phase space.
Let us start with the $SU(2)_L \times SU(2)_R$ NJL model \cite{Nambu:1961tp} for $N_C$ 2-flavored Dirac fermions $\psi$:
\begin{equation}
{\cal L}_{\rm NJL} = \bar \psi i\gamma^\mu\partial_\mu \psi + \frac{G}{2} \left[ (\bar \psi \psi)^2 +(\bar \psi i \gamma_5 \tau^a \psi)^2\right]\,.
\end{equation}
When the fermion-antifermion condensate in the vacuum takes place, $\langle 0|\bar \psi \psi|0\rangle \ne0$, it reads
\begin{equation}
{\cal L}_{\rm NJL} -\bar \psi i\gamma^\mu\partial_\mu \psi = \ G \langle \bar \psi \psi\rangle\, \bar \psi \psi +\cdots=
-m_F \, \bar \psi \psi +\cdots\,.
\end{equation}
At the $1/N_C$ leading order this yields the self-consistent
NJL gap equation for the dynamical mass $m_F$ of $\psi$:
\begin{equation}
m_F= -G \langle \bar \psi \psi\rangle=G\, {\rm Tr}[
S_F(p)] =G \cdot 4 N_C \int \frac{d^4p}{i (2\pi)^4} \frac{m_F}{m_F^2-p^2}
\label{NJLgap}
\end{equation}
which has an SSB solution $m_F\ne 0$:
\begin{equation}
\frac{1}{G}-\frac{1}{G_{\rm cr}} =\frac{\Lambda^2}{4\pi^2} \left(\frac{1}{g} -\frac{1}{g_{\rm cr}}\right)= - \frac{1}{4\pi^2} N_C m_F^2 \ln \left(\frac{\Lambda^2}{m_F^2}
\right) <0 \,,
\label{SSBNJL}
\end{equation}
only for the strong coupling
\begin{equation}
G>G_{\rm cr}=\frac{4\pi^2}{N_C \Lambda^2}\ne 0
\quad
(g\equiv\frac{G\Lambda^2}{4\pi^2} >g_{\rm cr}=\frac{1}{N_c}\ne 0)\,.
\label{strong}
\end{equation}
We shall later discuss that this in fact corresponds to the tachyon mass $\mu_0^2<0$ in Eq.(\ref{sigma}): $\mu_0^2 = (\frac{1}{G} - \frac{1}{G_{\rm cr}})\cdot Z_\phi^{-1}= - 2m_F^2<0$, where
$Z_\phi=\frac{N_C}{8\pi^2} \ln \frac{\Lambda^2}{m_F^2}$. The (composite) tachyon has been induced dynamically by the $-1/G_{\rm cr}$ term due to the loop effects in the large $N_C$ limit. Of course the tachyon is not a physical particle, which simply implies instability of the trivial vacuum with $m_F=0$
(no SSB).
For the weak coupling $G<G_{\rm cr}$ there exists only the non SSB solution $m_F\equiv 0$ where no tachyon exists.
Note that ``strong coupling'' as defined by non-zero critical coupling
does not necessarily mean
numerically strong, particularly in the large $N_C$ limit $g_{\rm cr} = 1/N_C \ll 1$. However, the attractive forces in the condensate channel are not from a single fermion but actually from sum of all the $N_c$ fermions coherently, which ends up with really strong $N_C g_{\rm cr} = {\cal O}(1)$. [As we discuss later,
this also applies to the strong coupling gauge theory where $N_C \alpha_{\rm cr} ={\cal O} (1)$, while the gauge coupling criticality itself $\alpha_{\rm cr} \sim 1/N_C \ll 1$
is negligibly small (but non-zero) in the large $N_C$ limit].
In contrast to the non-zero critical coupling of the NJL model, the BCS theory for the superconductor has the zero critical coupling (``weak coupling theory'')
due to the electron Fermi surface $E_F= \frac{p_F^2}{2m_e}$ ($m_e$: electron mass in the free space), which affects the fermion-fermion condensate $\langle \psi \psi\rangle\ne 0$ (dynamical Majorana mass $\Delta$ as the gap) instead of fermion-antifermion condensate $\langle \bar \psi \psi\rangle\ne 0 $ . The essence can be read from the effective dimension of the momentum in the integral of the gap equation of the BCS $\int \frac{d^4 p}{i (2\pi)^4}=\int \frac{d^3 p}{(2\pi)^3} \frac{d \omega}{i (2\pi)}$:
\begin{equation}
\int \frac{d^3 p}{(2\pi)^3}
=\int \frac{(4\pi p^2) d p}{(2\pi)^3}
\Rightarrow
4\pi p_F^2 \int_{|E(p)-E_F|<\omega/2} \frac{d p}{(2\pi)^3} = \frac{N}{2} \int_{E_F-\omega_D/2}^{E_F+\omega_D/2} d E(p) \,,
\end{equation}
where $E(p)\equiv \frac{p^2}{2m_e}$ and $N\equiv \frac{m_e p_F}{\pi^2}=$ constant:
The $3-$dimensional electron momentum $\overrightarrow{p}$ is confined to a one-dimensional direction normal to the Fermi surface $E_F=\frac{p_F^2}{2m_e}$ in the
narrow energy shell bounded by the Debye energy $\omega_D$ (cutoff). After integral $\int d \omega$, with the fermion propagator $S_F(p)^{(Majorana)}={\cal F.F.} \langle T(\psi(x) \psi(0)\rangle
= \Delta/(\omega^2 -|\Delta|^2 -E(p)^2)$ (
instead of ${\cal F.F.} \langle T(\bar \psi(x) \psi(0)\rangle$) ), the BCS gap equation
corresponding to Eq.(\ref{NJLgap}) (Majorana mass without
factor 4) reads
\begin{equation}
|\Delta| = G\frac{N}{2} \int_0^{E_F+\omega_D/2} d E(p) \frac{|\Delta| } {\sqrt{|\Delta|^2+E(p)^2}}\sim |\Delta| \left[ \frac{N G}{2} \ln\frac{|\Delta|}{\omega_D}
\right]\,.
\label{gapBCS}
\end{equation}
Then the SSB solution with $|\Delta|\ne 0$ exists even for infinitesimal coupling $1\gg N G >N G_{\rm crit}=0$:
\begin{equation}
|\Delta| \sim \omega_D \exp\left(-\frac{2}{N G}\right)\,,\quad \left(1\gg N G >N G_{\rm cr}=0\right)\,,
\label{BCSgap}
\end{equation}
in contrast to the SSB solution in NJL model in Eq.(\ref{SSBNJL}) with $N_C g>N_C g_{\rm cr} ={\cal O} (1) $ in Eq.(\ref{strong}).
The result is intuitively obvious: The fermion pair in the one dimensional space is ``bound'' even for infinitesimal coupling, since there is no way
to escape from each other, while that in the higher dimensional space can freely move from each other and hence needs strong attractive forces
to bind it together. This is the effective dimensional reduction.
The situation that the lower dimensional theory lowers the critical coupling can be viewed explicitly by the $D (1+1<D<3+1)$ dimensional four-fermion theory, the Gross-Neveu model, with
$D$ changed continuously \cite{Kikukawa:1989fw}. The gap equation is simply changed as $\int \frac{d^4p}{i(2\pi)^4} \Rightarrow \int \frac{d^D p}{i (2\pi)^D}$ in Eq. (\ref{NJLgap}). Similarly to Eq.(\ref{SSBNJL}), the SSB solution exists:\cite{Kikukawa:1989fw}
\begin{equation}
\frac{1}{g} -\frac{1}{g_{\rm cr}}
=- \frac{N_C\xi_D}{2- \frac{D}{2}} \cdot \left(\frac{m_F}{\Lambda}\right)^{D-2} <0
\,,
\label{SSBDNJL}
\end{equation}
only for the strong coupling;\footnote{
In $D>4$ dimensions, the form $g_{\rm cr}= \frac{1}{N_C} \left(\frac{D}{2}-1\right) $ remains the same, in accord with the above intuitive picture for the
required binding force strength depending on the phase volume,
while the gap equation takes a similar but different form:
$ 1/g -1/g_{\rm cr}
=- \frac{N_C}{D/2-2} \cdot \left(\frac{m_F}{\Lambda}\right)^2$ \cite{Hashimoto:2000uk}\,.
}
\begin{equation}
g \,\, > \,\, g_{\rm cr}= \frac{1}{N_C} \left(\frac{D}{2}-1\right) \rightarrow 0 \quad (D\rightarrow 2),
\end{equation}
where $g \equiv G\Lambda^{D-2} \, \frac{2^{D/2}}{(4\pi)^{D/2} \Gamma(D/2)}$ and
$\xi_D= B(\frac{D}{2}-1,3-\frac{D}{2}) \rightarrow \frac{1}{D/2-1}=\frac{1}{N_C g_{\rm cr}}$ for $D\rightarrow 2$ ($\rightarrow 1 $ for $D\rightarrow 4$) \footnote{
If we take the $D\rightarrow 4$ limit, on the other hand, the gap equation is reduced to Eq.(\ref{SSBNJL}) except for the logarithmic factor. This log factor is a crucial difference between the $2< D<4$ and the $D=4$ four-fermion theories. As we discuss later, the former is renormalizable in $1/N_C$ expansion having the nontrivial fixed point at $g=g_{\rm cr}$ in the beta function, $\beta(g) =- \frac{D-2}{g_{\rm cr}} \, g\,(g-g_{\rm cr})$ \cite{Kikukawa:1989fw}, while the latter is not, a trivial theory, with the beta function having the Gaussian fixed point at $g=g_{\rm cr}$. ($g=0$ is an infrared fixed point defining the infrared free theory, with $g<0$ being the repulsive
forces.)
}.
The critical coupling $g_{\rm cr}$ indeed decreases as $D$ does to vanish at $D=2$.
This yields for $D\rightarrow2$ the well-known result:
\begin{equation}
m_F=\Lambda \exp \left(-\frac{1}{2N_C g} \right)\,,\quad \left(N_C g > N_C g_{\rm cr}=0\right)\,,
\label{D2NJL}
\end{equation}
which is of the same form as Eq.(\ref{BCSgap}).
Thus the BCS dynamics in some sense is similar to the $D=1+1$ four-fermion theories
such as the Thirring model and the Gross-Neveu model.
There is a caveat \cite{Witten:1978qu}, however: the genuine $D=1+1$ dimensional theory is not actually in the SSB phase in accord with the Merwin-Wagner-Coleman theorem, although it has a massless bound state and a massive
fermion with mass of the form of Eq.(\ref{D2NJL}) in the large $N_C$ limit, similarly to the SSB phase. However, the absence of the NG boson and lack of SSB does not apply to the BCS theory in contrast to
the Thirring model and Gross-Neveu model, since the BCS theory
is not a genuine $1+1$ dimensional model but rather a brane model: only fermions (not anti-fermions) are confined to the $1+1$-brane, the Fermi surface, a consequence of the Fermi statistics, while
the fermion-fermion pair composite NG boson as a boson lives freely from the Fermi surface in the full $3+1$ dimensional bulk,
and hence SSB and NG boson do exist, in accord with the superfluidity and superconductor.
To summarize the Nambu's approach to the origin of mass,
the theory having intrinsic mass scale $\Lambda\sim G^{-1/2}$ may or may not produce the particle mass $m_F$, depending on the coupling strength:
the strong coupling dynamics for $G>G_{\rm cr}\ne 0$ creates the composite tachyon
with negative ${\rm mass}^2$ $\mu_0^2 \sim 1/G -1/G_{\rm cr} = -\frac{N_C}{4\pi^2} m_F^2 \ln \frac{\Lambda^2}{m_F^2} <0$ in such a way that the particle mass $m_F$ is generated from the intrinsic mass scale
$\Lambda$. By
fine tuning the strong coupling $G (>G_{\rm cr}\ne 0)$ as $G\simeq G_{\rm cr}$, we can arrange a big hierarchy $m_F\ll \Lambda$ (near chiral symmetry restoration).
On the other hand, for the weak coupling $G<G_{\rm cr}$ there
exists no particle mass $m_F\equiv 0$, although the theory has an intrinsic mass scale $\Lambda$.
This is an essential difference from the BCS theory
which has a zero critical coupling, producing always a non-zero gap $\Delta\ne 0$ even for the infinitesimal coupling.
As discussed later, this non-BCS phase structure of the NJL dynamics
was in fact the original motivation of
the top quark condensate model of Ref. \cite{Miransky:1988xi}, where the top
quark having a coupling larger than the critical coupling is discriminated from others having those smaller than the critical coupling, so that
only the top has mass of order of weak scale in such a way as to produce only three NG bosons
responsible for the electroweak symmetry breaking.
This is in contrast to the ``bootstrap symmetry breaking'' \cite{Nambu:1989jt} which is based on the BCS dynamics without the notion of the
non-zero criticality.
\section{Strong Coupling Gauge Theories for the Origin of Mass}
\label{SCGT}
The dynamical mass of the fermion $m_F$ picks up the intrinsic scale $\Lambda$ (cutoff) which regularizes the theory and brings the explicit breaking of the scale symmetry
corresponding to the trace anomaly in the renormalized quantum theory. In the asymptotically free gauge theory $\Lambda$ can be identified with
the renormalization-group invariant intrinsic scale such as $\Lambda_{\rm QCD}$ induced by the perturbative trace anomaly, as we discuss later.
There also exists a non-zero critical coupling for SSB in the gauge theory with massless fermion,
as first noted \cite{Maskawa:1974vs} in the ladder SD equation,
with non-running coupling $\alpha(\mu^2)\equiv \alpha=g^2/(4\pi)$ in the Landau gauge,
a straightforward extension of the NJL gap equation Eq. (\ref{NJLgap}), this time for the fermion mass function $\Sigma(-p^2)$ instead of the constant mass $m_F$
(For details see e.g., Ref.\cite{Matsuzaki:2015sya}):
\begin{equation}
S_F^{-1}(p) \ =\ S^{-1}(p) + \int \frac{d^4 k}{(2 \pi)^4}\
C_2\,
g^2
D_{\mu\nu} (p-k)
\gamma^\mu \ S_F(k) \ \gamma^\nu,
\label{eq:SDeq}
\end{equation}
where $i S_F^{-1}(p)=Z^{-1} (-p^2) (\xbar{p} - \Sigma(-p^2))$ and $i S^{-1}(p)=(\xbar{p} - m_0)$ are the full and bare fermion inverse propagators, respectively, and
$i D_{\mu\nu}(p)= (g_{\mu\nu} -p_\mu p_\nu/p^2)/p^2$ the bare gauge boson propagator in the Landau gauge, and $C_2$ is the quadratic Casimir of the fermion of the gauge theory, with $C_2= (N_C^2 -1)/(2N_C)$ for
the fundamental representation in $SU(N_C)$.
After the angular integration, the
ladder SD equation in Landau gauge for $\Sigma (x\equiv -p^2)$
reads:
\begin{equation}
\Sigma(x) = m_0 +
\frac{3 C_2}{4\pi} \alpha \int^{\Lambda^2}
dy\, \left[\frac{\theta(x-y)}{x} + \frac{\theta(y-x)}{y}\right] \frac{y \Sigma(y)}{ y +\Sigma^2(y)}, \quad (Z^{-1}(x) \equiv 1).
\label{SDeq}
\end{equation}
This form is reduced back to the form of the NJL gap equation with $\Sigma(x)\equiv m_F$, Eq.(\ref{NJLgap}), if the kernel is local: $\theta(x-y)/x + \theta(y-x)/y \rightarrow 1/\Lambda^2$, such as in the case of the massive gauge boson, $i D_{\mu\nu} \sim g_{\mu\nu}/\Lambda^2$ (See also Eq.(\ref{gaugedNJLSD})).
Eq.(\ref{SDeq}) is converted into a differential equation plus IR and UV boundary conditions \cite{Fukuda:1976zb}:
\begin{eqnarray}
\left( x \Sigma(x) \right)''
+ \alpha \frac{3 C_2}{4 \pi} \frac{\Sigma(x)}{x + \Sigma^2(x)} &=& 0,
\label{eq:diffSDo}\\
\lim_{x\rightarrow 0} x^2 \Sigma'(x) &=& 0,
\label{eq:IRBC}\\
\left.
\left( x \Sigma(x) \right)'\right|_{x=\Lambda^2} &=& m_0.
\label{eq:UVBC}
\end{eqnarray}
The asymptotic solution of Eq.(\ref{eq:diffSDo}) at $x \gg \Sigma^2(x)$ takes the form
$\Sigma (x)\sim m_F (x/m_F^2)^a$, with a conventional normalization $\Sigma(x=m_F^2) =m_F$, which is plugged back into the equation to yield
$(a+1) a +\alpha (3C_2)/(4\pi) =0$, i.e., $a= (-1 \pm \sqrt{1-3C_2\alpha/\pi})/2$.
For $\alpha<\frac{\pi}{3C_2}\equiv \alpha_{\rm cr}$, either solution, dominant ($a=(-1+\omega)/2$) or non-dominant ($a=(-1-\omega)/2$), has a power behavior,
which does not satisfy the UV boundary condition Eq.(\ref{eq:UVBC}) for the chiral limit $m_0=0$, where $\omega \equiv \sqrt{1-\alpha/\alpha_{\rm cr}}$.
The solution exists only at the presence of the explicit breaking $m_0$, namely the explicit breaking solution with the renormalized mass
$m_F=m_R= Z_m^{-1}\, m_0$, which yields the anomalous dimension $\gamma_m$ in the unbroken phase \cite{Leung:1985sn}:
\begin{equation}
m_0= m_R\left(\frac{\Lambda}{m_R}\right)^{-1+\omega},\quad \gamma_m =\Lambda \frac{\partial \ln Z_m^{-1}}{\partial \Lambda} =1-\sqrt{1-\frac{\alpha}{\alpha_{\rm cr}}}<1\quad \left(\alpha<\alpha_{\rm cr}=\frac{\pi}{3 C_2}\right)\,.
\label{gammamconformal}
\end{equation}
The result is written in terms of the one-loop anomalous dimension $\gamma_m^{(one-loop)}=3C_2\alpha/(2\pi)$ as
$\gamma_m=1-\sqrt{1-2\gamma_m^{(one-loop)}}$ which coincides with $\gamma_m^{(one-loop)} $ for $\alpha/\alpha_{\rm cr} \ll 1$:
On the other hand, the
SSB solution does exist for
\begin{equation}
\alpha >\frac{\pi}{3C_2}=\alpha_{\rm cr}\,,
\end{equation}
where $a=(-1\pm i \tilde \omega)/2$ with $\tilde \omega\equiv \sqrt{\alpha/\alpha_{\rm cr}-1}$, and the solution is of the oscillating form
$\Sigma(x) \sim \frac{m_F^2}{\sqrt{x}} \frac{1}{\tilde \omega} \sin \left(\tilde \omega [ \ln (\sqrt{x}/m_F) +\delta] \right)$, $\delta ={\cal O} (1)$, which satisfies the UV boundary condition
as
\begin{equation}
0=m_0\sim \frac{m_F^2}{\Lambda \tilde \omega } \sin \left(\tilde \omega \ln\left(\frac{4\Lambda}{m_F}\right)\right)\, ,
\end{equation}
for $\tilde \omega \ln\left(\frac{4\Lambda}{m_F}\right)=n \pi$ (numerically $e^\delta \simeq 4$).
In the large $N_C$ limit the critical coupling itself is numerically small, $\alpha_{\rm cr}\sim 1/N_C \ll 1$, although the effective coupling in the
condensate channel is $C_2 \alpha_{\rm cr} ={\cal O} (1)$ as was the case in the NJL coupling. This is the reason why the ladder approximation
yields reasonable result.
The ground state solution is $n=1$, which yields the dynamical mass $m_F$ of
the Berezinsky-Koterlitz-Thouless (BKT) form of essential-singularity (``Miransky scaling'') \cite{Miransky:1984ef}:
\begin{eqnarray}
m_F&\simeq& 4 \Lambda \exp\left(-\frac{\pi}{\sqrt{\frac{\alpha}{\alpha_{\rm cr}}-1}}\right)\,,\quad \left(\alpha> \alpha_{\rm cr}=\frac{\pi}{3 C_2} \ne 0\right)\,,
\label{Miransky}\\
&=&0 \,,\quad \left(\alpha<\alpha_{\rm cr}\right)\,.
\end{eqnarray}
This is compared with the NJL gap equation Eq. (\ref{SSBNJL}) and D-dimensional NJL Eq.(\ref{SSBDNJL}), and also with the
BCS Eq.(\ref{BCSgap}) and 2-dimensional model Eq.(\ref{D2NJL}). Again the large hierarchy
\begin{equation}
m_F\ll \Lambda\quad \left(\alpha/\alpha_{\rm cr} -1\ll 1\right)
\label{hierarchy}
\end{equation}
can be realized near criticality (near chiral symmetry restoration).
The essential-singularity scaling yields a peculiar phase transition, dubbed ``conformal phase transition'' \cite{Miransky:1996pd}, which is different from the typical 2nd order phase transition
as the Ginzburg-Landau phase transition. While the order parameter such as $m_F$
is continuously changed as $m_F\ne0$ to $m_F=0$ from $\alpha>\alpha_{\rm cr}$ to $\alpha<\alpha_{\rm cr}$, the spectrum changes discontinuously,
since there is no light spectrum in $\alpha<\alpha_{\rm cr}$ (conformal, unparticle) in contrast to the SSB phase where mass spectrum all
goes to zero as $\alpha\rightarrow \alpha_{\rm cr}+0$. It reflects the fact that the essential singularity is not analytic at $\alpha=\alpha_{\rm cr}$.
The light spectrum is possible for $\alpha<\alpha_{\rm cr}$ only when $m_0\ne0$ which violates the conformality. Thus all the mass spectrum $M$
for the conformal phase $\alpha<\alpha_{\rm cr}$ scales like the explicit breaking renormalized
mass $m_R$,
which is given by Eq.(\ref{gammamconformal}) as $M \sim m_R \sim m_0^{1/(1+\gamma_m)}$ \cite{Miransky:1998dh}, in conformity with the hyperscaling relation frequently used
in the lattice analyses for the conformal signals.
Eq.(\ref{Miransky}) implies that the coupling $\alpha$ is a function of $m_F/\Lambda$ with the nonperturbative beta function:
\begin{eqnarray}
\beta^{(NP)}(\alpha) &=&\Lambda \frac{\partial \alpha(\Lambda)}{\partial \Lambda}
= - \frac{2\pi^2\alpha_{\rm cr}}{\ln^3 (\frac{4\Lambda}{m_F})}
= - \frac{2\alpha_{\rm cr}}{\pi} \left(\frac{\alpha}{\alpha_{\rm cr}}-1\right)^{\frac{3}{2}}
\label{NPbeta},\\
\alpha(\mu) &=&\alpha_{\rm cr}
\left[1+ \frac{\pi^2}{\ln^2(\frac{4 \mu}{m_F})}
\right] \,,\label{NPrun}
\end{eqnarray}
with $\alpha_{\rm cr}$ being now regarded as a nontrivial ultraviolet fixed point (approached much faster than the asymptotic freedom $\sim 1/\ln \mu$).
The asymptotic form of the SSB solution is $\Sigma (x) \sim m_F^2/\sqrt{x}$ which is compared with the Operator Product Expansion
$\Sigma(x) \sim \frac{m_F^3}{x} (\frac{x}{m_F^2})^{\gamma_m/2}$ to yield a large anomalous dimension: \cite{Yamawaki:1985zg,Bando:1986bg}
\begin{equation}
\gamma_m = 1\quad \left(\alpha>\alpha_{\rm cr}\right).
\end{equation}
This ladder result is the characteristic feature of the walking technicolor.
Due to this mass generation which breaks the scale symmetry spontaneously, the ladder scale symmetry is also broken
explicitly producing the new nonperturbative trace anomaly besides the perturbative trace anomaly induced by the cutoff regularization $\Lambda$ (See Ref. \cite{Matsuzaki:2015sya} and references cited therein):
\begin{eqnarray}
\langle\partial_\mu D^\mu \rangle= \langle\theta_\mu^\mu\rangle^{(NP)}
&\equiv& \langle\theta_\mu^\mu\rangle^{(full)} - \langle\theta_\mu^\mu\rangle^{(perturbative)} =
\frac{\beta^{(NP)}(\alpha)}{4\alpha} \langle G_{\mu\nu}^2\rangle^{(NP)}
\,,\nonumber \\
&\simeq&
- N_F N_C \frac{4 \xi^2}{\pi^4}
m_F^4 \,, \quad \left(\xi \simeq 1.1\right)\,,
\label{NPanomaly}
\end{eqnarray}
where $\langle G_{\mu\nu}^2\rangle^{(NP)}
\equiv \langle G_{\mu\nu}^2\rangle^{(full)}
-\langle G_{\mu\nu}^2\rangle^{(perturbative)} $ is the nonpertubative gluon condensate and $N_F$ is a number of flavors of massless fermions
(besides color $N_C$).
Note that although $ \frac{\beta^{(NP)}(\alpha(\mu))}{4\alpha(\mu)} $ and
$\langle G_{\mu\nu}^2\rangle^{(NP)}_{(\mu)} $ are depending on the renormalization point $\mu$, the trace anomaly
$\langle\theta_\mu^\mu\rangle^{(NP)}$ is not as it should be (the energy-momentum tensor $\theta_{\mu\nu}$
is a conserved current and is not renormalized), with both dependence being cancelled each other precisely. \cite{Matsuzaki:2015sya}
This ladder dynamics was the basis for the walking technicolor \cite{Yamawaki:1985zg,Bando:1986bg} where the coupling is almost non-running $\alpha(\mu) \approx \alpha_{\rm cr}$ for $m_F<\mu<\Lambda$
even after the SSB takes place to produce the nonperturbative running.
The non-zero critical coupling also exists
in the asymptotically free gauge theory including the QCD in a more sophisticated way, in spite of no explicit non-zero critical coupling separating the SSB phase and the non-SSB phase, namely the
QCD is in one phase always in the SSB similarly to the BCS. The theory is classically scale-invariant but actually has an intrinsic mass scale
$\Lambda_{\rm QCD}$ due to the trace anomaly by the quantum effects (regulator). The intrinsic scale $\Lambda_{\rm QCD}$ is usually given by the one-loop beta function: $\Lambda_{\rm QCD} =\mu e^{-1/(b_0\alpha(\mu))}=\Lambda e^{-1/(b_0\alpha(\Lambda))}$ with $b_0$ given in Eq.(\ref{2loop}).
Although this looks like the BCS mass generation in Eq.(\ref{BCSgap}),
$\Lambda_{\rm QCD}$ should not be confused with the mass generation $m_F$.
The existence of the intrinsic scale $\Lambda_{\rm QCD}$ does not necessarily
imply the SSB $m_F\ne 0$ as in the NJL model where the intrinsic scale $1/\sqrt{G}$ does not necessarily imply the mass $m_F$.
The QCD coupling $\alpha(\mu)$ runs depending on the renormalization scale $\mu$ in units of $\Lambda_{\rm QCD}$ to grow in the infrared region.
The fermion mass $m_F$ is dynamically
generated due to the fermion-antifermion condensate which takes place in the infrared region $\mu < m_F$ where the coupling becomes strong as to exceed
the ``hidden'' critical coupling of order 1: $N_C \alpha(\mu<m_F) >N_C \alpha_{\rm cr} ={\cal O}(1)$. In the usual QCD it so happens that $m_F ={\cal O} (\Lambda_{\rm QCD})$.
In a wider parameter space, however, we can see the cases $m_F=0$ (chiral restoration) and $m_F\ll \Lambda_{\rm QCD}$ (near chiral restoration), where the non-zero critical coupling is actually essential.
The ``hidden'' non-zero critical coupling become ``visible'' when the system is put in the medium with finite temperature $T$ and density with baryon chemical potential $\mu_B$, where the running coupling
in the infrared region is no longer growing indefinitely and levels off at the relevant energy scale of order of $T$ or $\mu_B$. Then for $T,\mu_B$ such that $\alpha(T),\alpha(\mu_B) <\alpha_{\rm cr}$ the SSB would not take place, namely the chiral symmetry restoration occurs as has been studied actively. In contrast to the disappearance of the
fermion-antifermion condensate, the BCS dynamics for fermion-fermion condensate instead can be operative in
the finite density even with
the weakest coupling due to the Fermi surface, which is called color superconductor.
Here I discuss another case to visualize the non-zero critical coupling in
the QCD-like vector-like $SU(N_C)$ gauge theory with $N_F (\gg N_C)$ massless technifermions, still in the asymptotically free theory $N_F<11N_C/2$
with the running coupling vanishing in the ultraviolet region. This is the basis for the walking technicolor to be discussed later and I denote the intrinsic scale $\Lambda_{\rm QCD}$ as $\Lambda_{\rm TC}$ hereafter.
When one increases $N_F$, the vacuum polarization due to the
virtual fermion-antifermion pairs (loop effects) increases the screening of the charges in the long distance
(infrared energy region), which is
operative opposite to the asymptotically free anti-screening effects of the gluon loops: in the ultraviolet region $\mu\gg \Lambda_{\rm TC}$ the coupling is small and running is essentially one-loop dominated, while in the infrared region $\mu\ll \Lambda_{\rm TC}$ where the coupling grows, the higher loop effects particularly by the fermion loop screening effects are getting dominant, which
then balances the anti-screening effects to tend to make the coupling level off. Then the dynamical mass $m_F$ such that $\alpha(\mu=m_F) \simeq \alpha_{\rm cr}$ will be getting
smaller, as we increase $N_F/N_C$:
\begin{equation}
\frac{m_F}{\Lambda_{\rm TC}} \searrow \quad {\rm for} \quad \frac{N_F}{N_C} \nearrow\,,
\end{equation}
in contrast to the
ordinary QCD with $m_F={\cal O}(\Lambda_{\rm QCD})$ for $N_F=N_C=3$. It then
eventually could realize at certain large $r\equiv N_F/N_C\gg 1 $ an infrared fixed point $\alpha(\mu) < \alpha_*=\alpha(0)<\alpha_{\rm cr}$), which implies
that no SSB takes place and no bound states exist (``unparticle''), the phase called ``conformal window''.\footnote{ Here we are talking about the phase transition in the parameter $N_F/N_C$ by changing the theory.
It does not imply the existence of two phases in one theory with fixed $N_F/N_C$. See the discussions below and
Fig. \ref{beta:whole}
}
The approximate scale symmetry is operative with almost nonrunning coupling in the infrared region $\mu\ll \Lambda_{\rm TC}$, although it is violated explicitly by $\Lambda_{\rm TC}$ due to
the trace anomaly
in the ultraviolet region $\mu \gg \Lambda_{\rm TC}$ where the coupling is
running as in the usual asymptotically free theory.
The existence of the conformal window in fact can been seen explicitly at two-loop beta function, which is scheme-independent while higher loops are not:
\cite{Caswell:1974gg}.
\begin{eqnarray}
\beta^{(2-loop)}(\alpha) &=& -b_0 \alpha^2 -b_1 \alpha^3,\nonumber\\
b_0&=&\frac{1}{6\pi} (11N_C- 2N_F), \quad b_1= \frac{1}{24\pi^2}\left(34N_C^2-10N_C N_F -3 \frac{N_C^2-1}{N_C} N_F\right),\nonumber\\
\alpha_*&=&\alpha_*(N_F,N_C) = -\frac{b_0}{b_1} \,,
\label{2loop}
\end{eqnarray}
where we have $\beta^{(2-loop)}(\alpha=\alpha_*)=0$ by balancing the one-loop $-b_0 \alpha^2$ ($<0$ as far as asymptotically free, i.e., $N_F<11N_C/2$) with the two-loop contributions $-b_1 \alpha^3>0$ at the infrared limit $\mu=0$, which is realized only when $b_1<0$ s.t.
$N_F\gg N_C$ is satisfied. Note that $\alpha_*\searrow $ as $N_F/N_C\nearrow$ and $\alpha_*=\alpha_*(N_F,N_C) $ exists for $N_F^*< N_F < 11N_C/2 $ ($N_F^*\simeq 8$ for $N_C=3$).
In the context of large $N_C$ limit, such a situation corresponds to the
``anti-Veneziano limit'' (in distinction to the original Veneziano limit with $N_F/N_C\ll 1$):\cite{Matsuzaki:2015sya}
\begin{equation}
N_C \rightarrow \infty \quad {\rm and} \quad N_C \cdot \alpha = {\rm fixed},
\quad {\rm with} \quad r\equiv N_F/N_C ={\rm fixed} \,\, \gg 1\,.
\label{antiVeneziano}
\end{equation}
The anti-Veneziano limit in fact realizes a situation very close to the ladder approximation, with the $r=N_F/N_C$
behaving as a continuous parameter.
Then the theory has two phases in the parameter space $r$: SSB phase for $r>r_{\rm cr}$ such that $\alpha(\mu) >\alpha_{\rm cr}$ and the non-SSB phase otherwise.
In the case $\alpha_*<\alpha_{\rm cr}$, there in fact exists no
SSB $m_F\equiv 0$ and no bound states (``unparticle''). The coupling is almost constant for all the infrared region $\mu<\Lambda_{\rm TC}$ (infrared conformality), while it is running in the ultraviolet region $\mu>\Lambda_{\rm TC}$ essentially
as the one-loop running, in accord with the scale symmetry violation due to the
perturbative trace anomaly. \footnote{For the region $\alpha(\mu)>\alpha_{\rm cr} >\alpha_*$, there might exist
SSB, in which case there might exist two phases separated by the ultraviolet fixed point at $\alpha_{\rm cr}$ in the sense similar to the conjecture on the asymptotically non-free gauge theory
such as the strong coupling QED for $\alpha(\mu) >\alpha_{\rm cr}$,
although such a fixed point may be a Gaussian fixed point (trivial theory) \cite{Kogut:2005pm}.
}
On the other hand, for $\alpha_* > \alpha_{\rm cr}$ the SSB takes place with mass $m_F$ generated similarly to the ladder SD result in Eq. (\ref{Miransky}) with $\alpha$ replaced by $\alpha_*$ \cite{Appelquist:1996dq},
\begin{equation}
m_F\sim \Lambda_{\rm TC} \exp \left( - \frac{\pi}{\sqrt{\frac{\alpha_*}{\alpha_{\rm cr}}-1}}\right)\quad \left( \ll \Lambda_{\rm TC} \quad {\rm for} \quad \frac{N_F}{N_C}\quad {\rm s.t.} \quad \frac{\alpha_*(N_F,N_C)}{\alpha_{\rm cr} (N_C)}-1 \ll 1 \right)\,,
\end{equation}
where the phase transition in the parameter space $r$ has a characteristic essential singularity scaling (Miransky-BKT scaling), which
takes the same type of the ``conformal phase transition''
as the ladder one \cite{Miransky:1996pd}.
Once $m_F$ is generated, the scale symmetry is explicitly broken so as to yield the nonperturbative trace anomaly, Eq.(\ref{NPanomaly}), responsible for the nonperturbative running of the coupling, Eq.(\ref{NPrun}), and
the would-be infrared fixed point $\alpha_*$ at two-loop is actually washed out.
The resultant coupling would
have a form with (quasi) ultraviolet fixed point $\alpha_{\rm cr}$ similarly to Eq.(\ref{Miransky}) for
$\alpha(\mu) >\alpha_{\rm cr}$ ($\mu <\Lambda_{\rm TC}$), while it still has a remnant of infrared fixed point
(quasi fixed point) for $\alpha(\mu) <\alpha_{\rm cr}\simeq \alpha_*$ ($\mu>\Lambda_{\rm TC}$).
Thus the theory is in one phase, which is not separated by $\alpha_{\rm cr}$. The beta function has no exact zero at
$\alpha_{\rm cr}$ and the coupling runs through $\alpha_{\rm cr}$ continuously. See Fig.\ref{beta:whole}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=5.5cm]{walking-alpha-conj.eps}
\hspace{15pt}
\includegraphics[width=5.5cm]{RGE-conj.eps}
\vspace{15pt}
\caption{ Schematic picture of the running coupling (left) and the beta function (right) in the region,
with the (quasi-)infrared fixed point $\alpha_*$ and the critical coupling as a (quasi-)ultraviolet fixed point $\alpha_{\rm cr}$.
Perturbative coupling in $\alpha<\alpha_{\rm cr}$ is smoothly connected to the nonperturbative one in the
region $\alpha>\alpha_{\rm cr}$.
}
\label{beta:whole}
\end{center}
\end{figure}
To summarize the origin of mass in the strong coupling theories,
the mass $m_F$ originates from the intrinsic scale $\Lambda$
(Lagrangian parameter $G$ or the trace anomaly $\Lambda_{\rm QCD} (\Lambda_{\rm TC}))$ through SSB which takes place only in the strong coupling phase with the coupling larger than the non-zero critical coupling.
There is no mass generation $m_F\equiv 0$ in the weak coupling phase, even though the theory has intrinsic scale $\Lambda$.
\section{Hidden Symmetries in the SM Higgs Lagrangian
\cite{Fukano:2015zua}}
Here we recapitulate Ref. \cite{Fukano:2015zua} to show that the SM Higgs
Lagrangian Eq.(\ref{Higgs}) in the form of the {\it linear sigma model}, Eqs.(\ref{sigma}) and (\ref{sigmaMatrix}), is rewritten into precisely the form equivalent to the {\it scale-invariant} version of the
chiral $SU(2)_L\times SU(2)_R$ {\it nonlinear sigma model} based on the manifold $G/H$, with $G=SU(2)_L\times SU(2)_R$ and $H=SU(2)_{L+R}=SU(2)_V$, as far as it is in the broken phase, with {\it both the chiral and scale symmetries spontaneously broken} due to the same Higgs VEV
$v\ne 0$, and thus are {\it both nonlinearly realized}.
The SM Higgs Lagrangian is further shown to be {\it gauge equivalent} to the {\it scale-invariant}
version \cite{Kurachi:2014qma} of the Hidden Local Symmetry (HLS) Lagrangian \cite{Bando:1984ej,Bando:1987br,
Harada:2003jx}, which contains {\it possible new vector bosons}, analogues of the $\rho$ mesons, as the gauge bosons of the (spontaneously broken) HLS {\it hidden behind the SM Higgs Lagrangian}.
Let us discuss the Higgs Lagrangian in the form of Eqs.(\ref{sigma}) and (\ref{sigmaMatrix}):
The potential minimum exists at the chiral-invariant circle:
\begin{equation}
\langle \sigma^2(x)\rangle = \frac{-\mu_0^2}{\lambda}\equiv v^2\,, \quad \sigma^2(x) \equiv {\hat \sigma}^2(x) +{\hat \pi_a}^2(x)\,.
\label{chicircle}
\end{equation}
In Eq.(\ref{sigmaMatrix}) any complex matrix $M$ can be decomposed into the Hermitian (always diagnonalizable) matrix $H$ and unitary matrix $U$ as $M=HU$ ( ``polar decomposition'' ):
\begin{equation}
M(x) = H(x)\cdot U(x)\,, \quad H(x)=\frac{1}{\sqrt{2}} \left(\begin{array}{cc}
\sigma(x) & 0\\
0 &\sigma(x)
\end{array}\right)
\,, \quad U(x)= \exp\left(\frac{2i \pi(x)}{F_\pi}\right) \,,
\label{Polar}
\end{equation}
with $\pi(x)=\pi^a(x) \frac{\tau^a}{2} \,(a=1,2,3)$ and $F_\pi=v=\langle \sigma(x) \rangle$. The chiral transformation of $M$ is inherited by $U$,\footnote{The nonlinear realization was first introduced by K. Nishijima (then at Osaka City University) \cite{Nishijima} (in $G/H =U(1)_L\times U(1)_R/U(1)_V$ case) to make the nucleon massive in a chiral invariant way using the NG boson $\pi$ as
$M_N {\bar \Psi}_L \Psi_R$, where the physical (massive) nucleon $\Psi_{L/R} =(\xi^\dagger \psi_L, \xi \psi_R)$
transforms as $\Psi_{L/R} \rightarrow h(\pi(x),g_{L/R}) \cdot \Psi_{L/R}$, ($h\in H$), while the original nucleon field
$\psi_{L/R}$ does as $\psi_{L/R} \rightarrow g_{L/R} \cdot \psi_{L/R}$. Here the nonlinear base $(\xi^\dagger, \xi)$ is defined by $U(x)=\xi^2(x)$ with the transformation, $(\xi^\dagger,\xi)=(e^{-i \pi/F_\pi},e^{i\pi/F_\pi}) \rightarrow h(\pi(x),g_{L/R}) \cdot (\xi^\dagger,\xi) \cdot g_{L/R}^\dagger$, $h^\dagger \, h=1$, in accord with Eq.(\ref{transformation}).
See, e.g., Ref.\cite{Bando:1987br}.
} while $H$ is a chiral singlet such that:
\begin{equation}
U \rightarrow g_L \, U\, g_R^\dagger\,,\quad H \rightarrow H\,,
\label{transformation}
\end{equation}
where $g_{L/R} \in SU(2)_{L/R}$ and $U \, U^\dagger=1$ implies $\langle U\rangle =\langle \exp\left(\frac{2i \pi(x)}{F_\pi}\right)\rangle=1 \ne 0$, namely the spontaneous breaking of the chiral symmetry is taken granted in the polar decomposition.
Note that {\it the radial mode $\sigma$ is a chiral-singlet in contrast to $\hat \sigma$ which is a chiral non-singlet} transformed
into the chiral partner ${\hat \pi}_a$ by the chiral rotation. The {\it physical particles are $\sigma$ and $\pi$} which are
defined by the nonlinear realization, {\it in contrast to the tachyons $\hat \sigma$ and ${\hat \pi}_a$}.
We further parametrize $\sigma(x)$ as
\begin{equation}
\sigma(x) =v \cdot \chi(x)\,,\quad \chi(x)=\exp\left(\frac{\phi(x)}{F_\phi}
\right)
\,,
\label{NLscale}
\end{equation}
where $F_\phi=v$ is the decay constant of the dilaton $\phi$ as the Higgs.
The scale (dilatation) transformations for these fields are
\begin{equation}
\delta_D \sigma =(1 +x^\mu \partial_\mu) \sigma \,, \qquad
\delta_D \chi=(1+x^\mu \partial_\mu) \chi\,, \qquad
\delta_D \phi= F_\phi+x^\mu \partial_\mu\phi\,.
\end{equation}
Note that $\langle \sigma(x)\rangle= v \langle \chi(x) \rangle = v\ne 0$ breaks spontaneously the scale symmetry, but not the chiral symmetry, since
$ \sigma(x)$ ($\chi(x)$ as well) is a chiral singlet.
This is a nonlinear realization of the scale symmetry:
the $\phi(x)$ is a dilaton, NG boson of the spontaneously broken scale symmetry. Although $\chi$ is a dimensionless field,
it transforms as that of dimension 1, while $\phi$ having dimension 1 transforms as the dimension 0, instead.
Plugging Eqs.(\ref{Polar}) and (\ref{NLscale}) into the SM Higgs Lagrangian Eq.(\ref{sigmaMatrix}),
we straightforwardly arrive at the SM Higgs Lagrangian in the striking form:\cite{Fukano:2015zua}
\begin{eqnarray}
{\cal L}_{\rm Higgs}
&=&\left[ \frac{F_\phi^2}{2} \left(\partial_\mu \chi \right)^2+ \frac{F_\pi^2}{4}{\chi}^2\cdot {\rm tr} \left(\partial_\mu U \partial^\mu U^\dagger\right)\right]
- V(\phi)\nonumber\\
&=&\chi^2(x) \cdot \left[ \frac{1}{2} \left(\partial_\mu \phi\right)^2 +\frac{F_\pi^2}{4}{\rm tr} \left(\partial_\mu U \partial^\mu U^\dagger\right)\right] -V(\phi)\nonumber \,,\\
V(\phi)&=& \frac{\lambda}{4} v^4 \left[\left(\chi^2(x) -1\right)^2-1\right] =\frac{M_\phi^2 F_\phi^2}{8} \left[\left(\chi^2(x) -1\right)^2-1\right] \,,
\label{SNLSM}
\end{eqnarray}
which is nothing but {\it the scale-invariant nonlinear sigma model} with $F_\phi=F_\pi=v$, an effective theory of the walking technicolor~\cite{Matsuzaki:2012mk,Matsuzaki:2015sya}, apart from the form of the explicit scale-symmetry breaking potential $V(\phi)$ (see Eq.(\ref{WTC})).
The explicit scale-symmetry breaking comes only from the potential $V(\phi)$ such that $\delta_D V(\phi) = \lambda v^4\chi^2=-\theta_\mu^\mu$ whose scale dimension $d_{\theta}=2$ (originally the tachyon mass term) instead of 4 of the walking technicolor: namely, the scale symmetry is broken only by the dimension 2 operator.\footnote{Note that {\it mass term of all the SM particles except the Higgs is scale-invariant}. By the electro-weak gauging as usual; $\partial_\mu U\Rightarrow {\cal D}_\mu U= \partial_\mu U -i g_2 W_\mu U +i g_1 U B_\mu$
in Eq.(\ref{SNLSM}), we see that the mass term of $W/Z$ is scale-invariant thanks to the dilaton factor $\chi$, and so is the mass term of the SM fermions $f$: $g_Y \bar f h f
=(g_Y v/\sqrt{2}) (\chi \bar f f)$, all with the scale dimension 4.
}
This yields the mass of the (pseudo-)dilaton as the Higgs $M_\phi^2=2\lambda v^2$, which is in accord with the Partially Conserved Dilatation Current (PCDC) for $\partial^\mu D_\mu=\theta_\mu^\mu$:
\begin{equation}
M_\phi^2 F_\phi^2=-\langle0|\partial^\mu D_\mu|\phi\rangle F_\phi=-d_{\theta} \langle \theta_\mu^\mu\rangle =2\lambda v^4\langle \chi^2(x) \rangle=2\lambda v^4\,,
\label{PCDC}
\end{equation}
with $F_\phi=v$, where $D_\mu$ is the dilatation current: $\langle 0|D_\mu(x) |\phi\rangle=-i q_\mu F_\phi e^{-i q x}$, or equivalently $\langle 0| \theta_{\mu\nu}|\phi(q)\rangle= F_\phi (q_\mu q_\nu- q^2 g_{\mu\nu}/3)$.
Hence the SM Higgs
as it stands is a (pseudo) dilaton, with the {\it mass arising from the dimension 2 operator}
in the potential, which vanishes for $\lambda\rightarrow 0$:
\begin{equation}
M_\phi^2=2\lambda v^2
\rightarrow 0
\quad \left(\lambda\rightarrow 0\,, \,\,
v=\sqrt{\frac{-\mu^2_0}{\lambda}} =
{\rm fixed}\,\ne 0\right)
\label{conformallimit}
\end{equation}
(``conformal limit''\cite{Fukano:2015zua}).\footnote{With vanishing potential, $V(\phi) \rightarrow 0$, this limit still gives an {\it interacting theory where the physical particles $\pi$ and $\phi$ have derivative coupling} in the same sense as in the nonlinear chiral Lagrangian Eq.(\ref{NLS}). It should be contrasted to the triviality limit, $\lambda\rightarrow 0$ {\it without fixing $ v=\sqrt{\frac{-\mu^2_0}{\lambda}}\ne 0$}, which yields only a free theory of tachyons $\hat \pi$ and $\hat \sigma$. This limit should also be distinguished from the popular limit $\mu^2_0\rightarrow 0 $ with $\lambda=$fixed $\ne 0$, where the Coleman-Weinberg potential as the explicit scale symmetry breaking is generated by the trace anomaly (dimension 4 operator) due to the quantum loop. }
In fact the Higgs mass 125 GeV implies that the SM Higgs is in near conformal limit with $v=$ fixed:
\begin{equation}
\lambda=\frac{1}{2} \left(\frac{M_\phi}{v}\right)^2 \simeq \frac{1}{2} \left(\frac{125\,{\rm GeV}}{246\,{\rm GeV}}\right)^2 \simeq \frac{1}{8} \ll 1\,.
\end{equation}
It should be noted that {\it $\lambda\ll 1$ (with $v =$ fixed $\ne 0$) can be realized even when the underlying theory is strong coupling}, particularly when the {\it scale symmetry is operative},
as we discuss later in both NJL type theory, Eq.(\ref{NJLconformallimit}), and the strong coupling gauge theory (walking technicolor) in the anti-Veneziano limit, Eq.(\ref{weakcouplingTD}).
On the other hand, if we take the limit $\lambda \rightarrow \infty$,
then the SM Higgs Lagrangian goes over to the usual nonlinear sigma model {\it without scale symmetry}:
\begin{equation}
{\cal L}_{{\rm NL}\sigma}=\frac{F_\pi^2}{4} {\rm tr} \left(\partial^\mu U\partial_\mu U^\dagger\right)\,,\quad \left(\lambda\rightarrow \infty\,, \,\,
v=\sqrt{\frac{-\mu^2_0}{\lambda}} =
{\rm fixed}\,\ne 0\right)
\label{NLS}
\end{equation}
where the potential is decoupled with $\chi(x)$ frozen to the minimal point $\chi(x)\equiv 1$ ($\phi(x) \equiv \langle \phi(x) \rangle=v\ne 0$), so that the scale symmetry breaking is transferred from the potential to the kinetic term,
which is no longer transformed as the dimension 4 operator. This is known to be a good effective theory (chiral perturbation theory) of the ordinary QCD which in fact lacks the scale symmetry at all,
perfectly consistent with
the nonlinear sigma model, Eq.(\ref{NLS}). However, it cannot be true for
the walking technicolor which does have the scale symmetry, and the effective theory must respect the symmetry of the underlying theory, in a form of the scale-invariant nonlinear sigma model Eq.(\ref{SNLSM}) in the conformal limit $\lambda\rightarrow 0$.
Once rewritten in the form of Eq.(\ref{SNLSM}), it is easy to see \cite{Fukano:2015zua}
that the {\it SM Higgs Lagrangian is gauge equivalent to the ``scale-invariant HLS model'' (s-HLS)}\cite{Kurachi:2014qma}, a scale-invariant version of the HLS model \cite{Bando:1984ej,Bando:1987br,Harada:2003jx} \footnote{
The s-HLS model was also discussed in a different context, ordinary QCD in medium.\cite{Lee:2015qsa}
}
, which {\it contains massive spin-1 states}, spontaneously broken HLS gauge bosons, as possible yet other composite states in some
underlying theory hidden behind the
SM Higgs.
The HLS can be made explicit by dividing $U(x)$ into two parts:
\begin{equation}
U(x)= \xi_L^\dagger(x) \cdot \xi_R(x)\,,
\label{U:decomp}
\end{equation}
where $\xi_{R,L}(x)$ transform under $G_{\rm global} \times H_{\rm local}$ as
\begin{equation}
\xi_{R,L}(x) \rightarrow h(x) \cdot \xi_{R,L}(x) \cdot { g^\prime}_{R,L}^\dagger\,,\quad
U(x) \rightarrow {\hat g}_L U(x) { g^\prime}_R^\dagger \quad \quad
\left(h(x)\in H_{\rm local},\, {g^\prime}_{R,L}\in G_{\rm global} \right)
\end{equation}
The $H_{\rm local} $ is a gauge symmetry of group $H$ arising from the redundancy (gauge symmetry) how to divide
$U$ into two parts. Then we can introduce the HLS gauge boson $V_\mu(x)$ by covariant derivative as
\begin{equation}
D_\mu \xi_{R,L}(x) = \partial_\mu \xi_{R,L} (x)-i V_\mu(x) \xi_{R,L}(x) \,,
\label{HLScovariant}
\end{equation}
which transform in the same way as $\xi_{R,L}$. Then we have covariant objects transforming homogeneously under $H_{\rm local}$:
\begin{eqnarray}
{\hat \alpha}_{\mu,R,L}(x)&\equiv& \frac{1}{i}D_\mu \xi_{R,L}(x) \cdot \xi_{R,L}^\dagger(x) =\frac{1}{i}\partial_\mu \xi_{R,L}(x) \cdot \xi_{R,L}^\dagger(x)
- V_\mu(x)\,, \nonumber\\
{\hat \alpha}_{\mu, ||,\perp}(x)&\equiv & \frac{1}{2}\left({\hat \alpha}_{\mu,R}(x) \pm {\hat \alpha}_{\mu, L}(x)\right)
\nonumber \\
&=&
\Bigg\{ \begin{array}{c}
\frac{1}{2i} \left(\partial_\mu \xi_{R}(x) \cdot \xi_{R}^\dagger(x)+ \partial_\mu \xi_{L}(x) \cdot \xi_{L}^\dagger(x) \right)- V_\mu(x)= {\alpha}_{\mu,||}(x) - V_\mu(x)\\
\frac{1}{2i} \left(\partial_\mu \xi_{R}(x) \cdot \xi_{R}^\dagger(x)- \partial_\mu \xi_{L}(x) \cdot \xi_{L}^\dagger(x) \right) = {\alpha}_{\mu,\perp}(x) \end{array}\nonumber\,,\\
\{ {\hat \alpha}_{R,L}(x), {\hat \alpha}_{||,\perp}(x)\}
&\rightarrow& h(x) \cdot \{ {\hat \alpha}_{R,L}(x), {\hat \alpha}_{||,\perp}(x)\}\cdot h^\dagger(x)\,.
\end{eqnarray}
We thus have two independent invariants under the larger symmetry $G_{\rm global} \times H_{\rm local}$:
\begin{equation}
\,
{\cal L}_A=v^2\cdot {\rm tr} {\hat \alpha}_{\perp}^2(x)\,,\quad
\,\quad
{\cal L}_V=v^2\cdot {\rm tr} \,{\hat \alpha}_{||}^2(x) = v^2\cdot {\rm tr} \, \left(V_\mu(x) - {\alpha}_{\mu,||}(x)\right)^2 \,.
\end{equation}
Hence the scale-invariant version of the Higgs Lagrangian, Eq.(\ref{SNLSM}), in the conformal limit $\lambda \rightarrow 0\,, v={\rm fixed}$, can be extended to the scale-invariant version having the HLS (s-HLS):\cite{Kurachi:2014qma}
\begin{equation}
{\cal L}_{\rm s-HLS} = \chi^2(x) \cdot \left(\frac{1}{2} \left(\partial_\mu \phi\right)^2 + {\cal L}_A+ a {\cal L}_V\right)\,,
\label{SHLS}
\end{equation}
with $a$ being an arbitrary parameter.
We now fix the gauge of HLS as $\xi_L^\dagger=\xi_R=\xi=e^{i \pi/v}$ such that $U=\xi^2$.
Then $H_{\rm local}$ and $H_{\rm global} (\subset G_{\rm global})$
get simultaneously broken spontaneously (Higgs mechanism), leaving the
diagonal subgroup $H=H_{\rm local}+H_{\rm global}$, which is nothing but the subgroup $H$ of the original $G$ of $G/H$: $H\subset G$.
According to the Higgs mechanism, the HLS gauge boson $V_\mu(x)$ acquires the mass $\frac{1}{2} a (g_H \, v)^2\,(V^a_\mu(x))^2$ through the invariant ${\cal L}_V$ after rescaling the kinetic term of $V_\mu$ by the HLS gauge coupling $g_H$ as
$V_\mu(x) \rightarrow g_H\, V_\mu(x)$. {\it Obviously the vector boson mass terms are scale-invariant thanks to the nonlinear realization of the scale symmetry!}
For the low energy $p^2<M_V^2$ where the kinetic term can be ignored, the HLS gauge boson $V_\mu$ becomes just an auxiliary field to be solved away to yield ${\cal L}_V=0$.
Noting that ${\cal L}_A=v^2 \cdot{\rm tr}
{\hat \alpha}_{\mu,\perp}^2(x)=v^2 \cdot{\rm tr}{\alpha}_{\mu,\perp}^2(x)= \frac{v^2}{4}\cdot {\rm tr} \left(\partial_\mu U \partial^\mu U^\dagger\right) $ by a straightforward algebraic calculation, we see that
${\cal L}_{\rm s-HLS}$ in Eq. (\ref{SHLS}) is simply reduced back to the original SM Higgs Lagrangian ${\cal L}_{\rm Higgs}$ in nonlinear realization, Eq.(\ref{SNLSM}). Note that {\it the HLS gauge boson acquires the
scale-invariant mass term thanks to the dilaton factor $\chi^2$}, the nonlinear realization of the scale symmetry, in sharp contrast to the {\it Higgs (dilaton) which acquires mass only from the explicit breaking of the scale symmetry}.
The electroweak gauge bosons ($\in {\cal R}_\mu ({\cal L}_\mu)$) are introduced by extending the covariant derivative of Eq.(\ref{HLScovariant})
this time by gauging $G_{\rm global}$, which is {\it independent of $H_{\rm local}$} in the HLS extension:
\begin{equation}
D_\mu \xi_{R,L}(x)\Rightarrow {\hat D}_\mu \xi_{R,L}(x)\equiv \partial_\mu \xi_{R,L} (x)-i V_\mu(x) \, \xi_{R,L}(x) +i \xi_{R,L}(x)\, {\cal R}_\mu ( {\cal L}_\mu)\,.
\label{fullcovariant}
\end{equation}
As usual in the Higgs mechanism, the gauge bosons of ${\rm gauged-}H_{\rm global} (\subset
{\rm gauged-}G_{\rm global}$) get mixed with the gauge bosons of HLS, leaving only the gauge bosons of the unbroken diagonal subgroup $({\rm gauged-}H)=H_{\rm local} + ({\rm gauged-}H_{\rm global})$ be massless after mass diagonalization.
We then finally have a gauged s-HLS version of the Higgs Lagrangian (gauged-s-HLS):
\begin{equation}
{\cal L}^{\rm gauged}_{\rm s-HLS}= \chi^2(x)\cdot
\left[
\frac{1}{2} \left(\partial_\mu \phi\right)^2 + {\hat {\cal L}}_A+ a {\hat {\cal L}}_V
\right]\,,
\label{gaugedsHLS}
\end{equation}
with
\begin{equation}
{\hat {\cal L}}_{A,V}= {\cal L}_{A,V} \left( D_\mu \xi_{R,L}(x) \Rightarrow {\hat D}_\mu \xi_{R,L}(x)\right) .
\end{equation}
The new HLS boson may be identified with the LHC diboson events \cite{Aad:2015owa}, with the model parameter choice
consistent with the reported results \cite{MY16},
similarly to the walking technirho which is also the HLS
boson for the large chiral symmetry ($SU(8)_L\times SU(8)_R$ in the one-family model) \cite{Fukano:2015hga,Fukano:2015zua}.
A salient feature of the new vector boson of HLS in the scale-invariant SM Lagrangian is that
the scale-invariant vector boson mass terms in Eq. (\ref{gaugedsHLS}) having the $\phi$ (Higgs $H$) field in the overall conformal factor $\chi^2(x)$ yield
$\phi$ couplings only to the diagonal pairs of the (longitudinal) SM gauge bosons $H-W/Z-W/Z$ or of those of the new vector bosons $H-V-V$
after the mass diagonalization (as it should be done).
Thus the new HLS vector bosons hidden in the SM only couple to the Higgs in a pair of themselves as $V-V-H$ but not
in the off-diagonal combination with the SM weak bosons $W/Z$:
\begin{equation}
V- W/Z - H \, {\rm coupling} \, = \, 0 \,,
\label{Conformalbarrier}
\end{equation}
namely, the decay $V\rightarrow W/Z + H$ is forbidden by
the scale/conformal symmetry ({\it Conformal Barrier})~\cite{Fukano:2015zua}, in sharp contrast to the
popular ``equivalence theorem'' which implies comparable coupling of $V$ to $W/Z + W/Z$ and $W/Z +H$, based on the usual
(non scale-invariant) Higgs field identification $\hat \sigma=v+ H$, with the Higgs $H$ being
on the same footing as the NG modes $\hat \pi$ which are the longitudinal modes of $W/Z$ by the equivalence theorem.
Consequently, the $V$ predominantly decays to the weak boson pairs $WW/WZ$.
In other words, the popular consequence of the ``equivalence theorem'' is invalidated by the scale/conformal symmetry.
The absence of $V\to WH/ZH$ signatures at the LHC Run-II thus could indirectly probe the existence of the (approximate) scale/conformal invariance of the system involving
$V$, $W, Z$ and $H$.
It is straightforward to extend the internal symmetry group to $G_{\rm global}$ =$SU(N_F)_L\times SU(N_F)_R$ and $H_{\rm local}= SU(N_F)_V$.
The Lagrangian then takes the form
\begin{equation}
{\cal L}_{\rm s-HLS}
= \chi^2(x) \cdot \left(
\frac{1}{2} \left(\partial_\mu \phi\right)^2
+
F_\pi^2 \left[
{\rm tr}[ \hat{\alpha}_{\mu \perp}^2 ]
+
a \, {\rm tr}[ \hat{\alpha}_{\mu ||}^2 ]
\right]
\right)
+ \cdots
\,, \label{sHLS}
\end{equation}
where $F_\pi$ is related to $v=246$ GeV as $F_\pi = v/\sqrt{N_F/2}$.
This form of the Lagrangian is the same as that of the effective theory of the one-family ($N_F=8$) walking technicolor \cite{Kurachi:2014qma}, except for the shape of the scale-violating
potential $V(\phi)$ which has a scale dimension 4 (trace anomaly) in the case of the walking technicolor instead of 2 of the SM Higgs case (Lagrangian mass term). We shall come back to this later.
\section{Strong Coupling NJL as the UV completion of the Weak Coupling SM Higgs Lagrangian \cite{Yamawaki:2015tmu}}
\label{NJLvsSM}
Let us now recapitulate Ref. \cite{Yamawaki:2015tmu} which elaborated the composite Higgs model based on the strong coupling theory $G>G_{\rm cr}\ne 0$ pioneered by Nambu.
In the NJL model \cite{Nambu:1961tp} for the $N_C-$component 2-flavored fermion $\psi$ the Lagrangian takes the form:
\begin{eqnarray}
{\cal L}_{\rm NJL} &=& \bar \psi i\gamma^\mu\partial_\mu \psi + \frac{G}{2} \left[ (\bar \psi \psi)^2 +(\bar \psi i \gamma_5 \tau^a \psi)^2\right]\nonumber\\
&=& \bar \psi \left(i\gamma^\mu\partial_\mu +\hat \sigma +i\gamma_5 \tau^a \hat \pi_a \right)\psi -\frac{1}{2G} \left({\hat \sigma}^2 +{\hat \pi_a}^2\right)\,,
\label{NJL}
\end{eqnarray}
where the equations of motion of the auxiliary fields $\hat \sigma \sim G \bar \psi \psi$ and ${\hat \pi}^a \sim G \bar \psi i\gamma_5 \tau^a\psi$ are plugged back into the Lagrangian to get the original Lagrangian.
In the large $N_C$ limit ($N_C \rightarrow \infty$ with $N_C G\ne 0$ fixed), after rescaling the induced kinetic term to the canonical one, $Z_\phi^{1/2} \hat \sigma \rightarrow \hat \sigma$,
the quantum theory for $\hat \sigma$ and $\hat \pi$ sector yields
precisely the same form as the SM Higgs Eq.(\ref{Higgs}),
with: \cite{Eguchi:1976iz}
\begin{eqnarray}
\mu_0^2&=& \left(\frac{1}{G} - \frac{1}{G_{\rm cr}}\right) Z_\phi^{-1} =-2m_F^2=- v^2 Z_\phi^{-1} =- \lambda v^2 < 0 \quad (G>G_{\rm cr}=\frac{4\pi^2}{N_C \Lambda^2} )
\nonumber\\
\lambda&=& Z_\phi Z_\phi^{-2} = Z_\phi^{-1} = \left[\frac{N_C}{8\pi^2} \ln \frac{\Lambda^2}{m_F^2}\right]^{-1} \sim \left[\frac{N_C}{8\pi^2} \ln \frac{\Lambda^2}{v^2}\right]^{-1} \,,
\label{tachyon}
\end{eqnarray}
where the gap equation has been used:
\begin{equation}
\frac{1}{G} - \frac{1}{G_{\rm cr}}= -\frac{N_C}{4\pi^2} m_F^2 \ln \frac{\Lambda^2}{m_F^2}=- 2 m_F^2 Z_\phi =-F_\pi^2=-v^2\,.
\label{gap}
\end{equation}
Eq.(\ref{tachyon}) shows that the {\it tachyon with $\mu_0^2<0$ is in fact generated} by the dynamical effects for the {\it strong coupling} $G>G_{\rm cr}\ne 0$,
corresponding to the generation of mass $m_F\ne0$ in the gap equation.
Or, we can explicitly see it by computing the $\bar \psi \psi$ bound state using the $m_F=0$ solution (wrong solution) of the gap equation at $G>G_{\rm cr}$. The correct spectrum $M_{\pi}^2=0, M_{\phi}^2 =2 \lambda v^2 =-2\mu_0^2=4m_F^2$ can be
obtained when we use the correct solution $m_F\ne 0$ in the gap equation.
The last equality $M_\phi^2=4 m_F^2$, often dubbed ``BCS mass relation'', is specific to the $N_C \rightarrow \infty$ (with $N_C\, G\ne 0$ fixed) limit of the NJL model ({\it ``weak coupling'' limit} $G >G_{\rm cr} \sim 1/N_C\rightarrow 0$ in the {\it strong coupling phase}), but not the general outcome of the NJL model nor the generic linear sigma model.
There are two extreme limits for $\lambda$ in Eq.(\ref{tachyon}) :
$\lambda \rightarrow0$ ($N_C\gg 1$ and/or $\Lambda/v^2 \gg 1$) reproduces precisely the conformal limit, or scale-invariant nonlinear sigma model limit, Eq.(\ref{SNLSM}), of the SM Higgs Lagrangian,
while $\lambda\rightarrow \infty$ ($N_C, \Lambda^2/v^2 ={\cal O}(1)$) does the nonlinear sigma model limit without scale symmetry, Eq.(\ref{NLS}).
We are particularly interested in the limit
\begin{equation}
\lambda =
\frac{1}{\frac{N_C}{8\pi^2} \ln \frac{\Lambda^2}{v^2}}
\rightarrow 0
\label{NJLconformallimit}
\end{equation}
(conformal limit in Eq.(\ref{conformallimit})), which is
realized in the strong coupling theory with $G>G_{\rm cr}\ne 0$ for $\Lambda/v\rightarrow \infty$ and/or $N_C\rightarrow \infty$, with $v=F_\pi=F_\phi \ne 0$ fixed.\footnote{
If $\Lambda$ is regarded as a physical cutoff in contrast to the nonperturbative renormalization arguments below, this argument would not be realistic for the 125 GeV Higgs with $\lambda\simeq 1/8$, corresponding to $\Lambda\simeq v\cdot e^{32\pi^2/N_C} \gg 10^{19}$ GeV. For the NJL model with $N_D$ doublets, however, we would have $\Lambda\simeq v \cdot e^{32\pi^2/(N_D N_C)} \sim 10^{11}$ GeV for $N_D=N_C=4$, somewhat realistic if the condensate is mainly triggered by the strong
(ETC-induced) four-fermion coupling rather than the technicolor gauge coupling \cite{Miransky:1988gk,Matumoto:1989hf}
in the one-family technicolor model \cite{Dimopoulos:1979sp} ($N_F=2 N_D=8$, see later discussions.).
}
Then the effective Lagrangian in the large $N_C$ limit takes precisely the same as the SM Higgs Lagrangian,
which is further equivalent to the scale-invariant nonlinear sigma model, Eq.(\ref{SNLSM}), as mentioned before. Now the SM
Higgs is identified with the composite (pseudo-)dilaton
with mass vanishing $M_\phi^2 =2 \lambda v^2 \rightarrow 0$.
The limit theory gives an {\it interacting (nontrivial) low energy effective theory even in the $\Lambda/v \rightarrow \infty$ limit}: a scale-invariant nonlinear sigma model ~\cite{Matsuzaki:2012mk,Matsuzaki:2015sya} where massless $\pi$ and $\phi$ are {\it interacting with each other} with
the (derivative) couplings $\sim (1/F_\pi, 1/F_\phi) \ne 0$. It is actually the basis for the {\it scale-invariant chiral perturbation theory} (sChPT) with the derivative expansion as a loop expansion \cite{Matsuzaki:2013eva}, although the Yukawa couplings of
$\pi,\phi$ to the fermions are vanishing $g_Y\sim m_F/F_\pi, m_F/F_\phi \rightarrow 0$ (The composite particles are still interacting due to the loop divergence compensation of the vanishing Yukawa coupling).
This limit should be sharply distinguished from a similar limit $\Lambda/m_F \rightarrow \infty$, $m_F=$fixed (not $\Lambda/v \rightarrow \infty$, $v=$fixed), which is the famous triviality limit (Gaussian fixed point) where the theory becomes a free theory: free massive scalar for $G<G_{\rm cr}$ and free tachyon for $G>G_{\rm cr}$, with not just the Yukawa couplings but all the couplings vanishing.
One might wonder why dilaton in NJL model? Obviously the NJL model has the explicit scale-breaking coupling $G$ having dimension $[M]^{-2}$. But this scale is an ultraviolet scale to which the low energy effective theory
is insensitive. This is in exactly the same sense as in the scale-invariant ladder gauge theory, Eqs.(\ref{Miransky}) and (\ref{hierarchy}), where the intrinsic scale $\Lambda_{\rm TC}$ generated by the trace anomaly can be far bigger than the infrared scale of spontaneous
symmetry breaking $F_\pi, F_\phi ={\cal O} (v) \ll \Lambda_{\rm TC}$ thanks to the approximate scale symmetry due to the almost nonrunning coupling.
In fact we can formulate the nonperturbative running of the (dimensionless) four-fermion coupling $g=\frac{\Lambda^2}{4\pi^2} G$ in the same way as the Miransky nonperturbative renormalization: The gap equation Eq. (\ref{gap}) reads
\begin{equation}
\left(\frac{1}{g_{\rm cr}} - \frac{1}{g}\right) \Lambda^2 =N_C m_F^2 \ln \frac{\Lambda^2}{m_F^2} \simeq 4\pi^2 v^2\,,\quad g_{\rm cr}=\frac{1}{N_C}\,,
\label{gap2}
\end{equation}
which leads to a nonperturbative beta function for $g>g_{\rm cr}$:
\begin{equation}
\beta (g) = \Lambda \frac{\partial g(\Lambda)}{\partial \Lambda}\Bigg|_{v=\rm fixed}=-\frac{2}{g_{\rm cr}} g\cdot (g-g_{\rm cr})\,,\quad g(\mu) = g_{\rm cr}\frac{1}{
1-\frac{4\pi^2 g_{\rm cr} v^2}{\mu^2}
}
\label{NJLbeta}
\end{equation}
by {\it fixing $v=$ constant} (instead of the conventional limit with $m_F=$ constant)
and taking $\Lambda \rightarrow \infty$.
Thus, without troublesome log factor, $g=g_{\rm cr}=1/N_C$ is the ultraviolet fixed point defining a nontrivial interacting theory in the continuum limit. As the running coupling $g(\mu)$
reaches $g_{\rm cr}$ even much faster than the walking coupling in Eq.(\ref{NPrun}),
the scale symmetry is operative
$g(\mu) \approx g_{\rm cr}$ for
the wide region $4\pi^2 g_{\rm cr}v^2\ll \mu^2< \Lambda^2$.
We also have $- \langle \bar \psi_i \psi_j \rangle = \delta_{i,j} \Lambda^2 m_F N_C/(4\pi^2)= Z_m^{-1}\,\delta_{i,j} v^3 N_C /(4\pi^2) $, where $Z_m^{-1}=Z_m^{-1}(\Lambda/v)= (\Lambda/v)^2 [N_C \ln (\Lambda^2/v^2)/(4\pi^2)]^{-1/2} $ is the mass renormalization constant, which
implies\footnote{
Hence the operators have scale dimension $d_{\bar \psi \psi}= 1$ and $d_{(\bar \psi \psi)^2}= 2$ in the large $N_C$ limit.
Note that $\gamma_m$ is actually slightly smaller by the $1/\ln(\Lambda^2/v^2)$ than ``$\gamma_m=2$'' in the conventional limit $m_F=$ constant, so that $d_{(\bar \psi \psi)^2}$ is slightly larger than 2 actually, i.e., possible eight-fermion operators corresponding to the $\lambda \phi^4$ would have dimension $d_{(\bar \psi \psi)^4} >4$, barely
irrelevant. This is contrasted to the conventional limit where the eight-fermion
operators are marginal and hence the NJL coupling without such counter terms would not be renormalizable nor interacting theory in the continuum limit (See the section 8 of the 3rd entry of Ref.\cite{Kondo:1991yk}).
}
\begin{equation}
\gamma_m = Z_m \Lambda \frac{\partial Z_m^{-1}}{\partial \Lambda} =2 - 1/\ln(\Lambda^2/v^2)
\longrightarrow 2 \quad (\Lambda/v \rightarrow \infty)\,.
\label{gamma}
\end{equation}
Thus we may write $\bar \psi_i \psi_j= -Z_m^{-1}\delta_{i,j} v^3 N_C/(4\pi^2) \cdot \chi $, or $(G/2)(\bar \psi \psi)^2=2 g N_C\Lambda^2 v^2\cdot \chi^2/[\ln(\Lambda^2/v^2)]$.
The gap equation implies $\beta(g)/g = -g (4\pi^2) (v^2/\Lambda^2)$.
Putting all together, we have $\beta(g)/g\cdot (G/2) \cdot
(\bar \psi \psi)^2|_{g\rightarrow g_{\rm cr}=1}= - \lambda v^4\chi^2$.
Then we get the {\it explicit scale symmetry breaking in the dimension 2 operator}: $\theta_\mu^\mu= \frac{\beta(g)}{g} \frac{G}{2} \left[
\left(\bar \psi \psi\right)^2 +\left(\bar \psi i\gamma_5 \tau^a \psi \right)^2
\right]= -\lambda v^4 \chi^2$,
where $\lambda = 8\pi^2/[N_C\ln(\Lambda^2/v^2)]\rightarrow 0$ as in Eq.(\ref{tachyon}).
The PCDC follows in precisely the same way as in the SM Higgs as $M_\phi^2 F_\phi^2 =-d_{(\bar \psi \psi)^2} \langle \theta_\mu^\mu\rangle =2 \lambda v^4$ (See below Eq.(\ref{SNLSM})).
In any case the trace of energy-momentum tensor vanishes in the limit $\lambda \sim 1/[N_C\ln (\Lambda^2/v^2)] \rightarrow 0$, and the dilaton mass should come from the trace anomaly in the $1/N_C$ sub-leading loop effects, or the chiral loops of the effective theory Eq.(\ref{SNLSM}).
Again the spin 1 composites can also be introduced via HLS, precisely in the same way as Eq.(\ref{SHLS}) for the SM Higgs
Lagrangian. This time it can be done more explicitly by introducing the vector/axialvector type four-fermion coupling which
in fact become the ``explicit'' composite HLS gauge bosons.(See section 5.3 of Ref.\cite{Bando:1987br}).
Incidentally, the above prescription to have an interacting nontrivial continuum theory of the NJL model
is similar to the renormalizability arguments of the $D$-dimensional NJL model ($2<D<4$) \cite{Kikukawa:1989fw}
and the gauged NJL model \cite{Kondo:1991yk} both without troublesome log factor,
although in the case at hand the explicit scale-breaking from the Lagrangian parameters, i.e., the four-fermion interaction and fermion mass term (if present), depend on the renormalization point (vanish at the UV limit).
The beta function Eq.(\ref{NJLbeta}) and anomalous dimension Eq.(\ref{gamma}) of the $D-$dimensional four-fermion theory renormalizable in $1/N_C$ expansion are given: \cite{Kikukawa:1989fw}\begin{equation}
\beta(g) =- \frac{D-2}{g_{\rm cr}} \, g\,(g-g_{\rm cr})\,,\quad \gamma_m=(D-2)\frac{g}{g_{\rm cr}} \rightarrow D-2 \,\,(g\rightarrow g_{\rm cr})\,,
\label{DNJLbetagamma}
\end{equation}
which follows from the gap equation Eq.(\ref{SSBDNJL}) and the condensate $\langle \bar \psi \psi\rangle \sim (\Lambda/m_F)^{D-2}$, respectively, while the renormalization can be done independently of the phase in this theory.
For $D=2$ the ultraviolet fixed point $g=0$ and the infrared fixed point $g=g_{\rm cr}$
coincide, i.e.,
\begin{equation}
\beta(\alpha)\rightarrow -2 N_C g^2,\quad \gamma_m\rightarrow 2 N_C g,
\label{2NJLbetagamma}
\end{equation}
which also follows directly from Eq.(\ref{D2NJL}). The essential singularity scaling in Eq.(\ref{D2NJL}) and the associated colliding ultraviolet and infrared fixed points at $D=2$ are
characteristic features of the conformal phase transition \cite{Miransky:1996pd} similarly to the BKT-Miransky scaling, where there exist no light bound states for $g< g_{\rm cr}=0$, while the bound states
have mass of order of ${\cal O} (m_F) \rightarrow 0$ ($g\searrow g_{\rm cr}=0$) for $g>g_{\rm cr}$.
It is also compared with the beta function of the
gauged NJL model with walking gauge coupling $\alpha(\mu)=\alpha=$ constant: \cite{Kondo:1991yk}
\begin{equation}
\beta(g)=\frac{\partial g}{\partial \ln \Lambda}\Bigg|_{\alpha, m_F}= - 2 N_C \left(g-g^{(+)}(\alpha)\right) \left(g-g^{(-)}(\alpha) \right)\,,
\label{gaugedNJLbeta}
\end{equation}
where \cite{Kondo:1988qd,Yamawaki:1988na}
\begin{equation}
N_C \, g^{(\pm)}(\alpha)=\frac{1}{4} \left(1\pm \omega\right)^2\,, \quad \omega \equiv \sqrt{1-\frac{\alpha}{\alpha_{\rm cr}}}\quad \left(0<\alpha<\alpha_{\rm cr}\right)\,.
\label{criticalline}
\end{equation}
This follows from the gap equation (for $m_0=0$):
\begin{equation}
\Sigma(x) = \frac{N_C G}{4\pi^2} \int^{\Lambda^2} d x \frac{x \Sigma(x)}{x+\Sigma(x)} +
\frac{3 C_2}{4\pi} \alpha \int^{\Lambda^2}
dy\, \left[\frac{\theta(x-y)}{x} + \frac{\theta(y-x)}{y}\right] \frac{y \Sigma(y)}{ y +\Sigma^2(y)},
\label{gaugedNJLSD}
\end{equation}
which is a combined gap equation of Eq.(\ref{NJLgap}) and Eq.(\ref{SDeq}).
Were it not for the NJL interaction $G=0$, the SSB solution would exist only for $\alpha>\alpha_{\rm cr}$ as we have explained in section 3. However, due to strong NJL coupling,
this time it has the SSB solution for $g=G\Lambda^2/(4\pi^2) > g^{(+)}(\alpha)$ even at $\alpha<\alpha_{\rm cr}$: \cite{Kondo:1988qd,Yamawaki:1988na}
\begin{equation}
m_F^{2\omega} = \Lambda^{2\omega} \left(
\frac{g-g^{(+)}(\alpha)}{g-g^{(-)}(\alpha)} \right) \,,\quad \left(g>g^{(+)}(\alpha)=g_{\rm cr}(\alpha)\,, \quad 0<\alpha<\alpha_{\rm cr}\right)\,
\label{gaugedNJLsol}
\end{equation}
as well as $\alpha>\alpha_{\rm cr}$, while $g<g^{(+)}(\alpha)$ is
the unbroken phase, $m_F= 0$,
where the criticality is now extended to the critical line $g_{\rm cr}(\alpha)=g^{(+)}(\alpha)$ in the $(g,\alpha)$ space instead of just the gauge coupling $\alpha$.
The beta function Eq.(\ref{gaugedNJLbeta}) is readily obtained from Eq.(\ref{gaugedNJLsol}), and
the anomalous dimension is given as:~\cite{Miransky:1988gk, Kondo:1991yk}
\begin{equation}
\gamma_m =2N_C g +\frac{\alpha}{2\alpha_{\rm cr}}\,, \quad
\gamma_m^{(\pm)} = \gamma_m\Bigg|_{g=g^{(\pm)}(\alpha)}=1\pm \omega= 1 \pm \sqrt{1-\frac{\alpha}{\alpha_{\rm cr}}} \,.
\label{gaugedNJLgamma}
\end{equation}
The critical line $g=g_{\rm cr}(\alpha)=g^{(+)}(\alpha)$, coming from the four-fermion coupling additional to the gauge dynamics (e.g., ETC coupling in the technicolor case), behaves as an ultraviolet fixed point, while the non-critical line $g^{(-)}(\alpha)$, coming from the four-fermion coupling induced by the gauge dynamics itself in a Wilsonian sense, does as an infrared fixed point, both colliding ${\bar g}_{\rm cr}=g^{(+)}(\alpha)=g^{(-)}(\alpha)=1/(4 N_C)$ at $\alpha=\alpha_{\rm cr}$,
where the
SD gap equation yields the Miransky-BKT scaling of the essential-singularity form (conformal phase transition):
\begin{equation}
m_F =\Lambda \exp \left( - \frac{g}{ g -{\bar g}_{\rm cr}} \right)\,, \quad \left(\alpha=\alpha_{\rm cr}, \, g>{\bar g}_{\rm cr}=\frac{1}{4 N_C}\right)\,.
\label{gNJLgapcritical}
\end{equation}
Accordingly, the
beta function and the anomalous dimension read
\begin{equation}
\beta(g)= - 2 N_C \left(g-{\bar g}_{\rm cr}\right)^2\,, \quad \gamma_m=2 N_C g+\frac{1}{2} \quad \left(\gamma_m\Big|_{{\bar g}_{\rm cr}}=1\right),
\label{gNJLbetacritical}
\end{equation}
which is compared with Eq.(\ref{2NJLbetagamma}).
The outstanding feature of the gauged NJL model with $0<\alpha \leq\alpha_{\rm cr},\, (1\leq\gamma_m<2)$ is the renormalizability
(in the sense of nontriviality, or no Landau pole) independently of the phase \cite{Kondo:1991yk}, when the gauge coupling is walking, $\alpha(\mu^2) \approx {\rm const.}$. (Similar results are obtained for the gauged Yukawa model \cite{Kondo:1993ty}.)
The four-fermion operators have the full dimension $2< d_{(\bar \psi \psi)^2} = 2(3-\gamma_m) = 4-2\omega \leq4$ (relevant operator, or super renormalizable), including $d\simeq 2(1+ A/\ln \mu^2)>2 (d \ne 2)$
with a moderately ``walking''
small coupling $\omega \simeq 1 -\frac{\alpha}{2\alpha_{\rm cr}} \simeq 1-\gamma_m \,\,(\gamma_m (\mu) \sim A/\ln \mu^2) $ with $A=18 C_2/(11 N_C -2 N_F) >1$ \footnote{For the NJL gauged by the actual QCD with 6 flavors ($u,\dots, t$), $A=8/7>1$, as in the original top quark condensate model \cite{Miransky:1988xi} satisfies
this renormalizability condition \cite{Kondo:1991yk}, though the electroweak gauge interaction invalidates it.
If the SM gauge groups are embedded into a GUT which is usually walking $A>1$, then the GUT-gauged NJL is renormalizable, having the interacting continuum limit.
}
, in sharp contrast to the pure (non-gauged) NJL model with $\gamma_m=2,\, d_{(\bar \psi \psi)^2}=2$, which is a trivial theory having a Landau pole in the conventional way of the continuum limit keeping $m_F=$ constant (not the way described in the above keeping $F_\pi=$ constant).\footnote{
For renormalizability of the gauge theories and gauged NJL model in $D>4$ dimensions, with the extra dimensions $\delta=D-4$ compactified, see
Ref.\cite{Hashimoto:2000uk}.
}
\section{Top Quark Condensate a la NJL dynamics, the simplest UV completion of the SM Higgs}
\label{TopQuarkCondensate}
One of the concrete composite Higgs models as the
straightforward
application of the NJL type theory is the top quark condensate model (Top-Mode Standard Model) \cite{Miransky:1988xi,Nambu:1989jt,Bardeen:1989ds}.
The model predicted that only the top quark among SM fermions has mass on the order of the weak scale (Fig.\ref{fig:SMLego}), at the time many people expected the top mass below 50 GeV.
\begin{figure} [h]
\begin{center}
\includegraphics[width=6cm]{SMLego.eps}
\caption{Lego version of the Standard Model. Top quark mass vs. other masses in linear scale. }
\label{fig:SMLego}
\end{center}
\end{figure}
The explicit four-fermion Lagrangian of the top quark condensate takes the form: \cite{Miransky:1988xi}
\begin{eqnarray}
{\cal L}_{\rm MTY}&=&
G^{(1)} \left({\bar \psi}_L^i \psi_R^j\right)\left({\bar \psi}_R^j \psi_L^i\right)
+\left[G^{(2)}\left({\bar \psi}_L^i \psi_R^j\right)\left(i\tau_2\right)^{ik} \left(i\tau_2\right)^{jl}
\left({\bar \psi}_L^k \psi_R^l\right)+h.c.\right]\nonumber\\
&+&
G^{(3)} \left({\bar \psi}_L^i\psi_R^j\right)
\left( \tau_3\right)^{jk} \left({\bar \psi}_R^k \psi_L^i\right)\,, \quad \left(G^{(i)}= g^{(i)} \frac{4\pi^2}{\Lambda^2}\right)\,.
\label{MTY}
\end{eqnarray}
The inclusion of other generations is straightforward \cite{Miransky:1988xi}.
In the realistic case the SM gauge interaction, particularly QCD, was included via the gauged NJL model \cite{Miransky:1988xi},
where the critical coupling is actually the critical line of the gauged NJL, $G_{\rm cr}(\alpha)= g_{\rm cr}(\alpha)\cdot (4\pi^2/\Lambda^2)$ with
$ g_{\rm cr}(\alpha)$ in Eq.(\ref{criticalline}),
while the $U(1)_Y$ coupling is numerically negligible (the chiral gauge $SU(2)_L$ does not contribute to the condensate channel).
The crucial ingredient of the model is again the {\it non-zero critical coupling} {\it in sharp contrast to the ``bootstrap symmetry breaking'' \cite{Nambu:1989jt}
based on the weakly-coupled BCS theory which has $G_{\rm cr}=0$} as already mentioned:
{\it only the top quark coupling is strong coupling larger than the critical coupling}
$G_t =G^{(1)}+G^{(3)} > G_{\rm cr}(\alpha)$ while $G_b=G^{(1)}-G^{(3)}$ and all others are less as well, $G_{c,s,d,u} <G_{\rm cr}(\alpha)$, so that only the top acquires the dynamical mass of order of weak scale ${\cal O}(v)$ to produce
only three NG bosons to be absorbed into the $W/Z$ bosons \cite{Miransky:1988xi,Bardeen:1989ds}.
We disregard $G^{(2)}$ and $G_b$ terms for the moment, then
Eq.(\ref{MTY}) simply reads:
\begin{eqnarray}
{\cal L}_{\rm MTY}&=& G_t \left( {\bar \psi}_L t_R\right) \left({\bar t}_R \psi_L\right) \nonumber \\
&=& \bar \psi_L h t_R +h.c. - \frac{1}{G_t} h^\dagger h \,,
\end{eqnarray}
where $h=G_t \bar \psi_L t_R$ with $\hat \sigma \sim G_t \bar t t, {\hat \pi}^0\sim G_t \bar t i\gamma_5 t, \hat \pi^{\pm}\sim G_t \bar t i\gamma_5 b, G_t
\bar b i\gamma_5 t$. This simplified version was also considered in \cite{Bardeen:1989ds}.
The effective theory of the pure bosonic sector at $1/N_C$ leading order is precisely the same as the SM Higgs Lagrangian Eq.(\ref{Higgs}) \cite{Bardeen:1989ds}
as already mentioned in Section \ref{NJLvsSM},
which happens to have the $SU(2)_L \times SU(2)_R$ global symmetry
not just the $SU(2)_L \times U(1)_R$ to be gauged by the electroweak gauge bosons. (The $SU(2)_R$ is explicitly broken only by the Yukawa term:
$\bar \psi_L h t_R$.)
Then the effective theory of the top quark condensate model is nothing but the SM Higgs as the scale-invariant HLS model, Eq.(\ref{SHLS}), which includes the 3 NG bosons $\pi^{\pm,0}$ to be absorbed into $W,Z$ and
the Higgs $\phi(x)$ as a pseudo-dilaton, in the nonlinear realization in Eq.(\ref{Polar}) and (\ref{NLscale}), and in addition the vector composites (``top-mode rho meson'')
which can be identified \cite{MY16} with the 2 TeV diboson events at LHC \cite{Aad:2015owa}.
If we further assume a small bottom condensate $\langle \bar b b \rangle$ by fine-tuning the bottom four-fermion coupling: $G_t>G_b >G_{\rm cr}(\alpha)$,
then we have
a Peccei-Quinn type axion (``top-mode axion'') \cite{Miransky:1988xi,Luty:1990bg}, which acquires mass from the $G^{(2)}$ term in Eq.(\ref{MTY}).
(It may be identified with the 750 GeV diphoton
events at LHC \cite{750}.)
The obvious phenomenological problem of the top-mode SM is the prediction of the top mass in the large $N_C$ limit relation (BCS mass relation) to the Higgs mass:
$M_\phi=2 m_t$.
which is actually modified by the effects of the SM gauge interactions,
$M_\phi \simeq \sqrt{2} m_t$ \cite{Bardeen:1989ds, Shuto:1989te}. It is further modified to $M_\phi\simeq m_t$
by some of the non-leading order in $1/N_C$ expansion \cite{Bardeen:1989ds} using the ultraviolet boundary condition, ``compositeness condition'' \cite{Bardeen:1989ds}, at say the GUT
scale,
where both effective couplings of top Yukawa $g_Y^t$ and the Higgs quartic coupling $\lambda$ diverge.
Instead of the compositeness condition, we may consider the renormalizability of the gauged NJL model mentioned in the previous section.
A possible renormalizable top quark condensate model would then
be to unify the SM gauge interactions into a walking GUT, ``Top-mode walking GUT'' \cite{Yamawaki:1996tj},
which determines the values of the top Yukawa coupling $g_Y^t$
and the Higgs coupling $\lambda$ in terms of the GUT gauge coupling $g_{\rm GUT}$ all at the GUT scale $\Lambda_{\rm GUT}$
as the Pendleton-Ross infrared fixed point of the effective theory of the GUT-gauged NJL model, typically as ${g_Y^t}^2(\Lambda_{\rm GUT})\simeq \lambda(\Lambda_{\rm GUT}) \simeq \frac{3}{2} \,g_{\rm GUT}^2(\Lambda_{\rm GUT})$, instead of the diverging couplings of the compositeness condition. This generally yields prediction of mass of $m_t$ and $M_\phi$ much smaller than that of compositeness condition.
Another possibility would be to include the near marginal operators: since the anomalous dimension
is very close to 2, the four-fermion operators (corresponding to the bare $\lambda \phi^4$ term) have the dimension $d\simeq 2$, so that formally irrelevant
eight-fermion operators could be near
marginal and compete with the higher order corrections in $1/N_C$ expansion, which may change the mass ratio
substantially. It is not known presently whether or not the relation $M_\phi \simeq m_t/2$ can be naturally realized by yet
higher order effects in $1/N_C$ as well as the eight-fermion operators.
Yet other different solutions have been considered, see e.g.
top seesaw \cite{Dobrescu:1997nm}, and its NG boson Higgs version \cite{Fukano:2013aea} where the Higgs is a pseudo NG boson
living in the coset space of the larger internal symmetry $G/H=U(3)\times U(1)/[U(2)\times U(1)^\prime]$ rather than the pseudo-dilaton.
The LHC Run II will tell us whether or not the basic idea of the top quark condensate is on the right track.
\section{Walking Technicolor and Technidilaton}
Yet another composite Higgs model in the spirit of Nambu is the strong coupling gauge theory, similar in some sense to the
QCD, a simple scale-up version dubbed technicolor \cite{TC}. However the original technicolor was excluded long time ago for
the problem of the Flavor-Changing Neutral Currents (FCNC) in a way to give mass to the SM fermions through four-fermion interaction from Extended Technicolor (ETC) \cite{Dimopoulos:1979es}
or some composite technicolor models \cite{Yamawaki:1982tg}. As the strong coupling gauge theory with {\it more explicit role of the non-zero critical coupling}, the walking technicolor \cite{Yamawaki:1985zg,Bando:1986bg} was proposed as a solution to the FCNC problem by a large anomalous dimension $\gamma_m =1$, based on the SSB solution of the scale-invariant dynamics, the
ladder SD equation, and at the same time predicted a light composite Higgs, dubbed technidilaton, as a pseudo NG boson of the approximate scale symmetry.\footnote{
Similar work on the FCNC problem \cite{Holdom:1984sk} was also done without notion of the anomalous dimension, the scale symmetry,
nor the technidilaton.
Solving FCNC by a large anomalous dimension was proposed earlier \cite{Holdom:1981rm}, based on a pure assumption of the existence of a gauge theory
having nontrivial ultraviolet fixed point.
}
Here I recapitulate the explanation \cite{Matsuzaki:2015sya} on how the
technidilaton is a {\it naturally light and weakly coupled composite Higgs} out of {\it strongly coupled} underlying conformal gauge theory, the walking technicolor, in the light of the anti-Veneziano limit
Eq.(\ref{antiVeneziano}).
The technidilaton particularly for the one-family technicolor, with $N_F=8$ and $N_C=4$ \cite{Dimopoulos:1979es}, as the walking technicolor is nicely fit to the current 125 GeV Higgs data at LHC \cite{Matsuzaki:2012mk,Matsuzaki:2012xx,Matsuzaki:2015sya}.
As we discussed in section \ref{SCGT}, ladder approximation is realized by the anti-Veneziano limit Eq.(\ref{antiVeneziano}) in the large $N_F$ QCD. The evaluation of the nonperturbative trace anomaly in the anti-Veneziano limit can be essentially given
by the ladder result Eq.(\ref{NPanomaly}).
It then yields the mass $M_\phi$ and decay constant $F_\phi$ of the technidilaton $\phi$
through PCDC \cite{Bando:1986bg} as in Eq.({\ref{PCDC}), this time in terms of the dimension 4 operator:\cite{Matsuzaki:2015sya,Hashimoto:2010nw}
\begin{eqnarray}
M_\phi^2 F_\phi^2&=&
-d_\theta \langle \theta_\mu^\mu \rangle
=- \frac{\beta^{(\rm NP)}(\alpha (\mu^2))}{\alpha(\mu^2)}
\, \langle G_{\mu \nu}^2(\mu^2)\rangle
\simeq N_C N_F\frac{16 \xi^2}{\pi^4} m_F^4\quad \left(d_\theta=4 \right)
\label{PCDC2}\\
&\simeq& 2.5\left[\frac{8}{N_F}\frac{4}{N_C}\right] v^4 \,. \quad \left(v= 246\,{\rm GeV}\right)
\label{PCDC3}
\end{eqnarray}
First, the rightmost value in Eq.(\ref{PCDC2}) can be obtained by two different ladder calculations:
one through direct evaluation of the vacuum energy by the effective potential at the stationary point (Solution of the SD equation, $\Sigma=\Sigma_{\rm sol}$)\cite{Gusynin:1987em}, $E=V_{\rm eff} (\Sigma =\Sigma_{\rm sol}) =\langle\theta^0_0\rangle
=(1/4)\langle\theta^\mu_\mu\rangle$, the other through the ladder evaluation of the trace anomaly \cite{Hashimoto:2010nw,Matsuzaki:2015sya}, i.e., the technigluon condensate $\langle G_{\mu\nu}^2\rangle$ times the nonperturbative beta function
Eq.(\ref{NPbeta}), both in precise agreement with each other. The agreement is in highly nontrivial manner, being {\it independent of the renormalization point $\mu$} as it should be:
$\langle G_{\mu \nu}^2(\mu^2)\rangle \sim \ln^3 (\mu^2/m_F^2)$, while $\beta^{({\rm NP})} (\alpha (\mu^2)) /\alpha(\mu^2) \sim 1/\ln^3 (\mu^2/m_F^2)$, precisely cancelled by each other.\cite{Matsuzaki:2015sya}
Second, Eq.(\ref{PCDC3}) is obtained by use of the Pagels-Stokar formula:
\begin{equation}
v^2=(246\, {\rm GeV})^2
= N_D F_\pi^2 \simeq
N_F N_C\frac{\xi^2}{4\pi^2} \, m_F^2
\simeq m_F^2 \left[\frac{N_F}{8}\frac{N_C}{4}\right],
\label{PS}
\end{equation}
and the result indicates {\it important $N_F,N_C-$ dependence of $M_\phi^2F_\phi^2$ in the anti-Veneziano limit when $v=$ fixed} \cite{Matsuzaki:2015sya}.
Since the technidilaton is a flavor-singlet bound state, its decay constant by definition scales like $F_\phi^2 \propto N_F N_C m_F^2\, (\propto v^2)$ (Actually $F_\phi^2 \simeq N_F N_C m_F^2$). Then
$M_\phi^2/F_\phi^2, M_\phi^2/v^2 \sim 1/(N_F N_C) \rightarrow 0$ in the anti-Veneziano limit, where the technidilaton becomes NG boson although no exact
massless limit exists: the situation is in the same sense as the
$\eta^\prime$ meson in the original Veneziano limit $N_C \rightarrow \infty$ with $N_C\, \alpha=$fixed, and $N_F/N_C \ll 1$.\footnote{
There exists no exact massless limit in the conformal phase transition at $\alpha=\alpha_{\rm cr}$ (this time $r(=N_F/N_C)= r_{\rm cr}$), with $m_F=0$, where no massless spectrum exists for $\alpha\leq \alpha_{\rm cr}$ (conformal phase),
in sharp contrast to the Ginzburg-Landau phase transition where the spectrum continuously passes through the phase transition point with massless particles. \cite{Miransky:1996pd}
}
Although the PCDC relation together with Pagels-Stokar formula does not give $M_\phi$ and $F_\pi$ separately,
Eq.(\ref{PCDC3}) well accommodates numerically the desired result:
\begin{equation}
F_\phi\simeq 5 \,v \quad {\rm for} \,\, M_\phi \simeq \frac{v}{2} \simeq 125\,{\rm GeV}\quad \left(N_F=8, N_C=4\right)\,,
\label{dilatonHiggs}
\end{equation}
in the one-family model, which is best fit to the current LHC data of the 125 GeV Higgs up to 30 \% uncertainty due to limitation of the ladder approximation \cite{Matsuzaki:2012mk,Matsuzaki:2015sya}.
Similar results are also obtained within 30\% uncertainty in the holographic model for the walking technicolor.\cite{Matsuzaki:2012xx}
Incidentally, at the criticality $\alpha=\alpha_{\rm cr}=\frac{\pi}{3 C_2}$, the anomalous dimension $\gamma_m=1$ implies that the induced four-fermion interaction
generated by the walking technicolor coupling itself (i.e., not the ETC-like gauge interaction additional to the technicolor interaction) also becomes
a marginal operator with $d=4$. Then the phase diagram should be considered in the wider coupling space
$(\alpha, g)$\cite{Leung:1985sn}, where $g$ is the dimensionless coupling of the induced four-fermion interaction in the form of
the gauged NJL model. See Eq.(\ref{gNJLgapcritical}) and (\ref{gNJLbetacritical}). Then we can predict $M_\phi$ independently of
$F_\phi$: the denominator of the renormalized $\sigma$ propagator $D^\sigma(p)$ can be evaluated at $p=0$ in the large $N_C$ limit \cite{
Nonoyama:1989dq}:
\begin{equation}
M_\phi^2=D^\sigma(0)=\frac{16\xi^2}{\pi^4}m_F^2\simeq \left(\frac{m_F}{2}\right)^2 \quad \left(\alpha=\alpha_{\rm cr}\,,\, g\searrow {\bar g}_{\rm cr}\right)\,,
\end{equation}
which yields $M_\phi \simeq \frac{v}{2}\simeq 125$ GeV (!) through the Pagels-Stokar formula Eq.(\ref{PS}) for $N_F=8, N_C=4$, i.e., $v\simeq m_F$,
and in turn predicts
$F_\phi^2 \simeq N_F N_C m_F^2$,
combined with the PCDC relation Eq.(\ref{PCDC2}). The result is quite consistent with Eq.(\ref{dilatonHiggs}) for the 125 GeV Higgs.
The effective theory of the walking technicolor with $N_F$ massless flavors takes {\it precisely the same scale-invariant form as the nonlinearly realized SM Higgs Lagrangian} in Eq.(\ref{SNLSM}), with $U=e^{i \pi^a\, T^a}$ being $N_F\times N_F$ unitary matrix (${\rm tr} \,T^a=0\,, {\rm tr} (T^aT^b)=\delta^{ab}/2,\,a=1,\cdots, N_F^2-1$), except that
the explicit scale breaking comes from the different potential $V^{(4)}(\phi)$~\footnote{
This potential is indeed obtained by the explicit ladder computation of the effective potential at the conformal phase transition point: $V^{(4)}(\phi)= -( 4N_F N_C m_F^4/\pi^4) \chi^4 (\ln \chi-1/4)$, in precise agreement with
Eq.(\ref{WTC}) through Eq.(\ref{PCDC2}). See Eq.(65) in Ref.~\cite{Miransky:1996pd}}
responsible for the {\it trace anomaly of dimension 4 operator} this time:~\cite{Matsuzaki:2012mk,Matsuzaki:2015sya}
\begin{eqnarray}
{\cal L}_{\rm WTC}
&=&\chi^2 \cdot \left[ \frac{1}{2} \left(\partial_\mu \phi\right)^2 +\frac{F_\pi^2}{4}{\rm tr} \left({\cal D}_\mu U {\cal D}^\mu U^\dagger\right)\right] -V^{(4)}(\phi)-V^{(\rm SM)}(\phi) \nonumber \,,\\
V^{(4)}(\phi)&=&-\ln \chi\cdot \frac{\beta^{(\rm NP)}(\alpha)}{4\alpha} G_{\mu\nu}^2= \frac{M_\phi^2 F_\phi^2}{4} \chi^4 \left(\ln\chi-\frac{1}{4}\right)\nonumber\\
V^{(\rm SM)}(\phi)&=& - \chi^{2-\gamma_m} \left(m_f \chi \bar f f\right) -\ln \chi \left[\frac{\beta_F(\alpha_s)}{4\alpha_s} {G^{({\rm gluon})}_{\mu\nu}}^2
+ \frac{\beta_F(\alpha_e)}{4\alpha_e} {F^{({\gamma})}_{\mu\nu}}^2\right]\,,\,\nonumber\\
\chi &= &\exp \left(\frac{\phi}{F_\phi}\right)\,,
\label{WTC}
\end{eqnarray}
where {\it $F_\phi\ne F_\pi=v/\sqrt{N_D}=v/\sqrt{N_F/2}$ in general} in contrast to the SM Higgs case $F_\phi=F_\pi=v$, the electroweak gauging was done as usual $\partial_\mu U\Rightarrow {\cal D}_\mu U= \partial_\mu U -i g_2 W_\mu U +i g_1 U B_\mu$, and we have
added $V^{(\rm SM)}(\phi)$, the scale symmetry breaking terms related to the SM particles arising from the technifermion contributions: mass term of the SM fermion $f$, (one loop) technifermion contributions to the
trace anomaly for the gluon and photon operators, with $\beta_F(g_s) = \frac{g_s^3}{(4\pi)^2} \frac{4}{3} N_C$ and $\beta_F(e) = \frac{e^3}{(4\pi)^2} \frac{16}{9} N_C$.
It is obvious that ${\theta_\mu^\mu}^{({\rm TC})}=- \delta_D V^{(4)}(\phi) = \beta^{(\rm NP)}(\alpha)/(4\alpha)\cdot G_{\mu\nu}^2=- (M_\phi^2F_\phi^2/4) \chi^4$ up to total derivative, corresponding to
the PCDC with $d_\theta=4$ ($\langle \chi\rangle=1$), Eq.(\ref{PCDC2}).
The technidilaton potential $V^{(4)}(\phi)$ is expanded in $\phi/F_\phi$:
\begin{equation}
V^{(4)}(\phi)
= - \frac{M_\phi^2 F_\phi^2}{16} +\frac{1}{2}M_\phi^2\,\phi^2 +\frac{4}{3} \frac{M_\phi^2}{F_\phi} \,\phi^3
+ 2 \frac{M_\phi^2}{F_\phi^2}\, \phi^4
+ \cdots
\,,
\label{dilatonpotential}
\end{equation}
which shows a remarkable fact that in the anti-Veneziano limit
the technidilaton self couplings (trilinear and quartic couplings) are highly suppressed:
\begin{equation}
\lambda_{\phi^3}=4M_\phi^2/(3F_\phi) \sim 1/\sqrt{N_F N_C}\rightarrow 0\,, \quad
\lambda_{\phi^4}=2 M_\phi^2/F_\phi^2 \sim 1/(N_F N_C)\rightarrow 0\,,
\label{weakcouplingTD}
\end{equation}
by $M_\phi/F_\phi \sim 1/\sqrt{N_F N_C}$ and $M_\phi \sim N_F^0 N_C^0$.
Numerically, we may compare the technidilaton self couplings
with those
of the SM Higgs with $m_h=M_\phi=125$ GeV for $v/F_\phi \simeq 1/5$ in the one-family model ($N_F=8, N_C=4)$: \cite{Matsuzaki:2015sya}
\begin{eqnarray}
\frac{\lambda_{\phi^3}}{\lambda_{h_{\rm SM}^3}}\Bigg|_{M_\phi=m_h}
&=& \frac{\frac{4 M_\phi^2}{3 F_\phi}}{\frac{m_h^2}{2 v}} \Bigg|_{M_\phi=m_h}
\simeq \frac{8}{3} \left( \frac{v}{F_\phi}\right) \simeq 0.5 \,, \nonumber \\
\frac{\lambda_{\phi^4}}{\lambda_{h_{\rm SM}^4}} \Bigg|_{M_\phi = m_h}
&=& \frac{\frac{2 M_\phi^2}{F_\phi^2}}{\frac{m_h^2}{8 v^2}}\Bigg|_{M_\phi=m_h}
= 16 \left( \frac{v}{F_\phi} \right)^2 \simeq 0.6
\,.
\label{selfcouplings:0}
\end{eqnarray}
This shows that the {\it technidilaton self couplings, although generated by the strongly coupled interactions, are even smaller than those of the
SM Higgs}, a salient feature of the approximate scale symmetry in the ant-Veneziano limit !!
The coupling of the technidilaton ($M_\phi=125$ GeV) to the SM particles
can be seen by expanding $\chi= 1+\phi/F_\phi +(1/2!)(\phi/F_\phi)^2+\cdots$ in Eq.(\ref{WTC}):
\begin{eqnarray}
\frac{g_{\phi WW/ZZ}}{g_{ h_{\rm SM} WW/ZZ }}
=
\frac{g_{\phi ff}}{g_{h_{\rm SM} ff}} \,
=
\frac{v}{F_\phi}
\,.
\label{WWZZ}
\end{eqnarray}
\begin{eqnarray}
\frac{g_{\phi gg}}{g_{h_{\rm SM} gg}}
\simeq
\frac{v}{F_\phi}
\cdot
\left( 1 + 2 N_C \right),\,
\frac{g_{\phi \gamma\gamma}}{g_{h_{\rm SM} \gamma\gamma}}
\simeq \frac{v}{F_\phi}
\cdot
\left( \frac{63 - 16}{47} - \frac{32}{47} N_C \right)
\,, \label{g-dip-dig}
\end{eqnarray}
where besides the technifermions loop, only the top and W of the SM contributions were included at one-loop.
Note the couplings in Eq.(\ref{WWZZ}) with $v/F_\phi \sim1/5$ are even weaker than the SM Higgs, which are however compensated by
those in Eq.(\ref{g-dip-dig}) for $gg$ and $\gamma\gamma$ rather enhanced
by the extra loop contributions of the technifermions other than the SM particles, particularly for large $N_C$, resulting in signal strength similar to the SM Higgs within the current experimental accuracy.
In fact the current LHC data for 125 GeV Higgs are fit by the technidilaton as good as by the SM Higgs,
particularly for $N_F=8, N_C=4$, i.e., near the anti-Veneziano limit.~\cite{Matsuzaki:2012mk}
Most recent detailed analyses are given in Ref. \cite{Matsuzaki:2015sya}.
It should be mentioned here that the one-family model will be most naturally imbedded into the ETC in the case for $N_C=4$ \cite{Kurachi:2015bva}.
More precise data at LHC Run II will discriminate among them, SM Higgs or technidilaton.
What about the technipions? In the walking technicolor with $N_D =N_F/2>1$, the spontaneous breaking of the chiral symmetry larger than $SU(2)_L\times SU(2)_R$ produces NG bosons (technipions) more than 3 to
be absorbed into W/Z. Let us take the one-family model with $N_F=8$, which has colored techniquarks (3 weak doublets) $Q^a_i$ and non-colored technileptons (one weak-doublet) $L_i$
($a=1,2,3;i=1,2$), the resultant chiral symmetry being $SU(8)_L \times SU(8)_R$ \cite{Dimopoulos:1979sp}. There are 63 technipions, 60 of which are unabsorbed technipions. All of them acquire the mass from the explicit chiral symmetry breaking due to the SM gauge and ETC gauge interactions. Due to the large anomalous dimension $\gamma_m\simeq 1$, the mass of them are all
enhanced to TeV region \cite{Kurachi:2014xla}, which will be discovered at LHC Run II.
(After this symposium, 750 GeV diphoton events were reported at LHC \cite{750}, which can be identified with the color-singlet and iso-singlet (not flavor-singlet) technipion
$P^0$ \cite{Matsuzaki:2015che,Lebiedowicz:2016lmn}.
If it is the case, $M_{P^0}=750$ GeV, the model predicts another nearby color-singlet technipion $P^i$ (iso-triplet one), with mass $M_{P^i}= \sqrt{\frac{8}{5}} M_{P^0} \simeq
950$ GeV, a salient prediction of the one-family model independently of the dynamical details \cite{Kurachi:2014xla}.)
Another signature of the walking technicolor is the prediction of higher resonances such as the spin 1 boson, the walking techni-$\rho$,
walking techni-$a_1$, etc..
The straightforward $N_F$ extension of Eq.(\ref{SHLS}) is also obvious: Eq.(\ref{WTC}) is gauge equivalent to the {\it scale-invariant HLS Lagrangian} explicitly constructed for one-family walking technicolor with $N_F=8$ \cite{Kurachi:2014qma}:
\begin{eqnarray}
{\cal L}_{\rm s-HLS} &=& {\cal L}_{\rm WTC}+ {\cal L}_{\rm Kinetic} \left(V_\mu\right)\,,\nonumber \\
{\cal L}_{\rm WTC}&=& \chi^2 \cdot \left(\frac{1}{2} \left(\partial_\mu \phi\right)^2 + {\cal L}_A+ a {\cal L}_V\right) -V^{(4)}(\phi) \,,
\label{SHLS2}
\end{eqnarray}
where the HLS gauge bosons $V_\mu$ in the mass term $ a \chi^2 {\cal L}_V=a e^{2\phi/F_\phi} {\rm tr} (g_{_{\rm HLS}} V_\mu +\cdots)^2$ are the bound states of the walking technicolor, the walking techni-$\rho$,
with {\it mass term $M_V^2=a g_{_{\rm HLS}}^2 F_\pi^2$ being scale-invariant} thanks to the overall technidilaton factor $\chi^2$, as mentioned before.
The loop expansion is formulated as the scale-invariant HLS perturbation theory \cite{Fukano:2015zua} in the same way as the scale-invariant chiral perturbation theory \cite{Matsuzaki:2013eva}, a straightforward extension of the (non-scale-invariant) HLS perturbation theory \cite{Harada:2003jx}.
In fact there has been reported an interesting 2 TeV diboson event at LHC \cite{Aad:2015owa}.
We have shown \cite{Fukano:2015hga}
that this would be
the most natural candidate for the Drell-Yan produced walking techni-$\rho$, the color-singlet iso-triplet $\rho^{i}$, as a gauge boson of the HLS described
by the scale-invariant HLS model in Eq.(\ref{SHLS2}) \cite{Kurachi:2014qma}.
We further found \cite{
Fukano:2015zua}
that a salient feature of this possibility is
the scale symmetry which forbids the decay of the walking techni-$\rho$ to the 125 GeV Higgs (technidilaton) plus $W/Z$ (what we called
``conformal barrier'') in the same way as the hidden vector in the SM Higgs, Eq. (\ref{Conformalbarrier}), in sharp contrast to the popular ``equivalence theorem''.
The HLS is readily extendable to include techni-$a_1$, etc. \cite{Bando:1984ej,Bando:1987br,Harada:2003jx}, with an infinite set of the HLS tower being equivalent to the deconstructed extra dimension \cite{ArkaniHamed:2001ca} and/or the holographic models \cite{Son:2003et}, and the scale-invariant version of them are also straightforward and {\it mass term of all the higher HLS vector bosons
are scale-invariant}, an outstanding characterization in sharp contrast to other formulations for the spin 1 bosons. The conformal barrier applies not only to the techni-$\rho$ but
also to all the higher vector/axialvector resonances as the HLS gauge bosons, having scale-invariant mass term. The LHC Run II will tell us whether or not it is the case.
\section{Walking Technicolor on the Lattice}
Finally, I would like to make a brief review on the lattice studies on the walking technicolor particularly done by our group, the
LatKMI Collaboration \cite{Aoki:2012eq,Aoki:2013xza,Aoki:2013zsa,Aoki:2014oha}.
Since the dynamics is essentially nonperturbative,
reliable calculations would be eventually done by the lattice simulations.
In fact there has been extensive activity in the lattice simulations of the candidate theories for the walking technicolor, particularly the large $N_F$ QCD, i.e.,
the $SU(N_C)$ gauge theory with $N_F$ degenerate flavors of fundamental representation fermions, eventually extrapolated to the
chiral limit \cite{Lattice}. Among others particular interest has been paid to the $N_f=8$ and $N_F=12$ for $N_C=3$, partly because the infrared fixed point $\alpha_*$ in the two-loop beta function exists for $N_F \gtrsim 8$ ($N_C=3$),
and the SSB criticality condition $\alpha_*=\alpha_{\rm cr}$ for the ladder
result $\alpha_{\rm cr}=\pi/(3 C_2)$ is fulfilled for $N_F \simeq 12$ \cite{Appelquist:1996dq}, so that it is expected that the walking theory might exist somewhere around $8<N_F<12$. Inexpensive simulations are mostly done in the staggered fermion, $N_F=4,8,12,16$, within the asymptotically free
cases.
The LatKMI Collaboration started in 2010 for the lattice simulations on the possible candidate for the walking technicolor by
systematic studies of the $N_F=16,12,8, 4$ on the same lattice set-up, using the HISQ (Highly Improved Staggered Quarks)
action with tree-level Symanzik gauge action. We have mainly focused on the low-lying fermionic bound states (plus some gluonic ones), i,e., pseudoscalar (denoted as $\pi$), scalar ($\sigma,a_0$), vector ($\rho$), axialvector mesons ($a_1$), and nucleon-like states ($N, N^*$), etc. and particularly the flavor-singlet scalar $\sigma$ as a candidate for the technidilaton.
We found \cite{Aoki:2012eq} that $N_F=12$ is consistent with the conformal window indicating no spontaneous chiral symmetry breaking, satisfying the universal hyperscaling relation $M\sim m_f^{1/(1+\gamma_m)}$
for all the observed quantities (with $\gamma_m \sim 0.4\ll 1$), which is in agreement with the results of many other groups \cite{Lattice}.
We further found \cite{Aoki:2013xza} that $N_F=8$, in comparison with $N_F=12$ and $N_F=4$ (up-dated in \cite{Aoki:2015zny}),
is consistent with the SSB phase with remnants of the conformality (approximately universal hyperscaling relation {\it except for the NG boson pion mass}) with a large anomalous dimension
\begin{equation}
\gamma_m \simeq 1\,,
\end{equation}
namely the walking theory, which was confirmed by other groups \cite{Appelquist:2014zsa}.
A remarkable result of the LatKMI Collaboration is the {\it discovery of a light flavor-singlet scalar on the lattice}, lighter than the pion
in $N_F=12$ \cite{Aoki:2013zsa}.
This $N_F=12$ result of us was confirmed by
other groups \cite{Fodor:2014pqa}. Since the theory is consistent with the conformal window without SSB, such a light flavor-singlet scalar in $N_F=12$ may have no
direct relevance to the technidilaton as a composite Higgs. Nevertheless, it is suggestive that the conformal dynamics may produce the dilatonic scalar, which was generated only by
the explicit breaking fermion bare mass $m_0$ put on the lattice.
Furthermore, we made an outstanding discovery that in $N_F=8$ there exists a light flavor-singlet scalar with mass comparable to the pion \cite{Aoki:2014oha}.
See Fig. \ref{fig:Nf8scalar}.
\begin{figure} [h]
\begin{center}
\includegraphics[width=6cm]{m_mf.eps}
\caption{Flavor-singlet scalar meson denoted as $\sigma$ in $N_f=8$ QCD HISQ with $\beta=6/g^2=3.8$, in comparison with the NG boson pion $\pi$ and the vector meson $\rho$, for various
values of the degenerate fermion bare mass $m_f$ on the lattice. Lattice volumes are $L^3\times T=36^3\times 48, 30^3\times 40, 24^3\times 32, 18^3\times 24$. }
\label{fig:Nf8scalar}
\end{center}
\end{figure}
Our $N_F=8$ results were also confirmed by other group \cite{Appelquist:2016viq}.
Since $N_F=8$ seems to be a walking theory in the SSB phase as mentioned above, the light flavor-singlet scalar is particularly attractive as a candidate for the technidilaton.
Also $N_F=8$ is of phenomenological
relevance to the LHC data and of direct relevance to the one-family model as the most natural model building.
Future confirmation of our results is highly desired. Also $N_C=4$ simulations should be studied for various reasons as mentioned before.
\section{Conclusion}
We have discussed the modern version of the Nambu's path to the origin of mass, namely the dynamical symmetry breaking of the electroweak symmetry, which
may account for the origin of the otherwise mysterious input mass parameter (tachyon mass ) of the SM Higgs Lagrangian. The dynamics changes the vacuum by the strong coupling
attractive forces so as to produce the mass scale $m_F (\ll \Lambda)$ smaller than the intrinsic scale $\Lambda (\Lambda_{\rm QCD}) $ which is given either at Lagrangian level (like NJL model)
or induced as the trace anomaly when regularizing the quantum theory (like QCD and technicolor). Crucial to the hierarchy $m_F\ll \Lambda$ is the non-zero critical coupling, which yields a large anomalous dimension and infrared conformality even in the
NJL type four-fermion theory having the coupling of explicit mass scale.
We have defined ``strong coupling theories'' as ``those having non-zero critical coupling'' $N_C g_{\rm cr}={\cal O}(1)$, even though its
value could be small $g \sim 1/N_C \ll 1$ in the typical large $N_C$ limit. The NJL model pioneered by Professor Nambu is
the first and a typical example of such, to be distinguished from its preceding, the BCS theory, which has a zero critical coupling
$g_{\rm cr}=0$. Existence of such a non-zero critical coupling in gauge theory
was discovered by Maskawa and Nakajima in the
scale-invariant dynamics, ladder approximation, and became crucial for the walking technicolor with the coupling
$N_C \alpha> N_C \alpha_{\rm cr} ={\cal O}(1)$ in the SSB phase of the scale symmetry as well as the chiral symmetry.
The existence of the non-zero critical coupling is actually ``hidden'' even in the QCD which is regarded to have only one phase in the ordinary
situation without signal of the no-zero critical coupling: it manifests itself in the extreme condition, such as the
large number of fermions $N_F\gg N_C$ (so as to keep the asymptotic freedom), high temperature, high density, etc..
Indeed, it is the large $N_F$ QCD that models the walking technicolor where the large number of fermions give the screening effects and level off of
the infrared coupling which otherwise blows up due to the gluon anti-screening effects (Caswell-Banks-Zaks infrared fixed point).
For large $N_F$ with the fixed point value smaller than the critical coupling, the SSB phase is gone (what we called conformal phase transition).
Then the infrared scale invariance becomes manifest, dubbed the conformal window. The waking technicolor is close to just outside of the conformal
window. Although $N_f$ and $N_C$ are integers, the anti-Veneziano limit makes the analyses of the phase in the almost continuous parameter $r=N_F/N_C$.
To see the relevance of the infrared conformality, we have argued that the 125 GeV Higgs itself
is a (pseudo-)dilaton, with mass coming from the trace anomaly of dimension 2 due to the Lagrangian parameter,
even
if it is described by the Standard Model Higgs Lagrangian (!!). The SM Higgs Lagarangian
was in fact shown to be equivalent to the scale-invariant nonlinear sigma model with both chiral and scale symmetries
being nonlinearly realized. The SM Higgs Lagrangian was further shown to be gauge equivalent to the scale-invariant Hidden Local Symmetry (HLS)
Lagrangian which include new massive vector bosons as the gauge bosons of the (spontaneously broken) HLS,
with the mass term being scale-invariant.
All these features of the SM Higgs Lagrangian are reminiscent of the conformal UV completion behind the Higgs, the underlying theory with (approximate) scale symmetry with the coupling so strong as to produce composite states
such as the Higgs (dilaton), new vector bosons (HLS gauge bosons), etc.. We have seen that even the NJL model, though not gauge theory,
can be regarded as such a
conformal UV completion.
The walking technicolor, conformal SCGT, is a gauge theory version of such a typical
underlying theory, where the 125 GeV Higgs is a composite pseudo-dilaton, technidilaton, with mass coming from the nonperturbative trace anomaly.
The walking technicolor in the
anti-Veneziano limit $N_C \rightarrow \infty$ with $N_C \alpha=$fixed $={\cal O}(1)$ and $N_F/N_C=$ fixed ($\gg 1)$ makes
the ladder approximation reasonable, which yields a naturally light and weakly coupled technidilaton through the PCDC:
\begin{equation}
M_\phi^2 F_\phi^2=
-4 \langle \theta_\mu^\mu \rangle
=- \frac{\beta(\alpha (\mu^2))}{\alpha(\mu^2)}
\, \langle G_{\mu \nu}^2(\mu^2)\rangle
\simeq N_C N_F\frac{16 }{\pi^4} m_F^4\,,
\end{equation}
independently of the renormalization point $\mu$,
where the scale symmetry is explicitly broken by the nonperturbative trace anomaly of the dimension 4 operator $G_{\mu\nu}^2$,
which is induced by $m_F$ the
dynamical mass of the technifermion arising from the simultaneous spontaneous breaking of the scale symmetry and the chiral symmetry.
The technidilaton with mass 125 GeV and coupling by the PCDC relation for the one-family model with $N_F=8, N_c=4$ is consistent with
the present data for the LHC 125 GeV Higgs.
Lattice results are also encouraging for the light composite Higgs as the technidilaton in the walking technicolor, particularly for $N_F=8$ QCD,
which correspond to the one-family technicolor as the most straightforward walking technicolor model building.
We will see the fate of the strong coupling theories at LHC Run II, hoping that the Nambu's way will continue forever.
\section*{Acknowledgment}
Main materials given here were developed during the discussions and collaborations with Shinya Matsuzaki.
I also thank all the members of LatKMI Collaboration for excellent collaborations on the strong coupling gauge theories on the lattice.
|
\section{Introduction.}
The Oberwolfach Problem was first posed by Ringel in 1967 during a conference in Oberwolfach. The question was whether it was possible to seat the $v$ conference attendees at $n$ round tables for dinner during $\frac{v-1}{2}$ nights, in such a way that every attendee sits next to every other attendee exactly once. This is equivalent to asking whether the complete graph $K_v$ can be decomposed into $\frac{v-1}{2}$ copies of a $2$-factor $F$ (in a $2$-factor every component is a cycle, which represents a round table). To achieve this decomposition $v$ needs to be odd, because the vertices (attendees) need to have even degree. Later a version with $v$ even was studied. In this case, the attendees will never sit next to their spouses (and we are assuming that every attendee has a spouse). This is equivalent to asking for a decomposition of $K_v$ into $\frac{v-2}{2}$ copies of a $2$-factor $F$, and one copy of a $1$-factor (each attendee together with their spouse).
In \cite{Liu1} Liu first worked on the generalization of the Oberwolfach problem, where instead of avoiding their spouses, the attendees avoid all the other members of their delegation. The assumption was that all the delegations had the same number of people. Thus we are seeking to decompose the complete equipartite graph $K_{(m:n)}$ with $n$ partite sets (delegations) of size $m$ each (members of a delegations) into $\frac{(n-1)m}{2}$ copies of a $2$-factor $F$. Here $(n-1)m$ has to be even. In \cite{HH} Hoffman and Holliday worked on the equipartite generaliztion of the Oberwolfach problem when $(n-1)m$ is odd, decomposing into $\frac{(n-1)m-1}{2}$ copies of a $2$-factor $F$, and one copy of a $1$-factor.
The Hamilton-Waterloo problem is a generalization of the Oberwolfach problem, in which the conference is being held at two different cities. Because the table arrangements are different, we have two $2$-factors, $F_1$ and $F_2$. The Hamilton-Waterloo problem then asks whether the complete graph $K_v$ can be decomposed into $r$ copies of the $2$-factor $F_1$ (tables at Hamilton) and $s$ copies of the $2$-factor $F_2$ (tables at Waterloo), such that $s+r=\frac{v-1}{2}$, when $v$ is odd, and having $s+r=\frac{v-2}{2}$ and a $1$-factor when $v$ is even.
The uniform Oberwolfach problem (when all the tables have the same size, i.e. all the cycles of the $2$-factor have the same size) has been completely solved by Alspach and Haagkvist \cite{AH} and Alspach, Schellenberg, Stinson and Wagner \cite{ASSW}. For the non-uniform Oberwolfach problem many results have been obtained, for a survey of results up to 2006 see \cite{BR}. The uniform Oberwolfach problem over equipartite graphs has been completely solved by Liu \cite{Liu2} and Hoffman and Holliday \cite{HH}. In the non-uniform case Bryant, Danziger and Pettersson \cite{BDP} completely solved the case when the $2$-factor is bipartite. For the Hamilton-Waterloo problem most of the results are uniform, see for example \cite{AKKPO} or \cite{BDT}. In particular, Burgess, Danziger and Traetta \cite{BDT} proved the following theorem.
\begin{theorem}{\normalfont\cite{BDT}}\label{theoremBDT}
If $m$ and $n$ are odd integers with $n\geq m\geq 3$ and $t>1$, then there is a decomposition of $K_{mnt}$ into $s$ $C_{m}$-factors and $r$ $C_n$-factors if and only if $t$ is odd, $s,r\geq 0$ and $s+r=(mnt-1)/2$, except possibly when $r=1$ or $3$, or $(m,n,r)=(5,9,5),(5,9,7),(7,9,5),(7,9,7),(3,13,5)$.
\end{theorem}
Theorem \ref{theoremBDT} covers most of the odd ordered uniform cases. The authors in \cite{BDT} point out that it is possible to have solutions where the number of vertices is not a multimple of $mn$. Thus if $l=\lcm(m,n)$ and the number of vertices is a multiple of $l$, but not divisible by $mn$, then Theorem \ref{theoremBDT} cannot be applied. The constructions given in this paper can be applied to cover some of these cases.
There are some results in the non-uniform case, some examples are Bryant, Danzinger \cite{BD}, Bryant, Danzinger, Dean \cite{BDD} and Haggkvist \cite{H}.
The Hamilton-Waterloo problem can be generalized for complete equipartite graphs in the same way as the Oberwolfach problem was generalized, but not much work has been done in this direction. Asplund, Kamin, Keranen, Pastine and \"{O}zkan \cite{AKKPO} gave some constructions for complete equipartite graphs with $3$ parts. Burgess, Danziger and Traetta \cite{BDT} studied the case when the graph consists of $m$ partite sets of size $n$, and the cycle sizes are $m$ and $n$. In both papers the constructions were done in order to get a result on the Hamilton-Waterloo problem for complete graphs. The focus of this paper is to give a generalization of the Hamilton-Waterloo problem for complete equipartite graphs with an odd number of partite sets. We obtain results both in the uniform and non-uniform cases.
\section{Basic Definitions and Results}
Let $G$ be a multipartite graph with $k$ partite sets, $G_0,G_1,\ldots, G_{k-1}$. We identify each vertex $g$ of $G$ as an ordered pair $(g,i)$, where $g\in G_i$.
\begin{example}
Consider the graph in Figure \ref{identificationfigure}.
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle}]
\draw (0,0) node [scale=.7, minimum size=.8cm](1){a};
\draw (1,0) node [scale=.7, minimum size=.8cm](2){b};
\draw (2,0) node [scale=.7, minimum size=.8cm](3){c};
\draw (0,-1) node[scale=.7, minimum size=.8cm] (4){d};
\draw (0,-2) node [scale=.7, minimum size=.8cm](5){f};
\draw (1,-1) node [scale=.7, minimum size=.8cm](6){e};
\draw (1) to (2);
\draw (2) to (4);
\draw (4) to (3);
\draw (3) to [bend left=30](5);
\draw (5) to (6);
\draw (6) to (1);
\end{tikzpicture}
\caption{}
\label{identificationfigure}
\end{center}
\end{figure}
If we consider each column as a partite set, we have $3$ partite sets; $G_0$, with vertices $a,d$ and $f$; $G_1$ with vertices $b$ and $e$; and $G_2$ with $c$ as its only vertex. Then the vertices are $(a,0)$, $(b,1)$, $(c,2)$, $(d,0)$, $(e,1)$, $(f,0)$.
When it is convenient, we will just denote the vertices $(0,0)$, $(0,1)$, $(0,2)$, $(1,0)$, $(1,1)$, $(2,0)$, as in Figure \ref{identificationfigure2}.
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle}]
\draw (0,0) node [scale=.7](1){0,0};
\draw (1,0) node [scale=.7](2){0,1};
\draw (2,0) node [scale=.7](3){0,2};
\draw (0,-1) node[scale=.7] (4){1,0};
\draw (0,-2) node [scale=.7](5){2,0};
\draw (1,-1) node [scale=.7](6){1,1};
\draw (1) to (2);
\draw (2) to (4);
\draw (4) to (3);
\draw (3) to [bend left=30](5);
\draw (5) to (6);
\draw (6) to (1);
\end{tikzpicture}
\caption{}
\label{identificationfigure2}
\end{center}
\end{figure}
\end{example}
\begin{definition}\label{def1}
Let $G$ and $H$ be multipartite graphs. Then we define the \textit{partite product} of $G$ and $H$, $G\otimes H$ as follows:
\begin{itemize}
\item $V(G\otimes H)=\{(g,h,i)| (g,i)\in V(G)$ and $(h,i)\in V(H)\}$.
\item $E(G\otimes H)=\{\{(g_1,h_1,i),(g_2,h_2,j)\}|\{(g_1,i),(g_2,j)\}\in E(G)$ and $\{(h_1,i),(h_2,j)\}\in E(H)\}$.
\end{itemize}
\end{definition}
Notice that this definition is quite similar to that of the direct product. The main difference is that we are doing this product ``just in $1$ coordinate''. To see that they are different it suffices to count the number of vertices in the product. If $G=H=K_{3,3,3}$ then $|V(G\times H)|=81$ but $|V(G\otimes H)|=27$.
Indeed, if the $k$ partite sets of $G$ and $H$ have sizes $g_0,g_1,\ldots,g_{k-1}$ and $h_0,h_1,\ldots, h_{k-1}$, respectively, then $|V(G\otimes H)|=g_0h_0+g_1h_1+\dots+g_{k-1}h_{k-1}$, whereas $|V(G\times H)|=\left(\sum g_i\right)\left(\sum h_i\right)$.
\begin{remark}
The partite product depends on the multipartite representation chosen for a graph. For example, the graphs $G$ and $H$ in Figure \ref{isomorphicgraphs} are isomorphic, but they behave differently in the product (where we understand that each column is a part of the multipartite graph).
\end{remark}
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle}]
\draw (0,0) node [scale=.5](1){$0,0$};
\draw (1,0) node [scale=.5](2){$0,1$};
\draw (2,0) node [scale=.5](3){$0,2$};
\draw (0,-1) node[scale=.5] (4){$1,0$};
\draw (0,-2) node [scale=.5](5){$2,0$};
\draw (1,-1) node [scale=.5](6){$1,1$};
\draw (1) to (2);
\draw (2) to (4);
\draw (4) to (3);
\draw (3) to [bend left=30](5);
\draw (5) to (6);
\draw (6) to (1);
\draw (4,0) node [scale=.5](11){$0,0$};
\draw (5,0) node [scale=.5](22){$0,1$};
\draw (6,0) node [scale=.5](33){$0,2$};
\draw (4,-1) node[scale=.5] (44){$1,0$};
\draw (5,-1) node [scale=.5](55){$1,1$};
\draw (6,-1) node [scale=.5](66){$1,2$};
\draw (11) to (22);
\draw (22) to (33);
\draw (44) to (33);
\draw (44) to (55);
\draw (55) to (66);
\draw (66) to (11);
\node[left=10pt,fill=none,draw=none] at (0,-.75) {$G=$};
\node[left=10pt,fill=none,draw=none] at (4,-.75) {$H=$};
\draw (-4,-4) node [scale=.5,inner sep=1pt](000){$0,0,0$};
\draw (-2,-4) node [scale=.5,inner sep=1pt](001){$0,0,1$};
\draw (0,-4) node [scale=.5,inner sep=1pt](002){$0,0,2$};
\draw (-4,-5) node [scale=.5,inner sep=1pt](010){$0,1,0$};
\draw (-2,-5) node [scale=.5,inner sep=1pt](011){$0,1,1$};
\draw (-4,-6) node [scale=.5,inner sep=1pt](020){$0,2,0$};
\draw (-4,-7) node [scale=.5,inner sep=1pt](100){$1,0,0$};
\draw (-2,-7) node [scale=.5,inner sep=1pt](101){$1,0,1$};
\draw (-4,-8) node [scale=.5,inner sep=1pt](110){$1,1,0$};
\draw (-2,-8) node [scale=.5,inner sep=1pt](111){$1,1,1$};
\draw (-4,-9) node [scale=.5,inner sep=1pt](120){$1,2,0$};
\draw (-4,-10) node [scale=.5,inner sep=1pt](200){$2,0,0$};
\draw (-4,-11) node [scale=.5,inner sep=1pt](210){$2,1,0$};
\draw (-4,-12) node [scale=.5,inner sep=1pt](220){$2,2,0$};
\draw (000) to (001);
\draw (000) to (011);
\draw (000) to (101);
\draw (000) to (111);
\draw (100) to (001);
\draw (100) to (011);
\draw (010) to (001);
\draw (010) to (101);
\draw (110) to (002);
\draw (220) [bend right=15] to (002);
\draw (020) to (111);
\draw (020) to (011);
\draw (110) to (001);
\draw (120) to (011);
\draw (200) to (101);
\draw (200) to (111);
\draw (210) to (101);
\draw (220) to (111);
\node[left=10pt,fill=none,draw=none] at (-4,-4.75) {$G\otimes G=$};
\draw (6,-4) node [scale=.5,inner sep=1pt](a000){$0,0,0$};
\draw (8,-4) node [scale=.5,inner sep=1pt](a001){$0,0,1$};
\draw (10,-4) node [scale=.5,inner sep=1pt](a002){$0,0,2$};
\draw (6,-5) node [scale=.5,inner sep=1pt](a010){$0,1,0$};
\draw (8,-5) node [scale=.5,inner sep=1pt](a011){$0,1,1$};
\draw (10,-5) node [scale=.5,inner sep=1pt](a012){$0,1,2$};
\draw (6,-6) node [scale=.5,inner sep=1pt](a100){$1,0,0$};
\draw (8,-6) node [scale=.5,inner sep=1pt](a101){$1,0,1$};
\draw (10,-6) node [scale=.5,inner sep=1pt](a102){$1,0,2$};
\draw (6,-7) node [scale=.5,inner sep=1pt](a110){$1,1,0$};
\draw (8,-7) node [scale=.5,inner sep=1pt](a111){$1,1,1$};
\draw (10,-7) node [scale=.5,inner sep=1pt](a112){$1,1,2$};
\draw (a000) to (a001);
\draw (a000) to (a112);
\draw (a010) to (a011);
\draw (a010) to (a102);
\draw (a100) to (a101);
\draw (a100) to (a012);
\draw (a110) to (a111);
\draw (a110) to (a002);
\draw (a001) to (a002);
\draw (a011) to (a012);
\draw (a101) to (a102);
\draw (a111) to (a112);
\node[left=10pt,fill=none,draw=none] at (6,-4.75) {$H\otimes H=$};
\end{tikzpicture}
\caption{}
\label{isomorphicgraphs}
\end{center}
\end{figure}
The next result follows directly from Definition \ref{def1}
\begin{lemma}\label{commutative}
The product is commutative, that is, $G\otimes H=H\otimes G$.
\end{lemma}
Most of our results will be concerning complete multipartite graphs. We will denote by $K_{(n:m)}$ the complete multipartite graph with $m$ parts, each of size $n$.
\begin{lemma}\label{k1kidentity}
If $G$ is $k$-partite, then $G\otimes K_{(1:k)}$ is isomorphic to $G$.
\end{lemma}
\begin{proof}
Here each part of $K_{(1:k)}$ has just 1 vertex and so $|V(G)|=|V\left(G\otimes K_{(1:k)})\right |$. Because all the vertices of $K_{(1:k)}$ are neighbors, two vertices $(g_1,k_1,i), (g_2,k_2,j)$ of $G\otimes K_{(1:k)}$ are neighbors if and only if $(g_1,i)$ and $(g_2,j)$ are neighbors. Therefore $G\otimes K_{(1:k)}$ is isomorphic to $G$.
\end{proof}
\begin{definition}
The \textit{complete cyclic multipartite graph} $C_{(x:k)}$ is the graph with $k$ parts of size $x$, where two vertices $(g,i)$ and $(h,j)$ are neighbors if and only if $i-j=\pm 1\pmod{k}$, with this subtraction being done modulo $k$. The \textit{directed complete cyclic multipartite graph} $\overrightarrow{C}_{(x:k)}$ is the graph with $k$ parts of size $x$, with arcs of the form $\big((g,i),(h,i+1)\big)$ for every $0\leq g,h\leq x-1$, $0\leq i\leq k-1$.
\end{definition}
It should be noted that any decomposition of $\overrightarrow{C}_{(x:k)}$ gives a decomposition of $C_{(x:k)}$. Notice that $C_{(1:k)}$ is the cycle with $k$ vertices and $C_{(x:3)}$ is isomorphic to $K_{(x:3)}$.
The next three results are easy to see, so the proofs are left to the reader.
\begin{lemma}
Let $G$ and $H$ be $k$-partite graphs. If each part of $G$ has $\frac{|V(G)|}{k}$ vertices and each part of $H$ has $\frac{|V(H)|}{k}$ vertices, then:
\begin{itemize}
\item Each part of $G\otimes H$ has $\frac{|V(G)|\times |V(H)|}{k^2}$ vertices.
\item $|V\left(G\otimes H\right)|=\frac{|V(G)|\times |V(H)|}{k}$.
\end{itemize}
\end{lemma}
\begin{lemma}
\begin{itemize}
\item $K_{(x:k)}\otimes K_{(y:k)}$ is isomorphic to $K_{(xy:k)}$.
\item $C_{(x:k)}\otimes C_{(y:k)}$ is isomorphic to $C_{(xy:k)}$.
\item $\overrightarrow{C}_{(x:k)}\otimes \overrightarrow{C}_{(y:k)}$ is isomorphic to $\overrightarrow{C}_{(xy:k)}$.
\end{itemize}
\end{lemma}
\begin{lemma}
The complete cyclic multipartite graph is the product of the complete multipartite graph by the cycle. This is: $K_{(x:k)}\otimes C_{(1:k)}=C_{(x:k)}$.
\end{lemma}
\section{Product and Decompositions.}
We can consider a decomposition of a graph as a partition of the edge set or as a union of edge disjoint subgraphs. This means that a decomposition of $G$ into $H_1,\ldots,H_s$ is given by $E(G)=\cup E(H_i)$ or by $G=\oplus H_i$. We will think of $\oplus$ as a boolean sum, which means that $H_i\oplus H_i=\emptyset$.
We have the following easy result.
\begin{theorem}[Distribution]\label{distributive}
Let $G=\oplus_i G_i$ and $H=\oplus_j H_j$ be $k$-partite graphs. Then $G\otimes H=\left(\oplus_i G_i\right)\otimes \left(\oplus_j H_j\right)$. Furthermore, the following distributive property holds:
\[
\left(\oplus_i G_i\right)\otimes \left(\oplus_j H_j\right)=\oplus_i\left(G_i\otimes \oplus_j H_j\right)=\oplus_i\oplus_j\left( G_i\otimes H_j\right)
\]
\end{theorem}
\begin{proof}
It is enough to prove that
\[
G\otimes \left(H_1 \oplus H_2\right)=\left(G\otimes H_1\right)\oplus\left(G\otimes H_2\right),\]
where $E(H_1)\cap E(H_2)=\emptyset$.
Let
\[
\{(g_1,h_1,i),(g_2,h_2,j)\}\in E\left(G\otimes \left(H_1 \oplus H_2\right)\right).
\]
This means that
\[
\{(h_1,i),(h_2,j)\}\in E(H_1)\cup E(H_2).\]
But since $E(H_1)\cap E(H_2)=\emptyset$, without loss of generality we may assume
\[
\{(h_1,i),(h_2,j)\}\in E(H_1),\]
and so
\[
\{(g_1,h_1,i),(g_2,h_2,j)\}\in E\left(G\otimes H_1\right)\subset E\left(\left(G\otimes H_1\right)\oplus\left(G\otimes H_2\right)\right).\]
Hence
\[
E\left(G\otimes \left(H_1 \oplus H_2\right)\right)\subset E\left(\left(G\otimes H_1\right)\oplus\left(G\otimes H_2\right)\right).
\]
Let now
\[
\{(g_1,h_1,i),(g_2,h_2,j)\}\in E\left(\left(G\otimes H_1\right)\oplus\left(G\otimes H_2\right)\right).\]
This means that
\[
\{(g_1,h_1,i),(g_2,h_2,j)\}\in E\left(\left(G\otimes H_1\right)\right) \cup E\left(\left(G\otimes H_2\right)\right).\]
Without loss of generality we may assume
\[
\{(g_1,h_1,i),(g_2,h_2,j)\}\in E\left(\left(G\otimes H_1\right)\right),
\]
which gives
\[
\{(h_1,i),(h_2,j)\}\in E(H_1)\subset E(H_1\oplus H_2),\text{ and}
\]
\[
\{(g_1,h_1,i),(g_2,h_2,j)\}\in E\left(G\otimes \left(H_1\oplus H_2\right)\right).
\]
Hence
\[
E\left(\left(G\otimes H_1\right)\oplus\left(G\otimes H_2\right)\right)\subset E\left(G\otimes \left(H_1 \oplus H_2\right)\right).\]
Therefore
\[
G\otimes \left(H_1 \oplus H_2\right)=\left(G\otimes H_1\right)\oplus\left(G\otimes H_2\right),
\]
and by induction we get that the product and additions in
\[
G\otimes H=\left(\oplus_i G_i\right)\otimes \left(\oplus_j H_j\right)\]
are distributive.
\end{proof}
\begin{corollary}
Let $G$ and $H$ be multipartite graphs with $k$ partite sets.
\begin{itemize}
\item If $G$ can be decomposed into isomorphic copies of $\Gamma$ and $H$ can be decomposed into isomorphic copies of $K_{(1:k)}=K_k$, then $G\otimes H$ can be decomposed into isomorphic copies of $\Gamma$.
\item If $G$ can be factored into isomorphic copies of $\Gamma$ and $H$ can be factored into unions of copies of $K_{(1:k)}=K_k$, then $G\otimes H$ can be factored into unions of copies of $\Gamma$.
\end{itemize}
\end{corollary}
\begin{proof}
If $G$ is decomposed into copies of $\Gamma$, it means that $G=\oplus G_i$, where each $G_i$ is isomorphic to $\Gamma$. If $H$ is decomposed into copies of $K_{(1:k)}$ (or union of them), it means that $H=\oplus H_i$, where each $H_i$ is isomorphic to $K_{(1:k)}$ (or union of them).
By the Distribution Theorem we only need to show that $G_i\otimes H_i$ is isomorphic to $\Gamma$ or to $G_i$. But by Lemma \ref{k1kidentity} we know this is true.
\end{proof}
It is interesting to notice that the set of $k$-partite graphs, with $\oplus$ as a sum and $\otimes $ as a product form a commutative ring. The empty graph is the $0$ element, and $K_{(1:k)}$ is the $1$ element. All of the elements are additive involutions, Theorem \ref{distributive} gives us the distribution laws, and Lemma \ref{commutative} shows that the product is commutative.
\section{Product of cycles.}
In this section we will concern ourselves with the product of two or more cycles.
Since our product depends on what kind of partition we are using, we need to ask something more from our cycles in order to get results.
\begin{definition}
Given a graph $G$ we will say that $C$ is a \textit{$C_n$-factor of $G$} if $C$ is a $2$-factor of $G$ where each connected component is of size $n$. This means that $C$ is a spanning subgraph of $G$ and $C$ is a union of disjoint cycles of size $n$.
When it is understood that the graph is $G$, then we will just call $C$ a $C_n$-factor (instead of a $C_n$-factor of $G$).
Similarly given a directed graph $\overrightarrow{G}$ we will say that $\overrightarrow{C}$ is a \textit{$\overrightarrow{C}_n$-factor of $G$} if $\overrightarrow{C}$ is a $2$-factor of $\overrightarrow{G}$ where each connected component is a directed cycle of size $n$.
When it is understood that the graph is $\overrightarrow{G}$, the we will just call $\overrightarrow{C}$ a $\overrightarrow{C}_n$-factor.
\end{definition}
The following lemmas give us an idea of how directed cycles work under the product. They also illustrate why we introduce $\overrightarrow{C}_{(x:k)}$ instead of just working with $C_{(x:k)}$.
\begin{lemma}\label{productdirectedcycles}
Let $\overrightarrow{C}$ be a directed cycle of length $n$ of $\overrightarrow{C}_{(x:k)}$, and let $\overrightarrow{C}'$ be a directed cycle of length $m$ of $\overrightarrow{C}_{(y:k)}$. Then $\overrightarrow{C}\otimes \overrightarrow{C}'$ is a set of $\frac{\gcd(n,m)}{k}$ disjoint directed cycles of $\overrightarrow{C}_{(xy:k)}$ of length $l=\frac{nm}{\gcd(n,m)}$ and $xyk-\frac{nm}{k}$ isolated vertices.
\end{lemma}
\begin{proof}
Notice that $\overrightarrow{C}$ has $xk-n$ isolated vertices, because it is a subgraph of $\overrightarrow{C}_{(x:k)}$. Likewise, $\overrightarrow{C}'$ has $yk-m$ isolated vertices.
Let $(x_0,y_0,i)$ be a vertex in $\overrightarrow{C}\otimes \overrightarrow{C}'$. If either $(x_0,i)$ or $(y_0,i)$ are isolated vertices, then $(x_0,y_0,i)$ is isolated. If neither $(x_0,i)$ and $(y_0,i)$ are isolated, then they respectively have an arrow coming from $(x_1,i-1)$ and $(y_1,i-1)$; and an arrow going to $(x_2,i+1)$ and $(y_2,i+1)$, for some $0\leq x_1,x_2\leq x-1$, $0\leq y_1,y_2\leq y-1$. Hence $(x_0,y_0,i)$ has an arrow coming from $(x_1,y_1,i-1)$ and an arrow going to $(x_2,y_2,i+1)$. So $\overrightarrow{C}\otimes \overrightarrow{C}'$ is composed of directed cycles and isolated vertices.
Assume without loss of generality $n\leq m$.
Let $i_0,i_1,\ldots, i_{n-1}$ be the non-isolated vertices of $\overrightarrow{C}$, in the order they appear, with $i_0\in \overrightarrow{C}_0$; and let $j_0,j_1,\ldots,j_{m-1}$ be the non-isolated vertices of $\overrightarrow{C}'$, in the order they appear with $j_0\in \overrightarrow{C}'_0$. Then the directed cycle starting at $(i_0,j_0)$ in $\overrightarrow{C}\otimes \overrightarrow{C}'$ consists of the vertices:
\[
(i_0,j_0),(i_1,j_1),\ldots, (i_{n-1},j_{n-1}), (i_0,j_{n}),\ldots (i_{n-1},i_{m-1})
\]
This directed cycle has length $l=\frac{nm}{gcd(n,m)}$.
Notice that $\overrightarrow{C}\otimes \overrightarrow{C}'$ has $\frac{nm}{k}$ non-isolated vertices, which means that the number of directed cycles is $\frac{nm}{k}\frac{gcd(n,m)}{nm}=\frac{\gcd(n,m)}{k}$. Thus
$\overrightarrow{C}\otimes \overrightarrow{C}'$ is a set of $\frac{\gcd(n,m)}{k}$ disjoint directed cycles of $\overrightarrow{C}_{(xy:k)}$ of length $l=\frac{nm}{\gcd(n,m)}$ and $xyk-\frac{nm}{k}$ isolated vertices.
\end{proof}
\begin{lemma}\label{productofbalanced}
Let $\overrightarrow{G}$ and $\overrightarrow{H}$ be a $\overrightarrow{C}_n$-factor and a $\overrightarrow{C}_m$-factor of $\overrightarrow{C}_{(x:k)}$ and $\overrightarrow{C}_{(y:k)}$, respectively. Then $\overrightarrow{G}\otimes \overrightarrow{H}$ is a $\overrightarrow{C}_{l}$-factor of $\overrightarrow{C}_{(xy:k)}$, where $l=\frac{nm}{gcd(n,m)}$.
\end{lemma}
\begin{proof}
Notice that neither $\overrightarrow{G}$ nor $\overrightarrow{H}$ have isolated vertices.
Let $(x_0,y_0,i)$ be a vertex in $\overrightarrow{G}\otimes \overrightarrow{H}$. We know that $(x_0,i)$ has an arrow coming from $(x_1,i-1)$ and an arrow going to $(x_2,i+1)$, for exactly one pair $0\leq x_1,x_2\leq x-1$, because $\overrightarrow{G}$ is a $\overrightarrow{C}_n$-factor. Likewise $(y_0,i)$ has has an arrow coming from $(y_1,i-1)$ and an arrow going to $(y_2,i+1)$, for exactly one pair $0\leq y_1,y_2\leq y-1$. Hence $(x_0,y_0,i)$ has an arrow coming from $(x_1,y_1,i-1)$ and an arrow going to $(x_2,y_2,i+1)$, so each vertex in $\overrightarrow{G}\otimes \overrightarrow{H}$ is in exactly one directed cycle.
Let $\overrightarrow{G}=\oplus_i \overrightarrow{C}(i)$ and $\overrightarrow{H}=\oplus_j \overrightarrow{C}'(j)$, where each $\overrightarrow{C}(i)$ is a directed cycle of length $n$, and each $\overrightarrow{C}'(j)$ is a directed cycle of length $m$.
Then by Theorem \ref{distributive} we get:
\begin{align*}
\overrightarrow{G}\otimes \overrightarrow{H}&=\left(\oplus_i \overrightarrow{C}(i)\right)\otimes\left(\oplus_j \overrightarrow{C}'(j)\right)\\
&=\oplus_i\oplus_j\left(\overrightarrow{C}(i)\otimes \overrightarrow{C}'(j)\right)\\
\end{align*}
But by Lemma \ref{productdirectedcycles} we know that $\overrightarrow{C}(i)\otimes \overrightarrow{C}'(j)$ is composed of $\frac{\gcd(n,m)}{k}$ directed cycles of length $l=\frac{nm}{gcd(n,m)}$.
Hence each directed cycle in $\overrightarrow{G}\otimes \overrightarrow{H}$ has size $l=\frac{nm}{gcd(n,m)}$ and $\overrightarrow{G}\otimes \overrightarrow{H}$ is a $\overrightarrow{C}_{l}$-factor.
\end{proof}
\begin{definition}
Given a graph $G$ we will say that $F$ is a $[n_1^{e_1},n_2^{e_2},\ldots,n_p^{e_p}]$-factor of $G$ if $F$ is a $2$-factor of $G$ with $e_i$ connected components of size $n_i$, $i=1,2,\ldots, p$.
\end{definition}
If $e_i$ is not listed, we will just assume that it is $1$. Also, we allow $n_i=n_j$, so that the number of cycles of a certain size is just the sum of the exponents of that number in the expression $[n_1^{e_1},\ldots,n_p^{e_p}]$.
\begin{example}
A $[3^2,3^3,5^2,11,13]$-factor is a subgraph of a graph on $59$ vertices, consisting of $5$ cycles of size $3$, $2$ cycles of size $5$, $1$ cycle of size $11$ and $1$ cycle of size $13$. This subgraph can also be written as a $[3,3,3,3,3,5,5,11,13]$-factor or as a $[3^5,5^2,11,13]$-factor.
\end{example}
Notice that a $C_n$-factor of a graph on $m$ vertices is a $[n^{\frac{m}{n}}]$-factor.
\section{From $C_{(v:n)}$ to $K_{(v:m)}$}
It is advantageous to find solutions to the the Hamilton-Waterloo problem on complete multipartite graphs because they can then be used to obtain solutions to the Hamilton-Waterloo problem on complete graphs:
\begin{lemma}\label{buildcompletegraph}
Let $m$, $n$, $x$, $y$ and $v$ be positive integers. Suppose the following conditions are satisfied:
\begin{itemize}
\item There exists a decomposition of $K_v$ into $s_\alpha$ $C_{xn}$-factors and $r_\alpha$ $C_{yn}$-factors.
\item There exists a decomposition of $K_{(v:m)}$ into $s_\beta$ $C_{xn}$-factors and $r_\beta$ $C_{yn}$-factors.
\end{itemize}
Then there exists a decomposition of $K_{vm}$ into $s=s_\alpha+s_\beta$ $C_{xn}$-factors and $r=r_\alpha+r_\beta$ $C_{yn}$-factors.
\end{lemma}
\begin{proof}
Partition the vertices of $K_{vm}$ into $m$ sets $A_1,\ldots, A_m$ of size $v$ each.
The graph that contains the edges between vertices belonging to a same partite set is the union of $m$ disjoint copies of $K_v$. We can decompose each copy of $K_v$ into $s_\alpha$ $C_{xn}$-factors and $r_\alpha$ $C_{yn}$-factors.
The graph that contains the edges between vertices belonging to different parts is isomorphic to $K_{(v:m)}$. We can decompose this graph into $s_\beta$ $C_{xn}$-factors and $r_\beta$ $C_{yn}$-factors.
Thus we have a decomposition of $K_{vm}$ into $s=s_\alpha+s_\beta$ $C_{xn}$-factors and $r=r_\alpha+r_\beta$ $C_{yn}$-factors.
\end{proof}
One could write a version of the lemma for non-uniform solutions as follows, where by $mK_v$ we understand the graph consisting of $m$ disconnected copies of $K_v$.
\begin{lemma}\label{buildcompletegraphnonuniform}
Let $m$, and $v$ be positive integers. Let $F_1$ and $F_2$ be two $2$-factors on $vm$ vertices. Suppose the following conditions are satisfied:
\begin{itemize}
\item There exists a decomposition of $mK_v$ into $s_\alpha$ copies of $F_1$ and $r_\alpha$ copies of $F_2$.
\item There exists a decomposition of $K_{(v:m)}$ into $s_\beta$ copies of $F_1$ and $r_\beta$ copies of $F_2$.
\end{itemize}
Then there exists a decomposition of $K_{vm}$ into $s=s_\alpha+s_\beta$ copies of $F_1$ and $r=r_\alpha+r_\beta$ copies of $F_2$.
\end{lemma}
In order to use these lemmas we need two types of ingredients. The first one is the decomposition of the complete graph $K_v$. In \cite{AH,ASSW} uniform decompositions were given:
\begin{theorem}
\label{OP}
\cite{AH,ASSW}
There exists a decomposition of $K_v$ into $C_n$-factors if and
only if $v \equiv 0 \pmod{n}$, $(v,n) \not = (6,3)$ and $(v,n) \not = (12,3)$.
\end{theorem}
The other ingredient is the decomposition of the complete multipartite graph $K_{(v:m)}$. In \cite{BDT} the authors used decompositions of $C_{(v:m)}$ to obtain decompositions of $K_{(v:m)}$. We describe this type of construction in a formal fashion in the following lemma.
\begin{lemma}\label{cvntokvm}
Let $m$, $x_1,\ldots,x_p$, $y_1,\ldots,y_p$, and $v$ be positive integers. Let $s_1,\ldots,s_{\frac{m-1}{2}}$ be non-negative integers. Suppose the following conditions are satisfied:
\begin{itemize}
\item There exists a decomposition of $K_m$ into $[n_1,\ldots,n_p]$-factors.
\item For every $1\leq i\leq p$, and for every $1\leq t \leq \frac{m-1}{2}$ there exists a decomposition of $C_{(v:n_i)}$ into $s_t$ $C_{x_in_i}$-factors and $r_t$ $C_{y_in_i}$-factors.
\end{itemize}
Let
\[
s=\sum_{t=1}^{\frac{(m-1)}{2}}s_t\text{\hspace{.3in} and \hspace{.3in}}r=\sum_{t=1}^{\frac{(m-1)}{2}}r_t
\]
Then there exists a decomposition of $K_{(v:m)}$ into $s$ $[(x_1n_1)^{\frac{v}{x_1}},\ldots,(x_pn_p)^{\frac{v}{x_p}}]$-factors and $r$ $[(y_1n_1)^{\frac{v}{y_1}},\ldots,(y_pn_p)^{\frac{v}{y_p}}]$-factors.
\end{lemma}
\begin{proof}
Using the decomposition of $K_m$ into $[n_1,n_2,\ldots,n_p]$-factors, we have:
\[
K_{(1:m)}=K_m=\bigoplus_{t=1}^{\frac{m-1}{2}}N_t
\]
where each $N_t$ is a $[n_1,n_2,\ldots,n_p]$-factor.
We know that $K_{(v:m)}=K_{(v:m)}\otimes K_{(1:m)}$. This gives:
\[
\begin{array}{ccl}
K_{(v:m)}&=&K_{(v:m)}\otimes K_{(1:m)}\\
&=&K_{(v:m)}\otimes \left(\bigoplus_{t=1}^{\frac{m-1}{2}}N_t\right)\\
&=&\bigoplus_{t=1}^{\frac{m-1}{2}}\left(K_{(v:m)}\otimes N_t\right)\\
\end{array}
\]
Notice that $K_{(v:m)}\otimes N_t$ is a spanning subgraph of $K_{(v:m)}$. Since $N_t$ is a $[n_1,n_2,\ldots,n_p]$-factor, we have $N_t=\bigoplus_{i=1}^p C(w_i)$, where $C(w_i)$ is a cycle of size $n_i$.
Therefore,
\[
\begin{array}{ccl}
K_{(v:m)}\otimes N_t&=K_{(v:m)}\otimes \left(\bigoplus_{i=1}^p C(w_i)\right)\\
&=\bigoplus_{i=1}^p \left(K_{(v:m)}\otimes C(w_i)\right)\\
\end{array}
\]
But $K{(v:m)}\otimes C(w_i)$ is isomorphic to $C_{(v:n_i)}$ because $C(w_i)$ is isomorphic to $C_{(1:n_i)}$. So for each $w_i$ we can decompose $K{(v:m)}\otimes C(w_i)$ into $s_t$ $C_{x_in_i}$-factors and $r_t$ $C_{y_in_i}$-factors. Since $C_{(v:n_i)}$ has $vn_i$ vertices, a $C_{x_in_i}$-factor has $\frac{v}{x_i}$ cycles, hence it is a $[(x_in_i)^{\frac{v}{x_i}}]$-factor. Likewise a $C_{y_in_i}$-factor is a $[(y_in_i)^{\frac{v}{y_i}}]$-factor.
Taking the union of one $C_{x_in_i}$-factor for each $i$ gives a $[(x_1n_1)^{\frac{v}{x_1}},\ldots,(x_pn_p)^{\frac{v}{x_p}}]$-factor. Thus we get a decomposition of $K{(v:m)}\otimes N_t$ into $s_t$ $[(x_1n_1)^{\frac{v}{x_1}},\ldots,(x_pn_p)^{\frac{v}{x_p}}]$-factors and $r_t$ $[(y_1n_1)^{\frac{v}{y_1}},\ldots,(y_pn_p)^{\frac{v}{y_p}}]$-factors.
Doing this for every $1\leq t \leq \frac{m-1}{2}$, we end up with a decomposition of $K_{(v:m)}$ into $s$ $[(x_1n_1)^{\frac{v}{x_1}},\ldots,(x_pn_p)^{\frac{v}{x_p}}]$-factors and $r$ $[(y_1n_1)^{\frac{v}{y_1}},\ldots,(y_pn_p)^{\frac{v}{y_p}}]$-factors.
\end{proof}
It is important to note that Theorem \ref{OP} can be used to obtain decompositions of $K_m$ into $C_n$-factors. Thus the focus of the next three sections is to find decompositions of $C_{(v:n)}$. Any decomposition of $\overrightarrow{C}_{(v:n)}$ is equivalent to a decomposition of $C_{(v:n)}$, by simply removing the direction of each edge. We will work with the partite product on directed graphs to find decompositions of $\overrightarrow{C}_{(v:n)}$, and thus obtain decompositions of $C_{(v:n)}$.
\section{Hamilton Waterloo problem on directed complete cyclic multipartite graphs.}
For the entirety of this section, we assume $x$ is odd. In this section we will decompose $\overrightarrow{C}_{(x:n)}$ into $\overrightarrow{C}_n$-factors and $\overrightarrow{C}_{xn}$-factors (Hamilton Cycles), and $\overrightarrow{C}_{(4x:n)}$ into $\overrightarrow{C}_n$-factors and $\overrightarrow{C}_{2xn}$-factors.
Suppose $G_{\alpha}$ and $G_{\beta}$ are two parts of size $x$ in an equipartite directed graph $G$. We say an arc in $G$ has difference $d$ if $\left((g_1,\alpha),(g_2,\beta)\right)\in E(G)$ and $g_2-g_1\equiv d \pmod{x}$. If $\left\lbrace\left((g_1,\alpha),(g_2,\beta)\right):g_2-g_1\equiv d \pmod x \right\rbrace \subset E(G)$, then we say that difference $d$ between parts $\alpha$ and $\beta$ is covered by $G$.
Let the partite sets of $\overrightarrow{C}_{(x:n)}$ be $G_0,G_1,\ldots,G_{n-1}$. Write $n-1=2^{e_1}+2^{e_2}+\ldots+2^{e_k}$ with $e_{\alpha} > e_{\beta}$ if $\alpha < \beta$. Notice that $k\leq \frac{n}{2}$. Working modulo $x$, let $T_x(i)$ be the directed subgraph of $\overrightarrow{C}_{(x:n)}$ obtained by taking differences:
\begin{itemize}
\item $2^{e_j}i$ between $G_{j-1}$ and $G_{j}$ for $1\leq j \leq k$.
\item $-2i$ between $G_{j-1}$ and $G_{j}$ for $k+1\leq j \leq 2k-1$.
\item $-i$ between $G_{j-1}$ and $G_{j}$ for $2k\leq j \leq n-1$.
\item $-i$ between $G_{n-1}$ and $G_{0}$.
\end{itemize}
\begin{example}
We construct $T_5(1)$ with $n=7$. We have $n-1=6=2^2+2^1$, and $k=2$.
This means that from the first column to the second one we add $2^2$, from the second to the third we add $2$; since $k=2$, from the third to the forth we subtract $2$, and for the rest of the arcs we just subtract $1$. Figure \ref{t51figure} shows one directed cycle of this construction.
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle}]
\draw (0,0) node [scale=.5] (000){0,0};
\draw (1,0) node [scale=.5] (010){0,1};
\draw (2,0) node [scale=.5] (020){0,2};
\draw (3,0) node [scale=.5] (030){0,3};
\draw (4,0) node [scale=.5] (040){0,4};
\draw (5,0) node [scale=.5] (050){0,5};
\draw (6,0) node [scale=.5] (060){0,6};
\draw (0,-1) node [scale=.5](001){1,0};
\draw (1,-1) node [scale=.5](011){1,1};
\draw (2,-1) node [scale=.5](021){1,2};
\draw (3,-1) node [scale=.5](031){1,3};
\draw (4,-1) node [scale=.5](041){1,4};
\draw (5,-1) node [scale=.5](051){1,5};
\draw (6,-1) node [scale=.5](061){1,6};
\draw (0,-2) node [scale=.5](002){2,0};
\draw (1,-2) node [scale=.5](012){2,1};
\draw (2,-2) node [scale=.5](022){2,2};
\draw (3,-2) node [scale=.5](032){2,3};
\draw (4,-2) node [scale=.5](042){2,4};
\draw (5,-2) node [scale=.5](052){2,5};
\draw (6,-2) node [scale=.5](062){2,6};
\draw (0,-3) node [scale=.5](003){3,0};
\draw (1,-3) node [scale=.5](013){3,1};
\draw (2,-3) node [scale=.5](023){3,2};
\draw (3,-3) node [scale=.5](033){3,3};
\draw (4,-3) node [scale=.5](043){3,4};
\draw (5,-3) node [scale=.5](053){3,5};
\draw (6,-3) node [scale=.5](063){3,6};
\draw (0,-4) node [scale=.5](004){4,0};
\draw (1,-4) node [scale=.5](014){4,1};
\draw (2,-4) node [scale=.5](024){4,2};
\draw (3,-4) node [scale=.5](034){4,3};
\draw (4,-4) node [scale=.5](044){4,4};
\draw (5,-4) node [scale=.5](054){4,5};
\draw (6,-4) node [scale=.5](064){4,6};
\draw (000) [->,out=300,in=120,looseness=1] to (014);
\draw (014) [->] to (021);
\draw (021) [->] to (034);
\draw (034) [->] to (043);
\draw (043) [->] to (052);
\draw (052) [->] to (061);
\draw[dashed,->] (061) .. controls +(135:1) and +(-45:1) .. (000);
\end{tikzpicture}
\caption{one directed cycle in $T_5(1)$, with $n=7$.}
\label{t51figure}
\end{center}
\end{figure}
The rest of the arcs are obtained by developing this base directed cycle modulo $x$, i.e. if $\left((\alpha,j),(\beta,j+1)\right)\in E(T_x(i))$ then $\left((\alpha+h,j),(\beta+h,j+1)\right)\in E(T_x(i))$ for all $h\in \mathbb{Z}_x$.
\end{example}
\begin{lemma}\label{bbb}
$T_x(i)$ is a $\overrightarrow{C}_n$-factor for any $i$.
\end{lemma}
\begin{proof}
It suffices to show that the construction gives a base directed cycle of length $n$. The directed cycle containing the vertex $(0,0)$ can be tracked by considering the first coordinate of each vertex that is visited while passing from $G_0$ to $G_1$ to $G_2 \ldots$ to $G_0$. If we add the respective differences of the edges between $G_0$ and $G_1$, $G_1$ and $G_2, \ldots$, $G_{n-1}$ and $G_0$, we must show that this total sum is $0$. Because $2^{e_1}+2^{e_2}+\ldots+2^{e_k}=n-1$, we have for the sum:
\[
i(n-1)-2i(k-1)-i(n-2k+1)=i(n-1)-i(n-1)=0.
\]
\end{proof}
Let $F_h(G)$ be the directed subgraph of the directed graph $G$ that contains only the arcs between parts $h-1$ and $h$. That is
\[
E(F_h(G))=\left\{\left((g_1,h-1)(g_2,h)\right)|\{(g_1,h-1)(g_2,h)\}\in E(G)\right\}.
\]
In particular $F_n(G)$ contains the arcs between $G_{n-1}$ and $G_0$.
Let $H_x(i,s)=T_x(i)\oplus F_n(T_x(i))\oplus F_n(T_x(s))$. This means that $H_x(i,s)$ is the directed subgraph of $\overrightarrow{C}_{(x:n)}$ obtained by taking the same arcs as $T_x(i)$ between $G_j$ and $G_{j+1}$ for $0\leq j \leq n-2$ and $T_x(s)$ between $G_{n-1}$ and $G_0$.
\begin{example}
Figure \ref{Hexample} ilustrates $T_3(0)$, $T_3(1)$, $H_3(0,1)$ and $H_3(1,0)$ with $n=3$.
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black}]
\draw (0,0) node [scale=.5] (000){};
\draw (1,0) node [scale=.5] (010){};
\draw (2,0) node [scale=.5] (020){};
\draw (0,-.5) node [scale=.5](001){};
\draw (1,-.5) node [scale=.5](011){};
\draw (2,-.5) node [scale=.5](021){};
\draw (0,-1) node [scale=.5](002){};
\draw (1,-1) node [scale=.5](012){};
\draw (2,-1) node [scale=.5](022){};
\draw (000) [->] to (010);
\draw (010) [->] to (020);
\draw (020) [bend right=25, dashed,->] to (000);
\draw (001) [->] to (011);
\draw (011) [->] to (021);
\draw (021) [bend right=25, dashed,->] to (001);
\draw (002) [->] to (012);
\draw (012) [->] to (022);
\draw (022) [bend right=25, dashed,->] to (002);
\draw (6,0) node [scale=.5] (100){};
\draw (7,0) node [scale=.5] (110){};
\draw (8,0) node [scale=.5] (120){};
\draw (6,-.5) node [scale=.5](101){};
\draw (7,-.5) node [scale=.5](111){};
\draw (8,-.5) node [scale=.5](121){};
\draw (6,-1) node [scale=.5](102){};
\draw (7,-1) node [scale=.5](112){};
\draw (8,-1) node [scale=.5](122){};
\draw (100) [->] to (112);
\draw (112) [->] to (121);
\draw (121) [dashed,->] to (100);
\draw (101) [->] to (110);
\draw (110) [->] to (122);
\draw (122) [dashed,->] to (101);
\draw (102) [->] to (111);
\draw (111) [->] to (120);
\draw (120) [bend left=13, dashed,->] to (102);
\draw (0,-3) node [scale=.5] (200){};
\draw (1,-3) node [scale=.5] (210){};
\draw (2,-3) node [scale=.5] (220){};
\draw (0,-3.5) node [scale=.5](201){};
\draw (1,-3.5) node [scale=.5](211){};
\draw (2,-3.5) node [scale=.5](221){};
\draw (0,-4) node [scale=.5](202){};
\draw (1,-4) node [scale=.5](212){};
\draw (2,-4) node [scale=.5](222){};
\draw (200) [->] to (210);
\draw (210) [->] to (220);
\draw (201) [->] to (211);
\draw (211) [->] to (221);
\draw (202) [->] to (212);
\draw (212) [->] to (222);
\draw (220) [bend left=13, dashed,->] to (202);
\draw (222) [dashed,->] to (201);
\draw (221) [dashed,->] to (200);
\draw (6,-3) node [scale=.5] (300){};
\draw (7,-3) node [scale=.5] (310){};
\draw (8,-3) node [scale=.5] (320){};
\draw (6,-3.5) node [scale=.5](301){};
\draw (7,-3.5) node [scale=.5](311){};
\draw (8,-3.5) node [scale=.5](321){};
\draw (6,-4) node [scale=.5](302){};
\draw (7,-4) node [scale=.5](312){};
\draw (8,-4) node [scale=.5](322){};
\draw (320) [bend right=25, dashed,->] to (300);
\draw (321) [bend right=25, dashed,->] to (301);
\draw (322) [bend right=25, dashed,->] to (302);
\draw (300) [->] to (312);
\draw (312) [->] to (321);
\draw (301) [->] to (310);
\draw (310) [->] to (322);
\draw (302) [->] to (311);
\draw (311) [->] to (320);
\node[left=10pt,fill=none,draw=none] at (0,-.5) {$T_3(0)=$};
\node[left=10pt,fill=none,draw=none] at (6,-.5) {$T_3(1)=$};
\node[left=10pt,fill=none,draw=none] at (0,-3.5) {$H_3(0,1)=$};
\node[left=10pt,fill=none,draw=none] at (6,-3.5) {$H_3(1,0)=$};
\end{tikzpicture}
\caption{$T_3(0)$, $T_3(1)$, $H_3(0,1)$, $H_3(1,0)$.}
\label{Hexample}
\end{center}
\end{figure}
\end{example}
\begin{lemma}\label{HisforHamilton}
If $gcd(x,i-s)=1$ then $H_x(i,s)$ is a directed Hamiltonian cycle.
\end{lemma}
\begin{proof}
Because the arcs are given by differences it is clear that each vertex has in-degree and out-degree both equal to $1$.
We need to show that all of the vertices are connected. We will first show that there is a directed path between any 2 vertices of $G_0$.
Without loss of generality, we will show that $(0,0)$ is connected to $(\alpha,0)$ for any $\alpha$.
Because the arcs between groups $G_j$ and $G_{j+1}$ are the same as the arcs in $T_x(i)$ for $j=0,\ldots,n-2$ it is easy to see that there is a path from $(0,0)$ to $(i,n-1)$. Now the arc leaving from $(i,n-1)$ has its other end as $(i-s,0)$. So $(0,0)$ is connected to $(i-s,0)$. If we continue on this path, every time we arrive back in $G_0$, we arrive at the vertex $(\alpha'(i-s),0)$. Because $\gcd(x,i-s)=1$, the order of $i-s$ in the cyclic group $\mathbb{Z}_x$ is $x$. Thus any $\alpha$ modulo $x$ can be written as $\alpha'(i-s)$.
Hence $(0,0)$ is connected to all the vertices of $G_0$.
Because we are defining arcs by differences, every vertex in $G_1$ is connected to a vertex in $G_0$, every vertex in $G_2$ to a vertex in $G_1$, and so on. Therefore all the vertices are connected, and the directed cycle is Hamiltonian as we wanted to prove.
\end{proof}
Next we show how to decompose $\overrightarrow{C}_{(x:n)}$ by using the $H_x(i,j)$ graphs. First we will decompose $\overrightarrow{C}_{(x:n)}$ into $\overrightarrow{C}_n$-factors using the $T_x(i)$ graphs. Then we will show how to switch some edges in the $T_x(i)$ graphs to obtain $H_x(i,j)$ graphs. It is important to notice that $H_x(i,i)=T_x(i)$.
\begin{lemma}
Let $x$ be odd, and let $\phi$ be a bijection on $\{0,...,x-1\}$. Then
\[
\overrightarrow{C}_{(x:n)}=\bigoplus_{i=0}^{x-1} T_x(i)=\bigoplus_{i=0}^{x-1} H_x(i,\phi(i))
\]
\end{lemma}
\begin{proof}
To prove the first equality,
\[
\overrightarrow{C}_{(x:n)}=\bigoplus_{i=0}^{x-1} T_x(i),
\]
we need to show that any difference between consecutive parts is covered by one of the $T_x(i)$ graphs. Notice that all the differences in the $T_x$ graphs are given by a power of $2$ times $i$, or $-2i$ or $-i$. It is clear that between parts which use difference $-i$, we cover all the differences, with difference $\delta$ being covered in $T_x(x-\delta)$. Because $x$ is odd, $\gcd(x,2^e)=1$, so the order of $2^e$ in the cyclic group $C_x$ is $x$ for any $1\leq e\leq x-1$. This means that any $\delta\equiv 2^{e_j}i\pmod{x}$ can be written as $2^e\delta'$. Therefore, the difference $\delta$ between the remaining pairs of consecutive groups is covered in some $T_x(\delta')$. A similar calculation can be used for differences of the form $-2i$, because $\gcd(x,x-2)=1$. Here we write $\delta$ as $-2\delta'$, and the difference is covered in $T_x(\delta')$.
For the second equality we have
\begin{align*}
\bigoplus_{i=0}^{x-1} H_x(i,\phi(i))&=\bigoplus_{i=0}^{x-1}\left( T_x(i)\oplus F_n(T_x(i))\oplus F_n (T_x(\phi(i)))\right)\\
&=\bigoplus_{i=0}^{x-1} T_x(i) \bigoplus_{i=0}^{x-1} F_n(T_x(i)) \bigoplus_{i=0}^{x-1} F_n(T_x(\phi(i)))
\end{align*}
and
\begin{align*}
\bigoplus_{i=0}^{x-1} F_n(T_x(\phi(i)))&=\bigoplus_{i=0}^{x-1} F_n(T_x(i))
\end{align*} because $\phi$ is a bijection. So:
\begin{align*}
\bigoplus_{i=0}^{x-1} T_x(i) \bigoplus_{i=0}^{x-1} F_n(T_x(i)) \bigoplus_{i=0}^{x-1} F_n(T_x(\phi(i)))&=\bigoplus_{i=0}^{x-1} T_x(i) \bigoplus_{i=0}^{x-1} F_n(T_x(i)) \bigoplus_{i=0}^{x-1} F_n(T_x(i))\\
&=\bigoplus_{i=0}^{x-1} T_x(i)
\end{align*}
Hence:
\[
\bigoplus_{i=0}^{x-1} T_x(i)=\bigoplus_{i=0}^{x-1} H_x(i,\phi(i))
\]
\end{proof}
In some cases we have $H_x(i,i)$, notice that this is the same as $T_x(i)$. Decomposing $\overrightarrow{C}_{(x:n)}$ into $s$ directed Hamilton cycles and $x-s$ $\overrightarrow{C}_n$-factors is now equivalent to finding a bijection $\phi$ with $gcd(x,i-\phi(i))=1$ for $s$ elements of $\{0,...,x-1\}$ and $\phi(i)=i$ for the rest.
We will use these functions extensively throughout the paper, so we will refer to them as ``the phi-functions'':
Let $2\leq s \leq x$. Define $\phi_s:\mathbb{Z}_x\rightarrow \mathbb{Z}_x$ as follows:
\begin{center}
\underline{\textbf{The phi-functions}}
\[
\phi_s(i)=\left\lbrace \begin{array}{lcl}
i+1 & \text{for} & i\leq (s-3),\, i\text{ even}\\
i-1 & \text{for} & i\leq (s-3),\, i\text{ odd} \\
i+1 & \text{for} & i=s-2\\
i-2 & \text{for} & i=s-1,\, s\text{ odd} \\
i-1 & \text{for} & i=s-1,\, s\text{ even} \\
i &\text{for} & s\leq i \leq x-1\\
\end{array}\right.
\]
\end{center}
For example, if $s=7$ and $x=11$, then
\[
\phi_7=(01)(23)(456)(7)(8)(9)(10)
\]
Notice that $\phi_s$ has $x-s$ fixed points. For any non-fixed point we have $i-\phi(i)\in \{\pm 1,2\}$, and so $\gcd(x,i-\phi(i))=1$ if $x$ is odd.
\begin{theorem}\label{1varfun}
Let $x$ be odd. Let $s\in \{0,...,x\}$, $s\neq 1$. Then $\overrightarrow{C}_{(x:n)}$ can be decomposed into $s$ directed Hamiltonian cycles and $x-s$ $\overrightarrow{C}_n$-factors.
\end{theorem}
\begin{proof}
If $s=0$ we just use the identity mapping.
Otherwise we use the phi-function $\phi_s$.
Hence the discussion that precedes this theorem shows that
\[
\overrightarrow{C}_{(x:n)}=\bigoplus_{i=0}^{x-1} H_x(i,\phi_s(i))
\]
is a decomposition of $\overrightarrow{C}_{(x:n)}$ into $s$ directed Hamiltonian cycles and $x-s$ $\overrightarrow{C}_n$-factors.
\end{proof}
Next we turn to the case of $x$ even. We begin by considering $\overrightarrow{C}_{(4:n)}$.
In $\cite{AKKPO}$ the graphs in Figure \ref{trianglesk43} were used to decompose $K_{(4:3)}$ into triangle factors and $C_6$-factors.
We extend the ideas used in \cite{AKKPO} to decompose $\overrightarrow{C}_{(4:n)}$ into $\overrightarrow{C}_n$-factors and $\overrightarrow{C}_{2n}$-factors.
We will define directed subgraphs $\gamma_{i,j}$ and build the directed graphs $\Gamma(j)$ as the sum of some of these directed subgraphs. In each of these directed subgraphs, the vertices in the top row will be said to have height 0, in the second row height 1, and so on.
We begin with the base directed subgraphs from Figure \ref{babygammas} . In each subgraph, the rows are indexed by their height.
The following result is easy to verify by inspection of the graphs $\gamma_{i,j}$.
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black}]
\draw (0,0) node [scale=.5] (000){};
\draw (1,0) node [scale=.5] (010){};
\draw (2,0) node [scale=.5] (020){};
\draw (0,-.5) node [scale=.5](001){};
\draw (1,-.5) node [scale=.5](011){};
\draw (2,-.5) node [scale=.5](021){};
\draw (0,-1) node [scale=.5](002){};
\draw (1,-1) node [scale=.5](012){};
\draw (2,-1) node [scale=.5](022){};
\draw (0,-1.5) node [scale=.5](003){};
\draw (1,-1.5) node [scale=.5](013){};
\draw (2,-1.5) node [scale=.5](023){};
\draw (000) to (010);
\draw (010) to (020);
\draw (020) [bend right=25, dashed] to (000);
\draw (001) to (013);
\draw (013) to (022);
\draw (022) [dashed] to (001);
\draw (002) to (011);
\draw (011) to (023);
\draw (023) [dashed] to (002);
\draw (003) to (012);
\draw (012) to (021);
\draw (021) [bend left=15, dashed] to (003);
\draw (6,0) node [scale=.5] (100){};
\draw (7,0) node [scale=.5] (110){};
\draw (8,0) node [scale=.5] (120){};
\draw (6,-.5) node [scale=.5](101){};
\draw (7,-.5) node [scale=.5](111){};
\draw (8,-.5) node [scale=.5](121){};
\draw (6,-1) node [scale=.5](102){};
\draw (7,-1) node [scale=.5](112){};
\draw (8,-1) node [scale=.5](122){};
\draw (6,-1.5) node [scale=.5](103){};
\draw (7,-1.5) node [scale=.5](113){};
\draw (8,-1.5) node [scale=.5](123){};
\draw (101) to (111);
\draw (111) to (121);
\draw (121) [bend right=25, dashed] to (101);
\draw (100) to (112);
\draw (112) to (123);
\draw (123) [dashed] to (100);
\draw (102) to (113);
\draw (113) to (120);
\draw (120) [bend left=13, dashed] to (102);
\draw (103) to (110);
\draw (110) to (122);
\draw (122) [dashed] to (103);
\draw (0,-3) node [scale=.5] (200){};
\draw (1,-3) node [scale=.5] (210){};
\draw (2,-3) node [scale=.5] (220){};
\draw (0,-3.5) node [scale=.5](201){};
\draw (1,-3.5) node [scale=.5](211){};
\draw (2,-3.5) node [scale=.5](221){};
\draw (0,-4) node [scale=.5](202){};
\draw (1,-4) node [scale=.5](212){};
\draw (2,-4) node [scale=.5](222){};
\draw (0,-4.5) node [scale=.5](203){};
\draw (1,-4.5) node [scale=.5](213){};
\draw (2,-4.5) node [scale=.5](223){};
\draw (202) to (212);
\draw (212) to (222);
\draw (222) [bend right=25, dashed] to (202);
\draw (201) to (210);
\draw (210) to (223);
\draw (223) [bend left=15, dashed] to (201);
\draw (200) to (213);
\draw (213) to (221);
\draw (221) [dashed] to (200);
\draw (203) to (211);
\draw (211) to (220);
\draw (220) [bend left=25, dashed] to (203);
\draw (6,-3) node [scale=.5] (300){};
\draw (7,-3) node [scale=.5] (310){};
\draw (8,-3) node [scale=.5] (320){};
\draw (6,-3.5) node [scale=.5](301){};
\draw (7,-3.5) node [scale=.5](311){};
\draw (8,-3.5) node [scale=.5](321){};
\draw (6,-4) node [scale=.5](302){};
\draw (7,-4) node [scale=.5](312){};
\draw (8,-4) node [scale=.5](322){};
\draw (6,-4.5) node [scale=.5](303){};
\draw (7,-4.5) node [scale=.5](313){};
\draw (8,-4.5) node [scale=.5](323){};
\draw (303) to (313);
\draw (313) to (323);
\draw (323) [bend left=25, dashed] to (303);
\draw (301) to (312);
\draw (312) to (320);
\draw (320) [dashed] to (301);
\draw (302) to (310);
\draw (310) to (321);
\draw (321) [dashed] to (302);
\draw (300) to (311);
\draw (311) to (322);
\draw (322) [bend left=15, dashed] to (300);
\node[left=10pt,fill=none,draw=none] at (0,-.75) {$\Gamma(0)=$};
\node[left=10pt,fill=none,draw=none] at (6,-.75) {$\Gamma(1)=$};
\node[left=10pt,fill=none,draw=none] at (0,-3.75) {$\Gamma(2)=$};
\node[left=10pt,fill=none,draw=none] at (6,-3.75) {$\Gamma(3)=$};
\end{tikzpicture}
\caption{Triangle factors for $K_{(4:3)}$.}
\label{trianglesk43}
\end{center}
\end{figure}
\begin{figure}
\begin{center
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black}, every path/.style={->}]
\draw (0,0) node [scale=.5] (000){};
\draw (1,0) node [scale=.5] (010){};
\draw (2,0) node [scale=.5] (020){};
\draw (3,0) node [scale=.5] (030){};
\draw (0,-.5) node [scale=.5](001){};
\draw (1,-.5) node [scale=.5](011){};
\draw (2,-.5) node [scale=.5](021){};
\draw (3,-.5) node [scale=.5](031){};
\draw (0,-1) node [scale=.5](002){};
\draw (1,-1) node [scale=.5](012){};
\draw (2,-1) node [scale=.5](022){};
\draw (3,-1) node [scale=.5](032){};
\draw (0,-1.5) node [scale=.5](003){};
\draw (1,-1.5) node [scale=.5](013){};
\draw (2,-1.5) node [scale=.5](023){};
\draw (3,-1.5) node [scale=.5](033){};
\draw (000) to (010);
\draw (010) to (020);
\draw (020) to (030);
\draw (001) to (013);
\draw (013) to (022);
\draw (022) to (031);
\draw (002) to (011);
\draw (011) to (023);
\draw (023) to (032);
\draw (003) to (012);
\draw (012) to (021);
\draw (021) to (033);
\draw (6,0) node [scale=.5] (100){};
\draw (7,0) node [scale=.5] (110){};
\draw (8,0) node [scale=.5] (120){};
\draw (9,0) node [scale=.5] (130){};
\draw (6,-.5) node [scale=.5](101){};
\draw (7,-.5) node [scale=.5](111){};
\draw (8,-.5) node [scale=.5](121){};
\draw (9,-.5) node [scale=.5](131){};
\draw (6,-1) node [scale=.5](102){};
\draw (7,-1) node [scale=.5](112){};
\draw (8,-1) node [scale=.5](122){};
\draw (9,-1) node [scale=.5](132){};
\draw (6,-1.5) node [scale=.5](103){};
\draw (7,-1.5) node [scale=.5](113){};
\draw (8,-1.5) node [scale=.5](123){};
\draw (9,-1.5) node [scale=.5](133){};
\draw (101) to (111);
\draw (111) to (121);
\draw (121) to (131);
\draw (100) to (112);
\draw (112) to (123);
\draw (123) to (130);
\draw (102) to (113);
\draw (113) to (120);
\draw (120) to (132);
\draw (103) to (110);
\draw (110) to (122);
\draw (122) to (133);
\draw (0,-3) node [scale=.5] (200){};
\draw (1,-3) node [scale=.5] (210){};
\draw (2,-3) node [scale=.5] (220){};
\draw (3,-3) node [scale=.5] (230){};
\draw (0,-3.5) node [scale=.5](201){};
\draw (1,-3.5) node [scale=.5](211){};
\draw (2,-3.5) node [scale=.5](221){};
\draw (3,-3.5) node [scale=.5](231){};
\draw (0,-4) node [scale=.5](202){};
\draw (1,-4) node [scale=.5](212){};
\draw (2,-4) node [scale=.5](222){};
\draw (3,-4) node [scale=.5](232){};
\draw (0,-4.5) node [scale=.5](203){};
\draw (1,-4.5) node [scale=.5](213){};
\draw (2,-4.5) node [scale=.5](223){};
\draw (3,-4.5) node [scale=.5](233){};
\draw (202) to (212);
\draw (212) to (222);
\draw (222) to (232);
\draw (201) to (210);
\draw (210) to (223);
\draw (223) to (231);
\draw (200) to (213);
\draw (213) to (221);
\draw (221) to (230);
\draw (203) to (211);
\draw (211) to (220);
\draw (220) to (233);
\draw (6,-3) node [scale=.5] (300){};
\draw (7,-3) node [scale=.5] (310){};
\draw (8,-3) node [scale=.5] (320){};
\draw (9,-3) node [scale=.5] (330){};
\draw (6,-3.5) node [scale=.5](301){};
\draw (7,-3.5) node [scale=.5](311){};
\draw (8,-3.5) node [scale=.5](321){};
\draw (9,-3.5) node [scale=.5](331){};
\draw (6,-4) node [scale=.5](302){};
\draw (7,-4) node [scale=.5](312){};
\draw (8,-4) node [scale=.5](322){};
\draw (9,-4) node [scale=.5](332){};
\draw (6,-4.5) node [scale=.5](303){};
\draw (7,-4.5) node [scale=.5](313){};
\draw (8,-4.5) node [scale=.5](323){};
\draw (9,-4.5) node [scale=.5](333){};
\draw (303) to (313);
\draw (313) to (323);
\draw (323) to (333);
\draw (301) to (312);
\draw (312) to (320);
\draw (320) to (331);
\draw (302) to (310);
\draw (310) to (321);
\draw (321) to (332);
\draw (300) to (311);
\draw (311) to (322);
\draw (322) to (330);
\draw (0,0) node [scale=.5] (000){};
\draw (1,0) node [scale=.5] (010){};
\draw (2,0) node [scale=.5] (020){};
\draw (3,0) node [scale=.5] (030){};
\node[left=5pt,fill=none,draw=none] at (0,0) {$0$};
\node[left=5pt,fill=none,draw=none] at (0,-.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (0,-1) {$2$};
\node[left=5pt,fill=none,draw=none] at (0,-1.5) {$3$};
\node[left=5pt,fill=none,draw=none] at (6,0) {$0$};
\node[left=5pt,fill=none,draw=none] at (6,-.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (6,-1) {$2$};
\node[left=5pt,fill=none,draw=none] at (6,-1.5) {$3$};
\node[left=5pt,fill=none,draw=none] at (6,-3) {$0$};
\node[left=5pt,fill=none,draw=none] at (6,-3.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (6,-4) {$2$};
\node[left=5pt,fill=none,draw=none] at (6,-4.5) {$3$};
\node[left=5pt,fill=none,draw=none] at (0,-3) {$0$};
\node[left=5pt,fill=none,draw=none] at (0,-3.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (0,-4) {$2$};
\node[left=5pt,fill=none,draw=none] at (0,-4.5) {$3$};
\node[left=20pt,fill=none,draw=none] at (0,-.75) {$\gamma_{0,0}=$};
\node[left=20pt,fill=none,draw=none] at (6,-.75) {$\gamma_{0,1}=$};
\node[left=20pt,fill=none,draw=none] at (0,-3.75) {$\gamma_{0,2}=$};
\node[left=20pt,fill=none,draw=none] at (6,-3.75) {$\gamma_{0,3}=$};
\end{tikzpicture}
\end{center}
\vspace{.25in}
\begin{center
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black}, every path/.style={->}]
\draw (2,0) node [scale=.5] (100){};
\draw (3,0) node [scale=.5] (110){};
\draw (4,0) node [scale=.5] (120){};
\draw (5,0) node [scale=.5] (130){};
\draw (6,0) node [scale=.5] (140){};
\draw (2,-.5) node [scale=.5](101){};
\draw (3,-.5) node [scale=.5](111){};
\draw (4,-.5) node [scale=.5](121){};
\draw (5,-.5) node [scale=.5](131){};
\draw (6,-.5) node [scale=.5](141){};
\draw (2,-1) node [scale=.5](102){};
\draw (3,-1) node [scale=.5](112){};
\draw (4,-1) node [scale=.5](122){};
\draw (5,-1) node [scale=.5](132){};
\draw (6,-1) node [scale=.5](142){};
\draw (2,-1.5) node [scale=.5](103){};
\draw (3,-1.5) node [scale=.5](113){};
\draw (4,-1.5) node [scale=.5](123){};
\draw (5,-1.5) node [scale=.5](133){};
\draw (6,-1.5) node [scale=.5](143){};
\draw (100) to (110);
\draw (110) to (120);
\draw (120) to (130);
\draw (130) to (140);
\draw (101) to (113);
\draw (113) to (122);
\draw (122) to (133);
\draw (133) to (141);
\draw (102) to (111);
\draw (111) to (123);
\draw (123) to (131);
\draw (131) to (142);
\draw (103) to (112);
\draw (112) to (121);
\draw (121) to (132);
\draw (132) to (143);
\draw (9,0) node [scale=.5] (200){};
\draw (10,0) node [scale=.5] (210){};
\draw (11,0) node [scale=.5] (220){};
\draw (12,0) node [scale=.5] (230){};
\draw (13,0) node [scale=.5] (240){};
\draw (9,-.5) node [scale=.5](201){};
\draw (10,-.5) node [scale=.5](211){};
\draw (11,-.5) node [scale=.5](221){};
\draw (12,-.5) node [scale=.5](231){};
\draw (13,-.5) node [scale=.5](241){};
\draw (9,-1) node [scale=.5](202){};
\draw (10,-1) node [scale=.5](212){};
\draw (11,-1) node [scale=.5](222){};
\draw (12,-1) node [scale=.5](232){};
\draw (13,-1) node [scale=.5](242){};
\draw (9,-1.5) node [scale=.5](203){};
\draw (10,-1.5) node [scale=.5](213){};
\draw (11,-1.5) node [scale=.5](223){};
\draw (12,-1.5) node [scale=.5](233){};
\draw (13,-1.5) node [scale=.5](243){};
\draw (201) to (211);
\draw (211) to (221);
\draw (221) to (231);
\draw (231) to (241);
\draw (200) to (212);
\draw (212) to (223);
\draw (223) to (232);
\draw (232) to (240);
\draw (202) to (213);
\draw (213) to (220);
\draw (220) to (233);
\draw (233) to (242);
\draw (203) to (210);
\draw (210) to (222);
\draw (222) to (230);
\draw (230) to (243);
\node[left=5pt,fill=none,draw=none] at (2,0) {$0$};
\node[left=5pt,fill=none,draw=none] at (2,-.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (2,-1) {$2$};
\node[left=5pt,fill=none,draw=none] at (2,-1.5) {$3$};
\node[left=5pt,fill=none,draw=none] at (9,0) {$0$};
\node[left=5pt,fill=none,draw=none] at (9,-.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (9,-1) {$2$};
\node[left=5pt,fill=none,draw=none] at (9,-1.5) {$3$};
\node[left=20pt,fill=none,draw=none] at (2,-0.75) {$\gamma_{1,0}=$};
\node[left=20pt,fill=none,draw=none] at (9,-0.75) {$\gamma_{1,1}=$};
\draw (2,-3) node [scale=.5] (300){};
\draw (3,-3) node [scale=.5] (310){};
\draw (4,-3) node [scale=.5] (320){};
\draw (5,-3) node [scale=.5] (330){};
\draw (6,-3) node [scale=.5] (340){};
\draw (2,-3.5) node [scale=.5](301){};
\draw (3,-3.5) node [scale=.5](311){};
\draw (4,-3.5) node [scale=.5](321){};
\draw (5,-3.5) node [scale=.5](331){};
\draw (6,-3.5) node [scale=.5](341){};
\draw (2,-4) node [scale=.5](302){};
\draw (3,-4) node [scale=.5](312){};
\draw (4,-4) node [scale=.5](322){};
\draw (5,-4) node [scale=.5](332){};
\draw (6,-4) node [scale=.5](342){};
\draw (2,-4.5) node [scale=.5](303){};
\draw (3,-4.5) node [scale=.5](313){};
\draw (4,-4.5) node [scale=.5](323){};
\draw (5,-4.5) node [scale=.5](333){};
\draw (6,-4.5) node [scale=.5](343){};
\draw (302) to (312);
\draw (312) to (322);
\draw (322) to (332);
\draw (332) to (342);
\draw (301) to (310);
\draw (310) to (323);
\draw (323) to (330);
\draw (330) to (341);
\draw (300) to (313);
\draw (313) to (321);
\draw (321) to (333);
\draw (333) to (340);
\draw (303) to (311);
\draw (311) to (320);
\draw (320) to (331);
\draw (331) to (343);
\draw (9,-3) node [scale=.5] (400){};
\draw (10,-3) node [scale=.5] (410){};
\draw (11,-3) node [scale=.5] (420){};
\draw (12,-3) node [scale=.5] (430){};
\draw (13,-3) node [scale=.5] (440){};
\draw (9,-3.5) node [scale=.5](401){};
\draw (10,-3.5) node [scale=.5](411){};
\draw (11,-3.5) node [scale=.5](421){};
\draw (12,-3.5) node [scale=.5](431){};
\draw (13,-3.5) node [scale=.5](441){};
\draw (9,-4) node [scale=.5](402){};
\draw (10,-4) node [scale=.5](412){};
\draw (11,-4) node [scale=.5](422){};
\draw (12,-4) node [scale=.5](432){};
\draw (13,-4) node [scale=.5](442){};
\draw (9,-4.5) node [scale=.5](403){};
\draw (10,-4.5) node [scale=.5](413){};
\draw (11,-4.5) node [scale=.5](423){};
\draw (12,-4.5) node [scale=.5](433){};
\draw (13,-4.5) node [scale=.5](443){};
\draw (403) to (413);
\draw (413) to (423);
\draw (423) to (433);
\draw (433) to (443);
\draw (401) to (412);
\draw (412) to (420);
\draw (420) to (432);
\draw (432) to (441);
\draw (402) to (410);
\draw (410) to (421);
\draw (421) to (430);
\draw (430) to (442);
\draw (400) to (411);
\draw (411) to (422);
\draw (422) to (431);
\draw (431) to (440);
\node[left=5pt,fill=none,draw=none] at (2,-3) {$0$};
\node[left=5pt,fill=none,draw=none] at (2,-3.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (2,-4) {$2$};
\node[left=5pt,fill=none,draw=none] at (2,-4.5) {$3$};
\node[left=5pt,fill=none,draw=none] at (9,-3) {$0$};
\node[left=5pt,fill=none,draw=none] at (9,-3.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (9,-4) {$2$};
\node[left=5pt,fill=none,draw=none] at (9,-4.5) {$3$};
\node[left=20pt,fill=none,draw=none] at (2,-3.75) {$\gamma_{1,2}=$};
\node[left=20pt,fill=none,draw=none] at (9,-3.75) {$\gamma_{1,3}=$};
\end{tikzpicture}
\end{center}
\vspace{.25in}
\begin{center
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black}, every path/.style={->}]
\draw (1,0) node [scale=.5] (100){};
\draw (2,0) node [scale=.5] (110){};
\draw (3,0) node [scale=.5] (120){};
\draw (4,0) node [scale=.5] (130){};
\draw (5,0) node [scale=.5] (140){};
\draw (6,0) node [scale=.5] (150){};
\draw (1,-.5) node [scale=.5](101){};
\draw (2,-.5) node [scale=.5](111){};
\draw (3,-.5) node [scale=.5](121){};
\draw (4,-.5) node [scale=.5](131){};
\draw (5,-.5) node [scale=.5](141){};
\draw (6,-.5) node [scale=.5](151){};
\draw (1,-1) node [scale=.5](102){};
\draw (2,-1) node [scale=.5](112){};
\draw (3,-1) node [scale=.5](122){};
\draw (4,-1) node [scale=.5](132){};
\draw (5,-1) node [scale=.5](142){};
\draw (6,-1) node [scale=.5](152){};
\draw (1,-1.5) node [scale=.5](103){};
\draw (2,-1.5) node [scale=.5](113){};
\draw (3,-1.5) node [scale=.5](123){};
\draw (4,-1.5) node [scale=.5](133){};
\draw (5,-1.5) node [scale=.5](143){};
\draw (6,-1.5) node [scale=.5](153){};
\draw (100) to (110);
\draw (110) to (120);
\draw (120) to (130);
\draw (130) to (140);
\draw (140) to (150);
\draw (101) to (113);
\draw (113) to (122);
\draw (122) to (131);
\draw (131) to (143);
\draw (143) to (151);
\draw (102) to (111);
\draw (111) to (123);
\draw (123) to (132);
\draw (132) to (141);
\draw (141) to (152);
\draw (103) to (112);
\draw (112) to (121);
\draw (121) to (133);
\draw (133) to (142);
\draw (142) to (153);
\draw (9,0) node [scale=.5] (200){};
\draw (10,0) node [scale=.5] (210){};
\draw (11,0) node [scale=.5] (220){};
\draw (12,0) node [scale=.5] (230){};
\draw (13,0) node [scale=.5] (240){};
\draw (14,0) node [scale=.5] (250){};
\draw (9,-.5) node [scale=.5](201){};
\draw (10,-.5) node [scale=.5](211){};
\draw (11,-.5) node [scale=.5](221){};
\draw (12,-.5) node [scale=.5](231){};
\draw (13,-.5) node [scale=.5](241){};
\draw (14,-.5) node [scale=.5](251){};
\draw (9,-1) node [scale=.5](202){};
\draw (10,-1) node [scale=.5](212){};
\draw (11,-1) node [scale=.5](222){};
\draw (12,-1) node [scale=.5](232){};
\draw (13,-1) node [scale=.5](242){};
\draw (14,-1) node [scale=.5](252){};
\draw (9,-1.5) node [scale=.5](203){};
\draw (10,-1.5) node [scale=.5](213){};
\draw (11,-1.5) node [scale=.5](223){};
\draw (12,-1.5) node [scale=.5](233){};
\draw (13,-1.5) node [scale=.5](243){};
\draw (14,-1.5) node [scale=.5](253){};
\draw (201) to (211);
\draw (211) to (221);
\draw (221) to (231);
\draw (231) to (241);
\draw (241) to (251);
\draw (200) to (212);
\draw (212) to (223);
\draw (223) to (230);
\draw (230) to (242);
\draw (242) to (250);
\draw (202) to (213);
\draw (213) to (220);
\draw (220) to (232);
\draw (232) to (243);
\draw (243) to (252);
\draw (203) to (210);
\draw (210) to (222);
\draw (222) to (233);
\draw (233) to (240);
\draw (240) to (253);
\node[left=5pt,fill=none,draw=none] at (1,0) {$0$};
\node[left=5pt,fill=none,draw=none] at (1,-.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (1,-1) {$2$};
\node[left=5pt,fill=none,draw=none] at (1,-1.5) {$3$};
\node[left=5pt,fill=none,draw=none] at (9,0) {$0$};
\node[left=5pt,fill=none,draw=none] at (9,-.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (9,-1) {$2$};
\node[left=5pt,fill=none,draw=none] at (9,-1.5) {$3$};
\node[left=20pt,fill=none,draw=none] at (1,-0.75) {$\gamma_{2,0}=$};
\node[left=20pt,fill=none,draw=none] at (9,-0.75) {$\gamma_{2,1}=$};
\draw (1,-3) node [scale=.5] (300){};
\draw (2,-3) node [scale=.5] (310){};
\draw (3,-3) node [scale=.5] (320){};
\draw (4,-3) node [scale=.5] (330){};
\draw (5,-3) node [scale=.5] (340){};
\draw (6,-3) node [scale=.5] (350){};
\draw (1,-3.5) node [scale=.5](301){};
\draw (2,-3.5) node [scale=.5](311){};
\draw (3,-3.5) node [scale=.5](321){};
\draw (4,-3.5) node [scale=.5](331){};
\draw (5,-3.5) node [scale=.5](341){};
\draw (6,-3.5) node [scale=.5](351){};
\draw (1,-4) node [scale=.5](302){};
\draw (2,-4) node [scale=.5](312){};
\draw (3,-4) node [scale=.5](322){};
\draw (4,-4) node [scale=.5](332){};
\draw (5,-4) node [scale=.5](342){};
\draw (6,-4) node [scale=.5](352){};
\draw (1,-4.5) node [scale=.5](303){};
\draw (2,-4.5) node [scale=.5](313){};
\draw (3,-4.5) node [scale=.5](323){};
\draw (4,-4.5) node [scale=.5](333){};
\draw (5,-4.5) node [scale=.5](343){};
\draw (6,-4.5) node [scale=.5](353){};
\draw (302) to (312);
\draw (312) to (322);
\draw (322) to (332);
\draw (332) to (342);
\draw (342) to (352);
\draw (301) to (310);
\draw (310) to (323);
\draw (323) to (331);
\draw (331) to (340);
\draw (340) to (351);
\draw (300) to (313);
\draw (313) to (321);
\draw (321) to (330);
\draw (330) to (343);
\draw (343) to (350);
\draw (303) to (311);
\draw (311) to (320);
\draw (320) to (333);
\draw (333) to (341);
\draw (341) to (353);
\draw (9,-3) node [scale=.5] (400){};
\draw (10,-3) node [scale=.5] (410){};
\draw (11,-3) node [scale=.5] (420){};
\draw (12,-3) node [scale=.5] (430){};
\draw (13,-3) node [scale=.5] (440){};
\draw (14,-3) node [scale=.5] (450){};
\draw (9,-3.5) node [scale=.5](401){};
\draw (10,-3.5) node [scale=.5](411){};
\draw (11,-3.5) node [scale=.5](421){};
\draw (12,-3.5) node [scale=.5](431){};
\draw (13,-3.5) node [scale=.5](441){};
\draw (14,-3.5) node [scale=.5](451){};
\draw (9,-4) node [scale=.5](402){};
\draw (10,-4) node [scale=.5](412){};
\draw (11,-4) node [scale=.5](422){};
\draw (12,-4) node [scale=.5](432){};
\draw (13,-4) node [scale=.5](442){};
\draw (14,-4) node [scale=.5](452){};
\draw (9,-4.5) node [scale=.5](403){};
\draw (10,-4.5) node [scale=.5](413){};
\draw (11,-4.5) node [scale=.5](423){};
\draw (12,-4.5) node [scale=.5](433){};
\draw (13,-4.5) node [scale=.5](443){};
\draw (14,-4.5) node [scale=.5](453){};
\draw (403) to (413);
\draw (413) to (423);
\draw (423) to (433);
\draw (433) to (443);
\draw (443) to (453);
\draw (401) to (412);
\draw (412) to (420);
\draw (420) to (431);
\draw (431) to (442);
\draw (442) to (451);
\draw (402) to (410);
\draw (410) to (421);
\draw (421) to (432);
\draw (432) to (440);
\draw (440) to (452);
\draw (400) to (411);
\draw (411) to (422);
\draw (422) to (430);
\draw (430) to (441);
\draw (441) to (450);
\node[left=5pt,fill=none,draw=none] at (1,-3) {$0$};
\node[left=5pt,fill=none,draw=none] at (1,-3.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (1,-4) {$2$};
\node[left=5pt,fill=none,draw=none] at (1,-4.5) {$3$};
\node[left=5pt,fill=none,draw=none] at (9,-3) {$0$};
\node[left=5pt,fill=none,draw=none] at (9,-3.5) {$1$};
\node[left=5pt,fill=none,draw=none] at (9,-4) {$2$};
\node[left=5pt,fill=none,draw=none] at (9,-4.5) {$3$};
\node[left=20pt,fill=none,draw=none] at (1,-3.75) {$\gamma_{2,2}=$};
\node[left=20pt,fill=none,draw=none] at (9,-3.75) {$\gamma_{2,3}=$};
\end{tikzpicture}
\caption{}
\label{babygammas}
\end{center}
\end{figure}
\begin{lemma}\label{fixedheight}
For any $i,j,h$, the directed path beginning at height $h$ in the first column of $\gamma_{i,j}$ finishes at height $h$ in the last column of $\gamma_{i,j}$.
\end{lemma}
Let $n=3b+a$ with $0\leq a < 3$, $b\geq 2$.
For $0\leq t \leq b-2$, we define $\gamma_{0,j}(t)$ as the directed graph $\gamma_{0,j}$ on the parts $G_{3t-1},G_{3t},G_{3t+1},G_{3t+2}$, with calculations done in $\mathbb{Z}_n$. We define $\gamma_{a,j}(n)$ as the directed graph $\gamma_{a,j}$ on the parts $G_{3b-4},G_{3b-3},\ldots,G_{3b+a-1}$.
Let $\Gamma(j)=\bigoplus_{t=0}^{b-2}\gamma_{0,j}(t)\oplus \gamma_{a,j}(n)$. Notice that $\gamma_{0,j}(0)$ is on the parts $G_{-1}, G_{0}, G_{1}$ and $G_{2}$. This means that $F_n(\Gamma(j))$ is the matching between the first and second columns in $\gamma_{0,j}(0)$.
\begin{example}
Let $n=7$. We will construct $\Gamma(0)$. Since $n=6+1$, we have $b=2$ and $a=1$. This means that $\Gamma(0)=\gamma_{0,0}(0)\oplus\gamma_{1,0}(7)$, where $\gamma_{0,0}(0)$ is on the parts $G_{-1}=G_{7-1}=G_6,G_0,G_1$ and $G_2$; and $\gamma_{1,0}(7)$ is on the parts $G_2,G_3,G_4,G_5$ and $G_6$. So we get the picture in Figure \ref{Gamma0a}, where the arcs from $\gamma_{0,0}(0)$ are dashed.
\begin{figure}
\begin{center}
\begin{minipage}[b]{.5\linewidth}
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black}, every path/.style={->}]
\draw (0,0) node [scale=.5] (000){};
\draw (1,0) node [scale=.5] (010){};
\draw (2,0) node [scale=.5] (020){};
\draw (3,0) node [scale=.5] (030){};
\draw (0,-.5) node [scale=.5](001){};
\draw (1,-.5) node [scale=.5](011){};
\draw (2,-.5) node [scale=.5](021){};
\draw (3,-.5) node [scale=.5](031){};
\draw (0,-1) node [scale=.5](002){};
\draw (1,-1) node [scale=.5](012){};
\draw (2,-1) node [scale=.5](022){};
\draw (3,-1) node [scale=.5](032){};
\draw (0,-1.5) node [scale=.5](003){};
\draw (1,-1.5) node [scale=.5](013){};
\draw (2,-1.5) node [scale=.5](023){};
\draw (3,-1.5) node [scale=.5](033){};
\draw (000)[dashed] to (010);
\draw (010)[dashed] to (020);
\draw (020)[dashed] to (030);
\draw (001)[dashed] to (013);
\draw (013) [dashed]to (022);
\draw (022)[dashed] to (031);
\draw (002)[dashed] to (011);
\draw (011)[dashed] to (023);
\draw (023)[dashed] to (032);
\draw (003)[dashed] to (012);
\draw (012)[dashed] to (021);
\draw (021)[dashed] to (033);
\draw (4,0) node [scale=.5] (110){};
\draw (5,0) node [scale=.5] (120){};
\draw (6,0) node [scale=.5] (130){};
\draw (7,0) node [scale=.5] (140){};
\draw (4,-.5) node [scale=.5](111){};
\draw (5,-.5) node [scale=.5](121){};
\draw (6,-.5) node [scale=.5](131){};
\draw (7,-.5) node [scale=.5](141){};
\draw (4,-1) node [scale=.5](112){};
\draw (5,-1) node [scale=.5](122){};
\draw (6,-1) node [scale=.5](132){};
\draw (7,-1) node [scale=.5](142){};
\draw (4,-1.5) node [scale=.5](113){};
\draw (5,-1.5) node [scale=.5](123){};
\draw (6,-1.5) node [scale=.5](133){};
\draw (7,-1.5) node [scale=.5](143){};
\draw (030) to (110);
\draw (110) to (120);
\draw (120) to (130);
\draw (130) to (140);
\draw (031) to (113);
\draw (113) to (122);
\draw (122) to (133);
\draw (133) to (141);
\draw (032) to (111);
\draw (111) to (123);
\draw (123) to (131);
\draw (131) to (142);
\draw (033) to (112);
\draw (112) to (121);
\draw (121) to (132);
\draw (132) to (143);
\node[above=10pt,fill=none,draw=none] at (0,0) {$G_6$};
\node[above=10pt,fill=none,draw=none] at (1,0) {$G_0$};
\node[above=10pt,fill=none,draw=none] at (2,0) {$G_1$};
\node[above=10pt,fill=none,draw=none] at (3,0) {$G_2$};
\node[above=10pt,fill=none,draw=none] at (4,0) {$G_3$};
\node[above=10pt,fill=none,draw=none] at (5,0) {$G_4$};
\node[above=10pt,fill=none,draw=none] at (6,0) {$G_5$};
\node[above=10pt,fill=none,draw=none] at (7,0) {$G_6$};
\end{tikzpicture}
\subcaption{}\label{Gamma0a}
\end{minipage}
\vspace{.5in}
\begin{minipage}[b]{.5\linewidth}
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black}, every path/.style={->}]
\draw (1,0) node [scale=.5] (010){};
\draw (2,0) node [scale=.5] (020){};
\draw (3,0) node [scale=.5] (030){};
\draw (1,-.5) node [scale=.5](011){};
\draw (2,-.5) node [scale=.5](021){};
\draw (3,-.5) node [scale=.5](031){};
\draw (1,-1) node [scale=.5](012){};
\draw (2,-1) node [scale=.5](022){};
\draw (3,-1) node [scale=.5](032){};
\draw (1,-1.5) node [scale=.5](013){};
\draw (2,-1.5) node [scale=.5](023){};
\draw (3,-1.5) node [scale=.5](033){};
\draw (4,0) node [scale=.5] (110){};
\draw (5,0) node [scale=.5] (120){};
\draw (6,0) node [scale=.5] (130){};
\draw (7,0) node [scale=.5] (140){};
\draw (4,-.5) node [scale=.5](111){};
\draw (5,-.5) node [scale=.5](121){};
\draw (6,-.5) node [scale=.5](131){};
\draw (7,-.5) node [scale=.5](141){};
\draw (4,-1) node [scale=.5](112){};
\draw (5,-1) node [scale=.5](122){};
\draw (6,-1) node [scale=.5](132){};
\draw (7,-1) node [scale=.5](142){};
\draw (4,-1.5) node [scale=.5](113){};
\draw (5,-1.5) node [scale=.5](123){};
\draw (6,-1.5) node [scale=.5](133){};
\draw (7,-1.5) node [scale=.5](143){};
\draw (140)[dashed, bend right=15] to (010);
\draw (010)[dashed] to (020);
\draw (020)[dashed] to (030);
\draw (141)[dashed, bend left=35] to (013);
\draw (013) [dashed]to (022);
\draw (022)[dashed] to (031);
\draw (142)[dashed, bend right=17] to (011);
\draw (011)[dashed] to (023);
\draw (023)[dashed] to (032);
\draw[dashed] (143) .. controls +(135:1) and +(-45:1) .. (012);
\draw (012)[dashed] to (021);
\draw (021)[dashed] to (033);
\draw (030) to (110);
\draw (110) to (120);
\draw (120) to (130);
\draw (130) to (140);
\draw (031) to (113);
\draw (113) to (122);
\draw (122) to (133);
\draw (133) to (141);
\draw (032) to (111);
\draw (111) to (123);
\draw (123) to (131);
\draw (131) to (142);
\draw (033) to (112);
\draw (112) to (121);
\draw (121) to (132);
\draw (132) to (143);
\node[above=10pt,fill=none,draw=none] at (1,0) {$G_0$};
\node[above=10pt,fill=none,draw=none] at (2,0) {$G_1$};
\node[above=10pt,fill=none,draw=none] at (3,0) {$G_2$};
\node[above=10pt,fill=none,draw=none] at (4,0) {$G_3$};
\node[above=10pt,fill=none,draw=none] at (5,0) {$G_4$};
\node[above=10pt,fill=none,draw=none] at (6,0) {$G_5$};
\node[above=10pt,fill=none,draw=none] at (7,0) {$G_6$};
\end{tikzpicture}
\subcaption{}
\label{Gamma0b}
\end{minipage}
\caption{}
\label{Gamma0}
\end{center}
\end{figure}
Notice that the first and last columns are both $G_6$. Because the directed graph $\overrightarrow{C}_{(4:n)}$ has $G_0$ as the first column, we connect the vertices from $G_0$ to the last column instead, obtaining the picture in Figure \ref{Gamma0b}; where the dashed arcs still belong to $\gamma_{0,0}(0)$.
\end{example}
Using Lemma \ref{fixedheight} we have the following result.
\begin{lemma}\label{GisforGamma}
\[
\overrightarrow{C}_{(4:n)}=\Gamma(0)\oplus \Gamma(1) \oplus \Gamma(2) \oplus \Gamma(3)=\bigoplus_{j=0}^3 \Gamma(j)
\]
is a $\overrightarrow{C}_n$-factorization of $\overrightarrow{C}_{(4:n)}$.
\end{lemma}
\begin{proof}
It is easy to verify from the pictures that for any given $0\leq i\leq 2$, the directed graphs $\gamma_{i,0},\gamma_{i,1},\gamma_{i,2}$ and $\gamma_{i,3}$ are arc disjoint.
By Lemma \ref{fixedheight}, in each $\gamma_{i,j}$ the directed paths start and end at the same height. Thus when we connect all the directed paths in each factor $\Gamma(j)$, we obtain four directed cycles of length $n$.
Therefore $\bigoplus_{j=0}^3 \Gamma(j)$ is a $\overrightarrow{C}_n$-factorization of $\overrightarrow{C}_{(4:n)}$.
\end{proof}
To construct directed cycles of size $2n$, we perform switches on the edges between columns. Define $\lambda_{i,j}=\gamma_{0,i}(0)\oplus F_n(\gamma_{0,i}(0))\oplus F_n(\gamma_{0,j}(0))$. Keep in mind that $\gamma_{0,i}(0)$ is on the parts $G_{n-1},G_0,G_1,G_2$, and so $F_n(\gamma_{0,i}(0))$ only consists of the edges between parts $G_{n-1}$ and $G_0$.
\begin{lemma}\label{2heights}
If the directed path that starts at height $h_1$ in part $G_{n-1}$ in $\lambda_{i,j}$ ends at height $h_2$ in $G_2$, then the directed path that starts at height $h_2$ in $G_{n-1}$ ends at height $h_1$ in $G_2$. Even more, if $i\neq j$ then no directed path starts and ends at the same height.
\end{lemma}
\begin{proof}
We build tables that show for each possible combination of $i$ and $j$, the starting and ending heights of the directed paths $\lambda_{i,j}$. We have one table for each $i$, with the rows indexed by the options for $j$, and the columns indexed by the options for the starting height of each directed path. The entry in the table gives the finishing height.
\[
i=0\quad
\begin{array}{|c|c|c|c|c|}
\hline
&\text{height $0$} & \text{height $1$} & \text{height $2$} & \text{height $3$}\\
\hline
j=0 & 0 & 1 & 2 & 3\\
\hline
j=1 & 3 & 2 & 1 & 0\\
\hline
j=2 & 1 & 0 & 3 & 2\\
\hline
j=3 & 2 & 3 & 0 &1\\
\hline
\end{array}
\]
\[
i=1\quad
\begin{array}{|c|c|c|c|c|}
\hline
&\text{height $0$} & \text{height $1$} & \text{height $2$} & \text{height $3$}\\
\hline
j=0 & 3 & 2 & 1 & 0\\
\hline
j=1 & 0 & 1 & 2 & 3\\
\hline
j=2 & 2 & 3 & 0 & 1\\
\hline
j=3 & 1 & 0 & 3 &2\\
\hline
\end{array}
\]
\[
i=2\quad
\begin{array}{|c|c|c|c|c|}
\hline
&\text{height $0$} & \text{height $1$} & \text{height $2$} & \text{height $3$}\\
\hline
j=0 & 1& 0 & 3 & 2\\
\hline
j=1 & 2 & 3 & 0 & 1\\
\hline
j=2 & 0 & 1 & 2 & 3\\
\hline
j=3 & 3 & 2 & 1 &0\\
\hline
\end{array}
\]
\[
i=3\quad
\begin{array}{|c|c|c|c|c|}
\hline
&\text{height $0$} & \text{height $1$} & \text{height $2$} & \text{height $3$}\\
\hline
j=0 & 2 & 3 & 0 & 1\\
\hline
j=1 & 1 & 0 & 3 & 2\\
\hline
j=2 & 3 & 2 & 1 & 0\\
\hline
j=3 & 0 & 1 & 2 & 3\\
\hline
\end{array}
\]
Notice that whenever $i=j$ we have $\lambda_{i,i}=\gamma_{0,i}(0)$, in which case we already know that the starting and ending heights of each directed path are the same. When $i\neq j$ the starting and ending heights are never the same, but if the starting height in $\lambda_{i,j}$ is $h_1$ and the ending height is $h_2$, then the directed path with starting height $h_2$ has ending height $h_1$.
Therefore the result is proven.
\end{proof}
Let $\Lambda(i,j)=\Gamma(i)\oplus F_n(\Gamma(i))\oplus F_n(\Gamma(j))$.
\begin{lemma}\label{LisforLambda}
If $i\neq j$,
$\Lambda(i,j)$ is a $\overrightarrow{C}_{2n}$-factor.
\end{lemma}
\begin{proof}
Notice that
\begin{align*}
\Lambda(i,j)&=\Gamma(i)\oplus F_n(\Gamma(i))\oplus F_n(\Gamma(j))\\
&=\left(\bigoplus_{t=0}^{b-2}\gamma_{0,i}(t)\right)\oplus \gamma_{a,i}(n)\oplus F_n(\Gamma(i))\oplus F_n(\Gamma(j))
\end{align*}
Because $F_n(\Gamma(i))$ is the matching between the first and second columns in $\gamma_{0,i}(0)$, we have $F_n(\Gamma(i))=F_n(\gamma_{0,i}(0))$. Therefore,
\begin{align*}
\Lambda(i,j)&=\left(\bigoplus_{t=0}^{b-2}\gamma_{0,i}(t)\right)\oplus \gamma_{a,i}(n)\oplus F_n(\Gamma(i))\oplus F_n(\Gamma(j))\\
&=\left(\bigoplus_{t=0}^{b-2}\gamma_{0,i}(t)\right)\oplus \gamma_{a,i}(n)\oplus F_n(\gamma_{0,i}(0))\oplus F_n(\gamma_{0,j}(0))\\
&=\left(\bigoplus_{t=1}^{b-2}\gamma_{0,i}(t)\right)\oplus \gamma_{a,i}(n)\oplus\left( \gamma_{0,i}(0)\oplus F_n(\gamma_{0,i}(0))\oplus F_n(\gamma_{0,j}(0))\right)\\
&=\left(\bigoplus_{t=1}^{b-2}\gamma_{0,i}(t)\right)\oplus \gamma_{a,i}(n)\oplus \lambda_{i,j}\\
\end{align*}
Consider the directed cycle that contains the vertex at height $h_1$ in $G_{n-1}$.
From Lemma \ref{2heights} we know that in $\lambda_{i,j}$ the directed path that starts at height $h_1$ in $G_{n-1}$ finishes at height $h_2$ in $G_2$. By Lemma \ref{fixedheight}, the directed paths through all the $\gamma_{i,j}(l)$, with $l\in\{1,2,\ldots, b-1,n\}$ start and end at the same heights. So when we reach $G_{n-1}$ again, it is at at height $h_2$. We leave $G_2$ at height $h_1$ this time, and as we move through all the $\gamma_{i,j}(l)$, with $l\in\{1,2,\ldots, b-1,n\}$, the heights never change. Therefore, we reach $G_{n-1}$ again at height $h_1$, closing the directed cycle.
This produces one directed cycle of size $2n$. By repeating the process with the directed cycle starting at one of the vertices that we have not used yet, we get the second directed cycle.
Therefore $\Lambda(i,j)$ consists of two directed cycles of length $2n$.
\end{proof}
Notice that if $i=j$, then $\Lambda(i,j)=\Gamma(i)$ consists of $4$ directed cycles of length $n$.
\begin{theorem}\label{even}
If $s\in \{0,2,3,4\}$, then $\overrightarrow{C}_{(4:n)}$ can be decomposed into $s$ $\overrightarrow{C}_{2n}$-factors and $4-s$ $\overrightarrow{C}_{n}$-factors.
\end{theorem}
\begin{proof}
Let $\pi$ be a permutation of the set $\{0,1,2,3\}$ with exactly $4-s$ fixed points. Then
\[
\overrightarrow{C}_{(4:n)}=\bigoplus_{j=0}^3 \Gamma(j)=\bigoplus_{j=0}^3 \left(\Gamma(j)\oplus F_n(\Gamma(j))\oplus F_n(\Gamma(\pi(j))\right)=\bigoplus_{j=0}^3 \Lambda(j,\pi(j))
\]
Since $\Lambda(j,\pi(j))$ is a $\overrightarrow{C}_{2n}$-factor if $j\neq \pi(j)$ and a $\overrightarrow{C}_n$-factor otherwise, the theorem is proven.
\end{proof}
\begin{remark}
Notice that if $n=5$, we have $b=1$ and $a=2$. This means we have $\Gamma(j)=\gamma_{a,j}(5)=\gamma_{2,j}(5)$, which is on the parts $G_{-1}=G_4,G_0,G_1,G_2,G_3,G_4$. This will actually close the directed cycle. The results given in Lemmas \ref{2heights} and \ref{LisforLambda} only apply to $b\geq 2$, but it can be shown that the same results are true with $b=1$ by applying similar techniques on $\gamma_{a,j}$ instead of $\gamma_{0,j}$.
\end{remark}
There is one more basic decomposition that we will use, based on the resolvable gregarious decomposition of $K_{(w:n)}$ from \cite{BHR}. We make use of the constructions given in Lemma $3.1$ and Corollary $3.2$ of \cite{BHR}, and apply them to $\overrightarrow{C}_{(w:n)}$ instead of $K_{(w:n)}$.
\begin{definition}
A quasigroup $(Q,*)$ is a set $Q$ with a binary operation $*$ such that for each $a$ and $b$ in $Q$, there exist unique elements $x$ and $y$ in $Q$ such that:
\begin{itemize}
\item $a*x=b$;
\item $y*a=b$.
\end{itemize}
\end{definition}
\begin{definition}
Two quasigroups on the same set $(Q,*)$, $(Q,\circ)$ are said to be orthogonal if $i*j\neq i\circ j$ for every $i,j$ in $Q$.
\end{definition}
The reader may be familiar with Latin Squares, which are the multiplication tables of quasigroups, and mutually orthogonal Latin Squares, which are the multiplication tables of orthogonal quasigroups.
In \cite{BS}, \cite{BSP} it was shown that if $|Q|\not\in\{1,2,6\}$ then there are at least $2$ orthogonal quasigroups on $Q$.
Again, the decomposition in the following theorem is obtained by modifying the construction from Lemma $3.2$ in \cite{BHR} to work with $\overrightarrow{C}_{(w:n)}$ instead of $K_{(w:n)}$:
\begin{theorem}\label{gregarious}
Let $w\not\in\{2,6\}$ and $n$ odd. Then there is a decomposition of $\overrightarrow{C}_{(w:n)}$ into $\overrightarrow{C}_n$-factors.
\end{theorem}
\begin{proof}
Since $w\not\in\{2,6\}$ there exist two orthogonal quasigroups (Latin Squares) $(Q,\circ)$ and $(Q,*)$ of order $w$, with $Q=\{0,1,2,\dots,w-1\}$. We take directed cycles of the form:
\[
(i,0)(j,1)(i,2)(j,3)\ldots(i,n-3)(j,n-2)(k,n-1),\quad \text{ where }\quad
0\leq i,j \leq w-1, k=i\circ j.
\]
This produces a decomposition of $\overrightarrow{C}_{(w:n)}$ into $w^2$ directed cycles of size $n$.
To form a $\overrightarrow{C}_n$-factor, given $l\in Q$ we take all cycles arising from the pairs $i,j$ with $i*j=l$ in the second quasigroup $(Q,*)$. Thus we have a decomposition of $\overrightarrow{C}_{(w:n)}$ into $w$ $\overrightarrow{C}_n$-factors.
\end{proof}
\section{Multivariable functions}
\begin{definition}\label{definition71}
Let $x$ and $y$ be odd. We define $T_{(xy)}(i,\alpha)$ to be the directed subgraph of $\overrightarrow{C}_{(xy:n)}$ obtained by taking $T_{(xy)}(i,\alpha)=T_{x}(i)\otimes T_y(\alpha)$. We also define
\[
H_{(xy)}(i,\alpha)(j,\beta)=T_{(xy)}(i,\alpha)\oplus F_n(T_{(xy)}(i,\alpha))\oplus F_n(T_{(xy)}(j,\beta))\]
This means that $H_{(xy)}(i,\alpha)(j,\beta)$ is the directed graph obtained by taking the arcs of $T_{(xy)}(i,\alpha)$ between parts $t$ and $t+1$ for $0\leq t\leq n-2$, and the arcs between parts $n-1$ and $0$ from $T_{(xy)}(j,\beta)$.
\end{definition}
\begin{example}\label{examplemultipartite}
Figure \ref{figdef71} illustrates the first part of Definition \ref{definition71} by showing $T_{x}(i)$, $T_y(\alpha)$ and $T_{(xy)}(i,\alpha)$, for $x=3$, $y=5$, $i=1$ and $\alpha=2$, with $3$ partite sets.
Figure \ref{figlemma72} illustrates the second part of Definition \ref{definition71} by showing $H_{(xy)}(i,\alpha)(j,\beta)$, for $x=3$, $y=5$, $i=1$, $\alpha=2$, $j=2$, $\beta=4$, with $3$ partite sets. Figure \ref{figlemma72} also shows $H_x(i,j)$ and $H_y(\alpha,\beta)$, to illustrate Lemma \ref{lemma72}.
Notice that in both figures instead of giving all the coordinates in each vertex, we give the first two coordinates of all the vertices in each row (the third coordinate would specify which partite set the vertex belongs to).
\end{example}
\begin{lemma}\label{lemma72}
Let $x$, $y$ and $n$ be odd. Then:
\[
H_{(xy)}(i,\alpha)(j,\beta)=H_x(i,j)\otimes H_y(\alpha,\beta)
\]
\end{lemma}
\begin{proof}
Notice that
\[
F_n(T_{(xy)}(i,\alpha))=F_n(T_x(i)\otimes T_y(\alpha))=F_n(T_x(i))\otimes F_n(T_y(\alpha))
\]
Notice also that
\[
F_n(T_x(i)\otimes T_y(\alpha))=F_n(T_x(i))\otimes T_y(\alpha)=T_x(i)\otimes F_n(T_y(\alpha))
\]
Then we have
\begin{align*}
H_x(i,j)\otimes H_y(\alpha,\beta)&=\left(T_x(i)\oplus F_n(T_x(i))\oplus F_n(T_x(j))\right) \otimes \left(T_y(\alpha)\oplus F_n(T_y(\alpha))\oplus F_n(T_y(\beta))\right)\\
&=T_x(i)\otimes T_y(\alpha)\oplus F_n(T_x(i))\otimes F_n(T_y(\alpha))\oplus F_n(T_x(j))\otimes F_n(T_y(\beta))\\
&=T_{(xy)}(i,\alpha)\oplus F_n(T_{(xy)}(i,\alpha))\oplus F_n( T_{(xy)}(j,\beta))\\
&=H_{(xy)}(i,\alpha)(j,\beta)
\end{align*}
\end{proof}
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black},scale=.885,every path/.style={->}]
\draw (-5,-7) node [scale=.5](1){};
\draw (-4,-7) node [scale=.5](2){};
\draw (-3,-7) node [scale=.5](3){};
\draw (-5,-8) node[scale=.5] (4){};
\draw (-4,-8) node [scale=.5](5){};
\draw (-3,-8) node [scale=.5](6){};
\draw (-5,-9) node[scale=.5] (7){};
\draw (-4,-9) node [scale=.5](8){};
\draw (-3,-9) node [scale=.5](9){};
\draw (1) to (8);
\draw (8) to (6);
\draw (6) to (1);
\draw (4) [dashed] to (2);
\draw (2) [dashed] to (9);
\draw (9) [dashed] to (4);
\draw (7) [line width=1.5] to (5);
\draw (5) [line width=1.5] to (3);
\draw (3) [bend left=25,line width=1.5] to (7);
\draw (-5,-12) node [scale=.5](a1){};
\draw (-4,-12) node [scale=.5](a2){};
\draw (-3,-12) node [scale=.5](a3){};
\draw (-5,-13) node[scale=.5] (a4){};
\draw (-4,-13) node [scale=.5](a5){};
\draw (-3,-13) node [scale=.5](a6){};
\draw (-5,-14) node[scale=.5] (a7){};
\draw (-4,-14) node [scale=.5](a8){};
\draw (-3,-14) node [scale=.5](a9){};
\draw (-5,-15) node[scale=.5] (a10){};
\draw (-4,-15) node [scale=.5](a11){};
\draw (-3,-15) node [scale=.5](a12){};
\draw (-5,-16) node[scale=.5] (a13){};
\draw (-4,-16) node [scale=.5](a14){};
\draw (-3,-16) node [scale=.5](a15){};
\draw (a1) to (a14);
\draw (a14) to (a9);
\draw (a9) [bend right=11] to (a1);
\draw (a4) [blue] to (a2);
\draw (a2) [blue] to (a12);
\draw (a12) [blue,bend right=11] to (a4);
\draw (a7) [Castano] to (a5);
\draw (a5) [Castano] to (a15);
\draw (a15) [Castano, bend right=11] to (a7);
\draw (a10) [Green] to (a8);
\draw (a8) [Green] to (a3);
\draw (a3) [Green] to (a10);
\draw (a13) [Amarillo] to (a11);
\draw (a11) [Amarillo] to (a6);
\draw (a6) [Amarillo] to (a13);
\node[above=10pt,fill=none,draw=none] at (-4,-7) {$T_3(1)$};
\node[above=10pt,fill=none,draw=none] at (-4,-12) {$T_5(2)$};
\node[left=10pt,fill=none,draw=none] at (-5,-7) {$0$};
\node[left=10pt,fill=none,draw=none] at (-5,-12) {$0$};
\node[left=10pt,fill=none,draw=none] at (-5,-8) {$1$};
\node[left=10pt,fill=none,draw=none] at (-5,-13) {$1$};
\node[left=10pt,fill=none,draw=none] at (-5,-9) {$2$};
\node[left=10pt,fill=none,draw=none] at (-5,-14) {$2$};
\node[left=10pt,fill=none,draw=none] at (-5,-15) {$3$};
\node[left=10pt,fill=none,draw=none] at (-5,-16) {$4$};
\draw (1.5,-6) node [scale=.5](11){};
\draw (4.5,-6) node [scale=.5](22){};
\draw (7.5,-6) node [scale=.5](33){};
\draw (1.5,-7) node [scale=.5](14){};
\draw (4.5,-7) node [scale=.5](25){};
\draw (7.5,-7) node [scale=.5](36){};
\draw (1.5,-8) node [scale=.5](17){};
\draw (4.5,-8) node [scale=.5](28){};
\draw (7.5,-8) node [scale=.5](39){};
\draw (1.5,-9) node [scale=.5](110){};
\draw (4.5,-9) node [scale=.5](211){};
\draw (7.5,-9) node [scale=.5](312){};
\draw (1.5,-10) node [scale=.5](113){};
\draw (4.5,-10) node [scale=.5](214){};
\draw (7.5,-10) node [scale=.5](315){};
\draw (1.5,-12) node [scale=.5](41){};
\draw (4.5,-12) node [scale=.5](52){};
\draw (7.5,-12) node [scale=.5](63){};
\draw (1.5,-13) node [scale=.5](44){};
\draw (4.5,-13) node [scale=.5](55){};
\draw (7.5,-13) node [scale=.5](66){};
\draw (1.5,-14) node [scale=.5](47){};
\draw (4.5,-14) node [scale=.5](58){};
\draw (7.5,-14) node [scale=.5](69){};
\draw (1.5,-15) node [scale=.5](410){};
\draw (4.5,-15) node [scale=.5](511){};
\draw (7.5,-15) node [scale=.5](612){};
\draw (1.5,-16) node [scale=.5](413){};
\draw (4.5,-16) node [scale=.5](514){};
\draw (7.5,-16) node [scale=.5](615){};
\draw (1.5,-18) node [scale=.5](71){};
\draw (4.5,-18) node [scale=.5](82){};
\draw (7.5,-18) node [scale=.5](93){};
\draw (1.5,-19) node [scale=.5](74){};
\draw (4.5,-19) node [scale=.5](85){};
\draw (7.5,-19) node [scale=.5](96){};
\draw (1.5,-20) node [scale=.5](77){};
\draw (4.5,-20) node [scale=.5](88){};
\draw (7.5,-20) node [scale=.5](99){};
\draw (1.5,-21) node [scale=.5](710){};
\draw (4.5,-21) node [scale=.5](811){};
\draw (7.5,-21) node [scale=.5](912){};
\draw (1.5,-22) node [scale=.5](713){};
\draw (4.5,-22) node [scale=.5](814){};
\draw (7.5,-22) node [scale=.5](915){};
\node[above=10pt,fill=none,draw=none] at (4.5,-6) {$T_{(3,5)}(1,2)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-6) {$(0,0)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-7) {$(0,1)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-8) {$(0,2)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-9) {$(0,3)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-10) {$(0,4)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-12) {$(1,0)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-13) {$(1,1)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-14) {$(1,2)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-15) {$(1,3)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-16) {$(1,4)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-18) {$(2,0)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-19) {$(2,1)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-20) {$(2,2)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-21) {$(2,3)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-22) {$(2,4)$};
\draw (11) to (814);
\draw (814) to (69);
\draw (69) to (11);
\draw (14) [blue] to (82);
\draw (82) [blue] to (612);
\draw (612) [blue] to (14);
\draw (17) [Castano] to (85);
\draw (85) [Castano] to (615);
\draw (615) [Castano,bend right=5] to (17);
\draw (110) [Green] to (88);
\draw (88) [Green] to (63);
\draw (63) [Green] to (110);
\draw (113) [Amarillo] to (811);
\draw (811) [Amarillo] to (66);
\draw (66) [Amarillo] to (113);
\draw (41) [dashed] to (214);
\draw (214)[dashed] to (99);
\draw (99) [dashed,bend right=5]to (41);
\draw (44)[dashed,blue] to (22);
\draw (22)[dashed,blue] to (912);
\draw (912)[dashed,blue] to (44);
\draw (47) [dashed,Castano]to (25);
\draw (25) [dashed,Castano]to (915);
\draw (915) [dashed,Castano,bend right=5]to (47);
\draw (410)[dashed,Green] to (28);
\draw (28) [dashed,Green]to (93);
\draw (93) [dashed,Green]to (410);
\draw (413)[dashed,Amarillo] to (211);
\draw (211)[dashed,Amarillo] to (96);
\draw (96) [dashed,Amarillo]to (413);
\draw (71)[line width=1.5] to (514);
\draw (514)[line width=1.5] to (39);
\draw (39) [line width=1.5,bend right=5]to (71);
\draw (74) [line width=1.5,blue]to (52);
\draw (52) [line width=1.5,blue]to (312);
\draw (312) [line width=1.5,blue,bend right=5]to (74);
\draw (77) [line width=1.5,Castano]to (55);
\draw (55) [line width=1.5,Castano]to (315);
\draw (315) [line width=1.5,Castano,bend right=5]to (77);
\draw (710) [line width=1.5,Green]to (58);
\draw (58) [line width=1.5,Green]to (33);
\draw (33) [line width=1.5,Green]to (710);
\draw (713) [line width=1.5,Amarillo]to (511);
\draw (511) [line width=1.5,Amarillo]to (36);
\draw (36) [line width=1.5,Amarillo,bend right=5]to (713);
\end{tikzpicture}
\caption{$T_3(1)$, $T_5(2)$ and $T_{(3,5)}(1,2)$}
\label{figdef71}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={draw,shape=circle,fill=black},scale=.885,every path/.style={->}]
\draw (-5,-7) node [scale=.5](1){};
\draw (-4,-7) node [scale=.5](2){};
\draw (-3,-7) node [scale=.5](3){};
\draw (-5,-8) node[scale=.5] (4){};
\draw (-4,-8) node [scale=.5](5){};
\draw (-3,-8) node [scale=.5](6){};
\draw (-5,-9) node[scale=.5] (7){};
\draw (-4,-9) node [scale=.5](8){};
\draw (-3,-9) node [scale=.5](9){};
\draw (1) to (8);
\draw (8) to (6);
\draw (6) to (7);
\draw (4) to (2);
\draw (2) to (9);
\draw (9) to [bend right=25] (1);
\draw (7) to (5);
\draw (5) to (3);
\draw (3) to (4);
\draw (-5,-12) node [scale=.5](a1){};
\draw (-4,-12) node [scale=.5](a2){};
\draw (-3,-12) node [scale=.5](a3){};
\draw (-5,-13) node[scale=.5] (a4){};
\draw (-4,-13) node [scale=.5](a5){};
\draw (-3,-13) node [scale=.5](a6){};
\draw (-5,-14) node[scale=.5] (a7){};
\draw (-4,-14) node [scale=.5](a8){};
\draw (-3,-14) node [scale=.5](a9){};
\draw (-5,-15) node[scale=.5] (a10){};
\draw (-4,-15) node [scale=.5](a11){};
\draw (-3,-15) node [scale=.5](a12){};
\draw (-5,-16) node[scale=.5] (a13){};
\draw (-4,-16) node [scale=.5](a14){};
\draw (-3,-16) node [scale=.5](a15){};
\draw (a1) to (a14);
\draw (a14) to (a9);
\draw (a9) to (a10);
\draw (a4) to (a2);
\draw (a2) to (a12);
\draw (a12) to (a13);
\draw (a7) to (a5);
\draw (a5) to (a15);
\draw (a15) to [bend right=14] (a1);
\draw (a10) to (a8);
\draw (a8) to (a3);
\draw (a3) to (a4);
\draw (a13) to (a11);
\draw (a11) to (a6);
\draw (a6) to (a7);
\node[above=10pt,fill=none,draw=none] at (-4,-7) {$H_3(1,2)$};
\node[above=10pt,fill=none,draw=none] at (-4,-12) {$H_5(2,4)$};
\node[left=10pt,fill=none,draw=none] at (-5,-7) {$0$};
\node[left=10pt,fill=none,draw=none] at (-5,-12) {$0$};
\node[left=10pt,fill=none,draw=none] at (-5,-8) {$1$};
\node[left=10pt,fill=none,draw=none] at (-5,-13) {$1$};
\node[left=10pt,fill=none,draw=none] at (-5,-9) {$2$};
\node[left=10pt,fill=none,draw=none] at (-5,-14) {$2$};
\node[left=10pt,fill=none,draw=none] at (-5,-15) {$3$};
\node[left=10pt,fill=none,draw=none] at (-5,-16) {$4$};
\draw (1.5,-6) node [scale=.5](11){};
\draw (4.5,-6) node [scale=.5](22){};
\draw (7.5,-6) node [scale=.5](33){};
\draw (1.5,-7) node [scale=.5](14){};
\draw (4.5,-7) node [scale=.5](25){};
\draw (7.5,-7) node [scale=.5](36){};
\draw (1.5,-8) node [scale=.5](17){};
\draw (4.5,-8) node [scale=.5](28){};
\draw (7.5,-8) node [scale=.5](39){};
\draw (1.5,-9) node [scale=.5](110){};
\draw (4.5,-9) node [scale=.5](211){};
\draw (7.5,-9) node [scale=.5](312){};
\draw (1.5,-10) node [scale=.5](113){};
\draw (4.5,-10) node [scale=.5](214){};
\draw (7.5,-10) node [scale=.5](315){};
\draw (1.5,-12) node [scale=.5](41){};
\draw (4.5,-12) node [scale=.5](52){};
\draw (7.5,-12) node [scale=.5](63){};
\draw (1.5,-13) node [scale=.5](44){};
\draw (4.5,-13) node [scale=.5](55){};
\draw (7.5,-13) node [scale=.5](66){};
\draw (1.5,-14) node [scale=.5](47){};
\draw (4.5,-14) node [scale=.5](58){};
\draw (7.5,-14) node [scale=.5](69){};
\draw (1.5,-15) node [scale=.5](410){};
\draw (4.5,-15) node [scale=.5](511){};
\draw (7.5,-15) node [scale=.5](612){};
\draw (1.5,-16) node [scale=.5](413){};
\draw (4.5,-16) node [scale=.5](514){};
\draw (7.5,-16) node [scale=.5](615){};
\draw (1.5,-18) node [scale=.5](71){};
\draw (4.5,-18) node [scale=.5](82){};
\draw (7.5,-18) node [scale=.5](93){};
\draw (1.5,-19) node [scale=.5](74){};
\draw (4.5,-19) node [scale=.5](85){};
\draw (7.5,-19) node [scale=.5](96){};
\draw (1.5,-20) node [scale=.5](77){};
\draw (4.5,-20) node [scale=.5](88){};
\draw (7.5,-20) node [scale=.5](99){};
\draw (1.5,-21) node [scale=.5](710){};
\draw (4.5,-21) node [scale=.5](811){};
\draw (7.5,-21) node [scale=.5](912){};
\draw (1.5,-22) node [scale=.5](713){};
\draw (4.5,-22) node [scale=.5](814){};
\draw (7.5,-22) node [scale=.5](915){};
\node[above=10pt,fill=none,draw=none] at (4.5,-6) {$H_{(3,5)}(1,2)(2,4)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-6) {$(0,0)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-7) {$(0,1)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-8) {$(0,2)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-9) {$(0,3)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-10) {$(0,4)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-12) {$(1,0)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-13) {$(1,1)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-14) {$(1,2)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-15) {$(1,3)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-16) {$(1,4)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-18) {$(2,0)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-19) {$(2,1)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-20) {$(2,2)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-21) {$(2,3)$};
\node[left=10pt,fill=none,draw=none] at (1.5,-22) {$(2,4)$};
\draw (11) to (814);
\draw (814) to (69);
\draw (69) to (710);
\draw (14) to (82);
\draw (82) to (612);
\draw (612) to (713);
\draw (17) to (85);
\draw (85) to (615);
\draw (615) to (71);
\draw (110) to (88);
\draw (88) to (63);
\draw (63) to (74);
\draw (113) to (811);
\draw (811) to (66);
\draw (66) to (77);
\draw (41) to (214);
\draw (214) to (99);
\draw (99) to (110);
\draw (44) to (22);
\draw (22) to (912);
\draw (912) to (113);
\draw (47) to (25);
\draw (25) to (915);
\draw (915) to [bend right=7](11);
\draw (410) to (28);
\draw (28) to (93);
\draw (93) to (14);
\draw (413) to (211);
\draw (211) to (96);
\draw (96) to (17);
\draw (71) to (514);
\draw (514) to (39);
\draw (39) to (410);
\draw (74) to (52);
\draw (52) to (312);
\draw (312) to (413);
\draw (77) to (55);
\draw (55) to (315);
\draw (315) to (41);
\draw (710) to (58);
\draw (58) to (33);
\draw (33) to (44);
\draw (713) to (511);
\draw (511) to (36);
\draw (36) to (47);
\end{tikzpicture}
\caption{$H_3(1,2)$, $H_5(2,4)$ and $H_{(3,5)}(1,2)(2,4)$}
\label{figlemma72}
\end{center}
\end{figure}
\begin{lemma}
Let $\psi$ be a bijection on the set $\{(i,\alpha)|0\leq i \leq x-1, 0\leq \alpha \leq y-1\}$. Then
\[
\overrightarrow{C}_{(xy:n)}=\bigoplus_{(i,\alpha)}H_{(xy)}(i,\alpha)\psi(i,\alpha).\]
\end{lemma}
\begin{proof}
We know that $\overrightarrow{C}_{(xy:n)}=\overrightarrow{C}_{(x:n)}\otimes \overrightarrow{C}_{(y:n)}=\left(\bigoplus_i T_{x}(i)\right)\otimes \left(\bigoplus_{\alpha}T_{y}(\alpha)\right)$. By definition of $T_{(xy)}(i,\alpha)$ we get
\[
\overrightarrow{C}_{(xy:n)}=\bigoplus_{(i,\alpha)}T_{(x,y)}(i,\alpha)\]
We also have
\[
\bigoplus_{(i,\alpha)}T_{(xy)}(i,\alpha)=\bigoplus_{(i,\alpha)}H_{(xy)}(i,\alpha)\psi(i,\alpha)
\]
Combining both we get:
\[
\overrightarrow{C}_{(xy:n)}=\bigoplus_{(i,\alpha)}H_{(xy)}(i,\alpha)\psi(i,\alpha)
\]
as we wanted to prove.
\end{proof}
If $\psi(i,\alpha)=(j,\beta)$ we will denote $\psi_1(i,\alpha)=j$ and $\psi_2(i,\alpha)=\beta$.
If $\gcd(x,i-j)=1$ and $\alpha=\beta$, then $H_{(xy)}(i,\alpha)(j,\beta)$ is a $\overrightarrow{C}_{xn}$-factor. This is because
\[
H_{(xy)}(i,\alpha)(j,\beta)=H_x(i,j)\otimes H_y(\alpha,\alpha)=H_x(i,j)\otimes T_y(\alpha)
\]
By Lemma \ref{HisforHamilton} $H_x(i,j)$ is a $\overrightarrow{C}_{xn}$-factor. By Lemma \ref{bbb} $T_y(\alpha)$ is a $\overrightarrow{C}_n$-factor. Then by Lemma \ref{productofbalanced} $H_x(i,j)\otimes T_y(\alpha)$ is a $\overrightarrow{C}_{xn}$-factor.
Thus to obtain a decomposition of $\overrightarrow{C}_{(xy:n)}$ into $\overrightarrow{C}_{xn}$-factors and $\overrightarrow{C}_{yn}$-factors we need a bijection $\psi$ that satisfies the following set of conditions
\begin{conditions}\label{doubleCheese}
\begin{enumerate}[a)]
\item For all $(i,\alpha)$, $\gcd(x,i-\psi_1(i,\alpha))=1$ and $\psi_2(i,\alpha)=\alpha$, or
\item $\gcd(y,\alpha-\psi_2(i,\alpha))=1$ and $\psi_1(i,\alpha)=i$.
\end{enumerate}
\end{conditions}
\begin{lemma}\label{xnyn}
Let $x$, $y$, and $n$ be odd. Let $s_p\neq 1,xy-1$. Then there is a decomposition of $\overrightarrow{C}_{(xy:n)}$ into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p=xy-s_p$ $\overrightarrow{C}_{yn}$-factors.
\end{lemma}
\begin{proof}
We will describe a bijection $\psi$ that satisfies conditions \ref{doubleCheese} with $r_p$ pairs $(i,\alpha)$ that satisfy $i=\psi_1(i,\alpha)$.
Let $r_\alpha$, $0\leq \alpha\leq y-1$ be such that:
\begin{itemize}
\item $\sum_\alpha r_\alpha=r_p$,
\item $r_i\geq r_j$ if $i\leq j$,
\item $r_{0}=r_{1}$,
\item $0\leq r_{\alpha}\leq x$, $r_\alpha\neq x-1$.
\end{itemize}
Define the function $\phi_\alpha (i)=\phi_s(i)$, with $\phi_s(i)$ as the phi-function over the set $\{0,1,\ldots, x-1\}$ with $s=x-r_\alpha$.
Let $\pi(i)=\left|\{\alpha|\phi_\alpha(i)=i\}\right|$.
Let $\sigma_i(\alpha)=\phi_s(\alpha)$, with $\phi_s(\alpha)$ as the phi-function over the set $\{0,1,\ldots, y-1\}$ with $s=\pi(i)$.
Notice that
\[
\psi(i,\alpha)=(\phi_{\alpha}(i),\sigma_i(\alpha))
\]
is a function satisfying conditions \ref{doubleCheese} because if $\alpha\leq \pi(i)$, then $\psi_{\alpha}(i)=i$ and $\gcd(y,\alpha-\sigma_i(\alpha))=1$. If, on the other hand $\alpha \geq \pi(i)$, then we have $\sigma_i(\alpha)=\alpha$, and $\gcd(x,i-\psi_{\alpha}(i))=1$. Finally, notice that there are $r_p$ pairs $(i,\alpha)$ that satisfy $i=\varphi_1(i,\alpha)$. Therefore there is a decomposition of $\overrightarrow{C}_{(xy:n)}$ into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p=xy-s_p$ $\overrightarrow{C}_{yn}$-factors.
\end{proof}
We can work with $\Gamma(i)$ and $\Lambda(i)$ in a similar fashion as to what we did with $T_{x}(i)$.
\begin{definition}
Let $x$ be odd. We define $T_{(2x)}(i,\alpha)$ to be the directed subgraph of $\overrightarrow{C}_{(4x:n)}$ obtained by taking $T_{(xy)}(i,\alpha)=T_x(i)\otimes \Gamma(\alpha)$.
We also define $H_{(2x)}(i,\alpha)(j,\beta)=T_{(2x)}(i,\alpha)\oplus F_n(T_{(2x)}(i,\alpha))\oplus F_n(T_{(2x)}(j,\beta))$. This is the directed graph obtained by taking the arcs of
$T_{(2x)}(i,\alpha)$ between parts
$t$ and $t+1$ for $0\leq t\leq n-2$, and the arcs between parts $n-1$ and $0$ from $T_{(2x)}(j,\beta)$.
\end{definition}
Now we can apply the same techniques that we did to $T_{(xy)}$ and $H_{(xy)}$.
\begin{lemma}
Let $x$ and $n$ be odd. Then:
\[
H_{(2x)}(i,\alpha)(j,\beta)=H_x(i,j)\otimes \Lambda(\alpha,\beta)
\]
\end{lemma}
\begin{proof}
Notice that
\[
F_n(T_{(2x)}(i,\alpha))=F_n(T_x(i)\otimes \Gamma(\alpha))=F_n(T_x(i))\otimes F_n(\Gamma(\alpha))
\]
and
\[
F_n(T_x(i)\otimes \Gamma(\alpha))=F_n(T_x(i))\otimes \Gamma(\alpha)=T_x(i)\otimes F_n(\Gamma(\alpha))
\]
Using this the result is trivial.
\end{proof}
\begin{lemma}
Let $\varphi$ by a bijection on the set $\{(i,\alpha)|0\leq i \leq x-1, 0\leq \alpha \leq 3\}$. Then
\[
\overrightarrow{C}_{(4x:n)}=\bigoplus_{(i,\alpha)}H_{(2x)}(i,\alpha)\varphi(i,\alpha).\]
\end{lemma}
\begin{proof}
We know that $\overrightarrow{C}_{(4x:n)}=\overrightarrow{C}_{(x:n)}\otimes \overrightarrow{C}_{(4:n)}=\left(\bigoplus_i T_x(i)\right)\otimes \left(\bigoplus_{\alpha}\Gamma(\alpha)\right)$. By definition of $T_{(2x)}(i,\alpha)$ we get
\[
\overrightarrow{C}_{(4x:n)}=\bigoplus_{(i,\alpha)}T_{(2x)}(i,\alpha)\]
We also have
\[
\bigoplus_{(i,\alpha)}T_{(2x)}(i,\alpha)=\bigoplus_{(i,\alpha)}H_{(2x)}(i,\alpha)\varphi(i,\alpha)
\]
Combining both we get:
\[
\overrightarrow{C}_{(4x:n)}=\bigoplus_{(i,\alpha)}H_{(2x)}(i,\alpha)\varphi(i,\alpha)
\]
as we wanted to prove.
\end{proof}
Next we develop the conditions needed for our decompositions. Recall that if
$\varphi(i,\alpha)=(j,\beta)$ we will denote $\varphi_1(i,\alpha)=j$ and $\varphi_2(i,\alpha)=\beta$.
\begin{itemize}
\item If $\alpha\neq \beta$ and $\gcd(x,i-j)=1$ then
\[
H_{(2x)}(i,\alpha)(j,\beta)=H_x(i,j)\otimes \Lambda(\alpha,\beta)
\]
is a $\overrightarrow{C}_{2xn}$-factor by Lemmas \ref{HisforHamilton}, \ref{LisforLambda}, and \ref{productofbalanced}.
\item If $i=j$ and $\alpha\neq \beta$, then $H_{(2x)}(i,\alpha)(j,\beta)$ is a $\overrightarrow{C}_{2n}$-factor by Lemmas \ref{bbb}, \ref{LisforLambda}, and \ref{productofbalanced}.
\item If $\alpha=\beta$ and $\gcd(x,i-j)=1$, then $H_{(2x)}(i,\alpha)(j,\beta)$ is a $\overrightarrow{C}_{xn}$-factor by Lemmas \ref{HisforHamilton}, \ref{GisforGamma}, and \ref{productofbalanced}.
\item If $i=j$ and $\alpha=\beta$, then $H_{(xy)}(i,\alpha)(j,\beta)$ is a $\overrightarrow{C}_{n}$-factor by Lemmas \ref{bbb}, \ref{GisforGamma}, and \ref{productofbalanced}.
\end{itemize}
So for a decomposition of $\overrightarrow{C}_{(4x:n)}$ into $\overrightarrow{C}_{2xn}$-factors and $\overrightarrow{C}_{n}$-factors we need a bijection $\varphi$ that satisfies:
\begin{conditions}\label{2HamandTrig}
For all $(i,\alpha)$ such that $\varphi(i,\alpha)\neq (i,\alpha)$, $\alpha\neq \varphi_2(i,\alpha)$ and $\gcd(x,i-\varphi_1(i,\alpha))=1$.
\end{conditions}
For a decomposition of $\overrightarrow{C}_{(4x:n)}$ into $\overrightarrow{C}_{2n}$-factors and $\overrightarrow{C}_{xn}$-factors we need a bijection $\varphi$ that satisfies:
\begin{conditions}\label{2doubleCheese}
\begin{enumerate}[a)]
\item For all $(i,\alpha)$ either $\alpha\neq \varphi_2(i,\alpha)$ and $i=j$, or
\item $\alpha=\varphi_2(i,\alpha)$ and $\gcd(x,i-\varphi_1(i,\alpha))= 1$.
\end{enumerate}
\end{conditions}
We define a new family of functions $\theta_s:\mathbb{Z}_x\times\mathbb{Z}_4\rightarrow \mathbb{Z}_x\times\mathbb{Z}_4$. These functions will be referred as \textit{theta-functions}.
Let $s\in \{0,1,\ldots,4x\}$, $s\notin \{1,4x-1\}$, and write $s=4k+2a+3b$, with $a,b\in\{0,1\}$, $k\leq x$. We define:
\begin{center}
\textbf{Theta-functions}:
\[
\theta_s(i,\alpha)\coloneqq\left\lbrace\begin{array}{ccc}
(i+1,\alpha+1)&\text{ if }& 0\leq i \leq k-1, \alpha= 0,2\\
(i-1,\alpha-1)&\text{ if }& 1\leq i \leq k, \alpha= 1,3 \\
(i+1,\alpha+1)&\text{ if }& i=k,\, a=1,\, \alpha=0\\
(i-1,\alpha-1)&\text{ if }& i=k+1,\, a=1,\, \alpha=1\\
(i+1,\alpha+1)&\text{ if }& i=k,\, b=1,\, \alpha=2\\
(i+1,\alpha-2)&\text{ if }& i=k+1,\, b=1,\, \alpha=3\\
(i-2,\alpha+1)&\text{ if }& i=k+2,\, b=1,\, \alpha=1\\
(i,\alpha)& &\text{ otherwise}\\
\end{array}\right.
\]
\end{center}
If $s=4x-1=4(x-1)+3$, we define $\theta_s$ in a similar way, with a small change:
\[
\theta_{4x-1}(i,\alpha)\coloneqq\left\lbrace\begin{array}{ccc}
(i+1,\alpha+1)&\text{ if }& 0\leq i \leq x-2, \alpha= 0,2 \\
(i-1,\alpha-1)&\text{ if }& 1\leq i \leq x-2, \alpha= 1,3 \\
(x-2,2)&\text{ if }& i=x-1,\, \alpha=1\\
(0,1)&\text{ if }& i=x-1,\, \alpha=3\\
(x-1,0)&\text{ if }& i=0,\, \alpha=1\\
(0,3)&\text{ if }& i=x-1,\, \alpha=0\\
(x-2,0)&\text{ if }& i=0,\, \alpha=3\\
(i,\alpha)& &\text{ otherwise}\\
\end{array}\right.
\]
We give a visual example of $\theta_9$ and $\theta_{19}$, for $x=5$ in Figure \ref{thetaexample}.
\begin{figure}
\begin{center}
\begin{tikzpicture}[every node/.style={shape=circle},outer sep=0pt, inner sep=0pt]
\draw (0,0) node [scale=.7](00){$(0,0)$};
\draw (1,5) node [scale=.7](01){$(0,1)$};
\draw (2,0) node [scale=.7](02){$(0,2)$};
\draw (3,5) node [scale=.7](03){$(0,3)$};
\draw (0,1) node [scale=.7](10){$(1,0)$};
\draw (1,1) node [scale=.7](11){$(1,1)$};
\draw (2,1) node [scale=.7](12){$(1,2)$};
\draw (3,1) node [scale=.7](13){$(1,3)$};
\draw (0,2) node [scale=.7](20){$(2,0)$};
\draw (1,2) node [scale=.7](21){$(2,1)$};
\draw (2,2) node [scale=.7](22){$(2,2)$};
\draw (3,2) node [scale=.7](23){$(2,3)$};
\draw (0,3) node [scale=.7](30){$(3,0)$};
\draw (1,3) node [scale=.7](31){$(3,1)$};
\draw (2,3) node [scale=.7](32){$(3,2)$};
\draw (3,3) node [scale=.7](33){$(3,3)$};
\draw (0,4) node [scale=.7](40){$(4,0)$};
\draw (1,4) node [scale=.7](41){$(4,1)$};
\draw (2,4) node [scale=.7](42){$(4,2)$};
\draw (3,4) node [scale=.7](43){$(4,3)$};
\draw (00) [->,thick] to (11);
\draw (11) [->,thick,bend right=15] to (00);
\draw (02) [->,thick] to (13);
\draw (13) [->,thick,bend right=15] to (02);
\draw (10) [->,thick] to (21);
\draw (21) [->,thick,bend right=15] to (10);
\draw (12) [->,thick] to (23);
\draw (23) [->,thick,bend right=10] to (31);
\draw (31) [->,thick] to (12);
\draw (22) [->,out=90,in=135,looseness=4] to (22);
\draw (30) [->,out=90,in=135,looseness=4] to (30);
\draw (32) [->,out=90,in=135,looseness=4] to (32);
\draw (33) [->,out=90,in=135,looseness=4] to (33);
\draw (40) [->,out=90,in=135,looseness=4] to (40);
\draw (41) [->,out=90,in=135,looseness=4] to (41);
\draw (42) [->,out=90,in=135,looseness=4] to (42);
\draw (43) [->,out=90,in=135,looseness=4] to (43);
\draw (01) [->,out=90,in=135,looseness=4] to (01);
\draw (03) [->,out=90,in=135,looseness=4] to (03);
\draw (20) [->,out=90,in=135,looseness=4] to (20);
\draw (8,0) node [scale=.7](100){$(0,0)$};
\draw (9,5) node [scale=.7](101){$(0,1)$};
\draw (10,0) node [scale=.7](102){$(0,2)$};
\draw (11,5) node [scale=.7](103){$(0,3)$};
\draw (8,1) node [scale=.7](110){$(1,0)$};
\draw (9,1) node [scale=.7](111){$(1,1)$};
\draw (10,1) node [scale=.7](112){$(1,2)$};
\draw (11,1) node [scale=.7](113){$(1,3)$};
\draw (8,2) node [scale=.7](120){$(2,0)$};
\draw (9,2) node [scale=.7](121){$(2,1)$};
\draw (10,2) node [scale=.7](122){$(2,2)$};
\draw (11,2) node [scale=.7](123){$(2,3)$};
\draw (8,3) node [scale=.7](130){$(3,0)$};
\draw (9,3) node [scale=.7](131){$(3,1)$};
\draw (10,3) node [scale=.7](132){$(3,2)$};
\draw (11,3) node [scale=.7](133){$(3,3)$};
\draw (8,4) node [scale=.7](140){$(4,0)$};
\draw (9,4) node [scale=.7](141){$(4,1)$};
\draw (10,4) node [scale=.7](142){$(4,2)$};
\draw (11,4) node [scale=.7](143){$(4,3)$};
\draw (100) [->,thick] to (111);
\draw (111) [->,thick,bend right=15] to (100);
\draw (102) [->,thick] to (113);
\draw (113) [->,thick,bend right=15] to (102);
\draw (110) [->,thick] to (121);
\draw (121) [->,thick,bend right=15] to (110);
\draw (112) [->,thick] to (123);
\draw (123) [->,thick,bend right=15] to (112);
\draw (120) [->,thick] to (131);
\draw (131) [->,thick,bend right=15] to (120);
\draw (122) [->,thick] to (133);
\draw (133) [->,thick,bend right=15] to (122);
\draw (120) [->,thick] to (131);
\draw (131) [->,thick,bend right=15] to (120);
\draw (122) [->,thick] to (133);
\draw (133) [->,thick,bend right=15] to (122);
\draw (130) [->,thick] to (141);
\draw (141) [->,thick] to (132);
\draw (132) [->,thick] to (143);
\draw (143) [->,thick] to (101);
\draw (101) [->,thick] to (140);
\draw (140) [->,thick] to (103);
\draw (103) [->,thick,bend right=35,dotted] to (130);
\draw (142) [->,out=90,in=135,looseness=4] to (142);
\node[below=10pt,fill=none,draw=none] at (1.5,2) (aaaa){$x=5$, $s=9$, $k=1,a=1,b=1$};
\node[right=190pt,fill=none,draw=none] at (aaaa) {$x=5$, $s=19$};
\end{tikzpicture}
\vspace*{-1in}
\caption{Example of Theta-functions}
\label{thetaexample}
\end{center}
\end{figure}
The following lemma is a generalization of a result given in $\cite{AKKPO}$.
\begin{lemma}\label{2xnandntheorem}
Let $s_p\in \{0,2,\ldots,4x-1,4x\}$, $x$ odd. Then there exists a decomposition of $C_{(4x:n)}$ into $s_p$ $C_{2xn}$-factors and $r_p=4x-s_p$ $C_n$-factors.
\end{lemma}
\begin{proof}
The bijection $\psi=\theta_{s_p}$ satisfies Conditions $7.9$. In particular if $\psi(i,\alpha)\neq (i,\alpha)$, then $\alpha\neq \psi_{2}(i,\alpha)$ and $i-\psi_1 (i,\alpha)\in\{\pm 1, \pm 2\}$; and as $x$ is odd, $\gcd(x,i-\psi_1(i,\alpha))=1$. Furthermore, $\psi$ has $s_p$ non-fixed points.
Therefore there exists a decomposition of $C_{(4x:n)}$ into $s_p$ $C_{2xn}$-factors and $4x-s_p$ $C_n$-factors.
\end{proof}
\begin{lemma}\label{xnand2ntheorem}
Let $s_p\in \{0,2,3,\ldots,4x-3,4x-2,4x\}$. Then there exists a decomposition of $\overrightarrow{C}_{(4x:n)}$ into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p=4x-s_p$ $\overrightarrow{C}_{2n}$-factors.
\end{lemma}
\begin{proof}
We provide a bijection $\varphi$ that satisfies Conditions \ref{2doubleCheese} with $r_p$ pairs $(i,\alpha)$ that satisfy $i=\varphi_1(i,\alpha)$.
Let $r_\alpha$, $0\leq \alpha\leq 3$ be such that:
\begin{itemize}
\item $\sum_\alpha r_\alpha=r_p$,
\item $r_i\geq r_j$ if $i\leq j$,
\item $r_0=r_1$,
\item $r_\alpha\leq x$, $r_\alpha\neq x-1$.
\end{itemize}
The only case where such a choice of $r_\alpha$ cannot be made is when $x=3$, $s=7$. This case is covered in Lemma \ref{theissuewith12} in the Appendix.
Define the function $\psi_\alpha (i)=\phi_s(i)$, with $\phi_s(i)$ as the phi-functions over the set $\{0,1,\ldots, x-1\}$ with $s=x-r_\alpha$.
Let $\pi(i)=\left|\{\alpha|\psi_\alpha(i)=i\}\right|$.
Let $\sigma_i(\alpha)$ be the permutation on the set $\{0,1,2,3\}$ that cyclically permutes the first $\pi(i)$ elements and fixes the rest.
Notice that
\[
\varphi(i,\alpha)=(\psi_{\alpha}(i),\sigma_i(\alpha))
\]
is a function satisfying conditions \ref{2doubleCheese} because if $\alpha\leq \pi(i)$, then $\psi_{\alpha}(i)=i$ and $\sigma_i(\alpha)\neq \alpha$. If, on the other hand $\alpha \geq \pi(i)$, then we have $\sigma_i(\alpha)=\alpha$, and $\gcd(x,i-\psi_{\alpha}(i))=1$. Finally, notice that there are $r_p$ pairs $(i,\alpha)$ that satisfy $i=\varphi_1(i,\alpha)$.
Therefore there exists a decomposition of $\overrightarrow{C}_{(4x:n)}$ into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p=4x-s_p$ $\overrightarrow{C}_{2n}$-factors.
\end{proof}
We are interested in one more type of decomposition, into $\overrightarrow{C}_{2xn}$ and $\overrightarrow{C}_{yn}$ factors. To do this we introduce the following:
\begin{definition}
Let $x$ and $y$ be odd. Define $T_{(2xy)}(i,\alpha,\gamma)$ to be the directed subgraph of $\overrightarrow{C}_{(4xy:n)}$ obtained by taking $T_{(2xy)}(i,\alpha,\gamma)=T_{(2x)}(i,\alpha)\otimes T_y(\gamma)$. We also define
\[
H_{(2xy)}(i,\alpha,\gamma)(j,\beta,\delta)=T_{(2xy)}(i,\alpha,\gamma)\oplus F_n(T_{(2xy)}(i,\alpha,\gamma))\oplus F_n(T_{(2xy)}(j,\beta,\delta)\]
This means that $H_{(2xy)}(i,\alpha,\gamma)(j,\beta,\delta)$ is the directed graph obtained by taking the arcs of $T_{(2xy)}(i,\alpha,\gamma)$ between parts $t$ and $t+1$ for $0\leq t\leq n-2$, and the arcs between parts $n-1$ and $0$ from $T_{(2xy)}(j,\beta,\delta)$.
\end{definition}
Now we have all the usual results:
\begin{lemma}
Let $x$, $y$ and $n$ be odd. Then:
\[
H_{(2xy)}(i,\alpha,\gamma)(j,\beta,\delta)=H_{(2x)}(i,\alpha)(j,\beta)\otimes H_y(\gamma,\delta)
\]
\end{lemma}
\begin{proof}
Notice that
\[
F_n(T_{(2xy)}(i,\alpha,\gamma))=F_n(T_{(2x)}(i,\alpha)\otimes T_y(\gamma))=F_n(T_{(2x)}(i,\alpha))\otimes F_n(T_y(\gamma))
\]
Using this the result is trivial.
\end{proof}
\begin{lemma}
Let $\varphi$ be a bijection on the set $\{(i,\alpha,\gamma)|0\leq i \leq x-1, 0\leq \alpha \leq 3,0\leq \gamma \leq y-1\}$. Then
\[
\overrightarrow{C}_{(4xy:n)}=\bigoplus_{(i,\alpha,\gamma)}H_{(2xy)}(i,\alpha,\gamma)\varphi(i,\alpha,\gamma).\]
\end{lemma}
\begin{proof}
We know that $\overrightarrow{C}_{(4xy:n)}=\overrightarrow{C}_{(4x:n)}\otimes \overrightarrow{C}_{(y:n)}=\left(\bigoplus_{(i,\alpha)} T_{(2x)}(i,\alpha)\right)\otimes \left(\bigoplus_{\gamma}T_{y}(\gamma)\right)$. By the definition of $T_{(2xy)}(i,\alpha,\gamma)$ we get
\[
\overrightarrow{C}_{(4xy:n)}=\bigoplus_{(i,\alpha,\gamma)}T_{(2xy)}(i,\alpha)\]
We also have
\[
\bigoplus_{(i,\alpha,\gamma)}T_{(2xy)}(i,\alpha,\gamma)=\bigoplus_{(i,\alpha,\gamma)}H_{(2xy)}(i,\alpha,\gamma)\varphi(i,\alpha,\gamma)
\]
Combining both we get:
\[
\overrightarrow{C}_{(4xy:n)}=\bigoplus_{(i,\alpha,\gamma)}H_{(2xy)}(i,\alpha,\gamma)\varphi(i,\alpha,\gamma)
\]
as we wanted to prove.
\end{proof}
We have the following properties:
\begin{itemize}
\item If $\alpha\neq \beta$, $\gamma=\delta$, and $\gcd(x,i-j)=1$, then
\begin{align*}
H_{(2xy)}(i,\alpha,\gamma)(j,\beta,\delta)&=H_{(2x)}(i,\alpha)(j,\beta)\otimes H_{y}(\gamma,\gamma)\\
&=H_x(i,j)\otimes \Lambda(\alpha,\beta) \otimes T_{y}(\gamma)
\end{align*}
is a $\overrightarrow{C}_{2xn}$-factor by Lemmas \ref{HisforHamilton}, \ref{LisforLambda}, \ref{bbb} and \ref{productofbalanced}.
\item If $i=j$, $\alpha=\beta$, and $\gcd(y,\gamma-\delta)=1$ then
\begin{align*}
H_{(2xy)}(i,\alpha,\gamma)(j,\beta,\delta)&=H_{(2x)}(i,\alpha)(i,\alpha)\otimes H_{y}(\gamma,\delta)\\
&=T_x(i)\otimes \Gamma(\alpha) \otimes H_{y}(\gamma,\delta)
\end{align*}
is a $\overrightarrow{C}_{yn}$-factor by Lemmas \ref{bbb}, \ref{GisforGamma}, \ref{HisforHamilton}, and \ref{productofbalanced}.
\end{itemize}
To get a decomposition of $\overrightarrow{C}_{(4xy:n)}$ into $\overrightarrow{C}_{2xn}$-factors and $\overrightarrow{C}_{yn}$-factors we need a bijection $\varphi$ that satisfies:
\begin{conditions}\label{2xnandyn}
\begin{enumerate}[a)]
\item For all $(i,\alpha,\gamma)$, $\gcd(y,\gamma-\varphi_3(i,\alpha,\gamma))=1$ or $\gamma=\varphi_3(i,\alpha,\gamma)$.
\item If $\gamma=\varphi_3(i,\alpha,\gamma)$, then $\gcd(x,i-\varphi_{1}(i,\alpha,\gamma))=1$ and $\alpha\neq \varphi_{2}(i,\alpha,\gamma)$.
\item If $\gcd(y,\gamma-\varphi_3(i,\alpha,\gamma))=1$, then $i=\varphi_1(i,\alpha,\gamma)$ and $\alpha=\varphi_2(i,\alpha,\gamma)$.
\end{enumerate}
\end{conditions}
Now we can write our lemma:
\begin{lemma}\label{2xnyn}
Let $x,y$, and $n$ be odd. Let $s_p\neq 1, 4xy-1$. Then there is a decomposition of $C_{(4xy:n)}$ into $s_p$ $C_{2xn}$-factors and $r_p=4xy-s_p$ $C_{yn}$-factors.
\end{lemma}
\begin{proof}
We give a bijection $\varphi$ that satisfies Conditions $7.16$ with $r_p$ elements $(i,\alpha,\gamma)$ that satisfy $i=\varphi_1(i,\alpha,\gamma)$.
Let $s_p=4xk+q$, with $0\leq q \leq 4x-1$. We have two cases, $k\leq y-3$, and $k\geq y-2$.
\begin{description}
\item [Case 1]
If $k\leq y-3$, let $s_p=4xk+a-\epsilon$, with $2\leq a \leq 4x-1$, $0\leq \epsilon \leq 2$. For $y-k+1\leq \gamma \leq y-1$, let $\psi_{\gamma}(i,\alpha)=\theta_{4x}(i,\alpha)$, the theta-function with $s=4x$. Let $\psi_{y-k}(i,\alpha)=\theta_{4x-\epsilon}(i,\alpha)$, the theta-function with $s=4x-\epsilon$. Let $\psi_{y-k-1}(i,\alpha)=\theta_{a}(i,\alpha)$, the theta function with $s=a$. For $0\leq \gamma \leq y-k-2$ let $\psi_{\gamma}(i,\alpha)=(i,\alpha)$, the identity function.
\item [Case 2]
If $k\geq y-2$, let $s_p=4xk'+2a-\epsilon$, with $2\leq a \leq 4x-\epsilon$, $0\leq \epsilon \leq 4$, where $k'\in\{y-2,y-3\}$ because $2a$ may be greater than $4x$. For $y-k+1\leq \gamma \leq y-1$, let $\psi_{\gamma}(i,\alpha)=\theta_{4x}(i,\alpha)$, the theta-function with $s=4x$. Let $\psi_{y-k}(i,\alpha)=\theta_{4x-\epsilon}(i,\alpha)$, the theta-function with $s=4x-\epsilon$.
Let $\psi_{y-k-1}(i,\alpha)=\psi_{y-k-2}(i,\alpha)=\theta_{a}(i,\alpha)$, the theta function with $s=a$. For $0\leq \gamma \leq y-k-3$ let $\psi_{\gamma}(i,\alpha)=(i,\alpha)$, the identity function.
\end{description}
Notice that the fixed point of $\theta_{4x-1}$ is $(x-1,2)$, the fixed points of $\theta_{4x-2}$ are $\{(x-1,2),(0,3)\}$, the fixed points of $\theta_{4x-3}$ are $\{(x-1,2),(0,3),(x-1,0)\}$, and the fixed points of $\theta_{4x-4}$ are $\{(x-1,2),(0,3),(x-1,0),(0,1)\}$. This means that if $0\leq \epsilon\leq 4$ and $a\leq 4x-\epsilon$, the fixed points of $\theta_{4x-\epsilon}$ are a subset of the fixed points of $\theta_{a}$. Hence if $\psi_{\delta}(i,\alpha)=(i,\alpha)$, then $\psi_{\gamma}(i,\alpha)=(i,\alpha)$ for all $\gamma\leq \delta$. Notice also that $\psi_{0}(i,\alpha)=\psi_{1}(i,\alpha)$, hence $\max\{\delta \in \{0,\ldots,y-1\}|\psi_{\delta}(i,\alpha)=(i,\alpha)\}\neq 1$. Therefore we can define $\sigma_{i,\alpha}(\gamma)=\phi_s(\gamma)$, the phi-function over the set $\{0,1,\ldots,y-1\}$ with $s=\max\{\delta \in \{0,\ldots,y-1\}|\psi_{\delta}(i,\alpha)=(i,\alpha)\}$. Then:
\[
\rho(i,\alpha,\gamma)=(\psi_{\gamma}(i,\alpha),\sigma_{i,\alpha}(\gamma))
\]
is a function satisfying conditions $7.16$.
\end{proof}
\begin{example}
We provide two visual examples, with $x=3$ and $y=5$. To make the picture easier to understand, the points satisfying $i=\varphi_1(i,\alpha,\gamma)$ have been boxed and underlined, and rearranged with their images at the right side.
In Figure \ref{717examplea} we have $s_p=25=2\cdot 12+1=2\cdot 12+2 -1$, giving us $r_p=60-25=35$, $k=2$, $a=2$, $\epsilon=1$.
In Figure \ref{717exampleb} we have $s_p=37=3\cdot 12+1=3\cdot 12+2\cdot2-3$, giving us $r_p=60-37=23$, $k'=3$, $a=2$, $\epsilon=3$.
\end{example}
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=.6,every node/.style={shape=circle},outer sep=0pt, inner sep=0pt]
\draw (0,0) node [scale=.7,shape=rectangle,draw,inner sep=2pt](000){$\underline{(0,0,0)}$};
\draw (2,3) node [scale=.7,shape=rectangle,draw,inner sep=2pt](001){$\underline{(0,1,0)}$};
\draw (4,0) node [scale=.7,shape=rectangle,draw,inner sep=2pt](002){$\underline{(0,2,0)}$};
\draw (6,3) node [scale=.7,shape=rectangle,draw,inner sep=2pt](003){$\underline{(0,3,0)}$};
\draw (0,1) node [scale=.7,shape=rectangle,draw,inner sep=2pt](010){$\underline{(1,0,0)}$};
\draw (2,1) node [scale=.7,shape=rectangle,draw,inner sep=2pt](011){$\underline{(1,1,0)}$};
\draw (4,1) node [scale=.7,shape=rectangle,draw,inner sep=2pt](012){$\underline{(1,2,0)}$};
\draw (6,1) node [scale=.7,shape=rectangle,draw,inner sep=2pt](013){$\underline{(1,3,0)}$};
\draw (0,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](020){$\underline{(2,0,0)}$};
\draw (2,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](021){$\underline{(2,1,0)}$};
\draw (4,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](022){$\underline{(2,2,0)}$};
\draw (6,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](023){$\underline{(2,3,0)}$};
\node[left=50pt,fill=none,draw=none] at (0,2) (aaaa){$\gamma=0$};
\node[left=50pt,fill=none,draw=none] at (0,1) (aaaa){$s=0$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,-.25) -- (-2,3.25)node [black,midway,xshift=9pt]{};
\draw (0,5) node [scale=.7,shape=rectangle,draw,inner sep=2pt](100){$\underline{(0,0,1)}$};
\draw (2,8) node [scale=.7,shape=rectangle,draw,inner sep=2pt](101){$\underline{(0,1,1)}$};
\draw (4,5) node [scale=.7,shape=rectangle,draw,inner sep=2pt](102){$\underline{(0,2,1)}$};
\draw (6,8) node [scale=.7,shape=rectangle,draw,inner sep=2pt](103){$\underline{(0,3,1)}$};
\draw (0,6) node [scale=.7,shape=rectangle,draw,inner sep=2pt](110){$\underline{(1,0,1)}$};
\draw (2,6) node [scale=.7,shape=rectangle,draw,inner sep=2pt](111){$\underline{(1,1,1)}$};
\draw (4,6) node [scale=.7,shape=rectangle,draw,inner sep=2pt](112){$\underline{(1,2,1)}$};
\draw (6,6) node [scale=.7,shape=rectangle,draw,inner sep=2pt](113){$\underline{(1,3,1)}$};
\draw (0,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](120){$\underline{(2,0,1)}$};
\draw (2,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](121){$\underline{(2,1,1)}$};
\draw (4,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](122){$\underline{(2,2,1)}$};
\draw (6,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](123){$\underline{(2,3,1)}$};
\node[left=50pt,fill=none,draw=none] at (0,7) (aaaa){$\gamma=1$};
\node[left=50pt,fill=none,draw=none] at (0,6) (aaaa){$s=0$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,4.75) -- (-2,8.25)node [black,midway,xshift=9pt]{};
\draw (0,10) node [scale=.7](200){$(0,0,2)$};
\draw (2,13) node [scale=.7,shape=rectangle,draw,inner sep=2pt](201){$\underline{(0,1,2)}$};
\draw (4,10) node [scale=.7,shape=rectangle,draw,inner sep=2pt](202){$\underline{(0,2,2)}$};
\draw (6,13) node [scale=.7,shape=rectangle,draw,inner sep=2pt](203){$\underline{(0,3,2)}$};
\draw (0,11) node [scale=.7,shape=rectangle,draw,inner sep=2pt](210){$\underline{(1,0,2)}$};
\draw (2,11) node [scale=.7](211){$(1,1,2)$};
\draw (4,11) node [scale=.7,shape=rectangle,draw,inner sep=2pt](212){$\underline{(1,2,2)}$};
\draw (6,11) node [scale=.7,shape=rectangle,draw,inner sep=2pt](213){$\underline{(1,3,2)}$};
\draw (0,12) node [scale=.7,shape=rectangle,draw,inner sep=2pt](220){$\underline{(2,0,2)}$};
\draw (2,12) node [scale=.7,shape=rectangle,draw,inner sep=2pt](221){$\underline{(2,1,2)}$};
\draw (4,12) node [scale=.7,shape=rectangle,draw,inner sep=2pt](222){$\underline{(2,2,2)}$};
\draw (6,12) node [scale=.7,shape=rectangle,draw,inner sep=2pt](223){$\underline{(2,3,2)}$};
\node[left=50pt,fill=none,draw=none] at (0,12) (aaaa){$\gamma=2$};
\node[left=50pt,fill=none,draw=none] at (0,11) (aaaa){$s=2$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,9.75) -- (-2,13.25)node [black,midway,xshift=9pt]{};
\draw (0,15) node [scale=.7](300){$(0,0,3)$};
\draw (2,18) node [scale=.7](301){$(0,1,3)$};
\draw (4,15) node [scale=.7](302){$(0,2,3)$};
\draw (6,18) node [scale=.7](303){$(0,3,3)$};
\draw (0,16) node [scale=.7](310){$(1,0,3)$};
\draw (2,16) node [scale=.7](311){$(1,1,3)$};
\draw (4,16) node [scale=.7](312){$(1,2,3)$};
\draw (6,16) node [scale=.7](313){$(1,3,3)$};
\draw (0,17) node [scale=.7](320){$(2,0,3)$};
\draw (2,17) node [scale=.7](321){$(2,1,3)$};
\draw (4,17) node [scale=.7,shape=rectangle,draw,inner sep=2pt](322){$\underline{(2,2,3)}$};
\draw (6,17) node [scale=.7](323){$(2,3,3)$};
\node[left=50pt,fill=none,draw=none] at (0,17) (aaaa){$\gamma=3$};
\node[left=50pt,fill=none,draw=none] at (0,16) (aaaa){$s=12$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,14.75) -- (-2,18.25)node [black,midway,xshift=9pt]{};
\draw (0,20) node [scale=.7](400){$(0,0,4)$};
\draw (2,23) node [scale=.7](401){$(0,1,4)$};
\draw (4,20) node [scale=.7](402){$(0,2,4)$};
\draw (6,23) node [scale=.7](403){$(0,3,4)$};
\draw (0,21) node [scale=.7](410){$(1,0,4)$};
\draw (2,21) node [scale=.7](411){$(1,1,4)$};
\draw (4,21) node [scale=.7](412){$(1,2,4)$};
\draw (6,21) node [scale=.7](413){$(1,3,4)$};
\draw (0,22) node [scale=.7](420){$(2,0,4)$};
\draw (2,22) node [scale=.7](421){$(2,1,4)$};
\draw (4,22) node [scale=.7](422){$(2,2,4)$};
\draw (6,22) node [scale=.7](423){$(2,3,4)$};
\node[left=50pt,fill=none,draw=none] at (0,22) (aaaa){$\gamma=4$};
\node[left=50pt,fill=none,draw=none] at (0,21) (aaaa){$s=12$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,19.75) -- (-2,23.25)node [black,midway,xshift=9pt]{};
\draw (400) [->] to (411);
\draw (411) [->,bend right=15] to (400);
\draw (402) [->] to (413);
\draw (413) [->,bend right=15] to (402);
\draw (410) [->] to (421);
\draw (421) [->,bend right=15] to (410);
\draw (412) [->] to (423);
\draw (423) [->,bend right=15] to (412);
\draw (420) [->] to (401);
\draw (401) [->,bend right=15] to (420);
\draw (422) [->] to (403);
\draw (403) [->,bend right=15] to (422);
\draw (420) [->] to (401);
\draw (401) [->,bend right=15] to (420);
\draw (422) [->] to (403);
\draw (403) [->,bend right=15] to (422);
\draw (300) [->] to (311);
\draw (311) [->,bend right=15] to (300);
\draw (302) [->] to (313);
\draw (313) [->,bend right=15] to (302);
\draw (310) [->] to (321);
\draw (321) [->] to (312);
\draw (312) [->] to (323);
\draw (323) [->,out=150,in=350,looseness=0] to (301);
\draw (301) [->] to (320);
\draw (320) [->] to (303);
\draw (303) [->,out=135, in=150, looseness=1,dotted] to (310);
\draw (200) [->] to (211);
\draw (211) [->,bend left=15] to (200);
\draw (12,14) node [scale=.7](a220){$\underline{(2,0,2)}$};
\draw (12,13) node [scale=.7](a120){$\underline{(2,0,1)}$};
\draw (12,12) node [scale=.7](a020){$\underline{(2,0,0)}$};
\draw (12,9) node [scale=.7](a210){$\underline{(1,0,2)}$};
\draw (12,8) node [scale=.7](a110){$\underline{(1,0,1)}$};
\draw (12,7) node [scale=.7](a010){$\underline{(1,0,0)}$};
\draw (12,3) node [scale=.7](a100){$\underline{(0,0,1)}$};
\draw (12,2) node [scale=.7](a000){$\underline{(0,0,0)}$};
\draw (15,14) node [scale=.7](a201){$\underline{(0,1,2)}$};
\draw (15,13) node [scale=.7](a101){$\underline{(0,1,1)}$};
\draw (15,12) node [scale=.7](a001){$\underline{(0,1,0)}$};
\draw (15,9) node [scale=.7](a221){$\underline{(2,1,2)}$};
\draw (15,8) node [scale=.7](a121){$\underline{(2,1,1)}$};
\draw (15,7) node [scale=.7](a021){$\underline{(2,1,0)}$};
\draw (15,3) node [scale=.7](a111){$\underline{(1,1,1)}$};
\draw (15,2) node [scale=.7](a011){$\underline{(1,1,0)}$};
\draw (18,15) node [scale=.7](a322){$\underline{(2,2,3)}$};
\draw (18,14) node [scale=.7](a222){$\underline{(2,2,2)}$};
\draw (18,13) node [scale=.7](a122){$\underline{(2,2,1)}$};
\draw (18,12) node [scale=.7](a022){$\underline{(2,2,0)}$};
\draw (18,9) node [scale=.7](a212){$\underline{(1,2,2)}$};
\draw (18,8) node [scale=.7](a112){$\underline{(1,2,1)}$};
\draw (18,7) node [scale=.7](a012){$\underline{(1,2,0)}$};
\draw (18,4) node [scale=.7](a202){$\underline{(0,2,2)}$};
\draw (18,3) node [scale=.7](a102){$\underline{(0,2,1)}$};
\draw (18,2) node [scale=.7](a002){$\underline{(0,2,0)}$};
\draw (21,14) node [scale=.7](a203){$\underline{(0,3,2)}$};
\draw (21,13) node [scale=.7](a103){$\underline{(0,3,1)}$};
\draw (21,12) node [scale=.7](a003){$\underline{(0,3,0)}$};
\draw (21,9) node [scale=.7](a223){$\underline{(2,3,2)}$};
\draw (21,8) node [scale=.7](a123){$\underline{(2,3,1)}$};
\draw (21,7) node [scale=.7](a023){$\underline{(2,3,0)}$};
\draw (21,4) node [scale=.7](a213){$\underline{(1,3,2)}$};
\draw (21,3) node [scale=.7](a113){$\underline{(1,3,1)}$};
\draw (21,2) node [scale=.7](a013){$\underline{(1,3,0)}$};
\draw (a020) [->,out=170,in=190, looseness=1,thick] to (a120);
\draw (a120) [->,out=170,in=190, looseness=1,thick] to (a220);
\draw (a220) [->,out=170,in=190, looseness=1,thick] to (a020);
\draw (a010) [->,out=170,in=190, looseness=1,thick] to (a110);
\draw (a110) [->,out=170,in=190, looseness=1,thick] to (a210);
\draw (a210) [->,out=170,in=190, looseness=1,thick] to (a010);
\draw (a000) [->,out=170,in=190, looseness=1,thick] to (a100);
\draw (a100) [->,out=170,in=190, looseness=1,thick] to (a000);
\draw (a021) [->,out=170,in=190, looseness=1,thick] to (a121);
\draw (a121) [->,out=170,in=190, looseness=1,thick] to (a221);
\draw (a221) [->,out=170,in=190, looseness=1,thick] to (a021);
\draw (a011) [->,out=170,in=190, looseness=1,thick] to (a111);
\draw (a111) [->,out=170,in=190, looseness=1,thick] to (a011);
\draw (a001) [->,out=170,in=190, looseness=1,thick] to (a101);
\draw (a101) [->,out=170,in=190, looseness=1,thick] to (a201);
\draw (a201) [->,out=170,in=190, looseness=1,thick] to (a001);
\draw (a012) [->,out=170,in=190, looseness=1,thick] to (a112);
\draw (a112) [->,out=170,in=190, looseness=1,thick] to (a212);
\draw (a212) [->,out=170,in=190, looseness=1,thick] to (a012);
\draw (a002) [->,out=170,in=190, looseness=1,thick] to (a102);
\draw (a102) [->,out=170,in=190, looseness=1,thick] to (a202);
\draw (a202) [->,out=170,in=190, looseness=1,thick] to (a002);
\draw (a023) [->,out=170,in=190, looseness=1,thick] to (a123);
\draw (a123) [->,out=170,in=190, looseness=1,thick] to (a223);
\draw (a223) [->,out=170,in=190, looseness=1,thick] to (a023);
\draw (a013) [->,out=170,in=190, looseness=1,thick] to (a113);
\draw (a113) [->,out=170,in=190, looseness=1,thick] to (a213);
\draw (a213) [->,out=170,in=190, looseness=1,thick] to (a013);
\draw (a003) [->,out=170,in=190, looseness=1,thick] to (a103);
\draw (a103) [->,out=170,in=190, looseness=1,thick] to (a203);
\draw (a203) [->,out=170,in=190, looseness=1,thick] to (a003);
\draw (a322) [->,out=170,in=190, looseness=1,thick] to (a222);
\draw (a222) [->,out=170,in=190, looseness=1,thick] to (a322);
\draw (a122) [->,out=170,in=190, looseness=1,thick] to (a022);
\draw (a022) [->,out=170,in=190, looseness=1,thick] to (a122);
\node[below=10pt,fill=none,draw=none] at (9,5.5) (aaaa){$x=3$, $y=5$, $s_p=25$, $r_p=35$, $k=2$, $a=2$, $\epsilon=1$};
\end{tikzpicture}
\vspace*{-1in}
\caption{Example of Lemma $7.17$ Case 1}
\label{717examplea}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=.6,every node/.style={shape=circle},outer sep=0pt, inner sep=0pt]
\draw (0,0) node [scale=.7](000){$(0,0,0)$};
\draw (2,3) node [scale=.7,shape=rectangle,draw,inner sep=2pt](001){$\underline{(0,1,0)}$};
\draw (4,0) node [scale=.7,shape=rectangle,draw,inner sep=2pt](002){$\underline{(0,2,0)}$};
\draw (6,3) node [scale=.7,shape=rectangle,draw,inner sep=2pt](003){$\underline{(0,3,0)}$};
\draw (0,1) node [scale=.7,shape=rectangle,draw,inner sep=2pt](010){$\underline{(1,0,0)}$};
\draw (2,1) node [scale=.7](011){$(1,1,0)$};
\draw (4,1) node [scale=.7,shape=rectangle,draw,inner sep=2pt](012){$\underline{(1,2,0)}$};
\draw (6,1) node [scale=.7,shape=rectangle,draw,inner sep=2pt](013){$\underline{(1,3,0)}$};
\draw (0,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](020){$\underline{(2,0,0)}$};
\draw (2,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](021){$\underline{(2,1,0)}$};
\draw (4,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](022){$\underline{(2,2,0)}$};
\draw (6,2) node [scale=.7,shape=rectangle,draw,inner sep=2pt](023){$\underline{(2,3,0)}$};
\node[left=50pt,fill=none,draw=none] at (0,2) (aaaa){$\gamma=0$};
\node[left=50pt,fill=none,draw=none] at (0,1) (aaaa){$s=2$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,-.25) -- (-2,3.25)node [black,midway,xshift=9pt]{};
\draw (0,5) node [scale=.7](100){$(0,0,1)$};
\draw (2,8) node [scale=.7,shape=rectangle,draw,inner sep=2pt](101){$\underline{(0,1,1)}$};
\draw (4,5) node [scale=.7,shape=rectangle,draw,inner sep=2pt](102){$\underline{(0,2,1)}$};
\draw (6,8) node [scale=.7,shape=rectangle,draw,inner sep=2pt](103){$\underline{(0,3,1)}$};
\draw (0,6) node [scale=.7,shape=rectangle,draw,inner sep=2pt](110){$\underline{(1,0,1)}$};
\draw (2,6) node [scale=.7](111){$(1,1,1)$};
\draw (4,6) node [scale=.7,shape=rectangle,draw,inner sep=2pt](112){$\underline{(1,2,1)}$};
\draw (6,6) node [scale=.7,shape=rectangle,draw,inner sep=2pt](113){$\underline{(1,3,1)}$};
\draw (0,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](120){$\underline{(2,0,1)}$};
\draw (2,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](121){$\underline{(2,1,1)}$};
\draw (4,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](122){$\underline{(2,2,1)}$};
\draw (6,7) node [scale=.7,shape=rectangle,draw,inner sep=2pt](123){$\underline{(2,3,1)}$};
\node[left=50pt,fill=none,draw=none] at (0,7) (aaaa){$\gamma=1$};
\node[left=50pt,fill=none,draw=none] at (0,6) (aaaa){$s=2$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,4.75) -- (-2,8.25)node [black,midway,xshift=9pt]{};
\draw (0,10) node [scale=.7](200){$(0,0,2)$};
\draw (2,13) node [scale=.7](201){$(0,1,2)$};
\draw (4,10) node [scale=.7](202){$(0,2,2)$};
\draw (6,13) node [scale=.7,shape=rectangle,draw,inner sep=2pt](203){$\underline{(0,3,2)}$};
\draw (0,11) node [scale=.7](210){$(1,0,2)$};
\draw (2,11) node [scale=.7](211){$(1,1,2)$};
\draw (4,11) node [scale=.7](212){$(1,2,2)$};
\draw (6,11) node [scale=.7](213){$(1,3,2)$};
\draw (0,12) node [scale=.7,shape=rectangle,draw,inner sep=2pt](220){$\underline{(2,0,2)}$};
\draw (2,12) node [scale=.7](221){$(2,1,2)$};
\draw (4,12) node [scale=.7,shape=rectangle,draw,inner sep=2pt](222){$\underline{(2,2,2)}$};
\draw (6,12) node [scale=.7](223){$(2,3,2)$};
\node[left=50pt,fill=none,draw=none] at (0,12) (aaaa){$\gamma=2$};
\node[left=50pt,fill=none,draw=none] at (0,11) (aaaa){$s=9$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,9.75) -- (-2,13.25)node [black,midway,xshift=9pt]{};
\draw (0,15) node [scale=.7](300){$(0,0,3)$};
\draw (2,18) node [scale=.7](301){$(0,1,3)$};
\draw (4,15) node [scale=.7](302){$(0,2,3)$};
\draw (6,18) node [scale=.7](303){$(0,3,3)$};
\draw (0,16) node [scale=.7](310){$(1,0,3)$};
\draw (2,16) node [scale=.7](311){$(1,1,3)$};
\draw (4,16) node [scale=.7](312){$(1,2,3)$};
\draw (6,16) node [scale=.7](313){$(1,3,3)$};
\draw (0,17) node [scale=.7](320){$(2,0,3)$};
\draw (2,17) node [scale=.7](321){$(2,1,3)$};
\draw (4,17) node [scale=.7](322){$(2,2,3)$};
\draw (6,17) node [scale=.7](323){$(2,3,3)$};
\node[left=50pt,fill=none,draw=none] at (0,17) (aaaa){$\gamma=3$};
\node[left=50pt,fill=none,draw=none] at (0,16) (aaaa){$s=12$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,14.75) -- (-2,18.25)node [black,midway,xshift=9pt]{};
\draw (0,20) node [scale=.7](400){$(0,0,4)$};
\draw (2,23) node [scale=.7](401){$(0,1,4)$};
\draw (4,20) node [scale=.7](402){$(0,2,4)$};
\draw (6,23) node [scale=.7](403){$(0,3,4)$};
\draw (0,21) node [scale=.7](410){$(1,0,4)$};
\draw (2,21) node [scale=.7](411){$(1,1,4)$};
\draw (4,21) node [scale=.7](412){$(1,2,4)$};
\draw (6,21) node [scale=.7](413){$(1,3,4)$};
\draw (0,22) node [scale=.7](420){$(2,0,4)$};
\draw (2,22) node [scale=.7](421){$(2,1,4)$};
\draw (4,22) node [scale=.7](422){$(2,2,4)$};
\draw (6,22) node [scale=.7](423){$(2,3,4)$};
\node[left=50pt,fill=none,draw=none] at (0,22) (aaaa){$\gamma=4$};
\node[left=50pt,fill=none,draw=none] at (0,21) (aaaa){$s=12$};
\draw [decorate,decoration={brace,amplitude=10pt},xshift=-4pt,yshift=0pt]
(-2,19.75) -- (-2,23.25)node [black,midway,xshift=9pt]{};
\draw (400) [->] to (411);
\draw (411) [->,bend right=15] to (400);
\draw (402) [->] to (413);
\draw (413) [->,bend right=15] to (402);
\draw (410) [->] to (421);
\draw (421) [->,bend right=15] to (410);
\draw (412) [->] to (423);
\draw (423) [->,bend right=15] to (412);
\draw (420) [->] to (401);
\draw (401) [->,bend right=15] to (420);
\draw (422) [->] to (403);
\draw (403) [->,bend right=15] to (422);
\draw (420) [->] to (401);
\draw (401) [->,bend right=15] to (420);
\draw (422) [->] to (403);
\draw (403) [->,bend right=15] to (422);
\draw (300) [->] to (311);
\draw (311) [->,bend right=15] to (300);
\draw (302) [->] to (313);
\draw (313) [->,bend right=15] to (302);
\draw (310) [->] to (321);
\draw (321) [->,bend right=15] to (310);
\draw (312) [->] to (323);
\draw (323) [->,bend right=15] to (312);
\draw (320) [->] to (301);
\draw (301) [->,bend right=15] to (320);
\draw (322) [->] to (303);
\draw (303) [->,bend right=15] to (322);
\draw (320) [->] to (301);
\draw (301) [->,bend right=15] to (320);
\draw (322) [->] to (303);
\draw (303) [->,bend right=15] to (322);
\draw (200) [->] to (211);
\draw (211) [->,bend right=15] to (200);
\draw (202) [->] to (213);
\draw (213) [->,bend right=15] to (202);
\draw (210) [->] to (221);
\draw (221) [->,bend left=15] to (210);
\draw (212) [->] to (223);
\draw (201) [->,out=330,in=150] to (212);
\draw (223) [->,out=150,in=350,looseness=0] to (201);
\draw (100) [->] to (111);
\draw (111) [->,bend left=15] to (100);
\draw (000) [->] to (011);
\draw (011) [->,bend left=15] to (000);
\draw (12,14) node [scale=.7](a220){$\underline{(2,0,2)}$};
\draw (12,13) node [scale=.7](a120){$\underline{(2,0,1)}$};
\draw (12,12) node [scale=.7](a020){$\underline{(2,0,0)}$};
\draw (12,8) node [scale=.7](a110){$\underline{(1,0,1)}$};
\draw (12,7) node [scale=.7](a010){$\underline{(1,0,0)}$};
\draw (15,13) node [scale=.7](a101){$\underline{(0,1,1)}$};
\draw (15,12) node [scale=.7](a001){$\underline{(0,1,0)}$};
\draw (15,8) node [scale=.7](a121){$\underline{(2,1,1)}$};
\draw (15,7) node [scale=.7](a021){$\underline{(2,1,0)}$};
\draw (18,14) node [scale=.7](a222){$\underline{(2,2,2)}$};
\draw (18,13) node [scale=.7](a122){$\underline{(2,2,1)}$};
\draw (18,12) node [scale=.7](a022){$\underline{(2,2,0)}$};
\draw (18,8) node [scale=.7](a112){$\underline{(1,2,1)}$};
\draw (18,7) node [scale=.7](a012){$\underline{(1,2,0)}$};
\draw (18,3) node [scale=.7](a102){$\underline{(0,2,1)}$};
\draw (18,2) node [scale=.7](a002){$\underline{(0,2,0)}$};
\draw (21,14) node [scale=.7](a203){$\underline{(0,3,2)}$};
\draw (21,13) node [scale=.7](a103){$\underline{(0,3,1)}$};
\draw (21,12) node [scale=.7](a003){$\underline{(0,3,0)}$};
\draw (21,8) node [scale=.7](a123){$\underline{(2,3,1)}$};
\draw (21,7) node [scale=.7](a023){$\underline{(2,3,0)}$};
\draw (21,3) node [scale=.7](a113){$\underline{(1,3,1)}$};
\draw (21,2) node [scale=.7](a013){$\underline{(1,3,0)}$};
\draw (a020) [->,out=170,in=190, looseness=1,thick] to (a120);
\draw (a120) [->,out=170,in=190, looseness=1,thick] to (a220);
\draw (a220) [->,out=170,in=190, looseness=1,thick] to (a020);
\draw (a010) [->,out=170,in=190, looseness=1,thick] to (a110);
\draw (a110) [->,out=170,in=190, looseness=1,thick] to (a010);
\draw (a021) [->,out=170,in=190, looseness=1,thick] to (a121);
\draw (a121) [->,out=170,in=190, looseness=1,thick] to (a021);
\draw (a001) [->,out=170,in=190, looseness=1,thick] to (a101);
\draw (a101) [->,out=170,in=190, looseness=1,thick] to (a001);
\draw (a012) [->,out=170,in=190, looseness=1,thick] to (a112);
\draw (a112) [->,out=170,in=190, looseness=1,thick] to (a012);
\draw (a002) [->,out=170,in=190, looseness=1,thick] to (a102);
\draw (a102) [->,out=170,in=190, looseness=1,thick] to (a002);
\draw (a023) [->,out=170,in=190, looseness=1,thick] to (a123);
\draw (a123) [->,out=170,in=190, looseness=1,thick] to (a023);
\draw (a013) [->,out=170,in=190, looseness=1,thick] to (a113);
\draw (a113) [->,out=170,in=190, looseness=1,thick] to (a013);
\draw (a003) [->,out=170,in=190, looseness=1,thick] to (a103);
\draw (a103) [->,out=170,in=190, looseness=1,thick] to (a203);
\draw (a203) [->,out=170,in=190, looseness=1,thick] to (a003);
\draw (a222) [->,out=170,in=190, looseness=1,thick] to (a022);
\draw (a122) [->,out=170,in=190, looseness=1,thick] to (a222);
\draw (a022) [->,out=170,in=190, looseness=1,thick] to (a122);
\node[below=10pt,fill=none,draw=none] at (9,5.5) (aaaa){$x=3$, $y=5$, $s_p=37$, $r_p=23$, $k=4$, $a=2$, $\epsilon=3$, $k'=3$};
\end{tikzpicture}
\vspace*{-1in}
\caption{Example of Lemma $7.17$ Case 2}
\label{717exampleb}
\end{center}
\end{figure}
\section{Product}
In this section the partite product will be applied to the decompositions obtained in the previous section, to obtain decompositions of larger graphs.
\begin{lemma}\label{zwlemma}
Let $m=zw$ with $z$ odd and $w\not\in\{2,6\}$. Then there is a decomposition of $\overrightarrow{C}_{(m:n)}$ into $\overrightarrow{C}_{zn}$-factors and a decomposition of $\overrightarrow{C}_{(4m:n)}$ into $\overrightarrow{C}_{2zn}$-factors.
\end{lemma}
\begin{proof}
We know that $\overrightarrow{C}_{(m:n)}=\overrightarrow{C}_{(z:n)}\otimes \overrightarrow{C}_{(w:n)}$.
By Theorem \ref{1varfun} we can decompose $\overrightarrow{C}_{(z:n)}=\bigoplus_i H_z(i,\phi(i))$ into $\overrightarrow{C}_{zn}$-factors, and by Lemma \ref{bbb} $\overrightarrow{C}_{(w:n)}=\bigoplus_j T_w(j,j)$ into $\overrightarrow{C}_n$-factors.
Then
\begin{align*}
\overrightarrow{C}_{(m:n)}&=\overrightarrow{C}_{(z:n)}\otimes \overrightarrow{C}_{(w:n)}\\
&=\left(\bigoplus_i H_z(i,\phi(i))\right)\otimes\left(\bigoplus_j H_w(j,j)\right)\\
&=\bigoplus_i \bigoplus_j H_z(i,\phi(i))\otimes H_w(j,j)
\end{align*}
By Lemma \ref{productofbalanced} $H_z(i,\phi(i))\otimes H_w(j,j)$ is a $\overrightarrow{C}_{zn}$-factor, and so $\overrightarrow{C}_{(m:n)}$ can be decomposed into $\overrightarrow{C}_{zn}$-factors.
For the result on $\overrightarrow{C}_{(4m:n)}$ we just multiply by $\overrightarrow{C}_{(4:n)}$:
\[
\overrightarrow{C}_{(4m:n)}=\overrightarrow{C}_{(m:n)}\otimes \overrightarrow{C}_{(4:n)}
\]
We can decompose $\overrightarrow{C}_{(4:n)}$ into $\overrightarrow{C}_{2n}$-factors by Theorem \ref{even} and so when multiplying by $\overrightarrow{C}_{(m:n)}$ we obtain a decomposition of $\overrightarrow{C}_{(4m:n)}$ into $\overrightarrow{C}_{2zn}$-factors.
\end{proof}
We may now use Lemma \ref{zwlemma} and the decompositions obtained in Section 7 to get a decomposition of $\overrightarrow{C}_{(xyzw:n)}$ into $\overrightarrow{C}_{xzn}$-factors and $\overrightarrow{C}_{yzn}$-factors:
\begin{lemma}\label{productcontruction}
Suppose $\overrightarrow{C}_{(m:n)}$ can be decomposed into $s_p$ $\overrightarrow{C}_{m_1n}$-factors and $r_p$ $\overrightarrow{C}_{m_2n}$ factors, with $s_p,r_p\neq 1$, $r_p+s_p=m$. Let $z$ be odd with $\gcd(m_1,z)=\gcd(m_2,z)=1$ and $w\not\in\{2,6\}$. Then there is a decomposition of $\overrightarrow{C}_{(mzw:n)}$ into $s$ $\overrightarrow{C}_{m_1zn}$-factors and $r$ $\overrightarrow{C}_{m_2zn}$-factors for any $s,r\neq 1$, $s+r=mzw$. Furthermore, if $m_1$ and $m_2$ are odd, there is a decomposition of $\overrightarrow{C}_{(4mzw:n)}$ into $s'$ $\overrightarrow{C}_{2m_1zn}$-factors and $r'$ $\overrightarrow{C}_{2m_2zn}$-factors for any $s',r'\neq 1$, $s'+r'=4mzw$.
\end{lemma}
\begin{proof}
We start with the product $\overrightarrow{C}_{(mzw:n)}=\overrightarrow{C}_{(m:n)}\otimes \overrightarrow{C}_{(zw:n)}$.
By Lemma \ref{zwlemma} we can decompose $\overrightarrow{C}_{(zw:n)}=\oplus_{i=1}^{zw} Z_i$, where each $Z_i$ is a $\overrightarrow{C}_{zn}$-factor.
Let $s=mt+u$, with $0\leq u \leq m$. If $u\neq 1,m-1$ we decompose as follows:
\begin{align*}
\overrightarrow{C}_{(mzw:n)}&=\overrightarrow{C}_{(m:n)}\otimes \overrightarrow{C}_{(zw:n)}\\
&=\overrightarrow{C}_{(m:n)}\otimes \left(\oplus_{i=1}^{zw} Z_i\right)\\
&= \oplus_{i=1}^{zw}\overrightarrow{C}_{(m:n)}\otimes Z_i\\
&= \left(\oplus_{i=1}^{t}\overrightarrow{C}_{(m:n)}\otimes Z_i\right)\oplus \overrightarrow{C}_{(m:n)}\otimes Z_{t+1}\oplus\left(\oplus_{i=t+2}^{zw}\overrightarrow{C}_{(m:n)}\otimes Z_i\right)
\end{align*}
From the theorem hypothesis on $\overrightarrow{C}_{(m:n)}$, we have the following decompositions:
\begin{itemize}
\item
We decompose $\overrightarrow{C}_{(m:n)}$ into $m$ $\overrightarrow{C}_{m_1n}$-factors for the product $\oplus_{i=1}^{t}\overrightarrow{C}_{(m:n)}\otimes Z_i$.
\item
We decompose $\overrightarrow{C}_{(m:n)}$ into $m$ $\overrightarrow{C}_{m_2n}$-factors for the product $\oplus_{i=t+2}^{zw}C_{(m:n)}\otimes Z_i$.
\item
We decompose $\overrightarrow{C}_{(m:n)}$ into $u$ $\overrightarrow{C}_{m_1n}$-factors and $m-u$ $\overrightarrow{C}_{m_2n}$-factors for the product $\overrightarrow{C}_{(m:n)}\otimes Z_{t+1}$.
\end{itemize}
Because $m_1$ and $m_2$ are coprime with $z$, by Lemma \ref{productofbalanced} there is a decomposition into $mt+u=s$ $\overrightarrow{C}_{m_1zn}$-factors and $r$ $\overrightarrow{C}_{m_2zn}$-factors.
If $u=1$, we decompose as follows:
\begin{align*}
\overrightarrow{C}_{(mzw:n)}=&\left(\oplus_{i=1}^{t-1}\overrightarrow{C}_{(m:n)}\otimes Z_i\right)\oplus \left(\overrightarrow{C}_{(m:n)}\otimes Z_{t}\right)\oplus \\
&\left(\overrightarrow{C}_{(m:n)}\otimes Z_{t+1}\right)\oplus\left(\oplus_{i=t+2}^{zw}\overrightarrow{C}_{(m:n)}\otimes Z_i\right)
\end{align*}
We also have the following decompositions:
\begin{itemize}
\item
We decompose $\overrightarrow{C}_{(m:n)}$ into $m$ $\overrightarrow{C}_{m_1n}$-factors for the product $\oplus_{i=1}^{t-1}\overrightarrow{C}_{(m:n)}\otimes Z_i$.
\item
We decompose $\overrightarrow{C}_{(m:n)}$ into $m$ $\overrightarrow{C}_{m_2n}$-factors for the product $\oplus_{i=t+2}^{zw}\overrightarrow{C}_{(m:n)}\otimes Z_i$.
\item
We decompose $\overrightarrow{C}_{(m:n)}$ into $m-2$ $\overrightarrow{C}_{m_1n}$-factors and $2$ $\overrightarrow{C}_{m_2n}$-factors for the product $\overrightarrow{C}_{(m:n)}\otimes Z_{t}$.
\item We decompose $\overrightarrow{C}_{(m:n)}$ into $3$ $\overrightarrow{C}_{m_1n}$-factors and $m-3$ $\overrightarrow{C}_{m_2n}$-factors for the product $\overrightarrow{C}_{(m:n)}\otimes Z_{t+1}$.
\end{itemize}
Because $m_1$ and $m_2$ are coprime with $z$, this gives a decomposition into $m(t-1)+m-2+3=mt+1=s$ $\overrightarrow{C}_{m_1zn}$-factors and $r$ $\overrightarrow{C}_{m_2zn}$-factors.
If $u=m-1$ we just change the roles of $m_1$ and $m_2$ and take the decomposition for $u=1$.
Therefore there is a decomposition of $\overrightarrow{C}_{(mzw:n)}$ into $s$ $\overrightarrow{C}_{m_1zn}$ factors and $r$ $\overrightarrow{C}_{m_2zn}$-factors for any $s,r\neq 1$, $r+s=mzw$.
The decomposition of $\overrightarrow{C}_{(4mzw:n)}$ into $s'$ $\overrightarrow{C}_{2m_1zn}$-factors and $r'$ $\overrightarrow{C}_{2m_2zn}$-factors works in the same way.
\end{proof}
We can now combine this result with our decompositions from Section 7 to obtain the following result, which we write using non-directed graphs, as we are getting ready to apply the results from Section 5.
\begin{theorem}\label{cvresult}
Let $x,y,z,n$ be odd numbers with $\gcd(x,z)=\gcd(y,z)=1$ and $w\not\in \{2,6\}$. Then we have the following decompositions:
\begin{enumerate}[a)]
\item $C_{(xyzw:n)}$ can be decomposed into $s$ $C_{xzn}$-factors and $r$ $C_{yzn}$-factors for any $s,r\neq 1$, $s+r=xyzw$.
\item $C_{(4xzw:n)}$ can be decomposed into $s$ $C_{2xzn}$-factors and $r$ $C_{2zn}$-factors for any $s,r\neq 1$, $s+r=4xzw$.
\item $C_{(4xzw:n)}$ can be decomposed into $s$ $C_{2xzn}$-factors and $r$ $C_{zn}$-factors for any $s,r\neq 1$, $s+r=4xzw$.
\item $C_{(4xzw:n)}$ can be decomposed into $s$ $C_{xzn}$-factors and $r$ $C_{2zn}$-factors for any $s,r\neq 1$, $s+r=4xzw$.
\item $C_{(4xyzw:n)}$ can be decomposed into $s$ $C_{2xzn}$-factors and $r$ $C_{yzn}$-factors for any $s,r\neq 1$, $s+r=4xyzw$.
\item $C_{(4xyzw:n)}$ can be decomposed into $s$ $C_{2xzn}$-factors and $r$ $C_{2yzn}$-factors for any $s,r\neq 1$, $s+r=4xyzw$.
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}[a)]
\item $\overrightarrow{C}_{(xy:n)}$ can be decomposed into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p$ $\overrightarrow{C}_{yn}$-factors by Lemma \ref{xnyn}. So by Lemma \ref{productcontruction}; $\overrightarrow{C}_{(xywz:n)}$ can be decomposed into $s$ $\overrightarrow{C}_{xzn}$-factors and $r$ $\overrightarrow{C}_{yzn}$-factors.
\item $\overrightarrow{C}_{(x:n)}$ can be decomposed into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p$ $\overrightarrow{C}_{n}$-factors by Lemma \ref{1varfun}. Now apply Lemma \ref{productcontruction}.
\item $\overrightarrow{C}_{(4x:n)}$ can be decomposed into $s_p$ $\overrightarrow{C}_{2xn}$-factors and $r_p$ $\overrightarrow{C}_{n}$-factors by Lemma \ref{2xnandntheorem}. Now apply Lemma \ref{productcontruction}.
\item $\overrightarrow{C}_{(4x:n)}$ can be decomposed into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p$ $\overrightarrow{C}_{2n}$-factors by Lemma \ref{xnand2ntheorem}. Now apply Lemma \ref{productcontruction}.
\item $\overrightarrow{C}_{(4xy:n)}$ can be decomposed into $s_p$ $\overrightarrow{C}_{2xn}$-factors and $r_p$ $\overrightarrow{C}_{yn}$-factors by Lemma \ref{2xnyn}. Now apply Lemma \ref{productcontruction}.
\item $\overrightarrow{C}_{(xy:n)}$ can be decomposed into $s_p$ $\overrightarrow{C}_{xn}$-factors and $r_p$ $\overrightarrow{C}_{yn}$-factors by Lemma \ref{xnyn}. Now apply Lemma \ref{productcontruction}.
\end{enumerate}
\end{proof}
\section{Main Result}
We now to use the decompositions that we obtained for $C_{(v:n)}$ to obtain decompositions of $K_{(v:m)}$ via Lemmas $\ref{cvntokvm}$ and $\ref{buildcompletegraph}$.
\begin{theorem}\label{maintheorem}
Let $m$ and $n$ be odd, such that $m\equiv 0 \pmod n$. Let $s$ and $r$ be such that $s,r\neq 1$ and $s+r=v\frac{m-1}{2}$. Let $x_1,\ldots x_{m/n}$, $y_1,\ldots y_{m/n}$, $z_1,\ldots,z_{m/n}$ and $w_1,\ldots,w_{m/n}$ be such that:
\begin{itemize}
\item $\gcd(x_i,z_i)=\gcd(y_i,z_i)=1$,
\item $w_i\not\in\{2,6\}$,
\item $2$ divides at most one of $x_i,y_i$ and $z_i$,
\item $v=x_iy_iz_iw_i$ if $2$ divides none of $x_i,y_i,z_i$,
\item $v=2x_iy_iz_iw_i$ if $2$ divides one of $x_i,y_i,z_i$.
\end{itemize}
Furthermore, let $F_1$ be a $[(x_1z_1n)^{\frac{v}{x_1z_1}},\ldots,(x_{m/n}z_{m/n}n)^{\frac{v}{x_{m/n}z_{m/n}}}]$-factor, and let $F_2$ be a $[(y_1z_1n)^{\frac{v}{y_1z_1}},\ldots,(y_{m/n}z_{m/n}n)^{\frac{v}{y_{m/n}z_{m/n}}}]$-factor.
Then there is a decomposition of $K_{(v:m)}$ into $s$ copies of $F_1$ and $r$ copies of $F_2$.
\end{theorem}
\begin{proof}
By Theorem \ref{OP} there is a decomposition of $K_m$ into $C_n$-factors.
Pick $s_1,\ldots, s_{\frac{m-1}{2}}$ such that $s=\bigoplus_{i=1}^{v\frac{m-1}{2}}s_i$, with $0\leq s_i \leq v$ and $s_i\not\in \{1,v-1\}$.
By Theorem \ref{cvresult} there is a decomposition of $C_{(v:n)}$ into $s_i$ $C_{x_iz_in}$-factors and $r_i=v-s_i$ $C_{y_iz_in}$-factors.
Therefore by Theorem \ref{cvntokvm} there is a decomposition of $K_{(v:m)}$ into $s$ copies of the 2-factor $F_1$ and $r$ copies of the $2$-factor $F_2$.
\end{proof}
\section{Applications}
We can use our results to solve many cases of the Hamilton-Waterloo Problem for complete graphs.
For some of them we will need the notion of resolvable group divisible design.
A \emph{resolvable group divisible design} $(k,\l)\-\ensuremath{\mbox{\sf RGDD}}(h^u)$ is a triple $(\mathcal{V},\mathcal{G},\mathcal{B})$ where $\mathcal{V}$ is a finite set of size $v=hu$, $\mathcal{G}$ is a partition of $\mathcal{V}$ into $u$ \emph{groups} each containing $h$ elements, and $\mathcal{B}$ is a collection of $k$ element subsets of $\mathcal{V}$ called \emph{blocks} which satisfy the following properties.
\begin{packed_item}
\item If $B\in \mathcal{B}$, then $|B|=k$.
\item If a pair of elements from $\mathcal{V}$ appear in the same group, then the pair cannot be in any block.
\item Two points that are not in the same group, called a \emph{transverse pair}, appear in exactly $\l$ blocks.
\item $|\mathcal{G}|>1$.
\item The blocks can be partitioned into parallel classes such that for each element of $\mathcal{V}$ there is exactly one block in each parallel class containing it.
\end{packed_item}
Here we use the term group to indicate an element of $\mathcal{G}$. In this context, group simply means a set of elements without any algebraic structure.
If $\l=1$, we refer to the RGDD as a $k\-\ensuremath{\mbox{\sf RGDD}}(h^u)$.
In \cite{R} the following characterization theorem was proven:
\begin{theorem}{\normalfont\cite{R}}\label{3rgdd}
A $(3,\l)\-\ensuremath{\mbox{\sf RGDD}}(h^u)$ exists if and only if $u\geq 3,\l h(u-1)$ is even, $hu\equiv 0\pmod{3}$, and $(\l,h,u)\not\in\{(1,2,6),(1,6,3) \}\bigcup\{(2j+1,2,3), (4j+2,1,6):j\geq 0\}$.
\end{theorem}
In \cite{AKKPO} the authors used resolvable group divisible designs together with Theorem \ref{3rgdd} to decompose complete graphs into $C_3$-factors and $C_{3x}$-factors.
\begin{lemma}{\normalfont\cite{AKKPO}}
\label{main}
Let $x \geq 3$, $y \geq 3$ and $m$ be positive integers such that both $x$ and $y$ divide $3m$. Suppose the following conditions are satisfied:
\begin{packed_item}
\item There exists a 3-$\ensuremath{\mbox{\sf RGDD}}(h^{u})$,
\item there exists a decomposition of $K_{(m:3)}$ into $s_{p}$ $C_{x}$-factors and $r_{p}$ $C_{y}$-factors, for\\
$p \in \{1,2, \ldots, \frac{h(u-1)}{2}\}$,
\item there exists a decomposition of $K_{hm}$ into $s_\beta$ $C_{x}$-factors and $r_\beta$ $C_{y}$-factors.
\end{packed_item}
Let
\[s_{\alpha}=\sum_{p=1}^{\frac{h(u-1)}{2}}s_{p} \mbox{ and } r_{\alpha}=\sum_{p=1}^{\frac{h(u-1)}{2}}r_{p}.\]
Then there exists a decomposition of $K_{hum}$ into $s_\alpha+s_\beta$ $C_{x}$-factors and $r_\alpha+r_\beta$ $C_{y}$-factors.
\end{lemma}
We can now apply our decompositions to extend the result from \cite{AKKPO}. We will be concerned with prime numbers whose greatest power that divides $x$ is the same as their greatest power that divides $y$. Thus we give the following definition:
\begin{definition}
Let $x$ and $y$ be natural numbers, with $x=p_1^{a_1}\ldots p_n^{a_n}$ and $y=p_1^{b_1}\ldots p_n^{b_n}$ their prime factorization. Then we define the special greatest common divisor of $x$ and $y$, $\sgcd{x,y}$ as the smallest number such that $p_i^{a_i}$ divides $\sgcd{x,y}$ if and only if $a_i=b_i$.
\end{definition}
\begin{example}
For example if $x=2^33^25^27^2$ and $y=3^25^37^211^4$, then $\sgcd{x,y}=3^27^2$.
\end{example}
\begin{corollary}
Let $x,y,n$ be integers such that
\begin{itemize}
\item $\frac{xy}{\sgcd{x,y}}$ divides $n$,
\item $x,y\not\equiv 0 \pmod 4$, and $4$ divides $n$ if $2$ divides $xy$;
\item $3n=\frac{huxyw}{\sgcd{x,y}}$, with $h\equiv 0 \pmod 3$, $u\geq 3$, $h(u-1)$ even, and $(h,u)\not\in\{(2,6),(6,3)\}$.
\end{itemize}
Then there exists a decomposition of $K_{3n}$ into $s$ $C_{3x}$-factors and $r$ $C_{3y}$-factors for every pair $(s,r)$ such that $s+r=\lfloor \frac{3n-1}{2}\rfloor$, $s,r\neq 1$.
\end{corollary}
\begin{proof}
Let $(s,r)$ be such that $s+r=\lfloor \frac{3n-1}{2}\rfloor$ and $s,r\neq 1$.
If $s\geq r$ let $s_0=\lfloor\frac{3n-1}{2}\rfloor$ and $r_0=0$. Otherwise let $s_0=0$ and $r_0=\lfloor\frac{3n-1}{2}\rfloor$.
Let $s_1,\ldots,s_{\frac{h(u-1)}{2}}$ and $r_1,\ldots,r_{\frac{h(u-1)}{2}}$ be such that $s_i+r_i=\frac{xyw}{\sgcd{x,y}}$, $s_i,r_i\neq 1$ for all $i$ and
\[r_{\alpha}=\sum_{i=0}^{\frac{h(u-1)}{2}}r_{i} \mbox{ and } s_{\alpha}=\sum_{i=0}^{\frac{h(u-1)}{2}}s_{i}.\]
From Theorem \ref{3rgdd} we know that there is a 3-$\ensuremath{\mbox{\sf RGDD}}(h^{u})$. Let $z=\sgcd{x,y}$, $x_1=\frac{x}{z}$ and $y_1=\frac{y}{z}$. Then
\[
\frac{xyw}{z}=\frac{x_1zy_1zw}{z}=x_1y_1zw.
\]
So we may apply Theorem \ref{maintheorem} to obtain a decomposition of $K_{\left(\frac{xyw}{z}:3\right)}$ into $s_i$ $C_{3x_1z}$-factors and $r_i$ $C_{3y_1z}$-factors for each $i$.
Because $hxyw=hx_1zyw=hxy_1zw$ we have that $3x_1z|(hxyw)$ and $3y_1z|(hxyw)$. From Theorem \ref{OP} there is a decomposition of $K_{\frac{hxyw}{z}}$ into $C_{3x_1z}$-factors, and there is also a decomposition of $K_{\frac{hxyw}{z}}$ into $C_{3y_1z}$-factors.
Thus we may apply Lemma \ref{main} to obtain a decomposition of $K_{\frac{huxyw}{z}}=K_{3n}$ into $s$ $C_{3x}$-factors and $r$ $C_{3y}$-factors.
\end{proof}
We can also use Lemma \ref{buildcompletegraph} to obtain decompositions of complete graphs into $C_x$-factors and $C_y$-factors:
\begin{corollary}
Let $m,x$, and $y$ be integers such that:
\begin{itemize}
\item $z=\sgcd{x,y}$, $w=\frac{\gcd(x,y)}{z}\geq 2$,
\item $\frac{xy}{z}$ divides $m$,
\item $4$ does not divide $x$ nor $y$.
\item Neither $x$ nor $y$ is $3$ if $\frac{m}{w}\in\{6,12\}$.
\end{itemize}
Then there exists a decomposition of $K_m$ into $s$ $C_{x}$-factors and $r$ $C_y$-factors for every $s,r\neq 1$.
\end{corollary}
\begin{proof}
Let $s,r$ be such that $s+r=\lfloor\frac{m-1}{2}\rfloor$ and $s,r\neq 1$.
Let
\[
k=\frac{mz}{xy}\quad
x'=\frac{x}{zw}\quad
y'=\frac{y}{zw}\quad
m'=\frac{m}{w}=\frac{xyk}{zw}=x'y'zwk
\]
Let $s_\alpha,s_\beta,r_\alpha,r_\beta$ be such that $s_\beta,r_\beta\neq 1$, $\{s_\alpha,r_\alpha\}=\{0,\
\lfloor\frac{m'-1}{2}\rfloor\}$, $s=s_\alpha+s_\beta$ and $r=r_\alpha+r_\beta$.
By Theorem \ref{maintheorem} there is a decomposition of $K_{(m':w)}$ into $s_\beta$ $C_{x'zw}$-factors
and $r_\beta$ $C_{y'zw}$-factors. This is a decomposition of $K_{(m':w)}$ into $s_\beta$ $C_x$-factors and
$r_\beta$ $C_y$-factors.
Because $\frac{xy}{z}$ divides $m$, it follows that both $x$ and $y$ divide $m'=\frac{m}{w}=m\frac{z}
{\gcd(x,y)}$. Thus by Theorem \ref{OP} there is decompositon of $K_{m'}$ into $s_\alpha$ $C_x$-factors and $r_\alpha$ $C_y$-factors (keep in mind
that one of $s_\alpha$ and $r_\alpha$ is $0$).
Then by Lemma \ref{buildcompletegraph} there is a decomposition of $K_{m'w}$ into $s$ $C_{x}$-factors and $r$ $C_{y}$-factors.
Therefore there is a decompostion of $K_{m}$ into $s$ $C_{x}$-factors and $r$ $C_{y}$-factors.
\end{proof}
Notice that by asking $\frac{xy}{z}$ to divide $m$ we cover some of the cases left open in \cite{BDT} for the odd order case.
\begin{example}
Let $m=3^35^3$, $x=3^15^2$ and $y=3^25^2$. We have:
\begin{align*}
z=\sgcd{x,y}&=5^2,&w&=3^1,&k&=5^1
\end{align*}
And $\frac{xy}{z}=3^35^2$ divides $m$. So we can decompose $K_{m}$ into $s$ $C_x$-factors and $r$ $C_y
$-factors for any $s,r\neq 1$. Note that $l=\lcm(x,y)=3^25^2$. The number of vertices, $m$, is a multiple of $l$, however $xy\!\!\not | \,m$. Thus, Theorem \ref{theoremBDT} cannot be applied here.
\end{example}
\begin{example}
Let $m=4^13^45^37^1$, $x=2^13^15^27^1$ and $y=3^35^2$. We have:
\begin{align*}
z=\sgcd{x,y}&=5^2,&w&=3^1,&k&=2^15^1
\end{align*}
And $\frac{xy}{z}=2^13^45^27^1$ divides $m$. So we can decompose $K_{m}$ into $s$ $C_x$-factors and $r$ $C_y
$-factors for any $s,r\neq 1$. Note that because $x$ is even Theorem \ref{theoremBDT} cannot be applied here.
\end{example}
We can also use our Lemma \ref{buildcompletegraphnonuniform} to obtain non-uniform decompositions of
complete graphs:
\begin{corollary}
Let $v,m,n,x_1,\ldots,x_k,y_1,\ldots,y_k$ be integers such that:
\begin{itemize}
\item $m\geq 3$ is odd,
\item $n$ divides $m$, $x_i$, and $y_i$ for every $i$,
\item $k=\frac{m}{n}$,
\item $z_i=\sgcd{x_i,y_i}$,
\item $\frac{x_iy_i}{z_in}$ divides $v$ for each $i$,
\item $x_i$ divides $v$ for each $i$,
\item $4$ does not divide $x_i$ nor $y_i$ for any $i$,
\item $3\not\in\{x_1,\ldots,x_k,y_1,\ldots,y_k\}$ if $k\in\{6,12\}$.
\end{itemize}
Then there exists a decomposition of $K_{vm}$ into $s$ $[x_1^{vn/x_1},\ldots,x_k^{vn/x_k}]
$-factors and $r$ $[y_1^{vn/y_1},\ldots,y_k^{vn/y_k}]$-factors for every $s,r\neq 1$.
\end{corollary}
\begin{proof}
Let $s,r$ be such that $s+r=\lfloor\frac{vm-1}{2}\rfloor$ and $s,r\neq 1$.
Let
\[
x_i'=\frac{x_i}{z_in}\quad
y_i'=\frac{y_i}{z_in}\quad k_i=\frac{vz_in}{x_iy_i}
\]
Then
\[
v=\frac{x_iy_ik_i}{z_in}=\frac{x_i'z_iny_i'z_ink_i}{z_in}=x_i'y_i'z_ink_i
\]
Let $s_\alpha,s_\beta,r_\alpha,r_\beta$ be such that $s_\beta,r_\beta\neq 1$, $\{s_\alpha,r_\alpha\}=\{0,
\lfloor\frac{v-1}{2}\rfloor\}$, $s=s_\alpha+s_\beta$ and $r=r_\alpha+r_\beta$.
By Theorem \ref{maintheorem} there is a decomposition of $K_{(v:m)}$ into $s_\beta$ $[x_1^{vn/x_1},\ldots,x_k^{vn/x_k}]$-factors and $r_\beta$ $[y_1^{vn/y_1},\ldots,y_k^{vn/y_k}]$-factors.
We partition $mK_{v}$ into $k$ copies of $nK_{v}$, labeled $\kappa_1,\ldots\kappa_k$. Because $\frac{x_iy_i}{z_in}$ and $x_i$ divide $v$, we get that both $x_i$ and $y_i$ divide $v$, by Theorem \ref{OP}
there is a decompostion of $K_v$ into $s_{\alpha}$ $C_{x_i}$-factors and $r_\alpha$ $C_{y_i}$-factors (keep in mind that one of $s_\alpha$ and $r_\alpha$ is $0$). This means that $\kappa_i$ can be decomposed into $s_\alpha$ $[x_i^{vn/x_i}]$-factors
and $r_\alpha$ $[y_i^{vn/y_i}]$-factors. Combining these decompositions we get a decomposition of
$mK_{v}$ into $s_\alpha$ $[x_1^{vn/x_1},\ldots,x_k^{vn/x_k}]$-factors and $r_\alpha$
$[y_1^{vn/y_1},\ldots,y_k^{vn/y_k}]$-factors
Then by Lemma \ref{buildcompletegraphnonuniform} there is a decomposition $K_{vm}$ into $s$ $
[x_1^{vn/x_1},\ldots,x_k^{vn/x_k}]$-factors and $r$ $[y_1^{vn/y_1},
\ldots,y_k^{vn/y_k}]$-factors for every $s,r\neq 1$.
\end{proof}
\begin{example}
Let $v=5^37^211^313^4$, $m=3^15^1$, $n=5$, $k=3$, $x_1=5^27^2$, $y_1=5^17^211^113^1$, $x_2=5^111^113^1$, $y_2=5^17^213^3$, $x_3=5^27^111^113^4$, and $y_3=5^1$. We have:
\begin{align*}
z_1=\sgcd{x_1,y_1}&=7^2, &\frac{x_1y_1}{z_1n}=5^27^2\\
z_2=\sgcd{x_2,y_2}&=5^1, &\frac{x_2y_2}{z_2n}=7^211^113^4\\
z_3=\sgcd{x_3,y_3}&=1, &\frac{x_3y_3}{z_3n}=5^27^111^113^4\\
\end{align*}
We can see that $\frac{x_iy_i}{z_in}$ and $x_i$ divide $v$ for each $i$.
Let $F_1$ be the $2$-factor consisting of:
\begin{itemize}
\item $\frac{vn}{x_1}=5^211^313^4$ cycles of size $x_1=5^27^2$,
\item $\frac{vn}{x_2}=5^37^211^213^3$ cycles of size $x_2=5^111^113^1$,
\item and $\frac{vn}{x_3}=5^26^111^2$ cycles of size $x_3=5^27^111^113^4$.
\end{itemize}
Let $F_2$ be the $2$-factor consisting of:
\begin{itemize}
\item $\frac{vn}{y_1}=5^311^213^3$ cycles of size $y_1=5^17^211^113^1$,
\item $\frac{vn}{y_2}=5^311^313^1$ cycles of size $y_2=5^17^213^3$,
\item and $\frac{vn}{y_3}=5^37^211^313^4$ cycles of size $y_3=5^1$.
\end{itemize}
Then we can decompose $K_{vm}$ into $s$ copies of $F_1$ and $r$ copies of $F_2$ for any $s,r\neq 1$.
\end{example}
\section{Bibliography}
|
\section{Introduction}
\label{sec:intro}
There are a variety of rotating objects in the universe.
We are living on the rotating earth which is revolving around the sun.
The sun is a part of our rotating galaxy.
The neutron star is often observed as a ``pulsar'' whose rotational period
can be as small as $10^{-3}$ s.
However, if we look into a microscopic world, we find
much faster rotating objects, such as nuclei.
The nuclear rotational period in heavy nuclei
is typically $\tau_{\rm rot}=10^{-19}\sim 10^{-20}$ s.
This time scale is $100-1,000$ times larger than
the period of the single-particle Fermi motion inside the nucleus,
$\tau_F\sim 10^{-22}$ s.
Thus, the rotational motion can be treated as ``slow'' motion
at low-spin states.
However, in high-spin states produced by the fusion reaction,
it could reach $\tau=10^{-21}\sim 10^{-22}$ s which is comparable to $\tau_F$.
Therefore, the nuclei provide a unique laboratory to study
rapidly rotating quantum systems under strong Coriolis and
centrifugal fields.
The nucleus is a finite quantum many-body system.
Since the Hamiltonian is rotationally invariant,
its energy eigenstate has a definite total angular momentum $I$.
In order to realize the nuclear rotation,
the nucleus needs to define its orientation.
Since it is impossible to do it for spherical systems,
a deformed {\it intrinsic} state, which is
produced by breaking the rotational symmetry, is necessary.
The word ``intrinsic'' means the degrees of freedom approximately
independent of the rotational motion.
The spontaneous breaking of symmetry (SBS) is an important concept
to constitute the unified model of Bohr and Mottelson \cite{BM75}.
Hereafter, we denote the textbook \cite{BM75} by ``BM2''.
The SBS is strictly defined only in infinite systems.
Therefore, in the beginning of this article
(sections~\ref{sec:unified_model}, \ref{sec:finite_size_effect},
and \ref{sec:collective_motion}),
we address the following basic questions.
\begin{enumerate}
\item What is the origin of nuclear deformation?
\item What is the meaning of the SBS in finite systems?
\item What kind of collective motion will emerge due to the SBS?
\end{enumerate}
These are important issues to understand the essence of
the nuclear structure physics.
We hope that these sections are useful, especially
for students and non-practitioners.
The rotational motion is a collective motion emerged from the SBS,
corresponding to the massless Anderson-Nambu-Goldstone (ANG) mode
\cite{And52,And58-1,And58-2,Nam60,NJL61,Gol61}
in the infinite system.
It is approximately decoupled from the intrinsic motions, however,
the decoupling is never exact.
Moreover, as we mentioned in the beginning, we may generate
a nucleus spinning extremely fast in experiments.
Coupling between the rotational and intrinsic motions
produces a variety of phenomena.
We will discuss several related topics in
sections~\ref{sec:coupling}, \ref{sec:SD}, and \ref{sec:wobbling}.
The coupling introduces the Coriolis mixing among different bands.
The angular momentum dependence ($I$-dependence) of
the transition matrix elements is very sensitive to this, even at
low spin.
The unified model predicts a form of the $I$-dependent
intensity relation,
however, a systematic way of calculating intrinsic moments
entering in the intensity relation was missing.
We present a feasible microscopic method for the calculation of
the intrinsic moments using the cranking model at
an infinitesimal rotational frequency (section~\ref{sec:coupling}).
Low-lying vibrational modes of excitation strongly reflect the
underlying shell structure.
Therefore, the new shell structure produced by rapid rotation
and large deformation may significantly change their properties,
and could lead to new soft modes and instability.
Octupole vibrations at large angular momentum and in superdeformed bands
are discussed.
We present a possible banana-shaped superdeformation as a
consequence of the superdeformed shell structure (section~\ref{sec:SD}).
Nuclear wobbling motion, predicted by Bohr and Mottelson
(section 4-5 in BM2),
has been observed in $^{163}$Lu and neighboring nuclei.
This mode corresponds to a non-uniform three-dimensional rotation
and provides a clear signature of nuclear deformation without the
axial symmetry.
Microscopic analysis reveals an important role played by
the quasiparticle alignment.
In fact, without the alignment,
the wobbling motion cannot exist in $^{163}$Lu.
These issues and precession motion of the high-$K$ isomers
are discussed in section~\ref{sec:wobbling}.
For most of these studies, we use
the quasiparticle-random-phase
approximation (QRPA) in the rotating shell model,
which we have developed for studies of rapidly rotating nuclei
\cite{SM82,MSM90,NMMS96,SS09}.
Further inclusion of the quasiparticle-vibration coupling
has been carried out for odd nuclei \cite{MSM88,NMM93,Mat14}.
The method is still very useful and illuminating to obtain
insights into nuclear structure in extreme conditions.
In the present article, we do not present details of the
theoretical models.
Instead, we would like to concentrate our discussion on
basic concepts and emergent phenomena.
\section{Unified model and spontaneous breaking of symmetry}
\label{sec:unified_model}
The atom is a finite-size quantum system,
composed of electrons bound by the Coulombic attraction of
the central nucleus.
The nucleus is also a quantum finite-size fermionic system
composed of nucleons.
In both systems, the independent-particle (single-particle) motion
is a prominent feature which leads to the ``shell model''.
In the first order approximation,
the constituent particles (electrons in the atom and nucleons in the nucleus)
freely move in the confining potential.
However, there is an obvious but important difference
between the atom and the nucleus.
Namely, the nuclear potential binding the nucleons is generated
by the nucleons themselves.
The electrons in the atom are bound by
the attractive Coulomb potential generated by the nucleus.
This potential is spherical, $-Ze^2/r$, in the atomic scale ($r\sim$\AA).
Although the repulsive interaction creates correlations
among the electrons, the strong attractive potential always produces a
restoring force which favors the spherical shape.
In contrast, the shape of the nuclear potential is determined by
the shape of the nucleus itself.
It is often referred to as {\it ``nuclear self-consistency''}.
Therefore, we expect that the nucleus may change its shape,
much easier than the atomic case.
In other words, the nucleus is rather ``soft'' and produces low-energy
``slow'' shape vibrations.
\subsection{Unified model Hamiltonian}
Bohr and Mottelson treated these shape degrees of freedom as collective
variables $\alpha$ in addition to the single-particle
degrees of freedom $\xi$.
In general, the shape dynamics described by $\alpha$ is considerably
slower than the single-particle motion.
Thus, we could adopt a picture that the nucleons move in a one-body
potential $V(\xi,\alpha)$ which is specified by the nuclear shape $\alpha$.
The idea ends up with the unified-model Hamiltonian,
\begin{equation}
H=H_{\rm coll}(\alpha) + H_{\rm sp}(\xi) + H_{\rm int}(\xi,\alpha) ,
\label{unified_model}
\end{equation}
where $H_{\rm coll}$ is the {\it collective} Hamiltonian to describe
the low-energy shape vibrations.
$H_{\rm sp}$ corresponds to the {\it single-particle} (shell model) Hamiltonian
at the spherical shape, $H_{\rm sp}(\xi)=T_{\rm kin}(\xi) + V(\xi,\alpha=0)$.
$T_{\rm kin}$ is the {\rm kinetic} energy term and
the nuclear self-consistency requires
the potential $V(\xi,\alpha)$ to vary with respect to the shape $\alpha$.
The {\it interaction} between the collective and single-particle motions,
given by the third term $H_{\rm int}(\xi,\alpha)$,
is indispensable to take into account this important property of
nuclear potential.
\subsection{Symmetry breaking mechanism}
The coupling term in equation~(\ref{unified_model}) could lead the nucleus
to deformation.
This is associated with the SBS mechanism.
To elucidate the idea,
let us adopt a simple adiabatic (Born-Oppenheimer) approximation.
First, we solve the eigenvalue problem for
the Schr\"odinger equation for the variables $\xi$
with a fixed value of $\alpha$,
\begin{equation}
H_{\rm def}(\alpha) \ket{\phi_n(\alpha)}
= \epsilon_n(\alpha) \ket{\phi_n(\alpha)} ,
\end{equation}
where $H_{\rm def}(\alpha)\equiv H_{\rm sp} + H_{\rm int}(\alpha)$.
This gives the adiabatic collective Hamiltonian,
$H_{\rm ad}^{(n)}(\alpha)=H_{\rm coll}(\alpha)+\epsilon_n(\alpha)$,
for each intrinsic eigenstate
$\phi_n(\xi;\alpha)\equiv\inproduct{\xi}{\phi_n(\alpha)}$.
$H_{\rm ad}^{(n)}$ is an effective Hamiltonian for
the collective variables $\alpha$.
The total wave function is given by a product of the intrinsic and
the collective parts \cite{Vil84},
$\Psi_n(\alpha,\xi)=\psi^{(n)}(\alpha)\phi_n(\xi;\alpha)$.
There are two possible mechanisms of the SBS in the unified model
to realize the deformed ground state with $\alpha\neq 0$.
When $\epsilon_0(\alpha)$ strongly favors the deformation,
even if $H_{\rm coll}(\alpha)$ has the potential minimum at $\alpha=0$,
the adiabatic potential in $H_{\rm ad}^{(0)}(\alpha)$
may have a deformed minimum.
Apparently, this mechanism requires
the deformation-driving nature of $H_{\rm int}(\xi,\alpha)$,
which we call ``{\it coupling-driven mechanism}''.
On the other hand, there is another mechanism
which can deform the nucleus even if
the spherical shape ($\alpha=0$) is favored by the adiabatic
ground state $\epsilon_0(\alpha)$.
This is due to the additional coupling caused by
the kinetic term of $H_{\rm coll}$;
$T_{\rm kin}(\alpha)=-(1/2)B\partial_\alpha^2$.
We adopt units of $\hbar=1$ throughout the present article.
Roughly speaking,
the SBS takes place when the level spacing in $\epsilon_n(\alpha)$
is smaller than the additional coupling.
It is analogous to the Jahn-Teller effect in the molecular physics
\cite{JT37},
which we call ``{\it degeneracy-driven mechanism}''.
\subsection{Degeneracy-driven SBS; diagonal approximation}
In order to understand the degeneracy-driven mechanism, we
make the argument simpler,
neglecting the off-diagonal elements,
$\bra{\phi_n(\alpha)}T_{\rm kin}(\alpha)\ket{\phi_0(\alpha)}$ ($n\neq 0$).
Integrating the intrinsic (single-particle) degrees $\xi$,
the effective Hamiltonian for the collective variable $\alpha$ is obtained as
\begin{equation}
H_{\rm eff}^{(0)}(\alpha)=\bra{\phi_0(\alpha)} H \ket{\phi_0(\alpha)}
=H_{\rm coll}^{(0)}(\alpha)+\epsilon_0(\alpha)+\Phi_0(\alpha),
\end{equation}
where $H_{\rm coll}^{(0)}$ is identical to $H_{\rm coll}$ except
that its kinetic energy is modified into $T_{\rm kin}^{(0)}(\alpha)= -(1/2)B
(\partial_\alpha + \inproduct{\phi_0}{\partial_\alpha\phi_0})^2$.
This is equivalent to introduction of a ``vector'' potential \cite{Ber84},
$A(\alpha)\equiv i\inproduct{\phi_0}{\partial_\alpha\phi_0}$.
If the coordinate $\alpha$ is one-dimensional, the ``vector'' potential
$A(\alpha)$ can be eliminated by a gauge transformation,
$\exp(i\int A(\alpha) d\alpha)$.
However, the following ``scalar'' potential remains.
\begin{eqnarray}
\Phi_0(\alpha)
&=&\frac{1}{2}B\bra{\partial_\alpha\phi_0}
(1-\ket{\phi_0}\bra{\phi_0})
\ket{\partial_\alpha\phi_0}
=\frac{1}{2}B\sum_{n\neq 0}
\inproduct{\partial_\alpha\phi_0}{\phi_n}
\inproduct{\phi_n}{\partial_\alpha\phi_0}
\nonumber \\
&=&\frac{1}{2}B\sum_{n\neq 0}
\left|
\frac{\bra{\phi_n(\alpha)}(\partial_\alpha H_{\rm def}(\alpha))
\ket{\phi_0(\alpha)}}
{\epsilon_0(\alpha)-\epsilon_n(\alpha)}
\right|^2 .
\label{Phi_0}
\end{eqnarray}
From equation~(\ref{Phi_0}), it is apparent that $\Phi_0(\alpha)$ is positive
and becomes large where the adiabatic ground state is approximately
degenerate in energy, $\epsilon_0\approx \epsilon_n$ ($n\neq 0$).
When the spherical ground state ($\alpha=0$) shows degeneracy,
it could be significantly unfavored by $\Phi_0(\alpha)$.
The system tends to avoid the degenerate ground state, which leads to
the SBS with nuclear deformation.
We would like to emphasize again that
the coupling between the collective (shape)
degrees of freedom $\alpha$ and the intrinsic (single-particle) motion
$\xi$ is essential to produce the nuclear deformation.
This is apparent for the coupling-driven mechanism, and is also true
for the degeneracy-driven case.
If the coupling term $H_{\rm int}(\xi,\alpha)$ is absent,
the adiabatic states $\phi_n(\xi)$ are independent of $\alpha$,
thus, produce no gauge potentials, $A(\alpha)=\Phi_0(\alpha)=0$.
We also note here that the present argument on the degeneracy-driven
(Jahn-Teller) mechanism explains why the instability of a spherical state
occurs, but not what kind of deformation takes place.
This will be discussed in sections~\ref{sec:shell_structure}
and \ref{sec:banana_SD}.
\subsection{Field coupling}
The oscillation of the variable $\alpha$ correspond to the shape vibration.
Thus, it can be quantized to a boson operator.
In order to describe the vibrational motion associated with $\alpha$,
we introduce a boson space with the $n$-phonon state $\ket{n}$.
When $\alpha$ is small,
we may linearize the coupling term in equation~(\ref{unified_model})
with respect to $\alpha$ as
\begin{equation}
H_{\rm int}(\xi,\alpha)=-\kappa\alpha F(\xi) ,
\label{H_int}
\end{equation}
where $\kappa$ is a coupling constant which depends on the
normalization of $\alpha$ and $F$.
If the operator $F$ is given, the normalization of $\alpha$ is
usually chosen as follows.
The action of the one-body operator $F$ on the ground state
(a Slater determinant) produces many one-particle-one-hole states;
$F\ket{\Phi_0}=\sum_{ph} \ket{\Phi_{ph}}\bra{\Phi_{ph}}F\ket{\Phi_0}$.
This is identified with the operation of $\alpha$
in the collective (boson) space:
\begin{equation}
\sum_{ph}|\bra{\Phi_{ph}}F\ket{\Phi_0}|^2 = |\bra{n=1}\alpha\ket{n=0}|^2 .
\label{normalization}
\end{equation}
The coupling constant $\kappa$ can be also determined by this
self-consistency.
See chapter 6 of BM2 for details of the field coupling techniques.
If the matrix elements of $F$ are identical to
those of $\alpha$ as in equation~(\ref{normalization}),
the field coupling~(\ref{H_int}) can be interpreted as
an residual two-body interaction
\begin{equation}
H_{\rm res}(\xi)=-\frac{1}{2}\kappa F^2 .
\label{separable_interaction}
\end{equation}
This kind of separable effective interactions have been extensively
adopted in nuclear structure studies.
Among them, the most famous one is the pairing-plus-quadrupole model
\cite{BK65,KB68-1,BS69},
which was originally proposed by Bohr, Mottelson, and their colleague.
It represents two kinds of important low-energy correlations in nuclei;
One is the quadrupole correlations, $F\sim r^2Y_{2\mu}$,
which are inspired by existence of low-lying $2^+$ vibrational
excitations in even-even nuclei.
Another correlation is the pairing,
$F\sim P+P^\dagger$ and $F\sim P-P^\dagger$
where $P$ is the pair annihilation operator.
This is important in heavy nuclei in which the nuclear superfluidity
associated with the pair condensation is well established \cite{BB05}.
We also adopt this separable form as a residual interaction for
the QRPA calculations in sections~\ref{sec:coupling}, \ref{sec:SD},
and \ref{sec:wobbling}.
\section{Finite-size effect}
\label{sec:finite_size_effect}
Nuclei on earth are of finite size ($R<10$ fm)
with finite number of nucleons ($A<300$).
Strictly speaking,
the SBS in the ground state is realized in the infinite system.
For finite systems, the quantum fluctuation associated with the zero-point
motion restores the broken symmetry.
Thus, the symmetry-broken state is not stable for finite systems,
in a rigorous sense.
However, the SBS is ubiquitous in macroscopic objects in nature,
which are made of big but {\it finite} number of particles.
Thus, everybody agrees that
the zero-point motion to restore the symmetry
can be safely neglected in
the macroscopic number, say $A\sim 10^{23}$.
Then, how about the case of $A\sim 200$?
\subsection{Finite correlation time}
\label{sec:finite_correlation_time}
Let us consider a deformed nucleus and the single-particle states
$\phi_i^0$ in the deformed Nilsson potential.
The deformed ground state is simply assumed to be a Slater determinant,
$\ket{\Phi_0}\equiv\det\{\phi_1^0\phi_2^0\cdots\phi_A^0\}$.
If we rotate the nucleus by angle $\theta$, we have a state
$\ket{\Phi_\theta}=
\det\{\phi_1^\theta\phi_2^\theta\cdots\phi_A^\theta\}$.
where $\phi_i^\theta=\hat{R}(\theta)\phi_i^0$ with the rotation operator
$\hat{R}(\theta)$.
Each single-particle state $\phi_i^\theta$ in the tilted Nilsson potential
can be expanded in terms of the untilted state $\phi_i^0$, as
$\phi_i^\theta=\sum_j^\infty c_{ij}^\theta \phi_j^0$.
When the angle $\theta$ is small,
we can estimate the diagonal coefficients as $|c_{ii}|\sim 1-c|\theta|$
with $c>0$, and the off-diagonal ones ($i\neq j$) as $|c_{ij}|\sim O(\theta)$.
As far as the nucleon number $A$ is finite,
the tilted ground state $\ket{\Phi_\theta}$ can be written in terms of
the untilted Nilsson basis, $\{ \phi_i^0 \}$.
This is due to the fact that $\ket{\Phi_0}$ and $\ket{\Phi_\theta}$
belong to the same Hilbert space.
However, in the limit of $A\rightarrow\infty$,
this is no longer true.
The tilted ground state is expanded in terms of
the untilted Slater determinants as
\begin{equation}
\ket{\Phi_\theta}=\det\{\phi_1^\theta\cdots\phi_A^\theta\}
=\sum_{j_1,\cdots, j_A} C_{j_1\cdots j_A}
\det\{\phi_{j_1}^0\cdots\phi_{j_A}^0 \} ,
\end{equation}
where $C_{j_1 j_2\cdots j_A}=c_{1j_1}c_{2j_2}\cdots c_{Aj_A}$.
For a small value of $\theta$,
the largest coefficient among $\{ |C_{j_1\cdots j_A}|\}$
is apparently $|C_{1\cdots A}|$ whose absolute magnitude is
$|C_{1\cdots A}|\sim (1-c|\theta|)^A$.
Therefore, all the coefficients $C_{j_1\cdots j_A}$
vanish exponentially as functions of $A$.
This means that $\ket{\Phi_0}$ and $\ket{\Phi_\theta}$ belong to
different Hilbert spaces at $A\rightarrow\infty$,
thus, $\ket{\Phi_\theta}$ is no longer expandable
in terms of the untilted Slater determinants.
In other words, the deformed {\it infinite} nucleus never
rotates.
The same issue can be examined in terms of the excitation spectra.
The rotational spectra of deformed nuclei show $E_I=I(I+1)/2{\cal J}$,
in which the moment of inertia $\mathcal{J}$
is approximately order of $A^{5/3}$.
The rotational motion is {\it quantized} due to the finiteness of ${\cal J}$.
In the limit of $A\rightarrow\infty$,
the excitation spectra becomes {\it gapless} and
the ground state ($I=0$) is degenerate with other states ($I\neq 0$).
Therefore, an infinitesimally weak external field can fix its orientation
by superposing states with different $I$.
Now, let us come back to the question,
"how about heavy deformed nuclei?".
As far as $A$ is finite, the ``tilted'' and ``untilted''
Hilbert spaces are equivalent.
The zero-point fluctuation may connect
$\ket{\Phi_0}$ and $\ket{\Phi_\theta}$,
thus, the wave packet $\ket{\Phi_0}$
loses its direction in finite correlation time.
If this time scale is significantly larger than that of the single-particle
motion $\tau_F\sim 10^{-22}$ s,
we can claim that the SBS takes place and the nucleus is deformed.
In fact, this condition is well satisfied for heavy nuclei.
Let us limit the orientation of the deformed nucleus to
an angle range of unity ($\Delta\theta\sim 1$),
then, the quantum fluctuation produces the angular momentum with the
magnitude of $\Delta I \sim (\Delta\theta)^{-1} \sim 1$.
This leads to the correlation time,
$\tau_{\rm cor}\sim {\cal J}/\Delta I \sim {\cal J}$,
that amounts to $10^{-19}$ s for typical deformed actinide nuclei.
This argument is consistent with the vanishing behavior of the
coefficients $C_{j_1\cdots j_A}$.
Suppose the overlaps $|\inproduct{\phi_i^0}{\phi_i^\theta}|\sim 0.9$ for
$i=1,\cdots,A$,
then, we have $C_{12\cdots A}\sim (0.9)^A \sim 7\times 10^{-10}$
for $A=200$.
Therefore, the rotational fluctuation is significantly hindered for
heavy deformed nuclei.
These simple exercises also tell us that,
the concept of SBS has a greater significance for
nuclei with larger $A$ and larger deformation.
From a similar argument
replacing the single-particle states $\phi_i^0$ in the Slater determinant
$\det\{\phi_{j_1}^0\cdots\phi_{j_A}^0 \}$ by those in a spherical potential,
we may understand why the description based on the ``symmetry-broken''
deformed basis is important.
It is apparent that,
if we adopt a spherical shell model basis for such heavy well-deformed nuclei,
we need to treat very small coefficients, $C_{j_1\cdots j_A}$,
with enormous number of basis states.
In the limit of $A\rightarrow\infty$, this treatment becomes impossible.
The impossibility here is in a strict sense,
not in a practical sense due to computational limitation.
Thus, instead of superposing the ``symmetry-preserving'' (spherical)
Slater determinants, the theories of restoring broken symmetry,
such as the projection method,
have been extensively developed in nuclear physics,
to take into account effects of the zero-point fluctuation \cite{RS80}.
The usefulness of these symmetry restoration approaches has been
recognized recently in other fields \cite{YL07}.
\subsection{Zero-point motion and shell effect}
As we have mentioned in section~\ref{sec:finite_correlation_time},
the finiteness leads to
the finite correlation time and the finite energy gap in the excitation
spectra.
In the symmetry restoration mechanism, the zero-point fluctuation
associated with the ANG mode is a key element.
In this subsection, we discuss effects of other zero-point motions in
finite systems, which could hinder the SBS.
The zero-point kinetic energy of nucleons is roughly given as
$T_\textrm{\sc zpe}/A \sim 1/(mR^2) \sim 10$ MeV.
This is comparable to the nuclear binding energy $B/A\sim 8$ MeV
and has a non-negligible effect.
In fact, since the nucleons are fermions, the Fermi energy is even larger,
$t_F\sim k_F^2/(2m)\sim 40$ MeV.
The zero-point (Fermi motion) kinetic energy
generally favors the ``symmetry-preserving'' state with
a uniform and spherical density distribution.
Since this competes with the SBS driving effect,
the SBS which occurs in the thermodynamical limit may not occur
in finite systems.
The interplay between the zero-point motion and the interaction
leads to interesting phenomena in nuclei.
The shell effect is a kind of finite-size effect in many fermion systems
and is an indispensable factor in the low-energy
nuclear structure.
The prominent deformation hindrance effect can be found
at the spherical magic numbers.
The ground states of those magic nuclei favor spherical shape.
Nevertheless,
most of the spherical nuclei show the shape coexistence phenomena.
For instance, the even-even spherical nuclei often have deformed excited
$0^+$ states at very low energy.
It is prevalent in many semi-magic nuclei, and even true for
some doubly magic nuclei.
In contrast, as far as we know, when the ground state is deformed,
excited $0^+$ spherical states have not been clearly identified.
In the Strutinsky shell correction method \cite{Bra72,RS80},
the shell effect is regarded as an origin of the nuclear deformation.
It might be proper to say this in an opposite way;
the heavy nuclei are ``genetically'' deformed,
and some special nuclei become spherical
because of the finite-size (spherical shell) effect to hinder the SBS.
\subsection{Shell structure and soft modes}
\label{sec:shell_structure}
When the symmetry breaking takes place and the nucleus is deformed,
what kind of shape is realized?
This depends on the underlying shell structure.
Let us present a simple argument based on the one given by
by Bohr and Mottelson (pp. 578$-$591 in BM2).
For a spin-independent spherical potential,
the single-particle energy is characterized
by the radial quantum number $n$ and the orbital angular momentum $l$,
$\epsilon(n,l)$.
When we change $n$ and $l$ from a certain value $(n_0,l_0)$.
\begin{equation}
\epsilon(n,l)
=
\epsilon(n_0,l_0)+ \Delta n
\left(\frac{\partial\epsilon}{\partial n}\right)_0
+\Delta l
\left(\frac{\partial\epsilon}{\partial l}\right)_0
+ \cdots ,
\label{epsilon}
\end{equation}
where $\Delta n=n-n_0$ and $\Delta l=l-l_0$.
Since $n$ and $l$ take only integer numbers,
the ratio,
$a:b\equiv
({\partial\epsilon}/{\partial n})_0:({\partial\epsilon}/{\partial l})_0$
plays a very important role.
If the ratio $a:b$ is rational, we can choose
$a$ and $b$ as the integer numbers.
Then, in the linear order (\ref{epsilon}),
$\epsilon(n,l)$ and $\epsilon(n\pm mb, l\mp ma)$ are degenerate,
where $m$ is an integer number.
Now, let us define the shell frequency as
\begin{equation}
\omega_{\rm sh}\equiv
\frac{1}{a}\left(\frac{\partial\epsilon}{\partial n}\right)_0
=\frac{1}{b}
\left(\frac{\partial\epsilon}{\partial l}\right)_0 .
\end{equation}
There are degenerate single-particle energies at intervals of
$\omega_{\rm sh}$.
Larger integers $a$ and $b$ correspond to a smaller $\omega_{\rm sh}$.
Therefore, the prominent shell structure with a large shell gap
$\omega_{\rm sh}$ should be associated with the small integers $(a,b)$.
For instance, the isotropic harmonic oscillator potential
has the $a:b=2:1$ shell structure, with the constant $\omega_{\rm sh}$.
The Coulomb potential has the strict $a:b=1:1$, with the energy-dependent
$\omega_{\rm sh}$.
In general, the degeneracy is approximate and the ratio $a:b$ may change
according to the location of the Fermi level.
The derivatives, $(\partial\epsilon/\partial n, \partial\epsilon/\partial l)$,
correspond to the (angular) frequencies in the classical mechanics;
$\partial\epsilon/\partial n$ is the frequency of the radial motion,
while
$\partial\epsilon/\partial l$ is that of the angular motion.
The integer ratio $(a,b)$ of the frequencies means that the classical orbit
is closed (periodic).
Therefore, the quantum shell structure is closely related to
the classical periodic orbits.
Since the nuclear potential somewhat resembles
the harmonic oscillator potential,
the shell structure associated with $a:b=2:1$
is prominent.
The $2:1$ periodic orbit in the harmonic oscillator potential
is the elliptical orbit.
When there are many valence nucleons in the degenerate levels,
the short-range attractive interaction favors their maximal overlap,
which eventually leads to the SBS
to an ellipsoidal (quadrupole) shape.
In the quantum mechanical terminology,
we may say that the coupling among the degenerate single-particle
levels with $\Delta l=2$ produces a soft mode.
If the number of valence nucleons becomes large,
this correlation may produce the quadrupole deformation.
The spin-orbit potential decreases
the frequency $\partial\epsilon/\partial l$
for the single-particle levels of $j=l+1/2$.
This could lead to a new shell structure of $a:b=3:1$
among the levels of the $j=l+1/2$.
The $3:1$ frequency ratio corresponds to classical periodic orbits of
the triangular shape,
since the radial motion oscillates three times during the single
circular motion.
Thus, for heavy nuclei in which the high-$j$ single-particle levels
($j=l+1/2$) are located near the Fermi level,
the approximate degeneracy of the $\Delta l=3$ levels
may result in the octupole instability
in open-shell configurations.
For example,
the neutron-deficient actinide nuclei show typical spectra of
the alternating parity band,
which are understood as a realization of the pear-shaped deformation
of $Y_{30}$ type \cite{BN96}.
The investigation of the classical periodic orbits is
useful to identify a soft mode and a favorable shape.
The SBS toward the quadrupole deformation in open-shell nuclei
is nicely explained
in this simple argument.
However, it is more difficult to explain the fact that most nuclei have
the prolate shape, not the oblate shape.
There have been a number of works on this issue \cite{Fri90,TS01,HM09,TOST11}.
According to the classical periodic orbits, a recent analysis
sheds new light on the prolate dominance in nuclei \cite{Ari12}.
\subsection{Fermi motion and nuclear self-consistency}
\label{sec:self-consistency}
In nuclei, the kinetic energy of nucleons' Fermi motion is very large.
Adopting the harmonic oscillator potential model,
a simple estimate of the total kinetic energy is given by
\begin{equation}
T_0=\frac{1}{2}E_{\rm HO}
=\frac{1}{2}\sum_{k=x,y,z}\omega_k\Sigma_k,
\quad\quad
\Sigma_k\equiv\sum_{i=1}^A (n_k+\frac{1}{2})_i ,
\end{equation}
where $n_k$ ($k=x,y,z$) are the oscillator quantum numbers of
the single-particle states.
For a spherical nucleus ($\omega_x=\omega_y=\omega_z=\omega_0$)
filling the levels up to $n_x+n_y+n_z=N_{\rm max}$,
this amounts to
\begin{equation}
T_0=(1/4)\omega_0
(N_{\rm max}+1)(N_{\rm max}+2)^2(N_{\rm max}+3) .
\end{equation}
Taking $Z=N=40$ ($^{80}$Zr, $N_{\rm max}=3$), this gives
$T_0=150\omega_0\sim 1.43$ GeV,
with a standard value of $\omega_0\approx 41$A$^{-1/3}\sim 9.5$ MeV.
If we deform the harmonic oscillator to a prolate/oblate shape with
$(\omega_x,\omega_y,\omega_z)=(e^\eta,e^\eta,e^{-2\eta})\omega_0$,
the kinetic energy becomes
\begin{equation}
T(\alpha)=\frac{2e^\eta+e^{-2\eta}}{3}T_0 ,
\label{T_alpha}
\end{equation}
which has the minimum value at the spherical shape $\eta=0$.
According to equation~(\ref{T_alpha}),
a moderate prolate deformation of $e^\eta=1.1$ will
produce the increase in the kinetic energy by about 1 \%.
However, since $T_0$ is very large, this 1 \%\ increase is
significant, such as 14 MeV for $Z=N=40$.
However, the deformed ground state in $^{80}$Zr
is suggested by experiments observing a ground-state rotational band
\cite{Lis87}.
How does the nucleus compensate this large increase in kinetic energy?
The solution to this problem
is again attributed to the nuclear self-consistency.
In the harmonic oscillator model, the self-consistency condition, that
the deformation of the potential is equal to that of the density
distribution,
can be simply expressed by equation (4-115) in BM2,
\begin{equation}
\omega_x\Sigma_x =
\omega_y\Sigma_y =
\omega_z\Sigma_z .
\label{self_consistency_ho}
\end{equation}
Namely, when the nuclear potential is deformed as
$(\omega_x,\omega_y,\omega_z)=(e^\eta,e^\eta,e^{-2\eta})\omega_0$,
the configuration of the ground state should change accordingly,
$(\Sigma_x,\Sigma_y,\Sigma_z)=(e^{-\eta},e^{-\eta},e^{2\eta})\Sigma_0$.
Since the momentum distribution in the harmonic oscillator potential
model can be calculated as
$\langle p_k^2 \rangle = m \omega_k \Sigma_k$ ($k=x,y,z$),
the self-consistency condition~(\ref{self_consistency_ho})
means the isotropic momentum (velocity) distribution (no deformation in the
Fermi sphere).
In other words, the shape of the nucleus is specified by
the minimization of the kinetic energy which is equal to the
isotropic velocity distribution.
This indicates the importance of configuration rearrangements
in low-energy collective dynamics.
When the nuclear deformation is changed as
$(\omega_x,\omega_y,\omega_z)\rightarrow
(e^\eta\omega_x,e^{\eta'}\omega_y,e^{-\eta-\eta'}\omega_z)$,
the configuration should follow as
$(e^{-\eta}\Sigma_x,e^{-\eta'}\Sigma_y,e^{\eta+\eta'}\Sigma_z)$,
to keep the Fermi sphere spherical.
In order to change the configuration, we need two-particle-two-hole
excitations, to annihilate
a time-reversal pair of nucleons in a certain single-particle orbit
and create a pair in another orbit.
Therefore, we expect that,
during the shape evolution at low energy,
the pairing interaction plays a dominant
role in dynamical change of the configuration.
This was supported by experimental data that the spontaneous fission
life times of even-even nuclei are much shorter than those of
odd and odd-odd nuclei \cite{Swi82}.
According to the nuclear self-consistency,
each configuration has its optimal shape.
We may think about possibilities of realizing
different shapes corresponding to different configurations
in the same nucleus.
This phenomenon is called ``shape coexistence''.
For instance, in the harmonic oscillator model of $^{80}$Zr with $N=Z=40$,
in addition to the spherical configuration
($\Sigma_x=\Sigma_y=\Sigma_z$, $\omega_x=\omega_y=\omega_z$),
the self-consistency condition~(\ref{self_consistency_ho})
is also satisfied with the superdeformed configuration
($2\Sigma_x=2\Sigma_y=\Sigma_z$, $\omega_x=\omega_y=2\omega_z$).
In fact, the shape coexistence phenomena have been observed in
many areas throughout the nuclear chart \cite{HW11}.
\subsection{Fermi sphere in the rotating frame}
This idea of the isotropic velocity distribution
can be extended into the one
in the rotating frame (p.79 in BM2).
The local velocity in the rotating frame,
\begin{equation}
\vec{v}\equiv \vec{p}/m - (\vec{\omega}_{\rm rot}\times\vec{r}),
\label{v_rotating_frame}
\end{equation}
has an isotropic distribution $\rho_{\vec{r}}(v)$ at each $\vec{r}$.
The isotropic velocity distribution means no net current relative to
the rotating frame, which ends up with a rigid-body value for
the moment of inertia.
The deformed nucleus
would have a rigid-body value of moment of inertia if the pairing
correlations were absent.
The transformation from the laboratory frame to the rotating frame
leads to the cranking Hamiltonian
\begin{equation}
H'=H-\vec{\omega}_{\rm rot} \cdot \vec{J} ,
\label{cranking_hamiltonian}
\end{equation}
where $\vec{J}$ is the total angular momentum.
The velocity-dependent terms (kinetic energy and the centrifugal potential)
in the rotating frame can be written as
$p^2/(2m) -\vec{\omega}_{\rm rot}\cdot(\vec{r}\times\vec{p})
= mv^2/2 - m(\vec{\omega}_{\rm rot}\times\vec{r})^2/2$
where $\vec{v}$ is given by equation~(\ref{v_rotating_frame}).
This confirms the $\vec{v}$-dependence in $H'$ is isotropic ($\propto v^2$).
This isotropic velocity distribution is still valid in
rotating nuclei \cite{SM84}.
The cranking model~(\ref{cranking_hamiltonian}) plays
a key role in physics of high-spin nuclear structure
(sections~\ref{sec:coupling}, \ref{sec:SD}, and \ref{sec:wobbling}).
\section{SBS and collective motions}
\label{sec:collective_motion}
A broken continuous symmetry leads to the emergence of two types of
collective excitations.
One is the massless ANG mode
and the other is the massive Higgs mode \cite{PV15}.
Therefore, properties of the collective motions significantly change
before and after the SBS takes place.
In nuclei, we can observe them in the excitation spectra.
We can even see how the ANG and Higgs modes appear and evolve
from soft modes.
\subsection{Rotational motion; ANG mode}
The ANG mode is a gapless (massless) mode in the infinite system.
For the case of nuclear deformation,
the ANG mode correspond to the rotational motion of the deformed nucleus.
Because of the finiteness,
the spectrum is not exactly gapless, however,
shows a gradual emergence of the ``quasi-degenerate'' rotational spectra.
In Fig. 6-31 of BM2,
a typical example for even-even Sm isotopes ($^{144-154}$Sm)
is presented.
The $^{144}$Sm nucleus has the magic neutron number $N=82$.
Its ground state ($0^+$) is spherical and
the first excited state is located at excitation energy of 1.63 MeV.
Keeping the proton number the same and increasing the neutron number
two by two, we clearly observe the following:
\begin{enumerate}
\item The first $2^+$, $4^+$, $\cdots$ ($2I$) states lower their
excitation energies.
Eventually, a rotational band is formed to present the
excitation spectra, $E_I\propto I(I+1)$.
\item The second $0^+$ and the second $2^+$ states
lower their energies in the beginning.
However, they stop decreasing at $N=88$ ($^{150}$Sm).
\item Additional rotational bands are formed on top of
the second $0^+$ and $2^+$ states for $N\geq 90$ ($^{152,154}$Sm).
\end{enumerate}
In $^{154}$Sm, the excitation energy of the first $2^+$ state
is only 82 keV.
This is 1/20 of that in $^{144}$Sm and we may say that it is
approximately degenerate in the ground state.
Moreover, there appear five members ($0^+,\cdots,8^+$)
of rotational bands below 1 MeV of excitation.
It should be noted that similar phenomena
are observed in many region of nuclear chart,
when the neutron (proton) numbers are going away from the
spherical magic number.
A regular pattern of rotational spectra allows us to distinguish
the intrinsic excitations and the rotational motion.
A rotational band is constructed based on each intrinsic excitation
from the ground state.
From these observation, we may think of the Hamiltonian subtracting
the rotational energy, $H'\equiv H-{\vec{J}^2}/(2{\cal J})$.
$H'$ conserves the rotational symmetry,
however, the member of the rotational bands ($0^+$, $2^+,\cdots$)
will be degenerate in energy.
Then, a deformed wave-packet state which violates
the rotational symmetry becomes an eigenstate of the Hamiltonian $H'$.
The number of activated rotational degrees of freedom depends
on the nuclear shape.
For axially symmetric spheroidal shape, there is no collective
rotation around the symmetry axis.
In other words, the angular momentum component along the symmetry
axis (called $K$ quantum number in the following)
is purely determined by the intrinsic motion.
In this case, the two rotational axes perpendicular to the symmetry
($z$) axis are possible,
but they are equivalent in the sense that
they have equal moment of inertia, ${\cal J}_x={\cal J}_y$.
In contrast,
if the nucleus has an equilibrium shape away from the axial symmetry
(triaxial shape),
the collective rotations about three axes are all activated,
and they can have different moments of inertia,
${\cal J}_x$, ${\cal J}_y$, and ${\cal J}_z$.
We may expect that the rotational spectra become richer and more complex.
The wobbling motion is known to be a typical mode of excitation
in the triaxial nuclei,
that will be discussed in section~\ref{sec:wobbling}.
\subsection{Beta and gamma vibrations; amplitude (Higgs) mode}
\label{sec:beta_gamma_vib}
The quadrupole ($\lambda=2$) vibrations
produce $2^+$ excitations in spherical nuclei.
When the SBS takes place to produce
the prolate (spheroidal) ground state,
among five $\alpha_{2\mu}$ ($\mu=-2,\cdots,2$),
the two shape degrees ($\beta$ and $\gamma$) remain,
and rest of the degrees of freedom are absorbed in the rotational motion
(Euler angle $\Theta$).
For an axially symmetric ground state,
the normal modes can be classified by the vibrational angular
momentum along the symmetry axis, which is often denoted by
the quantum number $K$.
The $\beta$ and $\gamma$ vibrations correspond to $K^\pi=0^+$ and $2^+$,
respectively.
Note that the $K^\pi=1^+$ low-lying vibration does not exist, because
it corresponds to the rotation of the whole nucleus.
The $\beta$ vibration around the SBS minimum is associated with
a collective amplitude mode of order parameter with a finite energy gap,
in contrast to the ``gapless'' rotational motion.
This type of excitation is often referred to as the
``amplitude (Higgs) mode''~\cite{PV15}.
A number of those candidates have been observed in well-known
deformed regions, such as the rare-earth and actinide regions.
For instance, in the rare-earth region, the excitation energies
of $\beta$ vibration candidates are found at $E_x\approx 1$ MeV.
However, their $B(E2)$ values from the ground states to the $2^+$ states
in the $\beta$-vibrational bands are not large in most cases,
typically a few Weisskopf units.
Instead, strong population by the pair transfer reaction has been
observed in many $\beta$-vibration candidates.
Therefore, their true nature is still mysterious
and currently under debate~\cite{Gar01}.
We should note that
an important role of the Coriolis coupling in the $\beta$ vibrations
has recently been pointed out \cite{MU16}.
See also figure \ref{fig:beta} in section \ref{sec:GIR}.
In contrast, the $\gamma$ vibrations,
whose excitation energies are also around 1 MeV,
show $B(E2)$ values significantly larger than
the Weisskopf units.
Thus, the collective nature of the $\gamma$ vibration
is well established.
Effects of their coupling to the rotational motion have been
also studied within the generalized intensity relation
(section~\ref{sec:GIR_applications}).
For nuclei with the prolate shape,
a naive geometric consideration may predict that the vibrational frequency
along the symmetry axis ($K=0$) is lower than that of the $K\neq 0$.
This is true for high-frequency giant quadrupole resonance \cite{HW01}.
However, the low-lying $\beta$ and $\gamma$ vibrations do not
follow this simple expectation (section 6-3b in BM2).
They are much more sensitive to underlying shell structure.
\subsection{Octupole vibrations; Negative-parity modes}
\label{sec:octupole_vib}
Octupole vibrations ($\lambda=3$)
with negative parity have been systematically
observed in spherical and deformed nuclei.
In spherical nuclei, it produces $3^-$ state.
The most typical example is perhaps that in $^{208}$Pb.
It is split into four different normal modes with $K^\pi=0^-\sim 3^-$
in deformed nuclei.
Again, the geometric expectation for their ordering is not applicable
to low-frequency octupole vibrations;
Namely, the $K^\pi=0^-$ vibrational state is not necessarily the lowest
among the multiplet.
The rotational band is formed on top of the bandhead with the spin
$I=K$ for $K=1\sim 3$ and $I=1$ for $K=0$.
\subsection{A microscopic tool;
quasiparticle-random-phase approximation (QRPA)}
In normal degenerate Fermi systems, the most basic mode of excitation
at low energy
corresponds to the one-particle-one-hole (1p1h) excitations.
When the pairing correlations produce the pair-condensed (BCS-like)
ground state, the 1p1h excitations should be replaced by
the two-quasiparticle (2qp) excitations.
The quasiparticle, which is mixture of particle and hole states,
is usually defined as an eigenstate of
the Hartree-Fock-Bogoliubov (HFB) equation \cite{RS80,BR86}.
The ground state corresponds to the quasiparticle vacuum state.
The 2qp excitations include not only 1p1h states, but also
two-particle and two-hole states
which correspond to states in neighboring nuclei with $A\pm 2$.
The odd-$A$ nuclei are expressed by one-quasiparticle states
based on the quasiparticle vacuum.
The collective excitations, such as $\beta$, $\gamma$, and
octupole vibrations in sections \ref{sec:beta_gamma_vib}
and \ref{sec:octupole_vib}, are approximately given by
superposition of many 2qp excitations.
The most successful theory for this purpose is the QRPA \cite{RS80,BR86},
which can describe both collective and non-collective modes
of excitation.
The QRPA contains backward amplitudes corresponding to
2qp annihilation on the correlated ground state,
and respects the symmetry of the Hamiltonian \cite{RS80,BR86}.
The limitation of QRPA is associated with its small-amplitude nature.
The HFB equations with the cranking Hamiltonian $H'$
(\ref{cranking_hamiltonian}) is often utilized for studies of
high-spin nuclear structure.
The QRPA calculation with the cranking Hamiltonian $H'$
is able to describe the rotational coupling effects,
such as the alignment and stretching, on
the collective and the non-collective excitations.
Some examples of the QRPA calculations with $H'$ are presented in
the following sections \ref{sec:coupling}, \ref{sec:SD},
and \ref{sec:wobbling}.
\section{Coriolis coupling to intrinsic motions}
\label{sec:coupling}
The SBS of the translational symmetry produces the ANG mode of
the center-of-mass motion.
Since it is exactly decoupled,
the intrinsic motions are not affected by the speed of
the nucleus in the accelerator.
On the other hand, the rotational motion is not exactly decoupled,
thus, the Coriolis and centrifugal effects influences
intrinsic structure.
In the unified model,
as is mentioned in section~\ref{sec:beta_gamma_vib},
the five quadrupole variables
$\alpha_{2\mu}$ ($\mu=-2,\cdots,2$) are represented by
$(\beta,\gamma)$ and three Euler angles $\Theta$.
Accordingly, the total wave function is given by a product
of the intrinsic, the vibrational, and the rotational parts.
If the shape fluctuation is neglected,
one can write the total wave function as a product of
the rotor and the intrinsic parts.
When the nucleus has the axially symmetric shape,
the intrinsic state has a good $K$-quantum number, $|K_n)$.
\begin{equation}
\ket{\Psi_{K_n IM}}=\ket{K_n IM}\otimes |K_n) ,
\label{unified_model_wf}
\end{equation}
where the rotor wave function is given by
\begin{equation}
\inproduct{\Theta}{KIM}=
\left(\frac{2I+1}{8\pi^2}\right)^{1/2}\mathcal{D}^I_{\lambda K}(\Theta) .
\end{equation}
The additional $\mathcal{R}$ invariance
requires the symmetrization of
equation~(\ref{unified_model_wf}) for $K_n\neq 0$;
$\{\ket{\Psi_{K_n IM}} + (-1)^{I+K} \ket{\Psi_{\bar{K}_n IM}}\}/\sqrt{2}$.
The quantum nature of the angular momentum is properly treated
in this rotor wave function.
The Coriolis coupling,
which mixes states with different $K$ quantum numbers,
can be treated in a perturbative manner.
Chapter 4 of BM2 presents extensive discussion on this subject.
On the other hand, in the high-spin limit $I\rightarrow\infty$,
the semiclassical approximation works well.
The rotational frequency $\vec\omega_{\rm rot}$ is introduced,
which leads to the cranking model~(\ref{cranking_hamiltonian}).
Especially, when the direction of $\vec\omega_{\rm rot}$ is
parallel to a body-fixed principal axis $x$,
we have a uniform rotation $\omega_{\rm rot}(t)=\textrm{const}$.
\begin{equation}
H'=H-\omega_{\rm rot} J_x .
\label{cranking_hamiltonian_1D}
\end{equation}
This one-dimensional cranking model has been extensively applied
to high-spin nuclear structure problems with a tremendous success.
The non-linear effects of rotation are automatically
taken into account in the intrinsic structure,
which reproduces a number of striking high-spin phenomena,
such as back-bending, alignment, and band termination.
A drawback is the missing quantum nature of rotation,
particularly important at low spin.
\subsection{Quantization of the cranking model at low spin}
\label{sec:GIR}
In the semiclassical approximation, the direction of
$\vec\omega_{\rm rot}\ (\vec{I})$
is assumed to be the $x$ axis of both the intrinsic
(body-fixed) and the laboratory (space-fixed) frames.
The multipole operator $\tilde{Q}_{\lambda\mu}$,
in which $\mu$ is defined with respect to the $x$ axis,
changes the angular momentum $I$ to $I+\mu$.
Thus, a transition matrix element
between states with the angular momenta $I_i$ and $I_f$ is simply given by
$\bra{f}\tilde{Q}_{\lambda \mu}\ket{i}$, where
$\mu$ should be equal to $\Delta I=I_f-I_i$.
This is a good approximation at the high-spin limit
($I,\omega_{\rm rot}\rightarrow\infty$).
In contrast, at the low-spin limit ($I\sim 0$),
the angular momentum is coupled to the deformation,
thus, the $K$ quantum number along the symmetry ($z$)
axis is a good quantum number.
In this limit, the multipole operator ${Q}_{\lambda \nu}$,
defined with respect to the $z$ axis,
changes the $K$ quantum number by $\nu=\Delta K=K_f-K_i$.
In addition, the quantum mechanical nature of rotation is important
at low spin.
A perturbative expansion with respect to $I$ in the unified model
produces a specific $I$-dependence for the transition matrix element
({\it generalized intensity relations} in BM2).
To complete the intensity relation beyond the leading order,
we need to determine matrix elements of intrinsic operators
which take into account the Coriolis and centrifugal effects.
There is no systematic method to calculate these intrinsic matrix elements
in the unified model.
The cranking model~(\ref{cranking_hamiltonian_1D}), on the other hand,
is capable of microscopic treatment of
the rotational coupling to the intrinsic structure.
However, the semiclassical nature of the cranking model
forbids us to obtain the correct $I$-dependent intensity relations
at low spin.
This is mostly due to missing kinematics of the angular momentum algebra.
We present here a feasible prescription to recover the quantum mechanical
effect, that
enables us to calculate matrix elements of the intrinsic moments in
the generalized intensity relations.
\subsubsection{Generalized intensity relations}
The main idea is as follows \cite{SN96}.
In the high-spin limit, the cranking treatment becomes accurate and
the matrix elements of a multipole operator
$Q^{\rm (lab)}_{\lambda\mu}$
between the highest-weight states
are given by a relation \cite{Mar76,Mar77}
\begin{equation}
\bra{I_f I_f} Q^{\rm (lab)}_{\lambda \Delta I} \ket{I_i I_i}
=
(I_f| \tilde{Q}_{\lambda\Delta I} |I_i) ,
\label{high_spin_formula}
\end{equation}
where the state $|I)$ is a symmetry-broken state;
for instance, a mean-field solution of the
cranking Hamiltonian~(\ref{cranking_hamiltonian_1D}) with
the constraint $(I| J_x |I)=I$.
$\tilde{Q}_{\lambda \Delta I}$ can be expanded in terms of
those defined with respect to intrinsic $z$ axis,
and the coefficients are given by the $d$ functions.
\begin{equation}
\tilde{Q}_{\lambda\mu}=
\sum_\nu {\cal D}^\lambda_{\mu \nu}
\left( -\frac{\pi}{2}, -\frac{\pi}{2},0\right)
Q_{\lambda \nu}
= i^{-\mu} \sum_\nu d^\lambda_{\mu \nu}\left(-\frac{\pi}{2}\right)
Q_{\lambda \nu} .
\label{mu_to_nu}
\end{equation}
Now, let us do something not entirely correct.
Although the equality in equation~(\ref{high_spin_formula})
holds only at high spin,
we take the opposite low-spin limit ($I,\omega_{\rm rot}\rightarrow 0$),
in which the state $|I)$ becomes a ``non-cranked''
$K$-good intrinsic state $|K)$.
Substituting equation~(\ref{mu_to_nu}) into~(\ref{high_spin_formula}),
we have
\begin{equation}
\bra{I_f I_f} Q^{\rm (lab)}_{\lambda \Delta I} \ket{I_i I_i}_{\rm LO}
\leftrightarrow (K_f| \tilde{Q}_{\lambda\Delta I} |K_i)
= i^{-\Delta I} d^\lambda_{\Delta I \Delta K}
(K_f| Q_{\lambda \Delta K} |K_i) .
\label{low_spin_formula}
\end{equation}
Here, we use the symbol $\leftrightarrow$ instead of $=$ because
it is obtained by applying
the high-spin formula~(\ref{high_spin_formula}) to the low-spin limit.
For simplicity, we omit the argument $(-\pi/2)$ of the $d$ function.
The suffix ``LO'' indicates the relation in the zeroth order
$O(\omega_{\rm rot}^0)$, with respect to $\omega_{\rm rot}$.
Equation~(\ref{low_spin_formula}) is, of course,
not directly applicable to low spin.
However, it has a proper correspondence to the
leading order (LO) intensity relation in the unified model,
\begin{equation}
\bra{I_f I_f} Q^{\rm (lab)}_{\lambda \Delta I} \ket{I_i I_i}_{\rm LO}
= \bra{K_f I_f I_f} \mathcal{D}^\lambda_{\Delta I \Delta K}
\ket{K_i I_i I_i}
(K_f| Q_{\lambda \Delta K} |K_i) ,
\label{LO_unified_model}
\end{equation}
which is obtained using the $K$-good wave function~(\ref{unified_model_wf})
and the LO transformation of the multipole operator,
$Q_{\lambda \mu}^{\rm (lab)} =
\sum_\nu \mathcal{D}^\lambda_{\mu\nu} Q_{\lambda\nu}$.
Comparing equations~(\ref{low_spin_formula}) and (\ref{LO_unified_model}),
we may think of a quantization prescription,
\begin{equation}
d^\lambda_{\Delta I\Delta K}
\quad
\rightarrow
\quad
\bra{K_f I_f I_f} {\cal D}^\lambda_{\Delta I\Delta K} \ket{K_i I_i I_i}
.
\label{LO_quantization}
\end{equation}
Then, the ``non-cranked'' limit of the cranking formula
reproduces the LO intensity relation in the unified model.
This quantization procedure is supported by the fact that
the quantities in both sides of equation~(\ref{LO_quantization}) become
identical to the Clebsch-Gordan (CG) coefficients,
$\inproduct{I_i K_i \lambda \Delta K}{I_f K_f}$,
at the high-spin limit ($I\rightarrow\infty$).
Decreasing $I$, the left hand side of equation~(\ref{LO_quantization})
is losing its validity because of its classical nature,
while the right hand side stays valid keeping its quantum nature.
The present quantization of the cranking model is applicable to
higher-order Coriolis coupling terms.
These terms are not easily provided in the unified model.
The next leading order (NLO) is given by the first order in $\omega_{\rm rot}$,
which produces non-zero contributions of $Q_{\lambda \nu=\Delta K\pm 1}$.
The NLO terms to equation~(\ref{low_spin_formula}) are given as
\begin{eqnarray}
\bra{I_f I_f}Q^{\rm (lab)}_{\lambda\Delta I}\ket{I_i I_i}_{\rm NLO}
&\leftrightarrow&
i^{-\Delta I} \omega_{\rm rot} \left( d^\lambda_{\Delta I\ \Delta K+1}
\frac{d (K_f| Q_{\lambda\ \Delta K+1} |K_i)}{d\omega_{\rm rot}}
\right.
\nonumber \\
&& \quad\quad\quad
+ \left. d^\lambda_{\Delta I\ \Delta K-1}
\frac{d (K_f| Q_{\lambda\ \Delta K-1} |K_i)}{d\omega_{\rm rot}}
\right)
,
\end{eqnarray}
where the derivatives are evaluated at $\omega_{\rm rot}=0$.
A prescription of the NLO quantization is given by
\begin{equation}
\omega_{\rm rot}
d^\lambda_{\Delta I\ \Delta K\pm 1}
\quad\rightarrow\quad
\bra{K_f I_f I_f} \frac{1}{2 \mathcal{J}}
\left\{ I_\pm, {\cal D}^\lambda_{\Delta I\ \Delta K\pm 1} \right\}
\ket{K_i I_i I_i} ,
\label{NLO_quantization}
\end{equation}
where $\{A,B\}=AB+BA$ and
$I_\pm\equiv \mp (I_x \pm iI_y)/\sqrt{2}$ in the intrinsic frame.
${\cal J}$ is the moment of inertia of the rotational band,
which can be also calculated in the cranking model
at $\omega_{\rm rot}\rightarrow 0$:
$\mathcal{J} = (1/2)(
d(K_i|J_x|K_i)/d\omega_{\rm rot} +d(K_f|J_x|K_f)/d\omega_{\rm rot})
\approx d(K_i|J_x|K_i)/d\omega_{\rm rot}
\approx d(K_f|J_x|K_f)/d\omega_{\rm rot}$.
Again, in the high-spin limit, the left and right hand sides of
equation~(\ref{NLO_quantization}) become identical,
if we assume $\omega_{\rm rot}\approx I_i/\mathcal{J}\approx I_f/\mathcal{J}$.
In summary, the generalized intensity relation up to the NLO
is obtained by calculating the matrix element
$\bra{I_f I_f} Q^{\rm (lab)}_{\lambda\Delta I} \ket{I_i I_i}=
\bra{K_f I_f I_f} Q^{({\rm LO+NLO})}_{\lambda\Delta I} \ket{K_i I_i I_i}$,
using the operator
\begin{equation}
\fl
Q^{\rm (LO+NLO)}_{\lambda\Delta I}
= m_{\lambda\ \Delta K}^{(0)}
\mathcal{D}^\lambda_{\Delta I \Delta K}
+ \frac{m_{\lambda\ \Delta K+1}^{(+1)}}{2}
\left\{ I_+, {\cal D}^\lambda_{\Delta I\ \Delta K+1} \right\}
+\frac{m_{\lambda\ \Delta K-1}^{(-1)}}{2}
\left\{ I_-, {\cal D}^\lambda_{\Delta I\ \Delta K-1} \right\} ,
\label{LO+NLO}
\end{equation}
where the intrinsic matrix elements are given by
\begin{equation}
m_{\lambda\ \Delta K}^{(0)}
= (K_f| Q_{\lambda \Delta K} |K_i) , \qquad
m_{\lambda\ \Delta K\pm 1}^{(\pm 1)}
= \frac{1}{\mathcal{J}}
\frac{d (K_f| Q_{\lambda\ \Delta K\pm 1} |K_i)}{d\omega_{\rm rot}} .
\label{intrinsic_moments_LO+NLO}
\end{equation}
The right hand sides of these equations can be calculated
with the cranking model~(\ref{cranking_hamiltonian_1D})
in the vicinity of $\omega_{\rm rot}\rightarrow 0$.
Note that the $\mathcal{R}$-conjugate terms should be added
in the right hand side of equation~(\ref{LO+NLO})
when the $\mathcal{R}$ invariance is present \cite{SN96}.
\subsubsection{Applications}
\label{sec:GIR_applications}
The cranking model~(\ref{cranking_hamiltonian_1D})
has been applied to calculation of the intrinsic moments
in equation~(\ref{intrinsic_moments_LO+NLO}).
For low-lying quadrupole vibrational excitations,
we use the QRPA to calculate the intrinsic matrix elements.
For even-even nuclei, the ground state is $|0)=|K=0)$
and the vibrational state is given by the QRPA normal-mode creation
operator $\hat{X}_K^\dagger$ as $|K)=\hat{X}_K^\dagger |0)$.
The QRPA calculation is based on the cranked-Nilsson-BCS model
with a residual multipole interaction of a separable form similar to
equation~(\ref{separable_interaction}).
We reported these results for quadrupole and
octupole vibrations in the even-even rare-earth nuclei
in reference~\cite{SN96} in which the details of the model can be found.
In the left panel of figure~\ref{fig:Er_Hf},
we present an example of the Mikhailov plot
for the $\gamma$ vibrations in $^{166}$Er.
The LO+NLO electric quadrupole operator in a form
of equation~(\ref{LO+NLO}) leads to the intensity relation
between the $K_i=2$ ($\gamma$) band and the $K_f=0$ (ground) band,
\begin{equation}
\fl\qquad
\frac{\bra{I_f K_f^\pi=0^+_g}|M(E2)|\ket{I_i K_i^\pi=2^+_\gamma}}
{\sqrt{2I_i+1}\inproduct{I_i 2 2 -2}{I_f 0}} =
Q_t \left[ 1+ q \left\{ I_f(I_f+1)-I_i(I_i+1) \right\} \right] ,\quad
\label{GIR}
\end{equation}
where $Q_t$ and $q$ are obtained from the intrinsic moments
(\ref{intrinsic_moments_LO+NLO}),
though some modification is necessary
because of the $\mathcal{R}$ invariance.
See reference~\cite{SN96} for their exact formulae.
\begin{figure}[tb]
\centerline{
\includegraphics[clip,width=0.85\textwidth]{Er-Hf.eps}
}
\caption{
(Left) $E2$ transition amplitudes for the $\gamma$ vibrational
band in $^{166}$Er.
The experimental data are taken from figure 4-30 in BM2.
Adapted from reference \cite{NS99-P}.
(Right)
Hindrance factors of $B(E2;6^+_{K=6} \rightarrow 4^+)$
for decay of $K^\pi=6^+$ isomers in Hf isotopes.
Calculated values are shown by circles, while
the squares are the experimental data.
See text for details.
Adapted from reference \cite{NS99-P}.
}
\label{fig:Er_Hf}
\end{figure}
\begin{figure}[htb]
\begin{center}
\includegraphics[clip,width=0.48\textwidth]{exp-cal_be2_gd_3_both_2.eps}
\includegraphics[clip,width=0.46\textwidth]{exp-cal_be2_gd_3_both_small_anti_2.eps}
\end{center}
\caption{
(Left) Calculated and experimental $B(E2;0^+_\beta\rightarrow 2^+_g)$
values in Gd isotopes.
Adapted from reference~\cite{MU16}.
(Right) $B(E2;0^+_g\rightarrow 2^+_\beta)$ values in Gd isotopes.
Experimental data are taken from
references~\cite{Gar01,Mar13,Rei09,Rei12,Hel04,Rei05}.
Note that the scale of the ordinate is 1/10 of that in
the left panel.
}
\label{fig:beta}
\end{figure}
A very similar figure is shown in Fig. 4-30 of BM2,
however, we note here that the parameters $(Q_t,q)$
in the left panel of figure~\ref{fig:Er_Hf}
is based on the microscopic calculation, while those in BM2 are
determined by fitting the experimental data.
The LO relation produces the same $Q_t$ but $q=0$.
The Coriolis coupling effect in the NLO is represented
by the parameter $q$ ($m^{(\pm 1)}_{2\ \mp 1}$).
For the $\gamma$ vibrations, the cranking calculation always
produces $q>0$ in the rare-earth nuclei \cite{SN96}, which suggests
the transitions $I_\gamma \rightarrow I_g$ are enhanced (hindered)
for $I_g>I_\gamma$ ($I_g < I_\gamma$).
This is consistent with experimental data (see reference \cite{SN96} and
references therein).
We also obtain a reasonable agreement for the $M1$ transitions.
However, the calculated sign of the $M1/E2$ mixing amplitudes changes
from nucleus to nucleus, while the observed values are always negative
\cite{SN96}.
The collectivity of the $\beta$ vibrations measured by the strengths of the
interband transitions to/from the ground band is
weaker than that of the $\gamma$ vibrations in most cases.
However, in rare-earth nuclei in (near) the transitional region,
the $\beta$ vibrations produce
very low excitation energies and
large $B(E2;0^+_\beta\rightarrow 2^+_g)$ values.
In Gd isotopes, for example, their excitation energies are
681 keV and 615 keV for $^{154}$Gd ($N=90$) and $^{152}$Gd ($N=88$),
respectively.
The $B(E2;0^+_\beta\rightarrow 2^+_g)$ values
amount to $52\pm 8$ and $178^{+53}_{-33}$ W.u. \cite{Rei09,Mar13}.
We expect similar values of $B(E2;0^+_g\rightarrow 2^+_\beta)$,
which are predicted to be identical to
$B(E2;0^+_\beta\rightarrow 2^+_g)$ in the LO relation.
Surprisingly, the observed $B(E2;0^+_g\rightarrow 2^+_\beta)$ values are
much smaller than $B(E2;0^+_\beta\rightarrow 2^+_g)$ \cite{Gar01,Rei09,Mar13}.
Figure~\ref{fig:beta} shows the calculated $B(E2)$ values using the
the LO+NLO intensity relation identical to equation (\ref{GIR})
with some trivial changes in the left hand side
($K_i^\pi=2_\gamma^+ \rightarrow K_i^\pi=0_\beta^+$,
$\inproduct{I_i 2 2 -2}{I_f 0}\rightarrow
\inproduct{I_i 0 2 0}{I_f 0}$).
The LO relation cannot account at all for
both large
$B(E2;0^+_\beta\rightarrow 2^+_g)$ and
small $B(E2;0^+_g\rightarrow 2^+_\beta)$ values.
Owing to relatively small moments of inertia
($\mathcal{J}$) for these transitional nuclei,
the inclusion of the NLO terms with large values of $q$
nicely reproduces both of them.
In BM2 (pp.168--175), the band mixing
between the ground and the $\beta$ bands in $^{174}$Hf are presented
to explain the observed intensity relations.
An effect of hindrance of the shape fluctuation
induced by the rotation,
suggested in references \cite{HNMM09,HSNMM10,HSYNMM11},
may also play an important role.
The Coriolis coupling effects may be a key ingredient
to understand the peculiar $B(E2)$ properties of the $\beta$-vibrational bands.
Generally speaking, the Coriolis-coupling effect for the quadrupole
vibrations is relatively weak, because the low-lying
$K^\pi=1^+$ collective state is missing
(section~\ref{sec:beta_gamma_vib}).
In contrast, all the members of the multiplet are present for
the negative-parity octupole vibrations.
Thus, we expect stronger Coriolis effects.
The cranking calculation actually predict the NLO parameters
of the octupole vibrations
($|q|\sim 0.1$) larger than those of $\gamma$ vibrations ($|q|\sim 0.01$)
\cite{SN96}.
The $K$ quantum number of the lowest mode of excitation
among the octupole multiplet ($K^\pi=0^-\sim 3^-$)
changes from nucleus to nucleus.
Nevertheless, the lowest mode always has $q<0$ for transitions from
the octupole band ($K_i$) to the ground band ($K_f=0$),
Thus, $B(E3; 3^-\rightarrow 0^+)$ is enhanced for the lowest mode.
This Coriolis effect is clearly seen in Gd isotopes in figure~\ref{fig:Gd}.
\begin{figure}[tb]
\centerline{\includegraphics[clip,width=0.7\textwidth]{Gd_2.eps}}
\caption{
$E3$ transition amplitudes for the octupole vibrational states
in Gd isotopes, $I_i^\pi=3^- \rightarrow I_f^\pi=0^+$.
Open and filled circles correspond to calculated values
in the LO and LO+NLO, respectively,
compared with experimental data (open squares)
\cite{MM81}.
The lowest mode of excitation among the octupole multiplet is
$K=1$ for $^{156,158}$Gd and $K=2$ for $^{160}$Gd.
From reference \cite{NS97-P}.
}
\label{fig:Gd}
\end{figure}
Another application is presented here for $K$-forbidden transitions
in decays of the high-$K$ isomers.
For $K$-forbidden transitions with $|\Delta K| > \lambda$,
we should extend the LO+NLO relation of equation~(\ref{LO+NLO})
because the LO term vanishes.
The order of $K$ forbiddenness is defined by $n=|\Delta K|-\lambda$.
The N$^n$LO+N$^{n+1}$LO intensity relation for $\Delta K > 0$ is given by
\cite{SN96}
\begin{equation}
\fl\qquad
Q^{{\rm (N}^n+{\rm N}^{n+1}{\rm LO)}}_{\lambda\Delta I}
= m_{\lambda\ \lambda}^{(-n)}
\mathcal{D}^\lambda_{\Delta I \lambda} (I_-)^n
+\frac{m_{\lambda\ \lambda-1}^{(-n-1)}}{2}
\left\{ (I_-)^{n+1}, {\cal D}^\lambda_{\Delta I\ \lambda-1} \right\}
+ \mathcal{R}{\rm -conj.} ,\quad
\end{equation}
where the intrinsic moments are
\begin{equation}
\eqalign{
m_{\lambda\ \pm\lambda}^{(\mp n)}
= \frac{1}{n! \mathcal{J}^n}
\frac{d^n (K_f| Q_{\lambda\ \pm\lambda} |K_i)}{d\omega_{\rm rot}^n} ,\cr
m_{\lambda\ \pm(\lambda-1)}^{(\mp(n+1))}
= \frac{1}{(n+1)! \mathcal{J}^{n+1}}
\frac{d^{n+1} (K_f| Q_{\lambda\ \pm(\lambda-1)} |K_i)}{d\omega_{\rm rot}^{n+1}}
.
}
\end{equation}
These formulae are applied to the two-quasiparticle (2qp)
$K^\pi=6^+$ isomers in Hf isotopes.
The configuration
of the initial state $|K_i=6)$
is assumed to be the proton 2qp
$[402\ 5/2]\otimes[404\ 7/2]$.
The hindrance factors are shown in the right panel of figure~\ref{fig:Er_Hf}.
This is defined by
\begin{equation}
F\equiv \frac{B(E2;6^+\rightarrow 4^+)_{W}}{B(E2;6^+_{K=6} \rightarrow 4^+_g)}
,
\end{equation}
where $B(E2)_W$ is the Weisskopf estimate of the reduced transition probability.
A large hindrance factor means a long life time of the high-$K$
isomer state.
The calculated values are shown
in the right panel of figure~\ref{fig:Er_Hf}
by filled symbols (circles and triangles),
which are compared with experiment (filled squares).
The calculation qualitatively reproduces the experimental trends.
However, these values turn out to be quite sensitive to the details of
the quasiparticle spectra.
For instance, the triangles are obtained with slightly larger values of
the proton chemical potentials (by 70 keV) than those of circles.
The calculated hindrance factors differ by about one order of magnitude.
The effect of the residual interaction is very significant too.
The open symbols in figure~\ref{fig:Er_Hf} (right) show results including
the spin-spin interaction, $V_0\vec{\sigma}\cdot\vec{\sigma}$,
with $V_0=100$ keV, which roughly accounts for the Gallagher-Moszkowski
splitting.
This could change $F$ by two order of magnitude.
Nonetheless, the neutron number dependence is rather universal,
that indicates the largest hindrance at $N=104$.
\subsection{Cranking model at high spin: QRPA in the rotating frame}
\label{sec:cranking_at_high_spin}
Rotating the nucleus very fast,
the perturbative treatment in section~\ref{sec:GIR}
becomes no longer valid.
Instead, the semiclassical treatment in the cranking model
is better justifiable, and
the direct application of the cranking model~(\ref{cranking_hamiltonian_1D})
has been extensively performed for a variety of high-spin phenomena.
For instance, the famous back-bending phenomena have been studied
and understood as the crossing between the ground band and
an aligned 2qp band.
This can be also interpreted as breaking a Cooper pair
condensed in the ground band by the Coriolis anti-pairing effect.
In order to investigate properties of elementary modes of excitation
at high spin, it is useful to use the ``routhian'' $E'(\omega_{\rm rot})$
as a function of the rotational frequency $\omega_{\rm rot}$.
The routhian here is defined as the eigenenergies of the
cranking Hamiltonian~(\ref{cranking_hamiltonian_1D}),
which can be interpreted as the energy in the rotating frame
with the rotational frequency $\omega_{\rm rot}$.
To make a comparison, we often convert
the experimental excitation energy as a function of $I$, $E(I)$,
into the ``routhian'' $E'(\omega_{\rm rot})$.
This is done as follows.
First, from experimental rotational spectra $E_b(I)$,
we calculate the frequency, $\omega_{\rm rot}(I)=dE_b(I)/dI$.
Here, $b$ is the index of the rotational band.
According to the cranking Hamiltonian~(\ref{cranking_hamiltonian_1D}),
the routhian is defined as
$E'_b(\omega_{\rm rot})=E_b(I)-\omega_{\rm rot}I_x(I)$
with $I_x(I)=\sqrt{(I+1/2)^2-K^2}$,
at discrete values of $\omega_{\rm rot}(I)$.
The reference routhian $E'_{\rm ref}(\omega_{\rm rot})$ is defined,
for instance, by fitting that of the ground-state band
(``{\it b}''=``g.s.'').
Then, the excitation routhian relative to the reference band
as a function of $\omega_{\rm rot}$ is obtained
as $E'_{\rm ex}(\omega_{\rm rot})=E'_b(\omega_{\rm rot}) -
E'_{\rm ref}(\omega_{\rm rot})$ for each band ``$b$''.
In Ref.~\cite{BF79}, the experimental routhians in odd nuclei,
which were obtained by adopting the reference band fitting the
ground-state band in neighboring even-even nuclei,
show nice agreement with the calculated quasiparticle routhians.
\begin{figure}[tb]
\centerline{
\includegraphics[clip,width=0.45\textwidth]{u238.eps}
\includegraphics[clip,width=0.4\textwidth]{u238_align.eps}
}
\vspace{-15pt}
\caption{(Left)
Excitation routhian plot as functions of rotational frequency
$\omega_{\rm rot}$ for negative-parity states in $^{238}$U.
Large, medium, and small circles indicate the QRPA solutions with
$E3$ transition amplitudes larger than 300 $e$fm$^3$,
larger than 150 $e$fm$^3$, and less than 150 $e$fm$^3$, respectively.
The red (blue) ones correspond to the signature even (odd) states
with even (odd) $I$.
Experimental routhians are plotted by open squares.
(Right)
Aligned angular momentum as a function of $\omega_{\rm rot}$
for the lowest ($K^\pi=0^-$) and the second lowest ($K^\pi=1^-$)
octupole bands
in the lower and upper panels, respectively.
Open squares indicate experimental data \cite{War96}.
Adapted from reference \cite{Nak96-P}.
}
\label{fig:U}
\end{figure}
The routhian plot for the octupole vibrational bands
in $^{238}$U are presented in figure~\ref{fig:U}.
See references~\cite{NMMS96,Nak96-P} for details of the calculation.
In the right panels, the alignment, defined by
$i\equiv -dE'/d\omega_{\rm rot}$, is shown.
The alignment indicates the aligned component of the angular
momentum carried by the vibrational excitation.
For the lowest octupole band with $K^\pi=0^-$,
the alignment quickly increases up to $i\sim 3$,
which suggests that the angular momentum of the octupole phonon
is almost fully aligned along the rotational axis.
Then, at high spin around $\omega=0.25$ MeV,
it suddenly jumps up, which suggests the breakdown of the
collective vibration by a strong Coriolis force.
At $\omega_{\rm rot}\gtrsim 0.25$ MeV,
it becomes an aligned 2qp state.
This is seen in the left panel too.
The octupole collectivity (size of the circles) suddenly decreases
around $\omega_{\rm rot}=0.25$ MeV.
In contrast to the lowest band, the second lowest $K^\pi=1^-$ band
with even $I$ shows a gradual increase in the alignment, which may
suggest the gradual change of the octupole phonon into the aligned
2qp structure.
The present calculation nicely agrees with the experimental data.
The argument here suggests that the vibrational excitations based
on the yrast (ground-state) band tend to lose their collective character
at high spin, due to the intrusion of the aligned 2qp
states at low energy.
In this respect, the nucleus with larger deformation may be
better suited for the observation of the rapidly rotating
vibrational bands.
This is because the large deformation tends to hinder the alignment
of the quasiparticles.
Next, let us discuss such a case, the high-spin superdeformed (SD) bands.
\section{Elementary excitations in superdeformed rotational bands}
\label{sec:SD}
The SD state is characterized by a large prolate
deformation of approximate two-to-one axis ratio.
The shell structure at the SD shape is very different
from that near the spherical shape.
Since the low-lying modes of excitation strongly depend on the
underlying shell structure,
we may expect some new features in their properties.
\subsection{Octupole vibrations with $K^\pi=0^-$ and $2^-$}
One of the most striking features in the SD shell
structure is that the single-particle levels with
opposite parity ($\pi=\pm$) coexist in a single shell.
Adopting a simple harmonic oscillator potential,
one can easily understand this fact:
Namely, for $\omega_x=\omega_y=2\omega_z$,
an orbital with the oscillator quanta $(n_x,n_y,n_z)$
is degenerate in energy with those of
$(n_x\mp 1,n_y,n_z\pm 2)$ and $(n_x,n_y\mp 1,n_z\pm 2)$.
Since the parity is determined by the total quanta $N=n_x+n_y+n_z$,
they have different parity.
Another feature is that the observed SD bands are located around
the closed shell configurations corresponding
to the SD magic numbers (See figure 6-50 in BM2).
In contrast, the normally deformed nucleus is a consequence of
the SBS and has an open-shell configuration away from the spherical
magic numbers.
From these simple analysis, we may expect that
the collective negative-parity modes of excitation appear at low energy.
The QRPA based on the cranked Nilsson-BCS model is applied to
SD bands in the $A=190$ region.
The calculation predicts that the $K^\pi=2^-$ octupole states
are particularly low in energy, around $E'_x\lesssim 1$ MeV.
Especially, in $^{194}$Hg and $^{196}$Pb with $N=114$,
very collective $K^\pi=2^-$ octupole vibrations appear well below
1 MeV and their excitation routhians are roughly constant
with very little signature splitting \cite{NMMS96}.
See the right panel of figure~\ref{fig:Hg}.
Later, the interband $E1$ transitions between the octupole and ground SD
bands have been measured for $^{194}$Hg \cite{Hac97}
and for $^{196}$Pb \cite{Ros01},
which confirms nice agreement with calculated routhians
and the strong octupole collectivity.
For $N=110$, the calculation predicts an aligned octupole phonon,
shown in the left panel of figure~\ref{fig:Hg},
similar to the lowest octupole band in $^{238}$U
in figure~\ref{fig:U}.
This also nicely reproduces the experimental routhians in an
excited SD band in $^{190}$Hg \cite{Wil96}.
Later, the linear polarization measurement confirms the
aligned octupole vibrations \cite{Kor01}.
\begin{figure}[tb]
\centerline{
\includegraphics[width=0.8\textwidth]{Hg_2.eps}
}
\caption{
Excitation routhian plot for negative-parity excitations in SD $^{190}$Hg
(left) and $^{194}$Hg (right).
Open and filled circles correspond to the states
with even and odd signatures, respectively.
The size of circles represent the $E3$ transition amplitudes
($>200$ $e$fm$^3$, $>100$ $e$fm$^3$, and $<100$ $e$fm$^3$)
Experimental excitation routhians in $^{190}$Hg are shown by stars ($*$).
From reference \cite{Nak96-P}.
}
\label{fig:Hg}
\end{figure}
In the $A=150$ region, we theoretically predicted
a possible candidate of $K^\pi=0^-$ octupole band in SD
$^{152}$Dy \cite{NMMN95}.
It shows a rather constant excitation routhians in a wide range
of $\omega_{\rm rot}=0\sim 0.7$ MeV.
Later in 2002, its octupole character has been confirmed by the
measurement of the interband transitions and the spin identification
\cite{Lau02-1,Lau02-2}.
The $\omega_{\rm rot}$-dependence of the routhian well agrees
with the theoretical prediction.
\subsection{Soft mode with $K^\pi=1^-$}
\label{sec:banana_SD}
\begin{figure}[tb]
\includegraphics[width=0.3\textwidth]{figure_of_eight.eps}
\includegraphics[clip,width=0.7\textwidth]{HO_2.eps}
\caption{
(Left)
Typical classical periodic orbits in a potential with the $a:b:c=2:2:1$
shell structure.
From reference \cite{MNA93-P}.
(Right)
Calculated excitation energies for octupole vibrations in
the SD harmonic oscillator potential, with
stronger pairing (a) and weak pairing interactions (b).
From reference \cite{NMM92}.
}
\label{fig:HO}
\end{figure}
So far, the octupole vibrational excitations in SD rotational bands
have been observed (confirmed) only for the $K^\pi=0^-$ and $2^-$ modes.
Theoretically, these modes are predicted to appear lower than other
modes ($K^\pi=1^-$ and $3^-$), near the SD magic numbers.
However, moving away from the magic closed configurations,
the $K^\pi=1^-$ modes become a soft mode.
In order to investigate the soft mode in the SD shape,
we again follow the discussion in BM2 (pp. 591$-$598),
extending the argument for spherical potentials
in section~\ref{sec:shell_structure} to a deformed one.
This is possible if the motion is separable in the three coordinates,
such as the harmonic oscillator potential.
In a deformed harmonic oscillator potential,
the single-particle energy is specified by there numbers of
the oscillator quanta, $(n_x,n_y,n_z)$.
Thus, the shell structure is characterized by the ratio of
three integers, $a:b:c$,
and the shell frequency given by
\begin{equation}
\fl\qquad
\omega_{\rm sh}\equiv \frac{1}{a}
\left(\frac{\partial\epsilon}{\partial n_x}\right)_0
= \frac{1}{b}
\left(\frac{\partial\epsilon}{\partial n_y}\right)_0
= \frac{1}{c}
\left(\frac{\partial\epsilon}{\partial n_z}\right)_0 .
\end{equation}
Since $a:b:c=1:1:1$ correspond to the spherical harmonic oscillator,
the simplest integer ratio next to $1:1:1$
is $a:b:c=1:1:2$ and $2:2:1$.
The SD shape we are discussing here corresponds to the latter one,
which has the prolate shape.
The frequency ratio of $a:b:c=2:2:1$
($\omega_x=\omega_y=2\omega_z=2\omega_{\rm sh}$)
creates periodic orbits
shown in the left panel of figure~\ref{fig:HO}.
These are orbits of ``bending figure of eight''.
Since the shape of the classical periodic orbits is related to the
soft mode,
the SD state, with many nucleons outside the closed shell,
may be unstable against the banana-shaped bending mode.
In figure~\ref{fig:HO}, we show the result of the QRPA calculation
with the separable octupole interaction, based on
the SD harmonic oscillator potential \cite{NMM92}.
The $K^\pi=0^-$ and $2^-$ modes are the lowest near the SD magic numbers.
These modes are rather insensitive to the number of nucleons outside
the closed shell.
However, the $K^\pi=1^-$ octupole mode dramatically decreases its energy as
increasing
the number of
valence nucleons.
With enough number of valence nucleons,
the bending $K^\pi=1^-$ mode leads to the instability.
According to the qualitative discussion on the SD shell structure,
Bohr and Mottelson have already pointed out the possibility of
this instability toward the bending shape, in the context of
fission path (p.598 in BM2).
As far as we know, this effect on the fission dynamics
has not been fully studied so far.
\section{Nuclear wobbling motion and precession}
\label{sec:wobbling}
Most of the existing experimental data are known to be
consistent with
the interpretation based on the axially symmetric deformation.
Even the octupole deformation (section~\ref{sec:shell_structure})
observed in heavy nuclei
is associated with the axially symmetric one ($Y_{30}$).
In section~\ref{sec:banana_SD}, we have presented
a possible exotic nuclear shape in SD nuclei away from
the closed shell
configuration,
that breaks both the axial and the parity symmetry.
However, it has not been observed in experiments.
Bohr and Mottelson gave extensive discussion on the spectra of
triaxial nuclei in BM2.
In the beginning of section 4-5, they said
``{\it Although, at present {\rm (1975)},
there are no well-established examples of nuclear
spectra corresponding to asymmetric equilibrium shapes, it appears likely
that such spectra will be encountered in the exploration of nuclei
under new conditions (large deformations, angular momentum, isospin, etc.).}''
They were absolutely right.
The identification of the static triaxial deformation
has been a longstanding issue
in the nuclear structure physics.
One of the difficulties is to confirm its ``static'' nature
clearly distinguished from the ``dynamic'' one.
The observation of a wobbling band was
a breakthrough that provided a clear indication
of the non-uniform three-dimensional rotation of a triaxial nucleus.
We have really encountered this new phenomenon at
{\it large deformation and angular momentum}.
The first observation of the wobbling band was in $^{163}$Lu
\cite{Ode01,Jen02}.
At high spin ($I\gtrsim 20$),
several regular rotational bands with large moments of inertia
come down to the yrast region.
The deformation of these bands have been speculated to be large
($\epsilon\sim 0.4$) and triaxial ($\gamma\sim 20^\circ$),
according to calculations of the total routhian surface (TRS)
with the Nilsson potential~\cite{Gor04}.
They are called the triaxial superdeformed (TSD) bands
and labeled as TSD1, TSD2, etc.
In addition to the stretched intraband transitions ($\Delta I=2$),
the linking transitions ($\Delta I=1$)
between TSD2 and TSD1 (yrast),
between TSD3 and TSD2, and
between TSD3 and TSD1, have been observed.
The $E2$ character of these interband transitions
is experimentally confirmed~\cite{Gor04} and
their large strengths
nicely correspond to the estimate by a simple triaxial rotor model.
The measured $B(E2)$ values for the interband transitions
are order of 100 W.u.
which is
considerably larger than those of
the most collective $\gamma$ vibrations.
\subsection{Rotor model analysis of the wobbling in the high-spin limit}
\label{sec:rotor_model}
The prediction based on the rotor model given by Bohr and Mottelson
(section 4-5e in BM2)
is recapitulated here.
The rotor Hamiltonian contains three different moments of inertia,
$\mathcal{J}_x > \mathcal{J}_y > \mathcal{J}_z$,
with respect to the principal axes in the body-fixed frame.
\begin{equation}
\fl\qquad
H_{\rm rot}=\frac{J_x^2}{2\mathcal{J}_x}
+\frac{J_y^2}{2\mathcal{J}_y}
+\frac{J_z^2}{2\mathcal{J}_z}
=\frac{\vec{J}^2}{2\mathcal{J}_x}
+\left(\frac{1}{2\mathcal{J}_y}-\frac{1}{2\mathcal{J}_x}\right)J_y^2
+\left(\frac{1}{2\mathcal{J}_z}-\frac{1}{2\mathcal{J}_x}\right)J_z^2
.
\label{triaxial_rotor}
\end{equation}
For the lowest energy (yrast) state at a given $I$,
the term proportional to $\vec{J}^2$ in this Hamiltonian
is dominant at high spin ($I\rightarrow\infty$).
This corresponds to a uniform rotation around the $x$ axis:
$E_I\approx I(I+1) /(2\mathcal{J}_x)$.
In this high-spin limit, we assume $J_x\approx I$
which can be treated as a $c$-number.
The remaining terms of the Hamiltonian~(\ref{triaxial_rotor})
can be diagonalized, $[H_{\rm rot},X_{\rm wob}^\dagger]=\omega_{\rm wob}
X_{\rm wob}^\dagger$, by a linear transformation.
The normal-mode (wobbling phonon) creation operator
\begin{equation}
X_{\rm wob}^\dagger \equiv a \frac{iJ_y}{\sqrt{2I}} - b \frac{J_z}{\sqrt{2I}} ,
\end{equation}
with the normalization
$\left[X_{\rm wob},X_{\rm wob}^\dagger \right] = 1$
leads to the following relations:
\begin{eqnarray}
\frac{a}{b}&=&
\sqrt{
\left(\frac{1}{\mathcal{J}_y}-\frac{1}{\mathcal{J}_x}\right)
\left(\frac{1}{\mathcal{J}_z}-\frac{1}{\mathcal{J}_x}\right)^{-1} } ,
\quad\quad ab=1 ,
\\
\omega_{\rm wob}
&=& I
\sqrt{
\left(\frac{1}{\mathcal{J}_y}-\frac{1}{\mathcal{J}_x} \right)
\left(\frac{1}{\mathcal{J}_z}-\frac{1}{\mathcal{J}_x} \right) }
.
\label{wobbling_frequency}
\end{eqnarray}
The operator for the wobbling phonon number is given by
$n\equiv X_{\rm wob}^\dagger X_{\rm wob}$.
In this way, the rotor Hamiltonian can be written as a sum of
the rotation around the $x$ axis and the wobbling phonon excitation:
$E_{In}\approx I(I+1)/(2\mathcal{J}_x) + \omega_{\rm wob} (n+1/2)$.
A schematic figure for these spectra is shown in figure~\ref{fig:wobbling}.
In order to realize this kind of multiple band structure
from one intrinsic configuration, the nucleus should be able
to rotate about all three principal axes.
Therefore, the nuclear shape must be triaxial.
In addition, among the three moments of inertia,
the one along the major rotational axis $\mathcal{J}_x$ must be the largest.
\begin{figure}[tb]
\centerline{
\includegraphics[clip,width=0.45\textwidth]{schmwobb1_2.eps}
\hspace*{0.05\textwidth}
\includegraphics[clip,width=0.18\textwidth]{wobangm_2.eps}
}
\caption{
A schematic illustration of the wobbling motion.
(Left) Excitation spectra with $\Delta I=2$ and $\Delta I=1$ sequences.
(Right) A wobbling phonon excitation tilts the direction of
the angular momentum from the $x$-axis.
From reference \cite{SS06}.
}
\label{fig:wobbling}
\end{figure}
According to the LO high-spin formula~(\ref{high_spin_formula}),
the intraband $B(E2;\Delta I=2)$ strengths are proportional to
the quadrupole deformation $(\tilde{Q}_{22})^2$.
The $\Delta I=1$ transitions, which are associated with the wobbling
transitions, appear in the NLO with respect to $1/I$;
$B(E2;\Delta I=1)/B(E2;\Delta I=2) \sim 1/I$.
These $E2$ transition strengths
were explicitly given in chapter 4-5e in BM2.
The TRS calculation predicts the
``positive'' $\gamma$ shape
($\gamma\sim 20^\circ$)
for the TSD bands in $^{163}$Lu.
Here we use the so-called Lund convention for
the triaxiality parameter $\gamma$
in relation to the main rotation axis~\cite{And76},
where for the positive $\gamma$ shape, $0 < \gamma < 60^\circ$,
the rotation ($x$) axis is the shortest principal axis,
while for the negative $\gamma$ shape, $-60^\circ < \gamma < 0$,
it is the intermediate principal axis.
Then the out-of-band $E2$ transitions for the positive $\gamma$ shape
satisfy
a relation, $B(E2;I\rightarrow I-1)> B(E2;I\rightarrow I+1)$.
This is consistent with the experiments, in which only
the $I\rightarrow I-1$ transitions have been observed.
The simple rotor model picture, however, disagrees with
the observed data with respect to the following
points:
\begin{itemize}
\item
At $\gamma=20^\circ$ which is supported by the TRS calculation,
the $\gamma$-dependence of the irrotational
moments of inertia, which are commonly
assumed in the rotor model,
produce $\mathcal{J}_y >\mathcal{J}_x >\mathcal{J}_z$.
This contradicts the basic assumption of
$\mathcal{J}_x >\mathcal{J}_y >\mathcal{J}_z$ and
the formula~(\ref{wobbling_frequency}).
\item
According to equation~(\ref{wobbling_frequency}),
the wobbling frequency $\omega_{\rm wob}$ increases as a function of $I$.
Conversely, the observed frequency decreases.
\end{itemize}
Solutions to these problems will be provided by microscopic treatments
in section~\ref{sec:QRPA_model}.
\subsection{Microscopic QRPA analysis for the wobbling motion}
\label{sec:QRPA_model}
A microscopic theory to treat the nuclear wobbling motion
in the small amplitude limit
is naturally provided by the QRPA
in the rotating frame,
or the self-consistent cranking plus QRPA~\cite{Mar79}.
Among the quadrupole tensors,
$Q_{ij}\propto x_i x_j - \delta_{ij} r^2/3$
($i,j=x,y,z$),
the negative signature operators of
$Q_y\equiv -Q_{zx}$ and $Q_z=iQ_{xy}$
are responsible for the wobbling motion.
Adopting the separable quadrupole interaction of
the form~(\ref{separable_interaction}),
the mean-field approximation simply replaces
one of the operators $QQ$ into its expectation value
$\mathcal{Q}(t)$, leading to the time-dependent mean field
\begin{equation}
h_{\rm UR}(t)=h_{\rm def} -\omega_{\rm rot} J_x
- \kappa_y \mathcal{Q}_y(t) Q_y - \kappa_z \mathcal{Q}_z(t) Q_z ,
\label{UR}
\end{equation}
where $h_{\rm def}$ is a deformed single-particle Hamiltonian
that contains the fields associated with diagonal tensors,
$-\kappa_{ii} \mathcal{Q}_{ii} Q_{ii}$.
Although, in general, $h_{\rm def}$ is time-dependent,
we hereafter focus our discussion on the wobbling motion, and
assume $h_{\rm def}$ is time-independent.
In this treatment of
equation~(\ref{UR}), the rotational axis stays along the $x$ axis
and the wobbling motion is represented by a fluctuation
in the orientation of deformed density distribution
induced by $\mathcal{Q}_y(t)$ and $\mathcal{Q}_z(t)$.
This picture corresponds to the uniformly rotating (UR) frame.
The small shape fluctuation induced by the off-diagonal quadrupole
tensors, $\mathcal{Q}_y$ and $\mathcal{Q}_z$, is not associated
with the real shape change from the equilibrium.
The same effect can be realized by rotating the reference frame to
the principal axis (PA) frame where the non-diagonal elements,
$Q_y$ and $Q_z$, of the quadrupole tensors vanish.
If we adopt this body-fixed frame,
the direction of the angular momentum fluctuates.
In the PA frame, since the rotation is no longer uniform,
the cranking model should be extended to a time-dependent one.
\begin{equation}
h_{\rm PA}(t)=h_{\rm def}-\vec{\omega}_{\rm rot}(t) \cdot \vec{J} .
\end{equation}
The rotor-model analysis of Bohr and Mottelson in section~\ref{sec:rotor_model}
has a direct connection to the PA picture.
In this picture, the frequency
$\vec{\omega}_{\rm rot}$
should be
treated as dynamical variables (operators).
In the small amplitude limit,
Marshalek proved the equivalence between the UR and the PA frames
and obtained the same expression~(\ref{wobbling_frequency}) for the
wobbling frequency,
with the moments of inertia calculated in the QRPA \cite{Mar79}.
It is generalized to arbitrary mean-field potentials and residual
interactions \cite{SS09}.
The two pictures are schematically illustrated in figure~\ref{fig:UR_vs_PA}.
\begin{figure}[tb]
\centerline{
\includegraphics[clip,width=0.7\textwidth]{UR_vs_PA.eps}
}
\caption{
Two equivalent pictures of the wobbling motion:
one in the uniformly rotating (UR) frame (left)
and the other in the principal axis (PA) frame (right).
From reference \cite{SS06}.
}
\label{fig:UR_vs_PA}
\end{figure}
The microscopic QRPA calculations were first performed with
the Nilsson potential and the separable quadrupole interactions
\cite{Mat90,SM95,MSM02,MSM04,SSM08}.
Later, it has been done with the Woods-Saxon potential and
an separable interaction which is determined by the symmetry
restoration condition \cite{SS09}.
In general, it is difficult to perform
the QRPA calculation for an odd-$A$ nucleus,
however,
this is not a problem at a finite $\omega_{\rm rot}$
because the Kramers degeneracy is lifted and the RPA vacuum
is uniquely identified.
The calculated QRPA moments of inertia indicate a proper ordering
of moments of inertia,
$\mathcal{J}_x >\mathcal{J}_y >\mathcal{J}_z$,
for the wobbling mode in $^{163}$Lu.
Why is this ordering different from a naive expectation based
on the irrotational flow?
To answer the question, it is important to distinguish
the dynamic and the kinematic moments of inertia~\cite{BM81}.
The dynamic moment of inertia is defined by the second
derivative of the rotational energy,
$1/\mathcal{J}^{(2)}= d^2 E_I/dI^2$,
which is considered to be
the one in the rotor model in equation~(\ref{triaxial_rotor}).
In contrast, the kinematic moment of inertia is defined by the first
derivative,
$1/\mathcal{J}^{(1)}= I^{-1} dE_I/dI =\omega_{\rm rot}/I$.
The largest moment of inertia, $\mathcal{J}_x$,
in the QRPA wobbling formalism of reference~\cite{Mar79} is the kinematic one,
more precisely,
$\mathcal{J}_x=\mathcal{J}^{(1)}_x \equiv \langle J_x \rangle/\omega_{\rm rot}$,
which is strongly influenced by
the alignment of the intrinsic angular momentum along the rotational ($x$) axis.
Generally the kinematic moment of inertia is larger than the dynamic one
because of the effect of alignment.
In $^{163}$Lu, the odd-proton quasiparticle mainly produces the alignment.
When the alignment is large enough, we could have
a rigid-body-like ordering,
$\mathcal{J}_x^{(1)} >\mathcal{J}_y >\mathcal{J}_z$,
for the positive $\gamma$ shape,
even if the dynamic moments of inertia are irrotational-like,
$\mathcal{J}_y >\mathcal{J}_x >\mathcal{J}_z$.
The QRPA moments of inertia automatically take into account
this effect.
Thus, {\it the alignment effect is crucial for the appearance of
the wobbling mode}, which was first pointed out in
references~\cite{MSM02,MSM04}.
The fact that the largest moment of inertia is the kinematic one can be
justified by the simple particle-rotor model as in reference~\cite{FD14}:
When the quasiparticle alignment $j$ is present,
$J_x$ is replaced by $J_x-j$ in equation~(\ref{triaxial_rotor}).
Using $J_x =[{\vec{J}^2-(J_y^2+J_z^2)}]^{1/2}
\approx (I+\frac{1}{2})-(J_y^2+J_z^2)/(2I+1)$
which is valid for in high-spin limit $I \gg 1$,
we obtain
\begin{eqnarray}
&&\hspace*{-20mm}
H_{\scriptsize\mbox{p-rot}}=\frac{(J_x-j)^2}{2\mathcal{J}_x}
+\frac{J_y^2}{2\mathcal{J}_y}+\frac{J_z^2}{2\mathcal{J}_z}
\approx\frac{(I-j)(I-j+1)}{2\mathcal{J}_x}
\label{H_prot}
\\
&&\hspace*{-13mm}
+\left(\frac{1}{2\mathcal{J}_y}-\frac{1}{2\mathcal{J}_x}
+\frac{1}{2\mathcal{J}_x}\frac{j}{I+\frac{1}{2}} \right)J_y^2
+\left(\frac{1}{2\mathcal{J}_z}-\frac{1}{2\mathcal{J}_x}
+\frac{1}{2\mathcal{J}_x}\frac{j}{I+\frac{1}{2}} \right)J_z^2 .
\label{triaxial_protor}
\end{eqnarray}
Namely the inverse of the kinematic moment of inertia,
\begin{equation}
\frac{1}{\mathcal{J}_x^{(1)}}
= \frac{\omega_{\rm rot}}{\langle J_x\rangle}
\approx \frac{1}{\mathcal{J}_x}\left( 1-\frac{j}{I+\frac{1}{2}}\right),
\label{J_x}
\end{equation}
appears in equation~(\ref{triaxial_protor}) in place of $1/\mathcal{J}_x$
in equation~(\ref{triaxial_rotor}).
The wobbling frequency (\ref{wobbling_frequency})
with $1/\mathcal{J}_x$ replaced by $1/\mathcal{J}_x^{(1)}$
of equation~(\ref{J_x})
first increases as spin increases
and then turns to decrease.
Thus the quasiparticle alignment also explains why the observed
wobbling frequency decreases as a function of $I$.
From equations~(\ref{triaxial_protor})
there is a critical angular momentum,
$I_c\equiv j (1-\mathcal{J}_x/\mathcal{J}_y )^{-1}-\frac{1}{2}$,
at which the wobbling frequency vanishes, $\omega_{\rm wob}=0$
(remember $\mathcal{J}_y>\mathcal{J}_x>\mathcal{J}_z$).
Beyond $I_c$, the wobbling mode ceases to exist,
because of the irrotational-like ordering,
$\mathcal{J}_y >\left. \mathcal{J}_x^{(1)}\right|_{I>I_c} >\mathcal{J}_z$.
In this way, for the case where the alignment
takes place along the axis of the intermediate dynamic moment of inertia,
the $I$-dependence of the original wobbling frequency
in section~\ref{sec:rotor_model} drastically changes.
Such a novel wobbling scheme was first pointed out in reference \cite{SMM04},
although the terms proportional to $j$ in equation~(\ref{triaxial_protor}),
i.e. the effect of alignment,
are interpreted as decreasing $\mathcal{J}_y$ and $\mathcal{J}_z$
instead of increasing $\mathcal{J}_x$.
The observed decreasing tendency of $\omega_{\rm wob}$ in the Lu isotopes
clearly suggests such a character.
In reference~\cite{FD14} it is called ``{\it transverse} wobbler''
in order to distinguish it from ``{\it longitudinal} wobbler''
where the quasiparticle aligns along the axis of the largest inertia,
$\mathcal{J}_x>\mathcal{J}_y>\mathcal{J}_z$.
In the longitudinal wobbler,
the frequency $\omega_{\rm wob}$ monotonically increases with $I$.
In fact, the microscopic QRPA calculations also predicted
the wobbling motion of increasing $\omega_{\rm wob}$ as a function of $I$
in nuclei of negative $\gamma$ shapes~\cite{Mat90,SM95},
in which the irrotational-type moments of inertia satisfy
the longitudinal condition,
$\mathcal{J}_x > \mathcal{J}_y > \mathcal{J}_z$.
A similar argument of the effect of alignment for the possible decrease
of $\omega_{\rm wob}$ has been discussed also in reference~\cite{HH03}.
It should be noticed that the three moments of inertia are assumed to be
independent of spin $I$ in the rotor model or the particle-rotor model.
In reality, however, the microscopically calculated QRPA moments of inertia
change
as functions of $I$, although their dependencies on $I$ are not so strong
in most cases~\cite{Mat90,SM95,MSM02,MSM04,SS09}.
One should take this into account in order to study precisely
how the wobbling frequency changes as a function of spin.
In reference~\cite{ST10}, introducing a rather strong spin-dependence
common to all three moments of inertia,
the decreasing tendency of $\omega_{\rm wob}$ is realized
in the particle-rotor coupling model with the inertia of
the rigid-body-like ordering, $\mathcal{J}_x >\mathcal{J}_y >\mathcal{J}_z$.
\begin{figure}[tb]
\centerline{
\includegraphics[clip,width=0.47\textwidth]{163Lu.eps}
\includegraphics[clip,width=0.47\textwidth]{163Lu_BE2_ratio.eps}
}
\caption{
(Left)
Calculated and experimental wobbling frequencies as functions of the
rotational frequency.
(Right)
Calculated and experimental inter- to intraband $B(E2)$ ratio
as functions of the
rotational frequency.
See text for details.
Adapted from reference \cite{SS09}.
}
\label{fig:163Lu}
\end{figure}
In figure~\ref{fig:163Lu}, we show results of the QRPA calculation
based on the deformed Woods-Saxon potential~\cite{SS09}.
Note that there are no adjustable parameters in the calculation
because the minimal symmetry restoring interaction is employed,
which is uniquely fixed once the deformed mean-field is given.
The deformation parameters are determined by minimizing the TRS.
The calculated wobbling frequency has a proper trend,
though the absolute magnitude is underestimated by a few hundred keV.
The large interband $B(E2)$ values are rather well reproduced in the
calculation.
However, the observed ratio,
$B(E2;I\rightarrow I-1)_{\rm out}/B(E2;I\rightarrow I-2)_{\rm in}$,
seems to increase as a function of $I$,
while the calculated ratio decreases because of
the $1/I$ dependence of the interband transition.
The dotted line in the right panel of figure~\ref{fig:163Lu} indicates
the result obtained by artificially increasing the triaxiality
($\gamma$) at higher spins.
In fact, in order to explain the experimental $B(E2)$ ratios,
the triaxial parameters $\gamma({\rm den})\approx 20^\circ$
are necessary.
The $\gamma({\rm den})$ is defined with respect to the intrinsic quadrupole
moments calculated from the density distribution.
Here, it should be noted that the triaxial parameter $\gamma$ for the
potential shape is significantly different from $\gamma({\rm den})$ \cite{SSM08}
(See \ref{sec:triaxial_def} for details).
As far as we know, at present,
none of the microscopic calculations are able to
reproduce the triaxiality of $\gamma({\rm den})\approx 20^\circ$.
Another unsolved problem is that the observed $B(M1)$ values are
significantly overestimated.
In reference \cite{FD15},
the inclusion of the isovector separable orbital angular momentum
interaction is suggested to improve the agreement.
The nuclear wobbling motion has not been fully understood yet.
\subsection{Precession: Rotational band built on a high-$K$ isomer}
\label{sec:Precession}
In section~\ref{sec:QRPA_model}, we show that
the alignment of quasiparticle
is crucial for the wobbling motion
to appear in the Lu isotopes with the positive $\gamma$ shape.
An interesting extreme case
of the alignments is that
the nuclear shape is axially symmetric about the alignment ($x$) axis;
i.e. $\gamma=60^\circ$ (oblate) or $\gamma=-120^\circ$ (prolate)
in the Lund convention and the angular momentum
is supplied only by the alignments of quasiparticles.
It is expected that the optimal configurations of aligned quasiparticles
in the states of such shapes make the high-spin isomers, or the ``yrast traps'',
along the yrast line, see e.g. reference~\cite{VDS83}.
Although the rotational bands built
on the oblate isomers are not yet observed,
those on the prolate isomers have been well known~\cite{BM75}.
They are nothing but the high-$K$ rotational bands widely observed
in the Hf and W region,
where many high-$j$ and high-$\Omega$ Nilsson
orbits are
concentrated near the Fermi surface.
Here, $\Omega$ is the component of
single-particle angular momentum along the symmetry axis.
\begin{figure}[tb]
\centerline{
\includegraphics[clip,width=0.45\textwidth]{schmprec1_2.eps}
\hspace*{0.05\textwidth}
\includegraphics[clip,width=0.18\textwidth]{precbody_2.eps}
}
\caption{
A schematic illustration of the precession motions.
(Left) Excitation spectra from reference~\cite{SMM05}.
There are no ${\Delta I}=2$ horizontal sequences
leaving only one ${\Delta I}=1$ vertical band for each high-$K$ state;
compare with figure~\ref{fig:wobbling}.
(Right) Superposition of the collective rotation about the perpendicular
axis makes the high-$K$ aligned angular momentum vector to precess around
the symmetry axis.
}
\label{fig:precession}
\end{figure}
We call this rotational motion ``precession'' because the aligned
angular momentum vector tilts like in the case of the wobbling motion
by superimposing the collective rotation about the perpendicular axis;
it is illustrated schematically in figure~\ref{fig:precession}.
Since the high-$K$ isomers have been known for many years,
they have been investigated by various methods;
e.g., by the cranked mean-field method~\cite{FP77},
by the RPA method based on
the sloping Fermi surface~\cite{Kur80,Kur82,AKLS81,Ska87},
or by the tiled axis cranking method~\cite{FNSW00}.
In reference~\cite{SMM05} it was considered as the axially symmetric limit
of the RPA wobbling formalism~\cite{Mar79} discussed in
section \ref{sec:QRPA_model}.
In fact, the wobbling frequency in equation~(\ref{wobbling_frequency}) becomes
$\omega_{\rm wob}=I/{\cal J}_\bot - \omega_{\rm rot}$, where
the perpendicular moment of inertia is denoted as
${\cal J}_\bot\equiv {\cal J}_y={\cal J}_z$ in the axially symmetric
limit and the rotational frequency about the main rotation axis
is $\omega_{\rm rot}=I/{\cal J}^{(1)}_x$.
Here, the dynamic moment of inertia (${\cal J}_x\rightarrow 0$) is
replaced by the kinematic inertia ${\cal J}^{(1)}_x$
with the aligned angular momentum $I_x=K$ (we here use $K$ in place of $j$).
On the other hand the rotor Hamiltonian~(\ref{triaxial_rotor}) in this case
reduces to $H_{\rm rot}=(I_y^2+I_z^2)/(2{\cal J}_\bot)$
so that the energy spectrum is given simply by
$E_{{\scriptsize\mbox{high-}}K}=[I(I+1)-K^2]/(2{\cal J}_\bot)$,
which can be rewritten as
\begin{equation}
E_{{\scriptsize\mbox{high-}}K}=
\omega_{\rm prec}\Bigl(n+\frac{1}{2}+\frac{n(n+1)}{2K}\Bigr) ,
\label{eq:prec}
\end{equation}
introducing the precession phonon number $n\equiv I-K$.
Here the precession frequency is given by
$\omega_{\rm prec}\equiv K/{\cal J}_\bot
=[\omega_{\rm wob}+\omega_{\rm rot}]_{I=K}$.
Thus the precession motion can be described by the harmonic excitation
of $n$-phonons as long as $n \ll K$. The difference of frequencies
$\omega_{\rm prec}$ from $\omega_{\rm wob}$ is due to the fact that
the wobbling motion is treated in the body-fixed frame while the precession
motion in the laboratory-frame.
Remember the transformation of the energy
into the routhian in the rotating frame
in section~\ref{sec:cranking_at_high_spin}, $E'=E-\omega_{\rm rot}I_x$,
and that the precession mode transfer the angular momentum by one unit
$\Delta I_x=1$.
Since there is no collective rotation about $x$ axis,
the rotational frequency $\omega_{\rm rot}$ is a redundant variable
and any observable quantities do not depend on it;
$\omega_{\rm wob}$ does depend while $\omega_{\rm prec}$ does not.
As it is discussed in reference~\cite{SMM05} not only the energy but also
the electromagnetic transitions, like $E2$ and $M1$,
can be treated with the multi-phonon picture
as long as the phonon number $n$ is much smaller than $K$.
\begin{figure}[tb]
\centerline{
\includegraphics[clip,width=0.8\textwidth]{precws_2.eps}
}
\caption{
(Left)
Calculated (filled circles) and experimental (crosses)
precession frequencies $\omega_{\rm prec}$
for various high-$K$ isomers in $^{178}$W,
where the experimental data is denoted by crosses.
(Right)
Calculated (filled circles) and experimental (crosses)
moments of inertia perpendicular to the symmetry axis,
estimated by ${\cal J}_\bot=K/\omega_{\rm prec}$.
Based on the result of reference~\cite{Sho09}.
}
\label{fig:precRPAws}
\end{figure}
We show in figure~\ref{fig:precRPAws} the result of precession frequencies
for a number of $K$ isomers in $^{178}$W calculated
by using the axially symmetric limit
of the Woods-Saxon QRPA wobbling formalism~\cite{SS09}.
Compared with the corresponding calculation of reference~\cite{SMM05},
in which the Nilsson mean-field potential is employed,
considerable improvement can be seen and a good agreement
with experimental data is obtained.
In the right panel of figure~\ref{fig:precRPAws},
the estimated moments of inertia ${\cal J}_\bot=K/\omega_{\rm prec}$
are also shown.
The agreement is much better than the simple mean-field calculation,
e.g.~\cite{FP77},
because the effect of residual interaction is taken into account in the QRPA.
It can be seen that the moments of inertia
take various values depending on the isomer configurations.
They are considerably larger than the moment of inertia of
the ground state rotational band estimated from the first $2^+$ state,
$\mathcal{J}_{\rm gr} \approx 28$ [$\hbar^2$/MeV].
They do not show a simple correlation with the $K$ quantum number,
and do not approach to
the rigid-body value (with $\epsilon_2=0.240$).
${\cal J}_{\rm rig} \approx 88$ [$\hbar^2$/MeV],
even at considerably high spin.
Their properties strongly depend on what kind of quasiparticles
contribute to those high-$K$ isomers
in which the numbers of quasiparticles are from four to ten.
See reference~\cite{SMM05} for precise configuration assignments.
At an extreme high spin, we can even
imagine possible existence of
torus-shape isomers and their precession motions \cite{IMMI14}.
\section{Summary}
Bohr and Mottelson have explored a variety of fields in the nuclear
structure physics.
Among them, we have discussed selected topics related to
the nuclear deformation and rotation.
First, we presented the concept of the
symmetry breaking in the unified model.
The symmetry broken state is not stable in finite systems, such as nuclei.
The correlation time induced by the quantum fluctuation
is a key to understand
the interplay between the symmetry breaking and restoration.
The finite-size effect associated with the zero-point motion may
hinder the symmetry breaking.
The coupling between intrinsic and rotational motions is
well described by the cranking model.
Since the model assumes a semiclassical treatment of the
nuclear rotation (angular momentum),
the model requires the quantization in the low-spin limit.
We show a possible quantization of the cranking model,
which is applicable to calculation of transition matrix elements
at low spin.
This can be regarded as a kind of hybrid model of
the unified model and the cranking model.
It is applied to electromagnetic decay properties
of vibrational bands and high-$K$ isomers.
In the high-spin region, the cranking model is a golden tool to study
the nuclear structure under a strong Coriolis and centrifugal
field.
We discussed effects of the rapid rotation on the octupole vibrations,
which are nicely treated with the QRPA
in the uniformly rotating frame (one-dimensional cranking).
The calculation reproduces the experimental data, showing the
phonon alignment and loss of collectivity (phonon breakdown).
The closed shell configurations of the superdeformed (SD) states
are characterized by the $2:2:1$ shell structure.
This shell structure has the $K^\pi=1^-$ octupole mode as a soft mode.
Away from the SD magic numbers with many valence nucleons,
the prolate SD nucleus could show instability toward a bending banana shape.
The triaxial deformation produces the three dimensional
non-uniform rotation, which is called wobbling motion.
The QRPA in the uniformly rotating frame provides a microscopic
tool to calculate the wobbling and precession
modes of excitation.
The experimental data are qualitatively reproduced.
This microscopic study
clearly indicates the importance of the quasiparticle alignment
for the existence of the wobbling mode.
The self-consistently calculated triaxial deformation seems
to be smaller than what experimental data indicate,
which is an important open problem.
The nucleus provides a wonderful opportunity to study a finite
system going through many kinds of symmetry breaking,
under a variety of extreme circumstances, such as
large angular momentum, deformation, and isospin.
The topics we have discussed in this paper were pioneered by
Bohr and Mottelson who gave us a deep insight into
nuclear structure and quantum many-body physics.
There are still many open issues in these fields
which are waiting for future studies.
\ack
This work was supported in part by JSPS KAKENHI Grant
No. 25287065 and
by Interdisciplinary Computational Science Program in CCS,
University of Tsukuba.
|
\subsection*{Appendix}
\subsubsection*{NP-completeness of the balanced edge partition problem}
It is easy to prove, by a reduction from the NP-hard {\sf \bf partition problem},
that the {\em k-way} balanced edge partition problem is NP-hard, even for $k=2$.
We briefly sketch the reduction. Recall that in the
partition problem we are given a
set $S$ of positive integers $a_1,a_2,...,a_n$, and our task is to decide if the numbers can be
partitioned into two subsets $S_1$ and $S_2$ such that the sum of numbers in $S_1$ equals the sum of
the numbers in $S_2$.
In the reduction we construct a graph with $n$ {\em cycles} $c_1,c_2,...,c_n$ such that
$|C_i|=a_i$. Obviously, a partition $(S_{1},S_{2})$ for the instance of the partition problem exists if and only if the
2$-$way balanced edge partition problem reduce from it has a minimal communication cost of 1.
\subsubsection*{The case that a real/original edge is cut}
\medskip
\noindent {\bf \sf \underline{Edge Fixing}}
\medskip
If the cut set of $D'$ contains a real edge, one end point of the this edge needs to be moved
to the cluster of the other end point.
Since the above manipulation will upset equality between clusters, in order to minimize (and eventually eliminate) the harm
caused by the moves,
we process real edges that cross between different clusters in the following order.
We cycle through all pairs $(A', B')$ of clusters and do as follows:
For the current pair $(A', B')$ of clusters we first find the set $R_{A',B'}$ of real edges that cross
between $A'$ and $B'$. Then we put half of the edges of $R_{A',B'}$ entirely into $A'$
and the other half entirely into $B'$. If $|R_{A',B'}|$ is odd, we choose a single edge from it and leave
its end points in the same clusters as they were originally (i.e. in $A'$ and $B'$).
It is easy to see that restructuring the clusters as above leaves them balanced.
Consider now every cluster as a ``super-node,'' and form an edge
between two super-nodes if after the restructuring there is an edge between the corresponding clusters.
A graph will arises on the super-nodes that we call $C'$.
For technical reasons we shall assume that the cluster size, $2m/k$, is even (this is a natural assumption, since if $k$ does not divide $m$,
the original problem cannot be solved).
We argue that in this case the
degree of every super-node in $C'$ is even: Take out the inner-cluster edges together with their
end-points in $D'$., and all we are left with are nodes of internal real edges (that are not
cut), with an even number of end points. Thus the number of edges going out of a cluster must be even too.
Since $C'$ has only even degree vertices, its edge set decomposes into Euler-tours that can be made directed in an obvious way.
In a final stage we can fix edges along these directed Euler tours by
restructuring the clusters once more, at this time putting each inter-cluster edge into the cluster associated with its starting point in the directed Euler tour.
It is easy to see that after this final restructuring of clusters they remain of equal size.
Edge-fixing process can be designed in a more sophisticated way rather than arbitrarily picking
pairs of clusters and cut real edges, which we will describe in Section \ref{sec:psf}.
\section{Conclusion}
We formulate a generic graph partitioning model that systematically exploits the
complexity and analytical bound for the optimal cache data reuse problem in GPU
computing.
Our experiments show that our method can improve data sharing and performance significantly for GPU benchmarks.
Our exploration in this paper demonstrates for the first time that it is feasible to use graph
partitioning models to enhance GPU cache performance.
\section{Evaluation}
\label{sec:eval}
\subsection{Experiment Setup}
\label{sect:eval:env}
Our hardware platform is a machine with an NVIDIA GeForce GTX680 GPU with CUDA computing capability 3.0. It includes 8 streaming multiprocessors (SMs) with
192 cores on each of them and 1536 cores in total. On each SM there is 64 KB of L1 cache and shared
memory (i.e., software cache), shared by its 192 cores. There are three size configurations of L1
cache and shared memory: 16KB/48KB, 32KB/32KB and 48KB/16KB. Since L1 cache in
GTX680 is allowed for local memory data only, we always configure shared memory to be 48KB.
There is also 48KB texture cache per SM.
The CPU is Intel Core i7-4790 at 3.6 GHz with 8MB last-level cache.
\begin{table*}[htb]
\footnotesize
\centering
\begin{tabular}{|l|c|c|c|c|}
\hline
Benchmark & Application Domain & Input & Cache type \\
\hline
b+tree \cite{Che+:IISWC09} & Tree data structure & One-million size database
query & Software\\
bfs \cite{Che+:IISWC09} & Breadth first search & One-million node graph & Texture\\
cfd \cite{Che+:IISWC09} & Computation fluid dynamics & 97K, 193K, and 0.2M node
meshes & Software\\
gaussian \cite{Che+:IISWC09} & Gaussian elimination & Linear system with 1024
unknown variables & Software \\
particlefilter \cite{Che+:IISWC09} & SMC for posterio density estimation &
Tracking of 1000 particles & Software \\
streamcluster \cite{Che+:IISWC09} & Data stream clustering & 65,536 data
points & Software \\
CG \cite{CUSP} & Sparse matrix multiplication & Various size input matrices & Both\\
\hline
\end{tabular}
\caption{Benchmark summary.}
\label{tbl:bench}
\end{table*}
We use seven applications from various domains including data
mining, computation biology, object tracking, scientific simulation, and graph
traversal. This set of benchmarks are representative of important contemporary
workloads that can benefit from cache locality enhancement. The benchmarks are listed in Table \ref{tbl:bench}.
Six applications are from the Rodinia benchmark suite \cite{Che+:IISWC09}.
We also evaluate an important kernel, the sparse matrix vector multiplication
(SPMV) kernel, which is widely used in different types of applications including
machine learning and numerical analysis. The SPMV kernel is also frequently used for evaluating graph partition models on parallel CPUs \cite{karypis1999parallel, karypis2000multilevel}.
We report the performance of SPMV kernel and the task scheduling overhead. We also evaluate the asynchronous
optimization method. We run the SPMV kernel in
the context of the conjugate gradient (\emph{CG}) \cite{CUSP} application. The
CG application calls SPMV iteratively until the solution converges.
All benchmarks are compiled on a system that runs 64-bit Linux with kernel version 3.1.10 and CUDA 5.5.
We use the input sets bundled with the benchmark suite or real-world input sets. For example, the \emph{cfd} benchmark from Rodinia benchmark suite \cite{Che+:IISWC09} has three input meshes: \emph{fvcorr.domn.097K}, \emph{fvcorr.domn.193K} and \emph{missile.domn.0.2M}.
The sparse matrix inputs to \emph{CG} come from the University of Florida sparse matrix collection \cite{FloridaMat} and matrix market \cite{MATRIXMARKET}.
In the rest of this section, we first evaluate the performance of sparse matrix
vector kernel over a large number of sparse matrices and various configurations.
Then we show the evaluation results of the six applications from Rodinia.
\subsection{Sparse Matrix Workloads}
Conjugate gradient (CG) method \cite{Hestenes+:NBS52} is an important algorithm for solving linear
systems, whose matrices are usually large and sparse. The conjugate gradient
method is an iterative process that repeated invokes sparse matrix vector
multiplication until the result converges. We focus on the total amount of time
for sparse matrix vector multiplication in CG in the following
discussion.
We partition the sparse matrix workload with our edge partition model and hypergraph model to
maximize data reuse and compare their performance with the popular SPMV implementation from CUSP \cite{CUSP} and CUSPARSE \cite{CUSPARSE}.
The CUSP spmv kernel orders the data layout such that all non-zero elements are sorted by row indices and then it distributes the non-zero elements evenly to threads.
We are not aware of CUSPARSE's task distribution scheme since it is not open source, however since it is faster than CUSP for most inputs, we include CUSPARSE for comparison.
In the data-affinity graph of SPMV, there is a vertex for every element in the input vector $x$ and output vector $y$.
For each non-zero element $A[i,j]$ in the input matrix $A$, an edge is added to connect vertex $j$ in the input vector and vertex $i$ in the output vector, since a non-zero $A[i,j]$ implies a multiplication with $x_j$ to get $y_i$.
The data-affinity graph of SPMV is naturally a bipartite graph.
With the data-affinity graph, we perform edge partition and let one thread block be mapped to one partition.
We use software cache for the corresponding shared input and output vector elements within each thread block.
We will also show the results of using texture cache later.
\begin{table*}[t]
\centering
\footnotesize
\begin{tabular}{|l|c|c|c|c|c|c|c|} \hline
Name & Dimension & Nnz & CUSPARSE time & EP time & EP partition & HP time & HP partition \\
\hline
cant & 62K*62K & 2.0M & 2.53 & 2.89 & 1.702 & 2.92 & 11.66 \\
circuit5M & 5.6M*5.6M & 59.5M & 21599 & 783.2 & 67.157 & 784.2 & 2250 \\
cop20k\_A & 121K*121K & 1.4M & 25.93 & 20.17 & 1.457 & 19.99 & 5.76 \\
Ga41As41H72 & 268K*268K & 9.4M & 19.37 & 15.29 & 17.922 & 15.26 & 194.5 \\
in-2004 & 1.4M*1.4M & 16.9M & 430.9 & 359.4 & 17.889 & 355.7 & 413.6 \\
mac\_econ\_fwd500 & 207K*207K & 1.3M & 31.54 & 18.29 & 1.342 & 18.14 & 5.04 \\
mc2depi & 526K*526K & 2.1M & 36.45 & 28.31 & 1.436 & 28.36 & 4.87 \\
scircuit & 171K*171K & 0.96M & 20.42 & 13.62 & 0.642 & 13.51 & 2.91 \\
\hline
\end{tabular}
\caption{Matrix Information.}
\label{tbl:matrices}
\end{table*}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figures/spmv_speedup.pdf}
\caption{Speedup of various spmv kernels.}
\label{fig:spmv_speedup}
\end{figure}
We use 8 matrices with different data access patterns to show the applicability of our EP model.
These matrices are shown in Table \ref{tbl:matrices}.
\emph{Nnz} represents the total number of nonzero elements.
\emph{CUSPARSE time} column shows total SPMV kernel execution time of CUSPARSE.
\emph{EP time} and \emph{HP time} show the total kernel execution time of our EP model and hypergraph model respectively.
\emph{EP partition} and \emph{HP partition} show their partition time.
All times are measured in seconds.
In Table \ref{tbl:matrices}, we first observe similar kernel execution time for
EP and HP models, which confirms our EP model can produce similar partition
quality compared to hypergraph partition model. (Note that in Table
\ref{tbl:matrices} we show the partition time and kernel execution time without
applying the asynchronous optimization method -- the overhead control. We will
show the results with overhead control later.)
On the other hand, hypergraph partitioning incurs much larger partition overhead compared to our EP model.
Its partition time occupies 205\% of the total kernel time of CUSPARSE on average, which means the hypergraph partition model could not finish optimization for most matrices before program ends.
On the other hand, the partition time of our EP model only occupies 22.7\% of total time on average.
Therefore, hypergraph partitioning is infeasible for GPU applications due to its large overhead.
Other edge partition methods have been proved to have worse quality than the default scheduling in Figure \ref{fig:comparison}.
Thus we focus on the study of our EP model in the rest of this paper.
Figure \ref{fig:spmv_speedup} compares the performance of four versions of SPMV kernel execution, including CUSPARSE, CUSP, our EP model that does not consider optimization overhead (EP-ideal), and the one that takes overhead into consideration (EP-adapt).
We set the thread block size to be 1024. We use
CUSPARSE kernel time as the baseline, since it is faster than CUSP for most matrices.
In Figure \ref{fig:spmv_speedup}, we can see that CUSP is slower than
CUSPARSE except in two cases, \emph{circuit5M} and \emph{in-2004}.
However, our EP model based approach is faster than both of them except for \emph{cant}, where the quality of default scheduling is similar to that of our EP model as shown in Figure \ref{fig:comparison}.
When using EP-adapt for \emph{cant}, there is almost no slowdown for \emph{cant}. That is because adaptive overhead control checks if optimization is beneficial and if not it would stop using the optimized kernel.
EP model is slightly worse than CUSP for \emph{in-2004} because there is very limited data reuse, which causes EP model to use a large amount of software cache, adversely affecting occupancy.
We also observe that in most cases, the performance of EP-adapt is similar to that of EP-ideal except for \emph{Ga41As41H72}, where the partition time occupies most kernel execution time and we can only optimize a small portion of SPMV invocations.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figures/spmv_trans.pdf}
\caption{Normalized transaction number of spmv.}
\label{fig:spmv_trans}
\end{figure}
Figure \ref{fig:spmv_trans} shows the normalized memory transaction number for three spmv kernels.
All results are normalized to that of CUSPARSE.
We observe that memory transaction number is reduced significantly for all matrices except \emph{cant} and \emph{Ga41As41H72}.
There is no memory traffic reduction for \emph{cant} and \emph{Ga41As41H72} due to the same reasons we have explained above.
Overall, the transaction reduction maps well to the performance results.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figures/spmv_speedup_tc.pdf}
\caption{Texture cache v.s. software cache.}
\label{fig:spmv_speedup_tc}
\end{figure}
\paragraph{Software cache v.s. texture cache}
Here we study the performance difference when using software and texture caches to optimize the data sharing of input vector in SPMV.
Since the output vector is write-shared, texture cache cannot be used to store it.
Figure \ref{fig:spmv_speedup_tc} compares their performance; EP-text represents the texture cache version and EP-smem represents the software cache (i.e., shared memory) version.
Again, the baseline is the performance of CUSPARSE.
Software cache version outperforms texture cache version for almost all matrices except \emph{in-2004}, in which case the large shared memory usage of EP-smem degrades thread level parallelism significantly as stated above.
Note that EP-text outperforms CUSP in Figure \ref{fig:spmv_speedup} where \emph{in-2004} is the only case our optimized kernel using software cache is not as good as CUSP.
Compared to CUSP and CUSPARSE, EP-text also has better performance.
In general, the texture cache based approach could potentially pollute cache by evicting data before it gets fully reused, while using software-managed cache does not have such a problem.
On the other hand, using texture cache can reduce programming overhead since it requires less effort for program and data transformation, and there is still sufficient performance improvement.
Thus there is a trade-off between programmability and performance.
In practice, we can choose either version based on programmability and performance preference.
\begin{table}
\centering
\footnotesize
\begin{tabular}{|l|c|c|c|c|c|c|} \hline
Block size & \multicolumn{2}{c|}{256} & \multicolumn{2}{c|}{512} &
\multicolumn{2}{c|}{1024} \\ \hline
time (s) & tex & smem & tex & smem & tex & smem
\\\hline\hline
cant & 310 & 279 & 313 & 289 & 330 & 305 \\
circuit5M & 8312 & 7349 & 8524 & 7531 & 8691 & 7831\\
cop20k\_A & 208 & 186 & 211 & 190 & 216 & 201 \\
Ga41A* & 1353 & 1196 & 1378 & 1225 & 1409 & 1288 \\
in-2004 & 2471 & 4472 & 2492 & 3490 & 2565 & 3594 \\
mac* & 190 & 174 & 191.4 & 175 & 196 & 182 \\
mc2depi & 332 & 284 & 307 & 277 & 307 & 283 \\
scircuit & 143 & 130 & 144 & 133 & 147 & 136 \\ \hline
\end{tabular}
\caption{Performance on different thread block sizes.}
\label{tbl:block}
\end{table}
\paragraph{Thread block size} Finally, we show the sensitivity of our graph partition approach with respect to different thread block sizes.
Table \ref{tbl:block} shows the performance of our EP model
without overhead control (EP-ideal) under different thread block sizes for both software and texture caches.
At every thread block size, the performance difference between software and texture caches is similar to the observation we made earlier.
The results suggest for all shared memory kernels, the performance at small thread block size is slightly better than (or the same as) that of the larger thread block size except \emph{in-2004}, where shared memory version generally performs badly because of limited data sharing and low concurrency.
However, smaller block size implies larger block number, and thus longer partition time of EP model.
Taking both kernel performance and partition overhead into consideration, the performance using smaller block size is similar to that using large block size.
Therefore, we still choose 1024 as the default block size in our experiments.
\begin{figure*} [htb]
\centering
\subfloat[bfs]{\includegraphics[width=0.25\textwidth]{figures/block_size_bfs.pdf}}
\subfloat[b+tree]{\includegraphics[width=0.25\textwidth]{figures/block_size_b+tree.pdf}}
\subfloat[cfd (f097K)]{\includegraphics[width=0.25\textwidth]{figures/block_size_cfd1.pdf}}
\subfloat[cfd (f193K)]{\includegraphics[width=0.25\textwidth]{figures/block_size_cfd2.pdf}} \hfill
\subfloat[cfd (m0.2M)]{\includegraphics[width=0.25\textwidth]{figures/block_size_cfd3.pdf}}
\subfloat[gaussian]{\includegraphics[width=0.25\textwidth]{figures/block_size_gaussian.pdf}}
\subfloat[particlefilter]{\includegraphics[width=0.25\textwidth]{figures/block_size_pf.pdf}}
\subfloat[streamcluster]{\includegraphics[width=0.25\textwidth]{figures/block_size_sc.pdf}}
\caption{Application performance on different block sizes.}
\label{fig:appspec}
\end{figure*}
\subsection{General Workloads}
\label{subsec:app}
Figure \ref{fig:appspec} shows the performance of the six Rodinia applications under various thread block sizes. For every benchmark, we test performance using four thread block sizes, 128, 256, 384 and 512, unless the program does not allow such a thread block size.
For instance, \emph{gaussian} only allows square thread block sizes.
The original runtime is denoted as {\em original} in Figure \ref{fig:appspec}. The runtime of our EP model with overhead control is denoted as \emph{EP-adapt}.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/speedup.pdf}
\caption{Application performance summary.}
\label{fig:appsummary}
\end{figure}
First, we can see that in most cases our optimized version outperforms the original version when the thread block size is fixed. The maximum speedup is 1.97 times for {\em gaussian} at the thread block size 256.
For {\em streamcluster}, the maximum speedup 1.7\% is for thread block
size 1024, which is lower than that of other benchmarks.
This is because \emph{streamcluster} has less data sharing compared to other benchmarks at runtime.
In \emph{streamcluster}, every thread holds a unique data point, and a set of threads access the same center point to perform distance calculation, which makes the average degree of data-affinity graph to be $\le 2$ (average data reuse).
Enlarging thread block size typically would enlarge the amount of sharing since more threads imply more data sharing opportunities.
This is particularly true for {\em cfd}. The larger the thread block size is, the better performance it is for the EP model.
However, it is not the case for all benchmarks.
For instance, the best performance of EP-adapt is not achieved at the largest thread block size for \emph{particlefilter} and \emph{b+tree}.
This is because the GPU performance is determined by many factors besides data sharing.
Many of these factors change with the thread block size, such as occupancy, cache contention, etc.
For instance, \emph{cfd} suffers from performance degradation at a block size of 384 because 384 is non-integral power of 2, and thus it is impossible to achieve the best potential occupancy of 2048 threads.
In Figure \ref{fig:appsummary}, we show the comparison between the best EP-adapt version and the best original version in terms of performance across different thread block sizes.
The data is normalized to the runtime of the best original version.
Note that the execution time of \emph{EP-adapt} version includes the optimization overhead.
We have significant performance gains, or at least no performance degradation, for all benchmarks with adaptive overhead control.
Figure \ref{fig:read_trans} shows the normalized off-chip read transaction number measured by CUDA profiler for this set of benchmarks.
The reason why we do not show write transaction number is that there is no write sharing in these benchmarks.
All results are normalized to that of the {\em original} version at the same thread block size.
The results show that our technique can reduce off-chip memory transactions significantly, since more memory requests are satisfied by on-chip software or texture caches.
We also observed that using texture cache does not perform as well as software cache in most cases, which is similar to the observation for SPMV.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figures/read.pdf}
\caption{Read transaction number.}
\label{fig:read_trans}
\end{figure}
\section{Evaluation}
\label{sec:eval}
\subsection{Experiment Setup}
Our hardware platform is a machine with NVIDIA GeForce GTX680 GPU. The GTX680
has CUDA computing capability 3.0. It has 8 streaming multiprocessors (SM) with
192 cores on each of them. On each SM there is 64 KB of L1 cache and shared
memory in total, shared by 192 cores. There are three size configurations of L1
cache and shared memory: 16KB/48KB, 32KB/32KB and 48KB/16KB. Since L1 cache in
GTX680 is allowed for local memory data only, we always configure L1 cache to be
16KB. There is also 48KB read-only (texture) cache. The CPU is Intel Core
i7-4790 at 3.6 GHz with 8MB cache.
We use seven representative applications including Rodinia benchmark suite
\cite{Che+:IISWC09} and the CUSP\cite{CUSP} library. The benchmarks are listed in Table
\ref{tbl:bench}. In these seven benchmarks, the \emph{CG} benchmark calls sparse matrix vector
multiplication kernel
extensively and we perform optimization on sparse matrix vector multiplication
kernel. We also compare our sparse matrix vector multiplication kernel
performance with that of CUSPARSE, which is a highly optimized and well adopted
GPU sparse linear algebra library \cite{CUSPARSE}.
We use the input sets coming with the benchmark suite or real-world input sets.
For example, the \emph{Cfd} benchmark from Rodinia benchmark suite \cite{Che+:IISWC09}
has three input meshes. For sparse matrix input to \emph{Cg}, we matrices from the University of Florida sparse matrix collection
\cite{FloridaMat} and matrix market \cite{MATRIXMARKET}. All programs are
compiled on a system that runs 64-bit Linux with kernel version 3.1.10 and CUDA 5.5.
\citexx
\begin{table}[t]
\centering
\footnotesize
\caption{Benchmarks}
\begin{tabular}{|l|l|l} \hline
Benchmark & Description & Input\\
\hline
b+tree \cite{Che+:IISWC09} & Tree data structure. & Queries to database of 1
million records. \\
bfs \cite{Che+:IISWC09} & Breadth first search & \\
cfd \cite{Che+:IISWC09} & Comp. fluid dynamics & 97K, 193K, 0.2M node meshes. \\
gaussian \cite{Che+:IISWC09} & Gaussian elimination \\
particlefilter \cite{Che+:IISWC09} & SMC for posterio density estimation \\
streamcluster \cite{Che+:IISWC09} & Data stream clustering \\
spmv \cite{CUSP} & Sparse matrix multiplication \\
\hline
\end{tabular}
\label{tbl:bench}
\end{table}
\begin{figure*}
\centering
\subfloat[B+tree]{\includegraphics[width=0.3\textwidth]{figures/B+tree.pdf}}
\subfloat[Bfs]{\includegraphics[width=0.3\textwidth]{figures/BFS.pdf}}
\subfloat[CFD]{\includegraphics[width=0.3\textwidth]{figures/CFD.pdf}} \hfill
\subfloat[Gaussian]{\includegraphics[width=0.3\textwidth]{figures/Gaussian.pdf}}
\subfloat[Particle
filter]{\includegraphics[width=0.3\textwidth]{figures/particleFilter.pdf}}
\subfloat[Streamcluster]{\includegraphics[width=0.3\textwidth]{figures/streamcluster.pdf}}
\caption{Application performance on different block sizes}
\label{fig:appspec}
\end{figure*}
\subsection{Application Workload}
\label{subsec:app}
We first show the results of applications with potential data sharing in Rodinia benchmark
suite in Fig. \ref{fig:appspec}. For every benchmark, we use four different thread block sizes.
We show the original running time at different thread block sizes, denoted as {\em Original} in the
figure. We also show the running time
after applying our work partitioning techniques, denoted as {\em Optimized} in Fig.
\ref{fig:appspec}. The thread block size specifies how many threads can share data in
software-cache, since threads within the same block can see each other's data in software cache.
\noindent {\bf B+tree} B+ tree is a tree data structure that keeps data sorted and allows
searches, sequential access, insertions, and deletions in logarithmic
time. It is commonly used in databases and file systems. A B+ tree can
be viewed as a B-tree in which each node contains only keys. The potential data sharing is the back and
forth movement over a path from root node to leaf node, which is handled by a set of threads in
parallel. The input is 10,000 queries into a database of 1 million records.
\noindent {\bf CFD} Computational fluid dynamics, usually abbreviated as CFD, is a branch
of fluid mechanics that uses numerical methods and algorithms to solve
and analyze problems that involve fluid flows. The fluid elements are simulated in an unstructured
grid. The potential data sharing occurs due to neighbour interactions on the grid. The input is an
initial state of fluid with 97,046 fluid elements.
\noindent {\bf Gaussian} Gaussian Elimination computes results row by row, solving for variables in a
linear system. The algorithm synchronizes between iterations and the parallelism is exploited for value calculation in every
iteration. We focus on the kernel called {\em Fan2()}
exhibits data sharing pattern within control flow statements. The input is a linear system with 1024
unknown variables.
\noindent {\bf Particlefilter} The particle filter (PF) is statistical estimator of the location of
a target object given noisy measurements of that target's location and an idea of the object's path in a Bayesian framework. Particle filters or Sequential Monte Carlo (SMC) methods are a set of
on-line posterior density estimation algorithms. The potential data reuse happens when every
particle accesses the same density distribution function. The input is a tracking of 1,000
particles.
\noindent {\bf Streamcluster} Data Stream clustering, given a sequence of n points in a metric space and an integer k, output k points in the sequence to minimize the sum
of the distance of every point to its nearest neighborhood center. Data sharing occurs when every
thread tries to access the center points. The input is a set of 65536 data points with at maximum
1,000 centers.
First, we can see from Fig. \ref{fig:appspec} that in most cases the optimized version performs better than the
original version when a thread block size is fixed. The maximum speedup is 2.2 times for {\em
particleFilter} at the thread block size 256. Every benchmark has significant speedup except {\em
streamcluster}. In {\em streamcluster}, the maximum speedup ~2\% is for thread block
size 1024. It is because streamcluster has less data sharing than other benchmarks at runtime. The
data that is reused is the center point. None of the other data points is reused, while the total number
of data points is 65535. Since only threads within a thread block can see each other's data, when
thread block size is smaller, the benefits are diminishing and there is no speedup, but also at
least no slowdown.
Enlarging thread block size typically would enlarge the amount of sharing since more threads
share data, as illustrated in the {\em streamcluster} example. However, it is not always the case in the
our experiments. Depending on the workload data reuse pattern, when partitioning, it might cut more
nodes at the cluster boundary as the cluster size increases. And thus more data needs to be loaded
for a cluster. Recall that for our virtual cache design, we load all data objects of interest before
we carry out any computation a work cluster. In Fig. \ref{fig:appspec}, we find that CFD has best
performance of the optimized version for thread block size
320, among the thread block sizes of 192,256,320 and 384.
It is noteworthy that even for the original program, different thread blocks imply performance
variation. The performance of GPU program depends on many execution parameters factors such as
occupancy, register usage per thread and original shared memory allocation. Regardless, not only our
transformed code usually performs better than the original at a fixed thread block size, but also
the best optimized version among all thread blocks performs better than the best original version
among all thread block sizes.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figures/summary.pdf}
\caption{Application performance summary}
\label{fig:appsummary}
\end{figure}
In Fig. \ref{fig:appsummary}, we show the comparison between the best optimized version and the best
original version across different thread block sizes, we still have significant performance gains or
at least no speedup for all benchmarks. In Fig. \ref{fig:appsummary}, we also show the comparison between the worst optimized version and
the worst original version, the baseline is the best original version.
\subsection{Scale-free Workload}
The application workload in Section \ref{subsec:app} has the data sharing pattern of maximum reuse
frequency (maximum degree in locality graph) close to average reuse frequency (average degree in
locality graph), for instance, in CFD, the maximum number of neighbours of a fluid element is
bound by 4 and average is ~3, or simple scale-free pattern, for instance, in {\em streamcluster}, only one center node has maximum degree of
65535 and all others have degree of 1. In the second part of our evaluation, we select a workload
that exhibits {\em scale-free} locality graph pattern. That is, the degree distribution of the
locality graph, follows a power-law distribution. In another word, a set of nodes have large reuse
while the rest of nodes have few reuse.
\begin{table}[t]
\centering
\footnotesize
\caption{Matrices. Nnz: number of nonzero elments. AvgR: average reuse of a vector element. MaxR:
maximum reuse of a vector element. }
\begin{tabular}{|l|l|l|l|l|} \hline
Matrix & Nnz & Size & AvgR & MaxR\\
\hline
citationCiteseer & 2.3M & 268K$\times$268K & 5.50 & 1078 \\
IMDB & 3.7M & 428K$\times$896K & 6.30 & 1590\\
web-Google & 5.1M & 916K$\times$916K & 7.02 & 6326\\
web-NotreDame & 1.4M & 325K$\times$325K & 6.45 & 10721\\
web-Stanford & 2.3M & 281K$\times$281K & 8.51 & 38606 \\
wiki-Talk & 5.0M & 2.4M$\times$2.4M & 3.99 & 100022\\
\hline
\end{tabular}
\label{tbl:spm}
\end{table}
We use the {\em spmv} kernel in \cite{CUSP} as baseline and use a set of sparse matrices with scale-free pattern
\cite{Tang+:CGO15}. We show the data reuse distribution of one of the six matrices in Fig.
\ref{fig:degdistribution} as they have similar pattern. We use {\em citatioCiteseer} as this
example. We use complement cumulative density function (CCDF) to represent the degree
distribution. A point (x,y) in CCDF figure represents that $y$ is the percentage of nodes that have
degree $>= x$. A graph that exhibits scale-free feature would see fast decaying of the CCDF
function. We also show the logarithmic scale of CCDF of {\em citationCiteseer} in
Fig. \ref{fig:degdistribution}. We also show the statistics for these six matrices in Table
\ref{tbl:spm}.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/spmv.pdf}
\caption{SPMV results for scale-free matrices}
\label{fig:spmv}
\end{figure}
We perform data-sharing aware work partitioning and show the results in Fig.
\ref{fig:spmv}. Unlike in Section \ref{subsec:app} we only report the performance of the kernel run without partition overhead.
For two reasons, first we transform the input sparse matrix file format so that the non-zero elements are
ordered in an order as if they would be executed by continuous logical threads. So we do not see
overhead when running the kernel. Secondly we only get the time of one kernel run (avg of many
repeated runs), while in practice, when the kernel is
called in a real application, it might run thousands of or millions of times, and thus the overhead of partitioning can be well
amortized. For the application workload in Section \ref{subsec:app}, the runtime partition overhead is
well amortized. If a kernel is run for only one time in an application, we do not suggest using the
work partitioning algorithm we presented in this paper. We will leave it as our future work
to further approximate the runtime partitioning algorithm.
In Fig. \ref{fig:spmv}, we show the performance improvement over the kernel implemented in
CUSP \cite{CUSP} library. The performance enhancement can be seen as a potential to improve
performance for these sparse matrices as we do not take overhead into consideration. We have speed
up between 1.05 and 1.94. The matrix {\em wiki-talk} has a large maximum reuse, which is 100022
(Table \ref{tbl:spm}),
which means one of the input vector or output vector elements is reused significantly. However, the
average reuse is small which is 3.99, indicating a strong scale-free pattern. Our technique shows
significant improvement for the case, so is that for {\em web-Stanford}. For other cases, the
speedup is not significant, but no slow either. The baseline we compared to, the CUSP\cite{CUSP}
implementation, pre-process the data so that non-zero elements are sorted by their row number, it is
possible that the best partition is close to the one sorted by row index. Overall, based on results, for locality graph with strong
scale-free pattern, $deg_{max} >> deg_{avg}$, our technique show advantage.
\section{Introduction}
\label{sec:intro}
A GPU is a massively parallel computational accelerator that is equipped with
hundreds or thousands of cores. A single GPU can provide more than 4 Teraflops
single precision performance at its peak, however, the maximum memory throughput
of a GPU card is around 200 GB/s. Such a gap usually prevents GPU's computation power from being fully harnessed.
A cache is a layer in between GPU's computation units and memory units, which
is a fast but (relatively) small storage for frequently accessed data. The
modern GPUs are equipped with cache to improve program performance
\cite{FermiWhitePaper, KeplerWhitePaper, MaxwellWhitePaper}. The
throughput of the first level cache on GPUs is close to the throughput of
computation unit and thus effective use of cache is critical to performance.
While regular applications can take advantage of cache by classical
transformations such as loop tiling and pipelining, instruction
scheduling and etc, it is not so straightforward how irregular
applications can best utilize cache.
In this paper, we look at irregular GPU applications and focus on improving
\emph{shared cache} performance. A GPU consists of a set of streaming multiple
processors (SMs or SMXs). Every SM has a fast shared cache. Threads that run on
a SM can maximize data sharing in cache and minimize communication to
off-chip memory. We show an example of how threads can be effectively scheduled
to minimize off-chip memory communication in Figure \ref{fig:mot}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\linewidth]{figures/cachesharing.pdf}
\caption{Mapping of \emph{cfd} interaction computation into two streaming multiprocessors. Assuming every SM has three threads. In (a) there are three particle needs to be loaded into two SMs while in (b) there is only one. }
\label{fig:mot}
\end{figure}
We use the computation fluid dynamics (\emph{CFD}) application
\cite{Corrigan+:AIAA09, Corrigan+:IJNMF11} as an example (Figure \ref{fig:mot}).
In \emph{cfd}, the main computation is to calculate interaction between two
adjacent particles and apply the interaction information to obtain the state of the
particles in the next time step.
Figure \ref{fig:mot} shows the interaction among six particles as a graph on the
left. Each node in the graph represents a particle. Each edge represents an
interaction, which is taken care of by one thread. Figure \ref{fig:mot}(a) shows
one way to schedule the threads, with interactions {\em e1,e2,e3} mapped to one
SM, requiring 9 loads from the memory (assuming one load is for every distinct
particle). In this case, three particles need to be fetched twice from memory. However, the thread schedule in figure
\ref{fig:mot}(b) indicates that, with {\em e1,e2,e4} mapped into one SM, only one
particle needs to be fetched from memory twice, and the particle that is shared
by {\em e1, e2, e4} only needs to be fetched into cache once. All
together the second case requires 7 loads from memory.
According to our study, for all three inputs of \emph{CFD} from the Rodinia benchmark suite \cite{Che+:IISWC09}, an average of 73.4\% of all particle loads are redundant under the default task scheduling.
This example demonstrates that there is a great potential to improve data sharing through careful task scheduling.
The graph-based model is used and extensively studied in the past \cite{Hendrickson+:PC00, gonzalez2012powergraph,
malewicz2010pregel, low2012distributed} for irregular
application on parallel CPU processors, due to its efficiency in capturing the
relationship between data and computation. In this paper, we also used a
graph-based model to tackle the shared cache problem for irregular GPU applications.
The GPU shared cache problem brings several new challenges which cannot be
handled well by previous graph-based methods.
First of all, the expected task partitioning algorithm must have low time overhead, due to the high computation throughput of GPU.
This constraint makes complex schemes like hypergraph partitioning \cite{Hendrickson+:PC00} infeasible.
Secondly, we expect high partition equality in terms of task schedule
(partitioning), since up to thousands of concurrent threads share only a few KB of cache space, making random or greedy partitioning methods (e.g., techniques in PowerGraph \cite{gonzalez2012powergraph}) impractical.
Thirdly, the task partition results should be highly balanced because of the GPU's
single instruction multiple thread (SIMT) execution model, in which every thread
executes the same code but on different input. The execution time of a parallel
program depends on the execution time of the longest task. Imbalanced task
schedule may cause significant slowdown. This is different from CPU processors,
which can typically tolerate a relatively higher degree of imbalance because of
limited amount of parallelism \cite{malewicz2010pregel}.
To address the above problems, we propose to use an \emph{edge-centric}
graph-based model and we propose a fast task partition
algorithm for the \emph{edge-centric} model. Our partition algorithms has low
overhead compared with \emph{vertex-centric} model and yet provides high quality
balanced partition results.
Our contributions can be summarized as follows:
\begin{itemize}
\item This is the first \emph{edge-centric} model for GPU cache performance. It
overcomes the problems of traditional \emph{vertex-centric} model and captures
communication volume accurately.
\item We propose a task partition algorithm for the \emph{edge-centric} model.
Our algorithm gives the best approximation bound we know so far.
\item Our task partition algorithm is fast, often of orders of magnitude faster
than previous \emph{hypergraph} model and other \emph{edge-centric} partition
algorithm.
\item Our algorithm is robust. It provides steady performance improvement
regardless of the type of cache that is used: software cache or hardware cache,
and regardless of the thread block sizes (work group size).
\end{itemize}
The rest of the paper is organized as follows.
Section \ref{sec:background} describes the background of this paper.
Section \ref{sec:tech} introduces our task partition model and algorithm.
Section \ref{sec:appl} describes how to apply our model in GPU programs.
Section \ref{sec:eval} shows our experimental environment and evaluation results.
Section \ref{sec:rel} discusses the related work, and Section \ref{sec:conclusion} concludes the paper.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we propose a task partition technique to improve data sharing among different GPU threads.
We use data-affinity graphs to model data sharing and map task scheduling problem to an edge partition problem.
This is the first time the \emph{edge-centric} model is used for GPU cache
performance modeling. We propose a novel partition algorithm based on the
\emph{edge-centric} model, and our algorithm provides high quality task schedule
and yet is low-overhead.
We also provide rigorous proof for the analytical bound of our algorithm.
Our experiments show that our method can improve data sharing and thus performance significantly for various GPU applications.
\bibliographystyle{abbrv}
\subsection{Model Extension}
In realistic scenarios, there are data-affinity graphs with different features.
Using our basic EP model could not always achieve the best efficiency for all kinds of graphs.
Therefore, we extend our basic model to deal with certain types of data-affinity graphs.
Note that these extensions are not essential to the EP model.
\paragraph{Preferential grouping}
In practice, a computation task may consist of more than two operations (edges),
assuming one operation is a unit computation that uses two data operands. This is
usually specified by the programmer, who will specify which operations (edges)
should be preferred to be grouped together. These agglomerate operations should
not be split during workload partition. We handle this problem with
\emph{preferential grouping} in which we coarsen the graph by merging original edges
that should belong to the same group.
In \emph{preferential grouping}, we still perform a clone-and-connect
transformation in order to have a
vertex partition problem. However, we make slight adaptation in the
clone-and-connect stage. In the cloning stage, we only clone the vertices that
appear in different preferential groups of edges. In the connecting stage, we
connect cloned vertices with auxiliary edges to form a path for every set of
cloned vertices. Then we merge all original edges in every preferential group into a vertex.
We show an example of preferential grouping in Figure \ref{fig:pregroup}. Figure \ref{fig:pregroup} (a) shows the
original data-affinity graph. Figure \ref{fig:pregroup} (b) show there are
four preferential groups and show the graph after vertex cloning. Figure
\ref{fig:pregroup} (c) shows the cloned vertices are connected, but this time, the
number of auxiliary edges added for every set of cloned vertices, corresponds to
the total number of preferential groups the cloned vertices fall into subtracted by
1. In Figure \ref{fig:pregroup} (d), we identify the preferential groups and
merge each group into one vertex. Figure \ref{fig:pregroup} (e) shows the
coarsened graph after edge merging.
The effect of preferential grouping is to make sure the edges that are specified
to be in the same group does not get assigned to different partitions. The
analytical bound still holds with slight modification on the $d_{max}-1$ factor.
Previously, the maximum degree $d_{max}$ is used to represent the worst case
scenario. Here, we use $g_{max}-1$ as the worst case approximation factor, where
$g_{max}$ is the maximum number of preferential edge groups a vertex falls into.
The proof is similar to the proof for Theorem \ref{thrm:bound}.
This preferential grouping technique also applies when one computation task
involves more than two data objects and we do not specify how the binary
operations are among these data objects. We still perform the above modified
clone-and-connect process followed by a merging of vertices in the same preferential
task group. The above bound holds as well.
\paragraph{Handling frequently reused data objects}
When one vertex in the data-affinity graph has a large degree such that its
incident edges consist of a large portion of the total number edges, for
instance $>50\%$, we can take advantage of this feature and use it to reduce
partition overhead. An example of such case is shown in Figure
\ref{fig:scalefree}. 8 out of the 9 vertices in the Figure \ref{fig:scalefree}(a) has
degree of 1 or 2 except one, which has degree of 7, and this vertex's 7 incident
edges comprise over 70\% of the total edges.
To reduce the partition overhead and maintain good partition results, we split
large-degree vertices first before the actual edge partition. We
define \emph{splitting} here as the process of duplicating a vertex the same number of times as its
degree and we do not add any auxiliary edge to the duplicated vertices. An example is
shown in Figure \ref{fig:scalefree} (b). Since a large-degree vertex's incident edges
constitute a significant portion of the total number of edges, assuming it is
$x\%$, then this vertex at least appears in x\% of the final work partitions by
pigeon hole rule. By
partitioning the split graph, in the worst case, a large-degree
vertex will appear in every work (edge) partition, but since it already has to
appear in a large percentage of clusters, the additional cost is small, which
is at most $100/x$ times the optimum.
Furthermore, we can reduce the worst case $100/x$
factor by placing this set of large-degree data objects into physical cache
(partitions) at once. Recall that in GPU execution model, one thread block can
be mapped to one SM only, but one SM can be mapped to multiple thread blocks,
which take turns to use cache (partitions), as if the cache is virtualized.
Since this set of data is shared by most of the thread blocks, we can load them
into physical cache once, and no matter how many thread blocks there are, we do
not have to reload this set of data whenever a thread block is scheduled. To
achieve this, we do not use software cache (shared memory), since in current programming model,
software cache (shared memory) is divided exclusively among thread blocks and different thread
blocks cannot share software cache (shared memory) data. We use hardware cache instead and we use texture cache or constant cache. These
cases may not happen often, but still exist in a type of important graphs
called scale-free graphs, in which the probability of the vertex degree follows a
power-law distribution. By splitting the graph first, we can reduce
the overhead of the ensuing edge partition since no auxiliary edges need to be
added for the set of duplicated large degree vertices. Figure
\ref{fig:scalefree} (b) and (c) show an example of 3-way edge partition on split
graphs.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{figures/scalefree.pdf}
\caption{Handling frequently reused data node}
\label{fig:scalefree}
\end{figure}
\fi
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\linewidth]{figures/code.pdf}
\caption{Example of code transformation.}
\label{fig:code_trans}
\end{figure}
\section{Program Transformation}
\label{sec:appl}
In this section, we will introduce how to apply the partition result of our EP model to GPU programs for data sharing optimization.
Figure \ref{fig:code_trans} shows an example of original and transformed code that enables data sharing optimization.
For the host function on the CPU side in Figure \ref{fig:code_trans}(b), it first spawns a separate CPU thread to perform data sharing optimization.
The optimization process will be introduced in Section \ref{sect:appl:process}.
We use a separate thread for optimization to prevent it from adversely affecting the performance of the main program.
Assume the GPU kernel is called multiple times.
Before every kernel call, we check if the optimization process has been completed.
If so, the optimized kernel will be called with the optimized input.
Otherwise, the original kernel is called as usual.
The details of how we control optimization overhead, as well as the situation when the GPU kernel is only called once, will be discussed in Section \ref{sect:appl:overhead}.
If the target cache for data sharing optimization is texture cache,
the host function binds the optimized data layout to texture memory using the CUDA built-in function {\em cudaBindTexture()}.
The GPU kernel prefixes every data reference using {\em tex1Dfetch()} as shown in Figure \ref{fig:code_trans}(c) so that referenced data will be cached automatically by texture cache.
If the target cache is software cache, the kernel function on the GPU side in Figure \ref{fig:code_trans}(d) requires three major modifications in order to leverage optimized input.
First, an array local\_arrayA is created in software cache to store shared data.
Then, each thread block loads its shared parts from the input array (opt\_arrayA) into the local\_arrayA.
To minimize memory divergence \cite{Zhang+:ASPLOS11, Wu+:PPoPP13}, memory
accesses are coalesced into as few contiguous memory segments as possible.
At last, we replace the reference of the original input array with that of the local array which resides in software cache.
\subsection{Optimization Process}
\label{sect:appl:process}
Figure \ref{fig:workprocess} illustrates the working process of data sharing optimization.
The data access pattern of a GPU program is determined by both program semantics and input data.
Therefore, we extract the data access pattern to build the data-affinity graph at first.
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{figures/optimize_process.pdf}
\caption{Optimization workflow.}
\label{fig:workprocess}
\end{figure}
After data-affinity graph is built, we examine the graph before actually partitioning it using EP model.
First, we check if there is enough data reuse by checking the degree frequency
of the data-affinity graph.
If there is little data reuse, then we omit the partition process and use the original input.
Secondly, we check if the graph follows a special pattern such as clique, path, complete bipartite, etc.
For these special graphs, we have a preset optimal partitioning schedule using
the EP model offline, and thus we use the preset partition pattern.
If we determine that task partitioning/reschedule is necessary for the generated
data-affinity graph, we apply the EP algorithm to perform task partitioning as described in Section \ref{sec:tech}.
At last, we reorganize tasks among thread blocks based on the partition result
and we reorganize data layout as well. We use the cpack algorithm
\cite{Ding+:PLDI99}
to perform data layout transformation based on our task schedule.
After the data layout is transformed (from arrayA to opt\_arrayA in Figure
\ref{fig:code_trans}), the data references need to be updated as well and the
new index array (opt\_indexA in Figure \ref{fig:code_trans}) is passed to the optimized kernel.
\subsection{Adaptive Overhead Control}
\label{sect:appl:overhead}
To reduce overhead at runtime, we perform data sharing optimization using a separate thread on the CPU while kernel is executed on the GPU, so that the optimization process can run asynchronously.
We utilize a CPU-GPU pipelining technique to overlap computation and optimization for GPU programs \cite{Zhang+:ASPLOS11, Wu+:PPoPP13}.
In real GPU programs, a kernel is usually called within a loop, especially in scientific computation programs.
In such scenarios, We check if the asynchronous optimization is completed before calling the kernel and apply the optimization if so.
If the optimization thread does not complete when the program finishes, we terminate it to guarantee no slowdown.
If a kernel is only called once, we perform \emph{kernel splitting} \cite{Zhang+:ASPLOS11} which breaks one execution of a kernel into multiple executions of the same kernel with a smaller number of threads.
Then we can apply the asynchronous optimization method.
It is, however, possible that the transformed kernel runs slower than the original kernel. To prevent such cases from degrading performance, we record the transformed kernel runtime the first time it runs, and compare it with the original kernel runtime.
If the first run of the transformed kernel is slower, then we fall back to the original kernel in the next iteration.
\subsection{Model Extension}
In realistic scenarios, there are data-affinity graphs with different features.
Using our basic EP model could not always achieve the best efficiency for all kinds of graphs.
Therefore, we extend our basic model to deal with certain types of data-affinity graphs.
Note that these extensions are not essential to the EP model.
\paragraph{Preferential grouping}
In practice, a computation task may consist of more than two operations (edges),
assuming one operation is a unit computation that uses two data operands. This is
usually specified by the programmer, who will specify which operations (edges)
should be preferred to be grouped together. These agglomerate operations should
not be split during workload partition. We handle this problem with
\emph{preferential grouping} in which we coarsen the graph by merging original edges
that should belong to the same group.
In \emph{preferential grouping}, we still perform a clone-and-connect
transformation in order to have a
vertex partition problem. However, we make slight adaptation in the
clone-and-connect stage. In the cloning stage, we only clone the vertices that
appear in different preferential groups of edges. In the connecting stage, we
connect cloned vertices with auxiliary edges to form a path for every set of
cloned vertices. Then we merge all original edges in every preferential group into a vertex.
We show an example of preferential grouping in Figure \ref{fig:pregroup}. Figure \ref{fig:pregroup} (a) shows the
original data-affinity graph. Figure \ref{fig:pregroup} (b) show there are
four preferential groups and show the graph after vertex cloning. Figure
\ref{fig:pregroup} (c) shows the cloned vertices are connected, but this time, the
number of auxiliary edges added for every set of cloned vertices, corresponds to
the total number of preferential groups the cloned vertices fall into subtracted by
1. In Figure \ref{fig:pregroup} (d), we identify the preferential groups and
merge each group into one vertex. Figure \ref{fig:pregroup} (e) shows the
coarsened graph after edge merging.
The effect of preferential grouping is to make sure the edges that are specified
to be in the same group does not get assigned to different partitions. The
analytical bound still holds with slight modification on the $d_{max}-1$ factor.
Previously, the maximum degree $d_{max}$ is used to represent the worst case
scenario. Here, we use $g_{max}-1$ as the worst case approximation factor, where
$g_{max}$ is the maximum number of preferential edge groups a vertex falls into.
The proof is similar to the proof for Theorem \ref{thrm:bound}.
This preferential grouping technique also applies when one computation task
involves more than two data objects and we do not specify how the binary
operations are among these data objects. We still perform the above modified
clone-and-connect process followed by a merging of vertices in the same preferential
task group. The above bound holds as well.
\paragraph{Handling frequently reused data objects}
When one vertex in the data-affinity graph has a large degree such that its
incident edges consist of a large portion of the total number edges, for
instance $>50\%$, we can take advantage of this feature and use it to reduce
partition overhead. An example of such case is shown in Figure
\ref{fig:scalefree}. 8 out of the 9 vertices in the Figure \ref{fig:scalefree}(a) has
degree of 1 or 2 except one, which has degree of 7, and this vertex's 7 incident
edges comprise over 70\% of the total edges.
To reduce the partition overhead and maintain good partition results, we split
large-degree vertices first before the actual edge partition. We
define \emph{splitting} here as the process of duplicating a vertex the same number of times as its
degree and we do not add any auxiliary edge to the duplicated vertices. An example is
shown in Figure \ref{fig:scalefree} (b). Since a large-degree vertex's incident edges
constitute a significant portion of the total number of edges, assuming it is
$x\%$, then this vertex at least appears in x\% of the final work partitions by
pigeon hole rule. By
partitioning the split graph, in the worst case, a large-degree
vertex will appear in every work (edge) partition, but since it already has to
appear in a large percentage of clusters, the additional cost is small, which
is at most $100/x$ times the optimum.
Furthermore, we can reduce the worst case $100/x$
factor by placing this set of large-degree data objects into physical cache
(partitions) at once. Recall that in GPU execution model, one thread block can
be mapped to one SM only, but one SM can be mapped to multiple thread blocks,
which take turns to use cache (partitions), as if the cache is virtualized.
Since this set of data is shared by most of the thread blocks, we can load them
into physical cache once, and no matter how many thread blocks there are, we do
not have to reload this set of data whenever a thread block is scheduled. To
achieve this, we do not use software cache (shared memory), since in current programming model,
software cache (shared memory) is divided exclusively among thread blocks and different thread
blocks cannot share software cache (shared memory) data. We use hardware cache instead and we use texture cache or constant cache. These
cases may not happen often, but still exist in a type of important graphs
called scale-free graphs, in which the probability of the vertex degree follows a
power-law distribution. By splitting the graph first, we can reduce
the overhead of the ensuing edge partition since no auxiliary edges need to be
added for the set of duplicated large degree vertices. Figure
\ref{fig:scalefree} (b) and (c) show an example of 3-way edge partition on split
graphs.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{figures/scalefree.pdf}
\caption{Handling frequently reused data node}
\label{fig:scalefree}
\end{figure}
\fi
\section{Applicability}
\label{sec:appl}
In this section, we will introduce how to apply our EP model to optimize data sharing for real GPU programs.
Figure \ref{fig:code_trans} shows an example of original and transformed code.
Software cache is used in this example to store shared data within thread blocks.
For the host function on CPU side (main() in Figure \ref{fig:code_trans}), first, it spawns another CPU thread to perform data sharing optimization.
The optimization process will be introduced in Section \ref{sect:appl:process}.
We use separate thread for optimization to prevent it from adversely affecting the performance of GPU kernel.
Assume the GPU kernel is called multiple times.
Before every kernel call, we check if the optimization process has been completed.
If the answer is yes, the optimized kernel will be called with the optimized input.
Otherwise, the original kernel is called as usual.
The detail of our overhead control scheme as well as the situation when the GPU kernel is only called once will be discussed in Section \ref{sect:appl:overhead}.
For the kernel function on GPU side, there are three major modifications in order to leverage optimized input.
First, an array (local\_arrayA) is created in software cache to store shared data.
Then, each thread block loads its shared parts from the input array (opt\_arrayA) into the software cache array.
To avoid memory divergence which could cause severe performance degradation \cite{Zhang+:ASPLOS11, Wu+:PPoPP13}, adjacent threads are made to load adjacent data elements in the input array.
At last, we replace the reference of the original input array with that of the local array which resides in software cache.
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{figures/process.pdf}
\caption{Example of code transformation.}
\label{fig:code_trans}
\end{figure}
For the optimized kernel using texture cache, the host function needs to bind the optimized data layout to texture memory using the CUDA build-in function {\em cudaBindTexture()}.
The GPU kernel prefixes every data reference using function {\em tex1Dfetch()}. Then texture memory data are cached automatically by texture cache.
\subsection{Optimization Process}
\label{sect:appl:process}
Figure \ref{fig:workprocess} illustrates the working process of data sharing optimization.
The data access pattern of GPU programs are determined by both program semantics and input data.
Therefore, we extract data access pattern to build the data-affinity graph at first.
After data-affinity graph is built, we pre-examine the graph before actually partitioning it using EP model.
First, we check if there is enough data reuse by scanning the vertex degrees of data-affinity graph.
If there is no data reuse, or there is data reuse but the program's default scheduling is near optimal, we omit the partition process and use the original input.
Second, we check if the graph fellows a special pattern such as clique, path, complete bipartite, and etc. during the graph building.
For these special graphs, we can directly get the optimal task partition result without calling the EP model to reduce unnecessary partition overhead.
If we identify that task partition is necessary for the generated data-affinity graph, our EP model is applied on the graph to perform task partition as described in Section \ref{sec:tech}.
At last, we reorganize tasks among thread blocks based on the partition result by reordering input data layout.
In common GPU program model, a certain thread gets its task by reading a fixed location in the input data, and the location depends on the id of thread.
Therefore, we reorganize the data layout in the input (from arrayA to opt\_arrayA in Figure \ref{fig:code_trans}) to achieve the goal of rearranging tasks.
The mapping relationship from original to optimized data layout is also recorded and passed to the optimized kernel (opt\_indexA in Figure \ref{fig:code_trans}) to ensure correctness.
Besides, if shared data are loaded into software cache in the optimized kernel, we calculate the size of shared data within each thread block (beginA) to facilitate data loading of software cache.
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{figures/optimize_process.pdf}
\caption{Optimization process.}
\label{fig:workprocess}
\end{figure}
\subsection{Adaptive Overhead Control}
\label{sect:appl:overhead}
To reduce overhead at runtime, we perform data sharing optimization using a separate thread on CPU while kernel is executed on GPU so that optimization process can run asynchronously.
We utilize a CPU-GPU pipelining technique to overlap computation and optimization for GPU programs \cite{Zhang+:ASPLOS11, Wu+:PPoPP13}.
In real GPU programs, a kernel is usually called multiple times, especially in scientific computation programs.
In such scenarios, We check if the asynchronous optimization is completed before calling the kernel and apply the optimization if so.
If the optimization work does not complete when the program finishes, we terminate it to guarantee no slowdown.
If a kernel is only called once, we perform \emph{kernel splitting} \cite{Zhang+:ASPLOS11} which breaks one execution of a kernel into multiple executions of the same kernel with smaller number of threads.
Then we can apply the same method described above to amortize optimization overhead.
It is however possible that transformed kernel runs slower than the original kernel.
To prevent such cases from degrading performance of the overall execution, we record the transformed kernel time the first time it runs and compare it with the original kernel time.
If the first run of the transformed kernel is slower, we fall back to the original kernel in the next iteration.
\section{Background}
\label{sec:background}
We first describe the abstract memory model for a GPU architecture. As mentioned
earlier, a GPU is composed of a number of streaming multiprocessors (SMs or
SMXs). There is a cache on every SM, and threads running on the same SM share
this cache. For a GPU program, the minimal unit of threads that run on one SM is
called a thread block, also called \emph{cooperative thread arrays (CTAs)} using
NVIDIA CUDA terminology. The GPU threads are divided into blocks of the same
size. One thread block can run at most on one SM at one time. Multiple thread
blocks can run simultaneously on one SM. A thread block
would acquire one SM and hold the SM until all its threads are finished. As one thread
block releases the SM, another thread block (if there is any) will acquire the
SM. As such, a thread block is also the minimal thread work group that share
cache. We show the abstract cache sharing model in Figure \ref{fig:cachearch}.
Every thread block uses a partition of the cache on one SM and would release it
when the thread block is finished. Therefore, there are as if as many virtual
cache as the number of thread blocks.
There are two types of caches on the SM: software cache and hardware cache.
Software cache is scratch-pad memory, which is also referred to as {\em shared memory} in NVIDIA CUDA terminology\footnote{We use NVIDIA CUDA terminology throughout the paper.}.
Software cache requires explicit load/store instructions to move frequently used data from/into
memory. Every thread block gets an exclusive and even partition of {\em shared
memory}. Similarly, the acquired partition of the software cache will not be released
until the owner thread block finishes. A thread block cannot see another thread
block's
software cache (shared memory) partition.
Hardware caches can be further classified into several categories based on their usage.
In recent GPU architectures, the L1 cache is only used to store local variables that are private to each thread \cite{KeplerWhitePaper, MaxwellWhitePaper}.
The texture cache can store shared data objects after the data objects are binded with texture memory.
Thus we use texture cache as the hardware cache to evaluate our task
schedule/partition algorithms. The hardware cache does not require software
to explicitly load frequently used data into cache, and has good
programmability. However, it may not always keep the right data in cache since
it does not have as much knowledge about the program as the programmer.
Therefore, there is the trade-off between programmability and performance. We
show the comparison between software cache and hardware cache in the evaluation
section.
Finally, there is a L2 cache which is shared by all SMs on a GPU. A L2 cache
has a typical hit latency of hundreds of cycles (close to memory access latency
on GPU).
Because of its long latency and also because it is unified across all SMs like
the main memory, we
only focus on optimization of the first level cache on every SM.
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{figures/cachearch.pdf}
\caption{Abstract cache model for GPU cache architecture. TBLOCK represents a
thread block. }
\label{fig:cachearch}
\end{figure}
\section{Related Work}
\label{sec:rel}
\paragraph{CPU task partition for communication reduction}
Some models have been proposed to model data communication in traditional CPU parallel systems, and these models are used to reduce communication among processors through task scheduling.
Hendrickson et al. use vertex-partition based graph models where vertices represent tasks \cite{Hendrickson+:SIAMJSC00, Hendrickson+:PC00}.
The vertex-partition model cannot measure data communication accurately, and thus we use edges to represent tasks instead.
Hypergraph models \cite{karypis1998hmetis, catalyurek1999patoh} can model communication cost accurately.
The main drawback of the hypergraph model is its large overhead, as we demonstrate in Section \ref{sect:tech:compare}, which makes it infeasible for GPU computing.
Pregel \cite{malewicz2010pregel} and GraphLab \cite{low2012distributed} introduce two parallel computation models based on message passing and shared memory, respectively.
They always assign all computation of one vertex to one processor due to limitations of the programming model, while our model allows computation to be scheduled arbitrarily to achieve better partition quality and balance.
PowerGraph \cite{gonzalez2012powergraph} can distribute the computation of one vertex into several processors and proposes two edge partition methods for task scheduling.
One is to randomly assign edges, and the other, greedy based method assigns an edge to those partitions which already possess its endpoints.
These simple edge partition schemes produce low-quality partitions and are thus inapplicable for GPU task partitioning, as demonstrated in Section \ref{sect:tech:compare}.
Ding and Kennedy propose using runtime transformation for improving memory performance of irregular programs \cite{Ding+:PLDI99}, but it is heuristics based and there is no rigorous model for data reuse.
Bondhugula et al. introduce an automatic source-to-source transformation framework to optimize data locality, and they formulate data locality problem with polyhedral model \cite{Bondhugula+:PLDI08}, which only works for regular applications with affine memory indices.
\paragraph{GPU task partition for divergence elimination}
The dynamic task scheduling work for GPU memory performance optimization is mainly in memory coalescing rather than data sharing.
Zhang et al. propose to dynamically reorganize data and thread layout to minimize irregular memory accesses \cite{Zhang+:ASPLOS11}.
Wu et al. also propose two data reorganization algorithms to reduce irregular accesses \cite{Wu+:PPoPP13}.
However, these papers do not address the data sharing problem in GPU caches.
\paragraph{Other GPU software optimization techniques}
Many compiler techniques are proposed to achieve better utilization of GPU memory.
For affine loops, Baskaran et al. use a polyhedral compiler model to reduce non-coalesced memory accesses and bank conflicts in shared memory \cite{Baskaran+:ICS08}.
Jia et al. propose to characterize data locality and then guide GPU caching \cite{Jia+:ICS12}.
The limitation of these compiler methods is that they cannot address global data sharing among threads.
Some research uses hints provided by programmers to help compilers improve GPU memory performance.
CUDA-lite tunes shared memory allocation via annotations \cite{Ueng+:LCPC08}.
hiCUDA seeks to automate shared memory allocation with the help of programmer specified directives \cite{Han+:TPDS11}.
There are also application-specific studies for sparse matrix vector multiplication.
Bell and Garland discuss data structures of SPMV for various sparse matrix formats \cite{Bell+:NVIDIATR08}.
Choi et al. propose an automatic performance tuning framework for SPMV \cite{Choi+:PPoPP10}.
Volkov and Demmel analyze the bottleneck in dense linear algebra and optimize its performance by improving on-chip memory utilization and etc
\cite{Volkov+:SC08}.
But none of this handles the fundamental task scheduling problem for SPMV locality enhancement.
Nonetheless, these techniques are complementary to our technique. For instance, we can use auto-tuning to find out whether to use texture cache or software cache.
\section{Edge-Centric Model}
\label{sec:tech}
\subsection{Problem Definition}
\label{subsec:model}
We build a edge-centric graph model to describe the relationship between data objects and tasks.
In our model, we describe a data object as a vertex and a task as
an edge.
\iffalse
\begin{figure}[htb]
\centering
\begin{lstlisting}
void kernel(float density[], float energy[],
float3 momentum[]) {
int i = blockDim.x*blockIdx.x + threadIdx.x;
float density_i = 0, energy_i = 0;
float3 momentum_i = {0, 0, 0};
foreach neighbor j of particle i
{
density_i += compute_density_impact(density[i],
density[j]);
energy_i += compute_energy_impact(energy[i],
energy[j]);
momentum_i += compute_momentum_impact(momentu
m[i], momentum[j]);
}
update(density_i, density_energy_i, momentum_i);
}
\end{lstlisting}
\caption{Pseudo code of cfd kernel.}
\label{fig:cfd_code}
\end{figure}
\fi
\begin{definition}
We define a data-affinity graph $D = (V, E)$ with the set of vertices $V$ and the set of
edges $E \subset V \times V$. Let $n$ and $m$ denote the number of vertices and the number of edges,
respectively. A vertex $v \in V$ represents a data object. An edge $e \in E$ with $e = (u,v)$
denotes a computation task that uses the two data objects $u$ and $v$.
\end{definition}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{figures/reduction.pdf}
\caption{Balanced edge partition problem converted to balanced vertex partition problem.}
\label{fig:epart}
\end{figure*}
Here every data object has the same attribute and size.
The definition of a data object typically depends on the semantics of the program.
One task could correspond to one or a group of operations in the program based on the semantics.
For instance, \emph{cfd} \cite{Che+:IISWC09} calculates the interaction of two
particles based on their density, energy, and momentum variation.
We group these three attributes of one particle into one data object. Thus one
task is defined as the calculation based on these two aggregate data objects
corresponding to this pair of particles.
Our goal is to partition all $m$ edges in the data-affinity graph evenly into $k$ clusters, such that every edge (task) is assigned to exactly one cluster (thread block).
In short, this is a balanced edge partitioning (EP) problem. The optimization objective is to maximize data reuse, which is equivalent to minimizing the number of times a vertex (data object) appears in different clusters (thread blocks).
\begin{definition}
\label{def:epproblem}
A vertex $v$ is cut if its incident edges belong to more than one cluster. We
define $c_v$ as the cost that $v$ is cut, which is the number of unique clusters
that $v$'s incident edges belong to $p_v$ subtracted by 1, $c_v = p_v - 1$. We define
a total vertex-cut cost $C$ which has to be minimized under the condition of edge balancing.
Let $C = \sum_{v \in V } (p_v - 1)$ and let $L_i$ denote the number of edges in
cluster i, the {\em k$-$way balanced edge partition (EP)} problem is defined as:
\begin{equation}\label{commcost}
\begin{array}{rlll}
\displaystyle \min & { C(x) = \sum_{v \in V} (p_v - 1)} \\
\textrm{s.t.} & \forall i \in [1...k] ~~ L_i(x) = \frac{m}{k} \\
& x~is~a~valid~edge~partitioning
\end{array}
\end{equation}
\end{definition}
We use the same \emph{cfd} example from Section \ref{sec:intro} to illustrate the
edge partition model. Figure \ref{fig:epart}(a) shows an example of the data affinity graph for \emph{cfd}.
In this example, there are 6 interactions to be computed.
Assume a two-way balanced edge partition, i.e., $k=2$.
The total vertex cut cost of the partition in Figure \ref{fig:epart}(e) is one, since one vertex (data object) appears in two edge clusters.
The number of times the data objects appear in different edge partitions represents how often a data object is loaded into different thread blocks.
In the ideal case, each data object appears in only one partition.
In such scenarios, threads sharing data all reside in one thread block, and thus there is no redundant data access.
Each time one data object is shared by another thread block, one extra redundant data access is required.
Therefore, the total vertex cut $C(x) = \sum_{v \in V} (p_v-1)$ represents the number of redundant data accesses.
We also refer to it as the \emph{data reuse cost}, which we use to evaluate the quality of partitioning.
\subsection{Problem Transformation}
\label{subsec:alg}
It can be easily proved that the problem of balanced edge partitioning in
Definition \ref{def:epproblem} is NP-complete by reduction from existing NP-hard partition problems. For lack of space, we omit the proof here.
The edge partition problem is a non-traditional graph partition problem, and it lacks sophisticated solutions.
On the other hand, there is a rich exploration on both a theoretical foundation and practical solutions for the vertex partition problem.
We will show that we can convert the proposed balanced edge partition problem into the balanced vertex partition problem to leverage sophisticated vertex partition algorithms for efficient edge partitioning.
As far as we know, our work is the first to map a balanced edge partition problem into a balanced vertex partition problem.
We will prove that, not only does our solution work fast in practice, but also guarantees a polynomial algorithm with an worst case approximation factor of $(d_{max}-1)\mathcal{O}(\sqrt{\log{}m\log{}k})$, which is the best bound we know by far\footnote{$d_{max}$ is the maximum vertex degree in the data-affinity graph.}.
To convert the edge partition problem to a vertex partition problem, we first transform the data-affinity graph.
We replace every vertex in the graph
with a set of new \emph{cloned vertices} such that every cloned vertex is connected to
one unique incident edge of the original vertex. The cloning process is shown in
Figure \ref{fig:epart}(b). Next, assuming the degree of the original vertex is $d$, we add $d-1$
\emph{auxiliary edges} to connect the set of $d$ cloned vertices to form a path.
This process of adding auxiliary edges is shown as an example in Figure \ref{fig:epart}(c).
Formally,
\begin{figure*}[htp]
\includegraphics[width=0.24\textwidth]{figures/cant_pdf.png}
\includegraphics[width=0.24\textwidth]{figures/circuit5M_pdf.png}
\includegraphics[width=0.24\textwidth]{figures/in2004_pdf.png}
\includegraphics[width=0.24\textwidth]{figures/scircuit_pdf.png}
\caption{Degree distribution: the frequency of a degree in the graphs for the
graphs \emph{cant}, \emph{circuit5M}, \emph{in2004}, \emph{scircuit}. For
readability, we only show part of the x range for circuit5M, in2004 and
scircuit. }
\label{fig:4pdf}
\end{figure*}
\begin{figure}[htp]
\includegraphics[width=0.23\textwidth]{figures/in2004_loglog.png}
\includegraphics[width=0.23\textwidth]{figures/scircuit_loglog.png}
\caption{Log-log scale of degree distribution for in2004 and scircuit}
\label{fig:loglog2}
\end{figure}
\begin{definition}
\label{def:graphconv}
We define a graph transformation as a clone-and-connect transformation if it
satisfies following conditions. Assume $t$ is a legal clone-and-connect
transformation and $D \xmapsto{t} D'$, where $D = (V, E)$ and $D' = (V',E')$.
The transformed graph $D'$ needs to maintain the following characteristics. For every
vertex $v \in V$ of degree $d$, there are $d$ corresponding cloned vertices $v'_{1},\ldots v'_{d} \in
V'$. For every edge $e \in E$ with endpoints $u, v \in V$, there is a corresponding edge $e' \in
E'$ with endpoints $u'_i, v'_j \in V'$ such that $u'_i$ is the i-th cloned vertices of $u$, $v'_j$ the
j-th cloned vertex of $v$, and no cloned vertex is connected to more than one
original edge.
For every original vertex $v$ of degree $d$, its cloned set of vertices $v_i$, $i=1...d$ in $D'$ are connected into form a path using $d-1$ auxiliary edges, where a path is defined as a tree that contains only vertices of degree 2 and 1.
\end{definition}
By Definition \ref{def:graphconv}, a cloned vertex is connected to exactly one original edge, and no two original edges share a cloned vertex as shown in Figure \ref{fig:epart}(b).
Therefore, we can get the total number of cloned vertices by doubling the total number of original edges.
If we evenly partition the vertices in the converted graph into $k$ clusters and make sure no original edge is cut, we can ensure the two endpoints of any original edge belong to the same partition, and thereby we can determine the corresponding original edge belongs to that partition, too.
In the meantime, we ensure that the number of original edges is balanced across all partitions, since every partition has the same number of vertices.
Formally,
\begin{definition}
Assume a clone-and-connect transformation t transforms graph $D = (V,E)$ to $D' =
(V',E')$, $D
\xmapsto{t} D'$, we perform vertex partition on $D'$ and obtain a
vertex partition solution $VP_{D'}$ such that no original edge is cut. From $VP_{D'}$, we can reconstruct an
edge partition solution of $D$, $EP_{D}$. We define a reconstruction process $m$
such that $VP_{D'} \xmapsto{m} EP_{D}$ as follows.
For every original edge $e'\in E'$ with $e' = (u',v')$ in $D'$, assume both $u'$ and
$v'$ fall into the i-th partition; then its corresponding edge $e \in D$, falls into the i-th partition as well.
\end{definition}
As shown in Figure \ref{fig:epart}(d), the edges $A$, $B$, $C$
are in the same cluster since their endpoints are all in the partition on the left.
Using this approach, we can map the vertex partition back to the edge partition
by checking every edge on which partition its endpoints fall into (shown in Figure \ref{fig:epart}(e)).
\iffalse
There are two ways to ensure no original edge is cut in the vertex partition: 1) We can
assign an infinite weight to every original edge in the converted graph, or 2)
we can merge the end vertices of every original edge into one vertex in the converted problem. Either
approach would work since in the vertex partition process, an infinite weight
edge would never be cut in the final solution (in practice we just have to set the weight of the original edge as a very large number) and a vertex cannot be cut either.
\fi
To ensure no original edge is cut in the vertex partition, we assign a very large weight to all original edges in the converted graph so that the vertex partition process will only cut auxiliary edges.
Also note that in the conversion process described in Definition \ref{def:graphconv},
there are several different ways to connect cloned vertices to form a path.
We choose to connect them in index order in practice.
Any other way to connect cloned vertices is fine, since auxiliary edges are used to represent data sharing relationship between original edges (tasks).
The balance degree of $D$'s edge partition is equal to that of $D'$'s vertex partition.
We find that existing vertex partition algorithms \cite{karypis1995metis} can produce rather balanced results.
The \emph{balance factor} is used to measure the balance degree of partition, which is calculated by dividing maximum partition size by average partition size \cite{karypis1995metis}.
It is typically less than 1.03 in practice.
\iffalse
In the example in Figure \ref{fig:epart}(c), we show one possible way to connect
the cloned vertices. In fact, no matter how we connect the cloned vertices for
this example, the optimum vertex partition would always map to the optimum edge
partition: \{A,B,C\}
in one cluster, and \{D,E,F\} in the other cluster.
\fi
We will show the mapping relationship between the optimum vertex partition and the optimum edge partition with rigorous analytical results in Section \ref{subsec:analytical}.
\subsection{Comparison with Other Methods}
\label{sect:tech:compare}
\begin{figure*}
\centering
\includegraphics[width=0.7\textwidth]{figures/comparison.pdf}
\caption{Our proposed EP model versus other partition methods. NEM represents
not-enough-memory errors. }
\label{fig:comparison}
\end{figure*}
Figure \ref{fig:comparison} compares our proposed conversion based edge partition model (EP model) with other existing task partition methods.
The experimental environment is introduced in Section \ref{sect:eval:env}.
For the vertex partition of the converted problem in our EP model, we use METIS \cite{karypis1995metis} library.
Five data-affinity graphs are constructed using representative input matrices
selected from the Florida matrix collection \cite{FloridaMat} and matrix market
\cite{MATRIXMARKET} (note that a sparse matrix usually represents a graph in or
is used in sparse matrix vector kernels).
We show the degree distribution for the data-affinity graphs in Figure
\ref{fig:4pdf}. We did not show the frequency of
degree for the \emph{mc2depi} data set, since its pattern is relatively simple,
with degree 2 of 5.70532e-04\% occurrence, degree 3 of 0.583654\% occurrence,
and degree 4 of 99.4158\% occurrence.
The selected data-affinity have different degree
distribution patterns. The degree of the \emph{cant} data-affinity graph is
between 0 and 40. The \emph{circuit5M} has a more random degree distribution and
we only show part of the x range for readability.
The \emph{mc2depi} data set has nodes of degree 2 to 4, resembling the meshes
(at most 4 neighbour particles ) in computation fluid dynamics. Two
of the data sets exhibit the power-law pattern: the \emph{in$-$2004} data set and
the \emph{scircuit} distribution. We further show the log$-$log scale for these
two data sets in Figure \ref{fig:loglog2}. Regardless of the type of degree
distribution, our algorithm always outperforms the classical
\emph{vertex-centric} algorithm.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/hyper.pdf}
\caption{Hypergraph model versus edge partition model.}
\label{fig:hyper}
\end{figure}
\paragraph{Comparison with hypergraph model \cite{Hendrickson+:PC00}}
A hypergraph is a special type of graph where an edge may connect more than two vertices.
Therefore an edge is also called a hyperedge.
In the hypergraph task partition model, unlike our EP model, a vertex represents a task, and a hyperedge represents a data object where it covers all vertices (tasks) that share that data object.
The goal of maximizing data sharing is equivalent to partitioning vertices (tasks) into $k$ clusters so that the number of times hyperedges are cut is minimized.
Figure \ref{fig:hyper} shows an example to compare the hypergraph model with the EP model.
We also show the optimum partition for both models.
In Figure \ref{fig:hyper}(a) one hyperedge is cut, which corresponds to the one cut vertex in Figure \ref{fig:hyper}(b).
\iffalse
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/modelcompare.pdf}
\caption{Hypergraph model versus basic EP model. EP model represents edge
partition model. NEM represents not enough memory. Quality is the data
reuse cost in Definition \ref{def:epproblem}.}
\label{fig:hyperperf}
\end{figure}
\fi
For the hypergraph partition, we use hMETIS \cite{karypis1998hmetis}, a multilevel hypergraph partition tool, and PaToH \cite{catalyurek1999patoh}, the fastest hypergraph partition implementation we are aware of.
From Figure \ref{fig:comparison}, we see that PaToH is faster than hMETIS in the hypergraph model, and our basic EP model is significantly faster than both of them in all cases.
The partition quality, measured as the data reuse cost in Definition \ref{def:epproblem}, shows that our EP model generates similar quality of partitions as PaToH and hMETIS.
The quality of both the hypergraph model and our EP model are better that of the default task scheduling shown in the ``default quality'' column, which indicates that data sharing within each thread block is improved significantly after partitioning.
Overall, our basic EP model is faster than the fastest hypergraph partition model and yet produces high quality results.
Another important thing to notice is that our EP model scales better compared to the hypergraph model with respect to graph size.
For small graphs such as \emph{scircuit}, our EP model is about 4 times faster,
while for large graphs such as \emph{circuit5M} and \emph{in-2004}, the EP model is tens of times faster.
\paragraph{Comparison with other edge partition methods}
PowerGraph \cite{gonzalez2012powergraph} proposes two edge partition based task scheduling methods.
Both methods go through all edges linearly to distribute them among partitions.
The random based method randomly assigns edges into partitions.
The greedy based method prioritizes choosing partitions that already possess the endpoints of the to-be-assigned edge.
If no such partition is found, then the partition with the fewest edges is selected to ensure balance.
Figure \ref{fig:comparison} shows the partition quality of both methods.
We do not show the partition time since both methods are expected to be fast.
Compared with hypergraph and our EP models, both of them have significantly worse partition quality because of their random or greedy nature.
In many cases, their quality is worse than that of the default scheduling.
In such scenarios, rescheduling tasks only degrades the performance.
Therefore, we conclude that these two partition methods are not applicable for complicated data sharing patterns exhibited in GPU programs.
\subsection{Approximation Bounds}
\label{subsec:analytical}
Our EP model is not only good in practice, but also strong in theory. We provide the proof for approximation factor between
$\mathcal{O}(\sqrt{\log{}m\log{}k})$ and
$(d_{max}-1)\mathcal{O}(\sqrt{\log{}m\log{}k})$. We describe it in two steps.
\begin{theorem}
With the clone-and-connect transformation t such that $D \xmapsto{t} D'$
where $D=(V,E)$ and $D'=(V',E')$. The edge cut cost of the vertex partition of
$D'$ is greater than or equal to the vertex cut cost (in Equation \ref{commcost}) of
the corresponding edge partition of $D$, $C_{vp}(D') \geq C_{ep}(D)$.
\label{lm:costconv}
\end{theorem}
First recall that in Section \ref{subsec:alg}, we have ensured in the
vertex partition of $D'$ that only auxiliary edges are cut. For every set of cloned
vertices $v'_i$, i $= 1...d$ in $D'$ that corresponds to the vertex $v$ of degree $d$
in $D$, they are connected with $d-1$ auxiliary edges to form a path. If $l$ out
of the $d-1$ auxiliary edges are cut, the total number of distinct edge
clusters $v$ fall into in $D$, is at most $l+1$. Therefore, the
cut cost of $v$ in the edge partition of $D$ is at most $l$. When
converting the vertex partition of $D'$ into the edge partition of $D$, we sum
up the total vertex cut cost and thus the
total number of auxiliary edges cut in $D'$ is greater than or equal to the vertex
cut cost in $D$.
\begin{theorem}
There exists a clone-and-connect transformation t, $D \xmapsto{t} D'$ such
that the edge cut cost of the optimal vertex partition of $D'$ is the same as
the vertex cut cost of the optimal edge partition of $D$. For all legal
clone-and-connect transformations $t \in T$, in the worst case, the
edge cut cost of the optimal vertex partition of $D'$ is $d_{max}-1$ times of
the vertex cut cost of the optimal edge partition of $D$, while $d_{max}$
represents the maximum degree of the graph $D$.
\label{thrm:bound}
\end{theorem}
The first task is to prove that there exists a clone-and-connect
transformation $t$ such that the transformed vertex partition problem is the
{\em dual} of the original edge partition problem. We construct such a
transformation using an optimal edge partition $EP_{opt}$ on the graph $D=(V,E)$. Using the
optimal edge partition solution, we know which edges are grouped to the same
partition. Then we determine the transformation $t$ as follows. First, the cloning
process is the same as in Definition \ref{def:graphconv}. In the connecting
phase, for every set of cloned vertices $v'_i$, $i = 1...d$ that connect to every
original edge, we first divide each set into $k$ groups with respect to which
partition their incident original edges belong to. Within every set of cloned vertices,
within every group, we connect the
cloned vertices to form a path, then we connect all $k$ groups into
one path. Such a transformation $t$ is constructed. It is easy to see that if we reverse this path
connecting process, we can obtain a balanced vertex partition of $D'$ that can
be mapped to the optimum edge partition solution of $D$, which is $EP_{opt}$ we
defined earlier.
Since we do not \emph{a priori} know the optimal edge partition of the
original graph $D$, when connecting the cloned vertices into a path, we connect
them arbitrarily. Assuming the converted graph with a priori knowledge of
optimal partition of $D$
is $D'_{opt}$, the second task is to prove that with an arbitrary connecting
process, the edge cut cost of the optimum vertex partition of the converted
graph $D'$ is at most
$d_{max}-1$ times of the edge cut cost of the vertex partition of $D'_{opt}$
since $D'_{opt}$ maps directly back to $D_{opt}$ ($D_{opt}$ is the optimal edge
partition of the original problem).
This step is easy to prove as well, since for every set of $d$ cloned vertices that
correspond to the cut vertex in the optimal edge partition solution of $D$, at
most all auxiliary edges are cut ($d-1$ cuts) and in the best case, at least one
auxiliary edge is cut, the edge cut cost in the vertex partition of $D'$ is thus
at most $d_{max}-1$ the edge cut cost in the vertex partition of $D'_{opt}$.
According to Theorem \ref{lm:costconv}, in the reconstructed edge partition of
$D$ from
the vertex partition of $D'$, the vertex cut cost of $D$ is less than or equal
to the
edge cut cost of $D'$, $C_{ep}(D) \leq C_{vp}(D')$. Thus the reconstructed edge
partition solution from the optimum vertex partition of $D'$ is at most
$d_{max}-1$ times the optimum solution of edge partition of $D$.
The vertex partition problem can be approximated with a factor of $\mathcal{O}(\sqrt{\log{}n\log{}k})$ for a graph with $n$
vertices and $k$ partitions \cite{Krauthgamer+:SODA09}. Therefore we can approximate our solution by a
factor of $(d_{max}-1)\mathcal{O}(\sqrt{\log{}m\log{}k})$ since our converted
graph has $2m$ vertices where $m$ is the number of edges in the original graph.
|
\section{Introduction}
For the last decade, {\em compressed sensing} (CS) has received considerable attention as a novel technology in signal processing research. The purpose of CS is to enhance signal processing performance by utilizing the notion of the {\em sparsity} of signals \cite{CScandes}--\cite{StarckMurtaghFadili2010}. Let us suppose that a sparse vector $\vm{x^{0}} \in\mathbb{R}^{N}$, many components of which are zero, is linearly transformed into vector $\vm{y} \in\mathbb{R}^{M}$ by an $M \times N$ measurement matrix $\vm{\Phi}$, where $\vm{y} = \vm{\Phi x^{0}}$. For a given pair of $\vm{\Phi}$ and $\vm{y}$, the reconstruction of $\vm{x^{0}}$ is required \cite{CandesWakin2008}. Many studies in CS research have shown that the sparsity of signals makes it possible to perfectly reconstruct $\vm{x}^0$ at a viable computational cost, even in the region of $\alpha = M/N <1$ \cite{CandesRombergTao2006}--\cite{Krzakala2012}. This has led to the hardware-level realization of accurate signal reconstruction that had hitherto been regarded as out of reach due to limitations on sampling rates \cite{bandlimit} and/or the number of sensors \cite{singlepixel}.
In the signal processing context, the CS framework eases the burden on analog-to-digital converters (ADCs) by reducing the sampling rate required to acquire and recover sparse signals. However, in practice, ADCs not only sample, but also quantize each measurement to a finite number of bits; moreover, there is an inverse relationship between achievable sampling rate and bit depth. Therefore, many discussions on CS have shifted emphasis from sampling rate to number of bits per measurement \cite{measurementsVSbits, ratedistortion}. In particular, we are here interested in the extreme case of 1-bit CS measurement, which captures just the sign as $\vm{y} = \text{sign}(\vm{\Phi x^{0}})$ \cite{1bitCSbaranuik}. Thus, the measurement operator is a mapping from $\mathbb{R}^{N}$ to the Boolean cube $B^{M}:= \{ -1, 1 \} ^M$. This is highly beneficial in practice due to the significant reduction in the size of data that are transmitted and stored.
It is obvious that the scale (absolute amplitude) of the signal is lost in 1-bit CS measurements. To compensate for this, past studies have proposed the imposition of an additional constraint whereby the $l_2$-norm of the signal is normalized to a fixed constant \cite{1bitCSbaranuik, 1bitCSL1XuKaba}. In other words, this can only reconstruct the directional information but not the true scale information of the signal. Moreover, it yields another drawback such that solving the reconstruction problem becomes nontrivial, since the problem is no longer formulated as a convex optimization. In order to address these issues, by introducing a set of finite thresholds $\vm{\lambda} = (\lambda_\mu)$ $(\mu = 1, 2, \ldots, M)$ to the quantizer as
\begin{equation}
\vm{y} = \text{sign}(\vm{\Phi x} + \vm{\lambda}),
\label{1-bit-measurement}
\end{equation}
and combining the knowledge of the thresholds, we are able to estimate the scale of the signal {{\cite{1bitCSNormEstKKR}}}.
Furthermore, as the feasible set provided by the constraint of (\ref{1-bit-measurement}) for given measurements $\vm{y}$ is a convex region of $\vm{x}$, one can reconstruct a sparse signal in polynomial time by solving the $l_1$-norm minimization problem
\begin{equation}
\vm{\hat{ x}} = \underset{\vm{x} \in\mathbb{R}^{N}}{\text{argmin }} ||\vm{x}||_{1} \text{ subject to } \vm{y} = \text{sign}(\vm{\Phi x} + \vm{\lambda})
\label{thresh1bitCS}
\end{equation}
by using versatile convex optimization algorithms \cite{cvx}.
A lingering, natural question is how should we set the values of $\lambda_\mu$. To partially answer this, we compare two strategies: one involves fixing the thresholds at a constant value $\lambda_\mu = \lambda$ for all measurements, and the other consists of independently selecting $\lambda_\mu$ from an identical Gaussian distribution.
{{In \cite{1bitCSNormEstKKR}, worst case bounds of the number of measurements necessary for achieving permissible reconstruction errors are evaluated for the two strategies. However, worst case evaluations, in general, do not necessarily well describe the performance actually observed in practical situations, and therefore, alternative investigations for probing the typical performance are also important. Having this perspective, we here analyze the typical performance of the thresholding 1-bit CS using statistical mechancis.}}
We will show that the fixing-value strategy statistically yields better mean squared error (MSE) performance than the random strategy when adjustable parameters are optimally tuned using the replica method \cite{replica} of statistical mechanics.
Unfortunately, the value of the optimal threshold depends on the statistical property of the target signal, which may not be known in advance. To cope with such situations, we focus here on the distribution of binary output y, which indirectly conveys the {amplitude information of the target signal $\vm{x^0}$} and can be estimated from measurements. We develop a heuristic that adaptively tunes the threshold parameter based on the frequency of positive (or negative) values in the binary outputs.
Numerical experiments show that our algorithm exhibits satisfactory {performance which is comparable to that} achieved by the optimally tuned threshold.
The rest of this paper is organized as follows: In Section II, we formulate the problem to be addressed in this study. In Section III, we evaluate the performance of the reconstruction method of (\ref{thresh1bitCS}). Section IV is devoted to a description of our learning algorithm to tune the threshold value, whereas Section V summarizes our work in this study.
\section{Problem set up}
Let us suppose a situation where entry $x_i^0$ $(i = 1, 2, \ldots, N)$ of $N$-dimensional signal $\vm{x}^0 \in \mathbb{R}^N$ is independently generated from an identical sparse distribution:
\begin{equation}
P\left(x\right) = \left(1 - \rho\right) \delta \left( x \right)
+ \rho\tilde{P} \left( x \right),
\label{sparse}
\end{equation}
where $\rho\in[0, 1]$ represents the density of nonzero entries in the signal, and $\tilde{P} (x)$ is a distribution function of $x \in \mathbb{R}$ that does not have finite mass at $x = 0$. In the thresholding 1-bit CS, the measurement is performed as
\begin{equation}
\textrm{\boldmath $y$} = \mathrm{sign} \left(\vm{\Phi x^{0}} + \vm{\lambda} \right),
\label{measurement}
\end{equation}
where we assume that each entry of the $M \times N$ measurement matrix $\vm{\Phi}$ is provided as an independent sample from a Gaussian distribution of mean zero and variance $N^{-1}$.
We consider two strategies for setting the thresholding vector $\vm{\lambda} = (\lambda_\mu)$. Case 1: entry $\lambda_\mu = \lambda$ is fixed for all $\mu = 1, 2, \ldots, M$. Case 2: $\lambda_\mu$ is independently sampled from a Gaussian distribution ${\cal N}(0, \sigma_{\lambda}^2)$. For both cases, the feasible set consistent with given outputs $\vm{y}$ is provided by a set of inequalities
\begin{equation}
y_\mu \left (\sum_{i = 1}^N \Phi_{\mu i} x_i + \lambda_\mu \right ) > 0
\label{feasible_region}
\end{equation}
$(\mu = 1, 2, \ldots, M)$, which defines a convex region of $\vm{x}$. Therefore, a sparse signal is reconstructed by the $l_1$-norm minimization (\ref{thresh1bitCS}) utilizing a certain convex optimization algorithm.
\section{Analysis}
\subsection{Method}
The key to finding the statistical properties of reconstruction (\ref{thresh1bitCS}) is the average free energy density
\begin{equation}
\bar{f} \equiv - \lim_{\beta, N \to \infty}\frac{1}{\beta N} \left [\ln Z(\beta; \vm{\Phi}, \vm{x}^0, \vm{\lambda}) \right]_{\vm{\Phi},\vm{x}^0, \vm{\lambda}},
\end{equation}
where
\begin{equation}
Z\left(\beta;\vm{\Phi},\vm{x^{0}},\vm{\lambda}\right)
\!=\!\int d \textrm{\boldmath $x$}e^{-\beta\vert\vert\textrm{\boldmath $x$}\vert\vert _{1}}
\prod_{\mu=1}^M \Theta\left ( (\vm{\Phi} \vm{x}^0 +\vm{\lambda})_\mu (\vm{\Phi} \vm{x}+\vm{\lambda})_\mu \right )
\label{eq:partition_func}
\end{equation}
is the partition function. {{We consider the large system size limit, $N\rightarrow\infty ,M\rightarrow\infty$, while keeping $\alpha=M/N$ finite.}} Here, $\Theta\left(x \right) = 1$ and $0$ for $x > 0$ and $x < 0$, respectively, offers the basis for our analysis.
$\left [\cdots \right]_{X}$ generally denotes the operation of the average with respect to the random variable $X$.
As $\beta$ tends to infinity, the integral of (\ref{eq:partition_func}) is dominated by the correct solution of (\ref{thresh1bitCS}), {which offers the minimum $l_1$-norm of $\vm{x}$}. One can therefore evaluate the performance of the solution by examining the macroscopic behavior of (\ref{eq:partition_func}) in the limit of $\beta\rightarrow \infty$. Because directly averaging the logarithm of the partition function is difficult, we employ the replica method \cite{replica}, which allows us to calculate the average free energy density as
\begin{equation}
\bar{f} = -\lim_{n \to +0}\frac{\partial}{\partial n}\lim_{\beta, N\to\infty}\frac{1}{\beta N} \ln \left [Z^{n}(\beta; \vm{\Phi}, \vm{x}^0,\vm{\lambda}) \right]_{\vm{\Phi}, \vm{x}^0,\vm{\lambda}}.
\end{equation}
For this, we first evaluate the $n$-th moment of the partition function $\left [Z^n \left(\beta; \vm{\Phi}, \textrm{\boldmath $x$}^{0},\vm{\lambda} \right) \right]_{\vm{\Phi}, \vm{x}^0, \vm{\lambda}}$ for $n = 1, 2, \ldots \in \mathbb{N}$ by using the formula
\begin{equation}
Z^n \left(\beta; \vm{\Phi}, \textrm{\boldmath $x$}^{0},\vm{\lambda} \right)
=
\int \prod_{a=1}^n \left (d \textrm{\boldmath $x$}^a
e^{-\beta\vert\vert\textrm{\boldmath $x^a$}\vert\vert _{1}} \right )
\times
\prod_{a=1}^n \prod_{\mu=1}^M \Theta\left ( (\vm{\Phi} \vm{x}^0 +\vm{\lambda})_\mu (\vm{\Phi} \vm{x}^a+\vm{\lambda})_\mu \right ),
\label{eq:expansion}
\end{equation}
which holds only for $n = 1, 2, \ldots \in \mathbb{N}$. Here, $\vm{x}^a$ ($a = 1, 2, \ldots, n$) denotes the $a$-th replicated signal. Averaging (\ref{eq:expansion}) with respect to $\vm{\Phi}$ and $\vm{x}^0$ results in the saddle point evaluation concerning macroscopic variables $q_{0a} = q_{a0}\equiv N^{-1} \vm{x}^0 \cdot \vm{x}^a$ and $q_{ab} = q_{ba} \equiv N^{-1} \vm{x}^a \cdot \vm{x}^b$ ($a, b = 1, 2, \ldots, n$). Although (\ref{eq:expansion}) holds only for $n \in \mathbb{N}$, the expression $(\beta N)^{-1} \ln \left [Z^n \left(\beta; \vm{\Phi}, \textrm{\boldmath $x$}^{0},\vm{\lambda} \right) \right]_{\vm{\Phi}, \vm{x}^0, \vm{\lambda}}$ obtained by the saddle point evaluation, under a certain assumption concerning the permutation symmetry with respect to the replica indices $a, b$, is obtained as an analytic function of $n$, which is likely to also hold for $n \in \mathbb{R}$. Therefore, we utilize the analytic function to evaluate the average of the logarithm of the partition function to obtain $\bar{f}$.
In particular, under the replica symmetric (RS) ansatz, where the dominant saddle point is assumed to be of the form
\begin{eqnarray}
q_{ab}=q_{ba}=\left \{
\begin{array}{ll}
Q_{0} & (a=b=0) \cr
m & (a=1,2,\ldots,n; \ b=0) \cr
Q & (a=b=1,2,\ldots,n) \cr
q & (a\ne b =1,2,\ldots,n)
\end{array}
\right . .
\label{RSanzats}
\end{eqnarray}
{{The problem setting so far is applicable generally for any distribution $\tilde{P}(x)$ in (\ref{sparse}).}} For simplicity, we hereafter assume that $x^0$ is distributed from (\ref{sparse}) with $\tilde{P}(x) = {\cal N} (0, \sigma_{0}^2)$; therefore $Q_{0} = \rho\sigma_{0}^{2}$.
\subsection{Resulting equations}
The above procedure (\ref{RSanzats}) offers an expression of the average free-energy density as
\begin{eqnarray}
\bar{f}
&=& \mathop{\rm extr}_{\omega} \Biggr\{\! \int\! {\rm D}z P(x^{0})\phi \left(\sqrt{\hat{q}}z+\hat{m} x^{0};\hat{Q}\right)
-\frac{1}{2}\hat{Q}q+\frac{1}{2}\hat{q}\chi+\hat{m}m {\sigma_{0}^2}\nonumber\\
&&+\frac{\alpha}{2\chi}\left[{H}\left (-\frac{\frac{m}{\sqrt{q}}t+\lambda}{\sqrt{\rho{\sigma_{0}^2}-\frac{m^2}{q}}} \right )\left(\sqrt{q}t+\lambda\right)^2\Theta\left(-\sqrt{q}t-\lambda\right)\right.\nonumber\\
&&\left.+{H}\left (\frac{\frac{m}{\sqrt{q}}t+\lambda}{\sqrt{\rho{\sigma_{0}^2}-\frac{m^2}{q}}} \right )\left(\sqrt{q}t+\lambda\right)^2\Theta\left(\sqrt{q}t\!+\!\lambda\!\right)\!\right]_{t,\lambda} \!\Biggr\}\!
\label{eq:free energy}
\end{eqnarray}
in the limit of $\beta \to \infty$. Here, $\alpha = M/N$, $\textrm{extr}_{X}\{g(X)\}$ denotes the extremization of function $g(X)$ with respect to $X$, $\omega = \{\chi, m, q, \hat{Q}, \hat{q}, \hat{m}\}$, ${H}(x) = \int_x^{+\infty} {\rm D}z$, $\textrm{D}z = \textrm{d}z \textrm{exp}(-z^2/2)/\sqrt{2\pi}$ is a Gaussian measure, $t$ and $z$ are independent and identically distributed (i.i.d) random variables from ${\cal N}(0, 1)$. The function $\phi(h; \hat{Q})$ is defined as
\begin{eqnarray}
\phi(h;\hat{Q})&=&\mathop{\rm min}_{x} \left \{
\frac{\hat{Q}}{2} x^2-h x + |x| \right \}
= -\frac{1}{2\hat{Q}} \left (|h|-1 \right )^2 \Theta\left ( |h| -1 \right ).
\end{eqnarray}
The derivation of $(\ref{eq:free energy})$ is provided in \ref{replicaderivation}.
For Case 1, which fixes the threshold for all measurements to a constant $\lambda$ as $\lambda_\mu = \lambda$ $(\mu = 1, 2, \ldots, M)$, the extremization of (\ref{eq:free energy}) is reduced to the following saddle point equations:
\begin{flushleft}
\begin{eqnarray}
\hat{q}&\!=\!&\frac{\alpha}{\chi^2}\left\{\left[ {H}\left(-\frac{\frac{mt}{\sqrt{q}}+\lambda}{\sqrt{\rho\sigma_{0}^2-\frac{m^2}{q}}}\right)u\left(-\sqrt{q}t-\lambda\right)\right.\right]_t \nonumber\\
&&\left.+\left[{H}\left(\frac{\frac{mt}{\sqrt{q}}+\lambda}{\sqrt{\rho\sigma_{0}^2-\frac{m^2}{q}}}\right)u\left(\sqrt{q}t+\lambda\right)\right]_{t}\right\},\label{qh_fix}\\
\hat{Q}&\!=\!&\frac{\alpha}{\chi} \left\{ \left[{H}\left(-\frac{\frac{mt}{\sqrt{q}}+\lambda}{\sqrt{\rho\sigma_{0}^2-\frac{m^2}{q}}}\right)u''\left(-\sqrt{q}t-\lambda\right)\right]_t\right.\nonumber\\
&&\left.+\left[{H}\left(\frac{\frac{mt}{\sqrt{q}}+\lambda}{\sqrt{\rho\sigma_{0}^2-\frac{m^2}{q}}}\right)u''\left(\sqrt{q}t+\lambda\right)\right]_t\right\},\label{Qh_fix}\\
\hat{m}&\!=\!&\frac{\alpha}{\chi\sqrt{2\pi\left(\rho\sigma_{0}^2-\frac{m^2}{q}\right)}}\left[\text{exp}\left(-\frac{\left(\frac{mt}{\sqrt{q}}+\lambda\right)^2}{2\left(\rho\sigma_{0}^2-\frac{m^2}{q}\right)}\right)\right.\nonumber\\
&&\left.\times\left(u'\left(\sqrt{q}t+\lambda\right)-u'\left(-\sqrt{q}t-\lambda\right)\right)\right]_t, \label{mh_fix}\\
q&\!=\!&\frac{2}{\hat{Q}^{2}}\left\{ \left(1-\rho\right)\left(\left(\hat{q}+1\right){H}\left(\frac{1}{\sqrt{\hat{q}}}\right)
-\sqrt{\frac{\hat{q}}{2\pi}}e^{-\frac{1}{2\hat{q}}} \right) \right . \nonumber\\
&&\left. +\rho\left(\left(\hat{q}+\hat{m}^2\sigma_{0}^2+1\right){H}\left(\frac{1}{\sqrt{\hat{q}+\hat{m}^2\sigma_{0}^2}}\right)\right.\right.\nonumber\\
&&\left.\left.-\sqrt{\frac{\hat{q}+\hat{m}^2\sigma_{0}^2}{2\pi}}e^{-\frac{1}{2\left(\hat{q}+\hat{m}^2\sigma_{0}^2\right)}}\right)\right\}, \label{q_fix}\\
\chi &\!=\!&\frac{2}{\hat{Q}}\left\{\left(1-\rho\right)
{H}\left(\frac{1}{\sqrt{\hat{q}}}\right)\!+\!\rho {H}\!\left(\!\frac{1}{\sqrt{\hat{q}+\hat{m}^2\sigma_{0}^2}}\!\right)\!\right\}\!,\! \label{chi_fix}\\
m&\!=\!&\frac{2\rho\hat{m}\sigma_{0}^2}{\hat{Q}}{H}\left(\frac{1}{\sqrt{\hat{q}+\hat{m}^2\sigma_{0}^2}}\right)\label{m_fix},
\end{eqnarray}
\end{flushleft}
where $u(x) = x^2\Theta(x)$, and $t$ obeys the standard normal distribution ${\cal N}(0, 1)$.
On the other hand, for Case 2, where $\lambda_\mu$ is sampled independently from ${\cal N}(0, \sigma^2_\lambda)$ for $\mu = 1, 2, \ldots, M$, the saddle point equations of $\hat{q}$, $\hat{Q}$, and $\hat{m}$ are modified to
\begin{eqnarray}
\hat{q}&\!=\!&\frac{2\alpha}{\chi^2}\left[{H}\left(\frac{\frac{mt}{\sqrt{q}}+\sigma_{\lambda}r}{\sqrt{\rho\sigma_{0}^2-\frac{m^2}{q}}}\right)u\left(\sqrt{q}t+\sigma_{\lambda}r\right)\right]_{r,t},\label{qh_Gauss}\\
\hat{Q}&\!=\!&\frac{2\alpha}{\chi}\left[{H}\left(\frac{\frac{mt}{\sqrt{q}}+\sigma_{\lambda}r}{\sqrt{\rho\sigma_{0}^2-\frac{m^2}{q}}}\right)u''\left(\sqrt{q}t+\sigma_{\lambda}r\right)\right]_{r,t},\label{Qh_Gauss}\\
\hat{m}&\!=\!&\frac{2\alpha}{\chi\sqrt{2\pi\left(\rho\sigma_{0}^2-\frac{m^2}{q}\right)}}\left[\text{exp}\left(-\frac{\left(\frac{mt}{\sqrt{q}}+\sigma_{\lambda}r\right)^2}{2\left(\rho\sigma_{0}^2-\frac{m^2}{q}\right)}\right)\right.\nonumber\\
&&\times\left.u'\left(\sqrt{q}t+\sigma_{\lambda}r\right)\right]_{r,t},\label{mh_Gauss}
\end{eqnarray}
where $r$ is a variable sampled from the standard normal distribution ${\cal N}(0, 1)$. The remaining equations for $q$, $\chi$, and $m$ are identical to (\ref{q_fix}), (\ref{chi_fix}), and (\ref{m_fix}), respectively.
\subsection{Simulations and observations}
The value of $m$ determined by these equations physically represents the typical overlap $N^{-1} \left [\vm{x}^0 \cdot \hat{\vm{x}} \right]_{\vm{\Phi}, \vm{x}^0, \vm{\lambda}} $ between the original signal $\vm{x}^0$ and the solution $\hat{\vm{x}}$ of (\ref{thresh1bitCS}). Therefore, the typical value of MSE between $\vm{x}^0$ and $\hat{\vm{x}}$, which serves as the performance measure of the reconstruction problem, is evaluated as
\begin{eqnarray}
{\rm MSE} = N^{-1}\left [ \left |\hat{\vm{x}}-\vm{x}^0\right |^2 \right ]_{\vm{\Phi,\vm{x}^0},\vm{\lambda}}
=q+ \rho\sigma_{0}^2-2m.
\label{MSE}
\end{eqnarray}
Note that in past studies on 1-bit CS, reconstruction performance was evaluated through directional MSE, which is defined by $|\frac{\hat{\vm{x}}}{|\hat{\vm{x}}|} - \frac{\vm{x}^0}{|\vm{x}^0|}|^2$ as scale information is lost.
\begin{figure}[h]
\begin{center}
\includegraphics[width=3.5in]{fix}
\caption{Replica prediction of MSE (in decibel) versus fixed threshold $\lambda$ for signal distribution $\rho = 0.25, \sigma_{0}^{2} = 1$, and ratio $\alpha = 3$.}
\label{fig:MSEdb_fix}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=3.5in]{Gauss}
\caption{Replica prediction of MSE (in decibel) versus $\sigma_{\lambda}$ for signal $\rho = 0.25, \sigma_{0}^{2} = 1$, and ratio $\alpha = 3$.}
\label{fig:MSEdb_Gauss}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=3.5in]{fixedVSGaussian}
\caption{Lowest MSE [dB] (envelop) for each ratio $\alpha$ of signal $\rho = 0.25$, $\sigma_0^2 = 1$. The blue and red curves represent threshold strategies 1 and 2, respectively. The circles stand for the experimental estimate obtained using the CVX algorithm \cite{cvx} averaged over $1,000$ experiments with signal size $N = 128$ for each parameter set. }
\label{fig:fixedVSGaussian}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=3.5in]{MSE_Q0_db_Py}
\caption{Optimal MSE: $\text{MSE}/(\rho\sigma_{0}^{2})$ in decibel versus the probability of $+1$ in $\vm{y}$ for fixed threshold 1-bit CS model. {{The plots are obtained by tuning $\lambda$ so as to minimize MSE and evaluating corresponding $P(y=1)$ from (\ref{eq:Py}) for each set of ($\rho, \sigma_{0}^{2}, \alpha$).}} Different colors represent varying signal sparsity; {red, blue, green and black mean $\rho=0.25, 0.125, 0.0625$ and $0.03125$, respectively. In each color, there are 6 circle and 6 cross symbols. Each of 6 symbols from left to right, corresponds to the result of $\alpha = 1, 2, \dots, 6$, respectively. Circles and crosses represent the results for $\rho\sigma_{0}^{2}=0.25$ and $0.125$, respectively, where $\rho\sigma_{0}^{2}$ means the ``power''(per component) of the original signal.}}
\label{fig:normalizedMSE_Py}
\end{center}
\end{figure}
We solved the saddle point equations for signal sparsity $\rho = 0.25$ and variance $\sigma_{0}^{2} = 1$ when ratio $\alpha = 3$. The curve in Fig. \ref{fig:MSEdb_fix} denotes the theoretical predictions of MSE as evaluated by (\ref{qh_fix})--(\ref{m_fix}) (strategy 1) and (\ref{MSE}) plotted against the threshold $\lambda$. Fig. \ref{fig:MSEdb_Gauss} represents the theoretical predictions of MSE evaluated by (\ref{qh_Gauss})--(\ref{mh_Gauss}), (\ref{q_fix})--(\ref{m_fix}) (strategy 2), and (\ref{MSE}) plotted against the standard deviation $\sigma_{\lambda}$ of the threshold. Figures \ref{fig:MSEdb_fix} and \ref{fig:MSEdb_Gauss} show that there is an optimal threshold distribution (red circle symbol) that minimizes MSE for each set of parameters. Similar features hold for various sets of values of $\alpha, \rho, \sigma_{0}^{2}$ for both strategies 1 and 2.
To compare the optimal MSE (changing threshold distribution) of strategy 1 and strategy 2, we plot the optimal MSE for the same signal distribution in Fig. \ref{fig:MSEdb_fix} and Fig. \ref{fig:MSEdb_Gauss} against $\alpha$ in Fig. \ref{fig:fixedVSGaussian}, which is referred to the envelope curve of MSE. The blue and red curves represent the envelope curves for strategies 1 and 2, respectively. From Fig. \ref{fig:fixedVSGaussian}, we can see that strategy 1 outperforms strategy 2 when parameters are optimally tuned. Therefore, we hereafter focus on strategy 1, for which the thresholds are fixed.
The optimal value of $\lambda$ depends on $\rho$ and $\sigma_0^2$, which are not necessarily available in practice. To cope with such situations, we focus here on the distribution of binary output y, which indirectly conveys the information of $\rho\sigma_0^2$ and can be estimated from measurements. Fig. \ref{fig:normalizedMSE_Py} shows the relation between the optimal MSE and $P(y = +1)$ for eight signal distributions. For given $\lambda$, the probability of positive output $y = +1$ is evaluated as
\begin{eqnarray}
P(y=+1)&=&\frac{1}{M}\left[\prod_{\mu=1}^{M}\Theta\left(\vm{\Phi x_0}+\vm{\lambda}\right)_{\mu} \right]_{\vm{\Phi}, \vm{x_0}}\cr
&=&\left[\Theta\left(\vm{\Phi x_0}+\vm{\lambda}\right)_{\mu} \right]_{\vm{\Phi}, \vm{x_0}}\cr
&=&\int_{-\infty}^{\infty} \textrm{D}t \Theta\left(\sqrt{\rho}\sigma_{0}t+\lambda\right)\cr
&=&{H}\left(-\frac{\lambda}{\sqrt{\rho}\sigma_0}\right).
\label{eq:Py}
\end{eqnarray}
The horizontal axis in Fig. \ref{fig:normalizedMSE_Py} is calculated from (\ref{eq:Py}) by inserting the optimal value of $\lambda$.
{{{The results indicate that when the signal is sparser (from red to black), corresponding $P(y = +1)$ is greater. Also, the value of $P(y = +1)$ that yields the optimal MSE monotonically increases with $\alpha$ when the signal distribution is fixed.}}
MSE is normalized by $\rho\sigma_0^2$ in order to eliminate its dependence on the scale of the original signal.
From these results, we can see that the normalized MSE, $\text{MSE}/(\rho\sigma_{0}^{2})$, is the same when signal sparsity is the same.
Although the optimal MSE depends on all system parameters $\rho$, $\sigma_0^2$, and compression rate $\alpha$, we can see that the corresponding $P(y = +1)$ is always placed in the range of $0.75 \sim 0.85$ for modest values of $1 \le \alpha \le 6$ in Fig. \ref{fig:normalizedMSE_Py}. In addition, the plots imply that although the optimal value of $P(y = +1)$ monotonically increases as $\alpha$ grows, it tends to converge to a value close to $0.85$.
\section{Learning algorithm for threshold}
The results of the last section suggest that for each parameter set, the optimal threshold that minimizes MSE is loosely characterized by the value of $P(y = +1)$, which can be statistically estimated from the outputs of measurements. This property may be utilized to adaptively tune the threshold for each measurement based on the results of previous measurements.
A few studies have been conducted in the past on adaptive tuning of the threshold to improve signal reconstruction performance. For example, in \cite{adaptiveBaye}, given past measurements, a threshold value was determined to partition the consistent region along its centroid computed by generalized approximate message passing \cite{GAMP, 1bitCSBayeXuKaba}. However, in many realistic situations, precise knowledge of the prior distribution is unavailable, even if we might reasonably expect the signal to be sparse. Therefore, we will here develop a learning algorithm that can be executed without knowledge of the prior distribution of the signal. There is another general adaptive algorithm called $\Sigma\Delta$ quantization \cite{sigmadelta}. However, its goal is to find a satisfactory quantized representation of real number measurement and requires preprocessing based on real number measurements. Instead, the algorithm we develop aims to directly minimize MSE, and needs no preprocessing.
\begin{figure}
\renewcommand{\thepseudocode}{\arabic{pseudocode}}
\setcounter{pseudocode}{0}
\begin{pseudocode}[ruled]{adaptive thresholding}{\gamma, \delta, \lambda_{0}, U_0, V_0}
1)\ \mbox{\bf Initialization}:\\
\hspace{15pt}\lambda \text{ seed}: \hspace{65pt}\lambda_0 \\
\hspace{15pt}U \text{seed}: \hspace{65pt} U_{0}\GETS0 \\
\hspace{15pt}V \text{seed}: \hspace{65pt} V_{0}\GETS0\\
\hspace{15pt}\text{Counter}: \hspace{62pt} k\GETS 0\\
2)\ \mbox{\bf Counter increase}:\\
\hspace{20pt}k \GETS k+1\\
3)\ \mbox{\bf Measurement of signal}:\\
\hspace{20pt}
y_{k} = \textrm{sign}\left(\sum_{i} \vm{\Phi}_{k i} \vm{x_0}_{i} + \lambda_{k} \right)\\
4)\ \mbox{\bf Update $T_k$}:\\
\hspace{20pt}
U_k \GETS (y_k > 0) + \gamma U_{k - 1}\\
\hspace{20pt}
V_k \GETS 1 +\gamma V_{k - 1}\\
\hspace{20pt}
T_k \GETS U_{k}/V_{k}\\
5)\ \mbox{\bf Update $\lambda$}:\\
\hspace{20pt}
\lambda_k \GETS \lambda_{k - 1} + \delta\textrm{sign}(T - T_k)\\
6)\ \bold{Iteration}: \mbox{Repeat from 2) until $k = M$.}
\end{pseudocode}
\protect
\caption{\protect\label{algorithm}Pseudocode for adaptive thresholding of 1-bit CS measurements. Here, $y_k$ and $\lambda_k$ for $k = 1, 2, ..., M$ represent each element of vector $\vm{y}$ and $\vm{\lambda}$, respectively. Signal reconstruction can be carried out by versatile convex optimization algorithms.
}
\label{algorithm}
\end{figure}
As shown in Fig. \ref{fig:normalizedMSE_Py}, MSE is minimized when $P(y = +1)$ takes a value of $0.75 \sim 0.85$ for various sets of parameters. To incorporate this property, we propose a strategy that first fixes a target value of $T$ for $P(y = +1)$, and tunes $\lambda$ so that an empirical distribution of $P(y = +1)$ approaches $T$. As we see in Fig. \ref{fig:normalizedMSE_Py}, for larger values of $\alpha$ or sparser signals, we should set $T$ as greater in the relevant range. There are various ways of estimating $P(y = +1)$ from the results of measurements. Of these, we use the damped average
\begin{equation}
T_\mu = \frac{\sum_{n = 0}^{\mu - 1}\gamma^{n}\delta_{y_{\mu -n}, +1}}{\sum_{n = 0}^{k - 1}\gamma^{n}},
\end{equation}
since it can be computed in an online manner as
\begin{equation}
T_{\mu +1} = \frac{\gamma \left (1 - \gamma^\mu \right )T_\mu + (1 - \gamma) \delta_{y_{\mu + 1}, +1}}{1 - \gamma^{\mu + 1}},
\end{equation}
which does not require referring to the details of previous measurements. Here, the damping factor $\gamma$ is a parameter that we have to set. In experiments, we set $\gamma = 0.8$; but as long as we tested it, the obtained performance was not particularly sensitive to the choice of this parameter. (\ref{measurement}) indicates that $P(y = +1)$ monotonically increases as $\lambda_\mu$ grows. This implies that $\lambda_\mu$ should be increased when $T > T_{\mu - 1}$, and decreased otherwise. To implement this idea, we design the learning algorithm of $\lambda_\mu$ as
\begin{equation}
\lambda_\mu = \lambda_{\mu - 1} + \delta \textrm{sign}(T - T_{\mu - 1}),
\end{equation}
where $\delta$ denotes the step size that is also set by users. The pseudocode for adaptive thresholding 1-bit CS measurements is shown in Fig. \ref{algorithm}. Following measurement, signal reconstruction can be carried out by versatile convex optimization algorithms \cite{cvx} by solving (\ref{thresh1bitCS}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=3.5in]{replica_cvx}
\caption{Experimental result from the adaptive thresholding algorithm for signal $\rho = 0.0625$, $\sigma_{0}^2 = 2$, and $N = 128$. The circles denote the average of $1,000$ experiments. The parameter settings were $T = 0.8$, $\gamma = 0.8$, $\lambda_0 = 0.5$, and $\delta = 0.01$. The broken line represents the replica prediction when $\lambda$ is set to offer $P(y = +1) = 0.8$ while the full curve denotes this for optimally tuned $\lambda$.}
\label{fig:cvxresult}
\end{center}
\end{figure}
Since we plan to apply the adaptive algorithm in situations involving a finite number of measurements, the extent to which the initial threshold $\lambda_0$ is remote from the optimal threshold $\lambda_{\text{opt}}$, which is unknown beforehand, and the variation in the step size $\delta$ may significantly influence reconstruction performance. In order to set an appropriate value of $\lambda_0$, we propose testing it by measuring the signal a few times. If the outputs are limited almost exclusively to $+1$ or $-1$, we change the threshold through the bisection method, which involves dividing or multiplying it by $2$ until the outputs are adequately mixed with $+1$ and $-1$. The resulting threshold should yield an appropriate value of $\lambda_0$ close to $\lambda_{\text{opt}}$. Having set $\lambda_0$, an appropriate value of $\delta$ should be in smaller order in order to tweak it to $\lambda_{\text{opt}}$.
The results of our numerical experiments are shown in Fig. \ref{fig:cvxresult} as circles. Each circle denotes the average of $1,000$ experiments for systems where $N = 128$. The parameter settings of the experiments were $T = 0.8$, $\gamma = 0.8$, $\lambda_0 = 0.5$, and $\delta = 0.01$ for signal distribution $\rho = 0.0625$ and $\sigma_{0}^2 = 2$. The solid line in Fig. \ref{fig:cvxresult} represents the envelop for MSE (dB) for each $\alpha$. On the other hand, the dashed curve represents the prediction of MSE (dB) using replica analysis when $P(y = +1) = 0.8$, which was achieved by $\lambda = 0.2976$ according to (\ref{eq:Py}). Fig.\ref{fig:cvxresult} shows that the adaptive thresholding algorithm in conjunction with the employment of CVX for signal reconstruction can achieve nearly the same performance in terms of MSE as the statistical prediction for $P(y = +1) = 0.8$, and the result is reasonably close to the envelope MSE.
\section{Conclusion}
{{In this paper, we analyzed the typical performance of the thresholding 1-bit compressed sensing, which can reconstruct both the scaling and the directional information of the signal.}}
Considering the most general situation, where no detailed prior knowledge of sparse signals is available, we employed the $l_1$-norm minimization approach. By utilizing the replica method from statistical mechanics, the mean squared error behavior of reconstruction for standard i.i.d measurement matrix and i.i.d Bernoulli-Gaussian signal was derived in the large system size limit. We compared two design strategies for the elements of the threshold vector, which corresponded to setting a fixed or random value as threshold. Our analysis showed that the fixed threshold strategy can achieve lower MSE than the random threshold strategy statistically.
Another observation from the replica results was that there is an optimal threshold that minimizes MSE for a set of signal distributions and measurement ratios. However, in order to evaluate the optimal threshold, we need to know the prior distribution of the signal, which is not necessarily available in practical situations. Therefore, we shifted our focus to the relation between the optimal threshold and the distribution of the binary outputs, which can be empirically evaluated from signal measurements. The replica analysis indicated that the MSE is minimized when $P(y = +1)$ is set in the vicinity of $0.75 \sim 0.85$ for a wide region of system parameters.
On the basis of this observation, an algorithm that adaptively tunes the threshold at each measurement in order to obtain $P(y = +1)$ close to our target value was proposed. Combined with versatile convex optimization algorithms, the adaptive thresholding algorithm offers a computationally feasible and widely applicable 1-bit CS scheme. Numerical experiments showed that it can yield nearly optimal performance, even when no detailed prior knowledge of sparse signals is available.
Improvements on the adaptive thresholding algorithm as well as the application of the algorithm to practical problems form part of our future research in the area.
\ack
YX is supported by JSPS Research Fellowships DC2. This study was partially supported by JSPS KAKENHI Nos. 26011287 (YX) and 25120013 (YK).
|
\section{Introduction}
The solar photosphere is characterized by a wide variety of magnetic structures of different spatial size, complexity and magnetic field strengths, ranging from sunspots, pores, and network to small internetwork bright points. The morphology of sunspot umbrae and pores is very similar what produces that the second one is sometimes called ``naked umbrae'' \citep{Solanki2003}. The reason behind is that pores, as the inner part of sunspots, are detected in white light observations as dark regions where the photospheric convection and plasma motions have been inhibited due to the presence of a strong magnetic field \citep[see, for instance,][]{Biermann1941,Cowling1953}.
Solar pores present themselves with diameters of a few Mm, much smaller than the usual size of a sunspot (up to 40 Mm), a field strength in the order of 2 kG, and a lifetime of typically less than a day. Moreover, the absence of filamentary penumbra surrounding pores suggests a simpler magnetic configuration, that can be represented by a magnetostatic flux tube with a predominantly vertical magnetic field \citep{Simon1970}. However, high spatial-resolution observations indicate that some pores contain a wide variety of fine bright features, such as bright dots or light bridges that may be signs of a convective energy transportation mechanism \citep[][]{Sobotka1999,Keil1999,Hirzberger2003,Giordano2008,Sobotka2013}. For additional information about the general properties of solar pores and sunspots, see the reviews of \cite{Sobotka2003}, \cite{Solanki2003}, \cite{Thomas2004} or a more modern one as \cite{Borrero2011}.
Recent studies like the one performed by \cite{Sobotka2012} examined the magnetic properties of pores in detail, performing inversions of the full Stokes vector in order to infer the atmospheric physical parameters. They observed downflows at the borders of the structure, in agreement with \cite{Keil1999}, \cite{Hirzberger2003}, \cite{Morinaga2008}, and \cite{Scharmer2011}, traces of the magnetic field with the shape of spines extending more than $3.5$ arcsec from the white light pore borders, similar to the results of \cite{Cameron2007}, and a linear correlation between the temperature and the vertical component of the magnetic field, with lower temperatures in the regions of stronger magnetic fields.
Consequently, the work of \cite{Sobotka2012} provides a big step in the understanding of the nature of solar pores although there are some limitations in their analysis. We aim to surpass these limitations in the present work aiming to complement their findings. One of the limitations is the configuration used in the inversion of the Stokes profiles, with constant stratifications with height for the atmospheric parameters, except the temperature that was perturbed as a linear function of the continuum optical depth. The authors explained that the reason behind this choice is that they only observed a single line and they wanted to reduce the number of free parameters as much as possible. In our case, we do not have this limitation as we are going to analyse two spectral lines with different heights of formation observed with high spectral sampling. Thus, we can infer the atmospheric parameters in a wider range of heights with good accuracy. The second limitation in their work is the uncorrected contamination of the stray light. This stray light produces a lack of contrast in continuum images, and also affects to the size of the structures, and the amplitude of the magnetic Stokes profiles \citep{RuizCobo2013,QuinteroNoda2015,QuinteroNoda2016}. We solved the latter limitation applying a spatial deconvolution technique \citep{QuinteroNoda2015} that removes the stray light contamination induced by the spatial point spread function of the telescope. However, there is also a limitation from our side respect to the work of \cite{Sobotka2012}, and is the lack of high cadence observations because we use a slit spectrograph and, thus, we cannot track the evolution of the solar pore and the surrounding structures as they did in their work. Moreover, taking advantage of local correlation tracking techniques, these authors could infer the horizontal velocities in the surrounding area of the solar pore.
Finally, our aim, similar to \cite{Sobotka2012}, is to provide a comprehensive model of solar pores that help us to understand these structures and also provides a better input for numerical simulations. In addition, we also want to evaluate the possibilities and limitations of the spatial deconvolution method applied to this region because the Stokes profiles are affected by large horizontal intensity variations in small spatial scales, e.g., Stokes $I$ intensity goes from 0.5~$I_c$ (with $I_c$ the continuum signal) at the center of the pore to 1.1~$I_c$ at the neighbouring granulation in less than 3 arcsec, and this large variation could induce artefacts in the deconvolved data.
\section{Data analysis}
We analysed in this work the magnetic properties of the active region NOAA 10949. It was detected for the first time by SOHO/MDI \citep{Scherrer1995} on 2007 March 28th as an individual dark spot in the continuum intensity images. The active region crossed the solar prime meridian on 2007 April 2nd, and lasted several days more, until 2007 April 4th. It was composed of a single solar pore during its lifetime and it showed low coronal activity. In addition, it was located relatively close to a filament although we cannot confirm if there is a connection between both structures due to the lack of data about the filament. We studied this solar pore using observations taken by the Spectropolarimeter ($SP$) \citep{Lites2013} on board $Hinode/SOT$ \citep{Kosugi2007,Tsuneta2008,Suematsu2008,Shimizu2008SOT} on 2007 April 1st between 18:56-19:57 UT. At this time, the active region was located at $(-150,220)$ arcsec, i.e. $\mu=0.96$. $Hinode/SP$ recorded the Stokes $I$, $Q$, $U$, and $V$ profiles for the Fe~{\sc i} 6301.5 and 6302.5 \AA \ spectral lines. It used a spectral sampling of 21.55 m\AA, and a pixel size and scanning step of about 0.16 arcsec providing a spatial resolution of 0.32 arcsec. The integration time was 4.8 s per slit position what produces a noise value of approximately $1\times10^{-3}$ of $I_c$ for the magnetic Stokes parameters \citep[for more information, see][]{Ichimoto2008,Lites2013SP_Prep}.
\begin{figure}
\hspace{-0.2cm}
\includegraphics[width=9.0cm]{cont.eps}
\vspace{-0.4cm}
\caption{Continuum signal from the observation of 2007 April~1st 18:56 UT. Red square depicts the region we examine in detail in this work.}
\label{cont}
\end{figure}
\begin{figure*}
\hspace{-0.3cm}
\includegraphics[width=18.0cm]{general2.eps}
\vspace{-0.5cm}
\caption{Comparison between original (upper row) and deconvolved (bottom row) observations. Different columns, from left to right, show the continuum intensity, line core intensity, and a Fe~{\sc i} 6302 \AA \ magnetogram.}
\label{compar}
\end{figure*}
Figure \ref{cont} shows the continuum signal of the observed map examined in this work. We analysed a small fragment, marked in red, of the total observed map because we aim to focus only on the central pore. This active region is relatively small, with a size of around 10 arcsec, i.e. around 7 Mm, and it is composed of several small dots joined by brighter structures that resemble the granulation. It seems that the magnetic activity is not high enough to produce a large spot. In addition, after examining context Ca~{\sc ii} images taking by $Hinode/BFI$ we did not find any activity at upper heights either, being quiet during more than the three hours of observation examined.
\subsection{Deconvolution method}
We have performed the spatial deconvolution of the observed data using the same method presented in \cite{QuinteroNoda2015}. A detailed description of the procedure used in this work can be found in the mentioned paper. The number of principal components of eigenvectors we used to reconstruct the Stokes profiles is (8, 5, 4, 8) for ($I$, $Q$, $U$, $V$). This number of eigenvectors is similar to the one used on \cite{RuizCobo2013}, and \cite{QuinteroNoda2015} and smaller than the one used on \cite{QuinteroNoda2016}. This is because the examined region is close to disk center and the variety of Stokes profiles is smaller than the one found in the latter work. In addition, the predominately vertical nature of the magnetic field inside the solar pore demands just a few eigenvectors to reproduce the linear polarization profiles.
We performed 25 iteration steps in the deconvolution process. The original continuum contrast was 6.26 per cent and the value obtained after the deconvolution process is 9.73 per cent. We stopped the iteration process in this step because it provides an increase factor of the continuum contrast similar to the one obtained in previous works \citep{QuinteroNoda2015}.
\begin{figure*}
\hspace{+0.2cm}
\includegraphics[width=16.0cm]{as.eps}
\caption{Comparison between asymmetry maps from original (upper row) and deconvolved (bottom row) data. Left column displays the area asymmetry map while the amplitude asymmetry is depicted in the right column. White indicates null asymmetry while red and blue signifies positive and negative asymmetry, respectively. Symbols in top-left panel indicate the position of the profiles plotted on Figure \ref{per}.}
\label{as}
\end{figure*}
\subsection{Deconvolution results}
Figure \ref{compar} shows the comparison between the original (upper row) and deconvolved (bottom row) data. The field of view displayed in each panel corresponds to the region enclosed by the red square in Figure \ref{cont}. We chose three different quantities to perform the comparison study: the continuum signal (leftmost column), line core intensity (middle column), and a magnetogram (rightmost column) built as the difference between Stokes $V$ maps taken at $\pm$100~m\AA \ from the line centre of Fe~{\sc i} 6302.5~\AA. As found in \cite{QuinteroNoda2015}, the continuum image looks sharper due to the increase in the continuum contrast. In addition, the abundance of intergranular bright points in the surroundings of the solar pore is more evident in the deconvolved map.
If we examine the line core intensity, middle column, we are roughly measuring the temperature at upper layers, a few hundred kilometres above the continuum region. We can see that the pore is still dark at this height and surrounding it there is a web of bright structures that indicates higher temperatures than those of the surrounding quiet Sun granulation and correspond to the bright intergranular points seen in the continuum image. In addition, the magnetic structures located south of the pore forming a plage region also appear as hot structures at these layers. Comparing the original and deconvolved line core intensity panels we find more defined structures in the latter one and also with higher intensity values.
Finally, magnetogram panels (right column) show a single polarity structure that corresponds to the solar pore, and a surrounding plage with the same polarity. In this case, the deconvolved panel shows narrower structures with larger amplitude values. Remarkably, it is easier to see the separation between the magnetic substructures that formed the pore because the blurred effect present in the original data has disappeared.
\begin{figure*}
\centering
\includegraphics[width=17.3cm]{per.eps}
\vspace{-0.3cm}
\caption{Stokes profiles from different regions in the observed map. Upper row shows a pixel from the central part of the pore, middle row from its edge, and bottom row displays a pixel from the solar plage. Black depicts the deconvolved profiles while red shows the results of the inversion. Symbols inside the panels of left column indicate their position in the observed map, see top-left panel of Figure~\ref{as}.}
\label{per}
\end{figure*}
\subsection{Analysis of Stokes profiles asymmetries}\label{asy}
As we mentioned before, a solar pore seems to be a simple structure composed of a vertical magnetostatic field. However, we want to analyze the amplitude and area asymmetries inside and at the edges of the pore to check if there is any type of pattern. In fact, we expect a change in the inclination of the magnetic field vector at these regions that could generate gradients along the line of sight, therefore asymmetries \citep{Illing1975}. We computed the asymmetries of the Stokes $V$ profile following the definition used in \cite{MartinezPillet1997}. The area asymmetry is obtained as
\begin{equation}
\delta A=s\frac{\sum_{i} V(\lambda_i)}{\sum_{i}\mid V(\lambda_i) \mid} ,
\end{equation}
where the sum is extended along the wavelength axis and $s$ is the sign of the Stokes $V$ blue lobe (chosen as $+1$ if the blue lobe is positive and $-1$ if the blue lobe is negative). The selected range of integration of the Stokes $V$ signals goes from $-0.43$ \AA\ to 0.43 \AA\ around the Fe~{\sc i} 6302.5 \AA\ line centre. Likewise, the amplitude asymmetry is defined as
\begin{equation}
\delta a=\frac{a_b-a_r}{a_b+a_r} ,
\end{equation}
where $a_b$ and $a_r$ are the unsigned maximum value of the blue and red lobe of Stokes $V$.
We computed these quantities over all the field of view enclosed by the red square in Figure \ref{cont} and we only considered pixels that show maximum Stokes $V$ amplitudes higher than $1\times10^{-2}$ of $I_{c}$. The results of this study are included in Figure \ref{as}. Top left panel shows that the area asymmetry is negative in the inner parts of the pore, dark in the continuum map, and with opposite polarity, positive, in the surrounding regions. The reason behind these asymmetry changes between the central part of the structure and its edges could be a change in the line of sight component of the magnetic field being less inclined, i.e. almost vertical, in the core of the pore and more inclined in surrounding areas. Another explanation could be a change in the line of sight velocities that are usually small in the center of the structure and show large downflows at its edges \citep[for instance,][]{Morinaga2007,Sobotka2012}. We will delve into this aspect in following sections where we infer the atmospheric information performing inversions of the deconvolved Stokes profiles.
Contrary to the area asymmetries, amplitude asymmetries (right column) are always positive in the surroundings of the structure and close to zero in the inner part of the solar pore. In this case, some regions with negative amplitude asymmetry can be found in the deconvolved map although they are relatively scarce and with low amplitude. In that sense, we can confirm, as in \cite{QuinteroNoda2015}, that the deconvolution process barely modifies the Stokes $V$ profiles asymmetries.
\section{Inversion of Stokes profiles}
We obtain the physical information of the atmospheric parameters inverting the deconvolved Stokes profiles from the observed region enclosed by the red square of Figure \ref{cont}. We carried out the inversion of the Stokes profiles using {\sc sir} \citep[Stokes Inversion based on Response functions;][]{RuizCobo1992} code, which allows us to infer the optical depth dependence of the atmospheric parameters at each pixel independently.
\subsection{Configuration}
We are going to briefly summarize the configuration we used in this paper because is similar to the one used by \cite{QuinteroNoda2015,QuinteroNoda2016}. As the deconvolution process eliminates the stray light contribution of the spatial PSF, we do not need to include a stray light component in the inversion process. Consequently, we only used a single component for reproducing the observed profiles. The selection of nodes for each quantity is done using an automatic algorithm, based on the number of zeros of the response function of the corresponding quantity
(see Quintero Noda et al. 2016b, \textit{under revision})
In our case, to avoid too complex solutions, we limited the maximum number of nodes. We allowed five for temperature T($\tau$) (where $\tau$ refers to the optical depth evaluated at 5000 \AA), three for the line of sight (LOS) component of the velocity v$_{los}$($\tau$), three for the magnetic intensity B($\tau$), two for the inclination of the magnetic field $\gamma$($\tau$), one for the azimuthal angle of the magnetic field $\phi$($\tau$), and one for the microturbulence. The algorithm determines the optimum number of nodes for each atmospheric parameter at each iterative step, where this optimum number of nodes is always lower or equal to the maximum number of nodes.
On the other hand, macroturbulence is null and not inverted. At each iteration, the synthetic profiles are convolved with the spectral transmission profile of $Hinode/SP$ \citep{Lites2013}. Additionally, given that the inferred physical parameters could be reliant on the initial atmosphere, we minimize this effect by inverting each individual pixel with 10 different initial atmospheric models. These initial random atmospheric models were created as in \cite{QuinteroNoda2015}. Therefore, since each node corresponds to a free parameter during the inversion, our model could include up to 15 free parameters in case the automatic algorithm selects the maximum number of nodes allowed for the atmospheric parameters.
Finally, the inversion code provides the uncertainties of the retrieved atmospheric parameters based on the response function to a given physical quantity. In that sense, if the response function is small, i.e. the Stokes profiles are not sensitive to changes in a given atmospheric parameter, the uncertainty of the solution for this parameter will be large. In addition, although the response functions depend on the atmospheric model itself \citep{Landi1977}, the height of formation of these lines only cover a small range of heights that usually goes from $\log \tau=0$ to $-2.5$. Therefore, we should restrict the analysis of the atmospheric parameters inside these general limits.
\subsection{Selected Stokes profiles}
We aim to examine the accuracy of the inversion at some spatial locations where the magnetic field configuration is complex. We chose a pixel that belongs to the inner part of the pore, a second one located at its edge, and a third pixel that belongs to the plage region that appears at the bottom part of the field of view (see Figure \ref{compar}). Deconvolved and fitted profiles are shown in Figure \ref{per}. If we start with the central part of the pore, upper row, we find that the fits are very accurate for Stokes $I$ and $V$, and relatively good for Stokes $Q$ and $U$. This is a tendency that is present for the rest of selected regions. We believe that this tendency produces fairly good results because the Stokes $V$ profiles, well fitted for almost all cases, display signals between 5-10 times larger than Stokes $Q$ and $U$ signals. Additionally, it is worth to mention the quality of the fitting for all Stokes $V$ profiles even when they show so dissimilar shapes, being with large widths and low amplitudes in the pore (upper row), with large area and amplitude asymmetries at its edges (middle row), or with extremely large amplitudes and moderate spectral redshifts as in the case of the solar plage (bottom row). Finally, we found a relatively good fit for the pixel belonging to the edge of the pore, middle panel, that could be related to the complex atmospheric stratification at the transition region between the pore and the surrounding granulation. To improve the accuracy of the fitting we would need more free parameters on the LOS velocity and magnetic field stratifications.
\subsection{Retrieved asymmetries}
The Stokes $V$ profile asymmetries depend on velocity and magnetic field gradients. Therefore, an accurate match of these asymmetries serves as a criterion to
check the reliability of the retrieved stratification of line of sight velocity and magnetic field. We plotted in Figure \ref{as_inver} the results for amplitude and area asymmetries obtained through the inverted profiles. If we compare these results with bottom panels of Figure \ref{as}, we can see that they are very similar, indicating that the accuracy of the fitting is high. However, there are minor differences in the central part of the pore, where the asymmetries are relatively smaller in the inverted profiles. We detected some pixels that can show a diminishing in area asymmetry from 13 per cent in the deconvolved profiles to 9.5 per cent in the inverted data. This reduction is smaller in the amplitude asymmetry although still present. However, we believe that these differences are sufficiently small to be certain that both, line of sight velocity and magnetic field gradients, are reliable.
\begin{figure}
\hspace{+0.2cm}
\includegraphics[width=16.0cm]{as_inver.eps}
\caption{Stokes $V$ area (top) and amplitude (bottom) asymmetries. They were computed using the fitted profiles from the inversion of the deconvolved data. The original asymmetries from the deconvolved profiles are shown on row of Figure \ref{as}.}
\label{as_inver}
\end{figure}
\begin{figure*}
\hspace{-0.5cm}
\includegraphics[width=18.3cm]{inver_context.eps}
\vspace{-0.3cm}
\caption{Spatial distribution of atmospheric parameters at different heights. From top to bottom row, temperature, LOS Velocity and, LOS magnetic field. From left to right column, optical depths $\log \tau=[0,-1,-2]$. Red LOS velocity indicates downflows while blue depicts upflows. White colors indicate a null velocity or a null longitudinal magnetic field.}
\label{inver_context}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=9.0cm]{context_small.eps}
\vspace{-0.4cm}
\caption{Continuum signal centred in the pore region. Horizontal lines indicate the location of the vertical cuts presented in Figure \ref{cuts}. Vertical ticks tentatively delimit the end of the pore substructures using the continuum map as reference.}
\label{context_small}
\end{figure}
\subsection{Spatial properties}
Figure \ref{inver_context} shows the spatial distribution of temperature, LOS velocity, and LOS magnetic field at different heights. Focusing first on the temperature, first row, we can see that its spatial distribution at $\log \tau=0$, leftmost panel, is similar to the continuum map, see Figure \ref{compar}, with hot granules, cool narrow regions, i.e. intergranules, and larger cool regions that conform the solar pore. If we move to upper layers, at $\log \tau=-1$ (middle panel), the landscape changes and resembles the line core information showed in middle column of Figure \ref{compar}. Plage regions are hotter than the surrounding quiet Sun while the pore is cooler. However, it seems that the height that is closer to the information displayed by the line core is $\log \tau=-2$, rightmost panel, because we can only see at this height a cool region at the left part of the pore structure, while the rest of the region shows temperatures similar to the quiet Sun, like the pattern found at line core wavelengths.
\begin{figure*}
\centering
\includegraphics[width=17.8cm]{cuts2.eps}
\vspace{-0.3cm}
\caption{Vertical cuts of different atmospheric parameters corresponding to the horizontal lines displayed in Figure \ref{context_small}. Each row contains, from left to right, temperature, LOS velocity, and LOS magnetic field. We chose a different colour table than in Figure \ref{inver_context} for the latter physical quantity because, in this case, there is no change of sign of the LOS magnetic field in the examined regions being always with positive polarity or almost null. Vertical lines correspond to the location of the tick marks on Figure \ref{context_small}.}
\label{cuts}
\end{figure*}
Concerning the LOS velocity, the granulation pattern is seen at $\log \tau=0$ in quiet Sun regions while the inner part of the pore shows almost null velocities, white areas. In addition, there is a presence of strong downflows, red colour, at the periphery of the pore structure and at plage regions. If we analyse upper layers, middle and right panels, we find that LOS velocities slightly resemble the granulation pattern in quiet Sun areas although with lower velocities. In addition, the size of the structures is larger because they have expanded with height. However, the presence of the previously mentioned downflows detected at plage regions and at the edges of the pore have disappeared, indicating that these downward motions take place only at the bottom of the photosphere. We computed the height where these downflows velocities become null and we obtained a mean optical depth value of $\log \tau=-1.1\pm0.4$.
Finally, regarding the longitudinal magnetic field, we detected field strengths larger than 2~kG at the bottom of the photosphere, left panel, that decrease with height being in the range of $1-1.5$~kG at the middle photosphere, middle and right panels. Moreover, in this case, we can also detect an expansion of the magnetic structures with height.
\begin{figure*}
\centering
\includegraphics[width=16.0cm]{scatter.eps}
\vspace{-0.3cm}
\caption{Scatter plots of three different atmospheric parameters; temperature, LOS velocity, and magnetic field strength. We show the spatial relation between these quantities at three different heights, i.e. $\log \tau=[0,-1,-2]$ (different columns). Different rows, from top to bottom, display the relation between temperature and LOS velocity, temperature and magnetic field strength, and LOS velocity and magnetic field strength.}
\label{scatter}
\end{figure*}
\subsection{Vertical cuts}
In order to study the optical depth stratification of the different physical quantities we selected three regions, see lines in Figure \ref{context_small}, that cross the pore. They are located at different positions in the south-north direction and have the same size in the east-west direction. We selected their location trying to examine the spatial differences between the substructures that compose the pore. The vertical stratification of the atmospheric parameters in these three regions is shown in Figure \ref{cuts} (see different rows). In addition, we added vertical lines in every panel that correspond to the vertical ticks of Figure \ref{context_small} in order to facilitate the visualization of pore substructures. We selected the location of these lines, or ticks in Figure \ref{context_small}, trying to delimited the end of the pore substructures using the continuum intensity as reference.
Starting with the temperature, first column, we can see strong fluctuations of its value at the base of photosphere, i.e., around $\log \tau=-1$, being at lower optical depths where the pore substructures are located (see the pixels in between the vertical lines). This property is present in all the different selected horizontal cuts (see different rows). The vertical stratification is relatively different between the inner part of the pore and the surrounding granulation, being hotter at middle heights (around $\log \tau=-1$) for the pixels outside the pore and cooler for the pixels belonging to the magnetic structure.
In the case of the LOS velocity, we also found a clear distinction between the pixels that belong to the pore structure and surrounding pixels. In the first case, we found null (or very small) LOS velocity values while, in the second case, we detected redshifted velocities around $3\sim4$ km/s at lower heights that go to zero or to small ($\sim1$~km/s) upflow velocities at middle layers (see the pixels near the vertical lines).
Regarding the line of sight magnetic field we found that is always positive in the selected areas, indicating that it does not change of polarity at the edges of the pore nor outside the magnetic structure. The magnetic field strength is large inside the structure with strengths around 2~kG that decreases with height up to $1$~$\sim$~$1.5$~kG at middle heights. Outside the magnetic structure, see vertical lines, the LOS magnetic field becomes almost null generating a large spatial contrast similar to the one observed in the continuum map (see Figure~\ref{context_small}).
Finally, we detected a large horizontal variation, even inside the pore substructures, in the three examined quantities. The reason could be that the pore substructures are constituted by a complex bundle of magnetic tubes that we are partially solving with the present spatial resolution, i.e. around 200 km. If we take a closer look of Figure~\ref{context_small}, we can see that the continuum intensity strongly fluctuates in the neighbouring granulation, but also in the pore substructures, being much darker at their central part than at their edges, located at just few pixels away.
\subsection{Spatial relations between atmospheric parameters}
We take advantage of the inversion results to examine the spatial relation between different atmospheric parameters, i.e. temperature, LOS velocity, and magnetic field strength, at different heights. In that sense, we aim to understand which is the relation between temperature and LOS velocity, temperature and magnetic field strength, and LOS velocity and magnetic field strength. We studied these relations in the pixels that belong to the solar pore as well as for its surrounding pixels. We selected a small region of $12\times12$ arcsec centred in the solar pore and displayed in Figure \ref{scatter} scatter plots between the mentioned atmospheric parameters at three different heights, $\log \tau=[0,-1,-2]$ (different columns in the figure).
We can start with the relation between the temperature and the LOS velocity (first row). At the bottom of the photosphere, left panel, we detect hot blueshifted pixels (negative velocities) that correspond to granules and slightly cooler redshifted pixels (positive velocities) that belong to intergranular lanes and to the pixels at the edges of the pore. We also see a set of pixels that covers a large range of temperatures and shows very low, or null, LOS velocities. We believe that the reason of this wide range of temperatures is that we have relatively hot pixels that can correspond to the interface between granules and intergranules and also very cool pixels that belong to the inner part of the solar pore. This large variation of temperature for similar velocities is the reason why the distribution of points in the scatter plot resembles the shape of the Greek character \textit{Y}. The same shape appears at higher heights but it is less clear as we move to upper layers because the pixels located at the edges of the pore now show almost null velocities. Lastly, we can see at all heights that blueshifted pixels usually correspond to hotter regions.
Middle row displays the relation between temperature and magnetic field strength. In this case, we can see at middle heights, where we are more sensitive to the magnetic field strength, a faint arcade shape indicating that the larger the magnetic field strength the cooler the temperature of the pixel (see also \cite{Sobotka2012}). In addition, we can also detect at middle layers, i.e. $\log \tau=-1$, that the pixels showing a large field strength (more than 1.5~kG) can reach temperature values as low as 4500~K.
Finally, bottom row displays the relation between LOS velocity and magnetic field strength. Left panel, i.e. $\log \tau=0$, reveals large positive velocities for intermediate magnetic field strengths that correspond to the pixels located at the edges of the pore. We can also detect very low (or null) velocities for the pixels with largest magnetic field strengths. These pixels are associated with the inner part of the magnetic structure. If we examine higher layers (middle and right panels) we see no trace of large downflows while the pixels harbouring a strong magnetic field strength still show almost null LOS velocities.
\section{Discussion and Conclusions}
We applied for the first time the spatial deconvolution technique on Hinode/SP observations of a solar pore. This region represents a challenge to the deconvolution code because it displays large continuum contrasts between the central part of the structure, with low intensities, and its surroundings, located at less than a few arcsec, that display the brighter granulation pattern. Thus, we faced this work with two main aims, the first was to test the capabilities of the spatial deconvolution code and the second was to analyse the physical properties of the magnetic structure. The first question reveals that the spatial deconvolution properly works without introducing ringing effects in the pore or artefacts in the Stokes profiles. In addition, the Stokes $V$ asymmetries are slightly modified by the process but only in their amplitude because the asymmetry sign does not change in most of the observed pixels. Moreover, we also found an improvement in the continuum contrast and a diminishing of the size of the magnetic structures in agreement with previous works on this topic.
Regarding the second aim of our study, i.e. the analysis of the physical properties of a solar pore, we performed inversions of the Stokes profiles in a large fragment of the original observed field of view. First, we checked that the retrieved atmospheric models lead to synthetic Stokes profiles that accurately fit the observed ones. These synthetic profiles even match the Stokes $V$ amplitude and area asymmetries (although, in some cases, asymmetries were slightly lower in the inverted profiles), supporting the reliability of the retrieved line of sight gradients of the atmospheric parameters.
Later, we proceeded to examine the spatial distribution of the atmospheric parameters at different heights. We found that the inner part of the pore is cooler than its quiet Sun surroundings at all heights. At the same time, the neighbouring plage is hotter than the quiet Sun at middle layers. Then we examined the LOS velocity finding that is null in the inner part of the magnetic structure and that shows downward motions at its edges. These downward motions are only detected at lower heights, below $\log \tau\sim-1$, indicating that they are low photospheric features. Regarding the LOS magnetic field, we found strong field strengths, larger than 2~kG in the central part of the pore, and with single polarity everywhere. In addition, the magnetic field strength diminishes with height as the spatial size of the magnetic structures increases as we examine upper layers.
We also studied the vertical stratification of several cuts that cross the magnetic substructures of the pore. We found a significant decrease of the temperature inside the pore probably due to the inhibition of the overturning convection produced by its large magnetic field strength. In addition, the temperature between magnetic substructures was larger than inside the magnetic ones at middle heights. This contrast between magnetic and non magnetic substructures was also found in the LOS velocity. We detected null velocities at all heights inside magnetic concentrations and downward velocities outside them that can go to zero or turn to small upflow velocities at middle layers. Regarding the magnetic field strength, it shows maximum values around 2~kG at the low photosphere that decrease with height up to $1$~$\sim$~$1.5$~kG at middle layers. The inferred LOS magnetic field is unipolar in and outside magnetic substructures although its amplitude is very low outside magnetic areas generating a high spatial contrast between both regions. Finally, we examined spatial relations between temperature, LOS velocity, and magnetic field strength at different heights, corroborating all the properties described before for the pixels of the inner part of the pore as well as for the surrounding ones.
We can conclude that the spatial deconvolution improves the overall quality of the observations without introducing any kind of artefacts and, at the same time, it allows us to infer the atmospheric information without relying on any type of stray light correction. In addition, our findings are in agreement with previous works as the ones presented by \cite{Morinaga2007}, \cite{Scharmer2011}, or \cite{Sobotka2012}. Moreover we also increase our knowledge about pores and their surroundings describing the height stratification of their atmospheric parameters. However, the examined structure does not show any kind of bright features, as bright dots or light bridges, that would be interesting to study after applying the spatial deconvolution method. In that sense, we need to check different solar pore regions to increase our knowledge about them. Additionally, this particular pore limits our study to the photosphere because it is quiet at chromospheric layers so we could not examine which type of photospheric activity can trigger an active chromospheric response. Maybe it is due to its simple configuration but we plan to examine other solar pores trying to understand if the lack of chromospheric activity is a general property of solar pores or if it is just a particular condition of the present case.
\section*{Acknowledgements}
We thank A. Asensio Ramos his suggestions and ideas that helped to develop this work. BRC acknowledges financial support through the Project No. ESP2014-56169-C6-2-R funded by the Spanish Ministry of Economy and Competitiveness. We also thank the anonymous referee for providing helpful comments and suggestions that allowed to improve this work. \textit{Hinode} is a Japanese mission developed and launched by ISAS/JAXA, collaborating with NAOJ as a domestic partner, NASA and STFC (UK) as international partners. Scientific operation of the Hinode mission is conducted by the Hinode science team organized at ISAS/JAXA. This team mainly consists of scientists from institutes in the partner countries. Support for the post-launch operation is provided by JAXA and NAOJ (Japan), STFC (U.K.), NASA, ESA, and NSC (Norway).
\bibliographystyle{mnras}
|
\section{Conclusion}
In this paper we presented a novel autonomous system for acquiring close-range high-resolution images
that maximize the quality of a later-on 3D reconstruction.
We demonstrated that this quality strongly depends on the planarity of
the scene structure (complex structures vs. smooth surfaces), the camera constellation and the chosen dense MVS algorithm.
We learn these properties from unordered image sets without any hard ground truth and use the acquired knowledge
to constrain the set of possible camera constellations in the planning phase.
In using these constraints, we can drastically improve the success of the image acquisition,
which finally results in a high-accuracy 3D reconstruction with a significantly higher scene coverage
compared to traditional acquisition techniques.
\section{Multi-View Stereo Confidence Prediction}
\label{sec:confidence_prediction}
Given a specific scene structure (e.g. vegetation) and a camera constellation,
the MVS confidence encodes
the likelihood that a dense reconstruction algorithm will work as intended.
With "work as intended" we mean
that if a scene part is observed by a sufficient number of cameras then
the algorithm should be able to produce a 3D measurement within the theoretical uncertainty bounds for each pixel that observes this scene part~\cite{mostegel16}.
The first matter we address in this section is how we can generate training data to predict the MVS confidence without any hard ground truth.
Therefore we extend our approach for stereo vision~\cite{mostegel16} to multi-view stereo.
Then we outline our machine learning setup and explain how we can use this setup to
predict the MVS confidence in real-time during the image acquisition.
\subsection{MVS Training Data Generation}
As it is extremely tedious to come by 3D ground truth, the basic idea of \cite{mostegel16} is to
use self-consistency and self-contradiction from different
view points for generating labeled training data.
This approach is related to depthmap fusion, but outputs 2D label images instead of depthmaps.
Pixels that are associated with consistent depth values become positive training data,
while inconsistent depth values lead to negative training data.
This data is then used for training a pixel-wise binary classification task.
The main challenge during the training data generation is to keep
the false positive rate (consistent but incorrect) and the false negative rate (correct but inconsistent)
as low as possible, while labeling as many pixels as possible.
In \cite{mostegel16}, we start by computing a depthmap for each stereo pair in the dataset.
A single depthmap can be interpreted as the 3D reconstruction of a camera cluster with two cameras and a fixed baseline.
In the case of multi-view stereo, we can choose an arbitrary number of cameras per cluster
in any constellation.
As this general case has too many degrees of freedom to be estimated efficiently,
we limit ourselves to three cameras per cluster,
which is also the standard minimum number of cameras for most MVS approaches (e.g.~\cite{furukawa10pmvs,rothermel12}).
Within this triplet of cameras, the most important factor is the baseline between the cameras
or more precisely the triangulation angle between the cameras and the scene.
This triangulation angle can be freely chosen.
We want to use this property to learn the relationship between
MVS confidence and the triangulation angle so that
we can choose the right camera constellation for the presented scene in our view planning approach.
In theory, a large triangulation angle between cameras is beneficial as it reduces the 3D uncertainty.
However, in practice a large triangulation makes it more difficult to find correspondences between the images,
especially when the 3D structure is highly complex.
To learn this relationship, we first generate a large variety of triangulation angles in the training data.
We randomly sample image triplets from a fixed number ($t$) of triangulation angle bins,
while ensuring that the images have sufficient overlap.
For each of these camera triplets, we execute the chosen dense MVS algorithm
and project the resulting 3D reconstruction back into the images to obtain one depthmap per image.
Using these depthmaps, we can proceed with the training data generation
in three stages~\cite{mostegel16}.
The first stage has the purpose of reducing the influence of all consistent but incorrect measurements.
In practice, we can observe that the likelihood that two measurements of independent 3D reconstructions\footnote{3D reconstructions that were produced with the same MVS algorithm from independent image sets.}
are consistent but incorrect at the same time decreases as the relative view point difference increases.
Thus, we analyze how well each measurement is supported by reference reconstructions from a sufficiently different view point.
We treat a reference measurement as sufficiently different if the view angle difference $\alpha_\text{diff} > \alpha_\text{min}$
or the scale difference $s_{\text{res,query}} > s_{\text{min}}$ is sufficiently large.
We compute these values as
$\alpha_{\text{diff}} = \measuredangle (\overrightarrow{\mathbf{p}_{\text{query}} \mathbf{c}_{\text{ref}} },\overrightarrow{\mathbf{p}_{\text{query}} \mathbf{c}_{\text{query}} })$
and
$s_{\text{res,query}} = \text{res}_\text{ref} / \text{res}_\text{query} $ with $\text{res}_\mathsf{x} = f_\mathsf{x} / \|\mathbf{c}_\mathsf{x} - \mathbf{p}_{\text{query}}\|$,
where $\mathbf{c}_{\text{x}}$ is the mean camera center and $f_{\text{x}}$ the mean focal length of a camera triplet.
If a reference
measurement fulfills one of the two conditions, we increment the \emph{support} of the query measurement by one.
Note that as in~\cite{mostegel16}, reference measurements from a similar view point are
only allowed to increment the \emph{support} once.
In the second stage, we let the parts of the depthmaps with at least one \emph{support} vote on the consistency of all depthmap values.
The voting process proceeds analog to~\cite{mostegel16}.
For each query measurement, we collect positive and negative votes as shown in Fig.~\ref{fig:supp_n_contra}.
The votes are weighted with their \emph{support} and their inverse 3D uncertainty~\cite{mostegel16}.
Based on the voting outcome, all pixels with at least one vote are then either assigned a positive or a negative label.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{supp_n_contra.pdf}
\caption{Consistency voting. A positive vote (center) is only cast
if the reference measurement is within the uncertainty boundary of the query measurement.
A negative vote is either cast if a reference measurement blocks the line of sight of the query camera (left) or
the other way around (right).}
\vspace{-15pt}
\label{fig:supp_n_contra}
\end{figure}
The third stage requires more changes to generalize to multi-view stereo.
While in~\cite{mostegel16} this stage only has the purpose of detecting outliers,
in our case we also have to detect missing measurements.
More precisely, we have to detect if the MVS algorithm
failed to produce any output in a region where it should have been geometrically possible
and use this case as a negative training sample.
For detecting these missing parts we use a combination of a depthmap augmentation~\cite{mostegel16} and
two surface meshes.
We use two meshes with slightly different object boundaries to account for errors in the meshes.
To construct these meshes, we first use all available images in the dataset
to compute a sparse point cloud~\cite{rumpler2014automated}.
From this point cloud we robustly extract a surface mesh~\cite{labatut07delaunay,vu12dense_mvs},
and then shrink and expand this mesh for our purpose.
The shrunken mesh is obtained by
performing three iterations of neighbor-based smoothing.
In each iteration
a vertex moves half the distance to the average position of the vertices that share an edge with this vertex.
For the second mesh, we expand the shrunken mesh again.
For this purpose, we compute a vector by averaging the motion vectors of a vertex
and its neighbors from the shrinking procedure.
Each vertex is then moved twice the vector length in the
opposite direction of this vector.
If the depthmap augmentations and the two meshes agree that some part of the scene is missing,
the corresponding pixels are used as negative training samples.
\subsection{Machine Learning Setup}
For view planning, we want to know which camera constellation will give us a good chance of
getting a complete and accurate 3D reconstruction.
To help with this task,
we want to use the already acquired images during the acquisition.
For training, we pose the problem as a pixel-wise classification task.
During run-time, we compute the MVS confidence depending on the triangulation angle
and the scene around the pixel of interest.
For this task, we chose Semantic Texton Forests (STFs)~\cite{shotton08textonforest}.
We selected this approach for three main reasons.
First, this approach is very fast in the execution phase as it operates directly
on the input image (without any feature extractions or filtering).
Second, STFs have shown a reasonable performance in semantic image segmentation.
Third, it is possible to store meta information in the leaves of the forest.
We use this property to store the triangulation angle under which a sample was obtain (or failed to obtain).
This does not influence the learning procedure, but allows us to predict the reconstruction
confidence in dependence of the triangulation angle at evaluation time.
During the image acquisition, we want to compute the MVS confidence in real-time
on a specific computer for a specific high-resolution camera.
Thus we provide two ways to reduce the prediction time to the operator's needs.
First, we restructure the STF leaf nodes to contain a fixed number ($b$) of angular
bins with one confidence value for each bin.
Second, we can make use of the property that the confidence prediction is in general
a smooth function for a specific type of object (see~Sec.~\ref{sss:valcamonica}).
Thus, we evaluate the MVS confidence on a regular grid and
compute a confidence image with $b$ channels for each input image.
\section{Experiments}
\label{sec:experiments}
We split our evaluation in two main parts.
The first part evaluates the performance and the information
which is stored by our confidence prediction approach.
The second part focuses on our autonomous acquisition system
and how it performs in a real world experiment.
\subsection{Confidence Prediction}
In the first part of this section we benchmark the performance of our training data generation and
the prediction performance of the Semantic Texton Forest (STF)~\cite{shotton08textonforest} on the KITTI dataset~\cite{geiger12}.
In the second part, we use a challenging multi-view dataset to evaluate what the system can learn about two different
multi-view stereo approaches in relation to scene structure and camera constellation.
In all our experiments, we used the same STF setup.
We implemented the STF in the random forest framework
of Schulter~et~al.~\cite{schulter14a}.
We only use STF in its basic form (without image-level prior~\cite{shotton08textonforest}).
This means that the split decision is made directly on the
image data (Lab color space) within a patch of the size $27\times 27$.
We trained 20 trees with a maximum depth of 20.
For the split evaluation, we used the Shannon Entropy, minimum leaf size for further splitting of 50,
5000 node tests, 100 thresholds and 1000 random training samples at each node.
For all our experiments, we extracted approximately 4 million training patches for each
class in training.
\subsubsection{KITTI2012 Dataset}
In this experiment, we apply our approach to the scenario of street-view dense stereo reconstruction
using the KITTI dataset~\cite{geiger12}, which provides a semi-dense depth ground truth recorded with a Lidar.
For learning,
we follow the same procedure as in \cite{mostegel16} and
use the 195 sequences of 21 stereo pairs of the testing
dataset for automatically generating our label images.
We treat each stereo pair as a distinct cluster and
use a semi-global matcher with left-right consistency check (SURE~\cite{rothermel12})
as the query algorithm.
As in \cite{mostegel16}, we evaluate the label accuracy and the average Area Under the Sparsification Curve (AUSC),
although with a slightly different setup.
While stereo confidence prediction \cite{mostegel16} tries to decide which depth values cannot be trusted from an already computed depthmap,
our aim is to predict which kind of structures cause more problems than others.
Thus, we remove all regions from the Lidar ground truth, which are not visible in both color images (including object occlusions).
With this setup we reach a labeling accuracy of \textbf{98.7\%}
while labeling 35\% of the ground truth pixels (which is very similar to the results in \cite{mostegel16}).
For the sparsification we obtain a relative AUSC of 3.15 (obtained AUSC divided by optimal AUSC).
This means that the AUSC is \textbf{39\%} lower than random sparsification with 5.15.
This is a strong indication that the system learned to predict regions which are difficult to reconstruct
for the semi-global matcher.
For the matter of completeness, we also analyze the sparsification performance
of the STF~\cite{shotton08textonforest} with the exact same setup as in \cite{mostegel16}
(including the training data generation).
With this setup STF reaches a relative AUSC of 6.63.
It is not surprising that STF cannot reach the sparsification performance of stereo specific sparsification approaches (e.g. left-right difference with 2.81),
as the STF only uses color information of a single image and thus has no chance to reason about occlusions.
Nevertheless, the STF was able to extract some high level knowledge in which regions the chances of failure are higher and
thus still obtains a 31.4\% lower AUSC value than random sparsification (9.65).
\subsubsection{Val Camonica Dataset}
\label{sss:valcamonica}
\begin{figure*}[top]
\vspace{-15pt}
\centering
\subfigur
{
\includegraphics[width=0.35\textwidth]{sure_recon_prob.pdf}
}\quad
\subfigur
{
\includegraphics[width=0.35\textwidth]{pmvs_recon_prob.pdf}
}\quad
\subfigur
{
\includegraphics[width=0.12\textwidth]{structures.png}
}
\caption{Dependence of the confidence prediction on the triangulation angle and the 3D structure.
On the right, we display the patches ($50\times50$px) which we used to produce the curves.
These curves show the confidence prediction within angular bins (20 bins between min and max).
The curves stop if less than 1\% of the collected triangulation angles fall within a bin.
For both approaches (SURE and PMVS), there is a significant difference between smooth surfaces (marker, bridge, stone) and
high frequency structures (tree, grass).
The predicted confidence is to some extent correlated with the degree of non-planarity of
a structure.
While grass viewed from far away is quite easy to reconstruct,
the same grass viewed close up becomes very hard to reconstruct.
For both approaches, the chance for reconstructing highly non-planar structures above $30^\circ$ is virtually zero.
}
\vspace{-10pt}
\label{fig:sure_structures}
\end{figure*}
For the second dataset, we have chosen a reconstruction scenario in a
closed real-world domain, where the task is the 3D reconstruction of prehistoric rock art sites
in the Italian valley of Val Camonica.
The recorded dataset consists of over 5000 images of 8 different sites (see supplement), which
contain a well-defined set of 3D structures (mainly rock, grass, trees, bridges, signs and markers).
These structures dominate nearly all sites in the region (hundreds), which makes
this a perfect example for learning and predicting domain specific properties of a query algorithm.
For generating camera triplets we used $t=5$ triangulation bins.
The lowest triangulation angle bin starts at a minimum angle of $4^\circ$ and
ranges to double that value, where the next bin starts.
On each resulting triplet we execute a query algorithm three times at different image resolutions (levels 1, 2 and 3 of an image pyramid).
We evaluate two query algorithms for the dense 3D reconstruction.
The first query algorithm is based on semi-global matching SURE~\cite{rothermel12},
but can use more than two views for improving the reconstruction accuracy.
In contrast, our second query algorithm PMVS~\cite{furukawa10pmvs}
tries to densify an initial sparse 3D reconstruction
through iterative expansion.
For the quantitative evaluation of this experiment, we performed leave-one-out cross validation
across the 8 sites, i.e. we train on 7 sites and test on the remaining.
This led to the following classification accuracies:
PMVS: 81.1\% (STD: 4.2\%) and SURE: 65.3\% (STD: 6.1\%).
Within this context, we also analyzed the influence of regular grid sampling on the prediction performance.
For small grid sizes the classification error stays nearly the same (relative error increase is below 1\% for 4 pixels),
while for larger grid sizes it declines gradually (below 3\% for 16 pixels and below 7\% for 64 pixels).
This means that regular sampling can drastically reduce the computational load of the prediction with only a small decrease of the prediction performance.
Now let us analyze what the system learned about the two algorithms
in relation to scene structures and triangulation angle.
In Fig.~\ref{fig:sure_structures} we show the confidence prediction for six different structures.
From this experiment we can draw several conclusions.
First, the 3D structure of the scene has a significant influence on how well
something can be reconstructed under a given triangulation angle.
The more non-planar a structure is, the harder it is to reconstruct at large triangulation angles.
Second, the two analyzed approaches react very differently to a change in triangulation angle.
While for SURE the confidence is always highest for very small angles,
PMVS' confidence stays constant
for smooth surfaces.
In the case of non-planarity, SURE is clearly more robust than PMVS.
\subsection{Autonomous Image Acquisition}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{error_dist-eps-converted-to.pdf}
\caption{Error histogram on the rock surface. We show the normalized histograms of the error distribution and the
1 $\sigma$ bound in which 68.3\% of all measurements lie. Grid and F1x20 share the same error bound.
}
\vspace{-15pt}
\label{fig:error_histo}
\end{figure}
\begin{table*}
\centering
\begin{tabular}[b]{|c||c|c|c|c|c||c|c|c|}
\hline
&Init & Grid & F5x4 & F1x20 & NP5x4 & G+F5x4 & G+F1x20 & G+NP5x4\\\hline
\emph{cov} & $ 53.5 \pm 1.2 $ & $ 56.0 \pm 1.2 $ & $ \mathbf{65.6} \pm 1.6 $ & $ \mathbf{66.6} \pm 1.4 $ & $ 56.7 \pm 1.4 $ & $\mathbf{69.5} \pm 1.5$ & $67.0 \pm 1.5 $ & $57.2 \pm 1.2$\\\hline
$f_{res}$ & $ 17.9 \pm 1.3$ & $ 43.9 \pm 2.6 $ & $ 42.2 \pm 2.6 $ & $ \mathbf{47.5} \pm 2.7 $ & $ 29.3 \pm 2.3 $ & $\mathbf{52.8} \pm 2.7$ & $ \mathbf{55.3} \pm 2.7$ & $46.8 \pm 2.6$ \\\hline
$f_{unc}$ & $ 15.5 \pm 0.3 $ & $ \mathbf{22.8} \pm 0.4 $ & $ 21.2 \pm 0.5 $ & $ 20.7 \pm 0.5 $ & $ 19.9 \pm 0.5 $ & $\mathbf{27.7} \pm 0.5$ & $26.2 \pm 0.4$ & $25.6 \pm 0.4$ \\\hline
$f$ & $ 16.7 \pm 0.8 $ & $ \mathbf{33.4} \pm 1.5 $ & $ 31.7 \pm 1.5 $ & $ \mathbf{34.1} \pm 1.6 $ & $ 24.6 \pm 1.4 $ & $\mathbf{40.2} \pm 1.6$ & $\mathbf{40.7} \pm 1.6$ & $ 36.2 \pm 1.5$ \\\hline
\end{tabular}
\caption{Fulfillment statistics in percent. We show the coverage of the region of interest \emph{cov}, the resolution fulfillment $f_{res}$ and the
uncertainty fulfillment $f_{unc}$, as well as the overall fulfillment $f$ as defined in Sec.~\ref{sec:confidence_prediction}.
We display the mean value and the standard deviation over the three surface meshes. We mark all results within the standard deviation of the best method
with a bold fond. In the first column we show the results with only the 19 initialization images, then we show the four standalone approaches.
The last three columns show a combination of the standard grid approach (Grid) with the other approaches.
}
\vspace{-10pt}
\label{tab:fulfillment}
\end{table*}
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{details.jpg}
\caption{Resulting 3D reconstructions. On the top left, we show one of the images acquired by the UAV with the region of interest in red.
The other four columns show all view plans.
The blue cameras are regular or triplet cameras, while the pink cameras ensure sufficient overlap for sequential registration.
In the bottom row we show all four reconstructions (the field of view is marked blue in the top left image).
Note that our approach has the best coverage underneath the trees. The point colors vary due to changing illumination conditions.
}
\vspace{-15pt}
\label{fig:details}
\end{figure*}
To evaluate our image-acquisition approach in this scenario, we first run different view planning algorithms
on-site and then analyze the effective reconstruction output, which is computed off-site.
As we also desire a reconstruction of the surrounding environment (which is dominated by vegetation),
we use SURE~\cite{rothermel12} as an MVS algorithm.
For this experiment, we run three versions of the proposed approach.
The first version is our full approach (F5x4), where we let the algorithm plan 4 camera triplets per iteration
for a total of 5 iterations.
In the second version (F1x20), we let our approach plan the same number of total triplets (20) but in a single iteration, i.e. we disable the incremental geometry updates.
The third version (NP5x4) is exactly the same as F5x4 but without the prediction to constrain the triangulation angle.
As a baseline method, we use grid planning with 80 percent overlap.
All approaches share the same set of parameters.
The fulfillment requirements were set to $c = 3$, $r_d = 8$\,mm and $a_d = 8$\,mm with $\alpha = 0.5$.
The safety distance was set to 5\,m at a maximum octree~\cite{lau13} resolution of 2\,m
and the minimum camera overlap for registration to $o_{min}=50\%$.
The triangulation angle was binned in $b=9$ steps of $5^\circ$ from $0^\circ$ to $\gamma_{max} = 45^\circ$.
For the inverse visibility estimation we set the parameters such that the planning approximately takes 5 seconds per planned triplet, i.e. $N_t = 2000$, $N_p = 5000$ and $N_v=200$ with $\phi = 120^\circ$.
This parameters resulted in an effective execution time per triplet of 5.98 seconds (STD: 2.19) over all experiments on a HP EliteBook 8570w.
The confidence was evaluate on a regular grid with a step size of 8 pixels, which resulted in a confidence prediction time of $\sim$2\,sec/image.
We acquire the images with a Sony Nex-5 16Mpx camera mounted on an Asctec Falcon8 octocopter.
For this experiment, we focus on one site in Val Camonica, namely Seradina Rock 12C.
The rock surface (17$\times$13\,m) is covered with prehistoric rock carvings and
is partly occluded by the surrounding vegetation (Fig.~\ref{fig:details}).
We placed 7 fiducial markers in circle around the rock of interest and measured them with
a Leica total station.
These markers can be automatically
detected in the images and are used for geo-referencing the offline reconstructions~\cite{rumpler2014automated}.
Additionally, a ground truth mesh of the rock (not the surroundings) was obtained through terrestrial laser scanning (TLS) in the same coordinate system two years before.
The mesh has a resolution of 8\,mm edge length and the accuracy of the laser scanner (Riegl VZ-400) is 5\,mm.
We use this mesh to evaluate the resulting 3D uncertainty.
To evaluate the coverage and the requirement fulfillment, we first obtain a geo-referenced sparse reconstruction from all flights on the day of the experiment ($\sim$500 images).
Then we obtain three meshes, one based on~\cite{labatut07delaunay,vu12dense_mvs} and the two others as described in Sec.~\ref{sec:confidence_prediction}.
As we know that these meshes will contain errors, we only use these meshes as a guideline for the evaluation.
Within the region of interest, we split all triangles to have a maximum edge length of 8\,cm.
For each taken image, we first compute the triangle visibility.
Then we produce a depthmap from all SURE 3D points linked to the image.
If the measured depth is either larger than or within 24\,cm of the triangle depth, we accept the 3D point
as a valid measurement of the triangle.
Based on the links of the 3D measurement, we then compute the fulfillment of the triangle analog to Sec.~\ref{sec:confidence_prediction}.
Finally, this results in a set of fulfillment and coverage scores over all triangles in the region of interest.
In field, all approaches were initialized with 19 images taken in grid at a height of 50\,m above the lowest point of the site.
The region of interest was marked in one of the initialization images, such
that it is centered on the rock and includes a few meters of the surrounding vegetation (Fig.~\ref{fig:details}).
Landing and take-off are performed manually, while the view plans are executed autonomously by the UAV.
\vspace{-10pt}
\paragraph{Results.}
For each of our approach variants, we executed SURE only on the three images of the triplets.
Like this we can evaluate the general success rate of view planning variants
in analyzing on which triplets SURE succeeded to produce any 3D output.
Without the confidence prediction the success rate is very low (\textbf{18\%} for \textbf{NP5x4}).
This shows the gap between theory and practice.
While in theory a large triangulation leads to a small 3D uncertainty,
the matching becomes much more difficult and only flat surfaces survive.
However, with the proposed confidence prediction we were able to reach a prefect
success rate for our full approach (\textbf{100\%} for \textbf{F5x4}),
and still reached an acceptable success rate without the reconstruction updates (\textbf{80\%} for \textbf{F1x20}).
In Table~\ref{tab:fulfillment} we display the effective fulfillment statistics of all approaches in the region of interest.
Of the standalone approaches, F1x20 and Grid take the lead, but are closely followed by F5x4.
The worst performance was reached by NP5x4.
While the dense grid performs well on the overall fulfillment,
we can see a \textbf{10\% gap} in the scene coverage, where F5x4 and F1x20 lead with nearly equal results.
F1x20 performs slightly better than F5x4, because F1x20 found a sweet spot in the center above the rock for a single triplet
where it was able to drop below the tree line and acquire a close up of the rock.
If we combine the results of the dense grid (Grid) with the proposed approach, we achieve the overall best results.
All evaluated measures improve significantly, which is an indication of a symbiosis between the approaches.
This suggests that for the given scene (which is quite flat for many scene parts) an initial grid reconstruction
with a subsequent refinement with the proposed approach is recommended.
Note that if the scene complexity increases and a grid plan can no longer be executed safely (e.g. underneath a forest canopy or indoors),
our planning approach is still applicable.
If we take a look at the error distribution in relation to the ground truth of the rock surface (Fig.~\ref{fig:error_histo}),
we can see that our approach and grid planning achieve very similar results.
Note the Grid only covered 87.4\% of the rock surface, while all others covered significantly more:
F5x4 covered 97.9\%, F1x20 94.7\% and NP5x4 94.0\%.
This is a very promising result, as we only allowed our approach to use the planned triplets and no combination between them,
while we put no such restrictions on the Grid approach.
Furthermore, many of the triplets focused on the surrounding vegetation and the overall number of acquired images by our approach is
lower than for the Grid approach (60 vs. 108 images).
Thus, our approach achieved a high accuracy at a higher coverage
with fewer images, which can also be observed visually in Fig.~\ref{fig:details} and the supplementary material.
\section{Introduction}
In this paper, we address the problem of UAV-based image acquisition for dense monocular 3D reconstruction
with high-resolution images at close range.
The aim is to acquire images in such a way that
they are suited for processing with an offline dense multi-view stereo (MVS) algorithm,
while at the same time fulfilling a set of quality requirements.
These requirements include coverage, ground resolution and 3D accuracy and
can be assessed geometrically.
However, determining how well the images are suited
for a specific MVS algorithm is much harder to model.
To extract depth from 2D images, MVS approaches have to
establish correspondences between the images.
To solve this challenging task every MVS approach has to
make some assumptions.
These assumptions vary from approach to approach, but
the most popular assumptions include saliency, local planarity and a static environment.
If some of these assumptions are violated the MVS algorithm will not be able to reconstruct the scene correctly.
Up to now,
this problem was widely ignored by monocular image acquisition approaches~\cite{dunn09nbv,hoppe12nbv,schmid12view_planning_mav,munkelt10mvp,hossein13,martin16,pistellato15},
which often leads to missing parts in the resulting 3D reconstructions~\cite{schmid12view_planning_mav,hoppe12nbv}.
In this work, we propose a solution for this problem via machine learning.
The main idea is to predict
how well the acquired images are suited for the dense MVS algorithm directly during the acquisition.
While this is already useful for quality assurance, we take this idea one step further
and use the acquired images to plan the optimal camera constellation with respect to the observed scene structure.
Within this context, we demonstrate that the likelihood of a successful 3D reconstruction
depends on the combination of scene structure,
triangulation angle and the used MVS algorithm.
We further refer to the prediction of this likelihood as \emph{MVS confidence prediction}.
\begin{figure}[top]
\vspace{-10pt}
\includegraphics[width=1\columnwidth]{system.pdf}
\caption{Autonomous Image Acquisition.
After a manual initialization, our system loops between view planning and autonomous execution.
Within the view planning procedure, we leverage machine learning to predict the best camera constellation
for the presented scene and a specific dense MVS algorithm.
This MVS algorithm will use the recorded high resolution images
to produce a highly accurate and complete 3D reconstruction off-site in the lab.
}
\vspace{-15pt}
\label{fig:system}
\end{figure}
This MVS confidence prediction is related but not equal to the (two-view) stereo confidence prediction,
which is a topic of increasing interest in the domain of stereo vision~\cite{haeusler13,spyro14,park15,mostegel16}.
In stereo vision, the confidence encodes the likelihood that an already computed depth value is correct,
whereas in our case it encodes the likelihood that we will be able to compute a correct 3D measurement later-on.
Despite this difference, the training of both tasks is closely related
and requires a large amount of data.
Up to now, obtaining this training data was a tedious and time consuming task,
evolving manual interaction~\cite{ladicky12,geiger12,menze15}, synthetic data~\cite{butler12,peris12,menze15}
and/or 3D ground truth acquisition with active depth sensors~\cite{strecha08dataset,geiger12,menze15,scharstein14}.
In \cite{mostegel16}, we present a new way of obtaining this training data for stereo vision.
The main idea is to use multiple depthmaps (computed with the same algorithm) from
different view points and evaluate consistencies and contradictions between them to collect training data.
In this work, we extend this completely automatic approach to multi-view stereo.
After training, our system operates completely on-site (Fig.~\ref{fig:system}).
For estimating the scene geometry, we use the already acquired images for performing
incremental structure-from-motion (SfM)~\cite{hoppe12onlinesfm}
and incremental updates of an evolving mesh~\cite{hoppe13incmeshing}.
Both modules run concurrently in real-time and deliver the camera poses of the acquired images
and a closed surface mesh representation of the scene.
Based on this information,
we plan future camera positions that maximize the quality of
a later-on dense 3D reconstruction.
This task falls in the domain of view planning, which has been shown to be NP-hard~\cite{tarbox95complete_coverage}.
Consequently,
a wide range of very task specific problem simplifications
and solutions were developed in the communities of
robotics~\cite{sadat15,rainville15,nieuwenhuisen16,heng15,galceran13,galceran15,forster14,dornhege13,bircher15AR,alexis15,mostegel14icra,vasquez_gomez13raytracing,scott03view_planning_survey,
hollinger12vp_underwater,englot11inspection,schmid12view_planning_mav},
photogrammetry~\cite{martin16,liu14,hossein14a,hossein14b} and computer vision~\cite{pistellato15, scott02pose_error,trummer10nbv,hoppe12nbv,dunn09nbv,munkelt10mvp,haner11vslam_viewplanning}.
What kind of simplification is chosen strongly depends on the used sensor,
the application scenario and the time constraints.
In this work,
we propose a set of simplifications that allows us
to compute a view plan in a fixed time-frame.
In contrast to active depth sensors,
a single 3D measurement in monocular 3D reconstruction requires multiple images to observe the same physical scene part.
Thus our first simplification is to
remove this inter-camera-dependency by planning triplets of cameras as independent measurement units.
Second, we introduce the concept of surrogate cameras (cameras without orientation)
to reduce the dimensionality of the search space.
Finally, we lower the visibility estimation time through inverse scene rendering.
In contrast to the works above, our formulation allows us to evaluate a large number of potential camera poses
at low cost,
while the run-time can be adjusted to the acquisition requirements.
In the following, we first describe the training and setup of our MVS confidence predictor.
Then we describe our fixed-time view planning strategy.
In our experiments, we
evaluate the performance and stored information of the confidence predictor
on a challenging outdoor UAV dataset.
In the same domain, we finally evaluate our autonomous image acquisition system with respect to
quality and completeness of the resulting 3D reconstructions.
\section{Supplementary Material}
In this document we provide further figures such that the reader
can have a broader perspective of the conducted experiments and their outcome.
In Figure~\ref{fig:sites} we show some examples of the Val Camonica Dataset which was used for training.
The whole dataset contains over 5000 images.
The images contain a great variety of viewing angles and acquisition scales.
The camera to scene distance varies from to 2 to 50 meters.
In Figures \ref{fig:rf_4_evo}, \ref{fig:no_rf_evo} and \ref{fig:rf_20_evo} we show the planning behavior of the
three variants of the proposed approach together with the growing sparse reconstruction during the flights.
Note the different planning behavior between the variants.
With the MVS confidence prediction (F5x4 and F1x20) we see that the view planning algorithm
strongly varies the triangulation angle.
For smooth surfaces (the gray rock in the center) the algorithm prefers large angles as this decreases
the 3D uncertainty, but it prefers a small triangulation angle for vegetation which
cannot be reconstructed under a large triangulation angle.
In contrast, if we execute our approach without the confidence prediction (NP5x4), we can see that
the algorithm always uses a large triangulation angle, which leads to an incomplete 3D reconstruction where
only large smooth surfaces survive (see Fig.~\ref{fig:details4}).
In Figure~\ref{fig:mesh_evo} we show the evolution of the surface mesh over multiple iterations for our
full approach (F5x4). We can see that the surface topology changes drastically for initially occluded
parts. If the system would simply trust the initial reconstruction, the quality estimation in the
marked region would be far from realistic.
In Figure~\ref{fig:details4} we show the resulting 3D reconstructions with and without color.
Note that the confidence prediction leads to a much higher scene coverage especially underneath the vegetation.
In Figure~\ref{fig:gt_comp} we compare the resulting reconstructions to the groundtruth mesh of the rock surface.
We can see that all approaches except NP5x4 have a very similar error distribution.
Further, we can see that our full approach (F5x4) has the highest coverage of the rock surface, while
at same time sharing the high 3D reconstruction accuracy with Grid and F1x20.
\thispagestyle{empty}
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{sites.jpg}
\caption{Examples of the Val Camonica Dataset. In the top row we show three images from far away,
in the middle three slanted views and in the bottom row three close-ups. All sites contain a limited set of 3D structures
(mainly rock, grass, trees, bridges and markers).}
\vspace{-15pt}
\label{fig:sites}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.6\textwidth]{rf_4_evo.jpg}
\caption{View plans of our full approach (F5x4). From top to bottom we show the five planning iterations.
On the left side, we show the current view plan together with the 3D points that were added to the sparse reconstruction with these images during the acquisition.
On the right side, we show the sparse reconstruction at the time of planning and all computed view plans up to this point.
The blue cameras belong to the planned camera triplets, whereas the pink cameras ensure a successful registration during the acquisition.}
\vspace{-15pt}
\label{fig:rf_4_evo}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.5\textwidth]{no_rf_evo.jpg}
\caption{View plans of the view planning approach without confidence prediction (NP5x4). From top to bottom we show the five planning iterations.
On the left side, we show the current view plan together with the 3D points that were added to the sparse reconstruction with these images during the acquisition.
On the right side, we show the sparse reconstruction at the time of planning and all computed view plans up to this point.
The blue cameras belong to the planned camera triplets, whereas the pink cameras ensure a successful registration during the acquisition.
Note that the algorithm constantly reduces the number of planned triplets as it has too much trust in the reconstruction algorithm.
Further, the algorithm tries to push the triangulation angle to the allowed limits ($45^\circ$) as a large triangulation angle
decreases the theoretic 3D uncertainty.
}
\vspace{-15pt}
\label{fig:no_rf_evo}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.5\textwidth]{rf_20_evo.png}
\caption{View plans of our approach without structure updates (F1x20).
In this version the approach only planned a single view plan.
We show the view plan together with all 3D points that were generated on-site during the acquisition.
The blue cameras belong to the planned camera triplets, whereas the pink cameras ensure a successful registration during the acquisition.}
\vspace{-15pt}
\label{fig:rf_20_evo}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{mesh_evo.jpg}
\caption{Mesh evolution during the flight of our full approach F5x4. The mesh is colored red inside the region of interest and gray otherwise.
The green circle highlights a region where the mesh topology changes drastically during the acquisition.
The biggest change can be observed in the first two iterations.
The images 1 to 5 show the mesh at planning time of the corresponding iteration, whereas image 6 shows the mesh after the last plan execution.
}
\vspace{-15pt}
\label{fig:mesh_evo}
\end{figure*}
\begin{figure*}[t]
\vspace{-15pt}
\centering
\includegraphics[width=0.98\textwidth]{details4.jpg}
\caption{We show the resulting reconstructions for the four different approaches.
To eliminate the influence of color in the depth perception, we computed the normals of the point cloud and
show the resulting point clouds also without color.
For all triplet based approaches (NP5x4, F1x20, F5x4) we only show the output of the planned triplets (without initialization images or additional images for registration).
Note that the coverage underneath the trees is significantly higher with confidence prediction (F1x20 and F5x4).
The bottom row (4a-b) shows a close-up of the rock surface. Note that Grid is overly smooth, while F1x20 and F5x4 have much sharper edges.
NP5x4 has a lot of missing parts caused by self-occlusion of the rock surface.
}
\vspace{-15pt}
\label{fig:details4}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{gt_comp.jpg}
\caption{Distance to the groundtruth.
On the top we show the groundtruth mesh which has been acquired with terrestrial laser scanning and was partly textured with UAV acquired images.
The other images show the color coded distance of the final reconstructions to the groundtruth mesh.
Note that the error of Grid, F1x20 and F5x4 is very similar, while NP5x4 has a much larger error in the bottom part of the rock.
Further notice that F5x4 has the largest coverage of the rock surface in the lower left corner of the rock.}
\vspace{-15pt}
\label{fig:gt_comp}
\end{figure*}
\end{document}
\section{View Planning}
\label{sec:planning}
The aim of our view planning approach is to plan
a set of useful camera poses in a fixed time frame.
As the view planning problem is NP-hard, we have to make several
simplifications to constrain the computation time.
One of our most prominent simplifications is that we plan equilateral camera triplets
instead of single cameras.
On the one hand, this lets us directly integrate our MVS confidence prediction and,
on the other hand, we can treat each camera triplet as an independent measurement unit.
In Fig.~\ref{fig:viewplanning} we show an overview of our approach,
which we use to guide the reader through our algorithm and its submodules.
As input our approach requires a snapshot of the estimated geometry (mesh and camera poses),
as well as the pre-computed MVS confidence images.
Further, the operator has to label a region of interest in one of the images (Fig.~\ref{fig:details}),
and define the desired quality constraints (ground resolution and 3D accuracy).
\begin{figure}[top]
\centering
\includegraphics[width=1\columnwidth]{viewplanning.pdf}
\caption{View planning. Our algorithm tries to find the $k$ next best camera triplets for improving the
acquisition quality.
Next to the arrows, we show the data communication between our submodules (M1-M4) in red and
in black we show how often this data is computed.
$S$ is the set of surrogate cameras, $T$ the set of considered unfulfilled triangles and $C^3$ the set of camera triplets generated from the surrogate cameras.
}
\vspace{-15pt}
\label{fig:viewplanning}
\end{figure}
\textbf{Estimation of quality fulfillment (M1).}
The aim of this submodule is to estimate how well our
desired quality constraints are currently fulfilled by the already captured images.
For this estimation, we need the
already acquired images and their camera poses $C_t$ as well as the surface mesh.
First, we bring all mesh triangles within the region of interest to approximately the same size
through iteratively splitting them until the maximum edge length equals the
average edge length before splitting.
Within the region of interest we then randomly select a fixed number $N_t$ of triangles.
Next we determine the visibility information between these triangles and $C_t$ through rendering the mesh.
Based on the information which cameras see which triangles, we evaluate
how well the desired quality constraints are currently fulfilled.
We compute the fulfillment separately for each triangle using
four fulfillment functions.
(1) The \emph{coverage} is modeled as a
Boolean with $f_{cov} = 1$ if a triangle is visible in a minimum of $c$ cameras and $f_{cov} = 0$ otherwise.
(2) The \emph{resolution} requirement ($px/m^2$) is defined as
$f_{res} = \frac{r}{r_d}$ (truncated above 1) for a desired resolution $r_d$.
(3) The fulfillment of the \emph{3D uncertainty} requirement is defined
as $f_{unc} = \frac{a_d}{\sqrt{u}}$ (truncated above 1)
for a desired accuracy $a_d$.
Here, $u$ stands for the maximum Eigen value of the covariance matrix related
to a triangle's centroid~\cite{hartley04multiview}.
(4) The last fulfillment function is the output of our MVS confidence prediction algorithm $f_{conf}$ (Sec.~\ref{sec:confidence_prediction}).
For evaluating these functions,
we generate all possible combinations of camera triplets from the cameras that observe a triangle $t$ ($c^3 \in C_t^3$).
We then evaluate the combined fulfillment function as:
\begin{equation}
f(t,c^3) = (\alpha f_{res} + (1-\alpha) f_{unc} ) \cdot f_{cov} \cdot f_{conf}
\end{equation}
This formulation allows the operator to define the relative weight $\alpha$ between
desired ground resolution and 3D accuracy, while the coverage and MVS confidence
encode the chances of a successful reconstruction.
The overall fulfillment of a triangle $t$ is computed as $f(t) = \max_{c^3 \in C_t^3}{f(t,c^3)}\inlineeqno$.
Based on the fulfillment information, we now further reduce the number of considered triangles
to a triangle set $T$.
We guide this reduction such that we end up with triangles that
have a low fulfillment but are well distributed over the scene of interest.
Thus, we randomly select a fixed number $N_v$ of triangles from a piece-wise constant distribution,
where the chance of selecting a triangle $t$ is weighted with $w(t) = 1 - f(t )/f_{conf} (t )$.
We remove $f_{conf}$ from the weight to
avoid bias towards structures that might not be reconstructible at all.
\textbf{Surrogate cameras (M2).}
In this submodule, we use the concept of surrogate cameras
to estimate the visibility
of mesh triangles from a large number of possible camera positions.
Thus, we first
randomly sample a fixed number $N_p$ of 3D positions in the free space of the scene.
These 3D positions represent the camera centers of surrogate cameras.
A surrogate camera has an unlimited field of view and thus also no orientation at this point
(later we will transform this surrogate camera into an equilateral camera triplet).
The usage of surrogate cameras allows us to reformulate the visibility estimation problem
and to estimate which surrogate cameras
are visible from a given triangle instead of the other way around.
The benefit of this formulation is that we are able to control execution time of the visibility estimation with the number of considered triangles
instead of the number of considered camera poses.
This enables us to evaluate a high number of camera positions at low cost.
For each triangle $t \in T$,
we place a virtual camera in the scene. The camera center of a virtual camera is set to the triangle's centroid and the
optical axis to the triangle's normal. We set the focal length of this camera such that we get a fixed field of view $\phi$.
Now we use the virtual cameras for rendering the scene, i.e. the mesh and the 3D points that define the centers
of the surrogate cameras.
The resulting visibility links are stored in the surrogate cameras.
\textbf{Finding the best camera triplet (M3).}
To find the best camera triplet at a low computational cost,
we guide the transformation from surrogate cameras to camera triplets
such that we only need to evaluate potentially useful and feasible camera constellations.
Thus, we first compute the potential fulfillment gain $g_{pot}(t)$ of a surrogate camera
with respect to a linked triangle $t$.
Formally, we define $g_{pot}(t) = max_\alpha\{ f(t,c^3_\alpha) - f(t) , 0 \}$,
for a hypothetical equilateral camera triplet $c^3_\alpha$, that has the surrogate camera in its center
and where each camera directly faces towards the triangle.
The triangulation angle $\alpha$ defines the distance between the cameras in the $b$ steps of
the predicted MVS confidence, which we evaluate with the confidence image
of the closest already captured image (with respect to the surrogate camera) that observes the triangle.
Using this potential gain information,
we determine in which direction the surrogate cameras should face.
Therefore, we perform a weighted mean
shift clustering on the rays towards the linked triangles.
As a weight we use the fulfillment gain and the bandwidth is set to the minimum camera opening angle. The winning cluster (i.e. the
cluster with the highest potential fulfillment gain) is chosen to define the general viewing direction of the surrogate camera.
Then we update the visibility information and the potential gains of the
now oriented surrogate cameras.
Given the orientation, we generate $b$ camera triplets for each
surrogate camera,
one for each confidence bin.
For each camera triplet $c^3$ we efficiently check the distance to obstacles~\cite{lau13}
and compute the fulfillment gain of $c^3$ as
\begin{equation}
g(c^3) = \sum_{t\in T_{c^3}} max\{ f(t,c^3) - f(t) , 0 \},
\end{equation}
where $T_{c^3}$ are the triangles that are visible from $c^3$.
Over all triplets,
we find the best camera triplet as
\begin{equation}
c^3_{\text{best}} = \text{arg}\max_{c^3 \in C^3} g(c^3),
\end{equation}
where $C^3$ is the set of all generated camera triplets\footnote{
For the implementation, we can drastically reduce the number of evaluations
by using the potential gain.
If we start with the surrogate camera with the highest potential gain,
we can stop if $\sum_t g_{pot}(t)$ of the evaluated surrogate camera is zero or
smaller than the current best gain.}.
If $g(c^3_{\text{best}})$ is greater than zero and we have not yet planned $k$ camera triplets,
we add $c^3_{\text{best}}$ to the set of already acquired images ($C_t$) and plan a new camera triplet.
Otherwise, we pass all planned camera triplets with positive gain on to the flight path optimization.
\textbf{Flight path optimization (M4).}
This module minimizes the travel distance between the camera poses and
ensures that the resulting images can be registered by the geometry estimation module.
First, we reorder the camera poses with a greedy distance minimization using the last captured
image as a starting point.
Then we check if the taken images can be connected to the given set of images respecting the capture sequence.
We assume that this is the case if an image has a minimum overlap $o_{min}$ with at least one of the previously captured images.
If this is not the case we sample camera poses which fulfill this property along the
trajectory from the closest previously captured camera pose to the target camera pose.
This results in a view plan that ensures a successful sequential registration of the planned image set.
|
\section{Introduction}
In wireless communications, separation of transmission and reception in time or frequency has been the standard practice so far. However, through simultaneous transmission and reception in the same frequency band, wireless full-duplex has the potential to double the spectral efficiency. Due to this substantial gain, full-duplex technology has recently attracted noticeable interest in both academic and industrial worlds. The main challenge in full-duplex (FD) bidirectional communication is self-interference (SI) cancellation. In recent years, many attempts have been made to cancel the self-interference signal. In \citep{bharadia2013full}, it is shown that $110$ dB SI cancellation is achievable, and by jointly exploiting analog and digital techniques, SI may be reduced to the noise floor.
A full-duplex physical layer in cellular communications calls for a re-design of higher layers of the protocol stack, including scheduling and resource allocation algorithms. In \cite{goyal2013analyzing}, the performance of an FD-based cellular system is investigated and an analytic model to derive the average uplink and downlink channel rate is provided. A resource allocation problem for an FD heterogeneous OFDMA network is considered in \cite{sultanmode}, in which the macro BS and small cell access points operate in either FD or HD MIMO mode, and all mobile nodes operate in HD single antenna mode. In \cite{di2014radio}, using matching theory, a sub-channel allocation algorithm for an FD OFDMA network is proposed. In both \cite{sultanmode} and \cite{di2014radio} only a single sub-channel is assigned to each of the uplink users in which they transmit with constant power. Resource allocation solutions are proposed in \cite{nam2015joint} and \cite{namradio} for FD OFDMA networks with perfect FD nodes (SI is canceled perfectly).
In this paper, we consider a general resource allocation problem in an OFDMA-based network consisting of an imperfect FD BS and both HD and imperfect FD users. We aim to maximize the weighted sum-rate of this network in the uplink and the downlink by joint sub-channel assignment and power allocation. To be more realistic, imperfect SI cancellation in FD devices is assumed and FD nodes suffer from their SI. Since our model is general in the sense that we assign both uplink and downlink weights to the users, HD users are allowed to transmit and receive in different sub-channels. Hence, when a node is only an uplink (downlink) user, its downlink (uplink) weight is set to zero. A contribution of our work is to consider the presence of a mixture of FD and HD users, which enables us to quantify the percentage of FD users needed to capture the full potential of FD technology in wireless OFDMA networks. We also investigate the effect of the SI cancellation level on the network performance, which has never been considered in related works.
The remainder of this paper is organized as follows. In Section \ref{system_model}, the system model is given and the optimization problem is formulated. In Section \ref{subchannel allocation}, a sub-channel allocation algorithm for selecting the best pair in each sub-channel is presented. Power allocation is considered in Section \ref{power_allocation}. Numerical results for the proposed resource allocation methods are shown in Section \ref{simulation_results}. Finally, the paper is concluded in Section \ref{conclusion}.
\section{System Model and Problem Statement} \label{system_model}
We consider a single cell network that consists of a full-duplex base-station (BS) and a total of $K$ half-duplex and full-duplex users. For communications between the nodes and the BS, we assume that an OFDMA system with $N$ sub-channels is used. All sub-carriers are assumed to be perfectly synchronized, and so there is no interference between different sub-channels. Since the base-station operates in full-duplex mode, it can transmit and receive simultaneously in each sub-channel. In each timeslot the base-station is to properly allocate the sub-channels to the downlink or uplink of appropriate users and also determine the associated transmission power in an optimized manner. We assume that the base-station and the FD users are imperfect full-duplex nodes that suffer from self-interference. We define a self-interference cancellation coefficient to take this into account in our model and denote it by $0\leq\beta\leq1$, where $\beta=0$ indicates that SI is canceled perfectly and $\beta=1$ means no SI cancellation. For simplicity, we assume the same self-interference cancellation coefficient for BS and FD users, but consideration of different coefficients would be possible. In this paper, the goal is to maximize the weighted sum-rate of downlink and uplink users with a total power constraint at the base-station and a transmission power constraint for each user.
\begin{figure}
\includegraphics[width=\linewidth, height=4cm]{Drawing1.png}
\caption{A single cell OFDMA full-duplex network that contains an imperfect full-duplex base-station and multiple half-duplex and full-duplex mobile nodes. Due to the full-duplex nature of this network, the base-station suffers from its self interference, and the uplink nodes cause interference to their co-channel downlink nodes}
\end{figure}
We define the downlink weighted sum-rate as
\vspace{-3mm}
\begin{equation}\label{DLRateQ}
R_d=\sum_{k=1}^{K}\sum_{n\in S_{k,d}}w_k\log \bigg(1+ \frac{g_k(n)p_{k,d}(n)}{N_k+I_{k,j}(n)p_{j,u}(n)} \bigg)
\end{equation}
And the uplink weighted sum-rate as
\begin{equation}
R_u=\sum_{j=1}^{K}\sum_{n\in S_{j,u}}v_j\log \bigg(1+ \frac{g_j(n)p_{j,u}(n)}{N_0+\beta p_{k,d}(n)}\bigg)
\end{equation}
The variables used in the above equations are introduced in Table I. We assume here that the channel is reciprocal, i.e., uplink and downlink channel gains are the same. We further assume that the receiver noise powers in different sub-channels are the same. The term $I_{k,j}(n)p_{j,u}(n)$ in \eqref{DLRateQ} denotes the interference: When user $k$ is a FD device and both downlink and uplink of sub-channel $n$ are allocated to it $(j=k)$, $I_{k,j}(n)=\beta$, else $I_{k,j}(n)=g_{k,j}(n)$ is the channel gain between uplink user $j$ and downlink user $k$. We assume that the base-station knows all the channel gains, noise powers, the SI cancellation coefficient and weights assigned to the downlink and uplink of all users.
Let $P_0$ and $P_k$ denote the maximum available transmit power for the base-station and for user $k$, respectively. Then the proposed design optimization problem, denoted by \textbf{\textit{P}1}, can be formulated as follows
\begin{table}[t!]
\caption{Main Parameters and Variables}
\begin{tabular}{|c | m{24em}|}
\hline
$w_k$ & weight assigned to the downlink of user $k$ \\
\hline
$v_k$ & weight assigned to the uplink of user $k$ \\
\hline
$p_{k,d}(n)$ & transmission power from BS to user $k$ on sub-channel $n$ \\
\hline
$p_{j,u}(n)$ & transmission power from user $j$ to BS on sub-channel $n$ \\
\hline
$ N_k$ & Gaussian noise variance at the receiver of user $k$ \\
\hline
$N_0$ & Gaussian noise variance at the base-station receiver\\
\hline
$S_{k,d}$ & set of sub-channels allocated to user $k$ for downlink transmission\\
\hline
$S_{j,u}$ & set of sub-channels allocated to user $j$ for uplink transmission \\
\hline
$\beta$ & self-interference cancellation coefficient \\
\hline
$g_{k}(n)$ & channel gain between BS and user $k$ on sub-channel $n$ \\
\hline
$g_{k,j}(n)$ & channel gain between user $j$ and $k$ on sub-channel $n$ \\
\hline
$I_{k,j}(n)$ & equal to $\beta$ when $j=k$, and to $g_{k,j}(n)$ otherwise\\
\hline
$P_0$ & maximum available transmit power at BS\\
\hline
$P_k$ & maximum available transmit power at user $k$ \\
\hline
\end{tabular}
\end{table}
\begin{align}
\textbf{\textit{P}1}: \operatorname*{maximize}_{p_{k,d},p_{j,u},S_{j,u},S_{k,d}} \qquad R_d+R_u\\
\text{subject to } \sum_{k=1}^{K}\sum_{n\in S_{k,d}}p_{k,d}(n) \leq P_0\\
\sum_{n\in S_{j,u}}p_{j,u}(n) \leq P_j \quad \forall j\\
p_{j,u}(n),p_{k,d}(n)\geq 0 \quad \forall j,k,n\\
S_{i,d}\cap S_{j,d}=\phi \quad \forall i\neq j\\
S_{i,u}\cap S_{j,u}=\phi \quad \forall i\neq j\\
\cup_{j=1}^{K} \ S_{j,u} \subseteq \{1,2,...,N\} \\
\cup_{k=1}^{K} \ S_{k,d} \subseteq \{1,2,...,N\} \\
S_{k,u} \cap S_{k,d}=\phi \quad \text{if user} \ k \ \text{is} \ \text{HD}
\end{align}
where (4) and (5) indicate the power constraint on the BS and the users, respectively. Constraint (6) shows the non-negativity feature of powers; (7) and (8) come from the fact that a sub-channel cannot be allocated to two distinct users simultaneously; (9) and (10) indicate that we have no more than $N$ sub-channels, and the last constraint accounts for the half-duplex nature of the HD users.
The general resource allocation problem presented is combinatorial in nature because of the channel allocation issue and addressing it together with power allocation in an optimal manner is challenging, especially as the number of users and sub-channels grow. Moreover, the non-convexity of the rate function makes the power allocation problem itself challenging even for a fixed sub-channel assignment. Here, we invoke a two step approximate solution. First, we determine the allocation of downlink and uplink sub-channels to users and then determine the transmit power of the users and the base-station on their allocated sub-channels. In other words, we first specify the sets $S_{k,d}$ and $S_{j,u}$ and then determine the variables $p_{j,u}(n)$, $p_{k,d}(n)$. In the next Section, we introduce our sub-channel allocation algorithm.
\section{Sub-channel Allocation} \label{subchannel allocation}
The sub-channel allocation problem, denoted by \textbf{\textit{P}2}, can be formulated as follows
\begin{align*}
\textbf{\textit{P}2}: & \operatorname*{maximize}_{S_{j,u},S_{k,d}} \qquad &R_d+R_u\\
& \text{subject to} \qquad &\text{(7)-(11)}
\end{align*}
To solve the problem \textbf{\textit{P}2}, we should first solve the following power allocation problem, denoted by \textbf{\textit{P}3}, to maximize the weighted sum-rate in a single sub-channel and for a fixed pair of uplink and downlink users. Since a single sub-channel is being considered in \textbf{\textit{P}3}, we have dropped the variable $n$ in the notation.
\begin{align*}
\textbf{\textit{P}3}: \operatorname*{max}_{p_{k,d},p_{j,u}} L(p_{k,d},p_{j,u})=&w_k\log(1+\frac{g_kp_{k,d}}{N_k+I_{k,j}p_{j,u}})
\\
+&v_j\log(1+ \frac{g_jp_{j,u}}{N_0+\beta p_{k,d}})
\end{align*}
\begin{align}
&0\leq p_{k,d}\leq P_{max1} \\
&0\leq p_{j,u}\leq P_{max2}
\end{align}
Here, $P_{max1}$ and $P_{max2}$ are the maximum allowable transmit powers.
\newtheorem{prop}{Proposition}
\begin{prop}
For a fixed downlink user $k$ and uplink user $j$, the optimal pair of powers $(p_{k,d}^*,p_{j,u}^*)$ that optimizes \textbf{\textit{P}3} belongs to the following set.
{\small
\begin{align*}
\textbf{S}=\lbrace(0,P_{max2}),(P_{max1},0),(P_{max1},P_{max2}),&(p_{k,d}^a,P_{max2}), \\
&(P_{max1},p_{j,u}^a)\rbrace
\end{align*}
{\normalsize where}
\begin{align}
p_{k,d}^a=\frac{-B-\sqrt{B^2-4AC}}{2A} , p_{j,u}^a=\frac{-E-\sqrt{E^2-4DF}}{2D}
\end{align}
{\normalsize and}
\begin{align}
A&=w_k g_k \beta^2 \\
B&=2w_kN_0g_k\beta+(w_k-v_j)\beta g_kg_jp_{j,u} \\
\nonumber C&=w_kg_kN_0^2+w_kg_kg_jN_0p_{j,u}-v_jN_kg_jp_{j,u}\beta \\
&-v_jg_j\beta I_{k,j}p_{j,u}^2 \\
D&=v_j g_j I_{k,j}^2 \\
E&=2v_jN_kg_jI_{k,j}+(v_j-w_k)I_{k,j} g_kg_jp_{k,d} \\
\nonumber F&=v_jg_jN_k^2+v_jg_kg_jN_kp_{k,d}-w_kN_0g_kp_{k,d}I_{k,j} \\
&-w_kg_k\beta I_{k,j}p_{k,d}^2
\end{align}
}
\end{prop}
\begin{proof}
Computing the derivative with respect to $p_{k,d}$ and setting it to zero we have:
\begin{equation*}
\frac{\partial L}{\partial p_{k,d}}=0\Longrightarrow Ap_{k,d}^2+Bp_{k,d}+C=0
\end{equation*}
where $A$, $B$ and $C$ are defined above.
It is evident that $A\geq 0$, and if $w_k\geq v_j$ then $B\geq 0$. When $A,B\geq 0$ the above quadratic equation either has no zeros in $[0,P_{max1}]$ or has only one zero where the function changes sign from $-$ to $+$ indicating a local minimum for $L$. Therefore, in both cases the maximum is attained at a boundary point $0$ or $P_{max1}$. But when $w_k \leq v_j$, $B$ could be negative, and the smaller root of the quadratic equation $p_{k,d}^a$ could be positive. In this case, the maximum is attained at $P_{max1}$ or $p_{k,d}^a$. By similar analysis for $p_{j,u}$ one sees that if $v_j\geq w_k$ then the maximum is attained at a boundary point $0$ or $P_{max2}$ and when $w_k\geq v_j$ the maximum is attained at $P_{max2}$ or $p_{j,u}^a$.
As a result, when $B\geq0$ the optimal transmission powers belong to the
following set,
\small{
\begin{equation*}
P_{opt1}=\left\{(0,P_{max2}),(P_{max1},0),(P_{max1},P_{max2}),(P_{max1},p_{j,u}^a)\right\}.
\end{equation*}
}
\normalsize
Otherwise, if $B<0$, they belong to the set below
{\small
\begin{equation*}
P_{opt2}=\left\{(0,P_{max2}),(P_{max1},0),(P_{max1},P_{max2}),(p_{k,d}^a,P_{max2})\right\}.
\end{equation*}
}
The cases {\small$(0,p_{j,u}^a)$} and {\small$(p_{k,d}^a,0)$} cannot be the optimal solutions of \textbf{\textit{P}3} , because they are dominated by {\small$(0,P_{max2})$} and {\small$(P_{max1},0)$} which give a larger $L$.
\normalsize
Therefore, optimal powers could be found by checking the members of the set \textbf{S} and picking the one that corresponds to the largest $L$.
\end{proof}
Based on the above Proposition one can find the best uplink-downlink pair in each sub-channel by choosing the one with the largest value of $L$. This involves only $O(K^2)$ operations.
Now we can present our sub-channel allocation algorithm to solve Problem \textbf{\textit{P}2}, in which we employ a sub-optimum power allocation scheme.
First, for each sub-channel $n$, we find the best channel gain among all users and denote it by $\tilde{g}({n})=\operatorname*{arg\,max}_k g_k(n) $. Then, we sort the sub-channels based on the value of $\tilde{g}({n})$. In other words. we find a sub-channel permutation $\{a_1,...,a_N\}$ such that
$\tilde{g}({a_1}) \geq \tilde{g}({a_2}) \geq . . .\geq \tilde{g}({a_N})$. Then, starting from sub-channel $a_1$, we seek $k$ and $j$ that maximize $L$. At the first iteration, we set $P_{max1}=P_0$ , $P_{max2}=P_k$ and for iteration $l\geq 2$ set $P_{max1}= \frac{P_0}{d_0(l)}$, $ \frac{P_k}{d_k(l)}$ where $d_0(l)$ and $d_k(l)$ indicate the number of sub-channels to be allocated to the BS's downlink transmission and to user $k$'s uplink transmission, respectively, in the $l$th iteration. The proposed sub-channel allocation algorithm is summarized below.
\begin{table}[h!]
\begin{tabular}{m{30em}}
\hline
\textbf{Sub-channel Allocation Algorithm} \\
\hline\hline
1.\textbf{for} $n=1$ to $N$ \textbf{do} \\
2.\quad $\tilde{g}({n})=\operatorname*{max}_k g_k(n) $
\\
3.\textbf{end for} \\
4.Find a sub-channel permutation $\{ a_1,...,a_N \}$, $a_i\in\{1,...,N\}$, $a_i\neq a_j$ such that\\
\hspace{0.5cm} $\tilde{g}({a_1}) \geq \tilde{g}({a_2}) \geq . . .\geq \tilde{g}({a_N})$ \\
5. set $d_k(l)=1$ for $0\leq k \leq K$ and $1 \leq l \leq N$ \\
6.\textbf{for} $l=1$ to $N$ \textbf{do} \\
7.\quad Set $P_{max1}=\dfrac{P_0}{d_0(l)}$ and $P_{max2}=\dfrac{P_k}{d_k(l)} \forall k$ \\
8. \quad \textbf{for} $k=1$ to $K$ \textbf{do} \\
9. \quad \quad \textbf{for} $j=1$ to $K$ (if $k$ is an HD user $j\neq k$) \\
10.\quad \quad \quad In sub-channel $a(l)$ solve the problem \textbf{\textit{P}3} \\
11.\quad \quad \textbf{end for}
\\
12.\quad \textbf{end for}
\\
13.\quad Using the obtained optimal powers, find the best pair $(k^*,j^*)$ in the \\
\quad \quad sub-channel $a_l^*$ that has the largest value of $L$
\\
14.\quad $S_{j^*,u} \leftarrow [{S_{j^*,u},i}]$ , $S_{k^*,d} \leftarrow [{S_{k^*,d},i}]$ \\
15.\quad \textbf{if} $p_{k^*}\neq 0$ then $d_0(n)=d_0(n)+1$; \\
16.\quad \textbf{if} $p_{j^*}\neq 0$ then $d_{j^*}(n)=d_{j^*}(n)+1$; \\
17.\textbf{end for} \\
\hline
\end{tabular}
\end{table}
The complexity of finding the best user in each sub-channel is $O(K)$ and for $N$ sub-channels is $O(KN)$. Similarly, the complexity of finding the best pair in each sub-channel is $O(K^2)$ and doing so for $N$ sub-channels requires $O(NK^2)$ operations. Since the complexity of sorting $N$ values is $O(N \log N)$, then the overall computational complexity of the proposed sub-channel allocation algorithm is $O(N \log N+NK^2)$.
\section{Power Allocation} \label{power_allocation}
The power allocation problem, denoted by \textbf{\textit{P}4}, can be formulated as follows
\begin{align*}
\textbf{\textit{P}4}: & \operatorname*{maximize}_{p_{k,d},p_{j,u}} \qquad &R_d+R_u\\
& \text{subject to} \qquad &\text{(4)-(6)}
\end{align*}
Due to the interference terms, the power allocation problem is non-convex. Here, we use the ``difference of two concave
functions/sets'' (DC) programming technique \cite{tuy2013convex} to convexify this problem. In this procedure, the non-concave
objective function is expressed as the difference of two concave
functions, and the discounted term is approximated by its first order
Taylor series. Hence, the objective becomes concave and can
be maximized by known convex optimization methods.
This procedure runs iteratively, and after each iteration the
optimal solution serves as an initial point for the next iteration
until the improvement diminishes in iterations. In \cite{kha2012fast}, the DC approach is used to formulate optimized power allocation in a multiuser interference channel. Here, we rewrite the objective function of \textbf{\textit{P}4} in DC form as follows
\vspace{-2mm}
\begin{align*}
\operatorname*{max}_{\mathbf{p}} \quad
f(\mathbf{p})-h(\mathbf{p})
\end{align*}
\vspace{-9mm}
\small{
\begin{align*}
f(\mathbf{p})&=\sum_{k=1}^{K}\sum_{n\in S_{k,d}}w_k\log(N_k+I_{k,j}(n)p_{j,u}(n)+ g_k(n)p_{k,d}(n)) \\
&+\sum_{j=1}^{K}\sum_{n\in S_{j,u}}v_j\log(N_0+\beta p_{k,d}(n)+ g_j(n)p_{j,u}(n))
\end{align*}
\vspace{-7mm}
\begin{align*}
h(\mathbf{p})&=\sum_{k=1}^{K}\sum_{n\in S_{k,d}}w_k\log(N_k+I_{k,j}(n)p_{j,u}(n))\\
&+\sum_{j=1}^{K}\sum_{n\in S_{j,u}}v_j\log(N_0+\beta p_{k,d}(n))
\end{align*}
}
\normalsize
where
\begin{align*}
\mathbf{p}=[p_{k_1,d}(1),...,p_{k_N,d}(N),p_{j_1,u}(1),,...,p_{j_N,u}(N)]^T
\end{align*} is the downlink and uplink transmitted power vector, and $k_i$ and $j_i$ denote the uplink and downlink users that has been selected for the $i$th sub-channel after the sub-channel allocation phase. Now, the objective $f(\mathbf{p})-h(\mathbf{p})$ is a DC function. To write the Taylor series of the discounted function $h(\mathbf{p}$), we need its gradient, that can be easily derived as follows.
\small{
\begin{align*}
&\nabla h(\mathbf{p})= \bigg[\frac{u_{j_1}\beta}{\ln(2)}\frac{1}{N_0+\beta p_{k_1,d}(1))},. . . ,\frac{u_{j_N}\beta}{\ln(2)}\frac{N}{N_0+\beta p_{k_N,d}(N))}, \\ &\frac{w_{k_1}I_{k_1,j_1}(1)}{\ln(2)}\frac{1}{N_{k_1}+I_{k_1,j_1}(1)p_{j_1,u}(1))},..., \\
&\frac{w_{k_N}I_{k_N,j_N}(N)}{\ln(2)}\frac{1}{N_{k_N}+I_{k_N,j_N}(N)p_{j_N,u}(N))} \bigg] ^T
\end{align*}
}
\normalsize
To make the problem convex, $h(\mathbf{p})$ is approximated with its first order approximation $h(\mathbf{p}^{(t)})+\nabla h^T(\mathbf{p}^{(t)})(\mathbf{p}-\mathbf{p}^{(t)})$ at point $\mathbf{p}^{(t)}$. We start from a feasible $\mathbf{p}^{(0)}$ at the first iteration, and $\mathbf{p}^{(t+1)}$ at the $t$th iteration is generated as the optimal solution of the following convex program
\begin{align*}
\mathbf{p}^{(t+1)}=\operatorname*{arg max}_{\mathbf{p}} \quad &f(\mathbf{p})- h(\mathbf{p}^{(t)})-\nabla h^T(\mathbf{p}^{(t)})(\mathbf{p}-\mathbf{p}^{(t)}) \\
&\text{subject to } \ (4)-(6)
\end{align*}
Since $h(\mathbf{p})$ is a concave function, its gradient is also its super gradient so we have
\begin{align*}
h(\mathbf{p})\leq h(\mathbf{p}^{(t)})+\nabla h^T(\mathbf{p}^{(t)})(\mathbf{p}-\mathbf{p}^{(t)})
\end{align*}
and we can deduce
\begin{align*}
h(\mathbf{p}^{(t+1)})\leq h(\mathbf{p}^{(t)})+\nabla h^T(\mathbf{p}^{(t)})(\mathbf{p}^{(t+1)}-\mathbf{p}^{(t)}).
\end{align*}
Then it can be proved that in each iteration the solution of problem \textbf{\textit{P}4} is improved as follows
\begin{align*}
&f(\mathbf{p}^{(t+1)})- h(\mathbf{p}^{(t+1)}) \geq \\ &f(\mathbf{p}^{(t+1)})- h(\mathbf{p}^{(t)})-\nabla h^T(\mathbf{p}^{(t)})(\mathbf{p}^{(t+1)}-\mathbf{p}^{(t)}) \\
&=\operatorname*{max}_{\mathbf{p}} \quad f(\mathbf{p})- h(\mathbf{p}^{(t)})-\nabla h^T(\mathbf{p}^{(t)})(\mathbf{p}-\mathbf{p}^{(t)}) \\
& \geq f(\mathbf{p}^{(t)})- h(\mathbf{p}^{(t)})-\nabla h^T(\mathbf{p}^{(t)})(\mathbf{p}^{(t)}-\mathbf{p}^{(t)}) \\
&=f(\mathbf{p}^{(t)})- h(\mathbf{p}^{(t)}).
\end{align*}
According to the above equations, the objective value after each iteration is either unchanged or improved and since the constraint set is compact it can be concluded that the above DC approach converges to a local maximum.
\section{Simulation Results}
\label{simulation_results}
In this Section, we evaluate the proposed resource allocation scheme for OFDMA networks with half-duplex and imperfect full-duplex nodes in indoor and outdoor scenarios. We assume a time-slotted system, where nodes are uniformly distributed within a given cell radius. Table II presents the details of the indoor and outdoor simulation setup and channel models. In addition to the pathloss, a Rayleigh block fading channel model with unit average power is considered. The channel gains remain constant in each time slot and vary independently from one time slot to the next.
For comparison, we consider six schemes: (i) An HD uplink system (HD-U), (ii) An HD downlink system (HD-D), (iii) a system that includes an FD BS and HD users (FD-HD), (iv) a system that contains an FD BS and FD users (FD-FD), (v) an upper bound which is HD uplink rate plus HD downlink rate; (vi) a Hybrid HD scheme (HHD), in which a hybrid HD BS could transmit data to downlink users and receive data from uplink users simultaneously in different sub-channels.
For the HD-D case, each sub-channel is allocated to the user with the best weighted channel SNR, and multi-level water filling \cite{seong2006optimal} is applied for power allocation. For the sub-channel assignment of the HD-U scheme the $SOA1 \: 4B \: 5A$ method presented in \cite{huang2009joint} is used, and for power allocation each user performs water filling in its dedicated sub-channels. In the HHD scheme, we use the proposed sub-channel allocation algorithm by changing the set $P_{opt}$ to:
\begin{align*}
P_{opt}=\left\{(0,P_{max2}),(P_{max1},0)\right\}
\end{align*}
and perform multi-level water filling and water filling for the power allocation in the selected downlink and uplink sub-channels, respectively.
Fig. \ref{fig:convergence} illustrates the convergence of the proposed resource allocation scheme in an OFDMA network with 10 HD nodes and 10 imperfect FD nodes. As can be seen, the sum-rate converges in just a few iterations.
Fig. \ref{fig:Compare_Optimal} compares the proposed algorithm with the optimal exhaustive search solution. Due to the high computational complexity of exhaustive search, only a small network with one HD and one FD user and a small number of subchannels can be considered. Uplink and downlink weight vectors are assumed to be $\mathbf{u}=[1/3, 2/3]^{T}$ and $\mathbf{w}=[2/3, 1/3]^{T}$ respectively, and the SI cancellation coefficient is set to $\beta = -90$ dB. Simulation results show that, at least for small size networks, our proposed algorithm achieves the performance of the optimal exhaustive search.
\begin{figure}
\includegraphics[width=\linewidth, height=4cm]{convergence.eps}
\caption{Convergence behavior of the proposed algorithm. Here $\beta=10^{-6}$ and other simulation parameters are the same as in the outdoor case}
\label{fig:convergence}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\linewidth, height=4cm]{Compare_Optimal.eps}
\caption{Performance comparison of the proposed algorithm with optimal exhaustive search in a small network}
\label{fig:Compare_Optimal}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\linewidth, height=4cm]{all_outdoor2.eps}
\caption{Performance comparison of six schemes in FD and HD networks in the outdoor scenario.}
\label{fig:all_outdoor}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\linewidth, height=4cm]{all_indoor2.eps}
\caption{Performance comparison of six schemes in FD and HD networks in the indoor scenario.}
\label{fig:all_indoor}
\end{figure}
\begin{table}[h!]
\caption{simulation parameters}
\begin{tabular}{|m{14em} | m{15em}|}
\hline
\textbf{PARAMETER} & \textbf{VALUE} \\
\hline
Maximum BS Power (outdoor) & $43$ dBm \\
\hline
Maximum BS Power (indoor) & $24$ dBm \\
\hline
Maximum UE Power $(P_k)$ & $23$ dBm \\
\hline
Thermal Noise Density & $-170$ dBm/Hz\\
\hline
Number of Sub-channels $N$ & $64$\\
\hline
Total Bandwidth & $10$ MHz \\
\hline
Sub-channel Bandwidth & $150$ KHz \\
\hline
Cell Radius (outdoor) & $1$ km\\
\hline
Cell Radius (indoor) & $20$ m\\
\hline
Center Frequency & $2$ GHz\\
\hline
BS to UE Pathloss (outdoor) & urban Hata model with parameters $h_m= 1.5$ m, $h_B= 30$ m \\
\hline
UE to UE Pathloss (outdoor) & urban Hata model with parameters $ h_m= 1.5$ m, $h_B= 1.5$ m \\
\hline
Pathloss Model (indoor) & ITU model for indoor attenuation with parameters N= $22$, $p_f(n)= 9$ \\
\hline
\end{tabular}
\end{table}
Fig. \ref{fig:all_outdoor} shows the sum-rate of the different schemes in the outdoor scenario with perfect SI cancellation ($\beta=0$). It can be seen that when the BS and all nodes are perfect FD devices the upper-bound could be attained, and when the nodes are HD but the BS is FD the sum-rate is still bigger than the cases with HD BS, but it can not reach the upper-bound because of inter-node interference.
Fig. \ref{fig:all_indoor} shows the sum-rate of the six presented schemes in an indoor scenario. If we compare the outdoor
and indoor scenarios we find that using an FD BS in an outdoor environment has much larger gain than using it in an indoor case. This result is intuitive because in the outdoor environment the distances between nodes are larger, and hence the inter-node interference is smaller. As a result, the FD BS could work in FD mode in more sub-channels, which helps increase the network throughput more significantly.
\begin{figure}[t!]
\includegraphics[width=\linewidth, height=4cm]{beta_threshold_indoor.eps}
\caption{Effect of the self interference cancellation coefficient on the FD network capacity in the indoor scenario.}
\label{fig:beta_threshold_indoor}
\end{figure}
Fig. \ref{fig:beta_threshold_indoor} compares the sum-rates of an FD-FD network and an FD-HD network for different values of $\beta$ in the indoor scenario. It can be seen that when $\beta$ is larger than a specified threshold, which is near $-90$ dB, there is no difference between the sum-rate of the all HD user case and the sum-rate of the all FD user case. The reason is that when $\beta$ is large relative to the inter-node interference, FD users prefer to work in HD mode in order to increase their rate, hence the sum-rates of FD-FD and FD-HD become equal.
In Fig. \ref{fig:beta_threshold_outdoor}, the same experiment is repeated for the outdoor scenario. Here the threshold $\beta$ is approximately $-110$ dB which is much smaller than in the indoor case. Since the inter-node interference in the outdoor environment is smaller, the SI cancellation coefficient should be very small to make the FD users work in the FD mode. Therefore, in the outdoor scenario a full-duplex user should be almost a perfect FD in order to be allowed to work in the FD mode.
Fig. \ref{fig:FDpercentage} shows the performance of a full-duplex OFDMA network with a mix of FD and HD users. A total of 20 users are considered, assuming perfect SI cancellation for FD devices. It can be seen that increasing the percentage of FD users in an outdoor environment does not increase the total sum-rate significantly, but in the indoor case by equipping only $10\%$ of the nodes with FD technology the network throughput greatly increases. The reason behind this is the large inter-node interference in the indoor environment that could be avoided by using FD users instead of HD ones.
\begin{figure}[t!]
\includegraphics[width=\linewidth, height=4cm]{beta_threshold_outdoor.eps}
\caption{Effect of the self interference cancellation coefficient on the FD network capacity in the outdoor scenario}
\label{fig:beta_threshold_outdoor}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\linewidth, height=4cm]{FDpercentage.eps}
\caption{Effect of the fraction of FD nodes on the network throughput in both indoor and outdoor scenarios}
\label{fig:FDpercentage}
\end{figure}
\section{conclusion}
\label{conclusion}
In order to fully exploit the advantages of FD technology in wireless networks, it is important to design an appropriate resource allocation algorithm. In this paper we considered a single cell OFDMA network that contains an FD BS and a mixture of HD and FD users, and also assumed that FD nodes are not necessarily perfect FD devices and may suffer from residual self-interference. For this model, we proposed a sub-channel allocation algorithm and power allocation method and showed that when all users and the BS are perfect FD we can double the capacity. Otherwise, because of inter-node interference and self-interference the spectral efficiency gain is smaller, but we showed that even by using an imperfect FD BS in a network, the total capacity could increase significantly. We also investigated FD operation in both outdoor and indoor scenarios and studied the effect of the self interference cancellation coefficient and of the percentage of FD users.
\bibliographystyle{IEEEtran}
{\small
|
\section{Introduction}
In the \emph{tree compatibility problem}, we are given a collection $\P = \{\ensuremath{\mathcal{T}}_1, \ldots, \ensuremath{\mathcal{T}}_k\}$ of rooted phylogenetic trees with partially overlapping taxon sets. $\P$ is called a \emph{profile} and the trees in $\P$ are the \emph{input trees}. The question is whether there exists a tree $\ensuremath{\mathcal{T}}$ whose taxon set is the union of the taxon sets of the input trees, such that $\ensuremath{\mathcal{T}}$ exhibits the clusterings implied by the input trees. That is, if two taxa are together in a subtree of some input tree, then they must also be together in some subtree of $\ensuremath{\mathcal{T}}$.
The tree compatibility problem has been studied for over three decades \cite{Aho81a,DengFB2016,HenzingerKingWarnow99,Steel92}.
In the original version of the tree compatibility problem, only the leaves of the input trees are labeled.
Here we study a generalization, called \emph{ancestral compatibility}, in which taxa may be \emph{nested}. That is, the internal nodes may also be labeled; these labels represent \emph{higher-order taxa}, which are, in effect, sets of taxa. Thus, for example,
an input tree may contain the taxon \emph{Glycine max} (soybean) nested within a subtree whose root is labeled Fabaceae (the legumes), itself nested within an Angiosperm subtree. Note that leaves themselves may be labeled by higher-order taxa. The question now is whether there is a tree $\ensuremath{\mathcal{T}}$ whose taxon set is the union of the taxon sets of the input trees, such that $\ensuremath{\mathcal{T}}$ exhibits not only the clusterings among the taxa, but also the ancestor/descendant relationships among taxa in the input trees. Our main result is a $\tilde{O}(\MP)$ algorithm for the compatibility problem for trees with nested taxa, where $\MP$ is the total number of nodes and edges in the trees in $\P$.
\vspace{-1.5\parsep}
\paragraph{Background.} The tree compatibility problem is a basic special case of the \emph{supertree problem}.
A supertree method is a way to synthesize a collection of phylogenetic trees with partially overlapping taxon sets into a single supertree that represents the information in the input trees. The supertree approach, proposed
in the early 90s \cite{Baum:1992,Ragan:1992}, has been used successfully to build large-scale phylogenies \cite{BinindaEmonds:Nature:07}.
The original supertree methods were limited to input trees where only the leaves are labeled.
Page \cite{Page2004} was among the first to note the need to handle phylogenies where internal nodes are labeled, and taxa are nested. A major motivation is the desire to incorporate \emph{taxonomies} as input trees in large-scale supertree analyses, as way to circumvent one of the obstacles to building comprehensive phylogenies: the limited taxonomic overlap among different phylogenetic studies \cite{Sanderson:2008}. Taxonomies group organisms according to a system of taxonomic rank (e.g., family, genus, and species); two examples are the NCBI taxonomy \cite{NCBI2009} and the Angiosperm taxonomy \cite{APG2016}. Taxonomies spanning a broad range of taxa provide structure and completeness that might be hard to obtain otherwise.
A recent example of the utility of taxonomies is the Open Tree of Life, a draft phylogeny for over 2.3 million species \cite{HinchliffPNAS2015}.
Taxonomies are not, strictly speaking, phylogenies. In particular, their internal nodes and some of their leaves are labeled with higher-order taxa. Nevertheless, taxonomies have many of the same mathematical characteristics as phylogenies. Indeed, both phylogenies and taxonomies are \emph{semi-labeled trees}
\cite{BordewichEvansSemple2006,SempleSteel03}. We will use this term throughout the rest of the paper to refer to trees with nested taxa.
The fastest previous algorithm for testing ancestral compatibility, based on earlier work by Daniel and Semple \cite{DanielSemple2004}, is due to Berry and Semple \cite{BerrySemple2006}. Their algorithm runs in $O \left (\log^2 n \cdot \tau_\P \right)$ time using $O \left ( \tau_\P \right)$ space. Here, $n$ is the number of distinct taxa in $\P$ and $\tau_\P = \sum_{i = 1}^k \sum_{v \in I(\ensuremath{\mathcal{T}}_i)} d(v)^2$, where $I(\ensuremath{\mathcal{T}}_i)$ is the set of internal nodes of $\ensuremath{\mathcal{T}}_i$, for each $i \in \{1, \dots , k\}$, and $d(v)$ is the degree of node $v$. While the algorithm is polynomial, its dependence on node degrees is problematic: semi-labeled trees can be highly unresolved (i.e., contain nodes of high degree), especially if they are taxonomies.
\vspace{-1.5\parsep}
\paragraph{Our contributions.}
The $\tilde{O}(\MP)$ running time of our ancestral compatibility algorithm is independent of the degrees of the nodes of the input trees, a valuable characteristic for large datasets that include taxonomies. To achieve this time bound, we extend ideas from our recent algorithm for testing the compatibility of ordinary phylogenetic trees \cite{DengFB2016}. As in that algorithm, a central notion in the current paper is the \emph{display graph} of profile $\P$, denoted $\DG$. This is the graph obtained from the disjoint union of the trees in $\P$ by identifying nodes that have the same label (see Section \ref{sec:testAC}).
The term ``display graph'' was introduced by Bryant and Lagergren \cite{BryantLagergren06}, but similar ideas have been used elsewhere. In particular, the display graph is closely related to Berry and Semple's \emph{restricted descendancy graph} \cite{BerrySemple2006}, a mixed graph whose directed edges correspond to the (undirected) edges of $\DG$ and whose undirected edges have no correspondence in $\DG$. The second kind of edges are the major component of the $\tau_\P$ term in the time and space complexity of Berry and Semple's algorithm. The absence of such edges makes $\DG$ significantly smaller than the restricted descendancy graph. Display graphs also bear some relation to \emph{tree alignment graphs}
\cite{Smith:PloSCB:2013}.
Here, we exploit the display graph more extensively and more directly than our previous work. Although the display graph of a collection of semi-labeled trees is more complex than that of a collection of ordinary phylogenies, we are able to extend several of the key ideas --- notably, that of a semi-universal label --- to the general setting of semi-labeled trees. As in \cite{DengFB2016}, the implementation relies on a dynamic graph data structure, but it requires a more careful amortized analysis based on a weighing scheme.
\vspace{-1.5\parsep}
\paragraph{Contents.} Section \ref{sec:prelims} presents basic definitions regarding semi-labeled trees and ancestral compatibility. . Section \ref{sec:dispGraph} introduces the display graph and discusses its properties. Section \ref{sec:testAC} presents \ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees, our algorithm for testing ancestral compatibility. Section \ref{sec:implementation} gives the implementation details for \ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees. Section \ref{sec:discussion} gives some concluding remarks.
\section{Preliminaries}\label{sec:prelims}
For each positive integer $r$, $[r]$ denotes the set $\{1, \dots , r\}$.
Let $G$ be a graph. $V(G)$ and $E(G)$ denote the node and edge sets of $G$. The \emph{degree} of a node $v \in V(G)$ is the number of edges incident on $v$. A \emph{tree} is an acyclic connected graph.
In this paper, all trees are assumed to be rooted. For a tree $T$, $r(T)$ denotes the root of $T$.
Suppose $u, v \in V(T)$. Then, $u$ is an \emph{ancestor} of $v$ in $T$, denoted $u \le_T v$, if $u$ lies on the path from $v$ to $r(T)$ in $T$. If $u \le_T v$, then $v$ is a \emph{descendant} of $u$. Node %
$u$ is a \emph{proper descendant} of $v$ if $u$ is a descendant of $v$ and $v \neq u$. If $\{u,v\} \in E(T)$ and $u \le_T v$, then $u$ is the \emph{parent} of $v$ and $v$ is a \emph{child} of $u$. If neither $u \le_T v$ nor $v \le_T u$ hold, then we write $u \parallel_T v$ and say that $u$ and $v$ are \emph{not comparable} in $T$.
\vspace{-1.5\parsep}
\paragraph{Semi-labeled trees.} A \emph{semi-labeled tree} is a pair $\ensuremath{\mathcal{T}} = (T,\phi)$ where $T$ is a tree and $\phi$ is a mapping from a set $L(\ensuremath{\mathcal{T}})$ to $V(T)$ such that, for every node $v \in V(T)$ of degree at most two, $v \in \phi(L(\ensuremath{\mathcal{T}}))$. $L(\ensuremath{\mathcal{T}})$ is the \emph{label set} of $\ensuremath{\mathcal{T}}$ and $\phi$ is the \emph{labeling function} of $\ensuremath{\mathcal{T}}$.
For every node $v \in V(T)$, $\phi^{-1}(v)$ denotes the (possibly empty) subset of $L(\ensuremath{\mathcal{T}})$ whose elements map into $v$; these elements as the \emph{labels of $v$} (thus, each label is a taxon). If $\phi^{-1}(v) \neq \emptyset$, then $v$ is \emph{labeled}; otherwise, $v$ is \emph{unlabeled}.
Note that, by definition, every leaf in a semi-labeled tree is labeled. Further, any node, including the root, that has a single child must be labeled. Nodes with two or more children may be labeled or unlabeled. A semi-labeled tree $\ensuremath{\mathcal{T}} = (T,\phi)$ is \emph{singularly labeled} if every node in $T$ has
at most one label; $\ensuremath{\mathcal{T}}$ is \emph{fully labeled} if every node in $T$ is labeled.
Semi-labeled trees, also known as \emph{$X$-trees}, generalize ordinary phylogenetic trees, also known as \emph{phylogenetic $X$-trees} \cite{SempleSteel03}. An ordinary phylogenetic tree is a semi-labeled tree $\ensuremath{\mathcal{T}} = (T,\phi)$ where $r(T)$ has degree at least two and $\phi$ is a bijection from $L(\ensuremath{\mathcal{T}})$ into leaf set of $T$ (thus, internal nodes are not labeled).
Let $\ensuremath{\mathcal{T}} = (T,\phi)$ be a semi-labeled tree and let $\ell$ and $\ell'$ be two labels in $L(\ensuremath{\mathcal{T}})$.
If $\phi(\ell) \le_T \phi(\ell')$, then we write $\ell \le_\ensuremath{\mathcal{T}} \ell'$, and say that $\ell'$ is a \emph{descendant} of $\ell$ in $\ensuremath{\mathcal{T}}$ and that $\ell$ is an \emph{ancestor} of $\ell'$. We write $\ell <_\ensuremath{\mathcal{T}} \ell'$ if $\phi(\ell')$ is a proper descendant of $\phi(\ell)$. If $\phi(\ell) \parallel_T \phi(\ell')$, then we write $\ell \parallel_\ensuremath{\mathcal{T}} \ell'$ and say that $\ell$ and $\ell'$ are \emph{not comparable} in $\ensuremath{\mathcal{T}}$. If $\ensuremath{\mathcal{T}}$ is fully labeled and
$\phi(\ell)$ is the parent of $\phi(\ell')$ in $T$, then $\ell$ is the \emph{parent} of $\ell'$ in $\ensuremath{\mathcal{T}}$ and $\ell'$ is a \emph{child} of $\ell$ in $\ensuremath{\mathcal{T}}$; two labels with the same parent are \emph{siblings}.
Two semi-labelled trees $\ensuremath{\mathcal{T}} = (T,\phi)$ and $\ensuremath{\mathcal{T}}' = (T', \phi')$ are \emph{isomorphic} if there exists a bijection $\psi : V(T) \rightarrow V(T')$ such that $\phi' = \psi \circ \phi$ and, for any two nodes $u, v \in V(T)$, $(u,v) \in E(T)$ if and only $(\psi(u), \psi(v)) \in E(T')$.
Let $\ensuremath{\mathcal{T}} = (T,\phi)$ be a semi-labeled tree. For each $u \in V(T)$, $X(u)$ denotes the set of all labels in the subtree of $T$ rooted at $u$; that is, $X(u) = \bigcup_{v: u \le_T v} \phi^{-1}(v)$. $X(u)$ is called a \emph{cluster} of $T$.
$\Cl(\ensuremath{\mathcal{T}})$ denotes the set of all clusters of $\ensuremath{\mathcal{T}}$.
It is well known \cite[Theorem 3.5.2]{SempleSteel03} that a semi-labeled tree $\ensuremath{\mathcal{T}}$ is completely determined by $\Cl(\ensuremath{\mathcal{T}} )$. That is, if $\Cl(\ensuremath{\mathcal{T}}) = \Cl(\ensuremath{\mathcal{T}}')$ for some other semi-labeled tree $\ensuremath{\mathcal{T}}'$, then $\ensuremath{\mathcal{T}}$ is isomorphic to $\ensuremath{\mathcal{T}}'$.
Suppose $A \subseteq L(\ensuremath{\mathcal{T}})$ for a semi-labeled tree $\ensuremath{\mathcal{T}} = (T,\phi)$. The \emph{restriction} of $\ensuremath{\mathcal{T}}$ to $A$, denoted $\ensuremath{\mathcal{T}} |A$, is the semi-labeled tree whose cluster set is
$\Cl(\ensuremath{\mathcal{T}} | A) = \{X \cap A : X \in \Cl(\ensuremath{\mathcal{T}}) \text{ and } X \cap A \neq \emptyset \}.$
Intuitively, $\ensuremath{\mathcal{T}} | A$ is obtained from the minimal rooted subtree of $T$ that connects the nodes in $\phi(A)$ by suppressing all vertices of degree two that are not in $\phi(A)$.
Let $\ensuremath{\mathcal{T}} = (T,\phi)$ and $\ensuremath{\mathcal{T}}' = (T', \phi')$ be semi-labeled trees such that $L(\ensuremath{\mathcal{T}}') \subseteq L(\ensuremath{\mathcal{T}})$. $\ensuremath{\mathcal{T}}$ \emph{ancestrally displays} $\ensuremath{\mathcal{T}}'$ if $\Cl(\ensuremath{\mathcal{T}}') \subseteq \Cl(\ensuremath{\mathcal{T}}|L(\ensuremath{\mathcal{T}}'))$. Equivalently, $\ensuremath{\mathcal{T}}$ ancestrally displays $\ensuremath{\mathcal{T}}'$ if $\ensuremath{\mathcal{T}}'$ can be obtained from $\ensuremath{\mathcal{T}} | L(\ensuremath{\mathcal{T}}')$ by contracting edges, and, for any $\ell_1, \ell_2 \in L(\ensuremath{\mathcal{T}}')$,
(i) if $\ell_1 <_\mathcal{\ensuremath{\mathcal{T}}'} \ell_2$, then $\ell_1 <_\ensuremath{\mathcal{T}} \ell_2$, and
(ii)
if $\ell_1 \parallel_\mathcal{\ensuremath{\mathcal{T}}'} \ell_2$, then $\ell_1 \parallel_\mathcal{\ensuremath{\mathcal{T}}} \ell_2$.
The notion of ``ancestrally displays'' for semi-labeled trees generalizes the well-known notion of ``displays'' for ordinary phylogenetic trees \cite{SempleSteel03}.
For a semi-labelled tree $\ensuremath{\mathcal{T}}$, let
\begin{equation}
D(\ensuremath{\mathcal{T}}) = \{(\ell,\ell'): \ell, \ell' \in L(\ensuremath{\mathcal{T}}) \text{ and } \ell <_\ensuremath{\mathcal{T}} \ell'\} \quad \text{ and } \quad
N(\ensuremath{\mathcal{T}} ) = \{\{\ell,\ell'\} : \ell, \ell' \in L(\ensuremath{\mathcal{T}}) \text{ and } \ell \parallel_\ensuremath{\mathcal{T}} \ell'\}.
\end{equation}
Note that $D(\ensuremath{\mathcal{T}})$ consists of \emph{ordered} pairs, while $N(\ensuremath{\mathcal{T}})$ consists of \emph{unordered} pairs.
\begin{lemma}[Bordewich et al.\ \cite{BordewichEvansSemple2006}]\label{lm:DN}
Let $\ensuremath{\mathcal{T}}$ and $\ensuremath{\mathcal{T}}'$ be semi-labelled trees such that $L(\ensuremath{\mathcal{T}}') \subseteq L(\ensuremath{\mathcal{T}})$. Then $\ensuremath{\mathcal{T}}$ ancestrally displays $\ensuremath{\mathcal{T}}'$ if and only if $D(\ensuremath{\mathcal{T}}') \subseteq D(\ensuremath{\mathcal{T}})$ and $N(\ensuremath{\mathcal{T}}') \subseteq N(\ensuremath{\mathcal{T}})$.
\end{lemma}
\vspace{-1.5\parsep}
\paragraph{Profiles and ancestral compatibility.}
Throughout the rest of this paper $\P = \{\ensuremath{\mathcal{T}}_1, \dots, \ensuremath{\mathcal{T}}_k\}$ denotes a set where, for each $i \in [k]$, $\ensuremath{\mathcal{T}}_i = (T_i, \phi_i)$ is a semi-labeled tree. We refer to $\P$ as a \emph{profile}, and write $L(\P)$ to denote $\bigcup_{i\in[k]} L(\ensuremath{\mathcal{T}}_i)$, the \emph{label set} of $\P$. Figure~\ref{fig:profile} shows a profile where $L(\P) = \{a,b,c,d,e,f,g,h,i\}$.
We write $V(\P)$ for $\bigcup_{i\in[k]} V(T_i)$ and $E(\P)$ for $\bigcup_{i\in[k]} E(T_i)$,
The \emph{size} of $\P$ is $\MP = |V(\P)| + |E(\P)|$.
\begin{figure
\centering
\begin{minipage}[b]{0.46\linewidth}
\includegraphics[scale=0.25]{profile.pdf}
\caption{A profile $\P= \{\ensuremath{\mathcal{T}}_1, \ensuremath{\mathcal{T}}_2, \ensuremath{\mathcal{T}}_3\}$ --- trees are ordered left-to-right.
The letters are the original labels; grey numbers are labels added to make the trees fully labeled. (Adapted from \cite{BerrySemple2006}.)}
\label{fig:profile}
\end{minipage}
\quad
\begin{minipage}[b]{0.24\linewidth}
\includegraphics[scale=0.25]{supertree.pdf}
\caption{A tree $\ensuremath{\mathcal{T}}$ that ancestrally displays the profile of Figure \ref{fig:profile}. (Adapted from \cite{BerrySemple2006}.)}
\label{fig:supertree}
\end{minipage}
\quad
\begin{minipage}[b]{0.23\linewidth}
\includegraphics[scale=0.25]{displayGraph.pdf}
\caption{The display graph $\DG$ for the profile of Figure \ref{fig:profile}.}
\label{fig:displayGraph}
\end{minipage}
\end{figure}
$\P$ is \emph{ancestrally compatible} if there is a rooted semi-labeled tree $\ensuremath{\mathcal{T}}$ that ancestrally displays each of the trees in $\P$. If $\ensuremath{\mathcal{T}}$ exists, we say that $\ensuremath{\mathcal{T}}$ \emph{ancestrally displays} $\P$ (see Figure \ref{fig:supertree}).
Given a subset $X$ of $L(\P)$, the \emph{restriction} of $\P$ to $X$, denoted $\P|X$, is the profile $\{\ensuremath{\mathcal{T}}_1|X \cap L(\ensuremath{\mathcal{T}}_1), \dots, \ensuremath{\mathcal{T}}_k|X \cap L(\ensuremath{\mathcal{T}}_k)\}$. The proof of the following lemma is straightforward.
\begin{lemma}\label{lm:compatSubprofile}
Suppose $\P$ is ancestrally compatible and let $\ensuremath{\mathcal{T}}$ be a tree that ancestrally displays $\P$. Then, for any $X \subseteq L(\P)$, $\ensuremath{\mathcal{T}} | X$ ancestrally displays $\P| X$.
\end{lemma}
A semi-labeled tree $\ensuremath{\mathcal{T}} = (T, \phi)$ is \emph{fully labeled} if
every node in $T$ is labeled. Suppose $\P$ contains trees that are not fully labeled. We can convert $\P$ into an equivalent profile $\P'$ of fully-labeled trees as follows. For each $i \in [k]$, let $l_i$ be the number of unlabeled nodes in $T_i$. Create a set $L'$ of $n' = \sum_{i \in [k]} l_i$ labels such that $L' \cap L(\P) = \emptyset$. For each $i \in [k]$ and each $v \in V(T_i)$ such that $\phi_i^{-1}(v) = \emptyset$, make $\phi_i^{-1}(v) = \{\ell\}$, where $\ell$ is a distinct element from $L'$.
We refer to $\P'$ as the \emph{profile obtained by adding distinct new labels to $\P$} (see Figure \ref{fig:profile}).
\begin{lemma}[Daniel and Semple \cite{DanielSemple2004}]\label{lm:fullyL}
Let $\P'$ be the profile obtained by adding distinct new labels to $\P$. Then, $\P$ is ancestrally compatible if and only if $\P'$ is ancestrally compatible. Further, if $\ensuremath{\mathcal{T}}$ is a semi-labeled phylogenetic tree that ancestrally displays $\P'$, then $\ensuremath{\mathcal{T}}$ ancestrally displays $\P$.
\end{lemma}
From this point forward, we shall assume that,
for each $i \in [k]$, $\ensuremath{\mathcal{T}}_i$ is fully and singularly labeled.
By Lemma \ref{lm:fullyL}, no generality is lost in assuming that all trees in $\P$ are fully labeled.
The assumption that the trees are singularly labeled is inessential; it is only for clarity. Note that, even with the latter assumption, a tree that ancestrally displays $\P$ is not necessarily singularly labeled. Figure \ref{fig:supertree} illustrates this fact.
\section{The Display Graph}\label{sec:dispGraph}
The \emph{display graph} of a profile $\P$, denoted $\DG$, is the graph obtained from the disjoint union of the underlying trees $T_1, \dots, T_k$ by identifying nodes that have the same label. Multiple edges between the same pair of nodes are replaced by a single edge. See Figure~\ref{fig:displayGraph}.
$\DG$ has $O(\MP)$ nodes and edges, and can be constructed in $O(\MP)$ time. By our assumption
that all the trees in $\P$ are fully and singularly labeled,
there is a bijection between the labels in $L(\P)$ and the nodes of $\DG$. Thus, from this point forward, we refer to the nodes of $\DG$ by their labels.
It is easy to see that if $\DG$ is not connected, then $\P$ decomposes into label-disjoint sub-profiles, and that $\P$ is compatible if and only if each sub-profile is compatible. Thus, we shall assume, without loss of generality, that $\DG$ is connected.
\vspace{-1.5\parsep}
\paragraph{Positions.} A \emph{position} (for $\P$) is a vector $U = (U(1), \dots, U(k))$, where $U(i) \subseteq L(\ensuremath{\mathcal{T}}_i)$, for each $i \in [k]$. Since labels may be shared among trees, we may have $U(i) \cap U(j) \neq \emptyset$, for $i, j \in [k]$ with $i \neq j$.
For each $i \in [k]$, let $\Desc_i(U) = \{\ell : \ell' \le_{\ensuremath{\mathcal{T}}_i} \ell, \text{ for some } \ell' \in U(i)\}$, and let $\Desc_\P(U) = \bigcup_{i \in [k]} \Desc_i(U)$.
A position $U$ is \emph{valid} if, for each $i \in [k]$,
\vspace{-0.1cm}
\begin{enumerate}[(V1)]
\vspace{-0.5\parskip}
\itemsep1pt \parskip0pt \parsep0pt
\item\label{item:v1}
if $|U(i)| \ge 2$, then the elements of $U(i)$ are siblings in $\ensuremath{\mathcal{T}}_i$ and
\item\label{item:v2}
$\Desc_i(U) = \Desc_\P(U) \cap L(\ensuremath{\mathcal{T}}_i)$.
\end{enumerate}
\vspace{-1\parsep}
\begin{lemma}\label{lem:V2}
For any valid position $U$, $\P | \Desc_\P(U) = \{\ensuremath{\mathcal{T}}_1|\Desc_1(U), \dots, \ensuremath{\mathcal{T}}_k|\Desc_k(U)\}$.
\end{lemma}
\begin{proof}
By (V\ref{item:v2}), we have that $\ensuremath{\mathcal{T}}_i|\Desc_i(U)$ and $\ensuremath{\mathcal{T}}_i| \Desc_\P(U) \cap L(\ensuremath{\mathcal{T}}_i)$ are isomorphic, for each $i \in [k]$. The lemma then follows from the definition of $\P | \Desc_\P(U)$.
\end{proof}
For any valid position $U$, $\DG(U)$ denotes the subgraph of $\DG$ induced by $\Desc_\P(U)$.
\begin{observation}\label{obs:update1}
For any valid position $U$, $\DG(U)$ is the subgraph of $\DG$ obtained by deleting all labels in $V(\DG) \setminus \Desc_\P(U)$, along with all incident edges.
\end{observation}
A valid position of special interest to us is $\Uinit
, where
$\Uinit(i) = \phi_i^{-1}(r(T_i))$, for each $i \in [k]$. That is, $\Uinit(i)$ is a singleton containing only the label of $r(T_i)$. Thus, in Figure \ref{fig:displayGraph}, $(\Uinit(1), \Uinit(2), \Uinit(3)) = (\{1\}, \{2\}, \{g\})$.
It is straightforward to verify that $\Uinit$ is indeed valid, that $\Desc_\P(\Uinit) = L(\P)$, and that $\DG(\Uinit) = \DG$.
\vspace{-1.5\parsep}
\paragraph{Semi-universal labels.} Let $U$ be a valid position, and let $\ell$ be a label in $U$. Then, $\ell$ is \emph{semi-universal in $U$} if $U(i) = \{\ell\}$, for every $i \in [k]$ such that $\ell \in L(\ensuremath{\mathcal{T}}_i)$.
It can be verified that in Figure \ref{fig:displayGraph}, labels $1$ and $2$ are semi-universal in $\Uinit$, but $g$ is not, since $g$ is in both $L(\ensuremath{\mathcal{T}}_2)$ and $L(\ensuremath{\mathcal{T}}_3)$, but $\Uinit(2) \neq \{g\}$.
The term ``semi-universal'', borrowed from Pe'er et al.\ \cite{PeerShamirSharan04}, derives from the following fact.
Suppose that $\P$ is ancestrally compatible, that $\ensuremath{\mathcal{T}}$ is a tree that ancestrally displays $\P$, and that $\ell$ is a semi-universal label for some valid position $U$. Then, as we shall see, $\ell$ must label the root $u_\ell$ of a subtree of $\ensuremath{\mathcal{T}}$ that contains all the descendants of $\ell$ in $\ensuremath{\mathcal{T}}_i$, for every $i$ such that $\ell \in L(\ensuremath{\mathcal{T}}_i)$. The qualifier ``semi'' is because this subtree may also contain labels that do not descend from $\ell$ in any input tree, but descend from some other semi-universal label $\ell'$ in $U$ instead. In this case, $\ell'$ also labels $u_\ell$. This property of semi-universal labels is exploited in both our ancestral compatibility algorithm and its proof of correctness (see Section \ref{sec:testAC}).
For each label $\ell \in L(\P)$, let $k_\ell$ denote the number of input trees that contain label $\ell$. We can obtain $k_\ell$ for every $\ell \in L(\P)$ in $O(\MP)$ time during the construction of $\DG$.
\begin{lemma}\label{lem:semiU}
Let $U = (U(1), \dots , U(k))$ be a valid position. Then, label $\ell$ is semi-universal in $U$ if the cardinality of the set $J_\ell = \{i \in [k] : U(i) = \{\ell\}\}$ equals $k_\ell$.
\end{lemma}
\begin{proof}
By definition, $U(i) = \{\ell\}$, for every $i \in J_\ell$. Since $|J_\ell| = k_\ell$, the lemma follows.
\end{proof}
\vspace{-1.5\parsep}
\paragraph{Successor positions.}
For every $i \in [k]$ and every $\ell \in L(\ensuremath{\mathcal{T}}_i)$, let $\ensuremath{{\mathrm{Ch}}}_i(\ell)$ denote the set of children of $\ell$ in $L(\ensuremath{\mathcal{T}}_i)$. For a subset $A$ of $L(\ensuremath{\mathcal{T}}_i)$, let $\ensuremath{{\mathrm{Ch}}}_i(A) = \bigcup_{\ell \in A} \ensuremath{{\mathrm{Ch}}}_i(\ell)$.
Let $U$ be a valid position, and $S$ be the set of semi-universal labels in $U$. The \emph{successor of $U$ with respect to $S$} is the position $U'$ defined as follows. For each $\ell \in S$ and each $i \in [k]$, if $U(i) = \{\ell\}$, then $U'(i) = \ensuremath{{\mathrm{Ch}}}_i(\ell)$; otherwise, $U'(i) = U(i)$.
In Figure \ref{fig:displayGraph}, the set of semi-universal labels in $\Uinit$ is $S = \{1, 2\}$. Since $\ensuremath{{\mathrm{Ch}}}_1(1) = \{3,f\}$ and $\ensuremath{{\mathrm{Ch}}}_2(2) = \{e,f,g\}$, the successor of $\Uinit$ is $U' = (\{3,f\}, \{e,f,g\}, \{g\})$.
\begin{observation}\label{obs:update2}
Let $U$ be a valid position, and let $U'$ be the successor of $U$ with respect to the set $S$ of semi-universal labels in $U$. Then, $\DG(U')$ can be obtained from $\DG(U)$ by doing the following for each $\ell \in S$: (1) for each $i \in [k]$ such that $U(i) = \{\ell\}$, delete all edges between $\ell$ and $\ensuremath{{\mathrm{Ch}}}_i(\ell)$; (2) delete $\ell$.
\end{observation}
Let $U$ be a valid position, and $W$ be a subset of $\Desc_\P(U)$. Then, $U | W$ denotes the position $(U(1) \cap W, \dots , U(k) \cap W)$.
In Figure \ref{fig:displayGraph}, the components of $\DG(U')$, where $U'$ is the successor of $\Uinit$, are $W_1 = \{3,4,a,b,c,d,e,g\}$ and $W_2 = \{f,h,i\}$. Thus, $U' | W_1 = (\{3\}, \{e,g\}, \{g\})$ and $U' | W_2 = (\{f\}, \{f\}, \emptyset)$. We have the following result.
\begin{lemma}\label{lm:childOfU}
Let $U$ be a valid position, and $S$ be the set of all semi-universal labels in $U$. Let $U'$ be the successor of $U$ with respect to $S$, and let $W_1, W_2, \dots, W_p$ be the label sets of the connected components of $\DG(U')$. Then, $U' | W_j$ is a valid position, for each $j \in [p]$.
\end{lemma}
\begin{proof}
It suffices to argue that $U'$ satisfies conditions (V\ref{item:v1}) and (V\ref{item:v2}). The lemma then follows from the fact that the connected components of $\DG(U')$ are label-disjoint.
$U'$ must satisfy condition (V\ref{item:v1}), since $U$ does. Suppose $\ell \in S$. Then, for each $i \in [k]$ such that $\ell \in L(\ensuremath{\mathcal{T}}_i)$, $\Desc_i(U') = \Desc_i(U) \setminus \{\ell\}$ and $ \Desc_\P(U') \cap L(\ensuremath{\mathcal{T}}_i) =(\Desc_\P(U) \cap L(\ensuremath{\mathcal{T}}_i)) \setminus \{\ell\}$. Thus, since (V\ref{item:v2}) holds for $U$, it also holds for $U'$.
\end{proof}
\section{Testing Ancestral Compatibility}\label{sec:testAC}
\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees\ (Algorithm \ref{alg:SLBuild}) is our algorithm for testing compatibility of semi-labeled trees. Its argument, $U$, is a valid position in $\P$ such that $H_\P(U)$ is connected. Line \ref{alg:findSemi} computes the set $S$ of semi-universal labels in $U$. If $S$ is empty, then, as argued in Theorem \ref{thm:buildCorrect} below, $\P|\Desc_\P(U)$ is incompatible, and, thus, so is $\P$. This fact is reported in Line \ref{alg:S0emptyIncompat}. Line \ref{alg:S0singleton} checks if $S$ contains exactly one label $\ell$, with no proper descendants. If so, by the connectivity assumption, $\ell$ must be the only element in $\Desc_\P(U)$. Therefore, Line \ref{alg:S0singletonReturn} simply returns the tree with a single node, labeled $\ell$. Line \ref{alg:updateU} updates $U$, replacing it by its successor with respect to $S$. Let $W_1, \dots, W_p$ be the connected components of $\DG(U)$ after updating $U$.
By Lemma \ref{lm:childOfU}, $U | W_j$ is a valid position, for each $j \in [p]$.
Lines \ref{alg:connected}--\ref{alg:recurseEnd} recursively invoke $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$ on $U | W_j$ for each $j \in [p]$, to determine if there is a tree $t_j$ that ancestrally displays $\P| \Desc_\P(U \cap W_j)$. If any subproblem is incompatible, Line \ref{alg:recurseEnd} reports that $\P$ is incompatible. Otherwise, Lines \ref{alg:rU}--\ref{alg:finalReturn} assemble the $t_j$s into a single tree that displays $\P|\Desc_\P(U)$, whose root is labeled by the semi-universal labels in the set $S$ of Line \ref{alg:findSemi}.
\begin{algorithm}[t]
\SetAlgoLined
\SetNoFillComment
\DontPrintSemicolon
\KwIn{A valid position $U$ for $\P$ such that $\DG(U)$ is connected.}
\KwOut{A semi-labeled tree that ancestrally displays $\P' = \P|\Desc_\P(U)$, if $\P'$ is ancestrally compatible; \texttt{incompatible} otherwise.
}
Let $S = \{\ell \in U: \ell \text{ is semi-universal in } U\}$ \label{alg:findSemi} \;
\If{$S = \emptyset$\label{alg:S0empty}}
{
\Return \texttt{incompatible}\label{alg:S0emptyIncompat}\;
}
\If{$|S| = 1$ and the single element, $\ell$, of $S$ has no proper descendants\label{alg:S0singleton}}
{
\Return the tree consisting of exactly one node, whose label set is $\{\ell\}$\label{alg:S0singletonReturn}
}
Replace $U$ by the successor of $U$ with respect to $S$.\label{alg:updateU} \;
Let $W_1, W_2, \dots, W_p$ be the connected components of $\DG(U)$ \label{alg:connected}
\ForEach{$j \in [p]$ \label{alg:recurseBegin}}
{
Let $t_j = \ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U | W_j)$\; \label{alg:recurse}
\If{$t_j$ is not a tree}
{
\Return \texttt{incompatible} \label{alg:recurseEnd}
}
}
Create a node $r_U$, whose label set is $S$ \label{alg:rU}\;
\Return the tree with root $r_U$ and subtrees $t_1, \dots , t_p$ \label{alg:finalReturn}\;
\caption{\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$(U)$}\label{alg:SLBuild}
\end{algorithm}
Next, we argue the correctness of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$.
\begin{theorem}\label{thm:buildCorrect}
Let $\P = \{\ensuremath{\mathcal{T}}_1, \dots , \ensuremath{\mathcal{T}}_k\}$ be a profile and let $\Uinit = (\Uinit(1), \dots, \Uinit(k))$, where, for each $i \in [k]$, $\Uinit(i) = \phi_i^{-1}(r(T_i))$. Then, $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$
returns either
(i) a semi-labeled tree $\ensuremath{\mathcal{T}}$ that ancestrally displays $\P$, if $\P$ is ancestrally compatible, or (ii) \texttt{incompatible} otherwise.
\end{theorem}
\begin{proof}
(i) Suppose that $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ outputs a semi-labeled tree $\ensuremath{\mathcal{T}}$. We prove that $\ensuremath{\mathcal{T}}$ ancestrally displays $\P$. By Lemma \ref{lm:DN}, it suffices to show that $D(\ensuremath{\mathcal{T}}_i) \subseteq D(\ensuremath{\mathcal{T}})$ and $N(\ensuremath{\mathcal{T}}_i) \subseteq N(\ensuremath{\mathcal{T}})$, for each $i \in [k]$.
Consider any $(\ell,\ell') \in D(\ensuremath{\mathcal{T}}_i)$. Then, $\ell$ has a child $\ell''$ in $\ensuremath{\mathcal{T}}_i$ such that $\ell'' \le_{\ensuremath{\mathcal{T}}_i} \ell'$. There must be a recursive call to $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U)$, for some valid position $U$, where $\ell$ is the set $S$ of semi-universal labels obtained in Line \ref{alg:findSemi}. By Observation \ref{obs:update2}, label $\ell''$, and thus $\ell'$, both lie in one of the connected components of the graph obtained by deleting all labels in $S$, including $\ell$, and their incident edges from $\DG(U)$. It now follows from the construction of $\ensuremath{\mathcal{T}}$ that $(\ell, \ell') \in D(\ensuremath{\mathcal{T}})$. Thus, $D(\ensuremath{\mathcal{T}}_i) \subseteq D(\ensuremath{\mathcal{T}})$.
Now, consider any $\{\ell,\ell' \} \in N(\ensuremath{\mathcal{T}}_i)$. Let $v$ be the lowest common ancestor of $\phi_i(\ell)$ and $\phi_i(\ell')$ in $\ensuremath{\mathcal{T}}_i$ and let $\ell_v$ be the label of $v$. Then, $\ell_v$ has a pair of children, $\ell_1$ and $\ell_2$ say, in $\ensuremath{\mathcal{T}}_i$ such that $\ell_1 \le_{\ensuremath{\mathcal{T}}_i} \ell$, and $\ell_2 \le_{\ensuremath{\mathcal{T}}_i} \ell'$. Because $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ returns a tree, there are recursive calls $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U_1)$ and $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U_2)$ for valid positions $U_1$ and $U_2$ such that $\ell_1$ is semi-universal for $U_1$ and $\ell_2$ is semi-universal for $U_2$. We must have $U_1 \neq U_2$; otherwise, $|U_1(i)| = |U_2(i)| \ge 2$, and, thus, neither $\ell_1$ nor $\ell_2$ is semi-universal, a contradiction. Further, it follows from the construction of $\ensuremath{\mathcal{T}}$ that we must have $\Desc_\P(U_1) \cap \Desc_\P(U_2) = \emptyset$. Hence, $\ell \parallel_\ensuremath{\mathcal{T}} \ell'$, and, therefore, $\{\ell,\ell'\} \in N(\ensuremath{\mathcal{T}})$.
(ii) Asssume, by way of contradiction, that $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ returns \texttt{incompatible}, but that $\P$ is ancestrally compatible. By assumption, there exists a semi-labeled tree $\ensuremath{\mathcal{T}}$ that ancestrally displays $\P$. Since $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ returns \texttt{incompatible}, there is a recursive call to $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U)$ for some valid position $U$ such that $U$ has no semi-universal label, and the set $S$ of Line \ref{alg:findSemi} is empty.
By Lemma~\ref{lm:compatSubprofile}, $\ensuremath{\mathcal{T}} | \Desc_\P(U)$ ancestrally displays $\P|\Desc_\P(U)$. Thus, by Lemma \ref{lem:V2}, $\ensuremath{\mathcal{T}} | \Desc_\P(U)$ ancestrally displays $\ensuremath{\mathcal{T}}_i | \Desc_i(U)$, for every $i \in [k]$. Let
$\ell$ be any label in the label set of the root of $\ensuremath{\mathcal{T}} | \Desc_\P(U)$. Then, for each $i \in [k]$ such that $\ell \in L(\ensuremath{\mathcal{T}}_{i})$, $\ell$ must be the label of the root of $\ensuremath{\mathcal{T}}_i|\Desc_i(U)$.
Thus, for each such $i$, $U(i) = \{\ell\}$. Hence, $\ell$ is semi-universal in $U$, a contradiction.
\end{proof}
\section{Implementation}\label{sec:implementation}
Here we describe an efficient implementation of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$. We focus on two key aspects: finding semi-universal labels in Line \ref{alg:findSemi}, and updating $U$ and $\DG(U)$ in Lines \ref{alg:updateU} and \ref{alg:connected}.
By Observation \ref{obs:update1}, at each recursive call, $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$ deals with a graph obtained from $\DG$ through edge and node deletions. To handle these deletions efficiently, we represent $\DG$ using the dynamic graph connectivity data structure of Holm et al.\ \cite{HolmLichtenbergThorup:2001}, which we refer to as \emph{HDT}. HDT allows us to maintain the list of nodes in each component, as well as the number of these nodes so that, if we start with no edges in a graph with $N$ nodes, the amortized cost of each update is $O(\log^2 N)$. Since $\DG$ has $O(\MP)$ nodes, each update takes $O(\log^2 \MP)$ time.
The total number of edge and node deletions performed by $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ --- including all deletions in the recursive calls --- is at most the total number of edges and nodes in $\DG$, which is $O(\MP)$. HDT allows us to maintain connectivity information throughout the entire algorithm in $O(\MP \log^2 \MP)$ time.
As deletions are performed on $\DG$, $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$ maintains three data fields for each connected component $Y$ that is created: $Y.\WEIGHT$, $Y.\MAP$, and $Y.
\SEMI$. It also maintains a field $\ell.\cnt$, for each $\ell \in L(\P)$.
\vspace{-0.1cm}
\begin{enumerate}
\vspace{-0.5\parskip}
\itemsep1pt \parskip0pt \parsep0pt
\item
$Y.\WEIGHT$ equals $\sum_{\ell \in Y} k_\ell$.
\item
$Y.\MAP$ is a map from a set $J_Y \subseteq [k]$ to a set of nonempty subsets of $Y \cap L(T_i)$. For each $i \in J_Y$, $Y.\MAP(i)$ denotes the set associated with $i$.
\item
$\ell.\cnt$ equals the cardinality of the set
$\{i \in [k] : Y.\MAP(i) \text{ is defined and } Y.\MAP(i) = \{\ell\}\}$. (Recall that $k_\ell$ is the number of input trees that contain $\ell$.)
\item
$Y.\SEMI$ is a set containing all labels $\ell \in Y$ such that $\ell.\cnt = k_\ell$.
\end{enumerate}
\vspace{-0.1cm}
Informally, each set $Y.\MAP(i)$ corresponds to a non-empty $U(i)$; $Y.\SEMI$ corresponds to the semi-universal labels in $Y$. Next, we formalize these ideas.
At the start of the execution of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U)$ for any valid position $U$, $\DG(U)$ has a single connected component, $Y_U = \Desc_\P(U)$. Our implementation maintains the following invariant.
\begin{description
\item[INV:]
At the beginning of the execution of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U)$, $Y_U.\MAP(i) = U(i)$ for each $i \in [k]$ such that $U(i) \neq \emptyset$, and $Y_U.\MAP(i)$ is undefined for each $i \in [k]$ such that $U(i) = \emptyset$.
\end{description}
Thus, $\ell.\cnt$ equals the number of indices $i \in [k]$ such that $U(i) = \{\ell\}$. Along with Lemma \ref{lem:semiU}, INV implies that, at the beginning of the execution of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U)$, $Y_U.\SEMI$ contains precisely the semi-universal labels of $U$.
Thus, the set $S$ of line \ref{alg:findSemi} of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(U)$ can be retrieved in $O(1)$ time.
To establish INV for the initial valid position $\Uinit$, we proceed as follows. By assumption, $\DG(\Uinit)$ has a single connected component, $\Yinit = L(\P)$.
Since $\DG(\Uinit)$ equals $\DG$, we initialize data fields 1--4 for $\Yinit$ during the construction of $\DG$.
$\Yinit.\WEIGHT$ is simply $\sum_{\ell \in L(\P)} k_\ell$. For each $i \in [k]$, $\Yinit.\MAP(i)$ is $\{\ell\}$, where $\ell$ is the label of the root of $T_i$. We initialize the \cnt\ fields as follows.
First, set $\ell.\cnt$ to $0$ for all $\ell \in L(\P)$. Then, iterate through each $i \in [k]$, incrementing $\ell.\cnt$ by one if $\Yinit.\MAP(i) = \{\ell\}$. Finally, $\Yinit.\SEMI$ consists of all $\ell \in \Uinit$ such that $\ell.\cnt = k_\ell$. All data fields can be initialized in $O(\MP)$ time.
We now focus on Lines \ref{alg:updateU} and \ref{alg:connected} of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$. By Observation \ref{obs:update2}, we can update $U$ and $\DG(U)$ jointly as follows. We use a temporary variable $\GBNT$. Prior to executing Line \ref{alg:updateU}, we set $\GBNT = \DG(U)$.
Then, we successively consider each label $\ell \in S$, and perform two steps: (i) initialize data fields 1--4 in preparation for the deletion of $\ell$ and (ii) delete from $\GBNT$ the edges incident on $\ell$ and then $\ell$ itself, updating data fields 1--4 as necessary, to maintain INV. After these steps are executed, $\GBNT$ will equal $\DG(U)$ for the new set $U$ created by Line \ref{alg:updateU}. Steps (i) and (ii) are done by $\ensuremath{\texttt{Initialize}}(\ell)$ (Algorithm \ref{alg:initMarks}) and $\ensuremath{\texttt{Delete}}(\ell)$ (Algorithm \ref{alg:delLabel}), respectively.
Lines \ref{alg:delSemi}--\ref{alg:addPairMap} of $\ensuremath{\texttt{Initialize}}(\ell)$ initialize $Y.\MAP$ and $Y.\SEMI$ to reflect the fact that label $\ell \in S$ is leaving $U(i)$, for each $i \in [k]$ such that $i \in L(\ensuremath{\mathcal{T}}_i)$, to be replaced by its children in $\ensuremath{\mathcal{T}}_i$, and will no longer be semi-universal. Lines \ref{alg:initSingleton}--\ref{alg:forAddAlphaSemi} are needed to update certain $\cnt$ fields due to the possibility that singleton sets $Y.\MAP(i)$ may be created in the preceding steps. The number of operations on $Y.\MAP$ performed by $\ensuremath{\texttt{Initialize}}(\ell)$ is $O(\sum_{i \in [k] : \ell \in L(\ensuremath{\mathcal{T}}_i)} |\ensuremath{{\mathrm{Ch}}}_i(\ell)|)$; i.e., it is proportional to the total number of children of $\ell$ in all the input trees. Since $\ell$ is considered only once, the total number of operations on $\MAP$ fields of the various sets $Y$ considered over the entire execution of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ is $O(\MP)$. The number of updates of $Y.\MAP$ done by $\ensuremath{\texttt{Initialize}}(\ell)$ is at most $k_\ell$; the total over all $\ell$ considered by $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ over its entire execution is $O(\MP)$.
\begin{algorithm
\SetAlgoLined\SetNoFillComment
\DontPrintSemicolon
Delete $\ell$ from $Y.\SEMI$ \label{alg:delSemi}\;
\ForEach {$i \in [k]$ such that $\ell \in L(\ensuremath{\mathcal{T}}_i)$\label{alg:forEachIContainingL}}
{
Delete $Y.\MAP(i)$ \;
\ForEach {$\alpha \in \ensuremath{{\mathrm{Ch}}}_i(\ell) $\label{alg:forEachAlpha}}
{
Add $\alpha$ to $Y.\MAP(i)$ \label{alg:addPairMap} \;
}
\If{$Y.\MAP(i)$ is a singleton\label{alg:initSingleton}}
{
Let $\beta$ be the single element in $Y.\MAP(i)$ \;
Set $\beta.\cnt = \beta.\cnt +1$ \;
\lIf{$\beta.\cnt = k_{\beta}$}
{
add $\beta$ to $Y.\SEMI$\label{alg:forAddAlphaSemi}
}
}
}
\caption{\ensuremath{\texttt{Initialize}}$(\ell)$}\label{alg:initMarks}
\end{algorithm}
$\ensuremath{\texttt{Delete}}(\ell)$ begins
by consulting HDT to identify the connected component $Y$ that currently contains $\ell$. The loop in Lines \ref{alg:forScanEdges}--\ref{alg:endForScanEdges} successively deletes each edge between $\ell$ and a child $\alpha$ of $\ell$, updating the appropriate data fields for the resulting connected components. Line \ref{alg:testConn} queries the HDT data structure to determine whether deleting $(\ell,\alpha)$ splits $Y$ into two components. If $Y$ remains connected, no updates are needed --- the \textbf{continue} statement skips the rest of the current iteration and proceeds directly to the next. Otherwise, $Y$ is split into two parts $Y_1$ and $Y_2$. $\ensuremath{\texttt{Delete}}$ uses a weighted version of the technique of scanning the smaller component \cite{EvenShiloach:1981}. Line \ref{alg:smallerWeight} identifies which of the two components has the smaller \WEIGHT\ field; without loss of generality, it assumes that $Y_1.\WEIGHT \le Y_2.\WEIGHT$. Lines \ref{alg:delInitY1} and \ref{alg:delInitY2} initialize $Y_1.\MAP$ and $Y_1.\SEMI$ to \NULL\ and $Y_2.\MAP$ and $Y_2.\SEMI$ to the corresponding fields of $Y$. Lines \ref{alg:forBetaY1}--\ref{alg:addToJ}, scan each label $\beta$ in $Y_1$, from $Y_2.\MAP(i)$ to $Y_1.\MAP(i)$, for every $i$ such that $\beta \in L(\ensuremath{\mathcal{T}}_i)$.
Set $J$, updated in Line \ref{alg:addToJ}, keeps track of the indices $i$ such that $Y_1.\MAP(i)$ and $Y_2.\MAP(i)$ are modified. Lines \ref{alg:forUpdateSemi}--\ref{alg:endForScanEdges} iterate through $J$ to determine if any new singleton sets were created in either $Y_1$ or $Y_2$.
This operation requires at most one update in each of $Y_1.\SEMI$ and $Y_2.\SEMI$; each update takes $O(1)$ time. After all edges incident on $\ell$ are deleted, $\ell$ itself is deleted (Line \ref{alg:delEll}).
The preceding description of $\ensuremath{\texttt{Delete}}(\ell)$ omits the updating of the $\WEIGHT$ fields of the connected components created by an edge deletion. This is done before Line \ref{alg:smallerWeight}, by (again) using the technique of scanning the smaller component. We consult HDT to determine which of $Y_1$ and $Y_2$ has fewer labels. Assuming, without loss of generality, that $|Y_1| < |Y_2|$, compute $Y_1.\WEIGHT$ in a sequential scan of $Y_1$. Then, $Y_2.\WEIGHT = Y.\WEIGHT - Y_1.\WEIGHT$.
\begin{algorithm}[t]
\SetAlgoLined\SetNoFillComment
\DontPrintSemicolon
Let $Y$ be the connected component of $\GBNT$ that contains $\ell$ \label{alg:findY}\;
\ForEach {$\alpha \in \ensuremath{{\mathrm{Ch}}}(\ell)$\label{alg:forScanEdges}}
{
Delete edge $\{\ell, \alpha\}$ from $\GBNT$ \label{alg:deleteEdge}\;
\lIf{$Y$ remains connected\label{alg:testConn}}{\continue}
Let $Y_1, Y_2$ be the connected components of $\GBNT$; assume that $Y_1.\WEIGHT \le Y_2.\WEIGHT$ \label{alg:smallerWeight} \;
Let $Y_1.\MAP = \NULL$ and $Y_1.\SEMI = \NULL$ \label{alg:delInitY1}\;
Let $Y_2.\MAP = Y.\MAP$ and $Y_2.\SEMI = Y.\SEMI$ \label{alg:delInitY2}\;
Let $J = \emptyset$ \;
\ForEach {$\beta \in Y_1$\label{alg:forBetaY1}}
{
\ForEach{$i \in [k]$ such that $\beta \in L(\ensuremath{\mathcal{T}}_i)$}
{
Move $\beta$ from $Y_2.\MAP(i)$ to $Y_1.\MAP(i)$ \label{alg:MoveBeta} \;
$J = J \cup \{i\}$ \label{alg:addToJ}
}
}
\ForEach {$i \in J$\label{alg:forUpdateSemi}}
{
\ForEach{$j \in \{1,2\}$}
{
\If{$Y_j.\MAP(i) = \emptyset$}
{
Delete $Y_j.\MAP(i)$
}
\ElseIf{$Y_j.\MAP(i)$ is a singleton}
{
Let $\gamma$ be the single element in $Y_j.\MAP(i)$ \;
$\gamma.\cnt = \gamma.\cnt + 1$ \;
\lIf{$\gamma.\cnt = k_\gamma$}
{
add $\gamma$ to $Y_j.\SEMI$ \label{alg:endForScanEdges}
}
}
}
}
}
Delete $\ell$ from $\GBNT$ \label{alg:delEll}
\caption{$\ensuremath{\texttt{Delete}}(\ell)$}\label{alg:delLabel}
\end{algorithm}
Let us track the number of operations on $\MAP$ fields in Lines \ref{alg:forBetaY1}--\ref{alg:addToJ} of $\ensuremath{\texttt{Delete}}(\ell)$ that can be attributed to some specific label $\beta \in L(\P)$ over the entire execution of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$.
Each execution of Lines \ref{alg:forBetaY1}--\ref{alg:addToJ} for $\beta$ performs $k_\beta$ operations on $\MAP$ fields.
Let $w_r(\beta)$ be the weight of the connected component containing $\beta$ at the beginning of the loop of Lines \ref{alg:forBetaY1}--\ref{alg:addToJ}, at the $r$th time that $\beta$ is considered in those lines; thus, $w_0(\beta) \le \sum_{\ell \in L(\P)} k_\ell$. Then, $w_r(\beta) \le w_0(\beta)/2^r$. The reason is that we only consider $\beta$ if (i) $\beta$ is contained in one of the two components that result from
deleting an edge in Line \ref{alg:deleteEdge} and (ii) the component containing $\beta$ has the smaller weight of the two components. Thus, the number of times $\beta$ is considered in Lines \ref{alg:forBetaY1}--\ref{alg:addToJ} over the entire execution of $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ is $O(\log w_0(\beta))$, which is $O(\log \MP)$, since $w_0(\beta) = O(\MP)$. Therefore, the total number of updates of $\MAP$ fields over all labels is $O(\log \MP \cdot \sum_{\ell \in L(\P)} k_\ell)$, which is $O(\MP \log \MP)$.
It can be verified that the number of updates to \cnt\ and \SEMI\ fields is also $O(\MP \log \MP)$.
A similar analysis shows that the total time to update $\WEIGHT$ fields over all edge deletions performed by $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ is $O(\MP \log \MP)$.
To summarize, the work done by $\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees(\Uinit)$ consists of three parts: (i) initialization, (ii) maintaining connected components, and (iii) maintaining the \WEIGHT, \MAP, and \SEMI, and fields for each connected component, as well as $\ell.\cnt$ for each label $\ell$. Part (i) takes $O(\MP)$ time. Part (ii) involves $O(\MP)$ edge and node deletions on the HDT data structure, at an amortized cost of $O(\log^2 \MP)$ per deletion. Part (iii) requires a total of $O(\MP \log \MP)$ updates to the various fields. Using data structures that take logarithmic time per update, leads to our main result.
\begin{theorem}\label{thm:SLBuildA}
\ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees\ can be implemented so that \ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees$(\Uinit)$ runs in $O(\MP \log^2 \MP)$ time.
\end{theorem}
\section{Discussion\label{sec:discussion}}
Like our earlier algorithm for compatibility of ordinary phylogenetic trees, the more general algorithm presented here, \ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees, is a polylogarithmic factor away from optimality (a trivial lower bound is $\Omega(\MP)$, the time to read the input). \ensuremath{\texttt{BuildNT}}} % Build for semi-labeled trees\ has a linear-space implementation, using the results of Thorup \cite{Thorup2000}. A question to be investigated next is the performance of the algorithm on real data. Another important issue is integrating our algorithm into a synthesis method that deals with incompatible profiles.
|
\section{Introduction}\label{sec:intro}
Planets are born in protoplanetary disks. Planets can produce radial and azimuthal features in the disks observable in resolved images at various wavelengths \citep[e.g., simulations by][]{fouchet10, pinilla12-diffcavsize, gonzalez12, dejuanovelar13, dong15-gaps, juhasz15,dong15-spiralarms,pohl15}. Sculpting by planets is a leading hypothesis for the origin of transitional disks, protoplanetary disks with depleted inner cavities or gaps (see \citealt{espaillat14} for a recent review; hereafter we will refer to the depleted inner regions as gaps although some may be complete cavities). The existence of transitional disks was first suggested by the distinctive dip at $\sim1-10\micron$ in the spectral energy distributions \citep[SED;][]{storm89, skrutskie90, calvet02, espaillat07} of some young stars and subsequently confirmed in resolved observations in near-infrared (NIR) scattered light \citep[e.g.,][]{mayama12, muto12, hashimoto12, canovas13, garufi13, grady13, avenhaus14, tsukagoshi14} and mm dust continuum and molecular line emission \citep[e.g.,][]{andrews11, casassus13, fukagawa13, vandermarel13, perez14, zhang14}.
The gaps in transitional disks are both deep and wide. In gas they typically extend several tens of AU \citep[e.g.,][]{zhang14,vandermarel15,vandermarel16}. Measuring their depths in gas has been more challenging, but recent ALMA observations have suggested that the gaps are depleted by at least a factor of 10--1000 \citep[e.g.,][]{vandermarel15, vandishoeck15, vandermarel16}. Non-dynamical mechanisms to open these gaps have been proposed, including photoevaporation \citep[e.g.,][]{clarke01,alexander06, owen12, rosotti13}, grain growth \citep[e.g.,][]{dullemond05, birnstiel12}, and the magnetorotational instability \citep[e.g.,][]{chiang07}. However, all have major drawbacks and cannot explain at least a subset of protoplanetary disks \citep{espaillat14}.
In this study, we will focus on the planetary sculpting scenario, in which the inner hole is a common gap opened by multiple planets \citep{dodsonrobinson11,zhu11}. Numerical simulations of the planetary sculpting scenario have reproduced all major observed aspects of transitional disks. \citet{dong15-gaps} showed that the width and depth of planet-opened common gaps can match observations at both NIR and mm wavelengths. In particular, observations have systematically found bigger gap sizes in the mm continuum than in the gas or scattered light \citep[e.g.,][]{dong12cavity, hashimoto15, zhang14, vandermarel15}. The planet-opening-gap scenario has successfully reproduced this feature by trapping the large dust at the pressure peak beyond the gas gap edge \citep{pinilla12-diffcavsize, zhu12, dejuanovelar13}. Millimeter observations have revealed large scale azimuthal asymmetries in a few systems, for which the mm dust ring is lopsided and most emission comes from one side \citep{vandermarel13, casassus13, isella13, perez14}. Vortex formation at the planet-induced gap edge has been proposed to explain these observations \citep[e.g.,][but also see \citealt{mittal15}]{zhu14stone, lyra13}. Finally, accretion rate analysis also suggests the presence of giant planets inside the gaps to alter the accretion flow onto the stars \citep{najita15}.
Despite the success of the planet sculpting hypothesis in explaining observations of transitional disks, only a handful of companion candidates have been discovered in direct imaging observations of transitional disks (e.g., \citealt{huelamo11, biller12, kraus12, quanz13-planet, brittain14, close14, reggiani14}). Direct imaging of planets is difficult due to the low contrast ratio between the brightness of planets and their host stars, and the proximity of the planets to the stars\footnote{A planet at 30~AU is only $\sim$0.2$\arcsec$ away from its star at 140 pc, a typical distance for nearby resolved protoplanetary disks.}. In addition, most discovered companions are likely too far away from the outer gap edges to be solely responsible for the gap. So far the strongest direct evidence for the multi-planet sculpting scenario is the detection of three companions inside the wide gap in LkCa 15 \citep{sallum15}.
If the gaps in the younger protostellar disk HL Tau \citep{brogan15} are opened by sub-Jupiter mass planets \citep{dong15-gaps,dipierro15-hltau}, it may be a transitional disk ``embryo'' that will develop a wide common gap once the planets grow larger. The HR~8799 system, home to at least 4 giant planets at tens of AU orbiting a 30~Myr old young star \citep{marois08,marois10}, may be a later stage in the post-gas-disk era. It remains an open question whether systems of massive giant planets are sufficiently common to account for the occurrence rate of transitional disks.
Furthermore, it is unclear whether the large masses and close spacings of planets required to open common gaps are stable over transitional disks' $\sim$Myr lifetimes. \citet{duffell15dong} showed that a common gap is shallower than the gap open by a single planet of the same mass and the material between the two planets' gaps cannot be eroded away. Under typical disk conditions, a wide common gap with a depletion factor larger than 10 throughout the gap requires several massive planets with $q=M_{\rm p}/M_\star\gtrsim2\times10^{-3}$ (or $\sim$2$M_{\rm J}$ planets around a 1~$M_\odot$ star) spanning several to several tens of AU. The tension between close packing and stability has been explored for the HR 8799 system \citep{fabrycky10} and the HL Tau protoplanetary disk \citep{brogan15, tamayo15}, which features a series of gaps that may be sculpted by $\sim$~Saturn mass planets \citep{dong15-gaps, dipierro15-hltau}. \citet{sallum15} also suggested that the planets in LkCa 15's gap have to be either less massive than $5M_{\rm J}$ or in 2:1 mean motion resonance in order for the system to be stable.
Here we assess the plausibility of the planetary sculpting hypothesis using constraints from stability and occurrence rates and explore the implications for the planetary systems' subsequent evolution. We start by defining the properties of observed gaps in transitional disks in Section~\ref{sec:gapproperties}. In Section~\ref{sec:planetrequirements} we synthesize the requirements for configurations of multi-planet systems that are able to open gaps consistent with transitional disk observations. We then carry out $N$-body simulations in Section~\ref{sec:stability} to explore whether such configurations can be stable. In Section~\ref{sec:statistics}, we compare the properties of the multi-planet systems that can create transitional disks with current statistics on the occurrence rate of giant planets at large separations. A summary and discussion are presented in Section~\ref{sec:summary}.
\section{Observed properties of the gaps}\label{sec:gapproperties}
Here we lay out the observed properties of gaps that planetary sculpting must account for. We focus primarily on the gaps opened in the gas (``gas gaps'') because the dynamics of dust is subject to dust-gas coupling in addition to disk-planet interactions.
\subsection{Gap extent}
High resolution and high sensitivity resolved molecular gas observations are needed to determine the gap sizes. \citet{vandermarel15,vandermarel16} performed such observations for 8 systems and found a typical gas gap extent of $\sim$15-45 AU. We will use 30 AU as the fiducial value in the models throughout the paper.
An important related quantity is the continuity of the gap, which constrains the spacing of the sculpting planets. The current image resolution does not rule out the possibility of undepleted narrow rings of material within the gap (e.g., between planets). However, because the total mass of gas in the gap is constrained (except the inner $\sim$~1AU or so, see below), undepleted rings cannot compromise a significant portion of the gap extent.
\subsection{Depth (depletion)}\label{sec:gapdepletion}
Here we parametrize the gas gap depletion as $\delta_{\rm gap}=\Sigma_0/\Sigma_{\rm gap}$, where $\Sigma_{\rm gap}$ is the gas surface density in the gap and $\Sigma_0$ is the undepleted value extrapolated from the outer disk. Observations use relatively rare species, such as $^{13}$CO and C$^{18}$O, to infer the total gas abundance at a given location\footnote{An additional layer of complication is that CO is condensable at the low temperatures in the outer parts of a disk, and this needs to be taken into account when using CO observations to constrain the density structure of the gas in total.}. To date, most studies have constrained only lower limits on $\delta_{\rm gap}$ (hereafter $\delta_{\rm gap,min}$). Getting robust constraints on $\Sigma_{\rm gap}$ has been difficult. The abundance of these rare species depends not only on the gas surface density but on the temperature conditions of the disk and chemical reactions with other species. Model degeneracies in the radial profile of the gas are also a challenge (e.g., \citealt{bruderer14}).
Recently $\delta_{\rm gap}$ (or $\delta_{\rm gap,min}$) has been measured in a few individual systems by combining the latest ALMA data with state of the art disk physical-chemical modeling tools. \citet{bruderer14} found that the IRS 48 system's mm data is consistent with $\delta_{\rm gap} \sim 10$ inside 60 AU and $\delta_{\rm gap} \sim 100$ inside 20 AU. \citet{perez15} found that in the HD 142527 system, the gas inside the $\sim100$~AU gap is depleted by $\delta_{\rm gap}\sim50$. Using ALMA CO observations, \citet{vandermarel15} determined $\delta_{\rm gap} \gtrsim 10$ for LkCa 15, $\delta_{\rm gap} \gtrsim 10^4$ for RXJ1615-3255, and $\delta_{\rm gap} \sim 10^5$ for J1604-2130; using $^{13}$CO and C$^{18}$O observations, \citet{vandermarel16} have recently pushed $\delta_{\rm gap}$ to $\sim5000$ for SAO 206462; $\sim100$ for SR 21; $\gtrsim10^4$ for DoAr 44; and $\gtrsim1000$ for IRS 48.
Alternatively, $\delta_{\rm gap}$ can be inferred from the gap depletion factor of the sub-$\micron$-sized grains -- which are traced by scattered light imaging -- because these fine grains are generally well-mixed with gas such that $\delta_{\rm gap} \approx \delta_{\rm fine\ grains}$. \citet{dong12pds70} found that the scattered light from the $\sim70$~AU gap in PDS~70 is consistent with $\delta_{\rm fine\ grains} \sim1000$.
All the estimations of $\delta_{\rm gap}$ described above generally assume that the gap is uniformly (i.e., azimuthally symmetrically) depleted from a fiducial profile (usually taking the form of $\Sigma\propto1/r$ and sometimes including an exponential cut off at large radius) that describes the undepleted disk and that the gap edge is sharp. Azimuthal variations inside the gap, for example accretion streamers connecting the gap edge to the planets or the star \citep{casassus13}, may account for some remaining gas, though current data generally lack the needed angular resolution to resolve these structures.
The current observed $\delta_{\rm gap,min}$ span 10-1000, but we do not yet have a large enough sample to understand the selection effects. Below we will use $\delta_{\rm gap,min}=10$ as a conservative requirement for our planet-opened gaps and also assess our results for $\delta_{\rm gap,min}=100$ and 1000.
\subsection{The inner disk}\label{sec:inner_disk}
The gas depletion in the very inner part of the disk (i.e., $\sim$AU scales) is practically unconstrained based on current ALMA gas observations and chemical disk modeling. Interpreting mm gas emissions from the inner disk is difficult due to the challenge of chemical disk modeling with the high but uncertain temperatures close to the star. CO ro-vibrational emissions at 4.7$\micron$ indicate the existence of gas at $\sim$AU scale in at least some cases \citep[e.g.,][]{pontoppidan08,salyk09}; however the amount of gas is not well determined. In the analysis below, we will simply consider the gas surface density $\Sigma_{\rm gas}$ in the inner $\sim$AU scale disk as unconstrained.
\subsection{Occurrence rate, $f_{\rm TD,observed}$}\label{sec:ftd}
Studies of young clusters have revealed that on the order of 10\% of protoplanetary disks are transitional disks with significant gap sizes. Based on analysis of Spitzer photometry, \citet{luhman10} found that the fraction of protoplanetary disks that are transitional disks, $f_{\rm TD,observed}$, is $\sim$13\% (15/113) in Taurus\footnote{\citet{luhman10} defined transitional disks as objects with weak or no excess at $\lambda<10\micron$, and large excess at longer wavelengths. SED modeling \citep[e.g.,][]{espaillat07,andrews11} has shown that this kind of SED is indicative of large gap sizes of typically about 30 AU.}, consistent with a number of other star forming regions at 2-10 Myr such as Chamaeleon I and IC 348. \citet{muzerolle10} found a similar $f_{\rm TD,observed}$ --- $\sim$12\% (51/417) --- for a sample of 7 clusters\footnote{We include both the ``warm'' and ``weak excess'' transitional disks in \citet{muzerolle10} in the fraction, as both are likely to comply with our definition of transitional disks.}. Note that these fractions are computed for samples of host stars with a range of properties, such as spectral type and age.
\citet{muzerolle10} found $f_{\rm TD,observed}$ may depend on age and spectral type but the sample size is too small to robustly test correlations; \citet{andrews11} found $f_{\rm TD,observed}$ is higher around mm bright disks ($\sim25\%$ for the upper quartile of the total mm flux distribution). It is not yet known whether this fraction primarily reflects nature ($\sim 10\%$ disks develop wide gaps) or nurture (most disks spend $\sim 10\%$ of their lifetime in the transitional phase). We will further explore the interpretation $f_{\rm TD,observed}$ in Section~\ref{sec:statistics}.
\subsection{Summary}\label{sec:gapsummary}
Based on the observed characteristics of transitional disks discussed above, we synthesize a set of properties for a typical transitional disk gap (illustrated in Figure~\ref{fig:illustration_gap}),
which we will use as fidicuial requirements in later explorations of planetary system configurations:
\begin{enumerate}
\item The gap is approximately continuous with no wide, interspersed un-depleted regions inside\footnote{Relaxing this requirement to a weaker one, that the total gas mass inside the gap is depleted by at least a factor of 10 (or 100, 1000) --- a more direct result from modeling ALMA observations --- makes little difference. If surface density scales with radius as $\Sigma_{\rm 0}\propto 1/r$ in the inner part of the disk as assumed in most models \citep[e.g.,][]{andrews11, vandermarel15}, a depletion factor of at least 10 in total gas mass means less than $10\%$ of radius inside the gap can be occupied by undepleted ring-like regions in between adjacent individual gaps opened by planets. On average, the distance between planets can increase by less than $10\%$ if the weaker condition is adopted, which has little effect on the configurations of planetary systems that can meet the condition.}.
\item The depletion factor, $\delta_{\rm gap}$, is at least 10 (or 100, 1000) everywhere in the common gap, except in the poorly constrained inner disk ($< r_{\rm inner\ disk}$).
\item The typical size of the gas gap (i.e., its outer edge) is $r_{\rm gap}=30$~AU.
\item The size of the inner disk $r_{\rm inner\ disk}$, i.e., the inner edge of the gap, is no bigger than 3 AU, corresponding to 10\% of the gap size.
\end{enumerate}
\section{Requirements for multi-planet systems to open deep and wide gaps}\label{sec:planetrequirements}
We investigate the requirements for multi-planet systems to open deep and wide gaps consistent with transitional disk observations. Then we construct a series of such systems for use in the dynamical studies in Section~\ref{sec:stability}.
The profile of a gap opened by a single planet depends on the aspect ratio of the disk $h/r$; the parametrized disk viscosity $\alpha$ where the viscosity $\nu = \alpha h c_s$, $h$ is the scale height, and $c_s$ is the sound speed \citep{shakura73}; and the mass of the planet $q$ \citep[e.g.,][]{fung14, duffell15gap}. We assume that both $h/r$ and $\alpha$ are constant in time and location. The properties of a common gap opened by multiple planets also depend on the total number and the separation of the planets \citep[e.g.,][]{duffell15dong}. In general, given a set of disk properties $h/r$ and $\alpha$, a multi-planet system must satisfy three requirements to open a common gap that resembles those observed:
\begin{enumerate}
\item The planets must be massive enough to carve out a deep gap.
\item The planets must be placed close enough together so that no wide, interspersed un-depleted regions exist.
\item The outermost planet has to be at the right location to account for the gap size, while the innermost planet has to be close enough to the star to extend the gap to $\sim$ 1 AU.\footnote{\citet[][see also \citealt{duffell14}]{durmann15} have shown that a gap opened by planets does not separate the inner and outer disk; instead, gas may cross the gap and reach the inner disk.}
\end{enumerate}
In the rest of this section, we will look at these requirements one by one. We assume that the planets' eccentricities do not significantly change the gap width (e.g., \citealt{duffell15e} find that the eccentricity damps rapidly once it begins large enough for the planet to collide with the gap wall). For simplicity, we assume equal mass planets; our results could be extended by mixing and matching planets of different masses and their gaps. For example, a configuration in which planet mass decreases with semi-major axis may avoid Type II migration in a flared disk \citep{crida09}.
\subsection{Planet masses}\label{sec:mp}
\citet{fung14} carried out extensive hydrodynamics simulations of gap opening in disks, and found that $\delta_{\rm gap,i}$, the depth of a gap opened by a single planet\footnote{Throughout the paper, we reserve the subscript $_{\rm gap}$ for the common gap, and subscript $_{\rm gap,i}$ for the individual gap opened by a single planet. We use $(r)$ after a quantity (e.g., $\delta_{\rm gap}(r)$) to emphasize the radial dependence of the function. A quantity without $(r)$ represents the value of this quantity in the bulk part of the gap (e.g., $\delta_{\rm gap}$ represents the depletion factor in the core of a gap that is nearly independent of $r$).}, can be described by a simple power law with $h/r$ and $\alpha$ (see also \citealt{duffell13} and \citealt{kanagawa15}), segmented by the planet mass $q$. In the low mass planet regime $10^{-4} \lesssim \, q \lesssim \, 5 \times 10^{-3}$ (Eqn. 12 in \citealt{fung14})
\begin{align}
\delta_{\rm gap,i}^{-1}=& \, 0.14 \left( \frac{q}{10^{-3}} \right)^{-2.16} \left( \frac{\alpha}{10^{-2}} \right)^{1.41} \left( \frac{h/r}{0.05} \right)^{6.61} \,,
\label{eqn:depthlowmp}
\end{align}
and for higher planet masses $5\times10^{-3} \lesssim \, q \lesssim \, 10^{-2}$ (Eqn. 14 in \citealt{fung14})
\begin{align}
\delta_{\rm gap,i}^{-1} =& \, 4.7 \times 10^{-3} \left( \frac{q}{5 \times 10^{-3}} \right)^{-1.00} \left( \frac{\alpha}{10^{-2}} \right)^{1.26} \left( \frac{h/r}{0.05} \right)^{6.12} \,.
\label{eqn:depthhighmp}
\end{align}
\citet{duffell15dong} studied the profiles of gaps opened in multi-planet systems using hydrodynamics simulations. Low mass and/or widely separated planets open individual gaps well-characterized by the same scaling relation found in \citet{fung14}. At fixed separations, individual gaps formally merge at a critical planet mass $q=q_{\rm crit}$, when the high density ($\Sigma_0$) ring structure between each pair of adjacent gaps is squeezed out and a common gap forms with a flat bottom (i.e., constant $\delta_{\rm gap}(r)$ in the core of the common gap). When $q$ increases above $q_{\rm crit}$, $\delta_{\rm gap}$ is a constant in between $q_{\rm crit}$ and $\sim$2~$q_{\rm crit}$, before increasing again at $q\gtrsim2q_{\rm crit}$. Therefore Eqn.~\ref{eqn:depthlowmp} and \ref{eqn:depthhighmp} yield the lowest planet mass, $q_{\rm gap}(\deltagapmin|\alpha,h/r)$, needed to open a common gap with multiple planets for a given $\delta_{\rm gap,min}$ at each [$\alpha$, $h/r$], which is same as the planet mass needed to open an individual gap with $\delta_{\rm gap,i}=\delta_{\rm gap,min}$ and [$\alpha$, $h/r$]. A computed $q_{\rm gap}(\deltagapmin|\alpha,h/r)$ is listed in Table~\ref{tab:mplimit} for various system parameters. All the planet masses exceed the viscous gap opening criteria for given $\alpha$ and $h/r$ \citep[e.g., Eqn. 25,][]{kratter10}. The viscosity, $\alpha$, has been estimated to be on the order of 0.01 based on the median accretion rate of T Tauri stars \citep{hartmann98}. At 30~AU, in a disk around a typical Herbig Ae/Be star with a stellar mass 2.5~$M_\odot$, a stellar radius 2~$R_\odot$, and a photosphere temperature $10,000$~K, the midplane temperature $T_{\rm mid}$ is about 65~K and $h/r$ is about 0.06; while in a disk around a typical T Tauri star with 0.5~$M_\odot$, 2~$R_\odot$, and $4000$~K, $T_{\rm mid}\sim30$~K and $h/r\sim0.08$. We will choose $\alpha$ of 0.1, 0.01 and 0.001 and $h/r$ of 0.05 and 0.1 as representative values. Finally, we note that the gap opening timescale considered here is generally much shorter than the age of transitional disks. \citet[][Fig. 5]{fung14} showed that with $\alpha=0.001$, the gap opened by a $1M_{\rm J}$ planet reaches roughly its final state in a few thousand orbits, which is on the order of 0.1 Myr for the orbital periods considered here. The gap opening timescale is shorter for more massive planets and more viscous disks.
Observed lower limits on $\delta_{\rm gap}$ do not provide an upper limit on planet mass. We therefore turn to the outcomes of companion searches in transitional disks to set a meaningful upper limit. Recent near-IR interferometry \citep{pott10}, adaptive optics imaging \citep{cieza12}, and aperture masking observations \citep{kraus11} have shown that the observed inner holes and gaps in transitional disks are rarely due to close stellar companions or brown dwarfs \citep{alexander14}, as typically these objects would have been detected. In contrast, planetary mass companions (i.e., objects with a mass lower than the deuterium burning limit $\sim$13$M_{\rm J}$) at tens of AU can be well hidden in current observations. Given these considerations, we limit our analysis to the planetary mass regime, $M_{\rm p}\lesssim13M_{\rm J}$.
\subsection{Planet separation}
At a given mass, planets have to be located close enough together to open a common gap instead of individual gaps. \citet{duffell15dong} found that when two neighboring gaps start to overlap, the overlapping region has an effective depletion factor
\begin{align}
\delta_{\rm overlapping}(r)=\delta_{\rm gap,1}(r)\times\delta_{\rm gap,2}(r),
\label{eqn:deltaoverlapping}
\end{align}
where $\delta_{\rm gap,1}(r)$ ($\delta_{\rm gap,2}(r)$) is the depletion factor at $r$ for an individual gap opened by planet 1 (2). Figure~\ref{fig:sigma_rp} shows two examples of the profile of a gap opened by two planets. This relation holds until $\delta_{\rm overlapping}(r)\approx\delta_{\rm gap,1}\approx\delta_{\rm gap,2}$, at which point the two gaps formally merge and are replaced by a common gap. Therefore, in order for two planets to open a common gap with $\delta_{\rm gap,min}$ throughout, the two gaps must join at $r_0$ where $\delta_{\rm gap,1}(r_0)\approx\delta_{\rm gap,2}(r_0)\approx\sqrt{\delta_{\rm gap,min}}$. The location $r_0$ defines the maximum separation between the two planets. A precondition $\delta_{\rm gap,1} (\delta_{\rm gap,2})\gtrsim\delta_{\rm gap,min}$ is also required, i.e., that each planet is able to open a gap with a depletion factor at least $\delta_{\rm gap,min}$ (i.e., the condition in Section~\ref{sec:mp}).
\subsection{Planet locations}
Similar to $\delta_{\rm gap,i}$, the width of a gap induced by an individual planet, $\Delta_{\rm gap,i}$,
depends on $\alpha$, $h/r$, $q$, and the definition of the gap edge. As described in Section~\ref{sec:gapdepletion},
we focus on three illustrative values of $\delta_{\rm gap,min}$: 100, 100, and 1000. Respectively for the three values of $\delta_{\rm gap,min}$, we use planets with $q\geq q_{\rm gap}(10|\alpha,h/r)$, $q\geq q_{\rm gap}(100|\alpha,h/r)$, and $q\geq q_{\rm gap}(10|\alpha,h/r)$. Applying Eqn. \ref{eqn:deltaoverlapping}, we define the edges of individual gaps as the location where $\delta_{\rm gap,i}(r) =$ 3, 10, and 30. The gap width, $\Delta_{\rm gap,i}$, is the distance between the two edges. We obtain $\Delta_{\rm gap,i}$ numerically using the gap profiles calculated by \citet{fung14} using the code {\tt PENGUIN} \citep{fung15thesis}. In general, $\Delta_{\rm gap,i}$ is a monotonically decreasing function of $\alpha$ and $h/r$, and a monotonically increasing function of $M_{\rm p}$ \citep[see also][]{crida06}.
Finally, we construct maximally-spaced planetary systems that can open common gaps satisfying the conditions listed in Section~\ref{sec:gapsummary} for a given set of $\delta_{\rm gap,min}$, $\alpha$, and $h/r$. Planets are placed so their gap edges just touch; therefore $\delta_{\rm gap}(r)$ reaches 10 (or 100, 1000) at the joint point between each pair of adjacent individual gaps, and $\delta_{\rm gap}(r)\geq10$ (or 100, 1000) everywhere else in the common gap. The outermost planet, which maintains the outer edge of the common gap at $r_{\rm gap}=30$~AU, is designated as planet 1. Inner planets are designated with increasing numbers. We use a minimum number of planets in the sense that the inner gap edge maintained by the innermost planet ($r_{\rm inner\ disk}$) is inside 3~AU. Table~\ref{tab:planetarysystems} lists the parameters of all systems.
Opening a deep and wide common gap requires a system of 3--6 massive and closely-spaced planets. The separation between a pair of adjacent planets $i$ and $o$, $r_o-r_i$, in mutual Hill radius $R_{\rm H}=\frac{r_i+r_o}{2}\sqrt{\frac{2q}{3}}$ (the last column in Table~\ref{tab:planetarysystems}) is generally in the range of 3--6. This kind of planetary systems may be dynamically unstable on time scale shorter than the typical lifetime of disks \citep[e.g.,][]{smith09}, which is the subject to be addressed in the next section.
\section{Stability analysis}\label{sec:stability}
In this section, we will show that without gas damping, many of the systems of planets closely-spaced and massive enough to open deep and wide gaps are unstable. However, damping forces from residual gap gas can stabilize the systems and drive them into mean motion resonances, which can further enhance the stability. We demonstrate these conclusions through $N$-body simulations that take into account both planet-planet interactions (Section~\ref{sec:nogasdamping}) and various forms of gas damping forces (Section~\ref{sec:damping}).\footnote{Our
approach -- synthesizing planet-opened gaps based on hydrodynamical simulations and following with $N$-body calculations of planet-planet interactions that incorporate prescriptions for gas damping -- allows us to carry out thousands of simulations to assess the statistics of the stability and resonant behavior of these systems. Our two stage approach is adequate to study whether the systems remain stable and compact because the gas damping depends on the bulk properties of the gaps (e.g., order of magnitude of the gas surface density inside), which are insensitive to small variations in the planets' orbits (e.g, gaps remain largely the same for $\lesssim10\%$ changes to the semi-major axes and eccentricities of the sculpting planets).}
\subsection{No gas damping}\label{sec:nogasdamping}
To explore the stability of the planetary systems constructed in Section~\ref{sec:planetrequirements}, we carry out $N$-body simulations using {\tt mercury6} with the hybrid symplectic integrator \citep{Cham96}. In the first set of simulations presented in this subsection, we use purely $N$-body gravitational forces and neglect the effects of gas (i.e., planetary migration and eccentricity/inclination damping caused by planet-disk interactions are not included). For each multi-planet system in Table~\ref{tab:planetarysystems}, we perform 10 $N$-body simulations for an initial assessment. All simulations are run for 27 Myr, much longer than the typical disk lifetime. In each trial, planets are assigned an initial eccentricity $e=0$; a mean anomaly, and longitude of ascending node drawn randomly between 0 to 360$^\circ$; a semi-major axis drawn from a normal distribution with a median from Table~\ref{tab:planetarysystems} and a standard deviation 10\% the gap width; and an inclination $i$ drawn from a normal distribution with median $0$ and standard deviation $0.01 c_s/v_{\rm kep}$ rad, where $c_s = 1.29 {\rm km/s} \left(\frac{a}{\rm AU}\right)^{-1/4}$ is the sound speed and $v_{\rm kep}$ is the Keplerian velocity. If planets in any of the ten initial runs undergo instability, an additional set of 40 trials are carried out to more precisely determine the median unstable time scale $\tau_{\rm u}$ and the 1$^{\rm st}$ quartile unstable time scale $\tau_{\rm u,q}$ (i.e., when 25\% of the runs become unstable). A system is considered unstable if a planet is ejected $(a > 1000 {\rm AU})$, a planet's semi-major axis changes by more than 20\% due to scattering \footnote{For changes in a planet's semi-major axis greater than 20\% but less than 50\%, we visually inspect the orbital elements to ensure that scattering occurred rather than high amplitude oscillations.}, or two planets collide.
The outcomes of the $N$-body simulations are listed in Table~\ref{tab:stability}. Less than half of systems are stable for at least 27 Myr without the aid of gas damping forces. Figure~\ref{fig:orbitalevolution_stable} shows an example of a system that remains stable for $>$ 27 Myr (one trial of system {\tt 3-10a} in Table~\ref{tab:planetarysystems}) and contains three $q=0.01$ planets. Figure~\ref{fig:orbitalevolution_unstable} (left panel) shows a trial of system {\tt 5-2b} (containing five $q=0.002$ planets) that goes unstable shortly after 0.1 Myr. The planets scatter and two are ejected within 1 Myr. As expected from studies of orbit crossing timescale vs. spacing (e.g., \citealt{Cham96,Yosh99,Zhou07,Smit09}), systems with closer spacings (in units of $R_{\rm H}$) between planets or more planets tend to be more unstable. About 1/3 to 1/2 of unstable systems feature collisions between planets and most feature ejections of one or more planets. Many systems are left with only one or two giant planets, even those that started out with six.
In terms of gap characteristics, systems constructed for disks with larger $\delta_{\rm gap,min}$, $\alpha$, or $h/r$ tend to be less stable. This is expected, because for a given $M_{\rm p}$, larger $\delta_{\rm gap,min}$, $\alpha$, or $h/r$ produces a smaller $\Delta_{\rm gap,i}$. Therefore to maintain a given common gap size a system requires more planets with tighter spacings and is thus less stable. Additionally, for a given $\delta_{\rm gap,min}$, $\alpha$, and $h/r$, systems composed of less massive planets tend to be less stable. This is because $\Delta_{\rm gap,i}$ often decreases faster with decreasing $M_{\rm p}$ than $R_{\rm H}$; therefore the distance between adjacent planet pair shrinks in units of $R_{\rm H}$ as $M_{\rm p}$ goes down (e.g., decreasing $rmh$ with $M_{\rm p}$ for configurations {\tt 3-10b, 4-5c, 5-2b}, which all have $\delta_{\rm gap,min}=10$, $h/r=0.05$, $\alpha=0.01$). This trend is only marginal and does not hold in all cases. In addition, a system may need more planets to maintain the same common gap size as $\Delta_{\rm gap,i}$ decreases.
\subsection{Effects of gas on stability}\label{sec:damping}
Although the gas inside the gap is depleted, the remaining gas can still affect the dynamics and stability of the multi-planet systems through disk-planet interactions. Disk-planet interactions can alter the planet's eccentricity $e$ and inclination $i$ (typically damping them, e.g., \citealt{dunhill13}, but see also \citealt{tsang14,tsang14b,duffell15e}), semi-major axis $a$ (migration), or orbital orientation angles --- the longitude of periapse, $\varpi$, and the longitude of ascending node, $\Omega$ (precession). Gas damping and planetary migration can capture planets into mean motion resonance, and precession can alter the locations of the resonances. The possibility of resonant capture further motivates additional simulations that include the effects of gas because a resonant configuration can increase the stability timescale of a system, as explored for HR 8799 \citep{fabrycky10}, HL Tau \citep{tamayo15}, and LkCa~15 \citep{sallum15}.
We will first examine the effect of a non-zero $\dot{e}$ on the stability (Section~\ref{sec:dote}) and resonance capture (Section~\ref{sec:mmr}), followed by the effects of non-zero $\dot{a}$, $\dot{\varpi}$ and $\dot{\Omega}$ (Section~\ref{sec:dota}). We will conclude that if planets are responsible for the gaps, we expect to see many young planets in mean motion resonant configurations at the beginning of their post-gas-disk era evolution, and these configurations may persist for the older systems observable by non-direct imaging techniques (which we will discuss in Section~\ref{sec:disrupt} and Appendix~\ref{subsec:resage}).
\subsubsection{Non-zero $\dot{e}$ only}\label{sec:dote}
The strength of these effects depends on $\Sigma_{\rm gas}$ inside the gap, which is not well known. As discussed in Section~\ref{sec:gapdepletion}, in many cases current gas observations can only put a lower limit of $\delta_{\rm gap,min}\sim10$. Under this limit, the total gas mass within the gap is usually on the order of $M_{\rm J}$ or less \citep{vandermarel15}, which is small compared with the combined mass of the planets (usually $\sim3-40M_{\rm J}$, Table~\ref{tab:planetarysystems}). If the planets' self-stirring is dominated by close encounters (e.g., \citealt{Gold04}), gas damping is expected to shut off for $\Sigma_{\rm gas} < \Sigma_{\rm planets}$, but gas damping can still be important for widely-spaced planets on low-eccentricity orbits. (See e.g., Section~2 of \citealt{Daws15} for a discussion and \citealt{dawson15b} for examples.) This motivates us to explore the effects of gas eccentricity and inclination damping using simulations. In contrast, migration of planets that have opened gaps tends to be slow: type II migration may occur on a timescale comparable to the disk viscous timescale \citep[][but see also \citealt{duffell14}]{ward97}, and the type I migration timescale is longer than the eccentricity damping timescale by a factor of $(v_{\rm kep}/c_s)^2$. We do not include migration in the simulations in this subsection.
To explore the effects of gas damping, we use a customized version of {\tt mercury6} containing gas-damping forces (\citealt{dawson15b}, which makes use user-defined velocities and accelerations described in Appendix A of \citealt{Wolf12}). We adopt a gas surface density profile inside the gap of $\Sigma_{\rm gas}=\Sigma_{30} \left(\frac{a}{30\rm AU}\right)^{-3/2}$, with a fiducial normalization of $\Sigma_{30} = $~1 g cm$^{-2}$ at the outer edge of the gap, the median value based on current ALMA transitional disk observations \citep{vandermarel15} and corresponding to a depletion factor of 10 relative to the minimum mass solar nebula. We run additional sets of simulations with $\Sigma_{30} = \{10, 0.1, 0.01, 0.001, 0.0001\}$ g cm$^{-2}$. Following \citet{dawson15b}, we use three damping timescales, based on the regimes described in \citet{Papa00,Komi02,Ford07} and \cite{Rein12}:
\begin{eqnarray}
\label{eqn:damp}
\tau = 0.029 \frac{{\rm g\ cm}^{-2}}{\Sigma_{30}} \left(\frac{a}{\rm AU}\right)^2 \frac{M_\odot}{M_{\rm p}} {\rm yr} \times & \nonumber\\
& 1&, v < c_s \nonumber \\
&\left(\frac{v}{c_s}\right)^3&, v > c_s, i < c_s/v_{\rm kep} \nonumber \\
&\left(\frac{v}{c_s}\right)^4&, i > c_s/v_{\rm kep} \nonumber \\
\end{eqnarray}
\noindent where $v = \sqrt{e^2+i^2} v_{\rm kep}$ and $v_{\rm kep}$ is the Keplerian velocity. We impose $\dot{e}/e = -1/\tau$ and $\dot{i}/i = -2/\tau$ \citep{Komi02}.
Damping stabilizes planetary systems. The left panel in Figure~\ref{fig:orbitalevolution_unstable} shows the evolution of a five-planet system that was unstable without gas damping. With sufficient gas damping (right panel), eccentricities remain small and no scattering, collisions, or ejections occur. We show examples for four and six planet systems in Figures \ref{fig:orbitalevolution_unstable1} and \ref{fig:orbitalevolution_unstable2} respectively.
Next we assess the stability requirements for all the configurations developed in Section~\ref{sec:planetrequirements}. First we consider the criterion that at least 50\% of trials are stable, which we parametrize in terms of the minimum gas surface density normalization for 1 Myr stability $\Sigma_{30s,50\%}$. We list $\Sigma_{30s,50\%}$ for each system in Table \ref{tab:stability}. About half of configurations, including all the three planet configurations, are stable without gas damping. For other configurations, $\Sigma_{30s,50\%}$ ranges from 0.0001 g cm$^{-2}$ to 10 g cm$^{-2}$. Next we consider the stricter criterion that at least 90\% of trials are stable (parametrized as $\Sigma_{30s,90\%}$). About 1/3 of systems meet this criterion without gas damping. For other configurations, $\Sigma_{30s,90\%}$ ranges from 0.0001 g cm$^{-2}$ to 10 g cm$^{-2}$.
Even without explicit migration forces imposed (by default we set $\dot{a} = 0$ in our implementation of \citealt{Wolf12}, Appendix A), planets' semi-major axes can change as they planets repel each other. Figure~\ref{fig:orbitalevolution_unstable2}, right panel shows an example of repulsion in a six planet system. The planets are originally spaced from 3.5 to 26 AU. After 1 Myr with gas damping imposed, they are spaced from 3.0 to 30 AU. The repulsion may be a manifestation of the resonant-repulsion proposed for driving \emph{Kepler\ } super-Earths to period ratios wide of commensurability (e.g., \citealt{lithwick12}). Here the dissipative force comes from gas instead of tides. The mechanism operates even when no explicit $\dot{a}$ is imposed (e.g., \citealt{lithwick12}, Eqn. 22-23).
For many configurations, the repulsion is so strong with the fiducial value of $\Sigma_{30} = 1 $g cm$^{-2}$
that the planets cannot maintain a common gap for 1 Myr. These systems require a lower surface density, which we parametrize as $\Sigma_{30c}$, to avoid excessive repulsion (Table \ref{tab:stability}). About 1/3 of configurations remain compact (which we define as moving no further than one gap width apart) over 1 Myr with $\Sigma_{30c}=1$g cm$^{-2}$. Others require $\Sigma_{30c}\le 0.1$g cm$^{-2}$ or $\Sigma_{30c}\le 0.01$g cm$^{-2}$; these configurations tend to contain more massive planets.
For a few configurations ({\tt 6-2, 4-10cd}), too little gas means the system quickly goes unstable and too much gas means the planets repel each other excessively. The former value is $\Sigma_{30}=\{1, 0.01,0.01\}$ g cm$^{-2}$ for configurations \{{\tt 6-2, 4-10c, 4-10d}\} respectively and the latter value respectively is $\Sigma_{30}=\{10,0.1,0.1\}$ g cm$^{-2}$. We did not explore whether an intermediate value would allow for stability without excessive repulsion but if so, the conditions would require fine-tuning.
Overall, the range of $\Sigma_{30}$ that stabilizes the system without excessively repelling the planets is consistent with the plausible range of gas densities in the gap. The undepleted gas surface density at 30 AU ranges from about 10--100 g cm$^{-2}$ in observed transitional disks, interpolating from the gas density in the outer region (\citealt{vandermarel15}, Fig. 5). We multiply the 10--100 g cm$^{-2}$ range by the depletion factor at the bottom of the gap from the \citet{fung14} models (the last column in Table~\ref{tab:stability}) and compare to various $\Sigma_{30}$. We caution $\Sigma_{30}$ from our simulations is correct only to an order of magnitude due to approximations in the coefficients of Eqn. \ref{eqn:damp}. A handful of systems exhibit potential inconsistencies. For configurations {\tt 4-10a, 5-5ac}, the $\Sigma_{30} \le 0.01$ g cm$^{-2}$ required to avoid excessive repulsion is an order of magnitude or more lower than the expected density inside the gap. For configurations {\tt 6-2} and {\tt 6-1}, the large amount gas necessary to stabilize the system is several orders of magnitude larger than the expected gas density inside the gap.
In addition, an alternative form of eccentricity damping can further solve several of the inconsistencies described above. In theory planet-disk interactions can excite eccentricities instead of damping them, but the excitation is limited. For example, \citet{duffell15e} found that torques from a gas disk can excite a giant planet's eccentricity up to $\sim c_s/v_{\rm kep}$, above which eccentricity damping dominates. We ran a new set of simulations for Configuration {\tt 5-2b} in which we set the eccentricity damping to 0 when $v<c_s$. We found this damping to be insufficient to stabilize the system for $\Sigma_{30} = 0.1$ g cm$^{-2}$ but sufficient for $\Sigma_{30} = 1$ g cm$^{-2}$, which is still consistent with $\Sigma_{\rm gap}$ in Table \ref{tab:stability}. This damping still results in capture in mean motion resonances, which are discussed further below. For configurations {\tt 4-10a, 5-5ac}, $\Sigma_{30} = 0.1, 1, 1$ g cm$^{-2}$ respectively stabilizes the system while avoiding excess repulsion; this higher surface density is consistent with the expected surface density inside the gap.
Potentially gas outside the common gap could resist the resonant repulsion, but we do not expect this mechanism to be effective for configurations explored here. We expect the timescale to clear a new gap to be much less than the repulsion timescale so that the gap effectively moves with the planet, rather than the gas pushing the planet back toward the gap center. Furthermore, only a thin ring of gas with width on the order of the distance moved by the planet, typically a few AU, is expected to strongly interacts with the planet. The total mass in that gas ring is expected to be much smaller than the combined mass of the planets responsible for the repulsion. However, hydrodynamical simulations are necessary to fully explore the effect of the gas outside the common gap on resonant repulsion and the dependence on disk parameters. Because resonant repulsion only causes potential inconsistencies for a few configurations and the inconsistencies are solved by the alternative form of eccentricity damping, we leave such an exploration for future studies.
\subsubsection{Eccentricity damping driven capture into mean motion resonance}\label{sec:mmr}
Eccentricity damping drives the planets into mean motion resonances, defined as the libration of a resonant argument, $j_1 \lambda_o + j_2 \lambda_i + j_3 \varpi_o +j_4\varpi_i+j_5\Omega_o+j_6\Omega_i$ where $\{j_1,j_2,j_3,j_4,j_5,j_5\}$ is a set of integers that sums to 0. In the context of wide common gaps opened by multi-giant-planets, capture into resonances has been preliminarily shown in restricted parameter spaces with a small number of hydrodynamical simulations that included planetary sculpting of the gap, gravitational interactions among planets, and back reaction from the gas disk to the planets \citep[e.g.,][]{zhu11}.
In many of our simulations that include gas damping, the planetary systems end up configured in a chain of mean motion where each pair of adjacent planets are in mean motion resonance. As posited for \emph{Kepler\ } super-Earths, the period ratios are wide of resonance due to dissipation. In Figure~\ref{fig:damping}, we plot the resonant arguments of pairs of planets for the same system featured in Figure~\ref{fig:orbitalevolution_unstable}. The planets are captured into resonance on a $\sim$ 0.1 Myr timescale and then the libration amplitude grows. In column 4, row 1, the period ratio approaches 2.4 at 1 Myr yet the libration of the resonant argument persists. Libration occurs when the time derivative of the resonant argument, e.g., $2n_o + \dot{\xi_o} - n_i - \dot{\xi_i} - \dot{\varpi_i}$ where $n$ is the mean motion and $\xi$ is the mean longitude at epoch, is 0 (or oscillates about 0). Eccentricity damping produces a $\dot{\varpi_i, \xi_o, \xi_i}$ that allows $2n_o - n_i$ to deviate from 0, leading to non-commensurate period ratios.
We show more examples of resonant behavior in Appendix A and summarize the types of behavior in Table \ref{tab:res}. For each configuration, we show an example where $\Sigma_{30}$ allows for stability, avoids excess resonant repulsion, and is consistent (to within an order of magnitude) with the expected gas density in the gap (Table \ref{tab:stability}). Here we will use the term libration to refer to both true libration where the amplitude is bounded and angles that spend most of their time bounded but technically circulate through all values. In Fig. \ref{fig:app13} row 3, column 4 is an example of the former and column 1 of the latter. For every configuration (Table \ref{tab:res}, column 3) gas damping drives one or more pairs into libration of the 2:1 resonance. Higher multiplicity configurations are more likely to have the 2:1 resonance librating for all pairs. Other resonances, such as the 3:2 mean motion resonance, librate for a subset of configurations (e.g., Fig. \ref{fig:app19}). In many of the higher multiplicity configurations, one or more three-body resonant argument librates, including the Laplace resonance (e.g., Fig. \ref{fig:app14}) and others (e.g., the 3:4:1 angle in Fig. \ref{fig:app10} and the 3:5:2 angle in Fig. \ref{fig:app18}). However, we caution that even when an example configuration shown here includes libration of a three-body angle, other random versions of that configuration do not necessary include libration of the angle, so capture into a three-body resonance is not guaranteed. We find that the mean motion and three-body resonances generally persist after gas damping shuts off (Table \ref{tab:res}, final column).
Many pairs exhibit libration of the separation of periapses $\varpi_o-\varpi_i$. Unlike libration of the 2:1 resonance, libration of $\varpi_o-\varpi_i$ is not necessarily a signature of dissipation, particularly for the more widely-spaced planets, because much of the parameter space can librate for hierarchical systems (e.g., \citealt{michtchenko04}).
Finally, we note that although all configurations feature libration of a resonant argument, not all feature period commensurabilities ({\tt 3-10cd, 4-2}) and for many, a different resonant argument is librating than the commensurability. For example, the inner pair in {\tt 3-5} (Fig. \ref{fig:app1}) have a 4:1 period commensurability, but the 4:1 resonant arguments do not librate; instead, a 2:1 resonant argument librates. Many systems featuring libration of the 2:1 argument do not feature 2:1 commensurability. This poses a challenge for comparing to observations because the damped eccentricities are small (a few percent or less), making it difficult to measure $\varpi$ and hence determine whether resonant arguments are librating\footnote{Resonance libration timescales are quite long so rather than directly measuring the libration, we would need to constrain the orbital elements precisely enough to constrain the libration amplitude by forward integration.}.
\subsubsection{Non-zero $\dot{a}$, $\dot{\varpi}$ and $\dot{\Omega}$ in addition to $\dot{e}$}\label{sec:dota}
We run additional simulations with $\dot{a}$ (migration) and $\dot{\varpi}$ and $\dot{\Omega}$ (precession) imposed in addition to $\dot{e}$ but find the same qualitative behavior. Migration can enhance or undo resonant repulsion but requires fine tuning. For example, recall that in Configuration {\tt 4-10a}, the system went unstable with too little gas but repelled each other excessively with too much gas. Migration can counteract the repulsion (Fig. \ref{fig:mig}). However, the migration rate ($\dot{a}/a$) must be positive for the innermost planet and negative for the other planets. We fine-tuned the migration timescale to 0.9 Myr for the outer three planets and 9 Myr for the inner planet. For either Type I or Type II migration, we generally expect the migration timescale to increase with semi-major axis, so these rates would require fine-tuning of the disk conditions to produce the required magnitude and direction of migration.
In principle, precession ($\dot{\varpi}$, $\dot{\Omega}$) caused by the undepleted outer disk can affect the stability of system. Primarily the outer disk would cause the planetary orbits to precess, which can stabilize them against their own planet-planet secular interactions. However, in practice the precession timescale is slow to make a difference: on a 1 Myr timescale, the systems are destabilized by resonant and synodic planet-planet interactions rather than secular planet-planet interactions. For example, without precession caused by an outer disk, a simulation in the {\tt 4-10d} configuration with $\Sigma_{30} = 0.01 $ gcm$^{-2}$ went unstable on a timescale of 0.95 Myr and thus barely failed our stability criterion. Next we approximated the outer disk as a 10 Jupiter mass ring at 50 AU and added the precession to each orbit caused by this ring\footnote{We apply precession by assigning an equivalent stellar $J_2$ to each planet.}. The system goes unstable on a similar timescale (Fig. \ref{fig:precess}). On a much longer timescale than considered here, a remnant planetesimal disk could affect the system's stability (e.g., \citealt{thommes08,moore13}), but this does not affect the requirements for the gas disk stage.
\section{Comparison with the statistics of giant planets at large separations}\label{sec:statistics}
In this section, we consider the occurrence rates of giant planets at large separations, i.e., those that could be the survivors of young planetary systems that sculpted transitional disks in the gas disk era. We compare their occurrence rates (Section~\ref{sec:fplanet}) to the occurrence rate of transitional disks to determine whether there are enough giant planets to account for the occurrence rate of transitional disks (Section~\ref{sec:fplanetvsftd}). Finally, we connect the initial conditions of the multi-planet systems at the beginning of the post-gas-disk era to the observed properties of these systems at older stages (Section~\ref{sec:disrupt}).
As shown in Section~\ref{sec:planetrequirements} and \ref{sec:stability}, multiple giant planets with $q\sim10^{-4}-10^{-2}$ separated by $\sim$3-6 $R_{\rm H}$ are necessary to open a common gap as seen in transitional disks and usually get locked into mean motion resonance and are stable for as long as the gas disk is around (a few Myr). Such configurations may go unstable sometime after the gas disk dissipates so we cannot assume that the resonant chains persist until today. Instead, we conservatively assume that systems with giant planets in a wide range of configurations today may have produced transitional disks early in their history.
Thoroughly following the evolution of systems similar to those we simulated in Section~\ref{sec:stability} for Gyr after the dissipation of the gas disk is beyond the scope of this paper. We explore the post gas-disk evolution briefly in Section~\ref{sec:disrupt} but mostly draw conclusions from the outcomes of the generically spaced multi-giant systems simulated by \citet{juric08}. \citet{juric08} find that the final stable configurations for planets initially spaced by $\lesssim8 R_{\rm H}$ typically contain 2--3 planets (see also \citealt{chatterjee08}) and have final semi-major axes similar to their initial. We therefore synthesize the following necessary (but not sufficient) conditions for judging whether a planet could have been part of a common-gap-opening multi-planet system at its infancy: $q\geqq_{\rm gap}(\deltagapmin|\alpha,h/r)$, and 1~AU$\leq a\leq$50~AU. Below we will call the occurrence rate of such planets $f_{\rm p}(\delta_{\rm gap,min}|\alpha,h/r)$ (the fraction of stars with planets). Note that the specific choices of the minimum and maximum semi-major axis are not important, because the occurrence rate of giant planets outside the above semi-major axis range is low.
\subsection{The occurrence rate of planets at large separations, $f_{\rm p}$}\label{sec:fplanet}
The occurrence rate of giant planets at tens of of AU is not well constrained by current observations. The best method to detect such planets is direct imaging. However, direct imaging observations are challenging due to the low planet/star contrast ratio at these separations, and the detection limit has only recently reached the planetary mass regime. So far, only a handful of planet candidates have been discovered roughly in the parameter space explored here \citep[e.g.,][]{marois08, lagrange10, marois10, kuzuhara13, rameau13, macintosh15}. Combining non-detections and a handful of detections (including both planets and brown dwarfs) from several direct imaging surveys containing a few hundred stars, \citet{brandt14} proposed a single power-law distribution as a function of $M_{\rm p}$ and $a$,
\begin{align}
dN=0.057\%\timesM_{\rm p}^{-0.65\pm0.60} a^{-0.85\pm0.39} dM_{\rm p} da,
\label{eqn:brandt}
\end{align}
\noindent to account for the data, and concluded that 1.7\% (maximum likelihood) of stars host a 5-70~$M_{\rm J}$ companion between 10-100 AU, the most applicable range of the statistics.
On the other hand,
$f_{\rm p}$ at smaller separations of up to a few AU has been well constrained based on the results of radial velocity (RV) surveys. Using 8 years of RV data, \citet{cumming08} fit a power law distribution of the planet occurrence rate as a function of $M_{\rm p}$ and $a$,
\begin{align}
dN=0.74\%\timesM_{\rm p}^{-1.31\pm0.2} a^{-0.61\pm0.15} dM_{\rm p} da,
\label{eqn:cumming}
\end{align}
\noindent and concluded that 7.5\%\footnote{Note that the actual number quoted in \citet{cumming08} is 10.5\%. However this number is inconsistent with the power law indexes and normalization constant given in the paper, as well as the quoted planet assurance rates in other parameter spaces, e.g., their Table 1. The 7.5\% number quoted here is calculated directly using the power law indexes and normalization constant given by \citet{cumming08}.} of solar type stars have a $0.3<M_{\rm p}<10M_{\rm J}$ planet at $0.03<a<3$~AU. Long term radial velocity surveys are pushing to longer orbital periods. For example, \citet{wittenmyer16} find that the giant planet occurrence rate from 3--7 AU is consistent with an extrapolation of \citet{cumming08}. \citet{bryan16} found that occurrence rate of long period, giant planetary companions to known planets declines beyond $\sim 3-10$ AU, but it is unclear whether the companions of shorter period planets have the same period distribution as the general population. Since these recent studies do not have a sufficiently large sample to estimate a power law that we can extrapolate to 30 AU, we use the power law from \citet{cumming08}.
It is unclear to how large separations the $f_{\rm p}$ at small separations probed by the RV surveys can be extrapolated. This largely depends on whether widely separated giant planets are formed in the same way as giant planets as $\sim1-5$ AU, in which case a smooth and homogeneous distribution across the disk is expected. Although there is a general consensus that giant planets in the RV sample are probably formed through the core accretion scenario \citep{pollack96}, the conditions in a protoplanetary disk far from the star may not support core accretion. Both observational \citep[e.g.,][]{brandt14} and theoretical studies \citep[e.g.,][]{dodsonrobinson09, rice15} have proposed that massive companions (including both giant planets and brown dwarfs) beyond $\sim30$ AU may not be the long-period tail of the RV/core accretion samples. Instead, they may be formed through disk fragmentation \citep{boss98, rafikov05, kratter10}. Theoretical expectations for how $f_{\rm p}$ changes over a wide range of $a$ are uncertain at the moment.
Table~\ref{tab:mplimit} shows the expected percentage of (solar type) stars with planets,
$f_{\rm p}$, using the \citet{cumming08} power law (Eqn.~\ref{eqn:cumming}) for planets with $q_{\rm gap}(\deltagapmin|\alpha,h/r)<q<0.013$ at 1~AU$<a<50$~AU for each disk property combination [$\alpha,h/r$] and a minimum gap depletion factor $\delta_{\rm gap,min}$ (i.e., for the range of planet masses that would open deep enough caps for each set of disk properties). For comparison, we also list the corresponding $f_{\rm p}$ using the \citet{brandt14} power law (Equation \ref{eqn:brandt}). In general, $f_{\rm p}$ based on the direct imaging statistics is 5--10 times smaller than based on the RV statistics. We caution that neither Eqn.~\ref{eqn:brandt} nor \ref{eqn:cumming} fully applies to our parameter space, which is too distant compared to the RV sample and too low mass compared to the direct imaging sample.
\subsection{$f_{\rm p}$ $vs$ $f_{\rm TD,observed}$}\label{sec:fplanetvsftd}
In this section we compare the percentage of stars with giant planets that can open gaps, $f_{\rm p}$ (Section~\ref{sec:fplanet}), to the percentage of proto-planetary disks that are transitional disks, $f_{\rm TD,observed}$ (Section~\ref{sec:ftd}).
The conventional interpretation of transitional disks, as suggested by the class name, is that they represent a transient inside-out disk clearing phase at the end of the primordial/full disk stage that almost all protoplanetary disks undergo. Under this interpretation, the occurrence rate of transitional disks $f_{\rm TD,observed}$ in a homogeneous disk sample across all ages is simply the ratio of the disk dispersal timescale $\tau_{\rm dispersal}$ to the disk life timescale $\tau_{\rm life}$: $f_{\rm TD,observed}\approx\tau_{\rm dispersal}/\tau_{\rm life}$. Thus, a small $f_{\rm TD,observed}$ ($\approx$10\%) suggests that the clearing process is rapid. This interpretation dates back to \citet{skrutskie90}, who concluded that $\tau_{\rm dispersal}\sim$0.3 Myr, 10\% of $\tau_{\rm life}$ ($\sim$3 Myr), and has been adopted in the literature (e.g., \citealt{luhman10,koepferl13}).
However, a more generalized interpretation of $f_{\rm TD,observed}$ is as follows. Only a fraction of disks, $f_{\rm TD,intrinsic}$, ever go through a transitional disk phase, which lasts for $\tau_{\rm TD}$. The occurrence rate of transitional disks is determined by both factors, $f_{\rm TD,observed}\approxf_{\rm TD,intrinsic}\times\tau_{\rm TD}/\tau_{\rm life}$. The conventional interpretation represents one extreme: $f_{\rm TD,intrinsic}\approx1$ and $\tau_{\rm TD}\approx\tau_{\rm dispersal}\approx f_{\rm TD,observed}\times\tau_{\rm life}$.
In the other extreme, $\tau_{\rm TD}$ can be comparable to $\tau_{\rm life}$, and $f_{\rm TD,observed}\approxf_{\rm TD,intrinsic}$, meaning only on the order of 10\% of disks ever go through a long-lasting transitional disk phase. This possibility was raised by \citet[][see also \citealt{owen12clarke}]{muzerolle10}, who noted that if transitional disks do not represent a universal phase of disk evolution, then the conventional clearing timescale estimate $\tau_{\rm dispersal}=f_{\rm TD,observed}\times\tau_{\rm life}$ may be an underestimate.
If we assume that the majority transitional disks are disks with deep and wide gaps opened by planetary mass companions, we require $f_{\rm p} \gtrsim f_{\rm TD,intrinsic}$ (and by definition $f_{\rm TD,intrinsic}\geqf_{\rm TD,observed}$), because otherwise there are not enough planets to open gaps. Comparing $f_{\rm TD,observed}$ (Section~\ref{sec:ftd}) with $f_{\rm p}$ (Section~\ref{sec:fplanet}), the condition $f_{\rm p}\geqf_{\rm TD,observed}$ can only be satisfied under the most favorable interpretations:
\begin{enumerate}
\item The occurrence rate of giant planets throughout most of the 1--50 AU radius range must be greater or equal to an extrapolation of the RV planet distribution to wider separations, because the direct imaging planet distribution extrapolated to lower masses predicts far too few planets in this range.
\item The gap region must have disk conditions $h/r<0.1$ and $\alpha\lesssim0.01$ ( $\alpha\lesssim0.001$ if future observations determine $\delta_{\rm gap,min}$ is typically $\sim 1000$), because $f_{\rm p}$ with $h/r=0.1$ or $\alpha\gtrsim0.01$ is too small (due to the large planet masses required).
\item The conventional interpretation, that $f_{\rm TD,intrinsic}\approx1$ and $\tau_{\rm TD}\approxf_{\rm TD,observed}\times\tau_{\rm life}$, cannot hold, because giant planets at large separations are rare. $f_{\rm p}$ is only 23\% or $\sim2\timesf_{\rm TD,observed}$ under the most favorable conditions in the disk parameter space explored here: $\delta_{\rm gap,min}=10$, $\alpha=0.001$, and $h/r=0.05$. Instead, only a small fraction of disks, $f_{\rm TD,intrinsic}$, can go through the transitional disk phase, which must last for a significant fraction of the disk lifetime, $\tau_{\rm TD}\approx\tau_{\rm life}(f_{\rm TD,observed}/f_{\rm TD,intrinsic})$, while $f_{\rm TD,observed}/f_{\rm TD,intrinsic}$ is close to unity. Meanwhile most, if not all, giant planets must participate in transitional disk sculpting process when they are in gaseous disks.
\end{enumerate}
A prediction can be made based the above conclusions. Since most giant planets need to be at work in transitional disk sculpting, stars with more (fewer) giant planets should show a higher (lower) $f_{\rm TD,intrinsic}$ and $f_{\rm TD,observed}$ when they are in the gas disk era. Current data tentatively support this prediction in the case of stellar mass. \citet{andrews11} found while $f_{\rm TD,observed}\sim10\%$ for their entire sample of 91 disks in Taurus and Ophiuchus star-forming regions, big holes preferentially exist in mm bright disks, indicating higher disk mass and higher mass of the central stars \citep{andrews13}. \citet{montet14} found the occurrence rate of giant planets ($1-13M_{\rm J}$) at $a<20$ AU around lower-mass stars ($M$ dwarfs) is lower $(6.5\pm0.5\%)$ than around their higher mass counterparts, and \citet{shvartzvald16} found an occurrence rate of $5.0^{+4.0}_{-2.4}\%$ for Jupiters orbiting M-dwarfs in the $\sim$ 1.5--6 AU range. (See also consistent results for giant planets in much closer in orbits, \citealt{bowler10, johnson10}.) \citet{lada06} and \citet{downes15} find the fraction of stars with full (non-transitional) disks decreases with stellar mass, possibly because more massive stars are more likely to have giant planets to clear large cavities.
If future ALMA observations ascertain that the typical $\delta_{\rm gap,min}$ for transitional disk is 1000 or even higher, a tension will emerge (i.e., planets massive enough to open such deep gaps are not sufficiently common), as $f_{\rm p}$ decreases with increasing $\delta_{\rm gap,min}$. We will come back to a potential mitigation to this issue in Section~\ref{sec:observations}. A final caveat is that the transitional disk statistics from which \citet{luhman10} estimate $f_{\rm TD,observed}$ are for stars with a median stellar mass of $\sim 0.7 M_\odot$, whereas here we do not consider the stellar mass dependence of the giant planet occurrence rate.
\subsection{Transitional disks as constraints on initial conditions}
\label{sec:disrupt}
Here we discuss how transitional disks constrain the initial conditions for the post-gas evolution that establishes the architecture of planetary systems, and explore the 10 Gyr evolution of a single configuration as a case study.
We also investigate the implications of the initial conditions established during the transitional disk stage for \citet{koriski11}'s statistical study of resonances vs. stellar age in Appendix~\ref{subsec:resage}.
We have argued that the relative prevalence of transitional disks and rarity of giant planets means that most giant planet systems must have carved a transitional disk in their youth. Therefore the planet configurations that account for the observed characteristics of transitional disks (Section~\ref{sec:planetrequirements}) must comprise the initial conditions for the subsequent, gas-free dynamical evolution. This subsequent evolution is thought to play a major role in establishing the observed eccentricity distribution (e.g., \citealt{juric08}) and spawning hot and warm Jupiters through subsequent tidal evolution from highly elliptical orbits (e.g., \citealt{rasio96,nagasawa11,beauge12,dawson13}). The initial conditions for this evolution have been a major source of uncertainty. For example, \citet{juric08} state that the ``the theory is still too crude to allow'' initial conditions ``based on the predictions of planetary formation theory'' and draw planets' initial semi-major axes randomly from a log uniform distribution. \citet{beauge12} take an ``uncertain leap of faith'' and place planets in chains of first order mean motion resonances, inspired by migration simulations that aim to account for resonances observed in exoplanet systems (e.g., \citealt{snellgrove01}) or necessary in the early Solar System (e.g., \citealt{morbidelli07}) in the Nice model \footnote{A popular hypothesis to account for the dynamical structure of the Kuiper belt, e.g., \citep{tsiganis05}.}. In the transitional disk sculpting scenario we have explored in this paper, the properties of transitional disks serve as a helpful check-point for the ``initial'' (i.e., post gas-disk) conditions of giant planet systems. As shown in Section~\ref{sec:stability}, most of the suitable configurations feature a chain of planets in or near orbital resonances.
We showed in Section~\ref{sec:stability} that the resonant librations occurring in our simulations continue their libration if we remove gas damping and integrate for 1 Myr (Table \ref{tab:res}). To get a sense for how long the resonant configurations can survive after the gas disk's dissipation, we integrated ten versions of Configuration {\tt 5-2b} with $\Sigma_{30} = 0.1$ g/cm$^2$ for 1 Myr (Fig. \ref{fig:damping} is an example) and then for 10 Gyr without gas. Of those ten, three went unstable during the gas disk stage, three went unstable throughout the star's lifetime (at 63 Myr, 541 Myr, and 1.0 Gyr), and four remained stable for 10 Gyr. The seven systems that did not go unstable during the gas disk stage each feature between one and three (out of four) pairs within 10\% of 2:1 commensurability (with the other pairs within 20\% of commensurability). A detailed study of the post-gas stability timescales established by the configuration during gas disk stage is beyond the scope of this paper, but we preliminarily conclude that the resonant configurations sometimes survive to the present day but sometimes are disrupted on a wide range of timescales. Future studies could account for the range of gas disk lifetimes and the gradual dissipation of the gas disk, thoroughly determine the distribution of resonant disruption timescales, and focus on configurations that fulfill the more stringent constraints on $\delta_{\rm gap,min}$ expected from ongoing ALMA observations.
\section{Summary and discussion}\label{sec:summary}
We investigated the hypothesis that transitional disks are common gaps opened by multiple giant planets. We first synthesized the properties of the extended gaps in transitional disks based on ALMA gas observations (Section~\ref{sec:gapproperties}) and defined the multi-planet systems that can account for these properties (Section~\ref{sec:planetrequirements}). A series of minimally-packed planetary systems were constructed (Table~\ref{tab:planetarysystems}), and their stability was explored using $N$-body simulations (Section~\ref{sec:stability}, Table~\ref{tab:stability}). We then compared the occurrence rate of transitional disks, $f_{\rm TD,observed}$, with the extrapolated occurrence of giant planets at large separations, $f_{\rm p}$ and considered the subsequent dynamical evolution of the systems beyond the gas disk era (Section~\ref{sec:statistics}). Our main conclusions are:
\begin{enumerate}
\item To open a wide and deep gas gap around a solar type star consistent with observations, a system of 3--6 giant planets is indeed. The minimum planet mass increases with increasing disk scale height $h/r$ and viscosity $\alpha$ (as gaps become more difficult open), and increasing gap depth (Table~\ref{tab:mplimit}). The total number of planets inside a gap depends on the gap size: the planets must be placed close enough together to open a common gap. For given set of gap and disk properties, the less massive the planets, the larger the number and the closer the placement (Table~\ref{tab:planetarysystems}).
\item In general, without the aid from the gas damping, systems with a smaller number of more massive planets tend to be dynamically stable for the typical disk lifetime, while systems with a larger number of less massive planets may be unstable. Eccentricity damping from the residual gas inside the gaps can help stabilize systems by locking planets into mean motion resonances, establishing a chain of pairs resembling the older HR 8799 system \citep{fabrycky10}. However, in some cases eccentricity damping induces resonant repulsion that can drive planets away from commensurate period ratios, resulting in planets that are technically in resonance but not easily identifiable as such.
\label{con:stability}
\item The giant planet occurrence rate at wide separations $f_{\rm p}$ must equal or exceed the occurrence rate of transitional disks $f_{\rm TD,observed}\sim10\%$ for the planet-opening-common-gap scenario to remain viable.
This can only be satisfied under some of the most favorable conditions explored here, namely, $h/r<0.1$ and $\alpha<0.01$ in the gap regions; more importantly, the occurrence rate of giant planets at $\sim$1--50 AU has to largely follow the radial velocity statistics \citep{cumming08}, not the direct imaging statistics \citep{brandt14}. This situation may be significantly mitigated if the disk viscosity is lower than $10^{-3}$, enabling low mass planets to open gaps through non-linear wave damping processes (see discussion in Section~\ref{sec:alpha}).
\item The fact that $f_{\rm p}\sim25\%\sim2\timesf_{\rm TD,observed}$ under the most favorable conditions here ($h/r=0.05$, $\alpha=0.001$, $\delta_{\rm gap,min}=10$) implies that transitional disks are not a universal, fast disk dispersal phase at the end of protoplanetary disks' lifetimes, as assumed by most previous work dating back to \citet{skrutskie90}. Instead, the rarity of giant planets at large separations requires (1) most (if not all) giant planets to be at work in transitional disk sculpting in gaseous disks, (2) the transitional disk phase to be long-lasting, with a timescale comparable to typical disk lifetime, and (3) that only a small fraction ($\sim10\%$) of protoplanetary disks undergo this phase. If $f_{\rm TD,observed}$ increases with cluster age, the dependence reflects the time scale of giant planet formation instead of disk clearing.\label{con:fpvsftd}
\item The formation time scale for giant planets must be short (i.e., compared to the gas disk lifetime) so that they carve and maintain gaps for most of the disk lifetime. HL Tau, if its gaps are opened by giant planets, would be an example of fast planet formation; the system is believed to be $\lesssim 1$ Myr based on the youth of the Taurus cluster \citep{briceno02}.
\item As a consequence of the dynamical evolution of the multi-giant-planet systems inside the gaps, and the fact that most giant planets at large separations must participate in sculpting transitional disks in the gas disk era, most multi-giant systems have to be in mean motion resonances at the end of the gas disk era. The fraction of systems in resonance should decrease with time in the post gas disk era.
\item The properties of transitional disks serve as a helpful check-point for the ``initial'' (i.e., post gas-disk) conditions of giant planet systems. This subsequent evolution, for which the initial conditions have been a major source of uncertainty, is thought to play a major role in establishing the observed eccentricity distribution and spawning hot and warm Jupiters through subsequent tidal evolution from highly elliptical orbits (e.g., \citealt{rasio96,nagasawa11,beauge12,dawson13}).
\end{enumerate}
We re-emphasize that these conclusions only hold if the planetary sculpting hypothesis is responsible for most transitional disks, a mechanism with footings in both theory and observations (Section~\ref{sec:intro}), and in particular, supported by the recent detection of three companions inside LkCa~15's wide gap \citep{sallum15}.
In drawing these conclusions, we made a couple assumptions that simplified the analysis without losing much generality. We assumed a single planet mass for each configuration and when comparing $f_{\rm p}$ to $f_{\rm TD,observed}$, we did not take into account the different distribution of stellar masses in the samples. The current limited sample sizes of transitional disks and giant planets at large separations do not permit a detailed $f_{\rm p}$--$f_{\rm TD,observed}$ comparison by subgroups.
\subsection{Disk viscosity and low mass planets}\label{sec:alpha}
\label{sec:visc}
The tension between $f_{\rm p}$ and $f_{\rm TD,observed}$ can be significantly mitigated if the disk viscosity is lower than $10^{-3}$. For a given $\delta_{\rm gap,min}$ and $h/r$, lower $\alpha$ leads to a lower gap opening planet mass limit and thus more planets potentially capable of opening gaps. At least at close separations, the planet occurrence rate increases with decreasing planet mass \citep{howard10,mayor11}. Limited by the availability of the large systematic sets of gap opening simulations used to constructed our multi-planet systems \citep{fung14}, we did not go below $\alpha=10^{-3}$ and $h/r=0.05$. However, while $h/r$ can hardly go any lower than 0.05, the tension between $f_{\rm p}$ and $f_{\rm TD,observed}$ suggests that $\alpha$, a quantity that is hard to measure observationally and whose nature is not well understood theoretically, may span lower values in reality. We will argue below that a lower $\alpha$ can be consistent with recent observational and theoretical work.
As discussed in Section~\ref{sec:intro}, ALMA observations have found that the mm continuum ring in a few transitional disks is asymmetric (lopsided), with most of the emission coming from only one side of the disk \citep[e.g.,][]{vandermarel13, casassus13, perez14}. One interpretation is that these asymetric features result from vortex formation at the edge of planet-induced gaps, generated by the Rossby wave instability \citep[e.g.,][]{li00, lin10, lin12}, and dust trapping in vortices \citep[e.g.,][]{lyra09, heng10, zhu14votices, zhu14stone, lyra13}. A key ingredient in this scenario is a low viscosity. As \citet{zhu14stone} pointed out, $\alpha\lesssim10^{-3}$ is required to form vortices at the gap edge.
Other observational evidence comes from ALMA gas disk observations. First, non-detections of non-thermal motions induced by disk turbulence in a few systems suggest low viscosity. For example, based on ALMA CO observations of HD~163296, \citet{flaherty15} determined that the level of turbulence in this disk is lower than $3\%$ of the local sound speed, which implies $\alpha<10^{-3}$. Second, through detailed modeling of the gas gap structures and comparing with simulations of gap opening, \citet{vandermarel16} also proposed $\alpha\lesssim10^{-3}$ inside the gap of a few transitional disks.
On the theory side, low $\alpha$ in protoplanetary disks has recently gained some footing. The Magnetorotational instability (MRI) has been put forward as a prime candidate to provide disk viscosity. The operation of the MRI requires the disk to be sufficiently ionized and well coupled to the magnetic field. In the part of the disk that these conditions are not satisfied, a deadzone with no MRI and low viscosity is expected \citep{gammie96}. Magnetohydrodynamics (MHD) simulations and chemical disk modeling have shown than non-ideal MHD effects, in particular ambipolar diffusion, can significantly suppress MRI in disks at tens of AU, resulting in very low viscosity equivalent to $\alpha<10^{-3}$ at the bulk of the disk \citep{bai11, bai11stone, bai11-grains,perezbecker11-td, perezbecker11, simon13, turner14, bai15}.
The gap opening process in an extremely low viscosity environment with $\alpha<10^{-3}$ is different from the viscous disk case (i.e., the \citet{fung14} models used here), as the gap is now opened by nonlinear evolution, instead of the viscous damping, of waves. Nonlinear evolution of waves was explored by \citet{goodman01} and \citet{rafikov02}, who predicted that even planets much less massive than the thermal mass, $M_{\rm th}=c_{\rm s}^3/G\Omega_{\rm p}$ (about Saturn mass at 15 AU), can open gaps. Their theory was later confirmed by numerical simulations of disk-planet interactions in (nearly) inviscid disks \citep{li09, muto10, dong11-linear, dong11-nonlinear, duffell13, zhu13}. We note that gap opening by low mass planets in nearly inviscid disks can be slow; it has not yet been explored whether the timescale is consistent with observed disk lifetimes. Finally, if future observational and numerical studies demonstrate that disk viscosity is generally very low (e.g., $\alpha\lesssim10^{-4}$), {\it and} this low viscosity enables Earth and super-Earth-like planets to participate in the transitional disk sculpting processes, {\it and} the abundance of such planets found by Kepler (e.g., \citealt{howard12}) extends to tens of AU, we caution that the conclusion made in Section~\ref{sec:fplanetvsftd} -- namely the $\sim10\%$ occurrence rate of transitional disks reflects nature not nurture -- may be altered.
\subsection{Connections to ongoing observations}\label{sec:observations}
In this paper we discussed gaps with $\delta_{\rm gap,min}=10$, a conservative limit set by most current observations, as well as $\delta_{\rm gap,min}=10$ and 100, more aggressive limits put forward by pioneering studies with the latest ALMA results \citep{vandermarel16}. As additional attenae are commissioned for ALMA,
future observations with finer angular resolution and better sensitivity may systematically push $\delta_{\rm gap}$ to as high as $\sim10^3$. If such a large depletion is confirmed, even with the most gap-friendly disk properties in our models ($\alpha=0.001$ and $h/r=0.05$) and RV planet statistics at large separations, there may not be enough giant planets capable of opening such deep gaps to account for $f_{\rm TD,observed}$ (Table~\ref{tab:mplimit}).
Therefore the planetary sculpting hypothesis would only be viable for the ultra-low viscosity $\alpha<10^{-3}$ discussed in Section~\ref{sec:visc} and not considered in the configurations explored in the paper.
As we concluded most giant planets have to be at work in gap opening, stars with more (fewer) giant planets should show a higher (lower) $f_{\rm TD,intrinsic}$ and $f_{\rm TD,observed}$. This expected correlation is tentatively supported by the data for different stellar masses (Section~\ref{sec:fplanetvsftd}). Future measurements of $f_{\rm TD,observed}$ and $f_{\rm p}$ for subgroups of stars can further test this prediction.
We concluded that most giant planets had to participate in gap opening when they were in protoplanetary disks and gas damping established stable resonant configurations. For some configurations (particular those with four or more massive planets), the period ratios are commensurate and the system could be identifiable as a resonant configuration. The young HR~8799, with all four planets likely in a chain of 2:1 resonances \citep{fabrycky10}, is an excellent example. For other configurations (particularly those with only three planets), we found resonant repulsion drives the planets away from recognizable commensurability. We found that each of the configurations we studied features libration of the 2:1 resonant argument and many feature libration of three-body resonances (e.g., Laplace). The resonance librations persist after gas damping shuts off but may be difficult to measure from observations because the eccentricities are small. After the protoplanetary disk phase, the fraction of multi-giant systems in resonance is expected to decrease over time (Section~\ref{sec:disrupt}), a prediction that can be more robustly tested in the future with a sample planets whose host stars have well-determined ages that span multiple orders of magnitude.
The two major ongoing direct imaging surveys with VLT/SPHERE \citep{beuzit08} and Gemini/GPI \citep{macintosh08} are expected to establish much better statistics of giant planets at tens of AU from the star, resolving the question of whether giant planets are sufficiently common for planetary sculpting to be predominant cause of transitional disks.
\section*{Acknowledgments}
We thank the referee, John E. Chambers, for a constructive report that improved the quality and the clarity of the paper. R.D. thanks Xue-Ning Bai for teaching him about MRI in disks, and Ewine van Dishoeck and Nienke van der Marel for educating him on the topic of ALMA disk observations. We also thank Sean Andrews, Tim Brandt, Sourav Chatterjee, Eugene Chiang, Paul Duffell, Misato Fukagawa, Andrea Isella, John Johnson, Heather Knutson, Chalie Lada, Renu Malhotra, Rebecca Martin, and Ben Montet for insightful discussions. We particularly thank Jeffery Fung and Paul Duffell for kindly sharing the simulation data in \citet{fung14} and \citet{duffell15dong} with us. This project is partially supported by NASA through Hubble Fellowship grant HST-HF-51320.01-A (first R.D.) awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555, and by the Miller Institute for Basic Research in Science (second R.D.). Simulations were run on the SAVIO computational cluster provided by Berkeley Research Computing.
|
\section{Introduction}
Direct detection of gravitational waves (GWs) was achieved for the first time in 2015 \cite{LIGO}
by the advanced Laser Interferometer Gravitational wave Observatory (aLIGO) \cite{TheLIGOScientific:2014jea}.
This discovery brought a great impact on science, announcing the onset of gravitational wave astronomy.
Following aLIGO, the large-scale cryogenic gravitational wave telescope (LCGT) now known as KAGRA, is being constructed in Kamioka, Japan \cite{KAGRA3}.
It is the first detector in deep underground and expected to provide useful knowledges for future detectors.
In addition, it will improve determination accuracy of a gravitational wave source direction on the sky together with aLIGO and Virgo, and help to obtain more information from GWs.
Because gravity is the weakest force among the four elementary
interactions known to the present, GWs have a high penetrating power.
Thanks to the smallness of their cross section,
GWs are hard to be influenced by interstellar medium
during the propagation unlike electromagnetic waves.
GWs bring information on deep inside of compact stars such as
neutron stars directly to us. By the same token,
this property makes its detection very difficult.
To extract weak GW signals buried in detector noises,
a commonly used method is the matched filtering technique
which is based on the maximum likelihood method
assuming stationary, Gaussian noises \cite{GAUSSIANmethod}.
However, non-Gaussian noises frequently appear in actually measured output data.
They would hinder the sensitivity of the matched filter,
which is known to be susceptible to non-Gaussian noises
stemming from instrumental and environmental artifacts, and results in false alarms.
One approach to deal with non-Gaussian noises is
to appropriately modify the functional form of the likelihood \cite{Rover:2008yp, Rover}.
In the previous paper, one of us introduced the likelihood function based on
the Edgeworth expansion for weak non-Gaussian noise
and the Gaussian mapping method for strong non-Gaussian noise \cite{first}.
In this paper, we propose another method to mitigate the effect of
non-Gaussian noises using the independent component analysis (ICA) \cite{ICA1,ICA2,ICA3}
(see \cite{icab1,icab2} for textbooks).
The ICA has been developed in the context of blind source separation
among which the cocktail party problem is well-known as a representative example.
The ICA enables us to decompose output data into statistically independent components
on the assumption that there is at most one Gaussian component in the data.
Here, we apply the ICA method to data analysis for burst-like GW signals
and investigate how well the ICA works to
separate stationary non-Gaussian noise from output data.
The rest of the paper is organized as follows. In \S 2.1, we consider the simplest case where the detector outputs are linear combinations of the inputs. Next, the time delay between the inputs and the outputs, which exists in real experiments, is taken into account in \S 2.2. Finally, we study the applicability of our method to nonlinearly correlated noises in \S 2.3. The last section \S 3 is devoted to our conclusion.
\section{Independent Component Analysis (ICA)}
As is seen in the previous paper \cite{first}, signal detection under non-Gaussian
noises is much involved than the case with Gaussian counter parts since the optimal
statistic has a more complicated form. Interestingly, however, there
have been some proposals to make use of non-Gaussian natures of both
signals and noises to separate
signals from non-Gaussian noises known as the
independent component analysis (ICA).
Here we consider applicability of this new approach for the detection of
GWs in a simple model.
Suppose that there exist $N$ statistically
independent sources of signals and $N$ outputs
and that there is at most one source which follows the Gaussian
distribution and all the other sources obey non-Gaussian
distributions.
The ICA is a method to identify
each independent source of signals making use of the non-Gaussianity and
statistical independence of sources. So here non-Gaussianity is not an obstacle
of the analysis but rather a necessary ingredient. Indeed, if
multiple sources obey Gaussian distributions, we cannot distinguish
them using ICA even if they are statistically independent.
{\bf 2.1 The simplest model.}~~~Here let us consider a simple problem as a first step of realistic
application of ICA to the detection of GWs.
To be specific, let us identify two sources of signals $s_1(t)$,
and $s_2(t)$ as a burst-like gravitational wave signal $h(t)$ and non-Gaussian
noise $k(t)$ such as seismic noises. That is,
\begin{equation}
{\vect s}(t)=\begin{pmatrix} s_1(t)\\ s_2(t)\end{pmatrix}=\begin{pmatrix} h(t)\\ k(t)\end{pmatrix}.
\end{equation}
In addition to the output from the laser interferometer $x_1(t)$,
we make use of the output from an environmental monitor such as a seismograph $x_2(t)$, and assume that
they are linear functions of the signal ${\vect s}(t)$ as
\begin{equation}
{\vect x}(t)=\begin{pmatrix} x_1(t)\\ x_2(t)\end{pmatrix}=A{\vect s}(t) \label{vs}
\end{equation}
where $A$ is assumed to be a time independent matrix.
By definition, the gravitational wave signal obeys a probability
distribution function (PDF)
\begin{equation}
r_1(h,t)=\delta (h-h(t,\theta))
\end{equation}
where $h(t,\theta)$ is the actual waveform of gravitational radiation
emitted from some source, say a binary neutron star coalescence, to be
observed at the position of a laser interferometer, where
$\theta$ collectively denotes parameters of the source.
On the other hand, we do not specify the PDF of $k(t)$, $r_2(s_2)$, except that
it is a super-Gaussian distribution such as a Student t-distribution
with a larger tail than Gaussian \cite{Characterization}.
The detector output of a laser interferometer, of course, suffers from
Gaussian noises $n(t)$ besides non-Gaussian noise $k(t)$.
Hence (\ref{vs}) should actually read
\begin{equation}
{\vect x}(t)=A{\vect s}(t)+{\vect n}(t),~~~{\vect n}(t)=\begin{pmatrix}
n(t)\\0\end{pmatrix}.
\label{linearmodel}
\end{equation}
Here we have not incorporated any Gaussian noise to the second line
where the signal $k(t)$ itself consists of (non-Gaussian) noises and
any Gaussian noise can be incorporated to a part of it.
We now introduce a trick to replace the original source $s_1(t)=h(t)$
by $s_1(t)=h(t)+n(t)$, that is, we regard the Gaussian noise as a part
of the original signal. Since $n(t)$ is a Gaussian noise with vanishing mean,
its statistical property is entirely characterized by the two-point
correlation function $K(t-t')=E[n(t)n(t')]$. Then the
marginal distribution function of $s_1(t)$ is given by
\begin{equation}
r_1[s_1(t)]=\frac{1}{\sqrt{2\pi}\sigma}\exp\left[ -\frac{1}{2\sigma^2}
\left( s_1(t)-h(t,\theta)\right)^2\right],~~~\sigma^2=K(0). \label{gauss}
\end{equation}
Thus $s_1(t)$ now satisfies a simple Gaussian distribution which is much
easier to handle with than the delta-function distribution (\ref{vs}),
and ${\vect s}(t)$ and ${\vect x}(t)$ are related by a simple
formula ${\vect x}(t)=A{\vect s}(t)$. Now our tentative goal is to find the inverse
matrix of $A$ whose components are not known precisely.
One may set it as
\begin{equation}
A=\begin{pmatrix} a_{11}& a_{12}\\ 0 & a_{22}\end{pmatrix}, \label{aform}
\end{equation}
since the gravitational wave is so weak that it will not affect any
seismograph.
The aim of ICA is to find a linear transformation
\begin{equation}
{\vect y} = W{\vect x},
\end{equation}
such that each component of the transformed variables ${\vect y}$
is mutually statistically independent.
Thanks to the assumption (\ref{aform}), the matrix $W$ also takes a form
\begin{equation}
W=\begin{pmatrix} w_{11}& w_{12}\\ 0 & w_{22}\end{pmatrix}. \label{wform}
\end{equation}
If we knew all the components of $A$, the matrix $W$ could simply be given
by the inverse matrix $W=A^{-1}$, in which case we would find
${\vect y} ={\vect s}$. However, since we do not know them
we attempt to determine $W$ in such a way that the components of ${\vect y}$,
$y_1(t)$ and $y_2(t)$ to be statistically independent with each other as much as
possible.
The mutual independence of statistical variables may be judged by
introducing a cost function $L(W)$ which represents a ``distance'' in the
space of statistical distribution functionals. As an example, we
adopt the Kullback-Leibler divergence \cite{KL} defined between two arbitrary
PDFs $p({\vect y})$ and $q({\vect y})$ as
\begin{equation}
D[p({\vect y}); q({\vect y})]=\int p({\vect y})\ln \frac{p({\vect y})}{q({\vect y})}dy
=E_p\left[ \ln \frac{p({\vect y})}{q({\vect y})}\right],
\end{equation}
where $E_p[\cdot]$ denotes an expectation value with respect to a PDF $p$.
We examine the distance between the real distribution function of
statistically independent variables ${\vect s}$,
$r({\vect s})=r_1[s_1(t)]r_2[s_2(t)]$,
and a distribution of ${\vect y}$, $p_y$,
constructed from the observed distribution function of ${\vect x}$ through
the linear transformation ${\vect y}=W{\vect x}$ as
\begin{equation}
p_y({\vect y})\equiv ||W^{-1}||p_x({\vect x}),
\end{equation}
where $||W^{-1}||$ denotes the determinant of $W^{-1}$.
The cost function of $p_y({\vect y})$ from $r({\vect s})$ is given by
\begin{align}
L_r(W) &= D[p_y({\vect y});r({\vect y})]=E_{p_y}[\ln p_y({\vect y})]-E_{p_y}[\ln r({\vect y})]
\nonumber \\
&= \int ||W^{-1}||p_x({\vect x})\ln \left[ ||W^{-1}||p_x({\vect x})\right] dy
-E_{p_y}[\ln r({\vect y})] \nonumber \\
&= -\ln ||W|| + \int p_x({\vect x})\ln \left[ p_x({\vect x})\right] dx
- E_{p_y}[\ln r({\vect y})] \nonumber \\
&= -H[x]- E_{p_y}[||W||\ln r({\vect y})] = -H[x]- E_{p_x}[\ln p({\vect x},W)], \label{56}
\end{align}
with
\begin{equation}
p({\vect x},W) \equiv ||W|| r({\vect y}),
\end{equation}
and
\begin{equation}
H[x]\equiv -\int p_x({\vect x})\ln \left[ p_x({\vect x})\right] dx.
\end{equation}
The PDF of ${\vect x}$ in the last expression of (\ref{56})
has $W$ dependence because $p({\vect x},W)$ is
a PDF of ${\vect x}$ which is made out of the PDF of ${\vect y}$ ($={\vect s}$ in this
particular case) through the relation ${\vect y}=W{\vect x}$.
The above formula shows that the matrix $W$ which minimizes the cost
function $L_r(W)$ also maximizes the log-likelihood ratio of ${\vect x}$.
Since we do not know $r({\vect y})$ a priori, we instead adopt an arbitrary
mutually independent distribution $q({\vect y})=q_1(y_1)q_2(y_2)$
in the cost function.
Defining a PDF consisting of marginal distribution functions
\begin{equation}
\tilde{p}({\vect y})\equiv \int p_y(y_1,y_2)dy_2\int p_y(y_1,y_2)dy_1=
\tilde{p}_{1}(y_1)\tilde{p}_{2}(y_2),
\end{equation}
we find the following relation
\begin{equation}
L_q(W)=D[p_y({\vect y});q({\vect y})]=D[p_y({\vect y});\tilde{p}({\vect y})]+
D[\tilde{p}({\vect y});q({\vect y})]
\end{equation}
holds. Since the Kullback-Leibler divergence is known to be
positive semi-definite, a distribution that minimizes the first term
in the right-hand-side yields the desired linear transformation
${\vect y}=W{\vect x}$ for which this term vanishes. In this case the second
term gives a discrepancy due to the possible incorrect choice of $q({\vect y})$.
In this sense it would be better to choose a realistic trial function
$q({\vect y})$ as much as possible.
It is known in fact that even for an arbitrary choice of $q({\vect y})$,
the correct $W$ gives an extremum of $L_q(W)$.
Hence we solve
\begin{equation}
\frac{\partial L_q(W)}{\partial w_{ij}}=0. \label{partial}
\end{equation}
From
\begin{equation}
L_q(W)=-H[x]-\ln||W||-E_{p_y}[\ln q({\vect y})]\equiv
-H[x]-\ln||W||-E_{p_y}[f({\vect y})],
\end{equation}
\begin{equation}
f({\vect y}) \equiv \ln q({\vect y}),
\end{equation}
\begin{align}
d_W\ln||W|| &\equiv \ln||W+dW||-\ln||W||=\ln||{\vect 1}+dWW^{-1}|| \nonumber \\
&={\rm Tr}(dWW^{-1})=(W^{-1})_{ji}dw_{ij}, \nonumber
\end{align}
and
\begin{align}
d_W f({\vect y}) &\equiv f\left( (W+dW){\vect x}\right) -f(W{\vect x})
=\frac{\partial f}{\partial y_i}dw_{ij}x_j, \nonumber
\end{align}
we find
\begin{align}
d_W L_q(W)=E_{p_y}\left[ -(W^{-1})_{ji}-x_j\frac{\partial f}{\partial
y_i}\right] dw_{ij}.
\end{align}
In order to satisfy (\ref{partial}) the above expectation value should
vanish for each index.
Multiplying $w_{kj}$ to the argument of the expectation value, we find
it equivalent to
\begin{equation}
E_{p_y}\left[ y_k\frac{\partial f}{\partial y_i}\right] +\delta_{ki}=0 .
\end{equation}
That is, we require
\begin{equation}
E_{p_y}[\varphi_i(y_i)y_j]=\delta_{ij}
\end{equation}
with
\begin{equation}
\varphi_i(y_i)\equiv -\frac{d~}{dy_i}\ln q_i(y_i). \label{independence}
\end{equation}
We determine $W$ so that (\ref{independence}) is satisfied for each
component choosing plausible forms of $q_i(y_i)$. For $q_1(y_1)$ we take
\begin{equation}
q_1(y_1)=\frac{1}{\sqrt{2\pi}\sigma}\exp\left[ -
\frac{(y_1-h(t,\theta))^2}{2\sigma^2}\right] ,
\end{equation}
based on (\ref{gauss}), so that
$\varphi_1(y_1)=(y_1-h(t,\theta))/\sigma^2$.
In real experiments, we do not know $h(t,\theta)$ a priori. However, as found later, when we take temporal average, the contributions from gravitational waves can be neglected. Therefore, we can set $h(t,\theta)=0$ when we apply $q_1(y_1)$ to real analysis.
As for $\varphi_2(y_2)$, it is recommended to take
\begin{equation}
\varphi_2(y_2)=c_2\tanh y_2
\end{equation}
to model a super-Gaussian distribution \cite{SA}.
Using these expressions in (\ref{independence}) we determine $W$
which relates each component of ${\vect y}$ and ${\vect x}$ as
$y_1=w_{11}x_1+w_{12}x_2$ and $y_2=w_{22}x_2$.
In doing so we replace the ensemble average $E[\cdot]$ by temporal
average of observed values of ${\vect x}$ which we denote by brackets.
Each component of (\ref{independence}) reads as follows.
\begin{align}
1&=
\sigma^{-2}E[(w_{11}x_1+w_{12}x_2-h)(w_{11}x_1+w_{12}x_2)]\nonumber \\
&=\sigma^{-2}\left[ w_{11}^2\langle x_1^2\rangle +2w_{11}w_{12}\langle x_1x_2\rangle
+w_{12}^2\langle x_2^2\rangle -(w_{11}\langle h x_1\rangle+ w_{12}\langle h x_2\rangle)\right], \label{11}\\
0&=\sigma^{-2}E[(w_{11}x_1+w_{12}x_2-h)w_{22}x_2] \nonumber \\
&= \sigma^{-2}\left[ w_{11}w_{22}\langle x_1x_2\rangle +w_{12}w_{22}\langle x_2^2\rangle
-w_{22}\langle h x_2\rangle\right], \label{12}\\
0&=c_2 E[y_1 \tanh y_2]=c_2 E\left[ (w_{11}x_1+w_{12}x_2)\tanh
w_{22}x_2\right]\nonumber \\
&= c_2w_{11} \langle x_1\tanh (w_{22}x_2) \rangle +c_2 w_{12}\langle x_2
\tanh( w_{22}x_2)\rangle, \label{21}\\
1&=c_2 E[y_2\tanh y_2]=c_2 w_{22}\langle x_2 \tanh( w_{22}x_2) \rangle. \label{22}
\end{align}
Because a gravitational wave with a detectable amplitude is a rare
event, long-time averages of $h x_1$ and $h x_2$ should vanish.
Then from (\ref{12}) we obtain
\begin{equation}
w_{12}= -\frac{\langle x_1x_2\rangle}{\langle x_2^2 \rangle}w_{11}, \label{1211}
\end{equation}
and from (\ref{11})
\begin{equation}
\sigma^2=w_{11}^2\langle x_1^2\rangle +w_{11}w_{12}\langle x_1 x_2\rangle,
\end{equation}
so that
\begin{equation}
w_{11}=\left(\frac{\langle x_2^2\rangle}{\langle x^2_1\rangle \langle x_2^2\rangle - \langle x_1
x_2\rangle^2}\right)^{\frac{1}{2}}\sigma~~~{\rm and}~~
w_{12}=-\frac{\langle x_1x_2\rangle \sigma}{(\langle x_1^2\rangle \langle x_2^2\rangle^2-
\langle x_1x_2\rangle \langle x_2^2\rangle)^{\frac{1}{2}}}. \label{1112}
\end{equation}
On the other hand, (\ref{21}) yields a relation
\begin{equation}
w_{12}=-\frac{\langle x_1\tanh(w_{22}x_2)\rangle}{\langle
x_2\tanh(w_{22}x_2)\rangle}w_{11}.\label{tanh}
\end{equation}
It will be found later that this condition is consistent with (\ref{1211}).
We now apply the likelihood ratio test to the above result.
The output of a laser interferometer $x_1$ consists of
\begin{equation}
x_1=a_{11}s_1+a_{12}s_2=a_{11}h+a_{11}n+a_{12}k. \label{78}
\end{equation}
If there were only Gaussian noises with $a_{12}=0$,
we would find
\[
E_{p_x}[x_1] =a_{11}h,~~ E_{p_x}[x_1^2] - (E_{p_x}[ x_1 ])^2=
a_{11}^2E[n^2]
\]
so that
\begin{equation}
\frac{S}{N}=\frac{a_{11}h}{a_{11}\sqrt{E[ n^2]}}
=\frac{h}{\sqrt{E[n^2]}}.
\end{equation}
In the presence of non-Gaussian noise
instead we find
\begin{equation}
E_{p_x}[ x_1^2 ] - (E_{p_x}[ x_1 ])^2=a_{11}^2E[ n^2 ]+a_{12}^2 E[ k^2]
\end{equation}
assuming $E[ nk ] =0$. As a result signal to noise ratio ($S/N$) gets worse
\begin{equation}
\frac{S}{N}=\frac{h}{\sqrt{E[ n^2]+\frac{a_{12}^2}{a_{11}^2}
E[k^2]}}.
\end{equation}
Now we calculate $S/N$ of $y_1$ variable from
\begin{equation}
E_{p_y}[ y_1 ] =w_{11}E_{p_x}[ x_1] +w_{12}E_{p_x}[ x_2] =w_{11}a_{11}h
,\nonumber
\end{equation}\begin{equation}
E_{p_y}[ y_1^2 ] -(E_{p_y}[ y_1])^2=w_{11}^2a_{11}^2E[ n^2] +
(w_{11}a_{12}+w_{12}a_{22})^2E[ k^2] .\nonumber
\end{equation}
If $W$ is solved exactly, it should be identical to the inverse matrix
of $A$, namely,
\begin{equation}
\begin{pmatrix} w_{11}&w_{12}\\ 0&w_{22}\end{pmatrix}
=\begin{pmatrix} \frac{1}{a_{11}}& -\frac{a_{12}}{a_{11}a_{22}}\\
0& \frac{1}{a_{22}}\end{pmatrix},
\end{equation}
then we find
\begin{equation}
E_{p_y}[ y_1 ] =w_{11}E_{p_x}[ x_1] +w_{12}E_{p_x}[ x_2] =h, \nonumber
\end{equation}\begin{equation}
E_{p_y}[ y_1^2] -(E_{p_y}[ y_1])^2=E[ n^2]. \nonumber
\end{equation}
But in fact we cannot hope to obtain $W$ as an inverse matrix, but
instead
estimate it from temporal average of the observed sample.
From (\ref{1211}), (\ref{tanh}), and (\ref{78}), we find
\begin{equation}
w_{12}=-\frac{\langle x_1x_2\rangle}{\langle
x_2^2\rangle}w_{11} \approx -\frac{a_{12}a_{22}\langle
k^2\rangle}{a_{22}^2\langle k^2\rangle}w_{11}=-\frac{a_{12}}{a_{22}}w_{11},
\end{equation}
and
\begin{eqnarray}
w_{12} &=& -\frac{\langle x_1\tanh(w_{22}x_2)\rangle}{\langle
x_2\tanh(w_{22}x_2)\rangle}w_{11} \nonumber \\
&\approx&
-\frac{ a_{12} \langle k \tanh(w_{22}a_{22}k) \rangle }{a_{22} \langle k \tanh(w_{22} a_{22} k) \rangle} w_{11}\nonumber \\
&=&
-\frac{a_{12}}{a_{22}} w_{11},
\end{eqnarray}
which means these two equations are consistent with each other.
We also find that, even from the observational data, we can deduce
$w_{11}a_{12}+w_{12}a_{22}=0$ and $y_1$ variable is indeed free from
non-Gaussian noise. The resultant $S/N$ of $y_1$ is
simply given by
\begin{equation}
\frac{S}{N} \approx \frac{h}{\sqrt{E[ n^2]}}.
\end{equation}
Thus, the non-Gaussian noise is effectively removed here.
{\bf 2.2 memory effect.}~~~The above simple model (\ref{vs}) is just the first step to analyze
realistic detectors. In fact, $x_1(t)$ would depend not only on
$s_2(t)$ but also on some retarded times as well thanks to operation
of a sophisticated anti-vibration system.
In order to incorporate such a memory effect, we employ the
following
model as the second step.
\begin{equation}
{\vect x}(t)=\sum_{\tau=0}^{\Theta-1} A(\tau){\vect s}(t-\tau). \label{memory}
\end{equation}
In this subsection, all the time variables are dimensionless
being expressed in unit of the measurement interval.
The above expression corresponds to
a model such that ${\vect x}(t)$ depends on ${\vect s}(t)$ up to $\Theta$ time steps before
the measurement time $t$.
That is, we assume that the correlation between ${\vect x}(t)$ and ${\vect s}(t-\tau)(\tau\ge\Theta)$ is negligible.
Taking data from $t=t_s$ to $t=t_s+T-1$, $\left\{{\vect x}(t)\lvert t=t_s,t_s+1,...,t_s+T-1\right\}$, and working in the Fourier space
\begin{equation}
{\vect x}(t)=\frac{1}{T} \sum_{N=0}^{T-1}\tilde{{\vect x}}(\omega_N;t_s)e^{i\omega_N t},~~~
\omega_N=\frac{2\pi}{T}N,
\end{equation}
we find
\begin{align}
\tilde{{\vect x}}(\omega_N;t_s)&=\sum_{t=t_s}^{t_s+T-1}{\vect x}(t)
e^{-i\omega_N t}
\nonumber \\
&=\sum_{t=t_s}^{t_s+T-1}\sum_{\tau=0}^{\Theta-1}A(\tau){\vect s}(t-\tau)e^{-i\omega_N t} \nonumber \\
&=\sum_{\tau=0}^{\Theta-1}A(\tau)e^{-i\omega_N \tau} \sum_{t=t_s}^{t_s+T-1}{\vect s}(t-\tau)e^{-i\omega_N(t-\tau)}.
\end{align}
If $T \gg \Theta$, the sum over $t$ can be approximated by
\begin{equation}
\sum_{t=t_s}^{t_s+T-1}{\vect s}(t-\tau)e^{-i\omega_N(t-\tau)} \thickapprox \sum_{t=t_s}^{t_s+T-1}{\vect s}(t)e^{-i\omega_N t} = \tilde{{\vect s}}(\omega_N;t_s).
\end{equation}
Thus, when we take a long time-series compared to $\Theta$ for the Fourier expansion, the following equality holds for each Fourier mode.
\begin{equation}
\tilde{{\vect x}}(\omega_N;t_s)=\tilde{A}(\omega_N)\tilde{{\vect s}}(\omega_N;t_s),~~~
\tilde{A}(\omega_N) \equiv \sum_{\tau=0}^{\Theta-1}A(\tau)e^{-i\omega_N \tau}.
\end{equation}
To derive this relation, we have assumed that $A$ is independent of $t_s$.
Then for each Fourier mode, $A$ and $W$ take the following form
\begin{equation}
{\tilde A}(\omega_N)=\begin{pmatrix} {\tilde a}_{11}(\omega_N) & {\tilde a}_{12}(\omega_N) \\
0 & {\tilde a}_{22}(\omega_N)\end{pmatrix},~~~
{\tilde W}(\omega_N)=\begin{pmatrix} {\tilde w}_{11}(\omega_N) & {\tilde w}_{12}(\omega_N) \\
0 & {\tilde w}_{22}(\omega_N)\end{pmatrix}.
\end{equation}
Since the normalization of $y_i$ is arbitrary, one can put
${\tilde w}_{11}(\omega_N)={\tilde w}_{22}(\omega_N)=1$.
Calculating the Fourier components for various values of $t_s$, for example $t_s=0,T,2T,...,MT (\mathrm{M~is~an~integer.})$, and
following the same argument
as in the case of the simplest model, we find
\begin{equation}
{\tilde w}_{12}(\omega_N)=-\frac{\langle x_1(\omega_N;t_s)x_2(\omega_N;t_s)\rangle_{t_s}}
{\langle x_2^2(\omega_N;t_s)\rangle_{t_s}},~~~
{\tilde w}_{11}(\omega_N)=-\frac{\langle x_1(\omega_N;t_s)x_2(\omega_N;t_s)\rangle_{t_s}}
{\langle x_2^2(\omega_N;t_s)\rangle_{t_s}} \label{w11}
\end{equation}
where $\langle \cdot \rangle_{t_s}$ denotes an average with respect to $t_s$.
So the variable
\begin{equation}
\tilde{y}_1(\omega_N;t_s)=\tilde{x}_1(\omega_N;t_s)+{\tilde w}_{12}(\omega_N)\tilde{x}_2(\omega_N;t_s)
\end{equation}
follows a Gaussian distribution, and we can find the increased $S/N$
for each frequency $f=\omega/2\pi$ as in the previous subsection.
{\bf 2.3 Nonlinear coupling.}~~~So far we have considered the linear models (\ref{linearmodel}) and (\ref{memory}). However, in real gravitational wave experiments, there exist non-linearly correlated noises \cite{nonlinearnoise}. In order to investigate the applicability of ICA to nonlinear cases, we consider the following simple nonlinear model as a first step:
\begin{equation}
\begin{pmatrix} x_1(t)\\ x_2(t)\end{pmatrix}=\begin{pmatrix}a& b\\ 0 & 1\end{pmatrix}\begin{pmatrix}h(t)+n(t)\\ k(t)\end{pmatrix} + \begin{pmatrix} c[h(t)+n(t)]k(t) \\ 0 \end{pmatrix} \label{NLcoupling}
\end{equation}
Without loss of generality, we can set the covariances of $n(t)$ and $k(t)$ to be unity by redefining $a$, $h(t)$, $b$, and $x_2(t)$:
\begin{equation}
E[n^2(t)]=E[k^2(t)]=1 \label{covariance} ,
\end{equation}
and we define $n(t)$ so that
\begin{equation}
a \ge 0. \label{a_condition}
\end{equation}
In addition, we consider the case where $|b|$ is not much larger than $a$ and $|c|$ (see the argument above (\ref{alpha_evo2}) and (\ref{gamma_evo2})).
By assumption, the marginal PDF of $n(t)$ is the normal distribution,
\begin{equation}
q_1(n)=\frac{1}{\sqrt{2\pi}}\exp\left( -
\frac{n^2}{2}\right),
\end{equation}
whereas the PDF of $k$, $q_2(k)$, is not known.
Our goal is to remove the non-Gaussian noise $k(t)$.
If we know the coefficients, $a$, $b$, and $c$, we can obtain the time series without non-Gaussian noise, $h(t)+n(t)$ by using the transformation
\begin{equation}
h(t)+n(t)=\frac{x_1(t)-b x_2(t)}{a + c x_2(t)}. \label{removeNG}
\end{equation}
Therefore we estimate the values of $a$, $b$ and $c$.
When we estimate them, we consider a long-time average of the data. Assuming that the detectable burst gravitational waves rarely come to our detectors, we can neglect them. Therefore we may consider a simplified model
\begin{equation}
\begin{pmatrix} x_1(t)\\ x_2(t)\end{pmatrix}=\begin{pmatrix}a& b\\ 0 & 1\end{pmatrix}\begin{pmatrix}n(t)\\ k(t)\end{pmatrix} + \begin{pmatrix} cn(t)k(t) \\ 0 \end{pmatrix} \label{NLcoupling2}
\end{equation}
to estimate the coefficients.
To begin with, the coefficient $b$ can be estimated easily by
\begin{equation}
b_{est} = \langle x_1(t)x_2(t) \rangle, \label{b_est}
\end{equation}
because
\begin{equation}
E[n(t)k(t)]=E[n(t)]E[k(t)]=0 ,
\end{equation}
\begin{equation}
E[n^2(t)k(t)] = E[n^2(t)]E[k(t)] =0 .
\end{equation}
The next targets are $a$ and $c$.
To find their values, we consider a transformation similar to (\ref{removeNG}),
\begin{equation}
y_1(t)=\frac{x_1(t)-b_{est} x_2(t)}{\alpha + \gamma x_2(t)},~~~y_2(t)=x_2(t) \label{NLtransf}
\end{equation}
and regard $y_1(t)$ as the reconstructed $h(t)+n(t)$. They are very close to each other when $\alpha = a$ and $\gamma = c$. As the Jacobian of the transformation (\ref{NLtransf}) is given by
\begin{equation}
J = \left| \frac{\partial (y_1,y_2)}{\partial (x_1,x_2)} \right| = \left| \frac{1}{\alpha + \gamma x_2} \right|,
\end{equation}
the cost function can be expressed as
\begin{eqnarray}
L(\alpha, \gamma) &=& D[p_y({\vect y});q_1(y_1)q_2(y_2)] \nonumber \\
&=&E_{p_y}\left[\ln\left(\frac{p_y({\vect y})}{q_1(y_1)q_2(y_2)}\right)\right] \nonumber \\
&=&E_{p_x}\left[\ln\left(\frac{J^{-1}p_x({\vect x})}{q_1(\frac{x_1-b_{est} x_2}{\alpha + \gamma x_2})q_2(x_2)}\right)\right] \nonumber \\
&=&\frac{1}{2}E_{p_x}\left[\left(\frac{x_1-b_{est} x_2}{\alpha + \gamma x_2}\right)^2\right]+E_{p_x}\left[\ln|\alpha + \gamma x_2|\right] + \mathrm{const.}. \label{NLcost}
\end{eqnarray}
We minimize it with respect to $\alpha$ and $\gamma$. The derivatives of the cost function read
\begin{equation}
\frac{\partial L}{\partial \alpha} = E_{p_x}\left[ \frac{(\alpha + \gamma x_2)^2 - (x_1 - b_{est} x_2)^2}{(\alpha + \gamma x_2)^3} \right], \label{alpha1}
\end{equation}
\begin{equation}
\frac{\partial L}{\partial \gamma} = E_{p_x}\left[ \frac{x_2(\alpha + \gamma x_2)^2 - x_2(x_
1 - b_{est} x_2)^2}{(\alpha + \gamma x_2)^3} \right], \label{gamma1}
\end{equation}
and their learning rules are $\Delta \alpha \propto - \frac{\partial L}{\partial \alpha}$ and $\Delta \gamma \propto - \frac{\partial L}{\partial \gamma}$. However, the cost function is mathematically ill-defined because of the singularity at $\alpha + \gamma x_2 = 0$. Therefore we must improve these rules. We consider the following modified learning rules,
\begin{equation}
\Delta \alpha \propto - E_{p_x}[(\alpha + \gamma x_2)^2 - (x_1 - b_{est} x_2)^2], \label{alpha2}
\end{equation}
\begin{equation}
\Delta \gamma \propto -E_{p_x}[x_2(\alpha + \gamma x_2)^2 - x_2(x_1 - b_{est} x_2)^2],
\label{gamma2}
\end{equation}
which are obtained by removing the denominators of (\ref{alpha1}) and (\ref{gamma1}). From (\ref{NLcoupling}) and (\ref{covariance}), the rules can be written as
\begin{equation}
\Delta \alpha \propto -[\alpha^2 - a^2 - (b-b_{est})^2 + \gamma^2 - c^2], \label{alpha_evo}
\end{equation}
\begin{equation}
\Delta \gamma \propto -2(\alpha \gamma - ac) + [(b - b_{est})^2 - \gamma^2 +c^2] \epsilon, \label{gamma_evo}
\end{equation}
where we have defined $\epsilon$ by $\epsilon = E_{p_x}[k^3(t)]$.
Now we assume $|b - b_{est}| \ll a,|c|$, which means that we do not consider the case where $|b|$ is too large and the number of the samples is too small for us to extract the information about $a$ and $c$. In this case, (\ref{alpha_evo}) and (\ref{gamma_evo}) can be written as
\begin{equation}
\Delta \alpha \propto - [ \alpha^2 + \gamma^2 - a^2 - c^2], \label{alpha_evo2}
\end{equation}
\begin{equation}
\Delta \gamma \propto - [ 2(\alpha \gamma - ac) + \epsilon (\gamma^2 - c^2)], \label{gamma_evo2}
\end{equation}
respectively.
We study these learning rules and show that they tell us where the point ($a$,$c$) is.
First, we locate their stationary points, which satisfy $\Delta \alpha = 0$ and $\Delta \gamma = 0$. The curves defined by $\Delta \alpha = 0$ and $\Delta \gamma = 0$ are depicted in Fig. \ref{POI}. They are symmetrical with respect to the lines $\ell^{\pm}$ defined by the following equations,
\begin{equation}
\ell^+: \gamma = \left(\frac{(\epsilon^2 + 4)^{\frac{1}{2}} + \epsilon}{(\epsilon^2 + 4)^{\frac{1}{2}} - \epsilon} \right)^{\frac{1}{2}} \alpha, \label{l+}
\end{equation}
\begin{equation}
\ell^-: \gamma = - \left(\frac{(\epsilon^2 + 4)^{\frac{1}{2}} - \epsilon}{(\epsilon^2 + 4)^{\frac{1}{2}} + \epsilon} \right)^{\frac{1}{2}} \alpha. \label{l-}
\end{equation}
We can easily check that $(\alpha, \gamma) = (a,c)$ satisfies $\Delta \alpha = \Delta \gamma = 0$. Therefore, the points satisfying $\Delta \alpha = \Delta \gamma = 0$ are $(a,c)$ and those which are symmetrical to $(a,c)$ with respect to $\ell^+$, $\ell^-$ and the origin.
\begin{figure}
\begin{center}
\includegraphics[width = 10cm]{Fig1.eps}
\caption{The curves defined by $\Delta \alpha = 0$ and $\Delta \gamma = 0$ with $a=1$, $c=0.5$, $\epsilon =1$. They are symmetrical with respect to $\ell^+$ and $\ell^-$. The intersections are the stationary points of (\ref{alpha_evo2}) and (\ref{gamma_evo2}). The convergence point is one of them.}
\label{POI}
\end{center}
\end{figure}
However, all these points are not necessarily stable points. In order to locate the convergence point of (\ref{alpha_evo2}) and (\ref{gamma_evo2}), we must also study the evolution of $(\alpha, \gamma)$. Focusing on the sign of $\Delta \alpha$ and $\Delta \gamma$, we depict the flow schematically in Fig. \ref{FLOW}. We refer to the stationary point on the region defined by
\begin{equation}
- \left(\frac{(\epsilon^2 + 4)^{\frac{1}{2}} - \epsilon}{(\epsilon^2 + 4)^{\frac{1}{2}} + \epsilon} \right)^{\frac{1}{2}} \alpha < \gamma < \left(\frac{(\epsilon^2 + 4)^{\frac{1}{2}} + \epsilon}{(\epsilon^2 + 4)^{\frac{1}{2}} - \epsilon} \right)^{\frac{1}{2}} \alpha \label{REGION}
\end{equation}
as C$(\alpha_c,\gamma_c)$, and to the symmetrical points to C with respect to $\ell^{\pm}$ as $\mathrm{C}^{\pm}$. We can easily find that the stationary points other than C are not stable. For example, if the point $(\alpha, \gamma)$ goes in the negative direction of $\alpha$-axis from $\mathrm{C}^{\pm}$, it goes to infinity. Therefore, the point $(\alpha, \gamma)$ never converges to them. On the other hand, C is a stable point. In order to show that, we substitute $\alpha = \alpha_c + \delta \alpha$ and $\gamma = \gamma_c + \delta \gamma$ into (\ref{alpha_evo2}) and (\ref{gamma_evo2}), and keep leading terms, to find
\begin{equation}
\begin{pmatrix} \Delta (\delta \alpha)\\ \Delta (\delta \gamma)\end{pmatrix} = -2 A \begin{pmatrix} \delta \alpha\\ \delta \gamma\end{pmatrix},
\end{equation}
\begin{equation}
\mathrm{with} ~~~~ A = \begin{pmatrix} \alpha_c & \gamma_c \\ \gamma_c & \alpha_c + \epsilon \gamma_c \end{pmatrix}.
\end{equation}
From (\ref{REGION}), we can easily show that the matrix $A$ has only positive eigenvalues and so C is a stable point.
In addition, if the point $(\alpha, \gamma)$ is constrained on the region defined by (\ref{REGION}), it does not go to infinity and tends to approach C.
Because $\epsilon$ can be estimated from the data, we can return the point onto that region by displacing it with respect to $\ell^{\pm}$ or the origin.
Therefore, with the point $(\alpha,\gamma)$ displaced properly while it is evolving, it converges to C.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[clip, width=7cm]{Fig2-a.eps}
\hspace{1.6cm} [a]$2ac + \epsilon c^2 > 0$ \\($a = 0.5$, $c = -1$, $\epsilon = 1.5$)
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[clip, width=7cm]{Fig2-b.eps}
\hspace{1.6cm} [b]$2ac + \epsilon c^2 < 0$ \\($a = 0.5$, $c = -1$, $\epsilon = 0.8$)
\end{center}
\end{minipage}
\end{tabular}
\caption{The flow of (\ref{alpha_evo2}) and (\ref{gamma_evo2}). It is drawn only by taking the signs of (\ref{alpha_evo2}) and (\ref{gamma_evo2}) into account, which means the lengths of the arrows are not important. The large dots represent the stationary points and the points other than C are not stable points. It can be seen that points in the region satisfying (\ref{REGION}) approach C.}
\label{FLOW}
\end{center}
\end{figure}
Because C is just the stationary point on the region defined by (\ref{REGION}) and not necessarily coincide with $(a,c)$, there are three patterns depending on the location of $(a,c)$ shown in Fig. \ref{where}.
\begin{itemize}
\item Region 1: converges to the point symmetrical to $(a,c)$ with respect to $\ell^-$.
\item Region 2: converges to $(a,c)$.
\item Region 3: converges to the point symmetrical to $(a,c)$ with respect to $\ell^+$.
\end{itemize}
In each of these cases, the desired point $(a,c)$ is, respectively, $\mathrm{C}^-$, C, $\mathrm{C}^+$.
The coordinates of $\mathrm{C}^{\pm}$ are given respectively by
\begin{equation}
\mathrm{C}^+: \frac{1}{(\epsilon^2 + 4)^{\frac{1}{2}}}(-\epsilon \alpha_c + 2 \gamma_c , 2 \alpha_c + \epsilon \gamma_c), \label{plus}
\end{equation}
\begin{equation}
\mathrm{C}^-: \frac{1}{(\epsilon^2 + 4)^{\frac{1}{2}}} (\epsilon \alpha_c - 2 \gamma_c , -2 \alpha_c - \epsilon \gamma_c). \label{minus}
\end{equation}
Because the point $(a,c)$ is on the region $a \ge 0$ (see (\ref{a_condition})), it is one of the points $\mathrm{C}$, $\mathrm{C}^+$ and $\mathrm{C}^-$ which is on the right-half plane of the $\alpha-\gamma$ plane. We can identify the coefficients $a$ and $c$ by investigating which point makes $y_1$ follow Gaussian distribution.
\begin{figure}
\begin{center}
\includegraphics[width = 8cm]{Fig3.eps}
\caption{The right-half plane devided by $\ell^+$ and $\ell^-$. From the condition (\ref{a_condition}), $(a,c)$ is on one of these regions. On the other hand, the convergence point C is on the Region 2. Therefore whether the point $(\alpha,\gamma)$ converges to $(a,c)$ depends on which region $(a,c)$ is on.}
\label{where}
\end{center}
\end{figure}
We summarize how we obtain $a$, $b$, and $c$.
\begin{enumerate}
\item After subtracting the means of the time series from them, transform $x_2(t)$ so that the equation (\ref{covariance}) is satisfied:
\begin{equation}
x_2(t) \to \frac{1}{\langle x_2(t)^2 \rangle^{\frac{1}{2}}} x_2(t).
\end{equation}
\item Calculate $b_{est}$ by (\ref{b_est}).
\item Evolve $(\alpha, \gamma)$ following (\ref{alpha_evo2}) and (\ref{gamma_evo2}) while constraining the point on the region defined by (\ref{REGION}).
\item Calculate the coordinates of $\mathrm{C}^+$ and $\mathrm{C}^-$ by (\ref{plus}) and (\ref{minus}), and choose the best point that is on the right-half plane of $\alpha-\gamma$ plane and maximizes the Gaussianity of $y_1(t)$.
\end{enumerate}
In order to demonstrate that the algorithm works well, we make statistically independent samples of $x_1$ and $x_2$ (ignoring their temporal correlations) and test it. We assume that the probability density function of the non-Gaussian noise $k$ is the Laplace distribution function whose variance is unity and mean is zero:
\begin{equation}
q_2(k)=\frac{1}{\sqrt{2}} \mathrm{e}^{-\sqrt{2} |k|}.
\end{equation}
We generate 10,000 samples of $n$ and $k$, and calculate
\begin{equation}
x_1 = an+bk+cnk
\end{equation}
\begin{equation}
x_2 = k
\end{equation}
for each of the samples to make 10,000 samples of $x_1$ and $x_2$.
When we set $a=1$, $b=1$, $c=2$, the evolutionary flow on the $\alpha-\gamma$ plane is shown in Fig. \ref{TEST}.
The coordinates of $\mathrm{C}$, $\mathrm{C}^{\pm}$ are $\mathrm{C}(1.98, 1.13)$, $\mathrm{C}^+(1.08, 2.01)$ and $\mathrm{C}^-(-1.08,-2.01)$. The estimated point of $(a,c)$ is $\mathrm{C}$ or $\mathrm{C}^+$.
We choose the point which makes the kurtosis of $y_1$ smaller, then find that the estimated values of $a$, $b$, and $c$ are $a_{est}=1.08$, $b_{est}=1.00$, and $c_{est}=2.01$, which are close to the real values.
\begin{figure}
\begin{center}
\includegraphics[width = 12cm]{Fig4.eps}
\caption{The evolutionary flows of (\ref{alpha_evo2}) and (\ref{gamma_evo2}) with $a=1$, $b=1$, $c=2$. The initial points are set to be (1,0), (2,0), (3,0). For each regeneration, we use only one sample to calculate the right hand side of (\ref{alpha_evo2}) and (\ref{gamma_evo2}), and do this procedure successively for each sample (This is called online algorithm.). For the $i$-th regeneration, the constant of proportionality is 0.01/(1 + $i$/1000), which begins to decrease at about the 1000th regeneration. The estimated value of $\epsilon$ is 0.05 and the lines, $\ell^{\pm}$, are drawn by using this value. As can be seen, the flows from all the initial locations converges to the same point. Because $\mathrm{C}^-$ is not on the right-half of the plane, the estimated point is $\mathrm{C}$ or $\mathrm{C}^+$. We use both values to construct $y_1$ according to (\ref{NLtransf}) and calculate each kurtosis of them, which is zero when the samples are normally distributed. We choose the point which makes the kurtosis smaller and find that the estimated values of $a$, $b$, and $c$ are $a_{est}=1.08$, $b_{est}=1.00$, and $c_{est}=2.01$.}
\label{TEST}
\end{center}
\end{figure}
Finally, we can remove the non-Gaussian noise $k(t)$ by using $y_1(t)$ instead of $x_1(t)$. By using (\ref{removeNG}), we can find that S/N is improved from
\begin{equation}
\frac{S}{N}=\frac{h}{\sqrt{E[ n^2 ]+ \frac{b^2}{a^2} E[ k^2 ] + \frac{c^2}{a^2} (h^2 + E[ k^2 ]) E[ n^2 ]}}
\end{equation}
to
\begin{equation}
\frac{S}{N}\approx\frac{h}{\sqrt{E[ n^2]}},
\end{equation}
as $y_1(t)$ reproduces $h(t)+n(t)$ well.
\section{Conclusion}
\label{conclusion}
In the present paper, we have provided a new way to handle non-Gaussian noises
for the detection of GWs using the independent component analysis.
While many other methods attempt to overcome the non-Gaussian nature of noises
by frontal attack, ICA makes use of non-Gaussianity as well as statistical independence,
to separate signals.
First, we have considered a simplified linear model to show that this method
may be useful to remove non-Gaussian noises that can be measured by
enviromental sensors at least partially. Next, we have checked the applicability of our method to
a more realistic case where there is time delay between the inputs and the outputs. We have shown that we can also remove non-Gaussian noises by
using Fourier components.
These methods can be applied to the case which has
outputs of more channels.
Finally, we have tried to extend our method to nonlinearly
correlated noises.
As a first step, we have considered a specific nonlinear model and show that we can identify the coupling coefficients and remove non-Gaussian noises in that case.
It is worth while to pursue further
analysis by considering more realistic situations.
Thus we shall prepare
for the completion of KAGRA detector toward gravitational wave astronomy.
\vskip 2cm
\noindent
{\large\bf Acknowledgements}
This work was partially supported by
a research program of the Advanced Leading Graduate Course for Photon Science (ALPS) at the University of Tokyo (SM), the JSPS Grants-in-Aid for Scientific Research (KAKENHI) 15H02082(JY), 24103005 and 15K05070 (YI), and JSPS Fellows Grants No. 26.8636 (KE).
|
\section{Introduction}
\label{sec:introduction}
We consider the multiterminal secret key agreement problem formulated by \cite{csiszar04}. A group of users, each observing a private correlated random source, discuss
in public so they can agree on a secret key. The key is a random variable that needs to be recoverable by
every user after the discussion. Furthermore, the key must be secured against a wiretapper who can
observe the entire discussion among the users but has no access to their private sources.
The maximum achievable secret key rate is called the \emph{secrecy capacity}. It was characterized by
\cite{csiszar04} as a linear program. Because the wiretapper can listen to the entire public
discussion, \emph{the randomness of the secret key can only come from the information mutual to the
private correlated source components.}
Indeed, in the two-user case, it was shown in \cite{csiszar04} that the capacity is equal to
\emph{Shannon's mutual information}:
\begin{align}
I(\RZ_1\wedge \RZ_2) &= D(P_{\RZ_1\RZ_2}\|P_{\RZ_1}P_{\RZ_2}),\label{eq:I}
\end{align}
where each user $i\in \Set{1,2}$ observes the discrete memoryless correlated private source $\RZ_i$. The mutual information above is written as the \emph{divergence} $D$ from the joint distribution $P_{\RZ_1\RZ_2}$ to the product of the marginal distributions $P_{\RZ_1}$ and $P_{\RZ_2}$.
In the multiterminal case, let $V$ be the set of (two or more) users, and $\RZ_V:=(\RZ_i\mid i\in V)$ be a random vector where $\RZ_i$ is a discrete memoryless source component privately observed by user $i\in V$. There was a divergence upper bound on the capacity in~\cite{csiszar04}, which was identified~\cite{chan2008tightness,chan10md} to be tight in the special case without helpers, giving rise to the alternative capacity characterization:
\begin{subequations}
\begin{align}
I(\RZ_V) &:= \min_{\mcP\in \Pi'(V)} I_{\mcP}(\RZ_V), \kern2em \text{where}\label{eq:MMI}\\
I_{\mcP}(\RZ_V)&:=\frac{D(P_{\RZ_V}\|\prod_{C\in \mcP} P_{\RZ_C})}{\abs {\mcP}-1},\label{eq:IP}
\end{align}
\end{subequations}
and $\Pi'(V)$ is the collection of partitions $\mcP$ of $V$ into two or more non-empty disjoint sets.
Following \cite{chan15mi},
we call \eqref{eq:MMI} the \emph{multivariate mutual information (MMI)}. It is easy
to see that \eqref{eq:I} is a special case of \eqref{eq:MMI} when $V=\Set{1,2}$. Indeed, the MMI was
formally regarded in \cite{chan15mi} as a measure of mutual information among multiple random
variables, thereby extending various interpretations and properties of Shannon's mutual information to the multivariate case. The MMI has other operational meanings, ranging from
tree-packing~\cite{nitinawarat10}, hypergraph connectivity~\cite{chan10md}, undirected network
coding~\cite{chan11isit}, vocality constraints~\cite{mukherjee14,mukherjee15,zhang15}, successive omniscience~\cite{chan16so,ding15} and data
clustering~\cite{chan15allerton,chan16cluster}.
In this work, we want to study how the MMI of a set of random variables changes with respect to
changes in the MMI of a subset of the random variables. We formulate two new problems, called the
\emph{incremental secret key agreement (ISKA)} and \emph{decremental secret key agreement (DSKA)}.
In ISKA, a subset of users is given an additional common randomness in the form of a random source of
certain entropy rate. The objective is to find an efficient resource allocation, i.e., to increase
the capacity as much as possible without requiring too much common randomness to be added to too
many users. In DSKA, we remove some common randomness from a subset of users. The objective is to simplify the source model, but without reducing the capacity much.
In particular, we want to identify redundant common randomness whose removal does not diminish the capacity.\footnote{The idea of redundant common randomness first appeared in \cite[Proposition~2.3]{chan10md}. It was called excess edge under a hypergraphical source model, and was related to the notion of partition connectivity for hypergraphs. The idea was also used in \cite{courtade16} in characterizing the minimum communication complexity for secret key agreement, but under a non-asymptotic hypergraphical source model when the communication protocol is restricted to be linear. \cite{MKS16,mukherjee16,chan16itw} further considered the non-asymptotic case and derived more efficiently computable bounds on the communication complexity.}
\section{Motivation}
\label{sec:motivation}
We first explain the idea using a simple example. Define the random source as
\begin{align*}
\RZ_1 := (\RX_a,\RX_b),\kern1em \RZ_2 := (\RX_a,\RX_b),\text{ and }
\RZ_3 := \RX_a,
\end{align*}
where $\RX_a$ and $\RX_b$ are independent uniformly random bits. The random bits $\RX_j$'s determine
the correlation, or joint distribution, of the sources $\RZ_i$'s. Let $V:=\Set{1,2,3}$ be the set of
users. Each user $i\in V$ observes the discrete memoryless source $\RZ_i$ privately. It is easy to
see that the users can agree on a secret key bit, namely, $\RX_a$, without any public discussion.
In fact, the users cannot agree on any more secret key bits, even with additional public discussion.
This is clear since $\RX_a$ is already the entire private observation of user $3$. \emph{The secrecy capacity is
therefore $1$~bit.}
For ISKA, we consider adding a common randomness to a subset of users. For example, we may add a new
independent bit $\RX_c$ to the private sources of users~$2$ and $3$ as follows:
\begin{align*}
\RZ'_1 := (\RX_a,\RX_b),\kern1em
\RZ'_2 := (\RX_a,\RX_b,\RX_c),\text{ and }
\RZ'_3 := (\RX_a,\RX_c),
\end{align*}
where user $i\in V$ observes the new source $\RZ_i'$ instead of $\RZ_i$.
With such an increment to the private sources, the bit $\RX_b$ can also be used as a secret key in
addition to $\RX_a$, i.e., the secret key can be chosen as $\RK=(\RX_a,\RX_b)$. To achieve this, user~$2$ can reveal
in public the XOR $\RF:=\RX_b\oplus \RX_c$, and so user $3$ can recover $\RX_b$ by subtracting
$\RX_c$ from the sum. It can be shown that $\RK$ is independent of $\RF$ and is therefore kept
secret from a wiretapper observing the entire public discussion. \emph{The secrecy capacity is now equal
to $2$~bits.}
In the above, the addition of the private common randomness $\RX_c$ increased the secrecy capacity by $1$ bit.
If we are allowed to choose who to give this common randomness to,
the current choice of users $2$ and
$3$ is in fact the most efficient (besides the equivalent choice of users~1 and 3).
For example, if $\RX_c$ were given to users~$1$ and $2$
instead, then it is evident that the capacity would not have increased.
Of course, one may choose to give $\RX_c$ to
user~$1$ in addition to users $2$ and $3$,
where the capacity would be 2 bits.
However,
such an allocation is not considered efficient since additional resources, e.g., private communication, may be needed to give $\RX_c$ to user~$1$.
For DSKA, we consider removing some common randomness from a subset of users, while trying not to
diminish the secrecy capacity. It is easy to see that removing $\RX_b$ from users~$1$ and $2$ does
not diminish the capacity, while removing $\RX_a$ from all the users does.\footnote{Note that we remove a common randomness from all the users observing it.
The source is simpler after such a removal since its joint entropy reduces strictly by the amount of
common randomness we remove. It is possible to extend this to more general multi-letter preprocessing of the random source, and doing so is potentially more useful for non-hypergraphical sources.}
In other words, the common randomness $\RX_b$ is redundant but $\RX_a$ is not. We can
therefore consider the following simpler source for the purpose of achieving the secrecy capacity of
$1$~bit:
\begin{align*}
\RZ''_i&:=\RX_a &&\text{for $i\in \Set{1,2,3}$.}
\end{align*}
Simplifying the source is useful because it simplifies the capacity-achieving scheme in \cite{csiszar04} by reducing the amount of discussion required for the \emph{communication for omniscience}.
\section{Problem formulation}
We will formulate ISKA and DSKA as extensions of the secret key agreement problem in
\cite{csiszar04} under the source model without wiretapper's side information nor helpers. Readers
may refer to \cite{csiszar04} for the detailed secret key agreement protocol. In our formulation, we
will only need the characterization~\eqref{eq:MMI}~\cite{chan10md} of the secrecy capacity. In this paper, we primarily
denote sets with capital letters, random variables with the sans-serif font, and families of sets with
script typeface letters. Furthermore, for a family $\mcF$ of sets, we will use $\op{minimal} \mcF$ and
$\op {maximal} \mcF$ to denote, respectively, the sets of inclusion-wise minimal and maximal
elements of $\mcF$.
\subsection{Incremental secret key agreement}
To formulate ISKA, we consider adding a common randomness $\RX$ of entropy $`e>0$ to a subset $S$ of users:
\begin{Definition}
For $S\subseteq V$ and $`e>0$, we say that $\RZ_V^{S,`e}$ is an $(S,`e)$-incremented source of $\RZ_V$ if it can be written as
\begin{align}
\RZ_i^{S,`e}:=\begin{cases}
(\RZ_i,\RX) & i\in S\\
\RZ_i & \text{otherwise,}
\end{cases}
\end{align}
where $\RX$ is independent of $\RZ_V$ and has entropy $H(\RX)=`e$.
\end{Definition}
We want to characterize the \emph{rate of increase} in the secrecy capacity~\eqref{eq:MMI} of the incremented source:
\begin{Definition}
The subderivative (one-sided limit)
\begin{align}
`r_{S}^+(\RZ_V) &:= \left.\tfrac{\partial I(\RZ_V^{S,`e})}{\partial `e}`2|_{`e=0^+} = \lim_{`e\downarrow 0} \frac{I(\RZ_V^{S,`e})-I(\RZ_V)}{`e}\label{eq:`r_S^+}
\end{align}
is the \emph{growth rate} of the secrecy capacity for the private source $\RZ_V$ with respect to an infinitesimal increment in the common randomness of the subset $S$. Since the additional common randomness $\RX$ is a valuable resource, we want to maximize the growth rate among all subsets $S$ of the same size as in
\begin{align}
`r_k^+(\RZ_V) &:= \max_{S\subseteq V:\abs{S}\leq k} `r_{S}^+(\RZ_V) \label{eq:`r_k^+}
\end{align}
for integer $k$. We refer to $`r_k^+$ as the \emph{growth rate of order} $k$.
\end{Definition}
For an efficient allocation of common randomness, we want to identify small subsets with strictly positive growth rate:
\begin{Definition}
\label{def:critical}
$S\subseteq V$ is said to be a \emph{critical (hyper)edge} if it is inclusion-wise \emph{minimal} with $`r_S^+(\RZ_V)>0$. We use
\begin{align}
\pzS_{\op{crit}}(\RZ_V) &:=\op{minimal}\Set*{S\subseteq V\mid `r_S^+(\RZ_V)>0}
\end{align}
to denote the set of all critical edges.\footnote{Unlike \cite{courtade16}, the word critical is associated with an edge rather than a set family for a non-asymptotic hypergraphical source. Furthermore, the notion of critical family in \cite{courtade16} is related to the notion of excess edge in DSKA rather than the notion of critical edge in the ISKA problem. We also consider an asymptotic source model that is not restricted to be hypergraphical.}
\end{Definition}
(We remark that a critical edge $S$ is an edge in the
$(S,\epsilon)$-incremented source.)
Out of all the critical edges, the subsets with minimum size require the least resource. It is easy to argue that the minimum critical edges are the optimal solutions to \eqref{eq:`r_k^+} for the smallest $k$ such that $`r_k^+(\RZ_V)>0$.
\subsection{Decremental secret key agreement}
To formulate the DSKA problem, we will consider a special kind of random sources:
\begin{Definition}
We say that a source $\RZ_V$ has an edge $S\subseteq V$ if there is a common randomness $\RX'$
that is observed only by the users in $S$,
i.e., if we can rewrite $\RZ_i$ (up to bijection) as
\begin{align}
\RZ_i = \begin{cases}
(\RZ'_i,\RX') & i\in S\\
\RZ'_i & \text{otherwise},
\end{cases}
\end{align}
where $\RZ'_V$ is independent of $\RX'$. We reduce such a source to the following $(S,`e)$-decremented source by removing an $`e\in (0, H(\RX')]$ amount of common randomness $\RX'$:
\begin{align}
\RZ_i^{S,-`e} &= \begin{cases}
(\RZ'_i,\RX) & i\in S\\
\RZ'_i
\end{cases}
\end{align}
for some common randomness $\RX$ independent of $\RZ_V'$ and with $H(\RX)=H(\RX')-`e$.
\end{Definition}
Contrary to ISKA, we are interested in the \emph{rate of decrease} in the secrecy capacity~\eqref{eq:MMI}:
\begin{Definition}
The subderivative (one-sided limit)
\begin{align}
\kern-1em `r_{S}^-(\RZ_V) &:= \left.-\tfrac{\partial I(\RZ_V^{S,-`e})}{\partial `e}`2|_{`e=0^+} \kern-1em= \lim_{`e\downarrow 0} \frac{I(\RZ_V)-I(\RZ_V ^{S,-`e})}{`e},\kern-.5em
\end{align}
is the \emph{loss rate} of the secrecy capacity for the private source $\RZ_V$ with edge $S$.
\end{Definition}
Unlike ISKA, we are interested in edges $S$ with zero loss rate.
\begin{Definition}
$S\subseteq V$ is said to be an \emph{excess} or \emph{redundant} edge if the corresponding loss
rate is $0$ for the source $\RZ_V$ with edge $S$. In this case, we can simplify the secret key
agreement by removing common randomness of the edge $S$ without diminishing the secrecy capacity.
\end{Definition}
\section{Main results}
As pointed out by \cite{chan11isit,milosavljevic11} and elaborated in \cite{chan15mi}, the MMI can be computed in polynomial time using
\emph{submodular function
minimization}~\cite{schrijver02} algorithms.
A
polynomial time algorithm was also given by an earlier work of
Fujishige~\cite{fujishige84,fujishige88} for a more general type of \emph{submodular functions}
and \emph{set family}.
It turns out that the characterization and computation of the growth rate, loss rate, critical
edges, and excess edges depend only on the optimal partitions that attain the MMI~\eqref{eq:MMI}. The set of the optimal partitions will be denoted by
\begin{align}
\Uppi^*(\RZ_V) :=\Set*{\mcP\in \Pi'(V)\mid I_{\mcP}(\RZ_V)=I(\RZ_V)}.
\end{align}
Using the combinatorial result of \cite{narayanan90}, the set $\Uppi^*(\RZ_V)\cup \Set{V}$
forms a lattice, and hence admits a unique finest partitionm
\begin{align}
\pzP^*(\RZ_V):= \min \Uppi^*(\RZ_V)\label{P^*}.
\end{align}
Here, the minimum is with respect to the partial order ``$\prec$'' of the partitions, which is defined
as $\mcP\prec \mcP'$ if the partition $\mcP$ \emph{is finer than} (or \emph{a refinement of}) the
partition $\mcP'$. In other words, $\mcP$ can be obtained by further partitioning one or more
subsets in $\mcP'$. Following \cite{chan15mi}, we will refer to the finest partition $\pzP^*(\RZ_V)$
as the \emph{fundamental partition}.
This partition has an elegant interpretation~\cite{chan15allerton} in
data clustering, and furthermore, can be computed in strongly polynomial time using algorithms
such as \cite{nagano10} applied to the \emph{minimum average cost clustering}.
For ISKA, we can characterize the growth rate and critical edges using the optimal partitions as follows:
\begin{Theorem}
\label{thm:rhopatch}
\begin{subequations}
For any $\RZ_V$ and $S\subseteq V$,
\begin{align}
`r^+_{S}(\RZ_V) &= \min_{\mcP \in \Uppi^*(\RZ_V)} \frac{\sum_{C\in
\mcP} `c_{\Set{C \cap S\neq `0}} - `c_{\Set{S\neq `0}}
}{\abs{\mcP}-1}, \label{eq:rhoS}
\end{align}
where $`c$ is the indicator function of the condition specified in the subscript.
It follows that
\begin{align}
\pzS_{\op{crit}}(\RZ_V) &= \op{minimal}\{S\subseteq V\mid \notag \\
& \kern4em S\not\subseteq C,\forall C\in
\mcP\in \Uppi^*(\RZ_V)\}. \label{eq:patch}
\end{align}
In other words, $S\subseteq V$ is critical iff it is a minimal set that overlaps at least two blocks of every optimal partition.
\end{subequations}
\end{Theorem}
\begin{Proof}
See Appendix~\ref{sec:A}.
\end{Proof}
Indeed, $\pzS_{\op{crit}}(\RZ_V)$ depends on $\RZ_V$ only through the coarsest optimal
partitions in $\Uppi^*(\RZ_V)$. This is because, if $S$ crosses a partition, i.e., overlaps at least
two blocks of the partition, then it must also cross any refinement of the partition.
For DSKA, we can similarly characterize the loss rate and excess edges as follows:
\begin{Theorem}
\label{thm:rho-}
\begin{subequations}
For any $\RZ_V$ with edge $S\subseteq V$,
\begin{align}
`r^-_{S}(\RZ_V) &= \max_{\mcP \in \Uppi^*(\RZ_V)} \frac{\sum_{C\in
\mcP} `c_{\Set{C \cap S\neq `0}} - `c_{\Set{S\neq `0}}
}{\abs{\mcP}-1}. \label{eq:rhoS-}
\end{align}
It follows that $S$ is an excess edge iff
\begin{align}
\exists C\in \pzP^*(\RZ_V), S\subseteq C,\label{eq:excess}
\end{align}
i.e., $S$ does not cross the fundamental partition.
\end{subequations}
\end{Theorem}
\begin{Proof}
See Appendix~\ref{sec:A}.
\end{Proof}
The condition for an excess edge depends only on the fundamental partition, and therefore can be
checked in strongly polynomial time.
Equations \eqref{eq:rhoS} and \eqref{eq:patch} imply the following simple properties of $`r^+_k$~\eqref{eq:`r_k^+} and $\pzS_{\op{crit}}$.
\begin{Proposition}
\label{pro:rhopatch}
$`r_k^+(\RZ_V)$ is non-decreasing in $k$ and equal to $0$ for
$k\leq 1$. Furthermore,
\begin{align*}
`r_k^+(\RZ_V) =1\kern1em \text{ iff }\kern1em k\geq \abs{\pzP^*(\RZ_V)},
\end{align*}
the size of the fundamental partition. Finally, we have at least one critical edge,
i.e. $\pzS_{\op{crit}}(\RZ_V)\neq `0$, and the minimum size of a critical edge is at least $2$,
i.e., $\min\Set{\abs {S}\mid S\in \pzS_{\op {crit}}(\RZ_V)}\geq 2$.
\end{Proposition}
\begin{Proof}
See Appendix~\ref{sec:A}.
\end{Proof}
Computing $`r_k^+$ in general can be quite difficult but some simple cases will be given in the next section. Surprisingly, it turns out that
computing a minimum critical edge can be done in strongly polynomial time, which
is due to the result below.
\begin{Theorem}
\label{thm:C}
All critical edges in $\pzS_{\op{crit}}(\RZ_V)$ have the same size, and are therefore minimum.
\end{Theorem}
In other words, all the critical edges are minimum. A critical edge can be obtained easily as follows: Starting with $S=V$, repeatedly remove an element from $S$
as long as doing so does not violate
$I(\RZ_V^{S,1})>I(\RZ_V)$. The condition can be checked
in strongly polynomial time for at most $O(\abs{V}^2)$ times.
Indeed, a stronger result can be proved. We will characterize the
critical edges completely using only
the maximal blocks from the optimal partitions,
\begin{align}
\label{eq:T}
\pzT_{\max}(\RZ_V) &:= \op {maximal} \bigcup \Uppi^*(\RZ_V)
\end{align}
which can also be computed in strongly polynomial time. More
precisely, we can show that:
\begin{Lemma}
\label{lem:T}
Either one of the following cases happen:
\begin{align}
\pzT_{\max}(\RZ_V) &\in \Uppi^*(\RZ_V) \label{eq:T1}\\
\bar{\pzT}_{\max}(\RZ_V) &:= \Set{V`/C:C\in
\pzT_{\max}(\RZ_V)}\in \Pi'(U)\label{eq:T2}
\end{align}
for some $U\subseteq V$. In words, either $\pzT_{\max}(\RZ_V)$ is an
optimal partition or its complement $\bar{\pzT}_{\max}(\RZ_V)$ is a
set of at least two non-empty disjoint subsets of $V$. Indeed, \eqref{eq:T1} means that $\pzT_{\max}(\RZ_V)$ is the unique coarsest optimal partition.
\end{Lemma}
\begin{Theorem}
\label{thm:TC}
If \eqref{eq:T1} happens,
\begin{align}
\pzS_{\op{crit}}(\RZ_V) &= \Set{\Set{i,j}\mid i\in C,j\in V`/C,C\in
\pzT_{\max}(\RZ_V)}\label{eq:C1}
\end{align}
and so all the critical edges have size $2$.
If \eqref{eq:T2} happens,
\begin{align}
\pzS_{\op{crit}}(\RZ_V) &= \Set{\Set{i_C\mid C\in \pzT_{\max}(\RZ_V)}\mid i_C\in V`/C},\label{eq:C2}
\end{align}
which is taken to mean the collection of the sets $\Set{i_C\mid C\in \pzT_{\max}(\RZ_V)}$ of representatives $i_C$ of subsets $V`/C$ for $C\in \pzT_{\max}(\RZ_V)$. It follows that $\abs{\pzT_{\max}(\RZ_V)}$ is the size of all the critical edges.
\end{Theorem}
Note that, Theorem~\ref{thm:TC} implies Theorem~\ref{thm:C} immediately.
\begin{Proof}
See Appendix~\ref{sec:B}.
\end{Proof}
\begin{Example}
\label{eg:patch}
Consider $V=\Set{1,2,3}$. Let $\RZ_1=\RZ_2$ be a uniformly random bit, and $\RZ_3=0$. It can be shown that
\begin{align*}
\Uppi^*(\RZ_V) &= \Set{\Set{\Set{1,2},\Set{3}}}\\
\pzS_{\op {crit}}(\RZ_V) &= \Set{\Set{1,3},\Set{2,3}}.
\end{align*}
There is a unique optimal partition and
so $\pzT_{\max}(\RZ_V)=\pzP^*(\RZ_V)=\Set{\Set{1,2},\Set{3}}\in
\Uppi^*(\RZ_V)$, satisfying \eqref{eq:T1}. The set of
critical edges by \eqref{eq:C1} is
$\pzS_{\op{crit}}(\RZ_V)=\Set{\Set{1,3},\Set{2,3}}$
Indeed, \eqref{eq:T1} may hold even when the optimal partition is
not unique. For instance, consider $V=\Set{1,2,3}$ and let
$\RZ_1\!:=\!(\RX_a,\RX_b,\RX_c)$, $\RZ_2\!:=\!(\RX_a,\RX_b,\RX_d)$ and $\RZ_3\!:=\!(\RX_c,\RX_d)$
where $\RX_i$'s are uniformly random and
independent bits. It follows that
\begin{align*}
\Uppi^*(\RZ_V) &=
\Set{\Set{\Set{1,2},\Set{3}},\Set{\Set{1},\Set{2},\Set{3}}}\\
\pzT_{\max}(\RZ_V) &= \Set{\Set{1,2},\Set{3}}\in \Uppi^*(\RZ_V)\\
\pzS_{\op{crit}}(\RZ_V)&= \Set{\Set{1,3},\Set{2,3}}
\end{align*}
which satisfies \eqref{eq:T1} but $\pzT_{\max}$ is the coarsest partition
rather than the
fundamental partition.
\end{Example}
\begin{Example}
\label{eg:tree}
\begin{figure*}
\centering
\tikzstyle{point}=[draw,circle,minimum size=.2em,inner sep=0,
outer sep=.2em]
\subcaptionbox{Pairwise independent network $\RZ_V$.\label{fig:egrk:source}}{
\begin{tikzpicture}[x=.6em,y=.6em,>=latex]
\foreach \x/\angle/\lb in {1/45/{$\RZ_\x=\RX_a$},2/135/{$\RZ_\x=(\RX_a,\RX_b)$},3/-135/{$\RZ_\x=(\RX_b,\RX_c)$},4/-45/{$\RZ_\x=\RX_c$}}
{
\path (\angle:5) node (\x) [point,label=\angle:$\x$,label={[label distance=1.5em]\angle:{\color{gray}\scriptsize\lb}}] {};
}
\foreach \x/\y/\lp/\lb/\lpp/\lbb in {3/4/above/$\RX_c$/below/~
1/2/below/$\RX_a$/above/~
2/3/right/$\RX_b$/left/~
\draw[-] (\x) to node [label=\lp:{\color{gray}\scriptsize\lb},label=\lpp:{\color{blue}\scriptsize\lbb}]{} (\y);
\end{tikzpicture}
}
\hfil
\subcaptionbox{Optimal partitions $\Uppi^*(\RZ_V)$.\label{fig:egrk:optpart}}{
\includegraphics{DTL.pdf}
}
\hfil
\subcaptionbox{Rate of increase $`r_k^+(\RZ_V)$.\label{fig:egrk:`r_k}}{
{\def2.2{2.2}
\def5{5}
\begin{tikzpicture}[x=1em,y=1em,>=latex]
\draw[->] (0,0) -- (0,1.3*5) node [label=above:$`r_k(\RZ_V)$] {};
\draw[->] (0,0) -- (5*2.2,0) node (k) [label=right:$k$] {};
\path (1*2.2,0*5) node (1) [point,thick,label=above:{\scriptsize$0$}] {};
\path (2*2.2,0.33*5) node (2) [point,thick,label=above:{\scriptsize$\frac13$}] {};
\path (3*2.2,0.5*5) node (3) [point,thick,label=above:{\scriptsize$\frac12$}] {};
\path (4*2.2,1*5) node (4) [point,thick,label=above:{\scriptsize$1$}] {};
\draw[dashed,blue,thick]
(1)--(2)--(3)--(4)--(4.5*2.2,1*5);
\foreach \x in {1,2,3,4}
\draw[dashed] (\x) -- (\x|-k) node [below] {\scriptsize $\x$};
\end{tikzpicture}
}
}
\caption{Optimal partitions and rate of increase of the tree
network in Example~\ref{eg:tree}.}\vspace*{-1.2em}
\label{fig:egrk}
\end{figure*}
Let $V=\Set{1,2,3,4}$, and
\begin{align*}
\RZ_1=\RX_a, \ \RZ_2=(\RX_a,\RX_b), \ \RZ_3=(\RX_b,\RX_c), \text{ and } \RZ_4=\RX_c
\end{align*}
where $\RX_i$'s are independent uniformly random
bits. The optimal partitions in $\Uppi^*(\RZ_V)$ are shown in
\figref{fig:egrk:optpart}, and $\pzT_{\max}(\RZ_V)=\Set{\Set{1,2,3},\Set{2,3,4}}$.
Although $\pzT_{\max}(\RZ_V)$ is not an optimal partition,
$\bar{\pzT}_{\max}(\RZ_V)=\Set{\Set{4},\Set{1}}$, which
satisfies~\eqref{eq:T2}. By \eqref{eq:C2}, the set of minimum
critical edges is $\pzS_{\op{crit}}(\RZ_V)=\Set{\Set{1,4}}$. It crosses at least two blocks of every coarsest
optimal partition, and therefore every optimal partition.
Example~\ref{eg:tree} is a special case of the pairwise independent network (PIN) source model
\cite{nitinawarat-ye10}, where in this example the source is simply a tree, see \figref{fig:egrk:source}. For tree PINs in general, it can be argued that a partition $\mcP\in \Pi'(V)$ is optimal iff each block $C\in \mcP$ induces a tree on $C$. For example,
the optimal partition $\Set{\Set{1,2},\Set{3,4}}$ in
\figref{fig:egrk:optpart} induces two
subtrees, one connecting $1$ and $2$, and the other connecting $3$
and $4$. Thus, the
fundamental partition is always the partition into singletons, but it
is not the only optimal partition for $\abs{V}\geq 3$. Nevertheless, we can show that there is only
one critical edge and such an edge is equal to the set of leaves.
\end{Example}
\section{Computing the growth rate of different orders}
In this section, we will illustrates the computation of $`r_k^+$ in \eqref{eq:`r_k^+}.
There are simple cases where $`r_k^+$ (and therefore $\pzS_{\op {crit}}$) can be computed
easily. For example, in the special case when the fundamental partition is the unique
optimal partition,\mnote{a:un} i.e.\ $\abs{\Uppi^*(\RZ_V)}=1$, it can be argued easily from
\eqref{eq:rhoS} that
\begin{align*}
`r_k^+(\RZ_V)=\frac{k-1}{\abs{\pzP^*(\RZ_V)}-1} \kern1em \text{for
$k\leq |\pzP^*(\RZ_V)|$},
\end{align*}
where an optimal solution $S$ to \eqref{eq:`r_k^+} is any set of $k$
elements, each from a different block of the fundamental
partition. In particular,
\begin{align*}
\pzS_{\op {crit}}(\RZ_V) = \Set{\Set{i,j}\mid i\in C, j\in V`/C, C\in \pzP^*(\RZ_V)}
\end{align*}
and so all the critical edges are minimum with size $2$. When the
partition into singletons is the unique optimal partition, then the
optimal solution $S$ to \eqref{eq:`r_k^+} is simply any $k$-subset of $V$.
In particular, the singleton partition is shown to be the unique optimal partition in \cite{mukherjee14} for any PIN model
that corresponds to a complete graph. We can also show the same result for any PIN model that corresponds to a cycle. In general, it is possible to
check in strongly
polynomial time whether the fundamental partition is the unique optimal partition. (See Appendix~\ref{sec:uniqueness}.) For instance, in Example~\ref{eg:patch}, the optimal partition is unique and all the critical edges have size $2$ as expected.
More details about the computation and interpretations of the fundamental partition can be found in \cite{chan15mi}.
When the optimal partition is not
unique, $`r_k^+(\RZ_V)$ may not be
linear in $k$. The marginal increase in growth rate may not be diminishing in $k$ either. This is
the case, for instance, for Example~\ref{eg:tree} as shown in \figref{fig:egrk:`r_k}.
If $H(\RZ_B)$ is an integer for every $B\subseteq V$, then
$`r_S^+(\RZ_V)$ can be computed in strongly polynomial time. To argue
this, choose $`e=\frac1{\abs{V}!}$. Note that $I_{\mcP}(\RZ_V)/`e$ is
an integer and so $I_{\mcP}(\RZ_V)$ for different $\mcP\in \Pi'(V)`/\Uppi^*(\RZ_V)$
is larger than $I(\RZ_V)$ by at least $`e$, while
$I(\RZ_V^{S,`e})$ is larger than $I(\RZ_V)$ by at most
$`e$. Thus, $\Uppi^*(\RZ_V^{S,`e})\subseteq \Uppi^*(\RZ_V)$ and so
$`r_S^+(\RZ_V)= \frac{I(\RZ_V^{S,`e})-I(\RZ_V)}{`e}$.
Each term on the R.H.S.\ can be computed in strongly polynomial
time. A more refined argument suggests that one can choose any $`e\leq
\frac1{(\abs{V}-1)(\abs{V}-2)}$.
\section{Conclusion}
In this work, we have formulated the ISKA and DSKA problems to study how the MMI of a set of random variables changes with respect to an infinitesimal increment and decrement in the MMI of a subset of the random variables.
We found that the set of opimal partitions that attain the MMI of a set of random variables completely characterizes the growth rate for ISKA and the loss rate for DSKA.
For ISKA, the growth rate can be computed easily in some special cases, e.g., when the optimal partition is unique. In general, however, it is not clear whether the computation can be done in polynomial time. The growth rate can be non-linear in the order, and the marginal return may even increase as we increase the order. Very surprisingly, however, a minimum critical edge can be computed in strongly polynomial time because all critical edges have the same size. In other words, one can easily identify a minimum subset of users to give an additional common randomness to, such that the secrecy capacity strictly increases.
For DSKA, the condition for an edge to be redundant can be characterized in strongly polynomial time using the fundamental partition. Identifying excess edges is useful in simplifying secret key agreement schemes. In particular, it is hopeful that further investigation can resolve the conjectures regarding the communication complexity for secret key agreement~\cite{MKS16,mukherjee16,chan16itw}.
\appendices
\makeatletter
\@addtoreset{equation}{section}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\@addtoreset{Theorem}{section}
\renewcommand{\theTheorem}{\thesection.\arabic{Theorem}}
\@addtoreset{Lemma}{section}
\renewcommand{\theLemma}{\thesection.\arabic{Lemma}}
\@addtoreset{Corollary}{section}
\renewcommand{\theCorollary}{\thesection.\arabic{Corollary}}
\@addtoreset{Example}{section}
\renewcommand{\theExample}{\thesection.\arabic{Example}}
\@addtoreset{Remark}{section}
\renewcommand{\theRemark}{\thesection.\arabic{Remark}}
\@addtoreset{Proposition}{section}
\renewcommand{\theProposition}{\thesection.\arabic{Proposition}}
\@addtoreset{Definition}{section}
\renewcommand{\theDefinition}{\thesection.\arabic{Definition}}
\@addtoreset{Subclaim}{Theorem}
\renewcommand{\theSubclaim}{\theLemma.\arabic{Subclaim}}
\makeatother
\section{Proofs of basic properties}
\label{sec:A}
\begin{Proof}[Theorem~\ref{thm:rhopatch}]
Clearly, $`r_{`0}^+(\RZ_V)=0$. Consider $S\subseteq V:S\neq `0$. Rewriting the divergence in terms of the entropy as $D(P_{\RZ_V}\|\prod_{C\in \mcP}P_{\RZ_C})=\sum_{C\in \mcP} H(\RZ_C)-H(\RZ_V)$, we have
\begin{align*}
I(\RZ^{`e,S}_V) &= \min_{\mcP\in \Pi'(V)}
\frac1{\abs{\mcP}-1} `1[\sum_{C\in \mcP} H(\RZ^{`e,S}_C) -
H(\RZ^{`e,S}_V) `2]\\
&= \min_{\mcP\in \Pi'(V)}
\frac1{\abs{\mcP}-1} \Bigg[\sum_{C\in \mcP}
H(\RZ_C)-H(\RZ_V)\\
&\kern1em +\sum_{C\in \mcP} H(\RX)`c_{\Set{S\cap C\neq `0} }
- H(\RX) \Bigg]\\
&= \min_{\mcP\in \Pi'(V)} `1[I_{\mcP}(\RZ_V) + `e \frac{\sum_{C\in \mcP} `c_{\Set{C \cap S\neq `0}} - 1}{\abs{\mcP}-1}`2].
\end{align*}
Suppose $I(\RZ^{`e,S}_V)< \min_{\mcP\in \Pi'(V)`/\Uppi^*(\RZ_V)}
I_{\mcP}(\RZ_V)$, which is possible for all $`e>0$ sufficiently
small since
$\abs{\Pi'(V)}$ is finite.
It does not lose optimality to restrict $\mcP$ to $\Uppi^*(\RZ_V)$ and so
we have for all $`e>0$ sufficiently small that
\begin{align*}
I(\RZ^{`e,S}_V) &= I(\RZ_V) + `e \min_{\mcP\in
\Uppi^*(\RZ_V)} \frac{\sum_{C\in \mcP^*} `c_{\Set{C \cap S\neq `0}} - 1}{\abs{\mcP}-1}
\end{align*}
which gives \eqref{eq:rhoS}. \eqref{eq:patch} follows from
\eqref{eq:rhoS} directly.
\end{Proof}
\begin{Proof}[Theorem~\ref{thm:rho-}]
The proof of \eqref{eq:rhoS-} is analogous to the proof of \eqref{eq:rhoS} above, but we have $\max$ instead of $\min$ due to a sign change because the rate is on the loss rather than growth of the MMI. To derive the condition~\eqref{eq:excess} for excess edge, notice that \eqref{eq:rhoS-} is zero iff $S$ does not cross any optimal partitions. It suffices to consider only the finest optimal partition, namely $\pzP^*(\RZ_V)$, because $S$ does not cross $\pzP^*(\RZ_V)$ implies it does not cross any coarser optimal partitions, which cover all the optimal partitions.
\end{Proof}
\begin{Proof}[Proposition~\ref{pro:rhopatch}]
$`r_k^+(\RZ_V)$ is non-decreasing in $k$ because $`r_S^+(\RZ_V)$ by \eqref{eq:rhoS} is
non-decreasing in $S$ with respect to set inclusion. It is equal to
$0$ for $k=1$ because $`r_{\Set{i}}^+(\RZ_V)=0$ for all $i\in V$. This also means that a critical edge, if any, must be non-singleton, with size at least two. $`r_k^+(\RZ_V)$ is at most
$1$ because $`r_V^+(\RZ_V)=1$. More precisely, $`r_S^+(\RZ_V)=1$ iff $S\cap C\neq `0$ for every $C\in \mcP^*$ and $\mcP^*\in
\Uppi^*(\RZ_V)$. Since the fundamental partition $\pzP^*(\RZ_V)$ is
the unique finest optimal partition, we have $`r_S^+(\RZ_V)=1$ iff $S\cap C\neq `0$ for every $C\in \pzP^*(\RZ_V)$. That means
$`r_S^+(\RZ_V)<1$ if $\abs{S}<abs {\pzP^*(\RZ_V)}$, and $`r_S^+(\RZ_V)=1$ for
any $S$ obtained by taking at least one element from each block in the
fundamental partition. The fact that $`r_V^+(\RZ_V)=1$ also means that there is at least one critical edge.
\end{Proof}
\section{Proof of Theorem~\ref{thm:TC}}
\label{sec:B}
\begin{Proof}[Theorem~\ref{thm:TC}]
From \eqref{eq:patch}, we have
$S\in \pzS_{\op {crit}}(\RZ_V)$ iff
\begin{align}
S\nsubseteq C\kern1em\text{or equiv.}\kern1em S`/C\neq `0 \kern1em
\forall C\in \pzT_{\max}(\RZ_V).\label{eq:TC}
\end{align}
From this, it can
be argued easily that the sets defined in \eqref{eq:C1} and \eqref{eq:C2}
are critical edges for the cases \eqref{eq:T1} and \eqref{eq:T2}
respectively. If $S$ is a critical edge under \eqref{eq:T1}, any
element, say $i\in S$, must be contained by some $C\in
\pzT_{\max}(\RZ_V)$ since $\pzT_{\max}(\RZ_V)$ is a partition of $V$. By
\eqref{eq:TC}, it must contain an element $j\in V`/C$ and so $S$
must be in \eqref{eq:C1} as desired. If $S$ is a critical edge under
\eqref{eq:T2}, it must contain an element from $V`/C$ for every
$C\in \pzT_{\max}(\RZ_V)$ by \eqref{eq:TC}. Thus, it must be in
\eqref{eq:C2} as desired.
\end{Proof}
It remains to prove Lemma~\ref{lem:T}. We do so using the idea of
\emph{zero-singleton-submodular} function in \cite{narayanan90}. Denote
the fundamental partition as
\begin{align}
\pzP^*(\RZ_V) = (C^*_1,\dots,C^*_{\ell})
\end{align}
by indexing the blocks from $1$ to $\ell=\abs{\pzP^*(\RZ_V)}$. Define
$g:2^{[\ell]}\mapsto `R$ as
\begin{align}
g(B) &:= h_{`g}`1(\bigcup\nolimits_{i\in B} C^*_i`2) - \sum_{i\in B} h_{`g}(C^*_i)
\kern1em\text{for $B\subseteq [\ell]$}\kern-.2em\label{eq:zss:g}
\end{align}
with $`g:=I(RZ_V)$ and $h_{`g}(C):=H(\RZ_C)-`g$ is the residual randomness defined in \cite{chan15mi}.
It follows immediately that $g$ is submodular with
\begin{align}
g(\Set{i})=0\kern1em \text{for $i\in [\ell]$,}\label{eq:zs}
\end{align}
and is therefore called a
zero-singleton-submodular function. It can also be shown to be
non-negative over non-empty sets.
\begin{Proposition}[\mbox{\cite[p.198]{narayanan90}}]
\label{pro:zss}
$g(B)\geq 0$ for all $B\subseteq [\ell]:B\neq `0$. The
\emph{zero sets} of $g$,
\begin{align}
\pzZ(g):=\Set{B\subseteq [\ell]:g(B)=0}\label{eq:zsets}
\end{align}
forms an \emph{intersecting family}, i.e.\
\begin{align}
U\cap W,U\cup W\in \pzZ(g)\text{ for all }U,W\in \pzZ(g):U\cap W\neq `0
\end{align}
Furthermore,
\begin{align}
\Set*{\bigcup\nolimits_{i\in B}C^*_i:B\in
\pzZ(g)}=\bigcup\Uppi^*(\RZ_V) \cup\Set{V},\label{eq:zsets->P}
\end{align}
and so $\bigcup\Uppi^*(\RZ_V) \cup\Set{V}$ is also an intersecting family.
\end{Proposition}
\begin{Proof}
For any partition $\mcP\in
\Pi([\ell])$,
\begin{align*}
g[\mcP]&:=\sum_{C\in \mcP} g(C)\\
&= \sum_{C\in \mcP} h_{`g}`1(\bigcup\nolimits_{i\in C} C^*_i`2) - \sum_{i=1}^\ell
h_{`g}(C^*_i)\\
&= h_{`g}`1[\Set*{\bigcup\nolimits_{i\in C} C^*_i:C\in \mcP}`2] -
h_{`g}[\pzP^*(\RZ_V)]\geq 0,
\end{align*}
because, by \cite[Theorem~5.1]{chan15mi}, $\pzP^*(\RZ_V)$ minimizes
$h_{`g}$ over all partitions of $V$, which include $\Set{\bigcup_{i\in C} C^*_i:C\in
\mcP}$. Furthermore, we have equality $g[\mcP]=0$ for $\mcP\in \Pi'([\ell])$ iff
\begin{align*}
\Set*{\bigcup\nolimits_{i\in C} C^*_i:C\in \mcP} \in \Uppi^*(\RZ_V).
\end{align*}
Suppose to
the contrary that $g(B)<0$ for some non-empty $B\subseteq [\ell]$. Then,
$g[\Set{B}\cup\Set{\Set{i},i\in [\ell]`/B}]<0$ by the zero-singleton
property~\eqref{eq:zs}, but this contradicts $g[\mcP]\geq 0$ above
for all $\mcP\in \Pi([\ell])$. This proves the non-negativity of $g$
over non-empty sets.
For $U,W\in \pzZ(g):U\cap W\neq `0$, we have
\begin{align*}
0\geq g(U)-g(U\cap W) \geq g(U\cup W) - g(W) \geq 0
\end{align*}
where the second inequality is by the submodularity of $g$, and the
first and last inequalities are because $g(U)=g(W)=0$ and $g(U\cap
W),g(U\cup W)\geq 0$ by the non-negativity of $g$ over non-empty
sets argued above. Thus, all inequalities are satisfied with
equality and so $g(U\cap W)=g(U\cup W)=0$ as desired for $\pzZ(g)$
to be an intersecting family.
It remains to prove \eqref{eq:zsets->P}. By \cite[Theorem~5.2]{chan15mi}, every partition in $\Uppi^*(\RZ_V)\cup \Set{\Set{V}}$
is coarser than $\pzP^*(\RZ_V)$ and can therefore be expressed as $\Set{\bigcup_{i\in C} C^*_i:C\in
\mcP}\in \Pi'(V)$ for some $\mcP\in \Pi'([\ell])$. By optimality,
$g[\mcP]=0$, and so $g(C)=0$ for all $C\in \mcP$ since $C$ is
non-empty and $g$ is non-negative over non-empty sets as argued
before. Conversely, suppose $g(B)=0$. Then $g[\Set{B}\cup\Set{\Set{i},i\in
[\ell]`/B}]=0$ by \eqref{eq:zs}, and so $\bigcup_{i\in B}C^*_i$ is
a block of an optimal partition, namely $\Set{\bigcup_{i\in
B}C^*_i}\cup\Set{C^*_i:i\in [\ell]`/B}$. This completes the proof
of \eqref{eq:zsets->P}. Since $C^*_i$'s are disjoint, the fact that
$\pzZ(g)$ is an intersecting family implies that $\bigcup\Uppi^*(\RZ_V)\cup
\Set{V}$ is.
\end{Proof}
\begin{Proof}[Lemma~\ref{lem:T}]
We first argue that: for any distinct $C_1,C_2\in
\pzT_{\max}(\RZ_V)$, we have
\begin{align}
C_1\cap C_2\neq `0 \text{ implies }C_1\cup C_2= V.\label{eq:lt1}
\end{align}
To show this, note that $C_1$ and $C_2$ are maximal sets from
$\bigcup \Uppi^*(\RZ_V)$ by definition~\eqref{eq:T}. By
Proposition~\ref{pro:zss}, $\bigcup\Uppi^*(\RZ_V)\cup \Set{V}$ is an
intersecting family and so $C_1\cap C_2\neq `0$ implies $C_1\cup
C_2$ is also in the family. Suppose to the contrary that $C_1\cup
C_2\neq V$, then $C_1\cup C_2$ is a strictly larger set in
$\bigcup\Uppi^*(\RZ_V)$ than the distinct sets $C_1$ and $C_2$, which
contradicts the maximality.
Next, we argue that if there exists distinct $C_1,C_2\in
\pzT_{\max}(\RZ_V)$ such that $C_1\cap C_2\neq `0$, then
\begin{align}
C\cap C'\neq `0\text{ for all }C,C'\in \pzT_{\max}(\RZ_V).\label{eq:lt2}
\end{align}
Indeed, we have a stronger statement that
\begin{align}
`0\neq (V`/C_1)\subseteq C\text{ for all }C\in
\pzT_{\max}(\RZ_V)`/\Set{C_1}.
\label{eq:lt3}
\end{align}
If $C=C_2$, we have \eqref{eq:lt3} directly from \eqref{eq:lt1}. Suppose $C\neq C_2$. Then,
$C_1\cap
C\neq `0$ because $C\cap V`/C_2\neq `0$ by the maximality of $C\neq
C_2$, and
$V`/C_2 \subseteq C_1$ by \eqref{eq:lt1} that $C_1\cup
C_2=V$. Applying \eqref{eq:lt1} again with $C_2$ replaced by $C$, we
have $C_1\cup C=V$ as desired.
By the contrapositive statement of \eqref{eq:lt3}, if $C_1\cap
C_2=`0$ for some distinct $C_1,C_2\in \pzT_{\max}(\RZ_V)$, then all
the sets in $\pzT_{\max}(\RZ_V)$ must be disjoint, and so we have
\eqref{eq:T1} since every element in $V$ must be covered by at least
one block of an optimal partition. In the other case when every
distinct $C_1,C_2\in \pzT_{\max}(\RZ_V)$ must intersect,
\eqref{eq:lt1} implies
$C_1\cup C_2=V$, or equivalently, $(V`/C_1)\cap (V`/C_2)=`0$, which
implies \eqref{eq:T2} as desired.
\end{Proof}
By \eqref{eq:zsets->P}, $\pzT_{\max}(\RZ_V)$ can be obtained from the maximal zero sets in
$\pzZ(g)$ together with the fundamental
partitions $\pzP^*(\RZ_V)$. The maximal zero set not containing an element
$i\in [\ell]$, can be computed
in strongly polynomial time as follows: Starting with $C=`0$, add an
element from $[\ell]`/\Set{i}$ to $C$
repeatedly as long as $\min_{C'\in 2^{[\ell]`/\Set{i}}: C\subseteq
C'} g(C')=0$. The minimization is repeated at most $O(\abs{V}^2)$
times and can be solved in strongly
polynomial time using some existing algorithm for submodular function minimization over a
lattice family~\cite{schrijver02}. Hence, $\pzT_{\max}(\RZ_V)$ and
therefore $\pzS_{\op{crit}}(\RZ_V)$ can be computed in strongly polynomial
time by Theorem~\ref{thm:TC}.
\pbox[Conjecture]{Does critical edges have the same rate of increase? We
conjecture that $`r_S^+=\frac{\abs{S}-1}{\abs{\pzP^*(\RZ_V)}-1}$ for all $S\in
\pzS_{\op {crit}}(\RZ_V)$.}
\section{Uniqueness of the optimal partition}
\label{sec:uniqueness}
To check whether there is a unique optimal partition in strongly polynomial time, we can make use of
the zero-singleton-submodular function $g$ defined in
\eqref{eq:zss:g} based on the fundamental partition
$\pzP^*(\RZ_V)$ and the MMI. In essense of \eqref{eq:zsets->P}, we only need
to check whether the zero sets $\pzZ(g)$ in \eqref{eq:zsets}
consists only of the singletons. This can computed in $O(\abs{V}^2)$
submodular function minimizations, namely $\min_{B\supseteq \Set{i,j}}
g(B)$ for all pair $(i,j)$ of distinct elements. $\pzP^*(\RZ_V)$
and $I(\RZ_V)$ can also be computed in $O(\abs{V}^2)$
submodular function minimizations~\cite{nagano10}.
\section*{Acknowledgments}
The authors would like to thank Prof.\ Navin Kashyap, Manuj Mukherjee, and the colleagues at INC for detailed discussion and valuable comments.
\bibliographystyle{IEEEtran}
|
\section{Derivation of (1.5)}
In order to derive \eqref{1.5}, let us denote its right-hand side by $S$ and expressing $_2F_1$ as a series, we have
$$S=\exp \left( x+\frac{y}{x} \right)\, \sum_{k=0}^\infty\,\sum_{m=0}^k\,\frac{(-k)_m\,(-k-c+b+1)_m\,(-1)^k}{(c)_m\, m!\,k!}\,x^{2m-k}\,y^{k-m}$$
Using
$$\frac{(-k)_m}{k!}=\frac{(-1)^m}{(k-m)!}$$
we have
$$S=\exp \left( x+\frac{y}{x} \right)\,\sum_{k=0}^\infty\,\sum_{m=0}^k\,\frac{(-k-c+b+1)_m}{(c)_m\,m!}\,\frac{(-1)^{k+m}\, x^{2m-k}\,y^{k-m}}{(k-m)!}$$
Replacing $k$ by $k+m$, and using the result[5, Lemma 10, equ. 2, p-57]
$$S= \exp\left( x+\frac{y}{x} \right)\,\sum_{k=0}^\infty\,\sum_{m=0}^\infty\,\frac{(-k-m-c+b+1)_m}{\,(c)_m\,m!}\,\frac{(-1)^{k}\,x^{m-k}\,y^k}{k!}$$
Using $(-k-m-c+b+1)_m=(-1)^m\;(k+c-b)_m$, we have
$$S=\exp\left( x+\frac{y}{x} \right)\,\sum_{k=0}^\infty\,\sum_{m=0}^\infty\,\frac{(-1)^m\;(k+c-b)_m}{\,(c)_m\;m!}\,\frac{(-1)^{k}\,x^{m-k}\;y^k}{k!}$$
which can be written as
$$S=\exp \left( x+\frac{y}{x} \right)\,\sum_{k=0}^\infty\,\frac{(-1)^{k}}{k!}\, \left(\frac{y}{x} \right)^k\,
\sum_{m=0}^\infty\,\frac{(-1)^m\;(k+c-b)_m}{\,(c)_m\;m!}\,x^m $$
Now, summing up the inner series, we have
\begin{align*}
&S= \exp \left( x+\frac{y}{x} \right) \,\sum_{k=0}^\infty\,\frac{(-1)^{k}}{k!}\,\left(\frac{y}{x} \right)^k
{}_1F_1\left[ \begin{array}{c} k+c-b \\ c \end{array} ;\,-x \right] \, .\\
&\,\,\,=\exp \left(\frac{y}{x} \right)\,\sum_{k=0}^\infty\,\frac{\left(-\frac{y}{x} \right)^k}{k!}\,\,\left\{\exp (x)\,
{}_1F_1\left[ \begin{array}{c} k+c-b \\ c \end{array} ;\,-x \right] \right\} \, .
\end{align*}
Using Kummer's first transformation \cite{Manako4}
\begin{equation}
\label{2.6}
\exp (x)\,\,{}_1F_1\left[\begin{array}{c} c-b \\ c \end{array} ;\, -x \right] \, =
{}_1F_1\left[\begin{array}{c} b \\ c \end{array} ;\,x \right] \, .
\end{equation}
we have
$$S=\exp\left(\frac{y}{x}\right)\,\sum_{k=0}^\infty\,\frac{\left(-\frac{y}{x}\right)^k}{k!}
{}_1F_1\left[ \begin{array}{c} b-k \\ c \end{array} ;\,x \right] \, .
$$
Again, expressing $_1F_1$ as a series, we have
$$S=\exp\left(\frac{y}{x}\right)\,\sum_{k=0}^\infty\,\sum_{n=0}^\infty\,\frac{\left(-\frac{y}{x}\right)^k}{k!}\;\frac{(b-k)_n}{(c)_n} \frac{x^n}{n!}$$
Since $(b-k)_k=\frac{(b)_n\,(1-b)_n}{(1-b-n)_k}$, we have, therefore
$$S=\exp\left(\frac{y}{x}\right)\,\sum_{n=0}^\infty\,\frac{(b)_n}{(c)_n}\;\frac{x^n}{n!}\,\sum_{k=0}^\infty\,\frac{(1-b)_k}{(1-b-n)_k}\;\frac{\left(-\frac{y}{x}\right)^k}{k!}$$
Summing up the inner series
\begin{align*}
& S=\exp\left(\frac{y}{x}\right)\;\sum_{n=0}^\infty\,\frac{(b)_n}{(c)_n}\;\frac{x^n}{n!}\,
{}_1F_1\left[ \begin{array}{c} 1-b \\ 1-b-n \end{array} ;\, -\frac{y}{x} \right] \\
&=\sum_{n=0}^\infty\,\frac{(b)_n}{(c)_n}\,\frac{x^n}{n!}\,\left\{\exp\left(\frac{y}{x}\right)\,
{}_1F_1\left[ \begin{array}{c} 1-b \\ 1-b-n \end{array} ;\,-\frac{y}{x} \right] \right\} \,.
\end{align*}
Using \eqref{2.6}, we have
$$S=\sum_{n=0}^\infty \,\frac{(b)_n}{(c)_n}\;\frac{x^n}{n!}\,
{}_1F_1\left[ \begin{array}{c} -n \\ 1-b-n \end{array} ;\,\frac{y}{x} \right] \,.
$$
Expressing $_1F_1$ as a series
$$ S=\sum_{n=0}^\infty\,\sum_{m=0}^n\,\frac{(b)_n}{(c)_n}\;\frac{(-n)_m}{(1-b-n)_m}\;\frac{x^n\,\left(\frac{y}{x}\right)^m}{m!\;n!}$$
Using $(-n)_m=\frac{(-1)^m\;n!}{(n-m)!}$
$$ S=\sum_{n=0}^\infty\,\sum_{m=0}^n\,\frac{(b)_n\;(-1)^m}{(c)_n\;(1-b-n)_m}\,\frac{x^n\;\left(\frac{y}{x}\right)^m}{m!\;(n-m)!}$$
Changing $n$ to $n+m$ and using [5, Lemma 10, equ. 2, p-57 ], we have
$$ S=\sum_{n=0}^\infty\,\sum_{m=0}^\infty\,\frac{(b)_{n+m}\;(-1)^m\;y^m\;x^n}{(c)_{n+m}\;(1-b-n-m)_m\;m!\;n!}$$
Using $(1-b-n-m)_m= \frac{(-1)^m\;(b)_{n+m}}{(b)_n}$
$$ S=\sum_{n=0}^\infty\,\sum_{m=0}^\infty\,\frac{(b)_n\;x^n\;y^m}{(c)_{n+m}\;m!\;n!}$$
Finally, using definition \eqref{1.2}, we have
$$ S=\Phi_3(b;\; c;\; x,\; y) $$
This completes the proof of \eqref{1.5}.
\section{SPECIAL CASES}
In this section, we shall mention two interesting special cases of our results \eqref{1.5}.
In \eqref{1.5}, if we take $y=x^2$, we have
\begin{equation}
\label{3.1}
\Phi_3(b;\,c;\,x,\,x^2)=\exp\left(2x \right)\; \sum_{k=0}^\infty\;\frac{(-x)^k}{k!}\; \left(-\frac{y}{x} \right)^k
{}_2F_1\left[ \begin{array}{c} -k,\; -k-c+b+1 \\ c \end{array} ; \; 1 \right]\,.
\end{equation}
The $_2F_1$ appearing on the right-hand side of \eqref{3.1} can be evaluated with the help of classical Gauss's summation theorem \cite{Manako5}
\begin{equation}
\label{3.2}
_2F_1 \left[\begin{array}{c} a, \, b \\ c \end{array} ; 1 \right]=\frac{\Gamma(c)\;\Gamma(c-a-b)}{\Gamma(c-a)\;\Gamma(c-b)}
\end{equation}
provided $\text{Re}(c-a-b)>0$.\\
and we get, after some simplification, the following new reduction formula
\begin{equation}
\label{3.3}
\Phi_3(b;\,c;\,x,\,x^2)=\exp\left(2x\right)\,{}_2F_2 \left[ \begin{array}{c} c-\frac{b}{2}, \, c - \frac{b}{2}-\frac{1}{2}\\ \,c,\,2c-b-1\end{array} \,;\, -4x\,\right]
\end{equation}
Further, in \eqref{3.3}, if we take $c=2b$, we have
$$\Phi_3(b;\,2b;\,x,\,x^2)=\exp\left(2x\right)\,{}_2F_2 \left[\begin{array}{c} \frac{3b}{2}, \; \frac{3b-1}{2} \\ 2b,\; 3b-1 \end{array};\;-4x\,\right] $$
and using \eqref{1.4} after simplification, we get
\begin{equation}
\label{3.4}
\Psi_2(b;\; b,\;2b;\;x,\; x)=\,_2F_2 \left[\begin{array}{c} \frac{3b}{2},\; \frac{3b-1}{2} \\ 2b, \; 3b-1\end{array};\; 4x\right]
\end{equation}
which is a special case of the following result
$$\Psi_2(b;\, b,\, c;\, x,\, x)=\,_3F_3 \left[\begin{array}{c} a,\; \frac{c+b}{2},\; \frac{c+b-1}{2} \\ b,\;c,\; c+b-1\end{array} ;\;4x\right]$$
given by Burchnall and Chaundy \cite{Manako5,Manako6}, also recorded in \cite{Manako1,Manako2}.
Similarly, other results can also be obtained.
\section*{Remark :}
In 2015, Choi and Rathie [8] obtained explicit expressions of
$\Phi_2( a,\; a+i;\; c;\; x, \; -x)\,$
and
$\,\Psi_2(a, \; c;\; c+i; \; x,\; -x)\,$
each for $\,i= 0, \pm 1, \pm 2, \cdots, \pm 5\,$.\\
and deduced interesting summation formulas.
\vspace{0.6cm}
|
\section{Introduction}
As is well known, the basic ingredients relevant to the strong interactions are quarks and gluons to be described by quantum chromodynamics (QCD).
However, we believe that they are confined and never observed in their isolated form and that only the color-singlet composite particles such as mesons, baryons and glueballs consisted of quarks and gluons can be observed in experiments.
In particular, glueballs were predicted in the middle of the 1970s by Fritzsch and Minkowski \cite{FM75}, in contrast to mesons and baryons to be explained by the quark model proposed by Gell-Mann in the middle of the 1960s \cite{Gell-Mann64}. Since then, the properties of glueballs have been investigated in various theoretical frameworks and also by experiments.
The equation for describing the bound states in the framework of the relativistic quantum field theory is known as the Nambu-Bethe--Salpeter equation or the Bethe--Salpeter (BS) equation \cite{Nambu50,BS51}.
The BS equation is a self-consistent equation written in terms of the BS amplitude as the wave function of the bound state. See e.g., \cite{Nakanishi69} for reviews.
The BS equation is specified for a given set of propagators and the interaction vertex functions describing the constituent particles.
In the non-perturbative framework, propagators and vertex functions can be obtained by solving the Schwinger-Dyson (SD) equation which is a self-consistent equation written in terms of propagators and vertex functions for the fundamental fields in a given field theory.
In order to describe the bound state in the non-perturbative framework, therefore, we must take into account the SD equation and the BS equation simultaneously:
The SD equation must be solved to give propagators and vertex functions as inputs for specifying the BS equation. Then the BS equation is to be solved in the same level of approximations as those used in solving the SD equation, as far as they cannot be solved exactly.
The purpose of this paper is to understand glueball formation from the underlying dynamics of the Yang-Mills theory in a way consistent with confinement and ultraviolet asymptotic freedom.
In order to avoid complications arising from the mixing of glueballs with mesons in QCD, we restrict our consideration to the pure glue dynamics, namely, pure Yang-Mills theory. Therefore, a glueball is supposed to be a bound state of gluons.
In this paper, especially, we study a glueball as a two-gluon bound state for simplifying the treatment.
Already in the late 1970s, Fukuda \cite{Fukuda78} studied a two-body bound state of gluons and ghosts as solution of the BS equation in the ladder approximation in the framework of the Yang-Mills theory with the covariant gauge fixing.
He has shown the existence of the tachyon bound state due to the dominating attractive force between two gauge (gluon) particles in the color singlet and spin singlet channel, for whatever small coupling constant.
The gauge particles interact among themselves through a three-gluon interaction and a contact four-gluon interaction, and interact with ghosts through the gluon-ghost interaction due to the covariant gauge fixing.
The three-gluon interaction gives an attractive force between two gauge particles in the color singlet channel by exchanging a gauge particle (This is the analog of the attraction of opposite charge in the Abelian gauge theory).
While the contact four-gluon interaction produces a repulsive force. However, if the attraction due to former dominates the repulsion from the latter, the two gauge particles can form a bound state by the net attractive force.
This is nothing but a pairing phenomena familiar in the BCS theory of superconductivity.
Since gauge particles are massless, however, the resulting bound state should necessarily correspond to a tachyon pole.
However, his observations play the very important role for obtaining glueballs consisted of massive gluons in this paper.
It was also shown \cite{GM78} that such a tachyonic bound state continues to exist even in the presence of $n$ flavors of quarks and gluons acquire a mass due to vacuum rearrangement, when $n$ is smaller than or equal to a critical value $n_c$, $n \le n_c$, while all particles remain massless for $n>n_c$.
In the classical level, indeed, the Yang-Mills theory in the four-dimensional spacetime is a massless theory with no scale. In the quantum level, however, the Yang-Mills theory should have a mass scale as suggested from the dimensional transmutation and could have the mass gap.
Recently, a solution corresponding to a confining gluon of the massive type was discovered as a solution of the coupled SD equation for the gluon and ghost propagators in the $SU(N)$ Yang-Mills theory with the Landau gauge fixing, which is called the decoupling solution \cite{decoupling}, in sharp contrast to the scaling solution \cite{scaling}. The existence of the decoupling solution is supported and its confining properties are investigated by various analytical and numerical frameworks.
However, it is not yet clear how to understand the decoupling solution as a solution to be consistent with gluon confinement and more general color confinement, see e.g., \cite{Kondo11}.
The {Brout-Englert-Higgs mechanism} is accepted as a unique way for providing the mass for gauge bosons in quantum field theories, because it is the only one established method which enables one to maintain both {renormalizability} and {physical unitarity}.
In this paper, we consider a specific model including a mass term for the Yang-Mills field, which is a special case of the Curci-Ferrari model \cite{CF76}. We call this model the massive Yang-Mills theory for later convenience.
We adopt this model to perform the analytical investigation for gluons of massive type, instead of using the decoupling solution obtained by solving the SD equation, since the decoupling solution has been so far obtained only in a numerical way, and the analytical expression for the decoupling solution is not yet known explicitly.
We started our analysis on massive Yang-Mills theory in the previous papers \cite{KSFNS13} without introducing the Higgs field which causes the Higgs mechanism.
Nevertheless, we are not plagued by the unitarity violation in high energy region of the massive Yang-Mills theory without the Higgs scalar.
This is because we regard our massive Yang-Mills theory just as a low-energy effective theory (valid below a certain cutoff $\Lambda$) of the quantum Yang-Mills theory with mass gap and confinement.
To mimic precisely the decoupling solution, we must introduce the momentum-dependent gluon mass $m(p)$ which vanishes in the high-energy region and remains non-zero in the low-energy region.
In this paper, we have adopted a constant mass $m$ valid up to a momentum cutoff $\Lambda$ to simplify the analyses without assuming the specific function of the momentum $p$.
Our standpoint mentioned above has been already given more explicitly in the second paper of \cite{KSFNS13} and is not repeated here.
It is also known that this model of massive Yang-Mills theory is still renormalizable, but loses the physical unitarity as least in the perturbation theory \cite{KSFNS13}.
We leave the issue how to realize it as a sound model in a quantum field theoretical framework.
We have a hope that this issue will be resolved in a non-perturbative framework.
One of the motivation of this paper is to obtain some information towards this direction to consider bound states in a {massive Yang-Mills theory without the Higgs field} as a low-energy effective theory of QCD.
The issue of a tachyon bound state will be avoided by starting from the massive Yang-Mills particles instead of the massless gauge particles. Consequently, we will obtain the massive bound states consisted of two massive gluons described by the massive Yang-Mills theory.
In obtaining such a solution, especially, we pay special attention to the aspect how our results for the bound state is consistent with the ultraviolet asymptotic freedom \cite{Fukamachi15}.
This is a feature that has not been discussed in the preceding works \cite{MS13,SAFKS15}.
In their works, the precise determinations of the glueball spectrum have been the main purpose of investigations.
Such an issue is not the aim of the present paper, although it is a nice goal of future investigations to be achieved by developing our method.
We give a systematic derivation of the coupled BS equation for gluon and ghost by starting from the the Cornwall-Jackiw-Toumboulis effective (CJT) action for the composite operator with the bilocal source term.
We show that the bound state obtained as a solution of the homogeneous BS equation for the gluon-gluon BS amplitude is consistent with the ultraviolet asymptotic freedom.
This paper is organized as follows.
In sec. II, we introduce the massive Yang-Mills theory from which we start the study of the bound state.
In section III, we give the CJT effective action for the bilocal composite operator, which gives a systematic derivation of the BS equation.
In sec. IV, we derive the BS equation by differentiating the CJT effective action with respect to the full Green functions.
In sec. V, we write down the coupled BS equation for gluon and ghost at vanishing total momentum $P=0$ in the massless case, and obtain the numerical solutions. This is a preliminary study of obtaining the bound state.
In sec. VI, we write down the homogeneous BS equation at the general total momentum $P$ in the massive case, but restricting to the gluon-gluon BS amplitude alone.
In sec. VII,
we solve the BS equation for the gluon-gluon BS amplitude at the general total momentum $P$ in the massive case.
The final section is devoted to conclusion and discussion.
Some technical materials are collected in the Appendices.
\section{The massive Yang-Mills theory}
In order to discuss the bound state, we adopt the massive $SU(N)$ Yang-Mills theory without the Higgs field, which is defined to be the usual massless $SU(N)$ Yang-Mills theory in the most general Lorenz gauge \cite{Baulieu85} plus the ``mass term'' $\mathscr{L}_m$ formulated in a manifestly Lorentz covariant way.
The total Lagrangian density $\mathscr{L}^{\rm{tot}}_{m\rm{YM}}$ is written in terms of the Yang-Mills field $\mathscr{A}_\mu$, the Faddeev-Popov ghost field $\mathscr{C}$, the antighost field $\bar{\mathscr{C}}$ and the Nakanishi-Lautrup auxiliary field $\mathscr{N}$, if we use the terminology adopted in the usual massless Yang-Mills theory:
\begin{subequations}
\begin{align}
\mathscr{L}^{\rm{tot}}_{m\rm{YM}} =& \mathscr{L}_{\rm{YM}} + \mathscr{L}_{\rm{GF+FP}} + \mathscr{L}_{m} ,
\\
\mathscr{L}_{\rm{YM}} =& - \frac{1}{4} \mathscr{F}_{\mu \nu} \cdot \mathscr{F}^{\mu \nu} ,
\\
\mathscr{L}_{\rm{GF+FP}}
=& \mathscr{N} \cdot \partial^{\mu} \mathscr{A}_{\mu} + i \bar{\mathscr{C}} \cdot \partial^{\mu} \mathscr{D}_{\mu}[\mathscr{A}] \mathscr{C}
+ \frac{\beta}{4} ( \bar{\mathscr{N}} \cdot \bar{\mathscr{N}} + \mathscr{N} \cdot \mathscr{N})
+ \frac{\alpha}{2} \mathscr{N} \cdot \mathscr{N}
\nonumber\\
=& \frac{\alpha}{2} \mathscr{N} \cdot \mathscr{N} + \frac{\beta}{2} \mathscr{N} \cdot \mathscr{N}
+ \mathscr{N} \cdot \partial^{\mu} \mathscr{A}_{\mu}
- \frac{\beta}{2} g \mathscr{N} \cdot (i \bar{\mathscr{C}} \times \mathscr{C})
\nonumber\\
& + i \bar{\mathscr{C}} \cdot \partial^{\mu} \mathscr{D}_{\mu}[\mathscr{A}] \mathscr{C}
+ \frac{\beta}{4} g^2 (i \bar{\mathscr{C}} \times \mathscr{C}) \cdot (i \bar{\mathscr{C}} \times \mathscr{C})
,
\\
\mathscr{L}_{m} =& \frac{1}{2} m^2 \mathscr{A}_{\mu} \cdot \mathscr{A}^{\mu} + \beta m^2 i \bar{\mathscr{C}} \cdot \mathscr{C} ,
\end{align}
\end{subequations}
where
$\mathscr{D}_{\mu}$ is the covariant derivative defined by
\begin{align}
\mathscr{D}_{\mu}[\mathscr{A}] \omega(x)
:= \partial_{\mu} \omega(x) + g \mathscr{A}_\mu(x) \times \omega(x) ,
\end{align}
and
$\bar{\mathscr{N}}$ is defined by
\begin{equation}
\bar{\mathscr{N}} :=-\mathscr{N}+gi\bar{\mathscr{C}} \times \mathscr{C} .
\end{equation}
Here we have adopted the notation:
\begin{align}
\mathscr{X} \cdot \mathscr{Y}
:=& \mathscr{X}^a \mathscr{Y}^a
\quad
(\mathscr{X} \times \mathscr{Y})^c
:= f_{abc} \mathscr{X}^a \mathscr{Y}^b ,
\end{align}
where $f_{abc}$ is the structure constants of the Lie algebra of the $SU(N)$ group.
Here $\alpha$ and $\beta$ are parameters which are identified with the gauge-fixing parameters in the $m \rightarrow 0$ limit.
The $\alpha=0$ case is the Curci-Ferrari model \cite{CF76} with the coupling constant $g$, the mass parameter $m$ and the parameter $\beta$.
In the Abelian limit with vanishing structure constants $f_{abc}=0$, the Faddeev-Popov ghosts decouple and the Curci-Ferrari model reduces to the Nakanishi model \cite{Nakanishi72}.
In what follows, we put $\beta=0$ to simulate the decoupling solution representing massive gluon and massless ghost.
By eliminating the Nakanishi-Lautrup field $\mathscr{N}$, $\mathscr{L}^{\rm{tot}}_{m\rm{YM}}$ reads
\begin{align}
\mathscr{L}^{\rm{tot}}_{m\rm{YM}}
=&-\frac{1}{4}(\partial _\mu \mathscr{A}_{\nu} -\partial _\nu \mathscr{A}_{\mu}) \cdot (\partial ^\mu \mathscr{A}^{\nu} -\partial ^\nu \mathscr{A}^{\mu})
- \frac{1}{2\alpha} \left( \partial ^\mu \mathscr{A}_{\mu} \right) ^2
+ i\bar{\mathscr{C}} \cdot \Box \mathscr{C}
\notag \\
-\frac{1}{2}g\left( \partial _\mu \mathscr{A}_{\nu} -\partial _\nu \mathscr{A}_{\mu} \right) \cdot \left( \mathscr{A}^{\mu} \times \mathscr{A}^{\nu} \right)
-\frac{1}{4}g^2\left( \mathscr{A}_{\mu} \times \mathscr{A}_{\nu} \right) ^2
+g i\bar{\mathscr{C}} \cdot \partial ^\mu \left( \mathscr{A}_{\mu} \times \mathscr{C} \right)
\notag \\
&+\frac{1}{2}m^2 \mathscr{A}_{\mu} \cdot \mathscr{A}^{\mu} .
\label{mYM}
\end{align}
By starting from the massive Yang-Mills theory given by the Lagrangian density (\ref{mYM}), we derive the BS equation which includes simultaneously gluon and ghost on equal footing.
This can be done using the CJT effective action for the composite operators, as explicitly shown in \cite{Fukamachi15}.
See \cite{KSFNS13} for more details on this model.
\section{Cornwall--Jackiw--Tomboulis effective action}
For the Yang--Mills theory, we introduce the \textbf{local sources} $J^\mu _a(x)$, $\bar{\eta}_a(x)$, and $\eta _a(x)$ for the \textbf{gluon field} $\mathscr{A}_\mu ^a(x)$, the \textbf{ghost field} $\mathscr{C}^a(x)$, and the \textbf{antighost field} $\mathscr{\bar{C}}^a(x)$ respectively.
Moreover, we introduce the \textbf{bilocal sources} $I_{ab}^{\mu \nu}(x,y)$ and $\theta _{ab}(x,y)$ for the bilocal composite operators $\mathscr{A}_\mu^a(x) \mathscr{A}_\nu^b(y)$ and $\mathscr{\bar{C}}^a(x) \mathscr{C}^b(y)$ respectively,
where the bilocal source $I$ is symmetric $I^{ab}_{\mu \nu}(x,y)=I^{ba}_{\nu \mu}(y,x)$ and $\theta ^{ab}(x,y)$ is a general matrix.
Then the generating functional $W[J,\bar{\eta},\eta ,I,\theta ]$ is defined by
\begin{align}
& Z[J,\bar{\eta},\eta ,I,\theta ]
= e^{iW[J,\bar{\eta},\eta ,I,\theta ]}
\notag \\
=&\int \mathcal{D}\Phi \exp \Bigg\{ iS^{\rm{tot}}_{m\rm{YM}}[\mathscr{A},\mathscr{\bar{C}},\mathscr{C}]+i\int_{x} \left[ J^\mu _a(x)\mathscr{A}_\mu ^a(x)+\bar{\eta}_a(x)\mathscr{C}^a(x)+ \mathscr{\bar{C}}^a(x)\eta _a(x) \right]
\notag \\
&\hspace{3cm}
+i \int_{x} \int_{y} \left[ \frac{1}{2}\mathscr{A}^a_\mu (x)I_{ab}^{\mu \nu}(x,y)\mathscr{A}^b_\nu (y)+ \mathscr{\bar{C}}^a(x)\theta _{ab}(x,y) \mathscr{C}^b(y)\right] \Bigg\} ,
\end{align}
where we have introduced the abbreviated notation:
$\int_{x}:=\int d^Dx$.
For the local source, we find that the (left) derivatives of $W$ with respect to the local sources are given by
\begin{subequations}
\begin{align}
\frac{\delta W}{\delta J_a^\mu (x)}&=\left< \mathscr{A}^a_\mu (x)\right> \equiv \bm{A}_\mu ^a(x) ,
\\
\quad
\frac{\delta W}{\delta \bar{\eta}_a(x')} &= \left< \mathscr{C}^a(x')\right> \equiv \bm{C}^a(x') ,
\quad
\frac{\delta W}{\delta \eta _a(x')} =-\left< \mathscr{\bar{C}}^a(x')\right> \equiv -\bar{\bm{C}}^a(x') ,
\\
\frac{\delta ^2W}{\delta J^\nu _b(y)\delta J_a^\mu (x)}&=i\left< \mathscr{A}^a_\mu (x); \mathscr{A}^b_\nu (y)\right> \equiv iD^{ab}_{\mu \nu}(x,y) ,
\\
\frac{\delta ^2W}{\delta \eta _b(y')\delta \bar{\eta}_a(x')} &= i\left< \mathscr{C}^a(x');\mathscr{\bar{C}}^b(y')\right> \equiv i\Delta ^{ab}(x',y') ,
\\
\frac{\delta ^2W}{\delta \bar{\eta}_b(y')\delta \eta _a(x')} &=i\left< \mathscr{\bar{C}}^a(x');\mathscr{C}^b(y)\right> \equiv i\tilde{\Delta}^{ab}(x',y') ,
\end{align}
\end{subequations}
where $D$ and $\tilde{\Delta}$ are the full gluon and ghost-antighost propagators.
For the bilocal source, we find that the derivatives of $W$ with respect to the bilocal sources are given by
\begin{subequations}
\begin{align}
\frac{\delta W}{\delta I_{ab}^{\mu \nu}(x,y)}&
=\left< \mathscr{A}_\mu ^a(x) \mathscr{A}_\nu ^b(y)\right> =\left< \mathscr{A}_\mu ^a(x); \mathscr{A}_\nu ^b(y)\right> +\left< \mathscr{A}_\mu ^a(x)\right> \left< \mathscr{A}_\nu ^b(y)\right> \notag \\
&=D_{\mu \nu}^{ab}(x,y)+\bm{A}_\mu ^a(x)\bm{A}_\nu ^b(y) ,
\\
\frac{\delta W}{\delta \theta _{ab}(x',y')}&
=\left< \mathscr{\bar{C}}^a(x')\mathscr{C}^b(y')\right> =\left< \mathscr{\bar{C}}^a(x');\mathscr{C}^b(y')\right> +\left< \mathscr{\bar{C}}^a(x')\right> \left< \mathscr{C}^b(y')\right> \notag \\
&=\tilde{\Delta} ^{ab}(x',y')+\bar{\bm{C}}^a(x')\bm{C}^b(y') .
\end{align}
\end{subequations}
We define the CJT effective action \cite{CJT74} of the Yang-Mills theory by the Legendre transform:
\begin{align}
\Gamma [ \bm{A},\bar{\bm{C}},\bm{C},D,\tilde{\Delta} ]=&W[J,\bar{\eta},\eta ,I,\theta ] - \int_{x} \left( J_a^\mu (x)\bm{A}^a_\mu (x)+\bar{\eta}_a(x)\bm{C}^a(x)+\bar{\bm{C}}^a(x)\eta _a(x) \right)
\notag \\
&-\frac{1}{2} \int_{x} \int_{y} \left( \bm{A}^a_\mu (x)I_{ab}^{\mu \nu}(x,y)\bm{A}^b_\nu (y)+I_{ab}^{\mu \nu}(x,y)D^{ab}_{\mu \nu}(x,y)\right) \notag \\
&- \int_{x'} \int_{y'} \left( \bar{\bm{C}}^a(x')\theta _{ab}(x',y')\bm{C}^b(y')+\theta _{ab}(x',y')\tilde{\Delta}^{ab} (x',y')\right) .
\label{effective_action}
\end{align}
Then we find the derivatives of $\Gamma$ with respect to the field expectation,
\begin{subequations}
\begin{align}
\frac{\delta \Gamma}{\delta \bm{A}_\mu ^a(x)}&=-J^\mu _a(x)- \int_{y} I_{ab}^{\mu \nu}(x,y)\bm{A}_\nu ^b(y) ,
\\
\frac{\delta \Gamma}{\delta \bar{\bm{C}}^a(x')}&= - \eta _a(x')- \int_{y'} \theta _{ab}(x',y')\bm{C}^b(y') ,
\quad
\frac{\delta \Gamma}{\delta \bm{C}^b(y')} = \bar{\eta}_b(y')+ \int_{y'} \bar{\bm{C}}^a(x')\theta _{ab}(x',y') ,
\end{align}
and the derivatives of $\Gamma$ with respect to the full propagators,
\begin{align}
\frac{\delta \Gamma}{\delta D_{\mu \nu}^{ab}(x,y)}&=-I^{\mu \nu}_{ab}(x,y) ,
\quad
\frac{\delta \Gamma}{\delta \tilde{\Delta} ^{ab}(x',y')} =-\theta _{ab}(x',y') .
\end{align}
\end{subequations}
It is shown that the CJT effective action is rewritten as
\begin{align}
\Gamma [\bm{A},\bm{C},\bar{\bm{C}},D,\tilde{\Delta}]
=& S^{\rm{tot}}_{m\rm{YM}}[\bm{A},\bm{C},\bar{\bm{C}}] + \frac{i}{2}{\text{Tr}} \left[ \mathscr{D}^{-1}_{AA}D\right] - i{\text{Tr}}\left[ \bar{\mathscr{D}}^{-1}_{C\bar{C}}\tilde{\Delta}\right]
\notag \\
&+ \frac{i}{2}{\text{Tr}}\ln D^{-1} + i{\text{Tr}}\ln \tilde{\Delta}^{-1}
+ \Gamma _2[\bm{A},\bm{C},\bar{\bm{C}},D,\tilde{\Delta}] ,
\end{align}
where $\Gamma_2$ is the two particle irreducible (2PI) contributions,
$D$ and $\tilde{\Delta}$ are respectively the full gluon and ghost propagators, $\mathscr{D}_{AA}$ and $\bar{\mathscr{D}}_{C\bar{C}}$ are respectively the tree-level gluon and ghost propagators defined by
\begin{align}
\frac{\delta ^2 \bm{S}}{\delta \bm{A}_\nu ^b(y)\delta \bm{A}_\mu ^a(x)}&
\equiv (i\mathscr{D}^{-1}_{AA})_{\mu\nu}^{ab}(x,y) ,
\quad
\frac{\delta ^2 \bm{S}}{\delta \bar{\bm{C}}^a(x) \delta \bm{C}^b(y)}
\equiv (i\bar{\mathscr{D}}^{-1}_{C\bar{C}})^{ab}(x,y) .
\end{align}
The original Yang-Mills theory is characterized by the bare gluon and ghost propagators, and the bare one particle irreducible (1PI) vertex functions: 3-gluon, 4-gluon, and gluon-ghost-antighost vertices given in Fig.~\ref{fig:propagator-vertex}.
The full gluon and ghost propagators, and the full 1PI vertex functions: 3-gluon, 4-gluon, and gluon-ghost-antighost vertices are given in Fig.~\ref{fig:propagator-vertex-full}.
It is shown that the 2PI vacuum diagrams in the Yang-Mills theory are written as Fig.~\ref{fig:vacuum-diagram}.
\begin{figure}[t]
\centering
\includegraphics[scale=0.54]{fig-ep218/YM_propagator}
\quad\quad
\includegraphics[scale=0.37]{fig-ep218/YM_interaction}
\caption{
Bare gluon and ghost propagators;
the bare vertex functions: 3-gluon, 4-gluon, gluon-ghost-antighost vertices in the Yang-Mills theory.
}
\label{fig:propagator-vertex}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.54]{fig-ep218/YM_full_propagator}
\quad\quad
\includegraphics[scale=0.37]{fig-ep218/YM_full_interaction}
\caption{
Full gluon and ghost propagators;
the full vertex functions: 3-gluon, 4-gluon, gluon-ghost-antighost vertices in the Yang-Mills theory.
}
\label{fig:propagator-vertex-full}
\end{figure}
\begin{figure}[t]
\includegraphics[scale=0.42]{fig-ep218/2PI}
\caption{
2PI vacuum diagrams in the Yang-Mills theory.
}
\label{fig:vacuum-diagram}
\end{figure}
\section{A systematic derivation of the Bethe--Salpeter equation}
\subsection{Derivation of the Schwinger--Dyson equation}
By taking the first derivative of the CJT effective action $\Gamma$ with respect to the full gluon and ghost propagators $D$, we obtain the SD equations for the gluon propagator $D$ and ghost propagator $\delta \tilde{\Delta}$ respectively (after the sources are setting to be zero $I, \theta \to 0$):
\begin{subequations}
\begin{align}
-I = \frac{\delta \Gamma}{\delta D} = -iD^{-1} + i\mathscr{D}_{AA}^{-1}+ \frac{\delta \Gamma_2}{\delta D}
\\
-\theta = \frac{\delta \Gamma}{\delta \tilde{\Delta}(x,y)} = -i\tilde{\Delta}^{-1} + i\tilde{\mathscr{D}}_{C\bar{C}}^{-1}+ \frac{\delta \Gamma_2}{\delta \tilde{\Delta}}
\end{align}
\end{subequations}
Fig.~\ref{fig:gluon-SD} is the gluon SD equation and Fig.~\ref{fig:ghost-SD} is the ghost SD equation.
\begin{figure}[h]
\centering
\includegraphics[scale=0.42]{fig-ep218/SDgluon_in_momentum2}
\caption{
Gluon-SD equation.
}
\label{fig:gluon-SD}
\includegraphics[scale=0.40]{fig-ep218/SDghost_in_momentum}
\caption{
Ghost-SD equation.
}
\label{fig:ghost-SD}
\end{figure}
\subsection{The equation for the scattering Green function}
The BS equation is obtained by taking the first derivative of the SD equation (i.e., the second derivative of the CJT effective action) with respect to the full propagator.
For the general field $\phi$, the \textbf{scattering Green function} $D_4$ with four external lines is defined by the second derivative of $W$ with respect to the bilocal source $I$:
\begin{align}
D_4 (x^\prime,y^\prime;x,y)
=& \left< 0|\phi (x)\phi (y);\phi (x^\prime)\phi (y^\prime)|0\right>
= \frac{1}{i} \frac{\delta ^2 W}{\delta I (x^\prime,y^\prime)\delta I (x,y)} ,
\end{align}
while the inverse $D_4^{-1}$ of the {scattering Green function} $D_4$ is given by the second derivative of $\Gamma$ with respect to the full propagator $D$:
\begin{align}
D_4^{-1} := \frac{\delta^2 \Gamma}{\delta D\delta D} .
\end{align}
Therefore, the first derivative of the SD equation yields the equation for the inverse $D_4^{-1}$ of the {scattering Green function} $D_4$:
\begin{align}
D_4^{-1} = D_4^{(0)}{}^{-1} + K ,
\quad D_4^{(0)}{}^{-1} := \frac{\delta^2 [-iD^{-1}]}{\delta D\delta D} ,
\quad K := \frac{\delta^2 \Gamma_2}{\delta D\delta D} ,
\end{align}
where $D_4^{(0)}$ is the tree part of the scattering Green function $D_4$ and $K$ denotes the kernel.
The diagrammatic representation is given as
\begin{equation}
\includegraphics[scale=0.36]{fig-ep218/BS1}
\end{equation}
We make use of the identities:
\begin{equation}
\includegraphics[scale=0.30]{fig-ep218/identity}\hspace{1.5cm} {\rm{and}} \hspace{1.5cm} \includegraphics[scale=0.30]{fig-ep218/identity2}
\end{equation}
to rewrite the equation in terms of $D_4$, not the inverse $D_4^{-1}$.
We multiply $D_4$ from the right to obtain
\begin{equation}
\includegraphics[scale=0.30]{fig-ep218/BS2}
\end{equation}
and then multiply $D_4^{(0)}$ from the left to obtain
\begin{equation}
\includegraphics[scale=0.30]{fig-ep218/BS3}
\end{equation}
Thus we obtain the self-consistent equation for the scattering Green function $D_4$ after multiplying $i$:
\begin{equation}
\includegraphics[scale=0.36]{fig-ep218/BS4}
\end{equation}
For the Yang-Mills theory, we obtain a set of the BS equations for the four BS amplitude: gluon-gluon, gluon-ghost, ghost-gluon, and ghost-ghost scattering Green function
as given in Fig.~\ref{fig:mYM-BS4}.
\begin{figure}[t]
\includegraphics[scale=0.44]{fig-ep218/BSE_gluon}
\includegraphics[scale=0.44]{fig-ep218/BSE_ghost}
\caption{
The BS equation for
gluon-gluon $D$,
gluon-ghost $G$,
ghost-gluon $G$,
and
ghost-ghost $\Delta$ scattering Green functions.
}
\label{fig:mYM-BS4}
\end{figure}
\subsection{The homogeneous BS equation for the (amputated) BS amplitude}
We can use the completeness relation:
\begin{equation}
\bm{1}=|0\big> \big< 0|+\sum _c \int \frac{d^d\bm{P}_c}{(2\pi )^d}\frac{1}{2E_{P_c}}|P_c\big> \big< P_c| +\cdots ,
\end{equation}
where the sum $\sum _c$ runs over all the bound states except for the continuum spectrum.
The {scattering Green function} $D_4$ for particles described by the field $\phi_a$ is defined by
\begin{align}
D_4{}^{a'b';ab}(x',y';x,y)
=& \left< 0|\phi^a (x)\phi^b (y);\phi^{a'}(x')\phi^{b'}(y')|0\right>
= \frac{1}{i} \frac{\delta ^2W}{\delta I^{a'b'}(x',y')\delta I^{ab}(x,y)} .
\end{align}
This is decomposed into the product of the \textbf{BS amplitude} $\chi$:
\begin{align}
D_4{}^{a'b';ab}(x',y';x,y)
= &\sum _c\int \frac{d^dP_c}{(2\pi )^d}\frac{1}{2E_{P_c}}\chi ^{ab}_{P}(x,y)\bar{\chi}^{a'b'}_{P}(x',y') + ... ,
\end{align}
where the BS amplitude $\chi$ is defined by
\begin{equation}
\chi ^{ab}_{P}(x,y)=\left< 0|\phi^a (x)\phi^b (y)|P_c\right> .
\end{equation}
The momentum space representation of $D_4$ has the pole:
\begin{align}
D_4{}^{a'b';ab} (q,p;P) =& \sum _c \frac{i}{(2\pi )^D}\frac{\chi ^{ab}_{P }(p)\bar{\chi}^{a'b'}_{P }(q)}{P^2-M^2} + ...,
\end{align}
the product $\bar{\chi} \chi$ of the BS amplitude corresponds to the residue at the pole where the squared momentum $P^2$ coincides with the bound state mass $M$ squared: $P^2=M^2=P_c^2$.
\begin{equation}
\includegraphics[scale=0.3]{fig-ep218/BS_2body_propagator}
\end{equation}
From the equation for the scattering Green function $D_4$,
\begin{equation}
\includegraphics[scale=0.35]{fig-ep218/BS5}
\end{equation}
we can extract the product $\bar{\chi} \chi$ of the BS amplitude by identifying it with the pole residue at the pole corresponding to the bound state, since $D_4^{(0)}$ does not have the pole:
\begin{equation}
\includegraphics[scale=0.3]{fig-ep218/BS_amputate1}
\end{equation}
By multiplying $D_4^{(0)}{}^{-1}$ from the left,
\begin{equation}
\includegraphics[scale=0.3]{fig-ep218/BS_amputate2}
\end{equation}
we finally obtain the \textbf{homogeneous BS equation} for the \textbf{amputated BS amplitude} $D^{(0)}{}^{-1}\bar\chi$:
\begin{equation}
\includegraphics[scale=0.3]{fig-ep218/BS_amputate3}
\end{equation}
\rule[0cm]{0cm}{0cm}
For the Yang-Mills theory, the {scattering Green function} $D_4$ for the gluon field is defined by
\begin{align}
D_4{}^{a'b';ab}_{\mu '\nu ';\mu \nu}(x',y';x,y)
=& \left< 0|\mathscr{A}^a_\mu (x)\mathscr{A}^b_\nu (y);\mathscr{A}^{a'}_{\mu '}(x')\mathscr{A}^{b'}_{\nu '}(y')|0\right>
= \frac{1}{i} \frac{\delta ^2W}{\delta I_{\mu '\nu '}^{a'b'}(x',y')\delta I_{\mu \nu}^{ab}(x,y)} .
\end{align}
This is decomposed into the product of the \textbf{BS amplitude} $\chi$:
\begin{align}
D_4{}^{a'b';ab}_{\mu '\nu ';\mu \nu}(x',y';x,y)
= &\sum _c\int \frac{d^dP_c}{(2\pi )^d}\frac{1}{2E_{P_c}}\chi ^{ab}_{P;\mu \nu}(x,y)\bar{\chi}^{a'b'}_{P;\mu '\nu '}(x',y') + ... .
\end{align}
In the momentum space representation, we have
\begin{align}
D_4{}^{a'b';ab}_{\mu '\nu ';\mu \nu}(q,p;P) =& \sum _c \frac{i}{(2\pi )^D}\frac{\chi ^{ab}_{P;\mu \nu}(p)\bar{\chi}^{a'b'}_{P;\mu '\nu '}(q)}{P^2-M^2} + ...,
\end{align}
We introduce the BS amplitudes $\chi ^{ab}_{P;\mu \nu}$ and $\chi ^{ab}_P$ for two-gluons and ghost-antighost respectively as
\begin{equation}
\chi ^{ab}_{P;\mu \nu}(x,y)=\left< 0|\mathscr{A}^a_\mu (x) \mathscr{A}^b_\nu (y)|P_c\right> , \quad
\chi ^{ab}_P (x,y)=\left< 0|\mathscr{C}^a(x)\bar{\mathscr{C}}^b(y)|P_c\right> .
\end{equation}
For the Yang-Mills theory, we define the amputated gluon BS amplitude $A$
and the amputated ghost BS amplitude $B$ by
\begin{subequations}
\begin{align}
A^{\mu \nu}_{ab}(p;P)&:=\int \frac{d^Dq}{(2\pi )^D} \left( D^{(0)-1}\right) _{ab;a^\prime b^\prime }^{\mu \nu ;\mu^\prime \nu^\prime }(p,q;P)\bar{\chi}^{a^\prime b^\prime }_{P;\mu^\prime \nu^\prime }(q;P) ,
\\
B_{ab}(p;P)&:=\int \frac{d^Dq}{(2\pi )^D}\left( \Delta ^{(0)-1}\right) _{ab;a^\prime b^\prime }(p,q;P)\bar{\chi}^{a^\prime b^\prime }_P(q;P) .
\end{align}
\end{subequations}
\begin{figure}[t]
\centering
\includegraphics[scale=0.4]{fig-ep218/amputatedBS}
\caption{
The homogeneous BS equation for the amputated gluon and ghost BS amplitudes for the Yang-Mills theory.
}
\label{fig:BS}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.30]{fig-ep218/2-loop_cut}
\caption{
The kernels of the BS equation.
}
\label{fig:mYM-BS-kernel}
\end{figure}
Thus, the amputated BS amplitude obeys the homogeneous BS equation:
\begin{equation}
\begin{pmatrix}
A_{ab}^{\mu \nu}(p^+,p^-) \\ B_{ab}(p^+,p^-)
\end{pmatrix}
= \int \frac{d^Dq}{(2\pi )^D}
\begin{pmatrix}
\mathcal{K}^{ab;cd}_{\mu \nu ;\mu^\prime \nu^\prime } & \mathcal{K}^{ab;cd}_{\mu \nu ;} \\
\mathcal{K}^{ab;cd}_{\ \ ;\mu^\prime \nu^\prime } & \mathcal{K}^{ab;cd}
\end{pmatrix}
\begin{pmatrix}
A_{cd}^{\mu^\prime \nu^\prime }(q^+,q^-) \\ B_{cd}(q^+,q^-)
\end{pmatrix} ,
\end{equation}
where we have defined the momenta $p^\pm$ by
\begin{align}
p^\pm :=p\pm \frac{P}{2} , \quad
q^\pm :=q\pm \frac{P}{2} ,
\end{align}
and $\mathcal{K}$ by
\begin{align}
\mathcal{K}^{ab;cd}_{\mu \nu ;\mu^\prime \nu ^\prime }=K^{ab;ef}_{\mu \nu ;\rho \sigma}D_{4}^{(0)}{}^{ef;cd}_{\rho \sigma ;\mu^\prime \nu^\prime } .
\end{align}
The diagram of the homogeneous BS equation is given in Fig.~\ref{fig:BS}.
Here the kernels of the BS equation are given as Fig.~\ref{fig:mYM-BS-kernel}.
The explicit form of the kernel is given in Appendix A.
This BS equation can be compared with that obtained by
Meyers and Swanson \cite{MS13}.
\subsection{BS amplitude}
In what follows, we consider the color-singlet and scalar (i.e., spin zero) bound state.
Fukuda considered the decomposition for the amplitude:
\begin{align}
A _{\mu \nu}(k;P)
=& \left( g_{\mu \nu}-\frac{k _\mu k _\nu}{k^2} \right) A_1 + \frac{k _\mu k _\nu}{k^2} A_2 + P_\mu P_\nu A_3 + (k_\mu P_\nu + k_\nu P_\mu) (k \cdot P) A_4
\nonumber\\&
+ (k_\mu P_\nu - k_\nu P_\mu) A_5 ,
\end{align}
where $A_1, ..., A_5$ are functions of $k^2, P^2$ and $(k \cdot P)^2$.
Here $A_1$ and $A_2$ are respectively the transverse and longitudinal parts in terms of $k$, since we can define the transverse projector $t_{\mu \nu}$ and longitudinal projector $\ell_{\mu \nu}$ by
\begin{align}
t_{\mu \nu} &:= g_{\mu \nu}-\frac{k_\mu k_\nu}{k^2} , \
\ell_{\mu \nu} :=\frac{k_\mu k_\nu}{k^2} , \
\label{projector}
\end{align}
with the properties:
\begin{subequations}
\begin{align}
t_{\mu \nu} + \ell_{\mu \nu} =& g_{\mu \nu} ,
\\
t_{\mu \rho} t_{\rho \nu} =& t_{\mu \nu} , \
\ell_{\mu \rho} \ell_{\rho \nu} = \ell_{\mu \nu} ,
\\
t_{\mu \nu} =& t_{\nu\mu } , \
\ell_{\mu \nu} = \ell_{\nu\mu } ,
\\
k_\mu t_{\mu \nu}=& 0=t_{\mu \nu}k_\nu ,
\label{transverse}
\\
k_\mu \ell_{\mu \nu}=& k_\nu, \ \ell_{\mu \nu}k_\nu = k_\mu .
\label{longitudinal}
\end{align}
\end{subequations}
In this paper, however, we regard the amplitude $A _{\mu \nu}$ as the function $A _{\mu \nu}(k^+;k^-)$ of $k^+:=k+P/2$ and $k^-:=k-P/2$, rather than the function $A _{\mu \nu}(k;P)$ of $k$ and $P$.
For this purpose, this decomposition in terms of $k$ and $P$ is not suited in the presence of the non-vanishing total momentum $P^\mu$.
We construct another decomposition of the amplitude in terms of $k^+$ and $k^-$.
In particular, we require that the gluon BS amplitude $A_{\mu \nu}(k^+,k^-)$ is transverse in the sense that
\begin{equation}
k^+_\mu A_{\mu \nu}=0=A_{\mu \nu}k^{-}_\nu .
\end{equation}
In view of these, we can define the modified transverse projector in the presence of $P$ by
\begin{align}
T_{\mu \nu}&:= g_{\mu \nu}-\frac{k^-_\mu k^+_\nu}{k^+ \cdot k^-} \ne
T_{\nu\mu } := g_{\mu \nu}-\frac{k^+_\mu k^-_\nu}{k^+ \cdot k^-}
.
\end{align}
which is subject to
\begin{align}
k^+_\mu T_{\mu \nu} &=0=T_{\mu \nu} k^-_\nu .
\end{align}
Indeed, the modified transverse projector $T_{\mu \nu}$ reduces to the usual transverse projector (\ref{projector}) with the property (\ref{transverse}) in the limit $P \to 0$.
In the presence of $P$, we can construct another type of transverse projector.
For this purpose, we can introduce the vectors $k^{+\bot }$ and $k^{-\bot }$ which are orthogonal to $k^+$ and $k^-$ respectively:
\begin{align}
k^+\cdot k^{+\bot} &=0=k^-\cdot k^{-\bot} .
\end{align}
Indeed, such vectors $k^{+\bot }$ and $k^{-\bot }$ are constructed from $k^+$ and $k^-$ as
\footnote{
This is obtained as follows.
We define the unit vectors $e^+$ and $e^-$ of $k^+$ and $k^-$:
\begin{align}
e^+ :=\frac{k^+}{|k^+|} , \
e^- :=\frac{k^-}{|k^-|} ,
\end{align}
and we can use the ortho-normalization method due to Gram--Schmidt to construct the vectors $e^{+\bot}$ and $e^{-\bot}$ which are respectively orthogonal to $e^+$ and $e^-$:
\begin{align}
e^{\pm \bot}&=e^\mp - (e^\mp, e^\pm )e^\pm
= \frac{k^\mp}{|k^\mp|}-\frac{k^\mp \cdot k^\pm }{|k^\mp||k^\pm |}\frac{k^\pm }{|k^\pm |}
= - \frac{k^\mp \cdot k^\pm }{|k^\mp||k^\pm |^2} \left( k^\pm - \frac{ |k^\pm |^2}{k^\mp \cdot k^\pm } k^\mp \right) .
\end{align}
}
\begin{align}
k^{+\bot}_\mu :=k^+_\mu -\frac{|k^+|^2}{k^-\cdot k^+}k^-_\mu , \
k^{-\bot}_\mu :=k^-_\mu -\frac{|k^-|^2}{k^+\cdot k^-}k^+_\mu .
\end{align}
We can also define another transverse projector
\begin{align}
L_{\mu \nu} :=\frac{k^{+\bot}_\mu k^{-\bot}_\nu}{k^{+\bot}\cdot k^{-\bot}} ,
\end{align}
with the property:
\begin{align}
k^+_\mu L_{\mu \nu}&=0=L_{\mu \nu}k^-_\nu .
\end{align}
Notice that $k^{+\bot }_\mu$ and $k^{-\bot }_\mu$ vanish in the limit $P^\mu \to 0$.
In the limit $P \to 0$, therefore, $L_{\mu \nu}$ becomes ill-defined and does not reduce to $\ell_{\mu \nu}$ which is longitudinal to $k$.
They satisfy the following properties:
\begin{subequations}
\begin{align}
T_{\mu \rho}T_{\rho \nu}&=g_{\mu \nu}-\frac{k^-_{\mu } k^+_\nu}{k^-\cdot k^+} = T_{\mu \nu},
\\
L_{\mu \rho}L_{\rho \nu}&=-\frac{\left( k^+_{\mu } (k^- \cdot k^+) -{k^+}^2 k^-_{\mu}\right) \left({k^-}^2 k^+_{\nu }-k^-_{\nu } ({k^-}\cdot {k^+})\right)}{({k^-}\cdot {k^+})^3-{k^-}^2 {k^+}^2({k^-}\cdot {k^+})} ,
\\
T_{\mu \rho}L_{\rho \nu}&=-\frac{\left( k^+_{\mu } (k^-\cdot k^+)-{k^+}^2 k^-_{\mu}\right) \left( {k^-}^2 k^+_{\nu }-k^-_{\nu } (k^-\cdot k^+\right)}{(k^-\cdot k^+)^3-{k^-}^2 {k^+}^2 (k^-\cdot k^+)} ,
\end{align}
\end{subequations}
which lead to
\begin{align}
T_{\mu \nu}T_{\nu \mu} =D-1 , \
L_{\mu \nu}L_{\nu \mu} =1 , \
T_{\mu \nu}L_{\nu \mu} =1 .
\label{TL-prop}
\end{align}
Now we proceed to construct the transverse amplitude.
First, $k^+_\mu A_{\mu \nu}=0$ is satisfied by the tensor of the form:
\begin{equation}
A_{\mu \nu}=f T_{\mu \nu} +g _\nu k^{+\bot}_\mu .
\end{equation}
Second, the condition $A_{\mu \nu}k^{-}_\nu =0$ yields
\begin{equation}
A_{\mu \nu}k^-_\nu
=f T_{\mu \nu}k^-_\nu +g _\nu k^{+\bot}_\mu k^-_\nu
=( g _\nu k^-_\nu ) k^{+\bot}_\mu =0 .
\end{equation}
$g _\nu k^-_\nu =0$ follows from $g _\nu =g k^{-\bot}_\nu$.
Thus, $A_{\mu \nu}$ has the form:
\begin{align}
A_{\mu \nu}&
=f T_{\mu \nu} +h L_{\mu \nu}.
\end{align}
Consequently, we define the gluon BS amplitude $A_{\mu \nu}$ by \begin{equation}
A_{\mu \nu}(k^+,k^-)=A_1(k^+,k^-)T_{\mu \nu}+A_2(k^+,k^-)L_{\mu \nu} .
\label{projection}
\end{equation}
Multiplying both sides of $A_{\mu \nu}(k^+,k^-)$ with
$g_{\nu \mu}$ and $L_{\nu \mu}$ from the left, we obtain using (\ref{TL-prop}):
\begin{align}
g_{\nu \mu} A_{\mu \nu}(k^+,k^-)&=(D-1)A_1(k^+,k^-)+A_2(k^+,k^-) ,
\nonumber\\
L_{\nu \mu} A_{\mu \nu}(k^+,k^-)&=A_1(k^+,k^-)+A_2(k^+,k^-) ,
\end{align}
which has the matrix form:
\begin{equation}
\begin{pmatrix}
g_{\nu \mu} \\ L_{\nu \mu}
\end{pmatrix}
A_{\mu \nu}(k^+,k^-)=
\begin{pmatrix}
D-1 & 1 \\
1 & 1
\end{pmatrix}
\begin{pmatrix}
A_1(k^+,k^-) \\ A_2(k^+,k^-)
\end{pmatrix} .
\label{IM}
\end{equation}
This is solved with respect to $A_1,A_2$:
\begin{equation}
\begin{pmatrix}
A_1(k^+,k^-) \\ A_2(k^+,k^-)
\end{pmatrix}
=
\begin{pmatrix}
\frac{1}{D-2} & \frac{1}{2-D} \\
\frac{1}{2-D} & \frac{D-1}{D-2} \\
\end{pmatrix}
\begin{pmatrix}
g_{\nu \mu} \\ L_{\nu \mu}
\end{pmatrix}
A_{\mu \nu}(k^+,k^-) .
\end{equation}
Thus, $A_1$ and $A_2$ are extracted from $A_{\mu \nu}$ by operating the projection operators $P^{A_1}_{\nu \mu}$ and $P^{A_2}_{\nu \mu}$ respectively:
\begin{subequations}
\begin{align}
A_1 = P^{A_1}_{\nu \mu} A_{\mu \nu}, \ P^{A_1}_{\nu \mu}&:=\frac{1}{D-2}g_{\nu \mu}+\frac{1}{2-D}L_{\nu \mu} ,
\\
A_2 = P^{A_2}_{\nu \mu} A_{\mu \nu}, \ P^{A_2}_{\nu \mu}&:=\frac{1}{2-D}g_{\nu \mu}+\frac{D-1}{D-2} L_{\nu \mu} .
\end{align}
\end{subequations}
Consequently, the homogeneous BS equation for the gluon amplitude is rewritten
using the abbreviated form:
\begin{subequations}
\begin{align}
A^{ab}_1(k^+,k^-)&=i\int \frac{d^Dq}{(2\pi )^D}P^{A_1}_{\nu \mu}\left( \mathcal{K}^{ab;cd}_{\mu \nu ;\mu '\nu '}A_{\mu '\nu '}^{cd}(q^+,q^-)+\mathcal{K}^{ab;cd}_{\mu \nu ;}B^{cd}(q^+,q^-)\right) ,
\\
A^{ab}_2(k^+,k^-)&=i\int \frac{d^Dq}{(2\pi )^D}P^{A_2}_{\nu \mu}\left( \mathcal{K}^{ab;cd}_{\mu \nu ;\mu '\nu '}A_{\mu '\nu '}^{cd}(q^+,q^-)+\mathcal{K}^{ab;cd}_{\mu \nu ;}B^{cd}(q^+,q^-)\right) .
\end{align}
\end{subequations}
The homogeneous BS equation for the gluon and ghost amplitudes is obtained
\begin{equation}
\begin{pmatrix}
A^{ab}_1(p^+,p^-) \\ A^{ab}_2(p^+,p^-) \\ B^{ab}(p^+,p^-)
\end{pmatrix}
=i\int \frac{d^Dq}{(2\pi )^D}
\begin{pmatrix}
a_{11} & a_{12} & a_{13} \\
a_{21} & a_{22} & a_{23} \\
a_{31} & a_{32} & a_{33} \\
\end{pmatrix}^{ab;cd}
\begin{pmatrix}
A_1^{cd}(q^+,q^-) \\ A_2^{cd}(q^+,q^-) \\ B^{cd}(q^+,q^-)
\end{pmatrix} ,
\label{BSE}
\end{equation}
where the matrix element is given as
\begin{subequations}
\begin{align}
P^{A_1}_{\nu \mu}\left( \mathcal{K}^{ab;cd}_{\mu \nu ;\mu '\nu '}A^{cd}_{\mu '\nu '}+\mathcal{K}^{ab;cd}_{\mu \nu ;}B^{cd}\right) &=a_{11}A_1+a_{12}A_2+a_{13}B ,
\\
P^{A_2}_{\nu \mu}\left( \mathcal{K}^{ab;cd}_{\mu \nu ;\mu '\nu '}A^{cd}_{\mu '\nu '}+\mathcal{K}^{ab;cd}_{\mu \nu ;}B^{cd}\right) &=a_{21}A_1+a_{22}A_2+a_{23}B ,
\\
\left( \mathcal{K}^{ab;cd}_{\ \ ;\mu '\nu '}A^{cd}_{\mu '\nu '}+\mathcal{K}^{ab;cd}B^{cd}\right) &=a_{31}A_1+a_{32}A_2+a_{33}B .
\end{align}
\end{subequations}
For the explicit forms of the kernels, see Appendix~\ref{section:kernel-general}.
\section{The coupled BS equation in the massless case}
In this section, we study the massless case $m=0$ to reproduce the Fukuda's results by correcting some errors \cite{Fukuda78} in our framework, before studying the massive case $m \ne 0$ in the subsequent sections.
We obtain the numerical solutions of the amputated gluon and ghost BS amplitudes for the coupled homogeneous BS equations in the massless case $m=0$.
To write down the manageable integral equations,
we perform the Wick rotation to the relative momenta $p$ and $q$ to obtain the Euclidean momenta $p_{E}$ and $q_{E}$ \cite{Wick54}.
For the total momentum $P:= (E, \bm{P})$, however, we restrict our analysis to the vanishing total momentum $P^\mu =0$.
Although this is unrealistic for the bound state,
this investigation is a preliminary step for studying the true bound sates dictated by $P^\mu \ne 0$ in the subsequent sections.
Even the special result obtained at $P^\mu =0$ is expected to give some helpful information on the mass spectrum $P^2$, provided that the Yang-Mills theory has some similarity to the Wick-Cutkosky model \cite{Cutkosky54,Silagadze98}.
In what follows, we adopt the Landau gauge in the Lorenz gauge fixing.
We replace the full propagators and the full vertices by the bare propagators and bare vertices in the BS equation.
Then the homogeneous BS equation reduces to the integral equation of the Fredholm type \cite{J.Kondo59}.
First of all, we assume that the gauge coupling constant $g$ does not run, i.e., is standing.
After the angular integration in the integration measure,
the homogeneous BS equation for the $SU(N)$ Yang-Mills theory in the Landau gauge at $P^\mu =0$ is cast into the coupled integral equation of the form:
\begin{equation}
\begin{pmatrix}
A_1(p_{E}^2) \\ B(p_{E}^2)
\end{pmatrix}
=C \int_{\Lambda _I^2}^{\Lambda _U^2} \frac{dq_{E}^2}{q_{E}^2}
\begin{pmatrix}
a_{11}(p_{E}^2,q_{E}^2) & a_{12}(p_{E}^2,q_{E}^2) \\
a_{21}(p_{E}^2,q_{E}^2) & a_{22}(p_{E}^2,q_{E}^2) \\
\end{pmatrix}
\begin{pmatrix}
A_1(q_{E}^2) \\ B(q_{E}^2)
\end{pmatrix} ,
\quad C=\frac{Ng^2}{16\pi ^2} ,
\label{BS-coupled1}
\end{equation}
with the kernel written in the matrix with the elements:
\begin{subequations}
\begin{align}
{\text{Gluon--Gluon:}}~a_{11}&=\left\{ \frac{3}{2 } + \frac{22}{3}\frac{q_{E}^2}{p_{E}^2} \right\} \theta (p_{E}^2-q_{E}^2)+\left\{ {2} + 8\frac{ p_{E}^2}{q_{E}^2} -\frac{7}{6}\frac{p_{E}^4}{q_{E}^4} \right\} \theta (q_{E}^2-p_{E}^2) >0 ,
\\
{\text{Gluon--Ghost:}}~a_{12}&=\left\{ - \frac{1}{3}\frac{q_{E}^2}{p_{E}^2}\right\} \theta (p_{E}^2-q_{E}^2)+\left\{ - \frac{ 2}{3} + \frac{1}{3}\frac{p_{E}^2}{ q_{E}^2}\right\} \theta (q_{E}^2-p_{E}^2) <0 ,
\\
{\text{Ghost--Gluon:}}~a_{21}&=\left\{ - \frac{3}{4} + \frac{1}{4} \frac{q_{E}^2}{p_{E}^2} \right\} \theta (p_{E}^2-q_{E}^2)+ \left\{
- \frac{3}{4} \frac{p_{E}^2}{q_{E}^2} + \frac{1}{4} \frac{p_{E}^4}{q_{E}^4} \right\} \theta (q_{E}^2-p_{E}^2) <0 ,
\\
{\text{Ghost--Ghost:}}~a_{22}&=\left\{ - \frac{3}{4} \frac{q_{E}^2}{ p_{E}^2}\right\} \theta (p_{E}^2-q_{E}^2)+\left\{ -\frac{3}{4} \frac{ p_{E}^2}{ q_{E}^2}\right\} \theta (q_{E}^2-p_{E}^2) <0 .
\end{align}
\label{BS-coupled1-kernel}
\end{subequations}
\begin{figure}[t]
\begin{minipage}{0.48 \hsize}
\centering
\includegraphics[scale=0.60]{fig-ep218/a11_a12_P=0_}
\end{minipage}
\begin{minipage}{0.48 \hsize}
\centering
\includegraphics[scale=0.60]{fig-ep218/a21_a22_P=0_}
\end{minipage}
\caption{
The matrix elements $a_{11},a_{12}$ and $a_{21},a_{22}$ of the kernel as functions of $x :=\frac{p^2}{q^2}$.
}
\label{fig:kernel-elements}
\end{figure}
Notice that the integration measure is $dq^2/q^2$, modified from the naive $dq^2$, which affects specifying the kernel. It turns out that this choice is efficient for the bound state as shown in the next section.
Fig.~\ref{fig:kernel-elements} is the plot of the matrix elements $a_{jk}$ of the kernel as functions of $x :=\frac{p_{E}^2}{q_{E}^2}$.
Only $a_{11}$ is positive, while the other elements are all negative. For the magnitude, the gluon-gluon component $a_{11}$ is large compared to the other elements.
This fact suggest that it is enough to include only the contribution of $a_{11}$ to obtain the qualitatively correct result for the bound state.
Therefore, it is expected that the positively large element $a_{11}$ dominantly produces the attractive force between two gluons to form the 2-body bound state.
This observation agrees with the result of Fukuda \cite{Fukuda78}, but the kernel elements are slightly different from those given in \cite{Fukuda78}.
See Appendix~\ref{section:kernel-Landau_P=0} for more details.
First, we take into account only the element $a_{11}$ by putting $a_{12} =0$
to solve the integral equation for the gluon amplitude $A_1$ alone.
The result for the gluon amplitude $A_1$ is given in the left panel of Fig.~\ref{fig:A1-B-1}.
Then the ghost amplitude $B$ is obtained by taking into account $a_{21}$ by putting $a_{22} =0$ from the solution for $A_1$.
The result for the ghost amplitude $B$ obtained in this way is given in the right panel of Fig.~\ref{fig:A1-B-1}.
The BS equation is regarded as the eigenvalue equation and a BS amplitude as a solution of the BS equation is identified with an eigenfunction associated to each eigenvalue of $1/C$.
To perform the numerical calculations, we need to divide the integration region $[\Lambda_I^2,\Lambda_U^2]$ into a sufficient number of small steps. Then the integral equation has the same number of eigenvalues (and the associated eigenfunctions) as the number of partitions.
Here we have adopted the infrared and ultraviolet cutoff to be $\Lambda_I^2=1$ and $\Lambda_U^2=101$ respectively, and tried to divide the interval $[\Lambda_I^2,\Lambda_U^2]$ to 10, 100, and 200 pieces to see whether the result converges or not as the number of partition increases.
The results are shown in Table~\ref{Table:eigen-1}.
Fig.~\ref{fig:A1-B-1} is the plot of the eigenfunction corresponding to an eigenvalue.
Here we present only the graphs obtained from the result of 200 partitions, which gives the convergent result.
For this choice of partitions, indeed, we observe that the largest eigenvalue of $1/C$ and the associated eigenfunction is convergent as confirmed according to Table~\ref{Table:eigen-1}, which are denoted by the red curves in Fig.~\ref{fig:A1-B-1}.
The largest eigenvalue of $1/C$ corresponds to the nodeless eigenfunction, and the node of the eigenfunction monotonically increases as the eigenvalue decreases.
Here we define the coupling constant for $SU(2)$ and $SU(3)$ as follows.
\begin{equation}
C=\frac{N g^2}{16\pi ^2}=\frac{N}{4\pi}\times \frac{g^2}{4\pi}=\frac{N}{4\pi} \alpha ,
\quad \alpha := \frac{g^2}{4\pi} \to \alpha_{\text{SU(2)}}=\frac{4\pi C}{2}=\frac{2\pi}{\frac{1}{C}} ,
\quad \alpha_{\text{SU(3)}}=\frac{4\pi C}{3}=\frac{4\pi}{3\frac{1}{C}} .
\end{equation}
\begin{table}[h]
\centering
\caption{
A set of eigenvalues for $\frac{1}{C}$ in the descending order obtained for various partitions.
}
\begin{tabular}{|c|c|c|c|c|c|c|c|} \hline
number of partitions & \multicolumn{7}{c|}{$\frac{1}{C}$} \\ \hline
10 & 46.14 & 13.17 & 4.756 & 2.063 & 0.9797 & 0.5191 & ... \\ \hline
100 & 19.86 & 7.180 & 4.301 & 2.575 & 1.802 & 1.249 & ... \\ \hline
200 & 19.76 & 7.119 & 4.156 & 2.363 & 1.544 & 1.050 & ... \\ \hline
\end{tabular}
\label{Table:eigen-1}
\end{table}
\begin{figure}[h]
\includegraphics[scale=0.55]{fig-ep218/BSE_Landau_Gauge_P=0,A2=0__gluon_Div200}
\quad\quad
\includegraphics[scale=0.55]{fig-ep218/BSE_Landau_Gauge_P=0,A2=0__ghost_Div200}
\caption{
(Left) Gluon amplitude $A_1$,
(Right) Ghost amplitude $B$
obtained by solving the integral equation when $a_{11}\not =0, a_{21}\not =0, a_{12}=a_{22}=0$.
Here we adopt $\Lambda_I^2=1$, $\Lambda_U^2=101$ and 200 partitions for obtaining numerical solutions.
The red lines correspond to the largest eigenvalue of $1/C$.
}
\label{fig:A1-B-1}
\end{figure}
Second, we take into account all the matrix elements.
The numerical solutions are given in Fig.~\ref{fig:A1-B-2}.
See also Table~\ref{Table:eigen-2}.
Fig.~\ref{fig:A1-B-2} shows no sizable difference from Fig.~\ref{fig:A1-B-1}. This confirms the above observation is correct also from the quantitative point of view:
It is enough to include the gluon-gluon contribution $a_{11}$ to study the gluon BS amplitude $A_1$.
\begin{table}[t]
\centering
\caption{
A set of eigenvalues for $\frac{1}{C}$ in the descending order obtained for various partitions.
}
\begin{tabular}{|c|c|c|c|c|c|c|c|} \hline
number of partitions & \multicolumn{7}{c|}{$\frac{1}{C}$} \\ \hline
10 & 46.38 & 13.20 & 4.770 & $-$3.864 & 2.063 & $-$1.234 & ... \\ \hline
100 & 20.04 & 7.255 & 4.311 & 2.588 & 1.803 & $-$1.456 & ... \\ \hline
200 & 19.94 & 7.191 & 4.165 & 2.374 & 1.546 & $-$1.450 & ... \\ \hline
\end{tabular}
\label{Table:eigen-2}
\end{table}
\begin{figure}[t]
\includegraphics[scale=0.55]{fig-ep218/BSE_Landau_Gauge_P=0,A2=0__mix_gluon_Div200}
\quad\quad
\includegraphics[scale=0.55]{fig-ep218/BSE_Landau_Gauge_P=0,A2=0__mix_ghost_Div200}
\caption{
(Left) Gluon amplitude $A_1$,
(Right) Ghost amplitude $B$
obtained by solving the integral equation when all $a_{jk}\not =0$.
Here we adopt $\Lambda_I^2=1$, $\Lambda_U^2=101$ and 200 partitions for obtaining numerical solutions.
The red lines correspond to the largest eigenvalue of $1/C$.
}
\label{fig:A1-B-2}
\end{figure}
The solution can also be analyzed by converting the integral equation to the equivalent differential equation, which is simplified by taking into account the dominant element $a_{11}$ alone.
The resulting equation becomes the fourth-order differential equation with the boundary conditions imposed at the infrared lower bound and the ultraviolet upper bound.
In fact, the solution is analyzed in \cite{Fukamachi15}, but the details are omitted here.
\section{Gluon-gluon BS equation in the massive case}
In what follows, we take into account the gluon-gluon contribution alone in the BS equation.
Then the homogeneous BS equation for the {amputated BS amplitude} $D_{4}^{(0)}{}^{-1}\bar\chi$ reads
\begin{equation}
\Gamma ^{(0) ac}_{\mu \rho}(p^+)\Gamma ^{(0) bd}_{\nu \sigma}(p^-) \bar{\chi}_{cd}^{\rho \sigma}(p;P)
= -\int \frac{d^4q}{(2\pi )^4} \left[ 2K_3\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P)-i K_4\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P) \right] \bar{\chi} _{cd}^{\rho \sigma}(q;P) ,
\label{BSE2}
\end{equation}
where
$\Gamma ^{(0)ab}_{\mu \nu}$ is the inverse propagator for the massive gluon defined by
\begin{align}
\Gamma ^{(0)ab}_{\mu \nu}(p)&= \delta ^{ab} \left[ -\left( p^2-m^2\right) g_{\mu \nu}+\left( 1-\frac{1}{\alpha}\right) p_\mu p_\nu \right] ,
\end{align}
$K_3$ and $K_4$ are the kernels defined by
\begin{align}
K_3\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P) &:=\Gamma ^{(0)ace}_{\mu \rho \alpha}(-p^+,q^+,k^-)D^{(0)ee^\prime}_{\alpha \alpha^\prime}(k^-) \Gamma ^{(0)bde^\prime}_{\nu \sigma \alpha ^\prime}(p^-,-q^-,-k^-) ,
\nonumber\\
K_4\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P) &:=\Gamma ^{(0)abdc}_{\mu \nu \sigma \rho}(-p^+,p^-,-q^-,q^+) ,
\qua
k^- :=p-q .
\end{align}
Here $\Gamma ^{(0)ace}_{\mu \rho \alpha}$ is the three-gluon vertex and $\Gamma ^{(0)abdc}_{\mu \nu \sigma \rho}$ is the four-gluon vertex,
while $D^{(0)ab}_{\mu \nu}$ is the ordinary massless gluon propagator defined by
\begin{align}
D^{(0)ab}_{\mu \nu}(p)&= \delta ^{ab} \frac{-i }{p^2}\left[ g_{\mu \nu}-(1-\alpha )\frac{p_\mu p_\nu}{p^2} \right] .
\end{align}
It should be remarked that we adopt the massless gluon (with the propagator $D^{(0)ab}_{\mu \nu}(p)$) to mediate the strong interactions, while the constituent gluons are massive with the mass $m$.
This approximation will be allowed to study the large and intermediate momentum region except the very small momentum region where the gluon propagator strongly depends on the approximation adopted.
It is possible to replace the massless gluon mediating the strong interactions by the massive gluon with the same mass $m$ as the constituent gluon, which however leads to the very complicated expression for the BS kernel.
This issue will be tackled in a subsequent paper \cite{FKNS}.
In what follows, we restrict our consideration to the BS amplitude $\bar{\chi}^{ab}_{\mu \nu}(p;P)$ of the transverse type:
\begin{equation}
\bar{\chi}^{ab}_{\mu \nu}(p;P)= \delta ^{ab} \left( g_{\mu \nu}-\frac{p^-_\mu p^+_\nu}{p^+\cdot p^-}\right) f(p;P) .
\end{equation}
Indeed, this amplitude is transverse in the sense that
\begin{equation}
p_+^\mu \bar{\chi}^{ab}_{\mu \nu}(p;P)= 0
= \bar{\chi}^{ab}_{\mu \nu}(p;P) p_-^\nu .
\end{equation}
Therefore, the left-hand side of (\ref{BSE2}) does not depend on the parameter $\alpha$ and reads
\begin{equation}
\Gamma ^{(0) ac}_{\mu \rho}(p^+)\Gamma ^{(0) bd}_{\nu \sigma}(p^-) \bar{\chi}_{cd}^{\rho \sigma}(p;P)
= \delta ^{ab} \frac{({p^-}^2-m^2)({p^+}^2-m^2)\left( p^-\cdot p^+ g_{\mu \nu}-p^-_\mu p^+_\nu \right)}{p^-\cdot p^+} f(p;P) .
\label{left}
\end{equation}
By taking the contraction on the Lorentz indices, the left-hand side reads
\begin{align}
g^{\mu \nu} \Gamma ^{(0) ac}_{\mu \rho}(p^+)\Gamma ^{(0) bd}_{\nu \sigma}(p^-) \bar{\chi}_{cd}^{\rho \sigma}(p;P)
=3 \delta ^{ab} ({p^-}^2-m^2)({p^+}^2-m^2)f(p;P) .
\label{BS-left}
\end{align}
By adopting the Landau gauge $\alpha=0$, the integrand of the right-hand side of (\ref{BSE2}) reads
\begin{subequations}
\begin{align}
&g^{\mu \nu} K_3\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P) \bar{\chi}_{cd}^{\rho \sigma}(q;P)
\notag \\
&= - \delta ^{ab} \frac{i g^2 f(q;P)}{{k^-}^4 {q^-}\cdot {q^+}}
\Big\{ 5 {k^-}^4 {q^-}\cdot {q^+}-{k^-}^2 \Big[ {k^-}\cdot {q^+} \left({p^-}\cdot {q^-}+2{p^+}\cdot {q^-}+{q^-}^2+2 {q^-}\cdot {q^+}\right)
\notag \\
&\hspace{3cm} +{k^-}\cdot {q^-} \left( 2 {k^-}\cdot {q^+}+2{p^-}\cdot {q^+}+{p^+}\cdot {q^+}+2 {q^-}\cdot{q^+}+{q^+}^2\right)
\notag \\
&\hspace{3cm} -3 {k^-}\cdot {p^-} {q^-}\cdot{q^+}-3 {k^-}\cdot {p^+} {q^-}\cdot {q^+}+{p^-}\cdot{q^+} {p^+}\cdot {q^-}-4 {p^-}\cdot {p^+} {q^-}\cdot{q^+} \notag \\
&\hspace{3cm} -{q^+}^2 {p^-}\cdot {q^-}-2 {p^-}\cdot {q^+}{q^-}\cdot {q^+}-{q^-}^2 {p^+}\cdot {q^+}-2{p^+}\cdot {q^-} {q^-}\cdot {q^+}-2 {q^-}^2{q^+}^2
\notag \\
&\hspace{3cm} -3 ({q^-}\cdot {q^+})^2\Big]
\notag \\
&\hspace{3cm} -{k^-}\cdot {p^+}\left[{k^-}\cdot {p^-} {q^-}\cdot {q^+}+{k^-}\cdot{q^-} ({p^-}\cdot {q^+}+2 {q^-}\cdot {q^+})+{q^-}^2{k^-}\cdot {q^+}\right]
\notag \\
&\hspace{3cm} -{k^-}\cdot {p^-}\left[{k^-}\cdot {q^+} ({p^+}\cdot {q^-}+2 {q^-}\cdot{q^+})+{q^+}^2 {k^-}\cdot {q^-}\right]-{q^-}^2({k^-}\cdot {q^+})^2
\notag \\
&\hspace{3cm} -{q^+}^2 ({k^-}\cdot{q^-})^2\Big\} ,
\\
& -i g^{\mu \nu} K_4\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P) \bar{\chi}_{cd}^{\rho \sigma}(q;P)
= 18 i g^2 f(q;P) \delta ^{ab} .
\end{align}
\label{BS-right}
\end{subequations}
\section{gluon-gluon BS amplitude}
We take the approximation in which the angular dependence between the total momentum $P$ and the relative momenta $p, q$, namely, the terms $P\cdot p$ and $P\cdot q$ are neglected. Then the left-hand side of (\ref{BS-left}) reduces to
\begin{align}
& g^{\mu \nu} \Gamma ^{(0) ac}_{\mu \rho}(p^+)\Gamma ^{(0) bd}_{\nu \sigma}(p^-) \bar{\chi}_{cd}^{\rho \sigma}(p;P)
\nonumber\\
&=3 \delta ^{ab} ({p^-}^2-m^2)({p^+}^2-m^2)f(p;P) .
\nonumber\\
&= \frac{3}{16}\left[ 8 p^2 \left( P^2-4m^2 \right) +\left( P^2-4m^2\right) ^2+16p^4-16 (p\cdot P)^2\right] f(p;P) \delta ^{ab}
\nonumber\\
&\rightarrow 3\left( p^2+\frac{P^2}{4}-m^2\right)^2 f(p ;P ) \delta ^{ab} .
\end{align}
In the same approximation as that for the left-hand side of (\ref{BS-left}), the right-hand side of (\ref{BS-right}) reduces to
\begin{subequations}
\begin{align}
& g^{\mu \nu}K_3\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;0) \bar{\chi}_{cd}^{\rho \sigma}(q;0)
\nonumber\\
&=
\delta ^{ab} \frac{ig^2 f(q;P)}{8(q^2-\frac{P^2}{4})(p-q)^4}
\Big\{ 28 p^4 \left(P^2-4 q^2\right)+\left(88 p^2+90 P^2-32 q^2\right) (p\cdot q)^2
\notag \\
& +p^2 \left(-7 P^4+72 P^2 q^2-280 q^4\right)
+2 \left[ p^2 \left(152 q^2-52 P^2\right)+7 P^4-62 P^2 q^2+144 q^4 \right] p\cdot q
\notag \\
& -160 (p\cdot q)^3-7 P^4 q^2+38 P^2
q^4-96 q^6 \Big\} ,
\\
& -i g^{\mu \nu} K_4\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P) \bar{\chi}_{cd}^{\rho \sigma}(q;P)
= 18 i g^2 f(q;P) \delta ^{ab} .
\end{align}
\end{subequations}
Here notice that
$g^{\mu \nu} K_4\,^{ab;cd}_{\mu \nu ;\rho \sigma}(p,q;P) \bar\chi^{cd}_{\rho \sigma}(q;P)$ does not depend on $P$ and agrees with the original expression (\ref{BS-right}).
We perform the Wick rotation to the Euclidean region for performing the momentum integration explicitly.
Notice that we apply the Wick rotation to $p$ and $q$ to convert them to the Euclidean momenta $p_{E}$ and $q_{E}$, while we leave the total momentum $P$ the Minkowski momentum.
After performing the integration over the angular variables in the integration measure $d^4q_{E}$ for the Euclidean momentum $q_{E}$, we obtain the integral equation with respect to the magnitude $\sqrt{q_{E}\cdot q_{E}}$ of the Euclidean momentum:
\begin{subequations}
\begin{align}
& 3\left( p_{E}^2-\frac{P^2}{4}+m^2 \right)^2 f(p_{E}^2;P^2) = C\int q_{E}^2 dq_{E}^2 K(p_{E}^2,q_{E}^2;P) f(q_{E}^2;P^2) ,
\nonumber\\
& K(p_{E}^2,q_{E}^2;P^2) = \frac{1}{P^2+4q_{E}^2}\left( \frac{7 P^4}{p_E^2}+\frac{37 P^2 q_E^2}{2 p_E^2}+\frac{88 q_E^4}{p_E^2}+10 P^2+18 q_E^2\right) \theta (p_{E}^2-q_{E}^2)
\notag \\ &
+\frac{1}{P^2+4q_{E}^2} \left( -\frac{14 p_E^4}{q_E^2}+\frac{17 p_E^2 P^2}{2 q_E^2}+96 p_E^2+\frac{7 P^4}{q_E^2}+20 P^2+24 q_E^2 \right) \theta (q_{E}^2-p_{E}^2) ,
\label{BS-kernel-1}
\end{align}
where we have defined
$
C := \frac{Ng^2}{16\pi^2}
$
We can rewrite the kernel into the form which facilitates selecting the dominant terms:
\begin{align}
& K(p_{E}^2,q_{E}^2;P^2) = \frac{q_E^2}{P^2+4q_{E}^2}\left(
18+88\frac{q_E^2}{p_E^2} +10 \frac{P^2}{q_E^2}
+\frac{37}{2}\frac{P^2 }{p_E^2}
+ 7\frac{P^2}{p_E^2}\frac{P^2}{q_E^2}
\right) \theta (p_{E}^2-q_{E}^2)
\notag \\ &
+\frac{q_E^2}{P^2+4q_{E}^2} \left(
24 +96 \frac{p_E^2}{q_E^2} -14\frac{p_E^4}{q_E^4}
+20 \frac{P^2}{q_E^2} +7\frac{P^4}{q_E^4}
+\frac{17}{2}\frac{p_E^2 }{ q_E^2}\frac{P^2}{q_E^2}
\right) \theta (q_{E}^2-p_{E}^2) ,
\label{BS-kernel-1b}
\end{align}
\end{subequations}
\subsection{$P^\mu=0$ case}
First, we consider the limit $P^\mu=0$ of (\ref{BS-kernel-1}).
This is a first step to examine the existence of the solution of the BS equation, although this limit is obviously unphysical for the bound state problem.
This helps us to check whether or not our method works in this problem.
The BS equation (\ref{BS-kernel-1}) reduces
\begin{align}
3 \left( {p_{E}^2}+m^2 \right) ^2 f({p_{E}^2})
=& C\Big[ \int _0^{p_{E}^2} dq_{E}^2 \left( \frac{9q_{E}^2}{2} + \frac{7 q_{E}^4}{4{p_{E}^2}} \right) f(q_{E}^2)
\nonumber\\&
+\int _{{p_{E}^2}}^\infty dq_{E}^2 \left( 6q_{E}^2 +24{p_{E}^2} - \frac{7{p_{E}^4} }{2q_{E}^2} \right) f(q_{E}^2) \Big] .
\label{int_eq_P0}
\end{align}
In order to solve the integral equation, we leave only the term in the integrand which is expected to give the most dominant contribution to the integral, and put the momentum cutoff $\Lambda$ as the upper limit of integration:
\begin{equation}
\left( p_{E}^2 + m^2 \right) ^2 f(p_{E}^2)=C \left[ \int _0^{p_{E}^2} dq_{E}^2 \frac{3q_{E}^2}{2} f(q_{E}^2)+\int _{p_{E}^2}^{\Lambda^2} dq_{E}^2 2q_{E}^2 f(q_{E}^2) \right] ,
\label{2.12}
\end{equation}
By defining the function $A({p_{E}^2})$ by
\begin{equation}
A({p_{E}^2}) := ({p_{E}^2}+m^2)^2 f({p_{E}^2}) ,
\end{equation}
we can rewrite (\ref{2.12}) into
\begin{equation}
A({p_{E}^2})=C\left[ \int _0^{p_{E}^2} dq_{E}^2 \frac{3q_{E}^2}{2(q_{E}^2+m^2)^2} A(q_{E}^2)+\int _{p_{E}^2}^{\Lambda^2} dq_{E}^2 \frac{2q_{E}^2}{(q_{E}^2+m^2)^2} A(q_{E}^2) \right] .
\label{2.15}
\end{equation}
In (\ref{2.15}), we have IR value at ${p_{E}^2} = 0$:
\begin{equation}
A(0)=2C \int _0^{\Lambda^2} dq_{E}^2 \frac{q_{E}^2}{(q_{E}^2+m^2)^2} A(q_{E}^2) ,
\end{equation}
while we have UV value at ${p_{E}^2} = \Lambda^2$:
\begin{equation}
A(\Lambda^2 )=\frac{3C}{2} \int _0^{\Lambda^2} dq_{E}^2 \frac{q_{E}^2}{(q_{E}^2+m^2)^2} A(q_{E}^2) .
\end{equation}
Therefore, we find that the ratio is given by
\begin{equation}
\frac{A({\Lambda^2} )}{A(0)}=\frac{3}{4} .
\label{ir-bc-r}
\end{equation}
By differentiating both sides of (\ref{2.15}) with respect to ${p_{E}^2}$, we obtain the differential equation which is equivalent to the original integral equation (\ref{2.15}):
\begin{equation}
\frac{d A({p_{E}^2})}{d{p_{E}^2}}=-\frac{C}{2}\frac{{p_{E}^2}}{({p_{E}^2}+m^2)^2}A({p_{E}^2}) .
\label{diff1}
\end{equation}
By solving this equation, we obtain the general solution and the BS amplitude as
\begin{align}
A({p_{E}^2})&={\text{Const.}}\times \frac{e^{-\frac{C}{2}\frac{m^2}{{p_{E}^2}+m^2}}}{({p_{E}^2}+m^2)^\frac{C}{2}} ,
\quad
f(p_{E}^2) ={\text{Const.}}\times \frac{e^{-\frac{C}{2}\frac{m^2}{p_{E}^2+m^2}}}{(p_{E}^2+m^2)^{\frac{C}{2}+2}} .
\label{sol-diff1}
\end{align}
For the solution (\ref{sol-diff1}) of the differential equation (\ref{diff1}) to be the solution of the original integral equation, the boundary condition (\ref{ir-bc-r}) must be satisfied:
\begin{align}
\frac{A({\Lambda^2})}{A(0)}&= \left(\frac{m^2}{m^2+{\Lambda^2}} \right)^{\frac{C}{2}} e^{ \frac{C}{2}\frac{\Lambda^2}{\Lambda^2+m^2}}
=\frac{3}{4} ,
\end{align}
which leads to the relation:
\begin{align}
\ln \frac{m^2}{m^2+{\Lambda^2}} + \frac{\Lambda^2}{{\Lambda^2} +m^2} = - \frac{2}{C} \ln \frac{4}{3} .
\end{align}
When $\Lambda$ is quite large compared with $m$, this relation reduces to
\begin{equation}
\ln \frac{m^2}{m^2+{\Lambda^2}}= -\gamma,
\quad
\gamma := 1+\frac{2}{C}\ln \frac{4}{3}.
\end{equation}
Then the mass $m$ obeys the relation:
\begin{gather}
m^2=\frac{{\Lambda^2}}{e^\gamma-1} .
\end{gather}
Provided that $C$ becomes $\Lambda$ dependent and that $C\to 0$ as $\Lambda \to \infty$, $\gamma$ is approximated as $\gamma \sim \frac{2}{C}\ln \frac{4}{3}$ and the gluon mass $m$ obeys the scaling law for large $\Lambda$,
\begin{equation}
m^2 \sim {\Lambda^2} e^{-\frac{2}{C}\ln \frac{4}{3}} .
\end{equation}
The coupling $C$ must go to zero as $\Lambda \to \infty$ so that $m$ converges to a finite and nonzero value. This is consistent with the (ultraviolet) asymptotic freedom.
\subsection{$P^\mu \not =0$ case}
In what follows, we consider the $P^\mu \not =0$ case.
\subsubsection{$P^\mu \ne 0$ case: analytical study}
First, we give an analytical treatment. For this purpose, we incorporate the leading terms and neglect the subleading terms among the $P^2$ dependent and $p^2$ dependent terms in the numerator of the kernel $K$ on the right-hand side of (\ref{BS-kernel-1b}) where the denominator of the kernel $K$ is regarded as $P^2+4q_{E}^2$.
Then the BS equation reads
\begin{align}
3 \left( p_E^2-\frac{P^2}{4}+m^2\right) ^2f(p_E^2;P^2)&=C\Bigg[ \int _0^{p_E^2}q_E^2dq_E^2 \frac{18q_E^2+10P^2}{P^2+4q_E^2} f(q_E^2;P^2) \notag \\
&\hspace{1cm}+\int _{p_E^2}^\infty q_E^2dq_E^2 \frac{24q_E^2+20P^2+\frac{7P^4}{q_E^2}}{P^2+4q_E^2} f(q_E^2;P^2) \Bigg] .
\label{BSE-kernel2}
\end{align}
The neglected terms are incorporated in the numerical treatment afterwards.
In the case of $P^\mu \ne 0$, we introduce the reduced variables $x,y,r$ in units of $P^2$ as
\begin{align}
x := \frac{p_{E}^2}{P^2} \ge 0 , \
y := \frac{q_{E}^2}{P^2} \ge 0 , \
r := \frac{m^2}{P^2} \ge 0 .
\end{align}
For the momentum cutoff $\Lambda$ of the Euclidean momentum $p_{E}$, we define the reduced variable $\lambda$ by
\begin{align}
\lambda := \frac{\Lambda^2}{P^2} > 0 .
\end{align}
In order to consider the two-gluon bound state, we restrict the region of $r$ to
\begin{align}
r> \frac{1}{4} \Longleftrightarrow P^2 < 4m^2=(2m)^2 .
\end{align}
Then (\ref{BSE-kernel2}) is rewritten as
\begin{equation}
\left( 4x+4r-1\right) ^2 f(x)=\frac{16C}{3}\left[ \int _0^x dy \frac{18y^2+10y}{4y+1} f(y) +\int _x^\lambda dy \frac{24y^2+20y+7}{4y+1} f(y) \right] .
\label{BSE-bb1}
\end{equation}
Moreover, we introduce the scaled amplitude $A$ by
\begin{equation}
A(x)\equiv \left( x-\frac{1}{4}+r \right) ^2 f(x) .
\label{fA}
\end{equation}
Then the integral equation (\ref{BSE-bb1}) is rewritten in terms of $A(x)$ as
\begin{equation}
A(x)=\frac{16C}{3}\left[ \int _0^x dy \frac{18y^2+10y}{(4y+1)(4y+4r-1)^2} A(y) +\int _x^\lambda dy \frac{24y^2+20y+7}{(4y+1)(4y+4r-1)^2} A(y) \right] .
\label{integ-diff1}
\end{equation}
In order to obtain the differential equation which is equivalent to the integral equation (\ref{integ-diff1}), we differentiate (\ref{integ-diff1}) with respect to $x$.
Then the equivalent differential equation is given by
\begin{equation}
\frac{dA(x)}{dx} =-\frac{16C(6x^2+10x+7)}{3(4x+1)(4x+4r-1)^2}A(x) .
\label{de-A-P1}
\end{equation}
The differential equation is solved:
\begin{align}
A(x)&=\frac{{\text{Const.}}}{(4x+1)^{\frac{13C}{8(2r-1)^2}} (4x+4r-1)^{\frac{C(16r^2-16r-9)}{8(2r-1)^2}}} \exp \left[ -\frac{C(48r^2-104r+79)}{12(2r-1)(4x+4r-1)}\right] .
\end{align}
Thus, we obtain the BS amplitude:
\begin{align}
f(x)&=\frac{{\text{Const.}}}{(4x+1)^{\frac{13C}{8(2r-1)^2}} (4x+4r-1)^{\frac{C(16r^2-16r-9)}{8(2r-1)^2}+2}} \exp \left[ -\frac{C(48r^2-104r+79)}{12(2r-1)(4x+4r-1)}\right] .
\label{BS-amplitude-ana}
\end{align}
For a solution of the differential equation to become the solution of the original integral equation, it must satisfy the boundary conditions.
The boundary conditions are obtained:
IRBC by putting $x \to 0$ in (\ref{integ-diff1}):
\begin{equation}
A(0)=\frac{16C}{3}\int _{0}^{\lambda} dy \frac{24y^2+20y+7}{(4y+1)(4y+4r-1)^2} A(y) ,
\label{IRbc}
\end{equation}
UVBC by putting $x \to \lambda$ in (\ref{integ-diff1}):
\begin{equation}
A(\lambda )=\frac{16C}{3}\int _{0}^{\lambda} dy \frac{18y^2+10y}{(4y+1)(4y+4r-1)^2} A(y) .
\label{UVbc}
\end{equation}
By substituting the solution (\ref{BS-amplitude-ana}) into (\ref{IRbc}) or (\ref{UVbc}), we can in principle obtain the relation for $C$, $r$ and $\lambda$, which we call the scaling relation.
However, the resulting expression in this case will be too complicated to extract the physical results we need.
Therefore, we move to the numerical investigations.
\subsubsection{$P^\mu \ne 0$ case: numerical study}
\begin{figure}[t]
\centering
\includegraphics[scale=0.70]{fig-ep218/wave_function_1e4_r15}
\caption{
The gluon-gluon BS amplitudes $A$ as a function of $p^2/P^2$ associated to four different eigenvalues obtained for $\lambda=10000$ and $P^2/m^2 =1$.
}
\label{fig:BS-amplitude-f1}
\end{figure}
From now on, we proceed to perform the numerical study of the BS equation with the original kernel (\ref{BS-kernel-1}).
We encounter new features once we take into account the $p^2$ dependence of the kernel (\ref{BS-kernel-1}).
In the numerical solution of the BS equation the eigenvalue of $C$ is not uniquely determined, even if the values of $\lambda$ and $r$ are fixed, namely, the ultraviolet cutoff $\Lambda$ and the constituent gluon mass $m$ are given in units of $P^2$.
In other words, $P^2$ can have a number of possible values for a given coupling constant $C$, even if the ultraviolet cutoff $\Lambda$ and the constituent gluon mass $m$ are fixed.
These values converge to a set of eigenvalues if the number of partitions for approximating the integral are increased.
Fig.~\ref{fig:BS-amplitude-f1} shows the real part of the BS amplitude as a function of $p^2/P^2$ (momentum in units of $P$) for the cutoff $\lambda=10^4$ and a small value $r=1$.
The plots are given for the gluon-gluon BS amplitudes $A$ associated to four different eigenvalues.
The resulting BS amplitudes associated to the different eigenvalues have the different number of nodes.
In general, the eigenvalue and the associated BS amplitude can be complex valued.
We have confirmed that the numerical solutions for the BS amplitude are real valued, when the largest eigenvalue is real valued. In fact, we have checked that the imaginary parts vanish for the numerical solutions.
The numerical solution of the nodeless BS amplitude with the largest eigenvalue plotted in Fig.~\ref{fig:BS-amplitude-f1} is consistent with the continuum solution (\ref{BS-amplitude-ana}).
Fig.~\ref{fig:C-1/r-2} is the plot of the coupling $C$ as a function of the ratio $1/r:=P^2/m^2$ for various but fixed values of the cutoff $\lambda$, which represents the scaling relation.
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{fig-ep218/C_vs_r}
\vskip -0.3cm
\caption{
The plot of the scaling relation: $C$ vs. $1/r=P^2/m^2$
for various values of the cutoff $\lambda$, $\lambda=100,1000,10000$ (from top to bottom).
This is obtained by solving the BS equation (\ref{BS-kernel-1}) in a numerical way.
}
\label{fig:C-1/r-2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{fig-ep218/C_vs_log_lambda}
\vskip -0.5cm
\caption{
The coupling constant $C$ as a function of the momentum scale $\lambda$ for a fixed value of $1/r=P^2/m^2$, which is required for the scaling relation to hold.
The plots from top to bottom correspond to $1/r=P^2/m^2=0.5,1.0,1.5,2.5,3.0$.
}
\label{fig:running-coupling2}
\end{figure}
Fig.~\ref{fig:running-coupling2} is the plot of the coupling constant $C$ as a function of the momentum scale $\lambda$ for various but fixed values of $1/r:=P^2/m^2$, which is required for the scaling relation represented by Fig.~\ref{fig:C-1/r-2} to hold.
This is the renormalization group flow which gives the same physics, $P/m=\text{const.}$, namely, the same bound state mass for a given gluon mass $m$.
This result will be valid for relatively large $\lambda$ due to the approximations adopted in this paper.
As the cutoff $\lambda$ increases, the coupling $C$ decreases like $1/\ln \lambda$ and finally vanishes as $\lambda \to \infty$:
\begin{align}
C \cong \frac{\rm{Const.}}{\ln \lambda} .
\end{align}
This behavior is consistent with the (ultraviolet) asymptotic freedom expected to hold in the Yang-Mills theory.
\begin{figure}[h]
\centering
\includegraphics[scale=0.75]{fig-ep218/beta_function}
\vskip -0.3cm
\caption{
The plot of the $\beta$ function $\beta(C):=\lambda \frac{dC}{d\lambda}$ obtained from the numerical solutions of the BS equation (\ref{BS-kernel-1}) for various values of $P/m$: $P^2/m^2=0.5,1.0,1.5,2.5,3.0$.
The solid line denotes the $\beta$ function $\beta(C)=-\frac{11}{3}C^2$ obtained from the perturbation theory to one loop.
}
\label{fig:beta_function3}
\end{figure}
Fig.~\ref{fig:beta_function3} shows the plot of the $\beta$ function $\beta(C):=\lambda \frac{dC}{d\lambda}$ calculated from the running of the coupling $C$ with respect to $\lambda$ (Fig.~\ref{fig:running-coupling2}) for various values of $P/m$.
We find that all the $\beta$ functions are negative and coincide for small $C$ or large $\lambda$ independently of $P/m$.
The resulting $\beta$ function can be compared with the beta function $\beta(C)=-\frac{11}{3}C^2$ obtained by one-loop calculation in perturbation theory, which is indicated by the solid curve in Fig.~\ref{fig:beta_function3}.
We find that all the points are well fitted to the single curve, although they are located slightly above the solid curve.
We observe that the $\beta$ function obtained from the numerical solution of the BS equation has the same sign and order as the one-loop $\beta$ function.
It should be remarked that the exact agreement of our result with the one loop prediction is not anticipated, since we have neglected some terms already in obtaining the kernel (\ref{BS-kernel-1}).
Our results indicate that the bound state solution of the BS equation is consistent with the asymptotic freedom in the Yang-Mills theory. This is one of the main results of this paper.
Finally, we give a comment how the derivative of $C$ with respect to $\lambda$ is calculated numerically. The derivative was obtained by taking the finite difference between the smallest eigenvalues of $C$ at $\lambda$ and $1.1\lambda$ for any fixed value of $r^{-1}=P^2/m^2$.
This approximation for the derivative becomes worse for larger $\lambda$.
In view of this, we have used the data for the cutoff $\lambda=10^5$ which is the largest possible value we can take due to the limitation of the computer memory available to us.
\section{Conclusion and discussion}
We summarize the results obtained in this paper.
First, we have given a systematic derivation of the BS equation for the gluon and ghost BS amplitude by starting from the CJT effective action for the massive Yang-Mills theory. The solutions of the BS equation represent the simultaneous bound states for gluons and ghosts.
Second, we have obtained the numerical solutions of the derived BS equation in the ladder approximation with the standing gauge coupling, and the improved ladder approximation with the running gauge coupling.
As a warming up problem, we restricted to the vanishing total momentum $P^\mu =0$.
The we have found that
(i) the gluon--gluon contribution is dominant in the solution,
(ii) the ghost BS amplitude is much smaller than (at most 10\% of) the gluon BS amplitude,
(iii) the infrared behavior is influenced by the effect of the running coupling constant.
Third, armed with these preliminary investigations, we proceeded to obtain an approximate solution for the gluon-gluon BS amplitude alone, which corresponds to a two-gluon bound state as a candidate for a color singlet scalar glueball.
We have calculated the beta functions through the running of the coupling constant by using the numerical solutions of the BS equation with an approximated kernel for different choices of the parameters, the ultraviolet cutoff and the constituent gluon mass.
We have shown that the numerical results converge to a single function which has the same sign and order of magnitude as the one-loop beta function,
although the exact agreement with the one-loop perturbation theory is not reproduced due to the approximation we adopted in obtaining the kernel of the BS equation.
Thus, we have shown that the solution for the gluon bound state obtained in our approximation for the BS equation is consistent with the (ultraviolet) asymptotic freedom of the Yang-Mills theory.
It is obvious that we have a number of unsolved problems to be tackled:
improvement of the kernel for the BS equation by incorporating more terms, examination of the high-energy behavior for performing the normalization of the BS amplitude, renormalization of the BS equation, improvement of the approximation to keep gauge invariance, e.g., through the Ward-Takahashi relations, confirmation of physical unitarity of the massive Yang-Mills theory, and so on.
We hope to report some of the results in the near future.
\section*{Acknowledgment}
This work is supported by Grants-in-Aid for Scientific Research (C) No.24540252 and (C) No.15K05042 from the Japan Society for the Promotion of Science (JSPS).
|
\section{Detecting the boosted photon jets in the collider}\label{sec:photonjet}
The photon-jet is a special object that consists of a cluster of (nearly) collinear photons which have a signature similar to that of a single photon.
This idea was initially proposed in~\cite{Ellis:2012sd} and has recently been applied to explain this di-photon excess~\cite{Agrawal:2015dbf,Chang:2015sdy,Bi:2015lcf,Aparicio:2016iwr,Ellwanger:2016qax,Dasgupta:2016wxw,Domingo:2016unq,Badziak:2016cfd}. It could be generated from the decay of a highly boosted light particle (sub-GeV).
In this section we discuss two technical issues, angular separation and decay length, which are very crucial in detecting a photon-jet in the collider machine.
Suppose a particle $X(750)$, once produced in the hadron collider, {\it instantly} decays to two highly boosted $Y$s. For each $Y$, it decays into a pair of photons with a branching ratio close to $100\%$. The angular separation of two outgoing photons (which forms a photon-jet) will be $\theta \sim \frac{2 m_{Y}}{p_{Y}} \sim \frac{2 m_{Y}}{375}$, assuming 375 GeV is the expected momentum for $X(750) \to Y Y$ process. If this angular separation is smaller than the resolution of the ECAL in the CMS/ATLAS, the photon jet could probably be mistagged as a single photon. Hence, the naive estimation from the parameters of the ECAL segmentation ($\Delta \phi = 0.0174$ in the CMS~\cite{Khachatryan:2015iwa} and $\Delta \phi = 0.025$ at the ATLAS~\cite{Aad:2009wy,Aad:2010sp}) places a bound on $Y$ mass: $m_{Y}\lesssim (0.0174\times375)/2=3.26$~GeV so that two photons will hit the same ECAL segment and eventually be recorded as a single photon. Of course, the actual photon jet conversion could be much more complicated and gives a stronger bound. However, this is not a big issue as $m_{Y}$ stays well below this bound in the following analysis.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.45\textwidth]{./LossRatio_New.pdf}
\caption{Assuming that the typical $R_{\text{ECAL}}=1.5$ m from the beam, the signal loss ratio in percentage versus the decay length ($L_{\text{decay}}$) of light particle $Y$. When the decay length of $Y$ exceeds 2 m, about half of the $Y$ generated would not be detected.}
\label{decaylength}
\end{center}
\vspace{1mm}
\end{figure}
The other concern about this topology in the language of photon-jet is the proper decay length of the light (pseudo)scalar $Y$. This is because a very light but highly boosted particle could have an unexpectedly long decay length. If its decay takes place after passing through the ECAL detector, it cannot be technically detected.
Quantitatively, the proper average decay length of such a highly boosted particle is given by
\begin{equation}
L_{\text{decay}}=c\tau\gamma \approx \frac{375c}{\Gamma_{Y} m_{Y}}
\end{equation}
in which $\tau=\Gamma^{-1}$ and $\gamma= \frac{E_{Y}}{m_{Y}} \approx \frac{p_{Y}}{m_{Y}}$ are employed. $\Gamma_{Y}$, the total decay width of $Y$, is generally proportional to $m_{Y}$ for the light $Y$ state.
Hence, a lighter $m_{Y}$ leads to a longer $L_{\text{decay}}$. Once the decay length is comparable with the size of detector radius ($R_{\text{ECAL}}$ in this case), the probability of $Y$ escaping the detection soars. The correlation of the proper decay length versus the percentage of signal loss is illustrated in Fig.~\ref{decaylength}.
We can see that for a decay length longer than $\sim 1$~m, less than $80\%$ of the $Y_1$ decay events would happen inside (or before reaching) the ECAL and eventually be captured by the detector. In this case, a larger production rate is needed to compensate the significant event loss. However, such a long $L_{\text{decay}}$ will cause another problem. For a considerable portion of $Y$ which pass through the ECAL, they will decay within the HCAL and their decay products will be tagged as displaced jets.
We expect that future experiments searching for resonances using the construction of displaced jets and/or photons~\cite{Aad:2013txa,Aad:2014yea,Aad:2015asa} could place an upper bound on $L_{\text{decay}}$.
On the other hand, it was argued in~\cite{Khachatryan:2015iwa} that, if $L_{\text{decay}}$ is close to $R_{\text{ECAL}}$, $Y$ would decay in a position near the ECAL layer, thereby, the two photons produced from a not-so-light $Y$ ($m_{Y} \gtrsim$ 3.26~GeV) can still hit the same ECAL segment simultaneously.
Nonetheless, the probability of $Y$ decay takes place drops off exponentially along the distances away from the beam.
This infers a larger probability of $Y$ decay occurring close to the beam rather than near the ECAL layer, giving rise to a pair of photons distinguishable when reaching the ECAL layer.
If this were true, one would have seen a significant amount of 3-photon or 4-photon signals.
In short, $L_{\text{decay}}(Y) \mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 1$~m and $m_{Y}\mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 3$~GeV are viable assumptions in our scenario in order to have sufficient di-photon signal detected in the ECAL.
\vspace{2mm}
\section{Gluon fusion production and di-photon decay}\label{sec:ggfusion}
\subsection{Doublet $X(750)$ and its production at the collider}
Let us begin with a numerical estimate to assess the possibility of realizing this scenario in the context of the 2HDM with the inclusion of a singlet scalar.
For simplicity, we assume that the $X(750)$ resonance behaves like the heavy CP-even Higgs such as the couplings and the production modes. Thus, the $X(750)$ is dominantly produced via gluon fusion or $b-$quark associated production (we focus on $\tan\beta \lesssim 4$ and thus do not consider this production mode in the present paper).
It is shown in~\cite{Bernon:2015qea} that the gluon-fusion production cross section decrease as $\tan\beta$ grows in Type~I, while this cross section minimizes at a modest value of $\tan\beta \sim 8$ in Type~II.
Hence, a maximal value of $\tan\beta$ that could yield the di-photon signal can be estimated by assuming the $X(750)$ decays inclusively into two photons with 100\% branching ratio.
As shown in Fig.~\ref{fig:ggBRplot}, we evaluate the di-photon signal,
$\sigma\left(gg\to X\right) \mbox{BR}\left(X\to \gamma\gamma\right)$, as a function of $\tan\beta$ at different levels of $\text{BR}(X\to 2\gamma + ... )$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.45\textwidth]{./Scalar_BR_Tb_1.pdf}
\hspace{4mm}
\includegraphics[width=0.45\textwidth]{./Scalar_BR_Tb_2.pdf}
\caption{The correlation of the predicted di-photon signal ratio (in-box numbers) to the size of the branching ratio of $X(750) \to 2 \gamma + \text{else}$. Such decay could be either a direct or cascade decay. The gluon fusion production cross section (indicated in the unit of fb on the top axis) is obtained in the assumption that $X(750)$ resonance behaves like a doublet whose Yukawa couplings given in Type~I (left) and Type~II (right) models.}
\label{fig:ggBRplot}
\end{center}
\vspace{1mm}
\end{figure}
As expected, the obtainable di-photon signal diminishes as $\tan\beta$ increases, while maintaining the level of $\text{BR}(X\to 2\gamma + ... )$. (This is also true for $\tan\beta \lesssim 20$ in Type~II.) This is the result of suppressing the $X(750)$ production at high $\tan\beta$ due to the fact the heavy Higgs coupling normalized to the SM value with the top quark, which dominantly mediates the gluon-fusion, is inversely proportional to $\tan\beta$. To offset this drop, the decay branching ratio $\text{BR}(X\to 2\gamma + ... )$ has to be increased to acquire a larger signal rate. For instance, in order to yield a $10$~fb signal, $t_\beta \leq 8(10)$ and $t_\beta \leq 6(7)$ are required in Type~I (Type~II) in the assumption of $\text{BR}(X\to 2\gamma + ... )$ at $100\%$ and more practically at $60\%$, respectively.
In particular, such an upper limit on the $\tan\beta$ would be relaxed in Type~II for a lower cross section of $6$~fb.
In both Type~I and Type~II, a relative large cross session ($\approx 700$ fb in the SM) of producing the $X(750)$ state is achieved when $t_\beta \lesssim 1$ as a result of the enhanced coupling to top-quark. Meanwhile, this will maximize the cross section in the $t\bar{t}$ and di-jet final states. Comparing with the current experiment bounds, $\sigma(pp \to X \to gg) \lesssim 10$~pb and $\sigma(pp \to X \to t\bar{t})\lesssim 700$~fb~\cite{CMS:2015neg,Aad:2015fna}, we
find that our model stays well below the di-jet bound while the di-top bound could be marginal in the case of $t_\beta \geq 1$. Thus, we shall limit $t_\beta \geq 1$ in the following analysis.
\subsection{Singlet $Y$: scalar vs. pseudoscalar}
\label{ssec:singlet}
Concerning the light state $Y$, there are two important requirements that must be satisfied to realize this scenario. These include the presence of decay into di-photon and the substantially large total width. To accomplish the first goal, unlike the situation in the NMSSM, this state cannot be a pseudoscalar in the present study. As explained in the Introduction, the CP-odd scalar from the doublet field cannot be too light, while the pseudoscalar arising from the newly introduced singlet field does not couple to SM particles in the assumption of CP-conservation and thus has no di-photon decay mode. On the other hand, requiring $Y$ a singlet scalar is not sufficient.
Though $Y$ can decay to di-photon via charged Higgs loop, or to $4f$ via two off-shell $H^*, A^*$ or $H^{\pm *}$, these decay channels are highly suppressed by kinematics. Tree-level decays to light quarks are absent due to its decoupling to the SM particles.
The combination of these effects result in a extremely narrow decay width. The proper decay length for a singlet scalar is typically at the order of kilometer. Therefore, $Y$ state must be a scalar and also gains an amount of doublet composition from the mixing. The content to which the mixing is needed will be analyzed in Sec.~\ref{sec:decayY1}. Meanwhile, the singlet pseudoscalar, if present, could serve as a dark matter candidate.
In addition, the mass of $Y$ is very crucial in determining the products of $Y$ decay. When $m_Y$ is $\mathcal{O}(1)$ GeV, its main decay products are jets.
In this case, the decay to di-photon is only mediated via loop diagrams and its branching ratio $\mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 0.1-1\%$ (see an example of $m_{Y}=5$ GeV in Fig.~\ref{fig:doubletfraction}). In contrast, the notorious jet background is well suppressed in the decay for a sub-GeV particle. In particular, considering a $m_Y$ below the $s\bar{s}$ and $\mu^+ \mu^-$ thresholds, it will only decay to $u\bar u,d\bar{d}$ and $e^+e^-$ at tree level. Notice that $m_{Y} < \Lambda_{\text{QCD}}$, so the outgoing quark pairs would not develop into di-jet but hadronize and cascade decay into di-photon. On the other hand, $BR(Y \to e^+e^-)$ is negligible compared to quark pair production due to the small electron Yukawa coupling. Therefore, one can expect the branching ratio of the decay to di-photon for such a light $Y$ is nearly $\sim 100\%$.
Of course, the situation becomes rather subtle when $m_{Y}$ exceeds the $\mu^+ \mu^-$ threshold, $\approx 210$ MeV. Since the muon has a much larger Yukawa coupling than $u$ and $d$ quarks, there will be a considerable portion of $Y$ decaying to muon pairs. This seems to generate a heavy suppression in the branching ratio of the decay to di-photon. However, strong dynamics may have a large correction to the decay width of $Y\to \gamma\gam$ when this decay is mediated via hadronic states.
Above the $\mu^+\mu^-$ threshold, once $m_{Y}$ reaches the $\pi\pi$ threshold, $\approx 280$ MeV,
$Y$ will hadronically decay into either $\pi^+\pi^-$ or $\pi^0 \pi^0$ since $\pi$ is the lightest hadron.
By virtue of isospin conservation, only about $1/3$ of outgoing pion pairs are $\pi^0\pi^0$ pair and cascade decay to di-photon for each $\pi^0$. Whereas, the remaining $2/3$ will be $\pi^+\pi^-$ and eventually decay to $\mu^+\mu^-+4\nu$, which can be detected by the muon chamber, leaving a large missing $p_T$.
To avoid the subtlety, we will consider the case throughout the paper in which $m_{Y}$ is below the di-muon threshold. In the numerical analysis unless specified, $m_{Y}=150~\text{MeV}$ is chosen and the assumption of BR$(Y\to \gamma\gamma)=100\%$ is globally adopted.
\section{Model I: real singlet}
\label{sec:realmodel}
We first consider adding to the 2HDM a real scalar gauge singlet $S$.\footnote{This model has been already constructed in earlier literatures, see an example~\cite{Drozd:2014yla} in which the singlet does not acquire a vev and also is $Z_2$ odd for the sake of having a dark matter candidate.} To eliminate the substantial FCNC, we assume a $\mathbb{Z}_2$ symmetry under which $\Phi_2$ is odd. For the singlet $S$ we impose a $\mathbb{Z'}_2$ symmetry under which $S$ is the only odd field.
The gauge invariant and renormalizable Lagrangian for this model is
\begin{equation}
\mathcal{L}_{\text{2HDMS}} = \mathcal{L}_{\text{2HDM}} + \mathcal{L}_{S}
\end{equation}
where the Lagrangian of CP-conserving but soft $Z_2$ breaking 2HDM can be found in~ \cite{Bernon:2014nxa,Drozd:2014yla,Bernon:2015qea} in which two Higgs-doublet fields are expanded as
\begin{equation}
\Phi_a =
\begin{pmatrix}
H_a^+ \\ \frac{1}{\sqrt{2}}(v_a+\rho_a+ i\eta_a)
\end{pmatrix} \quad (a=1,2), \qquad S=v_s+\chi
\end{equation}
with the ratio of two vevs given by $\tan\beta \equiv v_2/v_1$.
While the $S$-associated part reads
\begin{equation}
\mathcal{L}_{S} = \frac{1}{2} m_s^2 S^2 +\frac{1}{4!}\lambda_s S^4+\kappa_1 \Phi^\dagger_1 \Phi_1 S^2 + \kappa_2 \Phi^\dagger_2 \Phi_2 S^2
\end{equation}
In contrast to~\cite{Drozd:2014yla}, in the present study we allow the $S$ to acquire a vev so that it will mix with the Higgs doublets.
Three mass-squared parameters can be replaced by three vevs through the corresponding minimization conditions such that
\begin{align}
m_{11}^2 &= m^2_{12} \frac{v_2}{v_1}-(\frac{1}{2} \lambda_1 v_1^2 +\frac{1}{2} \lambda_{345}v_2^2 +\kappa_1 v_s^2) \label{eq: minh1}\\
m_{22}^2 &= m^2_{12} \frac{v_1}{v_2}-(\frac{1}{2} \lambda_2 v_2^2 +\frac{1}{2} \lambda_{345}v_1^2 +\kappa_2 v_s^2) \label{eq: minh2}\\
m_s^2 &= -\frac{1}{6} \lambda_s v_s^2 -\kappa_1 v_1^2 -\kappa_2 v_2^2 \label{eq: mins}
\end{align}
\subsection{Mass eigenstates and spectrum}
For the CP-even neutral states $(\rho_1, \rho_2, \chi)$ in the $Z_2$ basis, the mass matrix can be written as
\begin{equation}
\mathcal{M}^2 = \begin{pmatrix}
m^2_{12} \frac{v_2}{v_1}+\lambda_1 v_1^2 & -m^2_{12} +\lambda_{345}v_1 v_2 & 2 \kappa_1 v_1 v_s\\
-m^2_{12} +\lambda_{345}v_1 v_2 & m^2_{12} \frac{v_1}{v_2} +\lambda_2 v_2^2 &2 \kappa_2 v_2 v_s \\
2 \kappa_1 v_1 v_s & 2 \kappa_2 v_2 v_s & m^2_\chi
\end{pmatrix}\label{m square of real singlet}
\end{equation}
where $m^2_\chi=\frac{1}{3} \lambda_s v_s^2$ given Eq.~(\ref{eq: mins}) has been used. In the presence of non-negligible off-diagonal elements $(v_s \neq 0)$ in the above mass matrix, $(\rho_1, \rho_2, \chi)$ are apparently not the mass eigenstates.
We first rotate two doublet components (upper $2 \times 2$ block) into the basis $(\hat{h}, H)$ via an angle $\alpha$.
\begin{equation}
\label{2hdmHiggsstate}
\begin{pmatrix}
\hat{h} \\ H
\end{pmatrix}
=
\begin{pmatrix}
\begin{array}{ccc}
-s_\alpha & c_\alpha \\
c_\alpha & s_\alpha \\
\end{array}
\end{pmatrix}
\begin{pmatrix}
\rho_1 \\ \rho_2
\end{pmatrix}
\end{equation}
In fact, they are mass eigenstates in the pure 2HDM but no longer true in the model we consider here due to the doublet-singlet mixture induced by $v_s \neq 0$.
This can be seen explicitly from the full $3 \times 3$ mass matrix under the unitary rotation.
\begin{equation}\label{eq:massmatrix}
\hat{\mathcal{M}}^2=
\begin{pmatrix}
-s_\alpha & c_\alpha & 0 \\
c_\alpha & s_\alpha &0 \\
0 & 0 & 1
\end{pmatrix}
\mathcal{M}^2
\begin{pmatrix}
-s_\alpha & c_\alpha & 0 \\
c_\alpha & s_\alpha &0 \\
0 & 0 & 1
\end{pmatrix} \\
= \begin{pmatrix}
m^2_{\hat{h}} & 0 & \Delta\\
0 & m_H^2 & D\\
\Delta & D & m^2_\chi
\end{pmatrix}
\end{equation}
where
\begin{align}
D &= 2v v_s (\kappa_1 c_\alpha c_\beta + \kappa_2 s_\alpha s_\beta) \label{eq:Ddef}\\
\Delta &=2 v v_s (-\kappa_1 s_\alpha c_\beta + \kappa_2 c_\alpha s_\beta) \label{eq:deltadef}
\end{align}
Clearly, both off-diagonal elements, $\Delta$ and $D$, are not vanishing due to the presence of non-zero $v_s$.
To fit the LHC Higgs data, we expect the SM Higgs $h$ with $m_h=125$~GeV to be nearly pure doublet. This demands the mixing parameter $\Delta$ very small. For simplicity we will take $\Delta=0$ in the following discussion, which then gives us
\begin{equation}
\label{smallmixing}
\kappa_1= \kappa_2t_\beta/t_\alpha
\end{equation}
Whereas, we allow an arbitrary mixing between $H$ and $\chi$ for our purpose.
Applying diagonalization between them, we find three resulting mass eigenstates which are formed by:
\begin{equation}
\label{eq:eigenstatecomp}
\begin{pmatrix}
h \\ Y_2\\ Y_1
\end{pmatrix}
=
\begin{pmatrix}
\begin{array}{ccc}
1 & 0 & 0 \\
0 & c_\theta & s_\theta \\
0 & -s_\theta & c_\theta \\
\end{array}
\end{pmatrix}
\begin{pmatrix}
\hat{h} \\ H \\ \chi
\end{pmatrix}
\end{equation}
where the states $(\hat{h}, H)$ defined in Eq.~(\ref{2hdmHiggsstate}) are expressed in the $Z_2$ basis.
The mixing angle between $H$ and $\chi$ is given by
\begin{equation}
\label{mixingangle}
s_{2\theta}=\frac{2D}{\sqrt{4D^2 + \left( m^2_\chi -m_H^2\right)^2}}, \quad c_{2\theta} =\frac{ m^2_\chi-m_H^2}{\sqrt{4D^2 + \left( m^2_\chi -m_H^2\right)^2}}
\end{equation}
Alternatively, one can parameterize the mixing angle Eq.~(\ref{mixingangle})
as,
\begin{equation}\label{eq:mixinganglev2}
c_\theta=\sqrt{\frac{m_{Y_2}^2-m^2_\chi}{m_{Y_2}^2-m_{Y_1}^2}},\qquad s_\theta=\mbox{Sign}(\kappa_1 v c_\beta c_\alpha+\kappa_2 v s_\beta s_\alpha)\sqrt{\frac{m^2_\chi -m_{Y_1}^2}{m_{Y_2}^2-m_{Y_1}^2}}
\end{equation}
As discussed in the Introduction, to realize the idea of photon jet we consider the scenario where $Y_1$ is a sub-GeV singlet-like state and $Y_2$ a heavy doublet-like resonance at 750 GeV.
To this end, we examine the lower $2 \times 2$ block in the mass matrix, Eq.~(\ref{eq:massmatrix}). Since both $m_\chi$ and $D$ are proportional to the singlet VEV $v_s$, the mixing $s_\theta$ is small for most of the parameter space if ${Y_1}$ is required to be very light. (The exception occurs in the region where the potential stability is violated.) This suppression in $s_\theta$ is a fatal weakness for the phenomenology of the di-photon excess in this model as will be shown in Sec.~\ref{subsec:realdiphoton}.
Finally, we present the masses for three scalar mass eigenstates
\begin{eqnarray}
m_{h}^2 & = & m_{\hat{h}}^2 \label{eq:hmass
\\
m_{Y_{1,2}}^2 &=& \frac{1}{2}\left[m_H^2+m^2_\chi \pm \sqrt{4D^2 + \left( m^2_\chi -m_H^2\right)^2} \right] \label{mYeq}
\end{eqnarray}
where the expression of $m_{\hat{h}}^2$ and $m_H^2$ for two CP-even Higgs states in the 2HDM can be found in~\cite{Bernon:2015qea}. $m_{Y_2}$ and $m_{Y_1}$ in Eq.~(\ref{mYeq}) take $+$ and $-$ signs, respectively.
\subsection{Higgs couplings and decays}
Though the $\chi$ component does not directly couple to any SM fermions, both $Y_1$ and $Y_2$ that arise from the $H-\chi$ mixing couple to SM particles as well as the charged Higgs. Some relevant couplings for the three CP-even mass eigenstates, normalized to SM values, are listed in Table~\ref{table:scalarcoupling}, where the couplings of $h, H$ to the charge Higgs are defined in the 2HDM context as
\begin{eqnarray}
g_{hH^\pm H^\mp} &=& -\frac{1}{v} \left\lbrace [m_h^2 + 2(m_{H^\pm}^2-\bar{m}^2)]s_{\beta-\alpha} + 2 \cot2\beta(m_h^2-\bar{m}^2)c_{\beta-\alpha} \right\rbrace \\
g_{HH^\pm H^\mp} &=& -\frac{1}{v} \left\lbrace [m_H^2 + 2(m_{H^\pm}^2-\bar{m}^2)]c_{\beta-\alpha} - 2 \cot2\beta(m_H^2-\bar{m}^2)s_{\beta-\alpha} \right\rbrace \\
g_{\chi H^\pm H^\mp} &=& 2 v_s (\kappa_1 s_\beta^2 + \kappa_2 c_\beta^2) \label{eq:chiHH}
\end{eqnarray}
with $\bar{m}^2 = \frac{2m_{12}^2}{s_{2\beta}}$.
\begin{table}[t]
\caption{The couplings of scalars normalized to the SM values except for those to the charged Higgs.
}
\vspace*{-5mm}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Scalar states& SM gauge bosons & SM fermions & charged Higgs \cr
\hline
$h$ & $s_{\beta-\alpha}$ & $C^h_f$ & $g_{hH^\pm H^\mp}$ \cr
$Y_1$ & $-c_{\beta-\alpha} s_\theta$ & $-C^H_f s_\theta$ & $-s_\theta g_{HH^\pm H^\mp} + c_\theta g_{\chi H^\pm H^\mp} $ \cr
$Y_2$ & $c_{\beta-\alpha} c_\theta$ & $C^H_f c_\theta$ & $c_\theta g_{HH^\pm H^\mp} + s_\theta g_{\chi H^\pm H^\mp} $ \cr
\hline
\end{tabular}
\end{center}
\label{table:scalarcoupling}
\end{table}%
It is noticeable that the coupling of the Higgs $h$ to gauge bosons is proportional to $\sin(\beta-\alpha)$ while that for the other two states $Y_1$ and $Y_2$, after mixing, display the $\cos(\beta-\alpha)$ dependence.
This implies that most relations in the 2HDM, in particular the alignment limit, still holds in this extended model. To fit the 125~GeV Higgs data, for simplicity we take the alignment limit ($c_{\beta-\alpha} \to 0$), in which the $h$ has SM-like couplings while the $Y_1$ and $Y_2$ decouple with gauge bosons. This reduces Eq.~(\ref{smallmixing}) to
\begin{equation}\label{eq:k1k2v2}
\kappa_1 \sim - \kappa_2 \tan^2\beta
\end{equation}
Taken this relation, the coupling $hY_1Y_1$ vanishes at small mixing limit. Hence, the decay of SM Higgs $h$ to new light scalars such as $h\to Y_1Y_1\to 4\gamma$, which is already constrained by the Higgs invisible decay search~\cite{ATLAS:2012soa}, is switched off automatically.
As for Yukawa couplings, they display an overall dependence on $c_\theta$ and $s_\theta$ for the $Y_2$ and $Y_1$ states, respectively. This means the mixing between them is crucial in determining their decays to the SM fermions.
Other than the couplings listed in Table~\ref{table:scalarcoupling}, the $Y_2 Y_1 Y_1$ coupling is most relevant to our discussion, which reads in the alignment limit
\begin{align}\label{eq:gy2y1y1}
g_{Y_2 Y_1 Y_1} &=3c_{\theta } s_{\theta }^2 g_{HHH}+2 \kappa _2 v c_{\theta } s_{\theta }^2 s_\beta c_{\beta }\left(t^2_\beta+1\right) -\kappa _2 v s_\beta c_{\beta } c_{\theta }^3 \left(t_{\beta }^2+1\right)\nonumber\\
&+\kappa _2 v_s s_{\theta }^3
\left(c_{\beta }^2-s_\beta^2 t_\beta^2\right)+\frac{1}{2} c_{\theta }^2
s_{\theta } v_s \left(-4 \kappa _2 c_{\beta }^2+\lambda_s +4 \kappa _2 s_\beta^2 t_\beta^2\right)
\end{align}
In addition, $Y_2 \to hh$ should have been another important decay mode considering the fact that in the 2HDM context the heavy Higgs $H$ generally has a sizable decay branching ratio into a pair of SM-like Higgs $hh$. However, we examine that this coupling $g_{Y_2 hh}$ is vanishing at the exact alignment limit~\cite{Bernon:2015qea}. Therefore, we do not consider this decay mode in the present analysis.
It is also necessary to notice that the couplings of the singlet $\chi$ to massless Goldstone bosons are proportional to:
$v_s (\kappa_1 c_\beta^2 + \kappa_2 s_\beta^2)$. As a result, nearly singlet-like $Y_1$ does not couple to the longitudinal modes of $W^\pm/Z$ in the limit $\Delta\to0$.
\subsection{The decay of $Y_1$}
\label{sec:decayY1}
We now turn to study $Y_1$ decay. Recall that $Y_1$ is a mixture of CP-even heavy Higgs $H$ and real singlet $\chi$, so it has Yukawa couplings as a doublet (the fraction of which is described by $s_\theta$) and also couplings with scalars present in the Higgs sector. As a result, it can decay to SM light quarks. The presence of these tree level decays greatly increases the total width and in turns shorten the decay length $L_{\text{decay}}$ within the range of $R_{ECAL}$ scale. The influence of mixing $s_\theta$ on the decay length of $Y_1$ is illustrated in Fig.~\ref{fig:doubletfraction} for two choices of $m_{Y_1}$.
In order to maximally enhance the decay width of $Y_1$, we consider the Yukawa patterns obeying the Type~II model in the following numerical analysis.
First, as expected, BR$(Y_1\to \gamma\gamma)\mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 0.1\%$ for $m_{Y_1}=5$ GeV. This is the result of the presence of tree-level decay to light quarks.
To better understand the $Y_1$ decay, it is useful to analyze the explicit form of $Y_1$ coupling to fermions, $g_{Y_1 ff}$ (c.f.~Table~\ref{table:scalarcoupling}). This coupling is proportional to $s_\theta$, which depends monotonically on $D=-2v v_s \kappa_2(1+ \tan^2\beta)$. Thus, as $s_\theta$, or essentially $\kappa_2$, increases, the total decay width of $Y_1$ grows up. This can be easily understood from the fact that $Y_1$ acquires more doublet component in the large mixing. The remaining factors $C^H_d\sim t_\beta$, $C^H_u\sim t_\beta^{-1}$ are specified at the exact alignment limit in the Type~II model. In this case $Y_1 \to d\bar d$ decay is enhanced for $t_\beta>1$. Of course, $Y_1 \to u\bar u$ is simultaneously suppressed but not so efficiently as an offset since $m_d/m_u \simeq 2$. Therefore, it could be expected that the total decay width increases as $t_\beta$ becomes large, which in turn leads to a shorter decay length. All these behaviors are clearly reflected in Fig.~\ref{fig:doubletfraction}.
\begin{figure} [t]
\centering
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_mixingatms200mev.pdf}
\caption{Decay length of $Y_1$
as a function of the doublet fraction in $Y_1$ for $m_{Y_1}=200$~MeV
by choosing various values of $\tan\beta$. Red, orange and green curves correspond to $\tan\beta=2, 4, 6$, respectively.
}
\label{fig:doubletfraction}
\end{figure}
\subsection{Di-photon cross section}\label{subsec:realdiphoton}
According to the preceding discussion, there are five independent parameters in the numerical analysis, which include $m_H$, $v_s$, $\lambda_s$, $\kappa_2$ and $t_\beta$. However, since $m_{Y_2} = 750 $~GeV$\gg m_{Y_1}$, the mixing effect on the mass of the heavy eigenstate would be very small, allowing us to take $m_H=750$~GeV as a good approximation. While $v_s$ can be determined by using Eq.~(\ref{mYeq}) once the value of $m_{Y_1}$ is chosen. In the end, we are left with three free parameters $\lambda_s$, $\kappa_2$ and $t_\beta$.
To compute the di-photon cross section of our interest, we examine the decay modes of $Y_2$. As discussed in Sec.~\ref{sec:decayY1}, we know that a moderate mixing (small $s_\theta$) is required to render the decay length of $Y_1$ sufficiently short (also see Fig.~\ref{fig:doubletfraction}).
In this case, $|c_\theta|\approx\pm 1$ and the leading contribution to $g_{Y_2Y_1Y_1}$ in Eq.~(\ref{eq:gy2y1y1}) reads
\begin{equation}\label{eq:gy2y1y1realv2}
g_{Y_2Y_1Y_1} = -\text{sign}(s_\theta)vc_\beta s_\beta\kappa_2(1+t_\beta^2)
\end{equation}
It implies that this coupling, or equivalently the decay $Y_2 \to Y_1 Y_1$ tends to be important as either $|\kappa_2|$ or $\tan\beta$ becomes large. However, one cannot achieve both two large because of the constraint from the stability, as you will see shortly.
In the alignment limit, $Y_2$ decays mainly into $Y_1 Y_1$ and $t\bar t$, the latter one is sensitive to $t_\beta$ as in the 2HDM. In particular, when $t_\beta \lesssim 6$, $Y_2 \to t\bar{t}$ is a predominant decay channel.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5\textwidth]{Scalar2HDM_k2tb.pdf}
\caption{The di-photon cross section $\sigma(gg\to Y_2) \text{BR}(Y_2 \to Y_1 Y_1)$ (in the unit of fb) produced by the 750 GeV resonance $Y_2$ under the assumption that $\text{BR}(Y_1 \to \gamma\gam)=100\%$, see the contours with white boxes. Only the region covered by blue scattering points is allowed by the potential stability condition for $\lambda_s=2\pi$. The dashed green and red lines show the contours of the proper decay length of $Y_1$ (in the unit of meter) and the total decay width of $Y_2$ (in the unit of GeV), respectively.
}
\label{fig:realscalar_di-photon}
\end{center}
\vspace{1mm}
\end{figure}
The viability of this scenario is illustrated in Fig.~\ref{fig:realscalar_di-photon}. There, we vary two free parameters $t_\beta$ and $\kappa_2$, and draw the contours of the production cross session $\sigma(gg \to Y_2) \text{BR}(Y_2 \to Y_1 Y_1)$, assuming the $\text{BR}(Y_1 \to \gamma\gam)=100\%$ for each $Y_1$ as argued in Sec.~\ref{ssec:singlet}. The decay length of $Y_1$, $L_{\text{decay}}$ and the total width of $Y_2$ are also presented. In this figure, $m_{Y_1}=150$ MeV is adopted as a typical value and $\lambda_s=2\pi$ is chosen to make our model compatible with the unitarity conditions~\cite{Drozd:2014yla}.
Clearly, there exists parameter space which could yield the 750 GeV di-photon signal, comparable to those observed at the LHC.
The yellow shaded strip indicates the $\sigma(gg \to Y_2) \text{BR}(Y_2 \to Y_1 Y_1)$ within 4-10 fb.
However, it is not necessarily acceptable because the allowed value of $|\kappa_2|$ for a fixed $\tan\beta$ is constrained by the vacuum stability, which has been shown to play the most important role in eliminating the parameter space~\cite{Drozd:2014yla}.
In essence, the upper bound on $|\kappa_2|$ is determined by $t_\beta, \lambda_s$ as well as quartic couplings $\lambda_{1-5}$ in the Higgs sector. A simple derivation to obtain this bound can be found in the Appendix.
The allowed region is sketched by blue scattering points in the figure and this band displays a tendency of compression as $\tan\beta$ increases.
Though it is still possible to find a value of $\kappa_2$ for $\tan\beta \lesssim 4$ producing the desired signal rate which can fit the data,
the decay length of $Y_1$ for such value of $\kappa_2$ within the stability bound is incredibly long due to the insufficient mixing.
Therefore, we conclude that this model containing a real singlet is difficult to simultaneously yield the di-photon signal comparable to the observed data and also achieve a reasonable proper decay length.
\section{Model II: complex singlet embedding a pseudoscalar dark matter}
\label{sec:complexmodel}
To remedy the problem of too long decay length present in the real singlet model, we are now introducing a {\it complex} scalar gauge singlet field $\mathbb{S}$ to the 2HDM in this section. The $\mathbb{S}$-associated part is given by
\begin{equation}
\label{complexL}
\mathcal{L}_{\mathbb{S}} = \frac{1}{2} m_0^2 \mathbb{S}^2 +\frac{1}{4!}\lambda_s \mathbb{S}^4+\kappa_1 \Phi^\dagger_1 \Phi_1 \ \mathbb{S}^\dagger \mathbb{S} + \kappa_2 \Phi^\dagger_2 \Phi_2 \ \mathbb{S}^\dagger \mathbb{S}+\omega_1 \Phi^\dagger_1 \Phi_1 (\mathbb{S}^\dagger+\mathbb{S})+\omega_2 \Phi^\dagger_2 \Phi_2 (\mathbb{S}^\dagger+\mathbb{S})
\end{equation}
Here $\omega_1$ and $\omega_2$ are dimensionful parameters.~\footnote{We confine our analysis to the CP conserving model where the CP phases of these two interaction terms are taken zero, resulting in real $\omega_1$ and $\omega_2$.} Unlike the real singlet model discussed in Sec.~\ref{sec:realmodel}, we do not impose the $Z'_2$ on the complex filed $\mathbb{S}$ so that the linear terms of $\mathbb{S}$ are present in the above Lagrangian. The singlet scalar can be expanded as
\begin{equation}
\mathbb{S}=\chi_S+i\chi_A
\end{equation}
We stress that the complex singlet field $\mathbb{S}$ cannot acquire a VEV: $\langle \mathbb{S} \rangle =0 $, otherwise the CP-odd mode $\chi_A$ would be a massless Goldstone boson.
Consequently, the minimization conditions in the scalar potential takes the same form as Eqs.~(\ref{eq: minh1}) and (\ref{eq: minh2}) with the elimination of the $v_s$ term, while the one with respect to $v_s$ ({\it i.e.} Eq.~(\ref{eq: mins})) is absent.
After EWSB, this complex singlet model has an additional CP-odd state $\chi_A$, that could be a candidate for dark matter.
However, among the three CP-even states there are many similarities between Model~I and Model~II considered in this paper even if the $\mathbb{S}$ in Model~II does not acquire a vev. It is useful to comment that the single $\chi_S$ interactions with two doublets appearing in Eq.~(\ref{complexL}) are also present in Model~I when the real field $S$ gets vev. This implies that the coupling $g_{\chi_S H^\pm H^\mp}$ can be obtained from Eq.~(\ref{eq:chiHH}) with a simple substitution of $\kappa_a v_s$ by $\omega_a$. This replacement is also true in the $(\rho_1,\rho_2,\chi_S)$ mixing.
To avoid redundancy, we shall transmit the results which have been derived in the real singlet model to the complex singlet model.
A particularly important exception is the $Y_2 Y_1 Y_1$ coupling. In the present model, this coupling depends on both $\omega_a$ and $\kappa_a$ $(a=1,2)$. As a result. such an interplay provides us the possibility of simultaneously ensuring the potential stability which is crucially determined by $\kappa_a$ and achieving a desired doublet-singlet compound for the light scalar through an essential increment on the $\omega_2$. In the end, the difficulty of too-long decay length can be overcome.
\subsection{Spectrum and couplings}
Without the $Z'_2$ protection, $\mathbb{S}$ can singly couple to the doublets. This results in the doublet-singlet mixing which occurs only when the real singlet gets vev as discussed in the previous section.
As already argued, the mass matrix in the $(\rho_1, \rho_2, \chi_S)$ basis can be obtained from Eq.~(\ref{m square of real singlet}) by replacing $k_a v_s$ with $\omega_a$
and $m_\chi^2=m_0^2+\kappa_1 v_1^2+\kappa_2 v_2^2$ in this model. By virtue of this similarity, one can simply follow the procedures described in Sec.~\ref{sec:realmodel} to derive the mixing parameters $\Delta$, $D$, the composition of the resulting mass eigenstates $h, Y_1, Y_2$ and their mass spectrum. The corresponding results are analogous to Eqs.~(\ref{eq:Ddef}), (\ref{eq:deltadef}), Eq.~(\ref{eq:eigenstatecomp}) and Eqs.~(\ref{eq:hmass}), (\ref{mYeq}) with the only substitution of $k_a v_s \to \omega_a$. This implies that in this scenario the mixing between CP-even fields as described by $s_\theta$ depends on $\omega_2$ as opposed to $\kappa_2$.
In accordance with the LHC data for 125~GeV Higgs, we also employ the alignment limit and require the zero singlet fraction for the $h$ as discussed in Sec.~\ref{sec:realmodel}. In this context, $D=-2 v\omega_2 t_\beta$ and $\Delta=0$ because $\omega_1=-\omega_2 t_\beta^2$.
In contrast to the CP-even fields, the mixing in the CP-odd sector is absent in this scenario and hence $\chi_A$ itself is a mass eigenstate with mass $m_{\chi_A}=m_\chi$. Once it is light enough, both the SM Higgs $h$ and the 750~GeV resonance state $Y_2$ in the model could decay into a pair of them. The $h$ to $\chi_A\chi_A$ decay, if kinematically allowed, can be switched off by the vanishing relevant coupling $g_{h \chi_A\chi_A}$. This actually gives rise to Eq.~(\ref{eq:k1k2v2}), $\kappa_1=-\kappa_2 t_\beta^2$, under which the couplings given in Table~\ref{table:scalarcoupling} are maintained except the substitution of $g_{\chi H^+ H^-}$ by
\begin{equation}
g_{\chi_S H^+ H^-} =2\omega_1 s_\beta^2+2 \omega_2 c_\beta^2=2\omega_2(1-t_\beta^2)
\end{equation}
In addition, similar to the real singlet case discussed in Sec.~\ref{sec:realmodel}, the decay channel $h\to Y_1 Y_1$ is also closed by Eq.~(\ref{eq:k1k2v2}) in this scenario.
\subsection{$Y_2$ decay}
\label{sec:compY2decay}
In addition to the main decay channels (to $Y_1 Y_1$ and $t \bar t$), $Y_2$ in this model can invisibly decays into $\chi_A\chi_A$. Of the three relevant couplings, $g_{Y_2 t t}$ is given in Table~\ref{table:scalarcoupling}, while the other two can be easily obtained at the exact alignment limit $c_{\beta-\alpha}=0$.
\begin{align}
g_{Y_2 Y_1 Y_1}
&=3 c_\theta s_\theta^2 g_{HHH}-\kappa_2 t_\beta (c_\theta^2-2s_\theta^2)c_\theta v+s_\theta(s_\theta^2-2c_\theta^2)(1-t_\beta^2)\omega_2\label{eq:gy211}\\
g_{Y_2 \chi_A\chi_A} &= (\kappa_1-\kappa_2)c_\beta s_\beta c_\theta v=-\kappa_2 t_\beta c_\theta v
\end{align}
It should be noted that the heavy Higgs $H$ in the 2HDM context generally has a sizable decay branching ratio into a pair of SM-like Higgs $h$. However, we examine that this coupling $g_{Y_2 hh}$ is vanishing at the exact alignment limit~\cite{Bernon:2015qea}. Therefore, we do not consider this decay mode in the present analysis.
For illustration, we display the branching ratios of $Y_2$ decay in Fig.~\ref{fig:Y2Br} by taking $\tan\beta=2,3,4$ (from left to right) in the range that could yield an observed di-photon cross session as we will see. In each graph, branching ratio curves are drawn in different colors corresponding to $\kappa_2=0$ and the maximal value of $|\kappa_2|$ such that the stability condition is obeyed for each $\tan\beta$. As seen from Fig.~\ref{fig:Y2Br}, in the small mixing case when $\omega_2$ is small, $t\bar t$ channel dominates the decay of $Y_2$ for small $\tan\beta$ as long as $\kappa_2$
stays within the stability bound, whereas $BR(Y_2\to Y_1Y_1)$ and $BR(Y_2\to \chi_A \chi_A)$ are not substantial.
Reversely, increasing $\omega_2$ will invoke a larger mixing, which leads to the reduction of the doublet fraction in the $Y_2$, while $Y_1$ gains more doublet component. This results in a quick grow on the coupling $g_{Y_2 Y_1 Y_1}$, but little change on the coupling $g_{Y_2 t t}$. Consequently, BR$(Y_2 \to t\bar t)$ drops drastically while BR($Y_2\to Y_1Y_1$) becomes important as the mixing increases.
Thus, large $\omega_2$ is favored in order to accomplish a sufficiently large BR($Y_2\to Y_1 Y_1$) for our purpose.
In the plots we also observe that BR($Y_2\to Y_1Y_1$) vanishes at a certain value of $\omega_2$ when $\kappa_2<0$.
This is the consequence of the interplay between $\kappa_2$ and $\omega_2$ terms in the coupling $g_{Y_2 Y_1 Y_1}$ given in Eq.~(\ref{eq:gy211}).
Besides, the particularly presented decay $Y_2$ into $\chi_A \chi_A$ is always sub-dominant. This can be attributed to the smallness of $\kappa_2$ as demanded by the stability.
\begin{figure} [t]
\centering
\includegraphics[height=0.3\textheight,width=0.32\textwidth]{PseudoScalar2HDM_BrY2_tb2.pdf}
\includegraphics[height=0.3\textheight,width=0.32\textwidth]{PseudoScalar2HDM_BrY2_tb3.pdf}
\includegraphics[height=0.3\textheight,width=0.32\textwidth]{PseudoScalar2HDM_BrY2_tb4.pdf}
\caption{The $\omega_2$ dependence of BR($Y_2\to Y_1 Y_1$) (solid), BR($Y_2\to t\bar t$) (dashed) and BR($Y_2\to \chi_A\chi_A$) (dotted) are indicated.
In each graph, the branching ratio curves are drawn corresponding to $\kappa_2=0$ (purple) and the minimal/maximal (blue/red) value of $\kappa_2$ (which are given on top of each graph) such that the stability condition is obeyed for each $\tan\beta$.
}
\label{fig:Y2Br}
\end{figure}
\subsection{Phenomenology of 750~GeV state}
\begin{figure}[t]
\begin{center}
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_o2_50.pdf}
\hspace{1mm}
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_o2_100.pdf}
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_o2_200.pdf}
\hspace{1mm}
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_o2_400.pdf}
\caption{The di-photon cross section $\sigma(gg\to Y_2) \text{BR}(Y_2 \to Y_1 Y_1)$ produced by the 750 GeV resonance $Y_2$ under the assumption that $\text{BR}(Y_1 \to \gamma\gam)=100\%$, see the contours with white box. $\omega_2$ is chosen different values specified on the top of each graph. Only the region which is covered by blue (cyan) scattering points for $\lambda_s=2\pi$ $(0.2)$ is allowed by the potential stability condition. The dashed lines with green and red box label the contours of the proper decay length of the light singlet $Y_1$ and the total decay width of $Y_2$, respectively. }
\label{fig:di-photon_o2}
\end{center}
\vspace{1mm}
\end{figure}
We begin our discussion for 750~GeV state by presenting in Fig.~\ref{fig:di-photon_o2} the contours of cross section $\sigma(gg\to Y_2\to Y_1Y_1)$. For illustration, we take $m_{Y_1}=150$~MeV and $\lambda_s=2\pi$~\footnote{Here we adopt the somewhat more conservative value of $\lambda_S=2\pi$ so that our model would remain valid to at least a moderately higher scale before additional new physics would need to be included to obtain a theory valid at all energy scales. The impact of $\lambda_S$ value on the stability was discussed in Ref.~\cite{Drozd:2014yla} in detail.}.
The four graphs are produced by choosing $\omega_2=50, 100, 200, 400$~GeV in sequence. The yellow shaded band corresponds to $\sigma(gg\to Y_2\to Y_1Y_1)$ within 4-10 fb that could fit the ALTAS+CMS data.
In the figure, the decay length of $Y_1$ and the total decay width of $Y_2$ are also shown in green and red dashed lines, respectively. The contour numbers are uniformly indicated in the corresponding colored boxes.
In addition, we examine the stability condition of this model and mark the allowed region which is covered by blue scattering points. The trapezoid shape indicates that a smaller value of $\tan\beta$ is able to accommodate a larger $|\kappa_2|$.
This can be understood from the fact that $\kappa_1=-\kappa_2 \tan\beta$ is employed and the stability condition essentially places upper bounds on $|\kappa_1|$ and $|\kappa_2|$ as well as their ratio~\cite{Drozd:2014yla}.
The presence of blue scattering points in the yellow shaded band tells us that
this complex singlet model could easily yield the observed cross section $\sigma(gg\to Y_2\to Y_1 Y_1)$, while obeying the stability condition. Recall that BR$(Y_1\to \gamma\gamma)=100\%$ is assumed in the estimate. Next, we examine the decay length of $Y_1$.
This figure shows the decay length of $Y_1$ (green dashed line) has no dependence on $\kappa_2$ but is very sensitive to $\tan\beta$ and $\omega_2$.
Particularly, the decay length of $Y_1$ increases as $\tan\beta$ goes down. Both phenomena are analogous to the real singlet model presented in Sec.~\ref{sec:decayY1}.
Apparently, the requirement that the decay length $L_{\text{decay}}\lesssim 1$~m has significant impact on eliminating the small $\tan\beta$ region.
Another important measurement that characterizes the potential 750~GeV di-photon resonance is its total width (red dashed lines), which varies from a few to tens of GeV seen from Fig.~\ref{fig:di-photon_o2}. Thus, this could be used as a critical signature in examining this scenario or determining the model parameter if confirmed.
Two interesting observations regarding the cross section are placed in order. First, the cross section contours display an asymmetry with respect to $\kappa_2=0$. This is actually a result of Eq.~(\ref{eq:gy211}).
Second, the magnitude of cross section becomes less sensitive to $\kappa_2$ as $\omega_2$ increases, as seen from the fact that the contours keep tilting to the left.
\begin{figure}[t]
\begin{center}
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_k2_p.pdf}
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_k2_n.pdf}\\
\includegraphics[height=0.3\textheight,width=0.48\textwidth]{PseudoScalar2HDM_k2small.pdf}
\caption{The di-photon cross section $\sigma(gg\to Y_2) \text{BR}(Y_2 \to Y_1 Y_1)$ in the $\omega_2 - \tan\beta$ plane with $\kappa_2= 0.1, -0.1, 0.01$, under the assumption that $\text{BR}(Y_1 \to \gamma\gam)=100\%$. The legend is the same as in Fig.~\ref{fig:di-photon_o2} except the blue/cyan horizontal dashed line which gives the approximate upper bound on $\tan\beta$ coming from the stability. }
\label{fig:di-photon_k2}
\end{center}
\vspace{1mm}
\end{figure}
Alternatively, the result can be projected onto $\kappa_2 - \omega_2$ plane. We show two examples of $\kappa_2=\pm 0.1$ in Fig.~\ref{fig:di-photon_k2}.
One can easily gain the additional information from this figure regarding the $\omega_2$ dependence.
As $|\omega_2|$ goes large, the decay width of $Y_1$ increases because $Y_1$ is composed of more doublet fraction. This eventually results in a shorter decay length.
On the other hand, for $\tan\beta \leq 2$, the total width of $Y_2$ is marginally sensitive to $\omega_2$. This is because $t\bar t$ channel dominates the decay of $Y_2$ in the small $\tan\beta$ region as seen in Fig.~\ref{fig:Y2Br}.
Finally, it is important to mention that the minimal width of $Y_2$ is $\gtrsim 1~\text{GeV}$ in this scenario. This is in contrast to many models where the width of 750 GeV state is of sub-GeV scale.
\section {Discovery prospects at the collider}
\label{sec:prospects}
\subsection{Searching for $Y_2 (750)$ in other channels}
Aside from into the $Y_1Y_1$, the $Y_2$ state of 750~GeV largely decays into $f\bar{f}$ final state. This has been shown in Sec.~\ref{sec:compY2decay}. The cross section of producing a $Y_2$ decaying to the $t\bar t$ or $b\bar b$ final states could be as large as hundreds of fb. Hence, they could be possible channels to search for $Y_2$ at the future run of the LHC. Compared to the tree-level $f\bar f$ decays, the branching ratio of the loop-induced decay $Y_2\to gg$ is negligible. This channel also suffers from a large QCD background at the LHC. Therefore, it is less likely to be a promising discovery channel.
\subsection{Pseudo-scalar Higgs and charged Higgs}
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{PseudoScalar2HDM_mamctypeII.pdf}
\caption{The allowed range on $m_A$ and $m_{H^\pm}$.
The gray points are compatible with both theoretical constraints and EWPO within the $\pm 3\sigma$ range. The green dashed line here indicates the mass relation $m_{A}=750-m_Z$. Thus, $Y_2$ cannot decay to $AZ$ on-shell on the right hand side of this line. Similarly, the orange and blue dashed line represent $m_{A}=750+m_Z$ and $m_{A}=m_{H^{\pm}}+m_W$
(The result is generically the same as in both Type~I and Type~II models.)
}
\label{fig:mAmHpmscan}
\end{figure}
The range of the rest of two scalars, the pseudo-scalar Higgs mass $m_A$ and charged Higgs mass $m_{H^{\pm}}$, is also interesting. In principle, $m_A$ can be above or below either of two CP-even Higgs bosons, and even $m_A < m_h/2$ is possible and consistent with the data~\cite{Bernon:2014nxa}. However, once the heavy Higgs mass ($m_{Y_2}$ in our model) is fixed, the allowed range of $m_A$ is limited.
The interrelation between $m_A$ and $m_{H^{\pm}}$ for the case of $m_{Y_2}=750$~GeV is illustrated in Fig.~\ref{fig:mAmHpmscan}.
There, we observe that $m_{H^{\pm}},m_A$ are bounded in the $400-950$~GeV. As expected, this result is identical to what is displayed in the 2HDM~\cite{Bernon:2015qea}. This is not surprising because the introduced complex singlet does not generate the mixing in the CP-odd states.
Due to kinematical suppression, $Y_2 \to AA/H^{\pm}H^{\mp}$ is always forbidden. The absence of these two decay modes is crucially important in the success of explaining the di-photon excess. Otherwise, they will eat a large amount of branching ratio of the $Y_2$ decay, so that the $Y_2 \to Y_1 Y_1$ decay will be heavily suppressed. While $Y_2 \to ZA$ and/or $Y_2 \to W^{\pm}H^{\mp}$ decay is kinematically possible for $m_{H^{\pm}},m_A \lesssim 670$~GeV, their contributions to the $Y_2$ decay are so small that could be neglected.
As regards for the $A$, $AZh$ coupling is vanishing in the exact alignment limit and thus the exotic decay $A \to Zh$ is not present.
However, $A\to Y_2 Z$ (right to the orange line) and $A\to W^\pm H^\mp$ (above the blue line) could be important in addition to the fermionic decay being the potential discovery channels for the $A$ including the $t\bar{t}$, $b\bar{b}$ and $\tau^- \tau^+$ final states. In general, the cross section of producing the $A$ via gluon-fusion is proportional to $\tan\beta$ and varies from few pb and 10 fb at the 13 TeV, depending on the exact mass.
Since the decay into $t\bar{t}$ dominates for moderate $\tan\beta$ we consider here, the cross section $gg\to A\to b\bar{b}$ can at most reach $\sim 10$~fb, an order which hardly enables them to compete with the large QCD background.
\section{Conclusions}\label{sec:conclusion}
First, we find that a pure doublet state at 750 GeV generally has a very limited branching ratio for the loop-induced di-photon decay when the tree-level decay to $t\bar{t}$ is present. As a result,
the process of $gg\to H/A \to \gamma\gamma$ with $CP$-even or $CP$-odd Higgs $H/A$ being identified as the 750 GeV resonance in the minimal version of 2HDM cannot reproduce the di-photon signal that is comparable to the observed level.
Other than through direct decay, we alternatively consider the di-photon signal that arises from two bunches of collimated photon jets emitting from a pair of highly boosted particles in this work. In particular, we studied the impact of $m_{Y_1}$ and its proper decay length on collider phenomenology. It turns out that $m_{Y_1}$ should be $\lesssim 210$~MeV to efficiently produce a photon jet as narrow as it must be. However, such a light boosted state cannot be the $A$ in the 2HDM, once the $H$ is identified as the 750 GeV resonance. Therefore, we extend the 2HDM by adding a gauge singlet scalar field.
Though the singlet scalar field generically mixes with two doublet fields, it is possible to accomplish a SM-like Higgs $h$ with another two mass eigenstates. They include one heavy doublet-like $Y_2$ and one light singlet-like $Y_1$ with mass of sub-GeV.
The $Y_2(750)$ can decay to a pair of $Y_1$, each of which can further decay into two photon jets. On the other hand, the presence of this mixing, although rather small, is the key of controlling the proper decay length of $Y_1$.
Two specific models containing an extra real or complex singlet scalar are studied. In both models, the SM-like $h$ decay to $Y_1Y_1$ or invisible final states are swiched off.
For the real singlet model,
the mixing between heavy Higgs and the singlet state is strongly constrained so that it is difficult to simultaneously yield the di-photon signal comparable to the observed data and also achieve a reasonable proper decay length. In contrast, in the absence of the $Z_2$ symmetry the complex model has linear singlet terms, which are irrelevant to the Higgs invisible decay but play an essential role in generating the mixing in the scalar sector. As a result, this model is easy to yield the 1-10~fb cross section in the di-photon final state with a decay length of $\mathcal{O}(1)$~m for the $Y_1$, meanwhile parametrically predicts the width of 750 GeV resonance $\gtrsim 1$~GeV. In addition, the pseudoscalar component of the singlet is naturally stable and hence could be a dark matter candidate.
Finally, we have discussed the discovery prospects of other scalar states such as the $CP-$odd $A$ and the charged Higgs $H^\pm$ at the future LHC run.
\section*{Acknowledgements}
We thank John Conway and Yong Yang for useful discussions regarding the details of collider tracker to verify the possibility of this scenario.
We are also grateful to John~F.~Gunion and Zhen~Liu for their particular attention as well as helpful comments on the manuscript.
YJ is supported by the Villum Foundation.
\section*{Appendix: Stability constraint}
To derive the stability bound we use the inequalities of $\kappa_{1,2}$ in \cite{Drozd:2014yla}. Given in Eq.~(\ref{eq:k1k2v2}), $\kappa_1$ and $\kappa_2$ are of the opposite signs: $\kappa_1=-t_\beta^2 \kappa_2$. Employing this relation, the upper limit of $|\kappa_2|$ reads
\begin{equation}
|\kappa_2|\leqslant \text{Min}(A_1,A_2)
\end{equation}
where $A_1=\sqrt{\lambda_1 \lambda_s /12}/ t_\beta^2$, $A_2=\sqrt{\lambda_2\lambda_s/12}$. When $\lambda_3 < 0$, one should ensure
\begin{eqnarray}\label{eq:k2constraint}
\left\lbrace \begin{array}{ll}
|\kappa_2|\leqslant \sqrt{\frac{A_1^2A_2^2-A_3^2}{A_1^2+A_2^2-2A_3}},& \text{if} \quad A_3 < A_1A_2\\
|\kappa_2|=0,& \text{if} \quad A_3 \geqslant A_1A_2
\end{array}\right.
\end{eqnarray}
where $A_3=-\lambda_3 \lambda_s/(12 t_\beta^2)$.
Additionally, if $\lambda_3+\lambda_4-|\lambda_5| < 0$ is satisfied, one more constraint is required which can be obtained by replacing $A_3$ by $A_4$ with
\begin{equation}
A_4=-(\lambda_3+\lambda_4-|\lambda_5|)\lambda_s/(12 t_\beta^2).
\end{equation}
|
\section*{Abstract}
The role of Network Theory in the study of the financial crisis has been widely spotted in the latest years. It has been shown how the network topology and the dynamics running on top of it can trigger the outbreak of large systemic crisis. Following this methodological perspective we introduce here the Accounting Network, i.e. the network we can extract through vector similarities techniques from companies' financial statements. We build the Accounting Network on a large database of worldwide banks in the period 2001-2013, covering the onset of the global financial crisis of mid-2007. After a careful data cleaning, we apply a quality check in the construction of the network, introducing a parameter (the Quality Ratio) capable of trading off the size of the sample (coverage) and the representativeness of the financial statements (accuracy). We compute several basic network statistics and check, with the Louvain community detection algorithm, for emerging communities of banks. Remarkably enough sensible regional aggregations show up with the Japanese and the US clusters dominating the community structure, although the presence of a geographically mixed community points to a gradual convergence of banks into similar supranational practices. Finally, a Principal Component Analysis procedure reveals the main economic components that influence communities' heterogeneity. Even using the most basic vector similarity hypotheses on the composition of the financial statements, the signature of the financial crisis clearly arises across the years around 2008. We finally discuss how the Accounting Networks can be improved to reflect the best practices in the financial statement analysis.
\section*{Introduction}
Network Theory has been used to establish how contagion, through a variety of channels (mutual exposures, social networks of board members, moral hazard from permissive regulations, financial instruments like swaps and derivatives, etc.), triggered the outbreak of the 2007-08 crisis. Scholars suggest that financial systems may affect positively economic development and its stability (\cite{Beck09, Beck11, Lev05}), although they may represent a source of distress which leads to bank failures and currency crises, or greater contraction for those sectors that depend more on external finance during banking crisis (\cite{arr, RR}). As a response to the recent financial turmoil, the banking sector has been affected by a substantial reorganization (\cite{BISANNUAL14}). For instance, as highlighted by the European Central Bank for the Euro area \textit{the main} \textit{findings reflect the efforts by banks to rationalize banking businesses, pressure to cut costs, and the deleveraging process that the banking sector has been undergoing since the start of the financial crisis in 2008} (\cite{ECB}). This implies that market pressure and regulatory amendments induce banks to reduce their levels of debt, through cost containment and stricter capital requirements. In addition, a gradual improvement in bank capital positions aims to enhance the capacity of the system to absorb shocks arising from financial and economic distresses. This limits the risk of spillover effects from the financial sector to the real economy and put the financial system in a better condition to reap the benefits of economic recovery. In particular, as the financial boom turned to a bust, banks' stability deteriorated abruptly and the economy entered a \textit{balance sheet recession}, which depressed spending levels through a reduction in consumption by households and investments by firms. Therefore, although at an uneven pace across regulations, the need to strengthen fundamentals has influenced the banking sector, and differences in banks' portfolio allocations, financial performances, and capitalizations might be interpreted as the combined results of policy decisions and sectoral responses to changes in the regulatory framework (see e.g. \cite{ALLEN}, \cite{DIARANJ}).\\
\noindent This paper relates to the literature on banking development and performance evaluation during the recent crisis (see e.g. \cite{ASHIN}, \cite{BERBOW},\cite{BRUNN}). We consider a large data set of worldwide banks retrieved from \textit{Bloomberg}, focusing on financial statements spanning from 2001 to 2013. We introduce a network based on similarities between banks' financial statement compositions (hereinafter \textit{Accounting Network}). Due to data limitations, the reference sample is restricted to banks for which a continuum and stable set of variables is available for the entire period. The introduction of a methodology (\textit{Quality Ratios}) to measure banks' data coverage aims to prevent that missing values for some variables or lack of annual financial statements for some banks affect the overall picture. We then exploit the maximum amount of available information from financial statements without further reducing the set of variables through an arbitrary selection of the financial statements fields. This choice aims to avoid any selection bias. Moreover, total assets (as a proxy for size) for each bank is applied to normalize banks' financial statements measures to prevent the emergence of ``size effects'' as the sizes of institutions are spanning for various orders of magnitude.\\
The introduction of Accounting Networks establish a bridge between the external perspective arising from market data and the internal one based on banking activities indicators. We study how Accounting Networks can be exploited to provide a description of the banking system during the crisis. This part sheds light on whether banks under different regulatory frameworks and diversification degrees have reacted to the crisis by strengthening their business peculiarities or by converging towards similar practices (\cite{BelStu},\cite{DeHui}). We rely on the assumption that market data alone, although highly representative of investors' perception of the banking sector, might be dis-informative during periods of distressed market conditions. This, in turn, stimulates a broader exploitation of the information on banking activities, thus pointing to a more comprehensive investigation which takes into account also the internal perspective arising from financial statements data. In addition, the use of accounting data allows a partition of business activities where banks are involved in, providing therefore an approximation of the state of the system related to several potential channels through which the financial distress might propagate. This is appealing also for regulators, since authorities are interested in a wide set of economic indicators in order to prevent the systemic relevance of financial institutions and they introduce specific requirements and constraints which affect directly financial statements measures. For these reasons, we believe that enriching the debate on financial stability by means of the Accounting Networks might give new clues about the resilience of the banking system. \\
Another important result is the possibility of getting a neutral partition of banks in "network communities" (i.e. clusters) that results from the analysis of the network through community detection algorithms like the \textit{Louvain} modularity maximization. The results indicates that regional communities evolve in time and the crisis has a clear role in weakening geographically determined structures. Furthermore, we focus on proxies for leverage, size and performance in order to understand if these variables have played a key role among the set of economic measures usually applied to classify banks (see e.g. \cite{BLU},\cite{HUI}). Hence, we aim to answer the question whether the collapse of financial markets has weakened these relationships, limiting therefore the power of traditional indicators to identify clusters of homogeneous banks. Correlation diagrams applied to show how network variables are related to economic measures suggest a turning point in correspondence of the outbreak of the crisis, which influenced the role of proxies for leverage, size or performance to group similar banks. This preliminary results motivated the last section, where by means of Principal Component Analysis we investigate which economic features are more likely to characterise the heterogeneity of the communities before, during and after the collapse of 2007-08.\\
The remaining part of the work discusses open issues and future lines of research, such as open questions on how to improve the building of the Accounting Networks. In particular, the effectiveness of this approach can be enhanced by means of a careful variable selection based on the best financial practices applied in the evaluation of the financial statements structures. In addition, a more accurate normalization of the variables and caring about national regulations may increase the usefulness of the methodology. Furthermore, matrix filtering techniques and missing data reconstruction for financial statements information can enhance the extraction of meaningful clusters. Then, more advanced and focused tools could be conceived to analyse banks evolution towards similar business configurations or, alternatively, their divergent patterns as a response to changing market conditions.
\section*{Methods}
\subsection*{Dataset preparation}
The dataset we analysed covers the set of banks provided by \textit{Bloomberg} which were active (i.e. with traded instruments) at the end of the first quarter of 2014. Although quarterly information is available, we prefer to focus on annual balance sheets and income statements for accounting standard reasons, as different countries can have different obligations in terms of the provision of quarterly financial statements and this can lead to a mismatch and a poor variables coverage. Data are collected during the reference period from 2001 to end of 2013.\\
As regards financial statements data, we select a large set of variables among those available in \textit{Bloomberg} and related to the current regulatory framework (\cite{BASEL}). We rely on the existing literature for the selection process, although providing a neutral approach. We focus the analysis on proxies for banking business models (see e.g. \cite{ALTU},\cite{CAL}). In particular, balance sheet data provide a year-by-year picture of stock variables in terms of assets and liabilities for different instruments and maturities, while income statement data describe annual economic performances by partitioning profits and losses according to banking activities ranging for instance from interests to fees. Since national regulations allow firms to fix a different end of fiscal year, we extend the ``end of year'' definition and the relative financial statements according to a window in the range between three months before and after the end of the solar year. Solving overlapping issues in variables definitions, as well as the base currency choice, constitute the first step in the data pre-processing procedure. Firstly we discard total and sub-total measures (as they are redundant measures), and secondly we choose US dollars as currency base, thus facilitating banks comparisons.\\
Working with financial statements data often leads to limitations in data coverage and completeness. Therefore, the starting point of our analysis is represented by the selection of a stable set of banks in terms of data availability during the sample period. In particular, banks might change the composition of their financial statements or they might be excluded by the \textit{Bloomberg} provider due to several reasons, such as for instance a new regulation or a change in the bank's economic activities. This, in turn, might cause \textit{missing values} for some variables or lack of financial statements for several banks in certain years. In order to limit the impact of these issues on our findings, we define a methodology to measure the coverage of available variables for each bank in the reference period. We refer to the \textit{Quality Ratios (QRs)} as the proportion of available and usable variables $V_{OK}$ over the maximum of all possible ones $V_{ALL}$ in the sample period: $QR=V_{OK}/V_{ALL}$. The tuning of this indicator, combined with two more filters on the frequency of financial reporting, provides a stable set of banks identified by their QR. The two additional criteria are: a minimum number of financial statements of ten out of thirteen possible fiscal years and a maximum gap period between two consecutive annual reports equal to seven hundred days. Once selected those banks that report almost continuously their financial statements, we study them according to their respective QR.\\ Actually, individual QRs, as empirically computed on the entire perimeter, lie in the range between 0.3 (low accuracy/coverage) and 0.8 (high accuracy/coverage). Interestingly, many measures computed on the sets of banks obtained by fixing the QR do not seem to be significantly affected by its choice (except, as expected, for high QRs, where the size of the sample reduces significantly). With greater values of the QR parameter we have less available banks to be considered, since only few of them have a large set of variables present in many of their financial statements. As the estimates are stable in a reasonable QR range, in this work we decide to use the set arising from the case of QR = 0.5 that, even if arbitrary, still represents a good compromise between the accuracy of the estimate and the size of the sample (see Figure \ref{figQR}).\\
\subsection*{Accounting networks}
For every year a vector of financial statement variables is assigned to each bank and used to compute the cosine similarities between pairs of banks/nodes. Here the intuition is that the most similar banks (as from their financial statements) must stay closer in the network and form a cluster. Then, the measure ``cosine similarity'' is transformed into a metrics (as triangular inequality must hold, the square root is used). The definition is the following: we compute the cosine of the angle between each pair of vectors with the dot product and then we apply the simple transformation $w_{i,j} = 1 - \sqrt{1 - C_{i,j}^2}$, where $w_{i,j} \in [0, 1]$ and $C_{i,j}$ is the cosine similarity between \textit{i} and \textit{j}. In network terms $w_{i,j}$ is the weight. This transformation (see \cite{DongenEnright} for an introduction to similarity measures and relative metrics) is used to move from the cosine similarities defined in the space [-1,1] to weights in the interval [0,1]. With this transformation the more two nodes are similar (or anti-similar) the larger is the weight, while a weight of 0 is assigned to a pair of nodes having totally dissimilar financial statements (actually, in our networks cosine similarities range mainly between 0 and +1).\\
\noindent In addition, before the computation of the metrics, we need to take care of the size distribution of banks, as it spans over several orders of magnitude. To avoid a bias toward large institutions, for each bank we divide all variables in its vector by the respective total assets in such a way that the attributes of the vector refer to economic and financial \textit{ratios}. This operation ensures that clusters will be formed by banks with similar business activities regardless their sizes.\\
An important methodological choice of our study is the ``neutral'' approach used for the selection of the variables within the financial statements. A part from removing related and redundant measures (total and subtotals), we used all the available information applying the same weight to each variable in the vectors. This agnostic approach is in line with the goal of the paper, i.e. introducing the concept of Accounting Network, although we are aware that practitioners can give a different importance to each variable of the financial statement. In our perspective we expect that the relevant information will emerge in a bottom up process, as a spontaneous feature selection carried by our methodology.
Finally, we introduce a confidence level (95\%) during the link formation. By using a Montecarlo sampling test, if the cosine similarity is statistically significant with 95\% of confidence we retain the link otherwise we discard it. As a result of this filtering procedure, we observe that the networks tend to be very dense and almost complete. The most of the information is carried by the weights of the links and less by the simple topology (degrees and other structural features).
\subsection*{Community detection}
A classical method to investigate the structure of a network is the search of communities, i.e. regions of the network with larger \textit{internal} links density. Intuitively, these regions are formed by clusters of nodes with higher degrees or, for weighted networks, with larger strengths. Several methods were proposed to find network communities without imposing a priori the number of communities but letting them emerging from the network itself. Among others we cite the optimization of the modularity that is a measure of how much the link structure differs from the random network where links are assigned with uniform probability and internal communities are not present (a part from fluctuation). For weighted networks, the modularity is defined by the following formula:\\
\begin{equation}
Q_{w} = \frac{1}{2W} \cdot \sum_{ij} \left( w_{ij} - \frac{s_i s_j}{2W} \right) \delta (c_i,c_j)
\end{equation}
\noindent where $s_i = \sum_{j} w_{ij}$ and $s_j = \sum_{i} w_{ij}$ are the strengths (sum of weights) of the nodes $i$ and $j$ respectively, $W=\sum_{ij} w_{ij}$ is the total sum of the weights and the function $\delta (c_i, c_j)$ is equal to 1 if $(i,j)$ belong to the same community or 0 if they are members of different communities. The maximum modularity value is 1 (an ideal case for which the communities are isolated) and can also take negative values. The 0 value coincides with a single partition that will correspond to the whole graph. A negative value means that there is no particular advantage in separating the nodes in that particular clusters and so there is not community structure whatsoever.\\
\noindent To study the presence of communities it is often necessary to prune the network cutting the links if their weight is below a certain threshold. In our case we intend to consider only the links formed by nodes having a large similarity/weight $w$ of their financial statement vectors. The procedure of pruning can be guided by the use of the tools related to the community detection methodology (\cite{Fortunato:2010}). In particular, working with the modularity optimization function (\cite{Newman:2004}), with the Louvain technique (\cite{Blondel:2008}), it is possible to look at the \textit{significance} associated to the threshold (as in \cite{Traag2013}), where the modularity is introduced as a parameter to check for the best resolved community structure. We use this parameter to help finding a reasonable pruning threshold range of values for the networks. A rule of thumb in this process is indeed avoiding network fragmentation, i.e. keeping the graph connected while removing not significant links. We made extensive tests computing quality/significance of the partitions (looking at the modularity parameter) using different pruning thresholds (i.e. removing the links having a low weight), determining a range of weights thresholds ($0.35< w_{i,j}<0.5$) that helps to prune the original networks to an optimal level. In this interval, communities are stable and the interpretation of each region can be seen as a result of the financial statement similarities across banks in different countries.
\subsection*{Network measures vs. economic indicators}
Comparisons among network measures and economic indicators are provided to describe the correlation between nodes' network topology and economic behavior. We study these features by means of extensive linear correlation tests (Pearson correlation) for the overall set of banks for each year and we verify the significance of the estimates by means of parametric tests. These estimates are based on the filtered networks, which are themselves based on the significance and the quality of the community detection algorithm. This analysis shows how nodes' network properties (e.g. \textit{Strength} or \textit{Clustering Coefficient}) are associated to basic economic indicators (e.g. \textit{Return on Assets}, \textit{Total Assets} and \textit{Total Debts to Total Assets}), thus showing whether nodes' topological properties are positively or negatively related to certain economic features and how these relationships have weakened or reinforced during the crisis.\\
\noindent Clustering coefficient is a measure of the local tendency of the nodes to form small regions of fully connected nodes, it is an average measure of the local clustering coefficient (actual number of triangles centered in each node over the total). Return on assets (ROA) is the net income over total assets and is a measure of the bank performance. Total debts to total assets is an indicator of the leverage of the bank and it is computed as the ratio between debts and its size (measured by total assets).
\subsection*{Principal Component Analysis}
Once communities are identified, we attempt to describe which financial statement variables are more likely to characterise these clusters. In order to facilitate comparability, we focus on those indicators more popular within the set of variables utilised to compute the cosine similarities (i.e. those indicators appearing with larger frequency in the entire dataset). In fact the inclusion of very poorly represented measures across different banks would have made the comparisons less effective with potential biases related to e.g. different regulations frameworks or geographical memberships. Hence, since we are interested in disentangling potential similarities/peculiarities across different communities, we prefer to rely on common and well-diffused measures of banking activities among those present in banks' financial statements. In addition, we enrich this set by means of indicators such as ratios (e.g. \textit{Return on cap} and \textit{Total debts to total assets}) and aggregated measures (e.g. \textit{Total assets}). Community detection identifies four main clusters, whose constituents are more numerous and stable in time. For the sake of conciseness, the \textit{Result} section will focus mainly on these communities. In particular, for each year we describe by means of Principal Component Analysis (PCA) which economic features are more (less) able to contribute to the explained variability of communities' members.\\
PCA is a multivariate technique that analyses observations described by several inter-correlated variables. PCA extracts the important information from the data and expresses it as a set of new orthogonal variables (principal components). In our exercise, since measures present different ranges of dispersion (e.g. by construction some ratios are bounded) we rely on a scaled version of PCA; finally, we consider only principal components with eigenvalues greater than $1$ (in almost all cases they correspond to the first $3$ components). Then, we compute the proportion of the variance of each original economic measure that can be explained by the selected principal components. This, in turn, leads to a ranking of the original economic measures in terms of their ability to describe a certain community's variability. In particular, since we are interested in how the onset of financial crisis has affected the banking system, we split this analysis in three periods: from $2001$ to $2006$ (before the crisis), from $2007$ to $2009$ (the onset of the crisis), and from $2010$ to $2013$ (after the breakdown of the markets). For each period we decided to characterise each community by the top and the bottom three measures, thus analysing how these ranks have evolved over time and across communities.
\section*{Results}
This section shows how Accounting Networks represent a complementary technique to traditional financial networks for the study of the banking system.\\
While financial networks reflect the view from the market, related to e.g. the pairwise correlations of stock prices, Accounting Networks capture the effects of business decisions on financial statements measures and on business models of different institutions. An ``ideal" investigation of the financial system would involve also a detailed analysis of the money flows among companies, which determine the so called ``mutual exposures" (an important contagion channel). Unfortunately, these high granular and detailed data are usually not available. However, financial statements provide an aggregated view of mutual exposures and obligations for different maturities and types of instrument. This is an important point in favour of Accounting Networks as they report summarised information for e.g. phenomena occurring with different time scales and contractual terms, as opposite to the financial networks that rely only on homogeneous (daily or intraday) market data.\\
\subsection*{Community Detection Results}
\noindent In this sub-section we focus our attention on the bottom up clusterization of the network from the application of the community detection algorithm and on the presence of geographical structures arising when we label each bank with its country. Therefore, we describe whether banks belonging to different countries (as a proxy for different regulations and/or level playing fields) have shown the tendency to be part of separate or, alternatively, common clusters and we verify, by analysing communities' evolution over time, whether the crisis influenced these configurations. In particular, our community detection analysis on Accounting Networks shows these main results.\\
It exhibits the presence over time of a clear community representing US banks and another one composed by Japanese banks, although for both regions there is also an additional smaller second group quite persistent in time. By contrast, it is not possible to identify a single and an unambiguous European community, since banks belonging to European countries seem to be likely to form national or sub-regional communities or to be included in a vast and geographically heterogeneous cluster (hereinafter the \textit{Mixed} community). In addition, Asian banks are fragmented in several sub-regions where, in particular, the Arab and the Indian-Pakistan groups emerge. Therefore, the detection of communities within Accounting Networks reveals the presence of two homogeneous clusters corresponding to US and Japanese banks surrounded by a more diversified cloud of banks belonging to different countries; remarkably, European banks are not able to clusterise together in a single community, while it persists over time a certain level of separation based also on national borders. Hence, an interesting contribution of the paper points to the presence of a large and geographically heterogeneous community, which can be related to the fact that the globally established regulatory framework might have indeed accelerated the tendency of banking activities of different countries to converge into more uniform banking practices. This is shown for instance in Figure \ref{figCD_PCA} where we also observe that the outbreak of financial markets contributed to make the Mixed community more cohesive; furthermore, although still representing separate communities, both US and JP clusters result topologically closer to the Mixed community after the breakdown of 2007-08, thus supporting the interpretation of a gradual convergence of different areas into more similar patterns. In addition, the application of the community detection on Accounting Networks allows to identify even small communities, such as those related to African or Scandinavian banks. This represents a quite promising aspect of the methodology, since it ensures the detection of local reliable communities although the approach taken so far is eminently agnostic.\\
It is not simple to explain the reasons behind the emergence and evolution of these communities, however it is possible to advance some intuitions based on the impact of globally recognized accounting standards (\cite{FASB}), the establishment of supranational supervisory and regulatory authorities, and on the role of the harmonization process of banking practices which have been implemented through e.g. the various Basel regulations (\cite{BASEL}). These contributions point to a common level playing field, which might have facilitated the emergence of a large and geographically heterogeneous community and its increasing topological proximity to both US and JP clusters. However, latter communities highlight the persistence of regional peculiarities. In Japan a deregulation process, known as the 'Japanese Big Bang', was formulated during the 1990s to transform the traditional bank-centered system into a market-centered financial system characterised by more transparent and liberalised financial markets (\cite{JP1}). In fact, peculiar features of Japanese banking sector were the over-reliance on intermediated bank lending, the absence of a sufficient corporate bond market and a marginal role for non-bank financial institutions, whose main consequences were an abundance of non-performing loans, excess in liquidity, scarce investments and low banks profitability (see e.g. \cite{BATTEN}). Although this program was intended to cover the period 1996-2001, the goals have not yet been achieved and policy makers' continuing reform efforts to remove past practices by market participants confirm the slowing convergence of the Japanese regulatory framework to a capital-market based financial system (\cite{JP2}). Thus, the presence of the JP community which gradually tends to the Mixed cluster is in line with evidences from the Japanese financial sector reforms aimed to change its reliance on indirect finance into a system of direct finance related to capital markets. Furthermore, it is remarkable the presence of a US community quite stable over time, which seems to be progressively attracted by the Mixed cluster. The US financial system presents peculiar features compared to other geographical areas. It is characterised by a relatively greater role of capital market-based intermediation, a higher importance of the 'shadow banking system', and differences in the accounting standards (\cite{ECB}). The impact of non-bank financial intermediation relates to the use of originate-to-distribute lending models, which determine the direct issuance of asset-backed securities and the transfers of loans to government-sponsored enterprises (GSEs, e.g. Fannie Mae and Freddie Mac). Financial innovation played a key role and the increasing use of securitisation explains the low percentage of loans to households on banks' balance sheets (\cite{ECB}). In addition, the US 'shadow banking system' is highly dependent on the presence of finance companies, money market funds, hedge funds and investment funds, which influenced the growth of total assets in the US financial sector during the last decades (\cite{shin2012global},\cite{shadow}). The presence of a distinct community is probably also due to differences in accounting standards which mainly involve the treatment of derivatives positions between the US Generally Accepted Accounting Principles (US GAAP) and the International Financial Reporting Standards (IFRSs). In particular, US GAAP allows to report the net value of derivative positions with the same counterparty under the presence of a single master agreement, thus impacting on the size representation of balance sheets items. However, in Figure \ref{figCD_PCA} we observe that the US community (similarly to the JP community) is gradually approaching the Mixed community, and the consequences of the breakdown of 2007-08 seem to have enhanced this behaviour. Among the possible several reasons, it is worthwhile to consider the impacts of the reform on the OTC derivatives market (embedded in the Dodd-Frank Act) and the Basel III new banking regulation, which may have facilitated similarities among US institutions and their peers in the Mixed cluster.
\subsection*{Relationships between Economic Indicators and Network Properties}
\noindent In this Section we provide a preliminary investigation of the relationships between banks' economic indicators and their network properties. In order to characterise banks, we consider three common proxies for their classification: \textit{Return on Assets} (for the \textit{Performance}), \textit{Total Assets} (for the \textit{Size}) and \textit{Total Debts to Total Assets} (for the \textit{Leverage}). Then, comparisons are presented against two basic network measures: the \textit{Strength} and the \textit{Clustering Coefficient}. For each year from 2001 to 2013, we provide some insights for these relationships by estimating for the overall sample the correlations between banks' economic indicators and network measures. As explained in the Method, in this exercise we consider the network filtered according to the quality/significance of the \textit{Louvain} community detection algorithm. This helps us in the assessment of the significance of our results. Below, we show some examples to discuss how these relationships have evolved over time.\\
\noindent In particular, we investigate whether once the effects of the crisis have spread throughout the financial sector, the capacity of traditional economic indicators (e.g. leverage, size, performance) to group banks could result undermined. For instance, the onset of the financial crisis clearly affects the relationships between \textit{Total Debts to Total Assets} and network properties. Although the correlation between \textit{Strength} and \textit{Total Debts to Total Assets} remains negative during the entire sample period, the breakdown of financial markets seems to further enhance this effect for subsequent years (Figure \ref{figCorrelations}, plot on the left). Thus, this relationship suggests that, after the onset of the crisis, the use of leverage became on average more anti-correlated to the \textit{Strength}. This implies that banks that are more dissimilar in terms of their financial statements (i.e. with lower values of \textit{Strength}) are those that turned out to be less capitalised (i.e. with higher values of \textit{Total Debts to Total Assets}). Furthermore, one might be interested in understanding the role played by the \textit{Size} which represents a typical indicator utilised to classify banks. The correlation between \textit{Strength} and \textit{Total Assets} is almost flat and negative even after the collapse of 2007-08, but it shows an increasing trend in the recent period (Figure \ref{figCorrelations}, plot on the middle). Hence, it seems that after the outbreak of the crisis the \textit{Size} became less correlated to the similarity among banks, as estimates pointing sharply to zero seem to suggest. We finally analyse the relationship between \textit{Performance} and network properties (Figure \ref{figCorrelations}, plot on the right). In particular, in order to mimic how the presence/absence of more connected groups of banks is related to economic results we consider the \textit{Clustering Coefficient} for determining the level of structure in the system. Although poorly statistically significant in the early 2000s, correlations with \textit{Return on Assets} exhibit a decreasing pattern before the onset of the crisis and then remain negative although slightly erratic. The negative relationship between \textit{Clustering Coefficient} and \textit{Return on Assets} seems to suggest that the presence of well connected areas in the network (nodes with higher clustering coefficients) do not foster economic performance.\\
\noindent These basic examples suggest that a clear investigation on the relationships between economic indicators and network properties might be not always conclusive. Moreover, once we consider the entire set of banks, there might be some cases where estimates are poorly significant. Still, some remarkable effects arise from this investigation strategy and preliminary results point to a turning point in the correlations across the outbreak of the financial crisis. In particular, diagrams confirm that leverage is an useful indicator for differentiating banks, hence deviations to a lower capitalization are associated to increasing dissimilarity with the rest of the system and the impact of the crisis suggests a reinforcement in this relationship. By contrast, it seems that size does not contribute too much on the similarity between banks after the breakdown of 2007-08, while it played a greater role before and during the crisis. Finally, the relationship between performance and the structure of the system is less clear and prevents straightforward conclusions.\\
The identification of economic features potentially able to characterise specific portions of the system is addressed in the next sub-section.
\subsection*{PCA results}
Community detection shows the presence of three large clusters (Mixed, US, and JP) and an additional quite stable and persistent but smaller community (mostly US+EU banks). In this Section we provide a way to describe how these communities can be represented in terms of economic features (see Figure \ref{figCD_PCA}). Given the multi-dimensionality of the set of measures utilised to build the networks, we adopt a Principal Component Analysis approach to identify those measures which contribute more (less) to the explained variance within each community. For the sake of simplicity, we propose the ranking of the top (bottom) three measures for each community during the following intervals: pre-crisis ($2001-2006$), crisis ($2007-2009$), and post-crisis ($2010-2013$). In particular, for each year we compute the contribution of the original measures to explained variance; then, we average within each sub-period and we determine the rankings based on the mean period values. Below, we name the community with a mixed geographical composition as \textit{C0}, while we refer to the communities with a prevalence of US, JP and European plus US banks as \textit{C1}, \textit{C2} and \textit{C3}, respectively.\\
This representation allows us to compare communities' features over time and across different groups. For instance, we observe that \textit{Total Assets} and \textit{Interest Income} are quite frequent among top measures contributors, while \textit{Total Debts to Total Assets} is recurrent among measures in the bottom rankings. This is not surprising given banks heterogeneity in terms of the size (\textit{Total Assets}) and the economic results (\textit{Interest Income}) distributions, in contrast with the tight constraints on leverage (\textit{Total Debts to Total Assets}) due to regulatory requirements. By focusing on the top rankings we notice that \textit{C0} and \textit{C1} have fairly stable top contributors, while communities \textit{C2} and \textit{C3} are more affected by the wave of financial turmoil. Furthermore, bottom rankings seem to be on average only slightly influenced by the choice of different sub-periods. In addition, differences between mean values among the set of top three and the set of bottom three contributors are quite stable over time with only few exceptions, while the middle part of the distribution of measures' contributions (not reported, available from authors upon request) is in general quite sparse. For these reasons, we prefer to focus on the top and the bottom rankings to describe communities' features.\\
One might be interested in how the outbreak of financial crisis have influenced these rankings. Top composition of \textit{C1} is unaffected by the 2007-08 financial breakdown, while \textit{C0} is only partially modified by the onset of the crisis (\textit{Interest Income} is replaced by \textit{Net Interest Income}). Conversely, \textit{C2} presents a quite different configuration during the crisis sub-period when it exhibits a relevant role for expenses measures (i.e. Non Interest Expenses and Operating Expenses). Similarly, income statement measures become more relevant among top contributors also within the \textit{C3} community. Interestingly, community \textit{C0}, which is characterised by a mixed geographical composition, and the US community (\textit{C1}) reach identical top contributors after the outbreak of 2007-08, while the JP community (\textit{C2}), which shows the same top contributors as community \textit{C0} in the first sub-period, seems to react differently during the crisis, although in the third sub-period it shows again top contributors similar to \textit{C0} (and to \textit{C1}). By contrast, community \textit{C3} seems to present a peculiar pattern over time.\\
Therefore, the crisis sub-period coincides with remarkable differences in the top contributors, while the recent sub-period points to a renewed tendency to get similar contributors for a wider set of banks (\textit{C0} and \textit{C1}, and partially \textit{C2}). This seems to be in line with the above discussion on community detection results, where we highlighted a gradual proximity between clusters over time. Hence, these results suggest that heterogeneity within clusters is driven by similar economic measures after the crisis, although specific differences persist. This is the case for instance of loans, which are not present among top contributors in the US community while they are in the top ranking of both the Mixed and the JP community (as expected according to the above discussion). We also notice that the crisis seems to suggest an increasing importance of income statement measures in terms of contribution to the explained variance within communities.
The breakdown of financial markets affected banks' results and this justifies the high level of heterogeneity expressed by income statements indicators. This can also be related to the impact of the crisis on financial statement measures and on the different ways banks update their balance sheet structures compared to the recognition of economic results as reported in the income statements items. Similar comparisons can involve also the bottom three measures, but for conciseness we omit this part.
\section*{Discussion}
\noindent In this paper, we depict the banking system through banks' financial statements. Our main contribution is represented by the introduction of a methodology to exploit balance sheets and income statements data to construct Accounting Networks. We show some relationships between economic indicators and network properties, which might provide some new useful insights for banking classification practices. Having depicted some effects of the recent financial crisis by using a simple framework is an encouraging sign for further extensions. We rely on ``neutral'' and ``naive'' techniques to build the Accounting Networks. In particular, among common approaches usually applied to describe similarities concepts, we adopt one of the basic method, i.e. the cosine similarity. Future works can exploit more advanced methodologies. Moreover, our selection of variables utilised to compute cosine similarities assumes that each component has the same importance. This is quite a naive hypothesis, which could be enriched by measures discrimination based on economic literature and/or practitioners practices. Finally, for accounting reasons we limit our study on annual financial statements, while a more detailed description of the system might easily involve the use of quarterly data. Despite these simplifying assumptions, our approach has the merit of introducing a novelty in the debate on banking networks, and we believe that future improvements in the directions outlined above will enforce Accounting Networks' ability to describe the evolution of banking systems.
\section*{Acknowledgments}
In this paper we thank the financial support of the Italian project CRISISLAB and the support of the Linkalab Laboratory for its open discussions and precious suggestions.
|
\section{Introduction}
Although not well understood, the phenomenon of Li-rich giants seems to be ubiquitous as they have been observed in different environments: open clusters, globular clusters, metal-rich and metal-poor field stars, the Galactic bulge, and also in dwarf galaxies \citep[see, e.g.,][and references therein]{1999A&A...348L..21H,2009A&A...508..289G,2011ApJ...730L..12K,2011ApJ...743..107R,2012ApJ...752L..16K,2016ApJ...819..135K,2015ApJ...801L..32D}.
Lithium-rich giants are usually defined as those that, after the first dredge-up, have A(Li) $\geq$ 1.50 dex. This limit is the post-dredge-up Li abundance of a low-mass star according to standard evolutionary models, i.e., those models that include only convection as a mixing mechanism. The first Li-rich giant was a fortuitous discovery by \citet{1982ApJ...255..577W}. Subsequent searches have shown that these stars comprise about 1--2\% of red giants \citep{1989ApJS...71..293B,2000AJ....119.2895P}. \citet{2000A&A...359..563C} suggested that these objects appear at the luminosity bump of the red giant branch (RGB) or at the early-asymptotic giant branch (AGB) for low- and intermediate-mass stars, respectively. Other recent results preferably classify these objects as core helium-burning giants \citep{2011ApJ...730L..12K,2014A&A...564L...6M,2014ApJ...784L..16S}. Nevertheless, some Li-rich giants have been found throughout the RGB \citep[see, e.g.,][]{2011A&A...531L..12A,2011A&A...529A..90M,2013MNRAS.430..611M}.
\begin{figure}
\centering
\includegraphics[height = 8.5cm]{trumpler20_spectra.eps}
\caption{Comparison of the spectra around the Li 6708 \AA\ line between the Li-rich giants and giants with similar atmospheric parameters.}\label{fig:spec}
\end{figure}
Lithium-rich giants have other noteworthy characteristics that add complexity to the puzzle. Some present a far-infrared excess, suggesting a connection with enhanced mass loss \citep{1996ApJ...456L.115D}. This mass-loss event can also explain the observation of complex organic and inorganic compounds detected in the infrared spectra of some Li-rich giants \citep{2015ApJ...806...86D}. In a few cases, the presence of circumstellar material has been confirmed by polarimetry \citep{2006A&A...449..211P}. Nevertheless, not all Li-rich giants have an infrared excess \citep[see, e.g.,][]{1999A&A...342..831J,2015A&A...577A..10B,2015AJ....150..123R}.
\citet{1993ApJ...403..708F} proposed a connection between Li enrichment, fast rotation, and chromospheric activity. Lithium-rich giants seem to be more common among fast rotating stars \citep[$\sim$ 50\%; see, e.g.,][]{2002AJ....123.2703D}. A strong magnetic field was detected in one Li-rich giant by \citet{2009A&A...504.1011L}. Further examples of Li-rich, fast-rotating, active giants exist \citep[e.g.,][]{2002AJ....123.1993R,2014A&A...571A..74K,2015A&A...574A..31S}.
In addition, a few Li-rich giants hosting planets have been found \citep[e.g.,][]{2012ApJ...754L..15A,2014A&A...569A..55A}. As proposed by \citet{1999MNRAS.308.1133S}, surface Li enrichment could be caused by planet engulfment, which also causes spin-up, magnetic field generation, and shell ejection. However, planet accretion would create a $^{9}$Be enhancement that has never been detected in Li-enriched objects \citep{1997A&A...321L..37D,1999A&A...345..249C,2005A&A...439..227M,2014A&A...563A...3P} with the exception of one F-type dwarf in the open cluster NGC 6633 \citep{2005MNRAS.363L..81A}. Alternatively, planet engulfment could activate internal Li production and induce its mixing to the surface \citep[][]{2000A&A...358L..49D}.
Indeed, the properties of many Li-rich giants discovered within the Gaia-ESO Survey \citep{2012Msngr.147...25G,2013Msngr.154...47R} seem to be consistent with those of giants that engulfed close-in giant planets before evolving up the RGB \citep{2016arXiv160303038C}. However, a small fraction of cases still require alternative explanations. Here, we report the discovery of two Li-rich giants that could be examples of such an alternative formation channel in the open cluster \object{Trumpler 20}, which is a system of $\sim$ 1.66 Gyr in age and [Fe/H] = +0.17 \citep[][]{2014A&A...561A..94D}.
\begin{figure}
\centering
\includegraphics[height = 7cm]{cmd_tr20_donati.eps}
\caption{Color magnitude diagram of Trumpler 20. Only stars within 3\arcmin of the cluster center are shown. The two Li-rich giants are shown as filled circles, other giants with Li detections as open circles, giants with Li upper limits as gray triangles, and the remaining stars in the field as dots. The solid line is an isochrone from \citet{2012MNRAS.427..127B} with age = 1.66 Gyr and [Fe/H] = +0.17, which is the best fit to the photometric data as determined by \citet{2014A&A...561A..94D}.}\label{fig:cmd}
\end{figure}
This paper is organized as follows. In Sect.\ \ref{sec:data} we briefly describe the data used here, the analysis, and the properties of our sample. In Sect.\ \ref{sec:extra} we discuss how extra mixing is needed to explain the surface Li abundances of the majority of the sample. In Sect.\ \ref{sec:noextra} we discuss the two Li-rich giants and the possibility that they have avoided extra-mixing mechanisms. Finally, Sect.\ \ref{sec:end} summarizes our findings and suggests new observations that could support our interpretation of the Li enrichment in these two giants.
\section{Data, analysis, and sample properties}\label{sec:data}
The high-resolution (R $\sim$ 47\,000) UVES \citep[Ultraviolet and Visual Echelle Spectrograph,][]{2000SPIE.4008..534D} spectra of 42 targets in Trumpler 20 were obtained in the context of the Gaia-ESO Survey. Data reduction is described in \citet{2014A&A...565A.113S}. Basic information on the observed giants is available online in Table \ref{tab:obs}.
The atmospheric parameters and abundances (see online Table \ref{tab:par}) are part of the fourth Gaia-ESO internal data release (iDR4). The spectra were analyzed using the Gaia-ESO multiple pipelines strategy \citep{2014A&A...570A.122S} with an updated methodology (Casey et al. 2016b, in preparation).
Membership was assigned using the radial velocities (RVs) as in \citet{2014A&A...561A..94D}. Likely members (40 giants in total) are those with RV within three standard deviations of the cluster average ($\overline{RV}$ $\pm \ \sigma$ = $-$40.2 $\pm$ 1.3 km s$^{-1}$). One star is a subgiant\footnote{Trumpler 20 MG 430. The numbering system we adopt is that defined in \citet{2005ApJS..161..118M}.} and one a probable nonmember (or binary) with deviant RV\footnote{Trumpler 20 MG 894 with RV = $-$35.3 km s$^{-1}$.}.
The Li abundances were determined from the 6708 \AA\ line. In Fig.\ \ref{fig:spec}, we compare the Li 6708 \AA\ lines of the two Li-rich giants to those of stars with similar atmospheric parameters to illustrate the Li enhancement. Corrections for nonlocal thermodynamical equilibrium (non-LTE) effects were applied using the grid of \citet{2009A&A...503..541L}. For the giants, the corrections range from 0.14 dex to 0.32 dex, depending on the atmospheric parameters.
The color magnitude diagram (CMD) of Trumpler 20 is shown in Fig.\ \ref{fig:cmd}. The photometry is originally from \citet{2010AJ....140..954C} corrected for differential reddening by \citet{2014A&A...561A..94D}. The uncertainties in the magnitudes are $\sim$0.02-0.05 mag. A noticeable feature in this CMD is the extended clump region of the cluster. Trumpler 20 is well known for its peculiar extended clump region \citep[see][]{2010AJ....140..954C,2012ApJ...751L...8P,2014A&A...561A..94D}. This feature is probably caused by the presence of two distinct clumps; the fainter clump comprises stars massive enough to start core He-burning in nondegenerate conditions and the brighter clump comprises stars with slightly lower mass that have been through the He-core flash \citep[see, e.g.,][]{1999MNRAS.308..818G,2000A&A...354..892G}.
Figure \ref{fig:tefflogg} shows the sample in the $T_{\rm eff}$-$\log~g$ plane. The group of giants with lower $\log~g$ are either at the luminosity bump of the RGB or at the early-AGB, as both stages are within the error bars of the parameters in Fig.\ \ref{fig:tefflogg} and are hard to separate in the CMD of Fig.\ \ref{fig:cmd}. We can be more confident about the evolutionary state of the group of giants with higher $\log~g$ because of their chemical abundances.
The evolutionary stage of the stars is an important source of information to interpret their Li abundances, as a high Li abundance could just be indicating that the giant is actually at the bottom of the RGB \citep[e.g.,][]{2015MNRAS.446.3562B}. Nevertheless, the C and N abundances of the giants demonstrate that they have all completed the standard first dredge-up. As shown in Fig. \ref{fig:cnteff}, according to the models of \citet{2012A&A...543A.108L}, giants of 1.5 and 2.0 M$_{\odot}$ after the dredge-up have C/N $\sim$ 1, as do all the giants in our sample. Giants at the bottom of the RGB with $T_{\rm eff}$ $\sim$ 5000 K would be in the stage before the end of the first dredge-up and thus would instead have C/N > 3. We can thus safely conclude that i) all the giants with $T_{\rm eff}$ $\sim$ 5000 K are in the red clump and not on the RGB and ii) that all the brightest and coolest giants have completed the Li dilution expected during the first dredge-up. The C and N abundances of the Trumpler 20 giants were discussed in \citep[][]{2015A&A...573A..55T}.
\begin{figure}
\centering
\includegraphics[height = 7cm]{teff_logg_iso_tr20_dr4.eps}
\caption{Trumpler 20 giants in the $T_{\rm eff}$-$\log g$ plane. The two Li-rich giants are shown as filled circles, other giants with Li detections as open circles, and giants with Li upper limits as gray triangles. The solid line is an isochrone from \citet{2012MNRAS.427..127B} of 1.66 Gyr in age and [Fe/H] = +0.17, which is the best fit to the photometric data \citep[by][]{2014A&A...561A..94D}. A typical error bar ($\pm$ 120 K and $\pm$ 0.23 dex for $T_{\rm eff}$ and $\log g$, respectively) is shown in the top left.}\label{fig:tefflogg}
\end{figure}
\section{Extra mixing in the majority of the giants}\label{sec:extra}
Lithium abundances have been extensively used as a tracer of mixing processes, as Li is rapidly destroyed in (p,$\alpha$) reactions at temperatures above 2.5 $\times 10^{6}$ K \citep{1951ApJ...113..536G}. Thus, Li only survives in the outermost layers of a star. As stars evolve to the RGB, the convective envelope deepens and Li from the surface is diluted.
Figure \ref{fig:liteff} shows the Li abundances as a function of the effective temperatures (T$_{\rm eff}$), for all cluster members, in comparison with the models of \citet{2012A&A...543A.108L}. The bulk of the stars fall in between the standard and extra-mixing models (solid and dashed lines, respectively). However, care is needed in interpreting the plot because of the evolutionary state of the stars.
\begin{figure}
\centering
\includegraphics[height = 7cm]{cn_old_teff_dr4_tr20.eps}
\caption{C/N ratio as a function of T$_{\rm eff}$. Solid lines are the predictions of standard models and dashed lines of models with rotation-induced mixing and thermohaline mixing \citep{2012A&A...543A.108L}. Lines in blue and red are for solar metallicity stars of 1.5 M$_{\odot}$ and 2.0 M$_{\odot}$, respectively. The two Li-rich giants are shown as full circles.}\label{fig:cnteff}
\end{figure}
The solid and dashed lines in Fig.\ \ref{fig:liteff} are predictions for first ascent RGB stars and not for clump giants. As we showed above, our giants with $T_{\rm eff}$ $\sim$ 5000 K are clump giants and not first ascent RGB stars. The observations should not be compared to this region of the models, but rather to the A(Li) level of core-He burning giants (the dotted lines). The clump giants have Li upper limits on average of about 0.3 dex below the two top dotted lines (standard models). For lower temperatures, the second group of stars has also Li upper limits well below the prediction of the standard models.
The clear exceptions to that are the two Li-rich giants, Trumpler 20 MG 340 and 591. This agrees with the findings of \citet{1989ApJS...71..293B} that giants with A(Li) $\sim$ 1.50 in agreement with standard models are a minority. Here, we are able to confirm this result in a large sample of giants of the same age, same initial chemical composition, and very similar masses. The enhanced Li depletion/dilution seen in the majority of the giants of Trumpler 20 is well documented in the literature \citep[e.g.,][]{1999A&A...345..936L,2001A&A...374.1017P,2004A&A...424..951P}, although the mechanism behind this extra mixing is still under debate.
We note another possible outlier, star MG 505, with A(Li)$_{\rm non-LTE}$ = 1.25 $\pm$ 0.21. However, because within the errors its abundance agrees with the highest upper limit at its temperature, we do not consider it among those that agree with the standard models.
\section{Inhibited extra mixing in two giants}\label{sec:noextra}
Star MG 340 has A(Li)$_{\rm non-LTE}$ = 1.54 $\pm$ 0.21 and T$_{\rm eff}$ = 4851 K, while in eight other stars with T$_{\rm eff}$ = 4850 $\pm$ 50 K there is one detection at A(Li)$_{\rm non-LTE}$ = 1.25 and seven upper limits below A(Li)$_{\rm non-LTE} \sim$ 1.05. Star MG 591 has A(Li)$_{\rm non-LTE}$ = 1.60 $\pm$ 0.21 and T$_{\rm eff}$ = 4412 K, while seven other stars with T$_{\rm eff}$ < 4600 K have upper limits below A(Li)$_{\rm non-LTE} \sim$ 0.76.
These two Li-rich giants do not show any additional chemical peculiarity when compared to the other cluster giants. Both stars seem to be single, but we do not have multiple epoch spectra to exclude long period companions. All sample giants seem to be slow rotators: $v~\sin~i < 4.0$ km s$^{-1}$. Therefore, the Li enhancement is probably not connected to rotation in the sense seen by \citet{2006A&A...450.1173L} and \citet{2012ApJ...757..109C}. These authors found that in a given sample of giants, those with higher Li abundance tend to be those giants that are rotating faster, however, our giants might seem to be slow rotators because of an unfavorable line of sight. Slow rotation also argues against, but does not fully exclude, planet accretion with transfer of angular momentum as the source of the Li enhancement \citep[see also][]{2016ApJ...818...25C,2016A&A...587A..66D}. Thus, external pollution as advocated by \citet{2016arXiv160303038C} to explain other Li-rich giants discovered within the Gaia-ESO Survey seems unlikely in our case.
\begin{figure}
\centering
\includegraphics[height = 7cm]{ali_teff_tr20_dr4.eps}
\caption{Lithium abundance as a function of T$_{\rm eff}$. The two Li-rich giants are shown as filled circles, other giants with Li detections as open circles, and giants with Li upper limits as gray triangles. Solid lines are the predictions of standard models and dashed lines of models with rotation-induced mixing and thermohaline mixing \citep{2012A&A...543A.108L}. Lines in blue and red are for solar metallicity stars of 1.5 M$_{\odot}$ and 2.0 M$_{\odot}$, respectively. The dotted lines are included as an eye guide to the Li abundance level of clump giants in the models (but we note here that the clump phase does not span the T$_{\rm eff}$ range of the dotted lines in the plot).}\label{fig:liteff}
\end{figure}
Internal Li production was also suggested to explain Li-rich giants. Fresh Li production might occur in the stellar interior through the $^{7}$Be mechanism \citep{1971ApJ...164..111C}. However, it is still unknown which transport mechanism would bring $^{7}$Be, which decays to $^{7}$Li, to the surface \citep{1999ApJ...510..217S,2001A&A...375L...9P}. \citet{2000A&A...359..563C} argued that Li-rich giants were preferentially found at the bump and the early-AGB, and connected the Li enrichment with an extra-mixing process that activates at these evolutionary stages. For low-mass stars at the bump, the extra mixing would be connected to the H-burning shell that is moving outward and removes the molecular weight discontinuity left by the receding convective layer. In intermediate-mass stars, the extra mixing would take place at the early-AGB when the convective envelope deepens again.
The two Li-rich clump giants found by \citet{2014ApJ...784L..16S}, with A(Li) = 2.71, and \citet{2014A&A...564L...6M}, with A(Li) = 3.75, showed that the above scenario is at least incomplete. For clump giants, the Li enrichment could be connected to the He flash, following an episode of H injection in deeper high temperature regions \citep{2011ApJ...743...55M}.
Because of the large surface convective layers of giants, the observed Li surface enrichment is likely to be short lived. A Li-rich low-mass giant that appears at the bump should not remain Li rich during its evolution to the clump. More likely, the two Li-rich giants discovered here have been freshly created during or close to their current evolutionary stages. Thus, to explain our Li-rich giants with internal Li production would require a combination of the two distinct scenarios above. One Li-rich giant would be created by mixing induced by the He flash, the other would be created by extra mixing at the bump. It also seems an odd coincidence that we would happen to observe both Li-rich giants at the moment in which their Li abundances are very close to the value expected by standard models.
While internal Li production could indeed be required to explain abundances above the meteoritic value (A(Li) $>$ 3.0), this is perhaps not necessary to explain stars MG 340 and 591. We instead suggest that their higher Li abundance is the result of the inhibition of extra-mixing processes. Without extra mixing, their surface Li abundance is at the level predicted by standard models. In addition, this single scenario would be able to explain both giants regardless of their different evolutionary states.
If this suggestion is correct, two instances of extra mixing must have been affected. The first is the extra Li dilution beyond the predicted first dredge-up dilution, which is observed as the star leaves the main sequence toward the RGB. Observationally, it is well known that an extra mixing causes the Li dilution to start earlier than predicted by standard models \citep[e.g.,][]{1999A&A...345..936L,2001A&A...374.1017P,2004A&A...424..951P}. This is because rotation-induced mixing creates a Li-free region inside real stars that is larger than predicted by these models \citep[]{2003A&A...399..603P}. In the two Li-rich giants, rotation-induced mixing must have been weak and the Li dilution proceeded as expected by standard models.
The second extra-mixing event to be avoided is the event taking place at the luminosity bump of the RGB \citep[see][and references therein]{2012A&A...543A.108L}. The luminosity bump \citep[see, e.g.,][]{2015MNRAS.453..666C} happens at the RGB of low-mass stars when the hydrogen-burning shell reaches the composition discontinuity left behind by the first dredge-up. The current best candidate for the mechanism behind this extra mixing seems to be thermohaline mixing \citep{2007A&A...467L..15C}, although there are discussions about the physical properties and efficiency of this mechanism \citep[e.g.,][and references therein]{2013A&A...553A...1M,2015ApJ...815...42G,2015MNRAS.446.2673L}.
Star MG 340 in the extended clump of the cluster is likely to be a low-mass star. As discussed by \citet{1999MNRAS.308..818G}, in such an extended clump, the less massive stars are actually the brighter stars. Both the CMD in Fig.\ \ref{fig:cmd} and the $T_{\rm eff}$-$\log~g$ diagram in Fig. \ref{fig:tefflogg} seem to indicate that MG 340 belongs to the group of brighter giants. Thus, in our scenario, for it to keep an unaltered Li surface abundance, thermohaline mixing must have been inhibited. On the other hand, star MG 591 is either on the bump or on the early-AGB. If on the early-AGB, then it is an intermediate-mass star that does not go through both the He-flash and the bump phase. However, it would still need to inhibit thermohaline mixing at the early-AGB \citep[see][and references therein]{2012A&A...543A.108L}. If on the bump, it either avoided thermohaline mixing or did not activate it yet. All other giants with similar $T_{\rm eff}$ and $\log g$ have lower Li abundances. This could indicate that star MG 591 is also after the moment where thermohaline mixing becomes efficient. As pointed out by the referee, however, at this phase stars ascend and reascend the RGB, crossing the same $T_{\rm eff}$ and $\log g$ region three times. Therefore, it is plausible that at least one star among the group at the bump has not yet activated thermohaline mixing. If this is the case, star MG 591 would not be an Li-rich giant, but a normal giant in a stage before extra mixing was activated.
Extra-mixing inhibition is not a new idea. Based on carbon isotopic ratios, $^{12}$C/$^{13}$C, \citet{1998A&A...336..915C} estimated that about 4\% of low-mass giants do not experience extra mixing on the RGB. \citet{2007A&A...476L..29C} suggested that extra mixing is avoided by giants that are descendant from Ap-type main-sequence stars. In these stars, fossil magnetic fields would be able to inhibit thermohaline mixing. Modern estimates of the percentage of Ap stars with respect to nonmagnetic A-type stars are between 1.7-3.5\% \citep{1993ASPC...44..577N,2007pms..conf...89P}. This is consistent with finding one or two stars in our sample of 40 giants of Trumpler 20 (a fraction of 2.5 or 5\%). In addition, the stellar mass of our giants is within the mass range of Ap stars \citep[$\sim$ 1.5 to 3.6 M$_{\odot}$,][]{2007pms..conf...89P}.
\section{Summary}\label{sec:end}
In this work, we presented the discovery of two Li-rich giants in the open cluster Trumpler 20. These two stars were identified in an analysis of a sample of 40 giants for which high-resolution spectra were obtained with UVES in the context of the Gaia-ESO public spectroscopic survey. This provides a unique large sample of giants that have the same age, same initial chemical composition, and very similar masses. The Li abundances in this sample clearly demonstrate that extra mixing is the norm in stars in this mass range. Giants with Li abundances in agreement with the predictions of standard models are the exception.
To explain the two Li-rich giants, we suggest that all instances of extra-mixing processes have been inhibited. Because of that, the surface Li abundance in these two stars remained at the level predicted by standard stellar evolution models, i.e., A(Li) $\sim$ 1.50. We argue that the fraction of Li-rich giants found in our sample is consistent with these giants being evolved counterparts of magnetic Ap-type dwarfs. In this case, the extra-mixing processes could have been inhibited by the action of magnetic fields, as suggested by \citet{2007A&A...476L..29C}.
Other explanations seem less likely, although they cannot be fully excluded. Extra Li from the accretion of external material should be accompanied by accretion of angular momentum, but there is no evidence of fast rotation in the two giants. Because the two Li-rich giants have different evolutionary stages, internal Li production would require two different mechanisms to bring the fresh Li to the surface. The extra-mixing inhibition hypothesis would instead be able to explain both giants at the same time.
Additional observations could help in providing extra support to our suggested scenario, or they could help to disprove it. First, if no extra mixing took place, the carbon isotopic ratio should be close to the prediction of standard models, $^{12}$C/$^{13}$C $\sim$ 30. We could not determine the carbon isotopic ratio because the region around 8000\AA, containing the CN bands preferred for this type of analysis in metal-rich giants, is not part of the Gaia-ESO spectra.
Second, there should likely be signs of magnetic activity in the Li-rich giants. Other candidates of Ap-type stars descendants were identified among giants and subgiants with magnetic activity \citep[e.g.,][]{2014psce.conf..444A}. However, fossil magnetic fields beneath the surface are hard to detect \citep{2015A&A...574A..90A,2016Natur.529..364S}. Nevertheless, the core-He burning star MG 340 is at one of the evolutionary phases where magnetic activity in giants is observed \citep{2015A&A...574A..90A}. We checked the H$\alpha$ line in our spectrum, but it shows no evident sign of activity. This is not inconsistent, as not all active giants display emission in H$\alpha$ \citep[see, e.g.,][]{1993ApJ...403..708F}. Emission should be clearer in the Ca H and K lines or in the UV, which are not part of our spectra.
Finally, if such a scenario of extra-mixing inhibition is correct, it would likely apply to many, if not all, Li-rich giants with A(Li) $\sim$ 1.50 (or slightly higher) and masses between $\sim$ 1.5 and 3.6 M$_{\odot}$. They would not have experienced fresh Li production, but would instead have preserved part of their original Li abundance.
\begin{acknowledgements}
We thank the anonymous referee for his/her suggestions and very fast report. R.S. acknowledges support by the National Science Center of Poland through grant 2012/07/B/ST9/04428. S.V. gratefully acknowledges the support provided by FONDECYT reg. n. 1130721. G.T. acknowledges support by the Research Council of Lithuania (MIP-082/2015). D.G. gratefully acknowledges support from the Chilean BASAL Centro de Excelencia en Astrof\'{\i}sica y Tecnolog\'{\i}as Afines (CATA) grant PFB-06/2007. S.G.S. acknowledges the support from FCT through Investigador FCT contract of reference IF/00028/2014. E.D.M. acknowledges the support from FCT in the form of the grant SFRH/BPD/76606/2011. S.G.S. and E.D.M. also acknowledge the support from FCT through the project PTDC/FIS-AST/7073/2014. Based on data products from observations made with ESO Telescopes at the La Silla Paranal Observatory under program ID 188.B-3002. These data products have been processed by the Cambridge Astronomy Survey Unit (CASU) at the Institute of Astronomy, University of Cambridge, and by the FLAMES/UVES reduction team at INAF/Osservatorio Astrofisico di Arcetri. These data have been obtained from the Gaia-ESO Survey Data Archive, prepared and hosted by the Wide Field Astronomy Unit, Institute for Astronomy, University of Edinburgh, which is funded by the UK Science and Technology Facilities Council. This work was partly supported by the European Union FP7 program through ERC grant number 320360 and by the Leverhulme Trust through grant RPG-2012-541. We acknowledge the support from INAF and Ministero dell' Istruzione, dell' Universit\`a' e della Ricerca (MIUR) in the form of the grant ``Premiale VLT 2012'' and ``The Chemical and Dynamical Evolution of the Milky Way and Local Group Galaxies''. The results presented here benefit from discussions held during the Gaia-ESO workshops and conferences supported by the ESF (European Science Foundation) through the GREAT Research Network Programme.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
One of the key challenges in stellar astrophysics is to understand
the formation of massive stars and their role in the evolution of
the universe. A lot of research has been done in astrophysics in the
last two decades and considerable progress has been seen in the
understanding of massive star formation (Zinnecker $\&$ Yorke 2007).
Even though their short lives and rare sources, massive stars play a
major role in determining physical, chemical and morphological
structures of galaxies through ionizing radiation, heating of dust
and expansions of their HII regions. These processes may also
trigger next generation of star formation by compressing neighboring
molecular clouds to the point of gravitational instability. It is
generally accepted that massive stars evolve through starless cores
in infrared dark clouds (IRDCs) to hot molecular cores (HMCs; e.g.
Kurtz et al. 2000). Then substantial UV photons and ionized stellar
winds rapidly ionize the surrounding hydrogen, forming a
hyper-compact HII region (HCHII) or ultra-compact HII region
(UCHII). The final stages are compact and classical HII regions.
However, many questions are still unclear: whether accretion could
be halted by the strong outward radiation and thermal pressure? Does
it continue in an ionized form? Does it continue through a molecular
or ionized disk? Chemical composition of molecular gas is thought to
evolve due to the physical changes that occur during the star
formation process. As material collapses and gets ionized by central
young stars, densities and temperature rise, leading to the
production and destruction of different molecular species.
Multi-wavelength observations are vital to our understanding of the
physical and chemical conditions where massive star formation
occurs.
IRDCs provide us with the possibility of investigating the
subsequent early stages of high mass star formation, and a lot of
research has also been done to investigate their physical and
chemical properties (e.g. Vasyunina et al. 2009; Vasyunina et al.
2011; Hoq et al. 2013; Liu et al. 2013; Miettinen 2014). In order to
characterize the different evolutionary stages of IRDCs, Chambers et
al. (2009) proposed an evolutionary sequence in which ``quiescent''
clumps evolve into ``intermediate'', ``active'' and ``red'' clumps,
based on the Spitzer observations. Sanhueza et al. (2012) and Hoq et
al. (2013) characterize the chemical properties of a sample of IRDCs
using the data of MALT90. Correlations and anti-correlations between
molecular abundances and the evolutionary stages have been found.
These studies suggest chemical processes in high-mass star
formations may differ from their low-mass counterparts, as their
temperature, densities and UV fluxes are very different from each
other. Establishing the chemical properties of massive star
formation requires that large surveys be conducted.
\begin{table*}
\begin{minipage}{13cm}
\caption{\label{tab:test}List of our sources.}
\begin{tabular}{lclclclclclclcl}
\hline
\hline
RMS & RA & Dec. & V$_{lsr}$ &$D$ & $S$$_{4.8GHz}$$^a$ & log($N_L$) & \\
name & (J2000) & (J2000) &(km s$^{-1}$) & (kpc) & (mJy) & (S$^{-1}$) & \\
\hline
G330.2838+00.4933 & 16:03:43.28 &-51:51:45.6 & -94.2 &5.32$\pm$0.06 & 44.9 &47.05$\pm$0.05 \\
G331.4181$-$00.3546 & 16:12:50.24 &-51:43:28.5 & -63.2 &3.90$\pm$0.05 & 83.9 &46.78$\pm$0.06 \\
G332.2944$-$00.0962 & 16:15:45.83 &-50:56:02.4 & -48.7 &3.64$\pm$0.03 & 175.6 &47.30$\pm$0.04\\
G332.5438$-$00.1277 & 16:17:02.33 &-50:47:03.3 & -47.1 &3.64$\pm$0.06 & 42.2 &46.69$\pm$0.07\\
G332.8256$-$00.5498A & 16:20:11.06 &-50:53:16.2 & -58.1 &3.64$\pm$0.05 & 1259.0 &48.16$\pm$0.07\\
G333.3401$-$00.1273 & 16:20:37.00 &-50:13:32.7 & -59.7 &3.64$\pm$0.03 & 3.3 &45.58$\pm$0.04\\
G337.0047+00.3226 & 16:34:04.73 &-47:16:29.2 & -62.4 &11.43$\pm$0.05 & 249.3 &48.46$\pm$0.02\\
G337.4032$-$00.4037 & 16:38:50.44 &-47:28:03.0 & -40.0 &3.14$\pm$0.04 & 90.3 &46.89$\pm$0.06\\
G337.7091+00.0932A & 16:37:51.80 &-46:54:33.4 & -75.9 &10.98$\pm$0.04 & 144.9 &48.19$\pm$0.02\\
G337.8442$-$00.3748 & 16:40:26.68 &-47:07:13.1 & -39.9 &3.06$\pm$0.06 & 11.1 &45.95$\pm$0.09\\
G339.1052+00.1490 & 16:42:59.58 &-45:49:43.6 & -78.2 &4.73$\pm$0.06 & 42.9 &46.93$\pm$0.03\\
G340.0708+00.9267 & 16:43:15.69 &-44:35:16.0 & -73.5 &4.69$\pm$0.04 & 93.8 &46.96$\pm$0.04\\
G343.5024$-$00.0145 & 16:59:20.78 &-42:32:37.5 & -27.9 &2.72$\pm$0.08 & 176.9 &47.06$\pm$0.12\\
G345.0034$-$00.2240A & 17:05:11.19 &-41:29:06.3 & -27.3 &2.87$\pm$0.08 & 140.0 &46.99$\pm$0.12\\
G345.4881+00.3148 & 17:04:28.04 &-40:46:23.3 & -17.9 &2.12$\pm$0.10 & 1980.0 &47.89$\pm$0.21\\
G346.0774$-$00.0562 & 17:07:53.91 &-40:31:34.2 & -83.9 &10.92$\pm$0.03 & 8.3 &46.90$\pm$0.01\\
G347.8707+00.0146 & 17:13:08.77 &-39:02:28.2 & -30.6 &3.42$\pm$0.04 & 206.1 &47.33$\pm$0.05\\
G348.8922$-$00.1787 & 17:17:00.10 &-38:19:26.4 & 8.13 &11.20$\pm$0.05 & 185.7 &48.32$\pm$0.02\\
\hline
\end{tabular}
\label{tb:rotn}
Notes. Columns are (1)source name, (2)right ascension,
(3)declination, (4)local standard of rest velocity, (5)distance,
(6)flux density at 4.8 GHz $S_{4.8}$, (7)number of UV ionizing
photons
a: Urquhart et al. (2007a)
\end{minipage}
\end{table*}
In this paper, we describe our physical and chemical studies of RMSs
(mainly HII regions) and compare our results with previous IRDCs
observations. The RMS survey was conceived at Leeds to
systematically search the entire Galaxy for MYSOs, by comparing the
colors of sources from the Midcourse Space Experiment (MSX) and
Two-Micron All-Sky Survey (2MASS) point sources (Lumsden et al.
2002). The RMS survey is an ongoing multi-wavelength observational
programme (e.g. Urquhart et al. 2007a; Urquhart et al. 2007b) and
will provide us with the largest MYSO sample for statistical
studies. We have analyzed 18 RMSs in the southern sky using archival
data from ATLASGAL, ATCA and MALT90. N$_2$H$^+$ and HCO$^+$ are both
good tracers of dense gases, as their critical density $n_{crit}$
$>$ $10^5$ cm$^{-3}$. N$_2$H$^+$ is more resistent to freeze-out on
grains than the carbon-bearing species (Hirota et al. 1998). So it
is an excellent tracer of cold and dense molecular clouds. In star
formation regions, it is primarily formed through the gas-phase
reaction H$_3$$^+$ + N$_2$ $\rightarrow$ N$_2$H$^+$ + H$_2$, and
destroyed by CO molecules in the gas-phase through N$_2$H$^+$ + CO
$\rightarrow$ HCO$^+$ + N$_2$. It can also be destroyed in the
electron recombination: N$_2$H$^+$ + e$^-$ $\rightarrow$ N$_2$ + H
or NH + N (e.g. Dislaire et al. 2012; Vigren et al. 2012). In dense
molecular clouds, HCO$^+$ is mainly formed through the gas-phase
ion-neutral reaction H$_3$$^+$ + CO $\rightarrow$ HCO$^+$ + H$_2$
(e.g. Herbst $\&$ Klemperer 1973). The HCO$^+$ abundance can also be
increased in regions where shocks are generated. In chemical models
of low mass star formation, a relative enhancement of N$_2$H$^+$ and
a depletion of HCO$^+$ abundances are expected in the cold
prestellar phase, as CO is thought to be depleted in starless cores
(e.g. Lee et al. 2003, 2004; Bergin \& Tafalla 2007). When the
central star evolves, the gas gets warmer and CO should be
evaporated from the dust grains if the dust temperature exceeds
about $\sim$ 20 K (Tobin et al. 2013). Thus the HCO$^+$ abundance
should increase while the N$_2$H$^+$ abundance drops.
\begin{table*}
\begin{minipage}{13cm}
\caption{\label{tab:test}Properties of the RMSs.}
\begin{tabular}{lclclclclclclcl}
\hline
\hline
RMS & $S_{peak}$/$\Omega$$^a$ & $S_{870}$$^a$ & $T_d$$^b$& R$_{eff}$$^a$ & $M_{tot}$ & $N_{H_2}$$^c$ & $<n_{H_2}>$ & \\
name & (Jy/beam) & (Jy) &(K) & ($^\prime$$^\prime$) & $(M_\odot$) & ($\times$ 10$^{22}$ cm$^{-2}$) & ($\times$ 10$^{4}$ cm$^{-3}$) & \\
\hline
G330.2838+00.4933 & 2.00 & 7.50 & 30.0 &30 & 1995$\pm$46 &2.2& 1.5$\pm$0.1 \\
G331.4181$-$00.3546 & 3.48 &39.81 & 18.9 &75 & 9099$\pm$238 &6.0& 0.3$\pm$0.1 \\
G332.2944$-$00.0962 & 7.04 &34.52 & 27.6 &62 & 4603$\pm$77 &8.2& 1.3$\pm$0.1 \\
G332.5438$-$00.1277 & 2.03 &26.30 & 30.5 &69 & 3174$\pm$97 &2.1& 0.6$\pm$0.1 \\
G332.8256$-$00.5498A &32.86 &139.92 & 31.7 &79 &16246$\pm$460 &33.5& 2.1$\pm$0.1\\
G333.3401$-$00.1273 & 1.01 & 3.43 & 40.5 &28 & 311$\pm$5 &0.8& 0.9$\pm$0.1 \\
G337.0047+00.3226 & 2.28 & 6.67 & 33.9 &25 & 6341$\pm$56 &2.2& 0.8$\pm$0.1 \\
G337.4032$-$00.4037 &18.35 &58.81 & 29.0 &47 & 5535$\pm$143 &20.4&5.4$\pm$0.2 \\
G337.7091+00.0932A & 9.73 &48.94 & 30.0(A) &66 &55045$\pm$424 &10.5&0.4$\pm$0.1 \\
G337.8442$-$00.3748 & 3.32 & 9.90 & 32.8 &31 & 771$\pm$30 &3.3 &2.9$\pm$0.1 \\
G339.1052+00.1490 & 1.98 &8.64 & 29.6 &34 & 1831$\pm$23 &2.2 &1.4$\pm$0.1 \\
G340.0708+00.9267 & 2.69 &15.10 & 33.4 &40 & 2716$\pm$48 &2.6 &1.3$\pm$0.1 \\
G343.5024$-$00.0145 & 6.21 &45.12 & 24.8 &71 & 3766$\pm$209 &8.1 &1.6$\pm$0.1 \\
G345.0034$-$00.2240A &18.51 &137.14 & 31.0 &113& 9850$\pm$528 &19.3&0.9$\pm$0.1 \\
G345.4881+00.3148 &17.27 &143.50 & 26.5 &77 & 6782$\pm$646 &21.0&4.8$\pm$0.4 \\
G346.0774$-$00.0562 & 2.38 &11.57 & 30.0(A) &43 &13013$\pm$76 &2.6 &0.4$\pm$0.1 \\
G347.8707+00.0146 & 3.16 &17.56 & 30.0(A) &45 & 1922$\pm$45 &3.4 &1.6$\pm$0.1 \\
G348.8922$-$00.1787 & 3.02 &7.80 & 30.0(A) &17 & 4177$\pm$19 &1.8 &1.8$\pm$0.1 \\
\hline
\end{tabular}
\label{tb:rotn}
Notes. Columns are (1)source name, (2)peak intensities $S_{peak}$,
(3)flux densities $S_{870}$, (4)dust temperature, (5)effective
radii, (6)clump mass, (7)H$_2$ column density, (8)volume-averaged
H$_2$ number density
a: Contreras et al. (2013)
b: The $T_{d}$ values marked with ``A'' represent the average value derived for other sources.
c: The data has been converted to the Mopra telescope beam size.
\end{minipage}
\end{table*}
\section{data sets and source selections}
The MALT90 is a large international project aimed at characterizing
the sites within our Galaxy where high-mass stars will form. The
target clumps are selected from the 870 $\mu$m ATLASGAL to host the
early stages of high-mass star formation and include pre-stellar
clumps, protostellar clumps, and HII regions. Exploiting the unique
broad frequency range and fast-mapping capabilities of the Mopra
22-m telescope, MALT90 maps 16 emission lines simultaneously at
frequencies near 90 GHz. These molecular lines will probe the
physical, chemical and evolutionary states of dense high-mass
star-forming cores. The survey covers a Galactic longitude range of
$\sim$ -60 to $\sim$ 15$^{\circ}$ and Galactic latitude range of -1
to +1 $^{\circ}$. The observations were carried out with the newly
upgraded Mopra Spectrometer (MOPS). The full 8 GHz bandwidth of MOPS
was split into 16 zoom bands of 138 MHz each providing a velocity
resolution of $\sim$ 0.11 km s$^{-1}$. The angular resolution of
Mopra is about 38 arcsec, with beam efficiency between 0.49 at 86
GHz and 0.42 at 115 GHz (Ladd et al. 2005). The mapping sizes are
3$^{\prime}$.4 $\times$ 3$^{\prime}$.4 with 9$^{\prime\prime}$
spacing between adjacent rows. The MALT90 data includes (\emph{l},
\emph{b}, \emph{v}) data cubes and (\emph{l}, \emph{b}) moment and
error maps, and is publicly available from the MALT90 Home
Page\footnote{See
http://atoa.atnf.csiro.au/MALT90}. More information
about this survey can be found in Foster et al. (2011) and Hoq et
al. (2013). The data processing was conducted using CLASS (Continuum
and Line Analysis Single-Disk Software) and GreG (Grenoble Graphic)
software packages.
The ATLASGAL is the first systematic survey of the inner Galactic
plane in the sub-millimeter. The observations were carried out with
the Large APEX Bolometer Camera (LABOCA), an array of 295 bolometers
observing at 870 $\mu$m (Contreras et al. 2013). The aim of this
survey is to obtain a complete census of cold dusty clumps in the
Galaxy, and to study their distributions across the Galaxy. The
half-power beam width (HPBW) of the telescope is
18.$^\prime$$^\prime$6 ($\sim$ 0.22 pc at 2.4 kpc) at the frequency
used. The total field of view (FoV) of LABOCA is 11.$^\prime$4. The
instrument and its observing modes are described in Siringo et al.
(2009). And the data reduction is described in Contreras et al.
(2013).
In order to choose a genuine sample of MYSOs from the RMSs, Urquhart
et al. (2007a) have completed the 5-GHz observations of 892 RMS
sources in the southern sky using the ATCA. According to the
observations, they divided these sources into three groups: real
MYSO candidates, HII regions (UCHII and HCHII) and others such as
evolved stars and planetary nebulae (PNe). The radio continuum
emissions allow us to derive the expected spectral type of the
exciting stars. Assuming an electron of $T_e$ = 10$^4$ K, the number
of UV ionizing photons needed to keep an HII region ionized is given
as (Chaisson 1976):
\begin{equation}
N_L = 7.6 \times 10^{46} s^{-1}
(\frac{S_\nu}{Jy})(\frac{D}{kpc})^2(\frac{\nu}{GHz})^{0.1}(\frac{T_e}{10^4
K})^{-0.45}
\end{equation}
The large number of Lyman continuum fluxes (listed in Table 1,
column 7) indicates they are probably massive O or early B type star
formation regions.
In order to study physical and chemical properties in massive star
formation regions, we selected 18 sources from the RMS survey by
applying the following criteria: (i)According to the observations of
ATCA (Urquhart et al. 2007a), sources should have simple
spherical/unresolved/shell-like/cometary radio emissions. (ii)The
N$_2$H$^+$ and HCO$^+$ emissions of our sources should be detected
by MALT90, and the local standard of rest (LSR) velocities of
HCO$^+$ and N$_2$H$^+$ should be similar to those of Urquhart et al.
(2007b, 2008). (iii)The center locations of RMSs and 870 $\mu$m peak
emissions are within 18$^\prime$$^\prime$, which is about the
angular resolution of MSX at 8.0 $\mu$m. (iv)Sources should not be
on the edge of known HII or supernova regions, considering the large
beam of the 22-m Mopra telescope. The information of our selected
sources is listed in Table 1. The detection of radio emissions in
our sources indicates they are much more evolved, at least be
similar to the ``active''/``red'' stages of IRDCs. However, sources
of Hoq et al. (2013) and Miettinen (2014) range from prestellar
(``quiescent'') to HII regions with bright IR emissions (``red'').
We suppose to find different chemical properties of our RMSs using
the above archival data.
\begin{table*}
\begin{minipage}{17cm}
\caption{\label{tab:test}Parameters for molecular lines N$_2$H$^+$ and HCO$^+$.}
\begin{tabular}{lclclclclclclcl}
\hline
\hline
RMS & Molecular & Width & $T_{ex}$$^a$ & $\tau$ & $\int$T$_{mb}$ d$v$ & $N$ & $\chi$ \\
name & line &(km s$^{-1}$) & (K) & & (K km s$^{-1}$) & ($\times$ 10$^{12}$ cm$^{-2}$) &($\times$ 10$^{-10}$) \\
\hline
G330.2838+00.4933 & N$_2$H$^+$ & 2.65$\pm$ 0.16 &11.2 $\pm$0.5 & 0.14$\pm$0.06 &3.6$\pm$ 0.5& 5.9$\pm$1.7&2.7$\pm$ 0.8\\
& HCO$^+$ & 3.54$\pm$ 0.17 &12.5$\pm$2.5 & 0.53$\pm$0.18 &13.3$\pm$ 1.0& 18.3$\pm$ 1.4&8.3$\pm$ 0.6\\
G331.4181$-$00.3546 & N$_2$H$^+$ & 2.34$\pm$ 0.21 &6.2$\pm$1.8 & 1.17$\pm$ 0.17&10.6$\pm$ 0.8& 13.8$\pm$3.2&2.3$\pm$ 0.5\\
& HCO$^+$ & 4.39$\pm$ 0.45 &12.5$\pm$2.5 & 0.47$\pm$0.16 &15.5$\pm$ 2.0& 20.9$\pm$ 2.7&3.5$\pm$ 0.5\\
G332.2944$-$00.0962 & N$_2$H$^+$ & 2.88$\pm$ 0.10 &7.2$\pm$0.5 & 0.83$\pm$0.12 &12.2$\pm$ 0.4& 16.4$\pm$2.1&2.0$\pm$ 0.3\\
& HCO$^+$ & 4.51$\pm$ 0.40 &12.5$\pm$2.5 & 0.23$\pm$0.07 &9.3$\pm$ 1.7& 11.7$\pm$ 2.1&1.4$\pm$ 0.2\\
G332.5438$-$00.1277 & N$_2$H$^+$ & 1.90$\pm$ 0.18 & 7.8$\pm$1.4 & 0.41$\pm$ 0.27 &4.5$\pm$ 0.3& 6.3$\pm$ 0.3&3.0$\pm$ 0.3\\
& HCO$^+$ & 2.70$\pm$ 0.14 &12.5$\pm$2.5 & 0.64$\pm$0.23 &4.1$\pm$ 0.2& 5.8$\pm$ 0.3&2.8$\pm$ 0.1\\
G332.8256$-$00.5498A & N$_2$H$^+$ & 4.19$\pm$ 0.21 &8.9$\pm$2.3(A) & ...&16.2$\pm$ 2.4& 23.6$\pm$ 3.9&0.8$\pm$ 0.1\\
& HCO$^+$ & 11.84$\pm$ 0.31 &12.5$\pm$2.5 & 0.49$\pm$0.16 &31.0$\pm$ 0.3& 42.2$\pm$ 0.4&1.3$\pm$ 0.1\\
G333.3401$-$00.1273 & N$_2$H$^+$ & 2.10$\pm$ 0.29 &8.9$\pm$2.3(A) & ... &2.8$\pm$ 0.4& 4.0$\pm$1.2&5.0$\pm$ 1.5\\
& HCO$^+$ & 2.47$\pm$ 0.26 &12.5$\pm$2.5 & 0.20$\pm$0.06 &4.1$\pm$ 1.2& 5.1$\pm$ 1.5&6.4$\pm$ 1.8\\
G337.0047+00.3226 & N$_2$H$^+$ & 3.76$\pm$ 0.21 &8.9$\pm$2.3(A) & ... &2.9$\pm$ 0.5& 4.1$\pm$ 1.0&1.5$\pm$0.5\\
& HCO$^+$ & 4.63$\pm$ 0.25 &12.5$\pm$2.5 & 0.40$\pm$0.13 &13.9$\pm$ 1.6& 18.4$\pm$ 2.1&8.4$\pm$ 1.0\\
G337.4032$-$00.4037 & N$_2$H$^+$ & 3.16$\pm$ 0.01 &10.4$\pm$1.3 & 0.47$\pm$0.05 &13.2$\pm$ 0.4& 20.7$\pm$1.2&1.0$\pm$ 0.1\\
& HCO$^+$ & 6.49$\pm$ 0.18 &12.5$\pm$2.5 & 0.51$\pm$0.17 &20.9$\pm$ 1.5& 28.6$\pm$ 2.1&1.4$\pm$ 0.1\\
G337.7091+00.0932A & N$_2$H$^+$ & 4.02$\pm$ 0.29 & 7.5$\pm$0.6 & 0.40$\pm$ 0.27 &8.9$\pm$ 0.7& 12.3$\pm$ 2.2&1.2$\pm$ 0.2\\
& HCO$^+$ & 5.40$\pm$ 0.31 &12.5$\pm$2.5 & 0.35$\pm$0.11 &15.1$\pm$ 0.7& 19.7$\pm$ 0.9&1.9$\pm$ 0.1\\
G337.8442$-$00.3748 & N$_2$H$^+$ & 2.36$\pm$ 0.12 &8.9$\pm$2.3(A) & ... &6.2$\pm$ 0.2& 9.1$\pm$1.2&2.8$\pm$ 0.4\\
& HCO$^+$ & 2.41$\pm$ 0.26 &12.5$\pm$2.5 & 0.25$\pm$0.08 &4.9$\pm$ 0.2& 6.2$\pm$ 0.3&1.9$\pm$ 0.1\\
G339.1052+00.1490 & N$_2$H$^+$ & 2.62$\pm$ 0.16 &12.1$\pm$0.5 & 0.2$\pm$0.07 &6.5$\pm$ 0.4& 11.1$\pm$2.8&5.0$\pm$0.3\\
& HCO$^+$ & 4.01$\pm$ 0.42 &12.5$\pm$2.5 & 0.21$\pm$0.06 &7.4$\pm$ 2.2& 9.2$\pm$ 2.7&4.2$\pm$ 1.2\\
G340.0708+00.9267 & N$_2$H$^+$ & 3.58$\pm$ 0.19 &4.1$\pm$0.5 & 2.1$\pm$0.35 &7.5$\pm$ 1.0& 10.0$\pm$1.2&3.8$\pm$ 0.5\\
& HCO$^+$ & 3.41$\pm$ 0.11 &12.5$\pm$2.5 & 0.93$\pm$0.38 &17.9$\pm$ 0.9& 27.0$\pm$ 1.4&10.4$\pm$ 0.5\\
G343.5024$-$00.0145 & N$_2$H$^+$ & 3.13$\pm$ 0.10 &9.9$\pm$0.5 & 0.76$\pm$0.01 &17.1$\pm$ 0.6& 26.2$\pm$2.5&3.2$\pm$ 0.3\\
& HCO$^+$ & 3.75$\pm$ 0.23 &12.5$\pm$2.5 & 0.55$\pm$0.19 &15.0$\pm$ 1.0& 20.7$\pm$ 1.4&2.6$\pm$ 0.2\\
G345.0034$-$00.2240A & N$_2$H$^+$ & 4.21$\pm$ 0.16 &4.4$\pm$0.5 & 2.0$\pm$0.30 &12.8$\pm$ 0.8& 16.7$\pm$1.7&0.9$\pm$ 0.1\\
& HCO$^+$ & 5.02$\pm$ 0.39 &12.5$\pm$2.5 & 0.20$\pm$0.04 &9.6$\pm$ 0.4& 12.1$\pm$ 0.5&0.6$\pm$ 0.1\\
G345.4881+00.3148 & N$_2$H$^+$ & 3.83$\pm$ 0.13 &5.6$\pm$0.5 & 1.7$\pm$0.10 &17.5$\pm$ 1.5& 22.6$\pm$3.3 &1.1$\pm$ 0.2\\
& HCO$^+$ & 5.21$\pm$ 0.19 &12.5$\pm$2.5 & 1.29$\pm$0.20 &43.2$\pm$ 2.8& 96.4$\pm$ 6.2&4.6$\pm$ 0.3\\
G346.0774$-$00.0562 & N$_2$H$^+$ & 3.12$\pm$ 0.24 & 9.5$\pm$0.5 & 0.46$\pm$ 0.19&10.6$\pm$ 0.6& 16.0$\pm$ 0.7&6.1$\pm$ 0.3\\
& HCO$^+$ & 2.59$\pm$ 0.24 &12.5$\pm$2.5 & 0.32$\pm$0.10 &6.2$\pm$ 1.1& 8.0$\pm$ 1.4&3.1$\pm$ 0.6\\
G347.8707+00.0146 & N$_2$H$^+$ & 4.40$\pm$ 0.29 &8.9$\pm$2.3(A) & ... &8.0$\pm$ 0.5& 11.6$\pm$1.1&3.4 $\pm$ 0.3\\
& HCO$^+$ & 9.02$\pm$ 0.31 &12.5$\pm$2.5 & 0.35$\pm$0.11 &17.5$\pm$ 1.0& 22.8$\pm$ 2.0&6.7$\pm$ 0.6\\
G348.8922$-$00.1787 & N$_2$H$^+$ & 3.23$\pm$ 0.23 &8.9$\pm$2.3(A) & ... &3.1$\pm$ 0.5& 4.4$\pm$1.6 &1.3$\pm$ 0.3\\
& HCO$^+$ & 5.70$\pm$ 0.20 &12.5$\pm$2.5 & 0.50$\pm$0.17 &20.2$\pm$ 1.4& 27.5$\pm$ 1.9&8.3$\pm$ 0.6\\
\hline
\end{tabular}
\label{tb:rotn}
Notes. Columns are (1)source name, (2)Molecular line name, (3)Line
width, (4)excitation temperature, (5)optical depth, (6)integrated
intensity, (7)total column density, (8)fractional abundance with
respect to H$_2$.
a: The $T_{ex}$ of N$_2$H$^+$ values marked with ``A'' represent the average value derived for other sources.
\end{minipage}
\end{table*}
\section{results and analysis }
\subsection{Physical parameters}
The dust temperature ($T_{d}$) is essential to study the physics and
chemistry of star formation. For those sources in our sample that
are associated with IRAS point sources, we estimated the dust
temperature to be the same as the 60/100 $\mu$m color temperature
defined by Henning et al. (1990) as
\begin{equation}
T_{d} \simeq T_c(\frac{60}{100}) = 96 [(3+\beta)ln(\frac{100}{60})
- ln(\frac{S_{60}}{S_{100}})]^{-1}
\end{equation}
where $\beta$ is the dust emissivity index set to be 1.8 to be
consistent with the Ossenkopf $\&$ Henning (1994) dust model.
$S_\lambda$ is the flux density at the wavelength $\lambda$. The
derived values are listed in Table 2. The mean dust temperature is
about 30 K, which is much larger than those in IRDCs (e.g. Miettinen
$\&$ Harju 2010; Hoq et al. 2013), indicating our sources are more
evolved. Four of our sources are not associated with any IRAS point
source. For these sources we assume their dust temperature to be 30
K. Masses will be underestimated if the temperature is lower than
the assumed value of 30 K. For example, at 20 K the masses will be
higher by a factor of 1.7.
Assuming that the dust emission at 870 $\mu$m is optically thin, the
clump masses and H$_2$ column densities could be estimated through
the following expressions:
\begin{equation}
M_{tot} = \frac{S_{870} D^2 R}{B_{870}(T_d) \kappa_{870}}
\end{equation}
\begin{equation}
N_{H_2} = \frac{S_{peak} R}{\Omega B_{870}(T_d) \kappa_{870}
m_{H_2}}
\end{equation}
where $S_{870}$ is the integrated flux density. $S_{peak}$ denotes
the peak flux density. $D$ is the distance to the RMS. $\Omega$ is
the beam solid angle. $R$ is the gas-to-dust mass ratio which is set
to be 100. $m_{H_2}$ is the mass of one hydrogen molecule.
$B_{870}(T_d)$ is the Planck function at the dust temperature $T_d$.
We assumed that $\kappa_{870}$ = 0.17 m$^2$ kg$^{-1}$ (Miettinen
$\&$ Harju 2010). The derived masses are inverse proportional to the
assumed value of the opacity $\kappa_{870}$, which has an
uncertainty of at least a factor 1.5.
The volume-averaged H$_2$ number densities ($<n(H_2)>$) could be
calculated assuming a spherical geometry for the clumps, using these
formula
\begin{equation}
<\rho> = \frac{3 M_{tot}}{4 \pi R_{eff}^3}
\end{equation}
\begin{equation}
<n(H_2)> = \frac{<\rho>}{\mu_{H_2} m_H}
\end{equation}
The linear clump effective radii is derived from the angular radii
and kinematic distances as R$_{eff}$ (pc) = R$_{eff}$(rad) $\times$
D(pc). The value of R$_{eff}$(rad) is from Contreras et al. (2013).
We adopt a mean molecular weight per H$_2$ molecule of $\mu$ = 2.8
to include helium. The derived parameters are listed in Table 2.
Taking into account the spatial resolution effects, the derived
column densities are all beyond the column density threshold
required by theoretical considerations for massive star formation.
\subsection{Chemical parameters}
Figures 3-20 show the N$_2$H$^+$ (1-0) and HCO$^+$ (1-0) spectra on
the peak emissions of 870 $\mu$m and their integrated intensities
superimposed on the 870 $\mu$m images of the RMSs. Most of the
detected N$_2$H$^+$ and HCO$^+$ emissions match well with the dust
emission, implying a close link to their chemical evolution in the
RMSs. The HCO$^+$ emission lines of some sources are far of having a
simple Gaussian shape, presenting asymmetries, and spectral wings or
shoulders, which suggest that the molecular gas is affected by the
dynamics of these star-forming regions. Yu $\&$ Wang (2014) made a
study of 19 RMSs and find these profiles are probably caused by
infall and/or outflow activities. Some of our sources are also
included in their study. The N$_2$H$^+$ (1-0) line has 15 hyperfine
structures (HFS) out of which seven have a different frequency (e.g.
Pagani et al. 2009; Keto $\&$ Rybicki 2010). These velocity
components blend into three groups (see figure 2 of Purcell et al.
2009). We followed the procedure outlined by Purcell et al. (2009)
to estimate the optical depth of N$_2$H$^+$. Assuming the line
widths of the individual hyperfine components are all equal, the
integrated intensities of the three blended groups should be in the
ratio of 1:5:3 under optically thin conditions. The optical depth
can then be derived from the ratio of the integrated intensities
($\int T_{MB} dv$) of any groups using the following equation:
\begin{equation}
\frac{\int T_{MB,1} dv}{\int T_{MB,2} dv} = \frac{1 -
exp(-\tau_1)}{1 - exp(-a\tau_2)}
\end{equation}
where $a$ is the expected ratio of $\tau_2/\tau_1$ under optically
thin conditions. We determined the optical depth only from the
intensity ratio of group 1/group 2, as anomalous excitation of the
$F_1F$ = 10-11 and 12-12 components (in our group 3) has been
reported by Caselli et al. (1995). The optical depth of group 2 is
listed in table 3 because it provides a better estimate for the
excitation temperature when $\tau
> 0.1$. Then the excitation temperature ($T_{ex}$) for N$_2$H$^+$
could be calculated with the following formula:
\begin{equation}
T_{ex} = 4.47 / ln(1+ (\frac{T_{MB}}{4.47(1 - exp(-\tau))} +
0.236)^{-1})
\end{equation}
The N$_2$H$^+$ optical depth of our six clumps could not be derived
through this method. For these cases, we adopted the mean derived
excitation temperature. Assuming local thermodynamic equilibrium
(LTE) and using Eq. (2) in Miettinen (2014), we derived the column
densities of N$_2$H$^+$. Finally, the fractional abundance of
N$_2$H$^+$ with respect to H$_2$ could be estimated by $\chi$
(N$_2$H$^+$) = N(N$_2$H$^+$)/N(H$_2$). The derived parameters are
listed in Table 3.
Generally speaking, the optical depth and $T_{ex}$ of HCO$^+$ could
be derived by comparing the intensities of HCO$^+$ and
H$^{13}$CO$^+$. However, the H$^{13}$CO$^+$ emission was not
detected in most cases. For the sources (G332.8256-00.5498A,
G337.4032-00.4037, G345.0034-00.2240A and G345.4881+00.3148) which
show distinct H$^{13}$CO$^+$ emissions, we derive the optical depths
of HCO$^+$ from
\begin{equation}
\frac{^{12}T_{mb}}{^{13}T_{mb}} =
\frac{1-exp(-\tau_{12})}{1-exp(-\tau_{12}/X)}
\end{equation}
where $\tau_{12}$ is the optical depth of the HCO$^+$ gas and $X$ is
the isotope abundance ratio
\begin{equation}
X = \frac{N_{HCO^+}}{N_{H^{13}CO^+}} = \frac{\chi (HCO^+)}{\chi
(H^{13}CO^+)} \simeq \frac{[^{12}C]}{[^{13}C]}
\end{equation}
The Galactic gradient in the $^{12}C/^{13}C$ ratio ranges from
$\sim$20 to $\sim$70 (see Savage et al. 2002, and references
therein). And the ratio of $\frac{\chi (HCO^+)}{\chi (H^{13}CO^+)}$
may also be affected by chemistry. Thus, we assumed a constant $X$ =
50. We then can calculate the excitation temperature according to
\begin{equation}
T_{ex} = 4.28(ln(1+
4.28[\frac{T_{mb}}{f(1-exp(-\tau_\nu))}+J(T_{bg})]^{-1}))^{-1}
\end{equation}
where $f$ is the filling factor, here we assume it to be 1. The
derived excitation temperature is 6.3K, 6.4K, 4.7K and 12.3K
respectively. Our values are considerably lower than the $\sim$ 10
$-$ 15K temperature in previous work (e.g. Girart et al. 2000).
Purcell et al. (2006) regard that the emission is beam diluted in a
significant fraction. If the emission is smaller than the beam, the
filling factor $f$ will be less than 1 and the excitation
temperature will be underestimated. In our final analysis, we have
assumed an excitation temperature of 10 $-$ 15 K and the column
densities are calculated by Eq. (4) in Liu et al. (2013). The
fractional abundance of HCO$^+$ with respect to H$_2$ could be
estimated by $\chi$ (HCO$^+$) = N(HCO$^+$)/N(H$_2$). The derived
parameters are also listed in Table 3. From Table 3 we find the
optical depths of HCO$^+$ are less than 1. Considering the fact that
HCO$^+$ emission should be optically thick, the optical depths in
our results are probably underestimated.
\section{discussions}
\subsection{Ionized accretion?}
The HCO$^+$ (1-0) lines show non-gaussian profiles in our four
sources (G332.8256-00.5498A, G337.4032-00.4037, G345.0034-00.2240A,
and G345.4881+00.3148).By position-velocity (PV) diagram and the
method described by Mardones et al. (1997), Yu $\&$ Wang (2014)
found recent outflow activities in G332.8256-00.5498A,
G345.0034-00.2240A and G345.4881+00.3148. Large scale infall has
also been discovered in G345.0034-00.2240A and G345.4881+00.3148.
From the 4.8 GHz radio observations of ATCA, we can see HCHII/UCHII
regions have already formed inside. The sound speed in $\sim$ 10$^4$
K photo-ionized gas is about 10 km s$^{-1}$, larger than the escape
speeds in galactic molecular cores (see Fig. 1 in Tan et al. 2014).
This means if the gas in a massive star formation region gets
ionized, the gas pressure will drive a thermal pressure that may
chock off accretion. It is still unclear whether accretion can
continue in some form such as through an ionized accretion flow. The
high resolution observations of H66$\alpha$ taken at the VLA have
directly shown an ionized accretion in G10.62-0.38 (Keto $\&$ Wood
2006), while accretion in G5.89-0.39 seems to have halted at the
onset of the UCHII region (Klassen et al. 2006). We here discuss
whether (ionized) accretion could continue in G345.4881+00.3148.
We can determine the infall velocity using the two layer radiative
transfer model of Myers et al. (1996). Using their equation (9):
\begin{equation}
V_{in} \approx \frac{\sigma^2}{v_{red}-v_{blue}} ln(\frac{1 + e
T_{BD}/T_D}{1 + e T_{RD}/T_D})
\end{equation}
where $T_D$, $T_{BD}$ and $T_{RD}$ are the brightness temperature of
the dip, the height of the blue peak above the dip and the height of
the red peak above the dip respectively. The velocity dispersion
$\sigma$ can be obtained from the FWHM of H$^{13}$CO$^+$ (1-0). The
gaussian-fitted lines of HCO$^+$ and H$^{13}$CO$^+$ is shown in
figure 17. We find G345.4881+00.3148 has an infall velocity of about
4.2 km s$^{-1}$ in large scale. To estimate the mass infall rate, we
use Eq. (3) of Klassen $\&$ Wilson (2007) $dM/dt$ = (4/3)$\pi$
$<n_{H_2}>$ $\mu$ $m_H$ $r^2$ V$_{in}$, where $\mu$ means molecular
weight, $m_H$ is the mass of Hydrogen, $r$ is the radius of the
emitting region (here is set to be the beam radius), V$_{in}$ is the
infall velocity, and $<n_{H_2}>$ is the ambient density calculated
above. From the analysis above, we estimate mass infall rate of
G345.4881+00.3148 is about 2.1 $\times$ 10$^{-3}$ M$_\odot$
yr$^{-1}$. This value is much higher than those generally observed
in low-mass star forming regions.
Walmsley (1995) considered the accretion flow in free-fall onto a
star, and showed that the escape speed from the edge of the ionized
region will exceed the ionized gas sound speed $c_i$ if the
accretion rate satisfies
\begin{equation}
\begin{split}
&\frac{dM}{dt} > [\frac{8 \pi m_H^2 G M S}{2.2 \alpha_B
ln(v_{esc}/c_i)}]\\
&= 4 \times 10^{-5} (\frac{M}{100
M_\odot}\frac{N_L}{10^{49}s^{-1}})^{1/2}M_\odot yr^{-1}
\end{split}
\end{equation}
where $M$ is the central stellar mass, $m_H$ is the mean mass per H
nucleus, $\alpha_B$ is the case B recombination coefficient and
$v_{esc}$ is the escape speed from the stellar surface. The
numerical evaluation uses $v_{esc}$ = 1000 km s$^{-1}$ and $c_i$ =
10 km s$^{-1}$ (The result is unsensitive to these parameters). Our
calculated mass infall rate is sufficient to allow continuing
accretion onto an B0 (G345.4881+00.3148) type star. Like the case of
G10.6-0.4, accretion flow in this region may be ionized. However,
given the low angular resolution of the data (at 2.1 kpc,
38$^{\prime\prime}$ is over half a parsec), it is also possible that
a lower mass star is forming in the vicinity. High angular
resolution observations are needed to verify our speculations.
\subsection{N$_2$H$^+$ and HCO$^+$}
N$_2$H$^+$ is widely detected in low- and high-mass prestellar and
protostellar cores, owing to its resistance of depletion at low
temperature and high densities. We find that its line widths for our
RMSs vary in the range from 1.9 to 4.4 km s$^{-1}$, much larger than
those found in low-mass prestellar cores (Lee et al. 2001). This
large line width can not be purely explained by thermal motions.
Internal turbulence may dominate in massive star formation regions
(Myers $\&$ Fuller 1992; Caselli $\&$ Myers 1995). Chambers et al.
(2006) and Vasyunina et al. (2011) found that the line widths and
the integrated intensities of N$_2$H$^+$ have a trend to larger
values, from ``quiescent'', ``middle(or intermediate)'' to
``active'' clumps in IRDCs. They regarded it was because that
``active'' sources are more evolved and present further evolutionary
stages compared with ``quiescent'' sources. The mean line width of
our RMSs is 3.2 km s$^{-1}$, larger than the mean value of their
``active'' clumps. It is probably that our sources are much more
evolved. However, our mean integrated intensity of N$_2$H$^+$ is
only 10 K km s$^{-1}$, between their ``middle'' and ``active'' clump
values. In our sources, we found N$_2$H$^+$ column densities in the
range of 4.0 $\times$ 10$^{12}$ $-$ 2.6 $\times$ 10$^{13}$
cm$^{-2}$, and abundances of 0.9 $\times$ 10$^{-10}$ $-$ 6.1
$\times$ 10$^{-10}$. The average values are 1.6 $\times$ 10$^{13}$
cm$^{-2}$ and 3.0 $\times$ 10$^{-10}$, respectively. Vasyunina et
al. (2011) derived fractional N$_2$H$^+$ abundances of 1.9 $\times$
10$^{-10}$ $-$ 8.5 $\times$ 10$^{-9}$ with an average of 2.8
$\times$ 10$^{-9}$ for their IRDCs. Miettien (2014) derived similar
fractional N$_2$H$^+$ abundances of 2.8 $\times$ 10$^{-10}$ $-$ 9.8
$\times$ 10$^{-9}$ with an average of 1.6 $\times$ 10$^{-9}$. Our
value of $\chi$ (N$_2$H$^+$) is nearly one order less than these
observations. Chemical models indicate when dust temperature exceeds
20 K, CO would be evaporated from the dust grains. The enhanced CO
abundance will destroy N$_2$H$^+$ producing HCO$^+$. The mean dust
temperature of our sources is about 30 K, much larger than that in
IRDCs ($\sim$ 15 K). Previous molecular-line observations have also
shown the typical gas kinetic temperature of IRDCs clumps lies in
the range of 10 and 20 K (e.g. Sridharan et al. 2005; Sakai et al.
2008; Ragan et al. 2011). This may be the reason why our values of
$\chi$ (N$_2$H$^+$) are much smaller.
HCO$^+$ can be formed through N$_2$H$^+$ + CO $\rightarrow$ HCO$^+$
+ N$_2$ when CO is evaporated from the dust grains. The HCO$^+$
abundance can also be increased in regions where shocks are
generated (e.g. the gas through which a protostellar outflow is
passing). Our HCO$^+$ column densities range from 5.1 $\times$
10$^{12}$ cm$^{-2}$ to 9.6 $\times$ 10$^{13}$ cm$^{-2}$ (2.4
$\times$ 10$^{13}$ cm$^{-2}$ on average). These values are very
close to those of Miettien (2014), and are higher by a factor of
about five on average than those derived by Liu et al. (2013). As
shown in figure 1, there is a hint that the fractional abundance of
HCO$^+$ increases as a function of the N$_2$H$^+$ abundance. A least
squares fit to the data points yields log[$\chi$ (HCO$^+$)] =
(0.76$\pm$0.51) + (1.07$\pm$0.05)log[$\chi$ (N$_2$H$^+$)], with the
linear Pearson correlation coefficient of $r$ = 0.61. The trend is
consistent with theoretical models. Because CO is both the main
supplier of HCO$^+$ and the main destroyer of N$_2$H$^+$, this
suggests that the HCO$^+$ abundance increases with respect to
N$_2$H$^+$ as sources evolve to a warmer phase. N$_2$H$^+$ can also
be destroyed in the electron recombination: N$_2$H$^+$ + e$^-$
$\rightarrow$ N$_2$ + H or NH + N (e.g. Dislaire et al. 2012; Vigren
et al. 2012). We tried to find a relationship between the N$_2$H$^+$
fractional abundance and Lyman continuum fluxes ($N_L$). Figure 2
seems to show that $\chi$(N$_2$H$^+$) decreases as a function of
$N_L$. UV radiation field may have an influence on the chemistry of
N$_2$H$^+$ in the RMSs. Since our limited data, more studies should
be done in the near future to check out this conclusion.
\section{summary}
Using archival data taken from ATLASGAL, ATCA, and MALT90, we have
studied the physical and chemical properties of 18 southern RMSs
(HII regions), and compared our results with previous observations
of IRDCs. Our sources are much more evolved than general IRDCs. The
H$_2$ column densities and gas number densities have been derived.
Ionized accretion may be taking place in G345.4881+00.3148. Most of
the detected N$_2$H$^+$ and HCO$^+$ emissions match well with the
dust emission, implying a close link to their chemical evolution in
the RMSs. We found the abundance of N$_2$H$^+$ is one order of
magnitude lower than in other surveys of IRDCs, and a positive
correlation between the abundances of N$_2$H$^+$ and HCO$^+$. The
fractional abundance of N$_2$H$^+$ with respect to H$_2$ decreases
as a function of $N_L$. It seems that the UV radiation field has an
influence on the chemistry of N$_2$H$^+$ in the RMSs. Like the
chemical model of low-mass star formation theory, these observed
trends could be interpreted as an indication of enhanced destruction
of N$_2$H$^+$, either by CO or through dissociative recombination
with electrons produced by central UV photons.
\begin{figure}
\centerline{\psfig{file=fig19.1.ps,width=2.6in,height=2.0in}}
\caption{ HCO$^+$ fractional abundances vs. N$_2$H$^+$ fractional
abundances. }
\end{figure}
\begin{figure}
\centerline{\psfig{file=fig20.ps,width=2.6in,height=2.0in}}
\caption{N$_2$H$^+$ fractional abundance plotted as a function of
$N_L$ in logarithmic scales.}
\end{figure}
\section*{ACKNOWLEDGEMENTS}
We thank the anonymous referee for his constructive suggestions.
This paper made use of information from the Red MSX Source (RMS)
survey database
http://rms.leeds.ac.uk/cgi-bin/public/RMS$_{-}$DATABASE.cgi which
was constructed with support from the Science and Technology
Facilities Council of the UK. This research made use of data
products from the Millimetre Astronomy Legacy Team 90 GHz (MALT90)
survey. The Mopra telescope is part of the Australia Telescope and
is funded by the Commonwealth of Australia for operation as National
Facility managed by CSIRO.
\section{Introduction}
One of the key challenges in stellar astrophysics is to understand
the formation of massive stars and their role in the evolution of
the universe. A lot of research has been done in astrophysics in the
last two decades and considerable progress has been seen in the
understanding of massive star formation (Zinnecker $\&$ Yorke 2007).
Even though their short lives and rare sources, massive stars play a
major role in determining physical, chemical and morphological
structures of galaxies through ionizing radiation, heating of dust
and expansions of their HII regions. These processes may also
trigger next generation of star formation by compressing neighboring
molecular clouds to the point of gravitational instability. It is
generally accepted that massive stars evolve through starless cores
in infrared dark clouds (IRDCs) to hot molecular cores (HMCs; e.g.
Kurtz et al. 2000). Then substantial UV photons and ionized stellar
winds rapidly ionize the surrounding hydrogen, forming a
hyper-compact HII region (HCHII) or ultra-compact HII region
(UCHII). The final stages are compact and classical HII regions.
However, many questions are still unclear: whether accretion could
be halted by the strong outward radiation and thermal pressure? Does
it continue in an ionized form? Does it continue through a molecular
or ionized disk? Chemical composition of molecular gas is thought to
evolve due to the physical changes that occur during the star
formation process. As material collapses and gets ionized by central
young stars, densities and temperature rise, leading to the
production and destruction of different molecular species.
Multi-wavelength observations are vital to our understanding of the
physical and chemical conditions where massive star formation
occurs.
IRDCs provide us with the possibility of investigating the
subsequent early stages of high mass star formation, and a lot of
research has also been done to investigate their physical and
chemical properties (e.g. Vasyunina et al. 2009; Vasyunina et al.
2011; Hoq et al. 2013; Liu et al. 2013; Miettinen 2014). In order to
characterize the different evolutionary stages of IRDCs, Chambers et
al. (2009) proposed an evolutionary sequence in which ``quiescent''
clumps evolve into ``intermediate'', ``active'' and ``red'' clumps,
based on the Spitzer observations. Sanhueza et al. (2012) and Hoq et
al. (2013) characterize the chemical properties of a sample of IRDCs
using the data of MALT90. Correlations and anti-correlations between
molecular abundances and the evolutionary stages have been found.
These studies suggest chemical processes in high-mass star
formations may differ from their low-mass counterparts, as their
temperature, densities and UV fluxes are very different from each
other. Establishing the chemical properties of massive star
formation requires that large surveys be conducted.
\begin{table*}
\begin{minipage}{13cm}
\caption{\label{tab:test}List of our sources.}
\begin{tabular}{lclclclclclclcl}
\hline
\hline
RMS & RA & Dec. & V$_{lsr}$ &$D$ & $S$$_{4.8GHz}$$^a$ & log($N_L$) & \\
name & (J2000) & (J2000) &(km s$^{-1}$) & (kpc) & (mJy) & (S$^{-1}$) & \\
\hline
G330.2838+00.4933 & 16:03:43.28 &-51:51:45.6 & -94.2 &5.32$\pm$0.06 & 44.9 &47.05$\pm$0.05 \\
G331.4181$-$00.3546 & 16:12:50.24 &-51:43:28.5 & -63.2 &3.90$\pm$0.05 & 83.9 &46.78$\pm$0.06 \\
G332.2944$-$00.0962 & 16:15:45.83 &-50:56:02.4 & -48.7 &3.64$\pm$0.03 & 175.6 &47.30$\pm$0.04\\
G332.5438$-$00.1277 & 16:17:02.33 &-50:47:03.3 & -47.1 &3.64$\pm$0.06 & 42.2 &46.69$\pm$0.07\\
G332.8256$-$00.5498A & 16:20:11.06 &-50:53:16.2 & -58.1 &3.64$\pm$0.05 & 1259.0 &48.16$\pm$0.07\\
G333.3401$-$00.1273 & 16:20:37.00 &-50:13:32.7 & -59.7 &3.64$\pm$0.03 & 3.3 &45.58$\pm$0.04\\
G337.0047+00.3226 & 16:34:04.73 &-47:16:29.2 & -62.4 &11.43$\pm$0.05 & 249.3 &48.46$\pm$0.02\\
G337.4032$-$00.4037 & 16:38:50.44 &-47:28:03.0 & -40.0 &3.14$\pm$0.04 & 90.3 &46.89$\pm$0.06\\
G337.7091+00.0932A & 16:37:51.80 &-46:54:33.4 & -75.9 &10.98$\pm$0.04 & 144.9 &48.19$\pm$0.02\\
G337.8442$-$00.3748 & 16:40:26.68 &-47:07:13.1 & -39.9 &3.06$\pm$0.06 & 11.1 &45.95$\pm$0.09\\
G339.1052+00.1490 & 16:42:59.58 &-45:49:43.6 & -78.2 &4.73$\pm$0.06 & 42.9 &46.93$\pm$0.03\\
G340.0708+00.9267 & 16:43:15.69 &-44:35:16.0 & -73.5 &4.69$\pm$0.04 & 93.8 &46.96$\pm$0.04\\
G343.5024$-$00.0145 & 16:59:20.78 &-42:32:37.5 & -27.9 &2.72$\pm$0.08 & 176.9 &47.06$\pm$0.12\\
G345.0034$-$00.2240A & 17:05:11.19 &-41:29:06.3 & -27.3 &2.87$\pm$0.08 & 140.0 &46.99$\pm$0.12\\
G345.4881+00.3148 & 17:04:28.04 &-40:46:23.3 & -17.9 &2.12$\pm$0.10 & 1980.0 &47.89$\pm$0.21\\
G346.0774$-$00.0562 & 17:07:53.91 &-40:31:34.2 & -83.9 &10.92$\pm$0.03 & 8.3 &46.90$\pm$0.01\\
G347.8707+00.0146 & 17:13:08.77 &-39:02:28.2 & -30.6 &3.42$\pm$0.04 & 206.1 &47.33$\pm$0.05\\
G348.8922$-$00.1787 & 17:17:00.10 &-38:19:26.4 & 8.13 &11.20$\pm$0.05 & 185.7 &48.32$\pm$0.02\\
\hline
\end{tabular}
\label{tb:rotn}
Notes. Columns are (1)source name, (2)right ascension,
(3)declination, (4)local standard of rest velocity, (5)distance,
(6)flux density at 4.8 GHz $S_{4.8}$, (7)number of UV ionizing
photons
a: Urquhart et al. (2007a)
\end{minipage}
\end{table*}
In this paper, we describe our physical and chemical studies of RMSs
(mainly HII regions) and compare our results with previous IRDCs
observations. The RMS survey was conceived at Leeds to
systematically search the entire Galaxy for MYSOs, by comparing the
colors of sources from the Midcourse Space Experiment (MSX) and
Two-Micron All-Sky Survey (2MASS) point sources (Lumsden et al.
2002). The RMS survey is an ongoing multi-wavelength observational
programme (e.g. Urquhart et al. 2007a; Urquhart et al. 2007b) and
will provide us with the largest MYSO sample for statistical
studies. We have analyzed 18 RMSs in the southern sky using archival
data from ATLASGAL, ATCA and MALT90. N$_2$H$^+$ and HCO$^+$ are both
good tracers of dense gases, as their critical density $n_{crit}$
$>$ $10^5$ cm$^{-3}$. N$_2$H$^+$ is more resistent to freeze-out on
grains than the carbon-bearing species (Hirota et al. 1998). So it
is an excellent tracer of cold and dense molecular clouds. In star
formation regions, it is primarily formed through the gas-phase
reaction H$_3$$^+$ + N$_2$ $\rightarrow$ N$_2$H$^+$ + H$_2$, and
destroyed by CO molecules in the gas-phase through N$_2$H$^+$ + CO
$\rightarrow$ HCO$^+$ + N$_2$. It can also be destroyed in the
electron recombination: N$_2$H$^+$ + e$^-$ $\rightarrow$ N$_2$ + H
or NH + N (e.g. Dislaire et al. 2012; Vigren et al. 2012). In dense
molecular clouds, HCO$^+$ is mainly formed through the gas-phase
ion-neutral reaction H$_3$$^+$ + CO $\rightarrow$ HCO$^+$ + H$_2$
(e.g. Herbst $\&$ Klemperer 1973). The HCO$^+$ abundance can also be
increased in regions where shocks are generated. In chemical models
of low mass star formation, a relative enhancement of N$_2$H$^+$ and
a depletion of HCO$^+$ abundances are expected in the cold
prestellar phase, as CO is thought to be depleted in starless cores
(e.g. Lee et al. 2003, 2004; Bergin \& Tafalla 2007). When the
central star evolves, the gas gets warmer and CO should be
evaporated from the dust grains if the dust temperature exceeds
about $\sim$ 20 K (Tobin et al. 2013). Thus the HCO$^+$ abundance
should increase while the N$_2$H$^+$ abundance drops.
\begin{table*}
\begin{minipage}{13cm}
\caption{\label{tab:test}Properties of the RMSs.}
\begin{tabular}{lclclclclclclcl}
\hline
\hline
RMS & $S_{peak}$/$\Omega$$^a$ & $S_{870}$$^a$ & $T_d$$^b$& R$_{eff}$$^a$ & $M_{tot}$ & $N_{H_2}$$^c$ & $<n_{H_2}>$ & \\
name & (Jy/beam) & (Jy) &(K) & ($^\prime$$^\prime$) & $(M_\odot$) & ($\times$ 10$^{22}$ cm$^{-2}$) & ($\times$ 10$^{4}$ cm$^{-3}$) & \\
\hline
G330.2838+00.4933 & 2.00 & 7.50 & 30.0 &30 & 1995$\pm$46 &2.2& 1.5$\pm$0.1 \\
G331.4181$-$00.3546 & 3.48 &39.81 & 18.9 &75 & 9099$\pm$238 &6.0& 0.3$\pm$0.1 \\
G332.2944$-$00.0962 & 7.04 &34.52 & 27.6 &62 & 4603$\pm$77 &8.2& 1.3$\pm$0.1 \\
G332.5438$-$00.1277 & 2.03 &26.30 & 30.5 &69 & 3174$\pm$97 &2.1& 0.6$\pm$0.1 \\
G332.8256$-$00.5498A &32.86 &139.92 & 31.7 &79 &16246$\pm$460 &33.5& 2.1$\pm$0.1\\
G333.3401$-$00.1273 & 1.01 & 3.43 & 40.5 &28 & 311$\pm$5 &0.8& 0.9$\pm$0.1 \\
G337.0047+00.3226 & 2.28 & 6.67 & 33.9 &25 & 6341$\pm$56 &2.2& 0.8$\pm$0.1 \\
G337.4032$-$00.4037 &18.35 &58.81 & 29.0 &47 & 5535$\pm$143 &20.4&5.4$\pm$0.2 \\
G337.7091+00.0932A & 9.73 &48.94 & 30.0(A) &66 &55045$\pm$424 &10.5&0.4$\pm$0.1 \\
G337.8442$-$00.3748 & 3.32 & 9.90 & 32.8 &31 & 771$\pm$30 &3.3 &2.9$\pm$0.1 \\
G339.1052+00.1490 & 1.98 &8.64 & 29.6 &34 & 1831$\pm$23 &2.2 &1.4$\pm$0.1 \\
G340.0708+00.9267 & 2.69 &15.10 & 33.4 &40 & 2716$\pm$48 &2.6 &1.3$\pm$0.1 \\
G343.5024$-$00.0145 & 6.21 &45.12 & 24.8 &71 & 3766$\pm$209 &8.1 &1.6$\pm$0.1 \\
G345.0034$-$00.2240A &18.51 &137.14 & 31.0 &113& 9850$\pm$528 &19.3&0.9$\pm$0.1 \\
G345.4881+00.3148 &17.27 &143.50 & 26.5 &77 & 6782$\pm$646 &21.0&4.8$\pm$0.4 \\
G346.0774$-$00.0562 & 2.38 &11.57 & 30.0(A) &43 &13013$\pm$76 &2.6 &0.4$\pm$0.1 \\
G347.8707+00.0146 & 3.16 &17.56 & 30.0(A) &45 & 1922$\pm$45 &3.4 &1.6$\pm$0.1 \\
G348.8922$-$00.1787 & 3.02 &7.80 & 30.0(A) &17 & 4177$\pm$19 &1.8 &1.8$\pm$0.1 \\
\hline
\end{tabular}
\label{tb:rotn}
Notes. Columns are (1)source name, (2)peak intensities $S_{peak}$,
(3)flux densities $S_{870}$, (4)dust temperature, (5)effective
radii, (6)clump mass, (7)H$_2$ column density, (8)volume-averaged
H$_2$ number density
a: Contreras et al. (2013)
b: The $T_{d}$ values marked with ``A'' represent the average value derived for other sources.
c: The data has been converted to the Mopra telescope beam size.
\end{minipage}
\end{table*}
\section{data sets and source selections}
The MALT90 is a large international project aimed at characterizing
the sites within our Galaxy where high-mass stars will form. The
target clumps are selected from the 870 $\mu$m ATLASGAL to host the
early stages of high-mass star formation and include pre-stellar
clumps, protostellar clumps, and HII regions. Exploiting the unique
broad frequency range and fast-mapping capabilities of the Mopra
22-m telescope, MALT90 maps 16 emission lines simultaneously at
frequencies near 90 GHz. These molecular lines will probe the
physical, chemical and evolutionary states of dense high-mass
star-forming cores. The survey covers a Galactic longitude range of
$\sim$ -60 to $\sim$ 15$^{\circ}$ and Galactic latitude range of -1
to +1 $^{\circ}$. The observations were carried out with the newly
upgraded Mopra Spectrometer (MOPS). The full 8 GHz bandwidth of MOPS
was split into 16 zoom bands of 138 MHz each providing a velocity
resolution of $\sim$ 0.11 km s$^{-1}$. The angular resolution of
Mopra is about 38 arcsec, with beam efficiency between 0.49 at 86
GHz and 0.42 at 115 GHz (Ladd et al. 2005). The mapping sizes are
3$^{\prime}$.4 $\times$ 3$^{\prime}$.4 with 9$^{\prime\prime}$
spacing between adjacent rows. The MALT90 data includes (\emph{l},
\emph{b}, \emph{v}) data cubes and (\emph{l}, \emph{b}) moment and
error maps, and is publicly available from the MALT90 Home
Page\footnote{See
http://atoa.atnf.csiro.au/MALT90}. More information
about this survey can be found in Foster et al. (2011) and Hoq et
al. (2013). The data processing was conducted using CLASS (Continuum
and Line Analysis Single-Disk Software) and GreG (Grenoble Graphic)
software packages.
The ATLASGAL is the first systematic survey of the inner Galactic
plane in the sub-millimeter. The observations were carried out with
the Large APEX Bolometer Camera (LABOCA), an array of 295 bolometers
observing at 870 $\mu$m (Contreras et al. 2013). The aim of this
survey is to obtain a complete census of cold dusty clumps in the
Galaxy, and to study their distributions across the Galaxy. The
half-power beam width (HPBW) of the telescope is
18.$^\prime$$^\prime$6 ($\sim$ 0.22 pc at 2.4 kpc) at the frequency
used. The total field of view (FoV) of LABOCA is 11.$^\prime$4. The
instrument and its observing modes are described in Siringo et al.
(2009). And the data reduction is described in Contreras et al.
(2013).
In order to choose a genuine sample of MYSOs from the RMSs, Urquhart
et al. (2007a) have completed the 5-GHz observations of 892 RMS
sources in the southern sky using the ATCA. According to the
observations, they divided these sources into three groups: real
MYSO candidates, HII regions (UCHII and HCHII) and others such as
evolved stars and planetary nebulae (PNe). The radio continuum
emissions allow us to derive the expected spectral type of the
exciting stars. Assuming an electron of $T_e$ = 10$^4$ K, the number
of UV ionizing photons needed to keep an HII region ionized is given
as (Chaisson 1976):
\begin{equation}
N_L = 7.6 \times 10^{46} s^{-1}
(\frac{S_\nu}{Jy})(\frac{D}{kpc})^2(\frac{\nu}{GHz})^{0.1}(\frac{T_e}{10^4
K})^{-0.45}
\end{equation}
The large number of Lyman continuum fluxes (listed in Table 1,
column 7) indicates they are probably massive O or early B type star
formation regions.
In order to study physical and chemical properties in massive star
formation regions, we selected 18 sources from the RMS survey by
applying the following criteria: (i)According to the observations of
ATCA (Urquhart et al. 2007a), sources should have simple
spherical/unresolved/shell-like/cometary radio emissions. (ii)The
N$_2$H$^+$ and HCO$^+$ emissions of our sources should be detected
by MALT90, and the local standard of rest (LSR) velocities of
HCO$^+$ and N$_2$H$^+$ should be similar to those of Urquhart et al.
(2007b, 2008). (iii)The center locations of RMSs and 870 $\mu$m peak
emissions are within 18$^\prime$$^\prime$, which is about the
angular resolution of MSX at 8.0 $\mu$m. (iv)Sources should not be
on the edge of known HII or supernova regions, considering the large
beam of the 22-m Mopra telescope. The information of our selected
sources is listed in Table 1. The detection of radio emissions in
our sources indicates they are much more evolved, at least be
similar to the ``active''/``red'' stages of IRDCs. However, sources
of Hoq et al. (2013) and Miettinen (2014) range from prestellar
(``quiescent'') to HII regions with bright IR emissions (``red'').
We suppose to find different chemical properties of our RMSs using
the above archival data.
\begin{table*}
\begin{minipage}{17cm}
\caption{\label{tab:test}Parameters for molecular lines N$_2$H$^+$ and HCO$^+$.}
\begin{tabular}{lclclclclclclcl}
\hline
\hline
RMS & Molecular & Width & $T_{ex}$$^a$ & $\tau$ & $\int$T$_{mb}$ d$v$ & $N$ & $\chi$ \\
name & line &(km s$^{-1}$) & (K) & & (K km s$^{-1}$) & ($\times$ 10$^{12}$ cm$^{-2}$) &($\times$ 10$^{-10}$) \\
\hline
G330.2838+00.4933 & N$_2$H$^+$ & 2.65$\pm$ 0.16 &11.2 $\pm$0.5 & 0.14$\pm$0.06 &3.6$\pm$ 0.5& 5.9$\pm$1.7&2.7$\pm$ 0.8\\
& HCO$^+$ & 3.54$\pm$ 0.17 &12.5$\pm$2.5 & 0.53$\pm$0.18 &13.3$\pm$ 1.0& 18.3$\pm$ 1.4&8.3$\pm$ 0.6\\
G331.4181$-$00.3546 & N$_2$H$^+$ & 2.34$\pm$ 0.21 &6.2$\pm$1.8 & 1.17$\pm$ 0.17&10.6$\pm$ 0.8& 13.8$\pm$3.2&2.3$\pm$ 0.5\\
& HCO$^+$ & 4.39$\pm$ 0.45 &12.5$\pm$2.5 & 0.47$\pm$0.16 &15.5$\pm$ 2.0& 20.9$\pm$ 2.7&3.5$\pm$ 0.5\\
G332.2944$-$00.0962 & N$_2$H$^+$ & 2.88$\pm$ 0.10 &7.2$\pm$0.5 & 0.83$\pm$0.12 &12.2$\pm$ 0.4& 16.4$\pm$2.1&2.0$\pm$ 0.3\\
& HCO$^+$ & 4.51$\pm$ 0.40 &12.5$\pm$2.5 & 0.23$\pm$0.07 &9.3$\pm$ 1.7& 11.7$\pm$ 2.1&1.4$\pm$ 0.2\\
G332.5438$-$00.1277 & N$_2$H$^+$ & 1.90$\pm$ 0.18 & 7.8$\pm$1.4 & 0.41$\pm$ 0.27 &4.5$\pm$ 0.3& 6.3$\pm$ 0.3&3.0$\pm$ 0.3\\
& HCO$^+$ & 2.70$\pm$ 0.14 &12.5$\pm$2.5 & 0.64$\pm$0.23 &4.1$\pm$ 0.2& 5.8$\pm$ 0.3&2.8$\pm$ 0.1\\
G332.8256$-$00.5498A & N$_2$H$^+$ & 4.19$\pm$ 0.21 &8.9$\pm$2.3(A) & ...&16.2$\pm$ 2.4& 23.6$\pm$ 3.9&0.8$\pm$ 0.1\\
& HCO$^+$ & 11.84$\pm$ 0.31 &12.5$\pm$2.5 & 0.49$\pm$0.16 &31.0$\pm$ 0.3& 42.2$\pm$ 0.4&1.3$\pm$ 0.1\\
G333.3401$-$00.1273 & N$_2$H$^+$ & 2.10$\pm$ 0.29 &8.9$\pm$2.3(A) & ... &2.8$\pm$ 0.4& 4.0$\pm$1.2&5.0$\pm$ 1.5\\
& HCO$^+$ & 2.47$\pm$ 0.26 &12.5$\pm$2.5 & 0.20$\pm$0.06 &4.1$\pm$ 1.2& 5.1$\pm$ 1.5&6.4$\pm$ 1.8\\
G337.0047+00.3226 & N$_2$H$^+$ & 3.76$\pm$ 0.21 &8.9$\pm$2.3(A) & ... &2.9$\pm$ 0.5& 4.1$\pm$ 1.0&1.5$\pm$0.5\\
& HCO$^+$ & 4.63$\pm$ 0.25 &12.5$\pm$2.5 & 0.40$\pm$0.13 &13.9$\pm$ 1.6& 18.4$\pm$ 2.1&8.4$\pm$ 1.0\\
G337.4032$-$00.4037 & N$_2$H$^+$ & 3.16$\pm$ 0.01 &10.4$\pm$1.3 & 0.47$\pm$0.05 &13.2$\pm$ 0.4& 20.7$\pm$1.2&1.0$\pm$ 0.1\\
& HCO$^+$ & 6.49$\pm$ 0.18 &12.5$\pm$2.5 & 0.51$\pm$0.17 &20.9$\pm$ 1.5& 28.6$\pm$ 2.1&1.4$\pm$ 0.1\\
G337.7091+00.0932A & N$_2$H$^+$ & 4.02$\pm$ 0.29 & 7.5$\pm$0.6 & 0.40$\pm$ 0.27 &8.9$\pm$ 0.7& 12.3$\pm$ 2.2&1.2$\pm$ 0.2\\
& HCO$^+$ & 5.40$\pm$ 0.31 &12.5$\pm$2.5 & 0.35$\pm$0.11 &15.1$\pm$ 0.7& 19.7$\pm$ 0.9&1.9$\pm$ 0.1\\
G337.8442$-$00.3748 & N$_2$H$^+$ & 2.36$\pm$ 0.12 &8.9$\pm$2.3(A) & ... &6.2$\pm$ 0.2& 9.1$\pm$1.2&2.8$\pm$ 0.4\\
& HCO$^+$ & 2.41$\pm$ 0.26 &12.5$\pm$2.5 & 0.25$\pm$0.08 &4.9$\pm$ 0.2& 6.2$\pm$ 0.3&1.9$\pm$ 0.1\\
G339.1052+00.1490 & N$_2$H$^+$ & 2.62$\pm$ 0.16 &12.1$\pm$0.5 & 0.2$\pm$0.07 &6.5$\pm$ 0.4& 11.1$\pm$2.8&5.0$\pm$0.3\\
& HCO$^+$ & 4.01$\pm$ 0.42 &12.5$\pm$2.5 & 0.21$\pm$0.06 &7.4$\pm$ 2.2& 9.2$\pm$ 2.7&4.2$\pm$ 1.2\\
G340.0708+00.9267 & N$_2$H$^+$ & 3.58$\pm$ 0.19 &4.1$\pm$0.5 & 2.1$\pm$0.35 &7.5$\pm$ 1.0& 10.0$\pm$1.2&3.8$\pm$ 0.5\\
& HCO$^+$ & 3.41$\pm$ 0.11 &12.5$\pm$2.5 & 0.93$\pm$0.38 &17.9$\pm$ 0.9& 27.0$\pm$ 1.4&10.4$\pm$ 0.5\\
G343.5024$-$00.0145 & N$_2$H$^+$ & 3.13$\pm$ 0.10 &9.9$\pm$0.5 & 0.76$\pm$0.01 &17.1$\pm$ 0.6& 26.2$\pm$2.5&3.2$\pm$ 0.3\\
& HCO$^+$ & 3.75$\pm$ 0.23 &12.5$\pm$2.5 & 0.55$\pm$0.19 &15.0$\pm$ 1.0& 20.7$\pm$ 1.4&2.6$\pm$ 0.2\\
G345.0034$-$00.2240A & N$_2$H$^+$ & 4.21$\pm$ 0.16 &4.4$\pm$0.5 & 2.0$\pm$0.30 &12.8$\pm$ 0.8& 16.7$\pm$1.7&0.9$\pm$ 0.1\\
& HCO$^+$ & 5.02$\pm$ 0.39 &12.5$\pm$2.5 & 0.20$\pm$0.04 &9.6$\pm$ 0.4& 12.1$\pm$ 0.5&0.6$\pm$ 0.1\\
G345.4881+00.3148 & N$_2$H$^+$ & 3.83$\pm$ 0.13 &5.6$\pm$0.5 & 1.7$\pm$0.10 &17.5$\pm$ 1.5& 22.6$\pm$3.3 &1.1$\pm$ 0.2\\
& HCO$^+$ & 5.21$\pm$ 0.19 &12.5$\pm$2.5 & 1.29$\pm$0.20 &43.2$\pm$ 2.8& 96.4$\pm$ 6.2&4.6$\pm$ 0.3\\
G346.0774$-$00.0562 & N$_2$H$^+$ & 3.12$\pm$ 0.24 & 9.5$\pm$0.5 & 0.46$\pm$ 0.19&10.6$\pm$ 0.6& 16.0$\pm$ 0.7&6.1$\pm$ 0.3\\
& HCO$^+$ & 2.59$\pm$ 0.24 &12.5$\pm$2.5 & 0.32$\pm$0.10 &6.2$\pm$ 1.1& 8.0$\pm$ 1.4&3.1$\pm$ 0.6\\
G347.8707+00.0146 & N$_2$H$^+$ & 4.40$\pm$ 0.29 &8.9$\pm$2.3(A) & ... &8.0$\pm$ 0.5& 11.6$\pm$1.1&3.4 $\pm$ 0.3\\
& HCO$^+$ & 9.02$\pm$ 0.31 &12.5$\pm$2.5 & 0.35$\pm$0.11 &17.5$\pm$ 1.0& 22.8$\pm$ 2.0&6.7$\pm$ 0.6\\
G348.8922$-$00.1787 & N$_2$H$^+$ & 3.23$\pm$ 0.23 &8.9$\pm$2.3(A) & ... &3.1$\pm$ 0.5& 4.4$\pm$1.6 &1.3$\pm$ 0.3\\
& HCO$^+$ & 5.70$\pm$ 0.20 &12.5$\pm$2.5 & 0.50$\pm$0.17 &20.2$\pm$ 1.4& 27.5$\pm$ 1.9&8.3$\pm$ 0.6\\
\hline
\end{tabular}
\label{tb:rotn}
Notes. Columns are (1)source name, (2)Molecular line name, (3)Line
width, (4)excitation temperature, (5)optical depth, (6)integrated
intensity, (7)total column density, (8)fractional abundance with
respect to H$_2$.
a: The $T_{ex}$ of N$_2$H$^+$ values marked with ``A'' represent the average value derived for other sources.
\end{minipage}
\end{table*}
\section{results and analysis }
\subsection{Physical parameters}
The dust temperature ($T_{d}$) is essential to study the physics and
chemistry of star formation. For those sources in our sample that
are associated with IRAS point sources, we estimated the dust
temperature to be the same as the 60/100 $\mu$m color temperature
defined by Henning et al. (1990) as
\begin{equation}
T_{d} \simeq T_c(\frac{60}{100}) = 96 [(3+\beta)ln(\frac{100}{60})
- ln(\frac{S_{60}}{S_{100}})]^{-1}
\end{equation}
where $\beta$ is the dust emissivity index set to be 1.8 to be
consistent with the Ossenkopf $\&$ Henning (1994) dust model.
$S_\lambda$ is the flux density at the wavelength $\lambda$. The
derived values are listed in Table 2. The mean dust temperature is
about 30 K, which is much larger than those in IRDCs (e.g. Miettinen
$\&$ Harju 2010; Hoq et al. 2013), indicating our sources are more
evolved. Four of our sources are not associated with any IRAS point
source. For these sources we assume their dust temperature to be 30
K. Masses will be underestimated if the temperature is lower than
the assumed value of 30 K. For example, at 20 K the masses will be
higher by a factor of 1.7.
Assuming that the dust emission at 870 $\mu$m is optically thin, the
clump masses and H$_2$ column densities could be estimated through
the following expressions:
\begin{equation}
M_{tot} = \frac{S_{870} D^2 R}{B_{870}(T_d) \kappa_{870}}
\end{equation}
\begin{equation}
N_{H_2} = \frac{S_{peak} R}{\Omega B_{870}(T_d) \kappa_{870}
m_{H_2}}
\end{equation}
where $S_{870}$ is the integrated flux density. $S_{peak}$ denotes
the peak flux density. $D$ is the distance to the RMS. $\Omega$ is
the beam solid angle. $R$ is the gas-to-dust mass ratio which is set
to be 100. $m_{H_2}$ is the mass of one hydrogen molecule.
$B_{870}(T_d)$ is the Planck function at the dust temperature $T_d$.
We assumed that $\kappa_{870}$ = 0.17 m$^2$ kg$^{-1}$ (Miettinen
$\&$ Harju 2010). The derived masses are inverse proportional to the
assumed value of the opacity $\kappa_{870}$, which has an
uncertainty of at least a factor 1.5.
The volume-averaged H$_2$ number densities ($<n(H_2)>$) could be
calculated assuming a spherical geometry for the clumps, using these
formula
\begin{equation}
<\rho> = \frac{3 M_{tot}}{4 \pi R_{eff}^3}
\end{equation}
\begin{equation}
<n(H_2)> = \frac{<\rho>}{\mu_{H_2} m_H}
\end{equation}
The linear clump effective radii is derived from the angular radii
and kinematic distances as R$_{eff}$ (pc) = R$_{eff}$(rad) $\times$
D(pc). The value of R$_{eff}$(rad) is from Contreras et al. (2013).
We adopt a mean molecular weight per H$_2$ molecule of $\mu$ = 2.8
to include helium. The derived parameters are listed in Table 2.
Taking into account the spatial resolution effects, the derived
column densities are all beyond the column density threshold
required by theoretical considerations for massive star formation.
\subsection{Chemical parameters}
Figures 3-20 show the N$_2$H$^+$ (1-0) and HCO$^+$ (1-0) spectra on
the peak emissions of 870 $\mu$m and their integrated intensities
superimposed on the 870 $\mu$m images of the RMSs. Most of the
detected N$_2$H$^+$ and HCO$^+$ emissions match well with the dust
emission, implying a close link to their chemical evolution in the
RMSs. The HCO$^+$ emission lines of some sources are far of having a
simple Gaussian shape, presenting asymmetries, and spectral wings or
shoulders, which suggest that the molecular gas is affected by the
dynamics of these star-forming regions. Yu $\&$ Wang (2014) made a
study of 19 RMSs and find these profiles are probably caused by
infall and/or outflow activities. Some of our sources are also
included in their study. The N$_2$H$^+$ (1-0) line has 15 hyperfine
structures (HFS) out of which seven have a different frequency (e.g.
Pagani et al. 2009; Keto $\&$ Rybicki 2010). These velocity
components blend into three groups (see figure 2 of Purcell et al.
2009). We followed the procedure outlined by Purcell et al. (2009)
to estimate the optical depth of N$_2$H$^+$. Assuming the line
widths of the individual hyperfine components are all equal, the
integrated intensities of the three blended groups should be in the
ratio of 1:5:3 under optically thin conditions. The optical depth
can then be derived from the ratio of the integrated intensities
($\int T_{MB} dv$) of any groups using the following equation:
\begin{equation}
\frac{\int T_{MB,1} dv}{\int T_{MB,2} dv} = \frac{1 -
exp(-\tau_1)}{1 - exp(-a\tau_2)}
\end{equation}
where $a$ is the expected ratio of $\tau_2/\tau_1$ under optically
thin conditions. We determined the optical depth only from the
intensity ratio of group 1/group 2, as anomalous excitation of the
$F_1F$ = 10-11 and 12-12 components (in our group 3) has been
reported by Caselli et al. (1995). The optical depth of group 2 is
listed in table 3 because it provides a better estimate for the
excitation temperature when $\tau
> 0.1$. Then the excitation temperature ($T_{ex}$) for N$_2$H$^+$
could be calculated with the following formula:
\begin{equation}
T_{ex} = 4.47 / ln(1+ (\frac{T_{MB}}{4.47(1 - exp(-\tau))} +
0.236)^{-1})
\end{equation}
The N$_2$H$^+$ optical depth of our six clumps could not be derived
through this method. For these cases, we adopted the mean derived
excitation temperature. Assuming local thermodynamic equilibrium
(LTE) and using Eq. (2) in Miettinen (2014), we derived the column
densities of N$_2$H$^+$. Finally, the fractional abundance of
N$_2$H$^+$ with respect to H$_2$ could be estimated by $\chi$
(N$_2$H$^+$) = N(N$_2$H$^+$)/N(H$_2$). The derived parameters are
listed in Table 3.
Generally speaking, the optical depth and $T_{ex}$ of HCO$^+$ could
be derived by comparing the intensities of HCO$^+$ and
H$^{13}$CO$^+$. However, the H$^{13}$CO$^+$ emission was not
detected in most cases. For the sources (G332.8256-00.5498A,
G337.4032-00.4037, G345.0034-00.2240A and G345.4881+00.3148) which
show distinct H$^{13}$CO$^+$ emissions, we derive the optical depths
of HCO$^+$ from
\begin{equation}
\frac{^{12}T_{mb}}{^{13}T_{mb}} =
\frac{1-exp(-\tau_{12})}{1-exp(-\tau_{12}/X)}
\end{equation}
where $\tau_{12}$ is the optical depth of the HCO$^+$ gas and $X$ is
the isotope abundance ratio
\begin{equation}
X = \frac{N_{HCO^+}}{N_{H^{13}CO^+}} = \frac{\chi (HCO^+)}{\chi
(H^{13}CO^+)} \simeq \frac{[^{12}C]}{[^{13}C]}
\end{equation}
The Galactic gradient in the $^{12}C/^{13}C$ ratio ranges from
$\sim$20 to $\sim$70 (see Savage et al. 2002, and references
therein). And the ratio of $\frac{\chi (HCO^+)}{\chi (H^{13}CO^+)}$
may also be affected by chemistry. Thus, we assumed a constant $X$ =
50. We then can calculate the excitation temperature according to
\begin{equation}
T_{ex} = 4.28(ln(1+
4.28[\frac{T_{mb}}{f(1-exp(-\tau_\nu))}+J(T_{bg})]^{-1}))^{-1}
\end{equation}
where $f$ is the filling factor, here we assume it to be 1. The
derived excitation temperature is 6.3K, 6.4K, 4.7K and 12.3K
respectively. Our values are considerably lower than the $\sim$ 10
$-$ 15K temperature in previous work (e.g. Girart et al. 2000).
Purcell et al. (2006) regard that the emission is beam diluted in a
significant fraction. If the emission is smaller than the beam, the
filling factor $f$ will be less than 1 and the excitation
temperature will be underestimated. In our final analysis, we have
assumed an excitation temperature of 10 $-$ 15 K and the column
densities are calculated by Eq. (4) in Liu et al. (2013). The
fractional abundance of HCO$^+$ with respect to H$_2$ could be
estimated by $\chi$ (HCO$^+$) = N(HCO$^+$)/N(H$_2$). The derived
parameters are also listed in Table 3. From Table 3 we find the
optical depths of HCO$^+$ are less than 1. Considering the fact that
HCO$^+$ emission should be optically thick, the optical depths in
our results are probably underestimated.
\section{discussions}
\subsection{Ionized accretion?}
The HCO$^+$ (1-0) lines show non-gaussian profiles in our four
sources (G332.8256-00.5498A, G337.4032-00.4037, G345.0034-00.2240A,
and G345.4881+00.3148).By position-velocity (PV) diagram and the
method described by Mardones et al. (1997), Yu $\&$ Wang (2014)
found recent outflow activities in G332.8256-00.5498A,
G345.0034-00.2240A and G345.4881+00.3148. Large scale infall has
also been discovered in G345.0034-00.2240A and G345.4881+00.3148.
From the 4.8 GHz radio observations of ATCA, we can see HCHII/UCHII
regions have already formed inside. The sound speed in $\sim$ 10$^4$
K photo-ionized gas is about 10 km s$^{-1}$, larger than the escape
speeds in galactic molecular cores (see Fig. 1 in Tan et al. 2014).
This means if the gas in a massive star formation region gets
ionized, the gas pressure will drive a thermal pressure that may
chock off accretion. It is still unclear whether accretion can
continue in some form such as through an ionized accretion flow. The
high resolution observations of H66$\alpha$ taken at the VLA have
directly shown an ionized accretion in G10.62-0.38 (Keto $\&$ Wood
2006), while accretion in G5.89-0.39 seems to have halted at the
onset of the UCHII region (Klassen et al. 2006). We here discuss
whether (ionized) accretion could continue in G345.4881+00.3148.
We can determine the infall velocity using the two layer radiative
transfer model of Myers et al. (1996). Using their equation (9):
\begin{equation}
V_{in} \approx \frac{\sigma^2}{v_{red}-v_{blue}} ln(\frac{1 + e
T_{BD}/T_D}{1 + e T_{RD}/T_D})
\end{equation}
where $T_D$, $T_{BD}$ and $T_{RD}$ are the brightness temperature of
the dip, the height of the blue peak above the dip and the height of
the red peak above the dip respectively. The velocity dispersion
$\sigma$ can be obtained from the FWHM of H$^{13}$CO$^+$ (1-0). The
gaussian-fitted lines of HCO$^+$ and H$^{13}$CO$^+$ is shown in
figure 17. We find G345.4881+00.3148 has an infall velocity of about
4.2 km s$^{-1}$ in large scale. To estimate the mass infall rate, we
use Eq. (3) of Klassen $\&$ Wilson (2007) $dM/dt$ = (4/3)$\pi$
$<n_{H_2}>$ $\mu$ $m_H$ $r^2$ V$_{in}$, where $\mu$ means molecular
weight, $m_H$ is the mass of Hydrogen, $r$ is the radius of the
emitting region (here is set to be the beam radius), V$_{in}$ is the
infall velocity, and $<n_{H_2}>$ is the ambient density calculated
above. From the analysis above, we estimate mass infall rate of
G345.4881+00.3148 is about 2.1 $\times$ 10$^{-3}$ M$_\odot$
yr$^{-1}$. This value is much higher than those generally observed
in low-mass star forming regions.
Walmsley (1995) considered the accretion flow in free-fall onto a
star, and showed that the escape speed from the edge of the ionized
region will exceed the ionized gas sound speed $c_i$ if the
accretion rate satisfies
\begin{equation}
\begin{split}
&\frac{dM}{dt} > [\frac{8 \pi m_H^2 G M S}{2.2 \alpha_B
ln(v_{esc}/c_i)}]\\
&= 4 \times 10^{-5} (\frac{M}{100
M_\odot}\frac{N_L}{10^{49}s^{-1}})^{1/2}M_\odot yr^{-1}
\end{split}
\end{equation}
where $M$ is the central stellar mass, $m_H$ is the mean mass per H
nucleus, $\alpha_B$ is the case B recombination coefficient and
$v_{esc}$ is the escape speed from the stellar surface. The
numerical evaluation uses $v_{esc}$ = 1000 km s$^{-1}$ and $c_i$ =
10 km s$^{-1}$ (The result is unsensitive to these parameters). Our
calculated mass infall rate is sufficient to allow continuing
accretion onto an B0 (G345.4881+00.3148) type star. Like the case of
G10.6-0.4, accretion flow in this region may be ionized. However,
given the low angular resolution of the data (at 2.1 kpc,
38$^{\prime\prime}$ is over half a parsec), it is also possible that
a lower mass star is forming in the vicinity. High angular
resolution observations are needed to verify our speculations.
\subsection{N$_2$H$^+$ and HCO$^+$}
N$_2$H$^+$ is widely detected in low- and high-mass prestellar and
protostellar cores, owing to its resistance of depletion at low
temperature and high densities. We find that its line widths for our
RMSs vary in the range from 1.9 to 4.4 km s$^{-1}$, much larger than
those found in low-mass prestellar cores (Lee et al. 2001). This
large line width can not be purely explained by thermal motions.
Internal turbulence may dominate in massive star formation regions
(Myers $\&$ Fuller 1992; Caselli $\&$ Myers 1995). Chambers et al.
(2006) and Vasyunina et al. (2011) found that the line widths and
the integrated intensities of N$_2$H$^+$ have a trend to larger
values, from ``quiescent'', ``middle(or intermediate)'' to
``active'' clumps in IRDCs. They regarded it was because that
``active'' sources are more evolved and present further evolutionary
stages compared with ``quiescent'' sources. The mean line width of
our RMSs is 3.2 km s$^{-1}$, larger than the mean value of their
``active'' clumps. It is probably that our sources are much more
evolved. However, our mean integrated intensity of N$_2$H$^+$ is
only 10 K km s$^{-1}$, between their ``middle'' and ``active'' clump
values. In our sources, we found N$_2$H$^+$ column densities in the
range of 4.0 $\times$ 10$^{12}$ $-$ 2.6 $\times$ 10$^{13}$
cm$^{-2}$, and abundances of 0.9 $\times$ 10$^{-10}$ $-$ 6.1
$\times$ 10$^{-10}$. The average values are 1.6 $\times$ 10$^{13}$
cm$^{-2}$ and 3.0 $\times$ 10$^{-10}$, respectively. Vasyunina et
al. (2011) derived fractional N$_2$H$^+$ abundances of 1.9 $\times$
10$^{-10}$ $-$ 8.5 $\times$ 10$^{-9}$ with an average of 2.8
$\times$ 10$^{-9}$ for their IRDCs. Miettien (2014) derived similar
fractional N$_2$H$^+$ abundances of 2.8 $\times$ 10$^{-10}$ $-$ 9.8
$\times$ 10$^{-9}$ with an average of 1.6 $\times$ 10$^{-9}$. Our
value of $\chi$ (N$_2$H$^+$) is nearly one order less than these
observations. Chemical models indicate when dust temperature exceeds
20 K, CO would be evaporated from the dust grains. The enhanced CO
abundance will destroy N$_2$H$^+$ producing HCO$^+$. The mean dust
temperature of our sources is about 30 K, much larger than that in
IRDCs ($\sim$ 15 K). Previous molecular-line observations have also
shown the typical gas kinetic temperature of IRDCs clumps lies in
the range of 10 and 20 K (e.g. Sridharan et al. 2005; Sakai et al.
2008; Ragan et al. 2011). This may be the reason why our values of
$\chi$ (N$_2$H$^+$) are much smaller.
HCO$^+$ can be formed through N$_2$H$^+$ + CO $\rightarrow$ HCO$^+$
+ N$_2$ when CO is evaporated from the dust grains. The HCO$^+$
abundance can also be increased in regions where shocks are
generated (e.g. the gas through which a protostellar outflow is
passing). Our HCO$^+$ column densities range from 5.1 $\times$
10$^{12}$ cm$^{-2}$ to 9.6 $\times$ 10$^{13}$ cm$^{-2}$ (2.4
$\times$ 10$^{13}$ cm$^{-2}$ on average). These values are very
close to those of Miettien (2014), and are higher by a factor of
about five on average than those derived by Liu et al. (2013). As
shown in figure 1, there is a hint that the fractional abundance of
HCO$^+$ increases as a function of the N$_2$H$^+$ abundance. A least
squares fit to the data points yields log[$\chi$ (HCO$^+$)] =
(0.76$\pm$0.51) + (1.07$\pm$0.05)log[$\chi$ (N$_2$H$^+$)], with the
linear Pearson correlation coefficient of $r$ = 0.61. The trend is
consistent with theoretical models. Because CO is both the main
supplier of HCO$^+$ and the main destroyer of N$_2$H$^+$, this
suggests that the HCO$^+$ abundance increases with respect to
N$_2$H$^+$ as sources evolve to a warmer phase. N$_2$H$^+$ can also
be destroyed in the electron recombination: N$_2$H$^+$ + e$^-$
$\rightarrow$ N$_2$ + H or NH + N (e.g. Dislaire et al. 2012; Vigren
et al. 2012). We tried to find a relationship between the N$_2$H$^+$
fractional abundance and Lyman continuum fluxes ($N_L$). Figure 2
seems to show that $\chi$(N$_2$H$^+$) decreases as a function of
$N_L$. UV radiation field may have an influence on the chemistry of
N$_2$H$^+$ in the RMSs. Since our limited data, more studies should
be done in the near future to check out this conclusion.
\section{summary}
Using archival data taken from ATLASGAL, ATCA, and MALT90, we have
studied the physical and chemical properties of 18 southern RMSs
(HII regions), and compared our results with previous observations
of IRDCs. Our sources are much more evolved than general IRDCs. The
H$_2$ column densities and gas number densities have been derived.
Ionized accretion may be taking place in G345.4881+00.3148. Most of
the detected N$_2$H$^+$ and HCO$^+$ emissions match well with the
dust emission, implying a close link to their chemical evolution in
the RMSs. We found the abundance of N$_2$H$^+$ is one order of
magnitude lower than in other surveys of IRDCs, and a positive
correlation between the abundances of N$_2$H$^+$ and HCO$^+$. The
fractional abundance of N$_2$H$^+$ with respect to H$_2$ decreases
as a function of $N_L$. It seems that the UV radiation field has an
influence on the chemistry of N$_2$H$^+$ in the RMSs. Like the
chemical model of low-mass star formation theory, these observed
trends could be interpreted as an indication of enhanced destruction
of N$_2$H$^+$, either by CO or through dissociative recombination
with electrons produced by central UV photons.
\begin{figure}
\centerline{\psfig{file=fig19.1.ps,width=2.6in,height=2.0in}}
\caption{ HCO$^+$ fractional abundances vs. N$_2$H$^+$ fractional
abundances. }
\end{figure}
\begin{figure}
\centerline{\psfig{file=fig20.ps,width=2.6in,height=2.0in}}
\caption{N$_2$H$^+$ fractional abundance plotted as a function of
$N_L$ in logarithmic scales.}
\end{figure}
\section*{ACKNOWLEDGEMENTS}
We thank the anonymous referee for his constructive suggestions.
This paper made use of information from the Red MSX Source (RMS)
survey database
http://rms.leeds.ac.uk/cgi-bin/public/RMS$_{-}$DATABASE.cgi which
was constructed with support from the Science and Technology
Facilities Council of the UK. This research made use of data
products from the Millimetre Astronomy Legacy Team 90 GHz (MALT90)
survey. The Mopra telescope is part of the Australia Telescope and
is funded by the Commonwealth of Australia for operation as National
Facility managed by CSIRO.
|
\section{Introduction and Main Results}
\label{Section:Introduction}
Reflected Brownian motion (RBM) in a wedge, as defined by Varadhan and Williams \cite{varadhan1985brownian}, is a 2-dimensional stochastic process $Z$ whose state space in $\mathbb{R}^2$ is given in polar coordinates by $S=\{(r,\theta): r \geq 0, 0 \leq \theta \leq \xi\}$ for some $0 < \xi < 2 \pi$. In the interior of $S$, RBM in a wedge behaves as a standard 2-d Brownian, whereas upon hitting the boundary of $S$, it is reflected back into the interior at an angle that depends upon which edge of the boundary has been hit. Let $\partial S_1=\{(r,\theta): r \ge 0, \theta = 0\}$ and $\partial S_2=\{(r,\theta): r \ge 0, \theta = \xi\}$. The angle of reflection off of $\partial S_1\setminus\{0\}$ is denoted by $ \theta_1$, and the angle of reflection off of $\partial S_2\setminus\{0\}$ is denoted by $ \theta_2 $. Both of these angles are measured with respect to the inward facing normal off of each edge, with angles directed toward the origin assumed to be positive. See Figure \ref{Figure:State:Space:Of:RBM:Wedge} below for an illustration of the above quantities. The angle of reflection at the vertex of the wedge is discussed later in the paper.
In \cite{varadhan1985brownian}, Varadhan and Williams provided the following rigorous definition of RBM in a wedge as the solution to a submartingale problem. For each $j=1,2,$ denote the direction of reflection off of the edge $\partial S_j\setminus\{0\}$ by $v_j$. Assume moreover that $v_j$ is normalized such that $v_j \cdot n_j =1$, where $n_j$ is the inward facing unit normal vector off of the edge $\partial S_j\setminus\{0\}$. Let $C_S$ denote the space of continuous functions with domain $\mathbb{R}_{+}=[0,\infty)$ and range $S$. For each $t \geq 0$ and $\omega \in C_S$, we denote by $Z(t):C_S \mapsto S$ the coordinate map $Z(t)(\omega)=Z(t,\omega)=\omega(t)$, and we also define the coordinate mapping process $Z=\{Z(t), t \geq 0\}$. Then, for each $t \geq 0$,
we set $\mathcal{M}_t = \sigma\{Z(s), 0 \leq s \leq t\}$, and set $\mathcal{M} = \sigma\{Z(s), s \geq 0\}$. Next, for each $n \geq 1$ and $F \subset \mathbb{R}^2$, denote by $C^n(F)$ the set of $n$-times continuously differentiable functions in some domain containing $F$ and let $C^n_b$ be the set of functions in $C^n(F)$ that have bounded partial derivatives up to and including order $n$ on $F$. Finally, define the differential operators $D_j = v_j \cdot \nabla$ for $j=1,2,$ and denote by $\triangle$ the Laplacian operator.
\begin{figure}[h]
\begin{center}
\scalebox{.75}{\input{RBMWedgeStateSpace.pstex_t}~~~~~~~~}
\caption{The state space $S$ of RBM in a wedge.}
\label{Figure:State:Space:Of:RBM:Wedge}
\end{center}
\end{figure}
\begin{definition}\label{vw}[Varadhan and Williams \cite{varadhan1985brownian}] A family of probability measures $\{P^z, z \in S\}$ on $(C_S,\mathcal{M})$ is said to solve
the submartingale problem if for each $z \in S$, the following 3 conditions hold,
\begin{enumerate}
\item~$P^z(Z(0)=z)~=~1$, \label{Condition:1:Submartingale}
\item For each $f \in C^2_b(S)$, the process
\begin{eqnarray*}
\left\{ f(Z(t)) - \frac{1}{2} \int_0^t \triangle f(Z(s))ds,~~t \geq 0 \right\}
\end{eqnarray*}
is a submartingale on $(C_S,\mathcal{M},\mathcal{M}_t,P^z)$ whenever $f$ is constant in a neighborhood of the origin and satisfies $D_i f \geq 0$ on $\partial S_i$ for $i=1,2,$ \label{Condition:2:Submartingale}
\item
$$E^z \left[ \int_0^{\infty} 1\{Z(t)=0\}dt\right] ~=~0.
$$ \label{Condition:3:Submartingale}
\end{enumerate}
\end{definition}
Now define
\begin{eqnarray*}
\alpha &=& \frac{\theta_1+\theta_2}{\xi}.
\end{eqnarray*}
In \cite{varadhan1985brownian}, it was shown that if $\alpha < 2$, then there exists a unique solution $\{P^z, z \in S\}$ to the submartingale problem. Moreover \cite{varadhan1985brownian}, the family $\{P^z, z \in S\}$ possess the strong Markov property. If $\alpha \geq 2$, then there is no solution \cite{varadhan1985brownian} to the submartingale problem but there is a unique family of probability measures $\{P^z, z \in S\}$ satisfying Conditions \ref{Condition:1:Submartingale} and \ref{Condition:2:Submartingale} of Definition \ref{vw}. In this case, we may continue to refer to $\{P^z, z \in S\}$ as the solution to the submartingale problem, with the understanding that only Conditions \ref{Condition:1:Submartingale} and \ref{Condition:2:Submartingale} of Definition \ref{vw} are satisfied.
Now, as in Williams \cite{williams1985recurrence}, let $\mathcal{B}_S$ denote the Borel $\sigma$-algebra on $S$ and for each probability measure $\mu$ on $(S,\mathcal{B}_S)$, define the measure $P^{\mu}$ on $(C_S,\mathcal{M})$ by setting
\begin{eqnarray*}
P^{\mu}(A)&=& \int_S \mu(dz)P^z(A),~A \in \mathcal{M}.
\end{eqnarray*}
We then denote by $(C_S,\mathcal{M}^{\mu},P^{\mu})$ the completion of $(C_S,\mathcal{M},P^{\mu})$
and for each $t \geq 0$, we let $\mathcal{M}^{\mu}_t$ be the augmentation of $\mathcal{M}_t$ with respect to $(C_S,\mathcal{M}^{\mu},P^{\mu})$. If $\mu=\delta_{\{z\}}$ for some $z\in S$ (the Dirac measure centered at $z$), then we shall write ${\mathcal M}^z$ and ${\mathcal M}_t^z$ instead of ${\mathcal M}^\mu$ and ${\mathcal M}_t^\mu$.
Next, define $\mathcal{F} = \bigcap_{\mu}{\mathcal M}^{\mu}$, where the intersection is taken over all $\mu$ which are probability measures on $(S,\mathcal{B}_S)$, and similarly set $\mathcal{F}_t = \bigcap_{\mu}M^{\mu}_t$. It then follows \cite{williams1985recurrence} that the filtration $\mathcal{F}_t$ is right-continuous and, moreover, the process $\mathcal{Z}= (C_S,\mathcal{F},\mathcal{F}_t,Z(t),\theta_t,P^z) $ is a Hunt process with state space $(S,\mathcal{B}_S)$, where $\theta_t$ is the usual shift operator.
\subsection{Dirichlet Process Result}
In \cite{williams1985reflected}, Williams proved that for each $z \in S$, $Z$ is a semi-martingale on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ if and only if $\alpha < 1$ or $\alpha \geq 2$ . More recently, Kang and Ramanan \cite{kang2010dirichlet} have proven that in the case of $\alpha=1$, $Z$ is a Dirichlet process (see Definition \ref{Definition:Dirichlet:Process} below) on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$. In Theorem \ref{Theorem:Main:Dirichlet:Process:Result} below, we complement the results of Williams \cite{williams1985reflected} and Kang and Ramanan \cite{kang2010dirichlet} by proving that in the case of $1 < \alpha < 2$, $Z$ is a Dirichlet process on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$.
For each $E \subseteq \mathbb{R}_{+}$, let
\begin{eqnarray*}
\pi(E) &=&\{\pi=\{t_i,i=0,...,n\} \subseteq E : n \in \mathbb{N}_0, ~t_{i-1} \leq t_i~,i=1,...,n\}
\end{eqnarray*}
denote the set of all partitions of $E$, and define the mesh of a partition $\pi \in \pi(E)$ by setting
\begin{eqnarray*}
\| \pi \| &=& \max\{|t_i-t_{i-1}|:i=1,...,n\}.
\end{eqnarray*}
We then recall the definition of a continuous process with zero energy (see, for instance, Definition 3.2 of Kang and Ramanan \cite{kang2010dirichlet}).
\begin{definition}\label{Definition:Zero:Energy}
Let $z \in S$. A $2$-dimensional continuous process $Y$ defined on $(C_S,\mathcal{F},\mathcal{F}_t,P^z) $ is said to be of zero energy if for each $T > 0$,
\begin{eqnarray*}
\sum_{t_i \in \pi^n}\|Y(t_i)-Y(t_{i-1})\|^2 &{\buildrel P \over \rightarrow} &0~\textrm{as}~n \rightarrow \infty,
\end{eqnarray*}
for any sequence $\{\pi^n, n \geq 1\}$ of partitions of $[0,T]$ with $\| \pi^n \| \rightarrow 0$ as $n \rightarrow \infty$.
\end{definition}
\noindent We now define a Dirichlet process as follows. (see Definition 2.4 of Coquet et al. \cite{coquet2006natural} or Definition 3.3 of \cite{kang2010dirichlet}).
\begin{definition}\label{Definition:Dirichlet:Process}
Let $z \in S$. The stochastic process $Z$ is said to be a Dirichlet process on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ if we may write
\begin{eqnarray}
Z &=& X + Y, \label{Display:Decomposition:Of:Z}
\end{eqnarray}
where $X$ is an $\mathcal{F}_t$-adapted local martingale and $Y$ is a continuous, $\mathcal{F}_t$-adapted zero energy process with $Y(0)=0$.
\end{definition}
\noindent The above definition of a Dirichlet process differs slightly from that of F\"{o}llmer \cite{follmer1981dirichlet}. However, it is sufficiently strong in order to enable one to obtain
formulas for performing certain operations on $Z$, such as a change-of-variables (see F\"{o}llmer \cite{follmer1981calcul}). Note also that if $Z$ is a semi-martingale, then it is also a Dirichlet process as defined above. However, the converse is not necessarily true. It may also be shown that the Doob-Meyer type decomposition of $Z$ given by \eqref{Display:Decomposition:Of:Z} is unique (see Remark 3.4 of \cite{kang2010dirichlet}).
\begin{theorem}\label{Theorem:Main:Dirichlet:Process:Result} Suppose that $1 < \alpha < 2$. Then, $Z$ has the decomposition
\begin{eqnarray}
Z&=&X+Y,\label{decomposition}
\end{eqnarray}
where $(X,Y)$ is a pair of processes on $(C_S,\mathcal{F},\mathcal{F}_t)$
such that for each $z \in S$, $X$ is a standard Brownian motion started from $z$ and $Y$ is a process of zero energy on $(C_S,\mathcal{F},\mathcal{F}_t, P^z)$. In particular, for each $z\in S$, the process $Z$ is a Dirichlet process on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$.
\end{theorem}
\noindent
\noindent We emphasize that the pair of processes $(X,Y)$ appearing in Theorem \ref{Theorem:Main:Dirichlet:Process:Result} do not depend on $z\in S$. That is, there is a single pair of processes $(X,Y)$ on $(C_S,\mathcal{F},\mathcal{F}_t)$ such that the statement of the above theorem hold for each $z\in S$.
\subsection{Roughness of the Paths of $Y$ Result}
Theorem \ref{Theorem:Main:Dirichlet:Process:Result} implies that in the case of $1 < \alpha < 2$, the paths of $Y$ are of zero energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$. We now turn our attention to providing a refinement of this result using the concept of strong $p$-variation.
\begin{definition}\label{Definition:Strong:P:Variation}
Let $E \subset \mathbb{R}_{+}$ and $p > 0$. The strong $p$-variation of a function $f: \mathbb{R}_{+} \mapsto \mathbb{R}^n$ on $E$ is defined by
\begin{eqnarray*}
V_p(f,E) &=& \sup \left\{ \sum_{t_i \in \pi}\|f(t_i)-f(t_{i-1})\|^p : \pi \in \pi(E)\right\}.
\end{eqnarray*}
\end{definition}
\begin{theorem}\label{Theorem:Converse:Dirichlet:Process:Result}Suppose that $1 < \alpha < 2$. Then, for each $p > \alpha $ and $z \in S$,
\begin{eqnarray}
P^z (V_p(Y,[0,T]) < + \infty) &=&1,~T \geq 0. \label{Display:Condition:Main:Dirichlet}
\end{eqnarray}
Furthermore, for each $0 < p \leq \alpha$,
\begin{eqnarray}
P^0 (V_p(Y,[0,T]) < + \infty) &=&0,~T \geq 0. \label{Display:Condition:Converse:Main:Dirichlet}
\end{eqnarray}
\end{theorem}
\noindent
\noindent Since $ \alpha < 2$ in the statement of Theorem \ref{Theorem:Converse:Dirichlet:Process:Result}, it is straightforward to show using the relationship between functions of finite strong $p$-variation and $1/p$-H\"{o}lder continuous functions \cite{chistyakov1998maps} that \eqref{Display:Condition:Main:Dirichlet} automatically implies that $Y$ is of zero energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$.
\subsection{Extended Skorokhod Map Result}
\label{Section:Extended:Skorokhod:Map:Results}
We now provide an additional connection between the pair of processes $(X,Y)$ appearing in Theorem \ref{Theorem:Main:Dirichlet:Process:Result}
and the process $Z$ itself. Let $D(\mathbb{R}_{+},\mathbb{R}^2)$ denote the space of $\mathbb{R}^2$-valued functions, with domain $\mathbb{R}_{+}$, that are right-continuous with left limits. Also, let $D_S(\mathbb{R}_{+},\mathbb{R}^2)$ be the set of $f \in D(\mathbb{R}_{+},\mathbb{R}^2)$ such that $f(0) \in S$. Next, let $d(\cdot)$ be a set-valued mapping defined on $ S$ such that $d(z)$ is a closed convex cone in ${\mathbb R}^2$ for every $z\in S$. In particular, we define
\begin{eqnarray}
d(z) &=&\begin{cases}
\{a v_1, a \geq 0\}, &\text{for $z \in \partial S_1\setminus\{0\}$,}\\
\{a v_2, a \geq 0\}, &\text{for $z \in \partial S_2\setminus\{0\}$,}\\
V, &\text{for}\ z=0,\\
\{0\},&\text{for}\ z\in\text{int}(S),\\
\end{cases} \label{Display:Definition:Of:Direction:Of:Reflection}
\end{eqnarray}
where the selection of the closed convex cone $V\subset{\mathbb R}^2$ will be discussed in a moment. For a set $A \subseteq \mathbb{R}^2$, let $\overline{\textrm{co}}(A)$ denote the closure of the convex hull of $A$. We then recall from Ramanan \cite{ramanan2006reflected} the definition
of the extended Skorokhod problem (ESP).
\begin{definition}\label{Definition:Extended:Skorohod:Problem}
The pair of processes $(\phi,\eta) \in D_S(\mathbb{R}_{+},\mathbb{R}^2) \times D(\mathbb{R}_{+},\mathbb{R}^2)$ solve the ESP $(S,d(\cdot))$ for $\psi \in D_S(\mathbb{R}_{+},\mathbb{R}^2)$ if $\phi(0)=\psi(0)$, and if for all $t \in \mathbb{R}_{+}$,
the following properties hold,
\begin{enumerate}
\item~$\phi(t) ~=~ \psi(t) + \eta(t)$, \label{Definition:ESP:Item:One}
\item~$\phi(t) \in S$, \label{Definition:ESP:Item:Two}
\item For every $s \in [0,t]$, \label{Definition:ESP:Item:Three}
\begin{eqnarray*}
\eta(t) - \eta(s) &\in& \overline{\mathrm{co}}[\cup_{u \in (s,t]}d(\phi(u))],
\end{eqnarray*}
\item $\eta(t)-\eta(t-) \in \overline{\mathrm{co}}[d(\phi(t))]$. \label{Definition:ESP:Item:Four}
\end{enumerate}
\end{definition}
\begin{theorem}\label{necessary}
Suppose that $1<\alpha<2$. Then, for each $z \in S$, the ESP $(S,d(\cdot))$ for the Brownian motion $X$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$
has a solution $P^z$-a.s. if and only if
\begin{equation}\overline{\mathrm{co}}(V\cup \{a v_1, a \geq 0\} \cup \{a v_2, a \geq 0\} )={\mathbb R}^2.\label{full:space}\end{equation}
In this case, $(Z,Y)$ solves the ESP $(S,d(\cdot))$ for $X$.
\end{theorem}
\noindent By trivially setting $V=\mathbb{R}^2$, it follows that one may always find a $V$ such that \eqref{full:space} holds. However, using the fact that $1 < \alpha < 2$, it is straightforward to verify that the smaller set $V=\{a v_0, a \geq 0 \}$ for any $v_0$ in the interior of $S$ satisfies \eqref{full:space} as well. Note also that we stop short of claiming in Theorem \ref{necessary} that $(Z,Y)$ is the unique solution to the ESP $(S,d(\cdot))$ for $X$.
\subsection{Organization of the Remainder of the Paper}
For the remainder of the paper, we assume that
\begin{eqnarray*}
&1 < \alpha < 2.&
\end{eqnarray*}
In Section \ref{Section:Excursion:Theory:Background}, we provide some useful results related to the zero set of RBM in a wedge. Much of the material found in Section \ref{Section:Excursion:Theory:Background} is based on the results of Williams \cite{williams1987local}. In Section \ref{Section:Identification:of:X:and:Skorokhod:Problem}, we identify the Brownian motion $X$ which appears in the Doob-Meyer decomposition of $Z$ given by Theorem \ref{Theorem:Main:Dirichlet:Process:Result}. Our proofs of Theorems \ref{Theorem:Main:Dirichlet:Process:Result}, \ref{Theorem:Converse:Dirichlet:Process:Result} and \ref{necessary} may be found in Sections \ref{section:dirichlet:process}, \ref{section:converse:dirichlet} and \ref{section:esp}, respectively.
\section{Results Related to the Zero Set of $Z$}
\label{Section:Excursion:Theory:Background}
The following results are helpful as we proceed throughout the paper. Let $\Lambda = \{t \in \mathbb{R}_{+} : Z(t)=0\}$ denote the zero set of $Z$ and note that since $Z$ is a continuous process for every $\omega
\in C_S$, it follows that $\Lambda$ is a closed set for every $\omega
\in C_S$. Moreover, since by Lemma 2.2 of Williams \cite{williams1987local} the origin is a regular point for $Z$, using standard arguments (see Theorem 2.9.6 of Karatzas and Shreve \cite{KaratzasShreve}) it may be shown that for each $z \in S$, the set $\Lambda$ contains no isolated points $P^z$-a.s. Next, let $\Lambda^C=\{t \in \mathbb{R}_{+} : Z(t) \neq 0\}$ denote the complement of $\Lambda$ in $\mathbb{R}_{+}$. It then follows that $\Lambda^C$ is an open set in $\mathbb{R}_{+}$, and in fact it may be written as
\begin{eqnarray}
\Lambda^C &=&[0,\tau_0) \cup \left( \bigcup_{i=1}^{\infty}(G_i,D_i) \right), \label{display:complement:of:lambda}
\end{eqnarray}
where $\tau_0$ is the first hitting time of the origin by $Z$, i.e.,
\begin{equation}\tau_0=\inf\{t>0: Z(t)=0\}.\label{tau:zero}\end{equation}
Note that $\tau_0=0$, $P^0$-a.s., in which case we interpret $[0,0)$ as the empty set.
For each $i \geq 1$, we refer to $(G_i,D_i)$ as an excursion interval of $Z$ away from the vertex and refer to the paths of $Z$ over such intervals as excursions from the vertex. It may easily be shown that the sequence of random times $\{(G_i,D_i), i \geq 1\}$ can be selected such that $(G_i,D_i):C_S \mapsto \mathbb{R}_{+}^2$ is a measurable function on $(C_S,\mathcal{F})$ for each $i \geq 1$. We assume for the remainder of the paper that such a selection has been made.
Closely related to the zero set $\Lambda$ is the concept of the local time of $Z$ at the origin. Since the origin is a regular point for $Z$, it follows by the results of Blumenthal and Getoor \cite{blumenthal2007markov} that there exists a continuous, non-decreasing, perfect, adapted functional $L$ with support $\{0\}$. We refer to this process as the local time of $Z$ at the origin. Moreover, the results of \cite{blumenthal2007markov} also guarantee that $L$ is uniquely determined up to a multiplicative constant. As in \cite{williams1987local}, we assume for the remainder of the paper that the multiplicative constant is chosen such that
\begin{eqnarray*}
E^0 \left[ \int_0^{\infty}e^{-t}dL(t) \right] &=&1.
\end{eqnarray*}
Now note that since $1 < \alpha < 2$, it is straightforward to show using Theorem 2.2 of Varadhan and Williams \cite{varadhan1985brownian}, together with the strong Markov property \cite{varadhan1985brownian} of $Z$, that $L(t) \rightarrow \infty$ as $t \rightarrow \infty$, $P^z$-a.s. for each $z \in $S. Hence, we may $P^z$-a.s. define the inverse local time process
\begin{eqnarray}
L^{-1}(a)&=&\inf\{t \geq 0 : L(t) > a\},~~a \geq 0, \label{def:inverse:local:time}
\end{eqnarray}
and we set $L^{-1}=\{L^{-1}(a), a \geq 0\}$. By Corollary 2.7 of Williams \cite{williams1987local}, under $P^0$, $L^{-1}$ is a stable subordinator of index $\alpha/2$. In particular, $\ln E^0[e^{-\lambda L^{-1}(a)}]=-a\lambda^{\alpha/2}$ for $a, \lambda \geq 0$. Hence, under $P^0$, $L^{-1}$ is a strictly increasing, right-continuous, pure jump process (see Bertoin \cite{bertoin1999subordinators}), and so we have that $P^0$-a.s.,
\begin{eqnarray*}
L^{-1}(a) &=& \sum_{0 \leq \kappa \leq a} \Delta L^{-1}(\kappa),~a \geq 0, \end{eqnarray*}
where
\begin{eqnarray*}
\Delta L^{-1}(\kappa) &=& L^{-1}(\kappa)-L^{-1}(\kappa -),~\kappa \geq 0,
\end{eqnarray*}
with the convention that $L^{-1}(0-)=0$. In particular, whenever $\Delta L^{-1}(\kappa)\not= 0$, then it is straightforward to show that $\Delta L^{-1}(\kappa)=D_i-G_i$ for the unique value of $i\ge 1$ such that $L(G_i)=L(D_i)=\kappa$. Thus, we also have that $P^0$-a.s.,
\begin{equation}
L^{-1}(a)~=~\sum_{i:L(G_i)\le a} (D_i-G_i),~a \geq 0.\label{l:inverse}\end{equation}
Finally, we note that using arguments similar to the proof of Corollary 5.11 (page 411) of \c{C}inlar \cite{cinlar2011probability}, it may be shown that $P^0$-a.s., $L^{-1}(\mathbb{R}_{+}) \subset \Lambda $ with
\begin{eqnarray}
\Lambda \backslash L^{-1}(\mathbb{R}_{+})&=& \bigcup_{i=1}^{\infty}\{G_i\}. \label{Display:Disjoint:Intervals}
\end{eqnarray}
\section{Identification of $X$ and The Standard Skorokhod Problem}
\label{Section:Identification:of:X:and:Skorokhod:Problem}
In this section, we prove the existence of a 2-dimensional process $X$ on $(C_s, {\mathcal F}, {\mathcal F}_t)$ which is a standard 2-dimensional Brownian motion started from $z$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$. Moreover, letting $Y=Z-X$, we show that over each excursion interval $[G_i,D_i]$ for $i\ge 1$ (and if $z\not= 0$, then also over the interval from zero to the first time that $Z$ hits the origin), the pair of processes $(Z,Y)$ solve the standard Skorokhod problem for $X$ given by Definition \ref{Definition:Skorohod:Problem} below.
In Section \ref{section:dirichlet:process}, it is proven
that the process $Y$ is of zero energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$, and hence $(X,Y)$
is the desired Doob-Meyer type decomposition of $Z$ given by Theorem \ref{Theorem:Main:Dirichlet:Process:Result}. We also note that our construction of the process $X$ is similar to that of Kang and Ramanan \cite{kang2014submartingale}, but differs in an important way in order to allow us to prove our results related the Skorokhod problem.
We now recall the definition of the standard Skorokhod problem (SP) as given in Ramanan \cite{ramanan2006reflected}.
Recall first from Section \ref{Section:Extended:Skorokhod:Map:Results} the definition of $D(\mathbb{R}_{+},\mathbb{R}^2)$ as the space of $\mathbb{R}^2$-valued functions, with domain $\mathbb{R}_{+}$, that are right-continuous with left limits, and the definition of $D_S(\mathbb{R}_{+},\mathbb{R}^2)$ as the set of $f \in D(\mathbb{R}_{+},\mathbb{R}^2)$ such that $f(0) \in S$. Also, let $\mathcal{B}V(\mathbb{R}_{+},\mathbb{R}^2) \subset
D(\mathbb{R}_{+},\mathbb{R}^2)$ denote the subset of functions in $D(\mathbb{R}_{+},\mathbb{R}^2)$ which have finite variation on each bounded interval. For each $f \in \mathcal{B}V(\mathbb{R}_{+},\mathbb{R}^2)$
and $t \geq 0$, we denote the variation of $f$ on $[0,t]$ by $\bar f(t)$, which is a shorthand notation for $V_1(f,[0,t])$. Next, recall from \eqref{Display:Definition:Of:Direction:Of:Reflection} of Section \ref{Section:Extended:Skorokhod:Map:Results}, the definition of the set-valued mapping $d(\cdot)$ defined on $S$,
and, for each $z \in S$, let $d^1(z)$ denote the intersection of $d(z)$ with the unit sphere $S_1(0)$ in $\mathbb{R}^2$ which is centered at the origin. We then recall from Ramanan \cite{ramanan2006reflected} the definition
of the standard Skorokhod problem (SP).
\begin{definition}\label{Definition:Skorohod:Problem}
The pair of processes $(\phi,\eta) \in D_S(\mathbb{R}_{+},\mathbb{R}^2) \times D(\mathbb{R}_{+},\mathbb{R}^2)$ solve the SP $(S,d(\cdot))$ for $\psi \in D_S(\mathbb{R}_{+},\mathbb{R}^2)$ if $\phi(0)=\psi(0)$, and if for all $t \in \mathbb{R}_{+}$,
the following properties hold,
\begin{enumerate}
\item~$\phi(t) ~=~ \psi(t) + \eta(t)$,
\item~$\phi(t) \in S$,
\item ~$ \eta\in \mathcal{B}V(\mathbb{R}_{+},\mathbb{R}^2) $,
\item ~$ \bar\eta(t)= \int_{[0,t]}1\{\phi(s) \in \partial S\}d\bar\eta(s)$ \label{Definition:SP:Item:Four}
\item There exists a measurable function $\gamma:\mathbb{R}_{+} \mapsto S_1(0)$ such that $d\bar\eta$-almost everywhere we have that $\gamma(t) \in d^1(\phi(t))$
and
\begin{eqnarray*}
\eta(t) &=& \int_0^t \gamma(s) d\bar\eta(s).
\end{eqnarray*}\label{Definition:SP:Item:Five}
\end{enumerate}
\end{definition}
Our main result of this section is the following.
\begin{theorem}\label{Proposition:Skorokhod:Problem}Suppose that $1 < \alpha < 2$.
\begin{enumerate}
\item \label{Proposition:Skorohod:Problem:Part:1}
There exists a 2-dimensional process $X$ on $(C_s, {\mathcal F}, {\mathcal F}_t)$ such that $X$ is a standard 2-dimensional Brownian motion started from $z$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$.
\item \label{Proposition:Skorohod:Problem:Part:2}
Let $X$ be as above, $Y=Z-X$, and let $V$ in \eqref{Display:Definition:Of:Direction:Of:Reflection} be an arbitrary, non-empty closed convex cone in ${\mathbb R}^2$.
Then, for each $z \in S$, $P^z$-a.s. the pair $(Z((G_i+\cdot)\wedge D_i),Y ((G_i+\cdot)\wedge D_i)-Y(G_i) )$ solves the SP $(S,d(\cdot))$ for $X ((G_i+\cdot)\wedge D_i) +Y(G_i)$ for each $i \geq 1$. In addition, for each $z\in S\setminus\{0\}$, $P^z$-a.s. the pair $(Z(\cdot\wedge \tau_0), Y(\cdot\wedge \tau_0))$ solves the SP $(S,d(\cdot))$ for $X (\cdot\wedge \tau_0)$.
\end{enumerate}
\end{theorem}
\noindent The two parts of Theorem \ref{Proposition:Skorokhod:Problem} are proven respectively in Sections \ref{Subsection:Proof:That:X:Exists} and \ref{Subsection:Semimartingale:Decomposition:Of:X:On:Excursion:Intervals} below. Also note that $d(0)=V$ in the definition of $d(\cdot)$, and the selection of $V$ turns out to be important in Theorem \ref{necessary}.
However, since $Z$ does not reach the origin on $(G_i,D_i)$, or on $[0,\tau_0)$ if $z\not= 0$, the selection of $V$ in the above theorem is irrelevant.
\subsection{Proof of Part \ref{Proposition:Skorohod:Problem:Part:1} of Theorem \ref{Proposition:Skorokhod:Problem} }\label{Subsection:Proof:That:X:Exists}
For each $ \delta > 0$, let $S_\delta \subset S$
be the closed set defined in polar coordinates by $S_\delta = S + \delta e^{i \xi/2}$. Next, set $\tau_0^{\delta}=0$, and, for each $k \geq 1$, recursively define
\begin{eqnarray*}
\sigma^{\delta}_k &=& \inf\{t \geq \tau^{\delta}_{k-1} : Z(t) \in S_{2 \delta}\}~~\textrm{and}~~\tau^{\delta}_k ~=~ \inf\{t \geq \sigma^{\delta}_{k} : Z(t) \in \partial S_{\delta}\}.
\end{eqnarray*}
Since $Z(\cdot,\omega)$ is continuous for every $\omega\in C_S$, we may equivalently set $\tau^{\delta}_k ~=~ \inf\{t \geq \sigma^{\delta}_{k} : Z(t) \in S_{\delta}^c \cap S^{o}\}$,
where $S_{\delta}^c$ denotes the complement of $S_{\delta}$ in $S$, and $S^{o}$ is the interior of $S$. It is immediate
by Proposition 2.1.5 of Ethier and Kurtz \cite{EK86} that $\sigma^{\delta}_k$
and $\tau^{\delta}_k$ are, for each $k \geq 1$, stopping times relative to the right-continuous filtration $\{\mathcal{F}_t, t \geq 0\}$.
Now, for each $k \geq 1$, let $W_{(k)}^{\delta}$ be the process defined by setting
\begin{eqnarray*}
W_{(k)}^{\delta}(t)&=& Z(t \wedge \tau_{k}^{\delta})-Z(t \wedge \sigma_{k}^{\delta}),~t \geq 0,
\end{eqnarray*}
and then define the process $W^{\delta}$ by setting
\begin{eqnarray*}
W^{\delta}(t) &=& \sum_{k=1}^{\infty}W_{(k)}^{\delta}(t),~t \geq 0.
\end{eqnarray*}
Recall next from Section \ref{Section:Introduction} the definition of the augmented sigma fields ${\mathcal M}^z$ and ${\mathcal M}_t^z$ for $z\in S$ and $t \geq 0$. For each $z \in S$, let $\Upsilon_2^{c,z}$ denote the space of continuous, square-integrable martingales $M$ with time horizon $\mathbb{R}_{+}$ on $(C_S,\mathcal{M}^z,\mathcal{M}^z_t,P^z)$, and such that $M(0)=0$. It is well-known (see for instance Proposition
1.5.23 of Karatzas and Shreve \cite{KaratzasShreve}) that the space $\Upsilon_2^{c,z}$ is complete under the norm
\begin{eqnarray}
\| M \|_z &=& \sum_{n=1}^{\infty} \frac{1}{2^n}\left(\sqrt{E^z[M^2(n)]} \wedge 1\right),~ M \in \Upsilon_2^{c,z}. \label{norm:def}
\end{eqnarray}
Now, for each $\delta > 0$ and $k \geq 1$ and $i=1,2,$ denote the $i$th component process of $W_{(k)}^{\delta}$ by $W_{(k),i}^{\delta}$. We begin our proof of Part \ref{Proposition:Skorohod:Problem:Part:1} of Theorem \ref{Proposition:Skorokhod:Problem} by recalling a result from Kang and Ramanan \cite{kang2014submartingale}.
\begin{lemma}\label{Lemma:Xk:Martingale}
For each $\delta > 0$, $k \geq 1,i=1,2$, and $z\in S$, the process $W_{(k),i}^{\delta} \in \Upsilon_2^{c,z}$ with quadratic variation processes
\begin{eqnarray*}
\langle W_{(k),i}^\delta \rangle_t &=& (t \wedge \tau_k^\delta) - (t \wedge \sigma_k^\delta)~,t \geq 0,
\end{eqnarray*}
for $i=1,2,$ and $\langle W_{(k),1}^\delta,W_{(k),2}^\delta \rangle_t = 0, t \geq 0$.
\end{lemma}
\begin{proof}
This follows from Lemma 5.1 in \cite{kang2014submartingale}.
\end{proof}
Next, we prove that $W_{(k)}^{\delta}$ and $W_{(\ell)}^{\delta}$ and orthogonal to one another for $k \neq \ell$.
\begin{lemma}\label{Lemma:Xk:Xl:Orthogonal}For each $\delta > 0$ and $k, \ell \geq 1$ with $k \neq \ell$, we have under $P^z$ for every $z\in S$ that
\begin{eqnarray*}
\langle W_{(k),i}^{\delta},W_{(\ell),j}^{\delta} \rangle_t &=& 0,~ t \geq 0,
\end{eqnarray*}
for $i,j=1,2$.
\end{lemma}
\begin{proof}Let $\delta > 0$ and assume without loss of generality that $\ell > k \geq 1$. The result now follows from the fact that $W_{(k)^,i}^\delta$ is flat outside of $(\sigma_k^\delta,\tau_k^\delta)$ for $i=1,2,$ and
$W_{(\ell),i}^\delta$ is flat outside of $(\sigma_{\ell}^\delta,\tau_{\ell}^\delta)$ for $i=1,2,$ together with the fact
that $(\sigma_k^\delta,\tau_k^\delta) \cap (\sigma_{\ell}^\delta,\tau_{\ell}^\delta) = \emptyset $.
Indeed, let $\{\pi_n=\{0=t^n_0 < t^n_1 < ... < t^n_{s(n)}=t\}, n \geq 1 \}$ be such that $\| \pi_n \| \rightarrow 0$ as $n \rightarrow \infty$,
and let
\begin{eqnarray*}
C_{k,i,\ell,j}(\pi_n) &=& \sum_{m=1}^{s(n)}(W_{(k),i}^\delta(t^n_m)-W_{(k),i}^\delta(t^n_{m-1}))(W_{(\ell),j}^\delta(t^n_m)-W_{(\ell),j}^\delta(t^n_{m-1}))
\end{eqnarray*}
for $i,j=1,2$. Then, $C_{k,i,\ell,j}(\pi_n)=0$ whenever $\| \pi_n \| < \sigma_l^\delta - \tau_k^\delta$. Thus, $P^z$-a.s. for each $z \in S$ we have $C_{k,i,\ell,j}(\pi_n) \rightarrow 0$
as $n \rightarrow \infty$.
\end{proof}
\begin{lemma}\label{Wdelta}
For each $z\in S$,\ $\delta>0$ and $i=1,2,$ we have that $W_i^\delta\in\Upsilon^{c,z}_2$ with quadratic and cross variations
\begin{equation}\langle W^{\delta}_i \rangle_t = \sum_{k=1}^{\infty}((t \wedge \tau_k^\delta)-(t \wedge \sigma_k^\delta)),~t \geq 0,\label{qv}\end{equation}
and $\langle W_1^\delta,X_2^\delta\rangle_t=0,t \geq 0$.
\end{lemma}
\begin{proof}
Let $\delta > 0$ and $z \in S$. We first identify the quadratic variation processes of $W^{\delta}$. Note that for each $n \geq 1$ we may write
\begin{eqnarray*}
W^{\delta}(t \wedge \tau_n^\delta) &=& \sum_{k=1}^{n}W_{(k)}^\delta(t),~t \geq 0,
\end{eqnarray*}
which by Lemma \ref{Lemma:Xk:Martingale} implies that $W^{\delta}$ is a local martingale on $(C_S,{\mathcal M},{\mathcal M}_t,P^z)$. Also by Lemma \ref{Lemma:Xk:Martingale} it
follows that $W^{\delta}(\cdot \wedge \tau_n^\delta) \in \Upsilon_2^{c,z}$. For $n \geq 1$ let
\begin{eqnarray*}
V_n(t) &=& \sum_{k=1}^{n}((t \wedge \tau_k^\delta)-(t \wedge \sigma_k^\delta))~~\textrm{and}~~V(t) ~=~ \sum_{k=1}^{\infty}((t \wedge \tau_k^\delta)-(t \wedge \sigma_k^\delta)),~t \geq 0.
\end{eqnarray*}
By Lemmas \ref{Lemma:Xk:Martingale} and \ref{Lemma:Xk:Xl:Orthogonal} and the definitions of $V_n(t)$ and $V(t)$ it now follows that
\begin{eqnarray*}
\langle W^{\delta}_i(\cdot \wedge \tau_n^\delta) \rangle_t &=&\sum_{k=1}^{n} \langle W^{\delta}_{(k),i} \rangle_t ~=~V_n(t)~=~V(t \wedge \tau_n^\delta),~t \geq 0,
\end{eqnarray*}
for $i=1,2,$ and hence $\{(W^{\delta}_i(t \wedge \tau_n^\delta))^2-V(t \wedge \tau_n^\delta), t \geq 0 \}$ is a continuous martingale on on $(C_S,{\mathcal M},{\mathcal M}_t,P^z)$
which implies \eqref{qv}. Also since $\langle W^{\delta}_i \rangle_t=V(t) \leq t$ we
have that $W^{\delta}_i \in \Upsilon_2^{c,z}$ for $i=1,2$. {The proof that $\langle W^{\delta}_1,W^{\delta}_2 \rangle_t=0,~t \geq 0$,
follows in a similar manner.}
\end{proof}
Let $\Gamma$ be the set of sequences $\{a(n), n\ge 1\}$ such that $a(n)\downarrow 0$ as $n\to\infty$ and $2a(n+1)<a(n)$ for every $n\ge 1$.
\begin{lemma}
For every sequence $\{\delta(n),n\ge 1\}\in\Gamma$ we have that $\{W^{\delta(n)}_i, n\ge 1 \}$ is Cauchy in $\Upsilon^{c,z}_2$ for every $z\in S$ and $i=1,2$.
\end{lemma}
\begin{proof} Let $\{\delta(n), n \geq 1\} \in\Gamma$. Clearly it is sufficient to prove that $\{W_i^{\delta(n)}(T), n \geq 1\}$ is Cauchy
in $L^2(C_S,{\mathcal F}, P^z)$ for each $T \geq 0$. Suppose that for each $n > m \geq 1$ we show that
\begin{equation}
E^z[(W_i^{\delta(n)}(T)-W_i^{\delta(m)}(T))^2 ] \leq E^z \int_0^T 1\{Z(t) \in S \backslash S_{2 \delta (m)}\}dt. \label{Display:Inequality:Using:Integral}
\end{equation}
Then by the Dominated Convergence Theorem and Lemma 4.2 of Williams \cite{williams1985recurrence} we have that
\begin{eqnarray}
\lim_{m \rightarrow \infty} E^z \left[ \int_0^T 1\{Z(t) \in S \backslash S_{2 \delta (m)}\}dt \right] ~=~E^z \left[ \int_0^T 1\{Z(t) \in \partial S \} dt
\right] &=&0, \label{Display:Integral:Goes:To:Zero}
\end{eqnarray}
and so $\{W_i^{\delta(n)}(T), n \geq 1\}$ is Cauchy in $L^2(C_S,{\mathcal F}, P^z)$ as desired.
Hence we need to show only \eqref{Display:Inequality:Using:Integral}, and that is what we shall do in the rest of this proof. We first claim that for each $k \geq 1$ and $n > m \geq 1$ there exists an $\ell \geq 1$ such that $[\sigma_k^{\delta(m)},\tau_k^{\delta(m)}] \subseteq [\sigma_{\ell}^{\delta(n)},\tau_{\ell}^{\delta(n)}] $.
Suppose that $t \in [\sigma_k^{\delta(m)},\tau_k^{\delta(m)}] $. Then $Z(t) \in S_{\delta(m)}$. Moreover since $2 \delta(n) < \delta(m)$, if
$Z(t) \in S_{\delta(m)}$ then $t \in [\sigma_{\ell}^{\delta(n)},\tau_{\ell}^{\delta(n)}] $ for some $\ell \geq 1$. We
now claim that the choice of $\ell \geq 1$ does not depend on the choice of $t \in [\sigma_k^{\delta(m)},\tau_k^{\delta(m)}]$.
Suppose that $t_1, t_2 \in [\sigma_k^{\delta(m)},\tau_k^{\delta(m)}]$ with $t_1 \in [\sigma_{\ell}^{\delta(n)},\tau_{\ell}^{\delta(n)}]$
and $t_2 \in [\sigma_{p}^{\delta(n)},\tau_{p}^{\delta(n)}]$ for some $p > \ell$. Then $\sigma_p^{\delta(n)} \in [\sigma_k^{\delta(m)},\tau_k^{\delta(m)}]$ which is impossible since $Z(\sigma_p^{\delta(n)}) \in \partial S_{2 \delta(n)}$
and so the result follows.
Now write
\begin{equation*}
E^z[(W_i^{\delta(n)}(T)-W_i^{\delta(m)}(T))^2 ]
= E^z[\langle W_i^{\delta(n)}-W_i^{\delta(m)} \rangle_T ]\end{equation*}
\begin{eqnarray*}&=&E^z \left[ \sum_{k=1}^{\infty} \left[ \langle W_i^{\delta(n)}-W_i^{\delta(m)} \rangle_{\tau_k^{\delta(m)} \wedge T}- \langle W_i^{\delta(n)}-W_i^{\delta(m)} \rangle_{\sigma_k^{\delta(m)} \wedge T} \right] \right]\\
&+&E^z \left[ \sum_{k=1}^{\infty} \left[ \langle W_i^{\delta(n)}-W_i^{\delta(m)} \rangle_{\sigma_k^{\delta(m)} \wedge T}- \langle W_i^{\delta(n)}-W_i^{\delta(m)} \rangle_{\tau_{k-1}^{\delta(m)} \wedge T} \right] \right].
\end{eqnarray*}
We analyze each of the terms on the right hand side above separately. Regarding the first term recall that for each $k \geq 1$ and $n > m \geq 1$ there exists an $\ell \geq 1$ such that $[\sigma_k^{\delta(m)},\tau_k^{\delta(m)}] \subseteq [\sigma_{\ell}^{\delta(n)},\tau_{\ell}^{\delta(n)}] $. It
then follows that for $k \geq 1$ and $n > m$,
\begin{eqnarray*}
W_i^{\delta(m)}(T \wedge \tau_k^{\delta(m)})- W_i^{\delta(m)}(T \wedge \sigma_k^{\delta(m)})&=& W_i^{\delta(n)}(T \wedge \tau_k^{\delta(m)})- W_i^{\delta(n)}(T \wedge \sigma_k^{\delta(m)}),
\end{eqnarray*}
hence by Proposition IV.1.13 in \cite{revuzyor} the first term is zero.
Regarding the second term note that by \eqref{qv} for every $\delta>0$, $0\le s<t$, and $i=1,2$ we have
\begin{equation*}\langle W^{\delta}_i \rangle_t - \langle W^{\delta}_i \rangle_s \le t-s,
\end{equation*}
and on the interval $[\tau^{\delta(m)}_{k-1}, \sigma_k^{\delta(m)}]$ the process $W^{\delta(m)}$ is flat. Then it follows that
for $i=1,2,$ we have
\begin{equation*}
\sum_{k=1}^{\infty} \left[ \langle W_i^{\delta(n)}-W_i^{\delta(m)} \rangle_{\sigma_k^{\delta(m)} \wedge T}- \langle W_i^{\delta(n)}-W_i^{\delta(m)} \rangle_{\tau_{k-1}^{\delta(m)} \wedge T} \right]
= \sum_{k=1}^{\infty} \left[ \langle W_i^{\delta(n)}\rangle_{\sigma_k^{\delta(m)} \wedge T}- \langle W_i^{\delta(n)}\rangle_{\tau_{k-1}^{\delta(m)} \wedge T} \right] \end{equation*}
\begin{equation*}\le\sum_{k=1}^{\infty} \left[ (\sigma_k^{\delta(m)} \wedge T) - (\tau_{k-1}^{\delta(m)} \wedge T) \right].
\end{equation*}
Now note that if $t \in [\tau_{k-1}^{\delta(m)},\sigma_k^{\delta(m)}]$ for some $k \geq 1$ then $Z(t) \in S \backslash S_{2\delta(m)}$, thus
\begin{equation}
\sum_{k=1}^{\infty} \left[ (\sigma_k^{\delta(m)} \wedge T) - (\tau_{k-1}^{\delta(m)} \wedge T) \right]=\int_0^T \sum_{k=1}^{\infty} 1\{t \in [\tau_{k-1}^{\delta(m)},\sigma_{k}^{\delta(m)} ]\} dt
\leq \int_0^T 1\{Z(t) \in S \backslash S_{2 \delta (m)}\}dt, \label{bound}
\end{equation}
and \eqref{Display:Inequality:Using:Integral} follows.
\end{proof}
\begin{proposition}\label{Proposition:Brownian:Motion}
There exists a process $W$ on $(C_S,{\mathcal F}, {\mathcal F_t})$ such that for every sequence $\{\delta(n),n\ge 1\}\in\Gamma$ and every $z\in S$, the process $W$ is a standard 2-dimensional Brownian motion under $P^z$ started at zero, and
\begin{equation} \lim_{n\to\infty} W_i^{\delta(n)}=W_i\label{L2limit},~i=1,2,\end{equation}
in the norm topology given in \eqref{norm:def}.
\end{proposition}
\begin{proof}
{Let $z \in S$ and for $i=1,2,$ let $W_i^z$ be the unique limit point of the Cauchy sequence $\{W_i^{\delta(n)}, n \geq 1\}$ in the complete space $\Upsilon_2^{c,z}$. We claim that $W^z=(W_1^z,W_2^z)$ is a 2-dimensional standard Brownian motion on $(C_S,{\mathcal M},{\mathcal M}_t,P^z)$.} Since $W^z \in {M}_2^{c,z}$ it suffices to characterize the finite-dimensional distributions of $W^z$. By Lemma \ref{Wdelta}, \eqref{bound},
\eqref{Display:Integral:Goes:To:Zero}, and the Martingale Invariance Principle (see Theorem 7.1.4 of \cite{EK86}) it follows that $W^{\delta(n)} \Rightarrow W^0$ as $n \rightarrow \infty$ under $P^z$, where $W^0$ is a standard 2-dimensional Brownian motion. This implies that for each $m \geq 1$ and
$0 \leq t_1 < t_2 < ... < t_m$, $(W^{\delta(n)}(t_1),...,W^{\delta(n)}(t_m)) \Rightarrow (W^{0}(t_1),...,W^{0}(t_m))$
as $n \rightarrow \infty$. On the other hand, for each $i=1,2,$ the sequence $\{W_i^{\delta(n)}, n \geq 1\}$ is Cauchy in $\Upsilon_2^{c,z}$
with unique limit point $W_i^z$ and so $W_i^{\delta(n)}(t_j) \rightarrow W_i^z(t_j)$ in $L^2(P^z)$ as $n \rightarrow \infty$
for each $j=1,...,m$. Thus there exists a subsequence $\{\delta(n(k)), k \geq 1\}$ such that $(W^{\delta(n(k))}(t_1),...,W^{\delta(n(k))}(t_m)) \rightarrow (W^z(t_1),...,W^z(t_m)), P^z$-a.s. as $k \rightarrow \infty$. But this then implies
that $W^z$ and $W^0$ have the same finite dimensional distributions, thus $W^z$ is indeed a standard Brownian motion under $P^z$. We still need to show the existence of a process $W$ on $(C_S,{\mathcal F}, {\mathcal F}_t)$ such that for every $z\in S$ the process $W^z$ is indistinguishable from $W$ under $P^z$. Notice that by \eqref{L2limit} for every $t\ge 0$ we have that $W^{\delta(n)}(t)\to W^z (t)$ in probability as $n\to \infty$, and the existence of such $W$ follows from Cinlar, Jacod, Protter, and Sharpe \cite{cinlar1980semimartingales}, Lemma 3.29. Note also that Hypothesis 3.1 and condition (iv) at the beginning of Section 3a in \cite{cinlar1980semimartingales} are satisfied by \cite{varadhan1985brownian}, Corollary 3.3 and Theorem 3.14.
\end{proof}
\begin{proof}[Proof of Part \ref{Proposition:Skorohod:Problem:Part:1} of Theorem \ref{Proposition:Skorokhod:Problem}]
We define
\begin{equation} X(t,\omega)=W(t,\omega)+\omega(0),\ t\ge 0,\label{X:def}\end{equation}
and note that by Proposition \ref{Proposition:Brownian:Motion} for every $z\in S$ the 2-dimensional process $X$ is indeed a standard Brownian motion on $(C_S,{\mathcal F}, {\mathcal F_t}, P^z)$ started at $z$, exactly as required.
\end{proof}
\subsection{Proof of Part \ref{Proposition:Skorohod:Problem:Part:2} of Theorem \ref{Proposition:Skorokhod:Problem}}
\label{Subsection:Semimartingale:Decomposition:Of:X:On:Excursion:Intervals}
Let the set of continuous functions $\omega:[0,\infty)\mapsto{\mathbb R}^2$ be denoted by $C_{\mathbb{R}^2}$, which is an extension of $C_S$.
On this space, for each $t\ge s \geq 0$, let $\tilde{\mathcal M}_t=\sigma\{\omega(s):0\le s\le t\}$, the sigma-field of subsets of $C_{{\mathbb R}^2}$ generated by the coordinate maps $\omega\mapsto\omega(s)$ for $0\le s\le t$. Similarly, define $\tilde{\mathcal M}=\sigma\{\omega(s): s\in[0,\infty) \}$. As explained in Williams \cite{williams1985reflected}, these $\sigma$-fields represent the natural extensions of ${\mathcal M}_t$ and ${\mathcal M}$ from $C_S$ to $C_{\mathbb{R}^2}$.
Let $\tilde X:[0,\infty)\times C_{\mathbb{R}^2}\mapsto\mathbb{R}^2$ be the coordinate-mapping process
\begin{eqnarray*}
\tilde X(t,\omega)&=&\omega(t),~~ t\ge 0,~w\in C_{\mathbb{R}^2}.
\end{eqnarray*}
For each $z\in S$, let $Q^z$ be a probability measure on $\tilde{\mathcal M}$ such that $\tilde X$ is a Brownian motion starting at $z$ under $Q^z$. Now let $R$ be the $2\times 2$ matrix whose column vectors are $v_1$ and $v_2$.
The set $\mathcal A$ defined below will play a central role in this subsection.
\begin{definition}\label{Definition:Of:The:Set:A}
Let ${\mathcal A}\subset C_S\times C_{\mathbb{R}^2}$ be the set that consists of all pairs $(z,y)\in C_S\times C_{\mathbb{R}^2}$ satisfying the following two properties:
\begin{enumerate}
\item There exists $u\in C_{\mathbb{R}^2}$ such that both components of $u$ are non-decreasing, and y(t)=Ru(t) for all $t\ge 0$, \label{Item:One:The:Set:A}
\item For $j=1,2,$ the $j^{ th }$ component $u_j$ of $u$ increases only at times $t \geq 0$ when $z(t)\in\partial S_j$, i.e., $\int_0^\infty 1\left\{z(v)\in S\setminus\partial S_j\right\}du_j(v)=0$. \label{Item:Two:The:Set:A}
\end{enumerate}
\end{definition}
\noindent We then have the following result.
\begin{lemma}
The set ${\mathcal A}$ is measurable, that is, ${\mathcal A}\in{\mathcal M}\times\tilde{\mathcal M}$.
\end{lemma}
\begin{proof} Let $\tilde {\mathcal A}\subset C_S\times C_{\mathbb{R}^2}$ be a set such that $(z,u)\in\tilde {\mathcal A}$ if both components of $u$ are non-decreasing, and part \ref{Item:Two:The:Set:A} of the above definition holds. We then have ${\mathcal A}=\{(z,Ru):(z,u)\in\tilde {\mathcal A}\}$ and thus it is sufficient to show that $\tilde {\mathcal A}$ is a measurable set.
{It is shown in Stroock and Varadhan \cite{stroock2007multidimensional}}, page 30, that ${\mathcal M}\times\tilde {\mathcal M}$ is exactly the class of Borel sets in $C_S\times C_{\mathbb{R}^2}$ under the topology of uniform convergence on compact sets. It is therefore sufficient to show that $\tilde {\mathcal A}$ is closed under the above topology. In order to show this, let $\{(z_n,u_n),n\ge 1\}\subset \tilde {\mathcal A}$ be such that
$(z_n,u_n)\to (z,u)$ uniformly on compact sets. Since both components of $u_n$ are non-decreasing for each $n
\geq 1$, both components of $u$ must be non-decreasing as well. Next we show that $(z,u)$ satisfies Condition \ref{Item:Two:The:Set:A} of Definition \ref{Definition:Of:The:Set:A}. Let $t > 0$ such that $z(t)\notin \partial S_j$, where $j$ is either 1 or 2. {We need to show that there exists a neighborhood $(a,b)$ of $t$ such that $u_j$ is flat on $(a,b)$, where $u_j$ is the $j^{\rm th}$ component of $u$. From $z(t)\notin \partial S_j$ follows that there exists a closed ball $B$ centered around $z(t)$ such that $B\cap \partial S_j=\emptyset$. By the continuity of $z$ there exists a neighborhood $(a,b)$ of $t$ such that $z(q)\in B$ for all $q\in(a,b)$. Then by the uniform convergence of $z_n$ to $z$ there exists another closed ball $C\supset B$ centered also around $z(t)$ and $N\in{\mathbb N}_+$such that $C\cap \partial S_j=\emptyset$, and for all $n\ge N$ we have $z_n(q)\in C$ for all $q\in(a,b)$. But $(z_n,u_n)\in \tilde {\mathcal A}$ implies that $u_{n,j}$ (the $j^{\rm th}$ component of $u_n$) must be flat on $(a,b)$ for $n\ge N$, hence $u_j$ is also flat on $(a,b)$.} Finally, if $t=0$ and $z(t)\notin \partial S_j$, then we already saw that there exists $b>0$ such that $u_j$ is flat on $(0,b)$. But then the continuity of $u_j$ implies that it is not increasing in $t=0$.
\end{proof}
The next proposition is a restatement of the results on page 163 and of Theorem 1 of Williams \cite{williams1985reflected}.
\begin{proposition}\label{williams}
There exists a unique pair $(\tilde Z, \tilde Y)$ of continuous, adapted processes on $(C_{\mathbb{R}^2},\tilde{\mathcal M},\tilde{\mathcal M}_t)$
(both $\tilde Z$ and $\tilde Y$ are two-dimensional) such that\\
\begin{equation} \tilde Y(0)=0,\ \tilde Z(t)\in S\ \hbox{for each}\ t\ge 0,\label{A4}\end{equation}
\begin{equation}\tilde Y(t\vee\tilde\tau_0)=\tilde Y(\tilde\tau_0),\quad t\ge 0,\label{A4.1}\end{equation}
\begin{equation}
\tilde Z(t)=\begin{cases} \tilde X(t)+\tilde Y(t) &\mbox{if }t\le\tilde\tau_0,\\
0 &\mbox{if } t\ge \tilde\tau_0,
\end{cases}\label{A5}
\end{equation}
where $\tilde \tau_0=\inf\{t\ge 0:\tilde Z(t)=0\}$,
\begin{equation*} (\tilde Z,\tilde Y)\in {\mathcal A}.
\end{equation*}
Furthermore, for every $z\in S$, $Z(\cdot\wedge\tau_0)$ has the same law under $P^z$ as $\tilde Z(\cdot\wedge\tilde \tau_0)=\tilde Z(\cdot)$ has under $Q^z$, i.e., for any $B\in{\mathcal M}$
\begin{equation*}P^z\left(Z(\cdot\wedge\tau_0)\in
B\right) = Q^z\left(\tilde Z(\cdot\wedge\tilde\tau_0)\in B\right).
\end{equation*}
\end{proposition}
Recall now the definitions of the stopping times $\tau_k^{\delta},\sigma_k^\delta$ from the beginning of Section \ref{Section:Identification:of:X:and:Skorokhod:Problem}. Since $Z$ is the coordinate-mapping process, these definitions could be cast in the form $\sigma_k^\delta(w)=\inf\{t\ge \tau_{k-1}^\delta: w(t)\in S_{2\delta}\}$ and $\tau_k^\delta(w)=\inf\{t\ge \sigma_k^\delta: w(t)\in \partial S_{\delta}\}$ for $w\in C_S$. We can also write $\tau_0=\inf\{t>0: w(t)=0\}$ for $w\in C_S$. In principle, some of these stopping times may be infinity for some $w\in S$, but by the continuity of $\omega$ we have
$\lim_{k\to\infty}\sigma_k^\delta(\omega) = \lim_{k\to\infty}\tau_k^\delta(\omega) = \infty$ for all $\omega\in C_S$.
For $\delta>0$, we define the measurable mapping $I_{\delta} : C_S\mapsto C_{\mathbb{R}^2}$ as
\begin{equation*} I_{\delta}(\omega)(\cdot)=\sum_{k=1}^\infty \left[ \omega\left(\cdot\wedge\tau_k^\delta(\omega)\right) -
\omega\left(\cdot\wedge\sigma_k^\delta(\omega)\right)\right].
\end{equation*}
By the continuity of $\omega$, for any $t\ge 0$, in the infinite sum defining $I_{\delta}(\omega)(t)$ all terms except finitely many are zero.
We also note that
\begin{equation}I_{\delta}(\omega)(\cdot)=W^\delta (\cdot,\omega).\label{I}\end{equation}
Now let $\underline 0$ be the zero function in $C_{\mathbb{R}^2}$.
\begin{definition} For any given sequence $\{\gamma(n),n\ge 1\} \in\Gamma$, we define the mapping $F_{\gamma}:C_S\mapsto C_{\mathbb{R}^2}$ in the following way: if there exists a function $\tilde \omega \in C_{\mathbb{R}^2}$ such that for each $m\in{\mathbb N}_+$,
\begin{equation*}\sup_{t\le m}\|I_{\gamma(n)}(\omega)(t)-\tilde \omega(t)\|^2 \to 0 \hbox{ as } n\to\infty,
\end{equation*}
then $F_{\gamma}(\omega)=\tilde\omega$. If such $\tilde\omega$ does not exist, then let $F_{\gamma}(\omega)$ be $\underline 0$.
\end{definition}
Notice that for $\delta>0$, $T\ge 0$, and $\omega\in C_S$,
\begin{equation*} I_{\delta}(\omega)(\cdot\wedge T)=I_{\delta}(\omega(\cdot\wedge T)),
\end{equation*}
hence for all $\omega\in C_S$ such that $F_\gamma(\omega)\not=\underline 0$, we have
\begin{equation}F_\gamma(\omega)(\cdot\wedge T)=F_\gamma(\omega(\cdot\wedge T)).\label{stopping}\end{equation}
\begin{lemma}
$F_{\gamma}$ is ${\mathcal M}/\tilde{\mathcal M}$ measurable.
\end{lemma}
\begin{proof}This follows immediately from \cite{brown:pearcy:operator:theory}, page 111, point R.\end{proof}
\begin{lemma}\label{shift:I} Let $\delta>0$, $T\ge 0$, $\omega\in C_S$ such that
\begin{equation}\omega(T)\in S_{2\delta}^0\cup(S\setminus S_\delta),\label{condition}\end{equation}
where $S_{2\delta}^0$ is the interior of $S_{2\delta}$.
Then
\begin{equation} I_\delta(\omega(T+\cdot))(t)= I_\delta(\omega)(T+t)-I_\delta(\omega)(T),\ t\ge 0.\label{I:shifted}\end{equation}
\end{lemma}
\begin{proof} Let $\tau_0^{\delta,T}(\omega)=T$, $\sigma_k^{\delta,T}(\omega)=\inf\{t\ge \tau_{k-1}^{\delta,T}(\omega):\omega(t)\in S_{2\delta}\}$ and
$\tau_k^{\delta,T}(\omega)=\inf\{t\ge \sigma_{k}^{\delta,T}(\omega):\omega(t)\in \partial S_{\delta}\}$ for $k\ge 1$.
Condition \eqref{condition} implies that there exists $n\in{\mathbb N}_+$ such that $\sigma_n^\delta(\omega)\le \sigma_1^{\delta,T}(\omega)$, $\tau_n^\delta(\omega)=\tau_1^{\delta,T}(\omega)$, and $[\sigma_{n+k}^\delta(\omega),\tau_{n+k}^\delta(\omega)]=[\sigma_{k+1}^{\delta,T}(\omega),\tau_{k+1}^{\delta,T}(\omega)]$ for $k\ge 1$ (this can be checked by looking at the cases $\omega(T)\in S_{2\delta}^0$ and $\omega(T)\in S\setminus S_\delta$ separately). Then clearly both sides of \eqref{I:shifted} are equal to
$$\sum_{k=n+1}^\infty \left\{\omega(\tau_k^\delta(\omega)\wedge t) - \omega(\sigma_k^\delta(\omega)\wedge t)\right\} + \omega(\tau_n^\delta(\omega)\wedge t) -\omega(T\vee\sigma_n^\delta).$$
\end{proof}
\begin{lemma} For every sequence $\gamma\in\Gamma$, $T\in[0,\infty)$, and $\omega\in C_S$ such that $F_\gamma(\omega)\not=\underline 0$, we have
\begin{equation}F_\gamma(\omega(T+\cdot))=F_\gamma(\omega)(T+\cdot) - F_\gamma(\omega)(T).\label{F:shifted}\end{equation}
\end{lemma}
\begin{proof}
For sufficiently large $n$ we have that $\omega(T)\in S_{2\delta(n)}^0\cup(S\setminus S_\delta)$, hence by Lemma \ref{shift:I} for these values of $n$ we have
\begin{equation} I_{\gamma(n)}(\omega(T+\cdot))= I_{\gamma(n)}(\omega)(T+\cdot)-I_{\gamma(n)}(\omega)(T).\label{shifted:n}
\end{equation}
By our assumption that $F_\gamma(\omega)\not=\underline 0$ we have that $I_{\gamma(n)}(\omega)(T+\cdot)\to F_\gamma(\omega)(T+\cdot)$ as $n\to\infty$, in the topology of uniform convergence on compacts. Also, $I_{\gamma(n)}(\omega)(T)\to F_\gamma(\omega)(T)$ as $n\to\infty$. Then by \eqref{shifted:n} $I_{\gamma(n)}(\omega(T+\cdot))$ also converges as $n\to \infty$ in the topology of uniform convergence on compact sets, and by the definition of $F_\gamma$, the limit must be $F_\gamma(\omega(T+\cdot))$.
\end{proof}
The relationship \eqref{F:shifted} can be cast in the form
\begin{equation}
F_\gamma(\omega)(t)-F_\gamma(\omega)(s)=F_\gamma(\omega(s+\cdot))(t-s),\ 0\le s \le t.\label{F:additive}\end{equation}
That is, $F_\gamma$ is additive on $\{\omega\in C_S: F_\gamma(\omega\not=\underline 0\}$. The combination of \eqref{F:shifted} and \eqref{stopping} yields that for $0\le T_1\le T_2$, $\omega\in C_S$, and $\{\gamma(n),n\ge 1\}\in\Gamma$ such that $F_\gamma(\omega)\not=\underline 0$, we have
\begin{equation}F_\gamma(\omega((T_1+\cdot)\wedge T_2))(t)=F_\gamma(\omega)((T_1+t)\wedge T_2) - F_\gamma(\omega)(T_1).\label{shiftedandstopped}\end{equation}
\begin{lemma} Let $a,\lambda>0$, $\{\gamma(n),n\ge 1\}\in\Gamma$ and $\omega\in C_S$ be such that $F_\gamma(\omega)\not=\underline 0$. Then
\begin{equation}aF_\gamma(\omega)(\lambda\cdot)
=F_{a\gamma}(a\omega(\lambda\cdot)),\label{linear}\end{equation}
where $a\gamma = \{a\gamma(n),n\ge 1\}$.
\end{lemma}
\begin{proof}
One can easily see that
\begin{equation*}aI_{\gamma(n)}(\omega)(\lambda\cdot)=I_{a\gamma(n)}(a\omega(\lambda\cdot)),\ n\ge 1.\end{equation*}
But the assumption $F_\gamma(\omega)\not=\underline 0$ implies that the expression on left-hand side converges to $aF_\gamma(\omega)(\lambda\cdot)$ as $n\to\infty$ in the topology of uniform convergence on compacts. Then the expression on the right-hand side also converges, and by the definition of $F_{a\gamma}$ the limit must be $F_{a\gamma}(a\omega(\lambda\cdot))$.
\end{proof}
Now recall the process $W$ from Proposition \ref{Proposition:Brownian:Motion} which also appears implicitly in \eqref{decomposition} via \eqref{X:def}. The following lemma is instrumental for the developments in this section. It specifies the exact form of the dependence of $W(\cdot,\omega)$ on $\omega$.
\begin{lemma}\label{sub} For every $z\in S$ and $\{\delta(n),n\ge 1\}\in\Gamma$ there exists a sequence $\{\gamma^z(n),n\ge 1\}\subset \{\delta(n),n\ge 1\}$ such that
\begin{equation}P^z\left(F_{\gamma^z}(\omega)=W(\cdot,\omega)\right)=1.\label{gamma:z}\end{equation}
In addition, the following three statements hold:\\
~\\
(a) If for some $z\in S$ and $\{\gamma(n), n\ge 1\}\in \Gamma$ we have $P^z(F_\gamma(\omega)=\underline 0)=0$ then
\begin{equation} P^z(F_\gamma(\omega)=W(\cdot,\omega))=1;\label{more}\end{equation}
(b) For every stopping time $\tau$ on $(C_s,{\mathcal F},{\mathcal F}_t)$ and $z\in S$,
\begin{equation}E^z\left[P^{Z(\tau)}\left(F_{\gamma^z}(\omega)=W(\cdot,\omega)\right)\right]=1;\label{stoppingtime}\end{equation}
(c) For every $z\in S$,
\begin{equation} P^0\left(F_{\gamma^z}(\omega)=W(\cdot,\omega)\right)=1.\label{z:is:zero}\end{equation}
\end{lemma}
\begin{proof}
Let $\{\delta(n),n\ge 1\}\in\Gamma$ be arbitrary.
By the concavity of the function $x\mapsto \sqrt{x}\wedge 1$, \eqref{I}, Doob's Maximal Inequality, and \eqref{L2limit}
\begin{equation}
\begin{aligned}
&\ \lim_{n\to \infty}\sum_{m=1}^\infty {1\over 2^m} E^z\left[\sup_{t\le m}\left\{\|I_{\delta(n)}(\omega)(t)-W(t,\omega)\|\wedge 1\right\}\right]\\
&\le \lim_{n\to \infty}\sum_{m=1}^\infty {1\over 2^m}\left\{\left( E^z\left[\sup_{t\le m}\left\{\|I_{\delta(n)}(\omega)(t)-W(t,\omega)\|^2\right\}\right]\right)^{1/2}\wedge 1\right\}\\
&\le \lim_{n\to \infty}\sum_{m=1}^\infty {1\over 2^m}\left\{\left( E^z\left[\sup_{t\le m}\left\{\|W^{\delta(n)}(t,\omega)-W(t,\omega)\|^2\right\}\right]\right)^{1/2}\wedge 1\right\}\\
&\le \lim_{n\to \infty}\sum_{m=1}^\infty {1\over 2^m}\left\{\left( 4\sum_{i=1,2}E^z\left[\left(W_i^{\delta(n)}(m)-W_i(m)\right)^2\right]\right)^{1/2}\wedge 1\right\}\\
&=0.\label{long}
\end{aligned}
\end{equation}
Then there exists a subsequence $\{\gamma^z(n),n\ge 1\}\subset \{\delta(n),n\ge 1\}$ such that
$$\lim_{n\to 0}\sum_{m=1}^\infty {1\over 2^m} \sup_{t\le m}\left\{\|I_{\gamma^z(n)}(\omega)(t)-W(t,\omega)\|\wedge 1\right\}= 0,$$
$P^z$-a.s., which implies \eqref{gamma:z}. Next we are going to show the statement in (a). From our assumption that $P^z(F_\gamma(\omega)=\underline 0)=0$, it follows that $\lim_{n\to\infty}I_{\gamma(n)}(\omega)= F_\gamma(\omega)$ in the topology of uniform convergence on compacts, $P^z$-a.s.
On the other hand, similarly to \eqref{long} one can show that
$$\lim_{n\to 0}\sum_{m=1}^\infty {1\over 2^m} E^z\left[\sup_{t\le m}\left\{\|I_{\gamma(n)}(\omega)(t)-W(t,\omega)\|\wedge 1\right\}\right]=0,$$
which implies \eqref{more}.
Next we are going to show part (b). By \eqref{F:shifted} and \eqref{gamma:z} we have $P^z$-a.s.
\begin{equation}F_{\gamma^z}(\omega(\tau+\cdot))
=F_{\gamma^z}(\omega)(\tau+\cdot)-F_{\gamma^z}(\omega)(\tau)=W(\tau+\cdot)-W(\tau)\not=\underline 0,\label{notzero}\end{equation}
thus by the strong Markov property
$$1=P^z\left(F_{\gamma^z}(\omega(\tau+\cdot))\not=\underline 0\right)=
P^z\left(F_{\gamma^z}(Z(\tau+\cdot))\not=\underline 0\right)=
E^z\left[P^{Z(\tau)}\left(F_{\gamma^z}(Z(\cdot))\not=\underline 0\right)\right].$$
Therefore there exists a set $S^\tau\in {\mathcal B}_S$ such that $P^z(Z(\tau)\in S^\tau)=1$ and
$$P^{z'}\left(F_{\gamma^z}(Z(\cdot))\not=\underline 0\right)=1 \ \hbox{for all}\ z'\in S^\tau$$
($S^\tau$ is a Borel set by Varadhan and Williams \cite{varadhan1985brownian}, Corollary 3.3). Then by part (a) we have $P^{z'}(F_{\gamma^z}(\omega)=W(\cdot,\omega))=1$ for all $z'\in S^\tau$, which implies \eqref{stoppingtime}. Finally we show part (c). By part (a) all we need to show that $P^0(F_{\gamma^z}(\omega)=\underline 0)=0$. By the strong Markov property the left-hand side is equal to $P^z(F_{\gamma^z}(Z(\tau_0+\cdot))=\underline 0)$, which is indeed 0 by \eqref{notzero}.
\end{proof}
For every $z\in S$, we fix a sequence $\{\gamma^z(n),n\ge 1\}$ such that \eqref{gamma:z} holds. Recall the class ${\mathcal A}$ from Definition \ref{Definition:Of:The:Set:A}.
\begin{proposition}\label{p3} For each $z\in S$, we have $P^z$-a.s. that $\left(Z(\cdot\wedge \tau_0),Y(\cdot\wedge\tau_0)\right)\in {\mathcal A}$.
\end{proposition}
\begin{proof} \eqref{gamma:z} implies that for any subsequence $\{\alpha(n), n\ge 1\}\subset \{\gamma^z(n), n\ge 1\}$ we have $F_\alpha(\omega)=W(\cdot,\omega)$, $P^z$-a.s., and by \eqref{stopping} and by $Z(\cdot,\omega)=\omega$ we have
$$F_\alpha(Z(\cdot\wedge\tau_0))=W(\cdot\wedge\tau_0)=X(\cdot\wedge\tau_0)-z,$$
$P^z$-a.s. Then it is sufficient to show that there exists a sequence $\{\alpha(n),n\ge 1\}\subset \{\gamma^z(n),n\ge 1\}$ such that
\begin{equation}F_\alpha(\tilde Z(\cdot))=\tilde X(\cdot\wedge\tilde\tau_0)-z,\ Q^z\hbox{-a.s.}\label{sufficient}\end{equation}
Indeed, suppose that \eqref{sufficient} holds, then
\begin{eqnarray*} (\tilde Z(\cdot),\tilde Y(\cdot))=(\tilde Z(\cdot), \tilde Z(\cdot)-\tilde X(\cdot\wedge\tilde\tau_0))&=&
(\tilde Z(\cdot), \tilde Z(\cdot)-F_\alpha(\tilde Z(\cdot))-z)
\buildrel \rm d\over =\\
( Z(\cdot\wedge\tau_0), Z(\cdot\wedge\tau_0)-F_\alpha(Z(\cdot\wedge\tau_0))-z) &=&
( Z(\cdot\wedge\tau_0), Y(\cdot\wedge\tau_0)),
\end{eqnarray*}
where the sign $\buildrel\rm d\over =$ means that the law of the expression on the left-hand side under $Q^z$ agrees with the law of the expression on the right-hand side under $P^z$. Then the statement of the present proposition follows from Proposition \ref{williams}. For the rest of this proof we shall prove \eqref{sufficient} which amounts to showing that for some $\{\alpha(n),n\ge 1\}\subset \{\gamma^z(n),n\ge 1\}$ and every $m\in{\mathbb N}_+$ we have
\begin{equation}\sup_{t\le m}\left\|I_{\alpha(n)}(\tilde Z(\cdot))(t) -\left(\tilde X(t\wedge\tilde\tau_0)-z\right)\right\|^2\to 0\label{also:sufficient}\end{equation}
as $n\to\infty$, $Q^z$-a.s.
Analog to $\tau_0^\delta, \tau_1^\delta,\sigma_1^\delta,\tau_2^\delta,\sigma_2^\delta \dots$ for $\delta>0$ we define the stopping times $\tilde\tau_0^\delta=0$,
$\tilde \sigma_k^\delta=\inf\{t\ge\tilde \tau_{k-1}^\delta: \tilde Z(t)\in S_{2\delta}\}$ and $\tilde\tau_k^\delta=\inf\{t\ge \tilde\sigma_k^\delta: \tilde Z(t)\in \partial S_{\delta}\}, k\ge 1$. In the rest of the proof we shall drop the superscript $z$ from $\gamma^z(n)$ whenever it leads to double superscript, and write $\tilde\sigma^{\gamma(n)}$ and $\tilde\tau^{\gamma(n)}$ instead.
By the fact that $\tilde Z$ is flat on $[\tilde \tau_0,\infty)$, we have
\begin{eqnarray*}
I_{\gamma^z(n)}(\tilde Z(\cdot))(t\wedge m)&=&\sum_{k=1}^\infty\left[\tilde Z(t\wedge\tilde\tau_0\wedge\tilde\tau_k^{\gamma(n)}\wedge m) - \tilde Z(t\wedge\tilde\tau_0 \wedge\tilde\sigma_k^{\gamma(n)}\wedge m) \right].
\end{eqnarray*}
Moreover, since $\tilde Y$ is flat on $[\tilde\sigma_k^\delta\wedge m,\tilde\tau_k^\delta\wedge m]$, we can cast this in the form
\begin{eqnarray*}
I_{\gamma^z(n)}(\tilde Z(\cdot))(t\wedge m)&=&\sum_{k=1}^\infty\left[\tilde X(t\wedge \tilde\tau_0\wedge\tilde\tau_k^{\gamma(n)}\wedge m) - \tilde X(t\wedge\tilde\tau_0\wedge\tilde\sigma_k^{\gamma(n)}\wedge m) \right] \nonumber \nonumber \\
&=&\int_{0}^t \sum_{k=1}^\infty 1\{u\in[ \tilde\sigma_k^{\gamma(n)}\wedge\tilde\tau_0\wedge m, \tilde\tau_k^{\gamma(n)}\wedge\tilde\tau_0\wedge m]\} d\tilde X(u).
\end{eqnarray*}
Hence by the the Dominated Convergence Theorem, the fact that $Z(\cdot\wedge\tau_0)$ has the same law as $\tilde Z(\cdot\wedge\tilde\tau_0)$, and by Williams \cite{williams1985recurrence}, Lemma 4.2, it follows that
\begin{eqnarray*}
&&\lim_{n\to\infty} E^z_Q\left[ \left\|I_{\gamma(n)}(\tilde Z(\cdot))(\tilde\tau_0\wedge m)-\left(\tilde X(\tilde\tau_0\wedge m)-z\right)\right\|^2\right] \nonumber\\
&=& 2\lim_{n\to\infty}E_Q^z\left[\sum_{k=0}^\infty \int_0^{\tilde\tau_0\wedge m} 1\{u\in [\tilde\tau_k^{\gamma(n)}, \tilde\sigma_{k+1}^{\gamma(n)}] \}du\right] \nonumber\\
&\leq& 2\lim_{n\to\infty}E_Q^z\left[ \int_0^{\tilde\tau_0\wedge m} 1\{\tilde Z(u)\in S\setminus S_{2\gamma^z(n)}\}du\right] \nonumber\\
&=& 2E_Q^z \left[\int_0^{\tilde\tau_0\wedge m} 1\{\tilde Z(u)\in\partial S\}du \right] \nonumber\\
& =&
2 E^z \left[\int_0^{\tau_0\wedge m} 1\{ Z(u)\in\partial S\}du \right] \nonumber \\
&=&0.
\end{eqnarray*}
Then Doob's inequality applied to the square-integrable martingale $I_{\gamma(n)}(\tilde Z\wedge m)(\cdot)-(\tilde X(\cdot\wedge\tilde\tau_0\wedge m)-z)$ gives
\begin{eqnarray*}
&&E_Q^z\left[\sup_{t\le m}\left\{\left\|I_{\gamma(n)}(\tilde Z(\cdot))(t\wedge\tilde\tau_0)-\left(\tilde X(t\wedge\tilde\tau_0)-z\right)\right\|^2\right\}\right]\\
&\leq&4E_Q^z\left[\left\|I_{\gamma(n)}(\tilde Z(\cdot))(\tilde\tau_0\wedge m )-\left(\tilde X(\tilde\tau_0\wedge m)-z\right)\right\|^2\right]\to 0, \hbox{ as }n\to\infty,
\end{eqnarray*}
and so the existence of a subsequence
$\{\alpha(n),n\geq 1\}\subset\{\gamma^z(n),n\geq 1\}$ follows such that \eqref{also:sufficient} holds.
\end{proof}
\begin{proposition}\label{main}
For each $z\in S$ , and each excursion interval $[G_i,D_i]$ of $Z$ away from the origin, we have $P^z$-a.s. that
\begin{equation*} \left(Z\left(\left(G_i+\cdot\right)\wedge D_i\right), Y\left(\left(G_i+\cdot\right)\wedge D_i\right)\right)\in{\mathcal A}.\end{equation*}
\end{proposition}
\begin{proof}
Let $\{(U^n_k,T^n_k), k\ge 1\}$ be the sequence of excursion intervals of $Z$ away from zero with length strictly larger than $1/n$, i.e., $T^n_k-U^n_k>1/n$ for every $n\in{\mathbb N}_+$, and let $V_k^n=U_k^n+1/n$.
Clearly it is sufficient to show that for all $k,n\in{\mathbb N}_+$ and $z \in S$,
\begin{equation*}P^z\left(\left(Z\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right), Y\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)\right)\in {\mathcal A}\right)=1,
\end{equation*}
which is the same as
\begin{equation}P^z\left(\left(Z\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right), Z\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)
-X\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)
\right)\in {\mathcal A}\right)=1.\label{new}\end{equation}
By \eqref{gamma:z} and \eqref{shiftedandstopped} we have $P^z$-a.s. that
$$X\left(\left(V_k^n+t\right)\wedge T_k^n\right)= F_{\gamma^z}(Z(\cdot)) \left(\left(V_k^n+t\right)\wedge T_k^n\right)+z =
F_{\gamma^z} \left(Z \left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)\right)(t) +F_{\gamma^z}\left(Z(\cdot)\right)(V_k^n)+z,$$
hence we can replace the $X\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)$ term in \eqref{new} by $F_{\gamma^z} \left(Z \left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)\right)$, since the terms $F_{\gamma^z}\left(Z(\cdot)\right)(V_k^n)+z$ have no impact on the membership in ${\mathcal A}$. Thus the left-hand side of \eqref{new} is equal to
\begin{equation*}P^z\left(\left(Z\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right), Z\left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)- F_{\gamma^z} \left(Z \left(\left(V_k^n+\cdot\right)\wedge T_k^n\right)\right)\in {\mathcal A}\right)\right),\end{equation*}
and by the strong Markov property and \eqref{stopping} this is equal to
$$E^z\left[P^{Z(V_k^n)}\left(\left(Z(\cdot\wedge\tau_0), Z(\cdot\wedge\tau_0)- F_{\gamma^z}\left(Z(\cdot\wedge\tau_0)\right)\right)\in {\mathcal A}\right)\right]$$
\begin{equation}=E^z\left[P^{Z(V_k^n)}\left(\left(Z(\cdot\wedge\tau_0), Z(\cdot\wedge\tau_0)- F_{\gamma^z}\left(Z(\cdot)\right)(t\wedge\tau_0)\right)\in {\mathcal A}\right)\right].\label{step}\end{equation}
When applying \eqref{stopping} above, we needed the condition $E^z\left[P^{Z(V_k^n)}\left(F_{\gamma^z}\left(Z(\cdot)\right)\not=\underline 0\right)\right]$, but this follows from \eqref{stoppingtime}. Applying \eqref{stoppingtime} and Proposition \ref{p3}, we get that \eqref{step} is equal to
$$E^z\left[P^{Z(V_k^n)}\left(\left(Z(\cdot\wedge\tau_0), Z(\cdot\wedge\tau_0)-X(\cdot\wedge\tau_0)\right)\in {\mathcal A}\right)\right]=1,$$
as desired.
\end{proof}
Now we have the following lemma.
\begin{lemma}Suppose that $(z,y) \in \mathcal{A}$ and $T > 0$ are such that $ z(s) \neq 0$ for $s \in (0,T)$.
Then, letting $u=(u_1,u_2)$ be the non-decreasing function from Definition \ref{Definition:Of:The:Set:A}, we have that for $t\in [0,T]$,
\begin{eqnarray}
V_1(y,[0,t]) &=& \| v_1 \|(u_1(t)-u_1(0)) + \|v_2\|(u_2(t)-u_2(0)). \label{display:variation:of:y:over:excursion}
\end{eqnarray}
\end{lemma}
\begin{proof} The main idea of this proof is that on a closed interval included in $(0,T)$ the process $Z$ can hit the boundaries $\partial S_1$ and $\partial S_2$ only finitely many times. This may not be true on the interval $[0,T]$, because $z(0)$ or $z(T)$ may possibly be equal to zero. In this spirit let $0 < r < t < T$. We first show that
\begin{eqnarray}
V_1(y,[r,t]) &=& \| v_1 \|(u_1(t)-u_1(r)) + \|v_2\|(u_2(t)-u_2(r)). \label{display:to:use:triangle:inequality}
\end{eqnarray}
A straightforward application of the triangle inequality guarantees that the left-hand side of \eqref{display:to:use:triangle:inequality} is dominated by the right-hand side. From the continuity of $z$ and by the assumption $z(s) \neq 0$ for $s \in [r,t]$, it follows that there exits a finite partition $\pi = \{r=t_0 < t_1 < ... < t_n=t\}$ of $[r,t]$ such that for each $i=0,...,n-1$, within the interval $[t_i,t_{i+1}]$, $z$ does not touch both boundaries $\partial S_1$ and $\partial S_2$. We can assume without loss of generality that $z$ does not reach $\partial S_1$ in $[t_i,t_{i+1}]$ if $i$ is odd, and
does not reach $\partial S_2$ in $[t_i,t_{i+1}]$ if $i$ is even. Then
\begin{eqnarray*}
\sum_{i=0}^{n-1}\|y(t_{i+1})-y(t_i)\| &=&\sum_{i=0}^{n-1} \| v_1 u_1(t_{i+1})-v_1u_1(t_i) + v_2u_2(t_{i+1})-v_2u_2(t_i)\| \\
&=&\sum_{i~ \textrm{even}} \| v_1 u_1(t_{i+1})-v_1u_1(t_i)\| +\sum_{i ~ \textrm{odd}} \| v_2 u_2(t_{i+1})-v_2u_2(t_i)\| \\
&=&\|v_1\|(u_1(t)-u_1(r))+\|v_2\|(u_2(t)-u_2(r)),
\end{eqnarray*}
which indeed proves \eqref{display:to:use:triangle:inequality}.
Now, from statement P6 in Section 2.2 of Chistyakov and Galkin \cite{chistyakov1998maps}, it follows that
\begin{eqnarray*}
V_1(y,(0,t]) &=& \| v_1 \|(u_1(t)-u_1(0)) + \|v_2\|(u_2(t)-u_2(0)), \end{eqnarray*}
for $0 < t < T$, and
\begin{eqnarray*}
V_1(y,(0,T)) &=& \| v_1 \|(u_1(T)-u_1(0)) + \|v_2\|(u_2(T)-u_2(0)). \end{eqnarray*}
However, based on the continuity of $y$ one can easily show that
$V_1(y,[0,t])=V_1(y,(0,t])$ and $V_1(y,[0,T])=V_1(y,(0,T))$, and hence \eqref{display:variation:of:y:over:excursion} follows.
\end{proof}
We now present the proof of Part \ref{Proposition:Skorohod:Problem:Part:2} of Theorem \ref{Proposition:Skorokhod:Problem}.
\begin{proof}[Proof of Part \ref{Proposition:Skorohod:Problem:Part:2} of Theorem \ref{Proposition:Skorokhod:Problem}:]
Let $[G_i,D_i]$ be an arbitrary excursion interval of $Z(\cdot, \omega)$
away from zero. By Proposition \ref{main} there exist two non-decreasing functions $U_1,U_2: \mathbb{R}_+ \mapsto \mathbb{R}$, such that $Y((G_i+t) \wedge D_i)=v_1 U_1(t) + v_2 U_2(t)$ for $t \geq 0$, and
\begin{eqnarray*}
\int_0^{\infty}1\{Z((G_i+p)\wedge D_i) \notin \partial S_j\}dU_j(p) &=& 0,
\end{eqnarray*}
for $j=1,2$. Then, by the previous lemma,
\begin{eqnarray}
V_1(Y((G_i+\cdot) \wedge D_i),[0,t])&=& \| v_1 \|(U_1(t)-U_1(0)) + \|v_2\|(U_2(t)-U_2(0)),~~t \geq 0. \label{display:to:be:used:with:item:4}
\end{eqnarray}
The first three requirements of Definition \ref{Definition:Skorohod:Problem} are obviously satisfied. Item \ref{Definition:SP:Item:Four}
of Definition \ref{Definition:Skorohod:Problem} follows from \eqref{display:to:be:used:with:item:4}. Moreover, from \eqref{display:to:be:used:with:item:4} it follows that
\begin{eqnarray*}
Y((G_i+t) \wedge D_i) -Y(G_i)&=&\int_0^t \gamma(p)d( \| v_1 \|U_1(p) +\|v_2\|U_2(p)),
\end{eqnarray*}
where
\begin{eqnarray*}
\gamma(p) &=& \frac{v_1}{\|v_1\|}1\{Z((G_i+p)\wedge D_i) \in \partial S_1\}
+ \frac{v_2}{\|v_2\|}1\{Z((G_i+p)\wedge D_i) \in \partial S_2\},
\end{eqnarray*}
and so Item \ref{Definition:SP:Item:Five} of Definition \ref{Definition:Skorohod:Problem} follows. Clearly the same proof applies if instead of $[G_i,D_i]$ we consider the interval $[0,\tau_0]$, whenever $\tau_0>0$; the only difference is that we have to use Proposition \ref{p3} instead of Proposition \ref{main}.
\end{proof}
\section{Proof of Theorem \ref{Theorem:Main:Dirichlet:Process:Result}}
\label{section:dirichlet:process}
Let the process $X$ be as in Theorem \ref{Proposition:Skorokhod:Problem}, so
for each $z \in S$ the process $X$ is a Brownian motion started from $z$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$, and let $Y$ be defined by $Z=X+Y$. In particular, $X$ is a local martingale on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$. Hence, by Definition \ref{Definition:Dirichlet:Process} of Section \ref{Section:Introduction}, in order to prove Theorem \ref{Theorem:Main:Dirichlet:Process:Result}
it suffices to prove that for each $z \in S$, $Y$ is a continuous, zero-energy process (see Definition \ref{Definition:Zero:Energy}) on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ with $P^z(Y(0)=0)=1$. The
fact that $Y$ is continuous with $P^z(Y(0)=0)=1$ is immediate since $Y=Z-X$,
and so the proof of Theorem \ref{Theorem:Main:Dirichlet:Process:Result} is reduced to proving that $Y$ is of zero-energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$.
In order to prove that $Y$ is of zero-energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ for each $z \in S$, we will rely on the relationship between functions of finite $p$-variation and $1/p$-H\"{o}lder continuous functions. In particular, let $E \subseteq \mathbb{R}_{+}$ and $f:E \mapsto \mathbb{R}^d$. Then, for each $\gamma > 0$ the function $f$ is said to be H\"{o}lder continuous
with exponent $\gamma$ and H\"{o}lder constant $H(f) \in \mathbb{R}_{+}$, if
\begin{eqnarray*}
\| f(t)-f(s)\| &\leq H(f) |t-s |^{\gamma}&~\textrm{for}~s,t \in E.
\end{eqnarray*}
In addition, we say that $f$ is locally H\"{o}lder continuous on $\mathbb{R}_{+}$ if it is H\"{o}lder continuous on compact $E \subset \mathbb{R}_{+}$. The following result now relates functions of finite $p$-variation to H\"{o}lder continuous functions with exponent $1/p$, see for instance Theorem 3.1 of Chistyakov and Galkin \cite{chistyakov1998maps} where it is proven for $p > 1$. The fact that it is true for $p > 0$ may be verified by mimicking their proofs.
\begin{proposition}\label{Proposition:Finite:Variation:Holder}
Let $E \subseteq \mathbb{R}_{+}$ and $f:E \mapsto \mathbb{R}^d$. Then for each $p > 0$, we have that $V_p(f,E) < \infty$
if and only if there exists a bounded, nondecreasing function $\varphi : E \mapsto \mathbb{R}$ and a map $g:\varphi(E) \mapsto \mathbb{R}^d$
which is H\"{o}lder continuous with exponent $1/p$ and H\"{o}lder constant $H(g) \leq 1$, such that $f = g \circ \varphi$ on $E$.
\end{proposition}
\noindent In the above, the quantity $\varphi(E) \subseteq \mathbb{R}$ is the range of the function $\varphi : E \mapsto \mathbb{R}$. The requirement that the H\"{o}lder constant $H(g) \leq 1$ is unnecessary in the if portion of Proposition \ref{Proposition:Finite:Variation:Holder}. Indeed, $V_p(f,E) < \infty$ if $f = g \circ \varphi$ on $E$ for a bounded, nondecreasing function $\varphi : E \mapsto \mathbb{R}$ and any $g:\varphi(E) \mapsto \mathbb{R}^d$
which is H\"{o}lder continuous with exponent $1/p$, regardless of its constant. We also note that given a $f$ with $V_p(f,E) < \infty$, one particular choice of $\varphi$ and $g$ satisfying the properties of Proposition \ref{Proposition:Finite:Variation:Holder} is as follows (see the proof of Theorem 3.1 of \cite{chistyakov1998maps} for details). For each $t \in E$, let $E_t^{-}=\{s \in E : s \leq t\}$. Also, for future reference, set $E_t^{+}=\{s \in E : t \leq s\}$, and for $a,b \in E$ with $a \leq b$, let $E^b_a=E^{+}_a \cap E^{-}_b$. Now set $\varphi(t)=V_p(f,E_t^{-})$ for $t \in E$. Next, for $\tau \in \varphi(E)$, define the multi-valued inverse function $\varphi^{-1}(\tau)=\{t \in E : \varphi(t)=\tau\}$. We may then define $g: \varphi(E) \mapsto \mathbb{R}$ by $g(\tau)=f(t)$ for any point $t \in \varphi^{-1}(\tau)$. The fact that $g$ is well-defined is proven in \cite{chistyakov1998maps}. Finally, we note that
in the event that $E=[0,T]$ for some $T > 0$ and the function $f$ is continuous, then we may take the function $\varphi$ to be continuous as well (see Proposition 5.14 of Friz and Victoir \cite{friz2010multidimensional}).
\begin{proposition}\label{Proposition:Variation:Time:Chanaged:Holder}
Suppose that $\alpha < p < 2$. Then, for $P^0$-a.e. $\omega \in C_S$, there exists a continuous, nondecreasing function $\varphi_p : \mathbb{R}_{+} \mapsto \mathbb{R}$ and a map $g_p:\varphi_p(E) \mapsto \mathbb{R}^2$, which is locally H\"{o}lder continuous with exponent $1/p$, such that $Y(\cdot,\omega) = g_p \circ \varphi_p$ on $\mathbb{R}_{+}$.
\end{proposition}
In the following two subsections, we provide the proof of Proposition \ref{Proposition:Variation:Time:Chanaged:Holder}. We close this subsection with the proof of Theorem \ref{Theorem:Main:Dirichlet:Process:Result} which follows as a result of Propositions \ref{Proposition:Finite:Variation:Holder} and \ref{Proposition:Variation:Time:Chanaged:Holder}, together with the results of Section \ref{Section:Identification:of:X:and:Skorokhod:Problem}.
\begin{proof}[Proof of Theorem \ref{Theorem:Main:Dirichlet:Process:Result}]We first treat the case of $z=0$. By the decomposition $Z=X+Y$, where $X$ is a Brownian motion started from $0$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^0)$, in order to show that $Z$ is a Dirichlet process on $(C_S,\mathcal{F},\mathcal{F}_t,P^0)$, it suffices by Definition \ref{Definition:Dirichlet:Process} to prove that $Y$ is a process of zero energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^0)$. Let $\alpha < p < 2$ and recall by Proposition \ref{Proposition:Variation:Time:Chanaged:Holder} that for $P^0$-a.e. $\omega \in C_S$, $Y(\cdot,\omega) = g_p \circ \varphi_p$ on $\mathbb{R}_{+}$, where $\varphi_p : \mathbb{R}_{+} \mapsto \mathbb{R}$ is a continuous, nondecreasing function and $g_p:\varphi_p(E) \mapsto \mathbb{R}^2$ is locally H\"{o}lder continuous on $\mathbb{R}_{+}$ with exponent $1/p$.
Now let $T \geq 0$ and let $H_{\varphi_p(T)}(g_p)$
denote the H\"{o}lder constant of $g_p$ on $[0,\varphi_p(T)]$. Then, for any partition $\pi$ of $[0,T]$,
\begin{eqnarray*}
\sum_{t_i \in \pi} \|Y(t_i)-Y(t_{i-1}) \|^2 &=&\sum_{t_i \in \pi} \|g_p \circ \varphi_p(t_i)-g_p \circ \varphi_p(t_{i-1}) \|^2 \\
&\leq& (H_{\varphi_p(T)}(g_p))^2\sum_{t_i \in \pi} |\varphi_p(t_i)-\varphi_p(t_{i-1}) |^{2/p} \\
&\leq&(H_{\varphi_p(T)}(g_p))^2(\varphi_p(T)-\varphi_p(0)) \max_{t_i \in \pi} |\varphi_p(t_i)-\varphi_p(t_{i-1}) |^{2/p-1}.
\end{eqnarray*}
Next let $\{\pi^n, n \geq 1\}$ be a sequence of partitions of $[0,T]$ with $\|\pi^n \| \rightarrow 0$ as $n \rightarrow \infty$, and note that since $ \alpha < p < 2$, $2/p-1 > 0$. Then, since $\varphi_p$
is $P^0$-a.s. continuous, it follows that $\max_{t_i \in \pi^n} |\varphi_p(t_i)-\varphi_p(t_{i-1}) |^{2/p-1} \rightarrow 0$ as $n \rightarrow \infty,~P^0$-a.s., and so by the above
\begin{eqnarray*}
\sum_{t_i \in \pi} \|Y(t_i)-Y(t_{i-1}) \|^2 &\rightarrow&0~~\textrm{as}~~n \rightarrow \infty,~P^0\textrm{-a.s.},
\end{eqnarray*}
which implies that $Y$ is of zero-energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^0)$. Thus, $Z$ is a Dirichlet process on $(C_S,\mathcal{F},\mathcal{F}_t,P^0)$.
Now let $z \in S$ arbitrary. Again by the decomposition $Z=X+Y$, where $X$ is a Brownian motion started from $z$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ \ it suffices by Definition \ref{Definition:Dirichlet:Process} to prove that $Y$ is a process of zero energy on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$. Using Proposition \ref{Proposition:Finite:Variation:Holder} above and similar arguments to the case of $z=0$ in the preceding paragraphs, it suffices to show that for $1 \leq \alpha < p < 2$,
\begin{eqnarray}
P^z(V_p(Y,[0,T]) < \infty)&=&1,~T \geq 0. \label{display:used:before:in:proof}
\end{eqnarray}
Recall from \eqref{tau:zero} of Section \ref{Section:Identification:of:X:and:Skorokhod:Problem} the definition of the stopping time $\tau_0=\inf\{t > 0 :Z(t)=0\}$. By Theorem 2.2 of Varadhan amd Williams \cite{varadhan1985brownian}, since $ \alpha > 0$, $P^z(\tau_0 < \infty)=1$.
By the monotonicity property (see (P2) in Chistyakov and Galkin \cite{chistyakov1998maps}),
$$P^z\left(V_p\left(Y,[0,T]\right)<\infty\right)\ge P^z\left(V_p\left(Y,[0,\tau_0+T]\right)<\infty\right).$$
By the semi-additivity property (see (P3) in Chistyakov and Galkin \cite{chistyakov1998maps})
the right-hand side in the above equation is equal to
$$P^z\left(V_p\left(Y,[0,\tau_0]\right)<\infty \ \hbox{and}\ V_p\left(Y,[\tau_0,\tau_0+T]\right)<\infty\right).$$
Now using the strong Markov property \cite{varadhan1985brownian}, this expression can be written as
\begin{eqnarray}
&&E^z\left[1\{V_p\left(Y,[0,\tau_0]\right)<\infty\}P^z(V_p\left(Y,[\tau_0,\tau_0+T]\right)<\infty|{\cal F}_{\tau_0})\right] \label{use:markov}\\
&=&P^z\left(V_p\left(Y,[0,\tau_0]\right)<\infty\right)P^0\left(V_p\left(Y,[0,T]\right)<\infty\right).
\nonumber
\end{eqnarray}
However, it has already been shown earlier in the proof that the second factor in \eqref{use:markov} is equal to 1. By Part \ref{Proposition:Skorohod:Problem:Part:2} of Theorem \ref{Proposition:Skorokhod:Problem}, $P^z\left(V_1\left(Y,[0,\tau_0]\right)<\infty\right)=1$, and this implies that the first term in \eqref{use:markov} is also equal to 1.
\end{proof}
\subsection{Properties of the Function $\mathbf{\varphi_{p,q}}$}
For each $p > \alpha$ and $q > \alpha/2$, define the function $\varphi_{p,q} : \mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ by setting
\begin{eqnarray}
\varphi_{p,q}(t) &=&\begin{cases}
V_{q}(L^{-1}, [0,L(t)] ), &\text{for $t \in L^{-1}(\mathbb{R}_{+})$,}\\
V_{q}(L^{-1}, [0,L(t)) ), &\text{for $t \in \Lambda \backslash L^{-1}(\mathbb{R}_{+})$,} \\
\varphi_{p,q}(G_i)+ \frac{(D_i-G_i)^q}{V_p(Y ,[G_i,D_i])} \cdot V_p(Y ,[G_i,t]) , &\text{for $t \in (G_i,D_i)$ for $i \geq 1$,} \\
\end{cases} \label{Display:Definition:Of:Varphi}
\end{eqnarray}
where we recall from Section \ref{Section:Excursion:Theory:Background} the definition of $\Lambda$ as the zero set of $Z$, and of $L$ as the local time of $Z$ at the origin. Our main result of this subsection is the following.
\begin{proposition}\label{Proposition:Main:Properties:Of:Varphi:P:Q}
For each $ p > \alpha $ and $1 > q > \alpha/2$, the function $\varphi_{p,q} : \mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. well-defined, non-decreasing and continuous with $\varphi_{p,q}(\mathbb{R}_{+})=\mathbb{R}_{+}$.
\end{proposition}
In preparation for the proof of Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}, we provide three lemmas.
\begin{lemma}\label{Lemma:Properties:Of:Var:Phi:P:Q:On:Lambda}
For each $p > \alpha$ and $q > \alpha/2$, the function $\varphi_{p,q} : \mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. well-defined and strictly increasing on the set $\Lambda$. Moreover, if in addition to the above assumptions for $p$ and $q$ it is also true that $q<1$, then $P^0$-a.s.,
\begin{equation} \varphi_{p,q}(t)=\sum_{i:G_i<t}(D_i-G_i)^q,~t \in \Lambda.\label{nice:rep}
\end{equation}
\end{lemma}
\begin{proof}Let $p > \alpha$ and $q > \alpha/2$. First note that by \eqref{display:complement:of:lambda}, it follows that $P^0$-a.s. the three cases provided in the definition of $\varphi_{p,q}$ are disjoint. Now recall from Theorem 2.6 of Williams \cite{williams1987local} that $L^{-1}$ is a stable subordinator of index $\alpha/2$ under $P^0$. Hence, by the results of Simon \cite{simon2004small} (see immediately below display (9) of \cite{simon2004small}), it follows that $P^0(V_q(L^{-1},[0,1]) < \infty)=1$. Using the scaling property \cite{williams1987local} under $P^0$,
$$L^{-1}(\lambda \cdot)~ \buildrel d \over =~ \lambda^{2/\alpha}L^{-1}(\cdot),~\lambda > 0, $$
it then follows that $P^0(V_q(L^{-1},[0,T]) < \infty)=1$ for each $T \geq 0$. Moreover, by the monotonicity property of strong $q$-variation (see (P2) of Chistyakov and Galkin \cite{chistyakov1998maps}), $V_q(L^{-1},[0,\cdot])$ is $P^0$-a.s. a non-decreasing function. Hence, $P^0(V_q(L^{-1},[0,T]) < \infty, T \geq 0)=1$. It now follows from \eqref{Display:Definition:Of:Varphi} that $\varphi_{p,q}$ is $P^0$-a.s. well-defined on the set $\Lambda$.
Next, we show \eqref{nice:rep}. It follows from \eqref{l:inverse} of Section \ref{Section:Excursion:Theory:Background} and (9) of Simon \cite{simon2004small} that for $\alpha/2<q<1$, $P^0$-a.s., \begin{equation}V_q(L^{-1},[0,a])=\sum_{i:L(G_i)\le a}(D_i-G_i)^q
\ \ \hbox{and}
\ \ V_q(L^{-1},[0,a))=\sum_{i:L(G_i)< a}(D_i-G_i)^q,~a \geq 0.\label{open}\end{equation}
\eqref{nice:rep} now follows from \eqref{Display:Definition:Of:Varphi} and \eqref{open}, combined with the identity \eqref{Display:Disjoint:Intervals}.
Finally, we show that $\varphi_{p,q}$ is $P^0$-a.s. strictly increasing on the set $\Lambda$. By Item \ref{Condition:3:Submartingale} of Definition \ref{vw}, it follows that $P^0$-a.s. for every $t,t'\in\Lambda$ such that $t<t'$, there exists some $s\in(t,t')$ such that $s\notin\Lambda$. But then by \eqref{display:complement:of:lambda}, $G_i\in[t,t')$ for some $i \geq 1$, and so \eqref{nice:rep} implies $\varphi_{p,q}(t)<\varphi_{p,q}(t')$.
\end{proof}
\begin{lemma}\label{Lemma:Y:Positive:Variation}
For each $i \geq 1$ and $\varepsilon > 0$,
\begin{eqnarray}
P^0(V_1(Y,[G_i,D_i \wedge (G_i+\varepsilon)]) > 0)&=&1. \label{Display:Y:Variation:Positive:On:Excursion}
\end{eqnarray}
\end{lemma}
\begin{proof}First, we need a slight generalization of Definition 2.9.10 and Theorem 2.9.12 in Karatzas and Shreve \cite{KaratzasShreve}. Let $f:\mathbb{R}_{+} \mapsto \mathbb{R}$ be given. A time $t\ge 0$ will be called a {\it point of local maximum (minimum) from the right} if there exists a $\delta>0$ such that $f(s)\le f(t)$ ($f(s)\ge f(t))$ for every $s\in(t,t+\delta)$. A time $t\ge0$ will be called a {\it point of strict local maximum (minimum) from the right} if there exists a $\delta>0$ such that $f(s)< f(t)$ ($f(s)>f(t)$) for every $s\in(t,t+\delta)$. Now let $w$ be a standard (1-dimensional) Brownian motion on some probability space. Then, following the proof of Theorem 2.9.12 in \cite{KaratzasShreve}, for almost every sample path of $w$ all points of local maximum from the right are strict.
In a similar manner, it is also true that for almost every path $w$ all points of local minimum from the right are strict.
Now we are ready to prove \eqref{Display:Y:Variation:Positive:On:Excursion}. We shall use the same notation as in Williams \cite{williams1987local}. In particular, we adjoin a point $\partial$ to $C_S$, and define the excursion process $\{U_t, t \geq 0\}$ as the $(C_S,{\cal M})$-valued point process given for $t \geq 0$ by
\begin{eqnarray*}
U_t(s)&=&\begin{cases}
Z\left(L^{-1}(t-)+s\right), &\textrm{if}~~ 0\le s <L^{-1}(t)-L^{-1}(t-), \\
0, &\textrm{if}~~ s\ge L^{-1}(t)-L^{-1}(t-)>0, \\
\partial, &\textrm{if}~~ L^{-1}(t)-L^{-1}(t-)=0.
\end{cases}
\end{eqnarray*}
Now let $N(A,U)$ be the number of times $t$ such that $(t,U_t)\in A$, where $A\in {\cal B}\times{\cal M}$ and ${\cal B}$ is the $\sigma$-field of Borel subsets of $(0,\infty)$. Then, from Theorem 3.3 of Williams \cite{williams1987local} it follows that
\begin{eqnarray*}
E^0\left[N\left(\left(0,t\right)\times\left\{\omega: T_{\partial S}(\omega)>0\right\},U\right)\right]&=&0,
\end{eqnarray*}
where $T_{\partial S}(\omega)=\inf\left\{t>0: \omega(t)\in\partial S\right\}$ for $\omega\in C_S$. Then, by the Monotone Convergence Theorem \cite{lieb2001analysis},
\begin{eqnarray*}
E^0\left[N\left(\left(0,\infty\right)\times\left\{\omega: T_{\partial S}(\omega)>0\right\},U\right)\right]&=&0.
\end{eqnarray*}
This implies that for $P^0$-a.e. sample path $T_{\partial S}^i=G_i$ for all $i\ge 1$, where
\begin{eqnarray*}
T_{\partial S}^i&=&\inf\left\{t>G_i: Z(t)\in\partial S\right\}.
\end{eqnarray*}
Suppose now that \eqref{Display:Y:Variation:Positive:On:Excursion} does not hold. That is, $Y$ is flat on $[G_i,(G_i+\epsilon)\wedge D_i]$ for some $i\ge 1$. Since $Z(G_i)=0$,
\begin{equation} Z(t)=X(t)+Y(G_i)=X(t)-X(G_i),\quad t\in[G_i,(G_i+\epsilon)\wedge D_i].\label{(2)}\end{equation}
Now $T_{\partial S}^i=G_i$ implies the existence of a sequence $t_n\downarrow G_i$ such that $Z(t_n)\in\partial S_1$ for all $n\ge 1$, or $Z(t_n)\in\partial S_2$ for all $n\ge 1$. In the first case, $Z_2(t_n)=0$ for all $n\ge 1$. Thus, by \eqref{(2)}, $G_i$ is a point of local minimum from the right for the Brownian motion $X_2$, but not strict, which is a contradiction. In the second case, we proceed similarly. $G_i$ is a point of local minimum from the right for the Brownian motion $\langle n_2, X\rangle$, but not strict, which is again a contradiction.
\end{proof}
\begin{lemma}\label{Lemma:Properties:Of:Phi:P:Q:On:Lambda:C}
For each $ p > \alpha $ and $q > \alpha/2$, the function $\varphi_{p,q} : \mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. well-defined, non-decreasing and continuous on the set $(G_i,D_i)$ for $i \geq 1$.
\end{lemma}
\begin{proof}Let $p > \alpha$ and $q > \alpha/2$. By \eqref{display:complement:of:lambda}, it follows that $P^0$-a.s. the three cases provided in the definition of $\varphi_{p,q}$ are disjoint. We now show that $\varphi_{p,q}$ is $P^0$-a.s. well-defined on $(G_i,D_i)$ for each $i \geq 1$. By \eqref{Display:Definition:Of:Varphi} and the fact that $V_p(Y,[G_i,\cdot])$ is non-decreasing on $(G_i,D_i)$, it suffices to show that $P^0$-a.s., $0 < V_p(Y,[G_i,D_i]) < +\infty$. The fact that $P^0$-a.s., $0 < V_p(Y,[G_i,D_i])$ follows by Lemma \ref{Lemma:Y:Positive:Variation} above. The fact that $P^0$-a.s., $V_p(Y,[G_i,D_i]) < +\infty$ follows since
by Part \ref{Proposition:Skorohod:Problem:Part:2} of Theorem \ref{Proposition:Skorokhod:Problem}, $V_1(Y,[G_i,D_i]) < +\infty$ $P^0$-a.s., and the fact that $V_p(Y,[G_i,D_i]) \leq (V_1(Y,[G_i,D_i]))^p $ (see Remark 2.5 of Chistyakov and Galkin \cite{chistyakov1998maps}) since by assumption $p > \alpha > 1$. The fact that $\varphi_{p,q} : \mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. continuous and non-decreasing on $(G_i,D_i)$ follows by \eqref{Display:Definition:Of:Varphi} and the facts that $V_p(Y,[G_i,\cdot])$ is non-decreasing on $(G_i,D_i)$ and,
since $Y$ is $P^0$-a.s. continuous, $V_p(Y,[G_i,\cdot])$ is $P^0$-a.s. continuous on $(G_i,D_i)$ as well.
\end{proof}
Combining Lemmas \ref{Lemma:Properties:Of:Var:Phi:P:Q:On:Lambda} and \ref{Lemma:Properties:Of:Phi:P:Q:On:Lambda:C}, we now provide the proof of Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}.
\begin{proof}[Proof of Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}]Let $p > \alpha$ and $1 > q > \alpha/2$. By \eqref{display:complement:of:lambda}, it follows that $P^0$-a.s. the three cases provided in the definition of $\varphi_{p,q}$ are disjoint and cover all of $\mathbb{R}_{+}$. The fact that $\varphi_{p,q} : \mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. well-defined and non-decreasing now follows immediately by Lemmas \ref{Lemma:Properties:Of:Var:Phi:P:Q:On:Lambda} and \ref{Lemma:Properties:Of:Phi:P:Q:On:Lambda:C}.
We next prove that $\varphi_{p,q} : \mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. continuous. By Lemma \ref{Lemma:Properties:Of:Phi:P:Q:On:Lambda:C}, it suffices to show that $P^0$-a.s. $\varphi_{p,q}$ is continuous at each point $t \in \Lambda$. Let $\mathcal{H}_1$ be the set of all $\omega \in C_S$ such that the following four conditions hold,
\begin{enumerate}
\item Identity \eqref{display:complement:of:lambda} holds,
\label{early:cond:display:complement}
\item Identity \eqref{nice:rep} holds,
\label{early:cond:nice:rep}
\item $V_p(Y(\cdot,\omega),[G_i,\cdot])$ is continuous on $(G_i,D_i)$ for each $i \geq 1$,
\label{early:cond:continuous}
\item $\varphi_{p,q}(\cdot,\omega):\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is well-defined and non-decreasing.
\label{early:cond:nondecreasing}
\end{enumerate}
By the discussion in Section \ref{Section:Excursion:Theory:Background}, together with the definition of $\varphi_{p,q}$ and Lemmas \ref{Lemma:Properties:Of:Var:Phi:P:Q:On:Lambda} and \ref{Lemma:Properties:Of:Phi:P:Q:On:Lambda:C}, and the discussion of the preceding paragraph, it follows that $P^0(\mathcal{H}_1)$=1. It therefore
suffices to show that $\varphi_{p,q}(\cdot,\omega):\mathbb{R}_{+} \mapsto \mathbb{R}_{+} $ is continuous
for each $\omega \in \mathcal{H}_1$. Hence, fix an arbitrary $\hat{\omega} \in \mathcal{H}_1$ for the remainder of this paragraph. Let $t \in \Lambda$ be arbitrary. We show that $\varphi_{p,q}$ is both left- and right-continuous at $t$. We start with left-continuity at $t$. Obviously then we can assume that $t>0$. If $t$ isolated from the left in $\Lambda$, then since identity \eqref{display:complement:of:lambda} holds by condition \ref{early:cond:display:complement}, $t=D_i$ for some $i \geq 1$, and so left-continuity follows from the fact that by condition \ref{early:cond:nondecreasing}, $\varphi_{p,q}$ is well-defined, the continuity of $V_p(Y,[G_i,\cdot])$ on $(G_i,D_i)$, which is guaranteed by condition \ref{early:cond:continuous}, and the identity \eqref{nice:rep} guaranteed by condition \ref{early:cond:nice:rep}. Suppose now that $t$ is not isolated from the left in $\Lambda$, i.e. there exists a sequence $\{t_n , n\ge 1\} \subset \Lambda$ such that $t_n\uparrow t$ as $n\to\infty$. Then, since \eqref{nice:rep} holds by condition \ref{early:cond:nice:rep}, it follows that
$$\lim_{n\to\infty}\left(\varphi_{p,q}(t) - \varphi_{p,q}(t_n)\right) ~=~ \lim_{n\to\infty}\sum_{i:t_n\le G_i<t}(D_i-G_i)^q~=~0,$$
and so left-continuity follows since by condition \ref{early:cond:nondecreasing}, $\varphi_{p,q}$ is non-decreasing. The proof of right-continuity of $\varphi_{p,q}$ at $t$ is quite similar and so we omit the details.
We now complete the proof by showing that $P^0$-a.s., $\varphi_{p,q}(\mathbb{R}_{+})=\mathbb{R}_{+}$. First note that using the fact from Section \ref{Section:Excursion:Theory:Background} that $L^{-1}$ is a subordinator under $P^0$, and the subadditivity property of strong $q$-variation (see (P3) in Section 2.2 of Chistyakov and Galkin \cite{chistyakov1998maps}), it is straightforward to show that $P^0$-a.s., $V_q(L^{-1},[0,A]) \rightarrow + \infty$ as $A \rightarrow +\infty$. Hence, since by the discussion in Section \ref{Section:Excursion:Theory:Background}, $P^0$-a.s., $L(t) \rightarrow + \infty$ as $t \rightarrow +\infty$, it follows by \eqref{Display:Definition:Of:Varphi} and the fact
that $\varphi_{p,q}:\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. non-decreasing, that $P^0$-a.s., $\varphi_{p,q}(t) \rightarrow + \infty$ as $t \rightarrow + \infty$. Moreover, it is clear from \eqref{Display:Definition:Of:Varphi} that $P^0$-a.s., $\varphi_{p,q}(0)=0$. The desired result now follows since $\varphi_{p,q}$ is $P^0$-a.s. continuous on $\mathbb{R}_{+}$.
\end{proof}
\subsection{Proof of Proposition \ref{Proposition:Variation:Time:Chanaged:Holder}}
By Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}, we have that for each $p > \alpha$ and $1 > q > \alpha/2$ the function $\varphi_{p,q}:\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is $P^0$-a.s. well-defined, non-decreasing and continuous with $\varphi_{p,q}(\mathbb{R}_{+})=\mathbb{R}_{+}$. Hence, we may $P^0$-a.s. define its right-continuous inverse function
\begin{eqnarray*}
\varphi_{p,q}^{-1}(\tau) &=& \inf\{t \geq 0 : \varphi_{p,q}(t) > \tau\},~\tau \geq 0.
\end{eqnarray*}
In preparation for the proof of Proposition \ref{Proposition:Variation:Time:Chanaged:Holder}, we provide two lemmas regarding $\varphi_{p,q}^{-1}$.
\begin{lemma}\label{Lemma:Inverse:Of:Varphi:On:Zero:Set}
For each $p > \alpha$ and $1 > q > \alpha/2$, it follows that $P^0$-a.s., $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in \Lambda$. Moreover, $\varphi_{p,q}^{-1}$ is $P^0$-a.s. H\"{o}lder continuous on $\varphi_{p,q}(\Lambda)$ with H\"{o}lder exponent $1/q$ and H\"{o}lder constant $H(\varphi_{p,q}^{-1}) \leq 1$.
\end{lemma}
\begin{proof}Let $p > \alpha$ and $1 > q > \alpha/2$. Next, let $\mathcal{H}_2$ be the set of all $\omega \in C_S$ such that the following five conditions hold,
\begin{enumerate}
\item There are no isolated points in $\Lambda$ and the identities \eqref{display:complement:of:lambda} and \eqref{Display:Disjoint:Intervals} hold,
\label{middle:cond:lambda:set}
\item $L^{-1}$ is strictly increasing and right-continuous,
\label{middle:cond:l:inverse}
\item $\varphi_{p,q}:\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is well-defined,
non-decreasing and continuous with
$\varphi_{p,q}(\mathbb{R}_{+})=\mathbb{R}_{+}$,
\label{middle:cond:phi:on:r}
\item $\varphi_{p,q}$ is strictly increasing on $\Lambda$,
\label{middle:cond:phi:increasing}
\item $V_1(Y,[G_i,D_i \wedge (G_i+2^{-j})]) > 0$ for $i,j \geq 1$.
\label{middle:cond:phi:excursion:intervals}
\end{enumerate}
By the discussion in Section \ref{Section:Excursion:Theory:Background}, Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}, and Lemmas \ref{Lemma:Properties:Of:Var:Phi:P:Q:On:Lambda} and \ref{Lemma:Y:Positive:Variation}, it follows that $P^{0}(\mathcal{H}_2)=1$. It therefore suffices to show that the results of the lemma hold for each $\omega \in \mathcal{H}_2$. Hence, fix an arbitrary $\hat{\omega} \in \mathcal{H}_2$
for the remainder of the proof.
We first prove that $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in \Lambda$. Since the identities \eqref{display:complement:of:lambda} and \eqref{Display:Disjoint:Intervals} are guaranteed by condition \ref{middle:cond:lambda:set}, no point $t \in L^{-1}(\mathbb{R}_{+})$ is isolated from the right in $\Lambda$. Hence, since by condition \ref{middle:cond:phi:increasing}, $\varphi_{p,q}$ is strictly increasing on $\Lambda $, it follows that $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in L^{1}(\mathbb{R}_{+})$. Next, note that by \eqref{Display:Disjoint:Intervals}, which is guaranteed by condition \ref{middle:cond:lambda:set}, it follows that if $t \in \Lambda \backslash L^{1}(\mathbb{R}_{+})$, then $t=G_i$ for some $i \geq 1$.
Hence, since by condition \ref{middle:cond:phi:excursion:intervals}, $V_1(Y,[G_i,D_i \wedge (G_i+2^{-j})]) > 0$ for $i,j \geq 1$, it follows since by condition \ref{middle:cond:phi:on:r}, $\varphi_{p,q}$ in well-defined and non-decreasing on $\mathbb{R}_{+}$, that using \eqref{Display:Definition:Of:Varphi}, $\varphi_{p,q}(t^{'}) > \varphi_{p,q}(t)$ for $t^{'} > t$. Hence, it follows that $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in \Lambda \backslash L^{1}(\mathbb{R}_{+})$, and so we have now shown that $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in \Lambda$, as desired.
Next, we show that $\varphi_{p,q}^{-1}$ is H\"{o}lder continuous on $\varphi_{p,q}(\Lambda)$ with H\"{o}lder exponent $1/q$ and H\"{o}lder constant $H(\varphi_{p,q}^{-1}) \leq 1$. First note that since by condition \ref{middle:cond:phi:on:r}, $\varphi_{p,q}$ is well-defined on $\mathbb{R}_{+}$, it follows using the definition of $\varphi_{p,q}$ in \eqref{Display:Definition:Of:Varphi} that
\begin{eqnarray*}
\varphi_{p,q}(t) &=& V_{q}(L^{-1},[0,L(t)]),~t \in L^{-1}(\mathbb{R}_{+}).
\end{eqnarray*}
Trivially, $L^{-1}=e \circ L^{-1}$, where $e:\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is the identity function, and so since $L^{-1}$ is non-decreasing by condition \ref{middle:cond:l:inverse}, the change-of-variable formula for strong $q$-variation (see (P4) of Chistyakov and Galkin \cite{chistyakov1998maps}) implies that \begin{eqnarray*}
\varphi_{p,q}(t)~=~ V_{q}(L^{-1},[0,L(t)])&=& V_{q}(e,L^{-1}([0,L(t)]))~=~V_{q}(e,(L^{-1}(\mathbb{R}_{+}))_{t}^{-} ),~t \geq 0,
\end{eqnarray*}
where the final equality follows since by condition \ref{middle:cond:l:inverse}, $L^{-1}$ is non-decreasing, and since $L^{-1} \circ L(t)=t$ for $t \in L^{-1}(\mathbb{R}_{+})$. Now note that since by condition \ref{middle:cond:phi:on:r}, $\varphi_{p,q}$ is continuous on $\mathbb{R}_{+}$ with $\varphi_{p,q}(\mathbb{R}_{+})=\mathbb{R}_{+}$, it follows that $\varphi_{p,q} \circ \varphi^{-1}_{p,q}(\tau)=\tau$ for $\tau \in \mathbb{R}_{+}$. Moreover, since it has been shown in the preceding discussion that $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in \Lambda$, it follows that $\varphi^{-1}_{p,q}(\tau) \in L^{-1}(\mathbb{R}_{+})$ for $\tau \in \varphi_{p,q}(L^{-1}(\mathbb{R}_{+}))$. Thus,
for $\tau, \tau^{'} \in \varphi_{p,q}(L^{-1}(\mathbb{R}_{+}))$ with $\tau \leq \tau^{'}$, by the minimality property and the semi-additivity property of strong $q$-variation (see (P1) and (P3) of Chistyakov and Galkin \cite{chistyakov1998maps}), it follows that
\begin{eqnarray*}
|\varphi_{p,q}^{-1}(\tau^{'})-\varphi_{p,q}^{-1}(\tau) |^{q} &\leq&V_{q}\left(e,\left(L^{-1}(\mathbb{R}_{+})\right)_{\varphi_{p,q}^{-1}(\tau)}^{\varphi_{p,q}^{-1}(\tau^{'})} \right)\\
&\leq&V_{q}\left(e,\left(L^{-1}(\mathbb{R}_{+})\right)_{\varphi_{p,q}^{-1}(\tau^{'})}^{-} \right)-V_{q}\left(e,\left(L^{-1}(\mathbb{R}_{+})\right)_{\varphi_{p,q}^{-1}(\tau)}^{-} \right)\\
&=&\tau^{'}-\tau.
\end{eqnarray*}
Thus, $\varphi_{p,q}^{-1}$ meets the desired H\"{o}lder conditions on $\varphi_{p,q}(L^{-1}(\mathbb{R}_{+}))$.
We now complete the proof by showing that $\varphi_{p,q}^{-1}$ meets the desired H\"{o}lder conditions on all of
$\varphi_{p,q}(\Lambda)$. Let $\tau,\tau^{'} \in \varphi_{p,q}(\Lambda)$.
If $\tau,\tau^{'} \in \varphi_{p,q}(L^{-1}(\mathbb{R}_{+})) $, then the H\"{o}lder conditions are met by the the preceding paragraph. On the other hand, suppose that $\tau \in \varphi_{p,q}(L^{-1}(\mathbb{R}_{+})) $
and $\tau^{'} \in \varphi_{p,q}(\Lambda) ~\backslash~ \varphi_{p,q}(L^{-1}(\mathbb{R}_{+})) $. Let $t \in L^{-1}(\mathbb{R}_{+})$ and $t^{'} \in \Lambda \backslash L^{-1}(\mathbb{R}_{+})$
be such that $\varphi_{p,q}(t)=\tau$ and $\varphi_{p,q}(t^{'})=\tau^{'}$. By the identity \eqref{Display:Disjoint:Intervals}, which holds by condition \ref{middle:cond:lambda:set}, it follows that $t^{'}=G_i$ for some $i \geq 1$. Thus, using the fact by condition \ref{middle:cond:lambda:set} that there are no isolated points in $\Lambda$ and that the identities \eqref{display:complement:of:lambda} and \eqref{Display:Disjoint:Intervals} hold, it is straightforward to show that there exists a non-decreasing sequence $\{t_k^{'}\}$ of elements of $L^{-1}(\mathbb{R}_{+})$ such that $t_k^{'} \uparrow G_i$ and $L(t_k^{'}) \uparrow L(G_i)$. Hence, by the continuity of $\varphi_{p,q}$ on $\mathbb{R}_{+}$, which is guaranteed by condition \ref{middle:cond:phi:on:r}, it follows that $\tau_k^{'}=\varphi_{p,q}(t^{'}_k) \uparrow \varphi_{p,q}(G_i)=\tau^{'}$. Thus, since $\tau_k^{'} \in \varphi_{p,q}(L^{-1}(\mathbb{R}_{+}))$, and noting that $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in \Lambda$ implies that $\varphi_{p,q}^{-1}(\tau_k^{'})=t_k^{'} \uparrow t^{'}=\varphi_{p,q}^{-1}(\tau^{'}) $, it follows that
\begin{eqnarray*}
|\varphi_{p,q}^{-1}(\tau)-\varphi_{p,q}^{-1}(\tau^{'})| ~=~\lim_{k \rightarrow \infty}|\varphi_{p,q}^{-1}(\tau)-\varphi_{p,q}^{-1}(\tau^{'}_k)| &\leq&
\lim_{k \rightarrow \infty}|\tau-\tau^{'}_k|^{1/q}~=~|\tau-\tau^{'}|^{1/q}.
\end{eqnarray*}
In a similar manner, it may be shown that $\varphi_{p,q}^{-1}$ meets the desired H\"{o}lder conditions if $\tau,\tau^{'} \in \varphi_{p,q}(\Lambda) ~\backslash~ \varphi_{p,q}(L^{-1}(\mathbb{R}_{+})) $.
\end{proof}
Now, for each $p > \alpha$ and $1 > q > \alpha/2$, let
\begin{eqnarray}
\Phi_{i} &=& \left(\frac{V_1(\varphi_{p,q} ,[G_i,D_i])}{V_p(Y ,[G_i,D_i])} \right)^{1/p},~~i \geq 1. \label{Display:Definition:Phi:I}
\end{eqnarray}
\begin{lemma}\label{Lemma:Inverse:Of:Varphi:On:Excursions}
For each $p > \alpha$ and $1 > q > \alpha/2$ and $i \geq 1$, it follows that $P^0$-a.s., $Y \circ \varphi_{p,q}^{-1} \circ \varphi_{p,q}(t)=Y(t)$ on $t \in [G_i,D_i]$ . Moreover, $Y \circ \varphi_{p,q}^{-1}$ is $P^0$-a.s. H\"{o}lder continuous on $[\varphi_{p,q}(G_i),\varphi_{p,q}(D_i)]$
with H\"{o}lder exponent $1/p$ and H\"{o}lder constant $H_i(Y \circ \varphi_{p,q}^{-1}) \leq 1/\Phi_i$.
\end{lemma}
\begin{proof} Let $p > \alpha$ and $1 > q > \alpha/2$ and $i \geq 1$. Since $\varphi_{p,q}:\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is, by Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}, $P^{0}$-a.s. well-defined, it follows by \eqref{Display:Definition:Of:Varphi} that $P^0$-a.s.,
\begin{eqnarray}
\frac{V_1(\varphi_{p,q} ,[G_i,D_i])}{V_p(Y ,[G_i,D_i])} \cdot V_p(Y ,[G_i,t])&=& V_p\left( \Phi_i Y ,[G_i,t]\right),~t \in [G_i,D_i],~~\label{Display:Linear:P:Variation:Y:Interval}
\end{eqnarray}
where $\Phi_i$ is as defined in \eqref{Display:Definition:Phi:I} above. Next, define the right-continuous inverse function $V_p^{-1}(\Phi_i Y, \cdot):[0, V_p\left( \Phi_i Y ,[G_i,D_i]\right)] \mapsto [G_i,D_i]$ by
\begin{eqnarray*}
V_p^{-1}(\Phi_i Y, \tau)&=& \inf\{t \in [G_i,D_i] : V_p( \Phi_i Y ,[G_i,t]) > \tau \},~\tau \in [0, V_p\left( \Phi_i Y ,[G_i,D_i]\right)],
\end{eqnarray*}
where $\inf \emptyset = D_i$. Since $Y$ is $P^0$-a.s. continuous, it follows that $V_p(\Phi_i Y, \cdot)$ is $P^0$-a.s. continuous as well,
and so $P^0$-a.s.,
\begin{eqnarray*}
V_p(\Phi_i Y, V_p^{-1}(\Phi_i Y,\tau ) )&=&\tau,~\tau \in [0, V_p\left( \Phi_i Y ,[G_i,D_i]\right)].
\end{eqnarray*}
It then follows that $P^0$-a.s.,
\begin{eqnarray*}
Y \circ V_p^{-1}(\Phi_i Y) \circ V_p(\Phi_i Y, [G_i,t])&=&Y(t),~t \in [G_i,D_i],
\end{eqnarray*}
and, by the discussion immediately following Proposition \ref{Proposition:Finite:Variation:Holder}, $Y \circ V_p^{-1}(\Phi_i Y, \cdot):[0, V_p\left( \Phi_i Y ,[G_i,D_i]\right)] \mapsto \mathbb{R}^2$
is $P^0$-a.s. H\"{o}lder continuous with H\"{o}lder exponent $1/p$ and H\"{o}lder constant $H(Y \circ V_p^{-1}(\Phi_i Y, \cdot)) \leq 1/\Phi_i$.
Now recall by \eqref{Display:Definition:Of:Varphi} that $P^0$-a.s.,
\begin{eqnarray}
\varphi_{p,q}(t) &=&
\varphi_{p,q}(G_i)+ V_p( \Phi_i Y ,[G_i,t]),~t \in [G_i,D_i]. \label{display:vp:time:changed}
\end{eqnarray}
Moreover, by \eqref{Display:Linear:P:Variation:Y:Interval}, $P^0$-a.s.,
\begin{eqnarray*}
V_p\left( \Phi_i Y ,[G_i,D_i]\right) &=& V_1\left( \varphi_{p,q} ,[G_i,D_i]\right)~=~\varphi_{p,q}(D_i)-\varphi_{p,q}(G_i).
\end{eqnarray*}
Hence, in order to complete the proof, it suffices by the preceding paragraph to show that $P^0$-a.s.,
\begin{eqnarray*}
V_p^{-1}(\Phi_i Y, \tau) &=&\varphi_{p,q}^{-1}(\varphi_{p,q}(G_i)+\tau),~\tau \in [0, V_p\left( \Phi_i Y ,[G_i,D_i]\right)]=[0,\varphi_{p,q}(D_i)-\varphi_{p,q}(G_i) ].
\end{eqnarray*}
However, the fact that by Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}, $\varphi_{p,q}$ is $P^0$-a.s. non-decreasing, implies that $P^0$-a.s.,
\begin{eqnarray*}
\varphi_{p,q}^{-1}(\varphi_{p,q}(G_i)+\tau) &=&\inf\{t \geq 0 : \varphi_{p,q}(t) > \varphi_{p,q}(G_i)+\tau \}\\
&=&\inf\{t \in [G_i,D_i] : \varphi_{p,q}(t) > \varphi_{p,q}(G_i)+\tau \},
\end{eqnarray*}
for $\tau \in [0,\varphi_{p,q}(D_i)-\varphi_{p,q}(G_i)]$.
Then, by \eqref{display:vp:time:changed}, $P^0$-a.s.,
\begin{eqnarray*}
\inf\{t \in [G_i,D_i] : \varphi_{p,q}(t) > \varphi_{p,q}(G_i)+\tau \}&=&\inf\{t \in [G_i,D_i] : V_p( \Phi_i Y ,[G_i,t]) > \tau \}\\
&=&V_p^{-1}(\Phi_i Y, \tau),
\end{eqnarray*}
for $\tau \in [0,\varphi_{p,q}(D_i)-\varphi_{p,q}(G_i)]$.
\end{proof}
The proof of Proposition \ref{Proposition:Variation:Time:Chanaged:Holder} is as follows.
\begin{proof}[Proof of Proposition \ref{Proposition:Variation:Time:Chanaged:Holder}]Combining Lemmas \ref{Lemma:Inverse:Of:Varphi:On:Zero:Set} and \ref{Lemma:Inverse:Of:Varphi:On:Excursions} above, it follows that for each $p > \alpha$ and $1 > q > \alpha/2$, $P^0$-a.s., $Y \circ \varphi^{-1}_{p,q} \circ \varphi_{p,q}(t)=Y(t)$ for $t \in \mathbb{R}_{+}$.
It is also immediate by Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q} that $\varphi_{p,q}$ is $P^0$-a.s. non-decreasing on $\mathbb{R}_{+}$, with $\varphi_{p,q}(\mathbb{R}_{+})=\mathbb{R}_{+}$. Hence, in order to complete the proof, it suffices by Proposition \ref{Proposition:Finite:Variation:Holder} and the discussion after it to show that for each $\alpha < p < 2$, there exists some $\alpha/2 < q < p/2 < 1$ such that $Y \circ \varphi^{-1}_{p,q} $ is $P^0$-a.s. locally H\"{o}lder continuous on $\mathbb{R}_{+}$ with H\"{o}lder exponent $1/p$.
Let $\mathcal{H}_3$ be the set of all $\omega \in C_S$ such that the following six conditions hold,
\begin{enumerate}
\item Identity \eqref{Display:Disjoint:Intervals} holds with $D_i \rightarrow + \infty$ as $i \rightarrow \infty$,
\label{final:cond:disjoint}
\item $\varphi_{p,q}:\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is well-defined, non-decreasing and continuous, with $\varphi_{p,q}(\mathbb{R}_{+})=\mathbb{R}_{+}$, \label{final:cond:phi:well:defined}
\item $\varphi_{p,q}^{-1} \circ \varphi_{p,q} (t)=t$ for $t \in \Lambda$, and $\varphi_{p,q}^{-1}$ H\"{o}lder continuous on $\varphi_{p,q}(\Lambda)$ with H\"{o}lder exponent $1/q$ and H\"{o}lder constant $H(\varphi_{p,q}^{-1}) \leq 1$,
\label{final:cond:holder:on:lambda}
\item For each $i \geq 1$, $Y \circ \varphi_{p,q}^{-1} \circ \varphi_{p,q}(t)=Y(t)$ on $t \in [G_i,D_i]$ . Moreover, $Y \circ \varphi_{p,q}^{-1}$ is H\"{o}lder continuous on $[\varphi_{p,q}(G_i),\varphi_{p,q}(D_i)]$
with H\"{o}lder exponent $1/p$ and H\"{o}lder constant $H_i(Y \circ \varphi_{p,q}^{-1}) \leq 1/\Phi_i$,
\label{final:cond:holder:on:excursion}
\item $X:\mathbb{R}_{+} \mapsto \mathbb{R}^2$ is locally H\"{o}lder continuous with H\"{o}lder exponent $\eta$ for each $0 \leq \eta < 1/2$, \label{final:cond:brownian:holder}
\item The components of $R^{-1}Y$ are non-decreasing on $[G_i,D_i]$ for each $i \geq 1$.
\label{final:cond:components:y}
\end{enumerate}
By the discussion in Section \ref{Section:Excursion:Theory:Background}, it is straightforward to show that condition \ref{final:cond:disjoint} holds $P^0$-a.s. Proposition \ref{Proposition:Main:Properties:Of:Varphi:P:Q}, and Lemmas \ref{Lemma:Y:Positive:Variation} and \ref{Lemma:Inverse:Of:Varphi:On:Zero:Set} imply that conditions \ref{final:cond:phi:well:defined} through \ref{final:cond:holder:on:excursion} hold $P^0$-a.s. By Part \ref{Proposition:Skorohod:Problem:Part:1} of Theorem \ref{Proposition:Skorokhod:Problem} of Section \ref{Section:Identification:of:X:and:Skorokhod:Problem}, $X$ is a standard 2-dimensional Brownian motion under $P^0$, and so condition \ref{final:cond:brownian:holder} holds $P^0$-a.s. by Remark 2.12 of \cite{KaratzasShreve}. Condition \ref{final:cond:components:y} holds $P^0$-a.s. by Part \ref{Proposition:Skorohod:Problem:Part:2} of Theorem \ref{Proposition:Skorokhod:Problem}. Thus, $P^{0}(\mathcal{H}_3)=1$. In order to complete the proof, it therefore suffices to show that for each $\omega \in \mathcal{H}_3$, for each $\alpha < p < 2$, there exists some $\alpha/2 < q < p/2 < 1$ such that $Y \circ \varphi^{-1}_{p,q}(\cdot,\omega) $ is locally H\"{o}lder continuous on $\mathbb{R}_{+}$ with H\"{o}lder exponent $1/p$. Hence, fix an arbitrary $\hat{\omega} \in \mathcal{H}_3$ for the remainder of the proof.
First note that since by condition \ref{final:cond:disjoint}, $D_i \rightarrow + \infty$ as $i \rightarrow \infty$, and since by condition \ref{final:cond:phi:well:defined}, $\varphi_{p,q}(t) \rightarrow +\infty$ as $t \rightarrow +\infty$, it suffices to show that
$Y \circ \varphi^{-1}_{p,q} $ is H\"{o}lder continuous on $[0,\varphi_{p,q}(D_i)]$ for each $i \geq 1$. Let $i \geq 1$ and $\tau,\tau^{'} \in [0,\varphi_{p,q}(D_i)]$ with $0 \leq \tau \leq \tau^{'} \leq \varphi_{p,q}(D_i)$ such that $\tau,\tau^{'} \in \varphi_{p,q}(\Lambda)$. Then, by condition \ref{final:cond:holder:on:lambda}, $\varphi_{p,q}^{-1}(\tau),\varphi_{p,q}^{-1}(\tau^{'}) \in \Lambda $, and so $Y \circ \varphi_{p,q}^{-1}(\tau)=-X \circ \varphi_{p,q}^{-1}(\tau)$ and $Y \circ \varphi_{p,q}^{-1}(\tau^{'})=-X \circ \varphi_{p,q}^{-1}(\tau^{'})$. This then implies that
\begin{eqnarray}
\| Y \circ \varphi_{p,q}^{-1}(\tau^{'})-Y \circ \varphi_{p,q}^{-1}(\tau) \|&=&\| X \circ \varphi_{p,q}^{-1}(\tau^{'})-X \circ \varphi_{p,q}^{-1}(\tau) \|. \label{Display:Increment:Y:On:Image:Lambda}
\end{eqnarray}
Now, by condition \ref{final:cond:brownian:holder}, for each $0 \leq \eta < 1/2$ and $T \geq 0$, there exists a constant $C_{\eta,T}$ such that
\begin{eqnarray}
\| X(t)-X(s)\| &\leq& C_{\eta,T} |t-s|^{\eta},~0 \leq s \leq t \leq T. \label{Display:Brownian:Local:Holder}
\end{eqnarray}
Moreover, by condition \ref{final:cond:holder:on:lambda} and the fact that $\tau,\tau^{'} \in [0,\varphi_{p,q}(D_i)]$ with $\tau,\tau^{'} \in \varphi_{p,q}(\Lambda)$, it follows since $\varphi_{p,q}:\mathbb{R}_{+} \mapsto \mathbb{R}_{+}$ is non-decreasing by condition \ref{final:cond:phi:well:defined}, that $0 \leq \varphi_{p,q}^{-1}(\tau) \leq \varphi_{p,q}^{-1}(\tau^{'}) \leq D_i$. Therefore, by \eqref{Display:Brownian:Local:Holder}, for each $0 \leq \eta < 1/2$, it follows again by condition \ref{final:cond:holder:on:lambda} that
\begin{eqnarray}
\| X \circ \varphi_{p,q}^{-1}(\tau^{'})-X \circ \varphi_{p,q}^{-1}(\tau) \| &\leq&C_{\eta,D_i}|\varphi_{p,q}^{-1}(\tau^{'})- \varphi_{p,q}^{-1}(\tau) |^{\eta} ~\leq~C_{\eta,D_i}|\tau^{'}- \tau |^{\eta/q}. \label{Display:Increment:X:On:Image:Lambda}
\end{eqnarray}
Now let $0 \leq \eta^{\star} < 1/2$ be such that $ q < \eta^{\star} p$. Note that such an $\eta^{\star}$ exists since by assumption $q < p/2$.
Then $\eta^{\star} /q > 1/p$, and since $0 \leq \tau \leq \tau^{'} \leq \varphi_{p,q}(D_i)$, it follows that
$|\tau^{'}- \tau |^{\eta^{\star}/q} \leq K_{\eta^{\star},i} |\tau^{'}- \tau |^{1/p}$, where $K_{\eta^{\star},i} = (\varphi_{p,q}(D_i))^{\eta^{\star}/q-1/p}$. Hence, by \eqref{Display:Increment:Y:On:Image:Lambda} and \eqref{Display:Increment:X:On:Image:Lambda},
\begin{eqnarray}
\| Y \circ \varphi_{p,q}^{-1}(\tau^{'})-Y \circ \varphi_{p,q}^{-1}(\tau) \| &\leq& K_{\eta^{\star},i} C_{\eta^{\star},D_i}|\tau^{'}-\tau |^{1/p}. \label{Display:Bound:Increment:Y:On:Image:Lambda}
\end{eqnarray}
Next, suppose that $\tau,\tau^{'} \in [0,\varphi_{p,q}(D_i)]$ with $0 \leq \tau \leq \tau^{'} \leq \varphi_{p,q}(D_i)$, and such that $\tau \in \varphi_{p,q}(\Lambda)$ and $\tau^{'} \notin \varphi_{p,q}(\Lambda)$.
In this case, by conditions \ref{final:cond:disjoint} and \ref{final:cond:phi:well:defined}, it follows that $\tau^{'} \in (\varphi_{p,q}(G_j),\varphi_{p,q}(D_j))$ for some $j \geq 1$. By condition \ref{final:cond:holder:on:excursion}, it follows that
\begin{eqnarray*}
\| Y \circ \varphi_{p,q}^{-1}(\tau^{'})-Y \circ \varphi_{p,q}^{-1}(\varphi_{p,q}(G_j)) \| &\leq (1/\Phi_j)|\tau^{'}-\varphi_{p,q}(G_j) |^{1/p}.
\end{eqnarray*}
Also since $\varphi_{p,q}(G_j) \in \varphi_{p,q}(\Lambda) \cap [0,\varphi_{p,q}(D_i)]$, it follows by \eqref{Display:Bound:Increment:Y:On:Image:Lambda} that
\begin{eqnarray*}
\| Y \circ \varphi_{p,q}^{-1}(\varphi_{p,q}(G_j))-Y \circ \varphi_{p,q}^{-1}(\tau) \| &\leq&K_{\eta^{\star},i} C_{\eta^{\star},D_i}|\varphi_{p,q}(G_j)- \tau |^{1/p}.
\end{eqnarray*}
Hence, since $0 \leq \tau \leq \varphi_{p,q}(G_j) < \tau^{'} \leq \varphi_{p,q}(D_i)$, it follows by the triangle inequality that
\begin{eqnarray}
\| Y \circ \varphi_{p,q}^{-1}(\tau^{'})-Y \circ \varphi_{p,q}^{-1}(\tau) \| &\leq ((1/\Phi_j)+ K_{\eta^{\star},i} C_{\eta^{\star},D_i})|\tau^{'}-\tau |^{1/p}. \label{Display:Bound:Increment:Y:On:Lambda:Not:Lambda}
\end{eqnarray}
Similar reasoning leads to the same inequality if we assume that $\tau,\tau^{'} \in [0,\varphi_{p,q}(D_i)]$ with $0 \leq \tau \leq \tau^{'} \leq \varphi_{p,q}(D_i)$, and such that $\tau \notin \varphi_{p,q}(\Lambda)$ and $\tau^{'} \in \varphi_{p,q}(\Lambda)$.
Next consider $\tau,\tau^{'} \in [0,\varphi_{p,q}(D_i)]$ with $0 \leq \tau \leq \tau^{'} \leq \varphi_{p,q}(D_i)$, and such that $\tau,\tau^{'} \notin \varphi_{p,q}(\Lambda)$. Suppose first that
$\tau \in (\varphi_{p,q}(G_j),\varphi_{p,q}(D_j))$ and $\tau \in (\varphi_{p,q}(G_k),\varphi_{p,q}(D_k))$ with $j \neq k$. In this case, similar reasoning to the above leads to the bound
\begin{eqnarray}
\| Y \circ \varphi_{p,q}^{-1}(\tau^{'})-Y \circ \varphi_{p,q}^{-1}(\tau) \| &\leq ((1/\Phi_j)+ K_{\eta^{\star},i} C_{\eta^{\star},D_i}+(1/\Phi_k))|\tau^{'}-\tau |^{1/p}. \label{Display:Bound:Increment:Y:Not:Lambda:Not:Lambda:Big:Diff}
\end{eqnarray}
On the other hand, if $\tau,\tau^{'} \in (\varphi_{p,q}(G_j),\varphi_{p,q}(D_j))$, it is immediate by condition \ref{final:cond:holder:on:excursion} that
\begin{eqnarray}
\| Y \circ \varphi_{p,q}^{-1}(\tau^{'})-Y \circ \varphi_{p,q}^{-1}(\varphi_{p,q}(\tau)) \| &\leq (1/\Phi_j)|\tau^{'}-\tau |^{1/p}. \label{Display:Bound:Increment:Y:Not:Lambda:Not:Lambda:Small:Diff}
\end{eqnarray}
Combining the cases \eqref{Display:Bound:Increment:Y:On:Image:Lambda},\eqref{Display:Bound:Increment:Y:On:Lambda:Not:Lambda},\eqref{Display:Bound:Increment:Y:Not:Lambda:Not:Lambda:Big:Diff} and \eqref{Display:Bound:Increment:Y:Not:Lambda:Not:Lambda:Small:Diff} we now obtain that
\begin{eqnarray*}
\| Y \circ \varphi_{p,q}^{-1}(\tau^{'})-Y \circ \varphi_{p,q}^{-1}(\tau) \| &\leq (2 \Theta_i+ K_{\eta^{\star},i} C_{\eta^{\star},D_i})|\tau^{'}-\tau |^{1/p}~\textrm{for}~\tau,\tau^{'} \in [0,\varphi_{p,q}(D_i)],
\end{eqnarray*}
where
\begin{eqnarray*}
\Theta_i &=& \sup \{(1/\Phi_k) : D_k \leq D_i\}.
\end{eqnarray*}
Hence, in order to show that $Y \circ \varphi^{-1}_{p,q} $ is H\"{o}lder continuous on $[0,\varphi_{p,q}(D_i)]$ with H\"{o}lder exponent $1/p$, it suffices to show that $\Theta_i < + \infty$.
Using the definition of $\Phi_i$ in \eqref{Display:Definition:Phi:I} it suffices to show that there exists $ 0 < K_i < + \infty$ such that for $k \geq 1$,
\begin{eqnarray}
1\{D_k \leq D_i\} \left(\frac{V_p(Y ,[G_k,D_k])}{V_1(\varphi_{p,q} ,[G_k,D_k])} \right)^{1/p} &<& K_i . \label{Display:Inequality:Between:Variations}
\end{eqnarray}
Recall the definition of the reflection matrix $R$ from Section \ref{Subsection:Semimartingale:Decomposition:Of:X:On:Excursion:Intervals}
and let $\bar{Y}=R^{-1}Y$. By condition \ref{final:cond:components:y}, the components of $\bar{Y} $ are non-decreasing on $[G_k,D_k]$ for each $k \geq 1$. Thus, since $Z = X + Y $ and recalling that $Z(G_k)=Z(D_k)=0$, it follows that
\begin{eqnarray*}
Y(D_k)-Y(G_k) &=& - (X(D_k)-X(G_k)).
\end{eqnarray*}
On the other hand, by the definition of $\bar{Y}$,
\begin{eqnarray*}
\bar{Y}(D_k)-\bar{Y}(G_k) &=&R^{-1}(Y(D_k)-Y(G_k)).
\end{eqnarray*}
Hence, using the local H\"{o}lder continuity \eqref{Display:Brownian:Local:Holder} of $X$ given by condition \ref{final:cond:brownian:holder}, it follows that for $k \geq 1$ such that $D_k \leq D_i$,
\begin{eqnarray}
\| \bar{Y}(D_k)-\bar{Y}(G_k) \|~\leq~ \| R^{-1}\| \|X(D_k)-X(G_k)\| &\leq& \| R^{-1}\| C_{\eta,D_i} (D_i-G_i)^{\eta},\label{Display:YBar:Less:Than:Holder}
\end{eqnarray}
for each $0 \leq \eta < 1/2$, where $\|R^{-1}\|$ denotes the operator norm of $R^{-1}$. Next, for $x = (x_1,x_2) \in \mathbb{R}^2$, it is straightforward to show that
\begin{eqnarray*}
\|x \|^{\rho}&\leq&|x_1|^{\rho}+|x_2|^{\rho}~\textrm{for}~0\leq \rho \leq 2.
\end{eqnarray*}
Hence, letting $\bar{Y}=(\bar{Y}_1,\bar{Y}_2)$ and using the definition of strong $p$-variation, it is immediate since $\alpha < p < 2$ that
\begin{eqnarray*}
V_p(\bar{Y},[G_k,D_k]) &\leq& V_p(\bar{Y}_1,[G_k,D_k])+V_p(\bar{Y}_2,[G_k,D_k]).
\end{eqnarray*}
However, since $\bar{Y} $ has components which are non-decreasing on $[G_k,D_k]$ and $1 < \alpha < p $, it follows that
\begin{eqnarray*}
V_p(\bar{Y}_{\ell},[G_k,D_k]) &=& (\bar{Y}_{\ell}(D_k)-\bar{Y}_{\ell}(G_k) )^{p}~\leq~ \|\bar{Y}(D_k)-\bar{Y}(G_k)\|^{p}~\textrm{for}~\ell=1,2.
\end{eqnarray*}
Thus, by \eqref{Display:YBar:Less:Than:Holder},
\begin{eqnarray*}
V_p(\bar{Y} ,[G_k,D_k]) &\leq&\bar{C}_{\eta,D_i}(D_k-G_k)^{\eta p},
\end{eqnarray*}
where $\bar{C}_{\eta,D_i}=2 (\| R^{-1}\|C_{\eta,D_i})^{p} $. Next, note that
\begin{eqnarray*}
V_p(Y ,[G_k,D_k])&=&V_p(R\bar{Y} ,[G_k,D_k])~\leq~\|R\|^{p} V_p(\bar{Y},[G_k,D_k]),
\end{eqnarray*}
where $\|R\|$ denotes the operator norm of $R$. Hence,
\begin{eqnarray*}
V_p(Y ,[G_k,D_k])&\leq&\bar{\bar{C}}_{\eta,D_i}(D_k-G_k)^{\eta p}, \end{eqnarray*}
where $\bar{\bar{C}}_{\eta,D_i}=2 (\|R\|\| R^{-1}\|C_{\eta,D_i})^{p} $. Now again let $0 \leq \eta^{\star} < 1/2$ be such that $ q < \eta^{\star} p$.
Then, since $(D_k-G_k) < D_i$, it follows that
\begin{eqnarray*}
\frac{V_p(Y ,[G_k,D_k])}{(D_k-G_k)^q }&\leq&\bar{\bar{\bar{C}}}_{\eta^{\star},D_i},
\end{eqnarray*}
where $\bar{\bar{\bar{C}}}_{\eta,D_i} =(D_i)^{\eta p-q} \bar{\bar{C}}_{\eta,D_i} $. Hence, \eqref{Display:Inequality:Between:Variations} holds with $K_i=(\bar{\bar{\bar{C}}}_{\eta^{\star},D_i})^{1/p}$.
\end{proof}
\section{Proof of Theorem \ref{Theorem:Converse:Dirichlet:Process:Result}}
\label{section:converse:dirichlet}
Theorem \ref{Theorem:Main:Dirichlet:Process:Result} implies that $Z$ has the unique decomposition $Z=X+Y$,
where for each $z \in S$, $X$ is a standard 2-dimensional Brownian motion started at $z$ and $Y$ is a process of zero energy on
$(C_S,\mathcal{F},\mathcal{F}_t,P^z)$. In the proof
of Theorem \ref{Theorem:Main:Dirichlet:Process:Result}, it was also shown (see \eqref{display:used:before:in:proof} and the ensuing arguments) that for each $2 > p > \alpha$ and $z \in S$,
\begin{eqnarray}
P^z(V_p(Y,[0,T]) < \infty)&=&1,~T \geq 0. \label{display:use:initally:prove:roughness}
\end{eqnarray}
The fact that \eqref{display:use:initally:prove:roughness} may be extended to all $p > \alpha$ is an immediate consequence of the identity $V_q(Y,[0,T]) \leq (V_p(Y,[0,T]))^{q/p} $ for $1 \leq p \leq q$ (see Remark 2.5 of Chistyakov and Galkin \cite{chistyakov1998maps}). Hence, in order prove Theorem \ref{Theorem:Converse:Dirichlet:Process:Result} it only remains to prove the partial converse result \eqref{Display:Condition:Converse:Main:Dirichlet}.
Recall from Section \ref{Section:Excursion:Theory:Background} the definition of $L$ as the local time of $Z$ at the origin and let $Y \circ L^{-1}$ be the process defined by $Y \circ L^{-1}=\{Y \circ L^{-1}(t), t \geq 0 \}$. For each $t \geq 0$, we have by \eqref{def:inverse:local:time} that $L^{-1}(t)$ is a stopping time relative to the filtration $\mathcal{F}_t$ and so we may define the new filtration $\mathcal{F}_{L^{-1}}=\{\mathcal{F}_{L^{-1}(t)}, t \geq 0 \} $.
\begin{proposition}\label{Lemma:Y:Composed:With:L:Inverse:Is:Levy} For each $z \in S$, the process $Y \circ L^{-1}$ is a L\'{e}vy process on $(C_S,\mathcal{F},\mathcal{F}_{L^{-1}},P^z)$. In particular, it is an $\alpha$-stable process on $(C_S,\mathcal{F},\mathcal{F}_{L^{-1}},P^0)$, i.e. under $P^0$,
\begin{equation}\lambda^{-{1\over\alpha}} Y \circ L^{-1}(\lambda\cdot) \buildrel d \over = Y \circ L^{-1}(\cdot),\quad \lambda>0.\label{stable}\end{equation}
\end{proposition}
\begin{proof}
We start with a remark on notation. It is customary to write $L(t)$ instead of $L(t,\omega)=L(t,Z(\cdot,\omega))$, as we did so far in this paper. However, in this proof occasionally we need to write down the local time at time $t$ calculated for an element of $C_s$ other than $\omega=Z(\cdot,\omega)$. For example, for the additivity property, we need to write down the local time at time $t-s$ applied to $Z(s+\cdot,\omega)$ for some $0\le s\le t$. In order to strike a compromise between clarity and brevity, we shall write $L(t)$ instead of $L(t,\omega)= L(t,Z(\cdot,\omega))$. However, for $u,s\ge 0$ we shall write $L(u,Z(s+\cdot))$ instead of $L(u, Z(s+\cdot,\omega))$.
In this spirit, we write the additivity property as
\begin{equation}L(t)=L(s)+ L(t-s,Z(s+\cdot)),\label{add}\end{equation}
$P^z$-a.s., for all $z\in S$, and $0\le s \le t$. The exceptional set on which \eqref{add} does not hold does not depend on $s$ and $t$ since $L$ is assumed to be perfect (see Section \ref{Section:Excursion:Theory:Background}).
Let $0\le a < b$ be arbitrary. Then applying \eqref{add} to $s=L^{-1}(a)$, we obtain
\begin{eqnarray}
L^{-1}(b) &=& \inf\Big\{t\ge L^{-1}(a): L(t)>b\Big\} \nonumber\\
&=&\inf\Big\{t\ge L^{-1}(a): a+ L \big(t-L^{-1}(a) ,Z(L^{-1}(a)+\cdot) \big) >b\Big\} \nonumber\\
&=&L^{-1}(a)+\inf\Big\{u\ge 0: L \big(u ,Z(L^{-1}(a)+\cdot) \big) >b-a\Big\} \nonumber \\
&=&L^{-1}(a) +L^{-1}(b-a, Z(L^{-1}(a)+\cdot)),\label{localtime}
\end{eqnarray}
$P^z$-a.s., for every $z\in S$.
Then, for any $B\in{\mathcal B}(S)$, by \eqref{gamma:z}, \eqref{F:additive}, \eqref{localtime}, \eqref{z:is:zero}, the Strong Markov Property (see Theorem 3.14 of Varadhan and Williams \cite{varadhan1985brownian}), and the fact that by \ref {Display:Disjoint:Intervals} $Z(L^{-1}(a))=0$, we have
\begin{eqnarray*}
&&P^z\left(Y(L^{-1}(b))- Y(L^{-1}(a))\in B\ |\ {\mathcal F}_{L^{-1}(a)}\right)\\
&=&
P^z\left(X(L^{-1}(a))- X(L^{-1}(b))\in B\ |\ {\mathcal F}_{L^{-1}(a)}\right)\\
&=&P^z\left(F_{\gamma^z}\left(Z(\cdot)\right)\left(L^{-1}(a)\right)- F_{\gamma^z}\left(Z(\cdot)\right)\left(L^{-1}(b)\right) \in B \ |\ {\mathcal F}_{L^{-1}(a)}\right)\\
&=&P^z\left(-F_{\gamma^z}\left(Z\left(L^{-1}(a)+\cdot\right)\right)\left(L^{-1}(b) - L^{-1}(a)\right)\in B\ |\ {\mathcal F}_{L^{-1}(a)}\right)\\
&=&P^z\left(-F_{\gamma^z}\left(Z\left(L^{-1}(a)+\cdot\right)\right)\left(L^{-1}\left(b-a,Z\left(L^{-1}(a)+\cdot\right)\right)\right)\in B \ |\ {\mathcal F}_{L^{-1}(a)}\right)\\
&=&P^0\left(-F_{\gamma^z}(Z(\cdot)) \left(L^{-1}(b-a)\right)\in B \right)\\
&=& P^0\left(-X(L^{-1}(b-a))\in B\right) =P^0\left(Y(L^{-1}(b-a))\in B\right),
\end{eqnarray*}
which shows that $Y \circ L^{-1}$ is indeed a L\'{e}vy process on $(C_S,\mathcal{F},\mathcal{F}_{L^{-1}},P^z)$. Next, we show that it is also an $\alpha$-stable process on $(C_S,\mathcal{F},\mathcal{F}_{L^{-1}},P^0)$.
Recall the function $\phi:S\mapsto{\mathbb R}$ from \cite{williams1987local} given by $\phi(r,\theta)=r^\alpha\cos(\alpha\theta-\theta_1)$, $r\ge 0$, $0\le\theta\le\xi$ ($r$ and $\theta$ are polar coordinates). It is shown in \cite{williams1987local} that $\{\phi(Z(t)), t \ge 0\}$ is a local submartingale on $(C_S, {\mathcal F},\mathcal {F}_t,P^z)$ for each $z\in S$, that can be uniquely decomposed as
\begin{equation}\phi(Z(t))=\phi(Z(0))+M(t)+A(t),\label{decompose}\end{equation}
where $M$ is a continuous local martingale, and $A$ is an adapted, measurable, non-decreasing, continuous process, and $M(0)=A(0)=0$, $P^z$-a.s. (see (2.6) in \cite{williams1987local}). It is also shown in \cite{williams1987local} that the processes $A$ and $M$ can be selected so that they do not depend on $z$. Moreover, by Theorem 2.6 in \cite{williams1987local} we have that
\begin{equation}L(t)=cA(t),\ t\ge 0,~~ P^z\textrm{-}\hbox{a.s.,}\label{la}\end{equation}
for some positive constant $c$.
Next for all $\lambda>0$ we introduce the transformed processes
$$A_\lambda(t)=\lambda^{-\alpha/2}A(\lambda t),\ \ X_\lambda(t)= \lambda^{-1/2}X(\lambda t),\ \ M_\lambda(t)=\lambda^{-\alpha/2}M(\lambda t),\ \ Z_\lambda(t)=\lambda^{-1/2}Z(\lambda t),~~ t \geq 0.$$
By Lemma 2.1 in \cite{williams1987local}, we have that $Z_\lambda(\cdot)\buildrel \rm d\over = Z(\cdot)$ under $P^0$. Consider now the following identities given in the proof of Theorem 2.6 of \cite{williams1987local}:
$$\phi( Z(\cdot))\buildrel d\over = \phi( Z_\lambda(\cdot))=\phi(\lambda^{-1/2} Z(\lambda\cdot))=
\lambda^{-{\alpha\over 2}}\phi( Z(\lambda\cdot))= $$
$$\lambda^{-{\alpha\over 2}}\left( M(\lambda\cdot)+ A(\lambda\cdot)\right) = M_\lambda(\cdot)+ A_\lambda(\cdot),$$
where the identity in law is understood to be under $P^0$.
By the uniqueness of the decomposition in \eqref{decompose}, $ A_\lambda(\cdot)$ must depend on $ Z_\lambda(\cdot)$ in the same way as $A(\cdot)$ depends on $ Z(\cdot)$. Since $A(\cdot)=A(\cdot)\circ Z(\cdot)$, so we have
$P^0$-almost surely
$ A_\lambda(\cdot)= A\circ Z_\lambda(\cdot)$. Furthermore, $\{\gamma^0(n),n\ge 1\}\in\Gamma$ implies $\{\lambda^{-1/2}\gamma^0(n),n\ge 1\}\in\Gamma$, thus by Lemma \ref{sub} there exists a sequence $\{\epsilon^\lambda(n), n\ge 1\}\subset \{\gamma^0(n), n\ge 1\}$ such that
\begin{equation*}F_{\lambda^{-1/2}\epsilon^\lambda}(\omega) =W(\cdot,\omega)=X(\cdot,\omega),\ P^0\textrm{-}\hbox{a.s.}
\end{equation*}
By \eqref{linear}, we have that $P^0$-almost surely
$$X_\lambda(t)=\lambda^{-1/2}X(\lambda t) = \lambda^{-1/2} F_{\epsilon^\lambda}(Z(\cdot))(\lambda t) = F_{\lambda^{-1/2}\epsilon^\lambda}(\lambda^{-1/2}Z(\lambda\cdot))( t) =F_{\lambda^{-1/2}\epsilon^\lambda}(Z_\lambda(\cdot))(t).$$
We summarize all these as
$$( X_\lambda(\cdot), A_\lambda(\cdot))= (F_{\lambda^{-1/2}\epsilon^\lambda}\circ Z_\lambda(\cdot), A(\cdot)\circ Z_\lambda(\cdot)) \buildrel \rm d\over = (F_{\lambda^{-1/2}\epsilon^\lambda}\circ Z(\cdot), A(\cdot)\circ Z(\cdot)) = ( X(\cdot), A(\cdot))$$
under $P^0$, that is,
$$\Bigl(\lambda^{-{1\over 2}}X(\lambda\cdot),\lambda^{-{\alpha\over 2}}A(\lambda\cdot)\Bigr)\buildrel \rm d\over = \Bigl(X(\cdot),A(\cdot)\Bigr).$$
However, the identity \eqref{la} implies
$$\lambda^{-{\alpha\over 2}}L(\lambda t)=c \lambda^{-{\alpha\over 2}}A(\lambda t),$$
and thus we have that
$$\Bigl(\lambda^{-{1\over 2}}X(\lambda\cdot),\lambda^{-{\alpha\over 2}}L(\lambda\cdot)\Bigr)\buildrel d\over = \Bigl(X(\cdot),L(\cdot)\Bigr).$$
We now conclude that
$$\lambda^{-{1\over 2}}X\Bigl(L^{-1}\left(\lambda^{\alpha\over 2}\cdot\right)\Bigr)\buildrel d\over = X\Bigl(L^{-1}(\cdot)\Bigr),$$
which implies \eqref{stable}.
\end{proof}
We now present the proof of Theorem \ref{Theorem:Converse:Dirichlet:Process:Result}.
\begin{proof}[Proof of Theorem \ref{Theorem:Converse:Dirichlet:Process:Result} ]Following the discussion at the outset of the section, it only remains to prove the identity \eqref{Display:Condition:Converse:Main:Dirichlet}.
Let $T \geq 0$. Since $Y=-X$ on $(L^{-1}(\mathbb{R}_{+}))_T^{-} \subset \Lambda_T^{-} \subset [0,T]$, it follows by Definition \ref{Definition:Strong:P:Variation} that
\begin{eqnarray*}
V_p(Y,[0,T]) &\geq&V_p(Y,(L^{-1}(\mathbb{R}_{+}))_T^{-} )~=~ V_p(X,(L^{-1}(\mathbb{R}_{+}))_T^{-} ),~p \geq 0.
\end{eqnarray*}
Hence it suffices to show that $P^0(V_p(X,(L^{-1}(\mathbb{R}_{+}))_T^{-}) < + \infty )=0$ for $0 < p \leq \alpha$. For each $A > 0$, let $T_A$ be defined by $L^{-1}(A)=T_A$. Since
$L^{-1}$ is $P^{0}$-a.s. right-continuous with $L^{-1}(0)=0$, it follows that $T_A \Rightarrow 0$ as $A \downarrow 0$. Also recall that $P^{0}(T_A > 0)=1$ for each $A > 0$. Moreover, since $L^{-1}$ is $P^0$-a.s. non-decreasing, it follows by the change-of-variables
for $p$-variation (see (P4) of Chistyakov and Galkin \cite{chistyakov1998maps}) that $V_p(X,(L^{-1}(\mathbb{R}_{+}))_{T_A}^{-})=V_p(X \circ L^{-1}, [0,A]),A > 0$. Thus, in order to complete the proof it suffices to show that for $0 < p \leq \alpha$,
\begin{eqnarray}
P^0( V_p(X \circ L^{-1},[0,A]) < + \infty) &=&0,~A > 0. \label{to:show}
\end{eqnarray}
Let $\nu$ be the Levy measure of $X\circ L^{-1}$ under $P^0$. According to Bretagnolle \cite{bretagnolle1972p} (see also Simon \cite{simon2004small}), \eqref{to:show} holds if and only if
\begin{equation}\int_{|u|<1}|u|^p\nu(du)=\infty.\label{integral}\end{equation}
However, since by Proposition \ref{Lemma:Y:Composed:With:L:Inverse:Is:Levy}, $Y \circ L^{-1} = -X \circ L^{-1} $ is an $\alpha$-stable process on
$(C_S,\mathcal{F},\mathcal{F}_{L^{-1}},P^0)$, it follows
by Theorem 14.3(ii) of \cite{sato1999levy} that $\nu(dB)=\int_C \int_0^\infty 1_B(r\xi)r^{-1-\alpha}dr\lambda(d\xi)$, where $C$ is the unit circle in ${\mathbb R}^2$ and $\lambda$ is a finite measure on $C$. \eqref{integral} now follows immediately. \end{proof}
\section{Proof of Theorem \ref{necessary}}
\label{section:esp}
Before providing the proof of Theorem \ref{necessary}, we first proceed with some propositions, starting with a result that condition \eqref{full:space} is necessary for the existence of a solution to the ESP.
\begin{proposition}\label{first:proposition}
Let $z\in S$. If the ESP $(S,d(\cdot))$ for the Brownian motion $X$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$
has a solution $P^z$-a.s., then \eqref{full:space} holds.
\end{proposition}
\begin{proof}
Let the polar coordinates of $v_i$ be $(\|v_i\|,\beta_i)$ for $i=1,2$. Then $\beta_1=\theta_1+{\pi/2}$ and $\beta_2=-(\pi/2-\xi+\theta_2)$. Thus, since by assumption $1 < \alpha < 2$, it follows that $\xi<\beta_1<\pi$ and $-(\pi-\xi)<\beta_2<0$. In addition, the condition $\alpha>1$ implies that the angle of $v_1$ and $v_2$ on the side which contains $S$ is $\beta_1-\beta_2>\pi$. Let $C=\overline{\rm co}(V \cup \{\alpha v_1, \alpha \geq 0\} \cup \{\alpha v_2, \alpha \geq 0\})$, which is a closed convex cone, and assume that $C\not={\mathbb R}^2$, which implies $C\cap S=\{0\}$.
Let $\tilde Z,\tilde Y$ be a solution of the ESP $(S,d(\cdot))$ for $X$. Let $B$ be an arbitrary open set in $C$ such that $0\notin B$, and let $t>0$ arbitrary. Now we have $P^z(X(t)\in B)>0$, and also $X(t)+\tilde Y(t)=\tilde Z(t)\in S$. However, whenever $X(t)\in B$, then $\tilde Y(t)\in C$ implies $X(t)+ \tilde Y(t)\in C$, which implies $X(t)+\tilde Y(t)=0$. But this is impossible, since a non-zero vector and its negative can not be in a closed convex cone that is neither ${\mathbb R}^2$, nor a single line.
\end{proof}
We now proceed to show that if the condition \eqref{full:space} holds, then $(Z,Y)$ solves the ESP $(S,d(\cdot))$ for $X$. We first provide some prepatory lemmas. For each $i=1,2,$ let
\begin{eqnarray*}
T_{\partial S_i} &=& \inf\{t > 0 : Z(t) \in \partial S_i\setminus\{0\}\}.
\end{eqnarray*}
\begin{lemma}\label{Proposition:Excursions:Hit:Both:Boundaries}
For each $i=1,2,$ $P^0(T_{\partial S_i} > 0)=0$.
\end{lemma}
\begin{proof}
Let $i=1$ or $2$, and
$$T_{\partial S_i}^c=\inf\left\{t\ge 0: {Z(ct)\over\sqrt c}\in\partial S_i\setminus\{0\}\right\},~c > 0.$$
Then, by Lemma 2.1 in \cite{williams1987local},
$$P^0\left(T_{\partial S_i}^c>t\right) = P^0\left(T_{\partial S_i}^1 >t\right),~t \geq 0.$$
On the other hand,
$$T_{\partial S_i}^c=\inf\left\{t\ge 0: {Z(ct)}\in\partial S_i\setminus\{0\}\right\}= \inf\left\{{u\over c}\ge 0: {Z(u)}\in\partial S_i\setminus\{0\}\right\}={1\over c}T_{\partial S_i}^1.$$
Hence,
$$P^0(T_{\partial S_i}^1>ct) = P^0(T_{\partial S_i}^1>t),$$
which implies $P^0(T_{\partial S_i}=0) + P^0(T_{\partial S_i}=\infty)=1$. By the Blumenthal $0$-$1$ law, $P^0(T_{\partial S_i}=0)$ is either 0 or 1. Hence, in order to complete the proof it suffices that $P^0(T_{\partial S_i}=\infty)<1$, and this is what we shall do.
By Lemma 4.2 in \cite{williams1985recurrence}, for Lebesgue-almost every $t\in(0,\infty)$,
\begin{equation} P^0(Z(t)\in\partial S)=0.\label{zero}\end{equation}
However, by Lemma 2.1 in \cite{williams1987local}, for every $c>0$
$$P^0(Z(t)\in\partial S)=P^0\left({Z(ct)\over\sqrt c}\in \partial S\right) = P^0\left({Z(ct)}\in \partial S\right), $$
and so \eqref{zero} holds for every $t>0$. Now let $t>0$ be arbitrary and
$$T_{\partial S_i}(t) = \inf\{s\ge t: Z(s)\in\partial S_i\setminus\{0\}\}.$$
Then, by the Markov property,
\begin{equation} P^0(T_{\partial S_i}(t)<\infty)= E^0 \left[P^0(T_{\partial S_i}(t)<\infty|Z(t))\right]=E^0\left[P^{Z(t)}(T_{\partial S_i}<\infty)\right].\label{markov}\end{equation}
Let $T_{\partial S}= \inf\{t\ge 0: Z(t)\in\partial S\}$. By Proposition \ref{williams}, for each $z\in \hbox{int}(S)$ the law of $Z(\cdot\wedge T_{\partial S})$ under $P^z$ coincides with the law of a Brownian motion started at $z$ and stopped at the first hitting time of $\partial S$. It follows that for all $z\in\hbox{int} (S)$ we have $P^z(T_{\partial S}<\infty)=1$ and $P^z(Z(T_{\partial S})\in\partial S_i\setminus\{0\})>0$,
which implies $P^z(T_{\partial S_i}<\infty)>0$. Then it follows from \eqref{zero} that the last expression in \eqref{markov} is strictly positive, and so also is the first expression there, i.e. $P^0(T_{\partial S_i}(t)<\infty)>0$. Thus, $P^0(T_{\partial S_i}<\infty)>0$.
\end{proof}
\begin{lemma}\label{endpoints:are:nice}
For each $z\in S$ and $P^z$-a.e. $\omega\in C_s$, the following holds: for the endpoint $D$ of each excursion interval of $Z(\omega)$ away from zero and each $\epsilon>0$, there exist $u_1,u_2,u\in(D,D+\epsilon)$ such that
\begin{equation} Z(u_1)\in\partial S_1\setminus\{0\}, \ Z(u_2)\in\partial S_2\setminus\{0\},\ \hbox{and}\ Z(u)=0.\label{all:the:v}\end{equation}
\end{lemma}
\begin{proof} Recall the sequence of stopping times $\{T_k^n,k\ge 1\}$ from the proof of Proposition \ref{main}, that is the sequence of endpoints of excursions away from zero with length larger than $1/n$. By the Strong Markov Property, Proposition \ref{Proposition:Excursions:Hit:Both:Boundaries}, and the regularity of the vertex (see Lemma 2.2 of \cite{williams1987local}), for every $k,n\ge 1$, we have that $P^z$-almost surely there exist $u_1,u_2,u\in(T_k^n,T_k^n+\epsilon)$ such that \eqref{all:the:v} holds. Since the set of excursions is countable, and $\cup_{n\ge 1}\{T_k^n,k\ge 1\}$ is exactly the set of endpoints of all excursion intervals, the statement follows.
\end{proof}
The following is an immediate corollary of Lemma \ref{endpoints:are:nice}.
\begin{cor}\label{no:isolated} For each $z\in S$ and $P^z$-a.e. $\omega\in C_s$, the set $\Lambda$ has no isolated point.
\end{cor}
We now present the proof that under \eqref{full:space}, $(Z,Y)$ solves the ESP $(S,d(\cdot))$ for $X$.
\begin{proposition}\label{second:proposition}
If \eqref{full:space} holds, then for each $z \in S$, $(Z,Y)$ $P^z$-a.s. solves the ESP $(S,d(\cdot))$ for the Brownian motion $X$ on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$.
\end{proposition}
\begin{proof} Let $z \in S$. We will verify that $P^z$-a.s. $(Z,Y)$ satisfies Conditions \ref{Definition:ESP:Item:One} through \ref{Definition:ESP:Item:Four} of Definition \ref{Definition:Extended:Skorohod:Problem} with $\psi=X$ and $(\phi,\eta)=(Z,Y)$. Condition \ref{Definition:ESP:Item:One} is clear
since by the Doob-Meyer type decomposition of Theorem \ref{Theorem:Main:Dirichlet:Process:Result}, we have that $Z=X+Y$. Next, Condition \ref{Definition:ESP:Item:Two} is immediate as well since $Z \in C_S$. Now recall that $X$ is a Brownian motion started at $z$
on $(C_S,\mathcal{F},\mathcal{F}_t,P^z)$ and so $X$ is $P^z$-a.s continuous. Hence $Y=Z-X$ is $P^z$-a.s. continuous as well and so Condition \ref{Definition:ESP:Item:Four} holds since $ 0 \in
\overline{\mathrm{co}}[d(\phi(t))]$ for $t \geq 0$.
In order to complete the proof, it suffices to show that $P^z$-a.s. Condition \ref{Definition:ESP:Item:Three} holds. That is, it suffices to show that $P^z$-a.s. for each $t \geq 0$ and $s \in [0,t]$,
\begin{eqnarray}
Y(t) - Y(s) &\in& \overline{\mathrm{co}}[\cup_{u \in (s,t]}d(Z(u))]. \label{Display:Condtion:Three:For:Z}
\end{eqnarray}
Let $0\le s <t$ arbitrary. We distinguish between two cases. In the first case, the interval $[s,t]$ lies entirely within a single excursion of $Z$ away from zero, i.e. $(s,t)\cap \Lambda=\emptyset.$ In this case, by Part \ref{Proposition:Skorohod:Problem:Part:2} of Theorem \ref{Proposition:Skorokhod:Problem}, $Y(v)=RU(v)$ where $U$ is non-decreasing on $[s,t]$, and $U_i$ is increasing on $[s,t]$ only at times $v\in[s,t]$ when $Z(v)\in \partial S_i$ for $i=i,2$, and then \eqref{Display:Condtion:Three:For:Z} follows.
Next, we show that \eqref{Display:Condtion:Three:For:Z} holds in the second case, where $(s,t)\cap\Lambda\not=\emptyset$. In this case, let $v\in (s,t)\cap\Lambda$. By Corollary \ref{no:isolated}, there exists another point $u$ besides $v$ in $(s,t)\cap\Lambda$, and, by Condition
\ref{Condition:3:Submartingale} of Definition \ref{vw}, there exists an excursion interval included in $[u,v]$ or in $[v,u]$, depending on whether $u<v$ or $v<u$. In either case, by Lemma \ref{endpoints:are:nice} this implies the existence of $u_1,u_2 \in (s,t)$ such that $Z(u_i)\in\partial S_i\setminus\{0\}$, $i=1,2$. Then, by \eqref{full:space},
\begin{eqnarray*}
\overline{\mathrm{co}}(\cup_{u\in(s,t]}d(Z(u)))~=~
\overline{\mathrm{co}}\left(V\cup\{a v_1,a\ge 0\}\cup\{a v_2,a\ge 0\}\right)~=~\mathbb{R}^2,
\end{eqnarray*}
and thus \eqref{Display:Condtion:Three:For:Z} is trivially satisfied.
\end{proof}
The proof of Theorem \ref{necessary} is now immediate.
\begin{proof} [Proof of Theorem \ref{necessary}]
The theorem follows by the combination of Propositions \ref{first:proposition} and \ref{second:proposition}.
\end{proof}
\bibliographystyle{plain}
|
\section{#1} \setcounter{equation}{0}}
\renewcommand{\dagger}{\#}
\newcommand{\proof}{{\sc Proof.} \quad}
\newcommand{\proofc}{{\sc Proof} \ }
\newcommand{\be}{\begin{equation} \label}
\newcommand{\ee}{\end{equation}}
\newcommand{\bea}{\begin{eqnarray}\label}
\newcommand{\eea}{\end{eqnarray}}
\newcommand{\bas}{\begin{eqnarray*}}
\newcommand{\eas}{\end{eqnarray*}}
\newcommand{\bit}{\begin{itemize}}
\newcommand{\eit}{\end{itemize}}
\newcommand{\qed}{\hfill$\Box$ \vskip.2cm}
\newcommand{\nn}{\nonumber}
\newcommand{\R}{\mathbb{R}}
\newcommand{\N}{\mathbb{N}}
\newcommand{\pO}{\partial\Omega}
\newcommand{\dN}{\partial_\nu}
\newcommand{\eps}{\varepsilon}
\newcommand{\dist}{{\rm dist} \, }
\newcommand{\supp}{{\rm supp} \, }
\newcommand{\subsubset}{\subset\subset}
\newcommand{\wto}{\rightharpoonup}
\newcommand{\wsto}{\stackrel{\star}{\rightharpoonup}}
\newcommand{\hra}{\hookrightarrow}
\newcommand{\ab}{\\[3mm]}
\newcommand{\abs}{\\[5pt]}
\newcommand{\Abs}{\\[5mm]}
\newcommand{\Aabs}{\\[7mm]}
\newcommand{\cb}{\color{blue}}
\newcommand{\cred}{\color{red}}
\newcommand{\io}{\int_\Omega}
\newcommand{\tm}{T_{max}}
\newcommand{\uu}{\underline{u}}
\newcommand{\ou}{\overline{u}}
\newcommand{\uw}{\underline{w}}
\newcommand{\ow}{\overline{w}}
\newcommand{\ue}{u_\eps}
\newcommand{\ve}{v_\eps}
\newcommand{\ophi}{\overline{\varphi}}
\newcommand{\uphi}{\underline{\varphi}}
\newcommand{\parab}{{\cal P}}
\newcommand{\qarab}{{\cal Q}}
\def\bw{{\bf w}}
\def\bv{{\bf v}}
\def\bx{{\bf x}}
\def\f{{\bf f}}
\def\bg{{\bf g}}
\def\vb{\, ,}
\def\pa{\cdotp}
\def\pb{\, .}
\def\va{\raise 2pt\hbox{,}}
\def\div{\mbox{{\rm div}}}
\def\Div{\mbox{{\rm Div}}}
\begin{document}
\title{A degenerate chemotaxis system with flux limitation:\\
Maximally extended solutions and absence of gradient blow-up}
\author{
Nicola Bellomo\footnote{<EMAIL>}\\
{\small Department of Mathematics, Faculty Sciences,}\\
{\small King Abdulaziz University, Jeddah, Saudi Arabia}\\
{\small Politecnico di Torino, 10129 Torino, Italy}\\
\and
Michael Winkler\footnote{<EMAIL>}\\
{\small Institut f\"ur Mathematik, Universit\"at Paderborn,}\\
{\small 33098 Paderborn, Germany} }
\date{}
\maketitle
\begin{abstract}
\noindent
This paper aims at providing a first step toward a qualitative theory for
a new class of chemotaxis models derived from the celebrated
Keller-Segel system, with the main novelty being that diffusion is nonlinear with flux delimiter features.
More precisely, as a prototypical representative of this class we study radially symmetric solutions
of the parabolic-elliptic system
\bas
\left\{ \begin{array}{l}
\displaystyle u_t=\nabla \cdot \Big(\frac{u\nabla u}{\sqrt{u^2+|\nabla u|^2}}\Big) - \chi \, \nabla \cdot \Big(\frac{u\nabla v}{\sqrt{1+|\nabla v|^2}}\Big), \\[1mm]
0=\Delta v - \mu + u,
\end{array} \right.
\eas
under the initial condition $u|_{t=0}=u_0>0$ and no-flux boundary conditions
in balls $\Omega\subset\R^n$, where $\chi>0$ and $\mu:=\frac{1}{|\Omega|} \io u_0$.\abs
The main results assert the existence of a unique classical solution, extensible in time up to
a maximal $\tm \in (0,\infty]$ which has the property that
\bas
\mbox{if} \quad \tm<\infty \quad \mbox{then} \quad
\limsup_{t\nearrow\tm} \|u(\cdot,t)\|_{L^\infty(\Omega)}=\infty.
\qquad \qquad (\star)
\eas
The proof therefor is mainly based on comparison methods,
which firstly relate pointwise lower and upper bounds for the spatial gradient $u_r$
to $L^\infty$ bounds for $u$ and to {\em upper bounds} for $z:=\frac{u_t}{u}$;
secondly, another comparison argument involving nonlocal nonlinearities provides an appropriate control of
$z_+$ in terms of bounds for $u$ and $|u_r|$,
with suitably mild dependence on the latter.\abs
As a consequence of ($\star$),
by means of suitable a priori estimates it is moreover shown that the above solutions are global and bounded
when either
\bas
n\ge 2 \ \mbox{ and } \chi<1,
\qquad \mbox{or} \qquad
n=1, \ \chi>0 \ \mbox{ and } m<m_c,
\eas
with $m_c:=\frac{1}{\sqrt{\chi^2-1}}$ if $\chi>1$ and $m_c:=\infty$ if $\chi\le 1$.\\
That these conditions are essentially optimal will be shown in a forthcoming paper in which ($\star$)
will be used to derive complementary results on the occurrence of solutions blowing up in finite time
with respect to the norm of $u$ in $L^\infty(\Omega)$.\abs
\noindent
{\bf Key words:} chemotaxis; flux limitation; degenerate diffusion\\
\noindent {\bf AMS Classification:} 35K65 (primary); 35B45, 35Q92, 92C17 (secondary)
\end{abstract}
\mysection{Introduction}
{\bf Keller-Segel systems with flux limitation.} \quad
The celebrated model by Keller and Segel \cite{Keller-Segel-70,Keller-Segel-71} was heuristically derived to model growth phenomena mediated by a chemoattractant, specifically the aggregation of dictyostelium discoideum due to an attractive chemical substance. The general structure of the model is as follows:
\begin{equation}
\label{KSO}
\left\{
\begin{array}{l}
u_t =
\nabla \, \big(D_u(u,v) \nabla u - S(u,v) u \nabla v\big)+H_1(u,v), {} \\ [15pt]
v_t = D_v \Delta v + H_2(u,v),
\end{array} \right.
\end{equation}
where $u=u(x,t)$ denotes the cell (or organism) density at position $x$ and time $t$, and $v=v(x,t)$ is the
density of the chemoattractant. Here the function $S$ measures the chemotactic sensitivity,
the positive functions $D_u$ and $D_v$ represent the diffusivity of the cells and of the
chemoattractant, respectively, and $H_1$ and $H_2$ model source terms related to interactions. In a more general framework in which diffusions are not isotropic, $D_u$ and $S$ can be positive definite matrices. See the survey by Hillen and Painter \cite{Hillen-Painter-09} for a review of modeling issues based on the classical approach of continuum mechanics closed by empirical models for the closure of conservation equations. The essay by Horstmann \cite{Horstmann-03} provides an additional source of information concerning modeling and applications in biology.
The recent survey \cite{BBYW} gives a review and qualitative analysis of a variety of mathematical problems and multiscale derivations of the
original model as well as of some recent developments such as the specific one treated in this paper. \abs
On the other hand, a natural question can be posed, namely if the use of parabolic models is consistent with the physics of the phenomena under consideration, or if, for instance, the use of hyperbolic models can be more appropriate. Or even within the approach by parabolic equations, if linear models are acceptable, while in the nonlinear case whether one should consider degenerate parabolic equations characterized by a finite propagation velocity. Intuitively, the answer is that phenomena with finite propagation velocity should be captured by an appropriate choice of nonlinear diffusion terms \cite{Vazquez-06}. A conceivable approach leads to consider functions $D_u(u,v)$ and $D_v(u,v)$ not only depending on $u$ and $v$, but also on their derivatives in space and time.
A recent study in this direction \cite{BBNS-10} has shown that macroscopic models can be obtained from the underlying description at the scale of cells delivered by suitable developments of kinetic theory methods. More in details, appropriate models of cell-cell interaction lead to macroscopic expressions for diffusion and cross-diffusion
with nonlinear limited flux terms of the type
\bas
\nabla \cdot \bigg( D_u(u,v) \frac{ u\nabla u}{\sqrt{u^2 + \frac{\nu^2}{c^2}|\nabla u|^2}}\bigg)
\qquad \hbox{and} \qquad
\nabla \cdot \bigg(S(u,v) \frac{u\nabla v}{\sqrt{1 + |\nabla v|^2}} \bigg)\va
\eas
respectively,
with $\nu$ denoting the kinematic viscosity and $c$ the maximum speed of propagation, so that
in combination with an adequate equation for the evolution of the chemoattractant, this appraoch suggests to consider
models of type
\begin{equation}\label{Flim-KS}
\left\{
\begin{array}{l}
\displaystyle u_t = \nabla \cdot \bigg(D_u (u,v)\frac{u \nabla u}{\sqrt{u^2 + \frac{\nu^2}{c^2}|\nabla u|^2}} - S(u,v) \frac{u \nabla v}{\sqrt{1+|\nabla {v}|^2}}\bigg)
+H_1(u,v)\va \\[5mm]
\displaystyle v_t = D_v \Delta v + H_2(u,v),
\end{array} \right.
\end{equation}
as consistent modifications of the classical Keller-Segel system.
This idea, which is somehow related to the optimal transport framework \cite{Brenier-09}, can be motivated by a natural assumption of cell dynamics, where overcrowding is naturally avoided \cite{Burger-06}.
Furthermore, the introduction of this type of terms is founded in the assumption that
particles do not diffuse arbitrarily in space but, on the contrary, move through some privileged ways such as the border of cells.
In this new approach the non-physical diffusion is eliminated and the population moves with a finite speed of propagation, which is one of the intrinsic characteristics.
Indeed, the qualitative analysis of related systems with limited flux \cite{ACM-05,ACM-06} as well as some extensions to biological contexts (transport of morphogens) has been recently explored \cite{ACMS-12},
inter alia confirming the expected movement of fronts at finite speeds.\Abs
{\bf Boundedness vs.~blow-up.} \quad
In the framework of chemotaxis systems, however, a different qualitative aspect seems even more important, namely
the ability of the respective system to spontaneously generate structures.
In this regard, the classical Keller-Segel system, as obtained from (\ref{KSO}) on letting
$D_u\equiv D_v\equiv S\equiv 1$, $H\equiv 0$ and $K(u,v)=u-v$, is known to have the property that
some solutions reflect such aggregation processes even in the extreme mathematical sense of finite-time
blow-up of some solutions when either the spatial dimension $n$ satisfies $n\ge 3$ \cite{win_JMPA},
or when $n=2$ and the total mass of cells is suitably large \cite{herrero_velazquez, mizoguchi_win};
on the other hand, if either $n\ge 3$ and the initial data fulfill appropriate smallness conditions,
or $n=2$ and $\io u(\cdot,0)$ is small, or if $n=1$, then for various types of initial-boundary value problems,
global bounded solutions are known to exist
\cite{corrias_perthame, nagai_senba_yoshida, win_JDE2010, osaki_yagi, cao_optspace}.\abs
As for Keller-Segel-type models with flux limitations, the corresponding problem appears to be unsolved,
and it is the goal of the present paper to present a first step into a qualitative theory for such systems,
with a particular focus on the question whether solutions exist globally, as conjectured in \cite{BBNS-12},
or whether blow-up in finite time may occur for some initial data.
Specifically, we will consider the apparently most prototypical among the systems (\ref{Flim-KS}) in its
parabolic-elliptic simplification, as suggested in \cite{jaeger_luckhaus};
more precisely, we shall be concerned with the initial-boundary value problem
\be{0}
\left\{ \begin{array}{l}
\displaystyle u_t=\nabla \cdot \bigg(\frac{u\nabla u}{\sqrt{u^2+|\nabla u|^2}}\Big) - \chi \, \nabla \cdot \Big(\frac{u\nabla v}{\sqrt{1+|\nabla v|^2}}\bigg)\va
\qquad x\in \Omega, \ t>0, \\[5mm]
0=\Delta v - \mu + u, \qquad x\in \Omega, \ t>0, \\[5mm]
\displaystyle \bigg(\frac{u\nabla u}{\sqrt{u^2+|\nabla u|^2}} - \chi \, \frac{u\nabla v}{\sqrt{1+|\nabla v|^2}} \bigg) \cdot \nu=0,
\qquad x\in \partial\Omega, \ t>0, \\[5mm]
u(x,0)=u_0(x), \qquad x\in\Omega,
\end{array} \right.
\ee
in a ball $\Omega=B_R(0)\subset \mathbb{R}^n$, $n\ge 1$, where
$\chi>0$ indicates the strength of chemotactic cross-diffusion.
In order to further simplify the analysis, we shall assume the initial data to satisfy
\be{init}
u_0\in C^3(\bar\Omega)
\quad \mbox{is radially symmetric and positive in $\bar\Omega$ with $\frac{\partial u_0}{\partial\nu}=0$ on } \pO,
\ee
so that the spatial average
\be{mu}
\mu:=\frac{1}{|\Omega|} \io u_0(x)dx
\ee
is positive.\Abs
{\bf Main results.} \quad
In this framework, the first of our main results asserts local existence of a uniquely determined
classical solution.
In its most crucial part, however, the following theorem furthermore provides the extensibility criterion
(\ref{43.1}) which will be of great importance
both for deriving global existence in Theorem \ref{theo55} below,
as well as for characterizing the asymptotic behavior of non-global solutions
near their blow-up time \cite{bellomo_winkler2}.
\begin{theo}\label{theo43}
Suppose that $u_0$ complies with (\ref{init}). Then there exist $\tm \in (0,\infty]$ and a uniquely determined
pair $(u,v)$ of positive radially symmetric functions
$u\in C^{2,1}(\bar\Omega\times [0,\tm))$ and $v\in C^{2,0}(\bar\Omega\times [0,\tm))$
which solve (\ref{0}) classically in $\Omega\times (0,\tm)$, and which are such that
\be{43.1}
\mbox{if} \quad \tm<\infty \quad \mbox{then} \quad
\limsup_{t\nearrow\tm} \|u(\cdot,t)\|_{L^\infty(\Omega)}=\infty.
\ee
\end{theo}
In particular, (\ref{43.1}) rules out
the occurrence of any {\em gradient blow-up} phenomenon in the present framework;
it is thus impossible that $\nabla u$ becomes unbounded in finite time, whereas $u$ itself remains bounded.
In view of the complex evolution mechanism in (\ref{0}), inter alia involving doubly degenerate diffusion,
this conclusion seems far from trivial;
indeed, various types of gradient-dependent nonlinearities and degeneracies are known to enforce unboundedness
of gradients for some solutions even in scalar reaction-diffusion equations
\cite{Angenent-96,Li-Souplet-10,Stinner-Winkler-08}.
Moreover, the additionally present cross-diffusive interaction apparently rules out the accessibility of (\ref{0})
to most of the techniques well-established in contexts of scalar parabolic equations with diffusion degeneracies
of related type, such as e.g.~the mean curvature flow equation and derivatives thereof, among others \cite{Evans-Spruck-91,Evans-Spruck-92}, or \cite{Bertsch-DalPasso-92}. \abs
A natural next goal consists in identifying circumstances under which the above solutions are global.
Going in this direction,
the second of our main results provides conditions on the parameter $\chi$ in (\ref{0}) and, when $n=1$,
on the mass level $m$, which turn out to be sufficient not only for global extensibility,
but also for uniform boundedness of all solutions emanating from initial data $u_0$ with $\io u_0=m$.
\begin{theo}\label{theo55}
Assume that $u_0$ satisfies (\ref{init}), and that either
\be{55.111}
n\ge 2
\qquad \mbox{and} \qquad
\chi<1,
\ee
or
\be{55.112}
n=1, \quad \chi>0
\qquad \mbox{and} \qquad
\io u_0<m_c,
\ee
where in the case $n=1$ we have set
\be{mc}
m_c:=\left\{ \begin{array}{ll}
\frac{1}{\sqrt{\chi^2-1}} \qquad & \mbox{if } \chi>1, \\[3mm]
+\infty & \mbox{if } \chi \le 1.
\end{array} \right.
\ee
Then the problem (\ref{0}) possesses a unique global classical solution $(u,v)\in C^{2,1}(\bar\Omega\times [0,\infty)) \times C^{2,0}(\bar\Omega\times [0,\infty))$ which is radially symmetric and such that for some $C>0$ we have
\be{55.1}
\|u(\cdot,t)\|_{L^\infty(\Omega)} \le C
\quad \mbox{and} \quad
\|v(\cdot,t)\|_{L^\infty(\Omega)} \le C
\quad \mbox{for all } t>0.
\ee
\end{theo}
As a first and
immediate conclusion thereof, we underline that when $\chi<1$, in stark contrast to the original
Keller-Segel model, the system (\ref{0}) does
not exhibit any critical mass phenomenon, nor any phenomenon of critical sizes of initial data with
respect to global existence of solutions.
Let us secondly mention that the conditions (\ref{55.111}) and (\ref{55.112}), as identified above,
are in fact essentially optimal for the obtained conclusion:
Indeed, in \cite{bellomo_winkler2} the picture will in this respect be basically completed
by showing that
if $\chi>1$ then in both cases $n=1$ with $m>m_c$, and $n\ge 2$,
some initial data can be constructed such that the corresponding solutions will blow up in finite time.
Together with the latter, our results thus indicate that in comparison to the original Keller-Segel system,
the occurrence of a critical mass phenomenon is shifted from the two-dimensional to the one-dimensional
setting, whereas in the case $n\ge 2$ we rather encounter a {\em critical sensitivity phenomenon}
in that the size of $\chi$ becomes the crucial quantity to determine whether or not blow-up may happen.\Abs
{\bf Main ideas. Excluding gradient blow-up.} \quad
In view of the doubly degenerate structure of the diffusion operator
$\nabla \cdot \Big(\frac{u\nabla u}{\sqrt{u^2+|\nabla u|^2}}\Big)$ in (\ref{0}),
standard theory yields local existence and extensibility as long as $u$ remains uniformly positive
and both $u$ and $\nabla u$ remain bounded (Lemma \ref{lem21}), where thanks to our positivity assumption
on $u_0$, a corresponding lower bound for $u$ can readily be obtained (Lemma \ref{lem23}).\abs
The crucial part in the derivation of Theorem \ref{theo43} will thus consist in ruling out
the possibility of gradient blow-up,
and in our approach toward this we will substantially make use of the radial symmetry of our solutions:
Based on
two different interpretations of the equation satisfied by $u_r$ as linear inhomogeneous parabolic
equations (Lemma 2.3), under the standing assumption that $u$ is non-global but remains bounded we will first obtain
a uniform lower bound for $u_r$ by a comparison argument (Lemma \ref{lem35}), and thereafter develop this
into a bound for $|u_r|$ in Section \ref{sect5}.\abs
The latter step itself will involve the quantity $z:=\frac{u_t}{u}$, as known to be of great importance
on various types of different nonlinear diffusion equations \cite{aronson, Vazquez-06}.
In the present context, we shall see that
$u_r$ can indeed be controlled in terms of the {\em positive part} $z_+$ of $z$ through an inequality
of the form
\be{est}
\|u_r(\cdot,t)\|_{L^\infty((0,R))} \le C \cdot \Big(1+\|z_+\|_{L^\infty((0,R)\times (0,t))} \Big)
\qquad \mbox{for all } t\in (0,\tm),
\ee
where $\tm\in (0,\infty)$ denotes the maximal existence time (Corollary \ref{cor40}).
This will be achieved by splitting the interval $(0,R)$ in two parts and
first performing a testing procedure to estimate $u_r$ in the corresponding inner region in certain weighted
Lebesgue spaces and taking limits appropriately (Lemma \ref{lem32}),
whereupon a comparison argument in the associated outer region will complete the proof of (\ref{est})
(Lemma \ref{lem37}).\abs
In order to complete the proof of Theorem \ref{theo43} by providing a suitable estimate for $z_+$,
we shall make use of the observation that $z$ satisfies the {\em one-sided} nonlocal parabolic inequality
\bas
z_t(r,t) \le {\cal L} z
+ d \cdot \Big( 1+\|z_+\|_{L^\infty((0,R)\times (0,t))} \Big)
\eas
with some $d>0$ and some homogeneous linear elliptic operator ${\cal L}$ (Lemma \ref{lem39}).
In fact, employing a maximum principle-type argument will show that this implies a pointwise upper bound
for $z$ (Lemma \ref{lem41}), which in conjunction with (\ref{est}) will prove Theorem \ref{theo43}.\abs
Thanks to the mild extensibility criterion (\ref{43.1}) thus gained, the proof of Theorem \ref{theo55}
thus actually reduces to the derivation of suitable a priori bounds for solutions with respect to
the norm of $u$ in $L^\infty(\Omega)$. This will be accomplished in the respective cases detailed in
Theorem \ref{theo55} by means of an essentially straightforward adaptation of the Moser-Alikakos iteration
technique to the present setting in Section \ref{sect6}.
\mysection{Preliminaries}
\subsection{Local existence and a first extensibility criterion}
To begin with, let us suitably reduce (\ref{0}), locally in time, so as to become accessible to standard
existence theory.
We thereby obtain the following result on local existence of a smooth solution to (\ref{0}), extensible
as long as such a reduction is possible.
As a by-product, this procedure yields the first basic extensibility criterion (\ref{ext_crit}) the improvement
of which will be the main objective of the subsequent Sections \ref{sect3}-\ref{sect5}.
\begin{lem}\label{lem21}
Suppose that $u_0$ satisfies (\ref{init}). Then there exist $\tm \in (0,\infty]$ and
a uniquely determined
pair $(u.v)$ of radially symmetric positive functions
\bea{21.0}
u \in C^{2,1}(\bar\Omega\times [0,\tm)), \qquad v\in C^{2,0}(\bar\Omega\times [0,\tm)),
\eea
which solve (\ref{0}) classically in $\Omega\times (0,\tm)$, and which are such that
\be{ext_crit}
\mbox{if $\tm<\infty$ \quad then either \quad }
\liminf_{t\nearrow\tm} \inf_{x\in\Omega} u(x,t)=0 \mbox{ \quad or \quad }
\limsup_{t\nearrow \tm} \|u(\cdot,t)\|_{W^{1,\infty}(\Omega)}=\infty.
\ee
\end{lem}
\proof
We let
\be{21.1}
\eps:=\min \bigg\{ \frac{1}{2} \inf_{x\in\Omega} u_0(x) \, , \, \frac{1}{2\|u_0\|_{L^\infty(\Omega)}} \, , \,
\frac{1}{2\|\nabla u_0\|_{L^\infty(\Omega)}} \bigg\}
\ee
and take cut-off functions $\psi_\eps \in C^\infty(\R)$ and $\phi_\eps\in C^\infty(\R)$ satisfying
\bas
\frac{\eps}{2} \le \psi_\eps(s) \le \frac{2}{\eps}
\quad \mbox{for all } s\in\R
\qquad \mbox{and} \qquad
\psi_\eps(s)=s
\quad \mbox{for all } s\in \Big( \eps \, , \, \frac{1}{\eps} \Big)\va
\eas
as well as
\bas
\phi_\eps(s)\le \frac{2}{\eps}
\quad \mbox{for all } s\in\R
\qquad \mbox{and} \qquad
\phi_\eps(s)=s
\quad \mbox{for all } s\le \frac{1}{\eps}\pa
\eas
Then
\bas
a_\eps(s,p):=\frac{\psi_\eps(s)}{\sqrt{\psi_\eps^2(s)+\phi_\eps^2(|p|)}},
\qquad s\in\R, \ p\in\R^n,
\eas
defines a function $a_\eps \in C^\infty(\R\times \R^n)$ fulfilling
\bas
a_\eps(s,p) \le \frac{\psi_\eps(s)}{\sqrt{\psi_\eps^2(s)}}=1
\quad \mbox{for all $s\in\R$ and } p\in\R^n
\eas
and
\bas
a_\eps(s,p) \ge \frac{\frac{\eps}{2}}{\sqrt{(\frac{2}{\eps})^2 + (\frac{2}{\eps})^2}} = \frac{\eps^2}{4\sqrt{2}}
\quad \mbox{for all $s\in\R$ and } p\in\R^n.
\eas
We can therefore adapt a fixed point argument which is well-established in the existence theory of parabolic-elliptic
chemotaxis systems (cf.~\cite{Djie-Winkler-10} or \cite{Fujie-Winkler-Yokota}, for instance) to find $T_\eps>0$ such that the problem
\bas
\left\{ \begin{array}{ll}
\displaystyle u_t=\nabla \cdot \Big( \frac{\psi_\eps(u)\nabla u}{\sqrt{\psi_\eps^2(u) + \phi_\eps^2(|\nabla u|)}}\Big)
- \chi \nabla \cdot \Big( \frac{u\nabla v}{\sqrt{1+|\nabla v|^2}}\Big)\va \qquad x\in\Omega, \ t\in (0,T_\eps), \\[5mm]
0=\Delta v - \mu + u, \qquad x\in\Omega, \ t\in (0,T_\eps), \\[5mm]
\displaystyle \frac{\partial u}{\partial\nu}=\frac{\partial v}{\partial\nu}=0, \qquad x\in\partial\Omega, \ t\in (0,T_\eps), \\[5mm]
u(x,0)=u_0(x),
\qquad x\in\Omega,
\end{array} \right.
\eas
possesses a unique classical solution $(u_\eps,v_\eps)$ such that $u_\eps \in C^{2,1}(\bar\Omega \times [0,T_\eps))$
and $v_\eps \in C^{2,0}(\bar\Omega\times [0,T_\eps))$, and such that both $u_\eps$ and $v_\eps$ are radially symmetric and
positive. Furthermore, since
$2\eps \le u_0 \le \frac{1}{2\eps}$ and $|\nabla u_0| \le \frac{1}{2\eps}$ in $\Omega$ according to our choice of $\eps$,
by continuity of $u_\eps$ and $\nabla u_\eps$ in $\bar\Omega \times [0,T_\eps)$ we can find $\tilde T_\eps \in (0,T_\eps)$
such that
\bas
\eps \le u_\eps \le \frac{1}{\eps}
\quad \mbox{and} \quad
|\nabla u_\eps| \le \frac{1}{\eps}
\qquad \mbox{in } \Omega \times (0,\tilde T_\eps).
\eas
In particular, this implies that $\psi_\eps(u_\eps)=u_\eps$ and $\phi_\eps(|\nabla u_\eps|)=|\nabla u_\eps|$ in
$\Omega\times (0,\tilde T_\eps)$, and that thus $a_\eps(u_\eps)=\frac{u_\eps}{\sqrt{u_\eps^2 + |\nabla u_\eps|^2}}$
in this region, meaning that $(u_\eps,v_\eps)$ actually solves the original problem (\ref{0}) in
$\Omega\times (0,\tilde T_\eps)$.
Finally, in view of the dependence of $\eps$ on $u_0$ as expressed in (\ref{21.1}),
a standard extensibility argument yields that
the above solution can be continued so as to exist up to some maximal time $\tm \le \infty$ in such a way that
(\ref{ext_crit}) is valid.
\qed
\subsection{Radial solutions}
Since all our solutions are radially symmetric, whenever this appears convenient me may without any danger of confusion
utilize the notation $u(r,t)$ and $v(r,t)$
instead of $u(x,t)$ and $v(x,t)$, respectively, where $r=|x|\in (0,R)$.\abs
In this particular radial setting, $u$ actually fulfills a favorable parabolic equation specified in the following lemma.
\begin{lem}\label{lem36}
Assume (\ref{init}). Then the solution of (\ref{0}) satisfies
\bea{36.1}
u_t &=& \frac{u^3 u_{rr}}{\sqrt{u^2+u_r^2}^3} + \frac{u_r^4}{\sqrt{u^2+u_r^2}^3}
+ \frac{n-1}{r} \cdot \frac{uu_r}{\sqrt{u^2+u_r^2}} \nn\\
& & - \chi \frac{u_r v_r}{\sqrt{1+v_r^2}} - \chi \frac{u(\mu-u)}{\sqrt{1+v_r^2}^3}
- \chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^3}{\sqrt{1+v_r^2}^3},
\eea
for all $r\in (0,R)$ and $t\in (0,\tm)$.
\end{lem}
\proof
We differentiate on the right-hand side of the first equation in (\ref{0}) to obtain
\bea{36.2}
u_t &=& \frac{1}{r^{n-1}} \cdot \bigg( r^{n-1} \frac{uu_r}{\sqrt{u^2+u_r^2}}\bigg)_r
- \frac{\chi}{r^{n-1}} \cdot \bigg( r^{n-1} \frac{uv_r}{\sqrt{1+v_r^2}}\bigg)_r \nn\\
&=& \frac{uu_{rr}}{\sqrt{u^2+u_r^2}} + \frac{u_r^2}{\sqrt{u^2+u_r^2}}
-\frac{1}{2} \cdot \frac{uu_r(2uu_r+2u_r u_{rr})}{\sqrt{u^2+u_r^2}^3}
+ \frac{n-1}{r} \cdot \frac{uu_r}{\sqrt{u^2+u_r^2}} \nn\\
& & - \chi \frac{uv_{rr}}{\sqrt{1+v_r^2}} - \chi \frac{u_r v_r}{\sqrt{1+v_r^2}}
+ \frac{1}{2} \cdot \chi \frac{uv_r \cdot 2v_r v_{rr}}{\sqrt{1+v_r^2}^3}
-\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r}{\sqrt{1+v_r^2}}
\eea
for all $r\in (0,R)$ and $t\in (0,\tm)$.
Here we can rearrange
\bas
& & \hspace*{-30mm}
\frac{uu_{rr}}{\sqrt{u^2+u_r^2}} + \frac{u_r^2}{\sqrt{u^2+u_r^2}}
-\frac{1}{2} \cdot \frac{uu_r(2uu_r+2u_r u_{rr})}{\sqrt{u^2+u_r^2}^3} \\
&=& \frac{uu_{rr}}{\sqrt{u^2+u_r^2}^3} \cdot \Big\{ (u^2+u_r^2) - u_r^2 \Big\}
+ \frac{u_r^2}{\sqrt{u^2+u_r^2}^3} \cdot \Big\{ (u^2+u_r^2)-u^2 \Big\} \\
&=& \frac{u^3 u_{rr}}{\sqrt{u^2+u_r^2}^3} + \frac{u_r^4}{\sqrt{u^2+u_r^2}^3}
\eas
and, similarly,
\bas
& & \hspace*{-30mm}
- \chi \frac{uv_{rr}}{\sqrt{1+v_r^2}}
+ \frac{1}{2} \cdot \chi \frac{uv_r \cdot 2v_r v_{rr}}{\sqrt{1+v_r^2}^3}
-\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r}{\sqrt{1+v_r^2}} \\
&=& - \chi \frac{uv_{rr}}{\sqrt{1+v_r^2}^3} \cdot \Big\{ (1+v_r^2)-v_r^2 \Big\}
- \chi \cdot \frac{n-1}{r} \cdot \frac{uv_r}{\sqrt{1+v_r^2}^3} \cdot (1+v_r^2) \\
&=& - \chi \frac{u(v_{rr}+\frac{n-1}{r}v_r)}{\sqrt{1+v_r^2}^3}
- \chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^3}{\sqrt{1+v_r^2}^3}
\eas
for $r\in (0,R)$ and $t\in (0,\tm)$. Since $v_{rr}+\frac{n-1}{r}v_r=\mu-u$ by (\ref{0}), the identity (\ref{36.1})
thus results from (\ref{36.2}).
\qed
We next differentiate (\ref{36.1}) to obtain a corresponding equation for $u_r$.
Here suitable arrangements will lead to the two alternative interpretations (\ref{35.01}) and (\ref{35.02})
thereof as linear inhomogeneous parabolic equations.
The first of these will be used to establish an estimate {\em from below} for $u_r$ in Lemma \ref{lem34} by a straightforward
comparison argument, whereas upon some more involved preparations, on the basis of the latter we will
apply another comparison procedure to derive a certain {\em upper} bound for $u_r$ in Lemma \ref{lem37}.
\begin{lem}\label{lem35}
Assume (\ref{init}). Then
\bea{35.1}
u_{rt}
&=& \frac{u^3 u_{rrr}}{\sqrt{u^2+u_r^2}^3}
+ 3\frac{u^2 u_r^3 u_{rr}}{\sqrt{u^2+u_r^2}^5}
- 3\frac{u^3 u_r u_{rr}^2}{\sqrt{u^2+u_r^2}^5} \nn\\
& & + 4\frac{u^2 u_r^3 u_{rr}}{\sqrt{u^2+u_r^2}^5}
+ \frac{u_r^5 u_{rr}}{\sqrt{u^2+u_r^2}^5}
-3\frac{uu_r^5}{\sqrt{u^2+u_r^2}^5} \nn\\
& & - \frac{n-1}{r^2} \cdot \frac{uu_r}{\sqrt{u^2+u_r^2}}
+ \frac{n-1}{r} \cdot \frac{u^3 u_{rr}}{\sqrt{u^2+u_r^2}^3}
+ \frac{n-1}{r} \frac{u_r^4}{\sqrt{u^2+u_r^2}^3} \nn\\
& & - \chi\mu \frac{u_r}{\sqrt{1+v_r^2}^3}
+ 2\chi \frac{uu_r}{\sqrt{1+v_r^2}^3}
+ 3\chi\mu \frac{uv_r v_{rr}}{\sqrt{1+v_r^2}^5}
- 3\chi \frac{u^2 v_r v_{rr}}{\sqrt{1+v_r^2}^5} \nn\\
& & -\chi \frac{u_{rr} v_r}{\sqrt{1+v_r^2}}
-\chi \frac{u_r v_{rr}}{\sqrt{1+v_r^2}}
+ \chi \frac{u_r v_r^2 v_{rr}}{\sqrt{1+v_r^2}^3} \nn\\
& & + \chi \cdot \frac{n-1}{r^2} \cdot \frac{uv_r^3}{\sqrt{1+v_r^2}^3}
- \chi \cdot \frac{n-1}{r} \cdot \frac{u_r v_r^3}{\sqrt{1+v_r^2}^3}
-3\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^2 v_{rr}}{\sqrt{1+v_r^2}^5}
\eea
for all $r\in (0,R)$ and $t\in (0,\tm)$. In particular,
\be{35.01}
(\parab u_r)(r,t)=0
\quad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\ee
where the inhomogeneous linear parabolic operator $\parab$ is defined by
\be{parab}
(\parab \varphi)(r,t)
:= \varphi_t - A_1(r,t)\varphi_{rr} - A_2(r,t) \varphi_r - A_3(r,t)\varphi - A_4(r,t),
\qquad r\in (0,R), \ t\in (0,\tm),
\ee
with
\bea{parab1}
A_1(r,t) &:=& \frac{u^3}{\sqrt{u^2+u_r^2}^3}, \nn\\[1mm]
A_2(r,t)&:=& 3\frac{u^2 u_r^3}{\sqrt{u^2+u_r^2}^5}
- 3\frac{u^3 u_r u_{rr}}{\sqrt{u^2+u_r^2}^5}
+ 4\frac{u^2 u_r^3}{\sqrt{u^2+u_r^2}^5}
+ \frac{u_r^5}{\sqrt{u^2+u_r^2}^5}
+ \frac{n-1}{r} \cdot \frac{u^3}{\sqrt{u^2+u_r^2}^3} \nn\\
& & - \chi \frac{v_r}{\sqrt{1+v_r^2}}, \nn\\[1mm]
A_3(r,t) &:=& -3\frac{uu_r^4}{\sqrt{u^2+u_r^2}^5}
- \frac{n-1}{r^2} \frac{u}{\sqrt{u^2+u_r^2}} \nn\\
& & -\chi\mu \frac{1}{\sqrt{1+v_r^2}^3}
+2\chi \frac{u}{\sqrt{1+v_r^2}^3}
- \chi \frac{v_{rr}}{\sqrt{1+v_r^2}}
+ \chi \frac{v_r^2 v_{rr}}{\sqrt{1+v_r^2}^3}
- \chi \cdot \frac{n-1}{r} \cdot \frac{v_r^3}{\sqrt{1+v_r^2}^3}
\qquad \mbox{and} \nn\\[1mm]
A_4(r,t) &:=& \frac{n-1}{r} \cdot \frac{u_r^4}{\sqrt{u^2+u_r^2}^3} \nn\\
& & + 3\chi\mu \frac{u v_r v_{rr}}{\sqrt{1+v_r^2}^5}
-3\chi \frac{u^2 v_r v_{rr}}{\sqrt{1+v_r^2}^5}
+ \chi \cdot \frac{n-1}{r^2} \cdot \frac{u v_r^3}{\sqrt{1+v_r^2}^3}
-3\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^2 v_{rr}}{\sqrt{1+v_r^2}^5}
\eea
for $r\in (0,R)$ and $t\in (0,\tm)$.
Likewise,
\be{35.02}
(\qarab u_r)(r,t)=0
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\ee
with $\qarab$ given by
\be{qarab}
(\qarab \varphi)(r,t):=\varphi_t - A_1(r,t) \varphi_{rr} - A_2(r,t)\varphi_r
-\tilde A_3(r,t) \varphi - \tilde A_4(r,t),
\qquad r\in (0,R), \ t\in (0,\tm),
\ee
where
\bea{qarab1}
\tilde A_3(r,t)
&:=& \frac{n-1}{r} \cdot \frac{u_r^3}{\sqrt{u^2+u_r^2}^3} \nn\\
& & - \chi \mu \frac{1}{\sqrt{1+v_r^2}^3}
+ 2\chi \frac{u}{\sqrt{1+v_r^2}^3}
-\chi \frac{v_{rr}}{\sqrt{1+v_r^2}}
+ \chi \frac{v_r^2 v_{rr}}{\sqrt{1+v_r^2}^3}
- \chi \cdot \frac{n-1}{r} \cdot \frac{v_r^3}{\sqrt{1+v_r^2}^3}
\qquad \mbox{and} \nn\\[1mm]
\tilde A_4(r,t)
&:=& -3\frac{uu_r^5}{\sqrt{u^2+u_r^2}^5}
- \frac{n-1}{r^2} \cdot \frac{uu_r}{\sqrt{u^2+u_r^2}} \nn\\
& & + 3\chi\mu \frac{uv_r v_{rr}}{\sqrt{1+v_r^2}^5}
-3\chi \frac{u^2 v_r v_{rr}}{\sqrt{1+v_r^2}^5}
+ \chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^3}{\sqrt{1+v_r^2}^3}
- 3\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^2 v_{rr}}{\sqrt{1+v_r^2}^5}
\eea
for $r\in (0,R)$ and $t\in (0,\tm)$.
\end{lem}
\proof
Differentiation of (\ref{36.1}) with respect to $r$ yields
\bea{35.2}
u_{rt}
&=& \frac{u^3 u_{rrr}}{\sqrt{u^2+u_r^2}^3}
+ 3\cdot \frac{u^2 u_r u_{rr}}{\sqrt{u^2+u_r^2}^3}
- \frac{3}{2} \cdot \frac{u^3 u_{rr} \cdot (2uu_r+2u_r u_{rr})}{\sqrt{u^2+u_r^2}^5} \nn\\
& & + 4\cdot \frac{u_r^3 u_{rr}}{\sqrt{u^2+u_r^2}^3}
-\frac{3}{2} \cdot \frac{u_r^4 \cdot (2uu_r+2u_r u_{rr})}{\sqrt{u^2+u_r^2}^5} \nn\\
& & -\frac{n-1}{r^2} \cdot \frac{uu_r}{\sqrt{u^2+u_r^2}}
+ \frac{n-1}{r} \cdot \frac{uu_{rr}}{\sqrt{u^2+u_r^2}} \nn\\
& & +\frac{n-1}{r} \cdot \frac{u_r^2}{\sqrt{u^2+u_r^2}}
-\frac{1}{2} \cdot \frac{n-1}{r} \cdot \frac{uu_r \cdot (2uu_r + 2u_r u_{rr})}{\sqrt{u^2+u_r^2}^3} \nn\\
& & -\chi\mu \cdot \frac{u_r}{\sqrt{1+v_r^2}^3}
+ 2\chi \cdot \frac{uu_r}{\sqrt{1+v_r^2}^3}
+\frac{3}{2} \chi \cdot\frac{u(\mu-u) \cdot 2v_r v_{rr}}{\sqrt{1+v_r^2}^5} \nn\\
& & -\chi \cdot \frac{u_{rr} v_r}{\sqrt{1+v_r^2}}
-\chi \cdot \frac{u_r v_{rr}}{\sqrt{1+v_r^2}}
+ \frac{1}{2} \chi \cdot \frac{u_r v_r \cdot 2v_r v_{rr}}{\sqrt{1+v_r^2}^3} \nn\\
& & +\chi \cdot \frac{n-1}{r^2} \cdot \frac{uv_r^3}{\sqrt{1+v_r^2}^3}
-\chi \cdot \frac{n-1}{r} \cdot \frac{u_r v_r^3}{\sqrt{1+v_r^2}^3} \nn\\
& & -3\chi \cdot \frac{n-1}{r} \cdot \frac{u v_r^2 v_{rr}}{\sqrt{1+v_r^2}^3}
+\frac{3}{2}\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^3 \cdot 2v_r v_{rr}}{\sqrt{1+v_r^2}^5}
\eea
for $r\in (0,R)$ and $t\in (0,\tm)$, where
\bas
3 \cdot\frac{u^2 u_r u_{rr}}{\sqrt{u^2+u_r^2}^3}
- \frac{3}{2} \cdot \frac{u^3 u_{rr} \cdot (2uu_r+2u_r u_{rr})}{\sqrt{u^2+u_r^2}^5}
&=& 3 \cdot\frac{u^2 u_r u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot
\Big\{ (u^2+u_r^2)-u^2\Big\}
-3 \cdot\frac{u^3 u_r u_{rr}^2}{\sqrt{u^2+u_r^2}^5} \\
&=& 3 \cdot\frac{u^2 u_r^3 u_{rr}}{\sqrt{u^2+u_r^2}^5}
-3 \cdot\frac{u^3 u_r u_{rr}^2}{\sqrt{u^2+u_r^2}^5}
\eas
and
\bas
4 \cdot\frac{u_r^3 u_{rr}}{\sqrt{u^2+u_r^2}^3}
-\frac{3}{2} \cdot \frac{u_r^4 \cdot (2uu_r+2u_r u_{rr})}{\sqrt{u^2+u_r^2}^5}
&=& \frac{u_r^3 u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot \Big\{ 4(u^2+u_r^2)-3u_r^2\Big\}
- 3 \cdot\frac{uu_r^5}{\sqrt{u^2+u_r^2}^5} \\
&=& 4 \cdot\frac{u^2 u_r^3 u_{rr}}{\sqrt{u^2+u_r^2}^5}
+ \frac{u_r^5 u_{rr}}{\sqrt{u^2+u_r^2}^5}
- 3 \cdot\frac{uu_r^5}{\sqrt{u^2+u_r^2}^5}
\eas
as well as
\bas
& & \hspace*{-20mm}
\frac{n-1}{r} \cdot \frac{uu_{rr}}{\sqrt{u^2+u_r^2}}
+\frac{n-1}{r} \cdot \frac{u_r^2}{\sqrt{u^2+u_r^2}}
-\frac{1}{2} \cdot \frac{n-1}{r} \cdot \frac{uu_r \cdot (2uu_r + 2u_r u_{rr})}{\sqrt{u^2+u_r^2}^3} \\
&=& \frac{n-1}{r} \cdot \frac{uu_{rr}}{\sqrt{u^2+u_r^2}^3} \cdot \Big\{ (u^2+u_r^2)-u_r^2 \Big\}
+ \frac{n-1}{r} \cdot \frac{u_r^2}{\sqrt{u^2+u_r^2}^3} \cdot \Big\{(u^2+u_r^2)-u^2\Big\} \\
&=& \frac{n-1}{r} \cdot \frac{u^3 u_{rr}}{\sqrt{u^2+u_r^2}^3}
+\frac{n-1}{r} \cdot \frac{u_r^4}{\sqrt{u^2+u_r^2}^3}
\eas
for $r\in (0,R)$ and $t\in (0,\tm)$.
Finally simplifying the last two summands in (\ref{35.2}) according to
\bas
& & \hspace*{-30mm}
-3\chi \cdot \frac{n-1}{r} \cdot \frac{u v_r^2 v_{rr}}{\sqrt{1+v_r^2}^3}
+\frac{3}{2}\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^3 \cdot 2v_r v_{rr}}{\sqrt{1+v_r^2}^5} \\
&=& -3\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^2 v_{rr}}{\sqrt{1+v_r^2}^5} \cdot \Big\{(1+v_r^2) - v_r^2 \Big\} \\
&=& -3\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^2 v_{rr}}{\sqrt{1+v_r^2}^5} \va
\eas
from (\ref{35.2}) we easily obtain (\ref{35.1}), and thus also (\ref{35.01}) and (\ref{35.02}).
\qed
Thanks to the favorable structure of the equation for $v$ in (\ref{0}), this second solution component
can be expressed explicitly in terms of $u$. This leads to the following observations which will frequently
be referred to throughout the sequel.
\begin{lem}\label{lem33}
Assume (\ref{init}). Then
\be{33.1}
v_r(r,t)=\frac{\mu r}{n} - r^{1-n} \cdot \int_0^r \rho^{n-1} u(\rho,t)d\rho
\quad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm)
\ee
and
\be{33.2}
v_{rr}(r,t)=\frac{\mu}{n} - u + \frac{n-1}{r^n} \cdot \int_0^r \rho^{n-1} u(\rho,t)d\rho
\quad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
Moreover we have
\be{33.3}
v_{rt}= - \frac{uu_r}{\sqrt{u^2+u_r^2}} + \chi \cdot \frac{uv_r}{\sqrt{1+v_r^2}}
\quad \mbox{in } (0,R)\times (0,\tm).
\ee
\end{lem}
\proof
Since by the second equation in (\ref{0}) we have
\bas
(r^{n-1}v_r)_r=r^{n-1}(\mu-u)
\quad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\eas
the identity (\ref{33.1}) easily results by integration, whereupon a differentiation of (\ref{33.1}) with respect
to $r$ yields (\ref{33.2}).\\
Next we differentiate (\ref{33.1}) with respect to $t$ and use the first equation in (\ref{0}) to see that
\bas
v_{rt}(r,t)
&=& - \frac{1}{r^{n-1}} \cdot \int_0^r \rho^{n-1} u_t(\rho,t)d\rho \\
&=& - \frac{1}{r^{n-1}} \cdot \int_0^r \rho^{n-1} \cdot
\bigg\{ \frac{1}{\rho^{n-1}} \Big( \rho^{n-1} \frac{uu_r}{\sqrt{u^2+u_r^2}}
-\chi \rho^{n-1} \frac{uv_r}{\sqrt{1+v_r^2}} \Big) \bigg\}_r (\rho,t) d\rho \\
&=& - \frac{1}{r^{n-1}} \cdot \Big\{ r^{n-1} \frac{uu_r}{\sqrt{u^2+u_r^2}}
-\chi r^{n-1} \frac{uv_r}{\sqrt{1+v_r^2}}\Big\}
\eas
for all $r\in (0,R)$ and $t\in (0,\tm)$, which shows (\ref{33.3}).
\qed
Let us note some pointwise estimates resulting from Lemma \ref{lem333} in a straightforward manner.
\begin{lem}\label{lem333}
Let (\ref{init}) hold. Then for each $t\in (0,\tm)$ and any $r\in (0,R)$ we have
\be{333.0}
-\frac{\mu R^n}{n} \cdot r^{1-n} \le v_r(r,t) \le \frac{\mu}{n} \cdot r
\ee
and
\be{333.1}
|v_r(r,t)| \le \frac{\|u(\cdot,t)\|_{L^\infty((0,R))}}{n} \cdot r
\ee
as well as
\be{333.2}
|v_{rr}(r,t)| \le \|u(\cdot,t)\|_{L^\infty((0,R))}.
\ee
\end{lem}
\proof
Fixing $t\in (0,\tm)$ and writing $M:=\|u(\cdot,t)\|_{L^\infty((0,R))}$, we clearly have $\mu \le M$, so that since from
Lemma \ref{lem33} we know that
\bas
v_r(r,t) \le \frac{\mu}{n} \cdot r
\quad \mbox{for all } r\in (0,R)
\eas
and
\bas
v_r(r,t) \ge - \frac{1}{r^{n-1}} \cdot \int_0^r \rho^{n-1} \cdot M d\rho
= - \frac{Mr}{n}
\quad \mbox{for all} r\in (0,R),
\eas
both the right inequality in (\ref{333.0}) as well as (\ref{333.1}) are immediate.
Similarly, using (\ref{33.2}) we can estimate
\bas
v_{rr}(r,t) &\le& \frac{\mu}{n} + \frac{n-1}{r^n} \cdot \int_0^r \rho^{n-1} \cdot M d\rho = \frac{\mu}{n} + \frac{(n-1)M}{n} \\
&\le& M
\quad \mbox{for all } r\in (0,R)
\eas
and
\bas
v_{rr}(r,t) \ge -M
\quad \mbox{for all } r\in (0,R),
\eas
which yields (\ref{333.2}).
Finally, to derive the left inequality in (\ref{333.0}) we observe that $\int_0^r \rho^{n-1} u(\rho,t) d\rho
\le \frac{m}{\omega_n}$ for all $r\in (0,R)$
and recall that $\frac{m}{\omega_n}=\frac{\mu R^n}{n}$ by (\ref{mu}) to obtain from (\ref{33.1}) that
\bas
v_r(r,t) \ge - r^{1-n} \int_0^r \rho^{n-1} u(\rho,t) d\rho
\ge - r^{1-n} \cdot \frac{m}{\omega_n}
= - r^{1-n} \cdot \frac{\mu R^n}{n}
\quad \mbox{for all } r\in (0,R).
\eas
This completes the proof.
\qed
\mysection{A pointwise estimate from below for $u$}\label{sect3}
In order to show that (\ref{ext_crit}) actually reduces to (\ref{43.1}), let us first rule out the occurrence
of the first alternative in (\ref{ext_crit}).
In proving this, we shall make use of the following elementary inequality.
\begin{lem}\label{lem24}
We have
\bas
\frac{\xi}{\sqrt{1+\xi}^3} \le \frac{2}{3\sqrt{3}}
\qquad \mbox{for all } \xi\ge 0.
\eas
\end{lem}
\proof
Since $\varphi(\xi):=\frac{\xi}{\sqrt{1+\xi}^3}, \ \xi\ge 0$, satisfies $\varphi(0)=0, \varphi(\xi)\to 0$ as
$\xi\to\infty$ and
$\varphi'(\xi)=(1+\xi)^{-\frac{5}{2}} \cdot (1-\frac{\xi}{2})$ for all $\xi>0$, it follows that
$\varphi(\xi) \le \varphi(2) = \frac{2}{\sqrt{3}^3}$ for all $\xi\ge 0$.
\qed
By means of a comparison argument applied to (\ref{36.1}),
we can now in fact exclude that solutions attain zeros within finite time.
\begin{lem}\label{lem23}
If (\ref{init}) holds, then
\be{23.1}
u(r,t) \ge \Big( \inf_{r\in (0,R)} u_0(r) \Big) \cdot e^{-\kappa t}
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\ee
where
\be{23.2}
\kappa:=\chi \mu + \frac{2(n-1)\chi\mu}{3\sqrt{3}n}.
\ee
\end{lem}
\proof
We rewrite (\ref{36.1}) in the form
\be{23.3}
u_t= a_1(r,t) u_{rr} + a_{21}(r,t)u_r + \frac{a_{22}(r,t)}{r} \cdot u_r
-\chi \cdot \frac{u(\mu-u)}{\sqrt{1+v_r^2}^3}
-\frac{n-1}{r} \cdot \chi \frac{uv_r^3}{\sqrt{1+v_r^2}^3}
\ee
for all $r\in (0,R)$ and $t\in (0,\tm)$, where
\bas
a_1(r,t):=\frac{u^3}{\sqrt{u^2+u_r^2}^3}
\eas
and
\bas
a_{21}(r,t):=\frac{u_r^3}{\sqrt{u^2+u_r^2}^3} - \chi \cdot \frac{v_r}{\sqrt{1+v_r^2}}
\eas
as well as
\bas
a_{22}(r,t):=(n-1)\cdot \frac{u}{\sqrt{u^2+u_r^2}}
\eas
define continuous functions in $[0,R] \times (0,\tm)$.
In (\ref{23.3}), we can estimate
\bas
- \chi \cdot\frac{u(\mu-u)}{\sqrt{1+v_r^2}^3}
\ge - \chi\mu \cdot \frac{u}{\sqrt{1+v_r^2}^3}
\ge - \chi\mu u
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\eas
and in order to control the last term in (\ref{23.3}) we
use the one-sided inequality $v_r \le \frac{\mu r}{n}$ provided by Lemma \ref{lem333}, which in conjunction with
Lemma \ref{lem24} entails that
\bas
- \frac{n-1}{r} \cdot \chi \frac{uv_r^3}{\sqrt{1+v_r^2}^3}
&=& - (n-1)\chi \cdot \frac{v_r^2}{\sqrt{1+v_r^2}^3} \cdot \frac{v_r}{r} \cdot u \\
&\ge& - (n-1)\chi \cdot \frac{2}{3\sqrt{3}} \cdot \frac{\mu}{n} \cdot u
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\eas
Accordingly, from (\ref{23.3}) we infer that with $\kappa$ as in (\ref{23.2}) we have
\bas
u_t \ge a_1(r,t) u_{rr} + a_{21}(r,t) u_r + \frac{a_{22}(r,t)}{r} \cdot u_r - \kappa u
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\eas
so that for all $\eps>0$, writing $\varphi(r,t):=e^{(\kappa+\eps)t} u(r,t)$ we see that
\bea{23.4}
\hspace*{-18mm}
\varphi_t &\ge& e^{(\kappa+\eps)t} \cdot \Big\{
a_1(r,t) u_{rr} + a_{21}(r,t) u_r + a_{22}(r,t) u_r - \kappa u + (\kappa+\eps)u \Big\} \nn\\
&=& a_1(r,t) \varphi_{rr} + a_{21}(r,t) \varphi_r + a_{22}(r,t) \varphi_r +\eps e^{(\kappa+\eps)t} u
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\eea
Now if for some $T\in (0,\tm)$, $\varphi$ attains its minimum over $[0,R] \times [0,T]$ at some
$(r_0,t_0) \in [0,R] \times [0,T]$, then necessarily
\be{23.5}
\varphi_r(r_0,t_0)=0, \quad \varphi_{rr}(r_0,t_0) \ge 0 \quad \mbox{and} \quad \varphi_t(r_0,t_0) \le 0.
\ee
Therefore, in the case $t_0>0$ and $r_0>0$ we may directly apply (\ref{23.4}) to obtain
\bas
0 \ge \varphi_t(r_0,t_0)
&\ge& a_1(r_0,t_0) \varphi_{rr}(r_0,t_0) + a_{21}(r_0,t_0) \varphi_r(r_0,t_0)
+ \frac{a_{22}(r_0,t_0)}{r_0} \cdot \varphi_r(r_0,t_0)
+ \eps e^{(\kappa+\eps)t_0} u(r_0,t_0) \\
&\ge& \eps e^{(\kappa+\eps)t_0} u(r_0,t_0)
>0,
\eas
which is impossible.
However, if $t_0>0$ and $r_0=0$, then there must exist a sequence
$(r_j)_{j\in\N}$ of numbers $r_j\in (0,R)$ such that $r_j\searrow 0$
as $j\to\infty$ and $\varphi_r(r_j,t_0)\ge 0$ for all $j\in\N$, because otherwise $\varphi(\cdot,t_0)$
would have a strict local maximum at $r=0$.
Since $a_3\ge 0$, evaluating (\ref{23.4}) at $r=r_j=$ we would thus obtain that
\bas
\varphi_t(r_j,t_0)
&\ge& a_1(r_j,t_0) \varphi_{rr}(r_j,t_0) + a_{21}(r_j,t_0) \varphi_r(r_j,t_0)
+ \frac{a_{22}(r_j,t_0)}{r_j} \cdot \varphi_r(r_j,t_0) + \eps e^{(\kappa+\eps)t_0} u(r_j,t_0) \\
&\ge& a_1(r_j,t_0) \varphi_{rr}(r_j,t_0) + a_{21}(r_j,t_0) \varphi_r(r_j,t_0)
+ \eps e^{(\kappa+\eps)t_0} u(r_j,t_0)
\qquad \mbox{for all } j\in\N,
\eas
so that since $\varphi(\cdot,t_0)$ is smooth in $[0,R]$,
we may let $j\to\infty$ here to infer using (\ref{23.5}) that
\bas
0 \ge \varphi_t(0,t_0)
&\ge& a_1(0,t_0) \varphi_{rr}(0,t_0) + a_{21}(0,t_0) \varphi_r(0,t_0)
+ \eps e^{(\kappa+\eps)t_0} u(0,t_0) \\
&\ge& \eps e^{(\kappa+\eps)t_0} u(0,t_0)
>0.
\eas
This absurd conclusion shows that actually $t_0=0$, which implies that
$\varphi\ge \inf_{r\in (0,R)} \varphi(r,0)=\inf_{r\in (0,R)} u_0(r)$
throughout $[0,R] \times [0,T]$ for any $T\in (0,\tm)$.
Taking $T\nearrow \tm$ and $\eps\searrow 0$ we thereby obtain (\ref{23.1}).
\qed
\mysection{A pointwise lower estimate for $u_r$}\label{sect4}
It remains to exclude the possibility of finite-time blow-up of $u_r$ despite boundedness of $u$.
A first step toward this can accomplished by invoking parabolic comparison to derive the following lower bound for
$u_r$ from (\ref{35.01}).
Let us emphasize that our argument makes essential use of the fact that on the right-hand side of (\ref{35.1}),
the most singular term $\frac{n-1}{r^2} \cdot \frac{uu_r}{\sqrt{u^2+u_r^2}}$ therein
appears with a negative sign, and that
in consequence a corresponding upper estimate for $u_r$ can apparently not be obtained by
a direct approach of the type pursued here, at least not when $n\ge 2$.
\begin{lem}\label{lem34}
Assume that $\tm<\infty$, but that $\sup_{(r,t)\in (0,R)\times (0,\tm)} u(r,t)<\infty$.
Then there exists $C>0$ such that
\be{34.1}
u_r(r,t) \ge -C
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
\end{lem}
\proof
According to our hypothesis, we can find $c_1>0$ such that
\be{34.2}
u(r,t)\le c_1
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\ee
so that Lemma \ref{lem333} provides $c_2>0$ and $c_3>0$ such that
\be{34.3}
|v_r(r,t)| \le c_2 r
\quad \mbox{and} \quad
|v_{rr}(r,t)| \le c_3
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
We now take $D\ge 1$ and $\alpha>0$ large enough fulfilling
\be{34.4}
u_{0r}(r) > -D
\qquad \mbox{for all } r\in (0,R)
\ee
and
\be{34.5}
\alpha>c_4 + \frac{c_5}{D}\va
\ee
where
\be{34.54}
c_4:=2c_1\chi + c_3\chi + c_2^2 c_3 \chi R^2 + (n-1) c_2^3 \chi R^2
\ee
and
\be{34.55}
c_5:=3c_1 c_2 c_3 \chi\mu + 3c_1^2 c_2 c_3 \chi R + (n-1) c_1 c_2^3 \chi R + 3(n-1) c_1 c_2^2 c_3 \chi R,
\ee
and define a comparison function $\uphi$ by letting
\bas
\uphi(r,t):=-D e^{\alpha t}
\qquad \mbox{for $r\in [0,R]$ and } t\ge 0.
\eas
Then since $\uphi_r=\uphi_{rr}\equiv 0$, with $\parab$ as in (\ref{parab}) we have
\bea{34.6}
(\parab \uphi)(r,t)
&=& -\alpha D e^{\alpha t} \nn\\
& & - 3\frac{u_r^4 \cdot D e^{\alpha t}}{\sqrt{u^2+u_r^2}^5}
- \frac{n-1}{r^2} \cdot \frac{u\cdot D e^{\alpha t}}{\sqrt{u^2+u_r^2}}
- \chi\mu \frac{D e^{\alpha t}}{\sqrt{1+v_r^2}^3} \nn\\
& & + 2\chi \frac{u\cdot D e^{\alpha t}}{\sqrt{1+v_r^2}^3}
- \chi \frac{v_{rr} \cdot D e^{\alpha t}}{\sqrt{1+v_r^2}}
+ \chi \frac{v_r^2 v_{rr} \cdot D e^{\alpha t}}{\sqrt{1+v_r^2}^3} \nn\\
& & - \chi \cdot \frac{n-1}{r} \cdot \frac{v_r^3 \cdot D e^{\alpha t}}{\sqrt{1+v_r^2}^3} \nn\\
& & - \frac{n-1}{r} \cdot \frac{u_r^4}{\sqrt{u^2+u_r^2}^3}
- 3\chi\mu \frac{u v_r v_{rr}}{\sqrt{1+v_r^2}^5}
+ 3\chi \frac{u^2 v_r v_{rr}}{\sqrt{1+v_r^2}^5} \nn\\
& & - \chi \cdot \frac{n-1}{r^2} \cdot \frac{uv_r^3}{\sqrt{1+v_r^2}^3}
+ 3\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^2 v_{rr}}{\sqrt{1+v_r^2}^5}
\eea
for all $r\in (0,R)$ and $t\in (0,\tm)$.
Here the second, third, fourth and ninth term on the right are nonpositive, and we claim that each of
the remaining summands containing $\chi$ can be controlled in modulus by the first term on the right-hand side
suitably.
Indeed, repeatedly using (\ref{34.2}), (\ref{34.3}) and (\ref{34.4}), we can estimate
\bas
\bigg| 2\chi \frac{u\cdot D e^{\alpha t}}{\sqrt{1+v_r^2}^3} \bigg|
\le 2\chi \cdot c_1 \cdot D e^{\alpha t},
\eas
\bas
\bigg| -\chi \frac{v_{rr} \cdot D e^{\alpha t}}{\sqrt{1+v_r^2}} \bigg|
\le \chi \cdot c_3 \cdot D e^{\alpha t},
\eas
\bas
\bigg| \chi \frac{v_r^2 v_{rr} \cdot D e^{\alpha t}}{\sqrt{1+v_r^2}^3} \bigg|
\le \chi \cdot c_2^2 r^2 \cdot c_3 \cdot D e^{\alpha t}
\le c_2^2 c_3 \chi R^2 \cdot D e^{\alpha t},
\eas
\bas
\bigg| -3\chi\mu \frac{uv_r v_{rr}}{\sqrt{1+v_r^2}^5} \bigg|
\le 3\chi\mu \cdot c_1 \cdot c_2 r \cdot c_3
\le 3c_1c_2c_3 \chi\mu R
\eas
as well as
\bas
\bigg| 3\chi \frac{u^2 v_r v_{rr}}{\sqrt{1+v_r^2}^5} \bigg|
\le 3\chi \cdot c_1^2 \cdot c_2 r \cdot c_3
\le 3c_1^2 c_2 c_3 \chi R
\eas
and finally,
\bas
\bigg| -\chi \cdot \frac{n-1}{r^2} \cdot \frac{uv_r^3}{\sqrt{1+v_r^2}^3} \bigg|
\le \chi \cdot \frac{n-1}{r^2} \cdot c_1 \cdot c_2^3 r^3
\le (n-1) c_1 c_2^3 \chi R,
\eas
as well as
\bas
\bigg| - \chi \cdot \frac{n-1}{r} \cdot \frac{v_r^3 \cdot D e^{\alpha t}}{\sqrt{1+v_r^2}^3} \bigg|
\le \chi \cdot \frac{n-1}{r} \cdot c_2^3 r^3 \cdot D e^{\alpha t}
\le (n-1) c_2^3 \chi R^2 \cdot D e^{\alpha t}
\eas
and
\bas
\bigg| 3\chi \cdot \frac{n-1}{r} \cdot \frac{uv_r^2 v_{rr}}{\sqrt{1+v_r^2}^5} \bigg|
\le 3\chi \cdot \frac{n-1}{r} \cdot c_1 \cdot c_2^2 r^2 \cdot c_3
\le 3(n-1) c_1 c_2^2 c_3 \chi R.
\eas
Therefore, (\ref{34.6}) implies that with $c_4$ and $c_5$ as in (\ref{34.54}) and (\ref{34.55}) we have
\bas
(\parab \uphi)(r,t)
&\le& -\alpha D e^{\alpha t} + c_4 \cdot D e^{\alpha t} + c_5 \\
&\le& -\alpha D e^{\alpha t} + \Big(c_4+\frac{c_5}{D}\Big) \cdot D e^{\alpha t}
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\eas
whence our assumption (\ref{34.5}) on $\alpha$ ensures that $(\parab \uphi)(r,t)< 0$
for all $r\in (0,R)$ and $t\in (0,\tm)$.
Since $(\parab u_r)(r,t)=0$ for all $(r,t)\in (0,R)\times (0,\tm)$ by Lemma \ref{lem35}, and since moreover
\bas
\uphi(r,0)=-D<u_{0r}(r)=u_r(0,r)
\qquad \mbox{for all } r\in [0,R]
\eas
and, clearly,
\bas
\uphi_r(0,t)=u_r(0,t)=0
\quad \mbox{as well as} \quad
\uphi_r(R,t)=u_r(R,t)=0
\qquad \mbox{for all } t\in (0,\tm),
\eas
from the comparison principle we conclude that $u_r(r,t)\ge \uphi(r,t)$ for all $r\in (0,R)$ and
$t\in (0,\tm)$, and that hence
\bas
u_r(r,t) \ge -D e^{\alpha \tm}
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\eas
which proves the claim.
\qed
\mysection{A bound for $|u_r|$. Proof of Theorem \ref{theo43}}\label{sect5}
The goal of this section is to complete the proof of Theorem \ref{theo43} by
further developing the one-sided inequality for $u_r$ from Lemma \ref{lem34}
into a bound for $|u_r|$ in modulus, provided that $\tm$ is finite but $u$ itself remains bounded
(Corollary \ref{cor42}).
An important role in our analysis in this direction
will be played by the function $z:=\frac{u_t}{u}$,
which is indeed well-defined and continuous in $[0,R]\times [0,\tm)$ by Lemma \ref{lem21}.
Furthermore, according to Lemma \ref{lem36} we have the representation
\bea{z}
z &=& \frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^3} + \frac{u_r^4}{u\sqrt{u^2+u_r^2}^3}
+ \frac{n-1}{r} \cdot \frac{u_r}{\sqrt{u^2+u_r^2}} \nn\\
& & - \chi \frac{u_r v_r}{u\sqrt{1+v_r^2}} - \chi \frac{\mu-u}{\sqrt{1+v_r^2}^3}
- \chi \cdot \frac{n-1}{r} \cdot \frac{v_r^3}{\sqrt{1+v_r^2}^3}\va
\eea
for $r\in (0,R)$ and $t\in (0,\tm)$.\abs
Now in a first key observation, to be presented in Corollary \ref{cor40},
we will establish a useful relationship between $u_r$ and the function $z$,
essentially controlling $\|u_r(\cdot,t)\|_{L^\infty((0,R))}$ for any fixed $t\in (0,\tm)$
by the maximum of the {\em positive part} $z_+$ of $z$ over the {\em whole memory region} $(0,R)\times (0,t)$.
To achieve this, we will foremost use an integral technique to estimate $|u_r|$ in terms of $z_+$ on the basis
of (\ref{z}) and Lemma \ref{lem34} in a suitably small subinterval $(0,R_0)$ of $(0,R)$ (Lemma \ref{lem32}).
This will in particular imply an upper bound for $u_r$ at $r=R_0$ and therefore allow for applying
a comparison argument to (\ref{35.02})
which will yield a pointwise upper estimate for $u_r$ in the corresponding outer region $(R_0,R)$
(Lemma \ref{lem37}).\abs
The second essential step will thereafter consist in deriving a nonlocal parabolic inequality for $z$
with a memory-type nonlinearity (Lemma \ref{lem39}).
Upon another comparison, this will entail a pointwise upper bound for $z$ (Lemma \ref{lem41})
and hence also for $u_r$.
\subsection{A bound for $|u_r|$ in terms of $z_+$}
\subsubsection{Estimating $|u_r|$ near the origin}
Let us first apply an appropriate testing procedure to (\ref{z}) to find some small $R_0\in (0,R)$ with the
property that $u_r(\cdot,t)$ can be bounded in certain weighted Lebesgue spaces over $(0,R_0)$
in such a way that on taking limits we can derive a respective $L^\infty$ estimate from this.
\begin{lem}\label{lem32}
Assume that $\tm<\infty$, but that $\sup_{(r,t)\in (0,R)\times (0,\tm)} u(r,t)<\infty$.
Then there exist $R_0\in (0,R)$ and $C>0$ such that
\be{32.01}
\|u_r(\cdot,t)\|_{L^\infty((0,R_0))} \le C \cdot \Big(1+\|z_+(\cdot,t)\|_{L^\infty((0,R_0))} \Big)
\qquad \mbox{for all } t\in (0,\tm).
\ee
\end{lem}
\proof
We first rearrange (\ref{z}) to obtain
\bea{32.1}
\frac{u_r^4}{u^3} &=& \frac{\sqrt{u^2+u_r^2}}{u^2} \cdot z
- u_{rr} - \frac{n-1}{r} \cdot \frac{\sqrt{u^2+u_r^2}^2 u_r}{u^2} \nn\\
& & + \chi \frac{(\mu-u)\sqrt{u^2+u_r^2}^3}{\sqrt{1+v_r^2}^3}
+ \chi \frac{\sqrt{u^2+u_r^2}^3 u_r v_r}{u^3 \sqrt{1+v_r^2}}
+ \chi \cdot \frac{n-1}{r} \cdot \frac{\sqrt{u^2+u_r^2}^3 v_r^3}{u^2 \sqrt{1+v_r^2}^3},
\eea
where
\bea{32.100}
- u_{rr} - \frac{n-1}{r} \cdot \frac{\sqrt{u^2+u_r^2}^2 u_r}{u^2}
&=& - \Big(u_{rr}+\frac{n-1}{r} u_r\Big) - \frac{n-1}{r} \cdot \frac{u_r^3}{u^2} \nn\\
&=& - \frac{1}{r^{n-1}} (r^{n-1} u_r)_r
-\frac{n-1}{r} \cdot \frac{u_r^3}{u^2}
\eea
for $r\in (0,R)$ and $t\in (0,\tm)$.
In order to choose $R_0$ appropriately, we use our boundedness assumption on $u$ to fix $c_1\ge \mu$ and
$c_2>0$ such that
\be{32.2}
u(r,t) \le c_1
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm)
\ee
and
\be{32.3}
\sqrt{u^2+u_r^2}^3 \le c_2 \cdot (1+|u_r|^3)
\qquad \mbox{in } (0,R)\times (0,\tm),
\ee
and recall Lemma \ref{lem23} to find $c_3>0$ fulfilling
\be{32.4}
u(r,t)\ge c_3
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
We claim that then the conclusion of the lemma holds if we pick any $R_0\in (0,R)$ satisfying
\be{32.5}
R_0 \le \frac{n c_3^3}{4c_1^3 c_2 \chi\mu}.
\ee
To see this, we take an arbitrary even integer $m\ge 0$, multiply (\ref{32.1}) by $r^{n-1} u_r^m$
and integrate over $(0,R_0)$ to see using (\ref{32.100}) that
\bea{32.6}
I(t) &:=&
\int_0^{R_0} r^{n-1} \frac{u_r^{m+4}}{u^3} dr \nn\\
&=& \int_0^{R_0} r^{n-1} \frac{\sqrt{u^2+u_r^2}^3}{u^2} \cdot u_r^m z dr
- \int_0^{R_0} (r^{n-1} u_r)_r \cdot u_r^m dr
- (n-1) \int_0^{R_0} r^{n-2} \frac{u_r^{m+3}}{u^2} dr \nn\\
& & + \chi \int_0^{R_0} r^{n-1} \frac{(\mu-u) \sqrt{u^2+u_r^2}^3 u_r^m}{u^2 \sqrt{1+v_r^2}^3} dr
+ \chi \int_0^{R_0} r^{n-1} \frac{\sqrt{u^2+u_r^2}^3 u_r^{m+1} v_r}{u^3\sqrt{1+v_r^2}} dr \nn\\
& & + (n-1)\chi \int_0^{R_0} r^{n-2} \frac{\sqrt{u^2+u_r^2}^3 u_r^m v_r^3}{u^2 \sqrt{1+v_r^2}^3} dr
\nn\\[2mm]
&=:& J_1(t)+...+J_6(t)
\qquad \mbox{for all } t\in (0,\tm).
\eea
Here by (\ref{32.2}) we have
\be{32.7}
I(t) \ge \frac{1}{c_1^3} \cdot \int_0^{R_0} r^{n-1} u_r^{m+4} dr
\qquad \mbox{for all } t\in (0,\tm),
\ee
and our goal is to show that the sum on the right-hand side of (\ref{32.6}) can be controlled adequately by
the term on the right of (\ref{32.7}).\\
For this purpose, we first use Lemma \ref{lem33} in rewriting $J_5(t)$ to obtain
\bea{32.8}
J_5(t)
&=& \frac{\chi \mu}{n} \int_0^{R_0} r^n \cdot \frac{\sqrt{u^2+u_r^2}^3 u_r^{m+1}}{u^3 \sqrt{1+v_r^2}} dr
- \chi \int_0^{R_0} \frac{\sqrt{u^2+u_r^2}^3 u_r^{m+1}}{u^3 \sqrt{1+v_r^2}}
\cdot \bigg( \int_0^r \rho^{n-1} u(\rho,t) d\rho \bigg) dr \nn\\
&=:& J_{51}(t)+J_{52}(t)
\qquad \mbox{for all } t\in (0,\tm),
\eea
where (\ref{32.3}), (\ref{32.4}) and Young's inequality enable us to infer that
\bas
J_{51}(t)
&\le& \frac{\chi\mu}{n} \int_0^{R_0} r^n \cdot \frac{c_2 \cdot (1+|u_r|^3) \cdot |u_r|^{m+1}}{c_3^3} dr \\
&=& \frac{c_2 \chi\mu}{nc_3^3} \int_0^{R_0} r^n |u_r|^{m+1} dr
+ \frac{c_2 \chi\mu}{nc_3^3} \int_0^{R_0} r^n u_r^{m+4} dr \\
&\le& \frac{c_2\chi\mu}{nc_3^3} \int_0^{R_0} r^n (1+u_r^{m+4}) dr
+ \frac{c_2 \chi\mu}{nc_3^3} \int_0^{R_0} r^n u_r^{m+4} dr \\
&=& \frac{c_2 \chi\mu R_0^{n+1}}{n(n+1)c_3^3}
+ \frac{2c_2 \chi\mu}{nc_3^3} \int_0^{R_0} r^n u_r^{m+4}dr
\qquad \mbox{for all } t\in (0,\tm).
\eas
Trivially estimating
\bas
\int_0^{R_0} r^n u_r^{m+4} dr
\le R_0 \int_0^{R_0} r^{n-1} u_r^{m+4},
\eas
according to (\ref{32.7}) and our restriction (\ref{32.5}) on $R_0$ we thus conclude that
\be{32.9}
J_{51}(t)
\le c_4 + \frac{1}{2c_1^3} \int_0^{R_0} r^{n-1} u_r^{m+4} dr
\qquad \mbox{for all } t\in (0,\tm)
\ee
with $c_4:=\frac{c_2 \chi\mu R^{n+1}}{n(n+1)c_3^3}$. \\
In order to derive an appropriate upper bound for the second term on the right of (\ref{32.8}), let us apply
Lemma \ref{lem34} to fix a constant $L\ge 1$ such that
\be{32.10}
u_r(r,t) \ge -L
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
Then since $-u_r^{m+1} = (-u_r)^{m+1}$ due to the fact that $m+1$ is odd, this in conjunction with (\ref{32.3}),
(\ref{32.4}) and (\ref{32.1}) implies that
\bea{32.11}
J_{52}(t)
&\le& \chi L^{m+1} \int_0^{R_0} \frac{\sqrt{u^2+u_r^2}^3}{u^3 \sqrt{1+v_r^2}}
\cdot \bigg(\int_0^r \rho^{n-1} u(\rho,t) d\rho\bigg) dr \nn\\
&\le& \frac{c_1 c_2 \chi}{c_3^3} L^{m+1} \int_0^{R_0} (1+|u_r|^3)
\cdot \Big( \int_0^r \rho^{n-1} d\rho\Big) dr \nn\\
&=& \frac{c_1 c_2 \chi}{nc_3^3} L^{m+1} \int_0^{R_0} r^n (1+|u_r|^3) dr \nn\\
&=& \frac{c_1 c_2 \chi R_0^{n+1}}{n(n+1) c_3^3} L^{m+1}
+ \frac{c_1 c_2 \chi}{nc_3^3} \int_0^{R_0} r^n L^{m+1} |u_r|^3 dr
\qquad \mbox{for all } t\in (0,\tm).
\eea
As by Young's inequality we have
\bas
\int_0^{R_0} r^n L^{m+1} |u_r|^3 dr
&\le& \frac{3}{m+4} \int_0^{R_0} r^n u_r^{m+4} dr
+ \frac{m+1}{m+4} \int_0^{R_0} r^n L^{m+4} dr \\
&\le& \frac{3R}{m+4} \int_0^{R_0} r^{n-1} u_r^{m+4} dr
+ \frac{R^{n+1}}{n+1} L^{m+4},
\eas
in view of the fact that $L\ge 1$ we obtain from (\ref{32.11}) that
\be{32.12}
J_{52}(t)
\le c_5 L^{m+4} + \frac{c_6}{m+4} \int_0^{R_0} r^{n-1} u_r^{m+4} dr
\qquad \mbox{for all } t\in (0,\tm)
\ee
with $c_5:=\frac{2c_1 c_2 \chi R^{n+1}}{n(n+1)c_3^3}$ and $c_6:=\frac{3c_1 c_2 \chi R}{nc_3^3}$. \\
Going back to (\ref{32.6}), we next apply (\ref{32.3}), (\ref{32.4}) and Young's and H\"older's inequalities
to estimate
\bas
J_1(t)
&\le& \frac{c_2}{c_3^2} \int_0^{R_0} r^{n-1} (1+|u_r|^3) \cdot u_r^m z_+ dr \\
&=& \frac{c_2}{c_3^2} \int_0^{R_0} r^{n-1} u_r^m z_+ dr
+ \frac{c_2}{c_3^2} \int_0^{R_0} r^{n-1} |u_r|^{m+3} z_+ dr \\
&\le& \frac{c_2}{c_3^2} \int_0^{R_0} r^{n-1} z_+ dr
+ \frac{2c_2}{c_3^2} \int_0^{R_0} r^{n-1} |u_r|^{m+3} z_+ dr \\
&\le& \frac{c_2}{c_3^2} \bigg( \int_0^{R_0} r^{n-1} z_+^{m+4} dr \bigg)^\frac{1}{m+4}
\cdot \Big(\frac{R_0^n}{n}\Big)^\frac{m+3}{m+4}
+\frac{2c_2}{c_3^2} \bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4}
\cdot \bigg(\int_0^{R_0} r^{n-1} z_+^{m+4} dr\bigg)^\frac{1}{m+4},
\eas
which means that if we let $R_1:=\max \{1,R\}$ and $c_7:=\max \{ \frac{c_2}{c_3^2} \cdot \frac{R_1^n}{n} \, , \,
\frac{2c_2}{c_3^2} \}$, then
\be{32.13}
J_1(t) \le c_7 \cdot \Bigg\{ 1 + \bigg( \int_0^{R_0} r^{n-1} u_r^{m+4}dr \bigg)^\frac{m+3}{m+4} \Bigg\}
\cdot \bigg( \int_0^{R_0} r^{n-1} z_+^{m+4} dr \bigg)^\frac{1}{m+4}
\qquad \mbox{for all } t\in (0,\tm).
\ee
As for the second term on the right of (\ref{32.6}), in order to remove second-order derivatives, in the case
$n\ge 2$ we twice integrate by parts to obtain
\bas
J_2(t)
&=& m \int_0^{R_0} r^{n-1} u_r^m u_{rr} dr
- R_0^{n-1} u_r^{m+1} (R_0,t) \\
&=& - \frac{(n-1)m}{m+1} \int_0^{R_0} r^{n-2} u_r^{m+1} dr
+ \frac{m}{m+1} R_0^{n-1} u_r^{m+1}(R_0,t)
- R_0^{n-1} u_r^{m+1} (R_0,t) \\
&=& - \frac{(n-1)m}{m+1} \int_0^{R_0} r^{n-2} u_r^{m+1} dr
- \frac{1}{m+1} R_0^{n-1} u_r^{m+1}(R_0,t)
\qquad \mbox{for all } t\in (0,\tm).
\eas
Once more since $m+1$ is odd, (\ref{32.10}) again becomes applicable to provide the one-sided estimate
\bea{32.14}
J_2(t)
&\le& \frac{(n-1)m}{m+1} L^{m+1} \int_0^{R_0} r^{n-2}
+\frac{1}{m+1} R_0^{n-1} L^{m+1} \nn\\
&\le& c_8 L^{m+1}
\qquad \mbox{for all } t\in (0,\tm)
\eea
with $c_8:=\frac{(n-1)m}{m+1} \cdot \frac{R^{n-1}}{n-1} + \frac{1}{m+1} R^{n-1} \equiv R^{n-1}$
when $n\ge 2$, and it can easily be verified that this conclusion can be extended so as to include the case $n=1$
as well.\\
Similarly, since also $m+3$ is odd, we may invoke (\ref{32.10}) and then (\ref{32.4}) to see that in the case
$n\ge 2$,
\bea{32.15}
J_3(t)
&\le& (n-1)\int_0^{R_0} r^{n-2} \frac{L^{m+3}}{u^2} dr \nn\\
&\le& \frac{n-1}{c_3^2} L^{m+3} \frac{R_0^{n-1}}{n-1} \nn\\
&\le& c_9 L^{m+3}
\qquad \mbox{for all } t\in (0,\tm)
\eea
with $c_9:=\frac{R^{n-1}}{c_3^2}$, and note that (\ref{32.15}) trivially holds when $n=1$.\\
We next estimate $J_4(t)$ by first using (\ref{32.3}), (\ref{32.4}) and Young's inequality according to
\bas
J_4(t)
&\le& \frac{c_2 \chi\mu}{c_3^2} \int_0^{R_0} r^{n-1} (1+|u_r|^3) u_r^m dr \\
&\le& \frac{c_2\chi\mu}{c_3^2} \int_0^{R_0} r^{n-1} dr
+ \frac{2c_2\chi\mu}{c_3^2} \int_0^{R_0} r^{n-1} |u_r|^{m+3} dr \\
&=& \frac{c_2\chi\mu R_0^n}{nc_3^2}
+ \frac{2c_2\chi\mu}{c_3^2} \int_0^{R_0} r^{n-1} |u_r|^{m+3} dr
\qquad \mbox{for all } t\in (0,\tm).
\eas
Since thanks to the H\"older inequality we know that
\bea{32.155}
\int_0^{R_0} r^{n-1} |u_r|^{m+3} dr
&\le& \Big(\frac{R_0^n}{n}\Big)^\frac{1}{m+4} \cdot
\bigg( \int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} \nn\\
&\le& \frac{R_1^n}{n} \cdot
\bigg( \int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4},
\eea
again with $R_1=\max \{1,R\}$, this entails that
\be{32.16}
J_4(t) \le c_{10} \cdot \Bigg\{
1 + \bigg( \int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} \Bigg\}
\qquad \mbox{for all } t\in (0,\tm)
\ee
if we let
$c_{10}:=\max\{ \frac{c_2\chi\mu R^n}{nc_3^2} \, , \, \frac{2c_2\chi\mu}{c_3^2} \cdot \frac{R_1^n}{n} \}$.\\
Finally, in treating the last integral in (\ref{32.6}) we make use of the upper estimate for $v_r$ in
(\ref{333.0}) to see, again recalling (\ref{32.3}), (\ref{32.4}) and using Young's inequality, that
\bas
J_6(t)
&\le& \frac{(n-1)\chi\mu^3}{n^3} \int_0^{R_0} \frac{r^{n+1} \sqrt{u^2+u_r^2}^3 u_r^m}{u^2\sqrt{1+v_r^2}^3}
dr \\
&\le& \frac{(n-1)c_2 \chi \mu^3}{n^3 c_3^2} \int_0^{R_0} r^{n+1} (1+|u_r|^3) u_r^m dr \\
&\le& \frac{(n-1)c_2 \chi \mu^3}{n^3 c_3^2} \int_0^{R_0} r^{n+1} dr
+ \frac{2(n-1)c_2 \chi \mu^3}{n^3 c_3^2} \int_0^{R_0} r^{n+1} |u_r|^{m+3} dr \\
&=& \frac{(n-1)c_2 \chi \mu^3 R_0^{n+2}}{n^3 (n+2) c_3^2}
+ \frac{2(n-1)c_2 \chi \mu^3}{n^3 c_3^2} \int_0^{R_0} r^{n+1} |u_r|^{m+3} dr
\qquad \mbox{for all } t\in (0,\tm).
\eas
Now due to (\ref{32.155}),
\bas
\int_0^{R_0} r^{n+1} |u_r|^{m+3} dr
&\le& R_0^2 \int_0^{R_0} r^{n-1} |u_r|^{m+3} dr \\
&\le& \frac{R_1^{n+2}}{n} \bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4},
\eas
whence we obtain that
\be{32.17}
J_6(t) \le c_{11} \cdot \Bigg\{ 1 + \bigg( \int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} \Bigg\}
\qquad \mbox{for all } t\in (0,\tm)
\ee
with $c_{11}:=\max\Big\{ \frac{(n-1) c_2 \chi \mu^3 R^{n+2}}{n^3(n+2)c_3^2} \, , \,
\frac{2(n-1) c_2 \chi\mu^3}{n^3 c_2^3} \cdot \frac{R_1^{n+2}}{n} \Big\}$.\abs
In summary, (\ref{32.7}), (\ref{32.13}), (\ref{32.14}), (\ref{32.15}), (\ref{32.8}), (\ref{32.9}), (\ref{32.12})
and (\ref{32.17}) combined with (\ref{32.6}) show that
\bas
\frac{1}{c_1^3} \int_0^{R_0} r^{n-1} u_r^{m+4} dr
&\le& c_7 \cdot \Bigg\{ 1+\bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} \Bigg\}
\cdot \bigg(\int_0^{R_0} r^{n-1} z_+^{m+4} dr \bigg)^\frac{1}{m+4} \\
& & + c_8 L^{m+1} + c_9 L^{m+3} \\
& & + c_{10} \cdot \Bigg\{ 1+\bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} \Bigg\} \\
& & + c_4 + \frac{1}{2c_1^3} \int_0^{R_0} r^{n-1} u_r^{m+4} dr \\
& & + c_5 L^{m+4} + \frac{c_6}{m+4} \int_0^{R_0} r^{n-1} u_r^{m+4} dr \\
& & + c_{11} \Bigg\{ 1+\bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} \Bigg\}
\qquad \mbox{for all } t\in (0,\tm).
\eas
Since $L\ge 1$, this means that if $m_\star$ is sufficiently large such that
\be{32.18}
\frac{c_6}{m_\star+4} \le \frac{1}{4c_1^3},
\ee
and if $m\ge m_\star$, then
\bea{32.19}
\frac{1}{4c_1^3} \int_0^{R_0} r^{n-1} u_r^{m+4} dr
&\le& c_7 \cdot \Bigg\{ 1+\bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} \Bigg\}
\cdot \bigg( \int_0^{R_0} r^{n-1} z_+^{m+4} dr \bigg)^\frac{1}{m+4} \nn\\
& & + c_{12} \bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr\bigg)^\frac{m+3}{m+4} \nn\\
& & + c_{13} L^{m+4}
\qquad \mbox{for all } t\in (0,\tm),
\eea
where $c_{12}:=c_{10}+c_{11}$ and $c_{13}:=c_8+c_9+c_{10}+c_4+c_5+c_{11}$.\abs
Now in order to prove (\ref{32.01}) for some suitably large $C>0$ independent of $t\in (0,\tm)$, we fix
any such $t$ and first consider the case when there exists a sequence of even numbers $m=m_j\ge m_\star, j\in\N$,
such that $m_j\to\infty$ as $j\to\infty$ and
\be{32.20}
\bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4}
\le L^{m+4}
\qquad \mbox{for all } m\in (m_j)_{j\in\N}.
\ee
Then taking $j\to\infty$ here, we directly obtain that
\be{32.21}
\|u_r(\cdot,t)\|_{L^\infty((0,R_0))}
= \lim_{j\to\infty} \bigg(\int_0^{R_0} r^{n-1} u_r^{m_j+4} (r,t) dr \bigg)^\frac{1}{m_j+4}
= \lim_{j\to\infty} L^\frac{m_j+4}{m_j+3} = L.
\ee
If conversely, such a sequence does not exist, then we can find $m_{\star\star} \ge m_\star$ such that for all
even $m\ge m_{\star\star}$,
\bas
\bigg(\int_0^{R_0} r^{n-1} u_r^{m+4} dr \bigg)^\frac{m+3}{m+4} > L^{m+4}.
\eas
Using that $L\ge 1$, we thus infer from (\ref{32.19}) that for any such $m$ we have
\bas
\frac{1}{4c_1^3} \int_0^{R_0} r^{n-1} u_r^{m+4}(r,t) dr
&<& 2c_7 \bigg(\int_0^{R_0} r^{n-1} u_r^{m+4}(r,t) dr\bigg)^\frac{m+3}{m+4} \cdot
\bigg(\int_0^{R_0} r^{n-1} z_+^{m+4} (r,t)dr \bigg)^\frac{1}{m+4} \\
& & + c_{12} \bigg(\int_0^{R_0} r^{n-1} u_r^{m+4}(r,t) dr\bigg)^\frac{m+3}{m+4}
+ c_{13} \bigg(\int_0^{R_0} r^{n-1} u_r^{m+4}(r,t) dr\bigg)^\frac{m+3}{m+4}
\eas
and hence
\bas
\frac{1}{4c_1^3} \bigg(\int_0^{R_0} r^{n-1} u_r^{m+4}(r,t) dr\bigg)^\frac{1}{m+4}
\le 2c_7 \bigg(\int_0^{R_0} r^{n-1} z_+^{m+4}(r,t) dr \bigg)^\frac{1}{m+4} + c_{12}+c_{13}.
\eas
In the limit $m\to\infty$, we therefore conclude that in this case,
\be{32.22}
\frac{1}{4c_1^3} \cdot \|u_r(\cdot,t)\|_{L^\infty((0,R_0))}
\le 2c_7 \cdot \|z_+(\cdot,t)\|_{L^\infty((0,R_0))} + c_{12}+c_{13}.
\ee
Since $c_7, c_{12}$ and $c_{13}$ as well as $L$ are independent of $t\in (0,\tm)$, (\ref{32.21}) and (\ref{32.22})
establish (\ref{32.01}).
\qed
\subsubsection{Estimating $|u_r|$ near the boundary}
For fixed $t\in (0,\tm)$, the above lemma in particular implies an upper bound for $u_r$ in terms of
$\|z_+\|_{L^\infty((0,t)\times (0,R_0))}$ on the lateral boundary line
$r=R_0$ of the parabolic cylinder $(R_0,R)\times (0,t)$.
This will enable us to apply a comparison argument to derive an estimate from above for $u_r$
in this region on the basis of (\ref{35.02}) to achieve the following.
\begin{lem}\label{lem37}
Assume that $\tm<\infty$, but that $\sup_{(r,t)\in (0,R)\times (0,\tm)} u(r,t)<\infty$.
Then with $R_0\in (0,R)$ taken from Lemma \ref{lem32}, we can find $C>0$ such that
\be{37.1}
\|u_r(\cdot,t)\|_{L^\infty((R_0,R))} \le C \cdot \Big(1+\|z_+\|_{L^\infty((0,R_0)\times (0,t))} \Big)
\qquad \mbox{for all } t\in (0,\tm).
\ee
\end{lem}
\proof
According to Lemma \ref{lem32}, we can pick $c_1>0$ such that
\bas
u_r(R_0,t) \le c_1 \cdot \Big(1+\|z_+(\cdot,t)\|_{L^\infty((0,R_0))} \Big)
\qquad \mbox{for all } t\in (0,\tm),
\eas
which in particular implies that given any $t_0\in (0,\tm)$ we have
\be{37.2}
u_r(R_0,t) \le D_1(t_0):= c_1 \cdot \Big( 1+\|z_+\|_{L^\infty((0,R_0)\times (0,t_0))} \Big)
\qquad \mbox{for all } t\in (0,t_0).
\ee
Let us next use our hypothesis and recall Lemma \ref{lem23} to pick $c_2>0$ and $c_3>0$ fulfilling
\be{37.3}
c_2 \le u(r,t) \le c_3
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\ee
and apply Lemma \ref{lem333} to find $c_4>0$ and $c_5>0$ such that
\be{37.4}
|v_r(r,t)| \le c_4 r
\quad \mbox{and} \quad
|v_{rr}(r,t)| \le c_5
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
Therefore, the coefficient functions $\tilde A_3$ and $\tilde A_4$ in (\ref{qarab1}) can be estimated according to
\be{37.5}
\tilde A_3(r,t)
\le c_6:=\frac{n-1}{R_0} + \frac{2\chi}{c_2^2} + \chi c_5 + \chi \cdot c_4^2 R^2 \cdot c_5
+ (n-1)\chi \cdot c_4^3 R^2
\ee
and
\bea{37.6}
\tilde A_4(r,t)
\le C_7 &:=& 3c_3 + \frac{n-1}{R_0^2} \cdot c_2
+ 3\chi\mu \cdot c_3 \cdot c_4 R \cdot c_5
+ 3\chi \cdot c_3^2 \cdot c_4 R \cdot c_5 \\
& & + (n-1)\chi \cdot c_3 \cdot c_4^3 R
+ 3(n-1) \cdot \cdot c_3 \cdot c_4^2 R \cdot c_5
\eea
for all $r\in (0,R)$ and $t\in (0,\tm)$.
We now fix $\alpha>0$ large such that
\be{37.7}
\alpha>c_6+\frac{c_7}{D}
\ee
and, given $t_0\in (0,\tm)$, define
\bas
\ophi(r,t):=D \, e^{\alpha t}
\qquad \mbox{for $r\in [R_0,R]$ and } t\in [0,t_0],
\eas
where
\be{37.8}
D:=\max \Big\{ D_1(t_0) \, , \sup_{r\in (R_0,R)} u_{0r}(r) \Big\} \, +1.
\ee
Then (\ref{37.2}) asserts that
\bas
u_r(R_0,t)\le \ophi(R_0,t)
\qquad \mbox{for all } t\in (0,t_0),
\eas
whereas clearly
\bas
u_r(R,t)=\ophi(R,t)=0
\qquad \mbox{for all } t\in (0,t_0)
\eas
and
\bas
u_r(r,0)=u_{0r}(r) < D =\ophi(r,0)
\qquad \mbox{for all } r\in [R_0,R].
\eas
Moreover, since $\ophi$ is positive and $\ophi_r=\ophi_{rr}\equiv 0$, we may use (\ref{37.5}), (\ref{37.6}) and (\ref{37.7})
to see that with $\qarab$ as in (\ref{qarab}) we have
\bas
\qarab \ophi
&=& \ophi_t - \tilde A_3(r,t) \ophi - \tilde A_4(r,t) \\
&=& \alpha D e^{\alpha t} - \tilde A_3(r,t) \cdot D e^{\alpha t} - \tilde A_4(r,t) \\
&\ge& (\alpha - c_6) \cdot D e^{\alpha t} - c_7 \\
&\ge& (\alpha - c_6) D - c_7 \\
&>& 0
\qquad \mbox{for all $r\in (R_0,R)$ and } t\in (0,t_0).
\eas
As $\qarab u_r \equiv 0$ due to Lemma \ref{lem35}, by comparison we conclude that $u_r \le \ophi$ in
$(R_0,R)\times (0,t_0)$, which in view of (\ref{37.8}) and (\ref{37.2}) readily entails (\ref{37.1}).
\qed
\subsubsection{A bound for $|u_r|$ in the entire domain}
Let us summarize the outcome of Lemma \ref{lem32} and Lemma \ref{lem37}:
\begin{cor}\label{cor40}
If $\tm<\infty$ but $\sup_{(r,t)\in (0,R)\times (0,\tm)} u(r,t)<\infty$,
then there exists $C>0$ such that
\bas
\|u_r(\cdot,t)\|_{L^\infty((0,R))} \le C \cdot \Big(1+\|z_+\|_{L^\infty((0,R)\times (0,t))} \Big)
\qquad \mbox{for all } t\in (0,\tm).
\eas
\end{cor}
\proof
We only need to combine Lemma \ref{lem32} with Lemma \ref{lem37}.
\qed
\subsection{A nonlocal parabolic inequality for $z$}
In order to bound $z$ from above, let us first identify a linear inhomogeneous parabolic equation satisfied
by this function.
\begin{lem}\label{lem38}
The function $z=\frac{u_t}{u}$ satisfies
\be{38.1}
z_t=B_1(r,t)z_{rr} + B_{21}(r,t) z_r + \frac{B_{22}(r,t)}{r} z_r + B_3(r,t)z + B_4(r,t)
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\ee
where
\be{38.2}
\left\{ \begin{array}{l}
\displaystyle B_1(r,t):=\frac{u^3}{\sqrt{u^2+u_r^2}^3}, \\[2mm]
\displaystyle B_{21}(r,t):=2\frac{u^2 u_r}{\sqrt{u^2+u_r^2}^3} - 3\frac{u^3 u_r u_{rr}}{\sqrt{u^2+u_r^2}^5}
+ 4\frac{u_r^3}{\sqrt{u^2+u_r^2}^3} - 3\frac{u_r^5}{\sqrt{u^2+u_r^2}^5} -\chi \frac{v_r}{\sqrt{1+v_r^2}}, \\[2mm]
\displaystyle B_{22}(r,t):=(n-1)\frac{u^3}{\sqrt{u^2+u_r^2}^3}, \\[2mm]
\displaystyle B_3(r,t):=\chi \frac{u}{\sqrt{1+v_r^2}^3} \qquad \qquad \mbox{and} \\[2mm]
\displaystyle B_4(r,t):=-3\chi \frac{u(\mu-u) u_r v_r}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^5}
+ 3\chi^2 \frac{u(\mu-u)v_r^2}{(1+v_r^2)^3} +\chi \frac{u_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^3}
-\chi^2 \frac{u_r v_r}{(1+v_r^2)^2} \\[1mm]
\hspace*{19mm}
+3\chi \cdot \frac{n-1}{r} \cdot \frac{uu_r v_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^5}
- 3\chi^2 \cdot \frac{n-1}{r} \cdot \frac{uv_r^3}{(1+v_r^2)^3}
\end{array} \right.
\ee
for $r\in (0,R)$ and $t\in (0,\tm)$.
\end{lem}
\proof
We divide (\ref{36.1}) by $u$ and differentiate each term on the right-hand side of the resulting identity separately.
Using that $u_t=uz$ and hence $u_{rt}=uz_r+u_rz$ and $u_{rrt}=uz_{rr}+2u_r z_r + u_{rr}z$, we first obtain
\bas
\bigg(\frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^3}\bigg)_t
&=& \frac{u^2 u_{rrt}}{\sqrt{u^2+u_r^2}^3}
+ 2\frac{uu_t u_{rr}}{\sqrt{u^2+u_r^2}^3}
- \frac{3}{2} \cdot \frac{u^2 u_{rr} \cdot (2uu_t + 2u_r u_{rt})}{\sqrt{u^2+u_r^2}^5} \\
&=& \frac{u^3}{\sqrt{u^2+u_r^2}^3} \cdot z_{rr}
+ 2\frac{u^2 u_r}{\sqrt{u^2+u_r^2}^3} \cdot z_r
+ \frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^3} \cdot z
+ 2\frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^3} \cdot z \\
& & - 3\frac{u^4 u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot z
- 3\frac{u^3 u_r u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot z_r
- 3\frac{u^2 u_r^2 u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot z.
\eas
Since
\bas
& & \hspace*{-30mm}
\frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^3} \cdot z
+ 2\frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^3} \cdot z
- 3\frac{u^4 u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot z
- 3\frac{u^2 u_r^2 u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot z \\
&=& \frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^5} \cdot \Big\{ (u^2+u_r^2)+2(u^2+u_r^2)-3u^2-3u_r^2\Big\} \\[2mm]
&=& 0,
\eas
this yields
\be{38.3}
\bigg(\frac{u^2 u_{rr}}{\sqrt{u^2+u_r^2}^3}\bigg)_t
= \frac{u^3}{\sqrt{u^2+u_r^2}^3} \cdot z_{rr}
+ \bigg\{ 2\frac{u^2 u_r}{\sqrt{u^2+u_r^2}^3} - 3\frac{u^3 u_r u_{rr}}{\sqrt{u^2+u_r^2}^5} \bigg\} \cdot z_r
\ee
for all $r\in (0,R)$ and $t\in (0,\tm)$.\\
Next,
\bas
\bigg( \frac{u_r^4}{u\sqrt{u^2+u_r^2}^3}\bigg)_t
&=& 4\frac{u_r^3 u_{rt}}{u\sqrt{u^2+u_r^2}^3}
- \frac{u_r^4 u_t}{u^2 \sqrt{u^2+u_r^2}^3}
-\frac{3}{2} \cdot \frac{u_r^4 \cdot (2uu_t+2u_r u_{rt})}{u\sqrt{u^2+u_r^2}^5} \\
&=& 4\frac{u_r^3}{\sqrt{u^2+u_r^2}^3} \cdot z_r
+ 4\frac{u_r^4}{u\sqrt{u^2+u_r^2}^3} \cdot z
- \frac{u_r^4}{u\sqrt{u^2+u_r^2}^3} \cdot z \\
& & -3\frac{uu_r^4}{\sqrt{u^2+u_r^2}^5} \cdot z
- 3\frac{u_r^5}{\sqrt{u^2+u_r^2}^5} \cdot z_r
- 3\frac{u_r^6}{u\sqrt{u^2+u_r^2}^5} \cdot z,
\eas
where again the zero-order terms have a vanishing sum in the sense that
\bas
& & \hspace*{-30mm}
4\frac{u_r^4}{u\sqrt{u^2+u_r^2}^3} \cdot z
- \frac{u_r^4}{u\sqrt{u^2+u_r^2}^3} \cdot z
-3\frac{uu_r^4}{\sqrt{u^2+u_r^2}^5} \cdot z
- 3\frac{u_r^6}{u\sqrt{u^2+u_r^2}^5} \cdot z \\
&=& \frac{u_r^4}{u\sqrt{u^2+u_r^2}^5} \cdot \Big\{ 4(u^2+u_r^2) - (u^2+u_r^2) - 3u^2 - 3u_r^2 \Big\} \cdot z \\[2mm]
&=& 0,
\eas
so that
\be{38.4}
\bigg(\frac{u_r^4}{u\sqrt{u^2+u_r^2}^3}\bigg)_t
= \bigg\{ 4\frac{u_r^3}{\sqrt{u^2+u_r^2}^3} - 3\frac{u_r^5}{\sqrt{u^2+u_r^2}^5} \bigg\} \cdot z_r
\ee
for all $r\in (0,R)$ and $t\in (0,\tm)$.\\
Likewise,
\bas
\bigg(\frac{u_r}{\sqrt{u^2+u_r^2}}\bigg)_t
&=& \frac{u_{rt}}{\sqrt{u^2+u_r^2}} - \frac{1}{2} \cdot \frac{u_r\cdot (2uu_t+2u_r u_{rt})}{\sqrt{u^2+u_r^2}^3} \\
&=& \frac{u}{\sqrt{u^2+u_r^2}} \cdot z_r + \frac{u_r}{\sqrt{u^2+u_r^2}} \cdot z
- \frac{u^2 u_r}{\sqrt{u^2+u_r^2}^3} \cdot z
- \frac{uu_r^2}{\sqrt{u^2+u_r^2}^3} \cdot z_r
- \frac{u_r^3}{\sqrt{u^2+u_r^2}^3} \cdot z \\
&=& \frac{u}{\sqrt{u^2+u_r^2}^3} \cdot \Big\{ (u^2+u_r^2)-u_r^2\Big\} \cdot z_r \\
& & + \frac{u_r}{\sqrt{u^2+u_r^2}^3} \cdot \Big\{(u^2+u_r^2)-u^2-u_r^2\Big\} \cdot z \\
&=& \frac{u^3}{\sqrt{u^2+u_r^2}^3} \cdot z_r,
\eas
whence
\be{38.5}
\bigg(\frac{n-1}{r} \frac{u_r}{\sqrt{u^2+u_r^2}} \bigg)_t
= \frac{n-1}{r} \cdot \frac{u^3}{\sqrt{u^2+u_r^2}^3} \cdot z_r
\ee
for $r\in (0,R)$ and $t\in (0,\tm)$.\\
As for the respective terms originating from the rightmost three summands in (\ref{36.1}), we make use of (\ref{33.3})
to express $v_{rt}$ conveniently. We thereby compute
\bea{38.6}
\bigg(-\chi \frac{\mu-u}{\sqrt{1+v_r^2}^3}\bigg)_t
&=& \chi\frac{u_t}{\sqrt{1+v_r^2}^3} + \frac{3}{2} \chi \frac{\mu-u}{\sqrt{1+v_r^2}^5} \cdot 2v_r v_{rt} \nn\\
&=& \chi\frac{u}{\sqrt{1+v_r^2}^3} \cdot z
+ 3\chi \frac{(\mu-u)v_r}{\sqrt{1+v_r^2}^5} \cdot
\Big\{ -\frac{uu_r}{\sqrt{u^2+u_r^2}} + \chi \frac{uv_r}{\sqrt{1+v_r^2}} \Big\} \nn\\
&=& \chi\frac{u}{\sqrt{1+v_r^2}^3} \cdot z
- 3\chi \frac{u(\mu-u)u_r v_r}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^5}
+ 3\chi^2 \frac{u(\mu-u)v_r^2}{(1+v_r^2)^3}
\eea
and
\bea{38.7}
\bigg( -\chi \frac{u_r v_r}{u\sqrt{1+v_r^2}}\bigg)_t
&=& - \chi \frac{u_{rt} v_r}{u\sqrt{1+v_r^2}} - \chi \frac{u_r v_{rt}}{u\sqrt{1+v_r^2}}
+ \chi \frac{u_r v_r u_t}{u^2 \sqrt{1+v_r^2}}
+ \frac{1}{2} \chi \frac{u_r v_r}{u\sqrt{1+v_r^2}^3} \cdot 2v_r v_{rt} \nn\\
&=& - \chi \frac{v_r}{\sqrt{1+v_r^2}} \cdot z_r
- \chi \frac{u_r v_r}{u\sqrt{1+v_r^2}} \cdot z \nn\\
& & - \chi \frac{u_r}{u\sqrt{1+v_r^2}} \cdot
\Big\{ -\frac{uu_r}{\sqrt{u^2+u_r^2}} + \chi \frac{uv_r}{\sqrt{1+v_r^2}} \Big\} \nn\\
& & + \chi \frac{u_r v_r}{u\sqrt{1+v_r^2}} \cdot z
+ \chi \frac{u_r v_r^2}{u\sqrt{1+v_r^2}^3} \cdot
\Big\{ -\frac{uu_r}{\sqrt{u^2+u_r^2}} + \chi \frac{uv_r}{\sqrt{1+v_r^2}} \Big\} \nn\\
&=& - \chi \frac{v_r}{\sqrt{1+v_r^2}} \cdot z_r
+ \chi \frac{u_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}}
-\chi^2 \frac{u_r v_r}{(1+v_r^2)^2} \nn\\
&=& - \chi \frac{v_r}{\sqrt{1+v_r^2}} \cdot z_r
+ \chi \frac{u_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^3}
-\chi^2 \frac{u_r v_r}{(1+v_r^2)^2},
\eea
and observe that
\bas
\bigg(\frac{v_r^3}{\sqrt{1+v_r^2}^3}\bigg)_t
&=& \Big\{ \frac{3v_r^2}{\sqrt{1+v_r^2}^3} - \frac{3}{2}
\cdot \frac{v_r^3 \cdot 2v_r}{\sqrt{1+v_r^2}^5} \Big\} \cdot v_{rt} \nn\\
&=& - 3\frac{uu_r v_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^5}
+3\chi \frac{uv_r^3}{(1+v_r^2)^3}
\eas
to obtain
\be{38.8}
\bigg( -\chi \cdot \frac{n-1}{r} \cdot \frac{v_r^3}{\sqrt{1+v_r^2}^3} \bigg)_t
= 3\chi\cdot \frac{n-1}{r} \cdot \frac{uu_r v_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^5}
-3\chi^2 \cdot \frac{n-1}{r} \cdot \frac{uv_r^3}{(1+v_r^2)^3}
\ee
for $r\in (0,R)$ and $t\in (0,\tm)$.
In light of (\ref{38.3})-(\ref{38.8}), (\ref{36.1}) easily yields (\ref{38.1}) with $B_1, B_{21}, B_{22}, B_3$
and $B_4$ as in (\ref{38.2}).
\qed
On suitably estimating the inhomogeneous term $B_4$ herein by means of Corollary \ref{cor40},
we can develop (\ref{38.1}) into a nonlocal parabolic inequality for $z$ as follows.
\begin{lem}\label{lem39}
Suppose that $\tm<\infty$, but that $\sup_{(r,t)\in (0,R)\times (0,\tm)} u(r,t)<\infty$.
Then there exist a constant $d>0$ and continuous functions $b_1, b_{21}, b_{22}$ and $b_3$ on $[0,R]\times [0,\tm)$
with the properties that $b_1$ and $b_{22}$ are nonnegative, that $b_3$ is bounded on $(0,R)\times (0,\tm)$, and such
that $z=\frac{u_t}{u}$ satisfies
\be{39.1}
z_t(r,t) \le b_1(r,t) z_{rr} + b_{21}(r,t) z_r + \frac{b_{22}(r,t)}{r} z_r + b_3(r,t)z
+ d \cdot \Big( 1+\|z_+\|_{L^\infty((0,R)\times (0,t))} \Big)
\ee
for all $r\in (0,R)$ and $t\in (0,\tm)$.
\end{lem}
\proof
With $B_1, B_{21}, B_{22}, B_3$ and $B_4$ taken from Lemma \ref{lem38}, we let $b_1:=B_1, b_{21}:=B_{21}, b_{22}:=B_{22}$
and $b_3:=B_3$. Then from (\ref{38.2}) we immediately obtain that $b_1, b_{21}, b_{22}$ and $b_3$ are continuous in
$[0,R] \times [0,\tm)$, and that $b_1 \ge 0$ and $b_{22} \ge 0$. Since our boundedness assumption on $u$ ensures that $b_3$
is bounded, it remains to control the inhomogeneity $B_4$ in (\ref{38.1}) adequately. To this end, we once more
use our hypothesis along with Lemma \ref{lem333} to pick positive constants $c_1, c_2$ and $c_3$ such that
\be{39.2}
u(r,t) \le c_1, \quad
|v_r(r,t)| \le c_2 r
\quad \mbox{and} \quad
|v_{rr}(r,t)| \le c_3
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
Then in (\ref{38.2}) we can estimate
\bas
-3\chi \frac{u(\mu-u) u_r v_r}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^5}
&\le& 3\chi \cdot c_1 (\mu+c_1) \cdot c_2 R \cdot \frac{|u_r|}{\sqrt{u^2+u_r^2}} \\
&\le& 3c_1(\mu+c_1)c_2 \chi R
\eas
and
\bas
3\chi^2 \frac{u(\mu-u) v_r v_{rr}}{(1+v_r^2)^3}
&\le& 3\chi^2 c_1 (\mu+c_1) \cdot c_2 R \cdot c_3 \\
&\le& 3c_1(\mu+c_1)c_2 c_3 \chi^2 R
\eas
as well as
\bas
3\chi \cdot \frac{n-1}{r} \cdot \frac{uu_r v_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^5}
&\le& 3\chi \cdot \frac{n-1}{r} \cdot c_1 \cdot c_2^2 r^2 \cdot \frac{|u_r|}{\sqrt{u^2+u_r^2}} \\
&\le& 3(n-1) c_1 c_2^2 \chi R
\eas
and
\bas
-3\chi^2 \cdot \frac{n-1}{r} \cdot \frac{uv_r^3}{(1+v_r^2)^3}
&\le& 3\chi^2 \cdot \frac{n-1}{r} \cdot c_1 \cdot c_2^3 r^3 \\
&\le& 3(n-1) c_1 c_2^3 \chi^2 R^2
\eas
for all $r\in (0,R)$ and $t\in (0,\tm)$.
In the third and fourth summands in the definition (\ref{38.2}) of $B_4$, however, apparently we can only estimate
\bas
\chi \frac{u_r^2}{\sqrt{u^2+u_r^2} \cdot \sqrt{1+v_r^2}^3}
\le \chi \frac{u_r^2}{\sqrt{u^2+u_r^2}} \le \chi |u_r|
\eas
and
\bas
- \chi^2 \frac{u_r v_r}{(1+v_r^2)^2} \le \chi^2 |u_r|
\eas
for all $r\in (0,R)$ and $t\in (0,\tm)$, with the possibly unbounded factors $|u_r|$ remaining.
Fortunately, applying Corollary \ref{cor40} yields $c_4>0$ such that
\bas
|u_r(r,t)| \le c_4 \cdot \Big(1+\|z_+\|_{L^\infty((0,R)\times (0,t))} \Big)
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\eas
Therefore, (\ref{39.1}) results from (\ref{38.1}) if we choose $d>0$ conveniently large.
\qed
\subsection{Boundedness of $z$ from above}
Apparently, nonlocal parabolic inequalities of type (\ref{39.1}) do not allow for general comparison priciples.
After all, the fact that here the memory term enjoys a certain linear boundedness property with respect to $z_+$
enables us to follow a maximum principle-type reasoning to establish an essentially exponential upper bound for $z$
and thereby obtain the following.
\begin{lem}\label{lem41}
Assume that $\tm<\infty$ and $\sup_{(r,t)\in (0,R)\times (0,\tm)} u(r,t)<\infty$.
Then there exists $C>0$ such that $z=\frac{u_t}{u}$ satisfies
\be{41.1}
z(r,t) \le C
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
\end{lem}
\proof
We let $b_1, b_{21}, b_{22}, b_3$ and $d$ be as provided by Lemma \ref{lem39}, so that by boundedness of $b_3$ we can
find $c_1>0$ such that
\be{41.2}
b_3(r,t) \le c_1
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm).
\ee
We than fix $\alpha>0$ large enough fulfilling
\be{41.3}
\alpha>c_1+d
\ee
and let
\bas
\varphi(r,t):=e^{-\alpha t} z(r,t) - dt
\qquad \mbox{for $r\in [0,R]$ and } t\in [0,\tm).
\eas
Then according to Lemma \ref{lem39},
\bea{41.4}
\varphi_t
&=& e^{-\alpha t} (z_t-\alpha z) -d \nn\\
&\le& e^{-\alpha t} \cdot \Big\{ b_1(r,t) z_{rr} + b_{21}(r,t) z_r + \frac{b_{22}}{r} z_r
+ b_3(r,t)z + d\|z_+\|_{L^\infty((0,R)\times (0,t))} + d - \alpha z \Big\} \, -d \nn\\
&=& b_1(r,t) \varphi_{rr} + b_{21}(r,t)\varphi_r + \frac{b_{22}(r,t)}{r} \varphi_r
+ \Big(b_3(r,t)-\alpha\Big) \cdot \Big(\varphi+dt \Big) \nn\\[1mm]
& & + d e^{-\alpha t} \|z_+\|_{L^\infty((0,R)\times (0,t))}
+ d e^{-\alpha t} - d
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\eea
and since
$z_r=\big(\frac{u_t}{u}\big)_r = \frac{u_{rt}}{u}-\frac{u_r u_t}{u^2}$ in $[0,R] \times [0,\tm)$,
the fact that $u_r(0,t)=u_r(R,t)=0$ for all $t\in (0,\tm)$ entails that
\be{41.5}
\varphi_r(0,t)=\varphi_r(R,t)=0
\qquad \mbox{for all } t\in (0,\tm).
\ee
Now if for some $T\in (0,\tm)$, the value $S:=\sup_{(r,t)\in (0,R)\times (0,T)} \varphi(r,t)$ was positive
and attained at some point $(r_0,t_0)\in [0,R]\times [0,T]$ with $t_0>0$, then necessarily
\be{41.6}
\varphi_t(r_0,t_0) \ge 0,
\ee
and (\ref{41.5}) ensures that in both cases $r_0\in (0,R)$ and $r_0\in \{0,R\}$ we moreover must have
\be{41.7}
\varphi_r(r_0,t_0)=0
\qquad \mbox{and} \varphi_{rr}(r_0,t_0) \le 0.
\ee
We claim that these properties imply that
\be{41.8}
0 \le \Big(b_3(r_0,t_0)-\alpha\Big) \cdot \Big( \varphi(r_0,t_0)+dt_0\Big)
+ d e^{-\alpha t_0} \|z_+\|_{L^\infty((0,R)\times (0,t_0))} + d e^{-\alpha t_0} - d.
\ee
Indeed, in the case $r_0\in (0,R)$ we may directly apply (\ref{41.4}) to easily deduce this from (\ref{41.6})
and (\ref{41.7}).
When $r_0=R$, by continuity of $\varphi, \varphi_t, \varphi_r$ and $\varphi_{rr}$ in $[0,R]\times (0,\tm)$
it is clear that (\ref{41.4}) actually remains valid at $(r_0,t_0)$, so that (\ref{41.8}) follows from the same argument.
If $r_0=0$, however, we make use of the favorable sign of the singular term $\frac{b_{22}}{r}$ in (\ref{41.4})
by first
choosing, once more relying on the extremal property of $\varphi(r_0,t_0)$, a sequence $(r_j)_{j\in\N} \subset (0,R)$
such that $r_j\searrow 0$ as $j\to\infty$ and
\bas
\varphi_r(r_j,t_0) \le 0
\qquad \mbox{for all } j\in\N,
\eas
and then evaluating (\ref{41.4}) at $r=r_j$ to see that
\bas
\varphi_t(r_j,t_0)
&\le& b_1(r_j,t_0) \varphi_{rr}(r_j,t_0) + b_{21}(r_j,t_0) \varphi_r(r_j,t_0)
+ \Big(b_3(r_j,t_0)-\alpha \Big) \cdot \Big( \varphi (r_j,t_0) + dt_0 \Big) \\
& & + d e^{-\alpha t_0} \|z_+\|_{((0,R) \times (0,t_0))}
+ d e^{-\alpha t_0} -d
\eas
for all $j\in\N$. Again by continuity of $\varphi,\varphi_t, \varphi_r$ and $\varphi_{rr}$, we may take $j\to\infty$
to conclude that
\bas
\varphi_t(0,t_0)
&\le& b_1(0,t_0) \varphi_{rr}(0,t_0) + b_{21}(0,t_0) \varphi_r(0,t_0)
+ \Big( b_3(0,t_0) -\alpha \Big) \cdot \Big( \varphi(0,t_0) + dt_0\Big) \\
& & + d e^{-\alpha t_0} \|z_+\|_{((0,R)\times (0,t_0))}
+ d e^{-\alpha t_0} - d,
\eas
whereupon one more application of (\ref{41.6}) and (\ref{41.7}) yields (\ref{41.8}) also in this case.\abs
Now observing that $e^{-\alpha t_0} \le 1$ and using that $S=\varphi(r_0,t_0)$ is positive, in view of (\ref{41.2})
we obtain from (\ref{41.8}) that
\be{41.9}
0 \le (c_1-\alpha) \cdot \Big( \varphi(r_0,t_0)+dt_0\Big)
+ d e^{-\alpha t_0} \|z_+\|_{L^\infty((0,R)\times (0,t_0))}.
\ee
Here we rewrite $z= e^{\alpha t} \varphi + dt e^{\alpha t}$ and use that if $f$ and $g$ are two functions on a set
$D\subset \R^N$, $N\ge 1$, then both $(f+g)_+$ and $\sup_D \{f+g\} \le \sup_D f + \sup_D g$. We thereby obtain that
\bas
d e^{-\alpha t_0} \|z_+\|_{L^\infty((0,R)\times (0,t_0))}
&=& d e^{-\alpha t_0} \cdot
\sup_{(r,s)\in (0,R)\times (0,t_0)} \Big\{ e^{\alpha s}\varphi(r,s) + ds e^{\alpha s} \Big\} \\
&\le& d e^{\alpha t_0} \cdot
\sup_{(r,s)\in (0,R)\times (0,t_0)} \Big\{ e^{\alpha s}\varphi_+(r,s) + ds e^{\alpha s} \Big\} \\
&\le& d e^{\alpha t_0} \cdot \bigg\{
\sup_{(r,s)\in (0,R)\times (0,t_0)} \Big\{ e^{\alpha s}\varphi_+(r,s) \Big\}
+ \sup_{s \in (0,t_0)} \Big\{ ds e^{\alpha s} \Big\} \bigg\}\\
&\le& d e^{\alpha t_0} \cdot
\sup_{(r,s)\in (0,R)\times (0,t_0)} \Big\{ e^{\alpha s}\varphi_+(r,s) \Big\}
+ d^2 t_0.
\eas
Since from the definition of $S$ we know that
\bas
\sup_{(r,s)\in (0,R)\times (0,t_0)} \Big\{ e^{\alpha s}\varphi_+(r,s) \Big\}
\le e^{\alpha t_0} \cdot
\sup_{(r,s)\in (0,R)\times (0,t_0)} \varphi_+(r,s)
= e^{\alpha t_0} \cdot \varphi(r_0,t_0),
\eas
this entails that
\bas
d e^{-\alpha t_0} \|z_+\|_{L^\infty((0,R)\times (0,t_0))}
\le d e^{-\alpha t_0} \|z_+\|_{L^\infty((0,R)\times (0,t_0))}
\le d \varphi(r_0,t_0) + d^2 t_0,
\eas
so that (\ref{41.9}) yields the inequality
\bas
0 &\le& (c_1-\alpha)\cdot \Big( \varphi(r_0,t_0)+dt_0 \Big) + d\varphi(r_0,t_0) + dt_0 \\
&=& (c_1-\alpha+d) \cdot \Big(\varphi(r_0,t_0)+dt_0\Big).
\eas
In light of our restriction (\ref{41.3}) on $\alpha$, however, this contradicts the positivity of $\varphi(r_0,t_0)$
and thereby proves that actually $\varphi$ cannot attain a positive maximum over any such region $[0,R]\times [0,T]$,
$T\in (0,\tm)$, at a positive time $t_0$. This means that in fact
\bas
\varphi(r,t) \le \|\varphi_+(\cdot,0)\|_{L^\infty((0,R))} = \|z_+(\cdot,0)\|_{L^\infty((0,R))}
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm)
\eas
and hence
\bas
z(r,t) &=& e^{\alpha t} \Big(\varphi(r,t) + dt \Big) \\
&\le& e^{\alpha \tm} \cdot \Big\{ \|z_+(\cdot,0)\|_{L^\infty((0,R))} + d\tm\Big\}
\qquad \mbox{for all $r\in (0,R)$ and } t\in (0,\tm),
\eas
which establishes (\ref{41.1}).
\qed
\subsection{Boundedness of $u$ implies extensibility. Proof of Theorem \ref{theo43}}
Combining the latter lemma with Corollary \ref{cor40} now directly yields the desired bound for $u_r$.
\begin{cor}\label{cor42}
If $\tm<\infty$ but $\sup_{(r,t)\in (0,R)\times (0,\tm)} u(r,t)<\infty$,
then there exists $C>0$ such that
\bas
\|u_r(\cdot,t)\|_{L^\infty((0,R))} \le C
\qquad \mbox{for all } t\in (0,\tm).
\eas
\end{cor}
\proof
Thanks to the upper estimate for $z$ obtained in Lemma \ref{lem41}, this is an immediate consequence of Corollary
\ref{cor40}.
\qed
We can thereby readily verify our main statement on local existence and extensibility.\abs
\proofc of Theorem \ref{theo43}.\quad
In view of the local existence result established in Lemma \ref{lem21}, we only need to verify (\ref{43.1}).
Indeed, if (\ref{43.1}) was false, then for some solution the respective maximal
existence time would satisfy $\tm<\infty$ but $\limsup_{t\nearrow\tm} \|u(\cdot,t)\|_{L^\infty(\Omega)}<\infty$.
Then, however, Corollary \ref{cor42} would apply to assert that also
$\limsup_{t\nearrow\tm} \|\nabla u(\cdot,t)\|_{L^\infty(\Omega)}$ would be finite.
Along with the lower bound for $u$ provided by Lemma \ref{lem23}, this would contradict the extensibility criterion
(\ref{ext_crit}) in Lemma \ref{lem21}.
\qed
\mysection{Boundedness for small $\chi$. Proof of Theorem \ref{theo55}}\label{sect6}
In light of the extensibility criterion provided by Theorem \ref{theo43}, in order to prove both global existence and
boundedness of a solution it is sufficient to derive an a priori estimate for $(u,v)$ in $(L^\infty(\Omega\times (0,T)))^2$
which does not explicitly depend on $T<\tm \le \infty$.
As a preparation for the proof of this in Theorem \ref{theo55} below, let us state the following elementary inequality.
\begin{lem}\label{lem51}
Let $p\ge 1$. Then
\be{51.1}
\io u^{p-1}|\nabla u| \le \io \frac{u^{p-1}|\nabla u|^2}{\sqrt{u^2+|\nabla u|^2}}
+ \io u^p
\qquad \mbox{for all } t\in (0,\tm).
\ee
\end{lem}
\proof
By means of Young's inequality, we see that
\be{51.2}
\io u^{p-1}|\nabla u|
\le \frac{1}{2} \io \frac{u^{p-1}|\nabla u|^2}{\sqrt{u^2+|\nabla u|^2}}
+ \frac{1}{2} \io u^{p-1}\sqrt{u^2+|\nabla u|^2}
\qquad \mbox{for all } t\in (0,\tm),
\ee
where using the elementary inequality $\sqrt{X+Y} \le \sqrt{X}+\sqrt{Y}$, valid for all $X\ge 0$ and $Y\ge 0$,
we can estimate
\bas
\frac{1}{2} \io u^{p-1}\sqrt{u^2+|\nabla u|^2}
\le \frac{1}{2} \io u^p + \frac{1}{2} \io u^{p-1}|\nabla u|.
\eas
Therefore, (\ref{51.1}) results from (\ref{51.2}).
\qed
We are now in the position to make sure that if either $n\ge 2, \chi<1$ and $u_0$ is an arbitrary function satisfying
(\ref{init}), or $n=1$, $\chi>0$ and $\io u_0<m_c$ with $m_c$ as in (\ref{mc}),
then the solution of (\ref{0}) in fact is global and remains bounded:\abs
\proofc of Theorem \ref{theo55}.\quad
We let $p_k:=2^k$ and, given $T\in (0,\tm)$, introduce
\be{55.22}
M_k:=\sup_{t\in (0,T)} \io u^{p_k}(x,t)dx
\ee
for nonnegative integers $k$.
Then clearly $M_k$ is well-defined for any such $T$ and $k$, and in order to control $M_k$ appropriately,
we fix $k\ge 1$ and multiply the first equation in (\ref{0}) by $pu^{p-1}$ for $p:=p_k$ to see upon integrating
by parts that
\bea{55.2}
\frac{d}{dt} \io u^p
+ p(p-1) \io \frac{u^{p-1}|\nabla u|^2}{\sqrt{u^2+|\nabla u|^2}}
&=& p(p-1)\chi \io \frac{u^{p-1}\nabla u\cdot \nabla v}{\sqrt{1+|\nabla v|^2}} \nn\\
&\le& p(p-1)\chi \io u^{p-1}|\nabla u| \cdot \frac{|\nabla v|}{\sqrt{1+|\nabla v|^2}}
\eea
for all $t\in (0,\tm)$.
Here in the multi-dimensional case, in which no evident uniform a priori bound for $|\nabla v|$ seems available,
we use the trivial pointwise inequality $\frac{|\nabla v|}{\sqrt{1+|\nabla v|^2}} \le 1$ to obtain
\be{55.211}
\frac{d}{dt} \io u^p
+ p(p-1) \io \frac{u^{p-1}|\nabla u|^2}{\sqrt{u^2+|\nabla u|^2}}
\le p(p-1)\chi \io u^{p-1}|\nabla u|
\quad \mbox{for all } t\in (0,\tm)
\qquad \mbox{if } n\ge 2.
\ee
In the one-dimensional setting, however, from (\ref{333.0}) and (\ref{mu}) we know that $|\nabla v|=|v_r| \le m$ throughout
$\Omega\times (0,\tm)$, whence by monotonicity of $0 \le \xi \mapsto \frac{\xi}{\sqrt{1+\xi^2}}$ we infer from
(\ref{55.2}) that
\be{55.212}
\frac{d}{dt} \io u^p
+ p(p-1) \io \frac{u^{p-1}|\nabla u|^2}{\sqrt{u^2+|\nabla u|^2}}
\le p(p-1)\chi \cdot \frac{m}{\sqrt{1+m^2}} \cdot \io u^{p-1}|\nabla u|
\quad \mbox{for all } t\in (0,\tm)
\quad \mbox{if } n=1.
\ee
In both (\ref{55.211}) and (\ref{55.212}) we now apply Lemma \ref{lem51} to estimate
\bas
p(p-1) \io \frac{u^{p-1}|\nabla u|^2}{\sqrt{u^2+|\nabla u|^2}}
\ge p(p-1) \io u^{p-1}|\nabla u| - p(p-1) \io u^p
\eas
and thus obtain on writing
\be{Lambda}
\Lambda:=\left\{ \begin{array}{ll}
\frac{m}{\sqrt{1+m^2}} \qquad & \mbox{if } n=1, \\[1mm]
1 & \mbox{if } n\ge 2,
\end{array} \right.
\ee
and adding $\io u^p$ on both sides of (\ref{55.211}) and (\ref{55.212}), respectively, that
\bea{55.3}
\frac{d}{dt} \io u^p + \io u^p + p(p-1)(1-\chi \Lambda) \io u^{p-1}|\nabla u|
&\le& \Big\{p(p-1)+1\Big\} \cdot \io u^p \nn\\
&\le& p^2 \io u^p
\qquad \mbox{for all } t\in (0,\tm).
\eea
We next invoke the Gagliardo-Nirenberg inequality \cite{win_critexp}
to find $c_1>0$ such that with $a:=\frac{n}{n+1}$ we have
\bas
\|\varphi\|_{L^1(\Omega)}
\le c_1 \|\nabla\varphi\|_{L^1(\Omega)}^a \|\varphi\|_{L^\frac{1}{2}(\Omega)}^{1-a}
+ c_1\|\varphi\|_{L^\frac{1}{2}(\Omega)}
\qquad \mbox{for all } \varphi\in W^{1,1}(\Omega),
\eas
and thereby obtain that
\bas
\io u^p \le c_1 \Big( \io |\nabla u^p| \Big)^a \cdot \Big(\io u^\frac{p}{2}\Big)^{2(1-a)}
\qquad \mbox{for all } t\in (0,\tm).
\eas
Since our specification of $p=p_k=2^k$ allows us to use the definition (\ref{55.22}) of $M_{k-1}$ in estimating
\bas
\io u^\frac{p}{2} \le M_{k-1}
\qquad \mbox{for all } t\in (0,T),
\eas
this implies that
\bas
\io u^p \le c_1 \Big( \io |\nabla u^p|\Big)^a \cdot M_{k-1}^{2(1-a)} + c_1 M_{k-1}^2
\qquad \mbox{for all } t\in (0,T).
\eas
Thanks to the fact that our assumptions ensure that $\chi\Lambda<1$, another application of Young's inequality therefore
provides $c_2>0$ fulfilling
\bas
p^2 \io u^p
\le (p-1)(1-\chi\Lambda) \io |\nabla u^p|
+ c_2 p^\frac{2}{1-a} M_{k-1}^2 + c_1 p^2 M_{k-1}^2
\qquad \mbox{for all } t\in (0,T),
\eas
from which due to the evident fact that $p^2 \le p^\frac{2}{1-a}$ we obtain that
\bas
p^2 \io u^p
\le p(p-1)(1-\chi\Lambda) \io u^{p-1}|\nabla u|
+ c_3 p^\frac{2}{1-a} M_{k-1}^2
\qquad \mbox{for all } t\in (0,T)
\eas
with $c_3:=c_1+c_2$.
Therefore, (\ref{55.3}) entails the autonomous ODI
\bas
\frac{d}{dt} \io u^p + \io u^p
\le c_3 p^\frac{2}{1-a} M_{k-1}^2,
\qquad t\in (0,T),
\eas
for $(0,T) \ni t \mapsto \io u^p(x,t)dx$, which upon a comparison argument implies that
\be{55.4}
M_k \le \max \bigg\{ \io u_0^{p_k} \, , \, c_3 p_k^\frac{2}{1-a} M_{k-1}^2 \bigg\}
\qquad \mbox{for all } k\ge 1.
\ee
Now if there exists a sequence $(k_j)_{j\in\N} \subset \N$ such that $k_j\to\infty$ as $j\to\infty$ and
\be{55.44}
M_{k_j} \le \io u_0^{p_{k_j}}
\qquad \mbox{for all } j\in\N,
\ee
we may take the $p_{k_j}$-th root on both sides here to see that according to the definition (\ref{55.22})
of $M_{k_j}$ we have
\bas
\sup_{t\in (0,T)} \|u(\cdot,t)\|_{L^{p_{k_j}}(\Omega)} \le \|u_0\|_{L^{p_{k_j}}(\Omega)},
\eas
which on letting $j\to\infty$ implies that
\be{55.5}
\sup_{t\in (0,T)} \|u(\cdot,t)\|_{L^\infty(\Omega)} \le \|u_0\|_{L^\infty(\Omega)}
\ee
in this case.\\
Conversely, if no such sequence exists, then (\ref{55.4}) means that with some suitably large $k_0\in\N$ we have
\bas
M_k \le c_3 p_k^\frac{2}{1-a} M_{k-1}
\qquad \mbox{for all } k\ge k_0.
\eas
Since $p_k^\frac{2}{1-a}=(2^\frac{2}{1-a})^k$, it is easy to see that this entails the existence of a number $b>1$
independent of $T$ which satisfies
\be{55.6}
M_k \le b^k M_{k-1}
\qquad \mbox{for all } k\ge 1.
\ee
By a straightforward induction, this warrants that
\bas
M_k \le b^{\sum_{j=0}^k j\cdot 2^{k-j}} \cdot M_0^{2^k}
\qquad \mbox{for all } k\ge 1,
\eas
where by an elementary computation,
\bas
\sum_{j=0}^k j\cdot 2^{k-j}
&=& 2^{k-1} \cdot \sum_{j=0}^k j\cdot \Big(\frac{1}{2}\Big)^{j-1} \\
&=& 2^{k-1} \cdot \frac{k\cdot(\frac{1}{2})^{k+1} -(k+1)\cdot (\frac{1}{2})^k + 1}{(\frac{1}{2})^2} \\
&=& k - 2(k+1) + 2^{k+1} \\[2mm]
&\le& 2^{k+1}
\qquad \mbox{for all } k\ge 1.
\eas
Thus,
\bas
M_k \le B^{2^{k+1}} \cdot M_0^{2^k}
\qquad \mbox{for all } k\ge 1,
\eas
by (\ref{55.22}) implying that
\be{55.7}
\sup_{t\in (0,T)} \|u(\cdot,t)\|_{L^{p_k}(\Omega)}
= M_k^\frac{1}{2^k}
\le b^2 M_0
\qquad \mbox{for all } k\ge 1.
\ee
Now as by the evident mass conservation property in (\ref{0}) we have $\io u(x,t)dx=\io u_0(x)dx$ for all
$t\in (0,\tm)$ and hence $M_0=\|u_0\|_{L^1(\Omega)}$, taking $k\to\infty$ in (\ref{55.7}) shows that in this
second case,
\be{55.8}
\sup_{t\in (0,T)} \|u(\cdot,t)\|_{L^\infty(\Omega)}
\le b^2 \|u_0\|_{L^1(\Omega)}.
\ee
Since all expressions on the right-hand sides of (\ref{55.5}) and (\ref{55.8}) do not depend on $T\in (0,\tm)$,
and since boundedness of $u$ clearly implies boundedness of $v$ by standard elliptic estimates, the proof
is complete.
\qed
|
\section{Introduction}
\label{sec:Introduction}
The India-based Neutrino Observatory (INO) collaboration has proposed to build a 50 kiloton magnetized Iron Calorimeter (ICAL) to precisely
measure the atmospheric neutrino oscillation parameters. The collaboration has chosen 2 $\times$ 2 m$^2$ size glass Resistive Plate Chambers
(RPCs) as the active detector elements and is going to deploy 28,800 of them in the ICAL detector \cite{1}. The RPCs will be operated in the
avalanche mode with an optimized gas mixture of C$_2$H$_2$F$_4$/iso-C$_{4}$H$_{10}$/SF$_{6}$ = 95.2/4.5/0.3 \cite{2}. The experiment is
expected to run for more than 10 years in order to record statistically significant number of neutrino interactions for the confirmation of
atmospheric neutrino oscillation. Therefore, long term stability and performance of the RPCs over the duration of the experiment is of prime
concern.
About 200,000 liters of gas is going to be circulating in the RPCs during the experiment. The gas lines running into about 135km in total
are going to supply/receive gas to/from the RPC detectors. In spite of stringent QC during the RPC gas gap making or gas lines plumbing,
it is impossible to prevent ambient air or water vapor entering into the gas circuit over these long periods of time. The contaminants are
known cause for serious degradation in the performance or permanent damage of the RPCs \cite{3,4,5}.
Considering the severe repercussions in the mammoth ICAL detector, a systematic study of this problem was undertaken. Two glass RPCs of
30 $\times$ 30 cm$^2$ size were fabricated and were simultaneously operated with standard gas mixture in one and with gas mixture along
with controlled and calibrated amount of water vapor in the other. A common cosmic ray muon telescope was set up for studying both the RPCs.
Ambient parameters such as temperature and relative humidity (RH) as well as RPCs operating and performance parameters such as currents,
efficiencies, singles rates, signal charges and time resolutions for cosmic ray muon detection were systematically recorded throughout the
experiment. It was observed significant deterioration in the performance of RPC in which gas with water vapor was flown.
\section{Experimental setup}
\label{sec:Experimental setup}
The RPCs of 30 $\times$ 30 cm$^2$ size were fabricated using 3 mm thick float glass plates. The outer surfaces of the glass plates were
coated with specially developed conductive graphite paint for the ICAL RPCs, which facilitates applying high voltage across the electrodes
\cite{6,7}. Readout strips of 2.8 cm width were orthogonally mounted on the external surfaces of the RPCs by keeping 0.2 cm gap between the
consecutive strips. The electrode and the readout strip are separated using a layer of mylar insulator.
A telescope was set up with three plastic scintillation counters to get a 3-fold (3F) coincidence for atmospheric muons. The dimensions
of scintillation counters in length $\times$ width $\times$ thickness are 30 cm $\times$ 2 cm $\times$ 1 cm (top),
30 cm $\times$ 3 cm $\times$ 1 cm (middle), and 30 cm $\times$ 5 cm $\times$ 1 cm (bottom). The block diagram of a stack of two RPCs and three
telescope counters, and the electronic circuit setup is shown in Figure~\ref{Fig:1}. An optimized gas mixture of C$_2$H$_2$F$_4$/
iso-C$_{4}$H$_{10}$/SF$_{6}$ = 95/4.5/0.5 was flown through the RPCs using polyurethane tubes. The block diagram of the gas flow system is
shown in Figure~\ref{Fig:2}.
\begin{figure}[ht]
\centering
\includegraphics[width=1\textwidth]{1.jpg}
\caption {Block diagram of the experimental setup developed for the studies.}
\label{Fig:1}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.85\textwidth]{2.jpg}
\caption {Block diagram of the gas flow system.}
\label{Fig:2}
\end{figure}
\section{Measurements and observations}
\label{sec:Results}
\subsection{Standard gas studies}
\label{sec:Standard gas studies}
{
The RPCs were operated with the standard gas at 10 SCCM rate and their currents, efficiencies and singles rates as a function of applied
voltage were measured. The measured efficiencies of RPC1 and RPC2 as a function of applied voltage are shown in Figure~\ref{Fig:3}. The
detectors showed more than 95\% efficiency on the plateau. The measured signal charges and time resolutions of the detectors are shown in
Figure~\ref{Fig:4}. The signal charges of RPC1 and RPC2 are 0.99 pC and 1.24 pC, and the time resolutions are 2.8 ns and 2.3 ns, respectively.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{3.pdf}
\caption {Efficiencies of RPC1 and RPC2 as a function of applied voltage.}
\label{Fig:3}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{4.pdf}
\caption {Signal charges: (a) and (b), and time resolutions: (c) and (d) of RPC1 and RPC2, respectively, with standard gas operation
at 12.2 kV.}
\label{Fig:4}
\end{figure}
\noindent {\bf Long-term stability studies:} \\
Then, the RPCs were operated at 12.2 kV and monitored their performances for 32 days. The efficiencies, singles rates and leakage currents
of the detectors for this time period are shown in Figure~\ref{Fig:5}. Their performances were stable throughout the period.
\begin{figure}[ht]
\centering
\includegraphics[width=1\textwidth]{5.pdf}
\caption {(a) Efficiencies, (b) singles rates and (c) currents of both the RPCs with standard gas operation for 32 days.}
\label{Fig:5}
\end{figure}
}
\subsection{Wet gas studies}
\label{sec:Wet gas studies}
{
The water vapor was started adding into RPC1 via the gas mixture using a water bubbler as shown in Figure~\ref{Fig:2}. The amount of water
vapor added to the detector was measured in a controlled way using RH sensor and is shown in Figure~\ref{Fig:6}. Throughout these studies,
RPC2 was continued with the standard gas flow itself and was a reference detector. Then, the performances of both the detectors were monitored
continuously.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{6.pdf}
\caption {The quantity of water vapor addition to the RPC1 for 10 SCCM and 30 SCCM gas flow rates.}
\label{Fig:6}
\end{figure}
With the wet gas operation for a few days, the efficiency and singles rate of RPC1 degraded to 0\% and 1 Hz, respectively. The currents drawn
by the detector increased gradually with time. The efficiency, singles rate and current drawn by the RPC1 during the wet gas operation in
comparison to RPC2 operated with the standard gas is shown in Figure~\ref{Fig:7}.
\begin{figure}[ht]
\centering
\includegraphics[width=1\textwidth]{7.pdf}
\caption {The (a) efficiencies, (b) singles rates and (c) currents of RPC1 with the wet gas operation, and that of RPC2 with
the standard gas operation.}
\label{Fig:7}
\end{figure}
\vskip 0.5cm
\noindent
{\bf Signal charge:} \\
The signal charge of RPC1 became smaller (0.36 pC) with the wet gas operation, whereas that of RPC2 with standard gas operation (1.19 pC)
remained similar to its measured value shown in Figure~\ref{Fig:4}b (1.24 pC). The signal charge distributions of RPC1 with wet
gas and RPC2 with standard gas operations are shown in Figures~\ref{Fig:8}a and~\ref{Fig:8}b, respectively.
\vskip 0.5cm
\noindent
{\bf Time resolution:} \\
The timing distribution of RPC1 with the wet gas operation got deteriorated as shown in Figure~\ref{Fig:8}c. Whereas, that of RPC2
with standard gas operation (1.9 ns) remained similar to its measured value shown in Figure~\ref{Fig:4}d (2.3 ns).
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{8.pdf}
\caption {Signal charge of: (a) RPC1 with the wet gas operation and (b) RPC2 with the standard gas operation, and time resolution of:
(c) RPC1 with the wet gas operation and (d) RPC2 with the standard gas operation.}
\label{Fig:8}
\end{figure}
}
\subsection{Recovery studies}
\label{sec:Recovery studies}
{
RPC1 was operated with wet gas at various flow rates and the deterioration in the detector's efficiency was observed as a function of time.
It was observed that higher the flow rate, faster the deterioration as shown in Figure~\ref{Fig:9}a. The detector was operated with wet gas
at 0\% efficiency for a day and then switched to standard gas. Then, the efficiency of detector recovered to more than 95\% as shown in
Figure~\ref{Fig:9}b.
\begin{figure}[ht]
\centering
\includegraphics[width=1\textwidth]{9.pdf}
\caption {(a) The deterioration in efficiency of RPC1 with wet gas operation at various flow rates and (b) its recovery with
standard gas operation.}
\label{Fig:9}
\end{figure}
}
\section{Conclusions}
\label{sec:Conclusions}
The effect of water vapor on the performance of glass RPC was studied. We used optimized gas mixture C$_2$H$_2$F$_4$/iso-C$_{4}$H$_{10}$/
SF$_{6}$ = 95/4.5/0.5 to operate the RPC in the avalanche mode. The detector showed more than 95\% efficiency with the standard gas operation,
its signal charge and time resolutions were 0.99 pC and 2.8 ns, respectively. Then, we started adding water vapor into the RPC. The detector's
efficiency deteriorated to 0\% after a few days of operation with wet gas. Its signal charge became 0.36 pC and timing distribution got
deteriorated. The RPC was operated at 0\% efficiency with wet gas for a day and then switched to standard gas. The detector's efficiency
recovered to more than 95\%.
\section{Acknowledgements}
\label{sec:Acknowledgements}
This work was supported by the Department of Atomic Energy (DAE), and the Department of Science and Technology (DST), Government of India.
The authors would like to gratefully acknowledge the help of their colleagues V. Asgolkar, S. Chavan, R. R. Shinde, P. Verma, S. D. Kalmani
and L. V. Reddy at TIFR, Mumbai, and V. Janarthanam and Jafar Sadiq at IIT Madras. K. Raveendrababu expresses gratitude to P. Fonte of LIP,
Coimbra for useful discussions during the RPC2016 workshop.
|
\section{Introduction}
A common challenge in science is to make predictions based on
incomplete information. Full details of the mechanism by which
correlations between two or more variables come about is often not
apparent and there may be several competing causal
explanations. Experimentation with interventions is one way to decide
between the candidate explanations~\cite{Pearl2009}. However, in many
situations such intervention is difficult (or unethical), for instance
if certain involved systems are outside our control.
Considering a particular causal structure generally imposes
restrictions on the set of correlations that can be produced. A
well-known example of such a constraint is a Bell
inequality~\cite{Bell1964}. That such relations can be violated using
measurements on quantum states motivates the consideration of more
general \emph{quantum} causal structures. Correlations that can be
generated in such structures but not in their classical analogue are
the basis for several important cryptographic tasks~\cite{Ekert1991},
in particular for device-independent protocols for key
distribution~\cite{Mayers1998,Barrett2005b,Acin2006,Vazirani2014} or
the generation of private
randomness~\cite{Colbeck2009,Pironio2010,Colbeck2011,Miller2014}.
In a cryptographic scenario, an adversary is usually able to exert
influence at particular points in the protocol, which can be
conveniently encoded using a causal structure.
Characterising the set of possible classical, quantum and post-quantum
correlations within a specific causal structure provides a basis to
understand further tasks and possible quantum and post-quantum
advantages, which were initially studied in specific
cases~\cite{Clauser1969,Braunstein1988,GHZ,Cerf1997, Collins2002}.
For a general causal structure with unobserved variables, deciding
whether a given set of correlations can be generated is
computationally difficult and only feasible for small
examples~\cite{Garcia2005, Lee2015}. One way to get around this, is
to use entropy to simplify the characterisation of the corresponding
sets of correlations~\cite{Chaves2012, Chaves2013, Fritz2013,
Chaves2014, Chaves2014b, Henson2014, Steudel2015, Chaves2015,
Chaves2016, Pienaar2016, Kela2017, Miklin2017}. Rather than looking
at the distributions themselves, we consider \emph{entropy vectors}
whose components are the joint entropies of each subset of the
observed variables.
This often\footnote{For classical causal structures, the set is always
convex, but for quantum causal structures it is not known whether
this is always the case.} has the advantage that the set of
entropies realisable in a given causal structure is convex, in
contrast to the set of compatible distributions. In addition, the
causal constraints can be represented by linear relations between
entropies instead of polynomial constraints. It is also significant
that entropic constraints on possible correlations in a causal
structure are independent of the dimension of the involved random
variables. Hence, the method enables the derivation of constraints
that are valid for arbitrarily large alphabet sizes of all involved
observed and unobserved systems. These properties make entropy vectors
a convenient means to distinguish different causal structures in many
situations.
In this paper we report the use of non-Shannon inequalities for
distinguishing causal structures. After a short outline of the
entropy vector approach and after introducing the necessary notation
in Section~\ref{sec:notation}, we go on to show in
Section~\ref{sec:non_shannon} that non-Shannon inequalities play a
central role for the distinction of causal structures. This is
illustrated with the triangle causal structure
(Section~\ref{sec:triangle}), one of the simplest causal structures in
which there is a separation between classical and quantum at the level
of correlations. For this example, we present numerous new entropic
constraints, which involve several infinite families of valid
inequalities, that together form the tightest entropic
characterisation of the classical triangle causal structure known to
date. This also leads us to disprove a claim that previously known
entropic approximations to this causal structure were
tight~\cite{Chaves2014, Chaves2015}. Whether our new inequalities are
sufficient to separate classical and quantum versions of causal
structures is left as an open problem.
In Section~\ref{sec:int}, we analyse a number of other causal
structures, taking into account non-Shannon inequalities for their
entropic characterisation. These inequalities are relevant for
distinguishing different classical causal structures as well as for
settling the question of whether there is a classical-quantum
separation in the entropy vector approach.
We further analyse the role of non-Shannon inequalities for the
entropic characterisation of the causal structure relevant in the
context of information causality~\cite{Pawlowski2009} in
Section~\ref{sec:post_selected_nonShan}, where the combination of
non-Shannon inequalities with post-selection allows us to derive
numerous new entropy inequalities.
In Section~\ref{sec:inner}, we provide the first inner approximations
to the entropy cones of causal structures. These are useful for
certifying that particular entropy vectors are realisable in a causal
structure as well as for showing tightness of an entropic outer
approximation in some cases (see Section~\ref{sec:int} for
examples). In cases where the outer approximation is not tight (or not
known to be tight), an inner approximation that shares some extremal
rays with the outer approximation allows the identification of parts
of the boundary of the true entropy cone as well as regions where
identifying the cone's boundary requires further analysis.
For comparison with the classical case, we also briefly consider
non-Shannon inequalities in the context of quantum and hybrid causal
structures in Section~\ref{sec:quantumtriangle}, which is illustrated
with the example of the triangle causal structure, before concluding
in Section~\ref{sec:conclusion}.
\section{Entropic cones and the entropy vector approach to causal structures}\label{sec:notation}
In this section, we briefly outline the entropy vector approach and
introduce the required notation. An elaborate introduction to the
topic can for instance be found in the review~\cite{our_review}.
\subsection{Entropic cones}\label{sec:classicalcone}
For a set of $n$ jointly distributed random variables
$\Omega= \left\{X_1,\ X_2,\ \ldots, X_n \right\}$ taking values in the
alphabet
$\mathcal{X}_\Omega= \mathcal{X}_1 \times \mathcal{X}_2 \times \cdots
\times \mathcal{X}_n$
we denote the set of all possible joint probability distributions as
$\mathcal{P}_n$. For a set of variables with joint distribution
$P_{\Omega} \in \mathcal{P}_n$ its \textit{Shannon
entropy}~\cite{Shannon1948} is
$$ H(\Omega):= - \sum_{x \in \mathcal{X}_\Omega} P_{\Omega}(x) \log_2{\left(P_{\Omega}(x)\right)}.$$
The Shannon entropy of $\Omega$ and of all its subsets can be
expressed in an \textit{entropy vector} in $\mathbb{R}^{2^{n}-1}$,
\begin{equation*}
\begin{split}
{\bf H}(P):=(H(X_1),H(X_2),\ldots,H(X_n), H(X_1X_2),\\
H(X_1X_3),\ldots,H(X_1 X_2 \ldots X_n)).
\end{split}
\end{equation*}
The closure of the set of all possible entropy vectors,
$\Gamma^{*}_n$, is a convex cone, denoted as
$\overline{\Gamma^{*}_n}$~\cite{Zhang1997}.\footnote{The closure is
taken because there isn't in general a good reason to put an upper
bound on the alphabet sizes and it is known that
$\overline{\Gamma^{*}_n}\neq\Gamma^{*}_n$ for
$n\geq3$~\cite{Zhang1997}.} While for $n\leq 3$, the \emph{entropy
cone} $\overline{\Gamma^{*}_n}$ is
polyhedral~\footnote{\label{ft:3var} In fact $\overline{\Gamma^{*}_3}$
equals the corresponding Shannon cone, $\Gamma_3$, introduced
below.}~\cite{Zhang1998}, an infinite number of linear inequalities
are required to characterise $\overline{\Gamma^{*}_n}$ for
$n \geq 4$~\cite{Matus2007}. Hence, considering approximations to
$\overline{\Gamma^{*}_n}$ is common practice.
\subsubsection{Approximations to $\overline{\Gamma^{*}_n}$}
Before specifying approximations to $\overline{\Gamma^{*}_n}$, we define a few quantities, that are relevant in the following. The \textit{conditional entropy} of two disjoint subsets $X_{S},\ X_{T} \subseteq \Omega$ is defined as
$$ H(X_S | X_T):= H(X_S \cup X_T)-H(X_{T})$$ and for three mutually disjoint subsets $X_{S},\ X_{T},\ X_U \subseteq \Omega$ the \textit{conditional mutual information} of $X_S$ and $X_T$ conditioned on $X_U$ is
$$ I(X_S \! : \! X_T | X_U):= H(X_S | X_U)-H(X_S | X_{T} \cup X_U).
$$
Note that the entropy of the empty set is $H(\emptyset)=0$, so that
$H(X_S)=H(X_S | \emptyset)$, for example.
Two other entropic quantities we will make use of in this article are the \textit{interaction information}~\cite{McGill1954} of three mutually disjoint subsets $X_{S},\ X_{T},\ X_U \subseteq \Omega$,
$$ I(X_S \! : \! X_T \! : \! X_U):= I(X_S \! : \! X_T )-I(X_S \! : \! X_T | X_U),$$
and the \textit{Ingleton quantity} of four mutually disjoint subsets $X_S,\ X_T,\ X_U,\ X_V \subseteq \Omega$,
\begin{multline} \label{eq:ingleton}
I_{\rm ING}(X_S,X_T ; \! X_U,X_V):= I(X_S \! : \! X_T|X_U)\\\!+\! I(X_S \! : \! X_T|X_V) + I(X_U \! : \! X_V) - I(X_S \! : \! X_T).
\end{multline}
For any entropy vector of a joint distribution of the random variables $\Omega$ the following \textit{Shannon inequalities} hold:
\begin{itemize}
\item \vspace{-0.1cm} For any $X_S \subseteq \Omega$, $H(X_S) \geq 0.$
\item \vspace{-0.1cm} For any disjoint $X_S, X_T \! \subseteq \! \Omega$, $H(X_S | X_T) \! \geq \! 0.$
\item \vspace{-0.1cm} For any disjoint $X_S,\ X_T,\ X_U \subseteq \Omega$, ${I(X_S \! : \! X_T | X_U)} \geq 0.$
\end{itemize}
They are known to constrain a convex polyhedral cone, the \textit{Shannon cone}, $\Gamma_n$~\cite{Yeung1997}.
Because the Shannon inequalities hold for any entropy vector we have
$\overline{\Gamma^{*}_n} \subseteq \Gamma_n$.
The first entropy inequality that is not of Shannon type was found in~\cite{Zhang1997} and is presented in the following.
\begin{prop}[Zhang \& Yeung] \label{prop:zhangyeung} For any four discrete random variables $X_1$, $X_2$, $X_3$ and $X_4$ the following inequality holds:
\begin{align*}
I(X_1\! : \! X_2|X_3) +I(X_1\! : \! X_2|X_4) + I(X_3\! : \! X_4) \\
- I(X_1\! : \! X_2) + I(X_1 \! : \! X_3|X_2) +I(X_2 \! : \! X_3|X_1)\\
+ I(X_1 \! : \! X_2|X_3) \geq 0.
\end{align*}
In the following the lhs of this inequality is abbreviated as
$\Diamond_{X_1 X_2 X_3 X_4}$.
\end{prop}
The first account of infinite families of inequalities was given in~\cite{Matus2007}.
\begin{prop}[Mat\'{u}\v{s}] \label{prop:matusfam}
Let $X_1$, $X_2$, $X_3$ and $X_4$ be random variables and let $s \in \mathbb{N}$. Then the following inequalities hold:
\begin{align}
&s \left[ I(X_1\! : \! X_2|X_3) + I(X_1\! : \!
X_2|X_4) + I(X_3\! : \! X_4)\right.\nonumber\\
&\left.\!-\!I(X_1\! : \! X_2) \right]\!+\!I(X_1\! : \! X_3|X_2)\!+\!
\frac{s(s\!+\!1)}{2} \left[ I(X_2\! : \!
X_3|X_1)\!\right.\nonumber\\ &\left. + I(X_1\! : \! X_2|X_3) \right] \geq 0,\label{eq:matus1}
\end{align}
\begin{align}
s&\left[ I(X_1\! : \! X_2|X_3) + I(X_1\! : \! X_2|X_4) + I(X_3\! : \! X_4) \right. \nonumber\\
&\left. - I(X_1\! : \! X_2) \right] +s \left[ I(X_2\! : \! X_3|X_1)+I(X_1\! : \! X_2|X_3) \right] \nonumber\\
&+ I(X_1\! : \! X_3|X_2) + \frac{s(s-1)}{2} \left[ I(X_2\! : \! X_4|X_1) \right. \nonumber\\
&\left. + I(X_1\! : \! X_2|X_4) \right] \geq 0.\label{eq:matus2}
\end{align}
\end{prop}
For $s=1$ both inequalities are equivalent to
$\Diamond_{X_1 X_2 X_3 X_4} \geq 0$. For the current state of the art
on non-Shannon inequalities we refer to~\cite{Dougherty2011}. To our
knowledge, all known non-Shannon entropy inequalities in four
variables that are not (known to be) rendered redundant by tighter ones
can be written as the sum of the Ingleton quantity and (conditional)
mutual information
terms~\cite{Zhang1998,Makarychev2002,Dougherty2006,Matus2007,Xu2008,Dougherty2011}.
\bigskip
Complementary to outer approximations (such as the Shannon cone,
$\Gamma_n$) it is also interesting to consider inner approximations,
$\Gamma^{I}_n$, to the $n$-variable entropy cone
$\overline{\Gamma_{n}^{*}}$. Such approximations can be defined in
terms of so-called linear rank inequalities, which are inequalities
that hold for the dimensions of subspaces of vector spaces~\cite{Ingleton}.
Entropic inequalities imply linear rank inequalities\footnote{For example, the
linear rank inequality implied by the fact that $H(AB)\geq H(A)$ for
all distributions $P_{AB}$ is $\dim(A\cup B)\geq\dim(A)$ for all
subspaces $A$ and $B$ of a (finite dimensional) vector space.} but
the converse does not hold~\cite{Hammer2000}, which is why using (the entropic analogue
of) linear rank inequalities gives an inner approximation. In the case
of $n=4$ the Shannon inequalities and the \emph{Ingleton inequality},
i.e.,
\begin{equation} I_{\rm ING}(X_1,
X_2 ; \! X_3, X_4) \geq 0 \label{eq:ingletoninequ}
\end{equation}
(and its permutations), define such an inner approximation
$\Gamma^{I}_4$~\cite{Ingleton}. $\Gamma^{I}_5$ is defined by the
Shannon inequalities, all instances of the Ingleton inequality and
$24$ additional classes of inequalities~\cite{Dougherty2009}. For $6$
or more variables, a complete list of all linear rank inequalities is
not known, nor is it known whether such a list would be finite. A list
of over a billion inequalities (counting permutations) has been
found~\cite{Dougherty2014}.\footnote{Note that we can always obtain
some inner approximations with other methods, e.g., by constructing
a set of achievable entropy vectors and taking their convex hull.}
\subsection{The entropy vector approach to causal structures}
A \textit{causal structure}, $C$, is a set of variables arranged in a
directed acyclic graph (DAG). The \emph{parents},
$X^{\downarrow_{1}}$, of a variable $X$ in a DAG are the variables
from which an arrow is directly pointing at $X$, and the
\textit{descendants}, $X^{\uparrow}$ of $X$ are all variables that may
be reached from $X$ along a directed path within the DAG. We use $C^{{\rm C}}$ and $C^{{\rm Q}}$ to denote the classical and the quantum version of a causal structure respectively.
\subsubsection{Classical causal structures}\label{sec:classical_method}
The graph of a classical causal structure, $C^{\mathrm{C}}$, with random variables $X_1,\ X_2,\ \ldots,\ X_n$, encodes the independence relations of $X_1,\ X_2,\ \ldots,\ X_n$ in the sense that the distribution $P_\mathrm{X_1 X_2 \ldots X_n}$ is said to be \textit{compatible} with $C^{\mathrm{C}}$ if it can be decomposed as
$$P_\mathrm{X_1 X_2 \ldots X_n}= \prod_{i} P_\mathrm{X_{i}|X^{\downarrow_{1}}_{i}} .$$
This interpretation of classical causal structures follows the theory of Bayesian networks~\cite{Pearl2009}. The set of all compatible distributions is in the following denoted $\mathcal{P}(C^{\mathrm{C}})$.
The compatibility requirement is equivalent to the condition that for each variable $X_i$,
\begin{equation}\label{eq:indepentr}
I(X_i\!:\!X_i^{{\centernot\uparrow}}|X^{\downarrow_{1}}_i)=0.
\end{equation}
$X_i^{{\centernot\uparrow}}$ denotes the \textit{non-descendants} of $X_i$, i.e., all variables in the causal structure except for the variable itself and its descendants.~\footnote{In particular, all other (conditional) independence relations of variables in the causal structure are implied by these $n$ equalities.}
The entropic description of causal structures has first been considered in~\cite{Chaves2012, Fritz2013}. The $n$ equalities \eqref{eq:indepentr} restrict the $n$-variable entropy cone $\overline{\Gamma_n^{*}}$ to the cone of all entropy vectors compatible with $C^{{\rm C}}$, denoted $\overline{\Gamma^{*}}\left(C^{\mathrm{C}}\right)$. An outer approximation to $\overline{\Gamma^{*}}(C^{\mathrm{C}})$ is constructed by supplementing $\Gamma_n$ with the same $n$ equalities, which leads to the cone $\Gamma\left(C^{{\rm C}} \right)$.
When $k$ out of the $n$ variables of $C^{{\rm C}}$ are
observed, we take these to be the first $k$ variables,
$X_1, X_2, \ldots, X_k$, without loss of generality. For $k <n$ we are then interested in deriving constraints for the observed variables only.
For a compatible distribution,
$P_{{\rm X_1 X_2} \cdots {\rm X_n}} \in \mathcal{P}\left(C^{{\rm C}}\right)$,
this is achieved by marginalising over the unobserved variables
$X_{k+1}, X_{k+2}, \ldots, X_n$ which yields a distribution in the set
of all compatible marginal distributions
$P_{{\rm X_1 X_2} \cdots {\rm X_k}} \in
\mathcal{P}_{\mathcal{M}}\left(C^{{\rm C}}\right)$.
Entropically, marginalisation corresponds to a projection of the
entropy cone to the corresponding $k$-variable marginal cone
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C^{\mathrm{C}}\right)
\subsetneq \mathbb{R}^{2^{k}-1}$,
which would be obtained by dropping all components involving any of
the $n-k$ unobserved variables from each vector in
$\overline{\Gamma^{*}}\left(C^{\mathrm{C}}\right)$. The outer
approximation $\Gamma\left(C^{\mathrm{C}}\right)$ can be analogously
projected to an approximation,
$\Gamma_{\mathcal{M}}\left(C^{\mathrm{C}}\right)$, of
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C^{\mathrm{C}}\right)$. Computationally,
$\Gamma_{\mathcal{M}}\left(C^{\mathrm{C}}\right)$ is usually found by considering $\Gamma\left(C^{\mathrm{C}}\right)$ characterised by means of bounding hyperplanes and applying a Fourier-Motzkin elimination algorithm to the system of linear inequalities~\cite{Williams1986}.
\subsubsection{Quantum causal structures} \label{sec:quantum_method} A
quantum causal structure $C^{{\rm Q}}$ differs from its classical analogue
in that the unobserved nodes correspond to quantum systems. Here, we
only consider causal structures with two generations of nodes, where
the nodes of the first generation are unobserved quantum systems
and the nodes of the second generation represent observed
(classical) variables. Note that this also allows for the description
of causal structures with observed input nodes, as is illustrated in
Figure~\ref{fig:inputs}.
\begin{figure}
\centering
\resizebox{0.8 \columnwidth}{!}{%
\begin{tikzpicture}[scale=0.9]
\node (a) at (-2,2.5) {$(a)$};
\node[draw=black,circle,scale=0.75] (X1) at (-1,2) {$X_2$};
\node[draw=black,circle,scale=0.75] (A1) at (-1,0) {$X_1$};
\node (Z1) at (-0.0,0.5) {};
\node (b) at (2,2.5) {$(b)$};
\node[draw=black,circle,scale=0.75] (X) at (3,2) {$X_2$};
\node[draw=black,circle,scale=0.75] (A) at (3,0) {$X_1$};
\node (ZZ) at (3,1) {$A$};
\node (Z) at (4,0.5) {};
\node (Y) at (5,0.5) {};
\draw [->,>=stealth] (A1)--(X1);
\draw [->,>=stealth] (ZZ)--(X);
\draw [->,>=stealth] (ZZ)--(A);
\draw [dotted,->,>=stealth] (Z1)--(X1);
\draw [dotted,->,>=stealth] (Z)--(X);
\end{tikzpicture}
}%
\caption{For a quantum causal structure with an observed input node,
$X_1$~--~meaning a parentless node from which there is only one
arrow to another observed node, $X_2$~--~there always exists another
(quantum) causal structure that allows for exactly the same
correlations and where the observed input is replaced by a shared
quantum parent of $X_1$ and $X_2$. To simulate any correlations in
(a) within scenario (b) we can use a quantum system that sends
perfectly correlated classical states to both nodes $X_1$ and $X_2$,
distributed as $X_1$. On the other hand, any correlations obtained
in scenario (b) can be created in scenario (a) by having a random
variable $X_1$ sent to node $X_2$, where the relevant quantum states
(the reduced states that would be present in (b) conditioned on the
value of $X_1$) are locally generated. Note that these
considerations are not restricted to quantum causal structures but
apply also in the classical case (or even if considering states from
a generalised probabilistic theory). }
\label{fig:inputs}
\end{figure}
For such causal structures, each edge has an associated Hilbert space,
which can be labelled by the parent and child, e.g., for a DAG with an
edge $X \rightarrow Y$, there is an associated
$\mathcal{H}_{X_{Y}}$. Each unobserved node is labelled by a quantum
state, a density operator on the tensor product of the Hilbert spaces
associated with the edges originating at that node. For each observed
node there is an associated POVM that acts on the tensor product of
the Hilbert spaces associated with the edges that meet at that node.
The distributions,
$P\in\mathcal{P}_{\mathcal{M}}\left(C^{{\rm Q}}\right)$, of the observed
variables that are \emph{compatible} with a causal structure
$C^{{\rm Q}}$, are those resulting from performing the specified POVMs on
the relevant systems via the Born rule.
A technique to analyse these sets entropically was proposed by Chaves et al.~\cite{Chaves2015} and is outlined in the following, where the idea of considering entropy cones of multi-party quantum states goes back to Pippenger~\cite{Pippenger2003}.
The set of compatible observed distributions $P \in \mathcal{P}_{\mathcal{M}}\left(C^{{\rm Q}}\right)$ can be mapped to a set of compatible entropy vectors, the closure of which is denoted $\overline{\Gamma^{*}_{\mathcal{M}}}\left(C^{{\rm Q}}\right)$.
To approximate this set, a system is assigned to each observed variable as well as to each outgoing edge of each unobserved node. As opposed to the classical case, where we can always define a joint distribution over all variables in a causal structure $C^{{\rm C}}$, there is in general no joint quantum state over all systems in $C^{{\rm Q}}$. In particular, the systems corresponding to the edges that meet at an observed node do not coexist with the outcome at that node and hence there is no joint quantum state from which a joint entropy could be derived.
The approach is therefore based on a notion of coexistence: two systems are said to \textit{coexist} if neither
is a quantum ancestor of the other in $C^{{\rm Q}}$, and a set of systems that pairwise coexist form a \textit{coexisting set}. For each coexisting set, $X_S \subseteq \Omega$, the von Neumann entropy $H(X_S):=- \tr(\rho_{X_S} \log_2 \rho_{X_S})$ of their joint state $\rho_{X_S}$ is defined; all of these von Neumann entropies are considered as components of an entropy vector.
For each coexisting set the entropies of all its subsets as well as
all conditional mutual informations of its systems are
positive~\cite{Lieb1973}. The conditional entropy may not be positive
in general, but for three mutually disjoint subsets of a coexisting
set, $X_T,\ X_U, \ X_V \subseteq X_S$,
$H(X_T)+H(X_U) \leq H(X_T \cup X_V)+H(X_U \cup X_V)$ holds
instead. These three types of inequality hold for the components of
any entropy vector. For the von Neumann entropy of a multi-party
quantum state no additional entropy inequalities are known. It has
been suggested, however, that any classical `balanced entropy
inequality'~\cite{Chan2003} (which includes all known non-Shannon
inequalities) may also hold for multi-party quantum
states~\cite{Cadney2012}. It is worth remarking that the lack of a
joint state for all nodes within a quantum causal structure would
restrict the applicability of such inequalities in the causal context
if they were to hold.\footnote{In the triangle causal structure, for
instance, even if the known non-Shannon inequalities held for
arbitrary quantum states, they would not allow us to derive any new
entropy inequalities for the quantum version of this causal structure.}
In many circumstances the conditional entropy of certain sets of
systems is known to be positive, e.g.\ if all systems in a coexisting
set are classical. Such constraints on the entropy vectors are also
added (see~\cite{our_review} for further details). The causal
restrictions encoded in the graph are accounted for by the condition
that two subsets of a coexisting set are independent (and hence have
zero mutual information between them) if they have no shared
ancestors.\footnote{Since a node and its (quantum) ancestors never
coexist, conditional inependences don't have to be taken into
account in two-generation causal structures.} To relate the
entropies of systems in different coexisting sets, data processing
inequalities (DPIs) are used: Let
$\rho_\mathrm{X_S X_T} \in \mathcal{S}(\mathcal{H}_{X_\mathrm{S}}
\otimes \mathcal{H}_{X_{\mathrm{T}}})$
and $\mathcal{E}$ be a completely positive trace preserving (CPTP) map
on $\mathcal{S}(\mathcal{H}_{X_{\mathrm{T}}})$ leading to a state
$\rho'_\mathrm{X_{\mathrm{S}} X_{\mathrm{T}}}$. Then
\begin{equation}\label{eq:DPI}
I(X_{\mathrm{S}} \! : \! X_{\mathrm{T}})_{\rho'_\mathrm{X_S X_T}} \leq I(X_{\mathrm{S}} \! : \! X_{\mathrm{T}})_{\rho_\mathrm{X_S X_T}}.
\end{equation}
Results on the redundancy of certain DPIs have been presented in~\cite{our_review}.
All constraints on the possible entropy vectors taken together define a polyhedral cone, which we denote $\Gamma\left( C^{{\rm Q}} \right)$. Its projection to the observed variables, $\Gamma_\mathcal{M}\left( C^{{\rm Q}} \right)$, is an outer approximation to $\overline{\Gamma^{*}_{\mathcal{M}}}\left(C^{{\rm Q}}\right)$, that can be computed from $\Gamma\left( C^{{\rm Q}} \right)$ with a Fourier-Motzkin elimination algorithm~\cite{Williams1986}.
\section{Improving current entropic characterisations with non-Shannon inequalities}\label{sec:non_shannon}
In this section we show how non-Shannon inequalities allow us to
improve the previous outer approximations to the entropy cones of
classical causal structures. We give an improved entropic description
of the triangle causal structure of Figure~\ref{fig:3variables}(e)
(Section~\ref{sec:triangle}), discuss the application of non-Shannon
inequalities to further causal structures (Section~\ref{sec:int}) and
demonstrate that non-Shannon inequalities are also applicable in
combination with post-selection using information causality as an
example (Section~\ref{sec:post_selected_nonShan}).
The computational procedure that we use in order to derive these new
inequalities is roughly outlined in the following. (1) We take the
Shannon inequalities for the joint distribution of all variables in a
causal structure $C^{{\rm C}}$, (2) we add a set of valid non-Shannon
inequalities to these, (3) we add all conditional independence
equalities that are implied by $C^{{\rm C}}$, (4) we eliminate all
entropies of unobserved variables from the full set of inequalities
(by means of a Fourier-Motzkin elimination
algorithm~\cite{Williams1986}), which leads to constraints on the
entropies of the observed variables only.
Note that the same procedure, but missing out step (2) corresponds to
the computation of $\Gamma_{\mathcal{M}}(C^{\rm C})$ as in~\cite{Chaves2012,
Fritz2013, Chaves2014b} and outlined in Section~\ref{sec:classical_method}.
Thus, the inclusion of (2) is responsible for the new constraints. In
addition to deriving entropy inequalities computationally, we also
provide analytic derivations of (infinite families of) new
inequalities.
\subsection{Improved outer approximation to the entropy cone of the
classical triangle scenario} \label{sec:triangle}
The triangle causal structure, called $C_3$, is one of the simplest
examples with interesting features~\cite{Branciard2012, Fritz2012,
Chaves2014, Henson2014, Chaves2015}. It can be used when three
parties make observations, $X$, $Y$ and $Z$ respectively, on systems,
$A$, $B$ and $C$, that are shared pairwise between them. This may for
instance be realised in a communication protocol where three parties
aim to obtain (correlated) data while interacting in pairs and without
ever having interacted as a group.
$C_3$ is one of only five distinct causal structures involving three
observed random variables that exhibit no ancestral relations between
the observed variables (cf.\ Figure~\ref{fig:3variables}).
\begin{figure}
\centering
\resizebox{0.95 \columnwidth}{!}{%
\begin{tikzpicture}[scale=0.48]
\node (a) at (-9.6,2.5) {$(a)$};
\node[draw=black,circle,scale=0.75] (X1) at (-8.5,2) {$X$};
\node[draw=black,circle,scale=0.75] (Y1) at (-4.5,2) {$Y$};
\node[draw=black,circle,scale=0.75] (Z1) at (-6.5,-1.46) {$Z$};
\node (b) at (-3.1,2.5) {$(b)$};
\node[draw=black,circle,scale=0.75] (X) at (-2,2) {$X$};
\node[draw=black,circle,scale=0.75] (Y) at (2,2) {$Y$};
\node[draw=black,circle,scale=0.75] (Z) at (0,-1.46) {$Z$};
\node (A) at (1,0.28) {$A$};
\node (c) at (3.4,2.5) {$(c)$};
\node[draw=black,circle,scale=0.75] (X2) at (4.5,2) {$X$};
\node[draw=black,circle,scale=0.75] (Y2) at (8.5,2) {$Y$};
\node[draw=black,circle,scale=0.75] (Z2) at (6.5,-1.46) {$Z$};
\node (A2) at (6.5,0.55) {$A$};
\node (d) at (-6.6,-3.0) {$(d)$};
\node[draw=black,circle,scale=0.75] (X3) at (-5.5,-3.46) {$X$};
\node[draw=black,circle,scale=0.75] (Y3) at (-1.5,-3.46) {$Y$};
\node[draw=black,circle,scale=0.75] (Z3) at (-3.5,-6.92) {$Z$};
\node (A3) at (-2.5,-5.18) {$A$};
\node (B3) at (-4.5,-5.18) {$B$};
\node (e) at (0.4,-3.0) {$(e)$};
\node[draw=black,circle,scale=0.75] (X4) at (1.5,-3.46) {$X$};
\node[draw=black,circle,scale=0.75] (Y4) at (5.5,-3.46) {$Y$};
\node[draw=black,circle,scale=0.75] (Z4) at (3.5,-6.92) {$Z$};
\node (A4) at (4.5,-5.18) {$A$};
\node (B4) at (2.5,-5.18) {$B$};
\node (C4) at (3.5,-3.46) {$C$};
\draw [->,>=stealth] (A)--(Y);
\draw [->,>=stealth] (A)--(Z);
\draw [->,>=stealth] (A2)--(Y2);
\draw [->,>=stealth] (A2)--(Z2);
\draw [->,>=stealth] (A2)--(X2);
\draw [->,>=stealth] (A3)--(Y3);
\draw [->,>=stealth] (A3)--(Z3);
\draw [->,>=stealth] (B3)--(X3);
\draw [->,>=stealth] (B3)--(Z3);
\draw [->,>=stealth] (A4)--(Y4);
\draw [->,>=stealth] (A4)--(Z4);
\draw [->,>=stealth] (B4)--(X4);
\draw [->,>=stealth] (B4)--(Z4);
\draw [->,>=stealth] (C4)--(X4);
\draw [->,>=stealth] (C4)--(Y4);
\end{tikzpicture}
}%
\caption{Assuming no ancestral relations between any of the three
observed variables $X$, $Y$ and $Z$ (i.e., no member of $\{X,Y,Z\}$
is an ancestor of any other), the above are the only possible causal
structures (up to relabelling). $A$, $B$ and $C$ correspond to
unobserved variables.}
\label{fig:3variables}
\end{figure}
All except for the causal structures (c) and (e) may be distinguished by looking at independences among the observed variables, $X$, $Y$ and $Z$, listed in Table~\ref{table:3variablescenarios}.
\begin{table}\scriptsize
\begingroup
\def\inlinedisplayeqn#1{\vspace*{0.1ex}$\displaystyle #1$\vspace*{0.1ex}}
\begin{tabular}{|m{0.55cm}|m{4.80cm}|m{1.52cm}|}
\hline
\hspace{0pt}\textbf{} &\hspace{0pt}\textbf{Compatible Distributions} &\hspace{0pt}\textbf{Observed $ \ \ $ Independence} \\
\hline
\hspace{0pt} (a) & \hspace{-3pt}$P_{XYZ} = P_{X}P_{Y}P_{Z}$ & $I(X\! \! : \! \! YZ) \! \! = \! \! 0$ \newline $I(Y\! \! : \! \! XZ)\!\! = \!\! 0$ \newline $I(Z \! \! : \! \! XY) \! \! =\! \! 0$ \\
\hline
\hspace{0pt} (b) & \hspace{-3pt}\vphantom{{\large H}}\inlinedisplayeqn{P_{XYZ}= \sum_{A}P_{X}P_{Y| A}P_{Z| A}P_{A}} &
$I(X\! \! : \! \! YZ) \! \! = \! \! 0$ \\
\hline
\hspace{0pt} (c) & \hspace{-3pt}\vphantom{{\large H}}\inlinedisplayeqn{P_{XYZ}= \sum_{A}P_{X| A}P_{Y| A}P_{Z| A}P_{A}} &
None \\
\hline
\hspace{0pt} (d) & \hspace{-3pt}\vphantom{{\large H}}\inlinedisplayeqn{P_{XYZ}= \sum_{A,B}P_{X| B}P_{Y\mid A}P_{Z| AB}P_{A}P_{B}} &
$I(X\! \! : \! \! Y)=0$ \\
\hline
\hspace{0pt} (e) & \hspace{-4pt}\vphantom{{\large H}}\inlinedisplayeqn{P_{X \! Y \! Z}\! \! = \! \!\! \! \! \! \sum_{A,B,C}\! \! \! \! \! P_{X \! | \! AC}P_{Y \! | \! AC}P_{Z \! | \! AB}P_{A}P_{B}P_{C}} &
None \\
\hline
\hline
\end{tabular}
\endgroup
\caption{Distributions compatible with the three-variable causal structures displayed in Figure~\ref{fig:3variables}.}
\label{table:3variablescenarios}
\end{table}
However, while the causal structure of Figure~\ref{fig:3variables}(c)
does not impose any restrictions on the compatible $P_{XYZ}$, the
distributions that are compatible with the triangle causal structure
of Figure~\ref{fig:3variables}(e) obey additional
constraints~\cite{Steudel2015}.\footnote{For instance, perfectly
correlated bits $X$, $Y$ and $Z$, i.e., those with joint
distribution
\begin{equation} \label{eq:perfcor}
P_{XYZ}(x,y,z)=
\begin{cases}
\frac{1}{2} &x=y=z\\
0 &\textrm{otherwise},
\end{cases}
\end{equation}
are not achievable in this causal structure. This is not only true
classically, but also in any generalised probabilistic
theory~\cite{Steudel2015, Henson2014}.}
This illustrates that causal structures encode more than the observed independences.
Furthermore, $C_3$ is unique among these five causal structures, it
being the only one that features quantum correlations that are not
classically reproducible, i.e.,
$P_{\mathcal{M}}\left(C_3^{{\rm C}}\right) \subsetneq
P_{\mathcal{M}}\left(C_3^{{\rm Q}}\right)$,
as proven in Ref.~\cite{Fritz2012} (see
Section~\ref{sec:quantcorr} for further details regarding the quantum
scenario).\footnote{In structures (a), (b) and (c) all joint
distributions are allowed for the variables that share a common
cause in the classical case. Hence, quantum systems do not enable
any stronger correlations. This is because, for any quantum
state $\rho_A$ shared at $A$ and measured later, the correlations can
be classically reproduced if $A$ sends out the same classical output
statistics to the parties directly.
In structure (d) no non-classical quantum correlations exist
either~\cite{Fritz2012}. This is also fairly intuitive: the quantum
measurements performed at $X$ and $Y$ could be equivalently
performed at the sources $B$ and $A$ respectively, such that these
sources distribute cq-states of the form
$\sum_{x} P_X(x) \ketbra{x}{x} \otimes \rho_{B_Z}^{x}$ and
$\sum_{y} P_Y(y) \ketbra{y}{y} \otimes \rho_{A_Z}^{y}$ instead. The
same correlations can be achieved classically by taking random
variables $B=X$ and $A=Y$ (these being distributed according to
$P_X$ and $P_Y$). Since $\rho_{B_Z}^{x}$ and $\rho_{A_Z}^{y}$ are
functions of $X$ and $Y$, the statistics formed by measuring such
states can be computed classically via a probabilistic function
(this function could be made deterministic by taking $B=(X,W)$,
where $W$ is distributed appropriately).}
\begin{figure}
\centering
\resizebox{0.45\columnwidth}{!}{%
\begin{tikzpicture}[scale=0.8]
\node[draw=black,circle,scale=0.75] (X) at (-2,2) {$X$};
\node[draw=black,circle,scale=0.75] (Y) at (2,2) {$Y$};
\node[draw=black,circle,scale=0.75] (Z) at (0,-1.46) {$Z$};
\node (A) at (1,0.28) {$A$};
\node (B) at (-1,0.28) {$B$};
\node (C) at (0,2) {$C$};
\draw [->,>=stealth] (A)--(Y) node [right,pos=0.4] {$A_\mathrm{Y}$};
\draw [->,>=stealth] (A)--(Z) node [right,pos=0.5] {$A_\mathrm{Z}$};
\draw [->,>=stealth] (B)--(X) node [left,pos=0.4] {$B_\mathrm{X}$};
\draw [->,>=stealth] (B)--(Z) node [left,pos=0.5] {$B_\mathrm{Z}$};
\draw [->,>=stealth] (C)--(X) node [above,pos=0.5] {$C_\mathrm{X}$};
\draw [->,>=stealth] (C)--(Y) node [above,pos=0.5] {$C_\mathrm{Y}$};
\end{tikzpicture}
}%
\caption{Triangle causal structure $C_3$. Three observed random
variables $X$, $Y$ and $Z$ have pairwise common causes. In the
classical case these common causes are random variables, $A$, $B$
and $C$, while in the quantum case these are replaced by quantum
systems, ($A_{\mathrm{Y}}$, $A_{\mathrm{Z}}$), ($B_{\mathrm{X}}$,
$B_{\mathrm{Z}}$) and ($C_{\mathrm{X}}$, $C_{\mathrm{Y}}$).}
\label{fig:triangle}
\end{figure}
In the following, we derive new and improved outer approximations
to $\overline{\Gamma^*_\mathcal{M}}(C_3^{{\rm C}})$ by using non-Shannon entropy
inequalities. These show that the Shannon approximation to
$\overline{\Gamma^*_\mathcal{M}}(C_3^{{\rm C}})$ is not tight, i.e., that
$ \overline{\Gamma^*_\mathcal{M}}(C_3^{{\rm C}}) \subsetneq
\Gamma_\mathcal{M}(C_3^{{\rm C}})$.
We remark that our findings contradict the considerations
of~\cite{Chaves2014,Chaves2015}, which together argue that in the
marginal scenario there is no separation between the Shannon
cone and the classical entropy cone, i.e., they argue that
$\Gamma_\mathcal{M}(C_3^{{\rm C}})=\overline{\Gamma^*_\mathcal{M}}(C_3^{{\rm C}})$, which
would imply that non-Shannon inequalities are
irrelevant.\footnote{The details of this are in the Supplementary
Information of~\cite{Chaves2015}.} For further discussion of the
discrepancy with~\cite{Chaves2014,Chaves2015}, see
Appendix~\ref{sec:prevwork}.
The set of all observed distributions compatible with $C_3^{{\rm C}}$ is\footnote{Note that this set is not convex, which can be seen by considering the perfect
correlations $P_{XYZ}$ of~\eqref{eq:perfcor} (which are not in
$\mathcal{P}_{\mathcal{M}}\left(C_3^{{\rm C}}\right)$) as a convex
combination of the distribution where $X$, $Y$ and $Z$ are always
$0$ and the distribution where $X$, $Y$ and $Z$ are always $1$
(both of which are).}
\begin{align*}
\mathcal{P}_{\mathcal{M}}
&\left(C_3^{{\rm C}}\right)=\left\{P_{XYZ}\in\mathcal{P}_{3}|\phantom{\sum_{A,B,C}}\right.\\
&\left. P_{XYZ}=\sum_{A,B,C}P_{A}P_{B}P_{C}P_{X|BC}P_{Y|AC}P_{Z|AB}\right\} \! .
\end{align*}
The compatible entropy vectors are,
\begin{equation*}
\Gamma^{*}_\mathcal{M} \! \left( \! C_3^{{\rm C}} \! \right)\! =\! \left\{ \! v \! \in \! \mathbb{R}^{7}_{\geq 0} | \exists P \! \in \! \mathcal{P}_\mathcal{M} \! \left( \! C_3^{{\rm C}} \! \right) \! \text{ s.t. } \! v \! = \! {\bf{H}}(P) \! \right\} \! ,
\end{equation*}
and $\overline{\Gamma^{*}_\mathcal{M}}\left(C_3^{{\rm C}}\right)$ is a
convex cone (cf.~\cite{our_review}).
The Shannon outer approximation\footnote{The outer approximation
obtained from all six variable Shannon inequalities and the
conditional independence equalities~\eqref{eq:indepentr}, which are
in this case $I(A\! : \! BCX)=0$, $I(X\! : \! AYZ| BC)=0$ and appropriate
permutations.},
$$\Gamma_\mathcal{M}\left(C_3^{{\rm C}}\right)=\left\{w\in
\Gamma_3 | M_\mathcal{M}\left(C_3^{{\rm C}}\right) \cdot w \geq 0
\right\},$$
was explicitly computed by Chaves et al.~\cite{Chaves2014,Chaves2015},
where $M_{\mathcal{M}}\left(C_3^{{\rm C}}\right)$ is the coefficient
matrix of the following three equivalence classes of inequalities
(where permutations of $X$, $Y$ and $Z$ lead to a total of $7$
inequalities):\footnote{Recall that an explicit linear description of
their entropy cone is generally only available for causal structures
with up to three nodes. In particular, such a description is not
available for $\overline{\Gamma^{*}}\left(C_3^{{\rm C}}\right)$, which
involves six nodes. Hence, it is impossible to directly compute
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$ with a
variable elimination algorithm.}
\begin{widetext}
\begin{equation}\label{eq:margindep}
\begin{split}
- H(X)-H(Y)-H(Z)+H(XY)+H(XZ) &\geq 0, \\
-5 H(X)-5H(Y)-5H(Z)+4H(XY)+4H(XZ)+4H(YZ)-2H(XYZ) &\geq 0, \\
-3 H(X)-3H(Y)-3H(Z)+2H(XY)+2H(XZ)+3H(YZ)-H(XYZ) &\geq 0.
\end{split}
\end{equation}
\end{widetext}
We now show that tighter outer approximations of the set of achievable
entropy vectors in the marginal scenario of the triangle,
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$, can be
derived by using non-Shannon type inequalities. However, there are
infinitely many such linear entropy inequalities. To restrict the
number of inequalities to be considered, the following reasoning can
be applied. As mentioned in Section~\ref{sec:classicalcone}, all known
non-Shannon entropy inequalities for four variables can be written as
the sum of the \emph{Ingleton quantity}~\eqref{eq:ingleton} and
(conditional) mutual information terms. Since the latter are always
positive, any non-Shannon inequality is irrelevant (i.e., implied by
existing ones) if the causal restrictions imply that the
Ingleton term is non-negative. This significantly reduces the choices
of variable sets for which the known additional inequalities may
be relevant.
\begin{example}
Consider Proposition~\ref{prop:zhangyeung} with $(X_1,~X_2,~X_3,~X_4)=(A,~B,~C,~X)$. The corresponding inequality is
\begin{equation*}
\begin{split}
I(A\! : \! B|C)+I(A\! : \! B|X)+I(C\! : \! X)-I(A\! : \! B)\\
+I(A\! : \! C|B)+I(B\! : \! C|A)+I(A\! : \! B|C) \geq 0.
\end{split}
\end{equation*}
Whenever a causal structure $C^{{\rm C}}$ implies $I(A\! : \! B)=0$, i.e.,
independence of $A$ and $B$, the above inequality is implied by the
Shannon inequalities and the independence constraint $I(A\! : \! B)=0$.
Hence it cannot improve our outer approximation.
\end{example}
The following proposition restricts the permutations of each non-Shannon inequality that may be relevant for the derivation of our improved approximations to $\overline{\Gamma^{*}_\mathcal{M}}\left(C_3^{{\rm C}}\right)$.
\begin{prop} \label{prop:ingletonperm}
Consider an entropy inequality on four variables
that enforces the non-negativity of a positive linear combination of
the Ingleton quantity~\eqref{eq:ingleton} and (conditional)
mutual information terms. This inequality is implied by the Shannon
inequalities
and the conditional independences of
$C_3^{{\rm C}}$ (i.e., $I(A\! : \! XBC)=0$, $I(X\! : \! YZA|BC)=0$ and appropriate permutations) for all choices of four out of the six involved random variables, except
\begin{equation*}
\begin{split}
\left( X_1,~X_2)~(X_3,~X_4 \right) &= \left( X,~Y)~(Z,~C \right)\\
&= \left( X,~Z)~(Y,~B \right) \\
&= \left( Y,~Z)~(X,~A \right),
\end{split}
\end{equation*}
up to exchange of $X_1$ and $X_2$ or exchange of $X_3$ and $X_4$.
\end{prop}
All known irredundant non-Shannon inequalities satisfy the conditions of this proposition. Note also that the application of non-Shannon inequalities to subsets of four out of the six random variables in $C_3^{{\rm C}}$ does not encompass all possible applications of these inequalities. Specifically, each inequality can also be applied to sets of five or to all six random variables, where the joint distribution of some sets of two or three random variables are interpreted as those of one of the four random variables in the non-Shannon inequality. We have not looked into such configurations.
\begin{proof}
For four random variables $X_1$, $X_2$, $X_3$ and $X_4$, the Ingleton inequality
\begin{eqnarray}
I(X_1\! \! : \! \! X_2|X_3) \! + \! I(X_1\! \! : \! \! X_2|X_4)&& + I(X_3\! \! : \! \! X_4) \nonumber\\
&&- \! I(X_1\! \! : \! \! X_2)\geq0,\label{eq:ingletproof}
\end{eqnarray}
can be equivalently rewritten in four more ways with the following equalities:
\begin{alignat}{2} \label{eq:rewritings}
\! I(X_1\! \! : \! \! X_2|X_3) \! - \! I(X_1\! \! : \! \! X_2)& \! = \! I(X_1\! \! : \! \! X_3|X_2)&& \! - \! I(X_1\! \! : \! \! X_3) \nonumber \\
& \! = \! I(X_2\! \! : \! \! X_3|X_1)&& \! - \! I(X_2\! \! : \! \! X_3), \nonumber \\
\! I(X_1\! \! : \! \! X_2|X_4) \! - \! I(X_1\! \! : \! \! X_2)& \! = \! I(X_1X_4|X_2)&& \! -\! I(X_1\! \! : \! \! X_4) \nonumber \\
&\! = \! I(X_2\! \! : \! \! X_4|X_1)&& \! - \! I(X_2\! \! : \! \! X_4) .
\end{alignat}
For the inequality~\eqref{eq:ingletproof} not to be implied by the Shannon inequalities and the conditional independences we need $X_1$, $X_2$, $X_3$ and $X_4$ to be such that
\begin{alignat}{3}\label{eq:infoterms}
&I(X_1\! : \! X_2)& &> 0, \nonumber \\
&I(X_1\! : \! X_3)& &> 0, \nonumber \\
&I(X_1\! : \! X_4)& &> 0, \\
&I(X_2\! : \! X_3)& &> 0, \nonumber \\
&I(X_2\! : \! X_4)& &> 0, \nonumber
\end{alignat}
hold simultaneously. If the conditional independences of $C_3^{{\rm C}}$
imply that one of these mutual informations is zero then the Ingleton
inequality can be expressed as a positive linear combination of
(conditional) mutual information terms in one of its five equivalent
forms and the corresponding non-Shannon inequality is redundant.
For the five constraints \eqref{eq:infoterms} to hold simultaneously,
$X_1$ and $X_2$ have to be correlated with one another as well as with
two further variables. This excludes the independent sources $A$, $B$
and $C$ as candidates for $X_1$ and $X_2$; therefore
$X_1,~X_2 \in \left\{X,~Y,~Z \right\}$. Furthermore, the variables
$X_3$ and $X_4$ have to be correlated with both, $X_1$ and $X_2$. This
excludes the two variables in $\left\{ A,~B,~C \right\}$ that do not
lie between $X_1$ and $X_2$ in $C_3^{{\rm C}}$. Hence, for each choice of
$X_1$ and $X_2$, the variables $X_3$ and $X_4$ have to be chosen as
the remaining element of $\left\{X,~Y,~Z \right\}$ and the variable
positioned opposite it in $C_3^{{\rm C}}$.
In summary, $\left(X_1,~X_2 \right) \left(X_3,~X_4 \right)$ can only be
$\left(X,~Y \right) \left(Z,~C \right)$, $\left(X,~Z \right)
\left(Y,~B \right)$ and $\left(Y,~Z \right) \left(X,~A \right)$ up to
permutations of the variables within a tuple.
\end{proof}
If we were to take one 4-variable non-Shannon inequality into account
and apply it to any subset of four out of the total of six random
variables in the causal structure, this would leave us with $360$
permutations of the inequality (if the inequality is not invariant
under the permutation of any of the four involved
variables). Proposition~\ref{prop:ingletonperm} reduces this to only
$12$ (potentially) irredundant permutations.
For each non-Shannon inequality, these $12$ permutations are
candidates for improving the outer approximation to
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$. We remark
here that for most known non-Shannon inequalities, several of these
$12$ permutations can be shown to be redundant\footnote{For instance,
if the non-Shannon inequality in question is invariant under the
permutation of some of its variables then some of the $12$
permutations are equivalent, or, if the marginalisation of different
permutations of the same inequality (that are not equal) imply the
same inequalities for the marginal scenario then some of these
inequalities may be redundant for our purposes.}. Despite
accounting for this reduction in the permutations of each inequality,
the number of different inequalities to be considered is infinite,
and any outer approximation to
$\overline{\Gamma^{*}_\mathcal{M}}\left(C_3^{{\rm C}}\right)$ could
(potentially) be tightened further by including additional
inequalities.
In principle, the more inequalities that are added, the better the
approximation to
$\overline{\Gamma^{*}_\mathcal{M}}\left(C_3^{{\rm C}}\right)$. However,
adding too many inequalities at a time renders the task of
marginalising infeasible. Applied to a system of $n_0$ inequalities
the Fourier-Motzkin algorithm can yield up to
$\left(\frac{n_0}{2}\right)^{2}$ inequalities in the first elimination
step. Iterating the procedure for $n$ steps produces up to
$4 \cdot \left(\frac{n_0}{4}\right)^{2^{n}}$ inequalities. To avoid
this double exponential behaviour the elimination algorithm can be
adapted by implementing a few rules to remove some of the many
redundant inequalities produced in each step. These rules are
collectively known as \u{C}ernikov
rules~\cite{Cernikov1960,Chernikov1965} and comprehensively explained
in~\cite{Bastrakov2015}. It is known, however, that the number of
necessary inequalities can still grow
exponentially~\cite{Monniaux2010}. That said, the worst case scaling
may not be exhibited in our case. In fact, the inequalities defining
$\Gamma\left(C_3^{{\rm C}}\right)$ contain few variables each and thus
lead to far fewer than the maximal number of inequalities. However,
computational resources still limit us to adding a relatively small
number of different supplementary inequalities to the standard Shannon
cone at a time.
We have used the previously outlined technique to compute tighter
outer approximations to
$\overline{\Gamma_\mathcal{M}^*}\left(C_3^{{\rm C}}\right)$, by including a
manageable number of non-Shannon inequalities at a time:\smallskip
\noindent \emph{Case 1}: We include the inequality
from Proposition~\ref{prop:zhangyeung} as well as all six inequalities
from~\cite{Dougherty2006} applied to all subsets of four out of the
six variables of $C_3^{{\rm C}}$. This leads to $45$ classes of
inequalities, of which $41$ are not part of the outer approximation $\Gamma_{\mathcal{M}}\left(C_3^{{\rm C}}\right)$.\smallskip
\noindent \emph{Case 2}: We include the inequalities of the form given
in~\eqref{eq:matus1} and~\eqref{eq:matus2} for $s=1,2,3$ and for all
subsets of four out of the six variables in $C_3^{{\rm C}}$. In this case,
we find $114$ classes of inequalities, of which $110$ are not part of
the outer approximation
$\Gamma_{\mathcal{M}}\left(C_3^{{\rm C}}\right)$.\smallskip
\noindent In each case, all classes (together with the number of
members in each class) are provided as Supplementary
Information. \smallskip
\noindent
We have compared our new approximations to the Shannon outer
approximation by sampling uniformly over the surface of the positive
sector of the unit hypersphere around ${\bf 0}$ in
$\mathbb{R}^7$~\cite{Muller1959}\footnote{I.e., from the set
$\{{\bf v}\in\mathbb{R}^7:v_i\geq 0, \sum_{i=1}^7v_i^2=1\}$.}. A
measure for the hyperdimensional solid angle included by these
approximations is given in terms of the fraction, $\alpha$, of points within the
respective cones. We have sampled $3.2\times10^{9}$ points
each, which led to the following estimates for $\alpha$: \smallskip
\noindent \emph{Shannon Cone}: $\alpha_S=(3.308 \pm 0.010)\times10^{-5}$.\smallskip
\noindent \emph{Case 1}: $\alpha_1=(3.090 \pm 0.010)\times10^{-5}$.\smallskip
\noindent \emph{Case 2}: $\alpha_2=(3.072 \pm 0.010)\times10^{-5}$. \smallskip
This shows that the difference between the three approximations it relatively small: the hyperdimensional solid angle encompassed by the cones of the Case~1 and Case~2 approximations are both roughly $93 \%$ of that of the Shannon cone. An explicit entropy vector that lies in the Shannon approximation, but not in either of the new outer approximations to $\overline{\Gamma^{*}_\mathcal{M}}\left(C_3^{{\rm C}}\right)$ is
$\left( \! H(X) \! , \! H(Y) \! , \! H(Z) \! , \! H(X\! Y) \! , \! H(X\! Z) \! , \! H(Y \! Z) \! , \! H(X \! Y \! Z) \! \right)\\
=\left( 11, 14, 14, 20, 20, 23, 28 \right)$.
We also derive some valid families of inequalities.
\begin{prop}\label{lemma:matus1}
All entropy vectors $v\in\Gamma^*_\mathcal{M}\left(C_3^{{\rm C}}\right)$ obey
\begin{align}
&\!\left(-\frac{1}{2} s^2 - \frac{3}{2} s \right) \left( H(X) +H(Z) \right)- \left(s+1\right) H(Y)\nonumber\\
+&\!\left( \frac{1}{2} s^{2} + \frac{3}{2} s + 1 \right) \left( H(XY) + H(YZ) \right)\nonumber\\
+&s(s+2)H(XZ)-\left(s+1\right)^2 H(XYZ) \! \geq \! 0,\label{eq:marginalmatus1}\\ \nonumber\\
&\!\left( \! \! - \frac{1}{2} s^2 \! - \! \frac{3}{2} s \! - \! 2 \! \right) \! \left( H(X) \! + \! H(Y) \! + \! H(Z) \! - \! H(XY) \right) \nonumber\\
+&\! \left(\frac{1}{2} s^2 + \frac{3}{2} s +1\right)H(XZ)
+\left( s + 2 \right) H(YZ) \nonumber\\
-&\! \left( s+1 \right) H(XYZ) \geq 0,\label{eq:mat1fam2}
\end{align}
\begin{align}
&\! \left(-\frac{1}{2} s^2 - \frac{3}{2} s -2 \right) \left( H(X) +H(Z)-H(XY) \right)\nonumber\\
-& \! \left(2s+2\right) H(Y)+\left( s^{2} + 2 \right) H(XZ) \nonumber\\
+& \! \left( \! \frac{1}{2} s^{2} \! + \! \frac{3}{2} s \! + \! 1 \! \right) H(YZ)
- \left( \! s^2 \! + \! 1 \! \right) H(XYZ) \! \geq \! 0,\label{eq:mat2}
\end{align}
for all $s\in\mathbb{N}$. The same holds for all permutations of $X$, $Y$ and $Z$.
\end{prop}
The proof of this proposition can be found in Appendix~\ref{sec:families}.
Further families of inequalities can be derived by separately considering different inequalities from a family, e.g.\ the same permutation of~\eqref{eq:matus1} for each $s\in\mathbb{N}$, and combining them with the same Shannon inequalities to obtain new constraints on the marginal scenario by means of the Fourier-Motzkin elimination algorithm. Tighter inequalities are often obtained by combining several permutations of an inequality~\eqref{eq:matus1}.
Combining instances of~\eqref{eq:matus1} for several $s\in\mathbb{N}$
leads to an even larger number of new inequalities, which render many
of the families derived with the previously explained method
redundant. For the few orders $s$ up to which we were able to run our
calculations, the families~\eqref{eq:marginalmatus1}
and~\eqref{eq:mat1fam2} from Proposition~\ref{lemma:matus1} were the
only two for which none of the inequalities were implied by others.
Similar considerations can be applied to \eqref{eq:matus2} (from
which~\eqref{eq:mat2} is derived) and to further families of
inequalities~\cite{Dougherty2011}.
One might imagine that adding genuine five and six variable
inequalities to $\Gamma\left(C_3^{{\rm C}}\right)$ leads to further
entropy inequalities for $C_3^{\rm C}$. It turns out that applying the
five and six variable inequalities from~\cite{Zhang1998,
Makarychev2002} to five and six variables of the triangle causal
structure respectively does not lead to a tighter outer approximation
to $C_3^{{\rm C}}$ than the inequality from Proposition~\ref{prop:zhangyeung}. This can be shown by expanding the inequalities into a
linear combination of mutual information terms and applying a similar
reasoning to that in the proof of
Proposition~\ref{prop:ingletonperm}. As they are not particularly
instructive, the technical details of these arguments are omitted
here. The same is not known to hold for the inequality derived
in~\cite{Zhang2003}.
\begin{conj}\label{conj:matusconj}
Infinitely many linear inequalities are needed to characterise
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$.
\end{conj}
Our main evidence for this is that the families of
inequalities~\eqref{eq:matus1}, used by Mat\'{u}\v{s} to prove that
the analogue of this conjecture holds for $\overline{\Gamma^{*}_4}$,
lead to infinite families of inequalities for $C_3^{{\rm C}}$ after
marginalising (cf.\ Proposition~\ref{lemma:matus1}). The curve
constructed by Mat\'{u}\v{s} in Ref.~\cite{Matus2007} to prove his
statement for $\overline{\Gamma^{*}_4}$ can be adapted to our
scenario, which can be used to show that the
inequalities~\eqref{eq:marginalmatus1} are independent. However, we
were not able to show that this curve can be realised with entropy
vectors that are compatible with the triangle causal structure, and
hence we cannot exclude the possibility that the marginal cone
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$ is polyhedral.
The infinite families of inequalities (cf.\
Proposition~\ref{lemma:matus1}) that we obtained from Mat\'{u}\v{s}'s
original family of inequalities may indicate that this region of
entropy space retains a non-polyhedral segment after the causal
constraints are included and the set is projected to the marginal
scenario. However, it could be that non-polyhedral boundary regions do
not survive the mapping to entropy vectors for $C_3^{{\rm C}}$. If this
were the case then (most of) our infinite set of inequalities would be
rendered redundant by another inequality.
\subsection{Application of non-Shannon inequalities to various causal structures} \label{sec:int}
The concept of a generalised DAG was introduced in~\cite{Henson2014},
the idea being to have a framework in which classical, quantum and
even more general systems can be shared by unobserved nodes. For the
details, we refer to the original paper. The part that is of interest
here is that the authors of Ref.~\cite{Henson2014} list $21$
generalised DAGs with up to six nodes for which there may be a
separation between the correlations realisable classically and quantum
mechanically, i.e., between $\mathcal{P}_{\mathcal{M}}(C^{\rm C})$ and
$\mathcal{P}_{\mathcal{M}}(C^{\rm Q})$~\cite{Henson2014,Pienaar2016}.\footnote{Note that
there are further causal structures with $5$ and $6$ nodes that have
this property, which can all be reduced to these $21$ examples with
rules specified in~\cite{Henson2014}.} We analyse
these from an entropic perspective, looking for a causal structure $C$
in which there is a separation between
$\overline{\Gamma^*_\mathcal{M}}\left(C^{{\rm C}}\right)$ and
$\overline{\Gamma^*_\mathcal{M}}\left(C^{{\rm Q}}\right)$.
Among these structures there are three that have fewer than six nodes, displayed in Figure~\ref{fig:HensonStructures}.
\begin{figure}
\centering
\resizebox{0.98 \columnwidth}{!}{%
\begin{tikzpicture}[scale=0.58]
\node (a) at (-10,1) {$(a)$};
\node[draw=black,circle,scale=0.75] (Z1) at (-5,-2) {$Y$};
\node[draw=black,circle,scale=0.75] (Y1) at (-7,-2) {$Z$};
\node[draw=black,circle,scale=0.75] (X1) at (-9,-2) {$X$};
\node (A1) at (-6,0.5) {$A$};
\node (b) at (-3,1) {$(b)$};
\node[draw=black,circle,scale=0.75] (W) at (-2,-2) {$W$};
\node[draw=black,circle,scale=0.75] (X) at (-1,0.5) {$X$};
\node (A) at (0,-2) {$A$};
\node[draw=black,circle,scale=0.75] (Y) at (1,0.5) {$Y$};
\node[draw=black,circle,scale=0.75] (Z) at (2,-2) {$Z$};
\node (c) at (3.5,1) {$(c)$};
\node[draw=black,circle,scale=0.75] (Y2) at (4.5,0.5) {$Y$};
\node[draw=black,circle,scale=0.75] (Z2) at (6.5,0.5) {$Z$};
\node[draw=black,circle,scale=0.75] (X2) at (5.5,-0.75) {$X$};
\node (A2) at (4.5,-2) {$A$};
\node (B2) at (6.5,-2) {$B$};
\node (00) at (7.5,-2) { };
\draw [->,>=stealth] (A1)--(Y1);
\draw [->,>=stealth] (A1)--(Z1);
\draw [->,>=stealth] (Y1)--(Z1);
\draw [->,>=stealth] (X1)--(Y1);
\draw [->,>=stealth] (W)--(X);
\draw [->,>=stealth] (A)--(X);
\draw [->,>=stealth] (A)--(Y);
\draw [->,>=stealth] (Z)--(Y);
\draw [->,>=stealth] (A2)--(X2);
\draw [->,>=stealth] (B2)--(Z2);
\draw [->,>=stealth] (B2)--(X2);
\draw [->,>=stealth] (A2)--(Y2);
\draw [->,>=stealth] (X2)--(Z2);
\draw [->,>=stealth] (X2)--(Y2);
\end{tikzpicture}
}%
\caption{Three causal structures, $C$, for which the outer approximation, $\Gamma_\mathcal{M}(C^{{\rm C}})$ tightly approximates the classical entropy cone $\overline{\Gamma^{*}_\mathcal{M}}(C^{{\rm C}})$, which also coincides with $\overline{\Gamma^{*}_\mathcal{M}}(C^{{\rm Q}})$. The observed variables are labelled $W$, $X$, $Y$ and $Z$, the unobserved nodes are called $A$ and $B$.}
\label{fig:HensonStructures}
\end{figure}
For these three, we find that the vertices of the corresponding
Shannon cone, $\Gamma_\mathcal{M}(C^{{\rm C}})$, are achievable with entropy
vectors of classical probability distributions compatible with the
causal structure, from which it follows that this cone is equal to the
entropy cone $\overline{\Gamma^*_\mathcal{M}}(C^{{\rm C}})$. (This can also be
shown by computing an inner approximation to the corresponding entropy
cones and showing that the inner and outer approximations coincide,
e.g.\ by employing linear rank inequalities as outlined in
Section~\ref{sec:inner}.) Our results also imply that the
consideration of non-Shannon inequalities cannot lead to any further
constraints in these three causal structures. In the following, we
furthermore show that there is no entropic separation between
classical and quantum versions of these causal structures.
\begin{prop}\label{prop:hen}
Let $C$ be any of the causal structures shown in
Figure~\ref{fig:HensonStructures}. Then
$\overline{\Gamma^*_\mathcal{M}}(C^{{\rm C}})=\overline{\Gamma^*_\mathcal{M}}(C^{{\rm Q}})$.
\end{prop}
\begin{remark}
Note that there are causal structures involving up to five variables
that reduce to those shown in Figure~\ref{fig:HensonStructures}
under the reduction rules from~\cite{Henson2014}. Our proof does
not rule out that these exhibit a classical to quantum
separation.
\end{remark}
\noindent Further details, including the proof of Proposition~\ref{prop:hen} are given in Appendix~\ref{sec:uptofive}.
The $18$ remaining example causal structures involve six
variables. For all of them we have found that several instances of the
non-Shannon inequality from Proposition~\ref{prop:zhangyeung} lead to
tighter entropic constraints for the classical marginal scenarios than
those listed in~\cite{Henson2014}.
For the causal structures with four observed variables, instances of
this inequality are relevant even without considering the unobserved
nodes. These instances thus hold whether or not the unobserved nodes
are classical or quantum. Hence, they allow us to tighten the outer
approximations to the sets of achievable entropy vectors in both
cases, in contrast to non-Shannon inequalities that are applied
to unobserved variables classically (for which the quantum analogue is not
known to hold).\footnote{Note that this reasoning is not restricted to
distinguishing classical and quantum, but, it may also apply to the
comparison of different causal structures with the same set of
observed variables. While non-Shannon inequalities derived from
unobserved variables may lead to a separation between the two causal
structures, non-Shannon inequalities valid only for the observed
variables may not.}
The above considerations have not enabled us to show a
separation between the achievable entropy vectors in the classical and
quantum cases, hence we are left with the following open problem.
\begin{prob}\label{prob:oooo}
Find a causal structure $C$ with a set of observed nodes $\mathcal{M}$ in
which the sets $\overline{\Gamma^*_\mathcal{M}}(C^{\rm C})$ and
$\overline{\Gamma^*_\mathcal{M}}(C^{{\rm Q}})$ are provably different, or show
that this can never occur.
\end{prob}
\subsection{Application of non-Shannon inequalities with post-selection}\label{sec:post_selected_nonShan}
In the discussion so far we have not considered a related technique
that allows for post-selection on particular outcomes of certain
variables. The idea of doing this first appeared
in~\cite{Braunstein1988} based on results by Fine~\cite{Fine1982,
Fine1982a} and was later generalised~\cite{Chaves2012, Fritz2013,
Chaves2013, Chaves2015, Pienaar2016, Chaves2016}. We refer
to~\cite{our_review} for an explanation of this technique.
Here we illustrate that non-Shannon inequalities can be used in
combination with post-selection by discussing a specific example
relevant for information causality~\cite{Pawlowski2009}. Information
causality is an information theoretic principle obeyed by classical
and quantum physics but not by general probabilistic theories in which
there are correlations that violate Tsirelson's
bound~\cite{Cirelson1980}, e.g.\ generalized no signalling
theory~\cite{Barrett07}, which allows PR-Boxes as a
resource~\cite{Tsirelson1993, Popescu1994FP}. The principle is stated
in terms of the optimal performance of two parties in a game, which we
describe below, and is quantified in terms of an entropic quantity.
Alice holds two pieces of information\footnote{In general the game is
formulated for more, but we restrict to two here for simplicity.},
$X_0$ and $X_1$, she can send classical information $Z$ to Bob, who is
later given a message $R$ indicating whether he should guess $X_0$ or
$X_1$. Bob's guess is denoted $Y$. Alice and Bob are able to use a
pre-shared resource (depicted as $A$) to help them. The relevant
causal structure of the game is displayed in
Figure~\ref{fig:IC_inner}(a) and it is often analysed after
post-selecting on the value of $R$, which can be done using the
causal structure of Figure~\ref{fig:IC_inner}(b) (note that in the
quantum case the variables $Y_{|R=0}$ and $Y_{|R=1}$ do not coexist,
so it doesn't make sense to consider ${\rm IC}^{\rm Q}_{\mathrm{R}}$; instead a
restricted set of entropies needs to be considered -- see later). A
theory is said to obey information causality if for all pre-shared
resources allowed by the theory,
$I(X_0\! \! : \! \! Y_{|R=0}) \! + \! I(X_1\! \! : \! \! Y_{|R=1}) \leq H(Z)+I(X_0\! : \! X_1)$.
A stronger set of entropic constraints for this causal structure were
found in~\cite{Chaves2015}, including the relation
\begin{align}
I(X_0\! : \! Z Y_{\rm |R=0}) \! + \! &I(X_1\! : \! Z Y_{\rm |R=1}) \! + \! I(X_0\! : \! X_1 | Z Y_{\rm |R=1}) \nonumber\\
&\leq H(Z) +I(X_0 \! : \! X_1) \, , \label{eq:IC_known}
\end{align}
which holds for both classical and quantum shared
resources.\footnote{Because the existence of a joint distribution of
$Y_{\rm | R=0}$ and $Y_{\rm | R=1}$ with appropriate marginals is
not clear in the quantum case, the two variables have
to be interpreted as alternatives and are part of different
coexisting sets. Therefore, the analysis of ${\rm IC}_{\rm R}^{\rm C}$ does not carry over to the quantum case, but a separate analysis is required there.}
\begin{figure}
\centering
\resizebox{1\columnwidth}{!}{%
\begin{tikzpicture} [scale=0.8]
\node (A0) at (-8.2,1.5) {$(a)$};
\node[draw=black,circle,scale=0.7] (A1a) at (-7.5,1) {$X_0$};
\node[draw=black,circle,scale=0.7] (A1) at (-7.5,-0.2) {$X_1$};
\node[draw=black,circle,scale=0.75] (A2) at (-5.5,-0.2) {$Z$};
\node[draw=black,circle,scale=0.75] (A3) at (-3.5,-0.2) {$Y$};
\node[draw=black,circle,scale=0.75] (A5) at (-3.5,1) {$R$};
\node (A4) at (-4.5,1.2) {$A$};
\draw [->,>=stealth] (A1)--(A2);
\draw [->,>=stealth] (A1a)--(A2);
\draw [->,>=stealth] (A2)--(A3);
\draw [->,>=stealth] (A4)--(A2);
\draw [->,>=stealth] (A4)--(A3);
\draw [->,>=stealth] (A5)--(A3);
\draw [->,>=stealth] (A1a)--(A1);
\node (B0) at (-1.7,1.5) {$(b)$};
\node[draw=black,circle,scale=0.7] (B1a) at (-1,1) {$X_0$};
\node[draw=black,circle,scale=0.7] (B1) at (-1,-0.2) {$X_1$};
\node[draw=black,circle,scale=0.75] (B2) at (1,-0.2) {$Z$};
\node (B4) at (2.3,1.5) {$A$};
\node[draw=black,circle,scale=0.6] (B3) at (3.4,-0.2) {$Y_{|R=0}$};
\node[draw=black,circle,scale=0.6] (B3a) at (2.3,0.4) {$Y_{|R=1}$};
\draw [->,>=stealth] (B1)--(B2);
\draw [->,>=stealth] (B1a)--(B2);
\draw [->,>=stealth] (B1a)--(B1);
\draw [->,>=stealth] (B4)--(B2);
\draw [->,>=stealth] (B4)--(B3);
\draw [->,>=stealth] (B4)--(B3a);
\draw [->,>=stealth] (B2)--(B3);
\draw [->,>=stealth] (B2)--(B3a);
\end{tikzpicture}
}%
\caption{(a) Causal structure underlying the Information Causality
game, ${\rm IC}$. Alice holds a database, here made up of two bits $X_0$
and $X_1$. These need not be independent, which is expressed by a
potential causal influence from $X_0$ to $X_1$. She is then allowed
to send a message $Z$ to Bob, who, depending on which bit $R$ a
referee asks for, takes a guess $Y$ of either $X_0$ or $X_1$. Alice
and Bob may have shared some resources (represented by $A$) before
performing the protocol, either some classical randomness, a quantum
system, or a resource from a more general non-signalling theory, which Alice may
use in order to choose her message and Bob may use to make his
guess. (b) The effective causal structure of the Information
Causality game after post-selecting on binary $R$, labelled ${\rm IC_R}$. This
causal structure shares some of its marginal distributions with
conditional distributions of ${\rm IC}$, i.e., if we use $P$ for the
distribution in ${\rm IC_R}$ and $Q$ for that in ${\rm IC}$ then
$P_{X_0 X_1 Z Y_{| R=r}}=Q_{X_0 X_1 Z Y | R=r}$ for $r=0,1$.}
\label{fig:IC_inner}
\end{figure}
We show that using non-Shannon inequalities leads to a tighter outer
approximation of the information causality scenario in the case of a
classical shared resource. Considering just the inequality from
Proposition~\ref{prop:zhangyeung} (and permutations) has led us to
derive a total of $265$ classes of entropy inequalities, including the
$52$ classes that were obtained without non-Shannon constraints
in~\cite{Chaves2015} (a list of all $265$ classes together with the
number of representatives of each class is available as Supplementary
Information). Moreover, we expect further non-Shannon inequalities to
lead to numerous additional constraints potentially rendering our
inequalities redundant. In principle, infinite families of
inequalities, similar to those found in Proposition~\ref{lemma:matus1}
for the triangle scenario could also be derived here.
In the quantum case, we can only apply the non-Shannon inequalities to
the two coexisting sets of exclusively classical variables
$\left\{ X_0, X_1, Z , Y_{\rm | R=0} \right\}$ and
$\left\{ X_0, X_1, Z , Y_{\rm | R=1} \right\}$, which means that we
can impose a set of $24$ additional constraints (including
permutations) just by adding all permutations of the inequality from
Proposition~\ref{prop:zhangyeung} to the outer approximation that is
obtained without these (no further variable elimination is required).
It is worth pointing out that although our results (in the form of new
inequalities) imply that previous entropic characterisations of
${\rm IC_R}$ were not tight, the inequality~\eqref{eq:IC_known} is not
rendered redundant by our new inequalities.
\section{Inner approximations to the entropy cones of causal structures}\label{sec:inner}
To complement the outer approximations, it is sometimes useful to
consider inner approximations to the entropy cones of causal
structures. This is particularly useful when one can show that inner
and outer approximations coincide, as they then identify the actual
boundary of the entropy cone. Examples for this are the three causal
structures of Figure~\ref{fig:HensonStructures}, also discussed in the
previous section. Hence, inner and outer approximation together serve
as a relatively simple means to identify the boundary of certain
entropy cones. Such findings also immediately imply that non-Shannon
inequalities are irrelevant for improving on the outer approximation
to the entropy cone for the causal structure in question.
Furthermore, we can often find inner approximations that share
extremal rays with the outer approximations derived from the Shannon
and independence constraints (even when the two do not coincide). They
hence allow us to identify the regions of entropy space where our
approximations are tight and those regions where there is a gap
between inner and outer approximation.\footnote{Such a comparison of
inner and outer approximations can be performed for the entropy cone
of a causal structure including its unobserved variables, i.e.,
before marginalisation, as well as for the respective approximations
to its marginal cone, which we are mainly interested in here.} Such
a gap can be explored, e.g. by using non-Shannon inequalities, as was
explained in the previous section.
Inner approximations also serve as a tool to decide whether entropy
vectors are suitable for certifying the unattainability of particular
distributions that are suspected not to be achievable within the
causal structure at hand. If such a distribution leads to an entropy
vector within an inner approximation to the entropy cone in question,
this means either that the distribution is in fact achievable within
the causal structure or that the causal structure allows for another
distribution with the same entropy vector (or an arbitrarily good
approximation of such). Hence, to determine whether the distribution
in question is achievable, switching to a more fine-grained method
(see for example~\cite{Pienaar2016, Wolfe2016}) is
necessary.
In the following we show how inner approximations can be found in
different scenarios.
\subsection{Techniques to find inner approximations for causal structures with up to five observed variables}
For a causal structure, $C$, that involves a total of four or five
variables, inner approximations to its entropy cone can be derived
from $\Gamma_4^{I}$ or $\Gamma_5^{I}$ respectively (as defined in
Section~\ref{sec:classicalcone}) combined with the conditional
independence constraints of $C^{{\rm C}}$, which together constrain a cone
$\Gamma^{I}\left(C^{{\rm C}}\right)$. An inner approximation to the
corresponding marginal scenarios,
$\Gamma^{I}_{\mathcal{M}}\left(C^{{\rm C}}\right)$, is then obtained from
$\Gamma^{I}\left(C^{{\rm C}}\right)$ with a Fourier-Motzkin elimination,
like for outer approximations. It is guaranteed that
$\Gamma_\mathcal{M}^{I}\left(C^{{\rm C}}\right)$ is an inner approximation
to $\overline{\Gamma_\mathcal{M}^{*}}\left(C^{{\rm C}}\right)$, as it is a
projection of an inner approximation
$\Gamma^{I}\left(C^{{\rm C}}\right) \subseteq
\overline{\Gamma^{*}}\left(C^{{\rm C}}\right)$.
Hence, inner approximations can be straightforwardly computed for such
causal structures. Examples where this applies are the three causal
structures of Figure~\ref{fig:HensonStructures}.
\begin{example}[Inner approximation to the instrumental
scenario.] \label{example:innerI} For the classical instrumental
scenario, $C_{\rm I}$ of Figure~\ref{fig:HensonStructures}(a), we
can compute an inner approximation by adding the conditional
independence constraints $I(A\! : \! X)=0$ and $I(X\! : \! Y|AZ)=0$ to the
Ingleton cone $\Gamma^I_4$, as prescribed above. We can, however,
also directly prove that
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_{\rm
I}^{{\rm C}}\right)=\Gamma_{\mathcal{M}}\left(C_{\rm I}^{{\rm C}}\right)$
by showing that all permutations of the Ingleton inequality are implied by Shannon and
conditional independence constraints and, hence, inner and outer
approximations coincide for $C_{\rm I}^{{\rm C}}$. Since $I(A\! : \! X)=0$,
$I_{\rm ING}\left(A,X; \! Y,Z \right) \geq 0$ is immediately implied by Shannon and
independence constraints. Furthermore, the rewritings of $I_{\rm ING}$
according to \eqref{eq:rewritings} imply that $I(A\! : \! X)=0$
which (together with the Shannon inequalities) implies all permutations of the
Ingleton inequality except for
$I_{\rm ING}\left(Y,Z; \! A,X \right) \geq 0$. We can rewrite
\begin{align*}
&I_{\rm ING}\left(Y,Z; \! A,X \right)\\
&= \! I(Y\! \! : \! \! Z|A) \! + \! I(Y\! \! : \! \! X|Z) \! + \! I(X\! \! : \! \! A|Y) \! - \! I(X\! \! : \! \! Y|A) \\
&= \! I(Y\! \! : \! \! X|Z) \! + \! I(X\! \! : \! \! A|Y) \! + \! I(Y\! \! : \! \! Z|A \! X) \! \! - \! \! I(X\! \! : \! \! Y|A \! Z),
\end{align*}
the positivity of which is hence implied by the Shannon inequalities and the independence constraint ${I(X\! : \! Y|AZ)=0}$.
\end{example}
Implementing all relevant linear rank inequalities of four and five
variables (which includes their permutations and the application of
the Ingleton inequality to each four variable subset as well as
grouping several variables to one)~\cite{Dougherty2009} and then
performing a variable elimination may be impractical and
computationally challenging for certain causal structures.
Furthermore, for causal structures that involve more than five nodes
not all possible linear rank inequalities are known and their number
may even be infinite~\cite{Dougherty2014}. It is therefore useful to
derive inner approximations by other methods. For a causal structure,
$C$, the following methods are examples of how to derive inner
approximations, $\Gamma_{\mathcal{M}}^{I}\left(C^{{\rm C}}\right)$:
\begin{itemize}
\item Construct (random) entropy vectors from distributions compatible with $C^{{\rm C}}$ and take their convex hull.
\item Take the vertices of $\Gamma_{\mathcal{M}}\left(C^{{\rm C}}\right)$ that are reproducible with distributions compatible with the causal structure, their convex hull is an inner approximation.
\item Take the outer approximation to the classical causal structure, $\Gamma\left(C^{{\rm C}}\right)$, as a starting point and add a manageable number of linear rank inequalities to derive further constraints. These inequalities may be employed either before or after marginalising, which leads to different cones.\footnote{If for instance all linear rank inequalities in up to $k$ observed variables are added after marginalisation, the resulting cone corresponds to the intersection $\Gamma_{\mathcal{M}}\left(C^{{\rm C}}\right) \cap \Gamma^{I}_{k}$, where $k$ is the number of observed variables.} The convex hull of the reproducible rays is an inner approximation.
\end{itemize}
For the three examples of Figure~\ref{fig:HensonStructures} it is rather straightforward to recover all extremal rays of the outer approximation to the marginal scenario, $\Gamma_{\mathcal{M}}\left(C^{{\rm C}}\right)$ (cf.\ also Appendix~\ref{sec:uptofive}), i.e., the second method above is effective.
Overall, we found that whenever the extremal rays are not all straightforwardly recovered, the third method is effective. This is our preferred technique because by starting out with extremal rays of the Shannon cone we obtain approximations that in some regions are already tight (as opposed to the first method), and, at the same time adding linear rank inequalities helps us identify those extremal rays that are likely to be reproducible with distributions in $C^{{\rm C}}$ (this may help us avoid dropping reproducible rays in some situations).
The entropy cones obtained in this way are not necessarily inner approximations, and, if they are, they have to be proven as such, for example by explicitly constructing distributions that reproduce entropy vectors on each of the extremal rays (as with the second method above). However, in all our examples this method allowed us to recover a cone of which all extremal rays were easily seen to be reproducible after adding only few linear rank inequalities to $\Gamma\left(C^{{\rm C}}\right)$.
(If this were not the case one could still drop several irreproducible rays from the resulting cones to obtain an inner approximation.)
The method is illustrated in the example below.
We also remark here that in order to improve on inner approximations obtained with the second or third method above, the first method is applicable.
\begin{example}\label{example:IC_inner}
Consider the classical causal structure of
Figure~\ref{fig:IC_inner}(a) and remove the node $R$ to give a
$5$-variable causal structure, $\hat{\rm IC}^{{\rm C}}$. We can in
principle consider all linear rank inequalities of five random
variables combined with all Shannon inequalities and the conditional
independence constraints, which would give us an inner
approximation, $\Gamma^{I}_{\mathcal{M}}\left(\hat{{\rm IC}}^{{\rm C}}\right)$,
to the entropy cone,
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(\hat{\rm IC}^{{\rm C}}\right)$. This
procedure would involve a (impractically) large number of
inequalities.
Instead, we can consider the outer approximation in terms of Shannon
inequalities and conditional independence constraints,
$\Gamma_{\mathcal{M}}\left(\hat{\rm IC}^{{\rm C}}\right)$, and intersect this
cone with the Ingleton cone for the four observed variables,
$\Gamma^{I}_4$, i.e., we add all permutations of the Ingleton
inequality for the four observed variables to
$\Gamma_{\mathcal{M}}\left(\hat{\rm IC}^{{\rm C}}\right)$. This is easily
obtained but does not result in any restrictions beyond those of the
Shannon outer approximation, which is characterised by $52$ extremal
rays.
Adding the Ingleton inequality for all subsets of four out of the
five random variables to $\Gamma\left(\hat{\rm IC}^{{\rm C}}\right)$
before performing the variable elimination, only $46$ extremal rays
are recovered. These are straightforward to reproduce with entropy
vectors in $\hat{\rm IC}^{{\rm C}}$.\footnote{We can show that the $6$
extremal rays of the Shannon cone that are not part of this inner
approximation are not achievable in $\hat{\rm IC}^{{\rm C}}$, because they
violate the entropy inequalities we obtain when taking non-Shannon
inequalities into account in the computation of the outer
approximations to
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(\hat{\rm IC}^{{\rm C}}\right)$.} A
detailed exposition of this is presented in
Appendix~\ref{sec:example}.
\end{example}
This method can also be applied to causal structures with more than five variables.
For the
first few causal structures
from~\cite{Henson2014} we have recovered inner approximations by adding the Ingleton inequality to $\Gamma_{\mathcal{M}}\left(C^{{\rm C}}\right)$, i.e., by taking the intersection $\Gamma_{\mathcal{M}}\left(C^{{\rm C}}\right) \cap \Gamma_k^{I}$ (the extremal rays as well as distributions
recovering entropy vectors on each of them are available
as Supplementary Information).
In the following we give a detailed analysis of the inner approximation to the triangle causal structure and compare this to the outer approximations presented in previous sections.
\subsection{Example: Inner approximation to $\overline{\Gamma^{*}_\mathcal{M}}\left(C_3^{{\rm C}}\right)$}
Here, we derive an inner approximation to the entropy cone compatible with $C_3^{{\rm C}}$. An inner approximation to $\overline{\Gamma^{*}_6}$ in terms of linear rank inequalities is not available (see also Section~\ref{sec:classicalcone}).
Nonetheless, we are able to derive an inner approximation to
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$ by relying on
Ingleton's inequality. In the following, we apply
\eqref{eq:ingletoninequ} to any subset of four out of the six random
variables of $C_3^{{\rm C}}$ and take all their permutations into account.
We concisely write these inequalities in a matrix $M_\mathrm{I}$ and
consider the cone
\begin{equation*}
\Gamma^\mathrm{I}\! \left(C_3^{{\rm C}}\right) \! := \! \left\{v \in \Gamma_6 | M_\mathrm{CI}\! \left(C_3^{{\rm C}}\right) \cdot v = 0, \ M_\mathrm{I} \cdot v \geq 0 \right\} \! .
\end{equation*}
When marginalising this cone we obtain
\begin{equation*}
\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right):= \left\{w \in \Gamma_3 | M_{\mathrm{I},\mathcal{M}}\left(C_3^{{\rm C}}\right) \cdot w \geq 0 \right\},
\end{equation*}
where $M_{\mathrm{I},\mathcal{M}}\left(C_3^{{\rm C}}\right)$ contains only
one inequality,\footnote{Inequality~\eqref{eq:intinfo} renders the
three Shannon inequalities of the form $I(X\! : \! Y|Z)\geq 0$
redundant. $\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)$ is
thus fully characterised by the six remaining three variable Shannon
inequalities (constraining $\Gamma_3$) and \eqref{eq:intinfo}. }
\begin{equation}\label{eq:intinfo}
-I(X\! : \! Y\! : \! Z) \geq 0.
\end{equation}
This relation can also be analytically derived from the Ingleton
inequality and the conditional independence constraints of
$C_3^{{\rm C}}$. \footnote{The proof proceeds as follows. There are only
three instances of the Ingleton inequality that are not implied by
the conditional independences and the Shannon inequalities (cf.\
also Proposition~\ref{prop:ingletonperm}). The independence
constraint ${I(X\! : \! Y|C)=0}$ and its permutations ${I(X\! : \! Z|B)=0}$ and
${I(Y\! : \! Z|A)=0}$ lead to \eqref{eq:intinfo} in all three cases.}
\begin{prop} \label{prop:triangleingleton}
$\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)$ is an inner approximation to the marginal entropy cone of the triangle causal structure,
\begin{equation*}
\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)\subsetneq \overline{\Gamma^*_\mathcal{M}}\left(C_3^{{\rm C}}\right).
\end{equation*}
\end{prop}
The proof of Proposition~\ref{prop:triangleingleton} is deferred to Appendix~\ref{sec:inner_appendix}. $\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)$ provides a certificate for vectors to be realisable as entropy vectors in $C_3^{{\rm C}}$: if a vector $v \in \mathbb{R}^{7}$ obeys all Shannon constraints as well as~\eqref{eq:intinfo}, then it lies in
$\overline{\Gamma_\mathcal{M}^*}\left(C_3^{{\rm C}}\right)$. Compared to
the different outer approximations to
$\overline{\Gamma_\mathcal{M}^*}\left(C_3^{{\rm C}}\right)$ analysed in
Section~\ref{sec:non_shannon} the hyperdimensional solid angle for
this inner approximation is considerably smaller. Sampling over the
unit hypersphere around ${\bf 0}$ in
$\mathbb{R}^7$ as before (meaning $3.2\times10^9$ samples), we obtain $\alpha_I=(2.147\pm0.008)\times10^{-5}$.
It is worth emphasising that not all correlations whose entropy
vectors lie in $\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)$
can be realised in $C_3^{{\rm C}}$. Instead, if ${\bf
H}(P)\in\Gamma_\mathcal{M}^\mathrm{I}$ then there exists
$P'\in\mathcal{P}_\mathcal{M}(C_3^{{\rm C}})$ such that ${\bf H}(P')={\bf H}(P)$. The correlations of Figure~\ref{fig:fritzproof}, realised in the quantum version of the triangle causal structure, $C_3^{{\rm Q}}$, which will be considered in detail in Section~\ref{sec:quantcorr}, are one such example.
\label{sec:w_correl}
These are not in $\mathcal{P}_\mathcal{M}(C_3^{{\rm C}})$, but their entropy
vector nevertheless satisfies~\eqref{eq:intinfo}. Our argument
implies that there must be another distribution realisable in $C_3^{{\rm C}}$ with the same entropy vector.\footnote{Another such example is given by the distribution
$$P_{XYZ}(x,y,z)=\left\{\begin{array}{cl}\frac{1}{3}&\text{ }
(x,y,z)=(1,0,0),(0,1,0),(0,0,1)\\
0&\text{ otherwise,}\end{array}\right.$$ which is not compatible
with $C_3^{{\rm C}}$, as shown in~\cite{Wolfe2016}.}
\section{Non-Shannon inequalities in the quantum and hybrid triangle causal structures} \label{sec:quantumtriangle}
In this section, we compare classical and quantum versions of the triangle
causal structure (the distinction reflecting the nature of the
unobserved nodes). We also consider hybrid scenarios, in which some of the unobserved systems are restricted to be classical
while others are quantum. These turn out to be insightful for
understanding the gap between classical and quantum causal structures.
We also analyse whether non-Shannon inequalities lead to improved entropic
characterisations in these cases.
\subsection{Quantum triangle scenario}\label{sec:quantcorr}
It was first shown in Ref.~\cite{Fritz2012}, that there are joint distributions among the three observed variables $X$, $Y$ and $Z$ in $C_3^{{\rm Q}}$ that cannot be reproduced in $C_3^{{\rm C}}$, based on the CHSH scenario (see Figure~\ref{fig:fritzproof} and Appendix~\ref{sec:fritzproof} for the details).
Hence $C_3^{{\rm Q}}$ might also lead to a larger set of compatible entropy vectors than $C_3^{{\rm C}}$.
Entropically, $C_3^{{\rm Q}}$ can be analysed with the technique outlined
in Section~\ref{sec:quantum_method}. An outer approximation,
$\Gamma_\mathcal{M}\left(C_3^{{\rm Q}}\right)$, to the set of achievable
entropy vectors,
$\overline{\Gamma_\mathcal{M}^{*}}\left(C_3^{{\rm Q}}\right)$, was
constructed in~\cite{Chaves2015}. It led to the Shannon
inequalities for the jointly distributed $X$, $Y$ and $Z$ and the
additional inequality
\begin{equation}
I(X\! : \! Y) + I(X\! : \! Z) \leq H(X),\label{eq:fritzhenson}
\end{equation}
as well as its permutations in $X$, $Y$ and $Z$~\cite{Chaves2015}.
It is natural to ask whether tighter approximations to
$\overline{\Gamma_\mathcal{M}^{*}}\left(C_3^{{\rm Q}}\right)$ can be
realised by a similar procedure to the one that led to tighter
approximations in the classical case. Unfortunately, we don't know of
any similar inequalities for the von Neumann entropy of multi-party
quantum states. Furthermore, even if the known non-Shannon
inequalities were to hold for von Neumann entropy we would not be able
to use them to add constraints to $C_3^{{\rm Q}}$ due to the lack of large
enough sets of coexisting, interdependent variables.\footnote{Note
that in causal structures that involve four or more observed
variables non-Shannon inequalities can be applied to these. However,
non-Shannon inequalities cannot be applied to unobserved quantum
systems (see also Section~\ref{sec:post_selected_nonShan}).}
\begin{prob}\label{op:1}
Do the closures of the sets of compatible entropy vectors coincide in the classical and the quantum triangle scenario, i.e., does $\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)=\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm Q}}\right)$ hold?
\end{prob}
Note that if this were to be answered in the affirmative, it would
point towards deficiencies of the current entropic techniques for
approximating
$\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm Q}}\right)$, which are
not able to recover any additional inequalities similar to the
non-Shannon inequalities found in the classical case.
One way to solve this problem would be to find an entropy vector
compatible with $C_3^{{\rm Q}}$ that lies outside one of our outer
approximations to
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$. Random
searches where the sources $A$, $B$ and $C$ distribute up to four
qubits each did not yield violations. However, the evidence from
these random searches against a separation of the classical and the
quantum sets is relatively weak. For one, our classical outer
approximations might be so loose that they contain
$\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm Q}}\right)$. To counter
this, we have attempted to randomly search for vectors that lie in
$\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm Q}}\right)$ but not in
the classical inner approximation
$\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)$. In spite of
the fact that we know such vectors exist, we were unable to randomly
find any. This shows the weakness of random searching, and also that
the region we are looking for (if it even exists) is small with
respect to our sampling measure.\footnote{This is not a statement
about the geometric extent of this region (for instance in terms of
a hyperdimensional solid angle as was previously considered for
inner and outer approximations to
$\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$). Instead,
since we are sampling quantum states here, and since these are not
in a one-to-one correspondence with the entropy vectors, it is a
statement about the fraction of states and measurements that may
produce entropy vectors outside
$\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)$ (in low
dimensions) according to our sampling distribution. This must be a
very small proportion of states and measurements (we didn't sample
any). Note also that if there is a gap between
$\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$ and
$\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm Q}}\right)$ it is
smaller than that between
$\Gamma^{\rm I}_{\mathcal{M}}\left(C_3^{{\rm C}}\right)$ and
$\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm Q}}\right)$. Hence,
constructing a vector in the first gap by sampling quantum states is
even more difficult than for the second.}
A natural candidate for an entropy vector that might violate some of our classical inequalities
is the one corresponding to the CHSH correlations that were shown not to be reproducible in $C_3^{{\rm C}}$ in Ref.~\cite{Fritz2012} (detailed in Figure~\ref{fig:fritzproof} where
$Z=(A',B')$ and in Appendix~\ref{sec:fritzproof}). However, the corresponding entropy vector lies inside $\Gamma_{\mathcal{M}}^{I}\left(C_3^{{\rm C}}\right)$ so is classically
reproducible. This particular distribution is also achievable in the
causal structure $P_4$ (a causal structure equivalent to the one in
Figure~\ref{fig:HensonStructures}(b)). Any distribution compatible
with $P_4$ may be realised in $C_3$ by choosing one of the variables,
e.g.\ $Z$, to have two outputs, one depending only on the input from
node $A$ and the other one depending on the input from $B$.
Distributions realisable in $P_4^{{\rm Q}}$ or $P_4^{{\rm C}}$ are thus always
realisable in $C_3^{{\rm Q}}$ or $C_3^{{\rm C}}$ respectively. According to
the results of~\cite{pnpaper}, all entropy vectors realised with
distributions in $P_4^{{\rm Q}}$ are also classically achievable, i.e.,
realisable in $P_4^{{\rm C}}$ (at least asymptotically).
Hence, no distribution in $P_4^{Q}$ can violate any of the classical entropy inequalities valid for $C_3^{{\rm C}}$. A way that might still allow us to use our knowledge about quantum correlations that are not classically reproducible in the Bell scenario to violate our entropic constraints to $\overline{\Gamma^*_{\mathcal{M}}}\left(C_3^{{\rm C}}\right)$, is by processing the inputs to all three nodes $X$, $Y$ and $Z$, so as to get around the results from~\cite{pnpaper}.\footnote{Two distributions that share the same entropy vector can
be very different and hence may be separated by local processing.}
In the following, we generalise the distribution that was utilised in Ref.~\cite{Fritz2012} to show that there is a separation between the achievable distributions in $C_3^{{\rm C}}$ and $C_3^{{\rm Q}}$, to a scenario where there is local processing at each output node. This also allows us reduce the required dimension of the output at $Z$ for which one can provably detect a difference between classical and quantum distributions from two bits to one bit.
\begin{prop} \label{prop:andextension}
There are non-classical quantum correlations in $C_3$ in the case where $X$ and $Y$ output two bits each while $Z$ outputs only one.
\end{prop}
A proof of Proposition~\ref{prop:andextension} can be found in
Appendix~\ref{sec:fritzproof}. It is interesting in so far
as the example in~\cite{Fritz2012} relies on a Bell inequality
violation. Given this, one might have expected that all information about the
measurement choices in the Bell setup, $\tilde{A}$ and $\tilde{B}$,
has to be exposed at the observed node $Z$.
Proposition~\ref{prop:andextension} shows that this is not the
case.
\begin{figure}
\centering
\resizebox{0.6 \columnwidth}{!}{%
\begin{tikzpicture}
\node (X) at (-2,2) {$(\tilde{X},\tilde{B})$};
\node (Y) at (2,2) {$(\tilde{Y},\tilde{A})$};
\node (Z) at (0,-1.46) {$Z$};
\node (A) at (1,0.28) {$A$};
\node (B) at (-1,0.28) {$B$};
\node (C) at (0,2) {$C$};
\draw [->,>=stealth] (A)--(Y);
\draw [->,>=stealth] (A)--(Z);
\draw [->,>=stealth] (B)--(X);
\draw [->,>=stealth] (B)--(Z);
\draw [->,>=stealth] (C)--(X);
\draw [->,>=stealth] (C)--(Y);
\end{tikzpicture}
}%
\caption{Scenario involving unobserved quantum systems, leading to a
distribution which is not reproducible with classical $A$, $B$ and
$C$~\cite{Fritz2012}. The observed variables $X=(\tilde{X},\tilde{B})$ and $Y=(\tilde{Y},\tilde{A})$ are chosen such that
$P_{\mathrm{\tilde{X}\tilde{Y}|AB}}$ maximally violates the CHSH
inequality~\cite{Clauser1969}.
$Z=(A',B')$ is such that $B'=\tilde{B}=B$ and
$A'=\tilde{A}=A$.
In essence the reason that this cannot be realised in the causal
structure $C_3^{{\rm C}}$ is the CHSH violation. Note though that it is
also important that information about $A$ is present in both $Y$ and
$Z$ (and analogously for $B$), otherwise the correlations could be mocked up.
In Proposition~\ref{prop:andextension}, we prove that a strategy where
$Z=\mathrm{AND}(A',B')$ also leads to correlations that cannot be
classically realised (see Appendix~\ref{sec:fritzproof} for further details).}
\label{fig:fritzproof}
\end{figure}
Nonetheless, we find that the entropy vector used to prove this proposition
does not violate our classical inequalities.
We have also taken $Z$ to be determined by different functions of $A$ and $B$ and have
additionally considered local processing of $X$ and $Y$. However,
even after such post-processing, for instance by applying all possible
functions from two bits to one, we have not been able to detect any
violations of the classical entropic bounds.
Note that vectors outside
$\Gamma_{\mathcal{M}}^{I}\left(C_3^{{\rm C}}\right)$ can be constructed
with appropriate post-processing of the (quantum) distribution. A
possible way to achieve this is applying \textsc{and} or \textsc{or}
functions appropriately. One may for instance consider the quantum
scenario detailed above, and take
$X=\operatorname{AND}(\tilde{X}, \tilde{B})$,
$Y=\operatorname{AND}(\tilde{Y}, \tilde{A})$ and
$Z=\operatorname{OR}(A',B')$. This renders the interaction information
of the entropy vector of the joint distribution of $X$, $Y$ and $Z$
positive, so the vector is not in
$\Gamma_{\mathcal{M}}^{I}\left(C_3^{{\rm C}}\right)$.
We have similarly tried to violate our entropy inequalities by relying on games other than the CHSH scenario, for which we know that there is distinctive quantum
behaviour (i.e., a separation at the level of
correlations); these include input states and
measurements known to lead to violations of the chained Bell
inequalities~\cite{Braunstein1988} or the Mermin-Peres magic square
game~\cite{Mermin1990, Peres1990}, all with post-processing at ($X$, $Y$ and) $Z$.
We have further considered scenarios where all three parties measure entangled states and use the
measurement outputs as inputs for further measurements. We have also attempted to incorporate functions known to lead to a positive interaction information in the classical case, as well as
functions from two to one bits in general, into these scenarios. None
of these attempts has led to a violation of the classical
inequalities so far. In a number of scenarios we have also considered
shared PR-boxes instead of entangled states, again without detecting
any violations of the inequalities. In most cases the corresponding
entropy vectors have a negative interaction information, and hence lie
in $\Gamma_\mathcal{M}^\mathrm{I}\left(C_3^{{\rm C}}\right)$, so can be
realised with a classical distribution as well, like in the case of
the correlations mentioned at the end of Section~\ref{sec:w_correl}.
\label{sec:op}
\subsection{Hybrid triangle scenarios}
In a \emph{hybrid causal structure} some of the unobserved nodes are
allowed to be quantum, whereas others are restricted to be
classical. One motivation for this is that sharing entanglement over
large distances is challenging due to noise, so two distant
observations might be assumed to have a classical cause while nearby
ones could have quantum causes. In the case of the causal structure $C_3$,
there are two such hybrid scenarios: either one or two of the three
unobserved variables can be chosen to be classical, whereas the others
are quantum. We call these two causal structures $C_3^{{\rm CQQ}}$ and
$C_3^{{\rm CCQ}}$ respectively. In the following, we will approximate the
sets of compatible entropy vectors for both scenarios. We show that
in hybrid scenarios of the triangle causal structure non-Shannon
inequalities are relevant.
\subsubsection{$C_3^{{\rm CQQ}}$ scenario}
In this scenario one of the unobserved variables is classical (we take
this to be $A$). The techniques introduced in
Sections~\ref{sec:classical_method} and~\ref{sec:quantum_method} allow
us to compute approximations of the set of allowed entropy vectors. We
find
\begin{equation*}
\Gamma_\mathcal{M}\left(C_3^{{\rm CQQ}}\right)=\Gamma_\mathcal{M}\left(C_3^{{\rm Q}}\right),
\end{equation*}
i.e., the outer approximation to
$\overline{\Gamma_\mathcal{M}^{*}}\left(C_3^{{\rm CQQ}}\right)$ obtained
without taking non-Shannon inequalities into account coincides with
the outer approximation to
$\overline{\Gamma_\mathcal{M}^{*}}\left(C_3^{{\rm Q}}\right)$. However,
unlike in the fully quantum case $C_3^{{\rm Q}}$, non-Shannon constraints
can be included for $C_3^{{\rm CQQ}}$, for instance the
inequality from Proposition~\ref{prop:zhangyeung} with variable
choices
\begin{align*}
\Diamond_{YZAX} \geq 0,\quad \Diamond_{YZXA} \geq 0.
\end{align*}
This results in a tighter approximation to
$\overline{\Gamma_\mathcal{M}^{*}}\left(C_3^{{\rm CQQ}}\right)$, which comprises the Shannon inequalities for three variables,
the constraint~\eqref{eq:fritzhenson} and\footnote{The second of
these inequalities can be easily derived from $\Diamond_{YZXA} \geq
0$ and the conditional independences, analogously to
Proposition~\ref{lemma:matus1}. To derive the first inequality, on the
other hand, several inequalities have to be combined.}
\begin{align*}
-3 H(X)&-3 H(Y)- 3 H(Z)+ 2H(XY)\\
&+ 2H(XZ)+ 3H(YZ)- H(XYZ) \geq 0, \\
-2 H(X)&-2 H(Y)- 2 H(Z)+ 3H(XY)\\
&+ 3H(XZ)+ 3H(YZ)- 4H(XYZ) \geq 0.
\end{align*}
Further non-Shannon constraints could also be exploited to improve
these approximations. Hence, some of the extremal rays of the Shannon
outer approximation $\Gamma_{\mathcal{M}}\left( C_3^{{\rm Q}} \right)$ are
provably not achievable if $A$, $B$ and $C$ do not all share entangled
states. Note that this does not imply that the sets of achievable entropy vectors in
$C_3^{{\rm Q}}$ and $C_3^{{\rm CQQ}}$ differ. However, the difference in their outer approximations
may prove useful for analysing whether there is a difference between the two.
If one were to prove by other means that there is no such difference the inequalities for $C_3^{{\rm CQQ}}$ would give us a way to better approximate the set of achievable entropy vectors of $C_3^{{\rm Q}}$.
\subsubsection{$C_3^{{\rm CCQ}}$ scenario}
In this scenario we take $A$ and $B$ to be classical. This scenario
can be understood as a Bell scenario, where the measurement choices of
the two parties are unobserved and processed to one single observed
output, $Z$.\footnote{Note that even though the sets of achievable
entropy vectors in classical and quantum case coincide in the Bell
scenario (cf.\ Section~\ref{sec:int} and~\cite{pnpaper}) this may
not be the case here as very different distributions may lead to the
same entropy vector in the classical and quantum case, which may be
separated by local processing.} The distributions from
Section~\ref{sec:quantcorr}, that are provably not reproducible in
$C_3^{{\rm C}}$ can be generated in this causal structure. Its entropic analysis thus
restricts the violations of our classical inequalities we may hope to achieve with such distributions.
To approximate the set of compatible entropy vectors of this scenario,
$\overline{\Gamma^{*}_{\mathcal{M}}}\left(C_3^{{\rm CCQ}}\right)$, we
proceed analogously to the $C_3^{{\rm Q}}$ and $C_3^{{\rm CQQ}}$ scenarios
before. However, the result differs and leads to a tighter cone, even
without considering non-Shannon inequalities, i.e.,
$\Gamma_{\mathcal{M}}\left(C_3^{{\rm CCQ}}\right) \subsetneq \Gamma_{\mathcal{M}}\left(C_3^{{\rm CQQ}}\right)$.
The approximation is given by the three variable Shannon inequalities and the following additional inequalities:
\begin{align}
&\!-H(X)\!-\!H(Y)\!-\!H(Z)\!+\!H(XY)\!+\!H(XZ)\!\geq\! 0,\nonumber \\
&-3 H(X)-3 H(Y)- 3 H(Z)+ 2H(XY)\nonumber\\
&+ 3H(XZ)+ 2H(YZ)- H(XYZ)\geq 0,\label{eq:ccq}
\end{align}
up to permutations of $X$, $Y$ and $Z$ in the first inequality and of
$Y$ and $Z$ in the second. Note that these five inequalities are a
subset of the seven inequalities \eqref{eq:margindep} delimiting
$\Gamma_{\mathcal{M}}\left( C_3^{{\rm C}}\right)$ and that
\begin{equation*}
\Gamma_{\mathcal{M}}\left(C_3^{{\rm C}}\right) \subsetneq \Gamma_{\mathcal{M}}\left(C_3^{{\rm CCQ}}\right).
\end{equation*}
\begin{widetext}
The inequalities $\Diamond_{XZBY} \geq 0$, $\Diamond_{YZAX} \geq 0$ and $\Diamond_{YZXA}\geq 0$
lead to the additional inequalities
\begin{equation}\label{eq:ccqadditional}
\begin{split}
-2H(X)-2H(Y)-2H(Z)+3H(XY)+3H(XZ)+3H(YZ)-4H(XYZ) &\geq 0, \\
-6H(X)-6H(Y)-6H(Z)+5H(XY)+5H(XZ)+5H(YZ)-3H(XYZ) &\geq 0, \\
-4H(X)-4H(Y)-4H(Z)+3H(XY)+4H(XZ)+3H(YZ)-2H(XYZ) &\geq 0,
\end{split}
\end{equation}
\end{widetext}
(including permutations of $X$ and $Y$ in the last inequality). They
render the second inequality (and its permutations) in \eqref{eq:ccq}
redundant, while the first remains (for all of its permutations). Note
that the first inequality of \eqref{eq:ccqadditional} is also present
in $\Gamma_{\mathcal{M}} \left(C_3^{{\rm CQQ}}\right)$.
As in the previous example, further constraints could likely be derived by considering additional non-Shannon inequalities.
\section{Conclusions} \label{sec:conclusion}
We have shown that
non-Shannon inequalities tighten the entropic approximations of the
classical entropy cones in many causal structures including the
triangle scenario and the causal structure relevant for information
causality.
Our newly derived inequalities improve on the entropic distinction of
these from other (classical) causal structures, which is of interest
for inferring (classical) causal relations. They also constitute a set
of restrictions on the classical entropy cones that we cannot derive
in the quantum case, which may point towards differences between the
sets of achievable entropy vectors in classical and quantum case.
Since it is known from the Bell scenario that quantum correlations can
be detected by considering the entropies of the variables in a post-selected causal structure~\cite{Braunstein1988}, our analysis of the information causality scenario is the one that is most likely to be useful for this purpose.
In this context, non-Shannon inequalities may also be important with regard to the discussion of whether entropic techniques may even be sufficient for certifying classical reproducibility in certain scenarios, a question that has previously been explored for the CHSH scenario in Ref.~\cite{Chaves2013}.
While the entropy vector approach is known to be a useful
means for distinguishing different classical causal structures, its
ability to differentiate between classical and quantum versions of the
same causal structure is known to be limited~\cite{pnpaper}. The
present work has unveiled further limitations of the approach:
for all causal structures classified
in~\cite{Henson2014} we found either that the sets of achievable
entropy vectors in classical and quantum case coincide (for the causal structures of Figure~\ref{fig:HensonStructures}), or that
non-Shannon inequalities play a role in their characterisation leaving
us unable to make such a statement.
One of the reasons why it is difficult to make such a statement when
non-Shannon inequalities play a role is our relatively poor
understanding of the structure of entropy space. Even in the absence
of a causal structure we lack a tight characterisation of the set of
allowed entropy vectors for four random variables. In the quantum
case, it is an important open problem whether any further general
constraints on the von Neumann entropy exist. This partly explains our inability to show whether there is some causal structure in which
the described entropy vector approach can be useful for distinguishing
classical and quantum.
Behind all this is the question, of whether there is a novel technique
that allows for an efficient and accurate way to distinguish classical
and quantum versions of the same causal structure. Such a technique
needs to simplify the description of the set of allowed distributions
but remain complex enough to retain the distinctive features of
classical, quantum and post-quantum probability
distributions. Identifying such a quantity would provide further
insight into the meaning of cause in quantum mechanics.
\vspace{-0.4cm}
\acknowledgments We thank Benjamin Chapman and Alex May for their
assistance with initial investigations, and we thank Luke Elliott for
carrying out random searches in Fortran. We are grateful to Matthew
Pusey for alerting us to an error in an earlier draft. This work was
supported by the Engineering and Physical Sciences Research Council
through a First Grant (no.\ EP/P016588/1) and the Quantum
Communications Hub (grant no.\ EP/M013472/1).
\bibliographystyle{plainnat}
|
\section{Introduction}
Non-topological solitons in the scalar field theory, which were initially proposed in \cite{Rosen0} and now are known as Q-balls \cite{Coleman:1985ki}, are widely discussed in the literature. However, among the variety of models providing Q-ball-type solutions, only a few of them admit of analytical treatment in four-dimensional space-time, at least for examining their main properties such as the energy--charge dependencies. The existence of analytical solutions simplifies the analysis considerably and allows one to perform a deeper study of the Q-ball properties. For such exceptions, one can recall the model with a very simple polynomial potential proposed in \cite{Anderson:1970et}, where the energy--charge dependence can be obtained analytically, as well as the models of \cite{Rosen1,Arodz:2008jk,Theodorakis:2000bz,Gulamov:2013ema} providing exact analytical solutions for Q-balls (the logarithmic scalar field potential of \cite{Rosen1} makes it possible to examine analytically even the linear perturbations above the Q-ball solution; see \cite{MarcVent}).
All the models mentioned above deal with a single complex scalar field. Meanwhile, there is another class of models, namely, with two different scalar fields (the complex one and the real one), which also admit the existence of Q-ball-like solutions. The best-known example of this type was proposed in \cite{Friedberg:1976me}, in which only an approximate analytical Q-ball solution\footnote{Although solutions in two-field models like the one of \cite{Friedberg:1976me} are not Q-balls in the sense of Coleman's definition of Q-balls \cite{Coleman:1985ki}, they are of the same kind, so from now on we call such soliton solutions ``Q-balls''.} can be obtained (using the trial functions), thus being demanding for numerical calculations. A simplification of this model by neglecting the potential of the real scalar field was performed in \cite{Levin:2010gp}. Q-ball solutions in such a theory have a rather interesting property -- they possess a (non-conserved) ``scalar charge'', which characterizes the long-range attraction between such Q-balls. However, exact analytical formulas for the energy--charge dependence cannot be obtained in this case too.
In the present paper we consider an even more simplified two-field model. Instead of the quartic interaction of the scalar fields in \cite{Levin:2010gp}, we consider the triple Yukawa interaction between the scalar fields, resulting in the well-known Wick--Cutkosky model \cite{Wick,Cutkosky}. Surprisingly, most of the analysis in this model can be performed analytically. In particular, the exact form of the corresponding energy--charge dependence can be obtained analytically (which simplifies the use of the well-know stability criteria), whereas the numerical solution is necessary only for obtaining the value of some universal dimensionless parameter (from this point of view the model is similar to the one-field model of \cite{Anderson:1970et}).
Q-balls are classical objects, but it is clear that they can be considered as bound states of scalar particles of the theory. However, it is not clear how to perform a consistent analysis for the case in which such a bound state consists of a large number of particles, which is exactly the case of a Q-ball. Indeed, even the problem of two bound scalar particles in the Wick--Cutkosky model is not trivial, though it can be solved analytically (see, for example, review \cite{Nakanishi:1988hp}). In the present paper we will propose a simple method that will allow one at least to roughly estimate the binding energy of a Q-ball. This method can be applied to Q-balls in different models, but we will test it on Q-balls of the Wick--Cutkosky model. For such an analysis the Wick--Cutkosky model is unique -- on the one hand, there exist Q-ball solutions in this model; on the other hand, the Bethe--Salpeter equation describing the bound state of two massive scalar particles can be solved analytically in the Wick--Cutkosky model. The latter will allow one to compare the results obtained with the help of the methods of classical and quantum field theories.
\section{Setup and equations of motion}
Let us start with the action of the Wick--Cutkosky model, describing the complex scalar field $\chi$ interacting with the real massless scalar field $\phi$ in the flat four-dimensional space-time with the coordinates $x^{\mu}=\{t,\vec x\}$, $\mu=0,1,2,3$; in the form
\begin{equation}
S=\int\left(\partial_\mu\chi^*\partial^\mu\chi+\frac{1}{2}\partial_\mu\phi\partial^\mu\phi-h\phi\chi^*\chi\right)d^4x,
\label{sys}
\end{equation}
where $h\ne 0$ is the coupling constant of the scalar Yukawa interaction.
Classical vacua of the theory correspond to the stationary points of the potential
\begin{equation}
V(\phi,\chi)=h\phi\chi^*\chi.
\label{potential}
\end{equation}
The corresponding vacuum solutions are just $\chi\equiv 0$ and $\phi=\phi_{0}\equiv\textrm{const}$, i.e., there is a flat direction along the real field. Now let us consider the quadratic part of action (\ref{sys}) for the fluctuations $\chi(t,x)$, $\phi(t,x)=\phi_{0}+\rho(t,x)$ above the vacuum solution. We get
\begin{equation}
S_{(2)}=\int\left(\partial_\mu\chi^*\partial^\mu\chi+\frac{1}{2}\partial_\mu\rho\partial^\mu\rho-h\phi_{0}\chi^*\chi\right)d^4x,
\end{equation}
We see that for $h\phi_{0}<0$ the effective mass $m=\sqrt{h\phi_{0}}$ of the field $\chi$ is imaginary, leading to tachyonic instability. For $h\phi_{0}>0$ the corresponding mass term has the proper sign and one expects that such vacua are stable. As for the case $\phi_{0}=0$, it is possible to show analytically that there are no Q-ball-type solutions for $h\phi_{0}\le 0$ (this topic will be discussed below). For these reasons, below we will consider only the case $h\phi_{0}>0$.
The fact that globally the scalar field potential is not bounded from below is not dangerous -- the vacuum solution is classically stable for $h\phi_{0}>0$.\footnote{Moreover, it will be shown below that the scalar field potential can be modified to become bounded from below while keeping in the theory the most interesting Q-ball solutions.} As for the flat direction, its existence is also not dangerous. Indeed, let us
consider a scalar field (not interacting with gauge fields) with the standard ``Mexican hat''-type potential. There is a class of vacuum solutions, all of them having the same zero energy –-- there exists a flat direction. But the existence of the massless Goldstone bosons does not indicate any instability of such a vacuum, because an infinite energy is necessary to change the vacuum solution of the scalar field in the whole space \cite{Rubakov:2002fi}. Our case is exactly the same –-- it is also necessary to have an infinite energy to change the vacuum solution with $h\phi_{0}>0$ in the whole space, whereas it is the corresponding massless mode that provides the attraction force forming Q-balls in the Wick--Cutkosky model.
The equations of motion, following from the action (\ref{sys}), take the form
\begin{eqnarray}
\Box\chi+h\phi\chi=0,\\
\Box\phi+h\chi^{*}\chi=0.
\end{eqnarray}
In the following, we will be looking for stationary spherically symmetric solutions without nodes, which take the form
\begin{eqnarray}\label{chiequiv}
\chi(t,\vec x)=e^{i\omega t}f(r),\\ \label{phiequiv}
\phi(t,\vec x)=\phi(r)
\end{eqnarray}
with $\partial_{r}f|_{r=0}=0$, $\lim\limits_{r\to\infty}f(r)=0$, $\partial_{r}\phi|_{r=0}=0$, $\lim\limits_{r\to\infty}\phi(r)=\phi_{0}$, where $r=\sqrt{{\vec x}^2}$ and $\omega$ is real.
For convenience, from the very beginning it is useful to represent the field $\phi$ as
\begin{equation}\label{vac-shift}
\phi=\phi_{0}+\tilde\phi
\end{equation}
and to consider the system of equations
\begin{eqnarray}\label{eq1}
-\omega^{2}f-\Delta f+m^2f+h\tilde\phi f=0,\\ \label{eq2}
-\Delta\tilde\phi+hf^2=0
\end{eqnarray}
with $m^2=h\phi_{0}>0$ and
\begin{eqnarray}\label{boundary1}
\partial_{r}f|_{r=0}=0,\qquad \lim\limits_{r\to\infty}f(r)=0,\\ \label{boundary2}
\partial_{r}\tilde\phi|_{r=0}=0, \qquad \lim\limits_{r\to\infty}\tilde\phi(r)=0.
\end{eqnarray}
Equations (\ref{eq1}) and (\ref{eq2}) follow from the action of the Wick--Cutkosky model in its common form in which the mass term of one of the scalar fields exists from the very beginning. In principle, the shift (\ref{vac-shift}) can be considered just as a redefinition of the field $\phi$ without any reference to vacuum, in this case one should consider the theory with the vacuum solution $\tilde\phi\equiv 0$.
In order to ensure that the field $f(r)$ falls off exponentially for $r\to\infty$, providing the finiteness of the Q-ball charge and energy, the frequency $\omega$ should be bounded as $|\omega|<m$ (moreover, it is not difficult to show that there are no Q-ball solutions for $\omega^{2}-m^{2}\ge 0$, including the case $m^{2}\le 0$; see Appendix~A). From the structure of these equations we expect that $\tilde\phi(r)\sim\frac{1}{r}$ for large $r$. We see that in the physically reasonable cases (i.e., in the cases with stable vacua) this system describes a massive charged scalar field in the long-range attractive potential provided by the field $\phi$.\footnote{The Schr\"{o}dinger--Poisson systems, which appear, for example, when one considers the Newtonian limit for boson stars made of scalar fields \cite{Ruffini:1969qy,Seidel:1990jh,Marsh:2015wka}, also provide equations of motion very similar to equations (\ref{eq1}) and (\ref{eq2}). But though the equations of motion in different models look similar from the mathematical point of view, the physical essence of different theories is completely different, starting from the origin of the coupling between the fields and ending with the definition of important physical characteristics of the solutions.}
\section{Q-ball solution and its properties}
\label{sec_charged}
Suppose that there exists a Q-ball solution to the system of equations (\ref{eq1}) and (\ref{eq2}), satisfying the boundary conditions (\ref{boundary1}) and (\ref{boundary2}). In this case, the $U(1)$ global charge of the Q-ball can be defined as
\begin{eqnarray}\label{chargedef}
Q=i\int\left(\chi\partial_{0}\chi^*-\chi^*\partial_{0}\chi\right)d^{3}x=4\pi\int\limits_{0}^{\infty}2\omega f^{2}r^{2}dr,
\end{eqnarray}
whereas the Q-ball energy takes the form
\begin{eqnarray}\label{energydef}
E=4\pi\int\limits_{0}^{\infty}\Bigl(\omega^{2}f^{2}+\partial_{r}f\partial_{r}f+m^{2}f^{2}+hf^{2}\tilde\phi+\frac{1}{2}\partial_{r}\tilde\phi\partial_{r}\tilde\phi\Bigr)r^{2}dr.
\end{eqnarray}
Using the equations of motion for the fields, it is not difficult to show that the Q-ball solution possesses the following properties:
\begin{eqnarray}
E&=&\omega Q+2\pi\int\limits_{0}^{\infty}(\partial_{r}\tilde\phi)^{2}r^{2}dr>\omega Q,\\ \label{dEdQ}
\frac{dE}{dQ}&=&\omega,
\end{eqnarray}
see also Appendix~B for details. The latter relation is well known for the one-field Q-balls.
The system of equations (\ref{eq1}) and (\ref{eq2}) can be brought into dimensionless form by means of the transformations
\begin{eqnarray}
R=r\sqrt{m^{2}-\omega^{2}},\qquad F(R)=\frac{h}{m^2-\omega^2}f(r),\qquad G(R)=\frac{h}{m^2-\omega^2}\tilde\phi(r),
\label{transformation}
\end{eqnarray}
resulting in
\begin{eqnarray}\label{eq1a}
-\Delta_{R}F+F+FG=0,\\ \label{eq2a}
-\Delta_{R}G+F^2=0
\end{eqnarray}
with the boundary conditions
\begin{eqnarray}\nonumber
\partial_{R}F|_{R=0}=0,\qquad \lim\limits_{R\to\infty}F(R)=0,\\ \label{boundary1a}
\partial_{R}G|_{R=0}=0, \qquad \lim\limits_{R\to\infty}G(R)=0.
\end{eqnarray}
Without loss of generality, we suppose that $F(R)>0$ for any $R$.
\begin{figure*}[!ht]
\centering
\includegraphics[width=0.95\linewidth]{fig1.eps}
\caption{Numerical solution for the functions $F(R)$ and $G(R)$.}\label{numsol}
\end{figure*}
In Fig.~\ref{numsol} the explicit numerical solution to equations (\ref{eq1a}), (\ref{eq2a}) with (\ref{boundary1a}) is presented (see Appendix~C for the details of the numerical analysis). In fact, this numerical solution is necessary only for the visualization of the Q-ball -- all the important characteristics of the Q-ball can be obtained either analytically or with the help of the auxiliary numerical solution presented in Appendix~C. The form of the solution implies that in general the thin-wall approximation cannot be used for describing solutions for the complex scalar field. Since the system of equations (\ref{eq1a}), (\ref{eq2a}) with (\ref{boundary1a}) is devoid of the parameters $\phi_{0}$, $h$, and $\omega$, the dimensionless solution presented in Fig.~\ref{numsol} can be used to restore the actual solution for the scalar fields $\chi$, $\phi$ for any physically reasonable choice of the coupling constant $h$, the vacuum expectation value $\phi_{0}$, and the frequency $\omega$.
We note that very often Q-ball solutions are obtained by minimization of the energy at a fixed charge. This procedure is fully equivalent to solving the corresponding equations of motion \cite{Coleman:1985ki,Friedberg:1976me}.
For the charge and the energy of the Q-ball we get
\begin{equation}\label{chargewickcutk}
Q=\frac{2\omega\sqrt{m^2-\omega^{2}}}{h^{2}}\,I, \qquad
E=\omega Q+\frac{(m^{2}-\omega^{2})^{\frac{3}{2}}}{h^{2}}\,J,
\end{equation}
where
\begin{equation}\label{Idef}
I=4\pi\int\limits_{0}^{\infty}F^{2}R^{2}dR,\qquad
J=2\pi\int\limits_{0}^{\infty}(\partial_{R}G)^{2}R^{2}dR.
\end{equation}
Using (\ref{dEdQ}), it is easy to show that $J=\frac{2}{3}I$, leading to
\begin{eqnarray}\label{Q}
Q&=&\frac{I}{h^2}2\,\omega\sqrt{m^2-\omega^{2}},\\ \label{E}
E&=&\frac{I}{h^2}\sqrt{m^2-\omega^{2}}\left(\frac{4}{3}\omega^{2}+\frac{2}{3}m^{2}\right).
\end{eqnarray}
In Fig.~\ref{figEQ} the $E(Q)$ dependence, corresponding to (\ref{Q}), (\ref{E}), is presented.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{fig2.eps}
\caption{$E(Q)$ dependence for Q-balls.}\label{figEQ}
\end{figure}
The parameter $I$ can be found numerically and turns out to be $I\approx 44.05$ (see Appendix~C for details). The cusps in Fig.~\ref{figEQ}, corresponding to the points with $\frac{dQ}{d\omega}=0$ (at $\omega=-\frac{m}{\sqrt{2}}$ and $\omega=\frac{m}{\sqrt{2}}$), are a direct consequence of the relation (\ref{dEdQ}). The form of the $E(Q)$ dependence shows that there exist solutions with the maximal and minimal charges $Q_{max}=-Q_{min}=\frac{Im^{2}}{h^2}$, both possessing maximal energy $E_{max}=\frac{2\sqrt{2}\,Im^{3}}{3h^2}$. It should be mentioned that an analogous form of the $E(Q)$ dependence for Q-balls appears in the model of \cite{Rosen1}, describing a single complex scalar field with the logarithmic scalar field potential, see \cite{MarcVent}. However, contrary to the case of the Wick--Cutkosky model, in the model of \cite{Rosen1} the complex scalar field formally has an infinite mass.
From (\ref{Q}) and (\ref{E}) it is not difficult to obtain explicit formulas for this $E(Q)$ dependence: the upper branch in Fig.~\ref{figEQ} is described by
\begin{eqnarray}
E=\frac{\sqrt{2}\,I}{3\,h^{2}}\left(2m^{2}-\sqrt{m^{4}-\frac{Q^{2}h^{4}}{I^{2}}}\right)\sqrt{m^{2}+\sqrt{m^{4}-\frac{Q^{2}h^{4}}{I^{2}}}},
\end{eqnarray}
whereas the lower branches are described by
\begin{eqnarray}\label{energylowerbr}
E=\frac{\sqrt{2}\,I}{3\,h^{2}}\left(2m^{2}+\sqrt{m^{4}-\frac{Q^{2}h^{4}}{I^{2}}}\right)\sqrt{m^{2}-\sqrt{m^{4}-\frac{Q^{2}h^{4}}{I^{2}}}}.
\end{eqnarray}
Let us briefly discuss the stability of the obtained solutions. First, the stability criterion proposed in \cite{Friedberg:1976me,LeePang} implies that solutions on the lower branches of the $E(Q)$ dependence in Fig.~\ref{figEQ} are classically stable, i.e., stable with respect to small perturbations of the fields. Indeed, for these solutions the relation $\frac{dQ}{d\omega}<0$, which is one of the conditions necessary for the classical stability of such a two-field model \cite{Friedberg:1976me}, holds. Second, since $d^{2}E/dQ^{2}<0$ for the lower branches in Fig.~\ref{figEQ}, the Q-balls, corresponding to these solutions, are stable against fission (a simple justification of this fact in the general case can be found in \cite{Gulamov:2013ema}). And third, it is not difficult to check that the lower branches of the $E(Q)$ dependence in Fig.~\ref{figEQ} lie below the lines $E=m|Q|$, standing for free particles of mass $m$. The latter means that Q-balls, corresponding to these branches, are also quantum mechanically stable, i.e., stable with respect to decay into free scalar particles of mass $m$ (of course, the statement about the quantum mechanical stability is valid if there are no interactions with other particles in the theory under consideration). Thus, Q-ball solutions from the lower branches of the $E(Q)$ dependence in Fig.~\ref{figEQ} can be thought of as absolutely stable.
There exists a classically unstable time-independent solution with $\omega=0$ (and, consequently, $Q=0$) with nonzero finite energy. Such solutions are usually called ``sphalerons'' and can play an interesting role in quantum theory \cite{Manton:2004tk,Rubakov:2002fi}.
In this connection it is interesting to check whether the values of the initial scalar field $\phi$ (not $\tilde\phi$!) in the Q-ball are such that $h\phi$ becomes negative at least for some $r$, which corresponds to the area of unstable vacua in the scalar field potential. The smallest value of $h\phi$ is attained at $r=0$, so we will consider $h\phi(0)$. Using (\ref{vac-shift}), (\ref{transformation}) and the results presented in Appendix~C, we get
\begin{equation}
h\phi(0)=h\phi_{0}\left(1-C\left(1-\frac{\omega^{2}}{m^{2}}\right)\right),
\end{equation}
where we have used the fact that $h\phi_{0}=m^{2}$. Here $C\approx 1.938$; see Appendix~C. It is easy to find that $h\phi(0)\ge 0$ (and, consequently, $h\phi(r)\ge 0$ for any $r$) for $|\omega|\ge m\sqrt{\frac{C-1}{C}}\approx 0.696\cdot m$. We see that Q-balls from the lower (``stable'') branches (i.e., Q-balls with $m>|\omega|\ge \frac{m}{\sqrt{2}}\approx 0.707\cdot m$) and from the small parts of the upper (``unstable'') branch reside in the area of the scalar field potential (\ref{potential}) which corresponds to stable vacua, i.e., $V(\phi,\chi)>0$ for these Q-balls. This also indicates that there is no direct and simple connection between the values of the Q-ball scalar fields, the form of the scalar field potential, and the Q-ball stability, as noted in \cite{Nugaev:2013poa}. As for the rest of the Q-balls (for which $h\phi(0)<0$), some parts of such Q-balls reside in the area of the scalar field potential for which $V(\phi,\chi)<0$ holds, corresponding to unstable vacua. An interesting observation is that all Q-ball solutions with $h\phi(r)>0$ for any $r$ (which includes all Q-balls from the lower branches) can be reproduced in a theory with the scalar field potential
\begin{equation}
V(\phi,\chi)=h|\phi|\chi^*\chi,
\label{potential-modif}
\end{equation}
where $h>0$, which is bounded from below.
Now let us compare the Wick--Cutkosky model with the model proposed and examined in \cite{Levin:2010gp} (as noted in the Introduction, the latter is a simplification of the two-field model of \cite{Friedberg:1976me}). First, for Q-ball solutions in both models the real scalar field behaves as $\sim\frac{1}{r}$ at large $r$, so different Q-balls undergo long-range Coulomb attraction between each other regardless of the sign of their $U(1)$ global charges. As noted in \cite{Levin:2010gp}, one can say that Q-balls possess a (non-conserved) ``scalar charge'', which characterizes the strength of this long-range attraction. It will be defined explicitly in the next section.
Second, there is no sphaleron solution in the model of \cite{Levin:2010gp}. Indeed, according to the Derrick theorem \cite{Derrick}, there are no time-independent localized solutions in a theory with nonnegative scalar field potential, which is exactly the case of \cite{Levin:2010gp}.
Finally, it is possible to show that, for $M-|\omega|\ll M$, where $M$ is the mass of the free charged scalar particle ($M=m=\sqrt{h\phi_{0}}$ in the Wick--Cutkosky model), both models have almost the same $E(Q)$ dependencies such that $Q\to 0$, $E\to 0$ for $|\omega|\to M$; see Appendix~D for details. However, in the Wick--Cutkosky model the Q-ball energy is bounded from above, whereas in the model of \cite{Levin:2010gp} the charge and the energy go to infinity for $\omega\to 0$ such that $E\sim\sqrt{|Q|}$ for large $|Q|$. The latter implies that all Q-balls in the model of \cite{Levin:2010gp} can be absolutely stable (because $\frac{dQ}{d\omega}<0$ and $E<M|Q|$), whereas in our case there may exist absolutely stable and unstable Q-balls.
\section{Q-ball as a bound state of scalar particles}
As noted in the Introduction, Q-balls can be considered as bound states of scalar particles of the theory. In order to calculate the binding energy of a Q-ball, it is necessary to know the number of scalar particles forming the Q-ball. In this case the binding energy $\epsilon$ is simply
\begin{equation}\label{bindenergy}
\epsilon=Nm-E,
\end{equation}
where $N$ is the number of scalar particles in the Q-ball, $E$ is the Q-ball energy and $m$ is the mass of the free scalar particle of the theory. An obvious problem is to estimate the number of particles $N$.
Indeed, Q-balls are formed not only from particles, but also from anti-particles, i.e., it is a bound state of particles and anti-particles (it is obvious for sphalerons -- solutions with $Q=0$ and $E\neq 0$). In quantum field theory the operator of charge $\hat Q$ gives
\begin{equation}
\langle N_{+},N_{-}|\hat Q|N_{+},N_{-}\rangle=N_{+}-N_{-},
\end{equation}
where $|N_{+},N_{-}\rangle$ defines the state with $N_{+}$ particles and $N_{-}$ anti-particles, whereas we need $N=N_{+}+N_{-}$. However, it was shown in \cite{Levin:2010gp} that there exists a ``scalar charge'', which characterizes the strength of the long-range attraction of Q-balls in the model of \cite{Levin:2010gp} regardless of the sign of their global charges. The latter suggests that in the general case this ``scalar charge'' can be somehow connected with the total number of particles in a Q-ball.
In order to examine such a possibility, let us define the scalar charge as
\begin{equation}
Q_{SC}=2m\int\chi^*\chi\,d^{3}x.
\end{equation}
It is clear that this charge is {\em non-conserved} in general. But for Q-balls at rest we get
\begin{equation}\label{QscQball}
Q_{SC}=8\pi m\int\limits_{0}^{\infty}f^{2}r^{2}dr,
\end{equation}
which obviously is conserved in time. It is not difficult to check that for the non-relativistic particles, i.e., for $k_{0}-m\ll m$, the following relation holds in quantum field theory:
\begin{equation}
\langle N_{+},N_{-}|\hat Q_{SC}|N_{+},N_{-}\rangle\approx N_{+}+N_{-},
\end{equation}
where $\hat Q_{SC}=2m:\int\chi^*\chi\,d^{3}x:$, which is exactly what we need. So, we can assume that in a Q-ball $N\approx Q_{SC}$. Of course, this estimate can be used for $N\gg 1$.
Now, under the assumption that the most of scalar particles in the Q-ball are {\em non-relativistic}, with the help of (\ref{QscQball}) we can estimate the binding energy of any Q-ball. For example, for the Q-balls in the Wick--Cutkosky model we get from (\ref{QscQball}) and (\ref{bindenergy})
\begin{eqnarray}\label{schargewickcutk}
Q_{SC}(\omega)=\frac{2m\sqrt{m^2-\omega^{2}}}{h^{2}}\,I,\\ \label{bindingenergy} \epsilon(\omega)=\frac{4I}{3h^{2}}(m^{2}-\omega^{2})^{\frac{3}{2}}>0.
\end{eqnarray}
The corresponding $\epsilon(Q)$ dependence is presented in Fig.~\ref{figEpsilonQ}. It is also illustrative to consider the contributions of particles and anti-particles to the total charge $Q$:
\begin{equation}\label{nplusnminus}
Q_{+}=N_{+}\approx\frac{Q+Q_{SC}}{2},\quad Q_{-}=-N_{-}\approx\frac{Q-Q_{SC}}{2}.
\end{equation}
For the Q-balls in the Wick--Cutkosky model the corresponding plots are presented in Fig.~\ref{figQplusminus}.
\begin{figure}[!ht]
\begin{minipage}[t]{.49\linewidth}
\centering
\includegraphics[width=0.93\linewidth]{fig3.eps}
\caption{$\epsilon(Q)$ dependence for Q-balls in the Wick--Cutkosky model.}\label{figEpsilonQ}
\end{minipage}
\begin{minipage}[t]{.49\linewidth}
\centering
\includegraphics[width=0.93\linewidth]{fig4.eps}
\caption{$Q(\omega)$ (thick line), $Q_{+}(\omega)$ (thin line) and $Q_{-}(\omega)$ (dashed line) for Q-balls in the Wick--Cutkosky model.}\label{figQplusminus}
\end{minipage}
\end{figure}
Several comments are in order here. As noted above, in the scalar field theory there are no conserved charges giving $N$, $N_{+}$ or $N_{-}$, so we cannot define the quantities $N$, $N_{+}$ or $N_{-}$ in a completely rigorous way. However, at the classical level (when $N$ is supposed to be very large) we may assume that we have {\em approximately} $N$ particles in a Q-ball -- since there are no corresponding conserved charges, the ``actual'' numbers of particles and anti-particles can vary in time accounting for virtual processes of creation, annihilation, etc. inside a Q-ball. Meanwhile, the mean values $N_{+}$ and $N_{-}$ can have the following meaning -- if we add the energy $\epsilon$ to the Q-ball, we expect that it can disassemble into approximately $N_{+}$ free particles and $N_{-}$ free anti-particles (i.e., it is the binding energy of the whole Q-ball). It is the scalar charge $Q_{SC}$ that determines how Q-balls in the Wick--Cutkosky model (as well as in the model of \cite{Levin:2010gp}) interact with each other by means of the field $\tilde\phi$. The field $\tilde\phi$ is also responsible for forming a Q-ball from the quanta of the field $\chi$, so the choice of $Q_{SC}$ for determining the approximate number of particles and anti-particles makes sense. Although this choice is motivated by the properties of the Wick--Cutkosky model and the model of \cite{Levin:2010gp}, we think that this approach can be used in the general case too.
It is clear that unstable Q-balls cannot decay into $N=N_{+}+N_{-}$ particles, it is energetically forbidden. A Q-ball of charge $Q$ may decay into $\tilde N$ particles and/or anti-particles (with $|Q|\le\tilde N<N$) plus quanta of the field $\tilde\phi$. For example, a possible decay channel of sphalerons is the decay only into quanta of the massless field $\tilde\phi$. The latter is similar to the case of, say, positronium -- its mass is a few $eV$ less than twice the electron mass, and it decays only into photons. Thus, the description of a Q-ball as a collection of, say, $|Q|$ particles or anti-particles is inappropriate -- it obviously fails for sphalerons with $Q=0$, $E\neq 0$.
Now let us look at this problem from another point of view. Suppose that we have a Q-ball with, say, charge $Q>0$ and we add a single particle of mass $m$ (and charge $+1$) to it. If $Q\gg 1$ (i.e., if $\frac{Im^{2}}{h^2}\gg 1$), then, according to (\ref{dEdQ}), we get the Q-ball of charge $Q+1$ and with the energy $E(Q+1)\approx E(Q)+\omega$ (here $\omega$ is similar to the chemical potential in thermodynamics). It is clear that in order to release a particle from this Q-ball, we should add the energy $m-\omega$ to the Q-ball.\footnote{When we add a particle to the Q-ball, the remaining energy $m-\omega$ should be radiated out, say, by the field $\phi$.} For the anti-particle, the corresponding ``chemical potential'' is $-\omega$, whereas the ``ionization'' energy is $m+\omega$. One may think that by adding a particle to a Q-ball, we just increase the corresponding number $N_{+}$ by unity. It is not so in general. For example, for $\omega=\frac{m}{2}$, using (\ref{Q}), (\ref{schargewickcutk}) and (\ref{nplusnminus}) we get $\Delta N_{+}\approx 0$, $\Delta N_{-}\approx -1$. In this case the interpretation is clear -- namely, the incoming particle annihilate with an anti-particle inside the Q-ball. However, for $\omega=0$, we get $\Delta N_{+}\approx\frac{1}{2}$, $\Delta N_{-}\approx -\frac{1}{2}$, which has no such a clear interpretation. For $\omega\to m$ we get $\Delta N_{+}\approx 1$, $\Delta N_{-}\approx 0$. These examples indicate that the numbers $N_{+}$, $N_{-}$, which are specific to each Q-ball, should be considered only as approximate values for estimates. Namely, if we increase the charge of a Q-ball by a finite value $\Delta Q>0$, the values $\Delta N_{+}=N_{+}(Q+\Delta Q)-N_{+}(Q)$ and $\Delta N_{-}=N_{-}(Q+\Delta Q)-N_{-}(Q)$ indicate, of course approximately, how many of the incoming particles are absorbed by the Q-ball and how many anti-particles inside the Q-ball are annihilated by the incoming particles. A fully analogous considerations can be made for $\Delta Q<0$, i.e., for the incoming anti-particles.
Finally, let us compare the Q-ball binding energy, obtained using the procedure presented in this section, with the energy of the bound state coming from the solution to the Bethe--Salpeter equation for $Q=2$, $\frac{h^{2}}{Im^{2}}\ll 1$. For the comparison, one should take Q-balls from the lower branches of Fig.~\ref{figEQ}, for which $N\approx 2$. Of course, Q-ball is a classical object and the use of our classical solution is not justified for such small values of $Q$ and $E$. However, an analogous comparison of the energy of a small Q-ball in the Minimal Supersymmetric Standard Model with the energy of the two-particle bound state in the Wick--Cutkosky model, which was performed in \cite{Kusenko:1998yj}, revealed a good qualitative agreement. Here we have the Q-ball solution in the Wick--Cutkosky model, so such a comparison can be even more illustrative.
The energy of the bound state in the Wick--Cutkosky model in the leading order in $\frac{h^{4}}{m^{4}}$ takes the form
\begin{equation}\label{BSWC}
E_{BS}=m\left(2-\left(\frac{h^{2}}{16\pi m^{2}}\right)^{2}\frac{1}{4n^{2}}\right),
\end{equation}
where $n=1,2,3,...$; see, for example, \cite{Feldman:1973jg,Darewych:1998mb}. As for the Q-balls, since the use of classical solutions is not justified for such small values of $Q$ and it is not clear what quantity should exactly correspond to $E_{BS}$, we will take both (\ref{energylowerbr}) and (\ref{bindingenergy}). Thus, in the leading order in $\frac{h^{4}}{m^{4}}$ and for $Q=2$ we get from (\ref{energylowerbr}) and (\ref{Q}), (\ref{bindingenergy})
\begin{eqnarray}
E&\approx&m\left(2-\frac{h^{4}}{3I^{2}m^{4}}\right),\\
\epsilon&\approx&m\frac{4h^{4}}{3I^{2}m^{4}}.
\end{eqnarray}
Using $I\approx 44.05$ and with $n=1$ in (\ref{BSWC}), we obtain
\begin{eqnarray}
2m-E_{BS}&\approx& \frac{h^{4}}{m^{3}}\,9.9\times 10^{-5},\\
2m-E&\approx& \frac{h^{4}}{m^{3}}\,1.7\times 10^{-4},\\
\epsilon&\approx& \frac{h^{4}}{m^{3}}\,6.9\times 10^{-4}.
\end{eqnarray}
Surprisingly, the agreement is very good taking into account the use of a classical solution for describing purely quantum effects. This implies that, for $N\gg 1$, for which the use of classical Q-ball solutions is justified, the method presented above indeed can give a result somewhat close to the real binding energy and it can be used at least for rough estimates.
\section*{Acknowledgements}
The authors are grateful to S.~Troitsky for valuable discussions. The work was supported by the Grant 16-12-10494 of the Russian Science Foundation.
\section*{Appendix~A: The absence of Q-balls for $\omega^{2}-m^{2}\ge 0$}
According to (\ref{chiequiv})--(\ref{vac-shift}), we can use the effective action
\begin{eqnarray}\label{effactapp}
S_{eff}=\int\Bigl((\omega^{2}-m^{2})f^{2}-\partial_{i}f\partial_{i}f-\frac{1}{2}\partial_{i}\tilde\phi\partial_{i}\tilde\phi-h\tilde\phi f^{2}\Bigr)d^3x,
\end{eqnarray}
where $i=1,2,3$, instead of the initial action. Suppose that there exists a Q-ball solution $f(\vec x)$, $\tilde\phi(\vec x)$. Let us apply the scale transformation $f(\vec x)\to f_{\lambda}(\vec x)=\lambda f(\lambda\vec x)$, $\tilde\phi(\vec x)\to \tilde\phi_{\lambda}(\vec x)=\lambda \tilde\phi(\lambda\vec x)$ to this solution and substitute the result into the effective action (\ref{effactapp}) instead of the original solution (in fact, it is just a generalization of the technique which was used in \cite{Derrick} to show the absence of time-independent soliton solutions in some nonlinear scalar field theories). We get
\begin{eqnarray}\nonumber
S_{eff}^{\lambda}=\int d^3y\frac{1}{\lambda^{3}}\Biggl(\lambda^{2}(\omega^{2}-m^{2})f^{2}(\vec y)-\lambda^{4}\frac{\partial f(\vec y)}{\partial y^{i}}\frac{\partial f(\vec y)}{\partial y^{i}}\\
-
\lambda^{4}\frac{1}{2}\frac{\partial \tilde\phi(\vec y)}{\partial y^{i}}\frac{\partial \tilde\phi(\vec y)}{\partial y^{i}}-\lambda^{3}h\tilde\phi(\vec y) f^{2}(\vec y)\Biggr),
\end{eqnarray}
where we have passed to the new coordinates $\vec y=\lambda \vec x$. According to the principle of least action
\begin{equation}
\frac{S_{eff}^{\lambda}}{d\lambda}\biggl|_{\lambda=1}=0,
\end{equation}
which results in
\begin{eqnarray}\label{scaletr}
(\omega^{2}-m^{2})\int f^{2}\,d^3x+\int\partial_{i}f\partial_{i}f\,d^3x+\frac{1}{2}\int\partial_{i}\tilde\phi\partial_{i}\tilde\phi\,d^3x=0.
\end{eqnarray}
It is clear that if $\omega^{2}-m^{2}\ge 0$ (which includes the case $m^{2}=h\phi_{0}\le 0$), then the only solution to (\ref{scaletr}) is
\begin{eqnarray}
f(\vec x)\equiv 0,\\ \tilde\phi(\vec x)\equiv 0.
\end{eqnarray}
\section*{Appendix~B: The relation $\frac{dE}{dQ}=\omega$}
It is reasonable to suppose that the only parameter, which
characterizes the charge and the energy of the Q-ball, is $\omega$. Thus, differentiating (\ref{energydef}) with respect to $\omega$, we get
\begin{eqnarray}\nonumber
\frac{dE}{d\omega}=4\pi\int\left(2\omega f^2+2\omega^2f\frac{df}{d\omega}+
2\partial_{r}f\partial_{r}\frac{df}{d\omega}+2m^{2}f\frac{df}{d\omega}\right.\\
\nonumber+\left.2hf\frac{df}{d\omega}\tilde\phi+hf^{2}\frac{d\tilde\phi}{d\omega}+\partial_{r}\tilde\phi\partial_{r}\frac{d\tilde\phi}{d\omega}
\right)r^{2}dr.
\end{eqnarray}
Integrating by parts the terms with derivatives in the latter formula (since it is supposed that $\tilde\phi(r)\sim\frac{1}{r}$ for $r\to\infty$ and consequently $\frac{d\tilde\phi(r)}{d\omega}\sim\frac{1}{r}$, the surface term, arising when an integration by parts is
performed, obviously vanishes), using equations of motion (\ref{eq1}), (\ref{eq2}) and the definition of the charge (\ref{chargedef}),
we get
\begin{eqnarray}\nonumber
\frac{dE}{d\omega}=4\pi\int\left(2\omega f^2+4\omega^2f\frac{df}{d\omega}\right)r^{2}dr=\omega\frac{dQ}{d\omega},
\end{eqnarray}
leading to
\begin{eqnarray}\label{dEdQgauged}
\frac{dE}{dQ}=\omega
\end{eqnarray}
for $\frac{dQ}{d\omega}\ne 0$. The points at which $\frac{dQ}{d\omega}=0$
(and, consequently, $\frac{dE}{d\omega}=0$) correspond to the cusps on the $E(Q)$ diagram and separate ``stable'' and ``unstable'' branches.
\section*{Appendix~C: Numerical solution}
In order to solve the system of equations (\ref{eq1a}), (\ref{eq2a}) numerically, it is convenient to pass to the new variables
\begin{eqnarray}\label{auxsol1}
F(R)=(C-1)\hat F(Y),\\ \label{auxsol2}
G(R)=-C+(C-1)\,\hat G(Y),\\
Y=\sqrt{C-1}\,R,
\end{eqnarray}
where $C=-G(0)>1$. In these notations the system of equations (\ref{eq1a}), (\ref{eq2a}) can be rewritten as
\begin{eqnarray}\label{eq1b}
-\Delta_{Y}\hat F-\hat F+\hat F\hat G=0,\\ \label{eq2b}
-\Delta_{Y}\hat G+{\hat F}^2=0
\end{eqnarray}
with the boundary conditions
\begin{eqnarray}\nonumber
\partial_{Y}\hat F|_{Y=0}=0,\qquad \lim\limits_{Y\to\infty}\hat F(Y)=0,\\ \label{boundary1b}
\partial_{Y}\hat G|_{Y=0}=0, \qquad \hat G|_{Y=0}=0.
\end{eqnarray}
Without loss of generality, we suppose that $\hat F(Y)>0$ for any $Y$.
Contrary to the case of initial equations (\ref{eq1a}), (\ref{eq2a}) with (\ref{boundary1a}), in which it is necessary to scan over the two parameters $F(0)$ and $G(0)$ in order to find a solution, in the case of the system of equations (\ref{eq1b}), (\ref{eq2b}) with (\ref{boundary1b}) one should scan over only the one parameter $\hat F(0)$ searching for such $\hat F(Y)$ that it falls off exponentially for $Y\to\infty$. The corresponding solution to equations (\ref{eq1b}), (\ref{eq2b}) with (\ref{boundary1b}) can easily be found numerically; the result is presented in Fig.~\ref{numsolaux}.
\begin{figure*}[!h]
\centering
\includegraphics[width=0.95\linewidth]{fig5.eps}
\caption{Numerical solution for the functions $\hat F(Y)$ and $\hat G(Y)$.}\label{numsolaux}
\end{figure*}
The value of $\hat F(0)$ for this solution turns out to be $\hat F(0)\approx 1.089$.
Now let us calculate the value of the constant $C$. From equation (\ref{eq2b}) it follows that
\begin{equation}
\hat G(\infty)=\int\limits_{0}^{\infty}{\hat F}^2(Y)YdY.
\end{equation}
On the other hand, from (\ref{auxsol2}) we get
\begin{equation}
C=(C-1)\,\hat G(\infty),
\end{equation}
leading to
\begin{equation}
C=\frac{\hat G(\infty)}{\hat G(\infty)-1}=\frac{\int_{0}^{\infty}{\hat F}^2(Y)YdY}{\int_{0}^{\infty}{\hat F}^2(Y)YdY-1}.
\end{equation}
Using the numerical solution for $\hat F(Y)$, the value of the constant $C$ was found to be $C\approx 1.938$. Using the explicit value of the constant $C$, it is not difficult to restore the solution to equations (\ref{eq1a}), (\ref{eq2a}) with (\ref{boundary1a}), which is presented in Fig.~\ref{numsol}, from the numerical solution for the functions $\hat F(Y)$ and $\hat G(Y)$.
The last step is to find the value of the parameter $I$ in formulas (\ref{Q}), (\ref{E}). According to the definition (\ref{Idef}) of the parameter $I$ and with the help of (\ref{auxsol1}), we get
\begin{equation}
I=4\pi\sqrt{C-1}\int\limits_{0}^{\infty}{\hat F}^{2}(Y)Y^{2}dY.
\end{equation}
The latter integral can also easily be evaluated numerically, resulting in $I\approx 44.05$.
\section*{Appendix~D: The model of \cite{Levin:2010gp} for $M-|\omega|\ll M$}
The scalar field potential in \cite{Levin:2010gp} has the form $h\phi^{2}\chi^*\chi$. For the vacuum solution $\chi\equiv 0$, $\phi\equiv\phi_{0}$ the mass of the free charged scalar particle is defined by $M^{2}=h\phi_{0}^{2}$. Using (\ref{chiequiv}), (\ref{phiequiv}) and the redefinition
\begin{eqnarray}
R=r\sqrt{M^{2}-\omega^{2}},\qquad F(R)=\frac{2h\phi_{0}}{M^2-\omega^2}f(r),\qquad G(R)=\frac{2h\phi_{0}}{M^2-\omega^2}\tilde\phi(r),
\end{eqnarray}
we get for the model of \cite{Levin:2010gp}
\begin{eqnarray}\label{eqLRa}
-\Delta_{R}F+F+FG+\frac{1}{4}\lambda(\omega) FG^{2}=0,\\ \label{eqLRb}
-\Delta_{R}G+F^2+\frac{1}{2}\lambda(\omega) F^{2}G=0,
\end{eqnarray}
where $\lambda(\omega)=1-\frac{\omega^{2}}{M^{2}}$. It is clear that $\lambda(\omega)\ll 1$ for $M-|\omega|\ll M$ and the terms proportional to $\lambda(\omega)$ in (\ref{eqLRa}), (\ref{eqLRb}) can be neglected in this case, leading to the system of equations (\ref{eq1a}), (\ref{eq2a}). Thus, the $E(Q)$ dependence for Q-balls in the model of \cite{Levin:2010gp} for very small charges is almost the same as the one for Q-balls with small charges from the lower branches of Fig.~\ref{figEQ}.
|
\section{\@startsection%
{section
{1
\z@{1.5\linespacing\@plus .2\linespacing
{.7\linespacing
{\normalfont\sc\centering}
\makeatother
\makeatletter
\renewenvironment{proof}[1][\proofname]{\par
\pushQED{\qed}%
\normalfont \topsep6\p@\@plus6\p@\relax
\trivlist
\itemindent\normalparindent
\item[\hskip\labelsep
\bfseries
#1\@addpunct{.}]\ignorespaces
}{%
\popQED\endtrivlist\@endpefalse
}
\providecommand{\proofname}{Proof}
\makeatother
\newcommand{\footnoteb}[1]{\footnote{~#1}}
\newenvironment{itemizeb}
{\begin{itemize}\itemsep=2pt\leftskip -10 pt}
{\end{itemize}}
\theoremstyle{remark}
\newtheorem{assumption}[subsection]{Assumption}
\theoremstyle{definition}
\newtheorem{definition}[subsection]{Definition}
\def\mint
{\mkern12mu\hbox{\vrule height4pt
depth-3.2pt width5pt}%
\mkern-16.5mu\int}
\def\smint
{\mkern3mu\hbox{\vrule height4pt
depth-3.2pt width5pt}%
\mkern-12.5mu\int}
\newcommand{\mathscr{Q}}{\mathscr{Q}}
\newcommand{\mathscr{T}}{\mathcal{T}}
\renewcommand{\SS}{\mathcal{S}}
\newcommand{\mathcal{F}}{\mathcal{F}}
\usepackage{amsmath,amsfonts,latexsym}
\usepackage{color}
\usepackage{cases}
\newcommand{\textcolor{blue}}{\textcolor{blue}}
\newcommand{\textcolor{red}}{\textcolor{red}}
\definecolor{Green}{rgb}{0.010,0.7,0.02}
\newcommand{\textcolor{Green}}{\textcolor{Green}}
\newcommand{\textcolor{black}}{\textcolor{black}}
\newcommand{\mathbf{u}}{\mathbf{u}}
\newcommand{\mathcal{D}}{\mathcal{D}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{N}}{\mathbb{N}}
\newcommand{\mathcal{O}}{\mathcal{O}}
\newcommand{\mathbb{T}}{\mathbb{T}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{i.\thinspace e.}{i.\thinspace e.}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\epsilon}{\epsilon}
\newcommand{\mathcal{D}}{\mathcal{D}}
\newcommand{\Tocomment}[1]{\textcolor{red}{#1}}
\begin{document}
\thispagestyle{empty}
~\vskip -1.1 cm
%
\vspace{1.7 cm}
%
{\LARGE\bf
\begin{center}
Exponential self-similar mixing \\
by incompressible flows
\end{center}
}
\vspace{.4 cm}
%
\centerline{\sc
Giovanni Alberti, Gianluca Crippa, Anna L.~Mazzucato}
\vspace{.7 cm}
\centerline{\small\sl
Dedicated to Alberto Bressan and Charles R.~Doering
on the occasion of their 60\,$^{th}$ birthdays}
\vspace{.9 cm}
{\rightskip 1 cm
\leftskip 1 cm
\parindent 0 pt
\footnotesize
%
{\sc Abstract.}
We study the problem of the optimal mixing of a passive scalar under the
action of an incompressible flow in two space dimensions. The scalar
solves the continuity equation with a divergence-free velocity field which
satisfies a bound in the Sobolev space $W^{s,p}$, where $s \geq 0$ and
$1\leq p\leq \infty$. The mixing properties are given in terms of a
characteristic length scale, called the mixing scale. We consider two notions of
mixing scale, one functional,
expressed in terms of the homogeneous Sobolev norm~$\dot H^{-1}$,
the other geometric, related to rearrangements of sets.
We study rates of decay in time of both scales under self-similar mixing. For
the case~$s=1$ and $1 \leq p \leq \infty$ (including the Lipschitz case, and
the case of physical interest of enstrophy-constrained flows), we present examples of
velocity fields and initial configurations for the scalar that saturate the exponential
lower bound established in previous works for the decay in time of both scales.
We also present several consequences for the geometry of regular Lagrangian
flows associated to Sobolev velocity fields.
\par
\medskip\noindent
{\sc Keywords:} mixing, continuity equation, negative Sobolev norms, incompressible flows,
self-similarity, potentials, regular Lagrangian flows.
\par
\medskip\noindent
{\sc MSC (2010):} 35Q35, 76F25
\par
}
\tableofcontents
\section{Introduction}
\label{intro}
We study the problem of optimal mixing of scalar, passive tracers by incompressible flows.
How well a quantity transported by a flow is mixed is an important problem in fluid mechanics
and in many applied fields, for instance in atmospheric and oceanographic science, in biology, and in
chemistry. In combustion, for example, fuel and air need to be well mixed for an
efficient reaction to take place. In many situations, the interaction between the tracer and the
flow can be neglected: mathematically, this results in the fact that the tracer solves a linear
continuity equation with a given velocity field (see~\eqref{e:continuity}).
This problem is also a surprisingly rich source of questions in analysis,
relating partial differential equations and dynamical systems with geometric measure theory in particular.
\medskip
There is a well-established fluid mechanics literature concerning mixing and turbulence,
especially with respect to statistical properties (see e.g.~\cite{BCCLV00,GW12} and
references therein). It is known, in fact, that turbulent advection enhances mixing, which in turn can enhance
diffusion and suppress concentration (see \cite{CKRZ} for steady ``relaxation enhancing" flows
and \cite{KXchemo} for an application to chemotaxis, for instance).
Enhanced dissipation occurs also in Euler flows as an effect of inviscid Landau damping (see \cite{BMV} and references therein).
Mixing has also long been studied in the context of chaotic dynamics \cite{Aref,Ottino,liverani}.
Indeed the decay to zero of the mixing scale defined in terms of negative Sobolev norms
corresponds to ergodic mixing by the flow
(as shown in~\cite{MMGVP}), and several well-known examples of discrete dynamical systems
exhibit an exponential decay of correlations, which essentially
means exponential mixing (however, these examples cannot be easily adapted to our context).
Recently there has been a renewed interest in quantifying the degree of mixing under an incompressible flow, and in producing examples that achieve optimal mixing. On the analytical side, progress has been possible in part due to the development of new tools to study transport and continuity equations under non-Lipschitz velocities~\cite{DPL,Amb, HW}, in particular quantitative estimates on regular Lagrangian flows \cite{CDL}. On the applied and computational side, optimal mixing has been approached from the point of view of homogenization and control with more realistic models~\cite{Liu08,FCS14}. More accurate experiments have also been performed (see for example~\cite{GDTR09,JCT00,Jul03}).
\subsection{The continuity equation}
\label{ss:cont}
We consider the two-dimensional setting, as it is the first dimension with non-trivial
divergence-free velocity fields and for comparison with computational and experimental
studies. Generally, dimension will not play a crucial role if what follows, except in setting
scaling laws.
The divergence-free condition is a strong
constraint that can be somewhat relaxed, but it is physically motivated in applications of mixing.
In fact, since we aim at producing examples of optimal mixing, the divergence-free
condition is a more restrictive requirement that must be satisfied in our constructions.
\medskip
We work on the two-dimensional torus $\mathbb{T}^2 := \mathbb{R}^2 / \mathbb{Z}^2$ or on the plane $\mathbb{R}^2$.
When considering the plane $\mathbb{R}^2$, both velocity fields and solutions eventually
resulting from our constructions will be supported in a fixed compact set.
Given a divergence-free, time-dependent velocity field $u = u(t,x)$,
we consider a passive scalar $\rho = \rho(t,x) \in L^\infty$ which is advected by $u$, i.e., a solution of
the continuity equation
\begin{equation}
\label{e:continuity}
\partial_t \rho + \mathop{\mathrm{div}} (u \rho) = 0 \,,
\end{equation}
with prescribed initial datum $\rho(0,\cdot) = \bar \rho$ at time $t=0$.
We always assume without loss of generality%
\footnoteb{\label{fn:aver} It is enough to subtract to the solution its total integral, and observe
that constants are solutions of \eqref{e:continuity}.}
that the initial datum $\bar \rho$ satisfies the zero-average
condition $\int \bar \rho = 0$, where the integration is performed over~$\mathbb{T}^2$ or
over~$\mathbb{R}^2$, depending on the case under consideration. Since the continuity
equation~\eqref{e:continuity} preserves the total integral of the solution along
the time evolution, it follows that $\rho(t,\cdot)$ has average zero for any time $t$
(this is relevant in connection with the measurement of the mixing scale with
negative Sobolev norms, see Definition~\ref{def:Functscale}, \S\ref{ss:defH}, and
Remark~\ref{r:fctsinH}). This assumption is also
physically consistent with $\rho$ representing a tracer.
\subsection{Functional and geometric mixing scales}
In order to discuss the mixing properties of solutions to the continuity
equation~\eqref{e:continuity} we need to define a notion of mixing scale,
which can give a quantification of the ``level of mixedness'' of the solution $\rho(t,\cdot)$
at time $t$. The continuity equation preserves (formally at least) all $L^p$
norms of the solution.%
\footnoteb{In fact, $L^p$ norms of the solutions are frequently used as a
measurement of the mixing scale for solutions of advection-diffusion
equations, i.e., in the case when $\rho$ solves $\partial_t \rho + \mathop{\mathrm{div}}(u\rho)
= \Delta \rho$. Due to the viscosity, $L^p$ norms of the solution are dissipated along the time evolution.}
Though, it is still possible for $\rho(t,\cdot)$ to converge to zero weakly.%
\footnoteb{Using characteristic functions of sets as test functions, it is not difficult
to prove that this will be the case for instance if the flow of $u$ is strongly mixing in the ergodic sense.}
In fact, the vanishing of the homogeneous negative Sobolev norm $\dot H^{-1}$ of $\rho(t,\cdot)$
(see \S\ref{ss:defH} for its definition) can be proved to be equivalent to the convergence
of $\rho(t,\cdot)$ to $0$ weakly in $L^2$ (see for instance~\cite{doering}). More generally,
the use of negative norms to measure mixing was proposed
in~\cite{MMP}, where the equivalence between the decay of the $\dot H^{-1/2}$
norm and mixing in the ergodic sense was established.
\medskip
In this paper, we will employ and compare two notions of mixing scales that are
considered in the literature. The first one is based on a negative
Sobolev norm of the solution $\rho(t,\cdot)$, more precisely the norm in the
homogeneous Sobolev space $\dot{H}^{-1}$ following \cite{doering}
(see Definition~\ref{def:Functscale} below), and will be referred to as
the {\em functional mixing scale}.%
\footnoteb{From a mathematical point of view there is nothing special with the
order $-1$ that has been chosen in the definition of functional mixing scale: every negative
Sobolev norm would behave in a similar way. However, from a physical point of view,
this choice is the most convenient, since the norm in $\dot{H}^{-1}$ scales
as a length.}
The second mixing scale arises from a conjecture of Bressan~\cite{bressan}
on the cost of rearrangements of sets and brings in a connection with geometric
measure theory. This second notion of scale is expresses in terms of
the relative proportion in suitably small balls of different level sets of the
solution (see Definition~\ref{d:geomix} below), and it will be referred to as
the {\em geometric mixing scale}.
The two scales are related though generally not equivalent.
In the next few paragraphs in this introduction, {\em we informally denote
any of the two mixing scales of the solution at time $t$ by
${\rm mix}\, \big( \rho(t,\cdot) \big)$}.
\medskip
Ideally, a flow that ``mixes optimally'' will achieve the largest decay rate in time for
${\rm mix}\, \big( \rho(t,\cdot) \big)$.
How fast ${\rm mix}\, \big( \rho(t,\cdot) \big)$ can decay in time depends
on properties of the flow. These, in turn, are in practice given in terms of constraints
on certain quantities of physical interest, typically energy, enstrophy, and palenstrophy.
These correspond respectively to uniform-in-time bounds on the $L^2$, $H^1$, and
$H^2$ norms of the velocity field $u$.
\subsection{Main results}
\label{ss:intromain}
In this article, we provide explicit examples of two-dimensional flows that mix certain
configurations optimally under a fixed enstrophy budget.
In fact, more generally our work gives examples of optimal mixers under a
constraint on the Sobolev norm $W^{1,p}$ of the velocity field, where $1\leq p\leq \infty$,
i.e., the Lipschitz case is included.
As described in detail in \S\ref{ss:literature} below, under uniform in time constraints
on the Sobolev norm $W^{1,p}$ of the velocity field, where $1 < p \leq \infty$, it has been
recently proved~\cite{CDL,ikx,seis} that the (functional or geometric) mixing scale
can decay at most exponentially in time:
\begin{equation}
\label{e:introbound}
{\rm mix}\, \big( \rho(t,\cdot) \big) \geq C \exp ( -ct) \,,
\end{equation}
where $C>0$ and $c>0$ are constant depending on the initial datum $\bar \rho$
and on the given bounds on the velocity field. The optimality of this bound in
the full range of summability exponents was previously unknown (see however~\cite{zlatos}
and the brief description in~\S\ref{ss:literature} below).
Our main results are contained%
\footnoteb{The verification of the assumption of the two theorems is
carried on in Sections~\ref{pinching} and \ref{snake} respectively.}
in Theorem~\ref{t:main} and Theorem~\ref{t:ass} and can be summarized as follows:
\medskip
\noindent{\em There exist a smooth, bounded, divergence-free velocity field
$u$, which is Lipschitz uniformly in time, and a bounded nontrivial solution $\rho$ of the continuity
equation~\eqref{e:continuity}, such that the (functional or geometric) mixing scale
of the solution decays exponentially in time:
\begin{equation}\label{e:intromain}
{\rm mix}\, \big( \rho(t,\cdot) \big) \leq C \exp ( -ct) \,,
\end{equation}
which shows the optimality of the lower bound~\eqref{e:introbound}.}
\begin{remark}
\label{r:remmain}
(i)~The optimal decay in~\eqref{e:intromain}
holds for both the functional and the geometric mixing scales.
\smallskip
(ii)~The solution $\rho$ can be required to be smooth, or to be the
characteristic function of a set which
evolves in time under the action of the flow.
\smallskip
(iii)~Our procedure extends to a full scaling
analysis: if the velocity field is required to be bounded in $W^{s,p}$ uniformly in time,
for some $s \geq 0$ real, then the constructed $u$ and $\rho$ are such that:
\begin{itemizeb}
\item If $s<1$ there is a time $t^*$ such that ${\rm mix}\big(\rho(t^*,\cdot)\big)=0$, i.e.,
there is perfect mixing in finite time;
\item If $s=1$ the exponential decay of the mixing scale in \eqref{e:intromain} holds;
\item If $s>1$ there is polynomial decay of the mixing scale: there exists an exponent
$\alpha=\alpha(s)>0$ such that ${\rm mix}\, \big( \rho(t,\cdot) \big) \leq C t^{-\alpha}$.
\end{itemizeb}
\smallskip
(iv)~It is very important to keep in mind the difference between the regularity
of the
velocity field, and the regularity bounds that it satisfies uniformly in time. For instance,
in the situation with exponential decay of the mixing scale of the solution
sketched in \S\ref{ss:intromain}, the velocity field is smooth in space and time;
however, only Sobolev norms $W^{s,p}$ with $0 \leq s \leq 1$ and $1 \leq p \leq \infty$
are bounded uniformly in time. If $s>1$, such norms are bounded for every given $t$,
but blow up when~$t \to \infty$.%
\footnoteb{In formulas, we have that $u \in L^\infty \big( [0,+\infty) ; W^{s,p}(\mathbb{R}^2) \big)$
if and only if $0 \leq s \leq 1$.}
\smallskip
Further remarks
are detailed in \S\ref{ss:introself}
and \S\ref{ss:introrem} below. The results presented in this article were announced in \cite{ACM}.
\end{remark}
Before making further observations on our results and techniques we make a
digression about the past literature on this topic.
\subsection{Past literature}
\label{ss:literature}
Mixing phenomena are studied in the literature under energetic constraints
on the velocity field, that is, assuming that the velocity field is bounded with
respect to some spatial norm, uniformly in time. This research area is related
in a very natural way to the study of transport and continuity equations under non-Lipschitz
velocities (see~\cite{HW} for a recent survey of this research area). We now make an
overview of some previous works on these topics
(most of the results hold indeed in any space dimension):
\begin{itemizeb}
\item[(a)]
The velocity field $u$ is bounded in $W^{s,p}$ uniformly in time for
some $s<1$ and $1\leq p\leq \infty$ (the case $s=0$, $p=2$, relevant for the
applications, is often referred to as energy-constrained flow).
In this case, there is in general no uniqueness for the Cauchy problem for
the continuity equation~\eqref{e:continuity} (see~\cite{ABC2,ABC1}). Hence, one can find
a velocity field and a bounded solution which is non-zero at the initial time, but is identically zero at some later time. Therefore it is possible to have perfect mixing in finite time, as
already observed in \cite{doering} and established in \cite{llnmd} for $s=0$,
building on an example from \cite{depauw,bressan}.
\item[(b)]
The velocity field $u$ is bounded in $W^{1,p}$ uniformly in time for
some~$1\leq p\leq \infty$ (the case $p=2$, relevant for the
applications, is often referred to as enstrophy-constrained flow).
The theory in \cite{DPL} guarantees uniqueness for the Cauchy problem~\eqref{e:continuity},
which in particular excludes perfect mixing in finite time. A quantification of the
maximal decay rate for the mixing scale has been achieved thanks to
the quantitative Lipschitz estimates for regular Lagrangian flows in~\cite{CDL}.
In detail, for $p>1$, the theory in~\cite{CDL} provides an exponential lower bound on the geometric mixing
scale (see~\eqref{e:introbound}). The extension to the borderline case $p=1$ is
still open (see, however, \cite{BC}). The same exponential lower bound~\eqref{e:introbound}
has been proved for the functional mixing scale in~\cite{ikx,seis}.
A new proof of these lower bounds based on harmonic analysis estimates has been proposed in~\cite{leger}.
\item[(c)]
The theory in~\cite{Amb} provides uniqueness for the Cauchy problem~\eqref{e:continuity}
for velocity fields bounded in $BV$ uniformly in time. Again, this excludes
perfect mixing in finite time. The validity of the bound~\eqref{e:introbound} is still
unknown in this context. However, in~\cite{bressan} it is observed that such
an exponential decay of the geometric mixing scale can indeed be attained for velocity fields bounded in $BV$ uniformly in time. Actually, the same example works also for the functional mixing scale.
\item[(d)]
The velocity field $u$ is bounded in $W^{s,p}$ uniformly in time for some $s>1$
and $1\leq p\leq \infty$ (the case $s=2$ and $p=2$, relevant for the
applications, goes under the name of palenstrophy-constrained flow).
No better bounds for the decay of the (functional or geometric) mixing scale
than~\eqref{e:introbound} are known. The common belief, supported also
by the numerical simulations in \cite{llnmd,ikx}, is that this bound is optimal.
See also the discussion in~\S\ref{ss:introrem}.
\end{itemizeb}
\medskip
Besides depending on the given bounds on the velocity field, the
constants $C$ and $c$ in~\eqref{e:introbound} depend on the initial datum
$\bar\rho$, and not only on its mixing scale. In fact, even for the functional
mixing scale, it is not clear that an estimate of the form
\[
\|\bar\rho(t,\cdot)\|_{\dot{H}^{-1}} \geq C
\|\bar\rho\|_{\dot{H}^{-1}} \exp (-ct),
\]
with $C$ and $c$ universal constants depending only on the given
bounds on the velocity field, can be achieved.
It is then natural to ask whether it is possible to obtain bounds on the rate of
decay with constants that only depend on the mixing scale of the initial tracer
configuration, and not on its geometry. Unfortunately, direct PDE methods, such
as energy estimate, do not seem to yield sharp bounds. For instance, they yields
a Gaussian bound for palenstrophy-constrained flows, when the optimal
bound is at least exponential (see~\cite{llnmd}).
\medskip
Besides the examples constructed in the present article (\S\ref{ss:intromain}),
there are other explicit examples in the literature of flows that saturate the
exponential decay rate for enstrophy-constrained flows. Yao and
Zlato\v s~\cite{zlatos} utilize a cellular flow to establish exponential mixing
for any bounded solution and with a bound on the $W^{1,p}$-norm of
the velocity field in the range $1 \leq p \leq \bar p$ for some $\bar p >2$.
They also give an interesting result on ``unmixing" a given configuration.
See also \S\ref{ss:introrem} for a comparison of our results with those
from~\cite{zlatos}.
\medskip
Before these analytical results, several numerical experiments have been
performed, that supported an exponential rate of decay for ${\rm mix}\, \big( \rho(t,\cdot) \big)$
under an enstrophy budget.
For instance, in~\cite{doering} a flow that instantaneously maximizes the rate
of decay of the mixing scale for a sinusoidal initial configuration of the passive
scalar was displayed. A global optimizer was computed numerically in~\cite{MMGVP}.
\subsection{Self-similarity ansatz and structure of our constructions}
\label{ss:introself}
In order to implement our constructions of optimal mixers we follow a
self-similar scheme. In few words, we assume that the solution
undergoes (on a suitable time scale) an evolution that is constructed
starting from a basic move and then successively rescaling this move on finer
and finer spatial scales
(the reader can glimpse at this construction in
Figure~\ref{f:self}).
A few comments are in order. We assume the self-similarity as an ansatz
for our examples. We do not claim that a self-similar evolution is more physical%
\footnoteb{Mixing patterns with decaying amplitude but invariant geometrical structure have
been experimentally observed~\cite{RHG99}.}
or more mixing-optimizing than other evolutions. We consider self-similar evolutions
since this choice has the substantial advantage of making computations
rather easy: given the existence of the basic move, all estimates for the
evolution can be derived through a fairly simple scaling analysis. The difficulty is
of course proving the existence of the basic move within the required regularity
setting.
\medskip
Following this line of thinking, our analysis
contains three main stages:
\begin{itemizeb}
\item[(1)] Scaling analysis: the existence of a basic move for time $0 \leq t \leq 1$
being assumed, we define the evolution for all times $t \geq 0$ and fully characterize the decay of
the mixing scale of the solution and the energetic constraints on the
velocity field (Section~\ref{selfsimilar});
\item[(2)] Geometric tools: we establish
a series of geometric lemmas that guarantee the existence of smooth, divergence-free
velocity fields with the property that the associated flows deform smooth sets
according to
a prescribed evolution in time (Section~\ref{basic});
\item[(3)] Existence of the basic move: we use such geometric tools to construct the basic move
(that is, we construct the velocity field and a weak solution of the
continuity equation, the level sets of which realize the geometric evolution).
The main technical point here is to deal
with the singularities that are present in the evolution. This last stage ``validates'' the
scaling analysis in~(1) (Section~\ref{pinching}).
\end{itemizeb}
We notice that, although self-similarity can be quite a rigid constraint, the
decay in time of
the mixing scale of an enstrophy-constrained self-similar evolution matches the
known exponential lower bound (recall~\eqref{e:intromain}
and~\S\ref{ss:literature}(b)).
This is not the case in the context of palenstrophy-constrained
flows, where the rate of decay ensuing from self-similar evolution is
polynomial in time, rather than the observed exponential rate
(recall~\S\ref{ss:literature}(d)).%
\medskip
One shortcoming of the strategy based on self-similarity is the limited
regularity available for velocity field and solution. In Section~\ref{pinching}, the velocity
field cannot have Lipschitz regularity and the solution is a characteristic function of a set evolving in time.
In order to overcome this problem, we introduce in Section~\ref{quasiselfsimilar}
the notion of quasi self-similar evolution. Instead of replicating rescaled copies of one
basic element at each step of the evolution, we rather consider a finite family of basic elements,
which are rescaled and rearranged at each step of the evolution according
to a combinatorial pattern (the reader can glimpse at this construction in
Figures~\ref{f:idea} and \ref{f:ideatwo}).
In Sections~\ref{quasiselfsimilar}, \ref{geo2}, and \ref{snake} we follow the
same line of
work described in (1)--(2)--(3) above: we perform a scaling analysis under the
condition that a finite family of basic moves exists, we develop geometric tools
well suited to deform
smooth profiles, and eventually we use such tools to construct the family of the basic moves.
In this way, we are able to obtain a velocity field and an associated solution
of the continuity equation that are actually smooth.
\subsection{Further remarks and open problems}
\label{ss:introrem}
Our main result is a proof of the optimality of \eqref{e:introbound}. In order
to do this, we construct one velocity field, which is bounded in $W^{1,p}$
uniformly in time for any~$1 \leq p \leq \infty$,
and one solution, whose mixing scale decays exponentially. In fact, with our strategy it
is possible to construct a large class of initial data whose mixing scale decays exponentially. However,
it is unclear whether this is the case for every initial datum. The following questions about the
existence of ``universal mixers'' are therefore very natural.
\begin{itemizeb}
\item[(a)] Given a bounded initial datum $\bar \rho$, does it exist a velocity
field, bounded in~$W^{1,p}$ uniformly in time and possibly dependent on $\bar \rho$,
such that ${\rm mix}\, \big( \rho(t,\cdot) \big)$ decays to zero?
\item[(b)] If the answer to Question~(a) is positive, does it exist
a velocity field, bounded in~$W^{1,p}$ uniformly in time, such that the decay to zero of
${\rm mix}\, \big( \rho(t,\cdot) \big)$ is precisely exponential?
\item[(c)] Does it exist a velocity field, bounded in~$W^{1,p}$ uniformly in time,
such that ${\rm mix}\, \big( \rho(t,\cdot) \big)$ decays to zero for every bounded initial datum $\bar \rho$?
\item[(d)] If the answer to Question~(c) is positive, does it exist a velocity field, bounded in~$W^{1,p}$ uniformly in time,
such that for any initial data $\bar \rho$ the decay to zero of
${\rm mix}\, \big( \rho(t,\cdot) \big)$ is precisely exponential?
\end{itemizeb}
We observed in Remark~\ref{r:remmain}(iii) that our construction provides an
example of palenstrophy-constrained flow such that ${\rm mix}\, \big( \rho(t,\cdot) \big)$
decays polynomially in time. More in general, the self-similarity ansatz implies polynomial
decay of ${\rm mix}\, \big( \rho(t,\cdot) \big)$ in the case the velocity field is bounded
in $W^{s,p}$ uniformly in time for some $s>1$. However, the numerical results
mentioned in~\S\ref{ss:literature}(d) support an exponential decay also
for $s>1$.
If this were the case, we would deduce that for $s>1$ self-similarity is too
strong a constraint, which only allows for sub-optimal decay rates (polynomial
vs.~exponential). This result would be in stark contrast with the case $s=1$,
for which the optimal decay rate can be obtained within self-similar evolutions.
We therefore point out the following important
question:
\begin{itemizeb}
\item[(e)] Does it exist a bounded initial datum $\bar \rho$ and a velocity field bounded
in $W^{s,p}$ uniformly in time for some $s>1$ such that
${\rm mix}\, \big( \rho(t,\cdot) \big)$ decays exponentially in time?
\end{itemizeb}
Such an example has necessarily to be non self-similar. The analysis in~\cite{schulze}
implies in addition that such an example cannot be realized with a ``localized'' flow:
roughly speaking, once the solution has been mixed to a certain scale, it can
be more convenient to ``unmix'' it before reaching a lower mixing scale.
\medskip
As mentioned in~\S\ref{ss:literature}, examples of enstrophy-constrained flows that saturate
the exponential decay rate~\eqref{e:introbound} have been constructed
in~\cite{zlatos}. There, the authors utilize a cellular flow consisting of
pseudo-rotations on a family
of nested tilings of the square, and are
able to obtain exponential mixing of every bounded initial datum by means of a
velocity field bounded in $W^{1,p}$ uniformly in time in the range $1 \leq p
\leq \bar p$ for some $\bar p > 2$. Therefore, this construction provides a
partial answer to Question~(b) above.
Their nice geometric argument is based on a ``stopping time'' for the
pseudo-rotation, which is determined by a clever application of the intermediate
value theorem for continuous functions. Their construction also applies in the
range $\bar p < p <\infty$, giving a mixing rate which is slightly slower
than exponential, thus answering Question~(a) above.
In comparison with~\cite{zlatos}, our construction apply only to specific
choices of initial data. On the other hand, we obtain
exponential mixing of the solution in the full range $1 \leq p \leq \infty$,
even in the
Lipschitz case. Moreover, our strategy has a clear geometric flavor, and generates
velocity fields and solutions which are smooth. This last fact is relevant for the
full scaling analysis (recall Remark~\ref{r:remmain}(iii)) and for the application to
the study of the loss of regularity for continuity equations, which will be
fully addressed in a forthcoming paper
(see~\S\ref{ss:loss} for a brief discussion). In addition, our examples provide
an important insight
into the geometrical properties of regular Lagrangian flows (see~\S\ref{ss:geomrlf}).
\subsection{Geometry of regular Lagrangian flows}
\label{ss:geomrlf}
An important aspect of our work concerns properties of solutions to the Cauchy
problem for the continuity equation~\eqref{e:continuity} and of the associated regular Lagrangian flows.%
\footnoteb{That of regular Lagrangian flows is the appropriate notion for the
flow generated by an ordinary differential equation (ODE for short) for which
the velocity field has low regularity. A regular Lagrangian flow in $\mathbb{R}^d$
solves the ODE for almost
every initial point, and additionally preserves the $d$-dimensional Lebesgue
measure
up to a bounded factor. The theory in~\cite{DPL,Amb} guarantees that, if the velocity field
is Sobolev or $BV$ and has bounded divergence, then there exists a unique
regular Lagrangian flow associated to it.}
In detail, the construction in Section~\ref{pinching} (see Remark~\ref{r:messrlf})
gives an example of a velocity field in two space dimensions that belongs to
the space $W^{s,p}$ for every $s,p$ such that $W^{s,p}$ does not embed in
the Lipschitz class, i.e., for $s<2/p+1$, with the following properties:
\begin{itemizeb}
\item The associated regular Lagrangian flow does not preserve the connectedness of sets;
\item The set of initial points for which the ODE that generates the flow has
more than one solution contains a full segment;
\item Such a segment is ``compressed'' by the regular Lagrangian flow to a
single point.%
\footnoteb{By definition, a regular Lagrangian flow in $\mathbb{R}^d$ does not compress
$d$-dimensional sets to null set. We see here that it can compress $1$-dimensional sets
to $0$-dimensional sets.}
\end{itemizeb}
Moreover, a time-reparametrization of our example in Section~\ref{snake}
gives an example of a bounded, compactly supported, divergence-free velocity field
$u$ such that $u(t,\cdot) \in {\rm Lip}(\mathbb{R}^2)$ for almost every $0 \leq t \leq 1$, and for which
there is no uniqueness of solutions for the associated continuity equation.
Indeed, such a velocity field fails to belong to $L^1([0,1]; {\rm Lip} (\mathbb{R}^2))$: the ${\rm Lip}$ norm blows
up as $t \downarrow 0$.
This example improves on the result in~\cite{depauw} in the $BV$ case.
\subsection{Follow up: loss of regularity for continuity equations}
\label{ss:loss}
Mixing by shearing and filamentation increases positive Sobolev norms
of the solution $\rho$, saturating the exponential growth which follows from the
classical Gr\"onwall inequality. Analytically, this can be seen by ``dualizing'' the
exponential decay of the negative Sobolev norms in~\eqref{e:intromain}.
In a forthcoming paper~\cite{loss}, we will present an
example of a velocity field in~$W^{1,p}$ for any $1 \leq p < \infty$ that is regular
except at a point and a smooth $\bar\rho$ such that for~$t>0$ the
corresponding solution of~\eqref{e:continuity}
leaves instantly any Sobolev space~$H^s$ with~$s>0$. Extensions to Sobolev
regularity of order higher than $1$ are also possible.
Lack of propagation of the continuity and of the $BV$ regularity for the continuity equation was
already observed in~\cite{ferruccio}.
More recently, lack of Sobolev regularity of order one for the flow of $W^{1,p}$ velocity fields was also
observed in~\cite{Jab15}, using a different construction that exploits a
randomization procedure on certain basic elements of the flow.
\subsection*{Acknowledgements}
This work was started during a visit of the first
and third authors at the University of Basel.
Their stay has been partially supported by the
Swiss National Science Foundation grants 140232 and 156112.
The visits of the second author to Pisa have been supported by the University of Pisa PRA
project ``Metodi variazionali per problemi geometrici [Variational Methods
for Geometric Problems]''.
The third author was partially supported by the
US~National Science Foundation grants
DMS~1009713, 1009714, and 1312727.
\section{Preliminaries}\label{s:prelim}
Throughout the whole paper we will make extensive use of homogeneous
Sobolev spaces with real (differentiation) order and of their properties. We
present here their definition and the main properties of interest for our
work, namely those regarding scaling, interpolation, and embeddings.
For a systematic exposition we refer the reader to~\cite{bergh,chemin,valdinoci,grafakos,triebel}.
In the last part of the section we define the two notions of mixing scale
that we will use in all our work.
\medskip
We limit our presentation to the two-dimensional case, however all definitions and
results can be extended with obvious changes to the case of higher space
dimensions.
We work both on the full plane $\mathbb{R}^2$ and on the two-dimensional torus,
which we denote by $\mathbb{T}^2 := \mathbb{R}^2 / \mathbb{Z}^2$. The Fourier transform of a
tempered distribution $f \in \SS'(\mathbb{R}^2)$ is denoted by $\hat f$ or by $\mathcal{F}(f)$, while
the inverse Fourier transform of $g$ is denoted by $\mathcal{F}^{-1}(g)$. Given
a distribution $f \in \mathcal{D}'(\mathbb{T}^2)$, for $k \in \mathbb{Z}^2$ we denote by $\hat{f}(k)$
its $k$-th Fourier coefficient.
\subsection{Homogeneous Sobolev spaces $\boldsymbol{\dot{H}^s}$}
\label{ss:defH}
For $s \in \mathbb{R}$, we say that a distribution $f \in \mathcal{D}'(\mathbb{T}^2)$ on the torus
belongs to the homogeneous Sobolev space $\dot{H}^s(\mathbb{T}^2)$ if
\begin{equation}\label{e:defHT}
\| f \|^2_{\dot{H}^s(\mathbb{T}^2)} := \sum_{k \in \mathbb{Z}^2} |k|^{2s} |\hat{f}(k)|^2 < \infty \,.
\end{equation}
If $f \in \SS'(\mathbb{R}^2)$ is a tempered distribution on the plane,
we say that $f$ belongs to the homogeneous Sobolev space
$\dot{H}^s(\mathbb{R}^2)$ if $\hat f \in L^1_{\rm loc}(\mathbb{R}^2)$ and
\begin{equation}\label{e:defH}
\| f \|^2_{\dot{H}^s(\mathbb{R}^2)} := \int_{\mathbb{R}^2} | \xi |^{2s} | \hat{f}(\xi) |^2 \, d\xi < \infty \,.
\end{equation}
We remark that homogeneous Sobolev spaces do not form a scale, due to the singularity
of the multiplier at the origin in frequency space. In particular, it is
generally not true that any square integrable function is automatically in
$\dot{H}^s$, for~$s<0$.
\begin{remark}\label{r:fctsinH}
(i)~From \eqref{e:defHT} we immediately recognize that, in order for a function $f \in L^1(\mathbb{T}^2)$
to belong to some $\dot{H}^{s}(\mathbb{T}^2)$ with $s<0$, it is necessary that $\hat{f}(0)=0$. This
corresponds to the zero average%
\footnoteb{Recalling that the functional mixing scale associated to a solution
of~\eqref{e:continuity} is given by its norm in $\dot{H}^{-1}$, we chose to
work with zero-average solutions in~\S\ref{ss:cont} (also recall Footnote~\ref{fn:aver}).}
condition $\int_{\mathbb{T}^2} f = 0$. Conversely, let
$f \in L^1(\mathbb{T}^2)$ be a
function with zero average. Since the sequence of its Fourier coefficients $\{ \hat{f}(k) \}_{k \in \mathbb{Z}^2}$
is bounded, we deduce that such a function necessarily belongs to $\dot{H}^{-s}(\mathbb{T}^2)$ for every
$s < -\frac{1}{2}$.
\smallskip
(ii)~Analogously, given a function $f \in L^1(\mathbb{R}^2)$, its Fourier transform $\hat{f}$ is continuous.
If $s \leq -1$ the singularity at $\xi=0$ in \eqref{e:defH} is not integrable, unless $\hat{f}(0)=0$,
which again corresponds to the zero average condition $\int_{\mathbb{R}^2} f = 0$.
\smallskip
(iii)~In the case when $f \in L^1(\mathbb{R}^2)$ has compact support something more can be said.
In this case, Paley-Wiener theorem implies that the Fourier transform $\hat{f}$ is an analytic
function. In particular, there is a constant $C>0$ for which $| \hat{f}(\xi) - \hat{f}(0)| \leq C |\xi|$
for any $|\xi|\leq 1$. If $f$ additionally has zero average, we deduce that the singularity
at $\xi=0$ in \eqref{e:defH} is integrable for every $s > -2$.
\smallskip
(iv)~Let $f \in L^2(\mathbb{R}^2)$ have compact support and zero average. Since $\hat{f} \in L^\infty(\mathbb{R}^2)$,
using item~(iii) above we find that $f \in \dot{H}^s(\mathbb{R}^2)$ for any $-2<s<-1$. Combining
this fact with the assumption $f \in L^2(\mathbb{R}^2)$, the interpolation property
\eqref{e:interpH} below gives
that $f \in \dot{H}^s(\mathbb{R}^2)$ for any $-2<s \leq 0$.
\end{remark}
\subsection{Homogeneous Sobolev spaces $\boldsymbol{\dot{W}^{s,p}}$}
In the particular case $s \geq 0$ we extend the definition in \S\ref{ss:defH} to an arbitrary
summability exponent $1<p<\infty$.%
\footnoteb{Only the spaces $\dot{W}^{s,p}$ with $s\geq 0$ will be needed for our
scopes (see~\cite{grafakos,triebel} for a discussion of homogeneous spaces
with regularity index $s\in \mathbb{R}$).}
We say that a distribution $f \in \mathcal{D}'(\mathbb{T}^2)$ on the torus belongs
to the homogeneous Sobolev space $\dot{W}^{s,p} (\mathbb{T}^2)$ if
\begin{equation}\label{e:defWT}
\sum_{k \in \mathbb{Z}^2} |k|^s \hat{f}(k) e^{ikx} \in L^p(\mathbb{T}^2) \,,
\end{equation}
and we let $\| f \|_{\dot{W}^{s,p}(\mathbb{T}^2)}$ be the $L^p(\mathbb{T}^2)$ norm of the function
in~\eqref{e:defWT}. In general, elements of $\dot{W}^{s,p}(\mathbb{T}^2)$ are
equivalence classes modulo constants.
If $f \in \SS'(\mathbb{R}^2)$ is a tempered distribution on the plane, we say that $f$ belongs
to the homogeneous Sobolev space $\dot{W}^{s,p}(\mathbb{R}^2)$ if $\hat{f} \in L^1_{\rm loc}(\mathbb{R}^2)$
and
\begin{equation}\label{e:defW}
\mathcal{F}^{-1} \big( | \xi |^s \hat{f}(\xi) \big) \in L^p(\mathbb{R}^2) \,,
\end{equation}
and we let $\| f \|_{\dot{W}^{s,p}(\mathbb{R}^2)}$ be the $L^p(\mathbb{R}^2)$ norm of the function
in~\eqref{e:defW}. The condition that $\hat{f} \in L^1_{\rm loc}(\mathbb{R}^2)$
guarantees that this quantity is a norm. In general, the spaces
$\dot{W}^{s,p}(\mathbb{R}^2)$ are spaces of equivalence classes of distributions
modulo polynomials.
In our work, homogeneous spaces will be used only to measure the ``size'' of
given functions and velocity fields, which will be typically be regular.
In particular, the issue of completeness (which holds only for specific values
of $s$ given our definition) and seminorms will not arise.
\begin{remark}
(i)~The non-homogeneous Sobolev spaces $H^s$ and $W^{s,p}$ are defined by replacing
in~\eqref{e:defHT}, \eqref{e:defH}, \eqref{e:defWT}, and~\eqref{e:defW} the symbols
$|k|$ and $|\xi|$ by the symbols $\langle k \rangle :=
\sqrt{1+|k|^2}$ and $\langle \xi \rangle := \sqrt{1+|\xi|^2}$ respectively.
The norms are equivalent on the torus for functions with zero average.
\smallskip
(ii)~It is possible to prove that, if $s=m \in \mathbb{N}$, then $\dot{W}^{m,p}(\mathbb{R}^2) \cap L^p(\mathbb{R}^2)$ coincides
with the usual Sobolev space $W^{m,p}(\mathbb{R}^2)$ consisting of those $L^p(\mathbb{R}^2)$ functions
possessing weak derivatives of order less or equal than $m$ in $L^p(\mathbb{R}^2)$. This however
does not hold in the borderline cases $p=1$ and $p=\infty$.
\smallskip
(iii)~In contrast to the case of negative spaces (see
Remark~\ref{r:fctsinH}(i)-(ii)), it is not necessary for functions to have zero
average to belong to $\dot{W}^{s,p}$ with $s \geq 0$. Since we will measure
the regularity of the velocity fields in terms of $\dot{W}^{s,p}$ norms with $s
\geq 0$, we will not need to prescribe conditions about the average of the velocity
fields. In addition, none of the velocity fields in this work will be spatially
constant, and hence we can assume that \eqref{e:defWT} defines a norm.
Moreover, in the case of the plane $\mathbb{R}^2$, we will deal with velocity fields that
are bounded and compactly supported, therefore the local integrability of the
Fourier transform will be guaranteed.
\end{remark}
\subsection{Lipschitz-H\"older spaces}
\label{ss:lhs}
For notational convenience, in this paper we denote by $\dot{W}^{s,\infty}(\mathbb{T}^2)$,
resp.~$\dot{W}^{s,\infty}(\mathbb{R}^2)$, the homogeneous Lipschitz-H\"older space
on $\mathbb{T}^2$, resp.~on $\mathbb{R}^2$, defined as follows. Since the definition of these
spaces on $\mathbb{T}^2$ and on $\mathbb{R}^2$ is the same, below we do not write explicitly
the domain.
If $s = k \in \mathbb{N}$ is greater of equal to $1$,
we say that $f \in \dot{W}^{k,\infty}$ if
$f \in C^{k-1}$ and there is a constant $C>0$
such that
\begin{equation}\label{e:LIP}
| f^{(k-1)}(x) - f^{(k-1)}(y)| \leq C |x-y|
\qquad \text{for every $x$ and $y$.}
\end{equation}
We let $\| f \|_{\dot{W}^{k,\infty}}$ be
the minimal constant $C$ for which \eqref{e:LIP} holds.
If $s$ is not an integer, we let $\lfloor s \rfloor$ be the largest integer smaller than $s$.
In this case, we say that $f \in \dot{W}^{s,\infty}$ if
$f \in C^{\lfloor s \rfloor}$ and there is a constant $C>0$
such that
\begin{equation}\label{e:HOLD}
| f^{(\lfloor s \rfloor)}(x) - f^{(\lfloor s \rfloor)}(y)| \leq C |x-y|^{s - \lfloor s \rfloor}
\qquad \text{for every $x$ and $y$.}
\end{equation}
We let $\| f \|_{\dot{W}^{s,\infty}}$ be
the minimal constant $C$ for which \eqref{e:HOLD} holds.
\begin{remark}
\label{r:embedding}
In two space dimensions, the Sobolev space $\dot{W}^{s,p}$ embeds in the
Lipschitz space $\dot{W}^{1,\infty}$ if $s>1$ and $p > \frac{2}{s-1}$.
\end{remark}
\subsection{Scaling properties}
We will be frequently interested in the behavior of homogeneous Sobolev norms
under rescaling of a given function $f$. Given $\lambda>0$, we set
\begin{equation}\label{e:frisc}
f_\lambda(x) := f \left( \frac{x}{\lambda} \right) \,.
\end{equation}
If $f$ is defined on the torus and $1/\lambda$ is an integer, then the function
$f_\lambda$ in \eqref{e:frisc} is well defined on the torus and it holds
\begin{equation}\label{e:scaleT}
\| f_\lambda \|_{\dot{W}^{s,p}(\mathbb{T}^2)}
=
\lambda^{-s} \, \| f \|_{\dot{W}^{s,p}(\mathbb{T}^2)} \,.
\end{equation}
If $f$ is defined on the plane, then the function $f_\lambda$ in \eqref{e:frisc}
is well defined for any $\lambda>0$ and it holds
\begin{equation}\label{e:scaleR}
\| f_\lambda \|_{\dot{W}^{s,p}(\mathbb{R}^2)}
=
\lambda^{\frac{2}{p}-s} \, \| f \|_{\dot{W}^{s,p}(\mathbb{R}^2)} \,.
\end{equation}
When $p=2$, both formulas~\eqref{e:scaleT} and~\eqref{e:scaleR}
hold for $s\in\mathbb{R}$ (with norms in $\dot{H}^s$).
\begin{remark}
The difference between the exponents in formulas~\eqref{e:scaleT}
and~\eqref{e:scaleR}
is due to the fact that the rescalings on the torus and on the plane are two
different
operations. The rescaling of a single bump on the plane remains a single bump,
while on the torus $1 / \lambda^2$ rescaled copies are produced.
\end{remark}
\subsection{Interpolation}
We will frequently rely on the following standard interpolation inequality.
If $s_1 < s < s_2$ and $s = \vartheta s_1 + (1-\vartheta) s_2$, then
\begin{equation}\label{e:interpH}
\| f \|_{\dot{H}^s} \leq \| f \|_{\dot{H}^{s_1}}^\vartheta \| f
\|_{\dot{H}^{s_2}}^{1-\vartheta}
\end{equation}
and
\begin{equation}\label{e:interpW}
\| f \|_{\dot{W}^{s,p}} \leq \| f \|_{\dot{W}^{s_1,p}}^\vartheta \| f
\|_{\dot{W}^{s_2,p}}^{1-\vartheta} \,,
\end{equation}
where both inequalities hold both with domain $\mathbb{T}^2$ and $\mathbb{R}^2$.
The same holds in the case $p=\infty$ in the context of Lipschitz-H\"older
spaces
(recall~\S\ref{ss:lhs}) and can be proved with a simple direct argument.
We next introduce the two notions of mixing scale that will be employed in this
paper to quantify
the level of mixedness of the solution $\rho$. Both definitions can be given on
the torus and
on the plane.
The first notion is based on homogeneous Sobolev norms. From a physical point of
view,
the choice of the order $-1$ is the most convenient, since the norm in
$\dot{H}^{-1}$ scales
as a length. However, from a mathematical point of view, every negative
homogeneous
Sobolev norm would behave in a similar way.
\subsection{Definition.~Functional mixing scale~\cite{doering,MMP}}
\label{def:Functscale}
The functional mixing scale of $\rho(t,\cdot)$ is $\| \rho(t,\cdot)
\|_{\dot{H}^{-1} (\mathbb{T}^2)}$.
The second notion is more geometric and has been originally introduced
in~\cite{bressan} for solutions with value $\pm 1$. We use the following
generalization introduced in~\cite{zlatos}, which has the advantage of making
sense for every bounded solution $\rho$, without any constraint on its values.
\subsection{Definition.~Geometric mixing scale~\cite{bressan}}
\label{d:geomix}
Given $0 < \kappa < 1$, the geometric mixing scale of $\rho(t,\cdot)$ is the
infimum $\varepsilon(t)$ of all $\varepsilon>0$ such that for every $x \in
{\mathbb T}^2$ there holds
\begin{equation}\label{e:geom}
\frac{1}{\| \rho(t,\cdot) \|_\infty}
\left| \mint_{B_\varepsilon(x)} \rho(t,y) \, dy \right| \leq \kappa \,.
\end{equation}
The parameter $\kappa$ is fixed and plays a minor role in the definition.
Informally,
in order for $\rho(t,\cdot)$ to have geometric mixing scale $\varepsilon(t)$, we
require
that the average of the solution on every ball of radius $\varepsilon(t)$ is
essentially zero.
In a certain sense, we require the property of having (approximately) zero
average to be
localizable to balls of radius $\varepsilon(t)$.
\begin{remark} In~\cite{bressan} the following definition was given for
solutions with
value $\pm 1$. Given $0<\tilde \kappa<1/2$, \eqref{e:geom} is replaced by the
requirement that
\begin{equation}\label{e:geomgeom}
\tilde \kappa \leq \frac{|\{ \rho(t,\cdot) = 1 \} \cap B_\varepsilon(x) |}{|
B_\varepsilon(x)|}
\leq 1 - \tilde \kappa \,.
\end{equation}
Informally, according to~\cite{bressan}, in order for $\rho(t,\cdot)$ to have
geometric
mixing scale~$\varepsilon(t)$, it is required that every ball of radius
$\varepsilon(t)$
contains a ``substantial portion'' of both level sets $\{ \rho(t,\cdot) = 1 \}$
and
$\{ \rho(t,\cdot) = -1 \}$. It is easily seen that the two definitions
correspond if
$\kappa = 1 - 2 \tilde \kappa$.
\end{remark}
As previously mentioned, the two notions of mixing scale are not equivalent,
though they are strongly related (we refer to \cite{llnmd} for a further
discussion on this point; see also Lemma~\ref{l:fg}).
\section{Scaling analysis in a self-similar construction}
\label{selfsimilar}
A conceivable procedure for mixing consists of a self-similar evolution.
Such a procedure, together with the related scaling analysis, has been presented
in~\cite{ACM}.
Let $s \geq 0$ and $1 \leq p \leq \infty$ be fixed and assume the following:
\subsection{Assumption.~Self-similar base element}\label{ass}
There exist a velocity field $u_0$ and a (not identically zero) solution
$\rho_0$ to \eqref{e:continuity}, both defined for $0 \leq t \leq 1$ and $x \in
\mathbb{T}^2$, such that:
\begin{itemize}
\item[(i)] $u_0$ is bounded in $\dot{W}^{s,p} ({\mathbb T}^2)$ uniformly in
time, bounded, and divergence-free;
\item[(ii)] $\rho_0$ is bounded and has zero average;
\item[(iii)] there exists a positive constant $\lambda$, with $1/\lambda$ an
integer greater or equal than $2$, such that
\[
\rho_0(1, x) = \rho_0 \left( 0, \frac{x}{\lambda} \right) \,.
\]
\end{itemize}
An explicit example of a $u_0$ and a $\rho_0$ satisfying this assumption will be
given
in Section~\ref{pinching} for $s=1$ and every $1\leq p<\infty$.
\medskip
For later use, we give the following definitions:
\begin{definition}\label{d:tiling}
(i)~Given $\lambda > 0$, with $1/\lambda$ an integer, we denote
by $\mathscr{T}_\lambda$ the tiling of $\mathbb{T}^2$ with squares of
sidelength $\lambda$, consisting of $1/\lambda^2$ open squares
in $\mathbb{T}^2$ of the form
$$
\left\{ (x,y) \in {\mathbb T}^2 \; : \; k \lambda < x < (k+1)\lambda
\; \text{ and } \; h\lambda < y < (h+1)\lambda \right\}\,,
$$
for $k,h = 0,\ldots, 1 / \lambda-1$.
\smallskip
(ii)~Denoting by $\mathscr{Q}$ the unit
open square $(-1/2,1/2)^2 \subset \mathbb{R}^2$
the tiling $\mathscr{T}_\lambda$ of $\mathscr{Q}$
is defined analogously. Given any square $Q \in \mathscr{T}_{\lambda^n}$, we denote by
$r_Q$
its center, so that $Q = \lambda^n \mathscr{Q} + r_Q$.
\end{definition}
\subsection{A self-similar construction}\label{ss:ssc}
Fix a positive parameter $\tau$ (that will be determined later).
For each integer $n=1,2,\ldots$ and for $t \in [0,\tau^n]$ we set
\[
u_n(t,x) := \frac{\lambda^n}{\tau^n} \; u_0 \left( \frac{t}{\tau^n} ,
\frac{x}{\lambda^n} \right) \,,
\quad \rho_n(t,x) := \rho_0 \left( \frac{t}{\tau^n} , \frac{x}{\lambda^n}
\right) \,.
\]
Notice that $\rho_n$ is a solution of~\eqref{e:continuity} corresponding
to the velocity field $u_n$. Moreover, because of Assumption~\ref{ass}(iii),
\begin{equation}\label{e:glue}
\rho_n (\tau^n , x) = \rho_{n+1} (0,x) \,.
\end{equation}
Let us now define $u$ and $\rho$ by patching together the velocity fields
$u_0, u_1, \ldots$ and the corresponding solutions
$\rho_0, \rho_1, \ldots$ In detail, we set
\[
u(t,x) := u_n (t - T_n , x) \,, \quad \rho(t,x) := \rho_n (t - T_n , x)
\]
for $T_n \leq t < T_{n+1}$, and $n=1,2,\ldots$, where
\[
T_n := \sum_{i=0}^{n-1} \tau^i \,, \quad \textrm{ for $n=1, 2, \ldots, \infty$.}
\]
Notice that $u$ and $\rho$ are defined for $0 \leq t < T_\infty$. Moreover,
recalling equation~\eqref{e:glue}, $\rho$ is a weak solution%
\footnoteb{\label{fn:glue} This can be easily seen by writing the weak
formulation of the continuity equation~\eqref{e:continuity} on each time
interval~$[T_n,T_{n+1}]$ with test functions in~$C^\infty([T_n,T_{n+1}])$.}
of the Cauchy
problem for~\eqref{e:continuity} with velocity field $u$ and
initial condition $\bar{\rho}(x) = \rho_0(0,x)$.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig3a}
\caption{Example of self-similar evolution for a function
$\rho$ taking only two values.}
\label{f:self}
\end{center}
\end{figure}
Using~\eqref{e:scaleT} we can compute
\[
\left\| u_n (t,\cdot) \right\|_{\dot{W}^{s,p}(\mathbb{T}^2)}
= \left( \frac{\lambda^{1-s}}{\tau} \right)^n \left\| u_0 \left(
\frac{t}{\tau^n} , \cdot \right) \right\|_{\dot{W}^{s,p}(\mathbb{T}^2)} \,.
\]
If we set $\tau = \lambda^{1-s}$, the formula above gives that $u$
is bounded in $\dot{W}^{s,p}({\mathbb T}^2)$ uniformly in time.
Moreover,
\[
\| \rho_n (t,\cdot) \|_{\dot{H}^{-1}(\mathbb{T}^2)} = \lambda^n \left\| \rho_0
\left( \frac{t}{\tau^n} , \cdot \right) \right\|_{\dot{H}^{-1}(\mathbb{T}^2)}
\leq M \lambda^n \,,
\]
where we have set
\[
M := \sup_{0 \leq t \leq 1} \| \rho_0 (t,\cdot) \|_{\dot{H}^{-1}(\mathbb{T}^2)}
\,.
\]
This can be rewritten as
\begin{equation}\label{e:decay}
\| \rho(t,\cdot) \|_{\dot{H}^{-1}(\mathbb{T}^2)} \leq M \lambda^n, \qquad
\textrm{ for $T_n \leq t < T_{n+1}$.}
\end{equation}
We can now consider the following three different scenarios:
\begin{itemize}
\item[(1)] $s<1$, then $\tau = \lambda^{1-s} < 1$. In this case, $T_\infty$ is
finite and
$$
\| \rho(t,\cdot) \|_{\dot{H}^{-1}(\mathbb{T}^2)} \to 0,
$$
as $t \to T_\infty$. That is, we have perfect mixing in finite time;
\item[(2)] $s=1$, then $\tau = 1$. In this case, $T_\infty = \infty$ and $T_n =
n$, therefore the inequality $t < T_{n+1}$ in~\eqref{e:decay} becomes $t-1<n$.
Hence, the estimate in~\eqref{e:decay} yields the following exponential decay of
the functional mixing scale:
\[
\| \rho(t,\cdot) \|_{\dot{H}^{-1}(\mathbb{T}^2)} \leq M \lambda^{t-1} \,;
\]
\item[(3)] $s>1$, then $\tau>1$. In this case, $T_\infty = \infty$ and
\[
T_n = \frac{\tau^n-1}{\tau - 1} = \frac{\lambda^{(1-s)n}-1}{\lambda^{1-s} - 1}.
\]
By the same argument as above, \eqref{e:decay} implies the following polynomial
decay:
\[
\| \rho(t,\cdot) \|_{\dot{H}^{-1}(\mathbb{T}^2)} \leq M \frac{\left[ 1 + t
(\lambda^{1-s} - 1) \right]^{-\frac{1}{s-1}}}{\lambda}
\simeq C(M,\lambda,s) \, t^{-\frac{1}{s-1}} \,.
\]
\end{itemize}
What we have just discussed can be summarized in the following theorem:
\begin{theorem}\label{t:main}
Given $s\geq 0$ and $1 \leq p \leq \infty$, under Assumption~\ref{ass}, there
exist a bounded divergence-free velocity field $u$ and a weak solution $\rho$ of
the Cauchy problem for~\eqref{e:continuity}, such that $u$ is bounded in
$\dot{W}^{s,p}(\mathbb{T}^2)$ uniformly in time and the functional mixing scale
of $\rho$ exhibits the following behavior depending on $s$:
\begin{itemize}
\item[(1)] Case $s<1$: perfect mixing in finite time;
\item[(2)] Case $s=1$: exponential decay;
\item[(3)] Case $s>1$: polynomial decay.
\end{itemize}
\end{theorem}
In fact, all homogeneous negative Sobolev norms $\|\rho(t)\|_{\dot{H}^{-r}}$
(with $r>0$) would exhibit the same behavior, with the only difference being in the constant
for the exponential decay and the exponent for the polynomial decay, which
depend on~$r$.
We observe that, in the case $s>1$, such self-similar scaling analysis does
not yield the known exponential lower bound for the (geometric and functional)
mixing scale, which is supposed to be optimal; rather, it yields a
slower polynomial decay in time.
\medskip
The following lemma proves that the geometric mixing scale $\varepsilon(t)$
exhibits the same behavior, detailed in Theorem~\ref{t:main}, as the functional
mixing scale.
\begin{lemma}\label{l:fg}
Let $\rho$ be a bounded function
such that
\begin{equation}\label{e:zerotil}
\mint_{Q} \rho \, dy = 0
\end{equation}
for every square $Q \in \mathscr{T}_{\lambda^n}$. Then the geometric mixing scale of $\rho$ as in \eqref{e:geom} is less or equal than
\begin{equation}\label{e:peri}
\frac{4 \sqrt{2} \, \lambda^n}{\kappa}\,.
\end{equation}
\end{lemma}
\begin{proof}
We fix an arbitrary ball $B_\varepsilon(x)$. Using \eqref{e:zerotil} we can
estimate
$$
\begin{aligned}
\frac{1}{\| \rho\|_\infty} \left| \int_{B_\varepsilon(x)} \rho \, dy \right|
& \leq \frac{1}{\| \rho\|_\infty}
\sum_{\substack{Q \in \mathscr{T}_{\lambda^n} \\ \text{ $Q \cap
\partial B_\varepsilon(x) \ne \emptyset$}}}
\int_Q | \rho| \, dy \\
& \leq
\frac{1}{\| \rho\|_\infty}
\int_{B_{\varepsilon+\sqrt{2}\lambda^n}(x)\setminus B_{\varepsilon-\sqrt{2}\lambda^n}(x)} |\rho|\, dy \\
&\leq
\pi \big[ (\varepsilon+\sqrt{2}\lambda^n)^2 - (\varepsilon-\sqrt{2}\lambda^n)^2 \big]
=
4 \sqrt{2} \pi \varepsilon \lambda^n\,,
\end{aligned}
$$
where the second inequality follows from elementary geometric considerations.
Hence,
$$
\frac{1}{\| \rho\|_\infty} \left| \mint_{B_\varepsilon} \rho \, dx \right|
\leq
\frac{4 \sqrt{2} \,\lambda^n }{\varepsilon} \,,
$$
and the right-hand side is less or equal than $\kappa$
for every $\varepsilon$ greater or equal than
the quantity in \eqref{e:peri}. Therefore, from \eqref{e:geom} the
desired estimate on the geometric mixing scale follows.
\end{proof}
\subsection{Regularity in time}\label{ss:timereg}
Under Assumption~\ref{ass}, the self-similar construction described above
ensures Sobolev regularity of the velocity field with respect to the
space variable, uniformly in time. No regularity with respect to the time variable is provided.
In fact, in all examples presented in this paper, the velocity field
is piecewise smooth with respect to both space and time variables.
If the velocity field is smooth in time on two adjacent time intervals,
and if it can be smoothly extended to the closure of each of them,
then the discontinuity across the interface of the two intervals can be eliminated.
Indeed, it is enough to replace in each interval $u$ and $\rho$ by
$$
\tilde{u}(t,x) := \eta'(t) \, u \big( \eta(t),x \big)\,,
\qquad
\tilde{\rho}(t,x) := \rho \big( \eta(t),x \big)\,,
$$
where in each interval the smooth function $\eta$ is chosen to be increasing,
surjective, and constant in a
small (left or right) neighborhood of each of endpoints of the interval.
It is immediate to check that $\tilde \rho$ solves the Cauchy problem for~\eqref{e:continuity}
with velocity field $\tilde u$, that $\tilde u$ is smooth on the union of the
closures of the two time intervals, and that the value of the solution at the endpoints of both
intervals has not changed.
We remark that the argument above does not apply in case the velocity field
lacks a smooth extension to the closure of (at least) one time interval. In such
a case the time discontinuity cannot be eliminated. Our example in
Section~\ref{pinching} indeed contains singularities, which is unavoidable,
since topological properties of the transported sets are changed along the time
evolution.
\section{First geometric construction}
\label{basic}
In the main result of this section (Proposition~\ref{s:basic}) we show that
given a regular set $E$ in the plane which evolves smoothly in
time, we can construct a smooth divergence-free velocity field $u$
such that the characteristic function of $E$ solves the continuity
equation \eqref{e:continuity} associated to $u$.
In the next section we use this result
to produce the building blocks for the examples of optimal mixers given in this
work.
\medskip
We begin by introducing some notation.
Given a vector $v=(v_1,v_2)\in\mathbb{R}^2$, we denote by
$v^\perp$ its rotation by $90^\circ$ counter-clockwise,
that is,
\[
v^\perp:=(-v_2,v_1)
\, .
\]
Given a set $E$ in $\mathbb{R}^2$ and a point $x\in\mathbb{R}^2$,
we denote by $\mathrm{dist}(x,E)$ the distance of $x$
from $E$, defined in the standard way by
\[
\mathrm{dist}(x,E) := \inf\big\{|x-y| \, \mid \, y \in E\big\}.
\]
If there exists exactly one point $y\in E$
where such infimum is attained, this point will be
called the \emph{projection} of $x$ onto $E$ and
denoted by $p_E(x)$.
For every $r>0$, we shall also denote
by $B(E,r)$ the open $r$-neighbourhood of $E$,
that is,
\[
B(E,r) := \big\{ x\in\mathbb{R}^2 \, : \ \mathrm{dist}(x,E)<r \big\}
\, .
\]
\subsection{Paths and curves}
\label{s:paths}
In this paper we consider only two kinds of paths.
A \emph{closed path} is a continuous map
$\gamma=\gamma(s)$ from the the circle,
which we identify with the one-dimensional
torus $\mathbb{T}^1 := \mathbb{R}/\mathbb{Z}$,
to the plane $\mathbb{R}^2$. Moreover we require
that $\gamma$ is injective, of class $C^1$,
and satisfies $\dot\gamma(s)\ne 0$ for all $s\in\mathbb{T}^1$.
A \emph{closed (oriented) curve} is the image $\Gamma=\gamma(\mathbb{T}^1)$
of a closed path $\gamma$.
A \emph{proper path} is a continuous map
$\gamma=\gamma(s)$ from the the real line $\mathbb{R}$
to the plane $\mathbb{R}^2$ which is proper, that is,
$|\gamma(s)|$ tends to $+\infty$ as $s\to\pm\infty$.
As before, we require that $\gamma$ is injective,
of class $C^1$, and satisfies $\dot\gamma(s)\ne 0$
for all $s\in\mathbb{R}$.
A \emph{proper (oriented) curve} is the image $\Gamma=\gamma(\mathbb{R})$
of a proper path $\gamma$.
When we do not want to distinguish between closed
and proper paths (or curves), we simply write path (or curve),
and denote the parametrization domain, which is either
$\mathbb{T}^1$ or $\mathbb{R}$, by the letter $J$.
As usual, the regularity of a curve $\Gamma$
refers to the regularity of the parametrization $\gamma$.
Let $\Gamma$ be a curve parametrized by $\gamma$.
A \emph{sub-arc} of $\Gamma$ is any set of the form
$\gamma(J')$ where $J'$ is an interval contained in $J$;
a sub-arc is \emph{proper} if it is strictly
contained in $\Gamma$.
The unit tangent vector $\tau(x)$ and
the unit normal vector $\eta(x)$
at a point $x=\gamma(s)$ in $\Gamma$ are given
by
\[
\tau:= \dot\gamma / |\dot\gamma|
\, , \quad
\eta:= -\tau^\perp = - \dot\gamma^\perp / |\dot\gamma|
\, .
\]
In particular if $|\dot\gamma|(s)$ is equal to
the constant $\ell$ for all $s$ then
$\tau:= \dot\gamma / \ell$,
$\eta:= \dot\gamma^\perp / \ell$, and
the curvature $\kappa(x)$ of $\Gamma$ at
the point $x=\gamma(s)$ satisfies the equation
\[
\kappa\, \eta = \ddot\gamma/\ell^2
\, .
\]
The \emph{tubular radius} of $\Gamma$
is the largest $r\ge 0$
such that the map $\Psi$ given by\,%
\footnoteb{With a slight abuse of notation, we sometimes
write the geometric quantities $\tau$, $\eta$ and $\kappa$
as functions of the parametrization variable $s$ instead of $x$.}
\begin{equation}
\label{e:param}
\Psi: (s,y) \mapsto \gamma(s) + y \, \eta(s)
\end{equation}
is injective on $J\times(-r,r)$.
If $\Gamma$ is of class $C^2$ then the tubular radius
is smaller than the \emph{curvature radius}
$1/\kappa(x)$ for every $x\in\Gamma$.
If $\Gamma$ is of class $C^k$ with
$k\ge 2$ and the tubular radius $r$ is strictly
positive, the map $\Psi$ is a diffeomorphism
of class $C^{k-1}$ from $J\times(-r,r)$
to the tubular neighbourhood $B(\Gamma,r)$;
the projection $p_\Gamma(x)$ is well-defined
for every point $x=\Psi(s,y)$ in $B(\Gamma,r)$,
and agrees with $\gamma(s)$.
If $\Gamma$ is closed and of class $C^2$ then the
tubular radius is strictly positive.
\subsection{Time-dependent paths and curves}
\label{s:tpaths}
In this paper we often consider paths
and curves which vary in time; in this case,
$\gamma$ is a map from the
rectangle $I\times J$
to the plane $\mathbb{R}^2$, where $I$ is a time interval
(open, closed, or else),
and $\Gamma$ is a map that to every $t\in I$
associate a curve $\Gamma(t)$ in $\mathbb{R}^2$; when we speak
of regularity for such objects we refer to the regularity
of the parametrization map $\gamma$ in both variables.
In the following
we reserve the letter $t$ for the time
variable in $I$ and the letter $s$ for
the parametrization variable in $J$,
we write $\dot\gamma$ for the partial derivative
w.r.t.\ ~$s$, and $\partial_t\gamma$ for the partial derivative
w.r.t.\ ~$t$.
The \emph{normal velocity} $v_\mathrm{n}=v_\mathrm{n}(t,x)$ of $\Gamma$
at time $t$ and point $x=\gamma(t,s)$ is the
normal component of the vector $\partial_t\gamma(t,s)$,
that is,
\[
v_\mathrm{n} := \partial_t\gamma \cdot \eta
\, .
\]
Note that the normal velocity does not change under
reparametrization of $\gamma$ in the variable $s$.
\subsection{Time-dependent domains}
\label{s:geomevol}
A time-dependent domain
is a map $E$
that to every $t$ in the time interval $I$
associates an open subset $E(t)$ of $\mathbb{R}^2$.
We say that $E$ is of class $C^k$ if there exist
finitely many time-dependent curves $\Gamma_i$,
parametrized by paths
$\gamma_i:I\times J_i\to\mathbb{R}^2$ of class $C^k$,
such that for every $t\in I$ the boundary $\partial E(t)$
can be written as disjoint union of
the curves $\Gamma_i(t)$.
We also require
that each parametrization is \emph{counter-clockwise},
which means that the normal vector $\eta$
defined in \S\ref{s:paths} agrees with the \emph{inner normal}
to the boundary $\partial E$ at every time $t$ and at every
point $x=\gamma_i(t,s)$.
Thus the normal velocity $v_\mathrm{n}$ defined in \S\ref{s:tpaths}
agrees with the \emph{outer normal velocity} of $\partial E$.
\subsection{Compatible velocity fields}
\label{s:compatible}
Let $u$ be a continuous, time-dependent velocity field on $\mathbb{R}^2$.
We say that $u$ is \emph{compatible} with a time-dependent curve
$\Gamma$ if, for every time $t$ and every
point $x\in\Gamma(t)$, the normal
velocity $v_\mathrm{n}$ of $\Gamma$ agrees with
the normal component of $u$,
that is
\begin{equation}
\label{e:compat}
v_\mathrm{n} = u \cdot \eta
\, .
\end{equation}
Accordingly, we say that $u$
is \emph{compatible} with a time-dependent domain $E$ of
class $C^1$ if the normal component of $u$ agrees with
the outer normal velocity $v_\mathrm{n}$ of $E$ at every time $t$
and at every point $x\in\partial E(t)$.
Fix $t_0\in I$, and let $\{\Phi(t,\cdot): \, t\in I\}$ be the flow
(if it exists) associated to $u$ with initial time $t_0$, which means that
each $\Phi(t,\cdot)$ is an homeomorphism from $\mathbb{R}^2$ into $\mathbb{R}^2$,
and that for every $x_0\in\mathbb{R}^2$ the map $t\mapsto \Phi(t,x_0)$
solves the ordinary differential equation $\dot x=u(t,x)$
with initial condition $x(t_0)=x_0$.
Then the compatibility of $u$ and $\Gamma$ means that
$\Gamma(t)=\Phi(t,\Gamma(t_0))$ for every $t\in I$;
in particular the compatibility of $u$ and $E$ means that
$\partial E(t)=\Phi(t,\partial E(t_0))$,
which yields
\[
E(t)=\Phi(t,E(t_0))
\quad\text{for every $t\in I$.}
\]
Finally, it is well-known that the last identity implies
that the characteristic function $\rho(t,x):= 1_{E(t)}(x)$
is a distributional solution of the transport equation
\begin{equation}
\label{e:transport}
\rho_t+u\cdot\nabla\rho=0
\, .
\end{equation}
\begin{remark}
When $u$ is divergence-free
equation \eqref{e:transport} agrees with
the continuity equation~\eqref{e:continuity} and therefore
the velocity field $u$ is compatible with the time-dependent
domain $E$ if and only if the characteristic function
$\rho(t,x):= 1_{E(t)}(x)$ is a distributional solution
of~\eqref{e:continuity}.
Note that the latter condition makes sense
also if $u$ and $E$ are not regular
(we just need that $u$ is locally in the class $L^1$,
and that $E$ is a measurable subset of $I\times\mathbb{R}^2$)
and therefore can be taken as a weak notion of compatibility.
\end{remark}
In the rest of this section we address the following question:
given a time-dependent curve $\Gamma$, or a time-dependent domain $E$,
under which assumptions there exists a compatible velocity field
$u$ which is divergence-free?
We begin with a basic statement,
which we then specialize
according to our specific needs.
\begin{proposition}
\label{s:basic}
Let $\Gamma$ be a time-dependent curve in $\mathbb{R}^2$
with time interval~$I$ and of class $C^k$ with $k\ge 2$,
and let $\bar r:I\to(0,+\infty)$ be a continuous function.
Assume that for every $t\in I$ the normal velocity
$v_\mathrm{n}(t,\cdot)$ has compact support,%
\footnoteb{This requirement is clearly redundant
when $\Gamma$ is closed.}
and satisfies
\begin{equation}
\label{e:constvol}
\int_{\Gamma(t)} v_\mathrm{n}(t,x) \, d\sigma(x) = 0
\, .
\end{equation}
Then there exists a divergence-free
velocity field $u:I\times\mathbb{R}^2\to\mathbb{R}^2$ of class
$C^{k-2}$ which is compatible with $\Gamma$,
and such that the support of $u(t,\cdot)$
is contained in $B(\Gamma(t),\bar r(t))$
for every $t\in I$.
Moreover, under the additional assumption that for
every $t\in I$ the support of $v_\mathrm{n}(t,\cdot)$ is
contained in a compact, proper sub-arc $G(t)$ of $\Gamma(t)$
which depends continuously in $t$,%
\footnoteb{The class of compact sub-arcs of $\Gamma$ is endowed with
the Hausdorff distance.}
we can choose $u$ so
that the support of $u(t,\cdot)$ is contained in $B(G(t),\bar r(t))$
for every $t\in I$.
\end{proposition}
\begin{remark}
\label{s:rembasic}
(i)~If $\Gamma$ is closed, assumption \eqref{e:constvol}
is necessary, in the sense that it is satisfied by every
time-dependent closed path $\Gamma$ which is compatible
with a divergence-free velocity field $u$.
Fix indeed $t\in I$, let $E(t)$
be the bounded open set with boundary $\Gamma(t)$,
and denote by $\eta_{E(t)}$ the outer normal to $\partial E(t)$;
then the divergence theorem yields
\[
\int_{\Gamma(t)} v_\mathrm{n} \, d\sigma
=\pm\int_{\partial E(t)} u\cdot \eta_{E(t)} \, d\sigma
=\pm\int_{E(t)} \mathop{\mathrm{div}} u \, dx
=0
\, ,
\]
where the sign $\pm$ depends on whether the normal
to $\Gamma(t)$ agrees with $\eta_{E(t)}$ or $-\eta_{E(t)}$.
\smallskip
(ii)~A modification of the previous argument shows that assumption
\eqref{e:constvol} is necessary if $\Gamma$ is proper,
and both $v_\mathrm{n}$ and $u$ have compact support.
However, if we do not require that $u$ has compact support,
then we can drop both the assumption that
$v_\mathrm{n}$ has compact support and \eqref{e:constvol}.
\smallskip
(iii)~If $\Gamma$ is a closed curve and agrees with the boundary
of a bounded, time-dependent domain $E$, then it is well-known that
\[
\int_{\Gamma(t)} v_\mathrm{n} \, d\sigma
= \frac{d}{dt} |E(t)|
\quad\text{for every $t\in I$.}
\]
Therefore assumption \eqref{e:constvol} is equivalent
to require that the area of $E(t)$ is constant in $t$.
\smallskip
(iv)~Proposition~\ref{s:basic} can be generalized
to higher dimensions, for instance to time-dependent surfaces with
codimension one in $\mathbb{R}^n$, but such extensions require quite
different proofs.
\end{remark}
We consider now the special case of a
curve which evolves homothetically in time.
We first give a definition and a few remarks.
\subsection{Homothetic curves}
\label{s:homcurve}
We say that a time-dependent curve
$\Gamma$ on the time interval $I$ is homothetic
in time if it can be represented as
\begin{equation}
\label{e:omot}
\Gamma(t)
= \lambda(t) \, \bar\Gamma
= \big\{ \lambda(t) \, x \, : \ x\in \bar\Gamma \big\}
\end{equation}
for some curve $\bar\Gamma$ and
some function $\lambda:I\to(0,+\infty)$.
Let $\bar\gamma:J\to\mathbb{R}^2$ be a path that parametrizes $\bar\Gamma$.
Then the time-dependent path $\gamma:I\times J\to\mathbb{R}^2$ given by
\begin{equation}
\label{e:omot1}
\gamma(t,s):=\lambda(t) \, \bar\gamma(s)
\end{equation}
is a parametrization of $\Gamma$;
thus $\Gamma$ is of class $C^k$
when $\bar\Gamma$ and $\lambda$ are of class~$C^k$.
Let $\bar\eta$ be the normal to $\bar\Gamma$ and let
$\bar v:\bar\Gamma\to\mathbb{R}$ be the function defined by
\begin{equation}
\label{e:omot1.1}
\bar v(x):= x\cdot \bar \eta(x)
\, .
\end{equation}
Then a simple computation starting from \eqref{e:omot1}
yields that the normal vector and the normal velocity
of $\Gamma$ (at $t\in I$ and $x\in\Gamma(t)$) are given by
\begin{equation}
\label{e:omot2}
\eta(t,x)
= \bar\eta \big( x/\lambda(t) \big)
\, , \quad
v_\mathrm{n}(t,x)
= \lambda'(t) \, \bar v \big( x/\lambda(t) \big)
\, .
\end{equation}
Finally, let $\bar u$ be an autonomous velocity field on $\mathbb{R}^2$
such that
\begin{equation}
\label{e:compat2}
\bar u(x) \cdot \bar \eta(x) = \bar v(x)
\end{equation}
for every $x\in \bar \Gamma$.
Then using \eqref{e:omot2} one readily checks that
the time-dependent velocity field $u:I\times\mathbb{R}^2\to\mathbb{R}^2$ defined by
\begin{equation}
\label{e:omot3}
u(t,x):=\lambda'(t) \, \bar u \big( x/\lambda(t) \big)
\end{equation}
is compatible with the time-dependent curve $\Gamma$.
The next result specializes the statement of
Proposition~\ref{s:basic}
to the case of homothetic curves.
\begin{proposition}
\label{s:basicomot}
Let be given a function $\lambda: I\to(0,+\infty)$ and
a proper curve $\bar \Gamma$, both of class $C^k$ with $k\ge 2$,
a positive number $\bar r$ and a compact sub-arc $\bar G$ of $\bar\Gamma$
which contains the support of the function $\bar v$
defined in \eqref{e:omot1.1}.
Assume moreover that
\begin{equation}
\label{e:constvol2}
\int_{\bar \Gamma} \bar v \, d\sigma = 0
\, .
\end{equation}
Then the following statements hold:
\begin{itemizeb}
\item[(i)]
there exists an autonomous velocity field
$\bar u$ on $\mathbb{R}^2$ of class $C^{k-2}$ which
satisfies \eqref{e:compat2}, is divergence-free,
and its support is contained in $B(\bar G,\bar r)$;
\item[(ii)]
if $\Gamma$ is the time-dependent (homothetic) curve
defined in \eqref{e:omot} and $u$ is the time-dependent
velocity field defined in \eqref{e:omot3},
then $u$ is of class $C^{k-2}$, divergence-free and compatible
with $\Gamma$, and the support of $u(t,\cdot)$
is contained in $B(\lambda(t)\, \bar G, \lambda(t)\, \bar r)$
for every $t\in I$.
\end{itemizeb}
\end{proposition}
\begin{remark}
\label{s:rembasic3}
(i)~The formula for the normal velocity in \eqref{e:omot2}
shows that assumption~\eqref{e:constvol2} in
Proposition~\ref{s:basicomot} plays the role
of assumption \eqref{e:constvol} in Proposition~\ref{s:basic}.
\smallskip
(ii)~Proposition~\ref{s:basicomot} does not apply to closed curves,
because condition \eqref{e:constvol2} is not verified if $\bar \Gamma$
is closed. Let indeed $E$ be a bounded open set with boundary $\bar \Gamma$;
then the divergence theorem yields
$$
\int_{\bar\Gamma} \bar v \, d\sigma
= \int_{\partial E} x \cdot \bar \eta (x) \, d\sigma(x)
= \pm\int_E \mathop{\mathrm{div}}(x) \, dx = \pm 2|E| \ne 0
\, ,
$$
where the sign $\pm$ depends on whether $\bar \eta$
is the inner or the outer normal of $E$.
\smallskip
(iii)~It is easy to check that
if the curve $\bar \Gamma$ agrees outside some open ball
$B=B(0,r)$ with two half-lines $L_-,L_+$ starting from the origin,
then the function $\bar v$ has compact support
(and the converse is also true).
If, in addition, $\bar \Gamma$ is the boundary of an open set $E$
and we denote by $T$ the open set delimited by the half-lines
$L_-,L_+$ which agrees with $E$ outside $B$ (see Figure~\ref{f:commarea}),
then a modification of the computation in~(ii) gives
\[
\int_{\bar \Gamma} \bar v \, d\sigma
= \pm \big( |E\cap B|-|T\cap B| \big)
\, .
\]
In particular assumption \eqref{e:constvol2}
is equivalent to saying that $E\cap B$ and $T\cap B$
have the same area, which means that
$(E\setminus T) \cap B$ and $(T\setminus E) \cap B$
have the same area (see Figure~\ref{f:commarea}).
\end{remark}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig4a}
\caption{The situation described in Remark~\ref{s:rembasic3}(iii).}
\label{f:commarea}
\end{center}
\end{figure}
The rest of this section is devoted to the proofs
of Propositions~\ref{s:basic} and \ref{s:basicomot}.
The key step is contained in Lemma~\ref{s:basiclemma}.
\subsection{Potential of a velocity field}
\label{s:potential}
Let $u:\mathbb{R}^2\to\mathbb{R}^2$ be a continuous velocity field
and let $\varphi:\mathbb{R}^2\to\mathbb{R}$ be a function of class $C^1$.
We say that $\varphi$ is a \emph{potential} for $u$ if
\[
u=\nabla^\perp \varphi
\, ,
\]
where $\nabla^\perp:=(-\partial_2,\partial_1)$.
Note that $u$ admits a potential if and only if it is divergence-free.
In the fluid dynamics literature, such $\varphi$ is called a stream
function for the flow generated by $u$.
\begin{lemma}
\label{s:basiclemma}
Let $\Gamma$ be a curve and $v$ be a function
on the curve $\Gamma$, both of class $C^k$ with $k\ge 2$. Given
a positive number $\bar r$ and a compact (not necessarily proper)
sub-arc $G$ of $\Gamma$,
which contains the support of $v$, assume that
\begin{equation}
\label{e:constvol3}
\int_\Gamma v \, d\sigma = 0
\, .
\end{equation}
Then there exists a divergence-free, autonomous velocity field
$u$ on $\mathbb{R}^2$ of class $C^{k-2}$, the normal component
of which on $\Gamma$, that is $u\cdot \eta$, agrees with $v$,
and the support of which is contained in $B(G,\bar r)$.
\end{lemma}
\begin{proof}
We describe the proof in the case $J=\mathbb{R}$, where we recall that $J$ is the
domain of the parametrization of the curve $\Gamma$; the case $J=\mathbb{T}^1$
requires few straightforward modifications.
In view of \S\ref{s:potential}, it suffices to find a potential
$\varphi:\mathbb{R}^2\to\mathbb{R}$ of class $C^{k-1}$ and with support contained
in $B(G,\bar r)$ such that
\begin{equation}
\label{e:compat4}
\partial_\tau\varphi = v
\quad\text{on $\Gamma$,}
\end{equation}
where $\tau$ is the tangent vector to $\Gamma$,
and then take $u:=\nabla^\perp\varphi$.
Let $\gamma:\mathbb{R}\to\mathbb{R}^2$ be a parametrization of $\Gamma$.
For the construction of $\varphi$
we choose
\begin{itemizeb}
\item
a point $x_0=\gamma(s_0)\in\Gamma$ and,
if $G$ is a proper sub-arc of~$\Gamma$, we
further require that $x_0$ does not belong to $G$;
\item
a smooth function $g:\mathbb{R}\to\mathbb{R}$ with support
contained in $[-1/2,1/2]$ such that $g(0)=1$;
\item
a number $0 < r < \bar r$ smaller
than the tubular radius of $\Gamma$.
\end{itemizeb}
\smallskip
Now we take the diffeomorphism $\Psi:\mathbb{R}\times(-r,r) \to B(\Gamma,r)$
defined in \eqref{e:param}, and for every $x=\Psi(s,y)\in B(\Gamma,r)$
we set
\begin{equation}
\label{e:defpot}
\varphi(x)
=\varphi(\Psi(s,y))
:= g(y/r) \, \int_{s_0}^s v(\gamma(s')) \, |\dot\gamma(s')| \, ds'
\, .
\end{equation}
\smallskip
If $x$ belongs to $\Gamma$ then $x=\gamma(s)=\Psi(s,0)$,
and therefore $\varphi(x)$ is the integral of $v$ along the (oriented)
sub-arc of $\Gamma$ starting from $x_0$ and ending at $x$.
Thus the restriction of $\varphi$ to $\Gamma$ is a primitive of $v$,
and therefore it satisfies \eqref{e:compat4}.
Now, formula \eqref{e:defpot} shows that $\varphi\circ\Psi$
is a function of class $C^k$ on $\mathbb{R}\times(-r,r)$
with support contained in $\mathbb{R}\times[-r/2,r/2]$,
and since $\Psi$ is a diffeomorphism of class $C^{k-1}$ and maps
$\mathbb{R}\times[-r/2,r/2]$ into the closure of $B(\Gamma,r/2)$, we deduce
that $\varphi$ is a function of class $C^{k-1}$ on $B(\Gamma,r)$
with support contained in the closure of $B(\Gamma,r/2)$.
\smallskip
We complete the construction
by extending $\varphi$ to $0$ in the rest of~$\mathbb{R}^2$.
\smallskip
It remains to check that the support of $\varphi$
is contained in $B(G,\bar r)$. When $G=\Gamma$,
this follows by the fact that the support of $\varphi$
is contained in the closure of $B(\Gamma,r/2)$, which
is contained in $B(\Gamma, \bar r)$. When $G=\gamma([s_1,s_2])$
is a proper sub-arc of $\Gamma$ we have that
\begin{itemizeb}
\item
$v(\gamma(s))=0$ for $s\notin [s_1,s_2]$ by assumption;
\item
$s_0\notin [s_1,s_2]$ by the choice of $x_0$;
\item
condition \eqref{e:constvol3} can be re-written as
$\int_{s_1}^{s_2} v(\gamma(s')) \, |\dot\gamma(s')|\, ds'=0$.
\end{itemizeb}
Putting together these facts and recalling the choice of $g$
one easily shows that $\varphi(\Psi(s,y))=0$
if $s\notin[s_1,s_2]$ or $y\notin[-r/2,r/2]$,
and then
\[
\mathrm{supp}(\varphi)
\subset \Psi \big( [s_1,s_2]\times[-r/2,r/2] \big)
\subset B(G,\bar r)
\, .
\qedhere
\]
\end{proof}
\begin{proof}[Proof of Proposition~\ref{s:basicomot}]
Statement~(i) follows from Lemma~\ref{s:basiclemma},
while statement~(ii) is an immediate consequence of
statement~(i) and \S\ref{s:homcurve}.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{s:basic}]
For every $t\in I$ we use Lemma~\ref{s:basiclemma}
to construct a divergence-free velocity field $u(t,\cdot)$
of class $C^{k-2}$ which satisfies the compatibility condition
\eqref{e:compat} at time $t$, and whose support is contained
in $B(G(t),\bar r(t))$.
However, this only gives that $u$ is of class $C^{k-2}$
in the variable $x$.
To show that $u$ can be taken of class $C^{k-2}$ in $t$ and $x$,
we must go through the proof of Lemma~\ref{s:basiclemma}.
In doing so we quickly realize that the key point is the
regularity of class $C^{k-1}$ in the variables $t,s,y$
of the right-hand side of formula \eqref{e:defpot},
that in our specific case is
\[
g( y/r(t) ) \, \int_{s_0(t)}^s
v_\mathrm{n}(t,\gamma(t,s')) \, |\dot\gamma(t,s')| \, ds'
\, .
\]
It is clear that this quantity has the required regularity
provided that we choose $r(t)$ and $s_0(t)$ at least of class $C^{k-1}$
in $t$.
Since both $\bar r(t)$ and the tubular radius of $\Gamma(t)$
are continuous, strictly positive functions of $t$,
it is always possible to choose $r(t)$ smaller than both,
strictly positive, and smooth in $t$.
If we only require that the support of $u$ is contained
in $B(\Gamma(t),\bar r(t))$, we can take $s_0(t)$ constant in $t$.
If we require that the support of $u$ is contained
in $B(G(t),\bar r(t))$, then we can again choose $s_0(t)$ smooth in $t$,
but this choice is more delicate, and relies on the fact that $G(t)$ is
a proper sub-arc for all $t\in I$.
\end{proof}
\section{First example: pinching}
\label{pinching}
In this section we verify Assumption~\ref{ass} for $s=1$ and for every $1 \leq p<\infty$.
In fact, we do this by constructing a velocity field and a solution
on $\mathbb{R}^2$, both compactly supported in the open unit square $\mathscr{Q}$,
for every $s$ and $p$ such that
$p$ is strictly below the critical exponent for the embedding of
$W^{s,p}$ in the Lipschitz class, that is,
\begin{equation}
\label{e:range}
\text{
$s<1$ and $p\le \infty$,
or $s\ge 1$ and $\displaystyle p < \frac{2}{s-1}$.
}
\end{equation}
More precisely, we give an example of a velocity field $u_0$
and a solution $\rho_0$ of the continuity equation \eqref{e:continuity},
both defined for $0 \le t \le 1$, such that
\begin{itemizeb}
\item[(a)]
$u_0$ is a time-dependent velocity field on $\mathbb{R}^2$,
compactly supported in the open unit square $\mathscr{Q}$, smooth
in both variables for $t\ne k/8$ with $k=1,\dots,7$,
bounded, divergence-free, and bounded in $\dot{W}^{s,p}(\mathbb{R}^2)$
uniformly in $t$ for all $s, p$ as above;
\item[(b)]
$\rho_0$ is of the form $\rho_0(t,\cdot) = 1_{E(t)} - \pi/16$
where $E(t)$ is a time-dependent domain in $\mathbb{R}^2$
defined for $0\le t \le 1$,
continuous for all $t$,%
\footnoteb{Or more precisely, $t\mapsto 1_{E(t)}$ is
continuous as a map with values, say, in $L^1(\mathbb{R}^2)$.}
smooth for $t\ne k/8$ with $k=1,\dots,7$,
with closure contained in $\mathscr{Q}$,
and area equal to $\pi/16$
(thus $\rho_0(t,\cdot)$ has average zero).
\item[(c)]
$E(0)$ is the disk with center $0$ and radius $1/4$
while $E(1)$ is the union of the four disks with
centers $(\pm 1/4,\pm 1/4)$ and radius $1/8$.
\end{itemizeb}
Since $u_0$ and $\rho_0$ have compact support
in the open square $\mathscr{Q}$ we can see them as a velocity field and
a solution on the torus~$\mathbb{T}^2$. For $s=1$, the locality
of the distributional derivative of order $1$ ensures that
$u_0$ is bounded in $\dot{W}^{1,p}(\mathbb{T}^2)$ for every $1 \leq p<\infty$.
Therefore Assumption~\ref{ass} is satisfied for $s=1$ and for every~$1 \leq p<\infty$.
\medskip
As a matter of fact, we construct a time-dependent domain
$E$ as above and a velocity field $u_0$ defined for
$t\ne k/8$ with $k=1,\dots,7$,
which is smooth and compatible with $E$.
This ensures that the characteristic function
$1_{E(t)}$ is a distributional solution of the
continuity equation \eqref{e:continuity}
in the open time intervals $((k-1)/8,k/8)$
with $k=1,\dots,8$; the fact that it is also a
solution on the time interval $[0,1]$ is ensured
by the continuity in $t$ (recall Footnote~\ref{fn:glue}).
The set $E(t)$ for $t=k/8$ with $k=0,\dots,4$ and $t=1$
is described in Figure~\ref{f:evol1}.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig5a}
\caption{The set $E(t)$ for $t=k/8$ with $k=0,\dots,4$ and $t=1$.}
\label{f:evol1}
\end{center}
\end{figure}
More precisely, we denote by $B$ the open disk with center
$0$ and radius $1/4$, by $T$ the cone of all $x\in\mathbb{R}^2$ such that
$|x_2| < |x_1|$, and for $t=1/8$, $t=1/4$, and $t=3/8$ we require that
\begin{itemizeb}
\item[(d)]
$E(t)$ is symmetric w.r.t.\ both axes;
\item[(e)]
the set $E(t)$ has area $\pi/16$;
\item[(f)]
the set $E(t)\setminus B$ is the same, that is, is left invariant by the flow;
\item[(g)]
the sets $(E(1/8)\setminus T) \cap B$ and $(T\setminus E(1/8))\cap B$ have the same area.
\end{itemizeb}
In the rest of this section we describe the construction
of $E(t)$ and $u_0(t,\cdot)$ for $t$
in the time intervals $[0,1/8]$ (Step~1 in Figure~\ref{f:evol1})
and $(1/8,1/4)$ (Step~2 in Figure~\ref{f:evol1}).
The construction in the remaining time intervals (steps)
is similar, and is omitted.
\bigskip
\emph{Step~1: construction of $E(t)$ and $u_0(t,\cdot)$ for $0 \le t \le 1/8$.}
Since $E(0)$ and $E(1/8)$ are both smooth and have area $\pi/16$, we can clearly
find $E(t)$ for $0<t<1/8$ with area $\pi/16$ so that $t\mapsto E(t)$
is smooth on $[0,1/8]$. Then by Proposition~\ref{s:basic} we can find
a smooth velocity field $u_0:[0,1/8]\times\mathbb{R}^2\to\mathbb{R}^2$
which is divergence-free and compatible with $E$.
Moreover, since $\partial E(t)$ is contained in $\mathscr{Q}$, we can assume that
the support of $u_0$ is contained in $\mathscr{Q}$ for all $t$.
In particular, all positive Sobolev norms of $u_0(t,\cdot)$ are uniformly bounded
in $t$.
\bigskip
\emph{Step~2: construction of $E(t)$ and $u_0(t,\cdot)$ for $1/8 < t < 1/4$.}
Let $\ove{\Gamma}$ be the proper curve drawn in Figure~\ref{f:evol3}.
More precisely, $\ove{\Gamma}$ is defined outside $B$ by the equation
$x_2=|x_1|$, and agrees in $B$ with the ``upper'' connected component
of $\partial E(1/8)\cap B$.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig5b}
\caption{The curve $\ove{\Gamma}$, the homothetic copies $\lambda\,\ove{\Gamma}$
with $\lambda=2/3$, $\lambda=1/3$, and their limit as $\lambda\to 0$.}
\label{f:evol3}
\end{center}
\end{figure}
We consider a decreasing smooth function
$\lambda$ on $[1/8,1/4)$ such that $\lambda(1/8)=1$
and $\lambda(t)$ tends to $0$ as $t\to 1/4$;
the function $\lambda$ will be explicitly chosen later
in order to satisfy further requirements.
Then we take $E(t)$ for all
$1/8<t<1/4$ satisfying the following requirements:
\begin{itemizeb}
\item[(h)]
$E(t)$ agrees with $E(1/8)$ outside $B$;
\item[(i)]
$\partial E(t) \cap B$ has two connected component which are symmetric
with respect to the $x_1$ axis, and the upper one agrees with
$\lambda(t) \, \ove{\Gamma}$ in $B$; thus $E(t)\cap B$ is described in
Figure~\ref{f:evol3} when $\lambda(t)=1$, $\lambda(t)=2/3$,
and $\lambda(t)=1/3$.
\end{itemizeb}
By Remark~\ref{s:rembasic3}(iii), assumption~(g) above implies that
the curve $\ove{\Gamma}$ satisfies condition \eqref{e:constvol2},
and therefore we can use Proposition~\ref{s:basicomot}(ii)
to obtain a smooth, divergence-free velocity field
$u:[1/8,1/4)\times\mathbb{R}^2\to\mathbb{R}^2$ which is compatible with the
homothetic curve $\lambda(t)\,\ove{\Gamma}$.
Moreover $u(t,\cdot)$ is compactly supported in
the upper half-plane $\mathbb{R}\times(0,+\infty)$ for all $t$
(more precisely, we can require that the support is contained
in the dashed rectangle $R$ in Figure~\ref{f:evol3}).
Finally we take $u_0: [1/8,1/4)\times\mathbb{R}^2\to\mathbb{R}^2$
equal to $u$ in the upper half-plane, and we extend it
to the lower half-plane by reflection.
Thus $u_0$ is still smooth and compactly supported,
and by assumption~(i) above is compatible with the time-dependent
domain $E$ inside the ball $B$.
On the other hand, $u_0$ vanishes outside the ball $B$
and therefore it is compatible with the set $E\setminus B$,
which is constant in time (recall assumption~(g)).
In conclusion, $u_0$ is compatible with $E$.
It remains to choose $\lambda$ so that $u_0(t,\cdot)$ is
bounded in $\dot{W}^{s,p}(\mathbb{R}^2)$ uniformly in
$t\in [1/8,1/4)$ for every $s,p$ as in \eqref{e:range}.
To this end, we recall that by Proposition~\ref{s:basicomot}(ii)
the field $u_0$ can be written in the form
\[
u_0(t,x) = \lambda'(t) \, \bar u \big( x/\lambda(t) \big)
\]
where $\bar u:\mathbb{R}^2\to\mathbb{R}^2$ is smooth and compactly supported.
Therefore, using~\eqref{e:scaleT}, for every $t$ we obtain
\begin{align*}
\| u_0(t,\cdot) \|_{\dot{W}^{s,p}(\mathbb{R}^2)}
= |\lambda'(t)| \, |\lambda(t)|^{2/p-s} \| \bar u \|_{\dot{W}^{s,p}(\mathbb{R}^2)}
\, .
\end{align*}
Now, a simple a computation shows that $u_0(t,\cdot)$ is
uniformly bounded in $\dot{W}^{s,p}(\mathbb{R}^2)$ if we take
\[
\lambda(t) := \exp \Big( 2- \frac{1}{1-4t} \Big)
\, .
\]
With this choice of the function $\lambda(t)$ the velocity
field $u_0(t,\cdot)$ is also bounded.
\begin{remark}
\label{r:messrlf}
(i)~The flow of the (non-Lipschitz) velocity field $u_0$ changes the topology
of sets: the ball at time $t=0$ is transformed into two balls at time $t=1/2$.
\smallskip
(ii)~Using symmetry considerations, it is possible to check that the flow of $u_0$ shrinks
from time $t=1/8$ to time $t=1/4$
a vertical segment to the origin ( which is represented by the intersection
of the two lines
in Figure~\ref{f:evol3}).
\smallskip
(iii)~Analogously, from time $t=1/4$ to time $t=3/8$ the flow of $u_0$ ``spreads''
the origin over an horizontal segment. In particular, the ODE
with velocity field $u_0$ and initial time $t=1/8$
does not have (pointwise) uniqueness for all initial data laying
on the vertical segment described in~(ii) above.
\end{remark}
\section{Scaling analysis in a quasi self-similar construction}
\label{quasiselfsimilar}
A velocity field behaving like the one constructed in
Section~\ref{pinching} cannot have Lipschitz regularity, since
the associated evolution
does not preserve the connectedness of sets.
Indeed, it is not evident that Assumption~\ref{ass} can be satisfied
with $s=1$ and $p=\infty$.
In this section we address this case by replacing the
(exactly) self-similar procedure in Section~\ref{selfsimilar} by a
{\em quasi self-similar} procedure.
That is, instead of replicating rescaled copies of
{\em one} basic element at each step of the evolution
(as in \S\ref{ss:ssc}), we consider a {\em finite family}
of basic elements, which are rescaled and rearranged at each step
of the evolution according to a combinatorial pattern.
We implement this construction on the plane $\mathbb{R}^2$.
\medskip
Given $s > 0$ and $1 \leq p \leq \infty$, Assumption~\ref{ass} is
replaced by the assumption below, where we denote
by $\lceil s \rceil$ the smallest integer greater or equal than $s$. We also
recall that, for $p=\infty$, the space $\dot{W}^{s,\infty}$ is a Lipschitz-H\"older space,
discussed in ~\S\ref{ss:lhs}.
\subsection{Assumption.~Almost self-similar basic family}\label{ass2}
For $j = 1,\ldots,N$ there exist velocity fields $u^j$ and corresponding (not identically zero) solutions $\rho^j$ to~\eqref{e:continuity}, all defined for $0 \leq t \leq 1$ and $x \in \mathbb{R}^2$, such that:
\begin{itemize}
\item[(i)] each $u^j$ is bounded in $\dot{W}^{\lceil s \rceil,p}(\mathbb{R}^2)$ uniformly in time,
bounded, divergence-free, and tangent to $\partial \mathscr{Q}$;%
\footnoteb{Notice that the normal trace of $u^j$ on $\partial \mathscr{Q}$
(both from inside and from outside) is well defined in distributional sense.}
\item[(ii)] each $\rho^j$ is bounded and has zero average on $\mathscr{Q}$;
\item[(iii)] there exists a positive constant $\lambda$, with $1/\lambda$ an
integer greater or equal than~$2$, such that each function $\rho^j(1,\cdot)$
agrees on each square of the tiling~$\mathscr{T}_\lambda$ (recall Definition~\ref{d:tiling})
with a rescaled translation of one of the functions $\{\rho^i(0,\cdot)\}_{1\leq i\leq N}$,
in the sense that for each $Q \in \mathscr{T}_\lambda$
$$
\rho^j (1,x) = \rho^{\gamma(j,Q)} \left( 0 , \frac{x-r_Q}{\lambda} \right)
\qquad \text{for $x \in Q$,}
$$
for suitable $\gamma = \gamma(j,Q) \in \{1,\ldots,N\}$.
\end{itemize}
For later use, for any $\sigma \geq 0$ we set
\begin{equation}\label{e:Msigma}
M_\sigma := \max_{j=1,\ldots,N} \sup_{0 \leq t \leq 1} \| \rho^j(t,\cdot) \, 1_\mathscr{Q} \|_{\dot{H}^{-\sigma}(\mathbb{R}^2)} \,.
\end{equation}
Since each $\rho^j$ is bounded and has zero average on $\mathscr{Q}$, it follows from
Remark~\ref{r:fctsinH}(iv) that $M_\sigma$ is finite for $0 \leq \sigma < 2$.
\subsection{Quasi self-similar construction}\label{ss:ass}
Under Assumption~\ref{ass2} we now define inductively a quasi self-similar evolution.
\emph{Initial step.} We start by choosing
a positive constant $\bar \lambda$, with $1 / \bar\lambda$
an integer greater or equal than~$1$.%
\footnoteb{Here we allow $\bar \lambda=1$.}
We define the evolution for
$0 \leq t \leq 1$ by patching together velocity fields and
solutions on the tiling $\mathscr{T}_{\bar \lambda}$ of $\mathscr{Q}$.
For every $Q \in \mathscr{T}_{\bar \lambda}$ we choose an index $\bar \jmath (Q) \in \{1,\ldots,N\}$
and we set for $x \in Q$ and $0 \leq t \leq 1$
\begin{equation}\label{e:gluefieldbase}
u(t,x) := \bar \lambda \, u^{\bar \jmath (Q)} \left( t , \frac{x-r_Q}{\bar \lambda} \right) \,,
\quad
\rho(t,x) := \rho^{\bar \jmath (Q)} \left( t , \frac{x-r_Q}{\bar \lambda} \right) \,.
\end{equation}
For $x \not \in \mathscr{Q}$ we set both $u$ and $\rho$ equal to zero.
We stress that, in this construction (as well as in the one in the
iterative step below), the resulting field is divergence-free, but it
does not necessarily have Sobolev regularity as the derivative may jump at the boundary
of the patch. In what follows, we will {\em assume} the needed regularity
(see Assumption \ref{ass3} below). We will prove later that this assumption is satisfied in the example
we give in Section \ref{snake}.
Since the velocity field $u$ in \eqref{e:gluefieldbase}
is tangent to all the interfaces of the tiling~$\mathscr{T}_{\bar \lambda}$ it follows that
$\rho$ in \eqref{e:gluefieldbase} solves the continuity equation
with velocity field $u$ globally in $\mathbb{R}^2$.
We notice that, by Assumption~\ref{ass2}(iii), on each square of the
tiling $\mathscr{T}_{\bar \lambda \lambda}$ the function $\rho(1,\cdot)$
equals a rescaled translation of one of the functions
$\{\rho^i(0,\cdot)\}_{1 \leq i \leq N}$.
\emph{Iterative step.}
For a given positive parameter $\tau$ to be chosen later, we define as in \S\ref{ss:ssc}
\[
T_n := \sum_{i=0}^{n-1} \tau^i \,, \quad \textrm{ for $n=1, 2, \ldots, \infty$.}
\]
We now assume that $u$ and $\rho$ have been defined for $0 \leq t \leq T_n$,
in such a way that on each square of the tiling $\mathscr{T}_{\bar \lambda \lambda^n}$
the function $\rho(T_n,\cdot)$ equals a rescaled translation
of one of the functions $\{\rho^i(0,\cdot)\}_{1 \leq i \leq N}$,
and we show how to define $u$
and $\rho$ for $T_n < t \leq T_{n+1}$.
Consider a square $Q \in \mathscr{T}_{\bar \lambda \lambda^n}$. By the inductive assumption there exists an index $j = j(n,Q)$ such that
\begin{equation}\label{e:choose}
\rho(T_n,x) = \rho^j \left(0 , \frac{x-r_Q}{\bar \lambda \lambda^n} \right) \qquad \text{ for $x \in Q$.}
\end{equation}
Accordingly we define for $x \in Q$ and $T_n< t \leq T_{n+1}$
\begin{equation}\label{e:gluefield}
u(t,x) := \frac{\bar \lambda \lambda^n}{\tau^n} u^j \left( \frac{t-T_n}{\tau^n} , \frac{x-r_Q}{\bar \lambda \lambda^n} \right) \,,
\quad
\rho(t,x) := \rho^j \left( \frac{t-T_n}{\tau^n} , \frac{x-r_Q}{\bar \lambda \lambda^n} \right) \,.
\end{equation}
As before, for $x \not \in \mathscr{Q}$ we set both $u$ and $\rho$ equal to zero.
By the same argument as in the initial step, we have that
for $T_n< t \leq T_{n+1}$ the function
$\rho$ in \eqref{e:gluefield} solves the continuity equation
with velocity field $u$ globally in $\mathbb{R}^2$.
By Assumption~\ref{ass2}(iii), on each square of the tiling
$\mathscr{T}_{\bar \lambda \lambda^{n+1}}$ the function $\rho(T_{n+1},\cdot)$
equals a rescaled translation of one of the functions
$\{ \rho^i(0,\cdot)\}_{1 \leq i \leq N}$. This concludes the inductive
procedure, which gives a velocity field $u$ and a solution
$\rho$ of \eqref{e:continuity} defined for a.e.~$x \in \mathbb{R}^2$
and for all $t \geq 0$.
We now make a further assumption on the velocity fields $u$ constructed in
the quasi-self similar manner. One drawback of our construction, in fact, is
that we do not control (at least a priori) the behavior of derivatives of the field at
the boundary of each patch. We are therefore forced at this stage to make a further
assumption on $u$, regarding its regularity. This assumption will be proved
to be satisfied in the example we give in Section \ref{snake}.
\subsection{Assumption.~Regularity of the patching}\label{ass3}
For a.e.~$t \geq 0$ we have that $u(t,\cdot) \in \dot{W}^{\lceil s \rceil ,p}(\mathbb{R}^2)$.
\begin{remark} (i) The fact that Assumptions~\ref{ass2} and \ref{ass3} entail regularity of order $\lceil s \rceil$ (rather than $s$) is technical and due to the fact that the norm on Sobolev spaces with integer order is local, property that we will exploit in the proof of Lemma~\ref{l:fieldqss}.
\smallskip
(ii) Assumption~\ref{ass3} is, in fact, the key structural condition to ensure a quasi
self-similar construction yields fields of the required regularity, as already observed above. It is indeed easy to construct families of velocity fields and solutions that satisfy
Assumption~\ref{ass2} without verifying Assumption~\ref{ass3}. We will construct
a family that satisfies both assumptions, for any $s$ and
$p$ in Section~\ref{snake}.
\smallskip
(iii) In the relevant case $s=1$ and $p=\infty$, i.e., in the Lipschitz case,
Assumption~\ref{ass3} can be rephrased as requiring that $u$ have a continuous
representative on $\mathbb{R}^2$ (and actually it is sufficient to check the continuity
of $u$ on the boundaries of the sub-squares in the tilings).
\end{remark}
We stress that we do not require any uniformity of
the Sobolev norm with respect to time in Assumption~\ref{ass3}. Under
this assumption, the uniformity in time is in fact guaranteed by the
following lemma, provided we choose $\tau$ as in \S\ref{ss:ssc}.
\begin{lemma}\label{l:fieldqss} Let $\tau = \lambda^{1-s}$. Under Assumptions~\ref{ass2} and \ref{ass3}, the velocity field $u$ constructed by the quasi self-similar procedure in \S\ref{ss:ass} is divergence-free and belongs to $\dot{W}^{s,p}(\mathbb{R}^2)$ uniformly in time.
\end{lemma}
\begin{proof}
First of all, since each velocity field $u^j$ is divergence-free in $\mathscr{Q}$ and tangent to the boundary $\partial \mathscr{Q}$, it follows that $u$ is divergence-free. It remains to prove the bound in $\dot{W}^{s,p}(\mathbb{R}^2)$.
\emph{Step~1: the case $s=k$ an integer.}
Consider $T_n < t \leq T_{n+1}$. Then $u(t,\cdot)$ is defined as in \eqref{e:gluefield} for some function $j = j(n,Q)$. Assumption~\ref{ass3} guarantees that $u(t,\cdot) \in \dot{W}^{k,p}(\mathbb{R}^2)$, therefore (by the absolute continuity of the derivative of a Sobolev function of integer order) we only need to estimate the sum of the $\dot{W}^{k,p}$ norms of the restriction of $u(t,\cdot)$ to the squares $Q$ in $\mathscr{T}_{\bar \lambda \lambda^n}$. We can compute as follows:
$$
\begin{aligned}
\| u(t,\cdot) \|_{\dot{W}^{k,p}(\mathbb{R}^2)}^p
& =
\int_{\mathscr{Q}} |\nabla^k u(t,x)|^p \, dx \\
& =
\sum_{Q \in \mathscr{T}_{\bar \lambda \lambda^n}} \int_{Q} \left| \nabla^k\left( \frac{\bar \lambda \lambda^n}{\tau^n} \, u^j \left( \frac{t-T_n}{\tau^n} , \frac{x-r_Q}{\bar \lambda \lambda^n} \right) \right) \right|^p \, dx \\
& =
\sum_{Q \in \mathscr{T}_{\bar \lambda \lambda^n}} \int_{Q} \left| \frac{\bar \lambda \lambda^n}{\tau^n \bar\lambda^k \lambda^{kn}} \left( \nabla^k u^j \right) \left( \frac{t-T_n}{\tau^n} , \frac{x-r_Q}{\bar \lambda \lambda^n} \right) \right|^p \, dx \\
&=
\left( \frac{\lambda^{1-k}}{\tau} \right)^{pn} \bar \lambda^{p(1-k)} \sum_{Q \in \mathscr{T}_{\bar \lambda \lambda^n}} \int_\mathscr{Q} \left| \left( \nabla^k u^j \right) \left( \frac{t-T_n}{\tau^n} , y \right) \right|^p (\bar \lambda\lambda^{n})^2 \, dy \\
& \leq
\left( \frac{\lambda^{1-k}}{\tau} \right)^{pn} \bar \lambda^{p(1-k)} \max_{1 \leq j \leq N} \sup_{0 \leq r \leq 1} \| u^j(r,\cdot) \|_{\dot{W}^{k,p}(\mathbb{R}^2)}^p \,.
\end{aligned}
$$
The computation for $p=\infty$ is analogous, considering the supremum norm, and gives
$$
\| u(t,\cdot) \|_{\dot{W}^{k,\infty}(\mathbb{R}^2)}
\leq
\left( \frac{\lambda^{1-k}}{\tau} \right)^n \bar \lambda^{(1-k)} \max_{1 \leq j \leq N} \sup_{0 \leq r \leq 1} \| u^j(r,\cdot) \|_{\dot{W}^{k,\infty}(\mathbb{R}^2)} \,.
$$
The choice $\tau=\lambda^{1-s}$ concludes the proof for $s=k$ integer.
\emph{Step~2: the general case $s \geq 0$ a real.} We rely on the previous step and
we use~\eqref{e:interpW} with
$s_1 = \lceil s \rceil$, $s_2 = 0$ and $\vartheta = s / \lceil s \rceil$, getting
$$
\begin{aligned}
\| u(t,\cdot) \|_{\dot{W}^{s,p}(\mathbb{R}^2)}
& \leq
\| u(t,\cdot) \|^{1-\vartheta}_{L^p(\mathbb{R}^2)} \,
\| u(t,\cdot) \|^\vartheta_{\dot{W}^{\lceil s \rceil,p}(\mathbb{R}^2)} \\
& \leq \left( \frac{\lambda}{\tau} \right)^{(1-\vartheta) n} \bar \lambda^{1-\vartheta}
\left( \frac{\lambda^{1-\lceil s \rceil}}{\tau} \right)^{\vartheta n} \bar \lambda^{(1-\lceil s \rceil)\vartheta}\\
& \quad \quad \quad \quad \times
\max_{1 \leq j \leq N} \sup_{0 \leq r \leq 1} \left[ \| u^j(r,\cdot) \|_{L^p(\mathbb{R}^2)}
+ \| u^j(r,\cdot) \|_{\dot{W}^{\lceil s \rceil ,p}(\mathbb{R}^2)} \right]
\\
& = \left( \frac{\lambda^{1-s}}{\tau} \right)^n \bar \lambda^{1-s}
\max_{1 \leq j \leq N} \sup_{0 \leq r \leq 1} \left[ \| u^j(r,\cdot) \|_{L^p(\mathbb{R}^2)}
+ \| u^j(r,\cdot) \|_{\dot{W}^{\lceil s \rceil ,p}(\mathbb{R}^2)} \right] \,,
\end{aligned}
$$
where we have used the estimate in Step 1 with $k=0$ and $k = \lceil s \rceil$.
Again the choice $\tau = \lambda^{1-s}$ allows to conclude.
\end{proof}
\subsection{Decay of the functional mixing norm}
We now analyze the behavior in time of negative Sobolev norms of the solution $\rho$ constructed in~\S\ref{ss:ass}.
For $T_n \leq t < T_{n+1}$ we have
$$
\rho(t,x) = \sum_{Q \in \mathscr{T}_{\bar \lambda \lambda^n}}
\rho^j \left( \frac{t-T_n}{\tau^n} , \frac{x-r_Q}{\bar \lambda\lambda^n} \right) \, 1_Q(x)
\qquad \text{for $x \in \mathbb{R}^2$,}
$$
for a suitable $j = j(n,Q)$. For any $r>0$, equation~\eqref{e:scaleR} implies
that
\begin{equation}\label{e:r}
\begin{aligned}
\| \rho(t,\cdot) \|_{\dot{H}^{-r}(\mathbb{R}^2)}
& \leq
\sum_{Q \in \mathscr{T}_{\bar \lambda \lambda^n}} \left\| \rho^j \left( \frac{t-T_n}{\tau^n} , \frac{x-r_Q}{\bar \lambda \lambda^n} \right) \, 1_Q(x) \right\|_{\dot{H}^{-r}(\mathbb{R}^2)} \\
& =
\sum_{Q \in \mathscr{T}_{\bar \lambda \lambda^n}} \bar \lambda^{1+r} \lambda^{n(1+r)} \left\| \rho^j \left( \frac{t-T_n}{\tau^n} , \cdot \right) \, 1_\mathscr{Q} \right\|_{\dot{H}^{-r}(\mathbb{R}^2)} \\
& \leq \bar \lambda^{1+r} \frac{\lambda^{n(1+r)}}{\lambda^{2n}} M_r = \bar \lambda^{1+r} \lambda^{n(r-1)} M_r \,,
\qquad
\text{ for $T_n \leq t < T_{n+1}$,}
\end{aligned}
\end{equation}
with $M_r$ as in \eqref{e:Msigma}. Using~\eqref{e:interpH} we find that,
for every $\gamma \geq 1$,
\begin{equation}\label{e:intgamma}
\| \rho (t,\cdot) \|_{\dot{H}^{-\sigma}(\mathbb{R}^2)}
\leq \| \rho (t,\cdot) \|^{1/\gamma}_{\dot{H}^{-\gamma \sigma}(\mathbb{R}^2)}
\| \rho (t,\cdot) \|^{1-1/\gamma}_{L^2(\mathbb{R}^2)}\,.
\end{equation}
Now, given $\sigma$, choose $\gamma \geq 1$ such that $1 < \sigma\gamma < 2$. This is indeed possible as long as $0 < \sigma < 2$. Then \eqref{e:intgamma}, together with
\eqref{e:r} for $r = \gamma \sigma$ gives
\begin{equation}\label{e:cg}
\| \rho(t,\cdot) \|_{\dot{H}^{-\sigma}(\mathbb{R}^2)}
\leq \bar \lambda^{\sigma + 1/\gamma}
M_0^{1-1/\gamma} M_{\gamma \sigma}^{1/\gamma} \lambda^{n(\sigma - 1/\gamma)} \,,
\qquad
\text{ for $T_n \leq t < T_{n+1}$,}
\end{equation}
owing to the fact that the $L^2$ norm of the solution is conserved for all
$t \geq 0$:
$$
\| \rho(t,\cdot) \|_{L^2(\mathbb{R}^2)} = M_0.
$$
Finally, setting
$$
c_\sigma := \sigma-1/\gamma > 0 \,,
\qquad
C_\sigma := \bar \lambda^{\sigma + 1/\gamma} M_0^{1-1/\gamma} M_{\gamma \sigma}^{1/\gamma} > 0 \,,
$$
Equation \eqref{e:cg} becomes
\begin{equation}\label{e:decaysigma}
\| \rho(t,\cdot) \|_{\dot{H}^{-\sigma}(\mathbb{R}^2)}
\leq C_\sigma \lambda^{n c_\sigma} \,,
\qquad
\text{ for $T_n \leq t < T_{n+1}$.}
\end{equation}
In particular, for $\sigma=1$, choosing $\gamma=3/2$ yields:
\begin{equation}\label{e:decay2}
\| \rho(t,\cdot) \|_{\dot{H}^{-1}(\mathbb{R}^2)}
\leq \bar \lambda^{5/3}
M^{1/3}_0 M^{2/3}_{3/2} \; \lambda^{n/3} \,,
\qquad
\text{ for $T_n \leq t < T_{n+1}$.}
\end{equation}
The above estimates \eqref{e:decaysigma} and \eqref{e:decay2}
corresponds to \eqref{e:decay} in \S\ref{ss:ssc}.
Therefore, arguing as in the final step of the proof
of Theorem~\ref{t:main}, we obtain the following result.
\begin{theorem}\label{t:ass}
Given $s>0$ and $1\leq p\leq \infty$, under Assumptions~\ref{ass2} and
\ref{ass3}, there exist a bounded, divergence-free velocity field $u$ and a
solution $\rho$ of the Cauchy problem for~\eqref{e:continuity} in $\mathbb{R}^2$, such
that $u$ is bounded in $\dot{W}^{s,p}(\mathbb{R}^2)$ uniformly in time, $u$ and $\rho$
are supported in $\mathscr{Q}$ for all times, and the functional mixing scale of
$\rho$ exhibits the following behavior depending on $s$:
\begin{itemize}
\item[(1)] Case $s<1$: perfect mixing in finite time;
\item[(2)] Case $s=1$: exponential decay;
\item[(3)] Case $s>1$: polynomial decay.
\end{itemize}
\end{theorem}
\begin{remark}\label{r:qssfinal}
(i)~In fact, in the case $s=1$, every $\dot{H}^{-\sigma}$ norm of $\rho$ decays
exponentially in time for $0 < \sigma < 2$ (see \eqref{e:decaysigma}). Moreover,
if $M_{\tilde \sigma} < \infty$ for some $\tilde \sigma \geq 2$, where
$M_\sigma$ is given in \eqref{e:Msigma}, then every $\dot{H}^{-\sigma}$
norm of $\rho$ decays exponentially in time for $0 < \sigma < \tilde \sigma$.
\smallskip
(ii)~Thanks to Lemma~\ref{l:fg} we can deduce that the geometric mixing scale
$\varepsilon(t)$ of $\rho(t,\cdot)$ exhibits the same behavior of the
functional mixing scale detailed in Theorem~\ref{t:ass}.
\smallskip
(iii)~For later use (in the companion paper~\cite{loss}) we make here the
following observation. In the case $s=1$, Theorem~\ref{t:ass} provides a
velocity field $u$ bounded in $\dot{W}^{1,p}(\mathbb{R}^2)$ uniformly in time and a
solution $\rho$ of the continuity equation for which the functional mixing
scale decays exponentially in time. In addition, for any real number $r \geq 0$,
such a velocity field satisfies
$$
\| u(t,\cdot) \|_{\dot{W}^{r,p}(\mathbb{R}^2)}
\leq
C_r \left( \lambda^{1-r} \right)^t \,.
$$
This estimates follows from the proof of
Lemma~\ref{l:fieldqss}
(recalling that in this case $\tau=1$).
In particular, the Sobolev norms of $u$ of order higher than one grow
exponentially in time, while the Sobolev norms of order lower than one decay
exponentially in time.
\smallskip
(iv)~In the case when $s$ is a positive integer, the fact that the
velocity field and the solution are supported in $\mathscr{Q}$ and the locality of the distributional derivative
of positive integer order imply the validity of Theorem~\ref{t:ass} with domain
the torus $\mathbb{T}^2$.
\end{remark}
\subsection{A combinatorial construction}
\label{ss:keycomb}
An explicit example of an almost self-similar basic family with an associated
evolution that satisfies Assumptions~\ref{ass2} and \ref{ass3} will be constructed
in Section~\ref{snake}. Here we want to illustrate, somewhat informally, the
``combinatorial properties'' on which our construction is based. This
illustration provides
a graphical sketch of the procedure described in~\S\ref{ss:ass} on a specific
example and substantiates the feasibility of our strategy of proof.
In this example, there are two basic moves, the action of which is depicted in
Figure~\ref{f:idea}.
The support of the functions $\rho^1(0,\cdot)$ and $\rho^2(0,\cdot)$
is contained in the grey ``straight strip'' and ``bent strip'' in the figure,
respectively.
\medskip
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1.3]{fig6a}
\caption{The two basic moves in the example of an
almost self-similar construction. The solutions $\rho^1$ (left) and $\rho^2$ (right).}
\label{f:idea}
\end{center}
\end{figure}
The two basic moves map
$\rho^1(0,\cdot) \leadsto \rho^1(1,\cdot)$ (Figure~\ref{f:idea}, left) and
$\rho^2(0,\cdot) \leadsto \rho^2(1,\cdot)$ (Figure~\ref{f:idea}, right),
respectively. The output at time $t=1$ consists of
configurations that
(up to rotations) only include rescaled translations of functions with
support contained in straight strips and in bent strips.
This transition will be implemented with velocity fields that are supported in the same sets.
We remark that, in the actual construction, we take
$\lambda = 1/5$ and $N=6$, since
we have to consider the two rotations of the straight channel and the four rotations
of the bent channel.
\medskip
Figure~\ref{f:ideatwo} illustrates how we can patch these basic moves (following
the procedure in~\S\ref{ss:ass}) in order to obtain a velocity field and a solution
that are compactly supported in $\mathbb{R}^2$. In detail, we implement the initial step
in~\S\ref{ss:ass} with $\bar \lambda = 1/2$ by starting at time $t=0$ with four
bent channels that altogether create a ``circular channel'' (Figure~\ref{f:ideatwo}, left).
Each of the four bent channels then evolves according to the description
in Figure~\ref{f:idea}: two steps in the evolution are
illustrated in Figure~\ref{f:ideatwo}.
The resulting velocity field and solution are smooth and
compactly supported, since we are able to control
how the basic elements match at the interfaces between the four sub-squares.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig6b}
\caption{Patching of four basic moves and further steps in the almost self-similar evolution.}
\label{f:ideatwo}
\end{center}
\end{figure}
The evolution in Figures~\ref{f:idea} and \ref{f:ideatwo} depicts the
combinatorics
of the example discussed in Section~\ref{snake} accurately.
However, the rigorous construction requires a careful analysis and
implementation of these ideas, as we need to produce smooth solutions that
are transported
by smooth, divergence-free velocity fields.
\section{Second geometric construction}
\label{geo2}
In this section we describe another geometric construction of
divergence-free velocity fields $u$ together with
(non-trivial) solutions $\rho$ of the continuity
equation~\eqref{e:continuity}.
For paths and curves we follow the notation introduced
in Section~\ref{basic}. The main improvement in comparison
to Section~\ref{basic} is that we construct here solutions that are
smooth.
\medskip
We begin with a simple remark. Let $I$ be a time interval
and $D$ an open subset of $\mathbb{R}^2$, and let
$\{\Phi(t,\cdot): \, t\in I\}$ be an area-preserving
flow on $D$ of class $C^k$ with $k\ge 2$, that is,
$\Phi:I\times D\to\mathbb{R}^2$ is a map of class $C^k$
such that, for every $t\in I$,
$\Phi(t,\cdot)$ is diffeomorphism from $D$
into its image $\Omega(t):=\Phi(t,D)$
which is area-preserving, which, for regular flows, is equivalent to imposing
for every $z\in D$,
\[
J\Phi(t,z) := \det(\nabla\Phi(t,z)) =1.
\]
We denote by $\Omega$ the (open) set of all points
$(t,x)$ with $t\in I$ and $x\in\Omega(t)$.
It is then well-known that the velocity field
$w:\Omega\to\mathbb{R}^2$ defined by
\begin{equation}
\label{e:velocity}
w(t,x) := \partial_t \Phi(t,z)
\quad\text{with }
x=\Phi(t,z)
\end{equation}
is of class $C^{k-1}$ and divergence-free.
Moreover, given a bounded function $\bar\rho$ on $D$,
the function $\rho:I\times \Omega\to\mathbb{R}$ obtained by
transporting $\bar\rho$ with the flow $\Phi$, that is,
$\rho(t,x) := \bar\rho(z)$ when $x=\Phi(t,z)$,
is a distributional solution of the continuity equation
$\partial_t\rho+\mathop{\mathrm{div}}(w\rho)=0$ on $\Omega$.%
\footnoteb{More precisely, $\rho$ solves the transport
equation $\partial_t\rho + w\cdot\nabla\rho=0$, which agrees
with the continuity equation $\partial_t\rho+\mathop{\mathrm{div}}(w\rho)=0$
because $w$ is divergence-free.}
\medskip
In the next proposition, we modify the previous construction
in order to obtain a velocity field and a solution
defined on $I\times\mathbb{R}^2$, rather than on $I\times \Omega$.
\begin{proposition}
\label{e:tronc}
Let $D$ and $\Phi$ be as above with $D$ a simply-connected domain, and
let $D'$ be a closed subset of $D$.
Then, there exists a divergence-free velocity field
$u:I\times\mathbb{R}^2\to\mathbb{R}^2$ of class $C^{k-1}$ such that
\begin{equation}
\label{e:compat8}
u(t,x)
= w(t,x)
= \partial_t \Phi(t,z),
\quad\text{if $x=\Phi(t,z)$ for some $z\in D'$.}
\end{equation}
Moreover, given a bounded function $\bar\rho:D'\to\mathbb{R}$,
the function $\rho:I\times\mathbb{R}^2\to\mathbb{R}$ defined by
\begin{equation}
\label{e:solution}
\rho(t,x) :=
\begin{cases}
\bar\rho(z) & \text{if $x=\Phi(t,z)$ for some $z\in D'$,} \\
0 & \text{otherwise,}
\end{cases}
\end{equation}
is a distributional solution of the continuity equation
\eqref{e:continuity}.
\end{proposition}
\begin{remark}
The assumption that $D$ is simply connected
can be weakened, but not entirely removed.
Take indeed $D:=\mathbb{R}^2\setminus\{0\}$ and let
$\{\Phi(t,\cdot): \, t\ge 0\}$ be the flow on $D$
associated with the (autonomous) velocity field $w(x):=x/|x|^2$.
Since $w$ is divergence-free on $D$,
the flow is area-preserving.
In addition, the flux of $w$ through any closed curve $\Gamma$
that winds around the origin once counterclockwise,
is $2\pi$, since the
distributional divergence of $w$ on $\mathbb{R}^2$ is $2\pi\, \delta_0$,
with $\delta_0$ the Dirac mass at the origin. On the other hand,
the flux through $\Gamma$ of any divergence-free
velocity field $u$ defined on $\mathbb{R}^2$ must be $0$.
This fact means that \eqref{e:compat8} cannot hold for any
set $D'$ such that $\Phi(t,D')$ contains
such a curve $\Gamma$ for some $t$.
\end{remark}
\begin{proof}[Proof of Proposition~\ref{e:tronc}]
The main idea is simple. We pick $w$ as in \eqref{e:velocity},
and we let $\phi$ be a potential of $w$. We then multiply
this potential by a suitable cut-off function, which agrees with
$1$ on $D'$, and finally take as $u$ the rotated gradient
of the truncated potential, which ensures that $u$ is
divergence-free.
We next present the proof in detail.
We choose a point $z_0\in D$ and smooth cut-off function
$g:\mathbb{R}^2\to [0,1]$ which agrees with $1$ on a open neighbourhood of
$D'$ and has support contained in $D$.
Since $D$ is simply connected, $\Omega(t)$ is simply
connected for every $t\in I$, and consequently the divergence-free
velocity field $w(t,\cdot)$ admits a unique potential
$\phi(t,\cdot)$ in the sense of \S\ref{s:potential} that satisfies the
normalization condition:
\begin{equation}
\label{e:renorm}
\phi(t,x_0(t))=0,
\quad\text{where}\quad
x_0(t):=\Phi(t,z_0)
\, .
\end{equation}
We then define the truncated potential
$\varphi(t,\cdot):\mathbb{R}^2\to \mathbb{R}$ by
\begin{equation}
\label{e:truncation-g}
\varphi(t,x) :=
\begin{cases}
\phi(t,x) \, g(z)
& \text{if $x=\Phi(t,z)$ for some $z\in D$, } \\
0 & \text{otherwise,}
\end{cases}
\end{equation}
and finally take $u:= \nabla^\perp\varphi$.
Since $\Phi$ is of class $C^k$, both $w$ and $\phi$ are of class
$C^{k-1}$ in both variables, and $\phi$ is of class $C^k$ in $x$.%
\footnoteb{Recall that $\phi$ is given by the formula
$\phi(x)=\int_\Gamma w\cdot \eta \, d\sigma$ where
$\Gamma$ is any curve that starts from $x_0(t)$
and ends at $x$.}
Clearly, the same holds for
$\varphi$, which in turn implies that $u$ is of class $C^{k-1}$.
Moreover, $\varphi$ agrees by construction with $\phi$
on an open neighborhood $U$ of the set
of all points of the form $\Phi(t,z)$ with $t\in I$, $z\in D'$,
and therefore $u$ agrees with $w$ on $U$;
in particular \eqref{e:compat8} holds.
Concerning the second part of the claim,
we note that $\rho$ is obtained by transporting $\bar\rho$
with the flow $\Phi$, and hence it solves the continuity
equation $\partial_t\rho +\mathop{\mathrm{div}}(w\rho)=0$ on $\Omega$.
On the other hand, $u$ and $v$ agree on $U$,
which contains the support of $\rho$, and therefore $\rho$
solves the continuity equation $\partial_t\rho +\mathop{\mathrm{div}}(u\rho)=0$
as well.
\end{proof}
Let $\Gamma$ be a curve in the plane.
In the next lemma, we modify the definition of
the parametrization $\Psi$ of the tubular
neighborhood $B(\Gamma,r)$ given in \eqref{e:param}, in order to obtain
an area-preserving map.
\begin{lemma}
\label{s:modtub}
Let $\Gamma$ be a proper curve
parametrized by a path $\gamma:\mathbb{R}\to\mathbb{R}^2$ of class
$C^k$ with $k\ge 3$, such that
$| \dot\gamma(\cdot) | =\ell $ for some constant~$\ell$.
Let $r$ be a positive number such that
$2r/\ell$ is smaller or
equal than the tubular radius $\bar r$ of $\Gamma$,
assumed to be strictly positive,
and let $\Phi:\mathbb{R}\times(-r,r)\to\mathbb{R}^2$
be the map defined by
\begin{equation}
\label{e:param2}
\Phi(s,y):= \gamma(s) + \alpha(s,y / \ell) \, \eta(s)
\quad\text{with}\quad
\alpha(s,y'):= \frac{ 2 y' }{ 1 + \sqrt{1 -2 y' \kappa(s)} }
\, .
\end{equation}
Then, $\gamma(\cdot)=\Phi(\cdot,0)$ and
$\Phi$ is an area-preserving diffeomorphism of
class $C^{k-2}$, the image of which is contained in the
tubular neighbourhood $B(\Gamma,2r/\ell)$ and contains
$B(\Gamma,r/(2\ell))$.
\end{lemma}
\begin{proof}
Using the assumption on $r$ and the fact that the
tubular radius $\bar r$ is smaller or equal than the
curvature radius $1/|\kappa|$ of the curve, it follows that
$r\le \ell/(2\kappa)$, which implies that
$\Phi$ is well defined on $\mathbb{R}\times(-r,r)$.
We observe that $\alpha$ is a function of class $C^{k-2}$
because $\kappa$ is of class $C^{k-2}$, and that
\begin{equation}
\label{e:param3}
\Phi(s,y) = \Psi\big(s,\alpha(s,y/\ell) \big)
\quad\text{for every $s,y$,}
\end{equation}
where $\Psi$ is defined in \eqref{e:param}.
Since $\Psi$ is a diffeomorphism of class $C^{k-1}$
on $\mathbb{R}\times (-\bar r,\bar r)$ and the function
$y \mapsto \alpha(s,y/\ell)$ has strictly positive
derivative for every $s$ and maps $(-r,r)$ into
$(-\bar r,\bar r)$, we have that $\Phi$ is a
diffeomorphism of class $C^{k-2}$.
The fact that $\Phi$ is area-preserving,
that is, $J\Phi=1$ everywhere, can be verified
by a direct computation.
For this purpose, it is convenient to write the gradient of
$\Phi$ at $(s,y)$ using the canonical basis
of $\mathbb{R}^2$ for the domain, and the orthonormal basis
$\tau(s),\eta(s)$ for the codomain, which gives
\[
\nabla \Phi(s,y) =
\begin{pmatrix}
\ell(1- \kappa\, \alpha) & 0 \\
\partial_s\alpha & \frac{1}{\ell} \partial_{y'}\alpha
\end{pmatrix}
\, ,
\]
where $\kappa = \kappa(s)$ and $\alpha=\alpha(s,y/\ell)$.
Finally, the fact that the image of $\Phi$ is contained
in $B(\Gamma,2r/\ell)$ and contains $B(\Gamma,r/(2\ell))$
follows from formula~\eqref{e:param3} and the estimate
$1/(2\ell) \le \alpha\le 2/\ell$.
\end{proof}
In the next subsections, we associate a velocity field $u$
and a solution $\rho$ of the continuity equation \eqref{e:continuity} to a given
time-dependent proper curve $\Gamma$.
This construction will provide the building blocks for the example
described in the next section.
\subsection{Canonical velocity field associated to a time-dependent curve}
\label{s:u}
Let $\Gamma$ be a time-dependent curve,
parametrized by the path $\gamma:I\times\mathbb{R}\to\mathbb{R}^2$
of class $C^k$ with $k\ge 3$, and let $r$ be a positive number
such that
\begin{itemizeb}
\item[(a)]
$|\dot\gamma(t,\cdot)|$ is equal to some $\ell(t)>0$
for every $t\in I$;
\item[(b)]
$2r/\ell(t)$ is smaller or
equal than the tubular radius of $\Gamma(t)$ for every $t\in I$.
\end{itemizeb}
Now, for every $t\in I$ we let $\Phi(t,\cdot)$ be the
diffeomorphism on $D:=\mathbb{R}\times(-r,r)$ defined by \eqref{e:param2},
and we take the velocity field $u:I\times\mathbb{R}^2\to\mathbb{R}^2$
as in Proposition~\ref{e:tronc},
having set $D':=\mathbb{R}\times[-r/2,r/2]$.
A close inspection of the proof of Proposition~\ref{e:tronc}
shows that the construction of $u$ depends on the choice of the
point $z_0$ in $\mathbb{R}\times(-r,r)$ used in the normalization
condition \eqref{e:renorm}, and the choice of the cut-off function $g$.
For the construction at hand, we choose:%
\footnoteb{In the present context, the variable $z$ agrees with $(s,y)$.}
\begin{itemizeb}
\item[(c)]
$z_0:=(0,0)$,
\item[(d)]
$g(s,y):=\bar g(y/r)$,
where $\bar g:\mathbb{R}\to[0,1]$ is a fixed smooth function
that is \emph{even}, takes value $1$ in a neighbourhood of $[-1/2,1/2]$,
and its support is contained in $(-1,1)$.
\end{itemizeb}
\subsection{Canonical solution associated to a time-dependent curve}
\label{s:rho}
We fix an \emph{even} bounded function $\bar\rho:[-1/2,1/2]\to\mathbb{R}$
with zero average and let $\rho:I\times\mathbb{R}^2\to\mathbb{R}$ be the solution of
the continuity equation \eqref{e:continuity} obtained by
replacing the function $\bar\rho(z)$ in formula \eqref{e:solution}
with $\bar\rho(y/r)$, that is,
\begin{equation}
\label{e:solution2}
\rho(t,x) :=
\begin{cases}
\bar\rho(y/r) & \text{if $x=\Phi(t,s,y)$ for some $s\in\mathbb{R}$, $y\in[-r/2,r/2]$,} \\
0 & \text{otherwise.}
\end{cases}
\end{equation}
\begin{remark}
\label{r:boh}
(i)~The velocity field $u$
constructed above is uniquely determined by the choice
of the parametrization $\gamma$,
the number $r$, and the function $\bar g$.
Since $\bar g$ is fixed for the
rest of the paper, the relevant parameters
are therefore $\gamma$ and $r$.
\smallskip
(ii)~The solution $\rho$ depends only on purely
geometrical quantities, and not on the choice of the
parametrization $\gamma$. More precisely, using
formulas~\eqref{e:param2} and \eqref{e:solution2}
one readily checks that the value $\rho(t,x)$ is zero if
$\mathrm{dist} (x,\Gamma(t)) > r/\ell(t)$, and otherwise it depends on:
\begin{itemizeb}
\item $r$, $t$, and $\ell(t)$;
\item the distance $\mathrm{dist} (x,\Gamma(t))$;
\item the curvature of $\Gamma(t)$ at the projection of $x$ on $\Gamma(t)$.
\end{itemizeb}
\smallskip
(iii)~By construction, for every $t\in I$,
the velocity field $u(t,\cdot)$ is supported
in $\Phi(t,\mathbb{R}\times(-r,r))$,
which is contained in
$B(\Gamma(t),2r/\ell(t))$,
while $\rho(t,\cdot)$ is supported in
$\Phi(t,\mathbb{R}\times(-r/2,r/2))$,
which is contained in
$B(\Gamma(t),r/\ell(t))$
(cf.~Lemma~\ref{s:modtub}).
\smallskip
(iv)~The requirement that $\bar \rho$ has zero average and
formula~\eqref{e:solution2}, together with the fact that the change
of variable $\Phi$ is area-preserving, imply that the solution
$\rho(t,\cdot)$ has zero average. This property will be used in
Section~\ref{snake}.
\end{remark}
We suppose now that we are given two time-dependent curves $\Gamma$ and
$\longtilde{\Gamma}$, and we let $u, \tilde u$ and $\rho, \tilde\rho$
be, respectively, the corresponding velocity fields and solutions
constructed in~\S\ref{s:u} and \S\ref{s:rho}.
In the next section, we will need
a sort of locality principle, informally stating that
if $\Gamma$ and $\longtilde{\Gamma}$ agree in a
neighbourhood of some point $x_0$,
then also $u$, $\tilde u$ and
$\rho$, $\tilde\rho$ agree in a neighbourhood of $x_0$.
In Lemma~\ref{s:locality} we give a precise statement of such a locality
principle that
is specifically designed for the application described in the next section.
We first introduce some additional notation.
\subsection{Centered sub-arcs and curved rectangles}
Let $\Gamma$ be a curve pa\-rame\-trized by a path
$\gamma$ such that $|\dot\gamma(\cdot)| = \ell$ constant,
and let $x_0=\gamma(s_0)$ be a point on $\Gamma$. For a given $\delta>0$,
we denote by $I(\Gamma,x_0,\delta)$ the
(centered) sub-arc of all $x\in\Gamma$ such that their geodesic
distance from $x_0$ is strictly less than $\delta$, that is,
\[
I(\Gamma,x_0,\delta)
:= \gamma \big( (s_0 - \delta/\ell,s_0 + \delta/\ell) \big)
\, .
\]
Moreover, given a $\delta'>0$ that is smaller or equal than the tubular
radius of $\Gamma$, we denote by $R(\Gamma,x_0,\delta,\delta')$
the (open, centered) curved rectangle of all $x\in\mathbb{R}^2$ such that their distance
from $\Gamma$ is strictly less than $\delta'$ and such that their
projection on $\Gamma$ belongs to $I(\Gamma,x_0,\delta)$,
that is
\[
R(\Gamma,x_0,\delta,\delta')
:= \Psi \big( (s_0-\delta/\ell,s_0+\delta/\ell)\times(-\delta',\delta') \big)
\, ,
\]
where $\Psi$ is given in \eqref{e:param}.
\begin{lemma}
\label{s:locality}
Let $\Gamma$ and $\longtilde\Gamma$ be two time-dependent, proper curves
of class $C^3$, parametrized by $\gamma, \tilde\gamma:I\times\mathbb{R}\to\mathbb{R}^2$
respectively.
Assume that conditions (a) and~(b) in \S\ref{s:u}
are verified by $\gamma$ and $\tilde\gamma$
with the same $\ell:I\to(0,+\infty)$ and the same $r>0$. Let then
$u, \tilde u$ be defined as in~\S\ref{s:u} and $\rho, \tilde\rho$
be defined as in~\S\ref{s:rho}.
Assume, in addition, that there exist $\delta>0$ and $s_0 \in \mathbb{R}$ such that
for every~$t\in I$,
\begin{itemizeb}
\item[(a)]
$\gamma(t,s_0) = \tilde\gamma(t,0) = : x_0(t)$;
\item[(b)]
the sub-arcs $I \big( \Gamma(t),x_0(t),\delta \big)$
and $I \big( \longtilde\Gamma(t),x_0(t),\delta \big)$
coincide and have the same orientation;
\item[(c)] the following condition is satisfied:
\[
\int\limits_{\gamma(t,[0,s_0])}
\hskip -5 pt v_\mathrm{n}(t,x) \, d\sigma(x) = 0
\,,
\]
where $v_\mathrm{n}$ denotes the normal velocity of $\Gamma$,
\end{itemizeb}
Then, $u(t,x) = \tilde u(t,x)$
and $\rho(t,x) = \tilde\rho(t,x)$
for every $t\in I$ and every $x$ in
the curved rectangle
$R(t) := R \big( \Gamma(t),x_0(t),\delta,2r/\ell(t) \big)$.
\end{lemma}
\begin{proof}
The proof is not difficult, but we must revisit
the entire construction of $u$ and $\rho$,
which is divided between \S\ref{s:u}, \S\ref{s:rho},
and the proofs of Proposition~\ref{e:tronc}
and Lemma~\ref{s:modtub}.
We shall fix $t\in I$.
Using assumptions (a) and (b), and the fact that
$\gamma(t,\cdot)$ and $\tilde\gamma(t,\cdot)$
have the same parametrization speed $\ell(t)$,
we obtain that
\[
\gamma(t,s+s_0) = \tilde\gamma(t,s)
\quad\text{when $|s|\le\delta/\ell$.}
\]
From this identity, it readily follows
that the flows $\Phi$ and $\longtilde{\Phi}$
defined by \eqref{e:param2} satisfy
\begin{equation}
\label{e:7.7.0g}
\Phi(t,s+s_0,y) = \longtilde{\Phi}(t,s,y)
\quad\text{when $|s|\le\delta/\ell$, $|y|<r$,}
\end{equation}
which implies that the velocity fields $w$
and $\longtilde w$ defined by \eqref{e:velocity} satisfy
\begin{equation}
\label{e:7.7g}
w(t,x) = \longtilde w(t,x)
\quad\text{when $x\in U(t)$,}
\end{equation}
where
\[
U(t):= \big\{ \longtilde \Phi(t,s,y) \, : \ |s|\le\delta/\ell \, , \ |y|<r \big\}
\, .
\]
\smallskip
Let now $\phi(t,\cdot)$ and $\longtilde\phi(t,\cdot)$ be the potentials
of $w(t,\cdot)$ and $\longtilde w(t,\cdot)$, respectively,
constructed in the proof of Proposition~\ref{e:tronc}.
We claim that
\begin{equation}
\label{e:7.8g}
\phi(t,x) = \longtilde\phi(t,x)
\quad\text{when $x\in U(t)$.}
\end{equation}
Since the corresponding fields agree on $U$, it suffices to show
that these potentials agree at one point of $U$, and we can actually
show that they both vanish at $x_0(t)$.
Indeed, formula \eqref{e:renorm}, statement~(c) in
\S\ref{s:u}, and the identities $\Phi(t,0,0) = \gamma(t,0)$
and $\longtilde\Phi(t,0,0) = \tilde\gamma(t,0)$
yield
\[
\phi \big( t,\gamma(t,0) \big)
= \longtilde\phi \Big( t, \tilde\gamma(t,0) \Big)
= 0
\,.
\]
Next, as $\tilde\gamma(t,0) = x_0(t)$,
we obtain
\[
\tilde\phi(t,x_0(t)) = 0
\, .
\]
Finally, since $x_0(t) = \gamma(t,s_0)$,
taking into account assumption~(c) and the identity
$v_\mathrm{n}=w\cdot\eta$, it follows that
\begin{align*}
\phi(t,x_0(t))
= \phi(t,\gamma(t,s_0))
& = \phi(t,\gamma(t,s_0)) - \phi(t,\gamma(t,0)) \\
& = \int\limits_{\gamma(t,[0,s_0])}
\hskip -5 pt w\cdot\eta \, d\sigma
= \int\limits_{\gamma(t,[0,s_0])}
\hskip -5 pt v_\mathrm{n} \, d\sigma
= 0 \, ,
\end{align*}
and the proof of \eqref{e:7.8g} is complete.
The rest of the proof is straightforward:
from \eqref{e:7.8g} and the choice of the cut-off function $g$
(see requirement~(d) in \S\ref{s:u}) we obtain that the
truncated potentials $\varphi(t,\cdot)$ and $\tilde\varphi(t,\cdot)$
defined by \eqref{e:truncation-g} agree on $U(t)$.
Moreover, one can show that both potentials vanish
on $R(t)\setminus U(t)$, and therefore they agree on
the whole $R(t)$, which implies that the corresponding
velocity fields $u(t,\cdot)$ and $\tilde u(t,\cdot)$
agree on~$R(t)$.
\smallskip
It remains to show that
\begin{equation}
\label{e:7.11g}
\rho(t,x)=\tilde\rho(t,x)
\quad\text{when $x\in R(t)$,}
\end{equation}
but this fact follows from Remark~\ref{r:boh}(ii).
\end{proof}
\section{Second example: Peano snake}
\label{snake}
In this final section we verify Assumptions~\ref{ass2} and \ref{ass3}
for any $s$ and $p$ for the specific example, already illustrated in~\S\ref{ss:keycomb}.
By doing this, we validate the assumptions
of Theorem~\ref{t:ass}; in particular, we establish the existence of a bounded,
divergence-free velocity field supported in the unit square,
which is Lipschitz uniformly in time, and the existence of a solution of the
continuity equation~\eqref{e:continuity}, the functional
and geometric mixing scales of which decay exponentially in time.
We call this example the ``Peano snake", since the construction is reminiscent of
the iterative construction of the Peano curve.
The velocity field and the solution provided by our construction
are, in fact, smooth in both time and space (see Remark~\ref{r:smoothC}).
However,
any Sobolev norm of order higher than one is not uniformly bounded in time.
We proceed as follows. Using the tools
provided in Section~\ref{geo2}, we prove in \S\ref{ss:contruC}, \S\ref{s:solse},
and \S\ref{s:regfin} that verifying the validity of Assumptions~\ref{ass2} and
\ref{ass3}
can be reduced to showing the existence of
two time-dependent, proper curves satisfying a certain number of geometric
conditions
described in \S\ref{p:basiccurvesC} and \S\ref{r:equivalenceC}.
In \S\ref{s:costruzione-g}, \S\ref{s:costruzione-g1}, and \S\ref{s:costruzione-g2},
we finally show how the two curves are actually constructed.
\subsection{Quasi self-similar basic curves: conditions}
\label{p:basiccurvesC}
The fundamental ingredients of our construction will be
two time-dependent proper curves $\Gamma_1(t)$ and $\Gamma_2(t)$, with
parametrizations $\gamma_1(t,s)$, $\gamma_2(t,s) : [0,1] \times \mathbb{R} \to \mathbb{R}^2$ of
class $C^\infty$, such that, for every $t\in [0,1]$:
\begin{itemizeb}
\item[(a)]
$\gamma_1(t,0) = \gamma_2(t,0) = (0,-1/2)$, $\gamma_1(t,1) = (0,1/2)$,
and $\gamma_2(t,1) = (1/2,0)$;
\item[(b)]
there exists a constant $\delta>0$ such that outside the square
$(1-\delta)\mathscr{Q}$, homothetic to $\mathscr{Q}$,
each of the curves $\Gamma_1(t)$ and $\Gamma_2(t)$ agrees with two unbounded
half-lines, orthogonal to $\partial\mathscr{Q}$ through the points defined in (a);
\item[(c)]
$| \dot \gamma_1 (t,s)| = | \dot \gamma_2 (t,s)| =: \ell(t)$,
and in particular the intersections of the curves $\Gamma_1(t)$
and $\Gamma_2(t)$ with $\mathscr{Q}$ have length $\ell(t)$ for every $t$;
\item[(d)]
denoting by $v_\mathrm{n}^1$, resp.~$v_\mathrm{n}^2$, the normal velocity of $\Gamma_1$, resp.~$\Gamma_2$,
\[
\int\limits_{\gamma_1(t,[0,1])} v_\mathrm{n}^1(t,x) \, d\sigma(x) = 0 \,,
\quad \text{resp.}\quad
\int\limits_{\gamma_2(t,[0,1])} v_\mathrm{n}^2(t,x) \, d\sigma(x) = 0 \,;
\]
\item[(e)]
for every square $Q$ that agrees with one the $25$ squares in the tiling
$\mathscr{T}_{1/5}$
of $\mathscr{Q}$, the sub-arc $\Gamma_1(1) \cap Q$ can be written as a translated,
rescaled, and possibly rotated copy of $\Gamma_1(0) \cap \mathscr{Q}$ or $\Gamma_2(0) \cap \mathscr{Q}$;
the same holds for $\Gamma_2(1) \cap Q$.
\end{itemizeb}
\subsection{Explanation of the combinatorics}
\label{s:combin}
An example of curves satisfying (some of) the conditions in \S\ref{p:basiccurvesC}
is illustrated in Figure~\ref{f:comb1}. In this figure, we can
see the evolution of $\Gamma_1$ starting from a straight
segment $\Gamma_1(0)$ inside $\mathscr{Q}$, and the evolution of
$\Gamma_2$ starting from a (slightly modified) quarter of circle
$\Gamma_2(0)$ inside~$\mathscr{Q}$. Both $\Gamma_1(1)$
and $\Gamma_2(1)$ consist of (rescaled, translated, and rotated)
straight segments and (slightly modified) quarters of circles.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig8a}
\caption{Toy example of basic curves satisfying (a), (b),
and (e) in \S\ref{p:basiccurvesC}.}
\label{f:comb1}
\end{center}
\end{figure}
From the geometry of $\Gamma_1$ and $\Gamma_2$ as depicted in Figure
\ref{f:comb1}, it follows that the curves satisfy conditions~(a), (b), and (e)
(which is the
most relevant), while they do not satisfy condition~(c),
and condition~(d) depends on the parametrization and
cannot be verified from the picture.
In fact, providing examples of basic curves
satisfying all the assumptions in \S\ref{p:basiccurvesC} is significantly more
complex and the construction will be carried out in
\S\ref{s:costruzione-g}, \S\ref{s:costruzione-g1}, and \S\ref{s:costruzione-g2}.
The combinatorial structure of our construction has been explained
in~\S\ref{ss:keycomb}. As illustrated in Figure~\ref{f:comb2}, we patch together
four (rotated) copies of the curve $\Gamma_2(0)$ and we let each of them
evolve as described in Figure~\ref{f:comb1}. The ``matching'' of the curves
at the interfaces between the four sub-squares will be essential in
establishing the smoothness of the velocity field and the solutions, which is
done in~\S\ref{s:regfin} and \S\ref{s:regsolfin}.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig8b}
\caption{Patching together four copies of $\Gamma_2$, and subsequent evolution in time.}
\label{f:comb2}
\end{center}
\end{figure}
\subsection{Reduction: simplified geometric conditions}
\label{r:equivalenceC}
We shall consider the following alternative assumptions~(c'),
(c''), and (d'), in which we denote by $\ell_1(t)$, resp.~$\ell_2(t)$,
the length of the intersection of the curve $\Gamma_1(t)$,
resp.~$\Gamma_2(t)$, with $\mathscr{Q}$:
\begin{itemizeb}
\item[(c')]
$\ell_1(0) = \ell_2(0)$;
\item[(c'')]
the derivatives in $t$ of the functions $\ell_1(t)$
and $\ell_2(t)$ are strictly positive;
\item[(d')]
the area of each of the two connected components of
$\mathscr{Q} \setminus \Gamma_1(t)$ equals $1/2$, and the same
holds for the area of the two connected components of
$\mathscr{Q} \setminus \Gamma_2(t)$.
\end{itemizeb}
We claim that assumptions (b)--(c')--(c'')--(d')--(e) imply the full set
of assumptions (a)--(b)--(c)--(d)--(e) in \S\ref{p:basiccurvesC}.
First of all, (a) follows from (b) by choosing suitable parametrizations
$\gamma_1(t,\cdot)$ and $\gamma_2(t,\cdot)$. Next, we can modify
such parametrizations in such a way that $|\dot \gamma_1(t,\cdot)|$
and $|\dot \gamma_2(t,\cdot)|$ are constant for all~$t$. This fact, together
with~(a), entails
that $|\dot \gamma_1(t,s)| = \ell_1(t)$ and $|\dot \gamma_2(t,s)| = \ell_2(t)$.
Assumption~(c'), together with assumption~(e), implies that
$\ell_1(1)=\ell_2(1)=5\ell_1(0)$. Then, assumption~(c'') implies that
with a change of variable in $t$ we can achieve $\ell_1(t) = \ell_2(t) =: \ell(t)$
for all $t$, that is, assumption~(c) holds. Finally, recalling
Remark~\ref{s:rembasic}(iii), we have that (d) and (d') are equivalent.
\begin{remark}
We will construct in \S\ref{s:costruzione-g}, \S\ref{s:costruzione-g1}, and \S\ref{s:costruzione-g2}
at the end of this section two curves $\Gamma_1$ and $\Gamma_2$
that do satisfy the set of conditions in \S\ref{r:equivalenceC}.
The advantage is that the conditions in \S\ref{r:equivalenceC}
are purely geometrical, in the sense
that they are written in terms of the curves $\Gamma_1$ and $\Gamma_2$,
and do not involve the parametrizations $\gamma_1$
and $\gamma_2$, which therefore will not be described.
\end{remark}
In the next subsections, we assume the existence of two quasi
self-similar basic curves as in \S\ref{p:basiccurvesC} that we then
patch together as in~\S\ref{s:combin}. We use
the results in Section~\ref{geo2} to construct the associated velocity fields
and solutions, and then prove they satisfy
Assumptions~\ref{ass2} and \ref{ass3}. Finally,
we establish additional regularity properties
of these velocity fields and solutions.
\subsection{Construction of velocity fields and solutions}
\label{ss:contruC}
Given the two time-dependent proper curves $\Gamma_1$ and $\Gamma_2$ as in
\S\ref{p:basiccurvesC}, we choose a small $r>0$ such that:
\begin{itemizeb}
\item $r < \min\{ r_1(t) \ell(t) /2 , r_2(t) \ell(t) /2 \} $ for every $t \in [0,1]$,
where $r_1(t)$, resp.~$r_2(t)$, is the tubular radius of $\Gamma_1(t)$, resp.~of $\Gamma_2(t)$,
and $\ell(t)$ is as in \S\ref{p:basiccurvesC}(c);
\item $B(\Gamma_1(t),2r) \subset \mathscr{Q}$ and $B(\Gamma_2(t),2r) \subset \mathscr{Q}$
for every $t \in [0,1]$.
\end{itemizeb}
Such an $r$ exists due to the smoothness of $\Gamma_1$ and
$\Gamma_2$, and assumption~(b) in \S\ref{p:basiccurvesC}.
Then, we apply the constructions in \S\ref{s:u} and \S\ref{s:rho} to
construct two associated time-dependent divergence-free velocity fields
$u_1$ and $u_2$ of class $C^\infty$, and two associated solutions $\rho_1$
and $\rho_2$ of class $C^\infty$ of the continuity equation~\eqref{e:continuity}.
One readily checks that:
\begin{itemizeb}
\item[(a)]
$u_1(t,\cdot)$ and $\rho_1(t,\cdot)$ are supported in $B(\Gamma_1(t),2r/\ell(t))$, while
$u_2(t,\cdot)$ and $\rho_2(t,\cdot)$ are supported in $B(\Gamma_2(t),2r/\ell(t))$;
\item[(b)]
$u_1(t,\cdot)$ and $\rho_1(t,\cdot)$ vanish on
$
\partial \mathscr{Q} \setminus \big( (\frac{-2r}{\ell(t)} , \frac{2r}{\ell(t)} ) \times \{ \frac{-1}{2} \}
\cup ( \frac{-2r}{\ell(t)} , \frac{2r}{\ell(t)} ) \times \{ \frac{1}{2} \} \big)
$,
while $u_2(t,\cdot)$ and $\rho_2(t,\cdot)$ vanish on
$
\partial \mathscr{Q} \setminus \big( ( \frac{-2r}{\ell(t)} , \frac{2r}{\ell(t)} ) \times \{ \frac{-1}{2} \}
\cup \{ \frac{1}{2} \} \times (\frac{-2r}{\ell(t)},\frac{2r}{\ell(t)} ) \big)
$;
\item[(c)]
$u_1(t,\cdot)$ and $u_2(t,\cdot)$ are tangent to the boundary $\partial\mathscr{Q}$.
\end{itemizeb}
Indeed, (a) and (b) follow at once from Remark~\ref{r:boh}(iii).
Let us prove (c) for the curve $\Gamma_1$, the proof for $\Gamma_2$ being analogous.
Consider the vertical line $\tilde\Gamma(t)$ with parametrization
$\tilde\gamma(t,s) = (0,\frac{1}{2}+ s \ell(t))$ and let $\tilde u$ be the velocity
field associated to $\tilde \Gamma$ as in \S\ref{s:u}. We use Lemma~\ref{s:locality} with
$s_0=1$ and $x_0(t)=(0,\frac{1}{2})$. We observe that assumption~(b) of the
lemma
is satisfied with the same $\delta>0$ as in condition~(b)
in \S\ref{p:basiccurvesC}, while assumption~(c)
follows from condition~(d) in \S\ref{p:basiccurvesC}; all other assumptions
of the lemma are clearly satisfied by our choice of~$\Tilde\Gamma$.
Lemma~\ref{s:locality}
then implies that for any $t\in [0,1]$, the velocity field $\tilde u(t,\cdot)$
coincides with $u_1(t,\cdot)$ in the curved rectangle
$R \big( \Gamma_1(t),(0,\frac{1}{2}),\delta,\frac{2r}{\ell(t)} \big)$, which
in this case coincides with the rectangle $\big( \frac{-2r}{\ell(t)} , \frac{2r}{\ell(t)} \big)
\times \big( \frac{1}{2} - \delta, \frac{1}{2}+\delta \big)$.
It is, therefore, sufficient to show that $\tilde u(t,\cdot)$ is tangent to
$\big( \frac{-2r}{\ell(t)} , \frac{2r}{\ell(t)} \big) \times \{\frac{1}{2} \}$
for any $t\in [0,1]$. This property follows
from the fact that the area-preserving diffeomorphism $\tilde \Phi$ in
\eqref{e:param2}
associated to $\tilde \Gamma$ has the form
$$
\tilde \Phi(t,y) = \left( \frac{y}{\ell(t)} \,,\, \frac{1}{2} + s \ell(t)
\right),
$$
and from the construction in \S\ref{s:u}.
The fact that $u_1(t,\cdot)$ is tangent to $\big( \frac{-2r}{\ell(t)} , \frac{2r}{\ell(t)} \big) \times \{ \frac{-1}{2} \}$
can be proved in an analogous manner, by considering the line $\tilde\Gamma$
parametrized by $\tilde\gamma (t,s) = \big(0, -\frac{1}{2} + s \ell(t)\big)$, $s_0=0$,
and $x_0(t)=(0, -\frac{1}{2})$.
\subsection{Verification of Assumption~\ref{ass2}}
\label{s:solse}
First, we recall that the velocity fields $u_1$ and $u_2$
are smooth, and in particular satisfy
Assumption~\ref{ass2}(i), while Assumption~\ref{ass2}(ii)
follows from Remark~\ref{r:boh}(iv).
Assumption~\ref{ass2}(iii) can be obtained combining the fact
that the solutions $\rho_1$ and $\rho_2$ can be described in
purely geometric terms (Remark~\ref{r:boh}(ii)) and the quasi
self-similarity of the curves $\Gamma_1$ and~$\Gamma_2$
in condition~(e) in \S\ref{p:basiccurvesC}.
\subsection{Verification of Assumption~\ref{ass3}}
\label{s:regfin}
According to the general scheme for a quasi self-similar construction
(\S\ref{ss:ass}), and to the
specific combinatorial structure of our example (\S\ref{s:combin}), we have
that, for times
$T_n < t \leq T_{n+1}$, on each square of the tiling $\mathscr{T}_{1/(2 \cdot 5^n)}$ the velocity field $u$
equals a rotated and rescaled translation of one of the velocity fields associated
(according to the procedure in \S\ref{ss:contruC}) to the curves $\Gamma_1$ and $\Gamma_2$
in \S\ref{p:basiccurvesC}.
We check that $u$ is in fact smooth on the entire $\mathbb{R}^2$. This fact implies
Assumption~\ref{ass3}
for any $s$ and $p$. The regularity in the interior of each square of the tiling $\mathscr{T}_{1/(2 \cdot 5^n)}$
follows from the regularity of $u_1$ and $u_2$ (\S\ref{ss:contruC}). We are therefore left with checking
the regularity across the boundary of $\mathscr{Q}$, and at the interface between each
pair of
squares of the tiling.
The combinatorial structure in \S\ref{s:combin} guarantees that, at every time,
the rotated and rescaled
translations of the curves $\Gamma_1$ and $\Gamma_2$ never intersect the boundary of $\mathscr{Q}$:
they intersect the boundary of squares of the tiling always at the interface with another square
of the tiling, and never at the boundary of $\mathscr{Q}$. Together
with~\S\ref{ss:contruC}(b), this observation
implies that the velocity field $u$ vanishes in a neighborhood of the boundary
of $\mathscr{Q}$, which ensures the
regularity across the boundary of $\mathscr{Q}$.
The proof of the regularity at the interface between each pair of squares of
the tiling
is based on Lemma~\ref{s:locality}. We fix $Q' \in \mathscr{T}_{1/(2 \cdot 5^n)}$ and
call $\Gamma'$
the rotated and rescaled translation of one of the curves $\Gamma_1$ and
$\Gamma_2$, which lies inside $Q'$ (see Figure~\ref{f:gluing}). We next fix one
of the sides of $Q'$ that is intersected by $\Gamma'$ at its middle point and
call it $\Sigma$.
We call $Q'' \in \mathscr{T}_{1/(2 \cdot 5^n)}$ the square which lies immediately
across $\Sigma$, and
call $\Gamma''$ the rotated and rescaled translation of one of the curves $\Gamma_1$ and
$\Gamma_2$ which lies inside $Q''$. Then, the combinatorial structure in \S\ref{s:combin}
guarantees that also $\Gamma''$ intersects $\Sigma$ at its middle point.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig8c}
\caption{The situation in \S\ref{s:regfin}.}
\label{f:gluing}
\end{center}
\end{figure}
We now recall (see \S\ref{ss:ass} and \S\ref{p:basiccurvesC}(b)) that the curve
$\Gamma'$,
resp.~$\Gamma''$, has been defined as the intersection with $Q'$, resp.~$Q''$,
of the curve $\hat{\Gamma}'$, resp.~$\hat{\Gamma}''$, which agrees with an half-line
outside of $\big(1-\delta/(2 \cdot 5^n)\big) Q'$, resp.~$\big(1-\delta/(2 \cdot
5^n)\big) Q''$.
Consequently, there is a rectangle $R$ of height $\delta / 5^n$ centered at the
middle point
of $\Sigma$ such that $\hat{\Gamma}'$ and $\hat{\Gamma}''$ coincide inside $R$ with
a straight line orthogonal to $\Sigma$.
We can, therefore, argue as in the proof of item~(c) in \S\ref{ss:contruC} and
prove that
the velocity fields $\hat{u}'$ and $\hat{u}''$ generated by $\hat{\Gamma}'$ and $\hat{\Gamma}''$
coincide inside $R$. Since both $\hat{u}'$ and $\hat{u}''$ are smooth, it is then clear
that the restriction of $\hat{u}'$ to $Q'$ patched with the restriction of $\hat{u}''$ to $Q''$
is smooth, given that the velocity fields are tangent to the boundary of the
squares.
We have therefore checked Assumption~\ref{ass3}.
\subsection{Regularity in space of solutions}
\label{s:regsolfin}
The same procedure in \S\ref{s:regfin} can be applied to the solution $\rho$, which is therefore
smooth in the whole $\mathbb{R}^2$ for every time $t$.
\subsection{Regularity in time of velocity fields and solutions}
\label{r:smoothC}
In \S\ref{s:regfin} and \S\ref{s:regsolfin}, we have proved that the velocity
fields and the solutions
are smooth with respect to the space. In fact, they are also smooth with
respect to the time for $0 \leq t \leq 1$: this property follows from the
smoothness
with respect to the time of the basic curves in~\S\ref{p:basiccurvesC}.
We can apply the procedure described in \S\ref{ss:timereg} and obtain
velocity fields and solutions that are smooth with respect to both space
and time, for $x \in \mathbb{R}^2$ and $t \geq 0$. We stress that we are {\em not}
asserting that Sobolev norms of order higher than one are bounded
{\em uniformly in time}: in fact, they blow up exponentially in time (see
Remark~\ref{r:qssfinal}(iii)). This construction will be relevant for the
companion
paper~\cite{loss}, which addresses the issue of the instantaneous loss
of fractional regularity of solutions of the continuity equation~\eqref{e:continuity}.
\medskip
As announced before, we conclude this section by constructing
two time-dependent curves $\Gamma_1, \Gamma_2$ on
the time interval $I=[0,1]$
that satisfy assumptions (b), (c'), (c''), (d') and (e)
in \S\ref{r:equivalenceC}.
Due to the complexity that a rigorous construction would entail, we only
give a precise description of the initial states $\Gamma_1(0)$, $\Gamma_2(0)$,
and of the final states~$\Gamma_1(1)$,~$\Gamma_2(1)$ (see Figure~\ref{f:gamma-i+f}) and sketch some of the intermediate states (Figures~\ref{f:gammauno} and~\ref{f:gammadue}).
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig8d}
\caption{The curves $\Gamma_1$ and $\Gamma_2$ for $t=0$ and $t=1$.}
\label{f:gamma-i+f}
\end{center}
\end{figure}
\subsection{Construction of $\boldsymbol{\Gamma_1}$
and $\boldsymbol{\Gamma_2}$ for $\boldsymbol{t=0, \, t=1}$}
\label{s:costruzione-g}
We take $\Gamma_1$ and $\Gamma_2$ at the initial time $t=0$
and at the final time $t=1$ as in Figure~\ref{f:gamma-i+f}.
Note that the curves $\Gamma_1(0)$ and $\Gamma_2(0)$
have been obtained by modifying the curves
described in Figure~\ref{f:comb1} (which do not satisfy
requirements~(c') and (d') in~\S\ref{r:equivalenceC}).
We specify that:
\begin{itemizeb}
\item
all the small squares drawn in the first and third picture
in Figure~\ref{f:gamma-i+f} have sidelength $1/5$;
\item
outside these small squares the curves $\Gamma_1(0)$ and $\Gamma_2(0)$
consists of segments and half-lines;
\item
the intersections of $\Gamma_1(0)$ and $\Gamma_2(0)$
with the small squares are not exactly quarters of circle,
but smooth curves chosen in such a way that $\Gamma_1(0)$
and $\Gamma_2(0)$ are smooth; these modified quarters
of circle agree up to reflections, rotations, and translations;
\item
the construction parameter $h_1$ and $h_2$ are chosen so that
$\Gamma_1(0)$ and $\Gamma_2(0)$ satisfy
requirements (c') and (d'):
we first choose $h_2$ so that $\Gamma_2(0)$
satisfies the equal-area requirement~(d')
(note that $\Gamma_1(0)$ satisfies (d') independently
of the choice of $h_1$ for symmetry reasons), and then
we choose $h_1$ so that $\Gamma_1(0)$ and $\Gamma_2(0)$ have
the same length in $\mathscr{Q}$, that is, (c') holds;
\item
finally, both $\Gamma_1(1)$ and $\Gamma_2(1)$ consist of $25$
copies of $\Gamma_1(0)$ and $\Gamma_2(0)$,
rotated, reflected, translated, and scaled by a factor $1/5$;
thus requirement~(e) is met.
\end{itemizeb}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig8e}
\caption{The evolution of $\Gamma_1$ from $t=0$ to $t=1$ in two steps.}
\label{f:gammauno}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig8f}
\caption{Details of the moves used in the second step
of the evolution of $\Gamma_1$.}
\label{f:finesteps}
\end{center}
\end{figure}
\subsection{Construction of $\boldsymbol{\Gamma_1(t)}$ for
$\boldsymbol{0 < t <1}$}
\label{s:costruzione-g1}
The transition from $\Gamma_1(0)$ to $\Gamma_1(1)$
is given in Figure~\ref{f:gammauno}.
More precisely, we specify that:
\begin{itemizeb}
\item
the small squares drawn in the fourth picture have sidelength $1/5$,
and within each of these
squares, $\Gamma_1(1/2)$ agrees with one of the modified quarters of
circle used in~\S\ref{s:costruzione-g};
\item
in the second step, we modify $\Gamma_1(1/2)$
within each square of $\mathscr{T}_{1/5}$ using one of ``moves'' described
in Figure~\ref{f:finesteps};
\item
it is clear that $\Gamma_1$ satisfies requirement~(c''), that is,
$\ell_1(t)$ is increasing in $t$
for $0 \le t \le 1$;
\item
$\Gamma_1$ satisfies the equal-area requirement~(d')
during the first step because of the symmetry of the evolution;
\item
$\Gamma_1$ satisfies the equal-area requirement~(d')
also during the second step: we remark indeed that
every move of the first type in Figure~\ref{f:finesteps} preserves
the equal-area condition, while the moves of the second type
can be coupled so that to each move that increases the area of
one connected component of $\mathscr{Q}\setminus\Gamma_1(t)$
(in some square of $\mathscr{T}_{1/5}$) corresponds a move that decreases
the area of that component (in another square).
\end{itemizeb}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{fig8g}
\caption{The evolution of $\Gamma_2$ from $t=0$ to $t=1$ in three steps.}
\label{f:gammadue}
\end{center}
\end{figure}
\subsection{Construction of $\boldsymbol{\Gamma_2(t)}$
for $\boldsymbol{0 < t <1}$}
\label{s:costruzione-g2}
The transition from $\Gamma_2(0)$ to $\Gamma_2(1)$
is described in Figure~\ref{f:gammadue}.
More precisely, we specify that:
\begin{itemizeb}
\item
the small squares drawn in the third picture have sidelength
$1/5$, and within each of these squares $\Gamma_2(1/4)$ agrees
with one of the modified quarters of circle used
in~\S\ref{s:costruzione-g};
\item
it is clear that requirement~(c'') is satisfied
during the entire evolution;
\item
the equal-area requirement~(d') is
satisfied during the first step of this evolution
for reasons of symmetry;
\item
most of $\Gamma_2$ is kept fixed by the evolution in the second step,
except the part contained in the two squares on top-left corner
of $\mathscr{Q}$, which moves downward, and the part contained in
the middle square of the right side, which evolves according to
the second move in Figure~\ref{f:finesteps};
the equal-area requirement~(d') is satisfied at time $t=1/2$,
and is satisfied also for the intermediate times $1/4 < t < 1/2$
provided that these two moves are suitably ``synchronized'', by a change of
time for one of the two basic moves;
\item
in the third step we modify $\Gamma_2(1/2)$
within each square in $\mathscr{T}_{1/5}$ (except the middle square on
in the right side of $\mathscr{Q}$) using one of the moves described
in Figure~\ref{f:finesteps}; note that this step satisfies the
equal-area requirement~(d') for the same reasons as the last step
of the evolution of $\Gamma_1$ does.
\end{itemizeb}
\bibliographystyle{plain}
|
\section{Introduction}
The Kazhdan-Lusztig polynomial $P_M(t)\in\mathbb{Z}[t]$ of a matroid $M$ was introduced in \cite{EPW}.
In the case where $M$ is realizable by a linear space $V\subset\mathbb{C}^n$, the coefficient of $t^i$ in $P_M(t)$
is equal to the dimension of the intersection cohomology group $I\! H^{2i}(X_V; \mathbb{C})$, where $X_V$
is the ``reciprocal plane" of $V$ \cite[Proposition 3.12]{EPW}. In particular, this implies that $P_M(t)\in \mathbb{N}[t]$
whenever $M$ is realizable. We conjectured \cite[Conjecture 2.3]{EPW} that $P_M(t)\in \mathbb{N}[t]$ for every matroid $M$.
We also gave some computations of $P_M(t)$ for uniform matroids and braid matroids of small rank.
The purpose of this paper is to define a more refined invariant.
Given a matroid $M$ equipped with an action of a finite group $W$, we define the equivariant Kazhdan-Lusztig polynomial
$P_M^W(t)$. The coefficients of this polynomial are not integers, but rather virtual representations of the group $W$.
If $W$ is the trivial group, the ring of virtual representations of $W$ is $\mathbb{Z}$, and $P_M^W(t)$ is equal to the ordinary polynomial $P_M(t)$.
More generally, the polynomial $P_M(t)$ may be obtained from $P_M^W(t)$ by sending a virtual representation to its dimension.
If $M$ is equivariantly realizable by a linear space $V\subset\mathbb{C}^n$, the coefficient of $t^i$ in $P_M^W(t)$
is equal to the intersection cohomology group $I\! H^{2i}(X_V; \mathbb{C})$, regarded as a representation of $W$ (Corollary \ref{rep}).
In particular, this implies that the coefficients of $P_M^W(t)$ are honest (rather than virtual) representations of $W$ whenever
$M$ is equivariantly realizable. We conjecture that this is the case even in the non-realizable case (Conjecture \ref{positivity}).
We compute the coefficients of $P_M^W(t)$ for arbitrary uniform matroids (Theorem \ref{uniform}) and for braid matroids
of small rank (Section \ref{sec:calculations}).\\
It is reasonable to ask why bother with an equivariant version of this invariant, especially since there are
still many things that we do not understand about the nonequivariant version.
We have four answers to this question, all of which are illustrated by the case of uniform matroids.
To set notation,
let $U_{m,d}$ be the uniform matroid of rank $d$ on a set of $m+d$ elements, which is equipped
with a natural action of the symmetric group $S_{m+d}$. Let $C_{i,m,d}$ be the coefficient of $t^i$ in the equivariant
Kazhdan-Lusztig polynomial of $U_{m,d}$, and let $c_{i,m,d} = \dim C_{i,m,d}$ be the coefficient
of $t^i$ in the nonequivariant Kazhdan-Lusztig polynomial.
\begin{itemize}
\item {\bf Nicer formulas:} Our formula for $C_{i,m,d}$ (Theorem \ref{uniform}) is very simple;
it is a multiplicity-free sum of irreducible representations that are easy to describe. We could of course
use the hook-length formula for the dimension of an irreducible representation of $S_{m+d}$
to derive a formula for $c_{i,m,d}$, but the resulting formula is messy and unenlightening.
Indeed, we computed a table in the appendix of \cite{EPW} consisting of the numbers $c_{i,m,d}$
for small values of $i$, $m$, and $d$, and at that time we were unable even to guess the general formula.
It was only by keeping track of the extra structure that we were able to see the essential pattern.
\item {\bf More powerful tools:} After we figured out the correct statement of Theorem \ref{uniform},
we attempted to prove the formula for $c_{i,m,d}$ directly (without going through Theorem \ref{uniform}), and we
failed. The Schubert calculus techniques that we employ in the proof of Theorem \ref{uniform}
are considerably more powerful than the tools to which we have access in the nonequivariant setting.
\item {\bf Representation stability:} The sequence of representations $C_{i,m,d}$ is uniformly
representation stable in the sense of Church and Farb \cite{CF}, which essentially means that
it admits a description that is independent of $d$, provided that $d\geq m+2i$ (Remark \ref{stable}).
This phenomenon cannot be seen by looking at the numbers $c_{i,m,d}$.
\item {\bf Non-realizable examples:} It is difficult to write down examples of
non-realizable irreducible matroids for which we can compute the Kazhdan-Lusztig polynomial, and therefore
we had no nontrivial checks of our non-negativity conjecture in the non-equivariant setting.
On the other hand, the uniform matroid $U_{m,d}$ is equivariantly non-realizable provided that both $d$ and $m$
are greater than 1. This means that Theorem \ref{uniform} provides good evidence for Conjecture \ref{positivity},
and therefore by extension for \cite[Conjecture 2.3]{EPW}.
\end{itemize}
The paper is structured as follows. In Section \ref{sec:def}, we define the equivariant characteristic polynomial
and use it to define the equivariant Kazhdan-Lusztig polynomial. This section closely mirrors Section 2 of \cite{EPW},
but some of the basic lemmas are much more technical in the equivariant setting. In particular, Lemma \ref{dual}
is an equivariant version of a well-known statement that is usually proved via M\"obius inversion. This proof
does not work in the equivariant context (due essentially to the fact that the equivariant analogue of the M\"obius
algebra is not associative), so we needed to find a different approach.
Section \ref{sec:uniform} is devoted to the study of uniform matroids, and in particular the statement
and proof of Theorem \ref{uniform}. Our main technique is to express everything in terms of generating
functions that encode all three parameters $i$, $m$, and $d$, and then to manipulate our functional equations
until they can be solved using repeated applications of the Pieri rule. Section \ref{sec:braid} treats the case
of braid matroids. In this case we are not able to give a general formula for the equivariant Kazhdan-Lusztig
polynomial, but we do derive generating function identities that allow us to compute the polynomial
explicitly in small rank.
Finally, in Section \ref{sec:elc} we introduce the notion of equivariant log concavity, which is a generalization
of the usual notion of log concavity to the equivariant setting. The statement that the coefficients of the characteristic
polynomial of a matroid form a log concave sequence goes back to the 1960s, and was only recently proved
by Adiprasito, Huh, and Katz \cite{AHK}. The statement that the coefficients of the Kazhdan-Lusztig polynomial of
a matroid form a log concave sequence was conjectured in \cite[Conjecture 2.5]{EPW}. Here we make the two
analogous conjectures in the equivariant setting (Conjecture \ref{lc}), and we prove equivariant log concavity
of the characteristic polynomial of a uniform matroid (Proposition \ref{uniform-lc}). The notion of equivariant
log concavity will be further developed in a future paper.
\vspace{\baselineskip}
\noindent
{\em Acknowledgments:}
The authors are grateful to Max Wakefield for his help in initiating
this project, and to June Huh and David Speyer for helpful conversations.
NP was supported by NSF grants DMS-0950383 and DMS-1565036.
\section{Definition}\label{sec:def}
Let $M$ be a matroid on the ground set $\mathcal{I}$, and let $W$ be a finite group acting on $\mathcal{I}$ and preserving $M$.
We will refer to this collection of data as an {\bf equivariant matroid} $W\curvearrowright M$.
Let $$\operatorname{grVRep}(W) := \operatorname{VRep}(W)\otimes_\mathbb{Z} \mathbb{Z}[t] \and \operatorname{grRep}(W) := \operatorname{Rep}(W)\otimes_\mathbb{N} \mathbb{N}[t].$$
Note that, for any group homomorphism $\varphi:W'\to W$, we obtain ring maps $$\varphi^*:\operatorname{VRep}(W)\to\operatorname{VRep}(W')
\and \varphi^*:\operatorname{grVRep}(W)\to\operatorname{grVRep}(W')$$
taking honest representations to honest representations.
\subsection{The equivariant characteristic polynomial}
Let $OS^W_{M,i}\in \operatorname{Rep}(W)$ be the degree $i$ part of the Orlik-Solomon algebra of $M$.
We define the {\bf equivariant characteristic polynomial}
$$H^W_M(t) := \sum_{p=0}^{\operatorname{rk} M} (-1)^p t^{\operatorname{rk} M - p} OS^W_{M,p} \in \operatorname{grVRep}(W).$$
Note that the graded dimension of $H_M^W(t)$ is just the usual characteristic polynomial $\chi_M(t)\in \mathbb{Z}[t]$.
The following lemma is an equivariant version of the statement that $\chi_M(1)=0$ for any matroid $M$ of positive rank.
\begin{lemma}\label{one}
For any equivariant matroid $W\curvearrowright M$ of positive rank, $H_M^W(1) = 0$.
\end{lemma}
\begin{proof}
Let $e = \sum_{i\in\mathcal{I}} e_i \in OS_{M,1}$, and consider the complex of $W$-representations
with $i^\text{th}$ term $OS^W_{M,i}$ and with differential given by multiplication by $e$.
Then $H_M^W(1)$ is equal to the Euler characteristic of this complex, which is equal to the Euler
characteristic of its homology. But the homology is zero provided that $M$ has positive rank \cite[2.1]{YuzBOS}.
\end{proof}
Let $L$ be the lattice of flats of $M$.
Given a flat $F\in L$, let $W_F\subset W$ be the stabilizer of $F$.
For any pair of flats $F,G\in L$, let $W_{FG} := W_F\cap W_G$.
Let $M_F$ be the {\bf localization} of $M$ at $F$; this is the matroid on the ground set $F$ whose lattice of flats is isomorphic
to $L_F := \{G\in L\mid G\leq F\}$. Dually, let $M^F$ be the {\bf restriction} of $M$ to $F$; this is the matroid on the ground
set $\mathcal{I}\smallsetminus F$ whose lattice of flats is isomorphic
to $L^F := \{G\in L\mid G\geq F\}$. The action of $W$ on $M$ induces an action of $W_F$ on both $M_F$ and $M^F$.
\begin{lemma}\label{Brieskorn}
For any equivariant matroid $W\curvearrowright M$,
\begin{eqnarray*}H_M^W(t) &=& \sum_{[F]\in L/W}(-1)^{\operatorname{rk} F} t^{\operatorname{crk} F}\operatorname{Ind}_{W_F}^W\!\left(OS^{W_F}_{M_F,\,\operatorname{rk} M_F}\right)\\
&=& \sum_{F\in L}\frac{|W_F|}{|W|}(-1)^{\operatorname{rk} F} t^{\operatorname{crk} F}\operatorname{Ind}_{W_F}^W\!\left(OS^{W_F}_{M_F,\,\operatorname{rk} M_F}\right).\end{eqnarray*}
\end{lemma}
\begin{proof}
Brieskorn's lemma says that the natural map $$\bigoplus_{\substack{F\in L\\ \operatorname{rk} F = p}}OS_{M_F,p}\to OS_{M,p}$$
is an isomorphism. When we incorporate the action of $W$, this map gives us the equation
$$OS_{M,p}^W \;\;= \sum_{\substack{[F]\in L/W\\ \operatorname{rk} F = p}}\operatorname{Ind}_{W_F}^W\!\left(OS^{W_F}_{M_F,p}\right)
= \sum_{\substack{F\in L\\ \operatorname{rk} F = p}}\frac{|W_F|}{|W|}\operatorname{Ind}_{W_F}^W\!\left(OS^{W_F}_{M_F,p}\right).$$
Our statement follows immediately from this.
\end{proof}
\begin{lemma}\label{easy}
For any equivariant matroid $W\curvearrowright M$ of positive rank,
$$\sum_{[F]\in L/W}(-1)^{\operatorname{rk} F} \operatorname{Ind}_{W_F}^W\!\left(OS^{W_F}_{M_F,\,\operatorname{rk} M_F}\right)
= \sum_{F\in L}\frac{|W_F|}{|W|}(-1)^{\operatorname{rk} F} t^{\operatorname{crk} F}\operatorname{Ind}_{W_F}^W\!\left(OS^{W_F}_{M_F,\,\operatorname{rk} M_F}\right) = 0.$$
\end{lemma}
\begin{proof}
This follows immediately from
Lemmas \ref{one} and \ref{Brieskorn}.
\end{proof}
Lemma \ref{easy} is an equivariant version of the statement that
\begin{equation*}\label{mob-left}\sum_{F\in L}\mu(\emptyset, F) = 0\end{equation*}
when $M$ has positive rank.
There is also a dual statement, which says
that \begin{equation*}\label{mob-right}\sum_{F\in L}\mu(F, \mathcal{I}) = 0\end{equation*}
when $M$ has positive rank. Lemma \ref{dual} is an equivariant version of this dual equation. Surprisingly, the proof of Lemma \ref{dual}
is much more difficult than the proof of Lemma \ref{easy}.
\begin{lemma}\label{dual}
For any equivariant matroid $W\curvearrowright M$ of positive rank,
$$\sum_{[F]\in L/W} (-1)^{\operatorname{crk} F} \operatorname{Ind}_{W_F}^W\left(OS^{W_F}_{M^F\!, \,\operatorname{crk} F}\right)
= \sum_{F\in L}\frac{|W_F|}{|W|} (-1)^{\operatorname{crk} F} \operatorname{Ind}_{W_F}^W\left(OS^{W_F}_{M^F\!, \,\operatorname{crk} F}\right) = 0.$$
\end{lemma}
\begin{proof}
For any flat $F$, we have $$OS_{M^F\!, \,\operatorname{crk} F} = \mathbb{C}\{e_S\mid \text{$S$ a basis for $M^F$}\}\;\Big{/}\;\mathbb{C}\{\partial e_C\mid \text{$C$ a circuit for $M^F$ of rank $\operatorname{crk} F$}\}.$$
If $F\leq G$ with $\operatorname{crk} F = p$ and $\operatorname{crk} G = p-1$, we define a map $$\varphi^F_G:OS_{M^F\!,\, p}\to OS_{M^G\!,\, p-1}$$
by the formula $\varphi^F_G(e_S) := \partial e_S$ for any basis $S$ of $M^F$, where we implicitly set $e_i=0$ for all $i\in G$.
More precisely, we note that $S$ can contain at most one element of $G$.
If $S$ contains no elements of $G$, then $\varphi^F_G(e_S) := \partial e_S = 0\in OS_{M^G\!,\, p-1}$. If $S = \{i_1,\ldots,i_r\}$ and $i_k\in G$, then
we put $S_k:= S\smallsetminus \{i_k\}$ and $\varphi^F_G(e_S) := (-1)^k e_{S_k}$. This is well defined because $\partial^2 = 0$.
Let $$C_p(M) := \bigoplus_{\operatorname{crk} F = p}OS_{M^F\!,\, p},$$
and combine the various maps $\varphi^F_G$ to obtain a map $\varphi_p:C_p(M)\to C_{p-1}(M)$. We claim that $(C_\bullet(M), \varphi_\bullet)$
is an exact sequence.
To show that $\varphi_{p}\circ\varphi_{p+1} = 0$, we need to show that, for all $E\leq G$ with $\operatorname{crk} E = p+1$ and $\operatorname{crk} G = p-1$,
we have $$\sum_{E<F<G}\varphi^F_G\circ\varphi^E_F = 0.$$
Let $S$ be a basis for $M^E$. Then $\varphi^F_G\circ\varphi^E_F(e_S) = 0$ unless $F$ contains exactly one element
of $S$ and $G$ contains exactly two elements of $S$. Thus we can reduce to the situation where $S = \{i_1,\ldots,i_r\}$, $F_k$ is the flat spanned
by $S$ and $i_k$, $F_\ell$ is the flat spanned by $S$ and $i_\ell$, and $G$ is the flat spanned by $S$, $i_k$, and $i_\ell$, and we need to show
that $\varphi^{F_k}_G\circ\varphi^E_{F_k}(e_S) + \varphi^{F_\ell}_G\circ\varphi^E_{F_\ell}(e_S) = 0$. This is easily checked by hand.
Thus $(C_\bullet(M), \varphi_\bullet)$ is a complex.
To prove that our complex is exact, we proceed by induction on $\operatorname{rk} M$. The case $\operatorname{rk} M = 1$ is trivial. Fix an $M$ of rank strictly greater than
1, and assume that the statement is proved for all smaller ranks. Choose an index $i\in\mathcal{I}$, and consider the sum
$$C'_\bullet(M):= \bigoplus_{i\in F}OS_{M^F\!,\, \operatorname{crk} F}\subset C_\bullet(M)$$
ranging over all flats $F$ that contain the index $i$.
It is clear that this is a subcomplex, and that $$C'_\bullet(M)\cong C_\bullet(M'),$$ where $M'$ is the restriction of $F$ to the unique flat of rank 1
containing $i$. Let $$C''_\bullet(M):= C_\bullet(M)/ C'_\bullet(M)$$ be the quotient complex.
As a vector space, we have $$C''_p(M) \cong \bigoplus_{\substack{\operatorname{crk} F = p\\ i\notin F}}OS_{M^F\!,\, p}.$$
Furthermore, for each flat $F$ of corank $p$ that does not contain $i$, we have an isomorphism
$$\bigoplus_{\substack{i\notin G\geq F\\ \operatorname{crk} G = 1}}OS_{M_G^F\!,\, p-1}\;\to \;OS_{M^F\!,\, p}$$
given by multiplication by $e_i$. (Indeed, if we choose an order on $\mathcal{I}$ such that $i$ is the maximal element, then
multiplication by $e_i$ gives a bijection from the nbc
basis for the left-hand side to the nbc basis for the right-hand side.\footnote{The abbreviation nbc stands for ``no broken circuit";
see \cite[Theorem 2.8]{YuzOS} for a discussion of this basis.})
These isomorphisms fit together into an isomorphism of complexes $$C''_\bullet(M)\cong \bigoplus_{\substack{i\notin G\\ \operatorname{crk} G = 1}} C_{\bullet-1}(M_G).$$
Now consider the short exact sequence of complexes
$$0\to C'_\bullet(M)\to C_{\bullet}(M)\to C''_{\bullet}(M)\to 0.$$
Since $\operatorname{rk} M' = \operatorname{rk} M_G = \operatorname{rk} M - 1 > 0$, our inductive hypotheses imply that $C'_\bullet(M)$ and $C''(M)$ both have
trivial homology. Then the long exact sequence in homology tells us that homology of $C_\bullet(M)$ vanishes, as well.
Finally, we note that the complex $C_\bullet(M)$ admits an action of $W$ with $$C_p(M) = \sum_{\substack{[F]\in L/W\\ \operatorname{crk} F = p}}
\operatorname{Ind}_{W_F}^W\left(OS^{W_F}_{M^F\!, \,p}\right)\in \operatorname{Rep}(W).$$
Since $C_\bullet(M)$ has trivial homology, its Euler characteristic is zero. This proves the lemma.
\end{proof}
The following lemma is an equivariant version of the statement that $\sum_{F\in L}\chi_{M^F}(t) = t^{\operatorname{rk} M}$.
\begin{lemma}\label{counting}
For any equivariant matroid $W\curvearrowright M$,
$$\sum_{[F]\in L/W} \operatorname{Ind}_{W_F}^W\left(H_{M^F}^{W_F}(t)\right) = \sum_{F\in L} \frac{|W_F|}{|W|}\operatorname{Ind}_{W_F}^W\left(H_{M^F}^{W_F}(t)\right) = t^{\operatorname{rk} M}\tau_W,$$
where $\tau_W$ is the trivial representation of $W$.
\end{lemma}
\begin{proof}
Applying Lemma \ref{Brieskorn} to $W_F\curvearrowright M^F$, we have
\begin{eqnarray*}&&
\sum_{F\in L} \frac{|W_F|}{|W|} \operatorname{Ind}_{W_F}^W\left(H_{M^F}^{W_F}(t)\right)\\
&=& \sum_{F\leq G}\frac{|W_{FG}|}{|W|}(-1)^{\operatorname{rk} G - \operatorname{rk} F}t^{\operatorname{crk} G} \operatorname{Ind}_{W_{FG}}^W\left(OS_{M^F_G,\,\operatorname{rk} G - \operatorname{rk} F}^{W_F^G}\right)\\
&=& \sum_{G\in L} \frac{|W_G|}{|W|}(-1)^{\operatorname{rk} G} t^{\operatorname{crk} G}\operatorname{Ind}_{W_G}^W\left(\sum_{F\leq G}\frac{|W_{FG}|}{|W_G|}(-1)^{\operatorname{rk} F}\operatorname{Ind}_{W_{FG}}^{W_G}\left(OS_{M^F_G, \,\operatorname{rk} G - \operatorname{rk} F}^{W_F^G}\right)\right).
\end{eqnarray*}
Applying Lemma \ref{dual} to $W_G\curvearrowright M_G$, we have
$$\sum_{F\leq G}\frac{|W_{FG}|}{|W_G|}(-1)^{\operatorname{rk} F}\operatorname{Ind}_{W_{FG}}^{W_G}\left(OS_{M^F_G, \,\operatorname{rk} G - \operatorname{rk} F}^{W_F^G}\right) = 0$$
unless $G$ is equal to the unique flat of rank 0, in which case it is equal to $\tau_{W}$.
\end{proof}
\begin{remark}
Suppose that $M$ is $W$-equivariantly realizable by a complex linear space $V$. More precisely, suppose that we are given
a linear subspace $V\subset \mathbb{C}^\mathcal{I}$, preserved by the action of $W$, such that a subset $B\subset \mathcal{I}$ is a basis for $M$
if and only if the projection of $V$ onto $\mathbb{C}^B$ is an isomorphism.
In this case, Lemma \ref{counting} has a nice geometric interpretation.
The right-hand side of the equation is clearly isomorphic to the compactly supported cohomology of $V$. It is possible to compute
this cohomology via a spectral sequence whose $E_1$ page consists of the compactly supported cohomology groups of the various
strata. By comparing the mixed Hodge structures on the various groups, we can conclude that this spectral sequence degenerates
at the $E_2$ page, which is given by the left-hand side of the equation.
\end{remark}
The following lemma is an equivariant version of \cite[Lemma 2.1]{EPW}.
\begin{lemma}\label{whatevernickwantstorenamethis}
For any equivariant matroid $W \curvearrowright M$ of positive rank, $$\sum_{[F]\in L/W} \operatorname{Ind}_{W_F}^{W} \left( t^{\operatorname{rk} F} H_{M_F}^{W_F} (t^{-1}) \otimes H_{M^F}^{W_F}(t)\right)
= \sum_{F\in L}\frac{|W_F|}{|W|} \operatorname{Ind}_{W_F}^{W} \left( t^{\operatorname{rk} F} H_{M_F}^{W_F} (t^{-1}) \otimes H_{M^F}^{W_F}(t)\right)
=0.$$
\end{lemma}
\begin{proof}
Applying Lemma \ref{Brieskorn} to $W_F\curvearrowright M_F$, we have
\begin{eqnarray*}
&&
\sum_{F\in L} \frac{|W_F|}{|W|}\operatorname{Ind}_{W_F}^{W} \left(t^{\operatorname{rk} F} H_{M_F}^{W_F} (t^{-1}) \otimes H_{M^F}^{W_F}(t)\right)\\
&=& \sum_{E\leq F}\frac{|W_{EF}|}{|W|}(-1)^{\operatorname{rk} E}t^{\operatorname{rk} E}\operatorname{Ind}_{W_F}^{W} \left(\operatorname{Ind}_{W_{EF}}^{W_F}
\left(OS_{M_E,\,\operatorname{rk} E}^{W_{EF}}\right)\otimes H_{M^F}^{W_F}(t)\right)\\
&=& \sum_{E\leq F}\frac{|W_{EF}|}{|W|}(-1)^{\operatorname{rk} E}t^{\operatorname{rk} E}\operatorname{Ind}_{W_{EF}}^{W}
\left(OS_{M_E,\,\operatorname{rk} E}^{W_{EF}}\otimes H_{M^F}^{W_{EF}}(t)\right)\\
&=& \sum_{E\leq F}\frac{|W_{EF}|}{|W|}(-1)^{\operatorname{rk} E}t^{\operatorname{rk} E}\operatorname{Ind}_{W_{E}}^{W}
\left(OS_{M_E,\,\operatorname{rk} E}^{W_{E}}\otimes \operatorname{Ind}_{W_{EF}}^{W_E}\left(H_{M^F}^{W_{EF}}(t)\right)\right)\\
&=& \sum_{E\in L}\frac{|W_{E}|}{|W|}(-1)^{\operatorname{rk} E}t^{\operatorname{rk} E}\operatorname{Ind}_{W_{E}}^{W}
\left(OS_{M_E,\,\operatorname{rk} E}^{W_{E}}\otimes \sum_{E\leq F}\frac{|W_{EF}|}{|W_{E}|}\operatorname{Ind}_{W_{EF}}^{W_E}\left(H_{M^F}^{W_{EF}}(t)\right)\right).
\end{eqnarray*}
Applying Lemma \ref{counting} to $W_E\curvearrowright M^E$, this becomes
\begin{eqnarray*}
&& \sum_{E\in L}\frac{|W_{E}|}{|W|}(-1)^{\operatorname{rk} E}t^{\operatorname{rk} E}\operatorname{Ind}_{W_{E}}^{W}
\left(OS_{M_E,\,\operatorname{rk} E}^{W_{E}}\otimes t^{\operatorname{crk} E}\tau_{W_E}\right)\\
&=& t^{\operatorname{rk} M}\sum_{E\in L}\frac{|W_{E}|}{|W|}(-1)^{\operatorname{rk} E}\operatorname{Ind}_{W_{E}}^{W}
\left(OS_{M_E,\,\operatorname{rk} E}^{W_{E}}\right)\\
&=& t^{\operatorname{rk} M} H_M^W(1),
\end{eqnarray*}
which vanishes by Lemma \ref{one}.
\end{proof}
\subsection{The equivariant Kazhdan-Lusztig polynomial}
Now that we have established some basic properties of the equivariant characteristic polynomial, we are ready
to define the equivariant Kazhdan-Lusztig polynomial. In the non-equivariant case, this polynomial is defined in \cite[Theorem 2.2]{EPW}.
The following theorem is a categorical version of that result.
\begin{theorem}\label{def}
There is a unique way to assign to each equivariant matroid $W \curvearrowright M$ an element $P^W_M(t)\in\operatorname{grVRep}(W)$,
called the {\bf equivariant Kazhdan-Lusztig polynomial},
such that the following conditions are satisfied:
\begin{enumerate}
\item If $\operatorname{rk} M = 0$, then $P^W_M(t)$ is equal to the trivial representation in degree 0.
\item If $\operatorname{rk} M > 0$, then $\deg P^W_M(t) < \tfrac{1}{2}\operatorname{rk} M$.
\item For every $M$, $\displaystyle t^{\operatorname{rk} M} P^W_M(t^{-1}) = \sum_{[F] \in L/W}\operatorname{Ind}_{W_F}^W\left(H^{W_F}_{M_F}(t)
\otimes P^{W_F}_{M^F}(t)\right).$
\item Given a homomorphism $\varphi:W'\to W$, $P^{W'}_M(t) = \varphi^* P^W_M(t)$.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $M$ be a matroid of positive rank. We may assume inductively that $P_{M'}^{W'}(t)$ has been defined for every
matroid $M'$ of rank strictly smaller than $\operatorname{rk} M$ and every group $W'$ acting on $M'$. In particular, $P^W_{M^F}(t)$ has been defined for all $\emptyset \neq F \in L(M)$. Let
$$R^W_M(t) \;\;:= \sum_{\emptyset \not = [F] \in L/W} \operatorname{Ind}_{W_F}^W \left( H_{M_F}^{W_F} (t) \otimes P_{M^F}^{W_F}(t) \right);$$
then item 3 says that we want
$$t^{\operatorname{rk} M}P^W_M(t^{-1}) - P_M^W(t) = R^W_M(t).$$
It is clear that there can be at most one element $P^W_M(t)\in\operatorname{grVRep}(W)$ of degree strictly less than $\frac 12 \operatorname{rk} M$ satisfying this condition. The existence of such a polynomial is equivalent to the statement that
$$t^{\operatorname{rk} M} R_M^W(t^{-1}) = - R^W_M(t),$$ so this is what we need to prove.
We have
\begin{align*}
t^{\operatorname{rk} M}R^W_M(t^{-1}) &= t^{\operatorname{rk} M} \sum_{\emptyset \neq [F] \in L/W} \operatorname{Ind}_{W_F}^W \left( H_{M_F}^{W_F}(t^{-1}) \otimes P_{M^F}^{W_F}(t^{-1}) \right)\\
&= \sum_{\emptyset \neq [F] \in L/W} \operatorname{Ind}_{W_F}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_F}(t^{-1}) \otimes t^{\operatorname{rk} M^F} P_{M^F}^{W_F}(t^{-1}) \right)\\
&=\sum_{\emptyset \neq [F] \in L/W} \operatorname{Ind}_{W_F}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_F}(t^{-1}) \otimes \sum_{[G] \in L^F/W_{F}} \operatorname{Ind}^{W_F}_{W_{FG}} \left( H^{W_{FG}}_{M^F_G} (t) \otimes P^{W_{FG}}_{M^G} (t) \right) \right)\\
&= \sum_{\emptyset \neq F \in L} \frac{|W_F|}{|W|} \operatorname{Ind}_{W_F}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_F}(t^{-1}) \otimes \sum_{G \in L^F} \frac{|W_{FG}|}{|W_F|}\operatorname{Ind}^{W_F}_{W_{FG}} \left( H^{W_{FG}}_{M^F_G} (t) \otimes P^{W_{FG}}_{M^G} (t) \right) \right)\\
&=
\sum_{\emptyset\neq F\leq G} \frac{|W_{FG}|}{|W|} \operatorname{Ind}_{W_F}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_F}(t^{-1}) \otimes \operatorname{Ind}^{W_F}_{W_{FG}} \left( H^{W_{FG}}_{M^F_G} (t) \otimes P^{W_{FG}}_{M^G} (t) \right) \right)\\
&= \sum_{\emptyset \neq F \leq G} \frac{|W_{FG}|}{|W|} \operatorname{Ind}_{W_{FG}}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_{FG}}(t^{-1}) \otimes H^{W_{FG}}_{M^F_G} (t) \otimes P^{W_{FG}}_{M^G} (t) \right)\\
&= \sum_{G\neq\emptyset}\sum_{F \leq G} \frac{|W_{FG}|}{|W|} \operatorname{Ind}_{W_{FG}}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_{FG}}(t^{-1}) \otimes H^{W_{FG}}_{M^F_G} (t) \otimes P^{W_{FG}}_{M^G} (t) \right) - R_M^W(t).
\end{align*}
Thus it will suffice to show that, for any flat $G\neq\emptyset$,
$$\sum_{F \leq G} \frac{|W_{FG}|}{|W|} \operatorname{Ind}_{W_{FG}}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_{FG}}(t^{-1}) \otimes H^{W_{FG}}_{M^F_G} (t) \otimes P^{W_{FG}}_{M^G} (t) \right)=0.$$
Indeed, fixing a flat $G\neq \emptyset$, we have
\begin{eqnarray*}
&& \sum_{F \leq G} \frac{|W_{FG}|}{|W|} \operatorname{Ind}_{W_{FG}}^W \left( t^{\operatorname{rk} F} H_{M_F}^{W_{FG}}(t^{-1}) \otimes H^{W_{FG}}_{M^F_G} (t) \otimes P^{W_{FG}}_{M^G} (t) \right)\\
&=&
\sum_{F\leq G} \frac{|W_{FG}|}{|W|} \operatorname{Ind}_{W_G}^W \left( P_{M^G}^{W_G}(t) \otimes \operatorname{Ind}_{W_{FG}}^{W_G} \left( t^{\operatorname{rk} F} H_{M_F}^{W_{FG}}(t^{-1}) \otimes H^{W_{FG}}_{M^F_G}(t) \right) \right)\\
&=& \frac{|W_G|}{|W|} \operatorname{Ind}_{W_G}^W \left( P_{M^G}^{W_G}(t) \otimes \sum_{F \in L_G} \frac{|W_{FG}|}{|W_G|} \operatorname{Ind}_{W_{FG}}^{W_G} \left( t^{\operatorname{rk} F} H_{M_F}^{W_{FG}}(t^{-1}) \otimes H^{W_{FG}}_{M^F_G}(t) \right) \right).
\end{eqnarray*}
Lemma \ref{whatevernickwantstorenamethis}, applied to $W_G \curvearrowright M_G$, says that the internal sum is zero,
as desired.
\end{proof}
The following result follows immediately by looking at the coefficient of $t^{\operatorname{rk} M - i}$ in item 3 above.
\begin{proposition}\label{coef}
Let $C^W_{M,i}\in \operatorname{VRep}(W)$ be the coefficient of $t^i$ in $P^W_M(t)$. If $i < \tfrac{1}{2}\operatorname{rk} M$, then
$$C^W_{M,i} = \sum_{\substack{[F]\in L/W\\ 0\leq j\leq \operatorname{rk} F}} (-1)^j\; \operatorname{Ind}_{W_F}^W\!\Big(OS^{W_F}_{M_F,j}\otimes\; C^{W_F}_{M^F\!,\, \,\operatorname{crk} F - i + j}\Big).$$
\end{proposition}
\begin{proof}
This follows immediately by looking at the coefficient of $t^{\operatorname{rk} M - i}$ on both sides of the equation in Theorem \ref{def}(3).
\end{proof}
\begin{corollary}\label{zero-one}
For any equivariant matroid $W\curvearrowright M$,
$$C^W_{M,0} = \tau_W\and C^W_{M,1}\; = \sum_{\substack{[F]\in L/W\\ \operatorname{crk} F = 1}} \operatorname{Ind}_{W_F}^W(\tau_{W_F}) - OS^W_{M,1}.$$
\end{corollary}
\begin{proof}
We apply Proposition \ref{coef}. When $i=0$, $C^{W_F}_{M^F\!,\, \,\operatorname{crk} F - i + j}\neq 0$ only if $j=0-\operatorname{crk} F$.
The proposition then says that $C^W_{M,0}$ is equal to $C^W_{M^F,0}$, where $F$ is the unique flat of corank 0.
By parts 1 and 4 of Theorem \ref{def}, this is equal to $\tau_W$.
When $i=1$, we have $C^{W_F}_{M^F\!,\, \,\operatorname{crk} F - i + j}\neq 0$ only if $j=0$ and $\operatorname{crk} F=1$ or $j=1$ and $\operatorname{crk} F = 0$.
The first case gives us a contribution of
$$\sum_{\substack{[F]\in L/W\\ \operatorname{crk} F = 1}} \operatorname{Ind}_{W_F}^W(\tau_{W_F})$$
(the permutation representation given by the action of $W$ on the set of corank 1 flats)
and the second case gives us a contribution of $-OS^W_{M,1}$.
\end{proof}
\begin{remark}
By taking dimensions of the representations in Corollary \ref{zero-one}, we recover Propositions 2.11 and 2.12 of \cite{EPW}.
\end{remark}
Suppose that $M$ is $W$-equivariantly realizable by a complex linear space $V\subset \mathbb{C}^{\mathcal{I}}$.
Let $X_V$ be the {\bf reciprocal plane}, which is defined as follows:
$$X_V := \overline{\{z\in (\mathbb{C}^\times)^\mathcal{I}\mid z^{-1}\in V\}} \subset \mathbb{C}^\mathcal{I}.$$
The action of $W$ on $\mathcal{I}$ induces an action on $X_V$. The following corollary is an equivariant
version of \cite[Proposition 3.12]{EPW}.
\begin{corollary}\label{rep}
If $M$ is $W$-equivariantly realizable by a linear subspace $V\subset\mathbb{C}^\mathcal{I}$, then we have $C^W_{M,i} = I\! H^{2i}(X_V; \mathbb{C})\in\operatorname{grRep}(W)$.
\end{corollary}
\begin{proof}
This follows from Proposition \ref{coef} and \cite[Remark 3.6]{PWY}.
\end{proof}
By definition, the coefficients $C^W_{M,i}$ are virtual representations of $W$. When $M$ is $W$-equivariantly realizable, however,
Corollary \ref{rep} implies that they are honest representations. We conjecture that this is always the case, even if $M$ is not
equivariantly realizable.
\begin{conjecture}\label{positivity}
For any equivariant matroid $W \curvearrowright M$, $P^W_M(t)\in\operatorname{grRep}(W)$.
\end{conjecture}
\begin{remark}
When $W$ is the trivial group, Conjecture \ref{positivity} says that the coefficients of the ordinary Kazhdan-Lusztig polynomial
are natural numbers rather than just integers. This conjecture appeared in \cite[Conjecture 2.3]{EPW}. We note, however,
that it is much easier to construct non-realizable examples of equivariant matroids than it is to construct non-realizable examples
of ordinary matroids. For example, let $U_{m,d}$ be the
the uniform matroid of rank $d$ on $m+d$ elements. This matroid is always realizable. However, it has an action of the symmetric
group $S_{m+d}$, and it is equivariantly realizable if and only if $d\in\{0,1\}$ or $m\in\{0,1\}$.
In the following section, we will prove Conjecture \ref{positivity} for arbitrary uniform matroids.
\end{remark}
\section{Uniform matroids}\label{sec:uniform}
Let $U_{m,d}$ be the
the uniform matroid of rank $d$ on $m+d$ elements, which admits an action of the symmetric group $S_{m+d}$.
Let $$H_{m,d}(t) := H_{U_{m,d}}^{S_{m+d}}(t),\qquad P_{m,d}(t) := P_{U_{m,d}}^{S_{m+d}}(t),\and
C_{m,d,i} := C_{U_{m,d},i}^{S_{m+d}}.$$
For any partition $\lambda$ of $m+d$, let $V[\lambda]$ be the irreducible representation of $S_{m+d}$ indexed by $\lambda$.
The purpose of this section is to prove the following result.
\begin{theorem}\label{uniform}For all $i>0$,
$${C_{m,d,i}\;
\;\; = \sum_{b=1}^{\min(m,d-2i)} V[d+m-2i-b+1,b+1,2^{i-1}]\;\in\;\operatorname{Rep}(S_{m+d}).}$$
\end{theorem}
\begin{corollary}
Conjecture \ref{positivity} holds for all uniform matroids.
\end{corollary}
\begin{remark}
When $m=1$, Theorem \ref{uniform} specializes to the main result of \cite{PWY}.
\end{remark}
\begin{remark}
One can use Theorem \ref{uniform}, along with the hook length formula for the dimension of $V_\lambda$,
to compute the coefficients of the ordinary Kazhdan-Lusztig polynomial of $U_{m,d}$.
(The formula is unenlightening, so we will not reproduce it here.) This is a computation
that we were unable to do in \cite{EPW}; see Section 2.4 and the appendix of that paper. Indeed, we still know of no way to compute
these numbers that does not go through Theorem \ref{uniform}.
\end{remark}
\begin{remark}
One immediate consequence of Theorem \ref{uniform} is that,
for any triple $(m,d,i)$, we have $C_{m,d,i} = C_{d-2i,m+2i,i}$.
This was first observed empirically in the non-equivariant setting by Max Wakefield,
based on computer calculations.
We still have no philosophical explanation for which this symmetry should exist.
\end{remark}
\begin{remark}\label{stable}
Another consequence of Theorem \ref{uniform} is that
the sequence $(C_{m,d,i})_{d\in\mathbb{N}}$ is uniformly representation stable in the sense of Church and Farb \cite[Definition 2.3]{CF},
with the stable range beginning at $d=m+2i$.
\end{remark}
\subsection{Translating to symmetric functions}
The {\bf Frobenius characteristic} is an isomorphism of vector spaces
$$\operatorname{ch}:\operatorname{grVRep}(S_n)\to \Lambda_n[t],$$
where $\Lambda_n$ is the space of symmetric functions of degree $n$
\cite[Section I.7]{Macdonald}.
It has the property that, given two graded virtual representations $V_1\in \operatorname{grVRep}(S_{n_1})$ and $V_2\in \operatorname{grVRep}(S_{n_2})$, we have
$$\operatorname{ch} \operatorname{Ind}_{S_{n_1}\times S_{n_2}}^{S_{n_1+n_2}}\!\big(V_1\boxtimes V_2\big) = \operatorname{ch}(V_1)\operatorname{ch}(V_2).$$
Let $$\cH_{m,d}(t) := \operatorname{ch} H_{m,d}(t), \qquad \mathcal{P}_{m,d}(t) := \operatorname{ch} P_{m,d}(t), \and \mathcal{C}_{m,d,i} := \operatorname{ch} C_{m,d,i}.$$
Applying the Frobenious characteristic to the equation in Theorem \ref{def}(3) (and applying Corollary \ref{zero-one}),
we obtain the statement
\begin{equation}\label{md}
t^d \mathcal{P}_{m,d}(t^{-1}) = \cH_{m,d}(t) + \sum_{k=1}^{d}
\cH_{0,d-k}(t) \mathcal{P}_{m,k}(t).\end{equation}
\subsection{Generating functions}
In this section we will work in the ring $\Lambda[[t,u,x]]$ of completed
symmetric functions with coefficients in the ring of formal power series in $t$, $u$, and $x$.
Let $$\cH(t,u,x) := \sum_{d=1}^\infty\sum_{m=0}^\infty \cH_{m,d}(t)u^dx^m
\and \mathcal{P}(t,u,x) := \sum_{d=1}^\infty\sum_{m=0}^\infty \mathcal{P}_{m,d}(t)u^dx^m.$$
Then Equation \eqref{md} for all values of $m$ and $d$ is equivalent to the generating function equation
\begin{equation}\label{gen}
\mathcal{P}(t^{-1},tu,x) = \cH(t,u,x) + \big(1 + \cH(t,u,0)\big)\mathcal{P}(t,u,x).
\end{equation}
\begin{remark}\label{eq-to-exp}
Once we have Equation \eqref{gen}, we obtain for free the corresponding functional equation involving the
(non-equivariant) exponential generating functions.
Let $$H(t,u,x) := \sum_{d=1}^\infty\sum_{m=0}^\infty \dim H_{m,d}(t) \frac{u^d x^m}{(d+m)!} = \sum_{d=1}^\infty\sum_{m=0}^\infty \chi_{U_{m,d}}(t)
\frac{u^d x^m}{(d+m)!}$$
and $$P(t,u,x) := \sum_{d=1}^\infty\sum_{m=0}^\infty \dim P_{m,d}(t) \frac{u^d x^m}{(d+m)!} = \sum_{d=1}^\infty\sum_{m=0}^\infty P_{U_{m,d}}(t)
\frac{u^d x^m}{(d+m)!}.$$
Then we have \begin{equation}\label{uniform exp}
P(t^{-1},tu,x) = H(t,u,x) + \big(1 + H(t,u,0)\big)P(t,u,x).
\end{equation}
Equation \eqref{uniform exp} follows from Equation \eqref{gen} using the following easy observation:
Let $V_i$ be a representation of $S_{n_i}$ for $i\in\{1,2\}$,
and let $V = \operatorname{Ind}_{S_{n_1}\times S_{n_2}}^{S_{n_1+n_2}}\!\left(V_1\boxtimes V_2\right)$.
Then $\frac{\dim V}{(n_1+n_2)!} = \frac{\dim V_1}{n_1!}\cdot\frac{\dim V_2}{n_2!}$.
\end{remark}
The remainder of this section is devoted to deriving explicit expressions for $\cH(t,u,x)$ and $H(t,u,x)$.
Given a partition $\lambda$ of $n$, let $s[\lambda] := \operatorname{ch} V[\lambda]$ be the Schur function associated with $\lambda$.
Let $$s(t) := \sum_{n=0}^\infty t^n s[n],$$
and recall the well-known identity $$s(-t)^{-1} = \sum_{n=0}^\infty t^n s[1^n].$$
\begin{lemma}\label{weird}
$\displaystyle \sum_{e=0}^\infty\sum_{m=0}^\infty t^{e}u^{m} s[m+1,1^{e}] = \frac{1}{t+u}\left(-1 + \frac{s(u)}{s(-t)}\right).$
\end{lemma}
\begin{proof}
We have
\begin{eqnarray*}-1 + \frac{s(u)}{s(-t)} &=& -1 + \left(\sum_{j=0}^\infty t^{j} s[1^{j}]\right)\left(\sum_{k=0}^\infty u^{k} s[k]\right)\\
&=& -1 + \sum_{j=0}^\infty\sum_{k=0}^\infty t^ju^k s[1^{j}]s[k]\\
&=& \sum_{j=1}^\infty\sum_{k=0}^\infty t^ju^ks[k+1,1^{j-1}] + \sum_{j=0}^\infty\sum_{k=1}^\infty t^ju^ks[k,1^j],
\end{eqnarray*}
where the last equality follows from the Pieri rule. Next, observe that
$$\sum_{j=1}^\infty\sum_{k=0}^\infty t^ju^k s[k+1,1^{j-1}] = t\sum_{e=0}^\infty\sum_{m=0}^\infty t^{e}u^{m} s[m+1,1^{e}]$$
and $$\sum_{j=0}^\infty\sum_{k=1}^\infty t^ju^k s[k,1^j] = u\sum_{e=0}^\infty\sum_{m=0}^\infty t^{e}u^{m} s[m+1,1^{e}].$$
Adding them together and dividing by $t+u$, we obtain the desired equation.
\end{proof}
\begin{proposition}\label{htux}
We have $$\cH(t,u,x) = \frac{u}{u-x}\left(-1 + \frac{s(x)}{s(u)}\right) + \frac{tu}{tu-x}\left(\frac{s(tu)}{s(u)} - \frac{s(x)}{s(u)}\right)$$
and $$1 + \cH(t,u,0) = \frac{s(tu)}{s(u)}.$$
\end{proposition}
\begin{proof}
If $i<d$, then
$$OS^{S_{m+d}}_{U_{m,d},i} = \wedge^i \mathbb{C}^{m+d} = V[m+d-i,1^i] + V[m+d-i+1,1^{i-1}].$$
Applying the Frobenius characteristic, we have
$$\operatorname{ch} OS^{S_{m+d}}_{U_{m,d},i} = s[m+d-i,1^i] + s[m+d-i+1,1^{i-1}] = s[1^i]s[m+d-i],$$
where the second equation follows from the Pieri rule.
We also have
$$OS^{S_{m+d}}_{U_{m,d},d} = V[m+1,1^{d-1}],$$
and therefore
$$\operatorname{ch} OS^{S_{m+d}}_{U_{m,d},d} = s[m+1,1^{d-1}].$$
By definition,
$$\cH(t,u,x) = \sum_{d=1}^\infty\sum_{m=0}^\infty\sum_{i=0}^\infty (-t^{-1})^i (tu)^d x^m \operatorname{ch} OS^{S_{m+d}}_{U_{m,d},i}.$$
Letting $k=d-i$, this tells us that
\begin{eqnarray*}
\cH(t,u,x)
&=& \sum_{d=1}^\infty\sum_{m=0}^\infty (-u)^d x^m \operatorname{ch} OS^{S_{m+d}}_{U_{m,d},d}
+ \sum_{k=1}^\infty\sum_{m=0}^\infty\sum_{i=0}^\infty (-t^{-1})^i (tu)^{i+k} x^m \operatorname{ch} OS^{S_{m+i+k}}_{U_{m,i+k},i}\\
&=& \sum_{d=1}^\infty\sum_{m=0}^\infty (-u)^d x^m s[m+1,1^{d-1}]
+ \sum_{k=1}^\infty\sum_{m=0}^\infty\sum_{i=0}^\infty (-u)^i (tu)^k x^m s[1^i]s[m+k]\\
&=& \sum_{d=1}^\infty\sum_{m=0}^\infty (-u)^d x^m s[m+1,1^{d-1}]
+ \left(\sum_{i=0}^\infty (-u)^i s[1^i]\right)\left(\sum_{k=1}^\infty\sum_{m=0}^\infty (tu)^k x^m s[m+k]\right)\\
&=& \frac{u}{u-x}\left(-1 + \frac{s(x)}{s(u)}\right)
+ \frac{tu}{tu-x}\cdot\frac{s(tu)-s(x)}{s(u)},
\end{eqnarray*}
where the last equation follows from
Lemma \ref{weird}. The second statement is obtained from the first by setting $x$ equal to zero.
\end{proof}
The following Corollary follows immediately from Proposition \ref{htux} as in Remark \ref{eq-to-exp}.
We use the fact that $s(t)$ is the Frobenius characteristic of the trivial representation in every degree,
so the exponential generating function for its dimensions is $e^t$.
\begin{corollary}\label{exp-htux}
We have $$H(t,u,x) = \frac{u}{u-x}\left(-1 + e^{x-u}\right) + \frac{tu}{tu-x}\left(e^{tu-u} - e^{x-u}\right)$$
and $$1 + H(t,u,0) = e^{tu-u}.$$
\end{corollary}
Proposition \ref{htux} combines with Equation \eqref{gen} to tell us that
$$\mathcal{P}(t^{-1},tu,x) = \frac{u}{u-x}\left(-1 + \frac{s(x)}{s(u)}\right) + \frac{tu}{tu-x}\left(\frac{s(tu)}{s(u)} - \frac{s(x)}{s(u)}\right) + \mathcal{P}(t,u,x) \frac{s(tu)}{s(u)}.$$
Rearranging terms, this is equivalent to the equation
\begin{equation}\label{sym}
\left(\frac{u}{u-x}+\mathcal{P}(t^{-1},tu,x)\right)s(u)
- \frac{u}{u-x}s(x) =
\left(\frac{tu}{tu-x} + \mathcal{P}(t,u,x)\right) s(tu)
- \frac{tu}{tu-x}s(x).\end{equation}
Let \begin{eqnarray*}\cR(t,u,x) &:=& \left(\frac{tu}{tu-x} + \mathcal{P}(t,u,x)\right)s(tu)
- \frac{tu}{tu-x} s(x)\\
&=& \left(\frac{tu}{tu-x}+ \mathcal{P}(t,u,x)\right)\sum_{n=0}^\infty (tu)^n s[n] - \frac{tu}{tu-x}\sum_{n=0}^\infty x^n s[n]\end{eqnarray*}
be the expression on the right-hand side of Equation \eqref{sym}.
Then Equation \eqref{sym} becomes $$\cR(t^{-1},tu,x) = \cR(t,u,x).$$
The results of this section can be summarized as follows.
\begin{proposition}\label{reformulation}
The element $\mathcal{P}(t,u,x)\in \Lambda[[t,u,x]]$ is uniquely characterized by the following properties:
\begin{itemize}
\item $\mathcal{P}(t,0,x) = 0$,
\item the coefficient of $t^iu^d$ in $\mathcal{P}(t,u,x)$ is zero if $2i\geq d$,
\item $\cR(t^{-1},tu,x) = \cR(t,u,x)$, where $\cR(t,u,x)$ is defined above.
\end{itemize}
\end{proposition}
\begin{remark}
It is interesting to observe exactly what our manipulations of generating functions has bought us.
The straightforward apporach to proving Theorem \ref{uniform} would have been to apply Proposition \ref{coef}
and proceed by induction on $d$. This works in theory, but it involves repeated applications of the Littlewood-Richardson
rule for hooks, and the combinatorics very quickly gets out of hand.
Instead, we will prove Theorem \ref{uniform} by taking our ``guess" for $\mathcal{P}(t,u,x)$ and verifying the equation
$\cR(t^{-1},tu,x) = \cR(t,u,x)$. From the definition of $\cR(t,u,x)$, we see that this will involve repeated
applications of the Pieri rule, which is much simpler than the general Littlewood-Richardson rule. This simplification
is exactly what makes our proof possible.
\end{remark}
\subsection{Proving the theorem}
We are now ready to prove Theorem \ref{uniform}.
Let $$\mathcal{P}'(t,u,x) := \sum_{d=1}^\infty u^ds[d] + \sum_{i=0}^\infty\sum_{d=1}^\infty\sum_{m=0}^\infty t^iu^dx^m \sum_{b=1}^{\min(m,d-2i)} s[d+m-2i-b+1,b+1,2^{i-1}].$$
Here and throughout this section we adopt the notational convention that $$s[a,2,2^{-1}] = s[a] \and s[a,b+1,2^{-1}] = 0\;\;\text{if $b>1$};$$
in particular, the coefficient of $t^0u^dx^m$ in $\mathcal{P}'(t,u,x)$ is equal to $s[d+m]$ for any $d\geq 1$ and $m\geq 0$.
Let $$\cR'(t,u,x) := \left(\frac{tu}{tu-x}+ \mathcal{P}'(t,u,x)\right)\sum_{n=0}^\infty (tu)^n s[n] - \frac{tu}{tu-x}\sum_{n=0}^\infty x^n s[n].$$
By Proposition \ref{reformulation}, Theorem \ref{uniform} is equivalent to the statement that
$\cR'(t^{-1},tu,x) = \cR'(t,u,x)$.
The coefficient of $t^i u^d x^m$ in $\cR'(t,u,x)$ is equal to
\begin{eqnarray*}
\sum_{b=1}^m\sum_{k=b-d+2i}^i s[k] s[d+k+m-2i-b+1,b+1,2^{i-k-1}]
\;\;&+&\;\; \begin{cases}
s[i]s[d-i]\;\;\;\text{if $m=0<d-i$}\\
0\;\;\;\text{otherwise}
\end{cases}\\
&+&\;\; \begin{cases}
s[m+d]\;\;\;\text{if $i=d>0$}\\
0\;\;\;\text{otherwise.}
\end{cases}\end{eqnarray*}
Thus we need to show that this expression is invariant under the substitution $i\leftrightarrow d-i$.
If $d=0$ or $i>d$, the expression is equal to zero. If $d>0$ and $i=0$ or $i=d$, then the expression is equal to $s[d+m]$.
Thus, we may assume that $0<i<d$.
If $m=0$, the expression is equal to $s[i]s[d-i]$, which is clearly invariant. Thus, we may further assume that $m\neq 0$, which means
that we can restrict our attention to the expression
$$\sum_{b=1}^m\sum_{k=b-d+2i}^i s[k] s[d+k+m-2i-b+1,b+1,2^{i-k-1}].$$
Letting $r=d-2i$, we can rewrite this expression as $$\Psi(i,r,m) := \sum_{b=1}^m\sum_{k=b-r}^i s[k] s[r+k+m-b+1,b+1,2^{i-k-1}],$$
and we want to show that it is equal to
$$\Phi(i,r,m) := \Psi(i+r,-r,m) = \sum_{b=1}^m\sum_{j=b}^{i} s[j+r] s[j+m-b+1,b+1,2^{i-j-1}].$$
The Pieri rule tells us that any Schur function appearing $\Psi(i,r,m)$
must be of the form $s[A,B,C,2^D]$ or $s[A,B,C,2^{D},1]$,
where we continue to adhere to our notational convention:
$$s[A,B,2,2^{-1}] = s[A,B],\qquad s[A,2,2,2^{-2}] = s[A],$$
and so on.
Let us focus on the case of $s[A,B,C,2^D]$ with $D\geq 0$ and $A+B+C+2D = 2i+r+m$.
The Schur function $s[A,B,C,2^D]$ can appear in the $k$ summand of $\Psi(i,r,m)$ when $k=i-D-2$ or $k=i-D-1$.
In each of these summands, the number of times that $s[A,B,C,2^D]$ is equal to the number of values
of $b\leq \min(m,k+r)$ satisfying the inequalities
\begin{equation*}\label{ineqs}
A \geq r+k+m-b+1\geq B\geq b+1 \geq C \geq 2,
\end{equation*}
which ensure that the partitions $[r+k+m-b+1,b+1,2^{i-k-1}]$ and $[A,B,C,2^D]$ interlace.
More precisely,
let $$\varepsilon_1 := \min(m,r+i-D-1+m-B,B-1,i-D-2+r),\qquad \varepsilon_2 := \min(m,r+i-D+m-B,B-1,i-D-1+r),$$
$$\Upsilon_1 := \text{max}(C-1,r+i-D-1+m-A),\and \Upsilon_2 := \text{max}(C-1,r+i-D+m-A).$$
Then the coefficient
of $s[A,B,C,2^D]$ in $\Psi(i,r,m)$ is equal to
$$\text{max}(0, \varepsilon_1-\Upsilon_1+1) + \text{max}(0, \varepsilon_2-\Upsilon_2+1),$$
where the first summand represents the number of possible values for $b$ when $k=i-D-2$, and the second summand
represents the number of possible values for $b$ when $k=i-D-1$.
Similarly, let
$$\varepsilon'_1 := \min(m,i-D-1+m-B,B-1,i-D-2),\qquad \varepsilon'_2 := \min(m,i-D+m-B,B-1,i-D-1),$$
$$\Upsilon'_1 := \text{max}(C-1,i-D-1+m-A),\and \Upsilon'_2 := \text{max}(C-1,i-D+m-A).$$
Then the coefficient
of $s[A,B,C,2^D]$ in $\Phi(i,r,m)$ is equal to
$$\text{max}(0, \varepsilon'_1-\Upsilon'_1+1) + \text{max}(0, \varepsilon'_2-\Upsilon'_2+1),$$
where the first summand represents the number of possible values for $b$ when $j=i-D-2$, and the second summand
represents the number of possible values for $b$ when $j=i-D-1$.
Our plan is to show that $$\varepsilon_1-\Upsilon_1 = \varepsilon'_2-\Upsilon'_2
\and
\varepsilon_2-\Upsilon_2 = \varepsilon'_1-\Upsilon'_1,$$
which will tell us that $s[A,B,C,2^D]$ appears with the same coefficient in $\Psi(i,r,m)$ and $\Phi(i,r,m)$.
\begin{lemma}\label{maxiff} We have
\begin{enumerate}
\item $\Upsilon_1 = C-1$ if and only if $\varepsilon'_2 = \min(m,B-1)$,
\item $\Upsilon_2 = C-1$ if and only if $\varepsilon'_1 =\min(m,B-1)$,
\item $\Upsilon'_1 = C-1$ if and only if $\varepsilon_2=\min(m,B-1)$,
\item $\Upsilon'_2 = C-1$ if and only if $\varepsilon_1=\min(m,B-1)$.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove the forward direction of Lemma \ref{maxiff}(1) and and note that the other cases are identical.
Since $\Upsilon_1 = C-1$, we have
\[
C-1 \geq r+i-D-1+m-A = B+C+D-i-1,
\]
which implies that
$i-D-1 \geq B-1$.
Adding $m+1$ to both sides and subtracting $B$ tells us that
$i-D+m-B \geq m$,
hence $\varepsilon'_2=m$ or $B-1$.
\end{proof}
Now consider the expression \begin{equation}\label{diff}\varepsilon_1-\Upsilon_1 - \varepsilon'_2 + \Upsilon'_2;\end{equation}
we will use a case-by-case analysis to prove that this expression is equal to zero.\\\\
{\em Case 1:} $\Upsilon_1 = C-1 = \Upsilon'_2$. Then $\varepsilon_1 =\min(m,B-1) = \varepsilon'_2$, and \eqref{diff} vanishes.\\\\
{\em Case 2:} $\Upsilon_1 \neq C-1 \neq \Upsilon'_2$. Then
$$\Upsilon_1 = r+i-D-1+m-A,\qquad \Upsilon'_2 = i-D+m-A = \Upsilon_1-(r-1),$$
$$\varepsilon'_2 = \min(i-D+m-B,i-D-1),\and \varepsilon_1 = \min(r+i-D-1+m-B,i-D-2+r) = \varepsilon'_2+(r-1),$$
and therefore \eqref{diff} vanishes.\\\\
{\em Case 3:} $\Upsilon_1 = C-1 \neq \Upsilon'_2$. Then $$\Upsilon'_2 = i-D+m-A,\qquad\varepsilon'_2 = \min(m,B-1),\and
\varepsilon_1 = \min(r+i-D-1+m-B,i-D-2+r).$$
Assume first that $\varepsilon'_2 = m$. This implies that $\varepsilon_1 = r+i-D-1+m-B$, and therefore that the expression \eqref{diff}
is equal to $r+2i+m - A - B - C - 2D = 0$. On the other hand, if $\varepsilon'_2 = B-1$, then $\varepsilon_1 = i-D-2+r$, and we reach
the same conclusion.\\\\
{\em Case 4:} $\Upsilon_1 \neq C-1 = \Upsilon'_2$. This is similar to Case 3.\\\\
We have now shown that the expression \eqref{diff} is equal to zero, and therefore that $\varepsilon_1-\Upsilon_1 = \varepsilon'_2-\Upsilon'_2$.
A similar argument allows us to conclude that $\varepsilon_2-\Upsilon_2 = \varepsilon'_1-\Upsilon'_1$.
This completes the proof that $s[A,B,C,2^D]$ appears with the same coefficient in $\Psi(i,r,m)$ and $\Phi(i,r,m)$.
The other cases, namely Schur functions of the form $s[A,B,C,2^D,1]$ and Schur functions of the form $s[A,B,C,2^D]$ with $D<0$,
can be analyzed in a similar fashion; we leave the details of these cases to the reader. This completes the proof of Theorem \ref{uniform}.\qed
\excise{
\subsection{Exponential generating functions}\label{sec:exp-braid}
\begin{proposition}\label{Htux}
We have $$H(t,u,x) = \frac{xe^{x-u} - tu e^{u(t-1)}}{x-tu} - \frac{xe^{x-u}-u}{x-u}
\and 1 + H(t,u,0) = e^{u(t-1)}.$$
\end{proposition}
\begin{proof}
We have
\begin{eqnarray*}H(t,u,x) &=& \sum_{m\geq 0}\sum_{d\geq 0}\frac{x^m u^d}{(m+d)!}\sum_{i=0}^d \binom{m+d}{i}(-1)^i\big(t^{d-i}-1\big)\\
&=& \sum_{m\geq 0}\sum_{d\geq 0}\sum_{i=0}^d(-1)^i\frac{x^m u^d(t^{d-i}-1)}{i!(m+d-i)!}\\
&=& \sum_{m\geq 0}\sum_{i\geq 0}\sum_{e\geq 0} (-1)^i\frac{x^m u^{i+e}(t^e-1)}{i!(m+e)!}\\
&=& \sum_{i\geq 0}(-1)^i \frac{u^i}{i!} \sum_{m\geq 0}\sum_{e\geq 0}\frac{x^m u^{e}(t^e-1)}{(m+e)!}\\
&=& e^{-u}\left(\sum_{m\geq 0}\sum_{e\geq 0}\frac{x^m (tu)^{e}}{(m+e)!} - \sum_{m\geq 0}\sum_{e\geq 0}\frac{x^m u^{e}}{(m+e)!}\right)\\
&=& e^{-u}\sum_{k\geq 0} \frac{1}{k!} \sum_{m+e=k} \Big( x^m (tu)^e - x^m u^e\Big)\\
&=& e^{-u}\sum_{k\geq 0} \frac{1}{k!} \left( \frac{x^{k+1} - (tu)^{k+1}}{x-tu} - \frac{x^{k+1} - u^{k+1}}{x-u}\right)\\
&=& e^{-u} \left(\frac{xe^x - tue^{tu}}{x-tu} - \frac{xe^x - ue^{tu}}{x-u}\right)\\
&=& \frac{xe^{x-u} - tu e^{u(t-1)}}{x-tu} - \frac{xe^{x-u}-u}{x-u}.
\end{eqnarray*}
The second statement follows from the first by setting $x$ equal to zero.
\end{proof}
}
\section{Braid matroids}\label{sec:braid}
Let $B_n$ be the braid matroid of rank $n-1$. Equivalently, $B_n$ is the matroid associated with the complete
graph on $n$ vertices. The ground set is equal to the set of all 2-element subsets of $[n]$, and the lattice of flats
is equal to the set of set-theoretic partitions of $n$; the rank of a flat is equal to $n$ minus the number of parts of
the partition. The group $S_n$ acts on $B_n$ in the obvious manner.
Let $$K_n(t) := H_{B_n}^{S_n}(t), \qquad Q_n(t) := P_{B_n}^{S_n}(t), \and D_{n,i} := C_{B_n,i}^{S_n}.$$
For any partition $\lambda\vdash n$, let $S_\lambda\subset S_n$ be the stabilizer of a set-theoretic partition of type $\lambda$.
Then Theorem \ref{def}(3) says
\begin{equation}\label{braid-recursion} t^{n-1}Q_n(t^{-1}) = \sum_{\lambda\vdash n}\operatorname{Ind}_{S_\lambda}^{S_n}\left(K_{\lambda_1}(t)
\otimes\cdots\otimes K_{\lambda_{\ell(\lambda)}}(t) \otimes Q_{\ell(\lambda)}(t)\right).\end{equation}
The polynomial $K_n(t)$ is well understood, going back to Lehrer and Solomon \cite{LS}.
An explicit formula for $\operatorname{ch} K_n(t)$ appears in \cite[Theorem 2.7]{HershReiner}.
We cannot give an explicit formula for $Q_n(t)$ in the same way that we did for uniform matroids,
but we will decribe the recursion that can be used to compute it and calculate some examples for small $n$.
\begin{remark}
As defined above, $B_n$ is not equivariantly realizable. However, let $B'_n$ be the analogously defined
matroid with ground set $\mathcal{I}_n$ equal the set of {\bf ordered} pairs of distinct elements of $[n]$, with $\{(i,j), (j,i)\}$
a dependent set.
Then $B_n$ is the underlying simple matroid of $B'_n$, so they have isomorphic lattices
of flats, the same equivariant characteristic polynomial, and the same equivariant Kazhdan-Lusztig polynomial.
Let $V_n = \mathbb{C}^n/\mathbb{C}_\Delta$, and consider the embedding of $V_n$ into $\mathbb{C}^{\mathcal{I}_n}$
given by $x_i-x_j$ in the $(i,j)$-coordinate. This embedding is an $S_n$-equivariant realization of $B'_n$.
It follows from Corollary \ref{rep} that $$Q_n(t) = P_{B_n}^{S_n}(t) = P_{B'_n}^{S_n}(t) \in\operatorname{grRep}(S_n).$$
\end{remark}
\subsection{Generating functions}
Let $K$ be the (virtual, graded) linear species that assigns to the set $[n]$ the (virtual, graded) representation $K_n(t)$ of $S_n$.
Similarly, let $Q$ be the (graded) linear species that assigns to the set $[n]$ the (graded) representation $Q_n(t)$ of $S_n$.
Finally, motivated by Equation \eqref{braid-recursion}, let $\bar Q$ be the
(graded) linear species that assigns to the set $[n]$ the (graded) representation $t^{n-1}Q_n(t^{-1})$ of $S_n$.
In the language of species, Equation \eqref{braid-recursion} says that $\bar Q = Q\circ P$.
Consider the following six generating functions:
$$K(t,z) := \sum_{n= 1}^\infty \dim K_n(t)\frac{z^n}{n!}, \;\; Q(t,z) := \sum_{n=1}^\infty \dim Q_n(t)\frac{z^n}{n!}, \;\;
\bar Q(t) := \sum_{n=1}^\infty \dim \bar Q_n(t)\frac{z^n}{n!}\;\;\in\;\;\mathbb{Q}[[t,z]],$$
$$\mathcal{K}(t) := \sum_{n= 1}^\infty \operatorname{ch} K_n(t), \;\; \mathcal{Q}(t) := \sum_{n= 1}^\infty \operatorname{ch} Q_n(t),
\;\; \bar\mathcal{Q}(t) := \sum_{n= 1}^\infty \operatorname{ch} \bar Q_n(t)\;\;\in\;\;\Lambda[[t]].$$
Note that $\dim K_n(t) = \chi_{B_n}(t) = (t-1)\cdots (t-n+1)$, $\dim Q_n(t) = P_{B_n}(t)$,
and $\dim \bar Q_n(t) = t^{n-1}P_{B_n}(t^{-1})$.
Since we have $\bar Q = Q\circ P$, the theory of species tells us that
\begin{equation}\label{braid-gen}\frac 1 t Q(t^{-1}, tz) = \bar Q(t,z) = Q(t,K(t,z))\and \bar\mathcal{Q}(t) = \mathcal{Q}(t)\big[\mathcal{K}(t)\big],\end{equation}
where square brackets denote plethysm. See, for example, \cite[Proposition 2.1]{Mendez},
which can be extended to virtual species as in \cite{Joyal}.
\begin{remark}
Equation \eqref{braid-gen} can be derived directly from Equation \eqref{braid-recursion}.
The second half of Equation \eqref{braid-gen} follows from the fact that, for any representation $V_i$ of $S_i$ ($i\in\{1,2\}$),
$\operatorname{ch} \operatorname{Ind}_{S_i\wr S_{j}}^{S_{ij}}\!\left(V_j\otimes V_i^{\otimes j}\right)$ is equal to $\operatorname{ch} V_j[\operatorname{ch} V_i]$.
The first half follows then from the second half by computing
the effect of this induction on the dimension of a (virtual, graded) representation.
The language of species simply provides a tidy formalism for these observations.
\end{remark}
\begin{remark}
The generating functions $K(t,z)$ and $\mathcal{K}(t)$ can be understood very explicitly.
We have $$(z+1)^{t} = \sum_{n=0}^\infty\binom{t}{n} z^n = 1 + t\cdot \sum_{n=1}^\infty \chi_{B_n}(t) \frac{z^n}{n}
= 1 + t K(t,z).$$
Based on the work of Lehrer and Solomon, Getzler \cite[Equation (2.5)]{Getzler} gives the following analogous formula
for $\mathcal{K}(t)$:
$$1 + t\mathcal{K}(t) = \prod_{k=1}^\infty(1+p_k)^{\frac{1}{k}\sum_{d|k}\mu(k/d)t^{d}},$$
where $p_k$ is the $k^\text{th}$ power sum symmetric function.
\end{remark}
\subsection{Linear term}
Though we have no general formula for $Q_n(t)$, we can use Corollary \ref{zero-one} to calculate $D_{n,1}$.
\begin{proposition}\label{linear braid}
When $n\leq 3$, $D_{n,1}=0$. When $n\geq 4$, we have $$D_{n,1} = V[n]^{\oplus \left\lfloor\frac{n-2}{2}\right\rfloor} \oplus V[n-1,1]^{\oplus\left\lfloor\frac{n-3}{2}\right\rfloor}
\oplus V[n-2,2]^{\oplus \left\lfloor\frac{n-4}{2}\right\rfloor} \oplus
\bigoplus_{3\leq i\leq \lfloor n/2\rfloor}V[n-i,i]^{d(n,i)},$$
where $$d(n,i) =
\begin{cases}
\quad n/2 - i\;\;\;\,\quad\text{if $n$ is even and $i$ is odd},\\
(n+1)/2 - i\;\;\text{if $n$ is odd},\\
(n+2)/2-i\;\;\text{if $n$ and $i$ are both even}.
\end{cases}
$$\end{proposition}
\begin{proof}
Let $W = S_n$.
By Proposition \ref{zero-one}, $$D_{n,1} = \sum_{\substack{[F]\in L/W\\ \operatorname{crk} F = 1}} \operatorname{Ind}_{W_F}^W(\tau_{W_F}) - OS^{S_n}_{B_n,1}.$$
We have $$OS^{S_n}_{B_n,1} = \operatorname{Sym}^2(\mathbb{C}^n) - \mathbb{C}^n = \operatorname{Sym}^2 V[n-1,1] = V[n-2,2] \oplus V[n-1,1] \oplus V[n].$$
Let $F_k$ be the partition of $[n]$ into $[k]$ and it's complement;
then $\{F_1,\ldots, F_{\lfloor n/2\rfloor}\}$ is a complete set of representatives of $S_n$-orbits of corank 1 flats.
Suppose that $n$ is odd. Then $W_{F_k} = S_k\times S_{n-k}$, so
\begin{eqnarray*}\operatorname{ch} D_{n,1} &=& \sum_{k=1}^{\frac{n-1}{2}} s[k]s[n-k] - \big(s[n-2,2] + s[n-1,1] + s[n]\big)\\
&=& \sum_{k=1}^{\frac{n-1}{2}}\sum_{i=0}^k s[n-i,i] - \big(s[n-2,2] + s[n-1,1] + s[n]\big)\\
&=& \frac{n-3}{2} s[n] + \frac{n-3}{2} s[n-1,1] + \frac{n-5}{2} s[n-2,2] + \sum_{i=3}^{\frac{n-1}{2}}\left(\frac{n+1}{2}-i\right)s[n-i,i].
\end{eqnarray*}
If $n$ is even, then $W_{F_k} = S_k\times S_{n-k}$ for all $k<n/2$, but $W_{F_{n/2}} = S_{n/2} \wr S_2$.
We therefore have
\begin{eqnarray*}\operatorname{ch} D_{n,1} &=& \sum_{k=1}^{\frac{n-2}{2}} s[k]s[n-k] + s[2]\Big[s[n/2]\Big] - \big(s[n-2,2] + s[n-1,1] + s[n]\big)\\
&=& \frac{n-4}{2} s[n] + \frac{n-4}{2} s[n-1,1] + \frac{n-6}{2} s[n-2,2] + \sum_{i=3}^{\frac{n-2}{2}}\left(\frac{n}{2}-i\right)s[n-i,i] + s[2]\Big[s[n/2]\Big].
\end{eqnarray*}
We also have $$s[2]\Big[s[n/2]\Big] = \sum_{j=0}^{n/2} s[n-2j, 2j],$$ and the proposition follows.
\end{proof}
\begin{remark}
Proposition \ref{linear braid} implies that the sequence $\{D_{n,1}\}$ of $S_n$-representations
is {\bf not} representation stable in the sense of Church and Farb.
\end{remark}
\subsection{Calculations}\label{sec:calculations}
We conclude our discussion of braid matroids with a calculation of $D_{n,i}$ for $n\leq 9$ and $i\geq 2$
(since $D_{n,0} = V[n]$ by Proposition \ref{zero-one} and $D_{n,1}$ is computed in Proposition \ref{linear braid}). These computations were
performed in SAGE \cite{sage}, using Equation \eqref{braid-recursion}.
\begin{eqnarray*}
D_{6,2} &=& V[2,2,2]\oplus V[4,2]\oplus V[6]\\\\
D_{7,2} &=& V[2,2,2,1]\oplus V[3,2,2]\oplus V[4,2,1]^{\oplus 2}\oplus V[4,3]^{\oplus 2}\oplus V[5,2]^{\oplus 2}\oplus V[6,1]^{\oplus 2}\oplus V[7]^{\oplus 2}\\\\
D_{8,2} &=& V[2,2,2,2]\oplus V[3,2,2,1]\oplus V[4,2,1,1]\oplus V[4,2,2]^{\oplus 4}\oplus V[4,3,1]^{\oplus 3}\oplus V[4,4]^{\oplus 4}\\ && \oplus V[5,2,1]^{\oplus 4}\oplus V[5,3]^{\oplus 4}\oplus V[6,1,1]\oplus V[6,2]^{\oplus 7}\oplus V[7,1]^{\oplus 4}\oplus V[8]^{\oplus 4}\\\\
D_{9,2} &=& V[3,2,2,2]\oplus V[4,2,2,1]^{\oplus 3}\oplus V[4,3,1,1]\oplus V[4,3,2]^{\oplus 4}\oplus V[4,4,1]^{\oplus 6}\oplus V[5,2,1,1]^{\oplus 2} \oplus V[5,2,2]^{\oplus 7} \\ &&\oplus V[5,3,1]^{\oplus 7}\oplus V[5,4]^{\oplus 7}\oplus V[6,2,1]^{\oplus 9}\oplus V[6,3]^{\oplus 12}\oplus V[7,1,1]^{\oplus 3}\oplus V[7,2]^{\oplus 12}\oplus V[8,1]^{\oplus 8}\oplus V[9]^{\oplus 6}\\\\
D_{8,3} &=& V[2,2,2,2]\oplus V[3,2,2,1]\oplus V[3,3,1,1]\oplus V[4,1,1,1,1]\oplus V[4,2,1,1]\oplus V[4,2,2]^{\oplus 3}\\ &&\oplus V[4,3,1] \oplus V[4,4]^{\oplus 2}\oplus V[5,2,1]^{\oplus 2}\oplus V[5,3]\oplus V[6,2]^{\oplus 2}\oplus V[7,1]\oplus V[8]\\\\
D_{9,3} &=& V[2,2,2,2,1]\oplus V[3,2,2,1,1]^{\oplus 2}\oplus V[3,2,2,2]^{\oplus 4}\oplus V[3,3,1,1,1]^{\oplus 2}\oplus V[3,3,2,1]^{\oplus 4}\oplus V[3,3,3]^{\oplus 2}\\ &&\oplus V[4,1,1,1,1,1]\oplus V[4,2,1,1,1]^{\oplus 3}\oplus V[4,2,2,1]^{\oplus 10}\oplus V[4,3,1,1]^{\oplus 7}\oplus V[4,3,2]^{\oplus 10}\oplus V[4,4,1]^{\oplus 8}\\ &&\oplus V[5,1,1,1,1]\oplus V[5,2,1,1]^{\oplus 8}\oplus V[5,2,2]^{\oplus 12}\oplus V[5,3,1]^{\oplus 13}\oplus V[5,4]^{\oplus 7}\oplus V[6,1,1,1]^{\oplus 2}\\ &&\oplus V[6,2,1]^{\oplus 12}\oplus V[6,3]^{\oplus 11}\oplus V[7,1,1]^{\oplus 4}\oplus V[7,2]^{\oplus 9}\oplus V[8,1]^{\oplus 5}\oplus V[9]^{\oplus 3}
\end{eqnarray*}
\excise{
\todo[inline,color=green!20]{This table could look much nicer.}
\begin{table}[ht
\caption{$D_{n,i}$}
\begin{tabular}{|c|p{2cm}|p{4cm}|p{4cm}|p{4cm}|}
\hline
$n\diagdown i$&0&1&2&3\\
\hline\hline
1&$V[1]$&&&\\
\hline
2&$V[2]$&&&\\
\hline
3&$V[3]$&&&\\
\hline
4&$V[4]$&$V[4]$&&\\
\hline
5&$V[5]$&$V[4,1]\oplus V[5]$&&\\
\hline
6&$V[6]$&$V[4,2]\oplus V[5,1]\oplus V[6]^{\oplus 2}$&$V[2,2,2]\oplus V[4,2]\oplus V[6]$&\\
\hline
7&$V[7]$&$V[4,3]\oplus V[5,2]\oplus V[6,1]^{\oplus 2}\oplus V[7]^{\oplus 2}$&$V[2,2,2,1]\oplus V[3,2,2]\oplus V[4,2,1]^{\oplus 2}\oplus V[4,3]^{\oplus 2}\oplus V[5,2]^{\oplus 2}\oplus V[6,1]^{\oplus 2}\oplus V[7]^{\oplus 2}$&\\
\hline
8&$V[8]$&$V[4,4]\oplus V[5,3]\oplus V[6,2]^{\oplus 2}\oplus V[7,1]^{\oplus 2}\oplus V[8]^{\oplus 3}$&
$V[2,2,2,2]\oplus V[3,2,2,1]\oplus V[4,2,1,1]\oplus V[4,2,2]^{\oplus 4}\oplus V[4,3,1]^{\oplus 3}\oplus V[4,4]^{\oplus 4}\oplus V[5,2,1]^{\oplus 4}\oplus V[5,3]^{\oplus 4}\oplus V[6,1,1]\oplus V[6,2]^{\oplus 7}\oplus V[7,1]^{\oplus 4}\oplus V[8]^{\oplus 4}$&
$V[2,2,2,2]\oplus V[3,2,2,1]\oplus V[3,3,1,1]\oplus V[4,1,1,1,1]\oplus V[4,2,1,1]\oplus V[4,2,2]^{\oplus 3}\oplus V[4,3,1]\oplus V[4,4]^{\oplus 2}\oplus V[5,2,1]^{\oplus 2}\oplus V[5,3]\oplus V[6,2]^{\oplus 2}\oplus V[7,1]\oplus V[8]$\\
\hline
9&$V[9]$&$V[5,4]\oplus V[6,3]^{\oplus 2}\oplus V[7,2]^{\oplus 2}\oplus V[8,1]^{\oplus 3}\oplus V[9]^{\oplus 3}$&
$V[3,2,2,2]\oplus V[4,2,2,1]^{\oplus 3}\oplus V[4,3,1,1]\oplus V[4,3,2]^{\oplus 4}\oplus V[4,4,1]^{\oplus 6}\oplus V[5,2,1,1]^{\oplus 2}\oplus V[5,2,2]^{\oplus 7}\oplus V[5,3,1]^{\oplus 7}\oplus V[5,4]^{\oplus 7}\oplus V[6,2,1]^{\oplus 9}\oplus V[6,3]^{\oplus 12}\oplus V[7,1,1]^{\oplus 3}\oplus V[7,2]^{\oplus 12}\oplus V[8,1]^{\oplus 8}\oplus V[9]^{\oplus 6}$&
$V[2,2,2,2,1]\oplus V[3,2,2,1,1]^{\oplus 2}\oplus V[3,2,2,2]^{\oplus 4}\oplus V[3,3,1,1,1]^{\oplus 2}\oplus V[3,3,2,1]^{\oplus 4}\oplus V[3,3,3]^{\oplus 2}\oplus V[4,1,1,1,1,1]\oplus V[4,2,1,1,1]^{\oplus 3}\oplus V[4,2,2,1]^{\oplus 10}\oplus V[4,3,1,1]^{\oplus 7}\oplus V[4,3,2]^{\oplus 10}\oplus V[4,4,1]^{\oplus 8}\oplus V[5,1,1,1,1]\oplus V[5,2,1,1]^{\oplus 8}\oplus V[5,2,2]^{\oplus 12}\oplus V[5,3,1]^{\oplus 13}\oplus V[5,4]^{\oplus 7}\oplus V[6,1,1,1]^{\oplus 2}\oplus V[6,2,1]^{\oplus 12}\oplus V[6,3]^{\oplus 11}\oplus V[7,1,1]^{\oplus 4}\oplus V[7,2]^{\oplus 9}\oplus V[8,1]^{\oplus 5}\oplus V[9]^{\oplus 3}$\\
\hline
\end{tabular}
\end{table}
}
\section{Equivariant log concavity}\label{sec:elc}
Fix a finite group $W$. We define a sequence $(C_0, C_1, C_2, \ldots)$ in $\operatorname{VRep}(W)$
to be {\bf log concave} if, for all $i>0$, $C_i^{\otimes 2} - C_{i-1}\otimes C_{i+1} \in \operatorname{Rep}(W)$.
We call an element of $\operatorname{grVRep}(W)$ log concave if its sequence of coefficients is log concave.
\begin{remark}
If $W$ is the trivial group, then $\operatorname{VRep}(W)\cong \mathbb{Z}$ and this is the usual notion of log concavity for a sequence of integers.
\end{remark}
\begin{remark}
More generally, we can replace $\operatorname{VRep}(W)$ by any partially ordered ring and have a reasonable definition
of a log concave sequence in that ring.
\end{remark}
\begin{conjecture}\label{lc}
Let $W\curvearrowright M$ be an equivariant matroid.
\begin{enumerate}
\item The equivariant characteristic polynomial $H^W_M(t)$ is log concave.
\item The equivariant Kazhdan-Lusztig polynomial $P^W_M(t)$ is log concave.
\end{enumerate}
\end{conjecture}
\begin{remark}
When $W$ is the trivial group, Conjecture \ref{lc}(1) has existed in various forms since the 1960s,
and was only recently proven by Adiprasito, Huh, and Katz \cite{AHK}. Conjecture \ref{lc}(2) for the trivial group
appeared in \cite[Conjecture 2.5]{EPW}.
\end{remark}
\begin{remark}
We have verified both parts of Conjecture \ref{lc} with a computer for the braid matroid $B_n$ for all $n\leq 9$,
as well as part 2 for the uniform matroid $U_{m,d}$ for all $m,d\leq 15$. Part 1 of the conjecture
for uniform matroids is proved below (Proposition \ref{uniform-lc}) for arbitrary values of $m$ and $d$.
\end{remark}
\begin{remark}
In a forthcoming paper, we will explore the notion of log concavity for $W$-representions in greater detail.
There we will give many more conjectural examples of naturally arising log concave sequences of representations.
\end{remark}
\begin{proposition}\label{uniform-lc}
The equivariant characteristic polynomial $H_{m,d}(t)$ of the uniform matroid $U_{m,d}$ is log concave.
\end{proposition}
\begin{proof}
First, we observe that $OS^{S_{m+d}}_{U_{m,d},i}$ is equal to $OS^{S_{m+d}}_{U_{0,m+d},i}$ when $i<d$,
it is equal to a quotient of $OS^{S_{m+d}}_{U_{0,m+d},i}$ when $i=d$, and it is equal to zero when $i>d$. For this reason,
it suffices to prove the proposition when $m=0$, in which case it says that
$$\left(\wedge^i\mathbb{C}^d\right)^{\otimes 2} - \left(\wedge^{i-1}\mathbb{C}^d\otimes\wedge^{i+1}\mathbb{C}^d\right)\in\operatorname{Rep}(S_d).$$
Let $V_i := V[d-i,1^i]$.
Since $\wedge^i\mathbb{C}^d = V_i\oplus V_{i-1}$, it is sufficient to prove that
\begin{equation}\label{ii}
V_i^{\otimes 2} - V_{i-1}\otimes V_{i+1} \in \operatorname{Rep}(S_d)
\end{equation}
and
\begin{equation}\label{iimo}
V_{i-1}\otimes V_{i} - V_{i-2}\otimes V_{i+1} \in \operatorname{Rep}(S_d).
\end{equation}
These tensor products may be computed using a formula of Remmel \cite[Theorem 2.1]{Remmel}.
We will only prove Equation \eqref{ii}; the proof of Equation \eqref{iimo} is similar.
First, Remmel observes that tensoring with the sign representation takes $V_i$ to $V_{d-i-1}$,
and we may use this to reduce to the case where $i<d/2$, which we will assume for the remainder of the proof.
For any partition $\lambda$ of $d$, Remmel computes $$c(\lambda, i, j):= \dim \operatorname{Hom}(V[\lambda], V_i\otimes V_j);$$
we need to show that $c(\lambda, i, i) \geq c(\lambda, i-1,i+1)$ for all $\lambda$.
The number $c(\lambda, i, j)$ is zero unless $\lambda = [r,1^{d-r}]$ for some $r$
or $\lambda = [q,p,2^k,1^\ell]$ for some $q\geq p\geq 2$ and $k,\ell\geq 0$.
When $\lambda = [r,1^{d-r}]$, Remmel tells us that
$$c\left(\lambda, i, i\right) = \chi(d-2i-1\leq r \leq d)\and
c\left(\lambda, i-1, i+1\right) = \chi(d-2i-1\leq r \leq d-2),$$
where $\chi$ of a statement is 1 if the statement is true and 0 if the statement is false.
In particular, we can see that $c\left([r,1^{d-r}], i, i\right)\geq c\left([r,1^{d-r}], i-1, i+1\right)$.
When $\lambda = [q,p,2^k,1^\ell]$, we put
$$u = \text{max}(p,d-2i),\qquad \omega = 2(d-i-k)-\ell,\qquad x =\lfloor\omega/2\rfloor,$$
$$v_1' = v_0 = \min(q,d-i-k-1),\qquad v_0' = \min(q,d-i-k-2),\and
v_1 = \min(q,d-i-k).$$
Remmel tells us that
$$c\left(\lambda, i, i\right) =
\begin{cases}
0 \qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad\;\text{if $p+k>d-i$}\\
\chi(u\leq x-1\leq v_0)+\chi(u\leq x\leq v_1) \;\;\text{if $p+k\leq d-i$ and $\ell$ is even}\\
\chi(u\leq x\leq v_0)+\chi(u\leq x\leq v_1) \;\qquad\text{if $p+k\leq d-i$ and $\ell$ is odd}
\end{cases}$$
and
$$c\left(\lambda, i-1, i+1\right) =
\begin{cases}
0 \qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad\;\text{if $p+k>d-i-1$}\\
\chi(u\leq x-1\leq v_0')+\chi(u\leq x\leq v_1') \;\;\text{if $p+k\leq d-i-1$ and $\ell$ is even}\\
\chi(u\leq x\leq v_0')+\chi(u\leq x\leq v_1') \;\qquad\text{if $p+k\leq d-i-1$ and $\ell$ is odd.}
\end{cases}$$
Since $v_0'\leq v_0$, $v_1'\leq v_1$, and $d-i-1<d-i$, this implies that $c\left(\lambda, i, i\right)\geq c\left(\lambda, i-1, i+1\right)$.
\end{proof}
\excise{
\todo[inline,color=green!20]{Notes:}
Note that our $V_i$ corresponds in Remmel's notation to $s = d-i$ (and $n=d$).
So we are interested in:
\begin{itemize}
\item $s=d-i=t$ for the first term in Equation \eqref{ii}
\item $s=d-i-1$ and $t=d-i+1$ for the second term in Equation \eqref{ii}
\item $s=d-i$ and $t=d-i+1$ for the first term in Equation \eqref{iimo}
\item $s=d-i-1$ and $t=d-i+2$ for the second term of Equation \eqref{iimo}.
\end{itemize}
}
\begin{remark}
We define a sequence $(C_0, C_1, C_2, \ldots)$ in $\operatorname{VRep}(W)$
to be {\bf strongly log concave} if, for all $0\leq k\leq i\leq j$, $C_i\otimes C_j - C_{i-k}\otimes C_{j+k} \in \operatorname{Rep}(W)$.
We call an element of $\operatorname{grVRep}(W)$ strongly log concave if its sequence of coefficients is strongly log concave.
When $W$ is the trivial group and $C_i\geq 0$ for all $i$, strong log concavity is equivalent to log concavity with no internal zeros.
When $W=S_2$, however, the element $$f(t) := (1+2t+2t^2+t^3)V[2] + (3+2t+2t^2+3t^3)V[1,1]\in\operatorname{grRep}(W)$$ is log concave with
no internal zeros but not strongly log concave.
The notion of strong log concavity may be more natural than the notion of ordinary log concavity,
since it has the property that strong log concavity of $f(t)$ and $g(t)$ implies
strong log concavity of $f(t)\otimes g(t)$.\footnote{The proof of this statement was communicated
to us by David Speyer, and will appear in a future paper.} This fails for ordinary log concavity, as we can see by taking
$f(t)$ as above and $g(t) = (1+t)V[2]$.
Conjecture \ref{lc} and Proposition \ref{uniform-lc} may both be generalized to the corresponding ``strong" versions.
\end{remark}
|
\section{Introduction}
Intensity of involvement among actors in a social network and the types of actions and interactions that arise between them are a long active topic in sociological research. The analysis of group structures has started with the study of dyads and triads that was pioneered by German sociologist Georg Simmel at the end of the nineteenth century \cite{Simmel1908}. In network analysis, the importance of triads has long been emphasized in many research studies, including the highly influential work of Granovetter \cite{Granovetter1973} and, more recently, the work of Watts and Strogatz \cite{Watts1999}, where the notion of clustering (formation of many triangles in networks) is an integral part of the analysis. Thus, both social and network analysis that represent actors as networks nodes and their interactions as (un)directed links have shown that the network structural properties like reciprocity and cycles are very important when trying to explain processes like information spreading and network evolution.
However, retaining the perspective on different connections, i.e. Granovetter's strong and weak ties, sheds additional light on the nature of connections between actors and provides deeper understanding of the triads formations and network clustering. Thus, deeper understanding of the network properties, and therefore the processes that run on top of it, can be gained if it is viewed as a collection of multiple types of links wherein each set of link types represents a separate layer of the network. This multilayered representation has led to the introduction of the concept of multiplex networks which refers to systems in which nodes are connected through more than one type of edges, and therefore, belong to multiple interacting and co-evolving networks. The importance of multiplex networks in sociology has been emphasized by many scholars. In the seminal treatment of multiple networks as the foundation of social structure, White, Boorman and Breiger \cite{Boorman1976} and Boorman and White \cite{White1976} argued that the patterning and interweaving of different types of ties are needed to describe and characterize social structures. It has been demonstrated that multiplexity is critical to diverse phenomena, such as the mobilization of social movements \cite{Gould1991}, the consolidation of political power \cite{Padgett1993}, the emergence of trust in economic relationships \cite{Granovetter1985}, the creation of social bonds within civic networks \cite{Baldassarri2007}, and the organization of party coalitions \cite{Grossman2009}. Multiplexity has been studied to understand scientific collaboration \cite{Maggioni2013}, structural logic of intra-organizational networks \cite{Rank2010}, formation of ties featuring both an economic and a social component in inter-organizational networks \cite{Ferriani2012}, and formation of relationships among producers in the multiplex triads \cite{Shipilov2012}.
Multiplex networks have also been recently subject of particularly intense research by the network science and physics communities. Szell, Lambiotte, and Thurner \cite{Szell2010} worked on correlations and overlap between different types of links and demonstrated the tendency of individuals to play different roles in different networks. Algorithmic detection of tightly connected groups of nodes known as communities in multiplex networks was studied in \cite{Mucha2010}. A framework for growing multiplexes where a node can belong to a different networks was developed by Nicosia et al \cite{Nicosia2013}, while Kim and Goh \cite{Kim2013} studied the possibility of growth of coevolving layers that can shape the network structure and showed analytically and numerically that the coevolution can induce strong degree correlations across layers, as well as modulate degree distributions. Evolutionary game dynamics on structured populations in which individuals take part in several layers of networks of interactions simultaneously which accounts for the different kind of social ties each individual has was studied by Gomez et al \cite{Gomez2012}.
In this paper we aim to study social relations among actors (strong and weak ties) as they appear in real life face-to face (offline) and virtual via social network sites (online) communications using the apparatus of multiplex network analysis/analytics. Real data was collected using an online survey/questionnaire given to two groups of students (two classroom based social networks). The students answers were used to map their own perception of strong vs. weak offline and online connections, thus constructing several offline and online directed friendship networks that constitute the layers of our multiplex network. The primary goal of this research is to study the interrelationship of different social structures represented as multiplex networks. For this reason, we develop normalized actor characteristics for multiplex networks, including metrics for dyads (such as reciprocity) and triads.
We found that normalized reciprocity, three cycles, and triplets of an actor can be expressed using an unifying framework that is based on the comparison of two sets associated with the actor: the set of her/his friends - out links and the set of actors that consider her/him as a friend - in links. By extending these metrics for multiplex networks, we were able to observe the relationship of strong and weak ties in the offline and online space. Analyzing the collected data from two groups of undergraduate students we found that in offline communication strong and weak ties are (almost) equally presented, while in online communication weak ties are dominant. Moreover, weak ties are a lot less reciprocal than strong ties. However, while reciprocities are preserved across layers, the triads (measured using normalized three cycles and triplets) are not significant on the different layers of the multiplex network.
This is the outline of the paper. First, we describe the participants involved in the study and the procedure for collecting data. Next, we address the network endogenous (structural) and exogenous factors. Endogenous factors include graph characteristics such as reciprocity, three-cycles, transitive triplets, together with their normalized versions, as well as their generalizations for multiplex graphs. The section Results summarizes our findings regarding the offline and online multiplex networks generated by collecting the data described in the section Materials and Methods. We conclude the paper with the section Conclusions where we also discuss our future work.
\section{Materials and Methods}
\subsection{Participants and the procedure}
Participants in this study were two groups of undergraduate students enrolled on two courses given at the Faculty of Computer Science and Engineering within the Cyril and Methodius University in Skopje, Macedonia. During the courses there was a dedicated lecture to explain the study and it's objectives in which the students were informed about the research experiment and only those who agreed to participate were part of the experiment. Thus, the two groups represent two social networks of students that attend the same class during one semester of their studies. The total number of students in the first group, Group 1, is 171, out of which 153 voluntarily participated in the study. The gender distribution is almost equal; there were 86 female and 85 male students. The total number of students in the second group, Group 2, is 150 (118 male, 32 female), while only 80 participated in the At the end we summarize our findings regarding the offline and online multiplex networks generated from the collected datastudy. The students age for both groups ranges from 19 years to 22 years. Additional data was also collected for each student via the university online electronic course enrollment system (such as: gender, study program, GPA, accumulated credits).
The online survey was developed by the research team. Each student (that participated in the study) was asked to select from the presented list of classmates those with whom she/he was engaged in face-to-face (offline) communication and Facebook (online) communication. Each chosen contact was described as weak or strong according to the student's own perception by the means of pre-defined categories for frequency of communication. In the case of face-to-face communication a contact (tie) is considered to be strong if the frequency of communication is higher than 5 interactions per month, while it is considered as weak contact (tie) if the number of interactions is larger than once in 3 months, but less than 5 interactions per month. When considering the online communication, strong contacts (ties) are those with whom the student interacts more than five times per week, while a weak contact (tie) is the one with whom there was Facebook activity more than once a month, but less than five times per week. Facebook was chosen as the online social network representative since it's use is extremely wide spread among the population.
\subsection{Friendship graphs as multiplex networks}
Graphs provide a powerful primitive for modeling data in social science. Nodes usually represent real world objects and edges indicate relationships between objects. In sociology, nodes may have attributes associated with them and graphs may contain many different types of relationships. The node attributes are used to describe the features of the objects that the nodes represent. For example, a node representing a student may have attributes that represent the student's gender and department. Different types of edges in a graph correspond to different types of relationships between nodes, such as friends and classmates relationships.
Here we study friendship relations among $n$ actors: the existence of a tie $i \to j$ will be described as $i$ calling $j$ a friend. The ties are represented as binary variables, denoted by $x_{ij}$. A tie from actor $i$ to actor $j$, $i \to j$, is either present or absent ($x_{ij}$ then having values 1 and 0, respectively). The tie variables constitute the network, represented by its $n \times n$ adjacency matrix $X = [x_{ij}]$ (self-ties are excluded). The graph is directed, where each tie $i \to j$ has a sender $i$, who will also be referred to as ego, and a receiver $j$, referred to as alter, as it is common in social network analysis.
We model different types of relationships among actors using the concept of multiplex graphs. In sociology, multiplex graphs (networks) refer to the case when nodes (actors) are connected through more than one type of (socially relevant) ties. In mathematics, such graphs are also called multi-graphs (a multi-graph is a graph that is allowed to have multiple edges, that is, edges that have the same end nodes). We give a definition of multiplex graphs adapted for the study reported in this paper. Let $V $ denote the set of nodes; nodes are connected via $L$ different type of connections (ties). Each type of connection (together with the set of nodes) forms a (directed) graph with $n$ vertices. We denote with $G^\alpha = (V, E^\alpha)$ the graph which represents the connection type $\alpha$, where $E^\alpha$ denotes the set of $\alpha$-type ties, $\alpha = 1, \ldots, L$. Let $ [x_{ij}(G^\alpha)]$ be $n \times n$ adjacency matrix of the graph $G^\alpha$. A multiplex graph $\cal{G}$ is then defined as a collection of all graphs $G^\alpha$ and all edge-aggregated graphs of the form $(V, \cup_{\alpha =\alpha_1}^{\alpha_m} E^\alpha)$, where $\alpha_1, \alpha_m \in \{1, \ldots, L\} $. We assume that the case $\alpha_1=\alpha_m$ is not excluded and thus write
$$
{\cal{G}} = \{ G^{\alpha_1 \alpha_2 \ldots \alpha_m} = (V, E^{\alpha_1} \cup E^{\alpha_2} \cup \ldots \cup E^{\alpha_m}): \alpha_1, \ldots \alpha_m \in \{1, \ldots, L\} \}
$$
for the multiplex graph $\cal{G}$. In the study discussed here, different types of edges correspond to the four types of friendship relations: online weak or strong connections and offline weak or strong connections.
\subsection{Jaccard similarity coefficient}
In order to introduce an unifying framework for discussing graph characteristics, we use the Jaccard index, also known as Jaccard similarity coefficient. It measures the similarity between finite sets and is defined as the size of the intersection divided by the size of the union of the sets:
$$
J(A,B) = \frac{|A \cap B|}{|A \cup B|} = \frac{|A \cap B|}{ |A| + |B| - |A \cap B| },
$$
where $\left\vert{S}\right\vert$ denotes the cardinality of the set $S$. If $A$ and $B$ are both empty, $J(A,B)$ is defined as 1. Note that
$$
0\le J(A,B)\le 1.
$$
The Jaccard distance, which measures dissimilarity between sets and is a metric on the collection of all finite sets, is complementary to the Jaccard coefficient and is obtained by subtracting the Jaccard coefficient from 1.
Using Jaccard index to introduce various actor-based characteristics is based on the following observation. Let $A_i = \{a: i \to a \}$ and $B_j = \{b: j \to b \}$. Then
\begin{eqnarray*}
|A_i \cap B_j| & = & \sum_h x_{ih} x_{jh} \\
|A_i \cup B_j| &= & \sum_h x_{ih} + \sum_h x_{jh} - \sum_h x_{ih} x_{jh},
\end{eqnarray*}
where $|A_i| = \sum_h x_{ih} $ and $ |B_j| = \sum_h x_{jh}$.
\subsection{Network (endogenous) characteristics}
We consider a number of social network characteristics; those depending only on the network are called structural or endogenous characteristics, while characteristics depending on externally given attributes are called covariate or exogenous characteristics.
\subsubsection{Simple-graph characteristics}
Let $G=(V,E)$ be a directed graph and let $i$ be an arbitrary node (actor) in the graph. We denote $S^{in}_i(k)$ and $S^{out}_i(k)$ for the $k$-in-neighborhood of $i$ and $k$-out-neighborhood of $i$, respectively, i.e., for the set of vertices from where vertex $i$ can be reached and for the set of vertices that can be reached from $i$ in $k$ steps (using $k$ directed edges). More formally, these sets are defined as
\begin{eqnarray}
S^{in}_i(k) & = & \{j: \mbox{ there is a path of length } k \mbox{ from } j \mbox{ to } i \} \label{set-in} \\
S^{out}_i(k) & = & \{j: \mbox{ there is a path of length } k \mbox{ from } i \mbox{ to } j \} \label{set-out}
\end{eqnarray}
The general case of arbitrary $k$ will be discussed elsewhere, here we consider only to the case when $k=1$. For $k=1$ we use the shorter notations $S^{in}_i(1) = S^{in}_i$ and $S^{out}_i(1) = S^{out}_i$. In other words, we consider the 1-hop-neighbor sets of node $i$ defined as:
\begin{eqnarray}
S^{out}_i &=& \{j: i \to j \}, \label{eq-s1} \\
S^{in}_i &=& \{j: j \to i \}, \label{eq-s2}
\end{eqnarray}
If the existence of a tie $i \to j$ is being interpreted as $j$ is a friend of $i$, then $S_i^{out}$ is the set of $i$'s friends, while $S_i^{in}$ is the set of nodes that consider $i$ as their friend. For the cardinalities of these sets, we write $d^{out}_i = \sum_j x_{ij} = |S^{out}_i|$ and $d^{in}_i = \sum_j x_{ji} = |S^{in}_i|$. Therefore, $d_i^{out}$ is the number of $i$'s friends and $d_i^{in}$ is the number of nodes that consider $i$ as their friend. These two metrics are known as activity, i.e. tendency to establish friendships, and popularity, i.e. ability to gain friends, in social science. In this paper, we analyse several graph characteristics that can all be expressed using the Jaccard similarity index for a given pair of node sets. Out of the many different possible characteristics, here the focus is set on those that we consider most relevant for this study: reciprocity, three-cycles, and triplets.
\textit{Reciprocity} --
One of the most basic properties of social networks is reciprocity, represented by the number of reciprocated ties of actor $i$ and defined as $\sum_j x_{ij} x_{ji}$, which can be rewritten as
$$
Rec_i = |S^{out}_i \cap S^{in}_i| = \sum_j x_{ij} x_{ji}
$$
Here we consider
normalized reciprocity defied as:
\begin{equation} \label{eq-r1}
r_i = J\left(S_i^{out},S_i^{in}\right) = \frac{\sum_j x_{ij}x_{ji}}{ d_i^{out} + d_i^{in} - \sum_j x_{ij}x_{ji} }.
\end{equation}
The normalized reciprocity enables one to decide on the quantity of reciprocated ties when compared to the total number of ties of both actors. This normalization provides a perspective and places the phenomenon of reciprocity into context by comparing it to the total number of possibilities for reciprocated ties under the given circumstances.
\textit{Three-cycles} -- Next to reciprocity, an essential feature in most social networks is transitivity, or transitive closure which is represented by two metrics: three-cycles and triplets. A cycle of length 3 in a given graph is defined as a sub-graph that consists of a sequence of directed edges $i \to j \to h \to i$ which connect a sequence of vertices $i, j, h$, all distinct from one another.
We discuss the first metric, three-cycles, defined as the number of three-cycles an actor $i$ is involved in, that is
$$
cyc_i = \sum_{j,h} x_{ij} x_{jh} x_{hi}
$$
We introduce a normalized characteristic for the number of three-cycles an actor $i$ is involved in as follows:
\begin{eqnarray*}
tc_i & = & \frac{1}{d_i^{in}} \sum_{h} x_{hi} J(S_i^{out},S_h^{in}) \\
& = & \frac{1}{d_i^{in}} \sum_{h} \frac{ x_{hi} \sum_j x_{ij} x_{jh} }{d_i^{out} +d_h^{in} - \sum_j x_{ij} x_{jh} }
\end{eqnarray*}
$| S_i^{out} \cap S_h^{in}| $ is the number of the common neighbors for both $i$ and $h$ or, the number of $i$'s friends that consider $h$ to be a friend as well.
The Jaccard index between these two sets $S_i^{out}$ and $S_h^{in}$ reflects how similar these sets are in terms of common versus non-common friends. Moreover,
$$
\sum_h x_{hi} | S_i^{out} \cap S_h^{in}| = \sum_h x_{hi} \sum_j x_{ij} x_{jh}
$$
is the number of three-cycles that actor $i$ is involved in. Therefore, by summarizing the Jaccard indexes $J(S_i^{out},S_h^{in})$ for all $h $
one can obtain a measure (metric) $tc_i$ that represents the status of common versus non-common friends of $i$ in the graph when considered in terms of transitive three-cycles.
\textit{Transitive triplets} --
A transitive triplet of length 3 in a given graph is a sub-graph that consists of a sequence of directed edges $i \to j \to h$ and $ i \to h$ which connect a sequence of vertices $i, j, h$ that are all distinct from one another. Recall, the metric triplets is defined as the number of triplets an actor $i$ is involved in, that is
$$
plt_i = \sum_{j.h} x_{ij} x_{jh} x_{ih}.
$$
In a similar fashion as for the three-cycles, one can define normalized metric for the number of triplets in the graph:
\begin{eqnarray*}
tp_i & = & \frac{1}{d_i^{out}} \sum_{j} x_{ij} J(S_i^{out},S_j^{out}) \\
& = & \frac{1}{d_i^{out}} \sum_{j} \frac{ x_{ij} \sum_h x_{ih} x_{jh} }{d_i^{out} + d_j^{out} - \sum_h x_{ih} x_{jh} }
\end{eqnarray*}
Note that the normalized reciprocity, normalized transitive triplets, and normalized three-cycles are related to reciprocity, transitive triplets, and three-cycles, respectively - quantities that are commonly used in social science and graph theory. The fact, shown here, that they can be expressed with Jaccard similarity index not only brings novel understanding of these quantities, but also suggests how they can be extended for multiplex networks, which will be done in the next section. However, take into consideration that while using the Jaccard similarity index to express the network characteristics, we define it as 0 when both considered sets are empty in order to reflect the social aspects of the interpretation, i.e. two nodes that have no friends at all are not reciprocated and are not part of any triplets or cycles.
\subsubsection{Multiplex-graph characteristics}
When considering a multiplex graph wherein $\alpha$ and $\beta$ ($\alpha = \beta $ is not excluded) represent two types of social relationships, the normalized reciprocity, normalized three-cycles, and normalized transitive triplets can be generalized as:
\begin{eqnarray*}
r_i(G^\alpha, G^\beta) &=& J\left(S_i^{out} (G^\alpha) ,S_i^{in} (G^\beta) \right) \\
tc_i(G^\alpha, G^\beta) &=& \frac{1}{d_i^{in}} \sum_{h} x_{hi} J(S_i^{out}(G^\alpha),S_h^{in}(G^\beta)) \\
tp_i (G^\alpha, G^\beta) &=& \frac{1}{d_i^{out}} \sum_{j} x_{ij} J(S_i^{out}(G^\alpha),S_j^{out}(G^\beta)).
\end{eqnarray*}
By using these metrics we can analyze how different types of links interact and form mixed dyad and triad formations. Recalling that in the general case a multiplex graph can consist of several different layers, the multiplex-graph extension of the reciprocity, cycles and triplets metrics enables us to study what type of links are stronger than others, as well as how the network characteristics change when considering only a subset of all existing interrelations among its nodes.
We also consider the following two characteristics called overlapping indexes and defined as:
\begin{eqnarray*}
oi_i^{out} (G^\alpha, G^\beta) &= & J(S_i^{out}(G^\alpha), S_i^{out}(G^\beta)) \\
oi_i^{in} (G^\alpha, G^\beta) & = & J(S_i^{in}(G^\alpha), S_i^{in}(G^\beta))
\end{eqnarray*}
These two metrics are introduced in order to create a contrasting view compared to the normalized multiplex graph characteristics as they are defined above. By analyzing the Jaccard similarity of sets of links for a given node that belong to different layers of the multiplex graph, we are able to infer the consistency of the relationship intensity for different types of links. In other words, does the node have the tendency to have different types of relationship with the same set of friends in different environments (for example, online and offline environment in our study case). We argue that the out overlapping index shows the node activity in creating and maintaining different types of links, while the in overlapping index represents the node popularity, i.e. ability to gain different types of relationships.
\subsection{Exogenous characteristics}
Nodes in a graph may have a set of associated attributes, also called exogenous actor covariates. In order to emphasize the completeness of the approach, in this section we show that our unified framework based on the Jaccard similarity index can also be used for the analysis of the exogenous characteristics of the multiplex graph. The results for the exogenous characteristics will be presented elsewhere.
Let $A_i$ be the set of attributes associated to the node $i$. There are two basic characteristics for the actor $i$: the out-attributes characteristic, measuring whether actors with higher similarity-index values tend to nominate more friends and hence have a higher out-degree; the in-attributes characteristic, measuring whether actors with higher similarity-index values will tend to be nominated by many others and hence have higher in-degrees. These two characteristics are defined as
\begin{eqnarray}
att^{out}_i &=& \frac{1}{d_i^{out}} \sum_j x_{ij} J(A_i, A_j) \label{eq-eq} \\
att^{in}_i &=&\frac{1}{d_i^{in}} \sum_j x_{ji} J(A_i, A_j) \label{eq-al},
\end{eqnarray}
which will be compared to the average un-networked similarity value computed as
$$
att = \frac{2}{n(n-1)} \sum_{i,j; i<j} J(A_i, A_j).
$$
The characteristic $att^{out}_i$ measures similarity of attributes of the pair of end-nodes $i$ and $j$ for all the friends of $i$, that is, for all nodes in the set $S_i^{out}$. The characteristic $att^{in}_i$ measures similarity of attributes of the pair of end-nodes $i$ and $j$ for those $j$ that consider $i$ as a friend, that is, for all nodes in the set $S_i^{in}$. The quantities (\ref{eq-eq}) and (\ref{eq-al}) can be extended for multiplex networks as
\begin{eqnarray*}
att^{out}_i(G^\alpha) &=& \frac{1}{d_i^{out}} \sum_j x_{ij} (G^\alpha) J(A_i, A_j) \\
att^{in}_i (G^\alpha) &=&\frac{1}{d_i^{in}} \sum_j x_{ji} (G^\alpha) J(A_i, A_j)
\end{eqnarray*}
With this extension once can also analyze the similarity of the attributes for different pairs of nodes that are connected across different layers of the multiplex network. Therefore, these metrics provide insight into how different types of links influence the attribute based node similarity.
\section{Results}
\begin{figure}[h] \label{fig-1}
\begin{center}
\includegraphics[width=\textwidth]{Group2-4layers.pdf}
\caption{{\bf Visualization of a part of the multiplex graph constructed according to the data from Group 2 with 2 basic and 2 aggregated graphs. Note that all nodes (if present) are in the same positions across the layers.}}
\end{center}
\end{figure}
Let $V$ be the set of all students enrolled on a given course (group). For each group of students, there are four distinct directed graphs that have been generated based on the answers collected from the online survey: offline/online social network with strong ties and offline/online social network with weak ties, defined as:
\begin{eqnarray}
G^{of}_s & = & (V, E^{of}_{s}), \hspace{0.5cm} E^{of}_{s} = \{ i \to j \mbox{ is an offline strong tie} \} \label{sOFF} \\
G^{of}_w & = & (V, E^{of}_{w}), \hspace{0.5cm} E^{of}_{w} = \{ i \to j \mbox{ is an offline weak tie} \} \label{wOFF} \\
G^{on}_s & = & (V, E^{on}_{s}), \hspace{0.5cm} E^{on}_{s} = \{ i \to j \mbox{ is an online strong tie} \} \label{sON} \\
G^{on}_w & = & (V, E^{on}_{w}), \hspace{0.5cm} E^{on}_{w} = \{ i \to j \mbox{ is an online weak tie} \} \label{wON}
\end{eqnarray}
From these four (basic) graphs, five more aggregated graphs are constructed:
\begin{eqnarray}
G^{of} & = & (V, E^{of}), \hspace{0.5cm} E^{of} = E^{of}_{s} \cup E^{of}_{w} \label{OFF} \\
G^{on} & = & (V, E^{on}), \hspace{0.5cm} E^{on} = E^{on}_{s} \cup E^{on}_{w} \label{ON} \\
G_s & = & (V, E_s), \hspace{0.5cm} E_s = E^{on}_{s} \cup E^{of}_{s} \label{s} \\
G_w & = & (V, E_w), \hspace{0.5cm} E_w = E^{on}_{w} \cup E^{of}_{w} \label{w} \\
G & = & (V,E) \hspace{0.5cm} E = E^{of} \cup E^{on} = E_s \cup E_w \label{all}
\end{eqnarray}
Two multiplex networks are studied for two groups. Each network consists of nine layers or nine sets of ties: four basic - strong OFF, weak OFF, strong ON, and weak ON and five aggregated sets - strong, weak, ON, OFF, and all, as described with the given equations. In Fig. 1 a partial visual representation of the multiplex network for Group 2 is presented aiming to conceptualize the different types of links in each layer together with the possible ways for aggregation.
\begin{table}[!ht] \label{tab-1}
\caption{{\bf Basic graph characteristics for all layers in the multiplex graph that represents the social interrelations between the students from Group 1}}
\begin{tabular}{|l|l|l|l|l|l|l|l|l|}
\hline
Group 1 & { $|V|$}&{ $|E|$}&{ TotDegree}&{ Assor}&{ $|V_1|$}&{ $|E_1|$ }&{ Path }&{ Diam}\\ \hline
{ strong OFF}&{ 153}&{ 675}&{ 4.412}&{ 0.431}&{ 122}&{ 630}&{ 4.632}&{ 14}\\ \hline
{ weak OFF}&{ 153}&{ 756}&{ 4.941}&{ 0.018}&{ 135}&{ 693}&{ 3.856}&{ 9}\\ \hline
{ OFF}&{ 153}&{ 1420}&{ 9.281}&{ 0.238}&{ 150}&{ 1402}&{ 2.994}&{ 7}\\ \hline
{ strong ON}&{ 153}&{ 428}&{ 2.797}&{ 0.235}&{ 87}&{ 319}&{ 5.196}&{ 14}\\ \hline
{ weak ON}&{ 153}&{ 783}&{ 5.118}&{ 0.154}&{ 138}&{ 742}&{ 3.782}&{ 9}\\ \hline
{ ON}&{ 153}&{ 1201}&{ 7.850}&{ 0.253}&{ 150}&{ 1192}&{ 3.251}&{ 7}\\ \hline
{ strong}&{ 153}&{ 730}&{ 4.771}&{ 0.392}&{ 128}&{ 689}&{ 4.339}&{ 11}\\ \hline
{ weak}&{ 153}&{ 1046}&{ 6.837}&{ 0.087}&{ 145}&{ 1016}&{ 3.212}&{ 7}\\ \hline
{ all}&{ 153}&{ 1487}&{ 9.719}&{ 0.195}&{ 153}&{ 1487}&{ 2.937}&{ 7}\\ \hline
\end{tabular}
\end{table}
\begin{table}[!ht] \label{tab-2}
\caption{{\bf Basic graph characteristics for all layers in the multiplex graph that represents the social interrelations between the students from Group 2}}
\begin{tabular}{|l|l|l|l|l|l|l|l|l|}
\hline
{ Group 2 }&{ $|V|$}&{ $|E|$}&{ TotDegree}&{ Assor}&{ $|V_1|$ }&{ $|E_1|$ }&{ Path }&{ Diam}\\ \hline
{ strong OFF}&{ 80}&{ 412}&{ 5.150}&{ 0.184}&{ 64}&{ 390}&{ 3.337}&{ 9}\\ \hline
{ weak OFF}&{ 80}&{ 521}&{ 6.513}&{ 0.297}&{ 57}&{ 420}&{ 2.697}&{ 5}\\ \hline
{ OFF}&{ 80}&{ 930}&{ 11.625}&{ 0.360}&{ 71}&{ 892}&{ 2.282}&{ 5}\\ \hline
{ strong ON}&{ 80}&{ 226}&{ 2.825}&{ 0.199}&{ 49}&{ 188}&{ 4.468}&{ 12}\\ \hline
{ weak ON}&{ 80}&{ 465}&{ 5.812}&{ 0.352}&{ 62}&{ 414}&{ 2.922}&{ 7}\\ \hline
{ ON}&{ 80}&{ 690}&{ 8.625}&{ 0.356}&{ 68}&{ 651}&{ 2.544}&{ 5}\\ \hline
{ strong}&{ 80}&{ 455}&{ 5.688}&{ 0.213}&{ 65}&{ 430}&{ 3.157}&{ 9}\\ \hline
{ weak}&{ 80}&{ 728}&{ 9.100}&{ 0.352}&{ 67}&{ 675}&{ 2.463}&{ 5}\\ \hline
{ all}&{ 80}&{ 1013}&{ 12.662}&{ 0.384}&{ 73}&{ 985}&{ 2.243}&{ 5}\\ \hline
\end{tabular}
\end{table}
Multiplex networks are particularly significant when they overlap and interact to create phenomena or processes that cannot be explained by a single network alone. Tables 1 and 2 summarize the basic graph characteristics for all graphs generated in the study (total of 9 graphs per group): number of actors $|V|$, number of ties $|E|$, average degree and assortativity. Recall, a directed graph is strongly connected if there is a directed path from each vertex to every other vertex. The strongly connected components (SCC) of a directed graph are its maximal strongly connected sub-graphs. The number of actors $|V_1|$ and ties $|E_1|$ for the largest SCCs for each of these graphs are also presented in the tables, including the average path length and the diameter.
Since the number of participants in the two groups (153 versus 80) is different, it is significant to confirm that the same conclusions (especially concerning the ratios) hold for both groups. Namely, the number of strong ties is smaller than the number of weak ties: 689 versus 1016 (40\% strong ties and 60\% weak ties) for the first group and 430 versus 685 (41\% strong ties and 59\% weak ties) for the second group. However, if we take a closer look at the strong ties graphs, one can notice that there are more strong offline ties than strong online ties. This leads us to the conclusion that the students have closer friendship relations in the offline real rather than the online virtual environment. The number of ON ties is smaller than the number of OFF ties: 1192 versus 1402 (46\% ON ties and 54\% OFF ties) for Group 1 and 651 versus 892 (42\% ON ties and 58\% OFF ties) for Group 2. Moreover, the number of strong OFF ties is almost the same as the number of weak OFF ties for both groups: 630 and 693 (48\% and 52\%) for the first group and 390 and 420 (48\% and 52\%) for the second group. This suggests that weak communications are (almost) equally presented in the online and real life communication. However, the number of strong ON ties is almost half the number of weak ON ties: 319 and 742 (30\% and 70\%) for the first group and 188 and 414 (31\% and 69\%) for the second group. This could be interpreted as the fact that students within one group interact with all colleagues no matter whether they consider them close or distant which could be due to the necessities of working together on different projects, homework or labs for example. However, looking at the number of online strong and online weak ties, we can infer that the students use the virtual world to spread weak and more common friendships most probably aiming to expand their circle of acquaintances.
Comparing the average total degrees between basic graphs and aggregated graphs in the multiplex network it is fairly straightforward to conclude that both groups exhibit similar patterns for the average degrees: strong OFF and weak OFF graphs have average degrees 4.4 and 4.9 for the first group and 5.1 and 6.5 for the second group, respectively. Strong ON and weak ON graphs have average degrees 2.8 and 5.1 for the first group and 2.8 and 5.8 for the second group, respectively. At the aggregated level, the average degrees for strong and weak graphs are 4.8 and 6.8 for the first group and 5.7 and 9.1 for the second group, while for the OFF and ON graphs these numbers are 9.3 and 7.8 for the first group and 11.6 and 8.6 for the second group. Note that the average degrees for all graphs in the multiplex network of the second group are greater than the average degrees for the corresponding graphs of the first group. One possible explanation is that the students which are part of smaller group are more friendly and associative between them. This is especially the case in this scenario since the the students belonging to Group 2 have a more diverse background (i.e. type of study program, year of study and alike) compared to Group 1. However, we do not have more data to confirm (or disconfirm) this conclusion (hypothesis).
Tables 3 and 4 provide summaries of the basic metrics: average values for the reciprocities, three cycles, and triplets for all (single) graphs in the multiplex networks that represent group 1 and 2, respectively. All weak graphs (weak OFF, weak ON, and weak) have smaller values for reciprocity, three cycles, and triplets for both groups indicating that weak ties are less socially significant. Both transitive triplets and three-cycles represent closed structures, however, triplets indicate hierarchical ordering in contrast to three-cycles which are against of such ordering. For all graphs studied here, the average values of transitive triplets is slightly larger than the average values of three-cycles showing that the elements of hierarchical ordering are present in these social networks. Also, the results given in both tables 3 and 4 indicate that there is a correlation between dyads (reciprocity) and triads (transitive triplet and three-cycle) such that larger (smaller) values of the former imply larger (smaller) values of the latter.
\begin{table}[!ht]
\caption{{\bf Basic endogenous characteristics for all layers in the multiplex network that represents social interactions in Group 1}}
\begin{tabular}{|l|l|l|l|}
\hline
{\bf Group 1 }& Average $r_i$ & Average $tc_i$ & Average $tp_i$ \\ \hline
{ strong OFF}&{ 0.445}&{ 0.153}&{ 0.168}\\ \hline
{ weak OFF}&{ 0.154}&{ 0.029}&{ 0.039}\\ \hline
{ OFF}&{ 0.467}&{ 0.147}&{ 0.190}\\ \hline
{ strong ON}&{ 0.465}&{ 0.141}&{ 0.136}\\ \hline
{ weak ON}&{ 0.204}&{ 0.045}&{ 0.049}\\ \hline
{ ON}&{ 0.442}&{ 0.133}&{ 0.173}\\ \hline
{ strong}&{ 0.482}&{ 0.159}&{ 0.177}\\ \hline
{ weak}&{ 0.278}&{ 0.061}&{ 0.074}\\ \hline
{ all}&{ 0.492}&{ 0.155}&{ 0.200}\\ \hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\caption{{\bf Basic endogenous characteristics for all layers in the multiplex network that represents social interactions in Group 2}}
\begin{tabular}{|l|l|l|l|}
\hline
{Group 2 }& Average $r_i$ & Average $tc_i$ & Average $tp_i$ \\ \hline
{ strong OFF}&{ 0.425}&{ 0.156}&{ 0.161}\\ \hline
{ weak OFF}&{ 0.074}&{ 0.041}&{ 0.049}\\ \hline
{ OFF}&{ 0.335}&{ 0.139}&{ 0.200}\\ \hline
{ strong ON}&{ 0.448}&{ 0.146}&{ 0.149}\\ \hline
{ weak ON}&{ 0.099}&{ 0.049}&{ 0.064}\\ \hline
{ ON}&{ 0.342}&{ 0.115}&{ 0.172}\\ \hline
{ strong}&{ 0.509}&{ 0.150}&{ 0.171}\\ \hline
{ weak}&{ 0.148}&{ 0.081}&{ 0.102}\\ \hline
{ all}&{ 0.360}&{ 0.153}&{ 0.216}\\ \hline
\end{tabular}
\end{table}
Tables 5 and 6 show the average values of 5 different additional characteristics that are focusing on the interrelationship of different parts of the multiplex graph. Here, in addition to reciprocity, three-cycle, and triplet, we also have the two overlapping indexes, all as they are defined in the subsection Multiplex graph characteristics. The given results indicate that reciprocity is preserved across different layers of the multiplex network; in particular the ties in strong OFF are reciprocal with ties in strong ON and vice versa (the normalized average values are $r_i(G_s^{OFF}, G_s^{ON}) = 0.417$ and $r _i(G_s^{ON}, G_s^{OFF}) = 0.406$, respectively. Similar values are also obtained for the pairs (OFF, ON) and (ON, OFF). However, for the considered social multiplex network, the triads (measured with normalized three cycles and triplets) are not significant. On the other hand, the values for the overlapping indexes show that activity and popularity patterns among some layers of the multiplex network are significant. For instance, the number of out-degree and in-degree friends in the strong OFF layer coincides with the out-degree and in-degree friends in the strong ON layer. Or, out-degree and in-degree friends in the weak OFF layer are also out-degree and in-degree friends in the weak ON layer. On the other hand, the overlapping indexes for both out-degree (activity) and in-degree (popularity) are small for the following combinations of two graphs (strong ON, weak ON) and (strong OFF, weak, OFF) for both Groups 1 and 2.
\begin{table}[!ht]
\caption{{\bf Multiplex-graph characteristics for different pairs of layers reflecting the combined types of ties for Group 1}}
\begin{tabular}{|l|l|l|l|l|l|}
\hline
{\bf Group 1}&{\bf Reciprocity}&{\bf tc2}&{\bf tp2} & oi1 & oi2 \\ \hline
{ strong OFF, strong ON}&{ 0.417}&{ 0.136}&{ 0.148} &{ 0.568}&{ 0.524} \\ \hline
{ weak OFF, weak ON}&{ 0.160}&{ 0.032}&{ 0.037} &{ 0.427}&{ 0.512} \\ \hline
{ strong OFF, weak OFF}&{ 0.091}&{ 0.061}&{ 0.065} &{ 0.025}&{ 0.005} \\ \hline
{ strong ON, weak ON}&{ 0.081}&{ 0.052}&{ 0.059} &{ 0.028}&{ 0.006} \\ \hline
{ strong, weak}&{ 0.175}&{ 0.093}&{ 0.102} &{ 0.195}&{ 0.180} \\ \hline
{ OFF, ON}&{ 0.434}&{ 0.141}&{ 0.182} &{ 0.735}&{ 0.775} \\ \hline
{ strong ON, strong OFF}&{ 0.406}&{ 0.145}&{ 0.136} &{ 0.568}&{ 0.524} \\ \hline
{ weak ON, weak OFF}&{ 0.165}&{ 0.038}&{ 0.042} &{ 0.427}&{ 0.512} \\ \hline
{ weak OFF, strong OFF}&{ 0.116}&{ 0.040}&{ 0.061} &{ 0.025}&{ 0.005} \\ \hline
{ weak ON, strong ON}&{ 0.098}&{ 0.046}&{ 0.064} &{ 0.028}&{ 0.006} \\ \hline
{ weak, strong}&{ 0.207}&{ 0.075}&{ 0.098} &{ 0.195}&{ 0.180} \\ \hline
{ ON, OFF}&{ 0.433}&{ 0.139}&{ 0.179} &{ 0.735}&{ 0.775} \\ \hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\caption{{\bf Multiplex-graph characteristics for different pairs of layers reflecting the combined types of ties for Group 2}}
\begin{tabular}{|l|l|l|l|l|l|}
\hline
{\bf Group 2}&{\bf Reciprocity}&{\bf tc2}&{\bf tp2} & oi1 & oi2 \\ \hline
{ strong OFF, strong ON}&{ 0.411}&{ 0.127}&{ 0.138} &{ 0.517}&{ 0.470} \\ \hline
{ weak OFF, weak ON}&{ 0.072}&{ 0.038}&{ 0.057} &{ 0.400}&{ 0.401} \\ \hline
{ strong OFF, weak OFF}&{ 0.071}&{ 0.054}&{ 0.056} &{ 0.093}&{ 0.003} \\ \hline
{ strong ON, weak ON}&{ 0.062}&{ 0.042}&{ 0.047} &{ 0.102}&{ 0.001} \\ \hline
{ strong, weak}&{ 0.132}&{ 0.086}&{ 0.095} &{ 0.269}&{ 0.141} \\ \hline
{ OFF, ON}&{ 0.334}&{ 0.117}&{ 0.181} &{ 0.708}&{ 0.604} \\ \hline
{ strong ON, strong OFF}&{ 0.362}&{ 0.130}&{ 0.138} &{ 0.517}&{ 0.470} \\ \hline
{ weak ON, weak OFF}&{ 0.067}&{ 0.040}&{ 0.047} &{ 0.400}&{ 0.401} \\ \hline
{ weak OFF, strong OFF}&{ 0.110}&{ 0.041}&{ 0.070} &{ 0.093}&{ 0.003} \\ \hline
{ weak ON, strong ON}&{ 0.135}&{ 0.039}&{ 0.065} &{ 0.102}&{ 0.001} \\ \hline
{ weak, strong}&{ 0.209}&{ 0.073}&{ 0.103} &{ 0.269}&{ 0.141} \\ \hline
{ ON, OFF}&{ 0.306}&{ 0.118}&{ 0.180} & { 0.708}&{ 0.604} \\ \hline
\end{tabular}
\end{table}
In social science structural equivalence is defined as ``two nodes are considered structurally equivalent if they share many of the same network neighbors.'' A possible operationalization of this definition could be done as follows: two nodes $i$ and $j$ are structurally equivalent if $r_i = r_j$, $tc_i = tc_j$ and $tp_i = tp_j$. We found that in each network some of the actors are structurally equivalent. For example in the strong ON graph for Group 1, the actors with id 2, 24, and 127 are structurally equivalent having $r_i=1$, $tc_i = 0.55$, and $tp_i = 0.532$. These actors have out-degree and in-degree $d_i^{out} = 4$ and $d_i^{in} = 4$, respectively. Table 7 shows characteristics for those actors in the strong OFF graph that have $d_i^{out} = 4$ and $d_i^{in} = 4$. Again some of those actors are structurally equivalent (2 and 127). These initial results are very promising and open up a way to formally mathematically define the concept of structural equivalence. However, further deeper study on this matter must be carefully conducted in order to confirm the viability of our proposed method. For instance, if we relax the condition to approximately equal then nodes 80 and 83 from Table 7 will also be considered equivalent.
Finally, we have also tested and confirmed the hypothesis of Granovetter. Consider two arbitrary selected individuals A and B and the set of all persons with ties to either or both of them. The hypothesis is: the stronger the tie between A and B, the larger the proportion of individuals in S to whom they will be both tied (connected by a weak or strong tie). For Group 1, there are a total of 5526 strong wedges out of which: 39.052 \% are closed by another strong link, 33.406 \% are closed by a weak link and 72.457 \% are closed by any link. Also, there are a total of 29860 weak wedges out of which: 8.279 \% are closed by a strong link, 9.752 \% are closed by another weak link and 18.031 \% are closed by any link. For Group 2, there are a total of 4811 strong wedges out of which: 39.243 \% are closed by another strong link, 45.022 \% are closed by a weak link and 84.265 \% are closed by any link. Also, there are a total of 25424 weak wedges out of which: 14.301 \% are closed by a strong link, 23.450 \% are closed by another weak link and 37.752 \% are closed by any link.
\begin{table}[!ht]
\caption{{\bf Structural equivalence. Graph characteristics of the nodes in Group 1 with $d_i^{out} = 4$ and $d_i^{in} = 4$ }}
\begin{tabular}{|l|l|l|l|}
\hline
{\bf strong OFF}& $r_i$ & $tc_i$ & $tp_i$ \\ \hline
$i=2$ & 1 & 0.424 & 0.442 \\ \hline
$i=40$ & 0.333 & 0.042 & 0.094 \\ \hline
$i=80$ & 0.6 & 0.343 & 0.353 \\ \hline
$i=83$ & 0.6 & 0.356 & 0.287 \\ \hline
$i=94$ & 0.6 & 0.053 & 0.056 \\ \hline
$i=122$ & 1 & 0.376 & 0.366 \\ \hline
$i=127$ & 1 & 0.424 & 0.442 \\ \hline
$i=148$ & 0.143 & 0.083 & 0.191 \\ \hline
\end{tabular}
\end{table}
\section{Conclusions}
By studying friendship relations among students enrolled on two different courses represented using a multiplex structure, a number of interesting conclusions regarding the strength of online and offline ties can be drawn: (1) strong ties are preferred in face-to-face (offline) communications; (2) weak ties are equally presented in online and offline communications; (3) in offline communication, strong and weak ties are (almost) equally included; (4) in online communication weak ties are dominant; (5) weak ties (in three layers of the multiplex network: weak offline, weak online, and weak) are much less reciprocal than strong ties; (6) dyads (that is, reciprocities) are preserved, however triads (measured with normalized three cycles and triplets) are not significant in different layers of a multiplex network; (7) activity and popularity patterns for some layers of the multiplex network are significant: out-degree and in-degree friends in one layer could also be out-degree and in-degree friends in another layer. These conclusions are supported by the data obtained from both groups. The number and consistency of the drawn conclusions have confirmed that by approaching the problem of different interrelationships between actors as a multiplex network problem, one can gain useful insight on the importance of each type of link within the social network as a whole, as well as on the interaction and overlapping between different link types, especially in the offline/online (real/virtual) environments as was our case of study.
In the future, we plan to study how exogenous characteristics influence the structure of the multiplex network and how the network structure dominates the actors (students) characteristics and their temporal evolution. In particular, we will address questions on: how student grades are distributed; students' commitment and progress to the studies; how one could empower the students to obtain better grades or determine their specific areas of interest with greater success. Are the students with better grades and habits more central to the network or not, and how changing the placement of these individuals could influence the structure of the network?
|
\section{Introduction}
Roughly speaking, a \emph{framed flow category} is a way of encoding the flow data associated to a Morse--Smale vector field or, more generally, the flow data of a Floer functional. Objects of the category correspond to critical points while the morphism set between any two objects has the structure of a framed manifold-with-corners corresponding to the space of flowlines between the objects.
A framed flow category gives rise to a CW-complex with one cell for each object of the category, and the attaching maps determined by the morphism spaces. When the framed flow category arises from a Floer functional, the (shifted) cohomology of the CW-complex agrees with the corresponding Floer cohomology. The idea is that the whole CW-complex (up to suspension and desuspension) should be considered as the Floer invariant, instead of just its cohomology. In other words, the Floer invariant should take the form of a \emph{stable homotopy type}.
In this paper we construct moves that change a framed flow category without changing the associated stable homotopy type. These are inspired by moves that can be performed in the Morse--Smale situation without altering the underlying smooth manifold. We posit that if two framed flow categories represent the same stable homotopy type then a finite sequence of these moves is sufficient to connect the two categories. This is directed towards the goal of reducing the study of framed flow categories (and their associated stable homotopy types) to combinatorics.
We show in Theorem \ref{thm:SNF} that after a finite sequence of our moves, we can arrange for the associated cochain complex to the framed flow category to be in primary Smith normal form. This immediately gives Corollary \ref{cor:moore}. This corollary implies, for example, that if two framed flow categories each represent the same Moore space, then the categories are indeed connected by a finite sequence of moves.
In a later paper we shall go further, developing the combinatorics of the moves and proving results based now on the classifications due to Chang and Baues--Hennes of stable homotopy types of small cohomological width.
\subsection{Plan of the paper}
\label{subsec:plan}
In Section \ref{sec:ffcs}, we give a condensed recap of the necessary definitions of a framed flow category. This is based on the formulation by Lipshitz and Sarkar \cite{MR3230817}, and is intended as a quick-reference guide for the rest of the paper. In Section \ref{sec:handleslide} we introduce an analogue for framed flow categories of the operation of \emph{handle sliding} in a handlebody decomposition of a smooth manifold. In Section \ref{sec:whitney} we similarly introduce an analogue of the \emph{Whitney trick}.
Handle sliding, the extended Whitney trick, and handle cancellation (already introduced in \cite{JLSmorse}) are each moves that can performed on framed flow categories while preserving the associated stable homotopy type. We give an example of their application in Section \ref{sec:tref}. Recently, Lipshitz and Sarkar \cite{MR3230817} have constructed a framed flow category from the input of a knot diagram. This gives a stable homotopy type whose cohomology recovers Khovanov cohomology (a knot invariant intimately connected with Floer cohomology). Our example is drawn from a particular quantum degree ($q = 21$) of the Khovanov stable homotopy type of the disjoint union of three right-handed trefoils.
Finally, in Section \ref{sec:smith} we prove general results about simplification of framed flow categories. Essentially, these results can be taken as saying that each framed flow category is move-equivalent to a framed flow category realizing the simplest possible cochain complex giving the correct cohomology.
\section{Framed flow categories}
\label{sec:ffcs}
The definition of a flow category was originally given by Cohen, Jones and Segal \cite{MR1338869} but the version we are working with is based on the definition by Lipshitz and Sarkar \cite{MR3230817}; the equivalence of the two approaches is shown in \cite[Prop.\ 3.27]{MR3230817}.
Before we can give the definition of a flow category, we will need a sharpening of the concept of a smooth manifold with corners. This refinement was originally defined by J{\"a}nich \cite{MR0226674} and was further devloped by Laures \cite{MR1781277}. We refer the reader to these sources for a fuller treatment.
\begin{definition}Let $M$ be a smooth $m$--dimensional manifold with corners. For each $p\in M$ there is an integer $l(p)$ such that an open neighbourhood of $p$ is homeomorphic to $(\mathbb{R}_+)^l\times \mathbb{R}^{m-l}$, with $p$ corresponding to $0\in (\mathbb{R}_+)^l\times \mathbb{R}^{m-l}$. We say $p$ is a \emph{corner point of codimension $l$}. The closure of any connected component of the set $\{p\in M\,|\,l(p)=1\}$ is called a \emph{face} of $M$, and the union of the faces is the \emph{boundary} $\partial M $ of $M$.
Now suppose each $p\in M$ belongs to precisely $l(p)$ faces. In this case, a \emph{smooth $m$--dimensional $\langle n\rangle$--manifold} is a manifold with corners $M$, together with an ordered $n$--tuple of faces $(\partial_1M,\dots,\partial_nM)$ which satisfy the following conditions:
\begin{itemize}
\item $\partial_1M\cup\dots\cup\partial_nM=\partial M$,
\item $\partial_iM\cap\partial_jM$ is a face of $\partial_iM$ and $\partial_jM$ for all $i\neq j$.
\end{itemize}In which case, for any non-empty subset $J\subset \{1,\dots,n\}$, we will write\[\partial_JM=\bigcap_{j\in J}\partial_jM.\]
\end{definition}
Now we introduce some more notation. Let $n$ be a non-negative integer and let $\textbf{d}=(d_0,\dots,d_n)$ be an $(n+1)$--tuple of non-negative integers. Write $\mathbb{R}_+=[0,\infty)$ and \[\begin{array}{rclll}\mathbb{E}^\textbf{d}&=&\mathbb{R}^{d_0}\times\mathbb{R}_+\times \mathbb{R}^{d_1}\times\mathbb{R}_+\times\dots\times\mathbb{R}_+\times\mathbb{R}^{d_n},&&\\ \mathbb{E}^\textbf{d}[a:b]&=&\mathbb{E}^{(d_a,\dots,d_{b-1})},&&\text{for $0\leq a<b\leq n+1$.}\end{array}\]We introduce a standard subdivision of the boundary of $\mathbb{E}^\textbf{d}$ into the following faces:
\[\partial_i\mathbb{E}^\textbf{d}=\mathbb{R}^{d_0}\times\mathbb{R}_+\times\dots\mathbb{R}^{d_{i-1}}\times\{0\}\times\mathbb{R}^{d_i}\times\dots\times \mathbb{R}_+\times\mathbb{R}^{d_n},\]for $i=1,\dots,n$, so that $\mathbb{E}^\textbf{d}$ is a $(d_0+\dots+d_n+n)$--dimensional $\langle n\rangle$--manifold.
\begin{definition}\label{def:neat}A \emph{neat immersion (embedding)} of a smooth $m$--dimensional $\langle n\rangle$--manifold $M$ is a smooth immersion (embedding) $\iota:M \looparrowright \mathbb{E}^\textbf{d}$ such that for all $l$, corner points of codimension $l$ in $M$ are sent to corner points of codimension $l$ in $\mathbb{E}^\textbf{d}$, and for any $J\subset I\subset \{1,\dots n\}$ the intersection $\partial_IM\cap\partial_J\mathbb{E}^\textbf{d}=\partial_J M$ is perpendicular.
\end{definition}
Any compact $m$--dimensional $\langle n\rangle$--manifold admits a neat embedding with respect to some $\mathbb{E}^\textbf{d}$ (see Laures \cite[2.1.7]{MR1781277}). Moreover, a neat immersion can always be improved to a neat embedding by perturbation, possibly after increasing the integers in the tuple $\textbf{d}$.
\begin{definition}A \em flow category \em $\mathscr{C}$ is a category with finitely many objects, and equipped with a function $\gr{\cdot}\colon \text{Ob}(\mathscr{C}) \to \mathbb{Z}$, called the \em grading\em, satisfying the following:
\begin{enumerate}
\item $\text{Hom}(x,x)=\{\text{id}\}$ for all $x\in\text{Ob}(\mathscr{C})$, and for $x\neq y\in\text{Ob}(\mathscr{C})$, $\text{Hom}(x,y)$ is a smooth, compact $(|x|-|y|-1)$--dimensional $\langle |x|-|y|-1\rangle$--manifold which we denote by $\mathcal{M}(x,y)$.
\item For $x,y,z\in \text{Ob}(\mathscr{C})$ with $\gr{z}-\gr{y}=m$, the composition map restricts to
\[
\circ\colon \mathcal{M}(z,y) \times \mathcal{M}(x,z) \to \mathcal{M}(x,y),
\]
which is an embedding into $\partial_m\mathcal{M}(x,y)$. Furthermore,
\[
\circ^{-1}(\partial_i \mathcal{M}(x,y)) = \left\{ \begin{array}{lr}
\partial_i \mathcal{M}(z,y)\times \mathcal{M}(x,z) & \mbox{for }i<m \\
\mathcal{M}(z,y)\times \partial_{i-m}\mathcal{M}(x,z) & \mbox{for }i>m
\end{array}
\right.
\]
\item For $x\not= y\in \text{Ob}(\mathscr{C})$, $\circ$ induces a diffeomorphism
\[
\partial_i\mathcal{M}(x,y) \cong \coprod_{z,\,\gr{z}=\gr{y}+i} \mathcal{M}(z,y) \times \mathcal{M}(x,z).
\]
\end{enumerate}
The manifold $\mathcal{M}(x,y)$ is called the \em moduli space from $x$ to $y$\em, and we also set $\mathcal{M}(x,x)=\emptyset$.
\end{definition}
\begin{definition}
Let $\mathscr{C}$ be a flow category and write \[k=\min\{|x|\,:\,x\in\text{Ob}(\mathscr{C})\},\quad n=\max\{|x|\,:\,x\in\text{Ob}(\mathscr{C})\}-k.\] Suppose $\textbf{d}=(d_k,\dots,d_{n+k})$ is an $(n+1)$--tuple of non-negative integers. A \emph{neat immersion (resp.\ neat embedding)} $\iota$ of a flow category $\mathscr{C}$ relative to $\textbf{d}$ is a collection of neat immersions (resp.\ neat embeddings) $\iota_{x,y}:\mathcal{M}(x,y)\looparrowright\mathbb{E}^\textbf{d}[|y|:|x|]$ such that for all objects $x,y,z$ and points $(p,q)\in \mathcal{M}(x,z)\times\mathcal{M}(z,y)$,\[\iota_{x,y}(q\circ p)=(\iota_{z,y}(q),0,\iota_{x,z}(p)).\]
If $\iota$ is a neat immersion of a flow category $\mathscr{C}$ relative to $\textbf{d}$, then a \emph{framing} $\phi$ of $(\mathscr{C},\iota)$ is a collection of immersions\[\phi_{x,y}:\mathcal{M}(x,y)\times[-\epsilon,\epsilon]^A\looparrowright \mathbb{E}^\textbf{d}[|y|:|x|],\qquad A=d_{|y|}+\dots+d_{|x|-1},\]extending $\iota_{x,y}$, such that for all objects $x,y,z$ and points $(p,q)\in \mathcal{M}(x,z)\times\mathcal{M}(z,y)$,\[\phi_{x,y}(q\circ p,t_1,t_2,\dots, t_A)=(\phi_{z,y}(q,t_1,t_2,\dots,t_B),0,\phi_{x,z}(p,t_{B+1},\dots,t_A)),\]where $B= d_{|y|}+\dots+d_{|z|-1}$.
\end{definition}
Note that a (framed) neat immersion of a flow category relative to some $\textbf{d}$ can be improved to a (framed) neat embedding relative to some $\textbf{d}'\geq\textbf{d}$ by perturbation.
\begin{definition} A \emph{framed flow category} is a triple $(\mathscr{C},\iota,\phi)$, where $(\mathscr{C},\iota)$ is a flow category with neat immersion, relative some $\textbf{d}$, and $\phi$ is a framing for $(\mathscr{C},\iota)$.
\end{definition}
One can associate to a framed flow category $(\mathscr{C}, \iota,\phi)$ a geometric realisation as a finite cell complex $|\mathscr{C}|_{\iota,\phi}=|\mathscr{C}|$ (see \cite[Definition 3.23]{MR3230817} for precise details of the version we are using). The construction of $|\mathscr{C}|$ requires several auxilliary choices, one such being an integer $C$ such that $C>|x|$ for all $x\in\text{Ob}(\mathscr{C})$. To construct $|\mathscr{C}|$, one takes for each object $x$ of $\mathscr{C}$ a $(C-m)$--dimensional cell $\mathcal{C}(x)$, where $|x|=m$. These cells are glued together using the data of the framed moduli spaces (essentially via the Thom construction).
The homotopy type of $|\mathscr{C}|$ depends on the framed embedding $(\iota,\phi)$, but modifying $(\iota,\phi)$ by an isotopy results in a homeomorphic cell complex (see \cite[Lemma 3.25]{MR3230817}). Modifying the other choices involved in constructing $|\mathscr{C}|$, including possibly increasing $\textbf{d}$ to $\textbf{d}'\geq\textbf{d}$, will result, up to homotopy, in some precise number of reduced suspensions of $|\mathscr{C}|$ (see \cite{MR3230817}). The following is a well-defined topological object arising from a framed flow category (and independent of choice of $\textbf{d}$).
\begin{definition}For a framed flow category $(\mathscr{C}, \iota, \phi)$ and any object $x$ of $\mathscr{C}$, let $C$ be the difference between the dimension of the associated cell $\mathcal{C}(x)\subset |\mathscr{C}|_{\iota,\phi}$ and the grading of $x$. The \emph{stable homotopy type of $(\mathscr{C}, \iota, \phi)$} is the desuspended suspension spectrum $\mathcal{X}(\mathscr{C}):=\Sigma^{-C}\Sigma^\infty|\mathscr{C}|_{\iota,\phi}$, considered up to homotopy of spectra.
\end{definition}
\section{Handle slides in framed flow categories}
\label{sec:handleslide}
In this section, we will define the framed flow category analogue for the process in handlebody theory of sliding one $i$--handle over a second $i$--handle. This results in a new framed flow category giving rise to the same stable homotopy type.
\begin{theorem}
\label{thm:handleslide}
If $(\mathscr{C}_S,\iota_S,\phi_S)$ is the result of a \emph{handle slide} in $(\mathscr{C}, \iota, \phi)$, then there is a homotopy equivalence
\[\mathcal{X}(\mathscr{C}_S) \simeq \mathcal{X}(\mathscr{C}) {\rm .}\]
\end{theorem}
We first describe a method for handle sliding, on which we base our flow category analogue. Given a handlebody $H$, which has $i$--handles $h_1$ and $h_2$, form a new handlebody $H'$ by attaching an $i$--handle $h_3$ and a cancelling $(i+1)$--handle $g$, attached as follows. The attaching sphere for $h_3$ will be glued along the sphere embedded in $H$ formed by taking an embedded connected sum of slightly pushed-off copies of the attaching spheres of $h_1$ and $h_2$. The attaching sphere of $g$ will be glued to intersect the belt spheres of each of our three $i$--handles precisely once positively each (and with no other $i$--handle interactions).
There are then exactly three $i$--handles against which $g$ can be cancelled. The handlebodies which are the effects of cancelling $g$ against $h_1$, $h_2$ and $h_3$ are homeomorphic to, respectively, the effect of $h_1$ slid over $h_2$, the effect of $h_2$ slid over $h_1$, and the original handlebody $H$. The reason for describing handle sliding with such an emphasis on handle cancellation is that it will allow us to build on the flow category handle cancellation technique described in the earlier paper \cite{JLSmorse}.
Let $\mathscr{C}$ denote a framed flow category $(\mathscr{C}, \iota, \phi)$ with respect to an $(n+1)$--tuple $\textbf{d}=(d_k,\dots,d_{n+k})$ of non-negative integers. Suppose $x,y$ are objects of $\mathscr{C}$ with grading $i$. To mimic the geometric process above, we will introduce a pair of cancelling objects $e$ and $f$ in degrees $i$ and $i+1$ respectively. Morphisms will be introduced so that cancelling $f$ against $e$ returns $\mathscr{C}$, and cancelling $f$ against $x$ returns a new category $\mathscr{C}_S$ which we will call the \emph{effect of sliding $x$ over $y$}. It is proved in \cite[Theorem 2.17]{JLSmorse} that handle cancellation in flow categories preserves the stable homotopy type of a framed flow category, and hence the stable homotopy types $\mathcal{X}(\mathscr{C})$ and $\mathcal{X}(\mathscr{C}_S)$ will agree.
The main task of this section therefore will be to correctly describe, embed, and frame the \emph{intermediate flow category} obtained by introducing $e$ and $f$. This is carried out in Subsection \ref{subsec:inter}, with Theorem \ref{thm:handleslide} deduced in Subsection \ref{subsec:sliding}. We end with a discussion of how handle sliding affects the framings of $1$-dimensional moduli spaces in Subsection \ref{subsec:framings}.
\subsection{The intermediate flow category}
\label{subsec:inter}
We begin with a small lemma that we will use to embed and frame moduli spaces later. Suppose for now that $M^m$ is a smooth $m$--dimensional $\langle n\rangle$--manifold. Then $M\times[0,1]$ has the structure of a smooth $(m+1)$--dimensional $\langle n+1\rangle$--manifold, given by setting $\partial_{n+1}(M\times[0,1])=M\times\{0,1\}$ and $\partial_i(M\times[0,1])=\partial_iM\times[0,1]$ otherwise.
\begin{lemma}\label{lem:arc}Suppose there is a neat immersion $\iota:M\looparrowright \mathbb{E}^{(d_k,\dots,d_{n+k})}$. For real numbers $u>0$, $v\geq 0$, let \[\gamma_{u,v}(t)=(s(t),2t-1)\in \mathbb{R}_+\times\mathbb{R}\] be such that $(t-u)^2+s^2=v^2$. Then the function\[\iota':M\times[0,1]\looparrowright \mathbb{E}^{(d_k,\dots,d_{n+k})}\times \mathbb{R}_+\times\mathbb{R};\qquad \iota'(p,t):=(\iota(p),\gamma_{u,v}(t))\] is a neat immersion.
Denote by $R:\mathbb{E}^\textbf{d}\to \mathbb{E}^\textbf{d}$ the reflection in the plane through the origin which is perpendicular to the first coordinate axis of $\mathbb{R}^{d_{n+k}}\subset \mathbb{E}^\textbf{d}$. Suppose $\iota$ extends to a neat immersion $\phi:M\times[-\epsilon,\epsilon]^A\looparrowright \mathbb{E}^\textbf{d}$, with $A=d_k+\dots+d_{n+k}$, then $-\phi:=R\circ\phi$ determines another possible framing for $\iota$. Define two possible boundary framings for $\iota'|_{M\times\{0,1\}}$by\begin{enumerate}
\item \begin{eqnarray*}\phi_1:M\times\{0,1\}\times[-\epsilon,\epsilon]^{A}\times[-\epsilon,\epsilon]&\looparrowright& \mathbb{E}^\textbf{d}\times\mathbb{R}_+\times\mathbb{R}\\
\quad \phi_1(p,0,\textbf{q},r)&=&(\phi(p,\textbf{q}),0, r-1)\\
\phi_1(p,1,\textbf{q},r)&=&(\phi(p,\textbf{q}),0,-r+1)\end{eqnarray*}
\item \begin{eqnarray*}\phi_2:M\times\{0,1\}\times[-\epsilon,\epsilon]^{A}\times[-\epsilon,\epsilon]&\looparrowright& \mathbb{E}^\textbf{d}\times\mathbb{R}_+\times\mathbb{R}\\
\quad \phi_2(p,0,\textbf{q},r)&=&(\phi(p,\textbf{q}),0, -r-1)\\
\phi_2(p,1,\textbf{q},r)&=&(-\phi(p,\textbf{q}),0,-r+1)\end{eqnarray*}
\end{enumerate}
Then $\phi_1$ and $\phi_2$ respectively extend to framings for the whole of $\iota'$:\[\phi''_1,\phi_2'':M\times[0,1]\times[-\epsilon,\epsilon]^{A+1}\looparrowright \mathbb{E}^\textbf{d}\times\mathbb{R}_+\times\mathbb{R}.\]
\end{lemma}
Actually there is more than one way to extend $\phi_2$ to the required framing $\phi_2''$. The proof below uses a specific choice, which we shall carry with us to make calculations later. Precisely, we make the choice which conforms to one of the cases $\overline{{}^+_+{}^+_-}, \overline{{}^+_-{}^+_+}$ in the \textit{coherent system of paths} described in \cite[Lemma 3.1]{MR3252965}, and in the preceding paragraphs there.
\begin{proof}For a manifold with corners $N$, denote by $N^{(l)}$ the set of corner points of codimension $l$. Then \[(M\times[0,1])^{(l)}=(M^{(l)}\times\{0,1\})\sqcup (M^{(l+1)}\times(0,1)).\]But \[\iota'(M^{(l)}\times\{0,1\})\subset (\mathbb{E}^{(d_k,\dots,d_{k+j})})^{(l)}\times\{0\}\times\mathbb{R},\]and \[\iota'(M^{(l+1)}\times(0,1))\subset (\mathbb{E}^{(d_k,\dots,d_{k+j})})^{(l+1)}\times(\mathbb{R}_+\setminus\{0\})\times\mathbb{R},\]so that both components of $M\times[0,1])^{(l)}$ are embedded in $(\mathbb{E}^{(d_k,\dots,d_{k+j})}\times\mathbb{R}_+\times\mathbb{R})^{(l)}$ as required. Moreover, as $\iota$ was a neat embedding, and the arc $\gamma_{u,v}$ meets the $\mathbb{R}$--axis transversely, $\iota'$ meets the boundary transversality condition (2) of Definition \ref{def:neat} above.
The two different boundary framings $\phi_1$, $\phi_2$ will now be extended to framings of $\iota'$. To build the required framing $\phi''_1$, take the embedding \[\phi':M\times[0,1]\times[-\epsilon,\epsilon]^A\looparrowright \mathbb{E}^\textbf{d}\times\mathbb{R}_+\times\mathbb{R}\] and extend it in a direction normal to $\gamma_{u,v}$ in $\mathbb{R}_+\times\mathbb{R}$ so that on the boundary we have $\phi_1$ as required. To build the required framing $\phi_2''$, we must look at the normal direction in $\mathbb{E}^\textbf{d}$ that was flipped by $R$ and turn that normal direction through $180^\circ$ as we move along $\gamma_{u,v}$. We do this by framing $\gamma_{u,v}\subset \mathbb{R}\times\mathbb{R}_+\times\mathbb{R}$ as follows. Referring to the coordinates of $\mathbb{R}\times\mathbb{R}_+\times\mathbb{R}$ as $e_1$, $\bar{e}$ and $e_2$, respectively, we will rotate $180^\circ$ around the $e_2$--axis, such that the first vector
equals $\bar{e}$ halfway through (compare \cite[Lemma 3.1]{MR3252965}). The rotation of the 2--dimensional frame as we move along $\gamma_{u,v}$ (extended trivially to the other normal directions) now defines the required embedding \[\phi_2'':M\times[0,1]\times[-\epsilon,\epsilon]^{A+1}\looparrowright \mathbb{E}^\textbf{d}\times\mathbb{R}_+\times\mathbb{R} {\rm .}\]
\end{proof}
We now add the objects $e$ and $f$ to $\mathscr{C}$, along with some new moduli spaces. Schematically, the old and the new morphisms are given by the following diagrams (with grading indicated by horizontal levels):
\[\xymatrix @R=1cm @C=0.5cm {\mathscr{C}:&\ar[d]&&\ar[d]&&&\bar{\mathscr{C}}:&\ar[d]&&\bullet_f\ar[d]\ar[drr]\ar[dll]&&\ar[d]\\
&\bullet_x\ar[d]&&\bullet_y\ar[d]&&&&\bullet_{\bar{x}}\ar[d]&&\bullet_e\ar[drr]\ar[dll]&&\bullet_{\bar{y}}\ar[d]\\
&&&&&&&&&&&}\]
Specifically, the morphisms and moduli spaces are given as follows.
\begin{definition}\label{def:int}The \emph{intermediate flow category} $\bar{\mathscr{C}}$ is the flow category with objects
\[\text{Ob}(\bar{\mathscr{C}})=\{\bar{a}\,|\,a\in\text{Ob}(\mathscr{C})\}\cup\{e,f\},\]
where $|e|=i$ and $|f|=i+1$ and whose moduli spaces are given by\begin{eqnarray*}
\mathcal{M}(e,\bar{b})&=&\mathcal{M}(x,b)\sqcup\mathcal{M}(y,b),\\
\mathcal{M}(f,\bar{b})&=&\left\{\begin{array}{lcl}\text{pt.}&&\text{if $\bar{b}=\bar{x},e,\bar{y}$},\\
\mathcal{M}(e,\bar{b})\times[0,1]&&\text{otherwise}.\end{array}\right.
\end{eqnarray*}In all other cases $\mathcal{M}(\bar{a},\bar{b})=\mathcal{M}(a,b)$.
\end{definition}
We now present a neat immersion and two choices of framing for $\bar{\mathscr{C}}$ relative to $\textbf{d}$. With some extra work, it is actually possible to neatly \emph{embed} and frame $\bar{\mathscr{C}}$ relative to $\textbf{d}$. But this is not necessary here as we are free to eventually increase $\textbf{d}$ and then perturb the immersion to an embedding, without changing the associated stable homotopy type.
\begin{proposition}\label{prop:immerse}There exists a neat immersion and two framings $(\bar{\iota},\bar{\phi}(+))$, $(\bar{\iota},\bar{\phi}(-))$ of $\bar{\mathscr{C}}$, relative to $\textbf{d}$, which extend the neat embedding and framing $(\iota,\phi)$ of $\mathscr{C}$.
\end{proposition}
\begin{proof}We will deal with the $(\bar{\iota},\bar{\phi}(+))$ version first and will suppress the ``+'' from the notation of $\bar{\phi}(+)$ while we work on this case.
When $\bar{a}\neq e,f$, define the embeddings and framings $\bar{\iota}$ and $\bar{\phi}$ of $\mathcal{M}(\bar{a},\bar{b})$ by the $\iota$ and $\phi$ as in $\mathscr{C}$.
For $\epsilon$ small, define the embeddings $\bar{\iota}_{f,\bar{x}}$ $\bar{\iota}_{f,e}$, $\bar{\iota}_{f,\bar{y}}$ as the points \[(4\epsilon,0,\dots,0), \quad(0,0,\dots,0), \quad(-4\epsilon,0,\dots,0)\in\mathbb{R}^{d_i}\]respectively. Extend these embeddings to framed embeddings $\bar{\phi}_{f,\bar{x}}$ $\bar{\phi}_{f,e}$, $\bar{\phi}_{f,\bar{y}}$ in the obvious way, by taking the product of $\epsilon$ neighbourhoods in each co-ordinate of $\mathbb{R}^{d_i}$. There are exactly two ways to frame an embedded point in any Euclidean space, which we will denote ``$+$'' and ``$-$''. Declare $\bar{\phi}_{f,\bar{x}}$ and $\bar{\phi}_{f,\bar{y}}$ to both be $+$ framings and $\bar{\phi}_{f,e}$ to be a $-$.
Fixing an object $\bar{b}$ with $|\bar{b}|<i$, we will now immerse and frame $\mathcal{M}(e,\bar{b})=\mathcal{M}(x,b)\sqcup\mathcal{M}(y,b)$. To do this we will simply superimpose the embeddings and framings given by $\iota$ and $\phi$. Note that this might introduce immersion points. Specifically, define the neat immersion\[\bar{\iota}_{e,\bar{b}}:\mathcal{M}(e,\bar{b})\looparrowright\mathbb{E}^{\textbf{d}}[|e|:|\bar{b}|];\qquad p\mapsto\left\{\begin{array}{lcl}\iota_{x,b}(p)&&\text{if $p\in \mathcal{M}(x,b)$,}\\ \iota_{y,b}(p)&&\text{if $p\in \mathcal{M}(y,b)$.}\end{array}\right.\]And similarly, for $\phi_{x,b},\phi_{y,b}$, we obtain the framing $\bar{\phi}_{e,\bar{b}}$.
We will now immerse and frame $\mathcal{M}(f,\bar{b})$, where $\bar{b}\neq \bar{x}, \bar{y},e$. First apply Lemma \ref{lem:arc}, case (1) to $\iota_{x,b}$, setting $j=i-k-1$ and $(u,v)=(2\epsilon,2\epsilon)$, to obtain neat embeddings $\iota_{x,b}'$, $\phi_{x,b}'$ and $\phi_{x,b}''$. Now apply Lemma \ref{lem:arc}, case (1) a second time, this time to $\iota_{y,b}$, setting $j=i-k-1$ and $(u,v)=(-2\epsilon,2\epsilon)$, to obtain neat embeddings $\iota_{y,b}'$, $\phi_{y,b}'$ and $\phi_{y,b}''$. In order to use these functions to define a neat immersion $\bar{\iota}_{f,\bar{b}}$, we will need to increase the range of $\iota_{y,b}', \iota_{y,b}'$ by a factor or $\mathbb{R}^{d_i-1}$. We do so by simply setting the new $d_i-1$ co-ordinates in the range of each neat embedding to 0. In an abuse of notation, use the same symbols for the new functions - i.e.\ from now on write\begin{eqnarray*}
\iota_{x,b}':&\mathcal{M}(x,b)\times[0,1]\looparrowright \mathbb{E}^{\textbf{d}}[|x|:|b|]\times\mathbb{R}_+\times\mathbb{R}\times\mathbb{R}^{d_i-1}= \mathbb{E}^{\textbf{d}}[|f|:|\bar{b}|],\\
\iota_{y,b}':&\mathcal{M}(y,b)\times[0,1]\looparrowright \mathbb{E}^{\textbf{d}}[|y|:|b|]\times\mathbb{R}_+\times\mathbb{R}\times\mathbb{R}^{d_i-1}= \mathbb{E}^{\textbf{d}}[|f|:|\bar{b}|].\end{eqnarray*} Using this, define a neat immersion of $\mathcal{M}(f,\bar{b})$\[
\bar{\iota}_{f,\bar{b}}:\mathcal{M}(e,\bar{b})\times[0,1]\looparrowright\mathbb{E}^{\textbf{d}}[|f|,|\bar{b}|];\quad (p,q)\mapsto\left\{\begin{array}{lcl}\iota_{x,b}'(p,q)&&\text{if $p\in \mathcal{M}(x,b)$,}\\ \iota_{y,b}'(p,q)&&\text{if $p\in \mathcal{M}(y,b)$.}\end{array}\right.\]To build a framing $\bar{\phi}_{f,\bar{b}}$, we will need to modify \emph{both} the domain and range of the neat embeddings $\phi_{x,b}''$, $\phi_{y,b}''$. Specifically, we must extend the domains of these neat embeddings by a factor of $[-\epsilon,\epsilon]^{d_i-1}$, and we must extend the ranges by a factor of $\mathbb{R}^{d_i-1}$. We do this by simply extending $\phi_{x,b}''$ and $\phi_{y,b}''$ by the obvious inclusion $[-\epsilon,\epsilon]^{d_i-1}\subset\mathbb{R}^{d_i-1}$. Again we abuse notation and use the same symbols for the extended functions:\begin{eqnarray*}
\phi_{x,b}'':&\mathcal{M}(x,b)\times[0,1]\times[-\epsilon,\epsilon]^A\looparrowright \mathbb{E}^{\textbf{d}}[|x|:|b|]\times\mathbb{R}_+\times\mathbb{R}\times\mathbb{R}^{d_i-1},\\
\phi_{y,b}'':&\mathcal{M}(y,b)\times[0,1]\times[-\epsilon,\epsilon]^A\looparrowright \mathbb{E}^{\textbf{d}}[|y|:|b|]\times\mathbb{R}_+\times\mathbb{R}\times\mathbb{R}^{d_i-1},\end{eqnarray*}where $A=d_{|\bar{b}|}+\dots+d_{|f|-1}$. So, at last, we have a neat immersion: \[\begin{array}{rcl}\bar{\phi}_{f,\bar{b}}:\mathcal{M}(e,\bar{b})\times[0,1]\times[-\epsilon,\epsilon]^A&\looparrowright&\mathbb{E}^{\textbf{d}}[|f|,|\bar{b}|]\\(p,q,\textbf{t})&\mapsto&\left\{\begin{array}{lcl}\phi_{x,b}''(p,q.\textbf{t})&&\text{if $p\in \mathcal{M}(x,b)$,}\\ \phi_{y,b}''(p,q,\textbf{t})&&\text{if $p\in \mathcal{M}(y,b)$.}\end{array}\right.\end{array}\]Note that the use of the construction from Lemma \ref{lem:arc} is consistent with our choice of $+$ and $-$ framings $\bar{\phi}_{f,\bar{x}}$ $\bar{\phi}_{f,e}$, $\bar{\phi}_{f,\bar{y}}$ from earlier. The ``+'' case is completed.
Now we must describe the required modifications to work on the case $(\bar{\iota},\bar{\phi}(-))$. Again we will suppress the ``$-$'' notation as we work on this case.
The first difference from the $(\bar{\iota},\bar{\phi}(+))$ case is that we declare $\bar{\phi}_{f,\bar{x}}$ be a $+$ framing and both of $\bar{\phi}_{f,\bar{y}}$ and $\bar{\phi}_{f,e}$ to be a $-$ framings. This, in turn, means that for each object $b$, we will want to use case (2) of Lemma \ref{lem:arc} when we come to embed and frame the $\mathcal{M}(y,b)\times[0,1]$ component of $\mathcal{M}(f,\bar{b})$. In order to get into case (2) of this lemma, we must endow the $\mathcal{M}(y,b)$ component of $\mathcal{M}(e,\bar{b})$ with the framing obtained by taking the image of $\phi_{y,b}$ in $\mathbb{E}^\textbf{d}[|y|:|b|]$ and performing the reflection $R$ in the plane through the origin and perpendicular to the first coordinate axis of the $\mathbb{R}^{d_{i-1}}$ factor. Other than these modifications, the construction of the immersions and framings in the $(\bar{\iota},\bar{\phi}(-))$ case proceeds exactly as the $(\bar{\iota},\bar{\phi}(+))$ case.
The $(\bar{\iota},\bar{\phi}(\pm))$ cases are illustrated in the following diagram, where $\alpha$, $\beta$, $\gamma$, $\delta$ denote framings and $-\delta$ represents the effect of reflecting $\delta$ by $R$ as described above.
\[\xymatrix @R=1cm @C=0.5cm {\mathscr{C}:&\ar[d]_-{\alpha}&&\ar[d]^-{\beta}&&&\bar{\mathscr{C}}:&\ar[d]_-\alpha&&\bullet_f\ar[d]^-{-}\ar[drr]^-{\pm}\ar[dll]_-{+}&&\ar[d]^-{\beta}\\
&\bullet_x\ar[d]_-{\gamma}&&\bullet_y\ar[d]^-{\delta}&&&&\bullet_{\bar{x}}\ar[d]_-{\gamma}&&\bullet_e\ar[drr]_-{\pm\delta}\ar[dll]^-{\gamma}&&\bullet_{\bar{y}}\ar[d]^-{\delta}\\
&&&&&&&&&&&}\]
\end{proof}
\begin{proposition}\label{prop:choices}Cancelling $e$ against $f$ in $(\bar{\mathscr{C}},\bar{\iota},\bar{\phi}(\pm))$, using the handle cancellation technique of \cite[\textsection 2.3]{JLSmorse}, results in precisely $(\mathscr{C}, \iota, \phi)$.
\end{proposition}
\begin{proof}Write $(\mathscr{C}_H,\iota_H,\phi_H)$ for the result of cancelling $e$ against $f$ as described in \cite[\textsection 2.3]{JLSmorse}. According to \cite[Definition 2.9]{JLSmorse}, the objects of $\mathscr{C}$ and of $\mathscr{C}_H$ are in 1:1 correspondence. Moreover, recall from \cite[\textsection 2.3]{JLSmorse} that if $a,b$ are objects of $\mathscr{C}$, then the moduli spaces of $\mathscr{C}_H$ are given by\[\mathcal{M}(a,b)=\mathcal{M}(\bar{a},\bar{b})\cup_f\left(\mathcal{M}(f,\bar{b})\times \mathcal{M}(\bar{a},e)\right),\]where $f$ is a certain identification described generally in \cite[\textsection 2.3]{JLSmorse}. However, when $a$ is an object of $\mathscr{C}$, there are no non-empty moduli spaces $\mathcal{M}(\bar{a},e)$, so in fact the moduli spaces of $\mathscr{C}_H$ are the same as those of $\mathscr{C}$. This simplification, carried through the construction in \cite[Theorem 2.17]{JLSmorse}, actually results in a precise agreement of embeddings and framings so that in fact $(\mathscr{C},\iota,\phi)=(\mathscr{C}_H,\iota_H,\phi_H)$.
\end{proof}
\subsection{Handle sliding in framed flow categories}\label{subsec:sliding}
\begin{definition}The framed flow category $(\mathscr{C}_S,\iota_S,\phi_S(\pm))$, called the \emph{effect of $(\pm)$--sliding $x$ over $y$ in $(\mathscr{C},\iota,\phi)$} is defined as the result of cancelling $f$ against $\bar{x}$ in $(\bar{\mathscr{C}},\bar{\iota},\bar{\phi}(\pm))$. We will often suppress the ``$+$'' and ``$-$'' from the notation.
\end{definition}
The following lemma is a straightforward consequence of the constructions so far and the definition of handle cancellation given in \cite[Definition 2.9]{JLSmorse}.
\begin{lemma}The framed flow category $\mathscr{C}_S$ has objects\[\text{Ob}(\mathscr{C}_S)=\{a'\,|\,a\in\text{Ob}(\mathscr{C})\},\]and moduli spaces given by\begin{eqnarray*}
\mathcal{M}(x',b')&=&\mathcal{M}(x,b)\sqcup\mathcal{M}(y,b),\\
\mathcal{M}(a',y')&=&\mathcal{M}(a,x)\sqcup\mathcal{M}(a,y),
\end{eqnarray*}In all other cases \[\mathcal{M}(a',b') = \mathcal{M}(a,b) \sqcup \left(\mathcal{M}(y,b) \times [0,1] \times \mathcal{M}(a,x)\right).\]
\end{lemma}
\begin{proof}The main thing to note is that after we cancel $f$ against $\bar{x}$ in the intermediate flow category, we are left with object set $\{\bar{a}\,|\,a\in\text{Ob}(\mathscr{C})\setminus\{x\}\}\cup\{e\}$. Hence the object $x'\in\text{Ob}(\mathscr{C}_S)$ in the statement of the lemma is a relabelling of $e$. The rest of the lemma follows directly from \cite[Definition 2.9]{JLSmorse}.
\end{proof}
The following is a schematic for sliding $x$ over $y$, where the framings determined by a $(\pm)$--slide are also shown. The signs are calculated by following the construction through the handle cancellation in \cite[Definition 2.9]{JLSmorse}.
\[\xymatrix @R=1cm @C=0.5cm {\mathscr{C}:&\ar[d]_-{\alpha}&&\ar[d]^-{\beta}&&&\mathscr{C}_S:&\ar[d]_-{\alpha}\ar[drr]^-{\mp\alpha}&&\ar[d]^-{\beta}\\
&\bullet_{x}\ar[d]_-{\gamma}&&\bullet_{y}\ar[d]^-{\delta}&&&&\bullet_{x'}\ar[d]_-{\gamma}\ar[drr]_-{\pm\delta}&&\bullet_{y'}\ar[d]^-{\delta}\\
&&&&&&&&&}\]
Now we are in a position to deduce the main result of this section.
\begin{proof}[Proof of Theorem \ref{thm:handleslide}] Define the intermediate framed flow category $(\bar{\mathscr{C}},\bar{\iota},\bar{\phi}(\pm))$ as in Definition \ref{def:int} and Proposition \ref{prop:immerse}. By Proposition \ref{prop:choices}, the framed flow category obtained from cancelling $e$ against $f$ is precisely $(\mathscr{C}, \iota,\phi)$. By sufficiently increasing $\textbf{d}$ to some $\bar{\textbf{d}}$ we may perturb the framed neat immersions $(\bar{\iota},\bar{\phi}(\pm))$, $(\iota_S,\phi_S)$ to framed neat embeddings and hence geometrically realise the CW-complexes $|\bar{\mathscr{C}}|$ and $|\mathscr{C}_S|$. It is immediately clear from \cite[Theorem 2.17]{JLSmorse} that there are homotopy equivalences \[\mathcal{X}(\mathscr{C})\simeq\mathcal{X}(\bar{\mathscr{C}})\simeq\mathcal{X}(\mathscr{C}_S).\]
\end{proof}
\subsection{Framings of 1--dimensional moduli spaces}\label{subsec:framings}
In order to compute cohomology operations such as $\Sq^2$ we need to understand how the framings of the 1--dimensional moduli spaces behave after a handle slide. If we keep in mind the cancellation from the intermediate flow category $\bar{\mathscr{C}}$ to $\mathscr{C}_S$ we get
\begin{align*}
\mathcal{M}(a',x') &= \mathcal{M}(f,e) \times \mathcal{M}(a,x), \\
\mathcal{M}(a',y') &= \mathcal{M}(a,y) \sqcup \left(\mathcal{M}(f,y) \times \mathcal{M}(a,x)\right).
\end{align*}
Furthermore,
\[
\mathcal{M}(x',b') = \mathcal{M}(x,b) \sqcup \mathcal{M}(y,b),
\]
but with the framing of $\mathcal{M}(y,b)$ depending on whether we do a $(+)$ or a $(-)$-handle slide.
Since $\mathcal{M}(f,e)$ is a negatively framed point, we get that the framing on $\mathcal{M}(a',x')$ corresponds to the old framing on $\mathcal{M}(a,x)$ by the framings in the cancelled category, compare \cite{JLSmorse}. Similarly, if we use the $(-)$--handle slide the framings on $\mathcal{M}(a',y')$ agree with the original framings on $\mathcal{M}(a,y)$ and $\mathcal{M}(a,x)$. In the case of the $(+)$--handle slide however, the copy of $\mathcal{M}(a,x)$ in $\mathcal{M}(a',y')$ has an extra reflection in one coordinate.
Recall that interval components $J$ in $\mathcal{M}(a,b)$ with $|a|=|b|+2$ can be framed in two ways (up to fixed boundary framing). The standard framings are described in \cite[\S 3.2]{JLSmorse} and \cite[\S 3]{MR3252965}, and we write $fr(J)\in \mathbb{Z}/2\mathbb{Z}$ with $fr(J)=0$ if the framing corresponds to the standard choice, and $fr(J)=1$ if not.
Similarly, if $|a|=|b|+1$ and $A\in \mathcal{M}(a,b)$ is a point, we write $\varepsilon_A\in \mathbb{Z}/2\mathbb{Z}$ for the framing sign of this point.
\begin{proposition}\label{prop:framechanges}
Let $\mathscr{C}$ be a framed flow category containing two objects $x,y$ with the same grading, and let $\mathscr{C}_S$ be the framed flow category obtained from $\mathscr{C}$ by sliding $x$ over $y$. Let $a,b$ be objects in $\mathscr{C}$ with $|a|=|b|+2$.
\begin{enumerate}
\item If $|a|=|x|+2$, then $\mathcal{M}(a',b')=\mathcal{M}(a,b)$ for all $b\not=y$ with the same framing. Furthermore
\[
\mathcal{M}(a',y') = \mathcal{M}(a,y) \sqcup \mathcal{M}(a,x)
\]
with components from $\mathcal{M}(a,y)$ identically framed. If $J\subset \mathcal{M}(a,x)$ is an interval component, denote $J'$ for the same component when viewed as a subset of $\mathcal{M}(a',y')$. Then
\begin{enumerate}
\item If $\mathscr{C}_S$ is obtained by a $(-)$--handle slide, then
\[
fr(J')= fr(J).
\]
\item If $\mathscr{C}_S$ is obtained by a $(+)$--handle slide, then
\[
fr(J')= 1 + fr(J).
\]
\end{enumerate}
\item If $|a|=|x|+1$, then
\[
\mathcal{M}(a',b') = \mathcal{M}(a,b) \sqcup \mathcal{M}(y,b)\times [0,1] \times \mathcal{M}(a,x)
\]
and every component of $\mathcal{M}(a,b)$ is framed the same way. For each pair of points $(B,A)\in \mathcal{M}(y,b)\times \mathcal{M}(a,x)$, denote $I_{B,A}$ the corresponding interval component in $\mathcal{M}(a',b')$.
\begin{enumerate}
\item If $\mathscr{C}_S$ is obtained by a $(-)$--handle slide, then
\[
fr(I_{B,A})=\varepsilon_B.
\]
\item If $\mathscr{C}_S$ is obtained by a $(+)$--handle slide, then
\[
fr(I_{B,A})=1+\varepsilon_B.
\]
\end{enumerate}
\item If $|a|=|x|$ and $a\not= x$, then the framings of components in $\mathcal{M}(a',b')$ are the same as in $\mathcal{M}(a,b)$. For
\[
\mathcal{M}(x',b')=\mathcal{M}(x,b) \sqcup \mathcal{M}(y,b)
\]
the framing values in the new moduli space agree with the framing values in the old moduli spaces.
\end{enumerate}
\end{proposition}
\begin{proof}
Assume $|a|=|x|+2$. In $\mathscr{C}_S$ the object $x'$ corresponds to $e$ from the intermediate flow category, and the intervals in $\mathcal{M}(\bar{a},e)$ are of the form $\{-\} \times J$ where $-\in \mathcal{M}(f,e)$ from the intermediate flow category, and $J$ is an interval in the original $\mathcal{M}(a,x)$. By \cite[Proposition 3.6.6]{JLSmorse}, $fr(\{-\}\times J)=fr(J)+1+\varepsilon_-$. This means $fr(\{-\}\times J)=fr(J)$.
For $\mathcal{M}(a',y')$ the intervals in $\mathcal{M}(a,y)$ do not change, but we also get intervals of the form $\{B\} \times J$ with $B\in \mathcal{M}(f,\bar{y})$ and $J$ in $\mathcal{M}(a,x)$. As before, $fr(\{B\}\times J) = fr(J) + 1 + \varepsilon_B$. Now if we use a $(-)$--slide, $\varepsilon_B=1$, and if we use a $(+)$--handle slide, $\varepsilon_B=0$. Hence we get (1).
Now assume $|a|=|x|+1$. In the intermediate flow category we get intervals $J_C$ and $J_B$ in $\mathcal{M}(f,\bar{b})$ coming from points $C\in \mathcal{M}(x,b)$ and $B\in \mathcal{M}(y,b)$. When passing to $\mathscr{C}_S$ the intervals $\{A\}\times J_C$ give a collar neighborhood to an interval in $\mathcal{M}(a,b)$ with endpoint $(C,A)\in \mathcal{M}(x,b)\times \mathcal{M}(a,x)$, which do not change the framing of the original interval. The intervals $J_B$ are framed depending on whether we have a $(+)$ or a $(-)$--slide. In the case of a $(+)$--slide we use Lemma \ref{lem:arc}.1.
Note that the sign of the second coordinate changes, while the first coordinate remains $\varepsilon_B$. But during the rotation described in the proof of Lemma \ref{lem:arc}, the second vector equals $-\bar{e}$. By the choice of a coherent system of paths as done in \cite{MR3252965} or \cite{JLSmorse}, this is a standard path if and only if $\varepsilon_B=1$. Hence $fr(J_B)=1+\varepsilon_B$. Passing to $\mathscr{C}_S$ gives intervals $I_{B,A}$ for each $A\in \mathcal{M}(a,x)$ with the same framing value as $J_B$ by \cite[Proposition 3.7.6]{JLSmorse}.
If we perform a $(-)$--slide, we use Lemma \ref{lem:arc}.2, and the sign of the first coordinate changes, while the second coordinate remains. In the proof of Lemma \ref{lem:arc} we note that there is a rotation around the $e_2$--axis chosen such that the first standard vector points in the positive direction of $\mathbb{R}_+$. Comparing with the choice of coherent system of paths we get a standard path if and only if the first vector of the framing points in the positive direction, that is, when $\varepsilon_B=0$. This implies $fr(I_{B,A})=\varepsilon_B$.
Finally, assume that $|a|=|x|$. If $a\not=x$, then $\mathcal{M}(a',b')=\mathcal{M}(a,b)$ and the framings do not change. If $a=x$, we get $\mathcal{M}(x',b')=\mathcal{M}(x,b)\sqcup \mathcal{M}(y,b)$ and the framing of $\mathcal{M}(x,b)$ does not change. If we perform a $(+)$--slide, the framing of $\mathcal{M}(y,b)$ does not change either. In the case of a $(-)$--slide, the first coordinate of $\mathbb{R}^{d_{i-1}}$ is reflected. This means that the second coordinate of the framing of an interval is changing sign. However, flipping the sign of the second coordinate maps standard paths to standard paths, compare also \cite[Lemma 3.3]{MR3252965}, although only half of this statement is proven there.
Hence the framing does not change either.
\end{proof}
\section{The extended Whitney trick in framed flow categories}
\label{sec:whitney}
In \cite[Theorem 3.1.5]{MR1781277}, Laures extends the Pontryagin--Thom Theorem to the setting of manifolds with corners using a suitably defined framed cobordism category. As the cell attachment maps in the Cohen--Jones--Segal construction are defined by a version of the Pontryagin--Thom collapse map, the homotopy classification of these attaching maps (hence the eventual homeomorphism type of the CW-complex) is closely related to the framed cobordism classes of the moduli spaces in a flow category.
In this section we will re-encode this relationship by defining a new technique for modifying framed flow categories, called \emph{the extended Whitney trick} (compare the Whitney trick of \cite[\textsection 1]{JLSmorse}). Suppose $(\mathscr{C}, \iota, \phi)$ is a framed flow category and that for some objects $x\neq y$ of $\mathscr{C}$, that $\mathcal{M}(x,y)=M$ is a manifold with boundary. Let $W$ be a framed cobordism (rel.\ boundary) between $M$ and another manifold with boundary $M'$. We will show how to use $W$ to define a new framed flow category $(\mathscr{C}_W,\iota_W,\phi_W)$ in which $M$ is replaced by $M'$ (and the other moduli spaces in $\mathscr{C}$ are modified appropriately). Using the ideas of the Pontryagin--Thom theorem we will deduce the following.
\begin{theorem}
\label{thm:whitney}
If $(\mathscr{C}_W,\iota_W,\phi_W)$ is the result of an \emph{extended Whitney trick} in $(\mathscr{C}, \iota, \phi)$, then there is a homotopy equivalence \[\mathcal{X}(\mathscr{C}_W) \simeq \mathcal{X}(\mathscr{C}) {\rm .}\]
\end{theorem}
\subsection{Framed cobordism of manifolds with corners rel.\ boundary}We will not need Laures's full machinery \cite{MR1781277} of cobordism of manifolds with corners in the sequel. The eventual complicated interactions between the moduli spaces in a framed flow category mean that allowing an unrestricted framed cobordism in the sense of \cite{MR1781277} becomes intractable. Instead we will work with the following.
\begin{definition}Suppose $M$ and $M'$ is are $m$--dimensional $\langle n\rangle$--manifolds with $\partial_iM=\partial_iM'$ for $i=1,\dots,n$. An $(m+1)$--dimensional $\langle n+1\rangle$--manifold $W$ is called a \emph{cobordism rel.\ boundary} between $M$ and $M'$ if $\partial_{n+1}W=M\sqcup M'$ and $\partial_iW=\partial_i\times[0,1]$ for $i\neq n+1$.
Suppose an embedding $\tilde{\iota}:W\hookrightarrow \mathbb{E}^{\textbf{d}}\times[0,1]$, of a cobordism rel.\ boundary, meets $\mathbb{E}^\textbf{d}\times\{0,1\}$ orthogonally in $M\sqcup M'$, and induces neat embeddings \[\begin{array}{rrcl}\tilde{\iota}|_M:&M&\hookrightarrow& \mathbb{E}^\textbf{d}\times\{1\},\\
\tilde{\iota}|_{M'}:&M'&\hookrightarrow& \mathbb{E}^\textbf{d}\times\{0\}.\end{array}\]Then $\tilde{\iota}$ is an \emph{embedded cobrdism rel.\ boundary} between the neat embeddings $(M',\tilde{\iota}|_{M'})$ and $(M,\tilde{\iota}|_{M})$. Suppose furthermore that there exists a framing $\tilde{\phi}$ of such a $(W,\tilde{\iota})$ and that $\tilde{\phi}$ meets $\mathbb{E}^\textbf{d}\times\{0,1\}$ orthogonally. Then the framing $\tilde{\phi}$ determines framings $\phi$ and $\phi'$ of $(M,\tilde{\iota}|_M)$ and $(M',\tilde{\iota}|_{M'})$ respectively. If such a $(W,\tilde{\iota}, \tilde{\phi})$ exists, it is called a \emph{framed cobordism rel.\ boundary} between $(M',\tilde{\iota}|_{M'},\phi')$ and $(M,\tilde{\iota}|_{M},\phi)$. $(M',\tilde{\iota}|_{M'},\phi')$ and $(M,\tilde{\iota}|_{M},\phi)$ are called \emph{framed cobordant rel.\ boundary} if there exists a framed cobordism rel.\ boundary between them, possibly after enlarging the $\textbf{d}\leq\textbf{d}'$.
\end{definition}
\begin{remark}Examples of framed cobordisms rel.\ boundary $W\subset \mathbb{E}^{\textbf{d}'}\times[0,1]$ whose framed boundary $M\sqcup M'$ can be framed embedded in a smaller space $\mathbb{E}^\textbf{d}$ are fairly common. This is why we have allowed the possibility of enlarging the ambient space in the final definition above. For instance, the generator $1\in\Omega_1^{fr}\cong\mathbb{Z}/2\mathbb{Z}$ can be embedded, along with a framed normal neighbourhood, as $\phi:S^1\times D^2\hookrightarrow\mathbb{R}^3$. But the minimum embedding dimension for a framed nullcobordism of $\phi\sqcup\phi$ is 5.\end{remark}
\subsection{Pushing $M$ out of the corner}
We now make a slight digression into an easy but technical construction we will need later. Suppose $(W,\tilde{\iota},\tilde{\phi})$ is a framed embedded cobordism rel.\ boundary between $m$--dimensional $\langle m\rangle$--manifolds $(M',\iota',\phi')$ and $(M, \iota,\phi)$, where $\iota, \iota'$ are with respect to some $\textbf{d}=(d_0,d_1,\dots,d_m)$. Denote by $U_\eta\subset (\mathbb{R}_+)^N$ the open ball at the origin with (small) radius $\eta>0$, $\overline{U_\eta}$ the corresponding closed ball, and by $H_\eta=\overline{U_\eta}\setminus U_\eta$. Suppose that \[\iota_X:X\hookrightarrow \mathbb{E}^{\textbf{d}}\times(\mathbb{R}_+)^N\]is a neat embedding of an $(m+N)$--dimensional $\langle m+N\rangle$--manifold, with framing $(X,\iota_X,\phi_X)$. Moreover, suppose that near the `corner' $\mathbb{E}^\textbf{d}\times\textbf{0}$ this embedding is\[\iota_X(X)\cap \mathbb{E}^\textbf{d}\times U_{2\eta}=\iota(M)\times U_{2\eta}\]and that here the framing $\phi_X$ agrees with the framing \[M\times U_{2\eta}\times[-\epsilon,\epsilon]^A\hookrightarrow \mathbb{E}^\textbf{d}\times U_{2\eta};\qquad (p,q,\textbf{t})\mapsto (\phi(p,\textbf{t}),q),\]where $A=d_0+\dots+d_m$.
Later on we will need a mechanism to `push $M$ out of the corner' and replace it with $M'$. Roughly speaking, this is achieved by glueing together the two spaces\begin{eqnarray*}X'&=&\iota_X(X)\setminus (M\times U_\eta),\\Y&=&(M'\times \overline{U_\eta})\cup_{M'\times H_\eta}(W\setminus(M'\times [0,\eta))\times H_\eta,\end{eqnarray*}and then embedding the result appropriately (compare Figure \ref{fig:push}). But to ensure the smooth structures can be made to agree, we will need to use the standard technique of overlaying open collar neighbourhoods.
\begin{figure}
\begin{tikzpicture}
\draw [<->] (0,5) -- (0,0) -- (-5,0);
\draw (-2,0) arc [radius=2, start angle=180, end angle= 90];
\draw (-4,0) arc [radius=4, start angle=180, end angle= 90];
\node at (-4.9,-0.4) {$\mathbb{R}_+$};
\node at (0.4,4.9) {$\mathbb{R}_+$};
\node at (0.3,4) {$2\eta$};
\node at (0.25,2) {$\eta$};
\node at (-1.5,2.5) {$M\times(U_{2\eta}\setminus U_\eta)$};
\node at (-0.9,0.7) {$M\times U_\eta$};
\node at (1.65,3.35) {Push $M$ out } ;
\node at (1.65,3) {of the corner};
\draw [->] [decorate, decoration={snake}] (0.7,2.5) -- (2.6,2.5);
\draw [<->] (7,5) -- (7,0) -- (2,0);
\draw (5,0) arc [radius=2, start angle=180, end angle= 90];
\draw (3,0) arc [radius=4, start angle=180, end angle= 90];
\draw [dashed] (3.3,0) arc [radius=3.7, start angle=180, end angle= 90];
\draw [dashed] (2.7,0) arc [radius=4.3, start angle=180, end angle= 90];
\draw [red] (5.0681,0.5176) -- (3.1362, 1.0352);
\draw [red] (5.2679,1) -- (3.5358, 2);
\draw [red] (5.5857,1.4142) -- (4.1715,2.8284);
\draw [red] (6,1.732) -- (5, 3.4641);
\draw [red] (6.4823,1.9318) -- (5.9647,3.8637);
\node at (2.1,-0.4) {$\mathbb{R}_+$};
\node at (7.4,4.9) {$\mathbb{R}_+$};
\node at (7.3,4) {$2\eta$};
\node at (7.25,2) {$\eta$};
\node at (6.1,0.7) {$M'\times U_\eta$};
\node [red] at (3.7,4.8) {$(W\setminus M'\times[0,\eta))\times H_\eta$};
\draw [->] (3.7,4.4) to [out=270,in=140] (5,2.6);
\draw [decorate,decoration={brace,amplitude=4pt,raise=2pt}] (3.3,-0.1) -- (2.7,-0.1);
\node at (4.15,-0.6) {\small $M\times H_\eta\times(-\eta',\eta')$};
\end{tikzpicture}
\caption{Example of pushing $M$ out of the corner. Here $N=2$ and $\mathbb{E}^\textbf{d}$ is thought of as the direction perpendicular to the page. Each radial red line is a copy of $W$ (minus a collar at the $M'$ boundary).}\label{fig:push}
\end{figure}
\begin{lemma}\label{lem:smooth}With notation as above, the $(m+N)$--dimensional topological manifold with corners $X'\cup_{M\times H_\eta} Y$ is homeomorphic to a smooth $(m+N)$--dimensional $\langle m+N\rangle$--manifold $Z$, which has a neat embedding $Z\hookrightarrow \mathbb{E}^\textbf{d}\times (\mathbb{R}_+)^N$ such that $M'$ is in the corner $\mathbb{E}^\textbf{d}\times \overline{U_\eta}$. Moreover the embedding of $Z$ can be framed compatibly with the framings of $(W,\tilde{\iota},\tilde{\phi})$ and of $(X, \iota_X,\phi_X)$.
\end{lemma}
In practice, we will only ever make use of the cases $N=1,2,3$ but it is no extra effort to prove Lemma \ref{lem:smooth} in generality. Also note that if $M$ is in a codimension $N$ corner then it is also in $N$ codimension $N-1$ corners. In fact the restriction of the construction in the proof to some choice of $(\mathbb{R}_+)^{N-1}\subset (\mathbb{R}_+)^N$ precisely reproduces the proof in this lower codimension. We will use this fact later.
\begin{proof}The topological manifold with corners $Y$ already carries a smooth structure as the two components that are glued to make $Y$ may be given their respective product smooth structures, and near the glueing boundary $M'\times H_\eta$ these product smooth structures agree precisely with one another.
Any radial direction of $(\mathbb{R}_+)^N$ determines a copy of $\mathbb{E}^\textbf{d}\times \mathbb{R}_+$. We may embed $\iota_Y:Y\hookrightarrow\mathbb{E}^\textbf{d}\times(\mathbb{R}_+)^N$ such that in any radial direction we have rescaled the embedding $\tilde{\iota}(W)\subset \mathbb{E}^\textbf{d}\times[0,1]\subset \mathbb{E}^\textbf{d}\times\mathbb{R}_+$ by a factor of $\eta$. Choose another small $\eta'<<\eta$. There is now a collar neighbourhood $\iota_Y(M\times H_\eta)\times(-\eta',0]$ of the embedded boundary component $M\times H_\eta\subset Y$, where the collar direction is radial in $(\mathbb{R}_+)^N$. Similarly, the boundary component $M\times H_\eta\subset X'$ has a radial collar neighbourhood $M\times H_\eta\times[0,\eta')\subset Y$. Writing $(p, t)\sim(p,(1-t)/2)$ for $p\in M\times H_\eta'$ and $t\in(0,\eta'/2)$, define the identification space\[Z=X'\sqcup Y/\sim.\]The smooth structures on $Y$ and $X'$ can now be made compatible on this open collar overlay, so that $Z$ has the structure of a smooth $\langle m+N\rangle$--manifold. Moreover, by radially dilating the embedding $\iota_X|_{X'}$ of $X'$, and combining with the embedding of $Y$, $Z$ is seen to have a (neat) embedding in $\mathbb{E}^\textbf{d}\times\mathbb{R}_+\times\mathbb{R}_+$ so that $M'$ is in the corner as required.
Now we turn to the framings. The framing of the corner embedding of $M'\times \bar{U_\eta}$ is given by extending the framing $\phi'$ of $\iota':M\hookrightarrow \mathbb{E}^\textbf{d}$ trivially to the product. Similarly we may frame the embedding of the product $(W\setminus(M'\times [0,\eta)))\times H_\eta$ by extending $\tilde{\phi}$. Note that these framings together form a framing of the embedded $Y$. Now, we assumed that $\phi_X$ agreed with $\phi$ near the corner $\mathbb{E}^\textbf{d}\times \textbf{0}$. Hence to frame $Z$, we can simply glue the framing on $X'$ coming from $X$ to the framing on $Y$ as they precisely agree on the collar overlap.
\end{proof}
\subsection{The extended Whitney trick}\label{subsec:whitney}
Let $(\mathscr{C},\iota,\phi)$ be a framed flow category with respect to $\textbf{d}=(d_k,\dots,d_{n+k})$. Suppose $\mathscr{C}$ has objects $x, y$ with $|x|-|y|=m+1$ and write $(M,\iota,\phi)=(\mathcal{M}(x,y),\iota_{x,y},\phi_{x,y})$. Suppose there is a framed cobordism rel.\ boundary $(W,\tilde{\iota},\tilde{\phi})$ between $(M,\iota,\phi)$ and some other embedded, framed $m$--dimensional $\langle m\rangle$--manifold $(M',\iota',\phi')$. We will define a new framed flow category $(\mathscr{C}_W,\iota_W,\phi_W)$ such that there is a homotopy equivalence $\mathcal{X}(\mathscr{C})\simeq\mathcal{X}(\mathscr{C}_W)$.
\begin{definition}\label{def:extW}With $(\mathscr{C},\iota,\phi)$ as above we will define $(\mathscr{C}_W,\iota_W,\phi_W)$. First, we define the object set $\text{Ob}(\mathscr{C}_W)=\{\bar{a}\,|\,a\in \text{Ob}(\mathscr{C})\}$. The moduli spaces of $\mathscr{C}_W$ are given as follows:\begin{enumerate}
\item $\mathcal{M}(\bar{x},\bar{y})=M'$.
\item If $a\in\text{Ob}(\mathscr{C})$ with $\mathcal{M}(a,x)\neq \emptyset$, then the moduli space $\mathcal{M}(\bar{a},\bar{y})$ is essentially given by glueing $W\times\mathcal{M}(a,x)$ to $\mathcal{M}(a,y)$ along the embedded boundary component\[M\times \mathcal{M}(a,x)\subset \partial_{m+1}\mathcal{M}(a,y).\]In order to ensure the smooth structures can be modified to match up when glueing, take small collar neighbourhoods $M\times\mathcal{M}(a,x)\times[0,\eta)$ and $M\times\mathcal{M}(a,x)\times(-\eta,0]$ inside $\mathcal{M}(a,y)$ and $W\times\mathcal{M}(a,x)$ respectively. Overlay the collars by setting $(p, q, t)\sim(p,q,(1-t)/2)$ for $t\in(0,\eta/2)$. Then define\[\begin{array}{l}\mathcal{M}(\bar{a},\bar{y}):=((\mathcal{M}(a,y)\setminus (M\times\mathcal{M}(a,x))\\
\qquad\qquad\qquad\qquad\qquad\qquad\sqcup (W\times \mathcal{M}(a,x)\setminus (M\times\mathcal{M}(a,x))))/\sim.\end{array}\]
\item If $b\in\text{Ob}(\mathscr{C})$ with $\mathcal{M}(y,b)\neq \emptyset$, then again take collar neighbourhoods $\mathcal{M}(y,b)\times M\times[0,\eta)$ and $\mathcal{M}(y,b)\times M\times(-\eta,0]$, now of \[\mathcal{M}(y,b)\times M\subset \partial_{|y|-|b|+1}\mathcal{M}(x,b)\] and $M\times \mathcal{M}(y,b)\subset W\times\mathcal{M}(y,b)$ respectively. Similarly to the overlay before, use $(q, p, t)\sim(q,p,(1-t)/2)$ for $t\in(0,\eta/2)$ to define\[\begin{array}{l}\mathcal{M}(\bar{x},\bar{b}):=((\mathcal{M}(x,b)\setminus (M\times\mathcal{M}(y,b))\\
\qquad\qquad\qquad\qquad\qquad\qquad\sqcup (W\times \mathcal{M}(y,b)\setminus (M\times\mathcal{M}(y,b))))/\sim.\end{array}\]
\item If $a,b\in\text{Ob}(\mathscr{C})$ with both $\mathcal{M}(a,x), \mathcal{M}(y,b)\neq \emptyset$, then the following is a subset of the boundary $\partial \mathcal{M}(a,b)$: \[\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)\subset \partial_{|y|-|b|+1}\mathcal{M}(a,b)\cap\partial_{|x|-|b|+1}\mathcal{M}(a,b).\]Take two 1--dimensional collars neighbourhoods of $\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)$. The first collar is taken inside $\partial_{|x|-|b|+1}\mathcal{M}(a,b)$, and perpendicular to the boundary $\partial_{|y|-|b|+1}\mathcal{M}(a,b)$, the second collar is taken \emph{vice versa}. Note that the product of the collars then defines a standard open neighbourhood inside the full moduli space $\mathcal{M}(a,b)$:\begin{equation}\label{eq:nhood}\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)\times[0,2\eta)\times[0,2\eta)\subset \mathcal{M}(a,b).\end{equation}In order to push $\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)$ out of the corner and replace it with $\mathcal{M}(y,b)\times M'\times\mathcal{M}(a,x)$, we will apply Lemma \ref{lem:smooth} to the standard open neighbourhood in line (\ref{eq:nhood}). Precisely, the embedded cobordism rel.\ boundary we will use as the input for the Lemma \ref{lem:smooth} is $(\mathcal{M}(y,b)\times W\times\mathcal{M}(a,x),\iota_{y,b}\times\tilde{\iota}\times\iota_{a,x})$ and the space $(X,\iota_X)$ is $(\mathcal{M}(a,b),\iota_{a,b})$. Now identify the two collar directions of $\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)$ with axes of $\mathbb{R}_+\times\mathbb{R}_+$ so that we have $U_{2\eta}\subset [0,2\eta)\times [0,2\eta)$, the open ball of radius $2\eta$ from Lemma \ref{lem:smooth}. Corresponding to our inputs $X$ and $\mathcal{M}(y,b)\times W\times\mathcal{M}(a,x)$, there are also spaces $X'$ and $Y$ defined by Lemma \ref{lem:smooth}. Hence we define \[\mathcal{M}(\bar{a},\bar{b}):=X'\sqcup Y/\sim,\]where $\sim$ is the equivalence relation defined in Lemma \ref{lem:smooth}. Note that this construction agrees with the constructions in (2) and (3), above, when considering the induced moduli spaces $\mathcal{M}(\bar{x},\bar{b})$ and $\mathcal{M}(\bar{a},\bar{y})$ in the boundary of $\mathcal{M}(\bar{a},\bar{b})$.
\item In all other cases define $\mathcal{M}(\bar{a},\bar{b})=\mathcal{M}(a,b)$. Note this includes the case of $c\in\text{Ob}({\mathscr{C}})$ with $\mathcal{M}(x,c)\neq\emptyset\neq\mathcal{M}(c,y)$.
\end{enumerate}
We now describe the embedding and framing $(\iota_W,\phi_W)$. For the case (1), we have $((\iota_W)_{\bar{x},\bar{y}}, (\phi_W)_{\bar{x},\bar{y}}):=(\iota',\phi')$, and for the case (5) $((\iota_W)_{\bar{a},\bar{b}}, (\phi_W)_{\bar{a},\bar{b}}):=(\iota_{{a},{b}}, \phi_{{a},{b}})$. In cases (2), (3) and (4), the embeddings and framings differ from those of $(\mathscr{C},\iota,\phi)$ only in a small neighbourhood of the glueing regions. The difference is determined by the embedding and framing of $Z$ which we defined in Lemma \ref{lem:smooth}.
\end{definition}
Finally we deduce the main result of this section.
\begin{proof}[Proof of Theorem \ref{thm:whitney}]First, we will assume $\textbf{d}$ has been enlarged enough that the framed cobordism rel.\ boundary $(W,\tilde{\iota},\tilde{\phi})$ is an embedding in $\mathbb{E}^\textbf{d}[|x|:|y|]\times[0,1]$. This does not affect the eventual stable homotopy type of the framed flow category.
Now, for each $a\in\text{Ob}(\mathscr{C})$, we will define a continuous map\[F_a:[0,1]\times\partial\mathcal{C}(a)\to X^{|a|-1};\qquad (t,p)\mapsto F_{a,t}(p),\]where $X^i$ is defined inductively for increasing $i$ by setting $X^0=\{pt\}$,\[X^i=X^{i-1}\cup_{F_a}\left([0,1]\times\mathcal{C}(a)\right),\]and with the union taken over all $a$ such that $|a|=i$. Furthermore, the maps $F_a$ will be defined such that $F_{a,0}$ and $F_{a,1}$ are the attaching maps for cells $\mathcal{C}(a)$ and $\mathcal{C}(\bar{a})$ in the CW-complexes $|\mathscr{C}|$ and $|\mathscr{C}_W|$ respectively. As such, the space $X:=\cup_iX^i$ can easily be seen to deformation retract onto each of $|\mathscr{C}|$ and $|\mathscr{C}_W|$, so that they are homotopy equivalent to one another.
It remains to define the $F_a$. Recall that $\mathcal{M}(a,b)\times\mathcal{C}(b)$ is embedded as \[\begin{array}{l}\mathcal{C}_b(a)=[0,R]\times[-R,R]^{d_B}\times\dots\times[0,R]\times[-R,R]^{|b|-1}\times\{0\}\times\mathcal{C}_{b,1}\\ \qquad\qquad\qquad\qquad\qquad \times\{0\}\times[-\epsilon,\epsilon]^{|a|}\times\dots\times\{0\}\times [-\epsilon,\epsilon]^{d_{A-1}}\subset\partial\mathcal{C}(a),\end{array}\]where $\mathcal{C}_{b,1}$ is the subset of $\mathbb{E}^\textbf{d}[|a|:|b|]$ given by the framed embedding of $\mathcal{M}(a,b)$. The cell attaching map for $\mathcal{C}(a)$ is given on $\mathcal{C}_b(a)\cong \mathcal{M}(a,b)\times\mathcal{C}(b)$ by the projection to $\mathcal{C}(b)$; and on $\partial\mathcal{C}(a)\setminus\bigcup_b\mathcal{C}_{b}(a)$ by mapping to the basepoint. To build $F_a$ we must take the framed embedding of each moduli space $\mathcal{M}(a,b)$ in turn, and `deform' the framed embedding to the framed embedding of $\mathcal{M}(\bar{a},\bar{b})$. Note that we do not need each time slice of the deformation to look precisely like a framed embedding whose exterior can be collapsed (indeed in general it will not). To show continuity of $F_a$ we will only need that the \emph{track} of the deformation has a framing. Precisely, for each $a,b\in\text{Ob}(\mathscr{C})$ we will define a framed embedded cobordism between $\mathcal{C}_{b,1}$ and $\mathcal{C}_{\bar{b},1}$, giving a subset $\mathcal{C}_{b,W}\subset\mathbb{E}^\textbf{d}[|a|:|b|]\times[0,1]$ which in turn defines an embedding \[\begin{array}{l}[0,R]\times[-R,R]^{d_B}\times\dots\times[0,R]\times[-R,R]^{|b|-1}\times\{0\}\times\mathcal{C}_{b,W}\\ \qquad\qquad\qquad\qquad \times\{0\}\times[-\epsilon,\epsilon]^{|a|}\times\dots\times\{0\}\times [-\epsilon,\epsilon]^{d_{A-1}}\subset[0,1]\times\partial\mathcal{C}(a),\end{array}\]which is homeomorphic to $Z_{a,b}\times\mathcal{C}(b)$ (for some space $Z_{a,b}$, to be defined below). We can then define $F_a:[0,1]\times\partial\mathcal{C}(a)\to X^{|a|-1}$ as follows. On $Z_{a,b}\times\mathcal{C}(b)$, it is the projection to $\mathcal{C}(b)$; and on $[0,1]\times\partial\mathcal{C}(a)\setminus\bigcup_b\left(Z_{a,b}\times\mathcal{C}(b)\right)$, we map to the basepoint.
So we proceed to modify $\mathcal{M}(a,b)$ in a series of case analyses based on (1), (2), (3), (4), of Definition \ref{def:extW}. Needless to say, the case of (5) requires no modification of $\mathcal{M}(a,b)$. In this case, define a subspace $\mathcal{C}_{b,W}=\mathcal{C}_{b,1}\times[0,1]\subset\mathbb{E}^\textbf{d}[|a|:|b|]\times[0,1]$ and hence an embedding $Z_{a,b}\times\mathcal{C}(b)\subset \partial \mathcal{C}(a)$ where $Z_{a,b}$ is simply $\mathcal{M}(a,b)\times[0,1]$.
For the first nontrivial case, set $a=x$ and $b=y$. Consider the embedding $\iota\times\text{id}: M\times[0,1]\hookrightarrow \mathbb{E}^\textbf{d}[|x|:|y|]\times[0,1]$. Now use Lemma \ref{lem:smooth} and the framed cobordism $W$ to push $M$ out of the codimension 1 corner $\mathbb{E}^\textbf{d}[|x|:|y|]\times\{1\}$ and replace it with $M'$. This determines a subset $\mathcal{C}_{b,W}$ of $[-R,R]^{|y|}\times[0,R]\times\dots\times[0,R]\times[-R,R]^{|x|-1}\times[0,1]$. Moreover, this results in an embedding of $Z_{a,b}\times\mathcal{C}(y)$ in $[0,1]\times\partial\mathcal{C}(x)$ (here $Z_{a,b}:=Z$ refers to the result of applying Lemma \ref{lem:smooth}).
For the second case, set $a=x$, $b\neq y$ and $\mathcal{M}(x,b)\neq\emptyset$. Then we have \[\mathcal{M}(y,b)\times M\subset\partial\mathcal{M}(x,b)\subset \mathbb{E}^\textbf{d}[|y|:|b|]\times\{0\}\times\mathbb{E}^\textbf{d}[|x|:|y|]\] and in the normal direction to this boundary we will take a $2\eta$ open collar neighbourhood. Consider that the product \[\mathcal{M}(y,b)\times M\times[0,2\eta)\times[0,1]\subset\mathcal{M}(x,b)\times[0,1]\]has a copy of $\mathcal{M}(y,b)\times M$ in the corner $\left(\mathbb{E}^\textbf{d}[|y|:|b|]\times\{0\}\times\mathbb{E}^\textbf{d}[|x|:|y|]\right)\times\{0\}\times\{1\}$. Using Lemma \ref{lem:smooth} and the framed cobordism $\mathcal{M}(y,b)\times W$, we push $\mathcal{M}(y,b)\times M$ out of the corner and replace it with $\mathcal{M}(y,b)\times M'$. Note that all embeddings were framed, so that resultant $Z_{a,b}$, is a framed embedding determining a subset $\mathcal{C}_{b,W}$ of $[-R,R]^{|b|}\times[0,R]\times\dots\times[0,R]\times[-R,R]^{|x|-1}\times[0,1]$ and hence an embedding of $Z_{a,b}\times\mathcal{C}(b)$ in $[0,1]\times\partial\mathcal{C}(x)$. Note that this agrees with the construction from the first case:\[\mathcal{C}_{y,W}\cap[-R,R]^{|y|}\times[0,R]\times\dots\times[0,R]\times[-R,R]^{|x|-1}\times[0,1]= \mathcal{C}_{b,W},\]as each case used Lemma \ref{lem:smooth}.
We must now look at cells $a\neq x$. There are two cases to consider. First suppose $a\neq x$, $b=y$ and that $\mathcal{M}(a,y)\neq\emptyset$ and we wish to deform the embedding of $M\times\mathcal{M}(a,x)$ appropriately. But this case proceeds identically to the case above where $a=x$, $b\neq y$ and $\mathcal{M}(x,b)\neq\emptyset$, so we omit the details.
Finally, there is the case that $a\neq x$, $b\neq y$ and that $\mathcal{M}(a,y)\neq\emptyset\neq\mathcal{M}(x,b)$. This is treated by the same pushing out technique as before (but now we need to do it in codimension 3). Precisely, we must consider $\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)\subset \partial \mathcal{M}(a,b)$. But observe there is a copy of $\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)$ in the codimension 3 corner $(0,0,1)\in[0,2\eta)\times[0,2\eta)\times[0,1]$ of the cylinder on the 2-way collar neighbourhood \[\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)\times[0,2\eta)\times[0,2\eta)\times[0,1]\subset \mathcal{M}(a,b)\times[0,1].\]We push $\mathcal{M}(y,b)\times M\times\mathcal{M}(a,x)$ out of the corner using $\mathcal{M}(y,b)\times W\times\mathcal{M}(a,x)$ and we obtain $\mathcal{C}_{b,W}\cong Z_{a,b}\times\mathcal{C}(b)$ as in previous cases. This completes the case analysis.
\end{proof}
\begin{example}We illustrate a homotopy from the proof Theorem \ref{thm:whitney} in the case where $M, M'$ are 0--dimensional. Let $M$ be a point, $M'$ be 3 points and $W$ be the 2--component cobordism from $M$ to $M'$ as in Figure \ref{pic:cob}. We suppose $M$ is actually a corner of a larger space $N$, so that after the homotopy, $M'$ is in the corner of this larger space. This is shown in Figure \ref{pic:whit}, where we have drawn $N$ extending in the vertical axis, and drawn the framed cobordism $Z$ constructed in Theorem \ref{thm:whitney} by pushing $M$ out of the corner of $N\times\{1\}$ of $N\times[0,1]$ using $W$.
\begin{figure}
\begin{tikzpicture}
\draw (0,0) -- (5,0);
\draw (0,3) -- (5,3);
\draw (1,0) arc [radius=1, start angle=180, end angle= 0];
\draw (4,0) -- (4,3);
\draw [fill] (1,0) circle [radius=0.05];
\draw [fill] (3,0) circle [radius=0.05];
\draw [fill] (4,0) circle [radius=0.05];
\draw [fill] (4,3) circle [radius=0.05];
\node at (4.1,3.3) {M};
\node at (5.35,3.15) {$\mathbb{E}^\textbf{d}$};
\node at (5.35,0.15) {$\mathbb{E}^\textbf{d}$};
\node at (2.7,1.8) {$W$};
\draw [->] (2.9,1.7) -- (3.8,1.4);
\draw [->] (2.6,1.55) -- (2.5,1.1);
\node at (2.8,-0.5) {$M'$};
\draw [decorate,decoration={brace,amplitude=5pt,raise=8pt}] (0,0) -- (0,3);
\node [rotate=90] at (-0.9,1.5) {$\mathbb{E}^\textbf{d}\times[0,1]$};
\end{tikzpicture}
\caption{The cobordism $W$ between $M$ and $M'$}\label{pic:cob}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{cobord.tikz}
\caption{The cobordism $Z$}\label{pic:whit}
\end{figure}
\end{example}
\section{The disjoint union of three trefoils}
\label{sec:tref}
We make a sample calculation of a stable homotopy type using our framed flow category moves. The example framed flow category $\mathcal{C}_1$ that we shall start with is depicted in Figure \ref{fig:smashofMoore}. We shall show explicitly that $\mathcal{C}_1$ is equivalent via our flow category moves to the framed flow category $\mathcal{C}_5$ depicted in Figure \ref{fig:smashofMooreIV}. The form of $\mathcal{C}_5$ corresponds to the Baues--Hennes classification of low homological width stable homotopy types and this is discussed further towards the end of the section.
We choose to consider the framed flow category $\mathcal{C}_1$ since it arises in a particular context, namely as a framed flow category associated to the disjoint union of three trefoils by the techniques of \cite{JLS}.
\begin{definition}
\label{defn:cC1}
There are eight objects in the category $\mathcal{C}_1$ depicted in Figure \ref{fig:smashofMoore}, each labelled by a 3--tuple. Any non-empty $0$--dimensional moduli space consists of two points each, as shown.
A letter `p' indicates that the point is framed positively, while a letter `m' indicates a negative framing.
The $1$--dimensional moduli spaces are then given by four intervals each, with all of them framed $0$.
Each such interval of course has two endpoints, and these are given as follows.
\begin{tikzpicture}[scale=0.8]
\node at (0,0) {$\mathcal{M}(333,223)=$};
\node[label=below:$P_0\mathscr{P}_0$, label=above:{\color{blue}$233$}] (5_5_1)at (1.9,0) {};
\node[label=below:$\tilde{P}_0\mathscr{M}_0$, label=above:{\color{blue}$323$}] (5_5_2)at (3.7,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$P_0\mathscr{P}_1$, label=above:{\color{blue}$233$}] (5_5_3)at (5.2,0) {};
\node[label=below:$\tilde{P}_1\mathscr{M}_0$, label=above:{\color{blue}$323$}] (5_5_4)at (7,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$P_1\mathscr{P}_0$, label=above:{\color{blue}$233$}] (5_5_5)at (8.5,0) {};
\node[label=below:$\tilde{P}_0\mathscr{M}_1$, label=above:{\color{blue}$323$}] (5_5_6)at (10.3,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$P_1\mathscr{P}_1$, label=above:{\color{blue}$233$}] (5_5_7)at (11.8,0) {};
\node[label=below:$\tilde{P}_1\mathscr{M}_1$, label=above:{\color{blue}$323$}] (5_5_8)at (13.6,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\node at (0,-2) {$\mathcal{M}(333,232)=$};
\node[label=below:$M_0\mathscr{P}_0$, label=above:{\color{blue}$233$}] (5_5_1)at (1.9,-2) {};
\node[label=below:$\bar{P}_0\tilde{\mathscr{P}}_0$, label=above:{\color{blue}$332$}] (5_5_2)at (3.7,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$M_0\mathscr{P}_1$, label=above:{\color{blue}$233$}] (5_5_3)at (5.2,-2) {};
\node[label=below:$\bar{P}_1\tilde{\mathscr{P}}_0$, label=above:{\color{blue}$332$}] (5_5_4)at (7,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$M_1\mathscr{P}_0$, label=above:{\color{blue}$233$}] (5_5_5)at (8.5,-2) {};
\node[label=below:$\bar{P}_0\tilde{\mathscr{P}}_1$, label=above:{\color{blue}$332$}] (5_5_6)at (10.3,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$M_1\mathscr{P}_1$, label=above:{\color{blue}$233$}] (5_5_7)at (11.8,-2) {};
\node[label=below:$\bar{P}_1\tilde{\mathscr{P}}_1$, label=above:{\color{blue}$332$}] (5_5_8)at (13.6,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\node at (0,-4) {$\mathcal{M}(333,322)=$};
\node[label=below:$\tilde{M}_0\mathscr{M}_0$, label=above:{\color{blue}$323$}] (5_5_1)at (1.9,-4) {};
\node[label=below:$\bar{M}_0\tilde{\mathscr{P}}_0$, label=above:{\color{blue}$332$}] (5_5_2)at (3.7,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$\tilde{M}_0\mathscr{M}_1$, label=above:{\color{blue}$323$}] (5_5_3)at (5.2,-4) {};
\node[label=below:$\bar{M}_1\tilde{\mathscr{P}}_0$, label=above:{\color{blue}$332$}] (5_5_4)at (7,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$\tilde{M}_1\mathscr{M}_0$, label=above:{\color{blue}$323$}] (5_5_5)at (8.5,-4) {};
\node[label=below:$\bar{M}_0\tilde{\mathscr{P}}_1$, label=above:{\color{blue}$332$}] (5_5_6)at (10.3,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$\tilde{M}_1\mathscr{M}_1$, label=above:{\color{blue}$323$}] (5_5_7)at (11.8,-4) {};
\node[label=below:$\bar{M}_1\tilde{\mathscr{P}}_1$, label=above:{\color{blue}$332$}] (5_5_8)at (13.6,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\node at (0,-6) {$\mathcal{M}(233,222)=$};
\node[label=below:$p_0P_0$, label=above:{\color{blue}$223$}] (5_5_1)at (1.9,-6) {};
\node[label=below:$\tilde{p}_0M_0$, label=above:{\color{blue}$232$}] (5_5_2)at (3.7,-6) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$p_0P_1$, label=above:{\color{blue}$223$}] (5_5_3)at (5.2,-6) {};
\node[label=below:$\tilde{p}_1M_0$, label=above:{\color{blue}$232$}] (5_5_4)at (7,-6) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$p_1P_0$, label=above:{\color{blue}$223$}] (5_5_5)at (8.5,-6) {};
\node[label=below:$\tilde{p}_0M_1$, label=above:{\color{blue}$232$}] (5_5_6)at (10.3,-6) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$p_1P_1$, label=above:{\color{blue}$223$}] (5_5_7)at (11.8,-6) {};
\node[label=below:$\tilde{p}_1M_1$, label=above:{\color{blue}$232$}] (5_5_8)at (13.6,-6) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\node at (0,-8) {$\mathcal{M}(323,222)=$};
\node[label=below:$p_0\tilde{P}_0$, label=above:{\color{blue}$223$}] (5_5_1)at (1.9,-8) {};
\node[label=below:$\bar{p}_0\tilde{M}_0$, label=above:{\color{blue}$322$}] (5_5_2)at (3.7,-8) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$p_0\tilde{P}_1$, label=above:{\color{blue}$223$}] (5_5_3)at (5.2,-8) {};
\node[label=below:$\bar{p}_1\tilde{M}_0$, label=above:{\color{blue}$322$}] (5_5_4)at (7,-8) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$p_1\tilde{P}_0$, label=above:{\color{blue}$223$}] (5_5_5)at (8.5,-8) {};
\node[label=below:$\bar{p}_0\tilde{M}_1$, label=above:{\color{blue}$322$}] (5_5_6)at (10.3,-8) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$p_1\tilde{P}_1$, label=above:{\color{blue}$223$}] (5_5_7)at (11.8,-8) {};
\node[label=below:$\bar{p}_1\tilde{M}_1$, label=above:{\color{blue}$322$}] (5_5_8)at (13.6,-8) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\end{tikzpicture}
\begin{tikzpicture}[scale=0.8]
\node at (0,-4) {$\mathcal{M}(332,222)=$};
\node[label=below:$\tilde{p}_0\bar{P}_0$, label=above:{\color{blue}$232$}] (5_5_1)at (1.9,-4) {};
\node[label=below:$\bar{p}_0\bar{M}_0$, label=above:{\color{blue}$322$}] (5_5_2)at (3.7,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$\tilde{p}_0\bar{P}_1$, label=above:{\color{blue}$232$}] (5_5_3)at (5.2,-4) {};
\node[label=below:$\bar{p}_1\bar{M}_0$, label=above:{\color{blue}$322$}] (5_5_4)at (7,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$\tilde{p}_1\bar{P}_0$, label=above:{\color{blue}$232$}] (5_5_5)at (8.5,-4) {};
\node[label=below:$\bar{p}_0\bar{M}_1$, label=above:{\color{blue}$322$}] (5_5_6)at (10.3,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$\tilde{p}_1\bar{P}_1$, label=above:{\color{blue}$232$}] (5_5_7)at (11.8,-4) {};
\node[label=below:$\bar{p}_1\bar{M}_1$, label=above:{\color{blue}$322$}] (5_5_8)at (13.6,-4) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\end{tikzpicture}
The $2$--dimensional moduli space $\mathcal{M}(333,222)$ consists of four hexagons. It will follow from the Baues--Hennes classification \cite{MR1113684} that the associated stable homotopy type is determined by the action of $\Sq^1$ and $\Sq^2$. These operations are in turn determined by the framed moduli spaces of dimensions $0$ and $1$, so we will not keep track of the $2$-dimensional moduli space.
\end{definition}
For those less interested in the Khovanov stable homotopy type, the next proposition may be skipped. We focus on quantum degree $q=21$ of the Khovanov stable homotopy type of the disjoint union of three right-handed trefoils, since we know that there is a non-trivial $\Sq^3$ in this degree. Indeed, the general formula of \cite{LawLipSar} confirms the existence of the smash of Moore spaces $$M(\mathbb{Z}/2\mathbb{Z},2)\wedge M(\mathbb{Z}/2\mathbb{Z},2) \wedge M(\mathbb{Z}/2\mathbb{Z},2)$$ as a wedge summand.
\begin{proposition}
\label{prop:eg_arises_from_tref}
Let $L$ be the disjoint union of three right-handed trefoils. The technique of \cite{JLS} constructs a framed flow category $\mathscr{L}^{\mathrm{Kh}}(L)$.
In quantum degree $q=21$, this framed flow category is the disjoint union of the framed flow category $\mathcal{C}_1$ described in Definition \ref{defn:cC1} with some other framed flow category.
\end{proposition}
\begin{proof}
The following calculation is similar to \cite[\S 4.3]{JLSmorse}. In homological degree $9$ we get exactly one object which is based at $(3,3,3)$. The smoothing of this object consists of three circles, each of which is decorated with a $-$. The objects of homological degree $8$ are based at $(2,3,3)$, $(3,2,3)$ and $(3,3,2)$. Again the smoothings are three circles, and one of them is decorated with a $+$. We therefore get $9$ objects of degree $8$. Note however that only three of those have non-empty moduli space with the object of degree $9$, namely those where the $+$ corresponds to the position of the $2$ in the base triple.
We will now only consider those objects that have non-empty moduli spaces with the object based at $(3,3,3)$, as these objects will give rise exactly to the product of Moore spaces predicted above. We then get three objects of homological degree $7$, based at $(2,2,3)$, $(2,3,2)$ and $(3,2,2)$, where the circle corresponding to the $3$ is decorated $-$, and one object of homological degree $6$ based at $(2,2,2)$ with all three circles decorated $+$. It is easy to see that these objects do not share non-empty moduli spaces with other objects in $\mathscr{L}^{\mathrm{Kh}}(L)$.
\end{proof}
\begin{figure}[ht]
\begin{tikzpicture}[scale=0.85]
\node at (0,6) {$333$};
\node at (-2,4) {$233$};
\node at (0,4) {$323$};
\node at (2,4) {$332$};
\node at (-2,2) {$223$};
\node at (0,2) {$232$};
\node at (2,2) {$322$};
\node at (0,0) {$222$};
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{P}_0\mathscr{P}_1$} (-2,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{M}_0\mathscr{M}_1$} (0,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\tilde{\mathscr{P}}_0\tilde{\mathscr{P}}_1$} (2,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$P_0P_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7,pos=0.25] {$M_0M_1$} (0,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) -- node [above,sloped,scale=0.7,pos=0.25] {$\tilde{P}_0\tilde{P}_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) -- node [above,sloped,scale=0.7,pos=0.25] {$\tilde{M}_0\tilde{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7,pos=0.25] {$\bar{P}_0\bar{P}_1$} (0,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7] {$\bar{M}_0\bar{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,2) -- node [above,sloped,scale=0.7] {$p_0p_1$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,2) -- node [above,sloped,scale=0.7] {$\tilde{p}_0\tilde{p}_1$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,2) -- node [above,sloped,scale=0.7] {$\bar{p}_0\bar{p}_1$} (0,0);
\end{tikzpicture}
\caption{A subcategory $\mathcal{C}_1$ of $\mathscr{L}^{\mathrm{Kh}}(L)$.}
\label{fig:smashofMoore}
\end{figure}
We are now going to perform handle slides and Whitney tricks, simplifying the framed flow category $\mathcal{C}_1$ in a sequence of four propositions until we arrive at the `Baues--Hennes' category $\mathcal{C}_5$.
\begin{proposition}
\label{prop:cC1_to_cC2}
The framed flow category $\mathcal{C}_1$ is move-equivalent to a framed flow category $\mathcal{C}_2$, depicted in Figure \ref{fig:smashofMooreI}. In this figure the $1$--dimensional moduli space denoted $\eta$ is a non-trivially framed circle.
\end{proposition}
\begin{proof}
To begin, we slide $223$ over $232$ via a $(-)$--handle slide. We will write an overline above all objects to indicate the resulting flow category, but these overlines will disappear again in time for the next move. The affected moduli spaces are
\[
\mathcal{M}(\overline{333},\overline{232})=\mathcal{M}(333,232) \sqcup \mathcal{M}(333,223)
\]
with no change in framings by Proposition \ref{prop:framechanges}.1(a). Furthermore, both $\mathcal{M}(\overline{233},\overline{222})$ and $\mathcal{M}(\overline{323},\overline{222})$ get four new intervals, each framed $0$ by Proposition \ref{prop:framechanges}.2(a). Note these new intervals correspond to $\mathcal{M}(232,222)\times \mathcal{M}(a,223)$ with $a=233$ or $a=323$. The points $\tilde{p}_0,\tilde{p}_1\in \mathcal{M}(232,222)$ create new points $m_0,m_1\in \mathcal{M}(\overline{223},\overline{222})$, $P_0,P_1\in \mathcal{M}(233,223)$ create $\hat{P}_0,\hat{P}_1\in \mathcal{M}(\overline{233},\overline{232})$, and $\tilde{P}_0,\tilde{P}_1\in \mathcal{M}(323,223)$ create $\check{P}_0,\check{P}_1\in \mathcal{M}(\overline{323},\overline{232})$.
The new intervals in $\mathcal{M}(\overline{233},\overline{222})$ are given by
\begin{tikzpicture}
\node[label=below:$m_0P_0$, label=above:{\color{blue}$\overline{223}$}] (5_5_1)at (0,0) {};
\node[label=below:$\tilde{p}_0\hat{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_2)at (1.6,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$m_0P_1$, label=above:{\color{blue}$\overline{223}$}] (5_5_3)at (3.3,0) {};
\node[label=below:$\tilde{p}_0\hat{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_4)at (4.9,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$m_1P_0$, label=above:{\color{blue}$\overline{223}$}] (5_5_5)at (6.6,0) {};
\node[label=below:$\tilde{p}_1\hat{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_6)at (8.2,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$m_1P_1$, label=above:{\color{blue}$\overline{223}$}] (5_5_7)at (9.9,0) {};
\node[label=below:$\tilde{p}_1\hat{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_8)at (11.5,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\end{tikzpicture}
and similarly in $\mathcal{M}(\overline{323},\overline{222})$ they are given by
\begin{tikzpicture}
\node[label=below:$m_0\tilde{P}_0$, label=above:{\color{blue}$\overline{223}$}] (5_5_1)at (0,0) {};
\node[label=below:$\tilde{p}_0\check{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_2)at (1.6,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$m_0\tilde{P}_1$, label=above:{\color{blue}$\overline{223}$}] (5_5_3)at (3.3,0) {};
\node[label=below:$\tilde{p}_0\check{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_4)at (4.9,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$m_1\tilde{P}_0$, label=above:{\color{blue}$\overline{223}$}] (5_5_5)at (6.6,0) {};
\node[label=below:$\tilde{p}_1\check{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_6)at (8.2,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$0$};
\node[label=below:$m_1\tilde{P}_1$, label=above:{\color{blue}$\overline{223}$}] (5_5_7)at (9.9,0) {};
\node[label=below:$\tilde{p}_1\check{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_8)at (11.5,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$0$};
\end{tikzpicture}
We now perform the Whitney trick in $\mathcal{M}(\overline{223},\overline{222})$ with $p_0,m_0$ and with $p_1,m_1$. The result is that in $\mathcal{M}(\overline{233},\overline{222})$ and $\mathcal{M}(\overline{323},\overline{222})$ intervals are glued together. For example, the endpoint $p_0P_0$ in the old $\mathcal{M}(233,222)$ is identified with the endpoint $m_0P_0$ in one of the new intervals. In fact, in each case two intervals are glued together to form a new interval. Furthermore, by \cite[Proposition 3.3]{JLSmorse}, the framing value of each new interval is $1$. We then get
\begin{tikzpicture}[scale=0.8]
\node at (0,0) {$\mathcal{M}(\overline{233},\overline{222})=$};
\node[label=below:$\tilde{p}_0\hat{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_1)at (1.9,0) {};
\node[label=below:$\tilde{p}_0M_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_2)at (3.7,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\tilde{p}_0\hat{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_3)at (5.2,0) {};
\node[label=below:$\tilde{p}_1M_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_4)at (7,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\tilde{p}_1\hat{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_5)at (8.5,0) {};
\node[label=below:$\tilde{p}_0M_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_6)at (10.3,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\tilde{p}_1\hat{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_7)at (11.8,0) {};
\node[label=below:$\tilde{p}_1M_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_8)at (13.6,0) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$1$};
\node at (0,-2) {$\mathcal{M}(\overline{323},\overline{222})=$};
\node[label=below:$\tilde{p}_0\check{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_1)at (1.9,-2) {};
\node[label=below:$\bar{p}_0\tilde{M}_0$, label=above:{\color{blue}$\overline{322}$}] (5_5_2)at (3.7,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\tilde{p}_0\check{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_3)at (5.2,-2) {};
\node[label=below:$\bar{p}_1\tilde{M}_0$, label=above:{\color{blue}$\overline{322}$}] (5_5_4)at (7,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\tilde{p}_1\check{P}_0$, label=above:{\color{blue}$\overline{232}$}] (5_5_5)at (8.5,-2) {};
\node[label=below:$\bar{p}_0\tilde{M}_1$, label=above:{\color{blue}$\overline{322}$}] (5_5_6)at (10.3,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\tilde{p}_1\check{P}_1$, label=above:{\color{blue}$\overline{232}$}] (5_5_7)at (11.8,-2) {};
\node[label=below:$\bar{p}_1\tilde{M}_1$, label=above:{\color{blue}$\overline{322}$}] (5_5_8)at (13.6,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$1$};
\end{tikzpicture}
In the next step we again perform the Whitney trick, this time using $\hat{P}_0,M_0$ and $\hat{P}_1,M_1$ in $\mathcal{M}(\overline{233},\overline{232})$. Note that we now remove the overline from the objects. The effect on $\mathcal{M}(333,232)$ is that the eight intervals turn into four similarly to the case above. More precisely, we get
\begin{tikzpicture}[scale=0.8]
\node at (0,-2) {$\mathcal{M}(333,232)=$};
\node[label=below:$\check{P}_0\mathscr{M}_0$, label=above:{\color{blue}$323$}] (5_5_1)at (1.9,-2) {};
\node[label=below:$\bar{P}_0\tilde{\mathscr{P}}_0$, label=above:{\color{blue}$332$}] (5_5_2)at (3.7,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_1)--(5_5_2) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\check{P}_1\mathscr{M}_0$, label=above:{\color{blue}$323$}] (5_5_3)at (5.2,-2) {};
\node[label=below:$\bar{P}_1\tilde{\mathscr{P}}_0$, label=above:{\color{blue}$332$}] (5_5_4)at (7,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_3)--(5_5_4) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\check{P}_0\mathscr{M}_1$, label=above:{\color{blue}$323$}] (5_5_5)at (8.5,-2) {};
\node[label=below:$\bar{P}_0\tilde{\mathscr{P}}_1$, label=above:{\color{blue}$332$}] (5_5_6)at (10.3,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_5)--(5_5_6) node[sloped, above,pos=0.5, red]{$1$};
\node[label=below:$\check{P}_1\mathscr{M}_1$, label=above:{\color{blue}$323$}] (5_5_7)at (11.8,-2) {};
\node[label=below:$\bar{P}_1\tilde{\mathscr{P}}_1$, label=above:{\color{blue}$332$}] (5_5_8)at (13.6,-2) {};
\draw[shorten >=-0.1cm,shorten <=-0.1cm,|-|] (5_5_7)--(5_5_8) node[sloped, above,pos=0.5, red]{$1$};
\end{tikzpicture}
The moduli space $\mathcal{M}(233,222)$ turns into a closed manifold. In fact, the outer intervals result in one circle each, and the inner two intervals are glued together along their endpoints to form a single circle. By \cite[Proposition 3.4]{JLSmorse} all circles are labelled with $0$, which means that each circle is \emph{non-trivially} framed (compare \cite{JLSmorse} for framing conventions). Using the extended Whitney trick, we can reduce this to one non-trivially framed circle, which we denote by $\mathcal{M}(233,222)=\eta$.
The result is the framed flow category $\mathcal{C}_2$ depicted in Figure \ref{fig:smashofMooreI}.
\end{proof}
\begin{proposition}
\label{prop:cC2_to_cC3}
The framed flow category $\mathcal{C}_2$ is move equivalent to the framed flow category $\mathcal{C}_3$ depicted in Figure \ref{fig:smashofMooreII}. In Figure \ref{fig:smashofMooreII} we denote non-trivially framed circles either by $\xi$ or $\eta$.
\end{proposition}
\begin{proof}
Starting from this category, we slide $322$ over $232$ with a $(-)$--handle slide. This introduces extra points $M_0,M_1$ in $\mathcal{M}(\overline{332},\overline{232})$ and $\check{M}_0,\check{M}_1$ in $\mathcal{M}(\overline{323},\overline{232})$, as well as $m_0,m_1$ in $\mathcal{M}(\overline{322},\overline{222})$. Similar to the last slide in Proposition \ref{prop:cC1_to_cC2}, the moduli spaces $\mathcal{M}(\overline{333},\overline{232})$, $\mathcal{M}(\overline{323},\overline{222})$ and $\mathcal{M}(\overline{332},\overline{222})$ each acquire four new intervals.
\begin{figure}[ht]
\begin{tikzpicture}[scale=0.85]
\node at (0,6) {$333$};
\node at (-2,4) {$233$};
\node at (0,4) {$323$};
\node at (2,4) {$332$};
\node at (-2,2) {$223$};
\node at (0,2) {$232$};
\node at (2,2) {$322$};
\node at (0,0) {$222$};
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{P}_0\mathscr{P}_1$} (-2,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{M}_0\mathscr{M}_1$} (0,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\tilde{\mathscr{P}}_0\tilde{\mathscr{P}}_1$} (2,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$P_0P_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) -- node [above,sloped,scale=0.7] {$\check{P}_0\check{P}_1$} (0,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) -- node [above,sloped,scale=0.7,pos=0.25] {$\tilde{P}_0\tilde{P}_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) -- node [above,sloped,scale=0.7,pos=0.25] {$\tilde{M}_0\tilde{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7,pos=0.25] {$\bar{P}_0\bar{P}_1$} (0,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7] {$\bar{M}_0\bar{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,2) -- node [above,sloped,scale=0.7] {$\tilde{p}_0\tilde{p}_1$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,2) -- node [above,sloped,scale=0.7] {$\bar{p}_0\bar{p}_1$} (0,0);
\end{tikzpicture}
\caption{The framed flow category $\mathcal{C}_2$.}
\label{fig:smashofMooreI}
\end{figure}
Performing all obvious extended Whitney tricks (0--dimensional and 1--dimensional) now leads to the flow category indicated in Figure \ref{fig:smashofMooreII}. One easily checks that $\mathcal{M}(333,232)$, $\mathcal{M}(323,222)$ and $\mathcal{M}(332,222)$ all turn into three circles, and after extended Whitney tricks all contain one non-trivially framed circle. We denote the non-trivially framed circle in $\mathcal{M}(333,232)$ by $\xi$, mainly to follow the conventions in \cite{MR1113684}.
\end{proof}
\begin{figure}[ht]
\begin{tikzpicture}[scale=0.8]
\node at (0,6) {$333$};
\node at (-2,4) {$233$};
\node at (0,4) {$323$};
\node at (2,4) {$332$};
\node at (-2,2) {$223$};
\node at (0,2) {$232$};
\node at (2,2) {$322$};
\node at (0,0) {$222$};
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{P}_0\mathscr{P}_1$} (-2,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{M}_0\mathscr{M}_1$} (0,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\tilde{\mathscr{P}}_0\tilde{\mathscr{P}}_1$} (2,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$P_0P_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) to [bend right=20] node [above,sloped,scale=0.7,pos=0.75] {$\xi$} (0,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) to [bend left=20] node [above,sloped,scale=0.7,pos=0.25] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) -- node [above,sloped,scale=0.7,pos=0.35] {$\tilde{P}_0\tilde{P}_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) -- node [above,sloped,scale=0.7,pos=0.35] {$\tilde{M}_0\tilde{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7] {$\bar{M}_0\bar{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,2) -- node [below,sloped,scale=0.7] {$\tilde{p}_0\tilde{p}_1$} (0,0);
\end{tikzpicture}
\caption{The framed flow category $\mathcal{C}_3$.}
\label{fig:smashofMooreII}
\end{figure}
\begin{proposition}
\label{prop:cC3_to_cC4}
The framed flow category $\mathcal{C}_3$ is move equivalent to the framed flow category $\mathcal{C}_4$ depicted in Figure \ref{fig:smashofMooreIII}.
\end{proposition}
\begin{proof}
Starting from Figure \ref{fig:smashofMooreII} we can perform two $(-)$--handle slides, sliding $323$ over both $233$ and $332$. The moduli spaces $\mathcal{M}(\overline{333},\overline{223})$ and $\mathcal{M}(\overline{333},\overline{322})$ each consist of eight intervals, and $\mathcal{M}(\overline{323},\overline{222})$ contains exactly three non-trivially framed circles, which are together Whitney trick equivalent to a single non-trivially framed circle.
Note that in $\mathcal{M}(\overline{333},\overline{322})$ the framing of the four new intervals is different from the framing of the new intervals in $\mathcal{M}(\overline{333},\overline{223})$, but after performing the obvious Whitney tricks we get the flow category $\mathcal{C}_4$ depicted in Figure \ref{fig:smashofMooreIII}.
\end{proof}
\begin{figure}[ht]
\begin{tikzpicture}[scale=0.8]
\node at (0,6) {$333$};
\node at (-2,4) {$233$};
\node at (0,4) {$323$};
\node at (2,4) {$332$};
\node at (-2,2) {$223$};
\node at (0,2) {$232$};
\node at (2,2) {$322$};
\node at (0,0) {$222$};
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{M}_0\mathscr{M}_1$} (0,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$P_0P_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) to [bend right=20] node [above,sloped,scale=0.7,pos=0.75] {$\xi$} (0,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) to [bend left=20] node [above,sloped,scale=0.7,pos=0.25] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\xi$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\xi$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7] {$\bar{M}_0\bar{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,2) -- node [below,sloped,scale=0.7] {$\tilde{p}_0\tilde{p}_1$} (0,0);
\end{tikzpicture}
\caption{The framed flow category $\mathcal{C}_4$. Note that the associated cochain complex is in Smith normal form (see Section \ref{sec:smith}).}
\label{fig:smashofMooreIII}
\end{figure}
\begin{proposition}
\label{prop:cC4_to_cC5}
The framed flow category $\mathcal{C}_4$ is move equivalent to the framed flow category $\mathcal{C}_5$ depicted in Figure \ref{fig:smashofMooreIV}.
\end{proposition}
\begin{proof}
Finally, we slide $332$ over $323$, $233$ over $323$, and then $232$ over $223$ and $322$. Again, using the extended Whitney trick on the 1-dimensional moduli spaces leads to the flow category $\mathcal{C}_5$ depicted in Figure \ref{fig:smashofMooreIV}.
\end{proof}
\begin{figure}[ht]
\begin{tikzpicture}[scale=0.8]
\node at (0,6) {$333$};
\node at (-2,4) {$233$};
\node at (0,4) {$323$};
\node at (2,4) {$332$};
\node at (-2,2) {$223$};
\node at (0,2) {$232$};
\node at (2,2) {$322$};
\node at (0,0) {$222$};
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) -- node [above,sloped,scale=0.7] {$\mathscr{M}_0\mathscr{M}_1$} (0,4);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (-2,4) -- node [above,sloped,scale=0.7] {$P_0P_1$} (-2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,6) to [bend right=20] node [above,sloped,scale=0.7,pos=0.75] {$\xi$} (0,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,4) to [bend left=20] node [above,sloped,scale=0.7,pos=0.25] {$\eta$} (0,0);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (2,4) -- node [above,sloped,scale=0.7] {$\bar{M}_0\bar{M}_1$} (2,2);
\draw [shorten >=0.3cm,shorten <=0.3cm,->] (0,2) -- node [below,sloped,scale=0.7] {$\tilde{p}_0\tilde{p}_1$} (0,0);
\end{tikzpicture}
\caption{The flow category $\mathcal{C}_5$. The middle summand corresponds to a \emph{special cyclic word} in the sense of Baues-Hennes \cite{MR1113684}.}
\label{fig:smashofMooreIV}
\end{figure}
The flow category $\mathcal{C}_5$ clearly has two wedge-summands $M(\mathbb{Z}/2\mathbb{Z},7)$, and the remaining four objects form a stable space with the property that $\Sq^2\Sq^1=\Sq^1\Sq^2$ is non-trivial. It follows from the Decomposition Theorem of \cite[Theorem 3.9]{MR1113684} that the space corresponding to this category is the space called $X(\xi^2\eta_2,\text{id})$, where $\text{id}\colon V \to V$ is the identity of the $1$--dimensional $\mathbb{Z}/2\mathbb{Z}$--vector space $V$.
We note that $(\xi^2\eta_2,\text{id})$ is a \emph{special cyclic word} in the sense of \cite{MR1113684}\footnote{In \cite{MR1113684}, the notation used is actually $\xi^1\eta_1$, while in \cite{MR1361886} it is $\xi^2\eta_2$. We use the notation of \cite{MR1361886}.}, and the only indecomposable space in their list with $\Sq^2\Sq^1=\Sq^1\Sq^2$ non-trivial.
\section{A space level Smith normal form and further reductions}
\label{sec:smith}
It is discussed in \cite[Lemmas 3.25, 3.26]{MR3230817} that the stable homotopy type of $(\mathscr{C}, \iota, \phi)$ is not affected by an isotopy of the framing $(\iota,\phi)$, or by increasing $\textbf{d}$ to $\textbf{d}'\geq \textbf{d}$. In addition to these basic modifications, we say the following are permissible \emph{flow category moves}:
\begin{enumerate}
\item Handle cancellation
\item Extended Whitney trick
\end{enumerate}(Note the handle slide of Subsection \ref{subsec:sliding} is based on handle cancellation, so does not need to be included as an extra flow category move.)
\begin{definition}$(\mathscr{C}, \iota, \phi)$ and $(\mathscr{C}', \iota',\phi')$ are \emph{directly move equivalent} if $(\mathscr{C}', \iota',\phi')$ is the effect of a flow category move on $(\mathscr{C}, \iota, \phi)$, or vice versa. We say $(\mathscr{C}, \iota, \phi)$ and $(\mathscr{C}', \iota',\phi')$ are \emph{move equivalent} if they are related by a finite sequence of directly move equivalent framed flow categories. Clearly, move equivalence is an equivalence relation on the set of framed flow categories.
\end{definition}
We now consider a general method to simplify a framed flow category within its move equivalence class.
Recall that a finite chain complex $C_*$ of finitely generated projective modules over a principal ideal domain $A$ can always be put into \emph{Smith normal form}. That is, there is a basis $C_r\cong U_r\oplus V_r\oplus W_r$ such that the differentials $d_r:C_{r+1}\to C_r$ vanish on $V_r\oplus W_r$ and the matrix of $d_r|_{U_r}$ has the form\[\left(\begin{matrix}0&D&0\end{matrix}\right)^t,\]where $D$ is injective and diagonal with $D_{ii}|D_{i+1\,i+1}$. Writing $m(r)=\text{rk}_AV_r$, $n(r)=\text{rk}_A W_r$, the Smith normal form basis presents the homology as \[H_r(C;A)\cong A/D_{11}A\oplus\dots\oplus A/D_{m(r)m(r)}A\oplus A^{n(r)}.\]Note that some diagonal entries may be equal to 1.
If we are moreover allowed to add or remove cancelling $A$--module generators in adjacent homological degrees (which will correspond to elementary expansions or contractions of the matrices of $d_r$), the Smith normal form can be changed so that the diagonal entries of $D$ are all prime powers. Elementary expansions and contractions result in chain homotopy equivalences, so the homology is not affected. In this changed form, the basis presents the unique primary decomposition of the homology modules. Call this modified type of basis the \emph{primary Smith normal form}.
Recall that any framed flow category $(\mathscr{C},\iota,\phi)$ determines a based chain complex $(C_*,d)$. The basis of $C_r$ is given by the objects $x$ of $\mathscr{C}$ with $|x|=r$. If $|x|=r$ and $|y|=r-1$ then the $(x,y)$ entry in the matrix of $d_{r-1}$ is given by the signed count of the points in $\mathcal{M}(x,y)$.
\begin{theorem}\label{thm:SNF}Any framed flow category $(\mathscr{C},\iota,\phi)$ is move equivalent to some framed flow category whose based chain complex $(C_*,d)$ is in primary Smith normal form and such that the number of points in any 0--dimensional moduli space is exactly the corresponding entry in the matrix of the differential $d$.
\end{theorem}
\begin{proof}Suppose $(\mathscr{C},\iota,\phi)$ is a framed flow category. Any 0--dimensional moduli space $\mathcal{M}(x,y)$ consists of a certain number of positively framed points and a certain number of negatively framed points. By pairing a positive point with a negative point we may cancel them against each other using an extended Whitney trick. This will not affect any other 0--dimensional moduli spaces or the other components of $\mathcal{M}(x,y)$. By repeatedly doing this, we may assume that every non-empty 0--dimensional moduli space consists entirely of positively framed points or entirely of negatively framed points. Write this as a positive or negative integer $n_{x,y}$, and if $\mathcal{M}(x,y)$ is empty, set $n_{x,y}=0$. The chain complex of the framed flow category is then $C_*$, where $C_n$ is freely generated by the objects $x$ in grading level $n$ and $d_{n-1}x=\Sigma_{|y|={n-1}}n_{x,y}y$.
We wish to first put $C_*$ into Smith normal form. To do so we will show that certain basis changes for the $C_*$ can be realised via flow category moves. If $|x|=n$ then write $r_x$ for the row vector corresponding to $x$ in the matrix of $d_n$, and $c_x$ for the column vector in $d_{n-1}$. Now if $|x|=|y|$, then the reader may check that if we $(\pm)$--slide $x$ over $y$ this does not affect $r_x$ or $c_y$, but has the effect $r_y\mapsto r_y\mp r_x$ and $c_x\mapsto c_x\pm c_y$. Set $k=\min\{|x|\,|\,x\in\text{Ob}(\mathscr{C})\}$. We may now perform the row and column operations required to put $d_k$ into Smith normal form. As a result, $C_k\cong V_k\oplus W_k$ and $C_k\cong U_k\oplus V_k\oplus W_k$ as required. Next, observe that $\text{im} (d_{k+1})=V_{k+1}\subset \ker d_k$, so that performing the handle slides required to put $d_{k+1}$ into Smith normal form will not destroy the Smith normal form of $d_k$ (as there are no 0--dimensional moduli spaces between $V_{k+1}$ and $C_k$). We may now repeat this process on $d_r$, for increasing $r$, so that the whole of $C_*$ is in Smith normal form.
To modify the flow category so that the chain complex is in primary Smith normal form, we only need to know that for each $r$, we can make elementary matrix expansions or contractions of $d_r$, using flow category moves. But to make a matrix expansion, we simply introduce a pair of objects $x,y$ with $|x|=r+1$ and $|y|=r$ with $\mathcal{M}(x,y)$ a single positively framed point. Clearly this has the required effect on the matrix of $d_r$, and $x$ can be cancelled against $y$ via handle cancellation, so introducing these points is a permissible flow category move. To make an elementary matrix contraction we perform a handle cancellation on the objects corresponding to the cancelling chain complex generators.
\end{proof}
Here is an easy corollary:
\begin{corollary}\label{cor:moore}Suppose $(\mathscr{C}_1,\iota_1,\phi_1)$ and $(\mathscr{C}_2,\iota_2,\phi_2)$ are framed flow categories such that there is a homotopy equivalence $\mathcal{X}(\mathscr{C}_1)\simeq \mathcal{X}(\mathscr{C}_2)$, and that for some $n\in\mathbb{Z}$ and all $k>0$ the reduced homology $\widetilde{H}_*(\mathscr{C}_1;\mathbb{Z}/k\mathbb{Z})$ is only supported in degrees $n$ and $n+1$ (in particular this means the flow categories have the stable homotopy type of a wedge of Moore spaces). Then $(\mathscr{C}_1,\iota_1,\phi_1)$ and $(\mathscr{C}_2,\iota_2,\phi_2)$ are move equivalent.
\end{corollary}
In fact we suggest that much more is true.
\begin{conjecture}\label{conj:main} If two framed flow categories determine the same stable homotopy type then they are move equivalent to one another.
\end{conjecture}
We provide more evidence for this conjecture in a forthcoming paper where we show that something similar to Corollary \ref{cor:moore} is true but with the reduced homology now possibly supported in degrees $n$, $n+1$, $n+2$ and $n+3$. To do this, we will show how to use move equivalence to reduce the stable homotopy types of such flow categories to those associated to the Chang \cite{MR0036508} and Baues-Hennes \cite{MR1113684} homotopy classifications.
\bibliographystyle{plain}
\def\MR#1{}
|
\subsubsection*{Subtask \arabic{st}: #1}} \newcommand{\subtasksbegin}{} \newcommand{\subtasksend}{}
\newcommand{\newtext}[1]{{\color{red} #1}}
\renewcommand{\newcommand}{\providecommand}
\usepackage{import}
\usepackage{ifthen}
\usepackage{url}
\usepackage{amssymb}
\usepackage{amsmath}
\usepackage{import}
\usepackage{xparse}
\usepackage{amsmath}
\usepackage[usenames,dvipsnames]{xcolor}
\usepackage{xspace}
\usepackage{datatool}
\usepackage{glossaries}
\providecommand\m[1]{\ensuremath{#1}\xspace}
\renewcommand{\m}[1]{\ensuremath{#1}\xspace}
\newcommand{\trval}[1]{\m{{\bf #1}}}
\newcommand{\limplies}{\m{\Rightarrow}}
\newcommand{\limpl}{\limplies}
\newcommand{\lequiv}{\m{\Leftrightarrow}}
\newcommand{\limpliedby}{\m{\Leftarrow}}
\newcommand{\limplied}{\limpliedby}
\newcommand{\lrule}{\m{\leftarrow}}
\newcommand{\cause}{\m{\stackrel{c}{\lrule}}}
\newcommand{\rul}{\m{\leftarrow}}
\newcommand{\ltrue}{\trval{t}}
\newcommand{\lfalse}{\trval{f}}
\newcommand{\lunkn}{\trval{u}}
\newcommand{\lincon}{\trval{i}}
\newcommand{\bigand}{\m{\bigwedge}}
\newcommand{\bigor}{\m{\bigvee}}
\newcommand{\true}{\m{\top}}
\newcommand{\false}{\m{\bot}}
\newcommand{\Lra}{\m{\Leftrightarrow}}
\newcommand{\lra}{\m{\leftrightarrow}}
\newcommand{\Ra}{\m{\Rightarrow}}
\newcommand{\La}{\m{\Leftarrow}}
\newcommand{\ra}{\m{\rightarrow}}
\newcommand{\la}{\m{\leftarrow}}
\newcommand{\mim}{\limplies}
\newcommand{\equi}{\lequiv}
\newcommand{\Tr}{\ltrue}
\newcommand{\Fa}{\lfalse}
\newcommand{\Un}{\lunkn}
\newcommand{\In}{\lincon}
\newcommand{\voc}{\m{\Sigma}}
\newcommand{\invoc}{\m{\sigma_{in}}}
\newcommand{\outvoc}{\m{\sigma_{out}}}
\newcommand{\struct}{\m{I}}
\newcommand{\structx}{\m{I}}
\newcommand{\I}{\m{\mathcal{I}}}
\newcommand{\Iin}{\m{\I_{in}}}
\newcommand{\J}{\m{\mathcal{J}}}
\newcommand{\instruct}{\m{I_{in}}}
\newcommand{\outstruct}{\m{I_{out}}}
\newcommand{\theory}{\m{\mathcal{T}}}
\newcommand{\PP}{\m{\mathcal{P}}}
\newcommand{\LL}{\m{\mathcal{L}}}
\newcommand{\WW}{\m{\mathcal{W}}}
\newcommand{\BB}{\m{\mathcal{B}}}
\newcommand{\D}{\m{\Delta}}
\newcommand{\f}{\m{\varphi}}
\newcommand{\atom}{\m{a}}
\newcommand{\lit}{\m{l}}
\newcommand{\rules}{\m{R}}
\newcommand{\set}{\m{S}}
\NewDocumentCommand\inter{g+g}
\IfNoValueTF{#1}
{\struct}
{\m{#1^{#2}}}}
\newcommand{\partinter}{\m{J}}
\newcommand{\model}{\m{M}}
\newcommand{\fone}{\m{\varphi}}
\newcommand{\ftwo}{\m{\psi}}
\newcommand{\defined}[1]{\ensuremath{\mbox{\it Def}({#1})}\xspace}
\newcommand{\open}[1]{\ensuremath{\mbox{\it Open}({#1})}\xspace}
\newcommand{\pars}[1]{\ensuremath{\mbox{\it Par}({#1})}\xspace}
\newcommand{\just}{\m{just}}
\newcommand{\mx}[3]{\m{<#1, #2, #3>}}
\newcommand{\xxx}{\m{\overline{x}}}
\newcommand{\XXX}{\m{\overline{X}}}
\newcommand{\yyy}{\m{\overline{y}}}
\newcommand{\zzz}{\m{\overline{z}}}
\newcommand{\ddd}{\m{\overline{d}}}
\newcommand{\eee}{\m{\overline{e}}}
\newcommand{\ccc}{\m{\overline{c}}}
\newcommand{\bracketddd}{\m{\big(\overline{d}\big)}}
\newcommand{\bddd}{\m{\big(\overline{d}\big)}}
\newcommand{\DDD}{\m{\overline{D}}}
\newcommand{\vvv}{\m{\overline{v}}}
\newcommand{\ttt}{\m{\overline{t}}}
\newcommand{\aaa}{\m{\overline{a}}}
\newcommand{\bbb}{\m{\overline{b}}}
\providecommand{\lll}{\m{\overline{l}}
\renewcommand{\lll}{\m{\overline{l}}}
\newcommand{\TTT}{\m{\overline{T}}}
\newcommand{\elim}{\m{\backslash}}
\newcommand{\transformarrow}{\pmb{\pmb\rightarrowtail}}
\newcommand{\rewrite}[2]{\m{#1 \ \transformarrow\ #2 }}
\newcommand{\bool}{\m{\mathbb{B}}}
\newcommand{\Bool}{\bool}
\newcommand{\nat}{\m{\mathbb{N}}}
\newcommand{\Nat}{\nat}
\renewcommand{\int}{\m{\mathbb{Z}}}
\newcommand{\real}{\m{\mathbb{R}}}
\newcommand{\rat}{\m{\mathbb{Q}}}
\newcommand{\leqp}{{\m{\,\leq_p\,}}}
\newcommand{\geqp}{{\m{\,\geq_p\,}}}
\newcommand{\leqk}{{\m{\,\leq_k\,}}}
\newcommand{\geqk}{{\m{\,\geq_k\,}}}
\newcommand{\leqt}{{\m{\,\leq_t\,}}}
\newcommand{\geqt}{{\m{\,\geq_t\,}}}
\DeclareMathOperator\glb{glb}
\DeclareMathOperator\lub{lub}
\DeclareMathOperator\lfp{lfp}
\DeclareMathOperator\gfp{gfp}
\newcommand{\val}{\m{\nu}}
\newcommand{\superval}{\m{sv}}
\newcommand{\kleeneval}{\m{Kl}}
\newcommand{\typed}[2]{\m{#1\in #2:}}
\newcommand{\hasmodel}{\mid\!\equiv}
\NewDocumentCommand\subs{g+g}
\IfNoValueTF{#1}
{\m{/}}
{\m{#1/ #2}}}
\newcommand{\substitute}[2]{\subs{#1}{#2}}
\newcommand{\func}[1]{\m{f(#1)}}
\newcommand{\setof}[1]{\m{\left \{ #1 \right \}}}
\newcommand{\tuple}[1]{\m{\left \langle #1 \right \rangle }}
\newcommand{\til}{\m{\sim}}
\newcommand\eqdef{\mathrel{\overset{\makebox[0pt]{\mbox{\normalfont\tiny\sffamily def}}}{=}}}
\newcommand{\logicname}[1]{\textsc{#1}\xspace}
\newcommand{\idp}{\logicname{IDP}}
\newcommand{\xsb}{\logicname{XSB}}
\newcommand{\idptwo}{\logicname{IDP$^2$}}
\newcommand{\idpthree}{\logicname{IDP3}}
\newcommand{\idpfour}{\logicname{IDP4}}
\newcommand{\idpdraw}{\logicname{ID$^{P}_{Draw}$}}
\newcommand{\idpide}{\logicname{ID$^{P}_{E}$}}
\newcommand{\minisat}{\logicname{MiniSAT}}
\newcommand{\minisatid}{\logicname{MiniSAT(ID)}}
\newcommand{\gidl}{\logicname{GidL}}
\newcommand{\sts}{\logicname{SAT-to-SAT}}
\newcommand{\fodotidp}{\logicname{FO(\ensuremath{\cdot})\ensuremath{^{\mathtt{IDP}}}}}
\newcommand{\foidp}{\fodotidp}
\newcommand{\fodot}{\logicname{FO(\ensuremath{\cdot})}}
\newcommand{\pcdot}{\logicname{PC(\ensuremath{\cdot})}}
\newcommand{\foid}{\logicname{FO(ID)}}
\newcommand{\foidaggpf}{\logicname{FO(ID,Agg,PF)}}
\newcommand{\foidaggpft}{\logicname{FO(ID,\allowbreak Agg,\allowbreak PF,\allowbreak T)}}
\newcommand{\foidplus}{\logicname{C-Log}}
\newcommand{\clog}{\logicname{C-Log}}
\newcommand{\foclog}{\logicname{FO(C)}}
\newcommand{\hoid}{\logicname{HO(ID)}}
\newcommand{\hopfid}{\logicname{HO(PF,ID)}}
\newcommand{\fo}{\FO}
\newcommand{\esoid}{\logicname{\ensuremath{\exists}SO(ID)}}
\newcommand{\cpl}{\logicname{CP}-logic\xspace}
\newcommand{\aspcore}{\text{\sc ASP-Core-2}\xspace}
\newcommand{\ouracronym}[3]
\newacronym{#1}{#2}{#3}
\expandafter\newcommand\csname #1\endcsname{\gls{#1}\xspace
}
\ouracronym{FO}{FO}{first-order logic}
\ouracronym{MX}{MX}{Model Expansion}
\ouracronym{MO}{MO}{Model Optimization}
\ouracronym{ASP}{ASP}{Answer Set Programming}
\ouracronym{TP}{TP}{Theorem Proving}
\ouracronym{LP}{LP}{Logic Programming}
\ouracronym{CP}{CP}{Constraint Programming}
\ouracronym{FP}{FP}{Functional Programming}
\ouracronym{KR}{KR}{Knowledge Representation}
\ouracronym{CSP}{CSP}{Constraint Satisfaction Problem}
\ouracronym{SMT}{SMT}{SAT Modulo Theories}
\ouracronym{KBS}{KBS}{Knowledge Base System}
\ouracronym{NNF}{NNF}{Negation Normal Form}
\ouracronym{FNNF}{FNNF}{Flat Negation Normal Form}
\ouracronym{DefNNF}{DefNNF}{Definition Negation Normal Form}
\newcommand{\DEFNNF}{\DefNNF}
\ouracronym{CDCL}{CDCL}{Conflict-Driven Clause-Learning}
\ouracronym{LCG}{LCG}{Lazy Clause Generation}
\ouracronym{AEL}{AEL}{Autoepistemic Logic}
\ouracronym{OEL}{OEL}{Ordered Epistemic Logic}
\ouracronym{AFT}{AFT}{Approximation Fixpoint Theory}
\makeatletter
\def\ifenv#1{
\def\@tempa{#1
\def\@ttempa{#1*
\ifx\@tempa\@currenvir
\expandafter\@firstoftwo
\else
\expandafter\@secondoftwo
\fi
}
\makeatother
\newcommand{\ddrule}[4]{\ensuremath{#1 \leftarrow #2 & \{#3\} & #4}}
\newcommand{\drule}[2]{\ensuremath{#1 & \leftarrow & #2}}
\newcommand{\darule}[4]{\ensuremath{#1 \leftarrow #2 & \{#3\} & #4}}
\newcommand{\arule}[2]{\ensuremath{#1 \, &\leftarrow \, #2}}
\newenvironment{ldef}{\renewcommand\arraystretch{1.25}\left\{\begin{array}{l@{ \,}l@{\,}l}}{\end{array}\right\}}
\newenvironment{ltheo}{\[\begin{array}{l}}{\end{array}\]\ignorespacesafterend}
\newcommand{\LNDRule}[2]{
\ifenv{array}
{\drule{#1}{#2}}
{ \ifenv{align}
{\arule{#1}{#2}}
{\ifenv{align*}
{\arule{#1}{#2}}
{ERROR: using LDRule in unsupported environment: \@currenvir}
}
}
}
\newcommand{\LDRule}[4]{
\ifenv{array}
{\ddrule{#1}{#2}{#3}{#4}}
{ \ifenv{align}
{\darule{#1}{#2}{#3}{#4}}
{\ifenv{align*}
{\darule{#1}{#2}{#3}{#4}}
{ERROR: using LDRule in unsupported environment: \@currenvir}
}
}
}
\NewDocumentCommand\LRule{m+g+g+g}
\IfNoValueTF{#2
{#1.&}
\IfNoValueTF{#3}
{\LNDRule{#1}{#2.}}
{\LDRule{#1}{#2.}{#3}{#4}
}
}
\NewDocumentCommand\CLRule{m+g}
\ifenv{array}
{\cdrule{#1}{#2}}
{ \ifenv{align}
{\carule{#1}{#2}}
{\ifenv{align*}
{\carule{#1}{#2}}
{ERROR: using CLRule in unsupported environment: \@currenvir}
}
}
}
\NewDocumentCommand\carule{m+g}
\IfNoValueTF{#2}
{\ensuremath{#1.}}
{\ensuremath{#1 \, &\cause \, #2}}}
\NewDocumentCommand\cdrule{m+g}
\IfNoValueTF{#2}
{\ensuremath{#1.}}
{\ensuremath{#1 & \cause & #2}}}
\newcommand{\algrule}[4]{
\hbox{{#1}:}&
\quad #2 ~\longrightarrow~ #3
\hbox{~ if } #4\\
}
\newenvironment{lprop}{\[\begin{array}{ll}}{\end{array}\]}
\newcommand{\AlgoRule}[4]{
\ifenv{array}
{\algrule{#1}{#2}{#3}{#4}}
{ERROR: using AlgoRule in unsupported environment: \@currenvir}
}
\newcommand{\commentstyle}{\color{Gray}}
\usepackage[final]{listings}
\lstdefinelanguage{idp}{
morekeywords=[1]{query(},
morekeywords=[2]{namespace,vocabulary,theory,structure,procedure,term,set,formula, spec, specification,query},
morekeywords=[3]{include,using,type,isa,contains,partial,extern,LFD,GFD,constructed,from,constraint,pred,supertype,of,subtype,define},
morekeywords=[4]{int,float,char,string,nat},
morekeywords=[5]{if,then,else,for,end},
morecomment=[s]{/*}{*/},
morecomment=[l]{//}
}
\lstset{
language=idp,
tabsize=3,
basicstyle=\small,
frame=none,
frame=single,
showstringspaces=false,
breaklines = true,
alsoletter=(,
commentstyle=\commentstyle,
keywordstyle=[1],
keywordstyle=[2]\color{BrickRed}\bfseries,
keywordstyle=[3]\color{OliveGreen}\bfseries,
keywordstyle=[4]\color{Blue}\bfseries,
keywordstyle=[5]\color{Violet}\bfseries,
literate={~} {$\sim$}{1}
}
\newcommand{\code}[1]{\texttt{#1}}
\newcommand{\ignore}[1]{}
\newboolean{nocomments}
\setboolean{nocomments}{false}
\newboolean{commentmargin}
\setboolean{commentmargin}{true}
\newcommand{\namedcomment}[3]{
\ifthenelse{\boolean{nocomments}}
{}
{
\ifthenelse{\boolean{commentmargin}}
{ {\color{#3} \marginpar{\color{#3}\sc #2}#1} }
{ {\color{#3} {\sc #2}: #1} }
}
}
\newcommand{\mnamedcomment}[3]{\ifthenelse{\boolean{nocomments}}{}{{\marginpar{ \color{#3}{\sc #2}:#1}}}}
\newcommand{\namedchange}[4]{\marginpar{\color{#4}\sc #3}\textcolor{#4}{#1}\textcolor{gray}{\st{#2}}}
\newcommand{\todo}[1]{\namedcomment{#1}{TODO}{blue}}
\newcommand{\todonm}[1]{{\color{blue}\sc TODO} #1}
\newcommand{\mtodo}[1]{\mnamedcomment{#1}{TODO}{blue}}
\newcommand{\old}[1]{\namedcomment{#1}{OLD}{gray}}
\newcommand{\bart}[1]{\namedcomment{#1}{bb}{red}}
\newcommand{\mbart}[1]{\mnamedcomment{#1}{bb}{red}}
\newcommand{\gerda}[1]{\namedcomment{#1}{gj}{orange}}
\newcommand{\marc}[1]{\namedcomment{#1}{md}{orange}}
\newcommand{\joost}[1]{\namedcomment{#1}{jv}{purple}}
\newcommand{\bdc}[1]{\namedcomment{#1}{bdc}{OliveGreen}}
\newcommand{\broes}[1]{\namedcomment{#1}{bdc}{OliveGreen}}
\newcommand{\mbroes}[1]{\mnamedcomment{#1}{bdc}{OliveGreen}}
\newcommand{\pieter}[1]{\namedcomment{#1}{pvh}{NavyBlue}}
\newcommand{\maurice}[1]{\namedcomment{#1}{mb}{orange}}
\newcommand{\ingmar}[1]{\namedcomment{#1}{id}{purple}}
\newcommand{\matthias}[1]{\namedcomment{#1}{mvdh}{blue}}
\newcommand{\mmaurice}[1]{\mnamedcomment{#1}{mb}{orange}}
\newcommand{\jo}[1]{ \namedcomment{#1}{jo}{Fuchsia}}
\newcommand{\joachim}[1]{ \namedcomment{#1}{jj}{Sepia}}
\newcommand{\mjoachim}[1]{ \mnamedcomment{#1}{jj}{Sepia}}
\usepackage{soul}
\newcommand{\bartch}[2]{\namedchange{#1}{#2}{bb}{red}}
\newcommand{\broesch}[2]{\namedchange{#1}{#2}{bdc}{OliveGreen}}
\newcommand{\pjs}[1]{\namedcomment{#1}{pjs}{orange}}
\newcommand{\jan}[1]{\namedcomment{#1}{jvdb}{orange}}
\usepackage{etoolbox}
\newcommand\setcitation[2]
\csdef{mycommoncitation#1}{#2}}
\newcommand\getcitation[1]
\csuse{mycommoncitation#1}}
\setcitation{IDP}{WarrenBook/DeCatBBD14}
\setcitation{idp}{WarrenBook/DeCatBBD14}
\setcitation{fodot}{tocl/DeneckerT08}
\setcitation{CP}{Apt03}
\setcitation{EZCSP}{lpnmr/Balduccini11}
\setcitation{KR}{Baral:2003}
\setcitation{ASPComp2}{lpnmr/DeneckerVBGT09}
\setcitation{ASPComp3}{journals/tplp/CalimeriIR14}
\setcitation{ASPComp4}{conf/lpnmr/AlvianoCCDDIKKOPPRRSSSWX13}
\setcitation{CPSupport}{ictai/DeCat13}
\setcitation{CPsupport}{ictai/DeCat13}
\setcitation{functionDetection}{iclp/DeCatB13}
\setcitation{fodot2asp}{DeneckerLTV12}
\setcitation{Tarskian}{DeneckerLTV12}
\setcitation{Inca}{iclp/DrescherW12}
\setcitation{csp2asp}{ijcai/DrescherW11}
\setcitation{DPLLT}{cav/GanzingerHNOT04}
\setcitation{AspInPractice}{synthesis/2012Gebser}
\setcitation{clasp}{ai/GebserKS12}
\setcitation{oclingo}{kr/GebserGKOSS12}
\setcitation{gringo}{lpnmr/GebserST07}
\setcitation{cmodels}{aaai/GiunchigliaLM04}
\setcitation{inputster}{tplp/Jansen13}
\setcitation{DLV}{tocl/LeonePFEGPS06}
\setcitation{LearningPaper}{TPLP/BruynoogheBBDDJLRDV}
\setcitation{clog}{iclp/BogaertsVDV14}
\setcitation{foc}{iclp/BogaertsVDV14}
\setcitation{inferenceClog}{ecai/BogaertsVDV14}
\setcitation{examplesClog}{nmr/BogaertsVDV14b}
\setcitation{AFT}{DeneckerBV12}
\setcitation{KBS}{iclp/DeneckerV08}
\setcitation{KBPE}{inap/DePooterWD11}
\setcitation{lazyGrounding}{jair/CatDBS15}
\setcitation{lazygrounding}{jair/CatDBS15}
\setcitation{ASP}{marek99stable}
\setcitation{satid}{sat/MarienWDB08}
\setcitation{lazyclausegeneration}{constraints/OhrimenkoSC09}
\setcitation{FP}{ACMCS/Hudak89}
\setcitation{GroundingWithBounds}{jair/WittocxMD10}
\setcitation{SAT}{faia/SilvaLM09}
\setcitation{LTC}{iclp/Bogaerts14}
\setcitation{SPSAT}{ictai/DevriendtBMDD12}
\setcitation{BreakID}{cspsat/DevriendtBB14}
\setcitation{LCG}{stuckeyLCG}
\setcitation{MiniZinc}{conf/cp/NethercoteSBBDT07}
\setcitation{minizinc}{conf/cp/NethercoteSBBDT07}
\setcitation{amadini}{cpaior/AmadiniGM13}
\setcitation{bootstrapping}{lash/BogaertsJDJBD14}
\setcitation{GroundedFixpoints}{ai/BogaertsVD15}
\setcitation{LogicBlox}{datalog/GreenAK12}
\setcitation{proB}{journals/sttt/LeuschelB08}
\setcitation{NaturalInductions}{KR/DeneckerV14}
\setcitation{LP}{jacm/EmdenK76}
\setcitation{SMT}{faia/BarrettSST09}
\setcitation{AF}{ai/Dung95}
\setcitation{ADF}{kr/BrewkaW10}
\setcitation{ADFRevisited}{ijcai/BrewkaSEWW13}
\setcitation{DefaultLogic}{ai/Reiter80}
\setcitation{DL}{ai/Reiter80}
\setcitation{AEL}{mo85}
\setcitation{minisat}{sat/EenS03}
\setcitation{completion}{adbt/Clark78}
\setcitation{wasp}{lpnmr/AlvianoDFLR13}
\setcitation{minisatid}{ictai/DeCat13}
\setcitation{lcg}{stuckeyLCG}
\setcitation{CEGAR}{jacm/ClarkeGJLV03}
\setcitation{cegar}{jacm/ClarkeGJLV03}
\setcitation{CuttingPlane}{or/DantzigFJ54}
\setcitation{kodkod}{tacas/TorlakJ07}
\setcitation{cdcl}{Marques-SilvaS99}
\setcitation{CDCL}{Marques-SilvaS99}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\setcitation{}{}
\newcommand\refto[1]
\getcitation{#1}}
\newcommand\mycite[1]
\ifcsname mycommoncitation#1\endcsnam
\cite{\getcitation{#1}
\els
\cite{#1}
\f
}
\newcommand\mycitet[1]
\ifcsname mycommoncitation#1\endcsnam
\citet{\getcitation{#1}
\els
\citet{#1}
\f
}
\usepackage{amsthm}
\theoremstyle{plain}
\newtheorem{thm}{Theorem}[section]
\newtheorem{theorem}[thm]{Theorem}
\newtheorem{prop}[thm]{Proposition}
\newtheorem{lem}[thm]{Lemma}
\newtheorem*{lem*}{Lemma}
\newtheorem{cor}[thm]{Corollary}
\newtheorem{conjecture}[thm]{Conjecture}
\newtheorem{corollary}[thm]{Corollary}
\newtheorem{proposition}[thm]{Proposition}
\newtheorem{lemma}[thm]{Lemma}
\theoremstyle{definition}
\newtheorem{defn}[thm]{Definition}
\newtheorem{definition}[thm]{Definition}
\newtheorem{nota}[thm]{Notation}
\newtheorem*{nota*}{Notation}
\newtheorem{note}[thm]{Note}
\newtheorem*{note*}{Note}
\newtheoremstyle{example_basic}
{\topsep}
{\topsep}
{\normalfont
{0pt
{\bfseries
{.
{5pt plus 1pt minus 1pt
{
\newtheoremstyle{example_contd}
{\topsep}
{\topsep}
{\normalfont
{0pt
{\bfseries
{.
{5pt plus 1pt minus 1pt
{\thmname{#1} \thmnumber{ #2}\thmnote{#3} (continued)
\theoremstyle{example_basic}
\newtheorem{ex}[thm]{Example}
\newtheorem{example}[thm]{Example}
\newtheorem{remark}[thm]{Remark}
\newtheorem{ex*}{Example}
\theoremstyle{example_contd}
\newtheorem*{ex_continued}{Example}
\theoremstyle{plain}
\renewcommand{\S}{S}
\renewcommand{\I}{I}
\newcommand{\citet[1]}{\citeN{#1}\xspace}
\setboolean{nocomments}{true}\renewcommand{\newtext}[1]{#1}
\newcommand\blfootnote[1]{ \begingroup
\renewcommand\thefootnote{}\footnote{#1} \addtocounter{footnote}{-1} \endgroup
}
\begin{document}
\title[The KB paradigm in interactive configuration]{The KB paradigm and its application to interactive configuration.}
\author[P. Van Hertum et al.]
{Pieter Van Hertum, Ingmar Dasseville, Gerda Janssens, Marc Denecker \\
Department of Computer Science\\
KU LEUVEN\\
Leuven, BELGIUM\\
<EMAIL>
}
\maketitle \blfootnote{This is an extended version of a paper presented at the international
symposium on Practical Aspects of Declarative Languages (PADL 2016),
invited as a rapid communication in TPLP. The authors acknowledge the
assistance of the conference program chairs Marco Gavanelli and John Reppy. This research was supported by the project GOA 13/010 Research Fund KU Leuven and projects G.0489.10, G.0357.12,
and G.0922.13 of the Research Foundation - Flanders.}
\begin{abstract}
The knowledge base paradigm aims to express domain knowledge in a rich formal language, and to use this domain knowledge as a knowledge base to solve various problems and tasks that arise in the domain by applying multiple forms of inference.
As such, the paradigm applies a strict separation of concerns between information and problem solving.
In this paper, we analyze the principles and feasibility of the knowledge base paradigm in the context of an important class of applications: interactive configuration problems.
In interactive configuration problems, a configuration of interrelated objects under constraints is searched, where the system assists the user in reaching an intended configuration.
It is widely recognized in industry that good software solutions for these problems are very difficult to develop.
We investigate such problems from the perspective of the KB paradigm.
We show that multiple functionalities in this domain can be achieved by applying different forms of logical inferences on a formal specification of the configuration domain.
We report on a proof of concept of this approach in a real-life application with a banking company.
To appear in Theory and Practice of Logic Programming (TPLP).
\end{abstract}
\begin{keywords}
Interactive Configuration, Knowledge Base Paradigm, Inferences, Applications of Declarative Systems
\end{keywords}
\section{Introduction}
In this paper, we investigate the application of knowledge representation and reasoning (KRR) to the problem of {\em interactive configuration}.
In the past decades enormous progress in many different areas of computational logic was obtained.
This resulted in a complex landscape with many declarative paradigms, languages and communities.
One issue that fragments computational logic more than anything else is the reasoning/inference task.
Computational logic is divided in different declarative paradigms, each with its own syntactical style, terminology and conceptuology, and designated form of inference (e.g, deductive logic, logic programming, abductive logic programming, databases (query inference), answer set programming (answer set generation), constraint programming, etc.).
Yet, in all of them declarative propositions need to be expressed.
Take, e.g., ``each lecture takes place at some time slot''.
This proposition could be an expression to be deduced from a formal specification if the task was a verification problem, or to be queried in a database, or it could be a constraint for a scheduling problem.
It is, in the first place, just a piece of information and we see no reason why depending on the task to be solved, it should be expressed in a different formalism (classical logic, SQL, ASP, MiniZinc, etc.).
The Knowledge Base (KB) paradigm \mycite{KBS} was proposed as an answer to this.
The KB paradigm applies a strict separation of concerns to information and problem solving.
A KB system allows information to be stored in a knowledge base, and provides a range of inference methods.
With these inference methods various types of problems and tasks can be solved using the {\em same knowledge base}.
As such the knowledge base is neither a program nor a description of a problem, it cannot be executed or run.
It is nothing but information. However, this information can be used to solve multiple sorts of problems.
Many declarative problem solving paradigms are mono-inferential: they are based on one form of inference.
In comparison, the KB-paradigm is multi-inferential.
We believe that this implements a more natural, pure view of what declarative logic is aimed to be.
The $\fodot$ KB project \mycite{KBS} is a research project that runs now for a number of years.
Its aim is to integrate different useful language constructs and forms of inference from different declarative paradigms in one rich declarative language and a KB system.
So far, it has led to the KB language $\fodot$~\mycite{fodot} and the KB system \idp~\mycite{IDP} which were used in the configuration experiment described in this paper.
An interactive configuration (IC) problem \cite{ai/McDermott82,ijcai/MittalF89,FleischanderlFHSS98,aiedam/JunkerM03,cp/Hadzic04} is an interactive version of a constraint solving problem.
One or more users search for a configuration of objects and relations between them that satisfies a set of constraints.
Industry abounds with interactive configuration problems: configuring composite physical systems such as cars and computers, insurances, loans, schedules involving human interaction, webshops (where clients choose composite objects), etc.
However, building such software is renowned in industry as difficult and no broadly accepted solution methods are available~\cite{Felfernig14,jlp/AxlingH96}.
Building software support using standard imperative programming is often a nightmare~\cite{cacm/BarkerO89,hicss/PillerHIS14}, due to the fact that (1) many functionalities need to be provided, (2) they are complex to implement, and (3) constraints on the configuration tend to get duplicated and spread out over the application, in the form of snippets of code performing various computations relative to the constraint (e.g., context dependent checks or propagations) which often leads to an unacceptable maintenance cost.
This makes interactive configuration an excellent domain to illustrate the advantages of declarative methods over standard imperative or object-oriented programming.
Our research question is: can we express the constraints of correct
configurations in a declarative logic and provide the required
functionalities by applying inference on this domain knowledge? This
is a KRR question albeit a difficult one. In the first place, some of the
domain knowledge may be complex. For an example in the context of a
computer configuration problem, take the following constraint: {\em
the total memory usage of different software processes that needs to be in main
memory simultaneously, may not exceed the available RAM memory}. It
takes an expressive knowledge representation language with aggregates to (compactly
and naturally) express such a constraint. Many interactive
configuration problems include complex constraints: various sorts
of quantification, aggregates, definitions
(sometimes inductive), etc. Moreover, an interactive configuration
system needs to provide many functionalities: checking the validity of
a fully specified configuration, correct and safe reasoning on a
partially specified configuration (this involves reasoning on
incomplete knowledge, sometimes with infinite or unknown domains),
computing impossible values or forced values for attributes,
generating sensible questions to the user, providing explanation why certain values are
impossible, backtracking if the user regrets some choices, supporting
the user by filling in his don't-cares while potentially taking into account a
cost function, etc.
That declarative methods are particularly suitable for solving this type of problem has been acknowledged before, and several systems and languages have been developed~\cite{cp/Hadzic04,inap/SchneeweissH11,TiihonenHAS13,ppdp/VlaeminckVD09}.
A first contribution of this paper
is the analysis of IC problems from a Knowledge Representation point of view.
We show that multiple functionalities in this domain can be achieved by applying different forms of logical inference on {\em the same} formal specification of the configuration domain. We define various sorts of inference and analyse them in terms of which different functionalities can be supplied.
The second contribution is the reverse: we study the feasibility and usefulness of the KB paradigm in this important class of applications.
The logic used in this experiment is
the logic \fodot~\mycite{fodot}, an extension of first-order logic
(FO), and the system is the IDP system~\mycite{IDP}. We discuss the
complexity of (the decision problems of) the inference problems and
why they are solvable, despite the high expressivity of the language
and the complexity of inference.
This research has its origin in an experimental IC system we developed in collaboration with industry.
We evaluated our approach using the evaluation criteria of the knowledge-based configuration research~\cite{Felfernig14}.
We conclude this paper with a discussion of related work in using knowledge-based systems for configuration and a comparison of our approach with these systems. \section{The FO(.) KB project}
\label{ssec:fodot}
\label{sec:kbs}
\paragraph{The language.}
$\fodot$ refers to the class of extensions of first order logic (FO) as is common in logic, e.g. FO(LFP) stands for the extension of FO with a least fixpoint construction \cite{cav/ImmermanV97}.
Currently, the language of the \idp system in the project is FO(T, ID, Agg, arit, PF) \cite{tocl/DeneckerT08,tplp/PelovDB07}: FO extended with types, definitions, aggregates, arithmetic and partial functions.
Abusing notation, we will use $\fodot$ as an abbreviation for this language. Below, we introduce the aspects of the logic and its syntax on which this paper relies.
\paragraph{A specification.} A \textit{vocabulary} is a set $\Sigma$ of type (denoted as $\Sigma_T$), predicate (denoted as $\Sigma_P$) and function symbols (denoted as $\Sigma_F$).
Variables $x, y$, atoms $A$, FO-formulas $\varphi$ are defined as usual.
A predicate $P$ of arity $n$ has a type $[\tau_1,\ldots,\tau_n]$, a $n$-tuple of type symbols.
A function of arity $n$ has a type $[\tau_1,\ldots,\tau_n]\rightarrow \tau_{n+1}$, a $(n+1)$-tuple of type symbols.
Aggregate terms are of the form $Agg(E)$, with $Agg$ an aggregate function symbol and $E$ an expression $\{(\overline{x},F(\overline{x}))|\varphi(\overline{x})\}$, where $\varphi$ is any FO-formula, $F$ a function symbol and
$\overline{x}$ a tuple of variables. Examples are the cardinality,
sum, product, maximum and minimum aggregate functions. For example $sum\{(x,F(x))|\varphi(x)\}$ is read as $\Sigma_{x\in\{y|\varphi(y)\}} F(x)$. A \textit{term}
in $\fodot$ can be an aggregate term or a term as defined in FO.
A \textit{theory} is a set of $\fodot$ formulas.
A \emph{partial set} on domain $D$ is a function from $D$ to $\{\Tr,\Un,\Fa\}$.
A partial set is two-valued (or total) if $\Un$ does not belong to its range.
A \textit{(partial) structure} $\pS$ consists of a domain $D_\tau$ for all types $\tau$ in $\Sigma_T$ and an assignment of a partial set $\sigma^\pS$ to each predicate or function symbol $\sigma\in(\Sigma_P\cup\Sigma_F)$, called the \emph{interpretation} of $\sigma$ in $\pS$.
The interpretation $P^\pS$ of a predicate symbol $P$ with type $[\tau_1,\ldots,\tau_n]$ in $\pS$ is a
partial set on domain $D_{\tau_1}\times \ldots \times D_{\tau_n}$. For
a function $F$ with type $[\tau_1, \ldots, \tau_n]\rightarrow
\tau_{n+1}$, the interpretation $F^\pS$ of $F$ in $\pS$ is a partial
set on domain $D_{\tau_1}\times \ldots \times D_{\tau_n} \times
D_{\tau_{n+1}}$. In case the interpretation of (a predicate or function symbol) $\sigma$ in $\pS$ is a
two-valued set, we abuse notation and use $\sigma^\pS$ as shorthand for
$\{\ddd|\sigma^\pS(\ddd)=\Tr\}$.
The precision-order on the truth values is given by $\Un<_p \Fa$ and $\Un<_p \Tr$.
It can be extended pointwise to partial sets and partial structures, denoted $\pS\leq_p\pS'$.
Informally, this means that an interpretation has become more precise if tuples of domain elements that were previously mapped to unknown now map to true or false.
Notice that total structures are the maximally precise ones. We will illustrate this precision relation in Example \ref{ex:specification}.
We say that $\pS'$ extends $\pS$ if $\pS\leq_p \pS'$.
We will sometimes use $\sigma^\pS_{x}$ as shorthand for the set $\{\ddd|\ddd \in D_{\tau_1}\times \ldots \times D_{\tau_n} \land \sigma^\pS(\ddd)=x\}$, with $x\in \{\Tr,\Fa,\Un\}$.
\newtext{
The associated theory $T_\pS$ of a partial structure $\pS$ is a representation of the information contained in $\pS$ as a theory, which will be used in Section \ref{sec:solution}.
It is defined by the following collection of constraints.
For every predicate symbol $P$, this collection contains two sets of constraints:
\begin{align*}
&\{P(\overline{d})|\ddd \in P^\pS_\Tr\}\\
&\{\neg P(\overline{d})|\ddd \in P^\pS_\Fa\}
\end{align*}
and two sets of constraints for every function symbol $F$:
\begin{align*}
&\{F(\overline{d})=e|(\ddd,e) \in F^\pS_\Tr\}\\
&\{\neg F(\overline{d})=e|(\ddd,e) \in F^\pS_\Fa\}
\end{align*}
Given a partial structure $\pS$, the domain structure $\pS_D$ is the structure containing only the domains of $\pS$. It is easy to see that $\pS$ contains the same information as $\T_\pS \cup \pS_D$.
}
A total structure\footnote{Note the difference in typography between a partial structure $\pS$ and a total structure $\S$.} $\S$ is called {\em functionally consistent} if for each function $F$ with type $[\tau_1, \ldots, \tau_n]\rightarrow
\tau_{n+1}$, the interpretation $F^\S$ is the graph of a function $D_{\tau_1} \times \ldots \times D_{\tau_n} \mapsto D_{\tau_{n+1}}$.
A partial structure $\pS$ is functionally consistent if it has a functionally consistent two-valued extension.
Unless stated otherwise, we will assume for the rest of this paper that all (partial) structures are functionally consistent.
A domain atom (domain term) is a tuple of a predicate symbol $P$ (a function symbol $F$) and a tuple of domain elements $(d_1,\ldots,d_n)$.
We will denote it as $P(d_1,\ldots,d_n)$ (respectively $F(d_1,\ldots,d_n)$).
We say a domain term $t$ of type $\tau$ is uninterpreted in $\pS$
if $\{d|d \in D_{\tau} \land (t=d)^\pS= \Un\}$ is non-empty.
To define the satisfaction relation on theories, we extend the interpretation of symbols to arbitrary terms and formulas using the Kleene truth assignments~\cite{Kleene52}.
For a theory $\T$ and a partial structure $\pS$, we say that $\pS$ is a model of $\T$ (or in symbols $\pS\vDash \T$) if $\T^\pS=\Tr$ and $\pS$ is two-valued.
We sometimes abuse notation and write $\T\vDash \varphi$ for the entailment relation, as a shorthand for ``For every structure $\S$ such that $\S\vDash \T$, we have $\S\vDash \varphi$.''.
\begin{example}\label{ex:specification}
To illustrate some of the concepts introduced above, assume a situation where we have some knowledge about printers, that have some type of connection.
A vocabulary to model such knowledge can look as follows:
\begin{align*}
&\voc =\{ \\
& \quad \Sigma_T=\{printer, connection\}\\
& \quad \Sigma_P=\{PrinterConnection(printer,connection)\}\\
& \quad \Sigma_F=\{ \}\\ \}
\end{align*}
A structure $\pS_0$ in which we have 2 printers $P_1$ and $P_2$ and 2 possible connections: $USB$ and $LAN$, where we have no additional information, looks like:
\begin{align*}
&\pS_0=\{ \\
& \quad printer = \{P_1,P_2\}\\
& \quad connection = \{USB, LAN\}\\
& \quad PrinterConnection = \{
(P_1,USB)\rightarrow \Un, (P_2,USB)\rightarrow \Un,\\
& \qquad (P_1,LAN)\rightarrow \Un, (P_2,LAN)\rightarrow \Un\}
\\ \}
\end{align*}
A more precise structure $\pS_1\geq_p \pS_0$, containing the partial information that $P_1$ has $USB$ and $P_2$ certainly has no $LAN$ connection looks like:
\begin{align*}
&\pS_1=\{ \\
& \quad printer = \{P_1,P_2\}\\
& \quad connection = \{USB, LAN\}\\
& \quad PrinterConnection = \{
(P_1,USB)\rightarrow \Tr, (P_2,USB)\rightarrow \Un,\\
& \qquad (P_1,LAN)\rightarrow \Un, (P_2,LAN)\rightarrow \Fa\}
\\ \}
\end{align*}
A total structure $\S_2 \geq_p \pS_1$ containing full information can look like:
\begin{align*}
&\S_2=\{ \\
& \quad printer = \{P_1,P_2\}\\
& \quad connection = \{USB, LAN\}\\
& \quad PrinterConnection = \{
(P_1,USB)\rightarrow \Tr, (P_2,USB)\rightarrow \Tr,\\
& \qquad (P_1,LAN)\rightarrow \Fa, (P_2,LAN)\rightarrow \Fa\}
\\ \}
\end{align*}
\end{example}
\paragraph{Inference tasks.} \label{ssec:inferences}
In the KB paradigm, a specification is a bag of information. This information can be used for solving various problems by applying a suitable form of inference on it.
FO is standardly associated with deduction inference: a deductive
inference task takes as input a pair of theory $\T$ and sentence
$\varphi$, and returns $\Tr$ if $\T\models\varphi$ and $\Fa$
otherwise. This is well-known to be undecidable for FO, and by extension
for $\fodot$. However, to provide the required functionality of an
interactive configuration system we can use simpler forms of
inference. Indeed, in many such domains a fixed finite domain is
associated with each unknown configuration parameter.
A natural format in logic to describe these finite domains is by a
partial structure with a finite domain. Also other data that are often
available in such problems can be represented in that structure. As
such various inference tasks are solvable by finite domain reasoning
and become decidable. Below, we give the base forms of inference for our KB system and
recall their complexity when using finite domain reasoning. We assume a fixed vocabulary $\Sigma$ and theory $\T$ and query. Our complexities are given in function of the domain size.
\begin{description}
\item[Modelexpand($\T,\pS$):] input: theory ${\T}$ and partial structure $\pS$;
output: a model $\I$ of $\T$ such that $\pS\leqp\I$ or \textit{UNSAT} if there is no such $\I$.
Modelexpand~\cite{lash08/WittocxMD08} is a generalization for $\fodot$ theories of the modelexpansion task as defined in Mitchell et al.~\cite{MitchellT05}.
Complexity of deciding the existence of a modelexpansion is in \textbf{NP}.
Structure $\S_2$ in Example \ref{ex:specification} is the output of Modelexpand($\T,\pS_1$), with $\pS_1$ as in Example \ref{ex:specification}, and $\T$ a theory consisting of the constraint that every printer has exactly one connection.
\item[Modelcheck($\T,\S$):] input: a total structure $\S$ and theory $\T$ over the vocabulary interpreted by $\S$; output is the boolean value $\S \models \T$.
Note that Modelcheck is a degenerate case of the Modelexpand inference, with $\pS$ a total structure. Complexity is in \textbf{P}.
\item[Minimize($\T,\pS,t$):] input: a theory $\T$, a partial structure
$\pS$ and a term $t$ of numerical type; output: a model $\I\geqp \pS$
of $\T$ such that the value $t^{\I}$ of $t$ is minimal. The term $t$ represents a numerical expression whose value has to be minimized.
This is an extension to the modelexpand inference.
The complexity of deciding that a certain $t^\I$ is minimal, is in $\mathbf{\Delta_2^P}$.
\item[Propagate($\T,\pS$):] input: theory ${\T}$ and partial structure $\pS$; output: the most precise partial structure $\pS_r$ such that for every model $\I\geqp\pS$ of ${\T}$ it is true that $I\geq_p \pS_r$.
The complexity of deciding that a partial structure $\mc{S}'$ is $\mc{S}_r$ is in $\mathbf{\Delta_2^P}$. Note that we assume that all partial structures are functionally consistent, which implies that we also propagate functional integrity constraints.
\item[Query($\pS,E$):] input: a (partial) structure $\pS$ and a set expression
$E=\{\overline{x}\mid \varphi(\overline{x})\}$; output: the set $A_Q=\{\overline{x}\mid\varphi(\overline x)^\pS=\Tr\}$. Complexity of deciding that a set $A$ is $A_Q$ is in
\textbf{P}.
\end{description}
Approximative versions exist for some of these inferences, with lower complexity \cite{ppdp/VlaeminckVD09}.
More inferences exist, such as simulation of temporal theories in
\fodot~\mycite{LTC}, but were not
used in the experiment.
\section{Interactive Configuration}
In an IC problem, one or more users search for a configuration of objects and relations between them that satisfies a set of constraints.
Typically, the user is not aware of all constraints.
There may be too many of them to keep track of.
Even if the human user can oversee all constraints that he needs to satisfy, he is not a perfect reasoner and cannot comprehend all consequences of his choices.
This in its own right makes such problems hard to solve.
The problems get worse if the user does not know about the relevant objects and relations or the constraints on them, or if the class of involved objects and relations is large, if the constraints get more complex and more ``irregular" (e.g., exceptions), if more users are involved, etc.
On top of that, the underlying constraints in such problems tend to evolve quickly.
All these complexities occur frequently, making the problem difficult for a human user.
In such cases, computer assistance is needed: the human user chooses and the system assists by guiding him through the search space.
For a given IC problem, an IC system has information on that problem.
There are a number of stringent rules to which a configuration should conform, and besides this there is a set of parameters.
Parameters are the open fields in the configuration that need to be filled in by the user or decided by the system.
\subsection{Running example: Domain knowledge}
A simplified version of the application in Section \ref{implementation} is used in Section \ref{sec:solution} as running example.
We introduce the domain knowledge of this example here.
\noindent \begin{example}
\label{MotEx}
Software on a computer has to be configured for different employees.
Table \ref{tab:ex} contains the information on the software, the requirements, the budgets of the employees and the prices of software.
Available software is Windows, Linux, \LaTeX, Office and a DualBoot system.
Each software item has a price, which can be seen in column \textbf{PriceOf}.
Column \textbf{PreReq} specifies which software is required for other software.
Every type of employee has a budget, provided in column \textbf{MaxCost}.
\textbf{IsOs} lists the pieces of software that are operating systems.
Next to the information in the table, we know that if more than one OS is installed, a DualBoot System is required.
\begin{table}
\caption{Example data \label{tab:ex}}
\centering
\begin{tabular}{llcllcllcl}
\toprule
\multicolumn{2}{c}{\textbf{PriceOf}} & \hspace{0.4cm} & \multicolumn{2}{c}{\textbf{PreReq}} & \hspace{0.4cm} & \multicolumn{2}{c}{\textbf{MaxCost}} &\hspace{0.4cm} &\multicolumn{1}{c}{\textbf{IsOS}}\\
\textit{software} & \textit{int} & \hspace{0.4cm} & \textit{software} & \textit{software} & \hspace{0.4cm} & \textit{employee} & \textit{int} &\hspace{0.4cm} &\textit{software}\\
\midrule
Windows & 60 & & Office & Windows & & Secretary & 100 & & Windows \\
Linux & 20 & & \LaTeX & Linux & & Manager & 150 & & Linux \\
\LaTeX & 10 & & & & & & & & \\
Office & 30 & & & & & & & & \\
DualBoot & 40 & & & & & & & & \\
\bottomrule
\end{tabular}
\end{table}
\end{example}
\newcounter{st}[section]
\subsection{Subtasks of an interactive configuration system}\label{ssec:subtasks}
Any system assisting a user in interactive configuration must be able to perform a set of subtasks.
We look at important subtasks that an interactive configuration system should support.
\subtasksbegin
\subtask{Acquiring information from the user}
The first task of an IC system is acquiring information from the user.
The system needs to get a value for a number of parameters of the configuration from the user.
There are several options: the system can ask questions to the user, it can make the user fill in a form containing open text fields, dropdown-menus, checkboxes, etc.
Desirable aspects would be to give the user the possibility to choose the order in which he gives values for parameters and to omit filling in certain parameters (because he does not know or does not care).
For example, in the running example a user might need a \LaTeX-package, but he does not care about which OS he uses.
In that case the system will decide in his place that a Linux system is required.
Since a user is not fully aware of all constraints, it is possible that he inputs conflicting information.
This needs to be handled or avoided.
\subtask{Generating consistent values for a parameter}
After a parameter is selected (by the user or the system) for which a value is needed, the system can assist the user in choosing these values.
A possibility is that the system presents the user with a list of all possible values, given the values for other parameters and the constraints of the configuration problem.
Limiting the user with this list makes that the user is unable to input inconsistent information.
\subtask{Propagation of information}
Assisting the user in choosing values for the parameters, a system can use the constraints to propagate the information that the user has communicated.
This can be used in several ways. A system can communicate propagations through a GUI, for example by coloring certain fields red or graying out certain checkboxes.
Another way is to give a user the possibility to explicitly ask ``what if''-questions to the system.
In Example \ref{MotEx}, a user can ask the system what the consequences are if he was a secretary choosing an Office installation.
The system answers that in this case a Windows installation is required, which results in a Linux installation becoming impossible (due to budget constraints) and as a consequence it also derives the impossibility of installing \LaTeX.
\subtask{Checking the consistency for a value}
When it is not possible/desirable to provide a list of possible values, the system checks that the value the user has provided is consistent with the known data and the constraints.
\subtask{Checking a configuration}
If a user makes manual changes to a configuration, the system provides him with the ability to check if his updated version of the configuration still conforms to all constraints.
\subtask{Autocompletion}
If a user has finished communicating all his preferences, the system autocompletes the partial configuration to a full configuration.
This can be done arbitrarily (a value for each parameter such that the constraints are satisfied) or the user can have some other parameters like total cost, that have to be optimized.
\subtask{Explanation}
If a supplied value for a parameter is not consistent with other parameters, the system can explain this inconsistency to the user.
This can be done by showing minimal sets of parameters with their values that are inconsistent, by showing (visualizations of) constraints that are violated or by combinations of both.
It can also explain to the user why certain automatic choices are made, or why certain choices are impossible.
\subtask{Backtracking}
It is not unthinkable that a user makes a mistake, or changes his mind after seeing consequences of choices he made.
Backtracking is an important subtask for a configuration system, and can be supported in numerous ways.
The simplest way is a simple back button, which reverts the last choice a user made.
A more involved option is a system where a user can select any parameter and erase his value for that parameter.
The user can then decide this parameter at a later timepoint.
Even more complex is a system where a user can supply a value for a parameter and if it is not consistent with other parameters the system shows him which parameters are in conflict and proposes other values for these parameters such that consistency can be maintained.
\subtasksend
\section{Interactive Configuration in the KB paradigm}\label{sec:solution}
To analyze the IC problem from the KB point of view, we aim at formalizing the subtasks of Section 3 as inferences.
In this paper we do not deal with user interface aspects.
For a given application, our knowledge base consists of a vocabulary $\voc$, a theory $\T$ expressing the configuration constraints and a partial structure $\pS$.
Initially, $\pS_0$ is the partial structure that contains the domains of the types and the input data.
During IC, $\pS_0$ will become more and more precise partial structures $\pS_i$ due to choices made by the user.
For IC, the KB also contains $L_{\pS_0}$, the set of all uninterpreted domain atoms/terms\footnote{In the rest of this paper, a domain atom is treated as a term that evaluates to true or false.} in $\pS_0$.
These domain terms are the logical formalization of the parameters of the IC problem.
$\voc$ and $\T$ are fixed.
As will be shown in this section, all subtasks can be formalized by (a combination of) inferences on this knowledge base consisting of $\Sigma, \T, \pS_0, L_{\pS_0}$ and information gathered from the user.
\begin{example}\label{ExD}
\newtext{Continuing Example \ref{MotEx}, use vocabulary $\voc$:
\begin{align*}
&\voc = \\
& \quad \Sigma_T=\{software,\ employee, \ int\}\\
& \quad \Sigma_P=\{ Install(software),\ IsOS(software),\ PreReq(software,software)\}\\
& \quad \Sigma_F=\{PriceOf(software):int,\ MaxCost(employee):int,\\
& \quad \quad \quad \quad \quad Cost:int,\ Requester:employee\}
\end{align*}
The initial partial structure $\pS_0$ consists of:
\begin{align*}
&employee \rightarrow \{Secretary, Manager\}\\
&software \rightarrow \{Windows, Linux, LaTeX, Office, DualBoot\}
\end{align*}
and interpretations for \textit{MaxCost (employee):int, IsOs(software), PriceOf(software): int} and \textit{PreReq(software, software)} as can be seen in Table \ref{tab:ex}.
All symbols from $\Sigma$ that are not specified above are assumed to be fully unknown in $\pS_0$.\\
The set of parameters $L_{\pS_0}$ is:
\begin{align*}
\{&Requester, Install(Windows), Install(Linux),\\ & Install(Office), Install(LaTeX), Install (DualBoot), Cost\}
\end{align*}
The theory $\T$ consists of the following constraints:
\[\begin{array}{l}
\forall s1\, s2:\ Install(s1)\land PreReq(s1,s2)\Rightarrow Install(s2).\\
\quad \text{// The total cost is the sum of the prices of all installed software.}\\
Cost = sum\{(s,PriceOf(s)) | Install(s)\}.\\
Cost \leq MaxCost(Requester).\\
\exists s:\ Install(s)\land IsOS(s). \\
Install(Windows)\land Install(Linux) \Rightarrow Install(DualBoot).
\end{array}\]}
\end{example}
\subtasksbegin
\subtask{Acquiring information from the user}
Key in IC is collecting information from the user on the parameters.
During the run of the system, the set of parameters that are still open changes.
In our KB system a derived inference (a combination of the inferences as introduced in Section \ref{ssec:inferences}) is used to calculate this set of parameters.
Complexity results of derived inferences stem from basic results formulated by \citet{MitchellT05} and the observation that modelchecking is polynomial in the size of the domain.
\begin{definition}
\textbf{Calculating uninterpreted terms.}\\ \textbf{GetOpenTerms($\T,\pS$)} is the derived inference with input a theory $\T$, a partial structure $\pS\geq_p\pS_0$ and the set $L_{\pS_0}$ of terms.
Output is a set of terms such that for every term $t$ in that set, there exist models $\I_1$ and $\I_2$ of $\T$ that extend $\pS$ ($\I_1,\I_2\geq_p\pS$) for which $t^{\I_1}\neq t^{\I_2}$.
Or formally:
$$\{l|l\in L_{\pS_0} \land \{d|(l=d)^{\pS'}=\Un\}\neq \emptyset \land \pS'=Propagate(\T,\pS)\}$$
\end{definition}
The complexity of deciding whether a given set of terms $A$ is the set of uninterpreted terms is in $\mathbf{\Delta_2^P}$.
An IC system can use this set of terms in a number of ways. It can use a metric to select a specific term, which it can pose as a direct question to the user.
It can also present a whole list of these terms at once and let the user pick one to supply a value for.
In Section \ref{implementation}, we discuss two different approaches we implemented for this project.
\begin{example}
\label{ex:questions}
In Example \ref{ExD}, the parameters and domains are already given.
Assume that the user has chosen the value \textit{Manager} for \textit{Requester}, true for \textit{Install(Windows)} and false for \textit{Install(Linux)}.
The system will return \textit{GetOpenTerms}$(\T,\pS)$ = $\{$\textit{Install(Office), Install(DualBoot), Cost}$\}$.
\end{example}
\subtask{Generating consistent values for a parameter}
A domain element $d$ is a possible value for term $t$ if there is a model $\I\geq_p\pS$ such that $(t=d)^\I=\Tr$.
\begin{definition}
\textbf{Calculating consistent values.}\\
\textbf{GetConsistentValues($\T,\pS,t$)} is the derived inference with input a theory $\T$, a partial structure $\pS$ and a term $t\in GetOpenTerms(\T,\pS)$.
Output is the set $$\{t^{\I}|\ \I \text{ is a model of }\T\text{ extending }\pS \}$$
\end{definition}
The complexity of deciding that a set $P$ is the set of consistent values for $t$ is in $\mathbf{\Delta^P_2}$.
\begin{example}
\label{ex:values}
The consistent values for $Requester$ given $\T$ and the initial partial structure $\pS_0$ from Example \ref{ExD} is:
\begin{align*}
&GetConsistentValues(\T,\pS, Requester)=\{Secretary, Manager\}\\
\end{align*}
Consistent values for other terms are the integers for \textit{Cost} and $\{$\textit{true, false}$\}$ for the others.
\end{example}
\subtask{Propagation of information}
It is informative for the user that he can see the consequences of assigning a particular value to a parameter.
\begin{definition}
\textbf{Calculating Consequences.}\\
\textbf{PosConsequences($\T,\pS,t,a$)} and \textbf{NegConsequences($\T,\pS,t,a$)} are derived inferences with input a theory $\T$, a partial structure $\pS$, an uninterpreted term $t$ $\in$ GetOpenTerms$(\T,\pS)$ and a domain element $a$ $\in$ GetConsistentValues$(\T,\pS,t)$. As output it has a set $C^+$, respectively $C^-$ of tuples $(q,b)$ of uninterpreted terms and domain elements.
$(q,b)\in C^+$, respectively $C^-$ means that the choice $a$ for $t$ entails that $q$ will be forced, respectively prohibited to be $b$.
Formally,
\begin{align*}
C^+=\{(q,b)\ |\ & (q=b)^{\pS'}=\Tr \land (q=b)^{\pS}=\Un \\
\land \ & \pS'=Propagate(\T,\pS\cup \{t=a\})\\% \land q \in Q_\pS \setminus \{t\} \ \}\\
\land \ & q \in GetOpenTerms(\T,\pS) \setminus \{t\} \ \}\\
C^-=\{(q,c)\ |\ & (q=c)^{\pS'}=\Fa \land (q=c)^{\pS}=\Un \\
\land \ & \pS'=Propagate(\T,\pS\cup \{t=a\})\\% \land q \in Q_\pS \setminus \{t\} \ \}
\land \ & q \in GetOpenTerms(\T,\pS) \setminus \{t\} \ \}\
\end{align*}
\end{definition}
The complexity of deciding whether a set $P$ is $C^+$ or $C^-$ is in $\mathbf{\Delta_2^P}$.
\begin{example}\label{ExAssisting}
Say the user has chosen $Requester= Secretary$ and wants to know the consequences of making $Install(Windows)$ true.
The output in this case contains
$(Install(LaTeX),\Fa)$ in $PosConsequences(\T,\pS,t,a)$ and $(Install(LaTeX),\Tr)$ in $NegConsequences(\T,\pS,t,a)$
since this combination is too expensive for a secretary.
Note that there is not always such a correspondence between the positive and negative consequences. For example, when deriving a negative consequence for $Cost$, this does not necessarily imply a positive consequence.
\end{example}
\subtask{Checking the consistency for a value}
A value $d$ for a term $t$ is consistent if there exists a model of $\T$ in which $t=d$ that extends the partial structure representing the current state.
\begin{definition}
\textbf{Consistency Checking.}\\ \textbf{CheckConsistency($\T,\pS,t,d$)} is the derived inference with input a theory $\T$, a partial structure $\pS$, an uninterpreted term $t$ and a domain element $d$.
Output is a boolean $b$ that represents whether $\pS$ extended with $t=d$ still satisfies $\T$.
Formally we return $\Tr$ if
$$(\pS\cup \{t^\pS=d\})\vDash \T$$
and $\Fa$ otherwise. Complexity of deciding if a value $d$ is consistent for a term $t$ is in \textbf{NP}.
\end{definition}
\begin{example}
If a user has chosen $Install(Windows)$ and $Install(LaTeX)$ to be true, then $Manager$ will and $Secretary$ will not be a consistent answer for $Requester$.
\end{example}
\subtask{Checking a configuration}
Once the user has constructed a 2-valued structure $\S$ and makes manual changes to it, he may need to check if all constraints are still satisfied. A theory $T$ is checked on a total structure $\S$ by calling $Modelcheck(T,S)$, with complexity in \textbf{P}.
\subtask{Autocompletion}
If a user is ready communicating his preferences (Subtask 1) and there are undecided terms left which he does not know or care about, the user may want to get a full configuration (i.e. a total structure). This is computed by modelexpand. In particular: $$\I=Modelexpand(\T,\pS)$$
In many of those situations the user wants to have a total structure with, for example, a minimal cost (given some term representing the cost $t$). This is computed by minimize:
$$\I=Minimize(\T,\pS,t)$$
\begin{example}\label{ExCompleting}
Assume the user is a secretary and all he knows is that he needs Office.
He chooses $Secretary$ for $Requester$ and true for $Install(Office)$ and calls autocompletion.
A possible output is a structure $\S$ where for the remaining parameters, a
choice is made that satisfies all constraints, e.g.,
$Install(Windows)^\S=\Tr$, $Install(DualBoot)^\S=\Tr$ and the other $Install$ atoms false. This is not a cheapest solution (lowest cost). By calling minimize using cost-term $Cost$, the DualBoot is dropped.
\end{example}
\subtask{Explanation}
\newtext{ It is clear that a whole variety of options can be developed to provide different kinds of explanations to a user.
If a user supplies an inconsistent value for a parameter, options can range from calculating an inconsistent subset of the theory $\T$ (1) to giving a proof of inconsistency as in \cite{iclp/PontelliS06} (2), to calculating a partial subconfiguration that has this inconsistency (3).
UnsatSubstructure is a logical inference for option 3.
\begin{definition}\label{def:inconstructure}
\textbf{Calculating inconsistent structures.}\\
\textbf{UnsatSubstructure($\T,\pS$)} is a derived inference with input a theory $\T$ and a partial structure $\pS$ that cannot be extended to a model of $\T$ and as output all (partial) structures $\pS'\leq_p \pS$ such that $\pS'$ cannot be extended to a model $\I$ of $\T$. Formally, we return:
$$\{\pS'|\pS' \leq_p \pS \land \neg (\exists \I \geq_p \pS' \land \I\vDash \T)\}$$
Complexity of deciding if a set is an inconsistent substructure is in $\mathbf{co-NP}$.
\end{definition}
The inference in Definition \ref{def:incontheoryzb} calculates an inconsistent subtheory.
\begin{definition}\label{def:incontheoryzb}
\textbf{Calculating inconsistent theories.}\\
\textbf{UnsatSubtheory($\T,\pS$)} is a derived inference with input theory $\T$ and a partial structure $\pS$ such that there does not exist a model $I$, extending $S$, satisfying $\T$. The inference has as output all theories $\T'$ such that $\T'\subseteq \T$ and there is no model satisfying $T$, extending $\pS$.
Formally, we return:
$$\{\T'|\T'\subseteq \T \land \neg (\exists \I\geq_p \pS \land \I\vDash \T')\}$$
Complexity of deciding if a theory is such an inconsistent theory is in $\mathbf{co-NP}$.
\end{definition}
Note that Definition \ref{def:inconstructure} and \ref{def:incontheoryzb} do not make any statements of minimality.
Using the associated theory $\T_\pS$ and domains structure $\pS_D$ of a partial structure $\pS$, it is possible to consider calculating minimally precise partial configurations as a special case of calculating a minimal inconsistent subset of the theory.
As in \cite{esws/ShchekotykhinFRF14}, we can introduce a ``background theory'' $B\subset \T\cup \T_\S$ (a subset of the theory in which there are assumed to be no conflicts).
We define multiple derived logical inferences, with different degrees of minimality (not-minimal, subset-minimal and minimal in size) of increasing complexity, able to provide explanations to the user.
\begin{definition}\label{def:incontheory}
\textbf{Calculating inconsistent theories with a background.}\\
\textbf{UnsatSubtheory($\T,\pS,B$)} is a derived inference with input theory $\T$, a partial structure $\pS$ and a background theory $B\subseteq \T \cup T_\pS$ such that there does not exist a model $I$, with the domains as in $\pS_D$ satisfying $\T\cup\T_\pS$ (or equivalently: extending $\pS$ and satisfying $\T$), but there is a model satisfying $B$. The inference has as output all theories $\T'$ such that $B\subseteq \T'\subseteq \T\cup\T_\pS$ and there is no model satisfying $T$.
Formally, we return:
$$\{\T'|B\subseteq \T'\subseteq (\T\cup\T_S) \land \neg (\exists \I\geq_p \pS_D \land \I\vDash \T')\}$$
Complexity of deciding if a theory is such an inconsistent theory is in $\mathbf{co-NP}$.
\end{definition}
\begin{definition}\label{def:sminincontheory}
\textbf{Calculating minimal inconsistent theories with a background.}\\
\textbf{MinimalUnsatTheory($\T,\pS,B$)} is a derived inference with input theory $\T$, a partial structure $\pS$ and a background theory $B$ as above.
Output is the subset of subset minimal theories from \textit{UnsatSubtheory($\T,\pS,B$)}.
Complexity of deciding if a set is a subset minimal inconsistent theory is in $\Delta^P_2$.
\end{definition}
\begin{definition}\label{def:cminincontheory}
\textbf{Calculating minimum inconsistent theories with a background.}\\
\textbf{MinimumUnsatTheory($\T,\pS,B$)} is a derived inference with input theory $\T$, a partial structure $\pS$ and a background theory $B$ as above.
Output is the subset of cardinality minimal theories from \textit{MinimalUnsatTheory($\T,\pS,B$)}.
Complexity of deciding if a set is a cardinality minimal inconsistent theory is $\Pi^P_2$.
\end{definition}
Note that Definition \ref{def:inconstructure} is equivalent to calculating a minimal inconsistent subset of a theory $\T\cup \T_\S$, with $B=\T$, if you translate the output back to a pair of a theory and a structure.
Definition \ref{def:incontheoryzb} is equivalent to calculating a minimal inconsistent subset of a theory $\T\cup \T_\pS$, with $B=\T_\pS$, if you translate the output back to a pair of a theory and a structure.
In recent literature multiple approaches are discussed, all mapping to one of our explanation-related inferences. QuickXPlain \cite{aaai/Junker04} is an algorithm that implements Definition \ref{def:incontheory}. The Hitting Set Directed Acyclic Graph (HSDAG) \cite{ai/Reiter87} algorithm calculates subset minimal inconsistent theories (Definition \ref{def:sminincontheory}, as in different ASP solvers \cite{kbse/ShlyakhterSJST03,nmr/Syrjanen06}. Implementations of Definition \ref{def:cminincontheory} have been described in \cite{conf/sat/LynceM04} and \cite{conf/ausai/ZhangLS06}. In our experiment, we have an implementation of Definition \ref{def:sminincontheory} \cite{iclp/WittocxVD09}, where we do however do not calculate the entire set of subset minimal theories. We only calculate one, which gives one explanation of the inconsistency. If the user resolves that problem, he can ask for a new explanation which will point to another reason of inconsistency. This process is reiterated untill all problems are resolved.
\begin{example}
We show a minimal inconsistent subtheory in a situation with $\T$ as in Example $\ref{ExD}$ and $\pS_i$, a partial structure representing an intermediate configuration where a user started with $\pS_0$ and has chosen $Secretary$ for $Requester$, and wants to Install \textit{Office} and \textit{Linux}. This is not possible, and as such, the user asks the system for an explanation in the form of a minimal inconsistent theory. A possible minimal inconsistent theory with $B=\emptyset$, is:
\begin{align*}
& (Install(\text{\textit{Office}}) \land PreReq(\text{\textit{Office}},Windows)) \Rightarrow Install(Windows).\\
&Cost = sum\{(s,PriceOf(s)) | Install(s)\}.\\
&Cost \leq MaxCost(Requester).
\end{align*}
This means that there is no valid configuration because Windows needs to be installed as prerequisite for Office, and the total cost then exceeds the budget of a Secretary.
\end{example}
}
\subtask{Backtracking}
If a value for a parameter is not consistent, the user has to choose a new value for this parameter, or backtrack to revise a value for another parameter.
In Section \ref{ssec:subtasks} we discussed three options of increasing complexity for implementing backtracking functionality.
Erasing a value for a parameter is easy to provide in our KB system, and since this is a generalization of a back button (erasing the last value) we have a formalization of the first two options.
Erasing a value $d$ for parameter $t$ in a partial structure $\pS$ is simply modifying $\pS$ such that $(t=d)^\pS=\Un$.
As with explanation, a number of more complex options can be developed. We look at one possibility.
Given a partial configuration $\pS$, a parameter $p$ and a value $d$ that is inconsistent for that parameter, calculate a minimal set of previous choices that need to be undone such that this value is possible for this parameter. The converse of this problem is very well known under the name of maximum satisfiability problems. In other words, you want to hold on to as much of the structure as possible while ensuring satisfiability.
This problem is closely related to the explanation subtask \cite{conf/aaai/HerasMM11,conf/date/Marques-SilvaP08}. You can imagine the explanation problem as asking the system to point out a mistake in your reasoning. However, solving this mistake will not guarantee you have not made any other mistake in the rest of the problem. What we actually need is a minimal set of things we can remove, so every problem is solved simultaneously.
So more formally, we can use Definition \ref{def:inconstructure} and calculate $UnsatStructure(\T\land (t=d), \pS)$. This inference calculates a set $A$ of sets of previous choices that together are inconsistent.
Undoing an arbitrary choice in all of these sets results in a partial subconfiguration $\pS'$ of $\pS$ such that $d$ is a possible value for $t$ in $\pS'$. To find the maximal partial subconfiguration $\pS'$ that satisfies that property, the minimal hitting set~\cite{ai/Reiter87} of all sets in $A$ has to be calculated.
\subtasksend
\section{Proof of Concept}
\subsection{Implementation}\label{implementation}
In this section we will describe the developed application and implementation.
Our application has a simple client-server architecture. The server plays the role of the reasoning engine, which is mainly a thin wrapper around the \idp system. The client consists of a GUI made in QML~\cite{url:qml} as front-end.
The server converts \idp into a stateless server which is accessible through the web. The client application sends the necessary information, consisting of theories, partial structures and choices, to this server and the server executes the needed inferences. This is a design which involves repeatedly sending over the choices a user has made, but it allows for a very simple architecture to show the feasibility of our design.
This implementation was done in cooperation with Adaptive Planet, a consulting company~\cite{url:adaptiveplanet}
that developed the user interface, and an international banking company that provided us with a substantial configuration problem for testing purposes.
More practical information about this implementation, some screenshots, a
downloadable demo and another example of a configuration system developed with \idp as a reasoning engine (a simpler course configuration demo) can be found at: \url{http://www.configuration.tk}.
\subsubsection{The Reasoning Engine}
As explained before, the application we developed was built on the knowledge base system \idp, which was not developed specifically with configuration problems in mind.
It provides the basic inferences listed at the end of Section \ref{sec:kbs}.
The goal of this experiment was to check if this general infrastructure could be readily applied to applications such as configuration.
In Section \ref{sec:solution} we showed how the tasks which are needed for configuration relate to the infrastructure provided by \idp.
Our main implementation task was to convert these specifications to code.
Some subtasks such as autocompletion did not require any extra work, as this functionality is directly available as the modelexpand inference.
Some functionality, e.g. calculating consequences, did require some work but the existing functionality provided almost all needed components.
We mainly use the existing forms of inference that are readily available in the \idp system. No dedicated or specialized algorithms are used for the configuration subtasks. This proves the point that the KB-paradigm is very flexible but this also means that we had relatively little impact upon the efficiency of our server.
However, the system ended up being quite responsive and we could conclude that \idp (and by extension the KB-paradigm) passed the test for usefulness in this application.
\subsubsection{User Interface}
Apart from a reasoning engine, it is also necessary to have an accessible front end so the user has easy access to the multitude of functionalities which are available.
The front end consists of an application written in the Qt framework using QML~\cite{url:qml} and connects to a configuration engine over the web.
For the purposes of our demo, we developed two different graphical interfaces:
\paragraph{Wizard}
In the wizard interface, the user is interrogated and he answers on subsequent questions selected by the system, using the $GetOpenTerms$ inference.
An important side note here is that the user can choose not to answer a specific question, for instance because he cannot decide as he is missing relevant information or because he is not interested in the actual value (at this point).
These parameters can be filled in at a later timepoint by the user, or by the system, using propagation, or in case the user calls autocompletion.
\paragraph{Drill-Down}
In the drill-down interface, the user sees a list of the still open parameters, and can pick which one he wants to fill in next.
This interface is useful if the user is a bit more knowledgeable about the specific configuration and wants to give the values in a specific order.
\vspace{1em}
In both interfaces the user is assisted in the same way when he enters data.
When he or the system selects a parameter, he is provided with a dropdown list of the possible values, using the $GetConsistentValues$ inference.
Before committing to a choice, he is presented with the consequences of his choice, using the calculate consequences inference.
The nature of the system guarantees a correct configuration and will automatically give the user support using all information it has (from the knowledge base, or received from the user).
\subsection{Evaluation}\label{sec:evaluation}
\subsubsection{Evaluation Criteria} \label{ssec:evaluationcriteria}
When evaluating the quality of software (especially when evaluating declarative methods), scalability (data complexity) is often seen as the most important quality metric.
Naturally when using an interactive configuration system, performance is important.
However, in the configuration community it is known that reasoning about typical configuration problems is relatively easy and does not exhibit real exponential behavior \cite{TiihonenHAS13}.
Also, depending on the application, it is reasonable to expect the number of parameters to be limited, since humans need to fill in the configuration in the end.
When developing a configuration system, challenges lie in the complexity of the knowledge, its high volatility and the complex functionalities to be built.
To get a more complete view of the performance of a configuration system, we chose to evaluate on a larger set of different evaluation criteria.
\newtext{In recent literature \cite{Felfernig14} nine evaluation criteria are used to differentiate between different paradigms used for configuration.
In Section \ref{sec:related}, ten other approaches will be discussed and compared to our solution using the same nine criteria.
\begin{description}
\item[Grapical Modeling Concepts (C1)] is supported if there are standard graphical modeling techniques available that visualize configuration knowledge. They improve understandability, development time and maintenance of new knowledge bases.
\item[Component Oriented modeling (C2)] is a criterion that states that the modeling language is a natural language that allows knowledge base design on the basis of real-world concepts: types, relations, hierarchies, etc.
\item[Automated Consistency Maintenance (C3)] can be broken down to two categories.
Firstly, a system can have support for a priori automated consistency maintenance.
This helps a developer write consistent constraints and verifying correctness while writing the knowledge base.
Secondly, runtime automated consistency maintenance supports the end user, by guaranteeing that every intermediate configuration he can make, can be extended to a valid configuration.
\item[Modularization concepts are available (C4)] if the modeling language is modular and has support for adding additional structure to the knowledge base, for example by organizing the constraints in blocks or groups.
\item[Maintainability (C5)] relates to the adaptability of the knowledge base if the background information changes.
This background information is volatile, it is for example depending on ever-changing company policies.
As such, it is vital that when that information changes, the system can be easily adapted.
When using custom software, all tasks using domain knowledge (like rules and policies) need their own program code.
The domain knowledge is scattered all over the program.
If this policy changes, a programmer has to find all snippets of program code that are relevant for guarding this policy and modify them.
This results in a system that is hard to maintain, hard to adapt and error-prone.
Every time the domain knowledge changes, a whole development cycle has to be run through again.
Some systems have support for intelligent knowledge base navigation tools for complex knowledge spaces.
\item[Model-based (C6)] means that a knowledge base in the system expresses exactly what it means for a configuration to be valid. This in contrast to rule-based configuration, where a knowledge base also contains problem solving knowledge (i.e. information on how the rules should be used/fired).
\item[Efficiency (C7)] relates to efficiency and scalability of the reasoning engine.
\item[Ability to solve generative problem settings (C8)] means that the language supports talking about component types instead of specific objects.
A system supports generic constraints if it allows for constraints that apply to every instance of a component type on which the constraint is defined.
For example, the first constraint of Theory $T$ in Example \ref{ExD} is a generic constraint about all software, without explicitly naming the individual pieces of software.
\item[Ability to provide explanations (C9)] means that the system is able to communicate reasons for inconsistencies or explain why certain choices are forced/prohibited.
\end{description}
}
\subsubsection{Evaluation}\label{ssec:evaluation}
\newtext{
The criteria discussed in previous section are a good way to evaluate the KB implementation of a configuration system.
We evaluate our implementation and the \idp system with these criteria.
\begin{description}
\item[Grapical Modeling Concepts (C1).]
\idp has no support for graphical modeling of domain knowledge and we did not develop any tools for this experiment.
However, it must be noted, that a highly expressive and readable modeling language often makes graphical modeling obsolete.
\item[Component Oriented modeling (C2).] The \fodot language used in this experiment is an extension of typed first-order logic.
First-order logic is about a small set of connectives: $\land, \lor, \neg, \Rightarrow, \Leftrightarrow, \exists, \forall$. These connectives are also the basic connectives of information used by humans.
Classical logic is a good KR language because it has a very clear informal semantics.
It does however not suffice for knowledge representation.
\fodot extends classical logic with a number of extensions that arise from research in AI and KR, such as aggregates, inductive definitions, types, \ldots
This makes \fodot a suited modeling language for a configuration system.
\item[Automated Consistency Maintenance (C3).]
A priori consistency maintenance is supported in the implementation by using the explanation inferences.
If the developer has a collection of constraints that is consistent, it is possible to evaluate if a new constraint leads to an inconsistency and ask the system what other constraints it conflicts with, using for example definition
\ref{def:sminincontheory}.
At runtime consistency maintenance is partially supported, by using the inferences in subtask 2, 3 and 4.
These inferences are theoretically able to guarantee consistency, but due to computational limitations, approximate versions can be used. These are not always able to give the same guarantees.
\item[Modularization concepts are available (C4). ]
The implemented configuration system is modular, since a knowledge base can consist of multiple theories and structures, that together make up the specification.
The explanation inference allows that a user selects background constraints, as in definition \ref{def:sminincontheory}, and in this way he can choose about which constraints he needs feedback.
\item[Maintainability (C5).]
The development of a KB system with a centrally maintained knowledge base makes the knowledge directly available, readable and adaptable.
A well-known advantage of this approach is in maintainability: if domain information changes, the developer can easily modify the knowledge base.
The current implementation does however have no additional support for knowledge base navigation tools.
\item[Model-based (C6).]
The \fodot modeling methodology is based on formulating the properties of a correct configuration in a natural way, such that the models of a specification correspond with configurations.
This is inherently a model-based approach.
\item[Efficiency (C7).] As explained in Section \ref{implementation}, we have only written a thin layer upon existing software which did not target configuration problems specifically.
The performance of the \idp system has been tested extensively in other contexts \cite{ppdp/JansenDDJ14,TPLP/BruynoogheBBDDJLRDV}.
The reasoning engine for \idp is very similar in performance to mainstream ASP solvers~\cite{journals/tplp/CalimeriIR14}. Their performance was tested more extensively in the context of configuration by \citet{TiihonenHAS13}.
It is also very difficult to reliably compare the response times for interactive systems.
Standard benchmarking techniques in software engineering traditionally use instances which need multiple minutes to solve.
In this setting we aim for subsecond response times, for which no standard benchmarks are available as far as we are aware.
In this experiment (a configuration task with 300 parameters and 650 constraints), our users reported a response time of a half second on average with outliers up to 2 seconds.
Note that the provided implementation was a naive prototype and optimizing the efficiency of the implemented algorithms is still possible in a number of ways.
\item[Ability to solve generative problem settings (C8).]
$\fodot$ is an extension of first-order logic, and as such has native support for quantification which is needed for generative problem settings.
\item[Ability to provide explanations (C9).] Subtask 7 and 8 in Section \ref{sec:solution} are inferences that are used to support giving explanations.
The implemented configuration system has an implementation of definition \ref{def:sminincontheory}.
\end{description}
}
\todo{PIETER: ASP resultaten vermelden}
\begin{comment}
\paragraph{Functionality.}
For evaluating functionality, industrial partners involved in this project have tested the proof of concept and compared with their conventional software solutions.
The most common approach to developing configuration tools is building custom software.
Other frequently used technology to handle interactive configuration problems are spreadsheets and business rules systems.
When starting this project, the users had the following major issues with these systems, for which conceptual, general solutions were given by our approach:
\begin{itemize}
\item \textbf{Unidirectional dataflow:}
All these systems have an obligatory unidirectional dataflow. This fixes beforehand which parameters are input and which parameters are output. However, given a problem statement, it is not natural to make a distinction between input and output.
Different users may have different information or different needs and regard different parameters as input.
In our approach, this distinction is not made at all by our inferences.
\item \textbf{Incomplete knowledge:}
These systems have problems reasoning with incomplete knowledge, i.e., rules and functions can only compute their result when their input is complete and they also cannot use partial knowledge to deduce (partial) new knowledge, e.g., to eliminate configuration options.
Our language does by nature accommodate for partial knowledge, and is able to represent every intermediate partial configuration.
These partial configurations are used by the inferences to calculate possible total configurations, consequences, etc.
\end{comment}
\section{Related Work}
\label{sec:related}
\subsection{Other approaches}
In different branches of AI research, people have been focusing on configuration software in different settings.
The following discussion of knowledge-based approaches is based on a book in recent literature~\cite{Felfernig14}.
After the discussion we will compare the ten approaches with our approach (\textbf{IDP}).
Historically, the first knowledge-based configuration systems were \emph{rule-based} (\textbf{RBS}) \cite{ai/McDermott82,cacm/BarkerO89}.
These systems operate on a working memory and if the condition of a rule is fulfilled, it fires and modifies the working memory, applying the conclusion of that rule.
Rule-based systems are sensitive to rule orderings. This complicates modification of the rule-base.
More importantly, inclusion of problem solving knowledge in the rule-base, makes a rule-base problem specific and focused towards one specific task.
This leads to the same problems as in imperative languages. To solve different tasks, more rule-bases have to be built, leading to duplication and fanning out of knowledge, giving issues in maintainability.
\emph{Constraint Satisfaction Problems} are widely used for tackling configuration problems~\cite{ijcai/MittalF89,FleischanderlFHSS98}.
A (static\footnote{In constrast to dynamic and generative constraint satisfaction problem.}) constraint satisfaction problem (\textbf{SCSP}) is a triple $(V,D,C)$ of a set of domain variables $V=\{v_1, v_2, \ldots, v_n\}$, a set of domains $\{dom(v_1), dom(v_2), \ldots, dom(v_n)\}$ and set of constraints $C$.
A solution for a SCSP is an assignment $S$ of domain elements $d_i\in dom(v_i)$ to variables $v_i$, such that each variable has a value in $S$ and constraints $C$ are satisfied by $S$.
A configuration task in SCSP is searching for a solution for a SCSP $(V,D,C)$, where $C$ contains the configuration constraints together with the user preferences.
To make efficient CSP configuration systems, different techniques have been used, such as local search \cite{Li2005}, symmetry breaking \cite{ki/KiziltanFH01} and knowledge compilation techniques such as binary decision diagrams \cite{cp/HadzicA05}.
In response to limitations of SCSP in configuration, extensions have been developed.
\emph{Dynamic Constraint Satisfaction Problems} (\textbf{DCSP})~\cite{aaai/MittalF90} allow for variables to be inactive or irrelevant.
If a variable is inactive, it does not need a value in a solution (for example, when configuring a smartphone, no camera resolution is needed if no camera is present).
\emph{Generative Constraint Satisfaction Problems} (\textbf{GCSP})~\cite{FleischanderlFHSS98} extends SCSP with component types and generative constraints.
\citet{splc/Janota08} studied a mapping of CSP to SAT to use a SAT solver to provide functionality for a configuration system.
There exist many graphical approaches for doing knowledge configuration, and visualizing a configuration model.
\citet{kang1990feature} used \emph{feature models} (\textbf{FM}) for modeling these concepts, while \textbf{UML} was proposed in \cite{aicom/FalknerH13}.
FM and UML configuration approaches have no reasoning algorithms, they need to be used with external algorithms.
\citet{splc/KaratasOD10} for example combined feature models with constraint logic programming (CLP) to provide reasoning and automated analysis.
Decidable subsets of first-order logic, \emph{description logics} (\textbf{DL}) are used often in context of the semantic web.
They have also been used for the development of configuration systems \cite{HotzKDSN06,McGuinnessW98}.
The trade-off for having decidable subsets of first-order logic is that they are limited in expressivity. This make domain knowledge in these systems less readable, less natural and harder to maintain.
An ontology based method was also proposed by Vanden Bossche et al. \citeyear{VandenBosscheRMVP07} using OWL.
Tiihonen et al. developed a configuration system WeCoTin~\cite{TiihonenHAS13}, based on \emph{Answer Set Programming} (\textbf{ASP}).
WeCoTin uses Smodels, an ASP system, as inference engine, for propagating consequences of choices.
Answer set programming (ASP) is a form of declarative programming based on the stable-model semantics~\cite{iclp/GelfondL88} for logic programs.
The architecture of their reasoning engine is closely related to the reasoning engine we use.
Also, in language, many similarities can be identified~\cite{DeneckerLTV12}, as they both have their roots in extended logic programming.
Combinations of the above approaches are also proposed in literature, called \emph{hybrid} (\textbf{HB}) configuration systems.
Typically, they use a DL-based representation for the ontology, together with constraints. They combine reasoning engines from these fields to provide inference \cite{HotzKDSN06}.
\subsection{Comparison of approaches}
\citet{Felfernig14} evaluated all these paradigms with respect to the evaluation criteria from Section \ref{ssec:evaluationcriteria}.
In Table \ref{tab:comparison}, we show this evaluation, together with scores for our implementation in the $\textbf{IDP}$ column, based on the discussion of Section \ref{ssec:evaluation}.
\begin{table}[h]
\centering
\caption{Comparison of systems from Section \ref{sec:related} using criteria from Section \ref{sec:evaluation} as in (Felfernig et al. 2014).
We use a \Checkmark to mark good support, a $\approx$ for partial support and a $-$ to denote that no support is available.}
\label{tab:comparison}
\begin{tabular}{@{}cccccccccccc@{}}
& RBS & SCSP & DCSP & GCSP & SAT & FM & UML & DL & ASP & HB & IDP \\
C1 & - & - & - & - & - & \Checkmark & \Checkmark & $\approx$ & - & $\approx$ & - \\
C2 & - & - & - & \Checkmark & - & - & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark \\
C3 & - & $\approx$ & $\approx$ & $\approx$ & $\approx$ & - & - & $\approx$ & $\approx$ & $\approx$ & $\approx$ \\
C4 & $\approx$ & - & - & \Checkmark & - & - & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark \\
C5 & - & $\approx$ & $\approx$ & $\approx$ & $\approx$ & $\approx$ & $\approx$ & $\approx$ & $\approx$ & $\approx$ & \Checkmark \\
C6 & - & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark \\
C7 & \Checkmark & \Checkmark & \Checkmark & \Checkmark & \Checkmark & - & - & $\approx$ & $\approx$ & $\approx$ & $\approx$ \\
C8 & $\approx$ & - & - & \Checkmark & - & - & - & - & $\approx$ & \Checkmark & \Checkmark \\
C9 & $\approx$ & \Checkmark & \Checkmark & \Checkmark & $\approx$ & - & - & \Checkmark & \Checkmark & \Checkmark & \Checkmark
\end{tabular}
\end{table}
All these approaches are focused towards one specific inference: ontologies are focused on deduction, rule systems are focused on backward/forward chaining, etc.
These approaches are less general then the KB paradigm, which is specifically designed to reuse the knowledge for different reasoning tasks.
The contributions of this paper are different from previously discussed approaches: we analyzed IC problems from a Knowledge Representation point of view.
This paper is a discussion of possible approaches and the importance of this point of view.
We made a study of desired functionalities for an IC system and how we can define logical reasoning tasks to supply these functionalities.
As far as we are aware, the language we used in this experiment is more expressive than earlier approaches.
The expressivity of the language is crucial for the usability of the approach.
It allows us to address a broader range of applications, moreover it is easier to formalize and maintain the domain knowledge.
Not discussed by \citet{Felfernig14} et al is work by \citet{ppdp/VlaeminckVD09}. They did a preliminary experiment using the KB approach for interactive configuration, also using the $\fodot$ \idp project.
It is on this work that we continue in this paper by analyzing a real-life application of a larger scale and discussing new functionalities and inferences.
This theoretical approach benefits from (1) the expressive language to express domain knowledge adequately and (2) the general basic inferences that realise derived inferences in an easy way, supporting the discussed functionalities, resulting in a IC system that scores very well with relation to the evaluation criteria (Table \ref{tab:comparison}).
An interesting remark in Table \ref{tab:comparison} is that the IDP column resembles the GCSP column, a generalisation of CSP, developed for configuration.
The IDP-system has better support for C5 (maintainability), due to the high level modeling language and the strict seperation between domain knowledge and reasoning.
GCSP has better efficiency results. This can be partly explained by the fact that CSP uses dedicated algorithms for reasoning over global constraints such as \textit{alldifferent}.
The goal of reusing knowledge makes that we typically do not make use of this kind of specific algorithms, since a dedicated algorithm can only be developed with one specific inference in mind.
\section{Challenges and Future Work} \label{sec:challenge}
Interactive configuration problems are part of a broader kind of problems, namely service provisioning problems.
Service provisioning is the problem domain of coupling service providers with end users, starting from the request until the delivery of the service.
Traditionally, such problems start with designing a configuration system that allows users to communicate their wishes, for which we provided a knowledge-based solution.
Once all the information is gathered from a user, it is still necessary to make a plan for the production and delivery of the selected configuration.
Hence the configuration problem is followed by a planning problem
that shares domain knowledge with the configuration problem but that
also has its own domain knowledge about providers of components,
production processes, etc.
This planning problem then leads to a monitoring problem.
Authorizations could be required, payments need to be checked, or it could be that the configuration becomes invalid mid-process.
In this case the configuration needs to be redone, but preferably without losing much of the work that is already done.
Companies need software that can manage and monitor the whole chain, from
initial configuration to final delivery and this without duplication
of domain knowledge. This is a problem area where the KB approach
holds great promise but where further research is needed to integrate
the KB system with the environment that the company uses to follow up its
processes.
Other future work may include language extensions to better support con\-figuration-like tasks.
A prime example of this are templates~\cite{tplp/DassevilleHJD15}. Oftentimes the theory of a configuration problem contains lots of constraints which are similar in structure. It seems natural to introduce a language construct to abstract away the common parts.
Another useful language extension is reification, to talk about the symbols in a specification rather than about their interpretation.
Reification allows the system to reason on a meta level about the symbol and for example assign symbols to a category like ``Technical'' or ``Administrative''.
\vspace{-1em}
\section{Conclusion}
\vspace{-0.5em}
The KB paradigm, in which a strict separation between knowledge and problem solving is proposed, was analyzed in a class of knowledge intensive problems: interactive configuration problems.
As we discussed why solutions for this class are hard to develop, we proposed a novel approach to the configuration problem based on an existing KB system.
We analyzed the functional requirements of an IC system and investigated how we can provide these, using logical inferences on a knowledge base.
We identified interesting new inference methods and applied them to the interactive configuration domain.
We studied this approach in context of a large application, for which we built a proof of concept, using the KB system as an engine, which we extended with the new inferences.
As proof of concept, we solved a configuration problem for a large banking company.
Results are convincing and open perspectives for further research in service provisioning.
\bibliographystyle{acmtrans}
|
\section{SUPPLEMENTAL MATERIAL}
\section{Proof of theorem 1 and generalization to arbitrary priors}
Here we provide a lower and an upper bound on the probability of correct state discrimination, valid for a generic set of linearly independent pure states $\{ |\Psi_k\rangle\}_{k=1}^n$ and for a generic choice of prior probabilities $\{ p_k\}_{k=1}^n$.
The bounds are expressed in terms of the Gram matrix of the weighted states
\begin{align}
|\widetilde \Psi_k\rangle : = \sqrt {p_k}\, |\Psi_k\rangle \, ,
\end{align}
that is, the matrix $W$ with elements
\begin{align}W_{ij} : = \langle \widetilde \Psi_i| \widetilde \Psi_j\rangle \, .
\end{align}
The maximum success probability can be estimated with the following Theorem, which generalizes Theorem~1 of the main text to arbitrary prior distributions:
\begin{theo}
Let $\{ |\Psi_k\rangle\}_{k=1}^n$ be a set of linearly independent pure states.
The maximum probability of correctly identifying a state drawn from the set $\{ |\Psi_k\rangle\}_{k=1}^n$ with probability $\{p_k\}_{k=1}^n$ satisfies the bounds
\begin{align}\label{app:lower}
P_{\max} \ge \frac{\left( {{\rm tr}\, \sqrt W}\right)^2 } n
\end{align}
and
\begin{align}\label{app:upper}
P_{\max} \le \frac{\left( {{\rm tr}\, \sqrt W}\right)^2 } n + \sqrt{n\, \lambda_{\max }} \, \| {\bf q} - {\bf u} \|_1 \, ,
\end{align}
where $\lambda_{\max}$ is the maximum eigenvalue of $W$, \mbox{${\bf q} = \{q_k\}$} is the probability distribution defined by \mbox{$q_k : = \big( \sqrt W\,\big)_{kk}/{\rm tr}\,\big(\sqrt W\,\big)$}, $ {\bf u} = \{u_k\}$ is the uniform distribution, and $\| {\bf q} - {\bf u}\|_1 : =\sum_k |q_k-u_k|$ is the trace norm.
\end{theo}
{\bf Proof.} Since the states $\{ |\Psi_k\rangle\}_{k=1}^n$ are linearly independent, the optimal measurement consists of orthogonal rank-one projectors~\cite{belavkin}. Let us denote the projectors by $M_k = |m_k\rangle\langle m_k|$, for a suitable orthonormal basis $\{ |m_k\rangle\}_{k=1}^n$. Then, the probability of correct discrimination can be written as
\begin{align}
\nonumber P_{\rm succ} &= \sum_k p_k \, |\langle m_k| \Psi_k\rangle|^2 \\
\nonumber & = \sum_k |\langle m_k| \widetilde \Psi_k\rangle|^2\\
& = \sum_k \left| B_{kk} \right|^2\, ,
\end{align}
where $B$ is the matrix defined by the relation
\begin{align} |\widetilde \Psi_k \rangle = \sum_i \, B_{ik} \, |m_i \rangle \, .
\end{align}
By definition, one has $B^\dag B = W$.
Hence, the polar decomposition yields the relation
\begin{align}\label{polar}
B = U \sqrt W \, ,
\end{align}
for a suitable unitary matrix $U$.
Note that a generic change of orthonormal basis,
\begin{align}
|m_i\rangle \to |m_i' \rangle = \sum_j V_{ji} \, |m_j\rangle
\end{align}
(where $V$ is a generic unitary matrix), results into the change of matrix
\begin{align}\label{changeV}
B \to B' = V^\dag B \, .
\end{align}
Combining Eqs. (\ref{polar}) and (\ref{changeV}), the maximum probability of correct discrimination can be expressed as
\begin{align}\label{pmax}
P_{\max} = \max_U \, \sum_k \left| \left( U \sqrt W \right )_{kk} \right|^2 \, .
\end{align}
Setting $U=\openone$, one has the lower bound
\begin{align}
P_{\max} \ge \sum_k \left( \sqrt W \right )^2_{kk} \ge \frac{\left( {\rm tr}\, \sqrt W \right)^2}n \, ,
\end{align}
the second inequality following from the convexity of the function $f(x)= x^2$. This proves the lower bound (\ref{app:lower}).
Let us prove the upper bound (\ref{app:upper}). Using the Cauchy-Schwarz inequality in Eq.~(\ref{pmax}), we obtain the upper bound
\begin{align*}
P_{\max} & = \max_U \, \sum_{k} \left| \sum_{s}\left(U W^{\frac{1}{4}}\right)_{ks} \left(W^{\frac{1}{4}}\right)_{sk} \right|^2 \\
& \le \max_U \, \sum_k \left( \sqrt W\right)_{kk} \, \left( U \sqrt W U^\dag \right)_{kk} \\
& = {\rm tr}\, \left(\sqrt W\right) \times \, \max_U \left[ \sum_k q_k \, \left( U \sqrt W U^\dag \right)_{kk} \right] \, .
\end{align*}
Moreover, the argument of the maximum can be upper bounded as
\begin{align*}
\sum_k q_k \, \left( U \sqrt W U^\dag \right)_{kk} &\le \sum_k \frac 1n \, \left( U \sqrt W U^\dag \right)_{kk} \\
& \quad + \sum_k \left|q_k - \frac 1n \right| \left( U \sqrt W U^\dag \right)_{kk} \\
& \le \frac { {\rm tr}\, \sqrt W }n + \| {\bf q}- {\bf u} \|_1 \, \sqrt{ \lambda_{\max} }\, .
\end{align*}
Finally, from Eq. (\ref{app:lower}) we have the bound
\begin{align}
{\rm tr}\, \sqrt W \le \sqrt {n \, P_{\max}} \le \sqrt n \, .
\end{align}
Combining the above inequalities we obtain the desired upper bound (\ref{app:upper}). $\blacksquare$
\medskip
When the prior distribution is uniform, the weighted Gram matrix $W$ is given by $W= G/n$, where $G$ is the unweighted Gram matrix used in the main text. Substituting this relation into the bound (\ref{app:upper}) one obtains Eq.~(5) of the main text.
\section {\boldmath Eigenvalues and eigenvectors of $H$}
In this section we derive explicit expressions for the eigenvectors and eigenvalues of the matrix $H$.
To this purpose, it is useful to first recall some properties of
the Chebyshev polynomials of the second kind, denoted by $U_n(x)$. The Chebyshev polynomials of the second kind can be defined as the characteristic polynomial of the tridiagonal matrix $T$ of size~$n$ whose entries are $T_{ij}=\delta_{i\,j+1}+\delta_{j\,i+1}$. Specifically, \mbox{$U_n(x)=\det(2x\,\openone-T)$}, i.e., the eigenvalues of~$T$ are defined to be twice the roots, $x_l$, of~$U_n(x)$. By expanding the determinant by the first row one readily obtains the well known recursion relation $U_n(x)=2x \,U_{n-1}(x)-U_{n-2}(x)$~\cite{abramowitz}. One can check that this recursion relation along with the initial conditions in standard form, $U_0(x)=1$, $U_{-1}(x)=0$, give the right characteristic polynomial for {\em any} size of $T$. It suffices to check the~$n=1,2$ cases. One has~$U_1(x)=2x$ and $U_2(x)=4x^2-1$, which are indeed the characteristic polynomials of $T$ of sizes 1 and~2.
We now turn to the eigenvalues and eigenvectors of~$H$, which we will compute using a different approach. The matrix $H$ is nothing but the matrix $T$ with the addition of two extra entries at each end of the principal diagonal, namely,
$H_{ij}= T_{ij}+c\,( \delta_{i \,1}\delta_{j\,1}+ \,\delta_{i\, n}\delta_{j\, n})$. Let us denote by~$2x_l$ the $l$-th eigenvalue of $H$ and by $\vec{w}^{\, l}$ the corresponding
unnormalized eigenvector, chosen with the convention~$w^l_1=1$. The equation $H\vec{w}^{\, l}=2x_l \vec{w}^{\, l}$
is equivalent to the following system of linear equations:
\begin{eqnarray}
\label{eigenvectors}
c w_1^l+w^l_2&=& 2x_l\, w_1^l;
\nonumber \\
w^l_{j-1}+w^l_{j+1} &=& 2x_l\, w_j ^l, \qquad 2\le j\le n-1; \\
w_{n-1}^l+ c w^l_{n}&=& 2x_l\, w_n^l.
\nonumber
\end{eqnarray}
The second line of this system can be viewed as the recursion relation $w^l_{j+1} = 2x_l\, w_j ^l - w^l_{j-1}$,
which is the recursion relation of the Chebyshev polynomials given above (with $n\to j+1$). It follows that the first and second line of Eq.~(\ref{eigenvectors}), along with the convention $w^l_1=1$, imply
\begin{equation}
w^l_j=U_{j-1}(x_l)-c\,U_{j-2}(x_l),\qquad j=1,2,\dots, n.
\label{wl}
\end{equation}
Since all the components of $\vec{w}^{\, l}$ have been determined, the third line in Eq.~(\ref{eigenvectors}) must give the eigenvalues of~$H$.
By substituting Eq.~(\ref{wl}) in the third line of Eq.~(\ref{eigenvectors}) and using the Chebyshev recursion relation again, one obtains
\begin{equation}
0=U_n(x_l)-2c\,U_{n-1}(x_l)+c^2 U_{n-2}(x_l):=P_n(x_l),
\label{P_n again}
\end{equation}
which must hold for $l=1,\dots,n$. The polynomial~$P_n(x)$ has degree $n$ and its $n$ roots, $x_l$, give the eigenvalues of $H$ as $2x_l$. Note that $P_n(x)$ has to be proportional to the characteristic polynomial of $H$, i.e., $P_n(x)\propto \det(2x\openone-H)$, as both polynomials have the same degree and the same zeroes.
\section {\boldmath Distribution of the eigenvalues of $H$}
Here we analyze the distribution of the zeroes of the polynomial $P_n(x)$ defined in Eq. (\ref{P_n again}).
Setting $x= \cos \theta$, we recall that the Chebyshev polynomial $U_n (\cos \theta)$ can be expressed as \cite{abramowitz}
\begin{equation}
U_n(\cos\theta)={\sin(n+1)\theta\over \sin\theta} \, .
\label{U_n(cos)}
\end{equation}
Then, a little bit of trigonometry yields the relation
\begin{equation}
P_n(\cos\theta)=A(\theta)\sin\left[n\theta+\delta(\theta)\right],
\label{beat}
\end{equation}
with
\begin{eqnarray}
\left \{
\begin{array}{lll} A(\theta)&:=&\displaystyle{1-2c\cos\theta+c^2\over\sin\theta}\\
&&\\
\delta(\theta)&:=&\displaystyle\arctan{(1-c^2)\sin\theta\over(1+c^2)\cos\theta-2c}
\end{array}
\right. \, ,
\end{eqnarray}
where $0<\delta(\theta)<\pi$.
{}From Eq. (\ref{beat}) we can see that every zero of $P_n(\cos\theta)$ must be the solution to one of the equations
\begin{align} n\theta+\delta(\theta)= l \pi \, , \qquad l=1,2,\dots,n \, .
\end{align}
Denote by $\theta_l$ the angle that solves the $l$-th equation. Since $\delta (\theta_l)$ is contained in the interval $(0,\pi)$, we have the bound
\begin{equation}
{\pi \over n} l \le \theta_l \le {\pi \over n} \left(l+1\right) \, .
\end{equation}
In other words, the interval $(0,\pi)$ can be divided into intervals of length $\pi/n$, with the $l$-th interval containing the zero $\theta_l$. For large $n$, this means that the zeros are uniformly distributed in the interval $(0,\pi)$.
\iffalse
As a result, we can approximate sums with integrals, according to the relation
\begin{equation}
\label{sum to int}
\lim_{n\to \infty} \sum_{l=1}^{n} \frac{1}{n}f(\cos\theta _l) = \frac{1}{\pi}\int_0^\pi d \theta f(\cos\theta ) \, .
\end{equation}
valid for any function $f$ that is continuous in the interval $[0,\pi]$.
\fi
\iffalse To make this statement even more explicit, let us take~$a$ and~$b$ such that $0<a<b<\pi$. Let
$l_a$ be the smallest integer such that $a/\pi<l_a/n$. Analogously,
let $l_b$ be the largest integer such that $l_b/n<b/\pi$. Let~${\mathscr N}(a,b)$ be the number of zeroes $\theta_l$ in $(a,b)$, which satisfies
\begin{equation}
l_b-l_a-2\le{\mathscr N}(a,b)\le l_b-l_a+2.
\end{equation}
It immediately follows from the inequalities above that
\begin{equation}
\label{uniformity}
\left|{{\mathscr N}(a,b)\over n}-{b-a\over \pi}\right|<{2\over n}.
\end{equation}
In words, for large $n$, the ratio of the number of zeroes in the interval $(a,b)$ and $(0,\pi)$ approaches the ratio of the corresponding interval lengths for {\em any} interval $(a,b)$ in $(0,\pi)$, which proves that the zeroes are inded uniformly distributed on $(0,\pi)$.
\fi
\iffalse
Alternatively, a general proof that the zeros of a monic polynomial
are uniformly distributed in the asymptotic limit [WHAT IS $n$ IN THE GENERAL CASE?] was given by Erd\" os and Tur\'an in~\cite{erdos}. However, our proof shows that for the problem at hand uniformity holds up to order $1/n$, whereas Erd\"os and Tur\'an's result can only ensure order $\sqrt{\log n/n}$. Our result is a necessary condition to show in turn that the corrections to the asymptotic success probability are of order $1/n$.
\fi
\section{\boldmath The trace and diagonal matrix elements of $\sqrt G$}
Here we evaluate the normalized trace $\sqrt G/n$ and we quantify its deviation from the limit value $\gamma: = \lim_{n\to\infty } {\rm tr}\, \sqrt G / n$. In the process of computing the trace, we also evaluate the diagonal matrix elements of $\sqrt G$, which will become useful in the next section.
We proceed along the following steps:
\begin{enumerate}
\item construct the normalized eigenvectors of $\sqrt G$
\item evaluate the diagonal elements
\item evaluate the trace.
\end{enumerate}
\subsection{The normalized eigenvectors of $\sqrt G$}
The eigenvectors of $\sqrt G$ coincide with the eigenvectors of the matrix $H$, provided in Eq. (\ref{wl}).
Setting $x_l = \cos \theta_l$, we can use the trigonometric representation of the Chebyshev polynomials given in Eq.~(\ref{U_n(cos)}). In this way, we obtain
\begin{equation}
{w}^l_j={\sin( j\theta_l)-c \sin[(j-1)\theta_l]\over \sin\theta_l}.
\end{equation}
Now, the norm $\vec w^l$ can be evaluated explicitly as
\begin{align}
\|\vec w^l\|^2&:= \sum_j | w^l_j|^2 \nonumber \\
& ={n\over2\sin^2\theta_l}
\left\{
1-2c \cos\theta_l+c^2\phantom{{\sin(2n\theta_l)\over2n\sin\theta_l}}\right.\nonumber\\
& \quad +{1-c^2\over2n}\left[1-\cos(2n\theta_l)\right]\nonumber\\
&\quad- \left. {\sin(2n\theta_l)\over2n\sin\theta_l}\left[(1+c^2)\cos\theta_l-2c\right] \right\} \\
& = {n\over2\sin^2\theta_l}
\left\{
1-2c \cos\theta_l+c^2 + \frac{ f_n(\theta_l)}n \right\} \, , \nonumber
\end{align}
having defined the function
\begin{align}
\nonumber f_n(x) &: = {1-c^2\over2}\left[1-\cos(2n x)\right] \\
& \quad - {\sin(2n x)\over2\sin x}\left[(1+c^2)\cos x -2c\right] \, .
\end{align}
Defining the normalized eigenvectors $\vec v^l:=\vec w^l/\|\vec w^l\|$, we then have
\begin{align}\label{vl}
\left| v^l_j \right|^2 = \frac 2 n \frac{ \left [\, \sin( j\theta_l)-c \sin(j-1)\theta_l\, \right]^2 }{ 1-2c \cos\theta_l+c^2 + f_n(\theta_l)/n} \, .
\end{align}
\iffalse
\begin{align}
\|\vec w^l\|^2&:= \sum_j | w^l_j|^2 \nonumber \\
& ={n\over2\sin^2\theta_l}
\left\{
1-2c \cos\theta_l+c^2\phantom{{\sin(2n\theta_l)\over2n\sin\theta_l}}\right.\nonumber\\
& \quad +{1-c^2\over2n}\left[1-\cos(2n\theta_l)\right]\nonumber\\
&\quad- \left. {\sin(2n\theta_l)\over2n\sin\theta_l}\left[(1+c^2)\cos\theta_l-2c\right] \right\} \, . \nonumber
\end{align}
For $c\not = 1$, we can neglect the second and third term in the sum, obtaining
\begin{align}
\|\vec w^l\|^2 & ={n\over2\sin^2\theta_l}
\left\{
1-2c \cos\theta_l+c^2 + O\left( \frac 1n\right) \right\} \, .
\end{align}
Hence, the normalized eigenvectors $\vec v^l:=\vec w^l/\|\vec w^l\|$ have the simple asymptotic expression
\begin{equation}\label{vl}
v^l_j=\sqrt{2\over n} \left\{ {\sin j\theta_l-c\sin(j-1)\theta_l\over \sqrt{1-2c \cos\theta_l+c^2}} + O\left(\frac 1{n}\right)\right\} \, .
\end{equation}
\fi
\subsection{The diagonal elements of $\sqrt G$}
Having computed the eigenvalues and eigenvectors of the Gram matrix $G$, we can now evaluate the diagonal elements of its square root $\sqrt G$.
We start from the expression
\begin{align}
\sqrt G = \sum_l \sqrt {\lambda_l} \, |v^l\rangle\langle v^l| \, ,
\end{align} recalling that the eigenvalues are given by
\begin{align}\label{lambdal}\lambda_l = {\frac{1-c^2}{ 1-2c \cos\theta_l+c^2}} \, .
\end{align}
Then, the diagonal elements of $\sqrt G$ are
\begin{align}
\left( \sqrt G\right)_{kk} & = \sum_l \sqrt{\lambda_l} \, |v^l_k|^2 \, ,
\end{align}
with $v^l_k$ given as in Eq. (\ref{vl}). Explicitly, the matrix element $\big(\sqrt G\,\big)_{kk}$ is given by
\begin{align}
\nonumber \left( \sqrt G\right)_{kk}
\nonumber & = \frac 1 n \, \sum_l \sqrt{ {\frac{1-c^2}{ 1-2c \cos\theta_l+c^2}}} \\
\label{sqrtGkk}& \qquad \times \, {{\left[ \sin k\theta_l-c\sin(k-1)\theta_l\right]^2}\over { 1-2c \cos\theta_l+c^2 + f_n(\theta_l )/n }}
\end{align}
We now show that most of the matrix elements $\big(\sqrt G\,\big)_{kk}$ are approximately equal to the limit value
$ \gamma: =\lim_{n\to \infty} {\rm tr}\, \sqrt G /n$.
Note that $\gamma$ can be computed explicitly in terms the eigenvalues: indeed, one has
\begin{align} \nonumber
\gamma & = \lim_{n\to \infty} \frac 1 n \, \sum_l \sqrt \lambda_l \\
\nonumber & = \lim_{n\to \infty} \frac 1 n \, \sum_l \sqrt{\frac{1-c^2}{ 1-2c \cos\theta_l+c^2}} \\
& = \frac 1 \pi \, \int_0^\pi {\rm d} \theta\, \sqrt{\frac{1-c^2}{ 1-2c \cos\theta+c^2}} \, .
\end{align}
We now show that the deviation vanishes for all values of $k$ in the interval $[ n^\epsilon , n- n^\epsilon]$.
To this purpose, we evaluate the matrix element $\big( \sqrt G\,\big)_{kk}$ at the leading order of the large $n$ asymptotics, obtained by replacing the sum in Eq. (\ref{sqrtGkk}) by an integral and by dropping the term~$f_n( \theta_l)/n$ in the denominator. In this way, we obtain the approximate equality
\begin{align}\label{approx}
\left( \sqrt G\right)_{\!kk} \!\! \approx \frac {\sqrt{1\!-\!c^2}}\pi \! \int_0^\pi \!\! {\rm d} \theta \, {{\left[ \sin k\theta-c\sin(k-1)\theta \right]^2}\over { (1-2c \cos\theta +c^2 )^{3/2} }} \, .
\end{align}
Then, some elementary algebra gives
\begin{equation}
\left(\sqrt G\right)_{kk}-\gamma \approx
{\sqrt{1\!-\!c^2}\over\pi}\left( 2c I_{2k-1}\!-\!I_{2k}\!-\!c^2 I_{2k-2}\right) \, ,
\label{deviation}
\end{equation}
where the integrals $I_r$ are defined as
\begin{equation}
I_r\! :=\!\int_0^\pi\! {\cos r\theta\; d\theta\over\left(1-2c \cos\theta+c^2\right)^{3/2}} \, .
\label{I_m}
\end{equation}
We then show that the integrals $I_r$ vanish exponentially with $r$:
\begin{prop}\label{prop:integral}
For $c<1$, the leading order of the integral $I_r$ in Eq. (\ref{I_m}) is given by
\begin{equation}
I_r={2 c^r\sqrt{\pi r}\over(1-c^2)^{3/2}} \, .
\label{I_m asympt}
\end{equation}
\end{prop}
The proof can be found in the end of this subsection. Inserting the asymptotic expression (\ref{I_m asympt}) into Eq. (\ref{deviation}) we obtain the relation
\begin{equation}\label{skkminusave}
\left( \sqrt G\right)_{kk}\!-\gamma \approx {1 \over {4 (1-c^2)}} ~ {{c^{2k}} \over {\sqrt{2\pi k^3}}} \, ,
\end{equation}
valid in the interval $[ n^\epsilon , n- n^\epsilon]$. In conclusion, the deviation $\big( \sqrt G\,\big)_{kk} - \gamma$ decays exponentially with $k$.
We stress that the error introduced by the approximation (\ref{approx}) is negligible with respect to the leading order, quantified by the r.h.s. of Eq. (\ref{skkminusave}). This point is illustrated in Fig.~\ref{fig:3}, which compares the the r.h.s. of Eq. (\ref{skkminusave}) with the exact values of the deviation, computed by direct numerical evaluation of $\sqrt G$ from $G$. Log-scale plots are shown for various values of the overlap~$c$, setting $n=30$ and letting~$k$ vary from 1 to 15.
The agreement is extremely good and backs up the validity of the approximation (\ref{approx}) even for small values of $k$.
\begin{figure}[bhtp]
$$
\includegraphics[width=27em]{deviation_figure}
$$
\caption{ Log-scale plots of the deviation $\big( \sqrt G\,\big)_{kk}-\gamma$, for $n=30$ and $k$ varying from 1 to 15. The solid lines are the asymptotic approximation on r.h.s of Eq.~(\ref{skkminusave}) and the dots are the result of numerical evaluation of $\sqrt G$ from $G$.}
\label{fig:3}
\end{figure}
In the next subsection we will use Eq. (\ref{skkminusave}) to quantify the deviation between the diagonal of the matrix $\sqrt G/{\rm tr}\,(\sqrt G)$ and the uniform distribution.
\iffalse A similar situation occurs for $ n/2 \le k \ll n$. In this case, the matrix element are evaluated through the identity
\begin{align}
\left(\sqrt G\right)_{n-k,n-l}= \left(\sqrt G\right)_{k,l} \, ,
\end{align}
which yields the relation
\begin{equation}\label{reciprocal}
\left( \sqrt G\right)_{kk}\!-\!\gamma= {1 \over {4 (1-c^2)}} ~ {{c^{2 (n-k)}} \over {\sqrt{2\pi (n-k)^3}}} \, ,
\end{equation}
valid for $n/2\le k\ll n$.
\fi
\medskip
{\bf Proof of Proposition \ref{prop:integral}.}
We start by noticing that the integral on the r.h.s of Eq.~(\ref{I_m}) can be expressed as a contour integral over the unit circle~$C$ on the complex plane:
\begin{equation}
I_r=
{1\over2i}\oint_C dz\;{z^{r+1/2}\over\left[z-c(z^2+1)+c^2 z\right]^{3/2}} .
\label{I_m complex}
\end{equation}
\begin{figure}[bhtp]
$$
\includegraphics[width=2.1in]{cuts_contours}
$$
\caption{The figure shows (gray) the branch cuts of the integrand in Eq.~(\ref{I_m complex}) and the contour $C'$ used to obtain Eq.~(\ref{I_m 2 int}).}
\label{fig:2}
\end{figure}
We can choose the branch of the integrand so that its branch cuts are the intervals~$[0,c]$ and~$[c^{-1},\infty)$ on the real axis. Since this branch is analytic elsewhere~and the cut~$[c^{-1},\infty)$ is outside $C$, we can deform the contour~$C$ to a new contour $C'$ around $[0,c]$. One readily sees that the integrand in Eq.~(\ref{I_m complex}) behaves as $(z-c)^{-3/2}$, so care must be taken to evaluate the new contour integral near the end point $z=c$, as some divergencies may arise because of this singular behavior. Specifically, we choose~$C'$ as in Fig.~\ref{fig:2}, where $\epsilon>0$ and the limit $\epsilon\to0$ is implicit. As a result, $I_r$ has a contribution coming from the discontinuity of the integrand along the interval $[0,c-\epsilon]$ and a contribution coming from the integration around the circle $C_\epsilon$ of radius $\epsilon$ and center at $z=c$:
\begin{eqnarray}
I_r
&=&-{1\over c^{3/2}}\int_{0}^{c-\epsilon}dx\;{x^{r+1/2}\over\left\{(c-x)[(1/c)-x]\right\}^{3/2}}\nonumber\\
&+&{1\over2ic^{3/2}}\oint_{C_\epsilon}dz{c^{r+1/2}\over(z-c)^{3/2}[(1/c)-c]^{3/2}}.
\label{I_m 2 int}
\end{eqnarray}
Note that the limit $\epsilon\to0$ of each separate line is ill-defined, as they both diverge as $\epsilon^{-1/2}$. To circumvent this problem,
we write $(c-x)^{-3/2}=2(d/dx)(c-x)^{-1/2}$ and integrate the first line of~Eq.~(\ref{I_m 2 int}) by parts. In doing so, we see that the $\epsilon^{-1/2}$ terms cancel and we obtain the simple expression
\begin{equation}
I_r=2\int_{0}^{c}{dx\over (c-x)^{1/2}}{d\over dx}{x^{r+1/2}\over[1- c x]^{3/2}}.
\end{equation}
We can further simplify this expression using the change of variable $x= c t$, which enables us to express $I_r$ in terms of hypergeometric functions. However, we are just interested in the asymptotic behavior of $I_r$. Keeping only the leading contribution as $r$ goes to infinity, we have
\begin{equation}
I_r=2r c^r\int_0^1 dt\, t^{r-1/2}(1-t)^{-1/2}(1-c^2 t)^{-3/2}.
\label{I_m t-int}
\end{equation}
The asymptotic behaviour of this integral can be easily evaluated by noticing that the leading contribution comes from the region near the upper limit of integration, so we can set~$t=1$ in the last factor in Eq.~(\ref{I_m t-int}) and write
\begin{eqnarray}
I_r\!\!&=&\!\!{2rc^r\over(1-c^2)^{3/2}}\int_0^1\!\! dt\, t^{r-1/2}(1-t)^{-1/2}\nonumber\\[.5em]
\!\!&=&\!\!{2rc^r B(\mbox{${1\over2}$},r\!+\!\mbox{${1\over2}$})\over(1\!-\!c^2)^{3/2}}\nonumber\\[.5em]
&=&{2\sqrt\pi c^r\,\Gamma(r\!+\!\mbox{${1\over2}$})\over (r\!-\!1)!(1\!-\!c^2)^{3/2}},
\label{I_m Gamma}
\end{eqnarray}
where $B(a,b)$ is the Euler Beta function,
\begin{eqnarray}
B(a,b)=\int_0^1 dt\; t^{a-1}(1-t)^{b-1},\\[-.1em]\nonumber
\end{eqnarray}
and we have used the relation,
\begin{equation}
B(a,b)={\Gamma(a)\Gamma(b)\over\Gamma(a+b)}.
\end{equation}
Using the Stirling formula in the third line of Eq.~(\ref{I_m Gamma}), we finally obtain Eq.~(\ref{I_m asympt}). $\blacksquare$
\subsection{The trace of $\sqrt G$}
Here we show that the normalized trace ${\rm tr}\, \sqrt G/n$ is close to its limit value $\gamma$, up to an error of size $1/n^{1-\epsilon}$, where $\epsilon$ is an arbitrary constant in the interval $(0,1)$.
For this purpose, we divide the values of $k$ into two subsets, defined as
\begin{align}
\nonumber {\mathsf S} &: = \{ k \in \mathbb N : \lceil n^{\epsilon} \rceil \le k \le n-\lceil n^{\epsilon} \rceil \ \} \\
\label{sets} \overline{\mathsf S} & : = \{ 1,\dots, n\}\setminus \mathsf S \, .
\end{align}
The trace of $\sqrt G$ can be evaluated as
\begin{align}
\nonumber {\rm tr}\, \sqrt G & = \sum_{k \in \mathsf S} \, \left( \sqrt G\right)_{kk} + \sum_{k \in \overline {\mathsf S}} \, \left( \sqrt G\right)_{kk} \\
\nonumber & = \sum_{k \in \mathsf S} \, \left( \gamma + {1 \over {4 (1-c^2)}} ~ {{c^{2k}} \over {\sqrt{2\pi k^3}}} \right) + \sum_{k \in \overline {\mathsf S}} \, \left( \sqrt G\right)_{kk} \\
& = | \mathsf{S} \, | \gamma + \sum_{k \in \overline {\mathsf S}} \, \left( \sqrt G\right)_{kk} + O \left( n^{1-3\epsilon/2}c^{2n^\epsilon}\right) \, ,
\end{align}
the second equality following from Eq. (\ref{skkminusave}).
Using the above expression it is easy to produce upper and lower bounds on ${\rm tr}\, \sqrt G$. An upper bound is obtained as follows:
\begin{align}\label{upperboundtrsqrt}
\nonumber{\rm tr}\,\sqrt G & \le | \mathsf{S}| \, \gamma + \sqrt{ \lambda_{\max}} \, |\overline {\mathsf S}| + O \left( n^{1-3\epsilon/2}c^{2n^\epsilon}\right) \\
& \le n \, \gamma + \sqrt{\lambda_{\max}} |\overline {\mathsf S}| + O \left( n^{1-3\epsilon/2}c^{2n^\epsilon}\right) \, .
\end{align}
Similarly, we have the lower bound
\begin{align}
\nonumber{\rm tr}\,\sqrt G & \ge | \mathsf{S}| \, \gamma + O \left( n^{1-3\epsilon/2}c^{2n^\epsilon}\right) \\
& = n \, \gamma - \gamma |\overline {\mathsf S}| + O \left( n^{1-3\epsilon/2}c^{2n^\epsilon}\right) \, .
\end{align}
Using the relation $ 2 n^{\epsilon} \le |\overline {\mathsf S}| \le 2 \, (n^{\epsilon}+1) $ we finally obtain
the bounds
\begin{align}
\gamma - \frac{ 2\gamma}{n^{1-\epsilon}} \le \frac{{\rm tr}\, \sqrt G}n \le \gamma + 2 \sqrt{\lambda_{\max}} \left( \frac{ 1}{n^{1-\epsilon}} + \frac 1n \right) \, ,
\end{align}
valid up to an exponentially small correction of size~$ O \left( n^{-3\epsilon/2}c^{2n^\epsilon}\right)$.
More compactly, the above bounds can be written as
\begin{align}
\left| \frac{ {\rm tr}\, \sqrt G} n - \gamma \right| \le \frac { 2 \, \max \{ \gamma, \sqrt{ \lambda_{\max}}\} }{n^{1-\epsilon}} + O \left( \frac 1n\right) \, .
\end{align}
Now, direct inspection shows that $\sqrt{\lambda_{\max}}$ is always larger than $\gamma$. Hence, the bound becomes
\begin{align}\label{normtracebound}
\left| \frac{ {\rm tr}\, \sqrt G} n - \gamma \right| \le \frac { 2 \sqrt{ \lambda_{\max}} }{n^{1-\epsilon}} + O \left( \frac 1n\right) \, .
\end{align}
\section{\boldmath Deviation of ${\bf q}$ from the uniform distribution}
In this section we consider the probability distribution $ {\bf q} = \{ \big( \sqrt G \, \big)_{kk}/{\rm tr}\, \sqrt G\}$ for the change point problem and we quantify the deviation of $ \bf q$ from the uniform distribution.
Here we upper bound the trace distance between the probability distribution ${\bf q }=\{ q_k\} $ defined by
\[ q_k : = \frac{ \left( \sqrt G \right)_{kk}}{{\rm tr}\, \sqrt G}\] and the uniform distribution, denoted by $\bf u$. Our strategy is to separately analyze the contributions to the trace distance coming from the two sets $\mathsf S$ and $\overline {\mathsf S}$ defined in Eq.~(\ref{sets}).
Let us consider first the contribution of the set $\mathsf S$. For~$k\in\mathsf S$, we have
\begin{align}
\nonumber
\left| q_k - \frac 1n \right| & = \left| \frac { \left( \sqrt G\right)_{kk} - {\rm tr}\, \sqrt G /n} { {\rm tr}\, \sqrt G} \right| \\
& \le \frac { \left| \left( \sqrt G\right)_{kk} - \gamma \right| + \left| \gamma - {\rm tr}\, \sqrt G /n \right| } { {\rm tr}\, \sqrt G} \, .
\end{align}
Now, the first term is upper bounded as
\begin{align}
\nonumber \left| \left( \sqrt G\right)_{kk} - \gamma \right| & \le {1 \over {4 (1-c^2)}} ~ {{c^{2n^\epsilon}} \over {\sqrt{2\pi n^{3\epsilon}}}} \\
\label{primo} & = O \left( n^{-3\epsilon/2}c^{2n^\epsilon}\right) \, ,
\end{align} due to Eq. (\ref{skkminusave}). The second term is upper bounded by Eq. (\ref{normtracebound}).
Hence, the contribution of $\mathsf S$ to the trace distance can be upper bounded as
\begin{align}
\nonumber
\sum_{k\in\mathsf S} \left| q_k - \frac 1n \right| & \le \sum_{k\in\mathsf S} \frac{ { 2 \sqrt{ \lambda_{\max}} }/{n^{1-\epsilon}} + O \left( 1/n \right) } {{\rm tr}\, \sqrt G} \\
\nonumber
& \le \sum_{k\in\mathsf S} \frac{ { 2 \sqrt{ \lambda_{\max}} }/{n^{1-\epsilon}} + O \left( 1/n\right) } {n\sqrt {\lambda_{\min}}} \\
\nonumber
& \le \sqrt{\frac{ \lambda_{\max}} {\lambda_{\min}}} \, \frac{2}{n^{1-\epsilon}} + O \left( \frac 1n\right)
\, , \label{vanish1}
\end{align}
having used the relation
\begin{align}
{\rm tr}\, \sqrt G \ge n \, \sqrt{\lambda_{\min}} \, ,
\end{align}
where $\lambda_{\min}$ is the minimum eigenvalue of $G$.
In conclusion, Eq. (\ref{vanish1}) shows that the contribution of the set $\mathsf S$ vanishes in the large $n$ limit.
Let us consider the contribution of the set $\overline {\mathsf S}$. For $k\in \overline{\mathsf S}$, we have the inequality
\begin{align}
\nonumber
\left| q_k - \frac 1n \right| & = \left| \frac { \left( \sqrt G\right)_{kk} - {\rm tr}\, \sqrt G /n} { {\rm tr}\, \sqrt G} \right| \\
\nonumber & \le { \sqrt{\lambda_{\max}} \over { {\rm tr}\, \sqrt G }} \\
& \le \sqrt{\frac{\lambda_{\max}}{\lambda_{\min}}} \, \frac 1n \, ,
\end{align}
which leads to the upper bound
\begin{align}
\label{vanish2} \sum_{k\in \overline{\mathsf S}} \left| q_k - \frac 1n \right| &
\le \sqrt{\frac{\lambda_{\max}}{\lambda_{\min}}} \, \frac2 {n^{1-\epsilon}} \, .
\end{align}
Using the bounds (\ref{vanish1}) and (\ref{vanish2}), the deviation between~$\bf q$ and the uniform distribution can be upper bounded as
\begin{align}
\nonumber \| {\bf q } - {\bf u}\|_1 & = \sum_{k\in {\mathsf S}} \left| q_k - \frac 1n \right| + \sum_{k\in \overline{\mathsf S}} \left| q_k - \frac 1n \right| \\
\nonumber & \le
\sqrt{\frac{ \lambda_{\max}} {\lambda_{\min}}} \, \frac{4}{n^{1-\epsilon}} + O \left( \frac1n\right) \\
& \le
{\frac{1+c} {1-c}} \, \frac{4}{n^{1-\epsilon}} + O \left( \frac 1n\right) \, ,
\end{align}
having used the bounds
\begin{align} \lambda_{\max} \le (1+c)/(1-c)
\end{align} and
\begin{align}\lambda_{\min} \ge (1-c)/(1+c) \, ,
\end{align} following from Eq. (\ref{lambdal}).
\section{Lower bound on the success probability of the square root measurement}
For a generic set of linearly independent pure states $\{ |\Psi_k\rangle\}_{k=1}^n$ and a generic choice of prior probabilities~$\{ p_k\}_{k=1}^n$, the success probability of the square root measurement can be expressed as \cite{pozza}
\begin{align}
P_{\rm SQ} = \sum_k \, \left( \sqrt W \right)_{kk}^2 \, .
\end{align}
The convexity of the function $f(x)= x^2$ then implies the bound $ P_{\rm SQ} \ge \big( {\rm tr}\, \sqrt W \, \big)^2/n $.
\section{Greedy strategy and Bayesian updating}
Here we show that Bayesian updating gives the optimal greedy strategy introduced in the main text. This follows from the observation that the optimal measurement (and the optimal guess) at step $s$ of the greedy strategy are determined solely by the posterior probability distribution after the measurement at step $s-1$, as will be explicitly shown at the end of this section.
To optimize the greedy strategy, we need to maximize ${\mathscr P}^{\rm G}_s=\sum_{r=1}^{n} \eta_{r}^{(s)} \langle \Psi_{r} |E_s(r) |\Psi_r\rangle$ over all POVM measurements on particle~$s$, $\{E_s(r)\}_{r=1}^n$.
Noticing that the source state $|\Psi_k\rangle$ restricted to particle~$s$ is $|\Psi_k\rangle_{\!s}=|0\rangle$ for $s< k$, and $|\Psi_k\rangle_{\!s}=|\phi\rangle$ for $s\ge k$, the following relations are self evident:
\begin{eqnarray}
\label{ps-t}
{\mathscr P}^{\mathrm{G}}_{s}\!\!\!&=&\!\!\!
\sum_{r=1}^{s} \eta_{r}^{(s)}\, \langle \phi | E_s(r) | \phi \rangle \! +\! \sum_{r=s+1}^n \eta_{r}^{(s)} \, \langle 0 | E_s(r)| 0 \rangle \nonumber\\
\!\!\!& \leq&\!\!p_{\phi}^{(s)} \!\sum_{r=1}^{s} \langle \phi | E_s(r) | \phi \rangle\! +\! p_{0}^{(s)}\! \sum_{r=s+1}^n \langle 0 | E_s(r)| 0 \rangle \nonumber \\[.5em]
\!\!\!&=&\!\! \! p_{\phi}^{(s)}\! \ \langle \phi | \Pi_s(\phi)| \phi \rangle + p_{0}^{(s)}\, \langle 0 | \Pi_s(0) | 0 \rangle ,
\end{eqnarray}
where
$p_{\phi}^{(s)} :=\max_r{\{\eta_{r}^{(s)}\}_{r=1}^{s}}$, $p_{0}^{(s)}=\max_r{\{\eta_{r}^{(s)}\}_{r=s+1}^n}$, $\Pi_{s}(\phi)=\sum_{r=1}^{s} E_s(r)$,
and $\Pi_s(0)=\openone -\Pi_s(\phi)$.
The inequality is saturated by choosing a new POVM $\{E'_s(r)\}_{r=1}^n$ whose elements are non-zero only in the two positions that maximize the prior probabilities:
\begin{equation}
r_\phi=\argmax_{r}{\{\eta_{r}^{(s)}\}_{r=1}^{s}},\; r_0=\argmax_{r}{\{\eta_{r}^{(s)}\}_{k=s+1}^n},
\end{equation}
so that
$E'_s ( r_0 ) = \Pi_s (0)$ and $E'_s ( r_1 ) = \Pi_s (\phi ) $. This justifies the choice of priors in Eq.~(12) of the main text.
The success probability can now we written in terms of the Helstrom matrix $\Gamma_s=p_{\phi}^{(s)} |\phi\rangle\langle\phi|-p_{0}^{(s)} |0\rangle\langle0|$ as:
\begin{eqnarray}
{\mathscr P}^{\mathrm{G}}_{s}&=&p_{0}^{(s)}+ {\rm tr}\,\left(\Pi_{s}(\phi) \Gamma_s\right)\leq p_{0}^{(s)}+ {\rm tr}\, \left(\Gamma_s^{+}\right)\nonumber \\
&=&\frac{1}{2}\left( p_{\phi}^{(s)}+p_{0}^{(s)}+{\rm tr}\,|\Gamma_{s}|\right),
\end{eqnarray}
where $\Gamma_s^{+}$ is the positive part of matrix $\Gamma$.
The inequality is saturated by choosing $\Pi_s(\phi)$ to be the projector onto the positive subspace of $\Gamma$~\cite{helstrom}.
We now show that the optimal measurement and guess at step $s$ of the greedy strategy
do not depend on the particular sequence measurement outcomes, but only on the posterior probability distribution after the measurement at step $s-1$.
Let us introduce the short-hand notation ${\mathbf{r}_{s}}:=\{r_{1},\ldots , r_{s}\}$ for a sequence of results obtained up to step~$s$.
The average success probability at each step $s$ is given by
\begin{equation}
\label{ps-tbayes}
\sum_{k=1}^n\sum_{\mathbf{r}_{s}}p(\mathbf{r}_{s},k) \delta_{k\,\hat{k}(\mathbf{r}_{s})}
\leq \sum_{\mathbf{r}_{s}}\max_{k}{p(\mathbf{r}_{s},k)},
\end{equation}
where $p(\mathbf{r}_{s},k)$ is the joint probability of obtaining the sequence $\mathbf{r}_{s}$ of results and the change point occuring at position $k$, and $\hat{k}(\mathbf{r}_{s})\in \{1,\ldots, n\}$ is the decision function that assigns to each $\mathbf{r}_{s}$ the guessed change point position~$k=\hat{k}(\mathbf{r}_{s})$. The inequality can be saturated by
$\hat{k}(\mathbf{r}_{s})=\argmax_{k} p(\mathbf{r}_{s},k)$. Since the source states $|\Psi_k\rangle$ are of product form, we can write
\begin{equation}
p(\mathbf{r}_{s},k)={1\over n}p(\mathbf{r}_{s-1}| k)\,\langle\Psi_k|E_{s}(r_s) |\Psi_k\rangle ,
\end{equation}
where we recall that $|\Psi_k\rangle$ restricted to particle $s$ is $|\Psi_k\rangle_{\!s}=|0\rangle$ for $s< k$, and $|\Psi_k\rangle_{\!s}=|\phi\rangle$ for $s\ge k$. The measurement over the $s$-th particle is represented by
the POVM $\{E_{s}(r)\}_{r=1}^n$ and it is understood that it may depend on the sequence~$\mathbf{r}_{s-1}$ of previous results. Hence,
the optimal greedy average success probability at step~$s$, can be written as
\begin{equation}
P^{\rm G}_{s}=
\sum_{\mathbf{r}_{s-1}}p(\mathbf{r}_{s-1}){\mathscr P}^{\rm G}_{s}({\mathbf{r}_{s-1}}),
\end{equation}
where the probability of successful identification of the change point at step $s$ conditioned to the occurrence of the sequence $\mathbf{r}_{s-1}$ [${\mathscr P}^{\rm G}_s$ in
Eq.~(\ref{ps-t}); we recall that the dependence on ${\bf r}_{s-1}$ is understood there]~is
\vspace{-.2cm}
\begin{eqnarray}
{\mathscr P}^{\rm G}_{s}({\mathbf{r}_{s-1}})\!\!&=&\!\!\max_{\{E_{s}(r)\}}\sum_{r=1}^n\max_{k} p(k|\mathbf{r}_{s-1})\nonumber
\\
\!\!&\times&\!\!\langle\Psi_k|E_{s}(r) |\Psi_k\rangle,
\label{only priors}
\\[-.5em]
\nonumber
\end{eqnarray}
\\[-1em]
and we have used Bayes' rule to obtain the relation $(1/n)\,p(\mathbf{r}_{s-1}|k)=p(k|\mathbf{r}_{s-1}) p(\mathbf{r}_{s-1})$. From Eq.~(\ref{only priors}), it is apparent that the optimal measurement can only depend on the updated priors $\eta_{k}^{(s)}:=p(k|\mathbf{r}_{s-1})$, rather than on the whole sequence of previous results, as the maximization is only subject to the POVM conditions $E_s(r)\ge0$ and $\sum_{r=1}^n E_s(r)=\openone$. Likewise, the optimal guess can only depend on $\eta_{r}^{(s)}$ [Eq.~(\ref{ps-tbayes}) and the paragraph below~it].
\bibliographystyle{unsrt}
|
\section{Introduction}
In the last few years, many-body physics has embraced a novel class of systems, the so-called quantum fluids of light \cite{Carusotto2013}. In these optical systems, light and matter combine to generate new photonlike particles that, differently from vacuum photons, are characterized by sizeable effective masses and mutual interactions and, therefore, may give rise to novel states of matter.
One of the most used platforms to study the physics of quantum fluids of light is the semiconductor planar microcavity, in which the cavity photons and the quantum-well excitons strongly couple to form mixed light-matter interacting bosonic quasiparticles called exciton polaritons \cite{Deng2010}. Numerous quantum-hydrodynamics collective phenomena have been investigated theoretically and successfully observed experimentally in such exciton-polariton fluids \cite{Carusotto2013}. Nevertheless, fluids of light in cavity-based systems are inevitably subject to losses, which is typically detrimental for the experimental observation of coherent quantum dynamical features. A more promising configuration for the study of quantum phenomena in fluids of light consists in the paraxial propagation of a quasimonochromatic beam of light in a nonabsorbing bulk nonlinear medium of Kerr type.
It is well known \cite{Boyd1992, Agrawal1995, Rosanov2002} that in such a cavityless, propagating, geometry the complex amplitude of the classical optical field is a slowly varying function of space and time which satisfies a nonlinear wave equation formally identical to the Gross--Pitaevskii (GP) equation of dilute Bose--Einstein (BE) condensates \cite{Pitaevskii2016} after exchanging the roles of the propagation coordinate and of the time parameter. This classical paraxial bulk dynamics may be regarded as the emerging mean-field description of an underlying quantum nonlinear Schr\"odinger dynamics, as formalized in full generality in a recent work by two of us \cite{Larre2015}.
In a recent experimental study \cite{Sun2012}, \textsc{C.~Sun} {\etal} have provided the first observation of classical-wave condensation using a beam of classical monochromatic light propagating in a nonlinear photorefractive crystal. The mechanism underlying this condensation of classical light finds its origin in the thermalization of the classical optical field \cite{Picozzi2007, Lagrange2007, Picozzi2008, Picozzi2008Bis, Barviau2008, Barviau2009, Suret2010, Klaers2010, Klaers2010Bis, Aschieri2011, Michel2011} towards an equilibrium state whose statistics obeys the Rayleigh--Jeans (RJ) thermal law, which corresponds to the classical (high-temperature and/or long-wavelength) limit of the BE distribution.
In this letter, we push this research line forward by investigating the very quantum aspects of the thermalization dynamics of the propagating fluid of light. Making use of the fully quantum theory developed in ref.~\cite{Larre2015}, we discuss the possibility of measuring the Boltzmann tails of the eventual BE distribution, which constitutes the hallmark of the particlelike, quantum, nature of the paraxial beam of light at thermal equilibrium. Inspired by recent advances towards atom-laser devices based on in-waveguide evaporative-cooling schemes \cite{Mandonnet2000, Castin2000, Lahaye2004, Lahaye2005}, we finally propose a mechanism leading to a complete BE condensation in the quantum fluid of light. If realized, such a process would offer a novel route to generate spontaneous optical coherence in a novel concept of coherent-light source.
\section{Quantum formalism}
We consider the propagation in the positive-$z$ direction of a paraxial, spectrally narrow beam of light of central angular frequency $\omega$ in a bulk, electrically neutral, nonmagnetic, nonabsorbing, nonlinear medium of real-valued intensity-dependent refractive index $n_{0}+n_{1}(\mathbf{r}_{\perp},z)+n_{2}\,|\mathcal{E}|^{2}$. Here, $n_{0}$ is the background refractive index, $n_{1}[\mathbf{r}_{\perp}=(x,y),z]$ describes the spatial profile of the refractive index, $n_{2}$ quantifies the strength of the ---spatially local and instantaneous--- Kerr nonlinearity of the medium and $\mathcal{E}$ is the slowly varying \cite{Boyd1992, Agrawal1995, Rosanov2002} envelope of the light wave's electric field $\mathrm{Re}[\mathcal{E}\,\mathrm{e}^{\mathrm{i}(\beta_{0}z-\omega t)}]$ of propagation constant $\beta_{0}=n_{0}\,\omega/c$ in the increasing-$z$ direction, where $c$ denotes the vacuum speed of light. For simplicity's sake, we neglect light polarization and we assume that Raman and Brillouin light-scattering processes on phonons in the optical medium occur at a negligible rate.
Following ref.~\cite{Larre2015}, it is possible to map the quantum propagation of the beam of light in the positive-$z$ direction onto a quantum nonlinear Schr\"odinger evolution of a closed system of many interacting photons in a three-dimensional space spanned by the two-dimensional transverse position vector $\mathbf{r}_{\perp}$ and by the physical time parameter $t$. Introducing the time parameter $\tau=\beta_{1}\,z$ and the three-dimensional position vector $\mathbf{r}=(\mathbf{r}_{\perp},\zeta=t/\beta_{1}-z)$, where $\beta_{1}=\mathrm{d}\beta_{0}/\mathrm{d}\omega=(n_{0}+\omega\,\mathrm{d}n_{0}/\mathrm{d}\omega)/c$ denotes the inverse of the group velocity of the photons in the medium at $\omega$, the quantum mechanical propagation equation of the light beam may be reformulated in the Heisenberg form $\mathrm{i}\,\hbar\,\partial\hat{\Psi}/\partial\tau=[\hat{\Psi},\hat{H}]$, where the quantum field operator $\hat{\Psi}=[c\,\varepsilon_{0}\,n_{0}\,\beta_{1}/(2\,\hbar\,\omega)]^{1/2}\,\hat{\mathcal{E}}$ is the second-quantized slowly varying envelope of the electric field, normalized ($\varepsilon_{0}$ is the vacuum permittivity) in a way to satisfy the usual equal-$\tau$ Bose commutation relations $[\hat{\Psi}(\mathbf{r}_{1},\tau),\hat{\Psi}^{\dag}(\mathbf{r}_{2},\tau)]=\delta^{(3)}(\mathbf{r}_{1}-\mathbf{r}_{2})$ and $[\hat{\Psi}(\mathbf{r}_{1},\tau),\hat{\Psi}(\mathbf{r}_{2},\tau)]=0$, and where
\begin{align}
\notag
\hat{H}&\left.=\int\mathrm{d}^{3}r\;\bigg[\frac{\hbar^{2}}{2\,m_{\perp}}\,\frac{\partial\hat{\Psi}^{\dag}}{\partial\mathbf{r}_{\perp}}\cdot\frac{\partial\hat{\Psi}}{\partial\mathbf{r}_{\perp}}+\frac{\hbar^{2}}{2\,m_{\zeta}}\,\frac{\partial\hat{\Psi}^{\dag}}{\partial\zeta}\,\frac{\partial\hat{\Psi}}{\partial\zeta}\right. \\
\label{Eq:Hamiltonian}
&\left.\hphantom{=}+U(\mathbf{r}_{\perp},\tau)\,\hat{\Psi}^{\dag}\,\hat{\Psi}+\frac{g}{2}\,\hat{\Psi}^{\dag}\,\hat{\Psi}^{\dag}\,\hat{\Psi}\,\hat{\Psi}\bigg]\right.
\end{align}
is the many-body Hamiltonian operator of the system.
In eq.~\eqref{Eq:Hamiltonian}, $U(\mathbf{r}_{\perp},\tau)=-\hbar\,\omega/(c\,\beta_{1})\,n_{1}(\mathbf{r}_{\perp},z)$ is the external potential experienced by the photons, due to the spatial variation of the refractive index, and $g=-2\,(\hbar\,\omega)^{2}/(c^{2}\,\varepsilon_{0}^{\hphantom{2}}\,n_{0}^{\hphantom{2}}\,\beta_{1}^{2})\,n_{2}^{\hphantom{2}}$ is the strength of the effective photon-photon interactions induced by the Kerr nonlinearity.
Even more importantly, $m_{\perp}=\hbar\,\beta_{0}\,\beta_{1}$ and $m_{\zeta}=-\hbar\,\beta_{1}^{3}/\beta_{2}^{\vphantom{3}}$ are the effective masses of the paraxial photons in, respectively, the transverse $\mathbf{r}_{\perp}$ plane and the $\zeta$ direction. In generic media, the values of $m_{\perp,\zeta}$ are typically very different, as they have completely different physical origins: the former originates from paraxial diffraction in the transverse plane while the latter, inversely proportional to the group-velocity-dispersion parameter $\beta_{2}=\mathrm{d}\beta_{1}/\mathrm{d}\omega=(2\,\mathrm{d}n_{0}/\mathrm{d}\omega+\omega\,\mathrm{d}^{2}n_{0}/\mathrm{d}\omega^{2})/c$ of the medium at $\omega$, starts playing a crucial role for nonmonochromatic optical fields having a nontrivial time dependence. Unless the carrier frequency $\omega$ lies in the neighborhood of some optical resonance where dispersion is strong, $m_{\perp}$ is generally much smaller than $m_{\zeta}$; as an example, using tabulated data for fused silica \cite{Malitson1965} around $1.55~\mu\mathrm{m}$ ($1~\mu\mathrm{m}$), one obtains a ratio $m_{\perp}/m_{\zeta}\simeq7\times10^{-3}$ ($m_{\perp}/m_{\zeta}\simeq-8\times10^{-3}$).
As the Hamiltonian \eqref{Eq:Hamiltonian} is only valid within a limited angular-frequency and wavevector range around $(\omega,\beta_{0})$, one has to ensure that photon-photon scattering induces no sizeable photon population outside this paraxial region. Thanks to the conservation of the energy \eqref{Eq:Hamiltonian}, a necessary and ---unless the chromatic dispersion has an unusually complex shape--- sufficient condition is that the two masses $m_{\perp,\zeta}$ have the same sign. The robustness of a coherent photon wave against modulational instabilities imposes further conditions that the longitudinal mass be positive, $m_{\zeta}>0$, and the photon-photon interactions be repulsive, $g>0$; by definition, this amounts to assume that the dielectric is characterized by an anomalous group-velocity dispersion, $\beta_{2}<0$, and a self-defocusing Kerr nonlinearity, $n_{2}<0$ \cite{Larre2015}.
\section{Thermalization time}
In this section, we provide an analytical estimate of the time $\tau_{\mathrm{th}}$ ---that is, of the propagation distance $z_{\mathrm{th}}=\tau_{\mathrm{th}}/\beta_{1}$ along the Kerr medium--- that is necessary for the isolated quantum fluid of light described by the Hamiltonian \eqref{Eq:Hamiltonian} to thermalize. It is worth stressing that the thermalization process is here assumed to occur \textit{via} photon-photon collisions within the fluid only, and not to involve any thermal equilibration with the underlying optical medium, \textit{e.g.}, by photon-phonon scattering or repeated absorption-emission cycles as it was instead the case in the experiment of refs.~\cite{Klaers2010, Klaers2010Bis}.
Assuming for the sake of simplicity that the dielectric is spatially homogeneous, $n_{1}(\mathbf{r}_{\perp},z)=0$, \textit{i.e.}, $U(\mathbf{r}_{\perp},\tau)=0$ in eq.~\eqref{Eq:Hamiltonian}, and that the total interaction energy is small with respect to the total kinetic one in the eventual thermal-equilibrium state, the latter has to be characterized by an occupation number in the plane-wave state of wavevector $\mathbf{k}=[\mathbf{k}_{\perp}=(k_{x},k_{y}),k_{\zeta}]$ and energy $E_{\mathbf{k}}=\hbar_{\vphantom{\perp}}^{2}\,\mathbf{k}_{\perp}^{2}/(2\,m_{\perp}^{\vphantom{2}})+\hbar_{\vphantom{\zeta}}^{2}\,k_{\zeta}^{2}/(2\,m_{\zeta}^{\vphantom{2}})$ of the BE form
\begin{equation}
\label{Eq:BoseEinsteinDistribution}
N_{\mathrm{BE}}(E_{\mathbf{k}},T,\mu)=\bigg[\!\exp\!\bigg(\frac{E_{\mathbf{k}}-\mu}{k_{\mathrm{B}}\,T}\bigg)-1\bigg]^{-1}
\end{equation}
($k_{\mathrm{B}}$ is the Boltzmann constant), where $T$ and $\mu$ are respectively the temperature and the chemical potential of the thermalized quantum fluid of light. As we have assumed there is no thermal contact with the underlying optical medium, $T$ is not related to the temperature of the latter as in refs.~\cite{Klaers2010, Klaers2010Bis}, and both $T$ and $\mu$ are fully determined as functions of the energy and number densities of the photon fluid entering the medium, as detailed in the next section.
A simple model ---based on the quantum nonlinear Schr\"odinger formalism \eqref{Eq:Hamiltonian}--- to investigate the relaxation dynamics of the initial state of the photon fluid, at $\tau=0$ (\textit{i.e.}, $z=0$), towards thermal equilibrium, at $\tau\gtrsim\tau_{\mathrm{th}}$ (\textit{i.e.}, $z\gtrsim z_{\mathrm{th}}$), is provided by the homogeneous [as $U(\mathbf{\mathbf{r}_{\perp}},\tau)=0$] Boltzmann kinetic equation \cite{Griffin2009}
\begin{align}
\notag
\frac{\partial N_{\mathbf{k}}}{\partial\tau}&\left.=\frac{2\,g^{2}}{\hbar}\int\frac{\mathrm{d}^{3}k_{2}}{(2\pi)^{3}}\,\frac{\mathrm{d}^{3}k_{3}}{(2\pi)^{3}}\,\frac{\mathrm{d}^{3}k_{4}}{(2\pi)^{3}}\right. \\
\notag
&\left.\hphantom{=}\times(2\pi)^{3}\,\delta^{(3)}(\mathbf{k}+\mathbf{k}_{2}-\mathbf{k}_{3}-\mathbf{k}_{4})\right. \\
\notag
&\left.\hphantom{=}\times\hphantom{(}2\pi\hphantom{)^{3}}\,\delta(E_{\mathbf{k}}+E_{\mathbf{k}_{2}}-E_{\mathbf{k}_{3}}-E_{\mathbf{k}_{4}})\right. \\
\notag
&\left.\hphantom{=}\times[(N_{\mathbf{k}}+1)\,(N_{\mathbf{k}_{2}}+1)\,N_{\mathbf{k}_{3}}\,N_{\mathbf{k}_{4}}\right. \\
\label{Eq:BoltzmannEquation}
&\left.\hphantom{=}-N_{\mathbf{k}}\,N_{\mathbf{k}_{2}}\,(N_{\mathbf{k}_{3}}+1)\,(N_{\mathbf{k}_{4}}+1)]\right.
\end{align}
for the uniform phase-space density $N_{\mathbf{k}}=N_{\mathbf{k}}(\tau)$ of the paraxial photons occupying the plane-wave state of wavevector $\mathbf{k}$ and energy $E_{\mathbf{k}}$ at the propagation time $\tau$. At long times, \textit{i.e.}, when $\tau\gtrsim\tau_{\mathrm{th}}$, the solution $N_{\mathbf{k}}$ of eq.~\eqref{Eq:BoltzmannEquation} approaches the stationary BE distribution \eqref{Eq:BoseEinsteinDistribution}. Equation \eqref{Eq:BoltzmannEquation} is valid (i) in the absence of condensate and (ii) in the weak-interaction regime. The constraint (i) is satisfied as long as one considers energies and densities yielding noncondensed equilibrium states; otherwise, one has to include the coherent dynamics of the condensate's order parameter in eq.~\eqref{Eq:BoltzmannEquation} \cite{Griffin2009}. The condition (ii) may be checked \textit{a posteriori} by requiring that, in the eventual thermal state, the total interaction energy is small compared to the total kinetic energy, as already supposed in the second paragraph of the present section.
To estimate the effective relaxation time $\tau_{\mathrm{th}}$ towards thermal equilibrium, we are going to mutuate well-known results from the theory of weakly interacting atomic Bose gases. A numerical study \cite{Wu1996} demonstrated that the thermalization time $\tau_{\mathrm{th}}$ of weakly interacting bosonic atoms not too far from thermal equilibrium is typically of the order of $3/\gamma$, where $\gamma$ denotes the average collision rate. This means that about three collisions per particle are sufficient to make the system thermalize. This collision rate may be expressed as \cite{Pitaevskii2016} $\gamma=\rho'\,v'\,\sigma'$, where $\rho'$ denotes the mean number density of the gas, $v'$ is the average norm of the velocity of ideal classical bosons at temperature $T$ and $\sigma'$ is the low-energy boson-boson-scattering cross section. This expression for $\gamma$ holds in the case of an isotropic three-dimensional system. In the present optical case, as highlighted in the previous section, the system is characterized by an anisotropic mass tensor. As a result, the above-given formula for $\gamma$ cannot be applied directly to estimate the time for the quantum fluid of light to relax towards thermal equilibrium.
In order to be able to safely use it, one has to make the kinetic contribution to the Hamiltonian \eqref{Eq:Hamiltonian} isotropic with a common mass $m$ in all the $x$, $y$, $\zeta$ directions. To do so, we introduce the mass parameter $m=(m_{\perp}^{2}\,m_{\zeta}^{\vphantom{2}})^{1/3}$ ---that corresponds to the geometric mean of the paraxial-photon effective masses in the transverse $x$, $y$ and longitudinal $\zeta$ directions--- and the rescaled position vector $\mathbf{r}'=[\mathbf{r}_{\perp}'=(m_{\perp}^{\vphantom{\prime}}/m)^{1/2}\,\mathbf{r}_{\perp}^{\vphantom{\prime}},\zeta'=(m_{\zeta}^{\vphantom{\prime}}/m)^{1/2}\,\zeta]$. As an inversed rescaling holds in momentum space, one readily verifies that such a transformation preserves the spatial as well as the phase-space densities. In the isotropic $\mathbf{r}'$ space, we are then allowed to use the estimate $\tau_{\mathrm{th}}\sim3/\gamma=3/(\rho'\,v'\,\sigma')$ for the thermalization time in terms of the mean number density $\rho'=\rho$, the Boltzmann-averaged velocity $v'=\overline{\hbar\,|\mathbf{k}'|/m}=[8\,k_{\mathrm{B}}\,T/(\pi\,m)]^{1/2}$ and the scattering cross section $\sigma'=8\pi\,a^{\prime2}$, where $a'=m\,g'/(4\pi\,\hbar^{2})$ is the $s$-wave scattering length written as a function of the two-body interaction parameter $g'$ in the isotropic $\mathbf{r}'$ space \cite{Pitaevskii2016}. As our coordinate change preserves both the spatial and the phase-space densities, it is immediate to check that $g'$ is equal to the original photon-photon coupling constant $g$ in the anisotropic $\mathbf{r}$ space, $g'=g$.
Combining the results of the previous paragraph, one eventually gets an explicit formula for the thermalization time $\tau_{\mathrm{th}}$,
\begin{equation}
\label{Eq:ThermalizationTime}
\tau_{\mathrm{th}}\sim3\,\bigg\{\rho\,\bigg[\frac{8\,k_{\mathrm{B}}\,T}{\pi\,(m_{\perp}^{2}\,m_{\zeta}^{\vphantom{2}})^{1/3}}\bigg]^{1/2}\,8\pi\,\bigg[\frac{(m_{\perp}^{2}\,m_{\zeta}^{\vphantom{2}})^{1/3}\,g}{4\pi\,\hbar^{2}}\bigg]^{2}\bigg\}^{-1},
\end{equation}
that corresponds to the usual expression of the thermalization time of a three-dimensional weakly interacting atomic Bose gas, with a mass $(m_{\perp}^{2}\,m_{\zeta}^{\vphantom{2}})^{1/3}$ given by the geometric average of the masses in the transverse $x$, $y$ and longitudinal $\zeta$ directions.
\section{Temperature and chemical potential at thermal equilibrium}
As the kinetic-energy density $E_{\mathrm{kin}}=\int\mathrm{d}^{3}k/(2\pi)^{3}\,N_{\mathbf{k}}\,E_{\mathbf{k}}$ and the photon number density $\rho=\int\mathrm{d}^{3}k/(2\pi)^{3}\,N_{\mathbf{k}}$ are quantities conserved during the evolution of the quantum fluid of light described by eq.~\eqref{Eq:BoltzmannEquation}, the temperature $T$ and the chemical potential $\mu$ characterizing the thermal-equilibrium, at $\tau\gtrsim\tau_{\mathrm{th}}$, BE distribution \eqref{Eq:BoseEinsteinDistribution} may be fixed by the initial, at $\tau=0$, values of $E_{\mathrm{kin}}$ and $\rho$: $E_{\mathrm{kin}}(\tau\gtrsim\tau_{\mathrm{th}})=E_{\mathrm{kin}}(\tau=0)$ and $\rho(\tau\gtrsim\tau_{\mathrm{th}})=\rho(\tau=0)$, where the left-hand sides depend on $T$ and $\mu$ and the right-hand sides are functions of the parameters of the incoming electromagnetic field.
In the final equilibrium state ($\tau\gtrsim\tau_{\mathrm{th}}$, \textit{i.e.}, $z\gtrsim z_{\mathrm{th}}$), using eq.~\eqref{Eq:BoseEinsteinDistribution}, one readily gets $E_{\mathrm{kin}}=\frac{3}{2}\,k_{\mathrm{B}}\,T\,g_{5/2}(f)/(\lambda_{\perp}^{2}\,\lambda_{\zeta}^{\vphantom{2}})$ and $\rho=g_{3/2}(f)/(\lambda_{\perp}^{2}\,\lambda_{\zeta}^{\vphantom{2}})$, where $f=\exp[\mu/(k_{\mathrm{B}}\,T)]$ is the fugacity and $g_{\nu}(f)=\Gamma^{-1}(\nu)\int_{0}^{\infty}\mathrm{d}u~u^{\nu-1}/(f^{-1}\,\mathrm{e}^{u}-1)$ refers to the Bose integral, with $\Gamma(\nu)$ the Euler gamma function. These equations are similar to the well-known results of the ideal Bose gas, with the difference that, in the present optical case, there are two different thermal de Broglie wavelengths $\lambda_{\perp,\zeta}=[2\pi\,\hbar^{2}/(m_{\perp,\zeta}\,k_{\mathrm{B}}\,T)]^{1/2}$ due to the anisotropy of the kinetic energy in eq.~\eqref{Eq:Hamiltonian}. By means of the equation for $E_{\mathrm{kin}}$, one finds that $E_{\mathrm{int}}/E_{\mathrm{kin}}=\frac{1}{3}\,g\,\rho/(k_{\mathrm{B}}\,T)\,g_{3/2}(f)/g_{5/2}(f)$, where $E_{\mathrm{int}}=g\,\rho^{2}/2$ is the mean-field interaction-energy density \cite{Pitaevskii2016} of the fluid of light at equilibrium. Thus, as the Bose integrals $g_{3/2}(f)$ and $g_{5/2}(f)$ are of the same order [$1\leqslant g_{3/2}(f)/g_{5/2}(f)\leqslant 1.94(7)$ for $0\leqslant f\leqslant1$], the weak-interaction condition $E_{\mathrm{int}}\ll E_{\mathrm{kin}}$ required for eq.~\eqref{Eq:BoltzmannEquation} to be valid reads $g\,\rho\ll k_{\mathrm{B}}\,T$, which may be reexpressed in terms of the $s$-wave scattering length $a'$ in the isotropic $\mathbf{r}'$ space as $\rho\,a^{\prime3}\ll(\rho\,\lambda_{\perp}^{2}\,\lambda_{\zeta}^{\vphantom{2}})^{-2}$. Note that this constraint directly implies the usual diluteness condition $\rho\,a^{\prime3}\ll1$ when one enters the quantum-degeneracy regime, $\rho\,\lambda_{\perp}^{2}\,\lambda_{\zeta}^{\vphantom{2}}\gg1$.
We assume that the initial fluid of light ($\tau=0$), \textit{i.e.}, the incident beam of light ($z=0$), is characterized by the following Gaussian distribution in real space:
\begin{equation}
\label{Eq:InitialDistribution}
\langle\hat{\Psi}^{\dag}(\mathbf{r},0)\,\hat{\Psi}(0,0)\rangle=\rho_{0}\,\mathrm{e}^{-\mathbf{r}_{\perp}^{2}/(2\,\ell_{\perp}^{2})}\,\mathrm{e}^{-\zeta^{2}/(2\,\ell_{\zeta}^{2})},
\end{equation}
with finite correlation lengths $\ell_{\perp}$ and $\ell_{\zeta}$ in the transverse $\mathbf{r}_{\perp}$ plane and the longitudinal $\zeta$ direction, respectively. In an actual experiment, the input density $\rho_{0}$ of the quantum fluid of light is tuned by varying the intensity $\mathcal{I}=\hbar\,\omega\,\rho_{0}/\beta_{1}$ of the incoming light beam, the transverse correlation length $\ell_{\perp}$ may be tuned by processing the input beam through spatial light modulators \cite{Sun2012} and the longitudinal correlation length $\ell_{\zeta}$ may in principle be varied by modifying the coherence time $\beta_{1}\,\ell_{\zeta}$ of the incident beam. Note that $\ell_{\perp}$ must be larger than the wavelength $2\pi/\beta_{0}$ of the carrier wave to ensure the paraxiality of the beam of light in the medium and $1/(\beta_{1}\,\ell_{\zeta})$ must be smaller than the frequency range within which the quadratic approximation of the dispersion relation of the medium is valid. Fourier transforming eq.~\eqref{Eq:InitialDistribution} yields the expression of the initial occupation number $N_{\mathbf{k}}(\tau=0)=(2\pi)^{3/2}\,\rho_{0}^{\vphantom{2}}\,\ell_{\perp}^{2}\,\ell_{\zeta}^{\vphantom{2}}\,\mathrm{e}^{-\ell_{\perp}^{2}\,\mathbf{k}_{\perp}^{2}/2}\,\mathrm{e}^{-\ell_{\zeta}^{2}\,k_{\zeta}^{2}/2}$ at $\mathbf{k}$. From this, one obtains, at $\tau=0$, $E_{\mathrm{kin}}=[\hbar_{\vphantom{\perp}}^{\vphantom{-}2}\,\ell_{\perp}^{-2}/m_{\perp}^{\vphantom{-2}}+\hbar_{\vphantom{\zeta}}^{\vphantom{-}2}\,\ell_{\zeta}^{-2}/(2\,m_{\zeta}^{\vphantom{-2}})]\,\rho_{0}$ and $\rho=\rho_{0}$.
Making use of the conservation laws $E_{\mathrm{kin}}(\tau\gtrsim\tau_{\mathrm{th}})=E_{\mathrm{kin}}(\tau=0)$ and $\rho(\tau\gtrsim\tau_{\mathrm{th}})=\rho(\tau=0)$, one eventually gets the following 2-by-2 system:
\begin{equation}
\label{Eq:ConservationLaws}
\frac{3}{2}\,k_{\mathrm{B}}\,T\,\frac{g_{5/2}(f)}{g_{3/2}(f)}=\frac{\hbar_{\vphantom{\perp}}^{\vphantom{-}2}\,\ell_{\perp}^{-2}}{m_{\perp}}+\frac{\hbar_{\vphantom{\zeta}}^{\vphantom{-}2}\,\ell_{\zeta}^{-2}}{2\,m_{\zeta}},\quad\frac{g_{3/2}(f)}{\lambda_{\perp}^{2}\,\lambda_{\zeta}^{\vphantom{2}}}=\rho_{0},
\end{equation}
the resolution of which makes it possible to obtain $T$ and $\mu$ in the final thermal-equilibrium state in terms of $\rho_{0}$, $\ell_{\perp}$, $\ell_{\zeta}$, $m_{\perp}$ and $m_{\zeta}$. Introducing the effective temperatures $T_{\perp,\zeta}=2\pi\,\hbar^{2}/(k_{\mathrm{B}}\,m_{\perp,\zeta}^{\vphantom{2}}\,\ell_{\perp,\zeta}^{2})$, the first of eqs.~\eqref{Eq:ConservationLaws} may be rewritten as $6\pi\,g_{5/2}(f)/g_{3/2}(f)\,T=2\,T_{\perp}+T_{\zeta}$, which shows that the transverse and longitudinal modes, initially distributed at different temperatures $T_{\perp}\neq T_{\zeta}$, eventually equilibrate at the same temperature $T$.
\section{Experimental considerations}
Reminding the definition of the spatial coordinate $\zeta$, the third component of the paraxial-photon wavevector $\mathbf{k}$ may be expressed as \cite{Larre2015, Larre2016} $k_{\zeta}=-\beta_{1}\,\Delta\omega$, where $\Delta\omega$ is the detuning from the angular frequency $\omega$ of the pump. As a result, the measurement of the BE distribution \eqref{Eq:BoseEinsteinDistribution} as a function of $\mathbf{k}=(\mathbf{k}_{\perp},-\beta_{1}\,\Delta\omega)$ requires a good angular resolution to isolate the light deflected with a transverse wavevector $\mathbf{k}_{\perp}$ as well as a good spectral resolution to isolate the angular-frequency component of the transmitted light at $\omega\pm\Delta\omega$.
On the other hand, to have access to the large-momentum, Boltzmann, tails of the BE distribution ---and so, in turn, to the whole BE distribution as a function of $\mathbf{k}$--- at the exit face of the nonlinear dielectric where the fluid of light is imaged, some conditions have to be satisfied.
The inverse of the de Broglie wavelengths $\lambda_{\perp}$ and $\lambda_{\zeta}$ being the natural scales of variation of $N_{\mathrm{BE}}(E_{\mathbf{k}},T,\mu)$ as a function of $\mathbf{k}_{\perp}$ and $k_{\zeta}$, a first condition for detecting the whole BE distribution in the transmitted beam of light is that $\lambda_{\perp}$ and $\lambda_{\zeta}$ must verify the constraints satisfied respectively by $\ell_{\perp}$ and $\ell_{\zeta}$ (see the third paragraph of the previous section).
A second, perhaps more challenging, condition concerns the length of the bulk nonlinear medium, which has to be at least of the order of the distance $z_{\mathrm{th}}=\tau_{\mathrm{th}}/\beta_{1}$ necessary for the quantum fluid of light to fully relax towards thermal equilibrium. Making use of the analytical result \eqref{Eq:ThermalizationTime} and of the first of eqs.~\eqref{Eq:ConservationLaws} with the reasonable estimate $g_{5/2}(f)\sim g_{3/2}(f)$ for $0\leqslant f\leqslant1$, one finds that $z_{\mathrm{th}}$ must behave at a given carrier wave at $(\omega,\beta_{0}=n_{0}\,\omega/c)$ as
\begin{equation}
\label{Eq:ThermalizationDistance}
z_{\mathrm{th}}=\frac{K}{|n_{2}|^{2}\,\mathcal{I}}\,\bigg[\frac{|\beta_{2}|}{\ell_{\perp}^{-2}+\beta_{0}^{\vphantom{2}}\,|\beta_{2}^{\vphantom{2}}|/(2\,\beta_{1}^{2})\,\ell_{\zeta}^{-2}}\bigg]^{1/2},
\end{equation}
where $K$ depends on $\hbar$, $c$, $\varepsilon_{0}$, $k_{\mathrm{B}}$ and on $\omega$, $n_{0}$, $\beta_{0}$. As a most important contribution, it is immediate to see that the stronger the Kerr nonlinearity is, the shorter the thermalization distance $z_{\mathrm{th}}$ is. Plugging explicit values into \eqref{Eq:ThermalizationDistance}, we estimate for a light beam of $1.55~\mu\mathrm{m}$ wavelength, $1~\mathrm{W}/\mu\mathrm{m}^{2}$ intensity and initial $\beta_{0}\,\ell_{\perp,\zeta}=10$ coherence lengths propagating in bulk silica [$2\,|n_{2}|/(c\,\varepsilon_{0}\,n_{0})\sim10^{-20}~\mathrm{m}^{2}/\mathrm{W}$] an unreasonably long $z_{\mathrm{th}}\sim10^{13}~\mathrm{m}$, \textit{i.e.}, of the order of the estimated radius of the solar system{\dots}
While an experiment using such standard bulk nonlinear media looks clearly unfeasable, very promising alternatives are offered by resonant media where photons are strongly mixed with matter excitations. In this way, very strong effective photon-photon interactions may be obtained, \textit{e.g.}, for polaritons in bulk semiconducting materials showing narrow exciton lines such as GaAs or ZnSe \cite{Carusotto2013}. This effect can be further reinforced by many orders of magnitude if the chosen material excitation involves spatially wide (even almost micron-sized) Rydberg states, either in optically dressed atomic gases in the so-called Rydberg-EIT regime \cite{Peyronel2012} or in highest-quality solid-state $\mathrm{Cu_{2}O}$ samples \cite{Jang2011}. A further advantage of resonant media is the wide tunability of the optical parameters simply by changing the carrier frequency $\omega$, which is of a great utility to ensure the dynamical stability of the photon fluid.
\section{Discussion of a recent experiment}
In ref.~\cite{Sun2012}, \textsc{C.~Sun} {\etal} reported having experimentally observed the relaxation of a classical, \textit{i.e.}, not quantum, fluid of interacting photons towards a thermal-equilibrium state. A beam of classical monochromatic light, initially prepared in a nonthermal state \textit{via} a suitable tayloring of the incident phase profile, was made to propagate in a photorefractive crystal whose optical nonlinearity was strong enough to make the transverse angular distribution of the beam of light fastly evolve towards a RJ-type, \textit{i.e.}, classical, thermal law. For small enough initial kinetic energies, a marked peak around $\mathbf{k}_{\perp}=0$ was observed in the transverse-momentum-$\mathbf{k}_{\perp}$ distribution, which was interpreted as a signature of the occurrence of a kinetic condensation of classical waves.
In order to fully understand the analogies and the differences with our quantum study, we can start by noting that a key conceptual assumption of the experiment \cite{Sun2012} is that the light beam remains perfectly monochromatic all along its propagation across the nonlinear crystal. Under a mean-field approximation and provided no spontaneous temporal modulations such as self-pulsing \cite{Larre2015} occur, monochromaticity at all distances is a trivial consequence of the classical GP form of the nonlinear Schr\"odinger field equation corresponding to the quantum Hamiltonian \eqref{Eq:Hamiltonian}.
On the other hand, monochromaticity corresponds within the framework of our quantum theory to having at all propagation times $\tau$ a factorized momentum distribution $N_{\mathbf{k}}(\tau)=N_{\mathbf{k}_{\perp}}(\tau)\,N_{k_{\zeta}}$, where the transverse-momentum distribution $N_{\mathbf{k}_{\perp}}(\tau)$ evolves with $\tau$ while the longitudinal one $N_{k_{\zeta}}$ remains constant and proportional to the Dirac function $\delta(k_{\zeta})$ at all $\tau$'s. Monochromaticity at all $\tau$'s then requires that no scattering process can change the $k_{\zeta}$'s of the colliding paraxial photons.
Most remarkably, the specific form of the optical nonlinearity of the photorefractive crystal used in the experiment \cite{Sun2012} automatically serves this purpose, as its slow response involves the time-$t$ average of the optical intensity and ---in many-body terms--- corresponds to infinite-range interactions along the $\zeta$ axis. As a result, all processes that would generate frequencies different from the incident one are suppressed. Keeping in mind that the population is sharply peaked on the only occupied states with $k_{\zeta}=0$, it is then straightforward to see that the kinetics will eventually relax to the classical RJ distribution $N_{\mathrm{RJ}}(E_{\mathbf{k}},T,\mu)=k_{\mathrm{B}}\,T/(E_{\mathbf{k}}-\mu)$ rather than to the BE one \eqref{Eq:BoseEinsteinDistribution}: because of the $\delta$-shaped factor $N_{k_{\zeta}}$ in the $N_{\mathbf{k}}$'s, all the quantum ``$+\,1$'' terms in the Boltzmann equation \eqref{Eq:BoltzmannEquation} are in fact irrelevant, so that the quantum kinetics reduces to a classical one.
The situation is of course completely different if a local and instantaneous nonlinearity is used in an experiment. Within our theory \cite{Larre2015}, this corresponds to a local interaction in the three-dimensional $x$, $y$, $\zeta$ space. As a result, wave-mixing processes can mix all the three components of the momentum, therefore allowing for a full three-dimensional thermalization of the photon gas in both its transverse-momentum-$\mathbf{k}_{\perp}$ distribution and its physical-frequency-$\Delta\omega$ distribution, where $\Delta\omega=-k_{\zeta}/\beta_{1}$ is measured from the carrier wave at $\omega$. Given the quantum nature of our model, the final result of this thermalization process will be a BE distribution of the form \eqref{Eq:BoseEinsteinDistribution}, which automatically solves all the ultraviolet black-body catastrophes that infest classical theories such as the one used in ref.~\cite{Sun2012}. As a final point, it is worth highlighting that thermalization to a quantum distribution is based on the quantum ``$+\,1$'' terms in the Boltzmann equation and thus does not benefit from the large Bose stimulation factor involved in the thermalization of classical waves. Together with the typically weaker Kerr optical nonlinearity of fast media, this explains why our prediction for $z_{\mathrm{th}}$ is dramatically longer than the experimental one of ref.~\cite{Sun2012}.
\section{Evaporative cooling and BE condensation of a beam of light}
An interesting consequence of the above-investigated thermalization process appears when the quantum fluid of light enters the BE-condensed phase. From the theory of the ideal Bose gas \cite{Pitaevskii2016}, the critical line for BE condensation in the $(\rho_{0},T)$ plane may be obtained by imposing $f=1$ in the second of the thermal-state equations \eqref{Eq:ConservationLaws}, which yields the usual formula for the BE-condensation critical temperature,
\begin{equation}
\label{Eq:BECTemperature}
T_{\mathrm{c}}=\frac{2\pi\,\hbar^{2}}{k_{\mathrm{B}}\,(m_{\perp}^{2}\,m_{\zeta}^{\vphantom{2}})^{1/3}}\,\bigg[\frac{\rho_{0}}{\zeta(3/2)}\bigg]^{2/3},
\end{equation}
in terms of the Riemann zeta function at $3/2$, $\zeta(3/2)=g_{3/2}(1)=2.61(2)$, and the before-introduced geometric mean $m=(m_{\perp}^{2}\,m_{\zeta}^{\vphantom{2}})^{1/3}$ of the paraxial-photon effective masses. To realize a BE condensate of light in a bulk geometry, the experimentalist has to choose the rescaled intensity $\rho_{0}$ and the correlation lengths $\ell_{\perp}$ and $\ell_{\zeta}$ of the incident beam in such a way that the temperature $T$ in the thermal state, solution of eqs.~\eqref{Eq:ConservationLaws}, is smaller than $T_{\mathrm{c}}$ given by eq.~\eqref{Eq:BECTemperature}.
Following the theoretical and experimental investigations \cite{Mandonnet2000, Castin2000, Lahaye2004, Lahaye2005} of the evaporative cooling of an atomic beam propagating in a magnetic trap, a promising way to facilitate BE condensation in the quantum fluid of light consists in progressively making the photon beam evaporate in the transverse $\mathbf{r}_{\perp}=(x,y)$ directions.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{EvaporativeCooling.pdf}
\caption{(Color online) Sketch of an optical platform allowing the evaporative cooling of a quantum fluid of light (red) to temperatures below the critical temperature for BE condensation. The core (light gray) of radius $R(\tau)$ and the cladding (dark gray) of the waveguide are designed so that the photons are trapped in an effective harmonic potential (thick gray curves) whose maximum amplitude $\propto R^{2}(\tau)$ diminishes as the propagation time $\tau$ increases. This removes the high-energy photons from the fluid of light, which then cools down.}
\label{Fig:EvaporativeCooling}
\end{figure}
This can be obtained by introducing a time-dependent trapping potential $U(\mathbf{r}_{\perp},\tau)\neq0$ into the Hamiltonian \eqref{Eq:Hamiltonian}, for instance of truncated harmonic form $U(\mathbf{r}_{\perp},\tau)=\frac{1}{2}\,m_{\perp}^{\vphantom{2}}\,\omega_{\perp}^{2}\,\mathbf{r}_{\perp}^{2}$ for $|\mathbf{r}_{\perp}|\leqslant R(\tau)$ and $U(\mathbf{r}_{\perp},\tau)=\frac{1}{2}\,m_{\perp}^{\vphantom{2}}\,\omega_{\perp}^{2}\,R^{2}(\tau)$ for $|\mathbf{r}_{\perp}|>R(\tau)$, where the radius $R(\tau)$ is a decreasing function of the propagation time $\tau$. Based on the relation $n_{1}(\mathbf{r}_{\perp},z)=-c\,\beta_{1}/(\hbar\,\omega)\,U(\mathbf{r}_{\perp},\tau)$ between the spatial profile of the refractive index and the effective potential in eq.~\eqref{Eq:Hamiltonian}, this truncated harmonic trap may be realized by means of a conically tapered multimode optical waveguide, as pictorially sketched in fig.~\ref{Fig:EvaporativeCooling}. The core [$|\mathbf{r}_{\perp}|\leqslant R(\beta_{1}\,z)$] is taken to have an inverse-parabolic refractive-index profile, while the cladding [$|\mathbf{r}_{\perp}|>R(\beta_{1}\,z)$] is homogeneous with a refractive index smoothly connecting the one of the core's edge.
As a result of this tapering, the maximum value of the trapping potential decreases as $\tau$ increases, so that the large-momentum (or large-energy) tails of the photon distribution are progressively removed. At the same time, the remaining photons keep reequilibrating to lower and lower temperatures under the effect of collisions, until the fluid of light eventually crosses the critical temperature for BE condensation.
Upon the $t\longleftrightarrow z$ mapping, BE condensation from an initially thermal photon gas corresponds to the appearance of spontaneous optical coherence when an initially incoherent beam of light is injected into the nonlinear medium: the long-range order of the BE condensate of light reflects into optical coherence extending for macroscopically long times $t$ and distances $x$, $y$. In contrast to trivial angular- and frequency-filtering processes, a key element of our proposal are the collisions between the photons, that allow the fluid of light to reestablish thermal equilibrium at lower and lower temperatures while the most energetic photons keep being removed.
\section{Conclusion}
In this letter, we have investigated the relaxation dynamics of a paraxial, quasimonochromatic beam of quantum light towards thermal equilibrium in a lossless bulk Kerr medium. Following ref.~\cite{Larre2015}, the propagation of the quantum light field has been mapped onto a quantum nonlinear Schr\"odinger evolution of a conservative quantum fluid of many interacting bosons. Correspondingly, in the weak-interaction regime, the evolution of the momentum distribution from an arbitrary nonthermal state towards a thermal state with a BE form can be modeled by the Boltzmann kinetic equation, which offers analytical formulas for the thermalization time and for the final temperature and chemical potential in terms of the parameters of the input beam and of the medium.
In addition to extending the concept of classical-light-wave condensation \cite{Sun2012} to a fully quantum level and solving well-known ultraviolet pathologies of existing classical theories, our results suggest an intriguing long-term application as a novel source of coherent light: taking inspiration from related advances in atom-laser devices \cite{Mandonnet2000, Castin2000, Lahaye2004, Lahaye2005}, we have pointed out a novel in-waveguide evaporative-cooling scheme to obtain spontaneous macroscopic optical coherence from an initially incoherent beam of light. As our proposal does not rely on population-inverted atomic transitions, it holds the promise of being implemented in an arbitrary domain of the electromagnetic spectrum.
\acknowledgments
We are grateful to \textsc{G.~Ferrari}, \textsc{A.~Gambassi}, \textsc{A.~Minguzzi}, \textsc{C.~Miniatura} and \textsc{M.~Richard} for helpful and stimulating discussions. \textsc{AC} acknowledges the kind hospitality of the BEC Center, where this work was partially done. \textsc{P-\'EL} and \textsc{IC} were funded by the EU-FET Proactive Grant AQuS, Project No.~640800, and by the Provincia Autonoma di Trento, partly through the Call ``Grandi Progetti 2012'', Project SiQuro.
|
\section{Introduction and statement of results}
A \emph{local Borcherds} product is a holomorphic function, which, like a Borcherds form has an absolutely convergent infinite product expansion
and an arithmetically defined divisor, called a local Heegner divisor. Here, \lq local\rq\ refers to boundary components of a modular variety.
Such products were first introduced by Bruinier and Freitag, who, in \cite{BrFr} studied the local divisor class groups of generic boundary components for the modular varieties of indefinite orthogonal groups $\group{O}(2,l)$, $l\geq 3$.
Since then, local Borcherds products have appeared in several places in the literature,
for example in \cite{B123}, for the Hilbert modular group, and in \cite{FS},
where they are introduced to study a specific problem in the geometry of Siegel three folds.
The aim of the present paper is to develop a theory similar to that of Bruinier and Freitag
for unitary groups of signature $(1, n+1)$, $n\geq 1$.
Let $\mathbf{k} = \mathbb{Q}(\sqrt{\Disc})$ be an imaginary quadratic number field with discriminant $\Disc$,
which we consider as a subset of $\mathbb{C}$.
Denote by $\mathcal{O}_{\fieldk}$ the ring of integers in $\mathbf{k}$, by $\mathfrak{d}_\fieldk^{-1}$ the inverse different ideal
and by $\delta_{\fieldk}$ the square-root of $\Disc$, with the principal branch of the complex square-root.
Let $V$ be an indefinite hermitian vector space over $\mathbf{k}$ of signature $(1, n+1)$,
equipped with a non-degenerate hermitian form $\hlf{\cdot}{\cdot}$.
Let $L$ be a lattice in $V$, of full rank as an $\mathcal{O}_{\fieldk}$-module,
so that $L\otimes_{\mathcal{O}_{\fieldk}} \mathbf{k} = V$. We assume that $L$ is an even and integral lattice,
hence $\hlf{\lambda}{\lambda} \in \mathbb{Z}$ for all $\lambda \in L$.
In this introductory section only, we additionally assume that $L$ is unimodular over $\mathbb{Z}$,
i.e.\ $L = L' = \left\{ \mu \in V;\; \hlf{\lambda}{\mu}\in\mathfrak{d}_\fieldk^{-1},\, \forall \lambda \in L \right\}$.
We denote by $\group{U}(V)$ the unitary group of $V$ and by $\group{U}(L) \subset \group{U}(V)$
the isometry group of $L$. Subgroups of finite index in $\group{U}(L)$ are called unitary modular groups.
We consider $\group{U}(V)$ as an algebraic group defined over $\mathbb{Q}$. Its set of real points, denoted $\group{U}(V)(\mathbb{R})$, is the unitary group of the complex hermitian space
$V \otimes_\mathbf{k} \mathbb{C}$. A symmetric domain for the operation of this group is given by the quotient
\[
\DomU = \group{U}(V)(\mathbb{R}) / \mathcal{K},
\]
where $\mathcal{K}$ is a maximal compact subgroup of $\group{U}(V)(\mathbb{R})$.
If $\Gamma \subset \group{U}(L)$ is a unitary modular group, we denote by $X_\Gamma$ the modular variety given by the quotient $\Gamma \backslash \DomU$.
Note that $X_\Gamma$ is non-compact.
The boundary points of $\DomU$ correspond one-to-one
to the elements $I$ of the set of rational one-dimensional isotropic subspaces of $V$, denoted $\mathrm{Iso}(V)$.
For every cusp of $X_{\Gamma}$ one can thus introduce a small open neighborhood $U_\epsilon(I)$.
These neighborhoods are then glued to $X_\Gamma$, furnishing a compactification.
We describe this procedure in section \ref{subsec:cmpct} both for the Baily-Borel compactification, in which singularities remain at the cusps,
and for a toroidal compactification, which turns $X_\Gamma$ into a normal complex space without singularities at the cusps.
We will study the Picard groups of such (suitably small) open neighborhoods $U_\epsilon(I)$.
Since the construction we carry out is local in nature, it suffices to examine only one fixed cusp.
For this purpose, we choose a primitive isotropic lattice vector $\ell \in L$. Fixing a vector $\ell' \in L$ with $\hlf{\ell}{\ell'} \neq 0$, denote
by $D$ the definite lattice $L\cap \ell^\perp \cap \ell'^\perp$.
The stabilizer $\operatorname{Stab}_\Gamma(\ell)$ of $\ell$ in $\Gamma$ contains a Heisenberg group, denoted $\Gamma_\ell$.
This group has finite index in the stabilizer. Its elements can be written as pairs $[h,t]$, with $h$ a rational number and $t$ a lattice vector. The set of all such $t$'s constitutes a sub-lattice $D_{\ell, \Gamma} \subseteq D$.
Following \cite{BrFr}, we define the Picard group $\mathrm{Pic}(X_\Gamma, \ell)$ as the direct limit $\varinjlim\mathrm{Pic}(U_{\epsilon}^{reg}(\ell))$,
where $U_{\epsilon}^{reg}(\ell)$ is the regular locus of $U_{\epsilon}(\ell)$ in the Baily-Borel compactification.
Up to torsion, this local Picard group can also be described by the direct limit
$\varinjlim \mathrm{Pic}(\Gamma_\ell \backslash U_{\epsilon}(\ell))$, see p.\ \pageref{sec:heegner} for details.
Thus, if we only want to describe the position of certain special divisors in $\mathrm{Pic}(X_\Gamma, \ell)$ up to torsion,
we can work in $\mathrm{Pic}(\Gamma_\ell \backslash U_\epsilon(\ell))$, with a sufficiently small $\epsilon >0$.
For a lattice vector $\lambda \in L$ of negative norm, i.e.\ $\hlf{\lambda}{\lambda} \in \mathbb{Z}_{< 0}$,
a primitive Heegner divisor $\mathbf{H}(\lambda)$ is defined by the orthogonal complement $\lambda^\perp$ with
respect to $\hlf{\cdot}{\cdot}$ of $\lambda$ in $\DomU$.
If $\ell$ lies in $\lambda^\perp$, we attach a local Heegner divisor to $\lambda$ by setting
$\mathbf{H}_\infty(\lambda) \vcentcolon= \sum_{ \alpha \in \mathfrak{d}_\fieldk^{-1}} \mathbf{H}(\lambda+ \alpha\ell)$.
A Heegner divisor of $\DomU$ is a $\Gamma$-invariant finite linear combination
of primitive Heegner divisors and the pre-image under the canonical projection of a divisor on $X_\Gamma$.
By a \emph{local Heegner divisor}, we mean a finite linear combination of local Heegner divisors of the form $\mathbf{H}_\infty(\lambda)$,
which corresponds to the pre-image of an element of the divisor group
$\mathrm{Div}( \Gamma_\ell \backslash U_\epsilon(\ell))$, see section \ref{subsec:heegner} for details.
We want to describe the position of local Heegner divisors in the local Picard group $\mathrm{Pic}(X_\Gamma, \ell)$ (up to torsion)
through their position in $\mathrm{Pic}(\Gamma_\ell\backslash U_\epsilon(\ell))$.
This is where local Borcherds products come into play:
For a negative-norm lattice vector $\lambda$ we define the local Borcherds product $\Psi_\lambda(z)$
as follows (see section \ref{subsec:localbp}):
\[
\Psi_\lambda(z) \vcentcolon= \prod_{\alpha \in \mathfrak{d}_\fieldk^{-1}}
\left( 1 - e\left(\sigma(\alpha) \hlf{z}{\lambda - \alpha\ell} \right) \right).
\]
Here, $\lambda - \alpha\ell$ runs over finitely many orbits under the operation of $\Gamma_\ell$, and
$\sigma(\mu)$ is a sign introduced to assure absolute convergence.
The product has divisor $\mathbf{H}_\infty(\lambda)$. However, because
of the sign $\sigma(\alpha)$, it is not invariant under $\Gamma_\ell$.
Instead, there is a non-trivial automorphy factor.
This is actually a desirable situation: By calculating the automorphy factor, we are able to determine
the Chern class of $\mathbf{H}_\infty(\lambda)$ in the cohomology group $\Htwo(\Gamma_\ell, \mathbb{Z})$ (see
sections \ref{subsec:localbp} and \ref{subsec:chern}).
It turns out to be given by the image $[c_\lambda]$ of a bilinear form in the cohomology:
\begin{equation}\label{eq:chern_intro}
\begin{gathered}
c_\lambda([h,t], [h', t']) = - \Im\left[ \abs{\delta_{\fieldk}} F_\lambda(t,t') \right] \quad(\text{for}\quad [h,t],[h',t'] \in \Gamma_\ell), \\
\text{where} \quad F_\lambda(x,y) \vcentcolon= \hlf{x}{\lambda}\hlf{y}{\lambda} + \hlf{\lambda}{x}\hlf{y}{\lambda} \quad(x,y\in D\otimes_{\mathcal{O}_{\fieldk}}\mathbb{C}).
\end{gathered}
\end{equation}
Through this, we know the Chern class of every local Heegner divisor
as a finite linear combination, and can thus describe its position in the cohomology.
For this we use results prepared in section \ref{sec:cohom}, from calculations in the group cohomology for $\Gamma_\ell$,
concerning the properties of cocycles in $\Htwo(\Gamma_\ell, \mathbb{Z})$.
We obtain an equivalent condition for the Chern class of a linear combination of Heegner divisors to be a torsion element,
in Lemma \ref{lemma:tors_lcH}. From the proof, we also obtain a further, necessary condition, see Corollary \ref{cor:nec_cond}.
Finally, our main result, Theorem \ref{thm:H_torsCond}, describes exactly
when Heegner divisors are torsion elements in the Picard group $\mathrm{Pic}(\Gamma_\ell\backslash U_\epsilon(\ell))$.
For a unimodular lattice $L$, the theorem can be formulated as follows, for the general version, see Theorem \ref{thm:H_torsCond} on p.\ \pageref{thm:H_torsCond}:
\begin{theorem}\label{thm:main_intro}
A finite linear combination of local Heegner divisors of the form
\[
\mathbf{H} = \frac{1}{2} \sum_{\substack{m \in \mathbb{Z} \\ m<0}} c(m)
\sum_{\substack{\lambda \in D \\ \hlf{\lambda}{\lambda} = m}} \mathbf{H}_\infty(\lambda)
\]
with coefficients $c(m) \in \mathbb{Z}$, is a torsion element
in the Picard group $\mathrm{Pic}(\Gamma_\ell\backslash U_\epsilon(\ell))$, if and only if
the equation
\[
\sum_{\substack{m \in \mathbb{Z} \\ m<0}} c(m) \sum_{\substack{\lambda \in D \\ \Qf{\lambda} = m }}
\left[ F_\lambda(t,t') - \abs{\Disc}\frac{\hlf{\lambda}{\lambda}}{n} \hlf{t'}{t} \right] = 0
\]
holds for all $t,t' \in D_{\ell, \Gamma}$. Here, $F_\lambda$ is the bilinear form from \eqref{eq:chern_intro} above.
\end{theorem}
As an application of the theorem, we study the obstructions for a (local) Heegner divisor to be a torsion element.
It turns out that they are given by certain spaces of cusp forms spanned by theta-series.
This result is Theorem \ref{thm:Obst} in section \ref{subsec:mf_lgkz}, which here can stated as follows,
with $G=\group{SL}_2(\mathbb{Z})$ and $k=n+2$:
\begin{theorem}\label{thm:obst_intro}
A finite linear combination of Heegner divisors
\[
\mathbf{H} = \frac{1}{2} \sum_{\substack{m \in \mathbb{Z} \\ m<0}} c(m)
\sum_{\substack{\lambda \in D \\ \hlf{\lambda}{\lambda} = m}} \mathbf{H}_\infty(\lambda)
\]
is a torsion element in $\mathrm{Pic}(\Gamma_\ell\backslash U_{\epsilon}(\ell))$ if and only if
\[
\sum_{\substack{m \in \mathbb{Z} \\ m<0}} c(m) a(-m) = 0
\]
for all cusp forms $f \in \mathcal{S}_{k}^\Theta(G)$ with Fourier
coefficients $a(m)$. Here, $\mathcal{S}_{k}^\Theta(G) \subset \mathcal{S}_{k}(G)$
denotes a space of cusp forms spanned by certain (positive-definite) theta-series, see p.\ \pageref{thm:Obst} for the precise definition.
\end{theorem}
Theorem \ref{thm:obst_intro} can be seen a local analog to the global obstruction result showed by the author in \citep[][Section 5]{Hof14},
which in turn is a unitary group version of the obstruction theory developed by Borcherds using Serre-duality \citep[see][Theorem 3.1]{Bo98}.
We discuss the relationship between the local and the global obstruction theories in section \ref{subsec:localglobal},
and also how the two theorems relate to the quite similar results obtained by Bruinier and Freitag in the setting of orthogonal groups
\citep[see][Proposition 5.2, Theorem 5.4]{BrFr}. Our results are also to some extent related to the results of Bruinier, Howard and Yang \cite{BHY15}
and to recent work of Funke and Millson.
\textit{The paper is structured as follows:}
In the first section, we present the set-up and notation used throughout.
We introduce a Siegel domain model of the symmetric domain, with the
fixed isotropic lattice vector $\ell$ corresponding to the cusp at infinity.
We then describe the stabilizer of this cusp and define the Heisenberg group $\Gamma_\ell$.
Also, we sketch the construction of the compactification used for $X_\Gamma$.
In section \ref{sec:cohom}, we study the cohomology of the Heisenberg group $\Gamma_\ell$ and
derive criteria describing when certain two-cocyles obtained from bilinear forms are torsion elements
in the cohomology group $\Htwo(\Gamma_\ell, \mathbb{Z})$. The following section \ref{sec:heegner} is the main part of the paper:
Here, we study Heegner divisors, we introduce the local Borcherds products and we determine their Chern classes.
Using the results established in the second section, we get an equivalent condition for a linear combination of Heegner divisors to be a torsion element in the cohomology, Lemma \ref{lemma:tors_lcH} on p.\ \pageref{lemma:tors_lcH}. A further, necessary condition follows from the proof, see Corollary \ref{cor:nec_cond}.
Finally, as our main result, we derive Theorem \ref{thm:H_torsCond}, part of which follows from the Lemma, while the converse is proved constructively.
The last section closes with the application to modular forms:
In Theorem \ref{thm:Obst} we find that cusp forms arising from certain theta-series constitute the obstructions for a local Heegner divisor to be a torsion element in the Picard group.
\section{Hermitian lattices and symmetric domains}
\subsection{Hermitian spaces and lattices} \label{subsec:ltcs}
Let $\mathbf{k} = \mathbb{Q}(\sqrt{\Disc})$ be an imaginary quadratic number field of discriminant $\Disc$, with $\Disc$ a square-free negative integer.
Let $\mathcal{O}_{\fieldk} \subset \mathbf{k}$ be the ring of integers in $\mathbf{k}$. Denote by $\mathfrak{d}_\fieldk$ the different
ideal and by $\mathfrak{d}_\fieldk^{-1}$ the inverse different ideal.
We shall consider $\mathbf{k}$ as a subset of the complex numbers $\mathbb{C}$ and denote
by $\delta_{\fieldk}$ the square-root of the discriminant, with the usual choice of the complex square-root.
Then, $\mathfrak{d}_\fieldk$ is given by $\delta_{\fieldk} \mathcal{O}_{\fieldk}$ and $\mathfrak{d}_\fieldk^{-1}$ by $\delta_{\fieldk}^{-1}\mathcal{O}_{\fieldk}$.
Let $V = V_\fieldk$ be an indefinite hermitian space over $\mathbf{k}$ of signature $(1,n+1)$,
endowed with a non-degenerate hermitian form denoted $\hlf{\cdot}{\cdot}$, linear in the left and conjugate linear in
the right argument.
A complex hermitian space $V_\mathbb{C} = V\otimes_\mathbf{k}\mathbb{C}$ is obtained by extension of
scalars.
We denote by $V_\mathbb{Q}$ the $\mathbb{Q}$-vector space underlying $V$, which bears the structure of a quadratic
space of signature $(2,2n+2)$ with the quadratic form $\Qf{\cdot}$ defined by $\Qf{x} \vcentcolon
= \hlf{x}{x}$. Similarly, the real quadratic space underlying $V_\mathbb{C}$ is denoted $V_\mathbb{R}$. We have $V_\mathbb{R} = V_\mathbb{Q} \otimes_\mathbb{Q} \mathbb{R}$.
Let $L$ be a lattice in $V$, with $L\otimes_{\mathcal{O}_{\fieldk}}\mathbf{k} = V$.
We denote by $L'$ the $\mathbb{Z}$-dual of $L$, defined as the set
\[
L' = \left\{ x\in V;\; \hlf{x}{y}\in \mathfrak{d}_\fieldk^{-1} \quad\text{for all $y \in L$}
\right\} = \left\{ x\in V;\; \tr_{\mathbf{k}/\mathbb{Q}} \hlf{x}{y}\in \mathbb{Z} \quad\text{for all $y \in
L$}\right\}.
\]
Naturally, $L'$ is a lattice in $V$, too.
If $L\subseteq L'$, the lattice $L$ is called integral. If further for all $x\in L$, $\hlf{x}{x} \in \mathbb{Z}$, then $L$ is called even. Finally, $L$ is unimodular, if $L' = L$.
The quotient $L'/L$ is referred to as the discriminant group of $L$.
More generally in the context of this paper,
by a hermitian lattice we mean a discrete subgroup $M$ of $V$,
for which the ring of multipliers $\order{M}$ is an order in $k$.
(A multiplier of $M$ is a complex number $\alpha$ with $\alpha M \subset M$.)
Most lattices will occur here as sub-lattices of a fixed lattice $L$,
with $L$ as above, of full rank, hermitian and even.
Denote by $\group{U}(V)$ the unitary group of $V$, and by $\group{SU}(V)$ the special unitary group. The
isometry group of a lattice $L$ in $\mathrm{U}(V)$ is denoted $\mathrm{U}(L)$, similarly for $\mathrm{SU}(L)$. The \emph{discriminant kernel} $\Gamma_L$ is the subgroup of finite index in $\group{SU}(L)$ which acts trivially on the discriminant group of $L$.
We refer to subgroups of finite index in $\Gamma_L$ as unitary modular groups.
In the following, $\Gamma$ will always denote a unitary modular group.
\subsection{A symmetric domain} \label{subsec:symm_domain}
Viewing $\group{U}(V)$ as an algebraic group, its set of real points, denoted $\group{U}(V)(\mathbb{R})$, is the
unitary group of $V_\mathbb{C}$.
A symmetric domain for the action of $\group{U}(V)(\mathbb{R})$ on $V_\mathbb{C}$ is given by the quotient $\DomU = \group{U}(V)(\mathbb{R}) / \mathcal{K}$ with a maximal compact subgroup $\mathcal{K}$.
Denote by $\mathbb{P}V_\mathbb{C}$ the projective space of $V_\mathbb{C}$. A projective model for $\DomU$ is
given by the \emph{positive cone}
\[
\coneU = \left\{ [v]\in \mathbb{P}V_\mathbb{C}\,;\, \hlf{v}{v} > 0 \right\}.
\]
We briefly review the construction of an affine model.
Denote by $\mathrm{Iso}(V)$ the set of one-dimensional isotropic subspaces of $V_\fieldk$. Its elements
are in one-to-one correspondence with the rational boundary components of the symmetric domain.
In particular, we fix an element $ I \in \mathrm{Iso}(V)$ by choosing a primitive isotropic
lattice vector $\ell \in L$ and setting $I = \mathbf{k} \ell$. Further, we choose a primitive vector
$\ell' \in L'$ such that $\hlf{\ell}{\ell'} \neq 0$. We shall assume that $\ell'$ is isotropic,
too. Note that this is a non-trivial assumption about the hermitian lattice $L$ and its
dual.
For $a \in V$, we denote by $a^\perp$ the orthogonal complement with respect to $\hlf{\cdot}{\cdot}$. We set
$D\vcentcolon = L\cap \ell^\perp \cap \ell'^\perp$. Equipped with the restriction of $\hlf{\cdot}{\cdot}$,
$D$ is a definite hermitian lattice of signature $(0,n)$.
Denote by $W = W_\mathbf{k}$ the subspace $D\otimes_{\mathcal{O}_{\fieldk}}\mathbf{k}$, and let $W_\mathbb{C} = W \otimes_\mathbf{k}
\mathbb{C}$.
Now, an affine model for $\DomU$, called the Siegel domain model,
is given by the following generalized upper-half-plane:
\[
\mathcal{H}_{\ell, \ell'} = \left\{ (\tau, \sigma) \in \mathbb{C} \times W_\mathbb{C}\,;\,
2\Im(\tau)\abs{\delta_{\fieldk}}\abs{\hlf{\ell}{\ell'}}^2 >
- \hlf{\sigma}{\sigma} \right\}.
\]
For $(\tau, \sigma)\in\mathcal{H}_{\ell, \ell'}$, we set
\[
z = z(\tau, \sigma) \vcentcolon = \ell' - \delta_{\fieldk} \tau\hlf{\ell'}{\ell}\ell + \sigma.
\]
Clearly, under the canonical projection $\projV: V_\mathbb{C} \rightarrow \mathbb{P}V_\mathbb{C}$, we have
$\projV(z) \in \coneU$ for all $(\tau, \sigma)\in\mathcal{H}_{\ell, \ell'}$. Conversely, every $[v] \in
\coneU$ contains a representative of the form $z(\tau, \sigma)$ for some pair $(\tau, \sigma)
\in \mathcal{H}_{\ell, \ell'}$. Usually, in the following, since $\ell$ and $\ell'$ are fixed, we shall
simply write $\mathcal{H} = \mathcal{H}_{\ell, \ell'}$.
The isotropic line $I_\mathbb{C} = I\otimes_\mathbf{k}\mathbb{C} = [\ell]$ corresponds to the cusp at infinity of
$\mathcal{H}$.
\subsection{Stabilizer of the cusp}
Next, we will describe the stabilizer in $\Gamma$ of the cusp $[\ell]$.
Consider the following transformations corresponding to elements of $\group{SU}(V)$:
\begin{align}
\label{def:translationsU} [h,0]:\;& v \; \mapsto \; v - \hlf{v}{\ell}\delta_{\fieldk}
h \ell & \quad\text{for $h\in\mathbb{Q}$}, \\
\label{def:eichlerU} [0,t]:\;& v \; \mapsto \; v + \hlf{v}{\ell}t -
\hlf{v}{t}\ell
- \frac{1}{2}\hlf{v}{\ell}\hlf{t}{t}\ell &
\quad\text{for $t\in W$}.
\end{align}
Clearly, these transformations stabilize the isotropic subspace $\mathbf{k}\ell$.
Their action on $\mathcal{H}$ is given as follows:
\[
[h,0]: (\tau,\sigma) \mapsto (\tau + h,\sigma), \quad
[0,t]: (\tau,\sigma) \mapsto \left( \tau +
\frac{\hlf{\sigma}{t}}{\delta_{\fieldk}\hlf{\ell'}{\ell}}
+\frac12 \frac{\hlf{t}{t}}{\delta_{\fieldk}}, \sigma + \hlf{\ell'}{\ell}t\right).
\]
The \emph{Heisenberg group} attached to $\ell$, denoted $\HeisU{\ell}$,
is the set of pairs $[h,t]$ with group law given by
\begin{equation}\label{eq:Hgrp_law}
[h,t]\circ [h',t'] = \bigl[h + h' +
\frac{\Im\hlf{t'}{t}}{\abs{\delta_{\fieldk}}}, t + t'\bigr].
\end{equation}
Here, we follow the convention that
$\left([h,t]\circ[h',t']\right)v = [h,t]\left( [h',t']\, v \right)$ for $v\in V_\mathbf{k}$.
The center of the Heisenberg group consists of transformations of type \eqref{def:translationsU}.
We denote by $\Gamma_\ell$ the subgroup of $\Gamma$ given by the intersection $\Gamma\cap \HeisU{\ell}$,
its center we denote by $\Gamma_{\ell, T}$.
The full stabilizer of the cusp in $\Gamma$ is given by the semi-direct product
\[
\Gamma_\ell \ltimes (\group{U}(W) \cap \Gamma)) = \operatorname{Stab}_\Gamma(\ell).
\]
Note that $\Gamma_\ell$ has finite index in the stabilizer.
The elements of $\Gamma_\ell$ can be described as follows (this is well-known):
\begin{rmk}\label{rmk:ParabN}
Suppose $\Gamma$ is a unitary modular group and let $\Gamma_\ell = \Gamma \cap
\HeisU{\ell}$. Then there exist a positive rational number $N_{\ell,\Gamma}$ and a lattice $D_{\ell, \Gamma}$ of finite
index in $D$, such that $[h,t] \in \Gamma_\ell$ for all $h \in N_{\ell,\Gamma}\mathbb{Z}$, $t\in D_{\ell, \Gamma}$, and that
$ \abs{\delta_{\fieldk}}^{-1}\Im\hlf{t'}{t} \in N_{\ell,\Gamma}\mathbb{Z}$ for all $t,t'\in D_{\ell, \Gamma}$.
\end{rmk}
\subsection{Boundary components}\label{subsec:cmpct}
The modular variety $X_\Gamma$ is given by the quotient
\[
\Gamma\backslash \DomU \simeq \Gamma \backslash \group{U}(V)(\mathbb{R}) / \mathcal{K}
\simeq \Gamma \backslash \mathcal{H}.
\]
Note that $X_\Gamma$ is non-compact. The usual Baily-Borel compactification $X_{\Gamma, {BB}}^*$ is obtained by
introducing a topology and a complex structure on the quotient
\[
\Gamma \backslash \left( \mathcal{H} \cup \left\{I_\mathbb{R};\; I \in \mathrm{Iso}(V)\right\} \right).
\]
We sketch this for the cusp at infinity, defined by $[\ell]$:
The following sets constitute a system of neighborhoods of the cusp
\begin{equation} \label{eq:defUeps}
U_\epsilon(\ell) = \left\{ [z]\in \coneU ;\, \frac{ \hlf{z}{z} }{ \abs{\hlf{z}{\ell}}^2 }
\abs{\hlf{\ell'}{\ell}}^2> \frac1{\epsilon} \right\} \quad\left( \epsilon>0 \right).
\end{equation}
A subset $V$ of $\coneU \cup \{[\ell]\}$ is called open if $V \cap \coneU$ is open in
the usual sense and further if $[\ell] \in V$ implies $U_\epsilon(\ell) \subset V$ for some $\epsilon>0$.
Through the quotient topology, this construction yields a topology on $\Gamma\backslash \left(
\coneU \cup \{ [\ell]\} \right)$. The complex structure is defined
though the pullback under
the canonical projection $\coneU \cap \{ I_\mathbb{R};\; I \in \mathrm{Iso}(V)\} \rightarrow X_{\Gamma,
{BB}}^*$, locally for each cusp, see \cite{Hof11} for details.
This way, one gets the structure of a normal complex space on $X_{\Gamma, {BB}}^*$. In general,
however, there are still singularities at the boundary points.
This difficulty can be avoided by using a toroidal compactification, instead. We recall the
construction briefly; see \citep[][Chapter 1.1.5]{Hof11} and, in particular \citep[][Section 4.3]{BHY15} for more
details.
In the following, identify the sets $U_\epsilon(\ell) \subset \coneU$ with the
corresponding sets of representatives in $\mathcal{H}_{\ell, \ell'}$.
Clearly, the Heisenberg group $\Gamma_\ell$ operates on
$U_\epsilon(\ell)$. For sufficiently small $\epsilon$, there is an open immersion
\[
\Gamma_\ell \backslash U_\epsilon(\ell) \rightarrow X_\Gamma.
\]
Recall that for the center $C(\Gamma_ \ell) = \Gamma_{\ell, T}$, we have $\Gamma_{\ell, T} \simeq
\mathbb{Z}N_{\ell,\Gamma}$. Set $q_\ell \vcentcolon = \exp( 2\pi i \tau/ N_{\ell,\Gamma} )$.
The quotient $\Gamma_{\ell, T}\backslash U_\epsilon(\ell)$
can now be viewed as bundle of punctured disks over $W_\mathbb{C}$:
\[
V_\epsilon(\ell) \vcentcolon= \Gamma_{\ell, T}\backslash U_\epsilon(\ell) \simeq \left\{ (q_\ell,
\sigma);\;
0< \abs{q_\ell} < \exp\left(\frac{\pi \hlf{\sigma}{\sigma} + \epsilon^{-1}}{\abs{\delta_{\fieldk}}^2
\abs{\hlf{\ell'}{\ell}}^2} \right) \right\}.
\]
Adding the center to each disk, we get the disk bundle
\[
\widetilde{V_\epsilon}(\ell) \vcentcolon= \left\{ (q_\ell, \sigma); \;
\abs{q_\ell} < \exp\left(\frac{\pi \hlf{\sigma}{\sigma} + \epsilon^{-1}}{\abs{\delta_{\fieldk}}^2
\abs{\hlf{\ell'}{\ell}}^2} \right) \right\}.
\]
The action of $\Gamma_\ell$ is well-defined at each center, leaving the divisor $q=0$ fixed.
Also, if $\Gamma$ is sufficiently small, the operation is free, hence we get an open
immersion
\begin{equation} \label{eq:immersionglue}
\Gamma_\ell \backslash U_\epsilon(\ell) \rightarrow \left( \Gamma_\ell / \Gamma_{\ell, T}
\right) \backslash \widetilde{V_\epsilon}(\ell),
\end{equation}
by which the right hand side can be glued to $X_\Gamma$, yielding a partial
compactification.
For a point $(0, \sigma_0) \in \widetilde{V_\epsilon}(\ell)$, we define a system of open sets
\[
B_\delta(0, \sigma_0) = \left\{ (q_\ell, \sigma)\in\widetilde{V_\epsilon}(\ell) \,;\,
\hlf{\sigma-\sigma_0}{\sigma - \sigma_0} < \delta, \abs{q_\ell} < \delta \right\}\qquad(\delta >
0).
\]
Under the immersion \eqref{eq:immersionglue} the images of these sets form a system of open
neighborhoods for the boundary point at $(0, \sigma_0)$.
Repeating this construction and the gluing procedure for every $I \in \Gamma \backslash \mathrm{Iso}(V)$ yields a
compactification of $X_\Gamma$, which we denote $X_{\Gamma, tor}^*$.
\section{The local cohomology group} \label{sec:cohom}
In the following, let $\Gamma$ be a unitary modular group and let $\Gamma_\ell \subset \Gamma$ be a
Heisenberg group of the form $\Gamma_\ell = N_{\ell,\Gamma}\mathbb{Z} \rtimes D_{\ell, \Gamma}$ with
$N_{\ell,\Gamma}\in\mathbb{Q}_{>0}$ and $D_{\ell, \Gamma}\subseteq D$ as introduced in Remark \ref{rmk:ParabN}.
We are interested in the cohomology of $\Gamma_\ell$, more
specifically the second cohomology group $\Htwo(\Gamma_\ell, \mathbb{Z})$.
As usual, if $G$ is a group acting on an abelian group $A$, the n-th cohomology group is defined as
the quotient
\[
\mathrm{H}^n(G, A) = \frac{\operatorname{ker}(\mathrm{C}^n(G, A)\xrightarrow{\partial}
\mathrm{C}^{n+1}(G,A))}{\operatorname{im}(\mathrm{C}^{n-1}(G,A)\xrightarrow{\partial}
\mathrm{C}^{n}(G,A))},
\]
wherein $\mathrm{C}^{n}$ is the set of $n$-cocycles, consisting of all functions $f: G^n\rightarrow A$, and $\partial$ is the coboundary operator. In the present setting, $G=\Gamma_\ell$, $A=\mathbb{Z}$ and the
action of $G$ is trivial.
Let $U_\epsilon(\ell)$ be a neighborhood of the cusp of infinity, as defined in
\eqref{eq:defUeps} above, with $\epsilon$ sufficiently small, so that the map in \eqref{eq:immersionglue}
is indeed an open immersion.
Further, denote by $\mathcal{O}_\epsilon = \mathcal{O}_\epsilon(U_\epsilon(\ell))$ the sheaf of holomorphic functions on $U_\epsilon(\ell)$ and by $\mathcal{O}_\epsilon^* = \mathcal{O}_\epsilon(U_\epsilon(\ell))^*$ the sheaf of invertible holomophic functions.
The action of $\Gamma_\ell$ on $U_\epsilon(\ell)$ naturally induces an action on $\mathcal{O}_\epsilon$ and $\mathcal{O}_\epsilon^*$.
The exact sequence
\[
\begin{tikzcd} 0 \arrow[r] & \mathbb{Z} \arrow[r, "\imath"] & \mathcal{O}_\epsilon \arrow[r, "e"]
& \mathcal{O}_\epsilon^* \arrow[r] & 0 \end{tikzcd}
\]
thus induces an exact sequence of cohomology groups:
\begin{equation}\label{eq:seqH2ZtoOps}
\begin{tikzcd} \mathrm{H}^1(\Gamma_\ell, \mathbb{Z}) \arrow[r] & \mathrm{H}^1(\Gamma_\ell, \mathcal{O}_\epsilon) \arrow[r] & \mathrm{H}^1(\Gamma_\ell, \mathcal{O}_\epsilon^*) \arrow[r, "\delta"] & \mathrm{H}^2(\Gamma_\ell, \mathbb{Z}) \arrow[r]
& \mathrm{H}^2(\Gamma_\ell, \mathcal{O}_\epsilon).
\end{tikzcd}
\end{equation}
The Picard group of $\Gamma_\ell \backslash U_{\epsilon}(\ell)$ is given by $\mathrm{H}^1(\Gamma_\ell \backslash U_{\epsilon}(\ell), \mathcal{O}_\epsilon^*)$. Since the open neighborhoods $U_{\epsilon}(\ell)$ are contractible, all analytic line bundles on $U_{\epsilon}(\ell)$ are trivial. Therefore,
\begin{equation}\label{eq:Pic_H1}
\mathrm{Pic}(\Gamma_\ell \backslash U_{\epsilon}(\ell)) = \mathrm{H}^1(\Gamma_\ell, \mathcal{O}_\epsilon^*).
\end{equation}
Further, let $\mathcal{P}_\epsilon$ denote the functions in $\mathcal{O}_\epsilon$ which are periodic for the action of $N_{\ell,\Gamma}\mathbb{Z}$.
As $N_{\ell,\Gamma}\mathbb{Z} = \Gamma_{\ell, T}$ is a normal subgroup with $\Gamma_\ell / N_{\ell,\Gamma}\mathbb{Z} = D_{\ell, \Gamma}$, and since $N_{\ell,\Gamma}\mathbb{Z}\backslashU_{\epsilon}(\ell)$ is contractible, we have
$\mathrm{H}^p(\Gamma_\ell, \mathcal{O}_\epsilon) = \mathrm{H}^p(D_{\ell, \Gamma}, \mathcal{P}_\epsilon)$ $(p=1,2,\dotsc)$.
Thus, from the exact sequences in \eqref{eq:seqH2ZtoOps} and \eqref{eq:Pic_H1}, we get the exact sequence
\begin{equation}\label{eq:KHwithP}
\begin{tikzcd} \frac{\displaystyle\operatorname{Hom}(D_{\ell, \Gamma}, \mathcal{P}_\epsilon)}{\displaystyle\operatorname{Hom}(D_{\ell, \Gamma}, \mathbb{Z})} \arrow[r] &
\mathrm{Pic}( \Gamma_\ell \backslash U_{\epsilon}(\ell)) \arrow[r] &
\mathrm{H}^2(\Gamma_\ell, \mathbb{Z}) \arrow[r]
& \mathrm{H}^2(\Gamma_\ell, \mathcal{O}_\epsilon).
\end{tikzcd}
\end{equation}
Further, since $D_{\ell, \Gamma}$ is a free group, the following sequence is exact:
\[
\begin{tikzcd} 0 \arrow[r] & \operatorname{Hom}(D_{\ell, \Gamma}, \mathbb{Z}) \arrow[r] & \operatorname{Hom} (D_{\ell, \Gamma}, \mathcal{P}_\epsilon) \arrow[r] & \operatorname{Hom}(D_{\ell, \Gamma}, \mathcal{P}_\epsilon^*) \arrow[r] & 0.
\end{tikzcd}
\]
Whence, from \eqref{eq:KHwithP} we find the exact sequence
\begin{equation}\label{eq:seqPictoH2}
\begin{tikzcd}
\operatorname{Hom}(D_{\ell, \Gamma}, \mathcal{P}_\epsilon^*) \arrow[r] &
\mathrm{Pic}( \Gamma_\ell \backslash U_{\epsilon}(\ell)) \arrow[r] &
\mathrm{H}^2(\Gamma_\ell, \mathbb{Z}) \arrow[r]
& \mathrm{H}^2(\Gamma_\ell, \mathcal{O}_\epsilon).
\end{tikzcd}
\end{equation}
Hence, to study $ \mathrm{Pic}( \Gamma_\ell \backslash U_{\epsilon}(\ell))$ we want to examine the structure of $\Htwo(\Gamma_\ell, \mathbb{Z})$.
\subsection{Bilinear forms in the cohomology}
In this subsection, we will examine the image of certain bilinear forms in the cohomology.
All calculations are carried out using the standard inhomogeneous complex of group
cohomology \citep[cf.][Chapter 8]{Shimura}.
\begin{definition}\label{def:BIL}
Consider the set of bilinear forms $B: W_\mathbb{C} \otimes W_\mathbb{C} \longrightarrow \mathbb{R}$,
for which there is either a hermitian form $H$ or a symmetric complex bilinear form $G$ such that
$B = \Im H$ or $B= \Im G$, respectively.
Such forms generate a vector space of real bilinear forms on $W_\mathbb{C}$, which we denote $\mathsf{BIL}$.
Further, let $\BIL_\Z$ denote the set of forms in $\mathsf{BIL}$ which are $\mathbb{Z}$-valued on the lattice $D_{\ell, \Gamma}$.
\end{definition}
To a bilinear form in $\mathsf{BIL}$ we can associate an element of $\Htwo(\Gamma_\ell, \mathcal{O}_\epsilon)$:
Define the two-cocycle in $C^2(\Gamma_\ell, \mathcal{O}_\epsilon)$ by setting
\begin{equation}\label{eq:BiltoH2e}
B([h,t], [h',t']) \vcentcolon = B(t,t')\qquad\left( [h,t], [h',t'] \in \Gamma_\ell \right).
\end{equation}
The class $[B]$ of this cocycle is the image of $B$ in the cohomology.
For $B \in \BIL_\Z$ we also define a two-coycle in $\mathrm{C}^2(\Gamma_\ell,\mathbb{Z})$ and the attached element in $\Htwo(\Gamma_\ell, \mathbb{Z})$.
Thus, composing with the natural map $\Htwo(\Gamma_\ell, \mathbb{Z}) \rightarrow \Htwo(\Gamma_\ell, \mathcal{O}_\epsilon)$ from \eqref{eq:seqH2ZtoOps} we have a sequence
\begin{equation}\label{eq:seqBilZ}
\begin{tikzcd}
{\BIL_\Z} \arrow{r} & \Htwo(\Gamma_\ell, \mathbb{Z}) \arrow{r} & \Htwo(\Gamma_\ell, \mathcal{O}_\epsilon).
\end{tikzcd}
\end{equation}
The composition of the two maps in \eqref{eq:seqBilZ} is just the restriction to $\BIL_\Z$ of the map $\mathsf{BIL} \rightarrow \Htwo(\Gamma_\ell, \mathcal{O}_\epsilon)$ defined by \eqref{eq:BiltoH2e}.
It turns out that the sequence is exact:
\begin{proposition}\label{prop:seqExact}
The image of $\mathsf{BIL}$ in $\Htwo(\Gamma_\ell, \mathcal{O}_\epsilon)$ vanishes.
\end{proposition}
\begin{proof}
In the following, let $B$ denote an element of $\mathsf{BIL}$.
Clearly, it suffices to consider the following two cases: Either $B$ arises from a hermitian form or $B$ arises from a bilinear form.
\begin{enumerate}
\item Let $H: W_\mathbb{C} \times W_\mathbb{C} \longrightarrow \mathbb{C}$ be a hermitian form.
Consider the following one-cocycle in $\mathrm{C}^1(\Gamma_\ell, \mathcal{O}_\epsilon)$:
\begin{gather*}
u( [h,t],z) = \frac1{2i}\left[ \frac{2}{\hlf{\ell'}{\ell}} H(\sigma, t) + H(t,t)\right]. \\
\shortintertext{Its image under the coboundary map it given by}
\begin{aligned}
\partial u([h,t],[h',t'], z) & = [h,t] u( [h', t'] , z) - u([h,t][h',t'], z) + u([h,t]) \\
& = \frac1{2i}\left( 2 H(t,t') - H(t,t') - H(t',t) \right).
\end{aligned}
\end{gather*}
Thus, we see that $B = \Im H$ is indeed trivialized by a cochain. Hence, its image $\Htwo(\Gamma_\ell, \mathcal{O}_\epsilon)$ vanishes
\item Let $G: W_\mathbb{C} \times W_\mathbb{C} \longrightarrow \mathbb{C}$ be a symmetric complex bilinear form.
We consider the following one-cocycle valued in $\mathcal{O}_\epsilon$:
\begin{gather*}
u( [h,t],z) = \frac{i}{2}\left( \frac{1}{\hlf{\ell'}{\ell}} G(\sigma, t) + \frac12 \overline{G(t,t)}\right).\\
\shortintertext{Its image under the coboundary map is given by }
\partial u([h,t],[h',t'], z) = \frac{1}{2i} \left( G(t',t) - \frac12\overline{G(t,t')} - \frac12 \overline{G(t',t)}\right) = \frac{1}{2i} \left( G(t',t) - \overline{G(t',t)}\right).
\end{gather*}
Thus $B = \Im G$ is trivialized by a cochain, and $[B] = 0$ in $\Htwo(\Gamma_\ell, \mathcal{O}_\epsilon)$.
\end{enumerate}
\end{proof}
\begin{rmk} We note that under a map of the type defined in \eqref{eq:BiltoH2e}, the real parts of sesquilinear forms have vanishing image in $\Htwo(\Gamma_\ell, \mathcal{O}_\epsilon)$, too. The proof is quite similar.
\end{rmk}
Now that we know the sequence \eqref{eq:seqBilZ} to be exact, we study the first map $\BIL_\Z \rightarrow \Htwo(\Gamma_\ell, \mathbb{Z})$.
It is far from being injective. The following Lemma and its proof are essentially due to Freitag,
a sketch is contained in \cite{FreitagNote07}.
\begin{lemma}\label{lemma:torsCrit}
The kernel of the map $\BIL_\Z \rightarrow \Htwo(\Gamma_\ell, \mathbb{Z})$ is the cyclic group generated by
the antisymmetric bilinear form
\[
\frac{1}{N_{\ell,\Gamma}}\frac{\Im\hlf{\cdot}{\cdot}}{\abs{\delta_{\fieldk}}}.
\]
In particular, the image of an element $B \in \BIL_\Z$ is a torsion element in $\Htwo(\Gamma_\ell, \mathbb{Z})$ if and if $B$ and
$\abs{\delta_{\fieldk}}^{-1}\Im\hlf{\cdot}{\cdot}$ are linear dependent over $\mathbb{Z}$.
\end{lemma}
\begin{proof} The proof uses a \emph{transgression map}, which we introduce next.
First note that, since the action of $\mathbb{Z}$ is trivial, the map $\BIL_\Z \rightarrow \Htwo(\Gamma_\ell, \mathbb{Z})$ factors over
$\Htwo(D_{\ell, \Gamma}, \mathbb{Z}) = \Htwo((\Gamma_\ell/ N_{\ell,\Gamma}\mathbb{Z}), \mathbb{Z})$. With Proposition \ref{prop:seqExact} we have:
\[
\begin{tikzcd}
\BIL_\Z \arrow{r} & \Htwo(D_{\ell, \Gamma}, \mathbb{Z}) \arrow{r} & \Htwo(\Gamma_\ell, \mathbb{Z}) \arrow{r} & {0.}
\end{tikzcd}
\]
Now, the transgression $\operatorname{tg}$ is defined as the map for which the sequence
\begin{equation} \label{eq:deftrg}
\begin{tikzcd}
{\mathbb{Z}} \arrow{r}{\operatorname{tg}} & { \Htwo(D_{\ell, \Gamma}, \mathbb{Z}) } \arrow{r} &{\Htwo(\Gamma_\ell, \mathbb{Z})} \arrow{r} & 0
\end{tikzcd}
\end{equation}
becomes exact. Thus, the kernel
of the map into $\Htwo(\Gamma_\ell, \mathbb{Z})$ is generated by the image
of the identity map $\mathbf{1}: \mathbb{Z} \rightarrow \mathbb{Z}$ under $\operatorname{tg}$.
The image $\operatorname{tg}(\mathbf{1})$ is represented by a coboundary
$(t,t') \mapsto (\partial u)([\cdot, t], [\cdot,t'])$,
with a one-cochain $u: \Gamma_\ell \rightarrow \mathbb{Z}$, which has to satisfy two conditions:
\begin{enumerate}
\item $u([N_{\ell,\Gamma}\,h,0]) = h$ for all $h \in \mathbb{Z}$ and
\item $(\partial u)([h,t],[h',t'])$ does not depend on $h$ or $h'$.
\end{enumerate}
A suitable $u$ is obtained by setting $u([N_{\ell,\Gamma} h, t]) \vcentcolon = h$. We get
\[
(\partial u)([h,t],[h',t']) = [h,t]\, u\left([h',t']\right) - u([h,t][h',t'])
+ u([h,t]) = -\frac{1}{ N_{\ell,\Gamma}}\frac{\Im\hlf{t'}{t}}{\abs{\delta_{\fieldk}}}.
\]
Hence, $\operatorname{tg}(\mathbf{1})$ is represented by the cocycle
\[
(t,t') \longmapsto \frac{1}{N_{\ell,\Gamma}} \frac{\Im\hlf{t}{t'}}{\abs{\delta_{\fieldk}}},
\]
any integer multiple of which is then contained in the kernel. Thus, for any $B \in \BIL_\Z$ the image
$[B]$ is a torsion element precisely if it is linear dependent to $\operatorname{tg}(\mathbf{1})$ over $\mathbb{Z}$.
\end{proof}
The linear dependence condition in the Lemma can more conveniently be formulated thus: If $B \in \BIL_\Z$,
the image is a torsion element if and only if there is a rational number $Q$, such that for all $t,t' \in D_{\ell, \Gamma}$, the following equation holds:
\begin{equation}\label{eq:lindep_cond}
B(t,t') - Q \frac{\Im\hlf{t}{t'}}{\abs{\delta_{\fieldk}}} = 0 .
\end{equation}
Since $D_{\ell, \Gamma}$ has full rank in $W_\mathbf{k}$, by linear extension, equivalently, the equation holds for all $t,t' \in W$; similarly for all $t,t' \in W_\mathbb{C}$.
As an example for this, we give an application to hermitian forms.
\begin{rmk}
Let $H$ be a $\mathbf{k}$-valued hermitian form on $W_\mathbf{k}$, and assume that $\frac{1}{\abs{\delta_{\fieldk}}}\Im H \in \BIL_\Z$.
Further assume that $H$ is linear in its left argument (otherwise, invert the sign in the equation below).
Then, the map $H \mapsto \frac{1}{\abs{\delta_{\fieldk}}}\Im H$ defines a torsion element in $\Htwo(\Gamma_\ell, \mathbb{Z})$ if and only if
the following equation holds for all $t,t' \in D_{\ell, \Gamma}$:
\begin{equation}\label{eq:extorsHerm}
H(t,t') + \frac{\operatorname{Tr}H}{n} \Im\hlf{t}{t'} = 0,
\end{equation}
where the trace $\operatorname{Tr}$ is taken over a normalized orthogonal basis for $\hlf{\cdot}{\cdot}$.
\end{rmk}
\begin{proof}
Taking the imaginary part of \eqref{eq:extorsHerm}, we see that indeed, if the equation holds,
$\Im H$ is a rational multiple of $\Im\hlf{\cdot}{\cdot}$ and thus $\frac{1}{\abs{\delta_{\fieldk}}}\Im H$ defines a torsion element by the Lemma.
Conversely, assume that the image is a torsion element in $\Htwo(\Gamma_\ell, \mathbb{Z})$.
Then, by the Lemma, the form has to be linear dependent to
$\frac{1}{\abs{\delta_{\fieldk}}}\Im\hlf{\cdot}{\cdot}$ and satisfies an equation of the form \eqref{eq:lindep_cond}.
Since by linear extension, the equation holds for all $t,t' \in W_\mathbb{C}$, we may replace $t$ by a purely imaginary multiple.
The resulting equation, equivalent to the first, is the following:
\[
\Re H(t,t') - Q\cdot \Re \hlf{t}{t'}= 0,
\]
valid for all $t,t' \in W_\mathbb{C}$. Whence by linear combination of the two equations, we find $H(t, t') = Q\hlf{t}{t'}$ for all $t,t' \in W_\mathbb{C}$.
To determine the factor of proportionality $Q$, we take the trace. With
$\operatorname{Tr}\hlf{\cdot}{\cdot}\mid_{W_\mathbb{C}} = - n$,
we get $Q = -\frac{1}{n}{\operatorname{Tr}H}$.
\end{proof}
\section{Local Heegner divisors and Borcherds products} \label{sec:heegner}
Our main interest here is to study the contribution of Heegner divisors to the local Picard group.
For this purpose, we will introduce local Borcherds products and, with their help, calculate the Chern classes of local Heegner divisors in $\Htwo(\Gamma_\ell, \mathbb{Z})$.
Then, we apply the cohomological results from section \ref{sec:cohom}.
The local Picard group $\mathrm{Pic}(X_\Gamma, \ell)$ is defined as the direct limit of the Picard groups on the regular loci (in the Baily-Borel compactification $X_{\Gamma, BB}^*$) of the open neighborhoods $U_{\epsilon}(\ell)$ of the cusp attached to $\ell$:
\begin{equation}
\label{eq:defPicLoc}
\mathrm{Pic}(X_\Gamma, \ell) = \varinjlim \mathrm{Pic}\bigl( U_{\epsilon}^{reg}\bigr).
\end{equation}
We can describe this local Picard group through the direct system $\mathrm{Pic}(\Gamma_\ell \backslash U_{\epsilon}(\ell))$ up to torsion, as $\Gamma_\ell$ has finite index in the stabilizer of the cusp, $\operatorname{Stab}_\Gamma(\ell)$. As the quotient $\operatorname{Stab}_\Gamma(\ell)/ \Gamma_\ell$ operates on the direct limit $\varinjlim \mathrm{Pic}(\Gamma_\ell \backslash U_{\epsilon}(\ell))$, for the invariant part, one has
\begin{equation}\label{eq:torEpstorLoc}
\mathrm{Pic}(X_\Gamma, \ell) \otimes \mathbb{Q} = \left(\varinjlim \mathrm{Pic}(\Gamma_\ell \backslash U_{\epsilon}(\ell)) \otimes \mathbb{Q} \right)^{\operatorname{Stab}_\Gamma(\ell) / \Gamma_\ell}.
\end{equation}
Thus, to describe the position of a local divisor up to torsion, it suffices to work with the Picard group $\mathrm{Pic}(\Gamma_\ell \backslash U_{\epsilon}(\ell))$ for a fixed (sufficiently small) $\epsilon>0$.
\begin{rmk}\label{rmk:toroidal_Pic}
Replacing the Baily-Borel compactification with the toroidal compactification $X_{\Gamma, tor}^*$,
the system of open neighborhoods $U_{\epsilon}$ is replaced by the system of open neighborhoods $\widetilde{V_\epsilon}(\ell)$ with the operation of $\Gamma_\ell / \Gamma_{\ell, T}$, and one can look at the Picard groups $\mathrm{Pic}\bigl(\Gamma_\ell / \Gamma_{\ell, T}\backslash \widetilde{V_\epsilon}(\ell)\bigr)$.
The main difference here is, that the divisor of $\{ q_\ell = 0 \}$ is now
a non-trivial element of the Picard group. A function with this divisor is given by $q_\ell = e(N_{\ell,\Gamma}^{-1}\tau)$. Note that the Chern class of $\{ q_\ell =0 \}$ is precisely $N_{\ell,\Gamma}^{-1}\frac{\Im\hlf{t}{t'}}{\abs{\delta_{\fieldk}}}$.
\end{rmk}
\subsection{Local Heegner divisors}\label{subsec:heegner}
First, we recall the usual definition of Heegner divisors on $\mathcal{H}$ \citep[cf.][Section 6]{Hof14},
and introduce local Heegner divisors in the neighborhoods
$U_\epsilon(\ell)$ of the cusp $[\ell]$.
Let $\lambda\in L'$ be a lattice vector of negative norm, i.e.\ $\hlf{\lambda}{\lambda}<0$.
The (primitive) \emph{Heegner divisor} $\mathbf{H}(\lambda)$ attached to $\lambda$ is a divisor on
$\mathcal{H}$ given by
\[
\begin{aligned}
\mathbf{H}(\lambda) & \vcentcolon =
\left\{ (\tau, \sigma) \in \mathcal{H}\,;\, \hlf{\lambda}{z(\tau,\sigma)} = 0 \right\}, \\
\end{aligned}
\]
with $z(\tau, \sigma) = \ell' - \tau\delta_{\fieldk}\hlf{\ell}{\ell'}\ell + \sigma$ (see section
\ref{subsec:symm_domain}).
Clearly, the divisor $\mathbf{H}(\lambda)$ intersects
$U_\epsilon(\ell)$ for every $\epsilon>0$, if and only if $\hlf{\lambda}{\ell} = 0$.
In the following, we denote by $\ell^\perp$ the (orthogonal) complement of $\ell$ with respect to $\hlf{\cdot}{\cdot}$.
Thus, let $\lambda \in L' \cap \ell^\perp$. Then, $\lambda = \lambda_\ell \ell + \lambda_D$ with $\lambda_D
\in W_\mathbf{k}$ and $\mathbf{H}(\lambda)$ is given by an equation of the form
\[
\lambda_\ell \hlf{\ell}{\ell'} + \hlf{\lambda_D}{\sigma} = 0.
\]
Consider the orbit of $\lambda$ under $\Gamma_\ell$.
Since $\Gamma$ is a modular group,
the Heisenberg group $\Gamma_\ell$ operates trivially on the discriminant group $L'/L$ and
thus $[h,t]\lambda \equiv \lambda\pmod{L}$ for all $[h,t] \in \Gamma_\ell$.
Also, since $\lambda \in \ell^\perp$, it remains fixed under $[h,0]$ for all $h\inN_{\ell,\Gamma}\mathbb{Z}$, and $\Gamma_{\ell, T}$ acts trivially.
For an Eichler element $[0,t]$ with $t\inD_{\ell, \Gamma}$, we have
\[
[0,t]\lambda = \lambda - \hlf{\lambda_D}{t} \ell =
(\lambda_\ell - \hlf{\lambda_D}{t}) \ell + \lambda_D.
\]
Thus, the orbit of $\lambda$ under $\Gamma_\ell/\Gamma_{\ell, T} \simeq D_{\ell, \Gamma}$ is given by
$\lambda - \mathfrak{T}\ell$, where $\mathfrak{T}$ denotes the set
\[
\mathfrak{T} = \mathfrak{T}(\lambda) \vcentcolon =
\left\{
\hlf{\lambda}{t}\, ;\, t\inD_{\ell, \Gamma}
\right\}.
\]
Note that $\mathfrak{T} \subseteq \mathfrak{d}_\fieldk^{-1}$ (as a fractional ideal), since $D_{\ell, \Gamma} \subseteq D$.
Hence, the group $\Gamma_\ell$ operates on the set $\lambda + \mathfrak{d}_\fieldk^{-1}\ell$ with only finitely many orbits and thus, the divisor
\begin{equation}\label{eq:defHinfty}
\mathbf{H}_\infty(\lambda) \vcentcolon = \sum_{\alpha \in \mathfrak{d}_\fieldk^{-1}}
\mathbf{H}(\lambda + \alpha\ell)
\end{equation}
is invariant under $\Gamma_\ell$ and defines an element of
$\operatorname{Div}{\left( \Gamma_\ell \backslash U_\epsilon(\ell)\right)}$.
\paragraph{Heegner divisors with index}
Now, let $\beta \in L'/L$ be an element of the discriminant group and
$m$ a negative integer. Then, the Heegner divisor of index $(\beta, m)$,
defined as the (locally finite) sum
\begin{equation}\label{eq:HeegUglob}
\mathbf{H}(\beta, m) = \sum_{\substack{\lambda \in L' \\ \QfNop(\lambda) = m \\ \lambda +
L = \beta}} \mathbf{H}(\lambda),
\end{equation}
is a $\Gamma$-invariant divisor on $\mathcal{H}$. Under the canonical projection
$\mathbf{H}(\beta, m)$ is the inverse image of the a divisor on $X_{\Gamma}$.
Also note that $\mathbf{H}(\beta, m) = \mathbf{H}(-\beta,m)$.
Through the open immersion $\Gamma_\ell \backslash U_\epsilon(\ell) \hookrightarrow
\Gamma\backslash \mathcal{H} = X_\Gamma$ from section \ref{subsec:cmpct}, the inclusion $U_\epsilon(\ell)
\subset \mathcal{H}$ and the projection maps, we get a commutative diagram
\[
\begin{tikzcd}
{\operatorname{Div}(X_\Gamma)} \arrow{r} \arrow{d} & { \operatorname{Div}(\Gamma_\ell \backslash
U_\epsilon(\ell)) } \arrow{d} \\
{\operatorname{Div}(\mathcal{H})} \arrow{r} & { \operatorname{Div}(U_\epsilon(\ell))\;. }
\end{tikzcd}
\]
We denote by $\mathbf{H}_\ell(\beta, m)$ the image in $\operatorname{Div}(\Gamma_\ell \backslash
U_\epsilon(\ell))$ of the divisor $\mathbf{H}(\beta, m) \in \operatorname{Div}(X_\Gamma)$.
The corresponding $\Gamma_\ell$-invariant divisor in $\operatorname{Div}(U_\epsilon(\ell))$ is
also denoted by $\mathbf{H}_\ell(\beta, m)$.
For sufficiently small $\epsilon$, the divisor $\mathbf{H}_\ell(\beta, m)$ is given by the restriction
to $U_\epsilon(\ell)$ of the sum on the right hand-side of \eqref{eq:HeegUglob}.
Then, only $\lambda$'s perpendicular to $\ell$ contribute.
In particular, if $\mathbf{H}_\ell(\beta, m)$ is non-zero, then $\beta$ is contained
in the subgroup
\[
\mathcal{L} \vcentcolon = \left\{
\gamma \in L'/L\,;\; 2\Re\hlf{\gamma}{\ell} \equiv 0
\bmod{M_1}\quad\text{and}\quad \abs{\delta_{\fieldk}}\Im\hlf{\gamma}{\ell} \equiv 0
\bmod{M_2} \right\} \subseteq L'/L,
\]
where $M_1$, $M_2$ are the unique integers given by
$2\Re\hlf{L}{\ell} = M_1 \mathbb{Z}$ and by $\abs{\delta_{\fieldk}}\Im\hlf{L}{\ell} =
M_2\mathbb{Z}$.
With $\beta \in \mathcal{L}$ the local divisor $\mathbf{H}_\ell(\beta, m)$ can be written in the form
\begin{equation}\label{eq:HeegUell_bm}
\begin{aligned}
\mathbf{H}_\ell(\beta, m)
= \sum_{\substack{ \kappa \in D \\ \QfNop(\kappa + \dot\beta) = m}}\mathbf{H}_\infty(\kappa +
\dot\beta).
\end{aligned}
\end{equation}
Here, we adopt the notation of \cite{BrFr} section 4, by which $\dot\beta$ denotes a
representative of $\beta$ with $\dot\beta \in L' \cap \ell^\perp$, fixed once and for all for
every $\beta \in \mathcal{L}$.
Note that a surjective homomorphism is given by
\[
\pi: \mathcal{L} \longrightarrow D'/D, \quad \beta \longmapsto \dot\beta_D,
\]
where $\dot\beta_D$ denotes the definite part of $\dot\beta$.
\subsection{Local Borcherds products}
\label{subsec:localbp}
In this section, our aim is to use local Borcherds products to describe the
position of Heegner divisors in the cohomology. Given a lattice
vector $\lambda$ of negative norm with $\lambda \in L'\cap \ell^\perp$
we can realize the local Heegner divisor attached to $\lambda$ through an infinite product with
factors of the form $\left(1 - e\left(\hlf{z}{[0,t]\lambda}\right)\right)$ with $[0,t] \in \Gamma_{\ell,T}$.
If for the Heegner divisor $\mathbf{H}_\infty(\lambda)$ as in \eqref{eq:defHinfty}, we set
\[
\prod_{\beta \in \mathfrak{d}_\fieldk^{-1}}
\bigl[ 1 - e\left( \sigma(\beta) \hlf{z}{\lambda - \beta\ell} \right)\bigr],
\qquad \text{with}\quad \sigma(\beta) \in \{\pm 1 \},
\]
we get an infinite product with (zero-)divisor $\mathbf{H}(\infty)$.
For $\sigma(\beta) \equiv 1$ the product would be $\Gamma_\ell$-invariant.
However, to assure absolute convergence, we must define the sign $\sigma(\beta)$ depending on $\Im\beta$.
Then, the product is no longer fully invariant. Instead, the operation of Eichler
transformations gives rise to a non-trivial automorphy factor, which we will use to determine the
position of $\mathbf{H}_\infty(\lambda)$ in the local Picard group.
\begin{assumption*}From here on, we shall require that $\hlf{\ell}{\ell'} = \delta_{\fieldk}^{-1}$.
\end{assumption*}
We remark that this is not a particularly serious restriction, as under the assumptions concerning $\ell$ and $\ell'$ from section \ref{subsec:ltcs}, it is always possible to choose $\ell'$ suitably.
Now, keeping in mind that $\mathfrak{d}_\fieldk = \delta_{\fieldk}^{-1}\mathcal{O}_{\fieldk}$ and $\overline{\mathcal{O}}_\mathbf{k} = \mathcal{O}_{\fieldk} = - \mathcal{O}_{\fieldk}$, we define the local Borcherds products as follows:
\begin{definition}\label{def:localBp}
Let $\lambda \in L'$ be a negative norm lattice vector in the orthogonal complement of $\ell$. The local
Borcherds product $\Psi_\lambda(z)$ attached to $\mathbf{H}_\infty(\lambda)$ is defined as
\[
\Psi_\lambda(z) \vcentcolon= \prod_{\alpha \in \mathcal{O}_{\fieldk}}
\left[1 - e\left(\sigma(\Im\alpha)\left(
\hlf{z}{\lambda} + \frac{\alpha}{\abs{\Disc}} \right)
\right) \right],
\]
with a sign $\sigma(\Im\alpha)$ defined as follows:
\[
\sigma(\Im\alpha) = \begin{cases}
\phantom{-}1 & \quad \text{if $\Im \alpha \geq 0$,} \\
-1 & \quad\text{otherwise.}
\end{cases}
\]
\end{definition}
Clearly, $\Psi_\lambda(z)$ is an absolutely convergent infinite product with divisor $\mathbf{H}_\infty(\lambda)$.
With $\mathfrak{d}_\fieldk^{-1} = \delta_{\fieldk}^{-1}\left( \mathbb{Z} + \zeta\mathbb{Z}\right)$, where $\Im \zeta = \frac12 \delta_{\fieldk}$ and $2\Re\zeta \equiv \Disc \pmod{4}$,
we can write $\Psi_\lambda(z)$ in the following form
\[
\Psi_\lambda(z) =
\prod_{\substack{p\bmod{ \abs{\Disc}} \\ q \in \mathbb{Z}}}
\left[1 - e\left(\sigma(q)\left(
\hlf{z}{\lambda} + \frac{1}{\abs{\Disc}} \left( p + \zeta\, q \right)\right)
\right) \right],
\]
with $\sigma(q) = \operatorname{sign}(q)$ if $q\neq 0$ and $\sigma(0) = +1$.
Note that $\Psi_\lambda$ is invariant under translations in $\Gamma_{\ell, T}$,
while the operation of Eichler transformations, $[0,t]$ with $t\in D_{\ell, \Gamma}$, gives rise to
the (non-trivial) automorphy factor
\begin{equation}\label{eq:J_def}
J_\lambda([h,t], z) = \frac{\Psi_\lambda([0,t]z)}{\Psi_\lambda(z)} \quad \left([h,t] \in
\Gamma_\ell\right).
\end{equation}
\begin{proposition}\label{prop:J_lambda}
The automorphy factor $J_\lambda$ attached to $\mathbf{H}_\infty(\lambda)$ takes the form
\[
J_\lambda([h,t], z) =
e\left( -2\abs{\Disc}\hlf{z}{\lambda}\Re\hlf{t}{\lambda} - 2\left(\Re\hlf{t}{\lambda}\right)^2\zeta + \Re\hlf{t}{\lambda}\left( \zeta +1\right) \right),
\]
with $\zeta$ such that $\mathcal{O}_{\fieldk} = Z + \zeta\mathbb{Z}$. Note that $J_\lambda$ is independent of the choice of $\zeta$.
\end{proposition}
\begin{proof}
Since $\hlf{\ell'}{\ell} =-{\delta_{\fieldk}}^{-1}$, by \eqref{def:eichlerU} we have
$\hlf{[0,t]z}{\lambda} = \hlf{z}{\lambda} - {\delta_{\fieldk}}^{-1} \hlf{t}{\lambda}$.
Since $\hlf{t}{\lambda} = \hlf{t}{\lambda_D} \in \mathfrak{d}_\fieldk^{-1}$ we can write
\[
\hlf{[0,t]z}{\lambda} = \hlf{z}{\lambda} + \frac{1}{\abs{\Disc}}(r + \zeta s), \qquad\text{with}\; r,s\in\mathbb{Z}.
\]
We note that $s = 2\Re\hlf{t}{\lambda}$.
Now, after permuting representatives modulo $\abs{\Disc}$
and a shift in the index $q$, the automorphy factor from \eqref{eq:J_def} takes the form
\begin{equation}\label{eq:Jintermed}
J_\lambda([h,t], z) =
\prod_{ p \bmod{\abs{\Disc}}}\prod_{q \in \mathbb{Z}}
\frac{1 - e\left( \sigma(q-s)\left(
\hlf{z}{\lambda} + \abs{\Disc}^{-1} \left( p + q \zeta \right)
\right) \right)}%
{1 - e\left( \sigma(q)\left(
\hlf{z}{\lambda} + \abs{\Disc}^{-1} \left( p + q\zeta \right) \right)
\right)}.
\end{equation}
Only factors with $\sigma(q-s) \neq \sigma(q)$ contribute to the
product. There are two cases: Either we have $s > q \geq 0$, or $s \leq q < 0$.
We examine the first case. By applying the elementary identity
\[
\frac{1 - e(-z)}{1-e(z)} = -e(-z).
\]
we get
\[
\begin{aligned}
J_\lambda([h,t], z)
& = \prod_{ p \bmod{\abs{\Disc}}}\prod_{0 \leq q < s}
- e\left( - \hlf{z}{\lambda} - \frac{1}{\abs{\Disc}} \left( p + q \zeta \right)
\right) \\
& = \prod_{ p \bmod{\abs{\Disc}}} (-1)^s
e\left( -s\hlf{z}{\lambda} - \frac{s}{\abs{\Disc}}\left( p +\frac{s-1}{2} \zeta\right)\right) \\
& = e\left( - s\abs{\Disc}\hlf{z}{\lambda} - \frac{s(s-1)}{2} \zeta - \frac{s(\abs{\Disc} - 1)}{2} + \frac{s\abs{\Disc}}{2} \right)\\
& = e\left( -s\abs{\Disc}\hlf{z}{\lambda} - \frac{s^2}{2}\zeta + \frac{s}{2}\zeta +\frac{s}{2} \right) \\
\end{aligned}
\]
Hence, recalling that $s = 2 \Re\hlf{t}{\lambda}$, we have
\begin{equation}\label{eq:thecocyclJ}
J_\lambda([h,t], z) =
e\left(
-2\abs{\Disc}\hlf{z}{\lambda}\Re\hlf{t}{\lambda}
- 2\left(\Re\hlf{t}{\lambda}\right)^2\zeta
+ \Re\hlf{t}{\lambda} \zeta + \Re\hlf{t}{\lambda} \right).
\end{equation}
We remark that the last term is determined only up to sign, since $2\Re\hlf{t}{\lambda} \in \mathbb{Z}$.
Finally, we note that as the second term in \eqref{eq:thecocyclJ}
is a quarter-integer while $2\Re\zeta$ is only determined modulo $4$,
the automorphy factor is independent of the choice for $\Re\zeta$.
The second case ($s\leq q < 0$) can be treated similarly, yielding the same result for the
automorphy factor $J_\lambda([h,t], z)$.
\end{proof}
\subsection{The Chern class of a Heegner divisor \texorpdfstring{$\mathbf{H}_\infty(\lambda)$}{}}\label{subsec:chern}
From the automorphy factor $J_\lambda$ we now determine a two-cocycle representing the Chern class of
the Heegner divisor $\mathbf{H}_\infty(\lambda)$.
\begin{proposition} \label{prop:ChernClassHl}
The Chern class $\delta(\mathbf{H}_\infty(\lambda))$ of the local Heegner divisor $\mathbf{H}_\infty(\lambda)$ in $\Htwo(\Gamma_\ell, \mathbb{Z})$ is determined by the cocycle
\[
[c_\lambda]: \quad ([h,t][h',t']) \longmapsto - 2\abs{\delta_{\fieldk}} \Re\hlf{t}{\lambda}\Im\hlf{t'}{\lambda}
= \Im\left( -\abs{\delta_{\fieldk}} F_\lambda(t,t') \right),
\]
where $F_\lambda(t,t') \vcentcolon = 2\Re\hlf{t}{\lambda} \hlf{t'}{\lambda}$.
\end{proposition}
\begin{proof}
To calculate the Chern class, we must realize the connecting homomorphism $\delta: \mathrm{H}^{1}(\Gamma_\ell, \mathcal{O}_\epsilon^*) \rightarrow \mathrm{H}^{2}(\Gamma_\ell, \mathbb{Z})$.
Thus, let $A(g,z)$ be a holomorphic function satisfying $ J_\lambda (g,z) = e\left(A(g,z)\right)$ and
set
\begin{equation}\label{eq:infcocycl}
c(g,g') = A(gg', z) - A(g,g'z) - A(g', z) \quad\text{for all $g,g' \in \Gamma_\ell$}.
\end{equation}
Then, the two-cocycle defined by the map $(g,g) \mapsto c(g,g')$ is a representative for the Chern class in $\mathrm{H}^2(\Gamma_\ell, \mathbb{Z})$.
Note that while $A(g,g')$ is not uniquely determined, $c(g,g')$ is independent of this choice; also, multiplying $J_\lambda$ with a trivial automorphy factor
changes $c(g,g')$ only by a coboundary.
Clearly, it suffices to calculate $c(g,g')$ for Eichler transformations $g = [0,t]$ and $g' = [0,t']$.
From \eqref{eq:infcocycl} we see that the last two terms in \eqref{eq:thecocyclJ}, being linear in $t$, cancel. We calculate
\begin{align*}
A([0, t+t'], z) & - A([0,t],[0,t']z) - A([0, t'], z) =\\
= &\; 2\abs\Disc \hlf{[0,t'] z - z}{\lambda} \Re\hlf{t}{\lambda} - 4\Re\hlf{t}{\lambda}\Re\hlf{t'}{\lambda}\zeta \\
= &\; 2\delta_{\fieldk}\Re\hlf{t}{\lambda} \hlf{t'}{\lambda} - 2\Re\hlf{t}{\lambda}\Re\hlf{t'}{\lambda} \delta_{\fieldk} - 4\Re\hlf{t}{\lambda}\Re\hlf{t'}{\lambda}\Re\zeta \\
= &\, -2\abs\delta_{\fieldk}\Re\hlf{t}{\lambda}\Im\hlf{t'}{\lambda} - 4\Re\hlf{t}{\lambda}\Re\hlf{t'}{\lambda}\Re\zeta,
\end{align*}
since $\Im\zeta =\frac12 \abs\delta_{\fieldk}$.
Now, consider the second term. We know $J_\lambda$ doesn't depend on the choice of $\Re\zeta$, thus
this term contributes at most a torsion element in the cohomology or vanishes entirely. It can hence be ignored.
Note also that the remaining first term is an integer for all $[0,t], [0,t'] \in \Gamma_{\ell, T}$.
\end{proof}
It is worth noting that the bilinear form $F_\lambda(\cdot, \cdot)$ introduced in Proposition \ref{prop:ChernClassHl} can be written in the form
\[
F_\lambda(a,b) = \Re\hlf{a}{\lambda} \hlf{b}{\lambda} = \hlf{b}{\lambda} \hlf{a}{\lambda} + \hlf{b}{\lambda} \hlf{\lambda}{a} \qquad(a,b \in W_\mathbf{k}).
\]
Clearly, the first term is a complex bilinear form, while the second term is a hermitian form, we denote them by $B_\lambda(a,b)$ and $H_\lambda(a,b)$, respectively. Note that $H_\lambda(a,b)$ is linear in its \emph{second} argument. Further, we remark that $F_\lambda(a,b) = F_{\lambda_D}(a,b)$ for all $a, b \in W_\mathbf{k}$.
\subsection{Torsion criteria for Heegner divisors}
Up to here, we have only worked on Heegner divisors attached to individual lattice
vectors, i.e.\ $\mathbf{H}_\infty(\lambda)$, for $\lambda \in L'$ with
$\Qf{\lambda}< 0$.
Next, we consider linear combinations of Heegner divisors. We will be
mainly interested in the Heegner divisors $\mathbf{H}_\ell(\beta, m)$.
For general linear combinations of Heegner divisors, we have the following Lemma:
\begin{lemma}\label{lemma:tors_lcH}
Let $\mathbf{H}$ be a finite linear combination of Heegner divisors of the form
\[
\mathbf{H} = \sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}} a(\lambda) \mathbf{H}_\infty(\lambda), \qquad \left( a(\lambda) \in \mathbb{Z} \quad \text{for every $\lambda$}\right).
\]
Then, the Chern class $\delta(\mathbf{H})$ of $\mathbf{H}$ is a a torsion element in $\Htwo(\Gamma_\ell, \mathbb{Z})$
if and only if for all $t,t' \in D_{\ell, \Gamma}$ the following equation holds
\[
\sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}} a(\lambda)
\left[ F_\lambda(t,t') - \frac{\hlf{\lambda}{\lambda}}{n} \hlf{t'}{t} \right] = 0.
\]
\end{lemma}
From the proof of this lemma we will also get the following necessary condition (where we use the same notation as in the Lemma):
\begin{corollary}\label{cor:nec_cond}
If $\delta(\mathbf{H})$ is a torsion element, then for the bilinear form $B_\lambda(a,b) = \hlf{a}{\lambda}\hlf{b}{\lambda}$ we have
\begin{equation}\label{eq:lemma_nec_cond}
\sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}}
a(\lambda) \tr B_\lambda = 0,
\end{equation}
where the trace is taken over a normal orthogonal basis with respect to $\hlf{\cdot}{\cdot}$.
\end{corollary}
\begin{proof}
The Chern class $\delta(\mathbf{H})$ is given by a linear combination
of cocycles $[c_\lambda]$ in $\Htwo(\Gamma_\ell, \mathbb{Z})$.
By Proposition \ref{prop:ChernClassHl}, each $[c_\lambda]$ is represented
by the two-cocycle
\[
(t,t') \mapsto -\Im\left[ \abs{\delta_{\fieldk}} F_\lambda(t,t')\right].
\]
Through \eqref{eq:Pic_H1} and the exactness of the sequence in \eqref{eq:seqH2ZtoOps}, the image of $[c_\lambda]$ in $\mathrm{H}^2(\Gamma_\ell, \mathcal{O}_\epsilon)$ vanishes.
By the results of section \ref{sec:cohom}, $\delta(\mathbf{H})$ is a torsion element in $\Htwo(\Gamma_\ell, \mathbb{Z})$ if and only if there is a rational number $Q$ such that the equation
\[
\sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}} a(\lambda)\abs{\delta_{\fieldk}} \cdot\Im F_\lambda(t,t')
= Q \frac{\Im\hlf{t'}{t} }{\abs{\delta_{\fieldk}}}
\]
holds for all $t, t' \in D_{\ell, \Gamma}$.
Since $D_{\ell, \Gamma}$ has full rank in $W_\mathbf{k}$, by extension of scalars, the equation holds for all pairs of vectors in $W_\mathbf{k}$.
Both sides of the equation are linear in $t'$. Thus replacing $t'$ with a purely imaginary multiple gives a second, equivalent equation:
\[
\sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}} a(\lambda)\abs{\delta_{\fieldk}} \cdot \Re F_\lambda(t,t')
= Q \frac{\Re\hlf{t'}{t} }{\abs{\delta_{\fieldk}}}.
\]
By linear combination of the two equations, we get
\begin{equation}\label{eq:Q_cmplxcond}
\sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}} a(\lambda)\abs{\delta_{\fieldk}}
F_\lambda(t,t') = Q \frac{\hlf{t'}{t} }{\abs\delta_{\fieldk}}.
\end{equation}
To determine $Q$, we take the trace of both sides of \eqref{eq:Q_cmplxcond}, using an orthogonal basis of $W_\mathbb{C}$ with respect to $\hlf{\cdot}{\cdot}$, say $\{e_l\}_{l=1,\dotsc, n}$ with $\hlf{e_l}{e_m} = -\delta_{l,m}$.
Now, $\tr \hlf{\cdot}{\cdot} = -n$ and the trace of $H_\lambda$ is $- \hlf{\lambda}{\lambda}$,
hence
\begin{equation}\label{eq:factorQ}
Q(-n) = \sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}} a(\lambda) \abs{\Disc}
\left( - \hlf{\lambda}{\lambda} + \tr_{\{e_l\}} B_\lambda \right).
\end{equation}
It turns out the trace of $B_\lambda$ does not contribute to $Q$.
Indeed, if we take the trace of \eqref{eq:Q_cmplxcond} over an orthogonal basis of $W_\mathbb{C}$ obtained from $\{ e_l\}$ by rescaling with the complex unit $i$, i.e.\ $\{ i e_l\}_{l=1,\dotsc, n}$,
the traces of the hermitian forms $\hlf{\cdot}{\cdot}$ and $H_\lambda$ remain unchanged while that of $B_\lambda$ switches sign.
Comparing this result with \eqref{eq:factorQ}, we obtain
\[
Q = \sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}}
a(\lambda)\cdot Q_\lambda
\quad\text{with}\quad
Q_\lambda \vcentcolon= \abs{\Disc} \frac{\hlf{\lambda}{\lambda}}{n}.
\]
Together with \eqref{eq:Q_cmplxcond} the statement follows.
Further, since the contribution of $B_\lambda$ to the trace vanishes, we get the necessary condition
\begin{equation*}
\sum_{\substack{\lambda \in L'\cap \ell^\perp \\ \QfNop(\lambda) < 0}}
a(\lambda) \tr B_\lambda = 0.
\end{equation*}
This proves the corollary, as well.
\end{proof}
\subsection{The Main Result}
We can now turn to the object of our main interest, Heegner divisors of the form $\mathbf{H}_\ell(\beta, m)$.
We want to describe their position in the local Picard group. Recall that by \eqref{eq:HeegUell_bm} the divisors $\mathbf{H}_\ell(\beta, m)$
can be written using divisors of the type $\mathbf{H}_\infty(\lambda)$. Thus, any finite linear combination $\mathbf{H}$ of Heegner divisors $\mathbf{H}_\ell(\beta, m)$,
can be written as a locally finite sum of Heegner divisors $\mathbf{H}_\infty(\lambda)$. Also, note that for a divisor of this type, the Chern
class $\delta(\mathbf{H}_\infty(\lambda))$ depends only on the projection $\lambda_D$.
With this notation, we formulate the following theorem.
\begin{theorem}\label{thm:H_torsCond}
Consider a finite linear combination of local Heegner divisors of the form
\begin{equation}
\mathbf{H} = \frac12 \sum_{\beta \in \mathcal{L}} \sum_{\substack{m \in \mathbb{Z} + \QfNop(\beta) \\ m <0 }}
c(\beta, m) \mathbf{H}_\ell(\beta, m),
\end{equation}
with integral coefficients $c(\beta, m)$, satisfying $c(\beta, m) = c(-\beta,m)$.
Then, $\mathbf{H}$ is torsion element in the Picard group $\mathrm{Pic}\left(\Gamma_\ell \backslashU_{\epsilon}(\ell)\right)$
if and only if for all $t, t' \in D_{\ell, \Gamma}$ the following equation holds
\begin{equation}\label{eq:tors_H}
\sum_{\beta \in \mathcal{L}}\; \sum_{\substack{m \in \mathbb{Z} + \QfNop(\beta) \\ m < 0 }} c(\beta, m)
\sum_{\substack{\lambda \in D' \\ \lambda + D \equiv\pi(\beta) \\
\QfNop(\lambda) = m}} \left[
F_\lambda(t,t') - \frac{\hlf{\lambda}{\lambda}}{n} \hlf{t'}{t}
\right] = 0.
\end{equation}
Further, a necessary conditions for this to be the case is
that the following identity holds, with $B_\lambda(x,y) = \hlf{x}{\lambda}\hlf{y}{\lambda}$:
\begin{equation}\label{eq:tr_cond}
\sum_{\beta \in \mathcal{L}}\; \sum_{\substack{m \in \mathbb{Z} + \QfNop(\beta) \\ m < 0 }} c(\beta, m)
\sum_{\substack{\lambda \in D' \\ \lambda + D \equiv\pi(\beta) \\
\QfNop(\lambda) = m}} \tr B_\lambda = 0.
\end{equation}
Here, the trace is taken over an orthogonal basis with respect to $\hlf{\cdot}{\cdot}$.
\end{theorem}
We note that by \eqref{eq:torEpstorLoc} a linear combination of Heegner divisors $\mathbf{H}$ is a torsion element in $\mathrm{Pic}(\Gamma_\ell\backslash U_\epsilon(\ell))$ if and only if it is a torsion element in the local Picard group $\mathrm{Pic}( X_\Gamma, \ell)$.
\begin{proof}
\emph{If $H$ is a torsion element}, the equation \eqref{eq:tors_H} follows from Lemma \ref{lemma:tors_lcH}.
Also, from the proof of that Lemma and Corollary \ref{cor:nec_cond}, it is clear that in this case, the identity
\eqref{eq:tr_cond} holds.
\emph{For the converse,} assume that \eqref{eq:tors_H}
holds for all $t, t' \in D_{\ell, \Gamma}$. We will show that $\mathbf{H}$ is a torsion element in the Picard group.
By extension of scalars, the equation remains valid for all $t,t' \in W_\mathbb{C}$.
Using \eqref{eq:thecocyclJ}, an automorphy factor describing $\mathbf{H}$ in
$\mathrm{Pic}\left(\Gamma_\ell\backslashU_{\epsilon}(\ell)\right)$ is given by the following (finite) product (for $g=[h,t]
\in \Gamma_\ell$, $z\inU_{\epsilon} = U_{\epsilon}(\ell)$):
\begin{multline}\label{eq:JH_factors}
J_{\mathbf{H}}(g, z) = \prod_{\substack{\beta \in \mathcal{L} \\ m \in \mathbb{Z} + \QfNop(\beta) \\ m<0}}
\prod_{\substack{\kappa \in D\\ \QfNop(\kappa + \dot\beta) = m }}
J_{\kappa + \dot\beta}(g, z)^{\frac12 c(\beta,m)} \\
= \prod_{\beta, m}\prod_{\kappa}
e\Bigl(
-2\abs{\Disc}\hlfa[big]{z , \kappa + \dot\beta}\Re\hlfa[big]{t,\kappa + \dot\beta_D} \\
- 2\zeta\left(\Re\hlfa[big]{t, \kappa + \dot\beta_D}\right)^2
+ \Re\hlfa[big]{t, \kappa + \dot\beta_D}\left( \zeta + 1\right) \Bigr)^{\frac{ c(\beta,m)}{2}}.
\end{multline}
Since $c(\beta,m) = c(-\beta,m)$, terms which are linear in the $\kappa +
\dot\beta_D$ cancel. The remaining factors are of the form
\[
e\left(-2\abs{\Disc}\hlfa[big]{z , \kappa + \dot\beta}\Re\hlfa[big]{t,\kappa + \dot\beta_D}
- 2\zeta\bigl(\Re\hlfa[big]{t, \kappa + \dot\beta_D}\bigr)^2
\right)^{\frac{c(\beta,m)}{2}}.
\]
Now, we write
$\bigl\langle {z, \kappa + \dot\beta} \bigr\rangle = \bigl\langle {z, \kappa +
\dot\beta_D}\bigr\rangle + \bigl\langle {z , \dot\beta - \dot\beta_D}\bigr\rangle$.
Since $\hlfa{\dot\beta, \ell} = 0$, the second part depends only on the
constant $\ell'$-component of $z$.
We get
\[
\left[ e\left( - 2\delta_{\fieldk}\overline{\dot\beta_\ell} \Re\hlfa[big]{t, \lambda_D}\right)
e\left( -2 \abs{\Disc}\hlfa[big]{\sigma, \kappa + \dot\beta_D} \Re\hlfa[big]{t, \kappa + \dot\beta_D}
-2 \zeta\bigl( \Re\hlfa[big]{t, \kappa + \dot\beta_D}\bigr)^2
\right)\right]^{\frac{c(\beta,m)}{2}}.
\]
We ignore the first factor for the time being and examine the second factor. There, the first term in the exponential is
$-2\abs\Disc F_{\kappa + \dot\beta_D}(t,\sigma)$ while the second term is equal to
$-2\zeta \Re F_{\kappa + \dot\beta_D}(t,t)$.
We apply \eqref{eq:tors_H} to both terms, and can rewrite this factor in the form
\begin{align}
e\left( 2\Disc\vphantom{\frac{\dot\beta_D}{n}} \right. & \left.\frac{\hlfa[big]{\kappa + \dot\beta_D, \kappa + \dot\beta_D}}{n}\Bigl[
\hlf{\sigma}{t} - \frac{\zeta}{\Disc}\hlf{t}{t} \Bigr] \right) \nonumber\\ \label{eq:somefactors}
&\qquad = e\left( 2\Disc\frac{\QfNop\bigl(\kappa + \dot\beta_D\bigr)}{n}\Bigl[
\hlf{\sigma}{t} - \frac{1}{2\delta_{\fieldk}}\Qf{t} \Bigr] \right) e\left( - 2\Re\zeta\cdot \frac{\Qf{\lambda}}{n} \Qf{t} \right).
\end{align}
Clearly, the last factor in \eqref{eq:somefactors} has finite order and is a torsion element in $\mathrm{Pic}(\Gamma_\ell\backslashU_{\epsilon})$.
Now, we claim that the first factor is actually a trivial automorphy factor. To see this, consider the invertible function
$f(z) = e(c\tau)$ with $c\in\mathbb{Q}^\times$; under the operation of $\Gamma_{\ell, T}$, it gives rise to the following trivial automorphy factor
\[
j_1([0,t], z) = \frac{f([h,t]z)}{f(z)} = e\left(c\left( - \hlf{\sigma}{t} +
\frac{1}{2\delta_{\fieldk}}\hlf{t}{t}\right)\right).
\]
Hence the first factor in \eqref{eq:somefactors} is indeed trivial.
Now, we return to the previously excluded factor
\begin{equation}\label{eq:furtherfactors}
e\left( - 2\delta_{\fieldk}\overline{\dot\beta_\ell} \Re\hlfa[big]{t, \kappa + \dot\beta_D}\right)
= e\left( - \left( \abs{\delta_{\fieldk}} \Im\dot\beta_\ell + \delta_{\fieldk}\Re\dot\beta_\ell\right) 2\Re\hlfa[big]{t, \kappa + \dot\beta_D} \right).
\end{equation}
Since $\abs{\delta_{\fieldk}} \Im\dot\beta_\ell$ is rational (actually, half-integer), this term contributes only a torsion element in the Picard-group.
Consider the invertible function $g(z) = e\left(\hlf{\sigma}{\mu}\right)$ with $\mu \in W_\mathbf{k}$ from which we get the trivial automorphy factor
\begin{equation*}
j_2(z,[0,t]) = e\left( \delta_{\fieldk}^{-1} \hlf{t}{\mu}\right).
\end{equation*}
Setting $\mu = \kappa + \dot\beta_D$, we multiply \eqref{eq:furtherfactors} with a suitable power of $j_2$
to kill the term in $\Re\hlfa[big]{t, \kappa + \dot\beta_D}$. Then, only torsion elements remain, as
$\abs{\delta_{\fieldk}} \Im\hlfa[big]{t, \kappa + \dot\beta_D}$ and $\Re\dot\beta_\ell$ are rational numbers.
Thus, we find that each of the finitely may factors of $J_\mathbf{H}$ from \eqref{eq:JH_factors}
can be expressed through suitable powers of trivial automorphy factors of the types $j_1$ and $j_2$ and factors of finite order.
Hence, it follows that $\mathbf{H}$ is a torsion element in
$\mathrm{Pic}\left(\Gamma_\ell\backslash U_{\epsilon}\right)$.
\end{proof}
\begin{rmk}
As in Remark \ref{rmk:toroidal_Pic}, if one looks at the neighborhoods $\widetilde{V_\epsilon}(\ell)$ from section \ref{subsec:cmpct} rather than $U_\epsilon(\ell)$,
in the proof, the function $f$ is no longer invertible and the automorphy factor $j_1$ becomes non-trivial,
since $f$ vanishes on the disk center $\{ q_\ell = 0 \}$. As mentioned before, the Chern class is given by $(t,t') \mapsto (N_{\ell,\Gamma}\abs{\delta_{\fieldk}})^{-1}\Im\hlf{t}{t'}$.
Thus, in the Theorem one would have to replace ''torsion element in $\mathrm{Pic}(\Gamma_\ell \backslash U_{\epsilon}(\ell))$`` by ''equivalent to the divisor of $\{ q_\ell = 0 \}$ in
$\mathrm{Pic}(\Gamma_\ell/\Gamma_{\ell, T}\backslash \widetilde{V_\epsilon}(\ell)\bigr)$ (up to torsion)``.
This kind of statement also carries over to the direct limit and describes the position (up to torsion) of $\mathbf{H}$ in
$\varinjlim \mathrm{Pic}(\widetilde{V_\epsilon}(\ell))$, which in analogy to \eqref{eq:defPicLoc} may be considered as a local Picard group for the cusp $[\ell]$ on $X_{\Gamma, tor}^*$.
\end{rmk}
\begin{rmk} \label{rmk:BrFr_thm}
Recall how the rational space $V_\mathbb{Q}$ underlying $V_\mathbf{k}$ has the structure of a quadratic space of signature $(2,2n+2)$.
Let $\group{O}(V)$ be the orthogonal group of $V_\mathbb{Q}$ and $\group{O}(V)(\mathbb{R})$ its set of real points.
In \cite{BrFr}, Bruinier and Freitag study local Heegner divisors at generic boundary components of the symmetric domain
for such indefinite orthogonal groups. The local Heegner divisors we consider here can be described as the restriction of their local Heegner
divisors:
For $\lambda \in D'$, $\mathbf{H}_\infty(\lambda)$ is
the restriction of a local Heegner divisor attached to $\lambda$ and, similarly, $\mathbf{H}_\ell(\beta,m)$ is
the restriction of a composite local Heegner divisor, in the local Picard group for a generic boundary component of the symmetric domain,
defined by $\mathbf{k}\ell$ as a two-dimensional isotropic subspace over $\mathbb{Q}$.
This follows from the embedding theory developed by the author in \cite{Hof11, Hof14}.
The relationship between Theorem \ref{thm:H_torsCond} and the results in \cite{BrFr} is the following:
By taking the real part of both sides of \eqref{eq:tors_H}, one gets precisely the torsion
condition from \citep[Theorem 4.5]{BrFr}. It follows that, under these assumptions,
if a local Heegner divisor $\mathbf{H}$ as in Theorem \ref{thm:H_torsCond} is a torsion element in $\mathrm{Pic}\left(\Gamma_\ell\backslashU_{\epsilon}(\ell)\right)$, there is a pre-image under restriction which satisfies the torsion criterion in \cite{BrFr} and hence is a torsion element in the local Picard group for a generic boundary component of the orthogonal modular variety.
Conversely, for every local Heegner divisor there which restricts to $\mathbf{H}$, the criterion of \cite{BrFr} implies that that \eqref{eq:tors_H} holds for $\mathbf{H}$.
\end{rmk}
\section{Application to modular forms}\label{subsec:mf_lgkz}
In this section as an application of Theorem \ref{thm:H_torsCond} we derive a
statement describing obstructions to local Borcherds products through certain vector valued cusp forms.
Our results are closely related to those obtained by Bruinier and Freitag in the context of orthogonal groups \citep[see][Section 5]{BrFr}.
Let us briefly recall some standard facts about the Weil representation and definition of vector valued modular forms.
The rational space $W_\mathbb{Q}$ underlying $W_\mathbf{k}$, equipped with the quadratic form $\Qf{\cdot}$, is negative definite with dimension $2n$,
and the definite lattice $D$ it contains has even $\mathbb{Z}$-rank $2n$.
Hence, the Weil representation of the metaplectic group $\group{Mp}_2(\mathbb{Z})$,
defined as the pre-image of $\group{SL}_2(\mathbb{Z})$ under the double covering map $\group{Mp}_2(\mathbb{R}) \twoheadrightarrow \group{SL}_2(\mathbb{R})$, factors over $\group{SL}_2(\mathbb{Z})$.
Thus, there is a unitary representation of $\group{SL}_2(\mathbb{Z})$ on the group algebra $\mathbb{C}[D'/D]$, denoted $\rho_D$.
The dual representation to $\rho_D$ is denoted by $\rho_D^*$.
The Weil-representation $\rho_D$ is defined through the action of the generators of $\group{SL}_2(\mathbb{Z})$,
$T = \left(\begin{smallmatrix} 1 & 1 \\ 0 & 1\end{smallmatrix} \right)$
and $S =\left( \begin{smallmatrix} 1 & 1 \\ 0 & 1\end{smallmatrix} \right)$.
Note that $\rho_D^*$ can be obtained from $\rho_D$ by complex conjugation of the matrix coefficients. Thus, we have
(cf.\ \cite{Shin75}):
\begin{equation*}
\begin{split}
\rho_D^*(T)\,\ebase_\gamma & = e\bigl(-\Qf{\gamma} \bigr)\ebase_\gamma, \\
\rho_D^*(S)\,\ebase_\gamma & = \frac{\sqrt{i}^{-2n}}{\sqrt{\abs{D'/D}}}
\sum_{\delta \in D'/D} e\bigl( \blf{\gamma}{\delta} \bigr) \ebase_\delta,
\end{split}
\end{equation*}
where $(\ebase_\gamma)_{\gamma\in D'/D}$ is the standard basis for the group algebra $\mathbb{C}[D'/D]$,
and $\blfempty$ is the bilinear form on $V_\mathbb{Q}$ given by $\blfempty\vcentcolon= \tr_{\mathbf{k}/\mathbb{Q}}\hlf{\cdot}{\cdot}$.
\begin{definition}
For $k\in\mathbb{Z}$, a function $f:\, \mathbb{H} \rightarrow \mathbb{C}[D'/D]$ is called a vector valued modular form of weight $k$ with respect to
$\rho_D^*$ if
\begin{enumerate}
\item $f( A\tau) = \left(c\tau + d\right)^{k} \rho_D^*(A) f(\tau)$ for all $A = \left( \begin{smallmatrix} a & b \\ c & d \end{smallmatrix}\right)\in\group{SL}_2(\mathbb{Z})$.
\item $f$ is holomorphic on $\mathbb{H}$,
\item $f$ is holomorphic at the cusp $i\infty$.
\end{enumerate}
\end{definition}
Here $\group{SL}_2(\mathbb{Z})$ acts on $\mathbb{H}$ as usual.
Thus, the first condition implies the existence of a Fourier expansion:
\[
f(\tau) = \sum_{\gamma \in D'/D} \sum_{m\in \mathbb{Z} - \QfNop(\gamma)}
a(\gamma, m) e(m\tau) \ebase_\gamma.
\]
The second condition means that all coefficients with $m<0$ vanish.
If $a(\gamma, m)=0$ for all $m\leq 0$, then $f$ is called a \emph{cusp form}.
We denote the space of cusp forms of weight $k$ transforming under $\rho_D^*$ by $\Cuspf_k(\rho_D^*)$.
In the following, set $k = n +2$. We will define certain $\mathbb{C}[D'/D]$-valued cusp forms in $\Cuspf_k(\rho_D^*)$
using theta-series with harmonic polynomials as coefficients:
\begin{equation}\label{eq:type_theta}
\Theta_p(\tau, v) = \sum_{\lambda \in D'} p(\lambda, v) e( - \Qf{\lambda}\tau)
\ebase_\lambda,
\end{equation}
for fixed $v \in W_\mathbb{C}$. If $p(\lambda, v)$ is harmonic in $\lambda$ and homogeneous of degree two, the theta series is a cusp form in $\Cuspf_k(\rho_D^*)$.
This is a well-known result in theory of theta functions which can be proved through Poisson-summation, see e.g.\ \citep[Theorem 4.1]{Bo98}.
The polynomials in question are obtained from the torsion condition in our main result, Theorem \ref{thm:H_torsCond}. This will allow us to identify
a space of cusp forms as the set of obstructions against the local Heegner $\mathbf{H}$ being torsion.
We rewrite \eqref{eq:tors_H} using polynomials $p_1(u,v,w), p_2(u,v,w)\in \mathbb{R}[u,v,w]$ defined as follows:
\begin{gather}\label{eq:tors_Hpoly}
\sum_{\beta \in \mathcal{L}}\; \sum_{\substack{m \in \mathbb{Z} + \QfNop(\beta) \\ m < 0 }} c(\beta, m)
\sum_{\substack{\lambda \in D' \\ \lambda + D \equiv\pi(\beta) \\
\QfNop(\lambda) = m}} \left[ p_1(\lambda, t, t') + i p_2(\lambda, t, t') \right] = 0, \\
\nonumber
\begin{aligned}
\qquad \text{with}\quad &
p_1(u,v,w) \vcentcolon = & \Re F_u(v,w) - \frac{\Qf{u}}{n} \Re\hlf{v}{w}, \\
& p_2(u,v,w) \vcentcolon = & \Im F_u(v,w) - \frac{\Qf{u}}{n} \Im\hlf{v}{w}.
\end{aligned}
\end{gather}
We note that for the real part of \eqref{eq:tors_H} to hold, by linearity, it suffices to verify for $t=t'$. Consequently, we set
\[
P(u,v) \vcentcolon = p_1(u,v,v) = 2\left(\Re\hlf{u}{v}\right)^2 - \frac{\Qf{u}}{n} \Qf{v}.
\]
It is easily seen that both $p_1$ and $p_2$ can be obtained from $P$ using the polarization identity, for example
\[
p_2(u,v,w) = \frac12\left( P(u,v) + P(u, -iw) - P(u, v - iw) \right).
\]
We also note that these polynomials are all harmonic and homogeneous in $u$. In fact, $P$ is harmonic in both indeterminates $u$ and $v$ and also homogeneous of the correct degree.
Thus, in particular, for every $v\in W_\mathbb{C}$, the theta series $\Theta_P(\tau, v)$ is a cusp form transforming under $\rho_D^*$ with the desired weight $k$. We rewrite \eqref{eq:type_theta} slightly to obtain the Fourier expansion of $\Theta_P(\tau, v)$:
\[
\Theta_P(\tau, v) = \sum_{\gamma \in D'/D} \sum_{\substack{ m\in \mathbb{Z} -
\Qf{\gamma} \\ m<0}} \biggl(
\sum_{\substack{\lambda \in D' \\ \lambda + D
\equiv\gamma \\ \QfNop(\lambda) = m}} P(\lambda,v) \biggr)
\cdot e( -m\tau) \ebase_\gamma.
\]
Now, the Fourier coefficients are precisely the real part of the inner sums in \eqref{eq:tors_Hpoly},
restricted to the diagonal with $t=t'=v$.
As $v$ varies over $W_\mathbb{C}$, these theta-series $\Theta_P(\tau,v)$ span a subspace of $\Cuspf_k(\rho_D^*)$ which we denote as $\Cuspf_k^\Theta(\rho_D^*)$.
We remark that the polynomials $p_1(\lambda, v,w)$ and $p_2(\lambda; v, w)$ also define theta series, but these are already contained in $\Cuspf_k^\Theta(\rho_D^*)$.
With these considerations, Theorem \ref{thm:H_torsCond} can be restated using modular forms.
\begin{theorem}\label{thm:Obst}
A finite linear combination of Heegner divisors
\[
\mathbf{H} = \frac12
\sum_{\beta \in \mathcal{L}}\sum_{\substack{m \in \mathbb{Z} + \QfNop(\beta) \\ m <0 }}
c(\beta, m) \mathbf{H}_\ell(\beta, m),
\]
with integer coefficients $c(\beta, m)$ satisfying $c(\beta, m) = c(\beta,-m)$ is a torsion element in the
Picard group $\mathrm{Pic}\left(\Gamma_\ell\backslashU_{\epsilon}(\ell)\right)$ if and only if
\[
\sum_{\beta \in \mathcal{L}} \sum_{\substack{m \in \mathbb{Z} + \Qf{\beta} \\ m<0}} c(\beta,
m) a(\pi(\beta), -m) = 0
\]
for every cusp form $f = \sum_{\gamma \in D'/D} \sum_{m\in \mathbb{Z} - \QfNop(\gamma)} a(\gamma, m) e( - m\tau) \ebase_\gamma \in
\Cuspf_k^\Theta(\rho_D^*)$.
\end{theorem}
\subsection{Relationship to global obstruction theory and the work of Bruinier and Freitag} \label{subsec:localglobal}
Since the statement of Theorem \ref{thm:Obst} holds for all sufficiently small $\epsilon$, passing to the direct limit we get the statement for the local Picard group at the cusp $\ell$. Now Theorem \ref{thm:Obst} formally resembles a global obstruction statement for unitary groups from \cite{Hof14} in the style of Borcherds \cite{Bo99}.
It can be states as follows (by \cite[Lemma 5, Theorem 4]{Hof14}):
\begin{theorem}[From \cite{Hof14}] \label{thm:obstglob}
A Heegner divisor of the form
\[
\mathbf{H} = \frac12 \sum_{\beta \in L'/L}\sum_{\substack{m \in \mathbb{Z} + \QfNop(\beta) \\ m <0 }} c(\beta,m) \mathbf{H}(\beta, m)
\]
is the divisor of a Borcherds product if and only if
\[
\sum_{\beta \in L'/L} \sum_{\substack{m \in \mathbb{Z} + \QfNop({\beta}) \\ m<0}}
c(\beta, m) b(\beta, -m) = 0,
\]
for every cusp form $g\in \Cuspf_k(\rho_L^*)$ with Fourier coefficients $b(\beta, -m)$.
\end{theorem}
Since by results of Bruinier \cite{Br02}, the local obstruction space $\Cuspf_k(\rho_D^*)$ can be embedded into the global obstruction space $\Cuspf_k(\rho_L^*)$ \citep[see also][Section 5]{BrFr}, the global obstruction equation implies the local one. Thus, if $\mathbf{H}$ is the divisor of a Borcherds product, then the local divisor
\[
\mathbf{H}_\ell \vcentcolon = \frac12 \sum_{\beta \in \mathcal{L}}\sum_{\substack{m \in \mathbb{Z} + \QfNop(\beta) \\ m <0 }} c(\beta,m) \mathbf{H}_\ell (\beta, m)
\]
is a torsion element in the local Picard group. In fact, the same argument applies for every cusp.
The results of Bruinier and Freitag, \citep[][Theorem 5.1]{BrFr}, in the setting of orthogonal groups are very similar to Theorem \ref{thm:Obst} above.
The definition of their theta-series is (essentially) the same. Indeed, if we look at the rational quadratic space $V_\mathbb{Q}$ underlying $V_\mathbf{k}$
and the lattices $L$ and $D$ as quadratic modules in $V_\mathbb{Q}$, the obstruction spaces are the same.
In this case, through the embedding theory from \cite{Hof11, Hof14} we can pull back Heegner divisors on the modular variety of the orthogonal group to Heegner divisors for the modular variety of the unitary group. As sketched in Remark \ref{rmk:BrFr_thm} above, this also works locally.
Thus if $H$ is the Heegner divisor of a Borcherds product for the orthogonal group, then it is \emph{trivial at generic boundary components} in the sense of \citep[][Definition 5.3]{BrFr}, i.e.\ locally torsion, and hence restricts to a torsion element in the local Picard group for every cusp $I$ of the unitary modular variety. Similarly, by pulling back the Borcherds product itself, one gets a Borcherds product for the unitary group with the pull-back of $H$ as its divisor, and through Theorem \ref{thm:obstglob}, again, the corresponding local Heegner divisors are torsion elements.
We also remark that the obstruction space in Theorem \ref{thm:obstglob} is the same as that from Borcherds' \cite{Bo99} in the orthogonal situation. Hence, if $\mathbf{H}$ is the Heegner divisor of a Borcherds product for the unitary group, one can find a Heegner divisor on the orthogonal side which restricts to $\mathbf{H}$ and is the divisor of a Borcherds product.
\begin{rmk} In \cite[][Theorem 5.4]{BrFr}, they were able to show that for a unimodular lattice $L$, the
triviality of a Heegner divisor at generic boundary components, conversely, implies the global obstruction equation of Borcherds from \cite{Bo99} and hence the existence of a Borcherds product for a Heegner divisor that kills all local obstructions.
Their argument depends on two results: The uniqueness of isomorphism classes of unimodular lattice, and a result of Waldspurger \citep[see][]{Wald78} on the generation of the space $\Cuspf_k(\rho_L^*)$ by theta-series for definite lattices.
Unfortunately, there is no obvious way to transfer this argument to hermitian lattices, since given a quadratic module over $\mathbb{Z}$ a complex structure need neither exist, nor need it be unique.
\end{rmk}
\textit{Acknowledegements:}\ I wish to thank Jan Bruinier for many helpful discussions and for his encouragement in the course of working on this paper.
Further, I would like to acknowledge the role of Prof. Freitag: His insights into group cohomology have helped me considerably to start out with this project.
Finally, I want to thank the anonymous referee, whose comments have led to enormous improvements in this paper.
\bibliographystyle{hplain}
|
\section{Introduction}
Applications of topology to scientific domains outside of pure mathematics are becoming increasingly common. Neuroscience, a field undergoing a golden age of progress in its own right, is no exception. The first reason for this is perhaps obvious -- at least to anyone familiar with topological data analysis. Like other areas of biology, neuroscience is generating a lot of new data, and some of these data can be better understood with the help of topological methods. A second reason is that a significant portion of neuroscience research involves studying networks, and networks are particularly amenable to topological tools. Although my talk will touch on a variety of such applications, most of my attention will be devoted to a third reason -- namely, that many interesting problems in neuroscience contain topological questions in disguise. This is especially true when it comes to understanding {\it neural codes}, and questions such as: how
do the collective activities of neurons represent information about the outside world?
I will begin this talk with some well-known examples of neural codes, and then use them to illustrate how topological ideas naturally arise in this context. Next, I'll take a brief detour to describe other uses of topology in neuroscience. Finally, I will return to neural codes and explain why topological methods are helpful for studying their intrinsic properties. Taken together, these developments suggest that topology is not only useful for analyzing neuroscience data, but may also play a fundamental role in the theory of how the brain works.
\section{Neurons: nodes in a network or autonomous sensors?}
It has been known for more than a century, since the time of Golgi and Ramon y Cajal, that the neurons in our brains are connected to each other in vast, intricate networks. Neurons are electrically active cells. They communicate with each other by firing action potentials (spikes) -- tiny messages that are only received by neighboring (synaptically-connected) neurons in the network. Suppose we were eavesdropping on a single neuron, carefully recording its electrical activity at each point in time. What governs the neuron's behavior? The obvious answer: it's the network, of course!
If we could monitor the activity of all the other neurons, and we knew exactly the pattern of connections between them, and were blessed with an excellent model describing all relevant dynamics, then (maybe?) we would be able to predict when our neuron will fire. If this seems hopeless now, imagine how unpredictable the activity of a single neuron in a large cortical network must have seemed in the 1950s, when Hodgkin and Huxley had just finished working out the complex nonlinear dynamics of action potentials for a simple, isolated cell \cite{rinzel1990HodgkinHuxley}.
And yet, around 1959, a miracle happened. It started when Hubel and Wiesel inserted a microelectrode into the primary visual cortex of an anesthetized cat, and eavesdropped on a single neuron.
They could neither monitor nor control the activity of any other neurons in the network -- they could only listen to one neuron at a time. What they {\it could} control was the visual stimulus. In an attempt to get the neuron to fire, they projected black and white patterns on a screen in front of the open-eyed cat. Remarkably, they found that the neuron they were listening to fired rapidly when the screen showed a black bar at a certain angle -- say, 45$^\circ$. Other neurons responded to different angles. It was as though each neuron was a sensor for a particular feature of the visual scene. Its activity could be predicted without knowing anything about the network, but by simply looking {\it outside} the cat's brain -- at the stimulus on the screen.
Hubel and Wiesel had discovered orientation-tuned neurons \cite{HubelWiesel59}, whose collective activity comprises a {\it neural code} for angles in the visual field (see Figure~\ref{fig:net-RF}B). Although they inhabit a large, densely-connected cortical network, these neurons do not behave as unpredictable units governed by complicated dynamics. Instead, they appear to be responding directly to stimuli in the outside world. Their activity has {\it meaning}.
\begin{figure}[!h]
\includegraphics[width = 5in]{Network_vs_RFs_CC-eps-converted-to.pdf}
\caption{\small The neural network and neural coding pictures. (A) Pyramidal neurons (triangles) are embedded in a recurrent network together with inhibitory interneurons (circles). (B) An orientation-tuned neuron in primary visual cortex with a preferred angle of 45$^\circ$. The neuron fires many spikes in response to a bar at a 45$^\circ$ angle in the animal's visual field, but few spikes in response to a horizontal bar. (C) Place cells in the hippocampus fire when the animal passes through the corresponding place field. The activity of three different neurons is shown (top), while the animal traces a trajectory starting at the top left corner of its environment (bottom). Each neuron's activity is highest when the animal passes through the corresponding place field (shaded disc).}
\label{fig:net-RF}
\end{figure}
A decade later, O'Keefe made a similar discovery, this time involving neurons in a different area of the brain -- the hippocampus. Unlike the visual cortex, there is no obvious sensory pathway to the hippocampus. This made it all the more mysterious when O'Keefe reported that his neurons were responding selectively to different locations in the animal's physical environment \cite{OKeefe}. These neurons, dubbed {\it place cells}, act as position sensors in space.
When an animal is exploring a particular environment, a place cell increases its firing rate as the animal passes through its corresponding {\it place field} -- that is, the localized region to which the neuron preferentially responds (see Figure~\ref{fig:net-RF}C).
Like Hubel and Wiesel, who received a Nobel prize for their work in 1981 \cite{nobelprize81}, O'Keefe's discovery of place cells had an enormous impact in neuroscience. In 2014, he shared the Nobel prize with Edvard and May-Britt Moser \cite{nobelprize14}, former postdocs of his who went on to discover an even stranger class of neurons that encode position, in a neighboring area of hippocampus called the entorhinal cortex. These neurons, called {\it grid cells}, display periodic place fields that are arranged in a hexagonal lattice. We'll come back to grid cells in the next section.
So, are neurons nodes in a network? or autonomous sensors of the outside world?
Both pictures are valid, and yet they lead to very different models of neural behavior. Neural network theory deals with the first picture, and seeks to understand how the activity of neurons emerges from properties of the network. In contrast, neural coding theory often treats the network as a black box, focusing instead on the relationship between neural activity
and external stimuli. Many of the most interesting problems in neuroscience are about {\it understanding the neural code}. This includes, but is not limited to, figuring out the basic principles by which neural activity represents sensory inputs to the eyes, nose, ears, whiskers, and tongue. Because of the discoveries of Hubel and Wiesel, O'Keefe, and many others, we often know more about the coding properties of single neurons than we do about the networks to which they belong.
But many open questions remain. And topology, as it turns out, is a natural tool for understanding the neural code.
\section{Topology of hippocampal place cell codes}
The term {\it hippocampal place cell code} refers to the neural code used by place cells in the hippocampus to encode the animal's position in space.
Most of the research about place cells, including O'Keefe's original discovery, has been performed in rodents (typically rats), and the experiments typically involve an animal moving around in a restricted environment (see Figure~\ref{fig:net-RF}C). It was immediately understood that a population of place cells, each having a different place field, could collectivity encode the animal's position in space \cite{OKeefe-book}, even though for a long time electrophysiologists could only monitor one neuron at a time. When simultaneous recordings of place cells became possible, it was shown via statistical inference (using previously measured place fields) that the animal's position could indeed be inferred from population place cell activity \cite{brown1998statistical}. Figure~\ref{fig:place-fields} shows four place fields corresponding to simultaneously recorded place cells in area CA1 of rat hippocampus.
\begin{figure}[!h]
\includegraphics[width = 5in]{placefields-eps-converted-to.pdf}
\caption{\small Place fields for four place cells, recorded while a rat explored a 2-dimensional square box environment. Place fields were computed from data provided by the Pastalkova lab.}
\label{fig:place-fields}
\end{figure}
The role of topology in place cell codes begins with a simple observation, which is perhaps obvious to anyone familiar with both place fields in neuroscience and elementary topology.
First, let's recall the standard definitions of an open cover and a good cover.
\begin{definition}
Let $X$ be a topological space. A collection of open sets, $\mathcal{U} = \{U_1,\ldots,U_n\}$, is an {\it open cover} of $X$ if $X = \bigcup_{i =1}^n U_i.$ We say that $\mathcal{U}$ is a {\it good cover} if every non-empty intersection $\bigcap_{i \in \sigma} U_i$, for $\sigma \subseteq \{1,\ldots, n\}$, is contractible.
\end{definition}
Next, observe that a collection of place fields in a particular environment looks strikingly like an open cover, with each $U_i$ corresponding a place field.
Figure~\ref{fig:environments} displays three different environments, typical of what is used in hippocampal experiments with rodents, together with schematic arrangements of place fields in each.
\begin{figure}[!h]
\includegraphics[width = 5in]{RF-topology-eps-converted-to.pdf}
\caption{\small Three environments for a rat: (A) A square box environment, also known as an ``open field''; (B) an environment with a hole or obstacle in the center;
and (C) a maze with two arms. Each environment displays a collection of place fields (shaded discs) that fully cover the underlying space.}
\label{fig:environments}
\end{figure}
Moreover, since place fields are approximately convex (see Figure~\ref{fig:place-fields}) it is not unreasonable to assume that they form a good cover of the underlying space.
This means the Nerve Lemma applies. Recall the notion of the {\it nerve}\footnote{Note that the name ``nerve'' here predated any connection to neuroscience!} of a cover:
$$\mathcal{N}(\mathcal{U}) \stackrel{\mathrm{def}}{=} \{\sigma \subset [n] \mid \bigcap_{i \in \sigma} U_i \neq \emptyset \},$$
where $[n] = \{1,\ldots,n\}$. Clearly, if $\sigma \in \mathcal{N}(\mathcal{U})$ and $\tau \subset \sigma$, then $\tau \in \mathcal{N}(\mathcal{U})$. This property shows that $\mathcal{N}(\mathcal{U})$ is an abstract {\it simplicial complex} on the vertex set $[n]$ -- that is, it is a set of subsets of $[n]$ that is closed under taking further subsets. If $X$ is a sufficiently ``nice'' topological space, then the following well-known lemma holds.
\begin{lemma}[Nerve Lemma]\label{lemma:nerve}
Let $\mathcal{U}$ be a good cover of $X$. Then $\mathcal{N}(\mathcal{U})$ is homotopy-equivalent to $X$. In particular, $\mathcal{N}(\mathcal{U})$ and $X$ have exactly the same homology groups.
\end{lemma}
It is important to note that the Nerve Lemma fails if the good cover assumption does not hold. Figure~\ref{fig:good-bad-cover}A depicts a good cover of an annulus by three open sets. The corresponding nerve (right) exhibits the topology of a circle, which is indeed homotopy-equivalent to the covered space. In Figure~\ref{fig:good-bad-cover}B, however, the cover is {\it not} good, because the intersection $U_1 \cap U_2$ consists of two disconnected components, and is thus not contractible. Here the nerve (right) is homotopy-equivalent to a point, in contradiction to the topology of the covered annulus.
\begin{figure}[!h]
\includegraphics[width=4.5in]{good-bad-cover-eps-converted-to.pdf}
\caption{Good and bad covers. (A) A good cover $\mathcal{U} = \{U_1, U_2, U_3\}$ of an annulus (left), and the corresponding nerve $\mathcal{N}(\mathcal{U})$ (right). (B) A ``bad'' cover of the annulus (left), and the corresponding nerve (right). Only the nerve of the good cover accurately reflects the topology of the annulus.}
\label{fig:good-bad-cover}
\end{figure}
The wonderful thing about the Nerve Lemma, when interpreted in the context of hippocampal place cells, is that $\mathcal{N}(\mathcal{U})$ can be inferred from the activity of place cells alone -- without actually knowing the place fields $\{U_i\}$. This is because the concurrent activity of a group of place cells, indexed by $\sigma \subset [n]$, indicates that the corresponding place fields have a non-empty intersection: $\bigcap_{i \in \sigma} U_i \neq \emptyset$.
In other words, if we were eavesdropping on the activity of a population of place cells as the animal fully explored its environment, then by finding which subsets of neurons co-fire (see Figure~\ref{fig:spikes2codewords}) we could in principle estimate $\mathcal{N}(\mathcal{U})$, even if the place fields themselves were unknown. Lemma~\ref{lemma:nerve} tells us that the homology of the simplicial complex $\mathcal{N}(\mathcal{U})$ precisely matches the homology of the environment $X$. The place cell code thus naturally reflects the topology of the represented space.\footnote{In particular, place cell activity from the environment in Figure~\ref{fig:environments}B could be used to detect the non-trivial first homology group of the underlying space, and thus distinguish this environment from that of Figure~\ref{fig:environments}A or~\ref{fig:environments}C.}
\begin{figure}[!h]
\includegraphics[width = 1.5in]{spikes2codewords-eps-converted-to.pdf}
\caption{\small By binning spike trains for a population of simultaneously-recorded neurons, one can infer subsets of neurons that co-fire. If these neurons were place cells,
then the first codeword 1110 indicates that $U_1\cap U_2 \cap U_3 \neq \emptyset$, while the third codeword 0101 tells us $U_2 \cap U_4 \neq \emptyset$.}
\label{fig:spikes2codewords}
\end{figure}
These and related observations have led some researchers to speculate that the hippocampal place cell code is fundamentally topological in nature \cite{dabaghian2012topological, chen2014neural}, while others (including this author) have argued that considerable geometric information is also present and can be extracted using topological methods \cite{curto2008cell, giusti2015clique}. In order to disambiguate topological and geometric features, Dabaghian et. al. performed an elegant experiment using linear tracks with flexible joints \cite{dabaghian2014reconceiving}. This allowed them to alter geometric features of the environment, while preserving the topological structure as reflected by the animal's place fields. They found that place fields recorded from an animal running along the morphing track moved together with the track, preserving the relative sequence of locations despite changes in angles and movement direction. In other words, the place fields respected topological aspects of the environment more than metric features \cite{dabaghian2014reconceiving}.
\begin{figure}[!h]
\includegraphics[width = 5in]{grid-cell-topology-eps-converted-to.pdf}
\caption{\small Firing fields for grid cells. (A) Firing fields for four entorhinal grid cells. Each grid field forms a hexagonal grid in the animal's two-dimensional environment, and each field thus has multiple disconnected regions. (B) A hexagonal fundamental domain contains just one disc-like region per grid cell. Pairs of edges with the same label (a, b, or c) are identified, with orientations specified by the arrows.}
\label{fig:grid-cells}
\end{figure}
What about the entorhinal grid cells? These neurons have firing fields with multiple disconnected components, forming a hexagonal grid (see Figure~\ref{fig:grid-cells}A). This means that grid fields violate the good cover assumption of the Nerve Lemma -- if we consider them as an open cover for the entire 2-dimensional environment. If, instead, we restrict
attention to a fundamental domain for these firing fields, as illustrated in Figure~\ref{fig:grid-cells}B, then each grid field has just one (convex) component, and the Nerve Lemma applies. From the spiking activity of grid cells we could thus infer the topology of this fundamental domain. The reader familiar with the topological classification of surfaces may recognize that this hexagonal domain, with the identification of opposite edges, is precisely a torus. To see this, first identify the edges labeled ``a'' to get a cylinder. Next, observe that the boundary circles on each end of the cylinder consist of the edges ``b'' and ``c'', but with a 180$^\circ$ twist between the two ends. By twisting the cylinder, the two ends can be made to match so that the ``b'' and ``c'' edges get identified. This indicates that the space represented by grid cells is not the full environment, but a torus.
\section{Topology in neuroscience: a bird's-eye view}
The examples from the previous section are by no means the only way that topology is being used in neuroscience. Before plunging into further details about what topology can tell us about neural codes, we now pause for a moment to acknowledge some other interesting applications. The main thing they all have in common is their recency. This is no doubt due to the rise of computational and applied algebraic topology, a relatively new development in applied math that was highlighted in the Current Events Bulletin nearly a decade ago
\cite{ghrist2007barcodes}.
Roughly speaking, the uses of topology in neuroscience can be categorized into three (overlapping) themes: (i) ``traditional'' topological data analysis applied to neuroscience; (ii) an upgrade to network science; and (iii) understanding the neural code. Here we briefly summarize work belonging to (i) and (ii). In the next section we'll return to (iii), which is the main focus of this talk.
\subsection{``Traditional'' TDA applied to neuroscience data sets.} The earliest and most familiar applications of topological data analysis (TDA) focused on the problem of estimating the ``shape'' of point-cloud data. This kind of data set is simply a collection of points, $x_1,\ldots,x_\ell \in \mathbb{R}^n$, where $n$ is the dimensionality of the data. A question one could ask is: do these points appear to have been sampled from a lower-dimensional manifold, such as a torus or a sphere? The strategy is to consider open balls $B_\varepsilon(x_i)$ of radius $\varepsilon$ around each data point, and then to construct a simplicial complex $\mathcal{K}_\varepsilon$ that captures information about how the balls intersect. This simplicial complex can either be the Cech complex (i.e., the nerve of the open cover defined by the balls), or the Vietoris-Rips complex (i.e., the clique complex of the graph obtained from pairwise intersections of the balls). By varying $\varepsilon$, one obtains a sequence of nested simplicial complexes $\{\mathcal{K}_\varepsilon\}$ together with natural inclusion maps. Persistent homology tracks homology cycles across these simplicial complexes, and allows one to determine whether there were homology classes that ``persisted'' for a long time. For example, if the data points were sampled from a 3-sphere, one would see a persistent 3-cycle.
There are many excellent reviews of persistent homology, including \cite{ghrist2007barcodes}, so I will not go into further details here. Instead, it is interesting to note that one of the early applications of these techniques was in neuroscience, to analyze population activity in primary visual cortex \cite{singh2008topological}. Here it was found that the topological structure of activity patterns is similar between spontaneous and evoked activity, and consistent with the topology of a two-sphere. Moreover, the results of this analysis were interpreted in the context of neural coding, making this work exemplary of both themes (i) and (iii). Another application of persistent homology to point cloud data in neuroscience was the analysis of the spatial structure of afferent neuron terminals in crickets \cite{brown2012structure}. Again, the results were interpreted in terms of the coding properties of the corresponding neurons, which are sensitive to air motion detected by thin mechanosensory hairs on the cricket. Finally, it is worth mentioning that these types of analyses are not confined to neural activity. For example, in \cite{bendich2014persistent} the statistics of persistent cycles were used to study brain artery trees.
\subsection {An upgrade to network science.} There are many ways of constructing networks in neuroscience, but the basic model that has been used for all of them is the graph. The vertices of a graph can represent neurons, cell types, brain regions, or fMRI voxels, while the edges reflect interactions between these units. Often, the graph is weighted and the edge weights correspond to correlations between adjacent nodes. For example, one can model a functional brain network from fMRI data as a weighted graph where the edge weights correspond to activity correlations between pairs of voxels. At the other extreme, a network where the vertices correspond to neurons could have edge weights that reflect either pairwise correlations in neural activity, or synaptic connections.
Network science is a relatively young discipline that focuses on analyzing networks, primarily using tools derived from graph theory. The results of a particular analysis could range from determining the structure of a network to identifying important subgraphs and/or graph-theoretic statistics (the distribution of in-degree or out-degree across nodes, number of cycles, etc.) that carry meaning for the network at hand. Sometimes, graph-theoretic features do not carry obvious meaning, but are nevertheless useful for distinguishing networks that belong to distinct classes. For example, a feature could be characteristic of functional brain networks derived from a subgroup of subjects, distinguishing them from a ``control'' group.
In this way graph features may be a useful diagnostic tool for distinguishing diseased states, pharmacologically-induced states, cognitive abilities, or uncovering systematic differences based on gender or age.
The recent emergence of topological methods in network science stems from the following ``upgrade'' to the network model: instead of a graph, one considers a simplicial complex. Sometimes this simplicial complex reflects higher-order interactions that are obtained from the data, and sometimes it is just the {\it clique complex} of the graph $G$:
$$X(G) = \{\sigma \subset [n] \mid (ij) \in G \text{ for all } i,j \in \sigma\}.$$
In other words, the higher-order simplices correspond to cliques (all-to-all connected subgraphs) of $G$.
Figure~\ref{fig:graph-cliques}A shows a graph (top) and the corresponding clique complex (bottom), with shaded simplices corresponding to two 3-cliques and a 4-clique. The clique complex fills in many of the 1-cycles in the original graph, but some 1-cycles remain (see the gold 4-gon), and higher-dimensional cycles may emerge. Computing homology groups for the clique complex is then a natural way to detect topological features that are determined by the graph. In the case of a weighted graph, one can obtain a sequence of clique complexes by considering a related sequence of simple graphs, where each graph is obtained from the previous one by adding the edge corresponding to the next-highest weight (see Figure~\ref{fig:graph-cliques}B). The corresponding sequence of clique complexes, $\{X(G_i)\}$, can then be analyzed using persistent homology. Other methods for obtaining a sequence of simplicial complexes from a network are also possible, and may reflect additional aspects of the data such as the temporal evolution of the network.
\begin{figure}[!h]
\includegraphics[width=4.5in]{graph-cliques-filtration-eps-converted-to.pdf}
\caption{Network science models: from graphs to clique complexes and filtrations.}
\label{fig:graph-cliques}
\end{figure}
For a more thorough survey of topological methods in network science, I recommend the forthcoming review article \cite{giusti2015two}. Here I will only mention
that topological network analyses have already been used in a variety of neuroscience applications, many of them medically-motivated: fMRI networks in patients with ADHD \cite{ellis2014describing}; FDG-PET based networks in children with autism and ADHD \cite{lee2011discriminative}; morphological networks in deaf adults \cite{kim2014morphological}; metabolic connectivity in epileptic rats \cite{choi2014abnormal}; and functional EEG connections in depressed mice \cite{khalid2014tracing}.
Other applications to fMRI data include human brain networks during learning \cite{stolz2014computational} and drug-induced states \cite{petri2014homological}. At a finer scale, recordings of neural activity can also give rise to functional connectivity networks among neurons (which are not the same as the neural networks defined by synaptic connections). These networks have also been analyzed with topological methods \cite{pirino2014topological, giusti2015clique, spreemann2015using}.
\section{The code of an open cover}
We now return to neural codes. We have already seen how the hippocampal place cell code reflects the topology of the underlying space, via the nerve $\mathcal{N}(\mathcal{U})$ of a place field cover. In this section, we will associate a binary code to an open cover. This notion is closer in spirit to a combinatorial neural code (see Figure~\ref{fig:spikes2codewords}), and carries more detailed information than the nerve. In the next section, we'll see how topology is being used to determine intrinsic features of neural codes, such as convexity and dimension.
First, a few definitions. A {\it binary pattern} on $n$ neurons is a string of $0$s and $1s$, with a $1$ for each active neuron and a $0$ denoting silence;
equivalently, it is a subset of (active) neurons
$\sigma \subset [n].$ (Recall that $[n] = \{1,\ldots,n\}$.)
We use both notations interchangeably. For example, $10110$ and $\sigma = \{1,3,4\}$ refer to the same pattern, or codeword, on $n = 5$ neurons. A {\it combinatorial neural code} on $n$ neurons is a collection of binary patterns $\mathcal{C} \subset 2^{[n]}$. In other words, it is a binary code of length $n$, where we interpret each binary digit as the ``on'' or ``off'' state of a neuron.
The {\it simplicial complex of a code}, $\Delta(\mathcal{C})$, is the smallest abstract simplicial complex on $[n]$ that contains all elements of $\mathcal{C}$. In keeping with the hippocampal place cell example, we are interested in codes that correspond to open covers of some topological space.
\begin{figure}[!h]
\includegraphics[width=5in]{cover-code-nerve-eps-converted-to.pdf}
\caption{Codes and nerves of open covers. (A) An open cover $\mathcal{U}$, with each region carved out by the cover labeled by its corresponding codeword. (B) The code $\mathcal{C}(\mathcal{U})$. (C) The nerve $\mathcal{N}(\mathcal{U})$.}
\label{fig:code-cover}
\end{figure}
\begin{definition}
Given an open cover $\mathcal{U}$, the {\it code of the cover} is the combinatorial neural code
$$\mathcal{C}(\mathcal{U}) \stackrel{\mathrm{def}}{=} \{ \sigma \subset [n] \mid \bigcap_{i \in \sigma} U_i \setminus \bigcup_{j \in [n] \setminus \sigma} U_j \neq \emptyset \}.$$
\end{definition}
Each codeword in $\mathcal{C}(\mathcal{U})$ corresponds to a region that is defined by the intersections of the open sets in $\mathcal{U}$ (Figure~\ref{fig:code-cover}A).
Note that the code $\mathcal{C}(\mathcal{U})$ is not the same as the nerve $\mathcal{N}(\mathcal{U})$. Figures~\ref{fig:code-cover}B and ~\ref{fig:code-cover}C display the code and the nerve of the open cover in Figure~\ref{fig:code-cover}A. While the nerve encodes which subsets of the $U_i$s have non-empty intersections, the code
also carries information about set containments. For example, the fact that $U_2 \subseteq U_1 \cup U_3$ can be inferred from $\mathcal{C}(\mathcal{U})$ because each codeword of the form
$*1**$ has an additional $1$ in position 1 or 3, indicating that if neuron 2 is firing then so is neuron 1 or 3. Similarly, the fact that $U_2 \cap U_4 \subseteq U_3$ can be
inferred from the code because any word of the form $*1*1$ necessarily has a 1 in position 3 as well. These containment relationships go beyond simple intersection data, and cannot be obtained from the nerve $\mathcal{N}(\mathcal{U})$. On the other hand, the nerve can easily be recovered from the code since $\mathcal{N}(\mathcal{U})$ is the smallest simplicial complex that contains it -- that is,
$$\mathcal{N}(\mathcal{U}) = \Delta(\mathcal{C}(\mathcal{U})).$$
$\mathcal{C}(\mathcal{U})$ thus carries more detailed information than what is available in $\mathcal{N}(\mathcal{U})$. The combinatorial data in $\mathcal{C}(\mathcal{U})$ can also be encoded algebraically via the {\it neural ideal} \cite{neural_ring}, much as simplicial complexes are algebraically encoded by Stanley-Reisner ideals \cite{miller-sturmfels}.
It is easy to see that any binary code, $\mathcal{C} \subseteq \{0,1\}^n$, can be realized as the code of an open cover.\footnote{For example,
if the size of the code is $|\mathcal{C}| = \ell$, we could choose disjoint open intervals $B_1, \ldots, B_\ell \subset \mathbb{R}$, one for each codeword, and define
the open sets $U_1,\ldots,U_n$ such that $U_i$ is the union of all open intervals $B_j$ corresponding to codewords in which neuron $i$ is ``on'' (that is, there is a $1$ in position $i$ of the codeword). Such a cover, however, consists of highly disconnected sets and its properties reflect very little of the underlying space -- in particular, the good cover assumption of the Nerve Lemma is violated.} It is not true, however, that any code can arise from a good cover or a {\it convex cover} -- that is, an open cover consisting of convex sets. The following lemma illustrates the simplest example of what can go wrong.
\begin{lemma}\label{lemma:disconnect}
Let $\mathcal{C} \subset \{0,1\}^3$ be a code that contains the codewords $110$ and $101$, but does \underline{not} contain $100$ and $111$. Then $\mathcal{C}$ is not the code of a good or convex cover.
\end{lemma}
The proof is very simple. Suppose $\mathcal{U} = \{U_1, U_2, U_3\}$ is a cover such that $\mathcal{C} = \mathcal{C}(\mathcal{U})$. Because neuron $2$ or $3$ is ``on'' in any codeword for which neuron $1$ is ``on,''
we must have that $U_1 \subset U_2 \cup U_3$. Moreover, we see from the code that $U_1 \cap U_2 \neq \emptyset$ and $U_1 \cap U_3 \neq \emptyset$, while $U_1 \cap U_2 \cap U_3 = \emptyset$. This means we can write $U_1$ as a disjoint union of two non-empty sets: $U_1 = (U_1 \cap U_2) \cup (U_1 \cap U_3).$ $U_1$ is thus disconnected, and hence $\mathcal{U}$ can be neither a good nor convex cover.
\section{Using topology to study intrinsic properties\\ of neural codes}
In our previous examples from neuroscience, the place cell and grid cell codes can be thought of as arising from convex sets covering an underlying space.
Because the spatial correlates of these neurons are already known, it is not difficult to infer what space is being represented by these codes. What could we say
if we were given just a code, $\mathcal{C} \subset \{0,1\}^n$, without {\it a priori} knowledge of what the neurons were encoding? Could we tell whether such a code
can be realized via a convex cover?
\subsection{What can go wrong.}
As seen in Lemma~\ref{lemma:disconnect}, not all codes can arise from convex covers. Moreover, the problem that prevents the code in Lemma~\ref{lemma:disconnect} from being convex is topological in nature. Specifically, what happens in the example of Lemma~\ref{lemma:disconnect} is that the code dictates there must be a set containment,
$$U_\sigma \subseteq \bigcup_{j \in \tau} U_j,$$
where $U_\sigma = \bigcap_{i \in \sigma} U_i$, but the nerve of the resulting cover of $U_\sigma$ by the sets $\{U_\sigma \cap U_j\}_{j \in \tau}$ is not contractible.
This leads to a contradiction if the sets $U_i$ are all assumed to be convex, because the sets $\{U_\sigma \cap U_j\}_{j \in \tau}$ are then also convex and thus form a good cover of $U_\sigma$. Since $U_\sigma$ itself is convex, and the Nerve Lemma holds, it follows that $\mathcal{N}(\{U_\sigma \cap U_j\}_{j \in \tau})$ must be contractible, contradicting the data of the code.
These observations lead to the notion of a {\it local obstruction} to convexity \cite{no-go}, which captures the topological problem that arises if certain codes
are assumed to have convex covers. The proof of the following lemma is essentially the argument outlined above.
\begin{lemma}[\cite{no-go}] \label{lemma:no-obs}
If $\mathcal{C}$ can be realized by a convex cover, then $\mathcal{C}$ has no local obstructions.
\end{lemma}
The idea of using local obstructions to determine whether or not a neural code has a convex realization has been recently followed up in a series of papers \cite{MRC, counterexample, intersection-complete}. In particular, local obstructions have been characterized in terms of links, $\mathrm{Lk}_{\Delta}(\sigma)$, corresponding to ``missing'' codewords that are not in the code, but are elements of the simplicial complex of the code.
\begin{theorem}[\cite{MRC}] \label{thm:loc-obs}
Let $\mathcal{C}$ be a neural code, and let $\Delta = \Delta(\mathcal{C})$. Then $\mathcal{C}$ has no local obstructions if and only if $\mathrm{Lk}_{\Delta}(\sigma)$ is contractible for all $\sigma \in \Delta \setminus \mathcal{C}$.
\end{theorem}
It was believed, until very recently, that the converse of Lemma~\ref{lemma:no-obs} might also be true. However, in \cite{counterexample} the following counterexample was discovered, showing that this is not the case. Here the term {\it convex code} refers to a code that can arise from a convex open cover.
\begin{example}[\cite{counterexample}]
The code $\mathcal{C} = \{2345, 123, 134, 145, 13, 14, 23, 34, 45, 3, 4\}$ is not a convex code, despite the fact that it has no local obstructions.
\end{example}
That this code has no local obstructions can be easily seen using Theorem~\ref{thm:loc-obs}. The fact that there is no convex open cover, however, relies on convexity arguments that are not obviously topological. Moreover, this code does have a good cover \cite{counterexample}, suggesting the existence of a new class of obstructions to convexity which may or may not be topological in nature.
\subsection{What can go right.}
Finally, it has been shown that several classes of neural codes are guaranteed to have convex realizations. {\it Intersection-complete codes} satisfy the property that for any $\sigma, \tau \in \mathcal{C}$ we also have $\sigma \cap \tau \in \mathcal{C}$. These codes (and some generalizations) were shown constructively to have convex covers in \cite{intersection-complete}. Additional classes of codes with convex realizations have been described in \cite{MRC}.
Despite these developments, a complete characterization of convex codes is still lacking. Finding the minimum dimension needed for a convex realization is also an open question.
\section{Codes from networks}
We end by coming back to the beginning.
Even if neural codes give us the illusion that neurons in cortical and hippocampal areas are directly sensing the outside world, we know that of course they are not. Their activity patterns are shaped by the networks in which they reside. What can we learn about the architecture of a network by studying its neural code? This question requires an improved understanding of neural networks, not just neural codes. While many candidate architectures have been proposed to explain, say, orientation-tuning in visual cortex, the interplay of neural network theory and neural coding is still in early stages of development.
Perhaps the simplest example of how the structure of a network can constrain the neural code is the case of simple feedforward networks. These networks have a single input layer of neurons, and a single output layer. The resulting codes are derived from hyperplane arrangements in the positive orthant of $\mathbb{R}^k$, where $k$ is the number of neurons in the input layer and each hyperplane corresponds to a neuron in the output layer (see Figure~\ref{fig:FF-code}). Every codeword in a {\it feedforward code} corresponds to a chamber in such a hyperplane arrangement.
\begin{figure}[!h]
\includegraphics[width=3in]{hyperplanearrangement-eps-converted-to.pdf}
\caption{A hyperplane arrangement in the positive orthant, and the corresponding feedforward code.}
\label{fig:FF-code}
\end{figure}
It is not difficult to see from this picture that all feedforward codes are realizable by convex covers -- specifically, they arise from overlapping half-spaces \cite{no-go}. On the other hand, not every convex code is the code of a feedforward network \cite{vladimir2015personal}. Moreover, the discrepancy between feedforward codes and convex codes is not due to restrictions on their simplicial complexes. As was shown in \cite{no-go}, every simplicial complex can arise as $\Delta(\mathcal{C})$ for a feedforward code. As with convex codes, a complete characterization of feedforward codes is still unknown. It seems clear, however, that topological tools will play an essential role.
\section{Acknowledgments}
I would like to thank Chad Giusti for his help in compiling a list of references for topology in neuroscience.
I am especially grateful to Katie Morrison for her generous help with the figures.
\bibliographystyle{amsplain}
|
\section{Introduction}
Surface structure of solids can be studied using penetrating hard X-rays without detection of the reflected radiation using grazing incidence X-ray photoemission spectroscopy (GIXPS) with angles of incidence in the vicinity of the critical angle for total external reflection. GIXPS was established by Henke~\cite{Henke72} as a method enabling determination of material constants and surface characterization. Further developments were performed \cite{Fadley74} including generalization to multilayer structures \cite{Chester93} followed by experimental effort \cite{Kawai95,Hayashi96}. The studies were mostly concentrated in the soft X-ray domain (photon energies $\lesssim$~5~keV) for applications in surface science. In spite of prior developments, applications of hard X-ray GIXPS remain limited to date~\cite{Fadley05,Fadley10,Kawai10}. Apart from the regime of grazing incidence hard X-ray photoemission spectroscopy has been successfully used to probe electronic properties of materials at depths consistent with the bulk environment (e.g., \cite{Sekiyama00,Gray12,DLFeng11}). Probing depths of about 100~$\rm \AA$ have been demonstrated~\cite{Dallera04}.
Contrary to studies of electronic properties resolving energies of photoelectrons is not essential for probing structure/composition of a multilayer. A substantial simplification of a typical X-ray photoemission spectroscopy setup can be accomplished using self-detection of integral electric charge generated in the exterior of the studied object. This self-detection approach has been utilized in X-ray absorption spectroscopy in the hard X-ray regime (e.g., \cite{Martens78,Martens79,Erbil88}).
For hard X-rays the detection technique can take advantage of negligible X-ray absorption in a light gas environment such as helium \cite{Kordesch84, Guo85}. Instead, helium is subject to efficient ionization by fast photoelectrons escaping the object, which provides enhancement in the quantum detection yield. The same approach is used in conversion electron M{\"{o}}ssbauer spectroscopy (e.g., \cite{Jones78}).
Self-detection of hard-X-ray induced photoelectron yield in the grazing incidence geometry was used recently to study an X-ray mirror enclosed in a flowing helium gas~\cite{Stoupin_APL16}. It was shown that the measured integral photoelectron yield as a function of the incidence angle contains structural information
i.e., Kiessig fringes, which originate from the layered structure of the mirror.
In this letter we show that the structural information extracted from the integral grazing-incidence yield photoemission curves is not limited by the escape depth of photoelectrons but rather is limited by the penetration depth of the X-ray wave. A bi-layer Pt-Cr system on a Si substrate was studied with a deeply buried Cr layer (116.5-$\rm \AA$-thick layer of Cr under 725-$\rm \AA$-thick layer of Pt). Detection of the layered structure using Fourier analysis of the integral photoemission yield is demonstrated at several different photon energies of the incident hard X-rays where the effective photoelectron escape depth is substantially smaller than the depth of the buried layer.
\begin{figure}[h]
\centering\includegraphics[width=0.47\textwidth]{fresnel2.eps}
\caption{Grazing incidence geometry illustrating the incident, the transmitted and the reflected waves. The energy flow through the layered structure generates ionization events due to X-ray photoabsorption. The escaping fast photoelectons (light blue) ionize the surrounding gas while the low-energy secondary photoelectrons (gray) do not participate in this ionization process. While the ionizing photoelectrons may escape only from a limited depth in the top layer the integral photoelectron yield is affected by the total energy flow in the multilayer and thus is sensitive to presence of buried layers.}
\label{fig:gzi}
\end{figure}
The energy flow through a multilayer system (Fig.\ref{fig:gzi}) and thus the number of generated photoelectrons is proportional to the fraction of energy attenuated by the system, $A = 1 - R - T$, where $R$ is the net reflectivity of the system and $T$ is the net transmissivity. If the sample is thick such that the fraction of the transmitted X-rays is negligible, then $A \approx 1 - R$ represents Fresnel transmissivity of the entrance interface.
Attentuation of x rays at depth $z$ within a layer of material having a thickness $dz$ is generally described by the difference between the incoming and outgoing mean energy flow. The energy flow through the layer is represented by the real part of projection $S_z (Q,z)$ of the Poynting vector as a function of the wavevector transfer $Q = 2 k \sin{\alpha}$, where $k$ is the absolute value of the wavevector of the incident wave and $\alpha$ is the glancing angle of incidence. The normalized total X-ray attenuation per unit depth $dz$ is
\begin{equation}
s_n(Q,z) = Re \bigg\{ \frac{1}{S^0_z(Q)}\frac{d S_z(Q,z)}{dz} \bigg\} ,
\end{equation}
where $S^0_z(Q)$ is the projection of the Poynting vector of the incident wave.
The dependence on the photon energy $E_X$ of the incident wave is omitted here for clarity.
For an infinitely thick mirror the penetration depth of the wave transmitted through the entrance interface is given by
\begin{equation}
\Lambda(Q) = \frac{1}{Im(Q_1)},
\label{eq:Lam}
\end{equation}
where
\begin{equation}
Q_1 = \sqrt{Q^2 - 8k^2\delta +i8k^2\beta}
\label{eq:Q1}
\end{equation}
is the wavevector transfer in the mirror material with an index of refraction $n = 1 - \delta +i \beta$. It can be shown that for the thick mirror
\begin{equation}
s_n(Q,z) = \frac{1}{\Lambda(Q)} T(Q) \exp[-z/\Lambda(Q)],
\label{eq:sn_tm}
\end{equation}
where $T (Q) = 1 - R(Q)$ is the Fresnel transmissivity.
If the thickness of the top layer of a multilayer X-ray mirror is such that the transmitted wave is preferentially attenuated in this layer the energy flow can be approximated with Eq.~\ref{eq:sn_tm}.
Generation of charge carriers above the surface of the mirror is initiated with the escape of photoelectrons. An exponential factor
$\exp[-z/L]$ (where $L$ is the effective photoelectron escape depth) can be used (e.g., \cite{Stohr_book}) to model propagation of photoelectrons towards the surface of the mirror prior to escape. Taking this factor into account, integration of Eq.~\ref{eq:sn_tm} results in the following expression for the integral electron quantum yield (yield normalized by the incident photon flux) \cite{Stoupin_APL16}.
\begin{equation}
Y (Q) = \frac{1}{2} \epsilon_q n_q G^e T(Q) \frac{L}{\Lambda(Q) + L},
\label{eq:Y}
\end{equation}
where $\epsilon_q$ is the charge collection efficiency, $n_q$ is the charge amplification factor, and $G^e$ is a proportionality factor, which represents a correction for photoelectron energy conversion. Strictly speaking, only a fraction of the attenuated intensity of the X-rays results in generation of photoelectrons.
This photon-energy-dependent fraction is ascribed to $G^e$ to avoid introduction of an additional factor in Eq.~\ref{eq:Y}.
If the X-ray mirror is enclosed in a flow chamber containing light gas (e.g., He) gas impact ionization events produced by the secondary photoelectrons can be neglected (Fig.~\ref{fig:gzi}) since the energy required to produce one ion pair is $W_{g} \simeq$~40.3~eV \cite{Weiss56} while the energies of the secondary electrons do not exceed $\approx$~20~eV \cite{Henke77}. In addition, absorption cross section for hard X-rays in He is negligible compared to the ionization cross section by photoelectron impact {e.g.,\cite{Shah88}}. Thus, the electric carriers generated in the gas flow chamber originate from the photoelectric response of the mirror material. It is convenient to ascribe the number of charge carriers $n_q$ generated by a single photoelectron in the gas flow chamber to the ratio of the maximum photoelectron energy $E_{pe} \simeq E_X$ and the ion pair production energy $W_{g}$, $n_q = 2 E_X/W_g$.
Remarkably, the integral electron yield Eq.~\ref{eq:Y} is represented by the photon-electron attenuation factor $L/(\Lambda(Q) + L)$ modulated by the Fresnel transmissivity. Precise derivation of the photon-electron attenuation factor for any given material requires modelling of the photoemission processes from various atomic sub-shells and integration of the resulting photoemission cross-sections using the geometry of the photoemission detector.
Such rigorous approach could be based on the existing theoretical developments in X-ray photoemission spectroscopy pertaining to
X-ray optical effects \cite{Yang13}.
It should be noted that the photon-electron attenuation factor is a slow varying function of $Q$ (or the angle of incidence $\alpha$). Thus, subtraction of a smooth function, which agrees with the overall shape of the experimental curve should isolate the modulating signal, which contains Kiessig fringes. Structural information can be extracted from the result of the subtraction using Fourier transform similarly to Fourier analysis of the interference structure in X-ray specular reflection \cite{Sakurai92}. This strategy is illustrated below applied to a bi-layer X-ray mirror.
Fresnel reflection coefficients of a multilayer system are described by the Parratt's recursive relation \cite{Parratt54,Hau-Riege_book,Windt98}:
\begin{equation}
r_i = \frac{r_{ij} + r_j e^{i Q_i \Delta_i}}{1 + r_{ij}r_j e^{i Q_i \Delta_i}}.
\label{eq:r_i}
\end{equation}
In our notation $\Delta_i$ is the thickness of layer "$i$ and $r_{ij}$ is the reflection coefficient of the interface between layers "$i$" and "$j$"
\begin{equation}
r_{ij} = \frac{Q_i-Q_j}{Q_i+Q_j},
\label{eq:r_ij}
\end{equation}
where zero interface roughness is assumed.
We note that any $|r_{ij}| \ll 1$ above the largest critical angle corresponding to the layer with the greatest refractive decrement $\delta$.
In this approximation the net Fresnel reflection coefficient of a bi-layer (Fig.~\ref{fig:gzi}) is given by
\begin{equation}
r(Q) = r_{01} + r_{12}e^{iQ_1\Delta_1} + r_{23}e^{i(Q_1\Delta_1+Q_2\Delta_2)} + O (r_{ij}^3).
\label{eq:rq}
\end{equation}
The Fresnel transmissivity is given by
\begin{equation}
\begin{split}
T(Q) \simeq & T_{01} + |r_{12}|^2 e^{-2Q"_1\Delta_1} + |r_{23}|^2 e^{-2(Q"_1\Delta_1 + Q"_2\Delta_2)} \\
& + r^*_{01}r_{12} e^{i Q_1 \Delta_1} + r^*_{12}r_{23} e^{i Q_2\Delta_2} e^{-2Q"_1\Delta_1} \\
& + r^*_{01}r_{23} e^{i(Q_1\Delta_1+Q_2\Delta_2)} + c.c.,
\end{split}
\label{eq:tt}
\end{equation}
where $T_{01} = 1 - |r_{01}|^2$ is the transmissivity of entrance interface of the thick mirror, the symbol "$^*$" denotes complex conjugation and "c.c." denotes conjugated components with the oscillating exponential factors. The reflection coefficients $r_{ij}$ are non-oscillating functions of the layer thicknesses (i.e., do not contain Kiessig fringes) \cite{Sakurai92}. If $\exp{[-2 Im(Q_1)\Delta_1]} \ll 1$ substantial attenuation of the transmitted wave occurs in the top layer (i.e., the second layer is deeply buried). Thus, according to Eq.~\ref{eq:tt} the Fresnel transmissivity of a bi-layer system will differ from that of the thick mirror by components oscillating with frequencies $\Delta_1$, $\Delta_2$, and the sum frequency $\Delta_1$ + $\Delta_2$. For the purpose of Fourier analysis in the ($Re(Q_i)$,$\Delta$)-space above the greatest critical angle one can assume $Re(Q_2) \approx Re(Q_1)$. Generalization of this analysis to the case of arbitrary number of layers is straightforward (see Supplemental Material).
To address the problem experimentally integral photoemission yield and X-ray reflectivity of a bi-layer Pt-Cr X-ray mirror were measured simultaneously at several different photon energies in the range 8~-~23~keV. Prior to the analysis of integral electron yield the mirror was fully characterized using X-ray reflectivity (XRR), a method commonly used for structural characterization of surfaces \cite{Parratt54}. As determined by XRR the thickness of the top Pt layer was $\Delta_1$~=~725.2~$\rm \AA$ and that of the buried Cr layer was $\Delta_2$~=~116.5~$\rm \AA$. The measured reflectivities and XRR fitting curves are shown in Fig.~\ref{fig:refl} (see Methods for details).
\begin{figure}[t]
\centering\includegraphics[width=0.45\textwidth]{all_reflectivities.eps}
\caption{Specular X-ray reflectivity (black squares) from the Pt/Cr bilayer deposited on the Si substrate, together with the fitting curves (solid red lines). The error bars indicate $\pm$ 1 standard deviation. The X-ray energy at which the corresponding data were taken is indicated to the right. The curves are offset by one order of magnitude for clarity.}
\label{fig:refl}
\end{figure}
\begin{figure*}[t]
\centering\includegraphics[width=0.9\textwidth]{aibcd.eps}
\caption{Integral photoelectron yield and Fourier power spectra showing the structure of the Pt-Cr-Si bi-layer X-ray mirror.
(a) Measured integral photoelectron yield as a function of the wavevector transfer $Q$ of the incident wave at 10~keV (circles) and fit to Eq.~\ref{eq:Y} corresponding to photoelectron yield of a thick Pt mirror (green solid line). The inset shows magnified region above the critical angle where the Kiessig fringes due to the layered structure of the mirror are clearly observed. (b) Differential photoelectron yield (circles, Y-error bars and solid line) obtained as the difference between the experimental data and the fit in (a). The differential yield is plotted as a function of the real part of $Q_1$, the wavevector transfer inside the Pt layer. (c) Forier power spectrum (circles, solid line) of the differential yield (b) showing peaks corresponding to thicknesses of individual layers ($\Delta_1$ and $\Delta_2$ and their sum. (d) Fourier power spectra (circles, solid lines) obtained at different photon energies (shifted for clarity).}
\label{fig:qy}
\end{figure*}
Figure~\ref{fig:qy} illustrates the strategy used to extract thicknesses of individual layers from the integral photoemission yield curve. In the first step the curves collected at different photon energies were renormalized in the units of quantum yield (electric current divided by the incident flux and unit electric charge) and fit using Eq.~\ref{eq:Y}. The effective photoelectron escape depths were found to be $L \approx$ 200 - 400 $\rm \AA$ with greater values obtained at higher photon energies (see Supplemental Material for summary of the fit parameters).
Figure~\ref{fig:qy}(a) shows the resulting experimental curve (blue circles) and the fit (solid green line) for the photon energy $E_X$~=~10~keV. The Kiessig fringes were clearly observed (magnified region shown in the inset of Fig.~\ref{fig:qy}(a)). In the next step, the fit was subtracted from the experimental data to isolate the differential yield containing the structural information. The differential yield $\Delta Y(Q_1)$ is shown in Fig.~\ref{fig:qy}(b) as a function of wavevector transfer $Re(Q_1)$ in the Pt layer. The power spectrum of a Fourier transform applied to $\Delta Y(Q_1)$ is shown in Fig.~\ref{fig:qy}(c). Peaks in the spectrum correspond to thicknesses of the individual layers with precision on the order of spectral resolution $2\pi/\Delta Q_1$~$\simeq$~10-18~$\rm \AA$ defined by the available range of
the wavevector transfer $\Delta Q_1$.
The low-$\Delta$ region of the spectrum with peak corresponding to the buried Cr layer (116.5 $\AA$) could be contaminated by non-ideal subtraction procedure using the baseline, which includes the approximated photon-electron attenuation factor and the transmissivity of the thick mirror. Thus, evaluation of the thickness of the buried layer from the position of this peak alone can raise doubts. However, the appearance of the peak corresponding to the sum frequency $\Delta_1 +\Delta_2$ in the same spectrum confirms the presence of the Cr layer and its thickness. Also, the relatively featureless high-$\Delta$ region of the spectrum validates the approximation on the smallness of $|r_{ij}|$ given by Eq.~\ref{eq:rq}.
Figure~\ref{fig:qy} shows power spectra of the differential yield of the bi-layer mirror at different photon energies. The peaks corresponding to individual thicknesses $\Delta_1$ and $\Delta_2$ as well as the sum $\Delta_1 + \Delta_2$ are clearly observed at photon energies 10~keV and 11~keV where Kiessig fringes are most intense.
At some other photon energies (e.g., 13.5~keV and 15~keV) the noise level in the differential yield exceeds the weights of the spectral components. However, at higher photon energies (18~keV and 23~keV) reliable spectral detection of the thicknesses is still possible. Overall, the magnitude of the spectral components at various photon energies is consistent with the intensity of Kiessig fringes observed in XRR (Fig.~\ref{fig:refl}), which is governed by the energy-dependent absorption losses in the Pt layer. At 10-11~keV just below the energy region, which includes Pt L-absorption edges ($L_3$ at 11.56 keV, $L_2$ at 13.27 keV and $L_1$ at 13.88 keV) the imaginary parts of the refractive index ($\beta$) for Pt and Cr become comparable resulting in the increased intensity of the fringes. At the intermediate photon energies (12~keV - 15~keV) $\beta$ for Pt increases substantially in comparison to that of Cr. Finally, at higher photon energies (18~keV, 23~keV) the absorption losses for Pt become sufficiently small and the intensity of the fringes again increases. However, the absorption contrast between Pt and Cr remains substantial which explains reduction in the intensity of spectral peaks at $\Delta_2$ and $\Delta_1 + \Delta_2$.
In summary, our findings demonstrate that structure of buried layers can be studied with hard X-rays, yet, without detection of the reflected radiation. Instead, integral photoelectron yield is detected using simple voltage-bias-driven collection of generated electric charges above the surface of the studied material.
It should be mentioned that the approach is rather general and is not limited to the case where the top layer represents a conductive material. Indeed, self-detection of X-ray standing wave-effects have been observed on non-conductive materials such as diamond (e.g.,~\cite{Stoupin_SPIE14}).
The approach could be used to advance the field of non-destructive evaluation where unique field applications can become feasible. In particular, it can be useful for quantitative evaluation of surfaces in enclosed geometries where implementation of the conventional $\theta - 2\theta$ XRR arrangement is not possible. Potentially, the use of very hard X-rays at very grazing incidence is feasible since contrary to XRR no effort is required to isolate the reflected radiation from the incident X-rays. Such approach could become indispensable for non-destructive evaluation of internal surfaces in nuclear reactors, hydrocarbon transport systems, internal combustion engines, buried transistors in semiconductor circuits and other systems with buried layers and interfaces.
\section*{Methods}
\subsection*{Experiment}
\begin{figure}[t]
\centering\includegraphics[width=0.45\textwidth]{experiment.eps}
\caption{(a) Experimental setup for simultaneous measurement of Fresnel reflectivity and the integral photoemission yield (see text for details). (b) IV-curves of the system at the angular peak of the photoemission yield measured at various photon energies.}
\label{fig:exper}
\end{figure}
An X-ray mirror was prepared by deposition of a Pt film with thickness $\gtrsim 500 \mathrm{\AA}$ on top of a $\approx$~100-$\rm \AA$-thick layer of Cr on a polished Si substrate. The length of the mirror was 80~mm and the width was 20~mm. The mirror was placed on a non-conductive surface and contained in a helium flow chamber having an entrance and exit windows made of Kapton$^{\textregistered}$ film. The metallic working surface of the mirror served as the first electrode while a separate second grounded electrode was placed inside the chamber above the mirror at a distance of about 20 mm. The experiment was performed at 1-BM beamline of the Advanced Photon Source. The X-ray beam incident on the mirror was delivered by a Si (111) double-crystal monochromator (DCM). The DCM was detuned to suppress high-order Si Bragg reflections at high photon energies. The electric current between the electrodes in the flow chamber was measured using a source meter with applied bias voltages of $\pm$ 200 V.
The observed maximum electric currents (at the photoemission peak) were $\simeq$~10~-~80 nA with greater values at higher photon energies. Electric current in the absence of x rays (i.e., dark current) was $\approx$~100~pA. The root mean square fluctuations in the measured signal were about 20~pA. A calibrated solid state detector was placed behind the flow chamber to measure the reflected X-ray flux (mirror in the beam) and the incident X-ray flux (mirror out of the beam).
The size of the incident beam was set to $0.1\times3.0$~mm$^2$ (vertical$\times$horizontal) using X-ray slits placed upstream of the mirror chamber. The incident beam was centered on the surface of the mirror. Simultaneous measurement of the reflectivity and the electric current were performed while scanning the mirror's grazing angle $\alpha$ at different photon energies selected by the double-crystal monochromator. Prior to each scan the incident photon flux was measured using the solid state detector. The measured values were in the range $7 \times 10^{8} - 2 \times 10^{9}$ photons/s. The experimental setup is shown in Fig.~\ref{fig:exper}(a).
Prior to collection of angular curves an IV curve of the system was measured at the photoemission peak angle for each photon energy. The resulting IV curves are shown in Fig.~\ref{fig:exper}(b). In all cases saturation of the electric current and thus linear ionization-chamber-like response (full charge collection $\epsilon_q \simeq 1$) was achieved at sufficiently large negative potentials (-~200~V) applied to the mirror surface. The absence of saturation at the positive potential (+~200~V ) suggests that the potential is insufficient to recapture all the electrons in the system. We note that in the case of Pd mirror studied earlier \cite{Stoupin_APL16} full charge collection was observed for both positive and negative potentials. Detailed explanation of this phenomenon requires consideration of energy spectra of the emitted photoelectrons and their energy transfer to He atoms via collision processes. For the purpose of linearity of detection of the integral photoelectron yield it was sufficient to ascertain the saturation at -~200~V in the present case.
\subsection*{Structural analysis using XRR}
XRR measures the intensity of the specularly reflected beam as a function of the wavevector transfer $Q$ to extract the scattering length density (SLD) profile of the sample. SLD yields the information about the chemical composition of each layer, its thickness, physical density, and interfacial roughness. One of the main disadvantages of XRR is that spatial separation of the incident and the specularly reflected radiation becomes problematic in very grazing incidence ($\alpha \lesssim$~1~mrad) at high photon energies ($E_X \gtrsim$~20~keV). Reliable detection of the specularly reflected radiation requires $\theta$-$2\theta$ geometry and a large dynamic range of the radiation detector. These factors restrict the use of XRR to an X-ray analytical laboratory or a specialized X-ray source facility. Thus, despite the high penetrating power of hard X-rays, non-invasive structural studies of internal surfaces remain limited.
In our case of a model system (bi-layer X-ray mirror) XRR was fully applicable. For the XRR modelling we used dynamic Parratt's formalism \cite{Parratt54} combined with Nevot--Croce interface roughness and genetic fitting algorithm using the MOTOFIT software \cite{Nelson06}. Literature SLD values for the Si substrate, Cr, and Pt layers were used for modelling and the individual thicknesses of Pt (705.2~$\rm \AA$) and Cr (116.5~$\rm \AA$) layers, as well as the interfacial roughness at Pt-Cr ($\sigma_{12}$~=~7.5~$\rm \AA$) and Cr-Si ($\sigma_{23}$~=~4.5~$\rm \AA$) interfaces, were kept the same for all energies. In this experiment, lateral projection of the coherence length $l_c$ (a few tens of microns) of the radiation is much smaller than the sample surface along the beam (80 mm). Therefore, the local surface roughness (within the length $l_c$) from different parts of the sample is averaged incoherently. Such averaging is represented in the model by an additional top layer of ~20 $\rm \AA$-thick with SLD $\simeq$~80-85\% of the literature value for Pt. In other words, this top layer can be considered as a long-scale mirror height profile with a peak-to-valley value of 20 \AA, consistent with the Pt deposition specification.
\section*{Author contributions}
S.S. and M.Z. performed the experiments, analyzed the data and wrote the paper.
B.S. designed and fabricated the x-ray mirror.
\section*{Acknowledgements}
K. Lang is acknowledged for technical support. Use of the Advanced Photon Source was supported by the U. S. Department of Energy, Office of Science, under Contract No. DE-AC02-06CH11357. The work at the National Synchrotron Light Source-II, Brookhaven National Laboratory, was supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under Contract No. DE-SC0012704.
|
\section{Introduction}
Contrary to their traditional position between ellipticals and spirals in morphological galaxy classification schemes, a number of independent studies have recently suggested that lenticular galaxies form a full sequence on their own, in parallel to spirals (S0a, S0b, etc.), which can be sorted based on the decreasing relevance of the spheroidal component in analogy to the sequence of spirals \citep{2010MNRAS.405.1089L,2011MNRAS.416.1680C,2012ApJS..198....2K}.
These parallel properties could well be revealing an underlying evolutionary connection: are lenticulars indeed spiral galaxies which have undergone a certain transformation process? And, if so, did all lenticular galaxies originate through the same mechanism, or are they a ``mixed bag'' that results from very different evolutionary tracks, as suggested by \citet{1990ApJ...348...57V}?
One of the most popular processes to explain the origin of lenticular galaxies in the cluster regime is \textit{ram pressure stripping}. It consists on the expulsion of gas from a spiral galaxy falling into a cluster due to the pressure exerted by the intra-cluster medium. The lack of gaseous fuel results in the suppression of star formation, while discs can be preserved \citep{2005AJ....130...65C,2006A&A...458..101A,2015MNRAS.447.1506M}.
Nevertheless, observations imply that a large fraction of the local population of lenticular galaxies is not associated with the dense environment of clusters, but rather belongs to less densely populated groups \citep{2009ApJ...692..298W}, where mergers become important \citep{2014ApJ...782...53M}. In the following pages, we summarise the main points of the talk given in the IAU Symposium 321 in Toledo (Spain), in March 2016.
The talk and these proceedings were largely based on \citet{2015A&A...573A..78Q} and \citet{2015A&A...579L...2Q}.
\section{Bulge--disc coupling and pseudobulges in lenticular galaxies}
Analysing photometric decompositions of deep near-infrared observations, \citet{2010MNRAS.405.1089L} found strong scaling relations between the discs and bulges of lenticular galaxies. We have performed equivalent photometric decompositions on simulated major mergers, a subset
of the GalMer database \citep{2010A&A...518A..61C}.
We find that, under favourable orbital configurations, discs are first destroyed but then rebuilt short ($\sim$1\,Gyr) after the merging process is over. The possibility of disc survival in simulations of major mergers of spirals had already been implied by some authors \citep[e.g.~][]{2005ApJ...622L...9S, 2009ApJ...691.1168H,2016arXiv160203189A}, but we have taken one step forward and probed the potential merger origin of lenticular galaxies by comparing simulations with observations. For that, we created mock photometric images of the resulting relaxed galaxies using a variable, physically motivated mass-to-light ratio, and we performed photometric structural decompositions on the mock images of the merger remnants.
The conclusion is that the lenticular-looking galaxies emerging from those simulated major mergers obey the photometric scaling relations of real S0s.
Figure\,\ref{fig:kinem} (left) illustrates the scaling relation between disc and bulge sizes in our simulation remnants compared to real lenticulars. For a given bulge radius, only some disc sizes are allowed by nature in lenticular galaxies, and the simulated major mergers can reproduce that bulge--disc coupling.
The simulation points cluster towards the top-right corner,
but this is because the initial spiral models of GalMer are already very massive; starting from smaller progenitors would, presumably, reproduce the scaling relation over a larger range.
We have analysed eight additional photometric planes in \citet{2015A&A...573A..78Q}, including bulge-to-total ratios, magnitudes, and S\'ersic indices; we have also shown that these mergers can reproduce the pseudobulges detected in real lenticulars. In a companion study, we have also proved the existence of antitruncations in these merger remnants, which follow the same scaling relations as observations \citep{2014A&A...570A.103B}.
\section{Simultaneous change in concentration and angular momentum}
\label{lambda}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=1.0\textwidth]{Fig1.pdf} \caption{\textit{Left}: the major merger remnants (black diamonds) follow the same scaling relation as real lenticulars (orange circles) from \citet{2010MNRAS.405.1089L}. \textit{Right}: specific angular momentum ($\lambda_\mathrm{Re}$) versus concentration (\textit{R}$_{90}$/\textit{R}$_{50}$) for our major merger simulations (diamonds), and observational CALIFA datapoints. All parameters correspond to an edge-on view.}
\label{fig:kinem}
\end{center}
\end{figure*}
Late-type spirals seem to be incompatible with lenticular galaxies in a plane of angular momentum ($\lambda_\mathrm{Re}$) versus concentration (\textit{R}$_{90}$/\textit{R}$_{50}$), according to recent observational findings from the CALIFA team (\citealt{2015IAUS..311...78F}; van de Ven et al.~in preparation).
Fig.\,\ref{fig:kinem} (right) reproduces this observational plane, and we overplot analogous measurements for the major merger progenitors and remnants that we introduced above. The simulated major mergers induce a simultaneous change in concentration and angular momentum, which can explain the stark offset between real spirals and lenticulars \citep{2015A&A...579L...2Q}; while early-type spirals (Sa) and lenticulars largely overlap, the change between late-type spirals and S0s becomes more obvious. $\lambda_\mathrm{Re}$ is a proxy for the stellar angular momentum, calculated within the effective radius of the galaxy (\textit{R}$\leq$\textit{R}$_\mathrm{e}$).
We estimate the light concentration through the Petrosian ratio \textit{R}$_{90}$/\textit{R}$_{50}$, measured on the 1D azimuthally averaged profile of our mock SDSS \textit{r} band images. The outliers in terms of concentration can be understood as the result of young stellar populations forming in the centre (the merger can funnel gas inwards), and because we are analysing those remnants short after coalescence ($\sim$1\,Gyr after the merger is completed), whereas most of the observed lenticulars have probably had much longer timescales to relax.
This analysis suggests that major mergers are able to transform spirals into realistic lenticular galaxies of lower $\lambda_\mathrm{Re}$ and higher concentration, in agreement with the observations from CALIFA. This observed evolution cannot be easily explained by simple fading mechanisms such as ram pressure stripping, which are not expected to change the angular momentum of the initial galaxy significantly.
Summing up, we have shown that major mergers of spirals can sometimes rebuild discs, leading to lenticular galaxies in good agreement with photometric and kinematic observations.
We have also emphasised the importance of adopting a correct mass-to-light ratio when comparing such galaxy merger simulations with observations. Of course, our results do not necessarily imply that major mergers are the leading mechanisms in producing lenticular galaxies, and other processes could act in concert with mergers, or operate preferentially over certain environments; their relative relevance still remains to be quantified.
|
\section{Introduction}
One of the key ingredient in the realization of topological superconductivity\cite{Kitaev2001} is to remove fermion doubling. The doubling is naturally absent in fully spin polarized systems, yet ferromagnetic spin arrangement is not compatible with a conventional s-wave superconductivity. It has been realized that spin doubling can be removed in spin-full systems if spin is locked to the carrier momentum\cite{fu08,Sau2010,Lutchyn2010a,Oreg2010,Alicea2010}. While signatures of Majorana fermions have been reported in hybrid semiconductor/superconductor nanowires \cite{Rokhinson2012a,Mourik2012,Churchill2013a}, removal of fermion doubling has been observed in electron transport only in the cleanest nanowires fabricated by cleaved edge overgrowth technique\cite{Quay2010}.
An elegant proposal to circumvent fermion doubling is to couple two two-dimensional electron gases (2DEGs) with different sign of Land\'e $g$-factor and subject then to a quantized magnetic field\cite{Clarke2012,Stern2013}. In a quantum Hall effect (QHE) regime two oppositely polarized counter-propagating edge channels at the boundary of two 2DEGs form a helical domain wall ({\it h}-DW), similar to helical channels at the edges of two-dimensional topological isolators \cite{Hasan2010}. Coupled to an s-wave superconductor, {\it h}-DW should support Majorana fermions in the integer QHE regime and parafermions in the fractional QHE regime\cite{Clarke2012}.
\begin{figure}[t]
\centering\includegraphics[width=0.95\columnwidth]{fig1.pdf}
\vspace{0in}
\caption{(a) In a QHE regime a potential barrier creates counter-propagating edge channels with the same polarizations, while (b) a filling factor gradient $\nu_2>\nu$ results in a formation of a chiral domain wall, \textit{c}-DW. (c) A local change of the topmost Landau level polarization results in the formation of a helical domain wall, \textit{h}-DW, where counter-propagating edge channels have opposite polarization. Coupled to a superconducting contact (green), these {\it h}-DW should support non-Abelian excitations (magenta dots). (d) Schematic of a reconfigurable {\it h}-DW in a multi-gate device.}
\label{f:helic}
\end{figure}
While bringing two different electron gases into a close proximity is an experimentally challenging proposition, we propose to use electrostatically controlled quantum Hall ferromagnetic (QHFm) transitions to form helical domain walls, see schematic in Fig.~\ref{f:helic}. In a QHE regime kinetic energy of electrons in a 2DEG is quantized into Landau levels (LL), which are further split due to the presence of spin. Polarization of a 2DEG and, more importantly, of the top filled energy level, depends on the number of occupied energy levels $\nu=n/n_\phi$ (the filling factor is a ratio of electron $n$ and magnetic flux $n_\phi=eB/h$ densities), and changes as the system undergoes phase transitions between QHE states with different filling factors. If a 2D gas is separated into regions with different $\nu$'s by, e.g., electrostatic gating, chiral current-caring states are formed at the boundary. The actual order of spin-split energy levels is determined by an intricate balance between Zeeman, cyclotron and exchange energies. By shifting the balance it is possible to induce magnetic phase transitions between different QHE states with the same filling factor. QHFm transitions in integer and fractional QHE regimes have been studied extensively in the past\cite{eisenstein90,DePoortere2000,Smet2001,Jaroszynski2002,Gusev2003,Betthausen2014}. The QHFm transition field $B^*(B_{||})$ in those experiments was adjusted by in-plane (Zeeman) magnetic field $B_{||}$, which does not afford local control of polarization.
In this article we report development and characterization of heterostructures where $B^*$ is sensitive to electrostatic gating, $B^*(V_g)$, and, thus, can be controlled locally, an enabling step toward experimental realization of theoretical concepts \cite{Clarke2012,Stern2013}. In devices with multiple gates a possibility to reconfigure a network of {\it h}-DW opens a new class of systems where non-Abelian excitation can be created and manipulated.
\section{Electrostatic control of quantum Hall ferromagnet}
\begin{figure}[t]
\centering\includegraphics[width=0.95\columnwidth]{fig2.pdf}
\vspace{0in}
\caption{(a) Energy spectrum of Landau levels in a CdTe QW with 1.5\% of Mn calculated for $T=25$ mK. For the filling factor $\nu=2$ (gray shadow) electron gas undergoes ferromagnetic phase transition at $B^*(V_g)$. Field dependence of spin subbands (Eq.~\ref{eq:Es}) is plotted in the inset. (b) Spectrum for composite fermions $\Lambda$ levels for $x_{eff}=0.15\%$. QHFm transitions at $\nu=5/3$ and 4/3 have been experimentally observed\cite{Betthausen2014}.}
\label{f:effZ}
\end{figure}
\subsection{QHFm transition in dilute magnetic semiconductor}
Electrostatic control of QHFm transitions is realized in a dilute magnetic semiconductor CdTe:Mn with engineered placement of paramagnetic impurities. Substitutional Mn is a neutral impurity in CdTe and fractional QHE has been observed in high mobility CdTe:Mn two-dimensional electron gases with $\sim1\%$ of Mn\cite{Betthausen2014}. Exchange interaction between d-electrons on Mn (spin $S=5/2$) and s-electrons in the QW modifies energy spectrum of a 2DEG and results in unusual spin splitting and level crossing at high magnetic fields\cite{Wojtowicz1999}. QHFm transition in both integer and fractional QHE regimes have been observed in tilted magnetic fields experiments in QWs with uniform Mn doping \cite{Jaroszynski2002,Betthausen2014}. In the presence of magnetic field $B$ spin-dependent energy in dilute magnetic semiconductors is \cite{Furdyna1988}:
\begin{equation}
\label{eq:Es}
E_s^{\uparrow\downarrow} = \pm\frac{1}{2}\left[g^*\mu_BB + x_{eff} E_{sd} S \mathfrak{B}_s\left(\frac{g^*\mu_BSB}{k_B(T+T_{AF})}\right)\right],
\end{equation}
where the first term is the Zeeman splitting and the second term is due to an s-d exchange. Here $g^*\approx-1.7$ in CdTe, $E_{sd}\approx 220$ meV \cite{Gaj1994,Kiselev1998}, $x_{eff}$ is an effective Mn concentration, and $T_{AF}$ is due to Mn-Mn antiferromagnetic interaction. At low fields spin splitting is dominated by a large positive exchange term, while at high fields and low temperatures the Brillouin function $\mathfrak{B}_s(B,T)\approx1$ and $B$-dependence is dominated by the negative Zeeman term. In Fig.~\ref{f:effZ} we plot spin splitting of energy levels (\ref{eq:Es}) and spectrum of Landau levels (LL) for electrons $(n+1/2)\hbar\omega_c+E_s^{\uparrow\downarrow}$ and composite fermions (CF) $E_p^{CF}+E_s^{\uparrow\downarrow}$, where energy gaps between CF levels\cite{Park1998} $E_{p+1}^{CF}-E_p^{CF}\approx \alpha_C E_c/(2p+1)\propto \sqrt{B}$. Here $\hbar$ is the reduced Plank's constant, $\omega_c$ is the cyclotron frequency, $E_c=e^2/\epsilon \ell$ is the charging energy, $\ell$ is the magnetic length, constant $\alpha_C\approx0.01-0.03$ depends on the confining potential \cite{Liu2014}, $n=0,1,2,...$ and $p=1,2,3...$. The field of spin subbands crossing $B^*$ for the same LL ($\ket{n,\uparrow}$ and $\ket{n,\downarrow}$) or neighboring LLs ($\ket{n,\uparrow}$ and $\ket{n\pm1,\downarrow}$) depends on the strength of the s-d exchange interaction $x_{eff} E_{sd}$. Thus, engineering heterostructures with gate-tunable s-d exchange will allow local control of spin polarization in both integer and fractional QHE regimes.
\subsection{Heterostructures with s-d exchange control}
The second term in (\ref{eq:Es}) is a mean-field approximation to the exchange Hamiltonian $J_{sd}\sum_{\vec{R}_i}\delta(\vec{r}-\vec{R}_i)\vec{S}_i\cdot \vec{\sigma}\propto \left[\int_{[Mn]}|\varphi(z)|dz\right] \langle\vec{S}\rangle$, where interaction of an electron at a position $\vec{r}$ with a large number of Mn ions at positions $\vec{R}_i$ is approximated as an overlap of the electron wavefunction $\varphi(z)$ with a uniform Mn background within $z\in[Mn]$ and an average magnetization $\langle\vec{S}\rangle=\langle S_z\rangle=S\mathfrak{B}_s(B,T)$. For quantum wells with homogeneous Mn distribution throughout the whole QW region an integral $\chi=\int_{[QW]}|\varphi(z)|dz$ has weak dependence on the shape of $\varphi(z)$ and level crossing field $B^*$ is almost independent of a gate voltage \cite{Jaroszynski2002}.
\begin{figure}[t]
\centering\includegraphics[width=0.95\columnwidth]{fig3.pdf}
\vspace{0in}
\caption{(a) Band diagram of a 30nm CdTe QW heterostructure device is modeled using next{\bf nano$^3$} package \cite{nextnano}. Electron wavefunction is calculated for different voltages on the top (b) and back (c) gates. In (d) an integral overlap $\chi(V_g)$ between Mn-doped regions $[Mn_1]$ and $[Mn_2]$, normalized to the value at zero gate voltage $\chi(0)$, is plotted as a function of the 2D gas density change for front (FG) and back (BG) gates. (e) Mn doping distribution (red regions) in different wafers.}
\label{bandeng}
\end{figure}
We now consider non-uniform distribution of Mn inside a QW, e.g. Mn is confined to regions $[Mn_1]$ or $[Mn_2]$ within the QW, see Fig.~\ref{bandeng}b,c. In these regions $\varphi(z)$ has strong dependence on the out-of-plane electric field and $\chi$ becomes gate dependent, $\chi=\chi(V_g)$. Application of positive (negative) voltage to the front gate shifts electron wavefunction closer to (away from) the surface, $d\chi/dV_{fg}>0$ for $[Mn_1]$ and $d\chi/dV_{fg}<0$ for $[Mn_2]$. Gate voltage also changes electron density $dn/dV_{fg}>0$, thus $d\chi/dn>0$ for $[Mn_1]$ and $d\chi/dn<0$ for $[Mn_2]$ for the front gate. Application of a back gate voltage results in a density change $dn/dV_{bg} > 0$ but electrical field shifts wavefunction in the opposite direction, thus $d\chi/dn<0$ for $[Mn_1]$ and $d\chi/dn>0$ for $[Mn_2]$ for the back gate. Described behaviour is summarized on a Fig.\ref{bandeng}d. For the formation of well defined {\it h}-DWs we want to control $B^*$ with a minimal change of $n$ in order to remain at the same filling factor $\nu$, or maximize $|d\chi/dn|$.
In order to demonstrate electrostatic control of QHFm transition several Cd$_{1-x}$Mn$_x$Te/ Cd$_{0.8}$Mg$_{0.2}$Te quantum well heterostructures were grown by molecular beam epitaxy (MBE), see \cite{Jaroszynski2002,Betthausen2014} for details. Iodine delta-doping layer is separated from the QW by a 30 nm Cd$_{0.8}$Mg$_{0.2}$Te spacer. Mn was introduced into the QW region either as a digital $\delta$-doping or as a continuous doping, see schematics in Fig.~\ref{bandeng}e. More than 35 wafers have been grown and characterized with different Mn placement and concentration, here we report data on 4 representative wafers with $x_{eff}=1.71\%$, 0.34\%, 0.20\% and 0.085\% (wafers A,B,C and D). Samples were patterned into $100 \mu$m-wide Hall bars. A semitransparent Ti front gate (10 nm thick) was thermally evaporated onto the central part of Hall bars. Ohmic contacts were produced by soldering freshly cut indium ingots similar to previous studies \cite{Jaroszynski2002, Betthausen2014}. Copper foil glued to the back of samples served as a back gate. Devices were illuminated with a red LED at 4 K, low temperature electron density and mobility were in the range of $3.2-3.5\cdot10^{11}$ cm$^{-2}$ and $2-3\cdot10^5$ cm$^2$/V$\cdot$s in different samples. Electron transport was measured in a dilution refrigerator using standard ac technique with 10 nA excitation.
\subsection{Smooth QHFm transition at $\nu=1$}
\begin{figure}[t]
\centering\includegraphics[width=0.95\columnwidth]{fig4.pdf}
\vspace{0in}
\caption{(a) Longitudinal ($R_{xx}$) and Hall ($R_{xy}$) magnetoresistances in wafer A measured at $T=400$ mK for $V_{fg}=V_{bg}=0$. A peak at $B=7$ T is a QHFm transition between $\ket{1\uparrow}$ and $\ket{0\downarrow}$ states. (b) magnetoresistance in wafer C measured at $T\approx30$ mK for various $V_{bg}$ from $-200$ V (bottom trace) to $+200$ V (top trace), the traces are offset proportional to $V_{bg}$. Blue arrow marks evolution of the $m=2$ node, red arrow marks evolution of SdH peaks. In (c) and (d) $R_{xx}$ in wafer A is plotted as a function of $V_{bg}$ or $V_{fg}$ at a fixed $V_{fg}=0$ or $V_{bg}=100$ V respectively. Position of the QHFm transition is highlighted by a white dotted line. For $B=7$ T polarization of the top LL can be switched between $\uparrow$ and $\downarrow$ by the gate. Both plots have the same color scale. Measurements are performed at $T=300$ mK.}
\label{f:gate}
\end{figure}
Spin levels crossing measured in optical experiments\cite{Wojtowicz1999} and QHFm transitions observed in the fractional QHE regime\cite{Betthausen2014} are well described by Eq.~\ref{eq:Es} and the values of $x_{eff}$ extracted from the beating of Shubnikov - de Haas (SdH) oscillations at low fields\cite{Teran2002}. Yet, we did not observe any re-entrant behavior at $\nu=1$. We conclude that the absence of a transport signature of the QHFm transition at $\nu=1$ is either due to a phase separation in the vicinity of the transition or strong e-e exchange interaction and anticrossing of levels with the same orbital wavefunction.
An ability to locally control exchange interaction for small $x_{eff}<0.01$ is crucial for the formation of {\it h}-DW in a fractional quantum Hall regime, a prerequisite for the creation of higher order non-Abelian excitations. The strength of the exchange interaction can be obtained from the beating in the SdH regime, where the $m$-th node is defined by the condition \cite{Teran2002} $(m+1/2)\hbar\omega_c=|E_s^{\uparrow}-E_s^{\downarrow}|$. Gate dependence of magnetoresistance in wafer C at low fields is shown in Fig.~\ref{f:gate}b. Nodes are shifted to lower fields as the voltage on the back gate increases, $d\chi/dV_{bg}<0$. At the same time SdH peaks shift to higher fields, $dn/dV_{bg}>0$, and $d\chi/dn<0$ as is expected for the [Mn$_1$] doping arrangement.
\subsection{Gate control of sharp QHFm transition at $\nu=2$}
Unlike $\ket{0\uparrow}\leftrightarrow\ket{0\downarrow}$ QHFm transition at $\nu=1$, the $\ket{0\uparrow}\leftrightarrow\ket{1\downarrow}$ transition at $\nu=2$ involves states from different Landau levels and e-e exchange is strongly suppressed. Also, at $\nu=2$ level crossing has much stronger $B$-dependence $\hbar\omega_c/B\approx1.6$ meV/T, as compared to $g\mu_B\approx0.057$ meV/T at $\nu=1$, which suppresses phase separation. As a result quantization is lifted in the vicinity of the QHFm transition and a prominent signature in magnetoresistance is observed\cite{Jaroszynski2002}.
Magnetoresistance in sample A is shown in Fig.~\ref{f:gate}a. A small peak at $B=7$ Tesla in the middle of the $\nu=2$ state is the QHFm phase transition between $\ket{1\downarrow}$ and $\ket{0\uparrow}$ states, polarization of the top filled energy level changes across the transition. In the color plots magnetoresistance is plotted as a function of voltage on the front and back gates (Fig.~\ref{f:gate}c,d), measurements are performed by sweeping magnetic field at constant gate voltages. Electron density increases with the increase of $V_{bg}$ and $V_{fg}$ and peaks between adjacent QHE states shift to higher $B$ in both plots. In contrast, the QHFm transition $B^*$ shifts in opposite directions as a function of $V_{fg}$ and $V_{bg}$, consistent with the modelling of the wavefunction-Mn$_1$ overlap $\chi(V_g)$ in Fig.~\ref{bandeng}d. Note that for $B=7$ Tesla polarization of the top level can be tuned between $\ket{1\downarrow}$ and $\ket{0\uparrow}$ states by electrostatic gating, thus realizing the theoretical concept of Fig.~\ref{f:effZ}a.
Gate control of the s-d exchange is summarized in Fig.~\ref{AC}a for several wafers. The absolute values of Mn concentration $x_{Mn}$ and s-d overlap $\chi$ cannot be measured independently with high accuracy, by a relative change of the exchange interaction can be obtained from the gate dependence of the experimentally measured $x_{eff}(V_g)/x_{eff}(0)=\chi(V_g)/\chi(0)$. Slopes $dx_{eff}(V_g)/dn(V_g)$ are in a good agreement with $d\chi(V_g)/dn(V_g)$ obtained from band simulations, Fig.~\ref{bandeng}d. Note that the efficiency of the s-d exchange control depends on the $|dx_{eff}/dn|$ slope: $dB^*/dn=d\chi/dn=dx_{eff}/dn$ for QHFm transitions in the integer QHE regime and $dB^*/dn\approx d\chi/dn$ in the fractional QHE regime for large fields.
\subsection{Spin-orbit-induced gap for $\nu=2$ QHFm transition}
\begin{figure}[t]
\centering\includegraphics[width=0.95\columnwidth]{fig5.pdf}
\vspace{0in}
\caption{(a) Gate dependence of the measured effective Mn concentration, $\chi_{eff}(V_G)$ for wafers A-D for front (open symbols) and back (solid symbols) gates. Efficiency of s-d exchange control depend on the $|dx_{eff}/dn|$ slope: for QHFm transition in the integer QHE regime $(B^*(V_g)-B^*(0))\propto (x_{eff}(V_g)-x_{eff}(0))$. (b) Arrhenius plot of the $R_{xx}$ T-dependence at the QHFm transition, the activation energy is 0.096 meV. Top inset: temperature dependence of $R_{xx}$ near $\nu=2$. Bottom plot: anticrossing of $\ket{0,\uparrow}$ and $\ket{1,\downarrow}$ levels calculated using spin-orbit Hamiltonian, see text.}
\label{AC}
\end{figure}
The height of the peak at $B^*$ has exponential $T$-dependence and vanishes at low temperatures with an activation energy $T_0\approx 1K$, see Fig.~\ref{AC}b. We attribute this small gap to the level anticrossing due to the spin-orbit (SO) coupling between neighboring LLs. Energy spectrum in the presence of SO interactions is calculated by adding Dresselhaus $\gamma_D \boldsymbol{\kappa}\cdot\sigma$ and Rashba $\gamma_R {\boldsymbol{\mathcal E}}
\cdot \left(\boldsymbol{\sigma} \times \bf{k}\right)$ spin-orbit terms to the single-particle Hamiltonian of a 2D gas in magnetic field in the presence of s-d coupling (\ref{eq:Es}), square well confinement potential in $z$ direction, and electric field potential $e\phi(z)\approx e\mathcal E_z z $, see Appendix for details. Here $\gamma_D$ and $\gamma_R$ are the Dresselhaus and Rashba constants, and $\boldsymbol{\kappa}$ is defined as $(\{\hat k_x,\hat k_y^2-\hat
k_z^2\}, \{\hat k_y,\hat k_z^2-\hat k_x^2\}, \{\hat k_z,\hat
k_x^2-\hat k_y^2\})$. The energy spectrum near $\ket{0\uparrow}$ and $\ket{1\downarrow}$ levels crossing is plotted in the insert in Fig.~\ref{AC}b. The value of the anticrossing gap is found to depend only on the Rashba spin-orbit coupling
\begin{equation}
\Delta_{SO}=\frac{2\sqrt{2} |\gamma_R \langle{\mathcal E}_z\rangle|}{ \ell}.
\label{SO}
\end{equation}
For an average electric field of $\langle{\mathcal E}_z\rangle=3.5\cdot 10^4~ {\rm{V/cm}} $, $B=7~ {\rm T}$ and $\gamma_R=6.9~ e{\rm{\AA}}^2$ the calculated gap $\Delta_{SO}=70$ $\mu$eV, in a good agreement with the experimentally measured activation gap of 96 $\mu$eV. We note that an ability to open a topologically trivial (spin-orbit) gap is required for the localization of non-Abelian excitations \cite{Clarke2012}.
\section{Conclusions}
In this paper we propose a new experimentally feasible platform to realize non-Abelian excitations. The platform is based on the ability to create ferromagnetic domains in a quantum Hall effect regime, where helical domain walls are formed at the domain boundaries. These domain walls, coupled to a superconductor with high critical field $B_c$, should support Majorana and higher order non-Abelian excitations. Topological protection of the QHE regime insures that only single channel with removed fermion doubling is formed, thus alleviating multi-channel complication encountered in nanowire-based devices. As a proof-of-concept we developed CdTe quantum well heterostructures with engineered placement of paramagnetic Mn impurities and demonstrated local control of the QHFm transition at $\nu=2$ by electrostatic gating. Further research is needed to develop superconducting contacts to CdTe, a possible path is to overgrow CdTe with HgCdTe/HgTe epilayers where ohmic contacts with a high-$B_c$ superconductor Nb have been demonstrated\cite{Oostinga2013}.
\section{Acknowledgements}
Authors acknowledge support by the U.S. Department of Energy, Office of Basic Energy Sciences, Division of Materials Sciences and Engineering under Awards DE-SC0008630 (A.K and L.P.R.), by the Department of Defence Office of Naval research Award N000141410339 (A.K, T.W., G.S. and Y. L-G.), by the National Science Centre (Poland) grant DEC-2012/06/A/ST3/00247 (V.K., Z.A., G.K., and T.W.), and by the Foundation for Polish Science (T.W.).
\begin{widetext}
|
\section{Motivation}\label{sec:motivation}
Most data analysis in X-ray astronomy concentrate on describing single datasets
or on characterizing samples with results of fits of individual datasets. Once a
good description of an example dataset is found, the analysis of comparable
datasets follows.
Finally, the results of all those individual analyses are compared and
interpreted.
For instance, a particular parameter is found to depend on other parameters.
Instead of going back to the data analysis and fitting this dependency
directly to enhance the parameter precision or break degeneracies (feasible
through reduced degrees of freedom), the dependency is then analyzed on its own. In
another way, the former analysis is indeed repeated but with this
parameter fixed according to the discovered dependency. Furthermore, if
parameters cannot be constrained well, it is common to keep those parameters
fixed to a certain standard value.
Thus, one cannot gain any physical information from this fixed parameter
and, more importantly, systematical effects might arise. The reason for not
following sophisticated ways is usually a lack of computation power. Implementing
parameter correlations or dependencies would require one to analyze all data
at the same time. However, since computer power has increased and parallel
computation using several computers is possible, this situation has changed
today. In other words, fitting data simultaneously has become feasible even when
large numbers of datasets (e.g., 50-100 pointings at a single source) are to be
considered.
\begin{table}
\caption{Advantages and disadvantages of fitting several datasets simultaneously.}
\centering
\begin{tabular}{p{.42\linewidth}p{.46\linewidth}}
\hline
advantages & disadvantages \\
\hline\hline
$\bullet$ \footnotesize fixed parameters can be determined correctly & $\bullet$ \footnotesize increased runtime of fits and uncertainty calculations \\
$\bullet$ \footnotesize complicated parameter correlations can be implemented and tested & $\bullet$ \footnotesize large memory is needed $\rightarrow$ multi-CPU calculations required \\
$\bullet$ \footnotesize combination of different types of data is possible & $\bullet$ \footnotesize statistical weights of datasets have to be choosen \\
$\bullet$ \footnotesize parameter degeneracies can be broken & $\bullet$ \footnotesize careful handling of fit-parameters required \\
$\bullet$ \footnotesize reduced number of degrees of freedom & \\
\hline
\end{tabular}
\label{tab:procontra}
\end{table}
In Section~\ref{sec:implementation} we introduce an implementation of
simultaneous data analysis into the \texttt{Interactive Spectral Interpretation System}
\citep[ISIS,][]{houck2000a}, which has been ``designed to facilitate the
interpretation and analysis of high resolution X-ray spectra''\footnote{http://space.mit.edu/CXC/ISIS/}.
In Section~\ref{sec:applications}, we present ideas for possible applications
of simultaneous data analysis and further demonstrate the power of this method on the
example of the transient X-ray binary GRO~J1008--57\xspace in Section~\ref{sec:gro1008}.
Finally, we discuss questions and issues, which arise by comparing
advantages and disadvantages of simultaneous fits (Table~\ref{tab:procontra}).
\section{Implementation into ISIS}\label{sec:implementation}
ISIS \cite{houck2000a} was developed to fit X-ray spectra, but it can also be used
to analyze nearly all kinds of data due to its strong customization capability
\cite{noble2008a} compared to, e.g., XSPEC \cite{schafer1991a,arnaud1996a}. For instance,
user-defined fit-functions, as well as complex correlations between data and models, can
be implemented. However, functions to handle these correlations for a large number
of parameters and datasets in an easy way are not yet available.
Before we describe the technical realization of simultaneous data analysis
in ISIS, we introduce new notations used by the implemented functions.
\subsection{Terminology}
The parameters of a model which is fitted to data either act on
all datasets loaded into ISIS, or on an individual dataset. By defining
parameters for each dataset and tying them to each other, parameters can be
linked to multiple datasets similar to the approach chosen, e.g., in XSPEC.
We call multiple datasets, which should be fitted with the same set of
parameters, a \textit{datagroup}. The corresponding parameters are called
\textit{group parameters}. \textit{Global parameters} denote
parameters which act on all datagroups.
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{terminology}
\caption{Terminology of simultaneous fits in ISIS. There are $n$
and $m$ simultaneous datasets, forming the datagroup A and B, respectively.
Both datasets have their own group parameters, resulting from a model with
$p$ parameters. Some of the group parameters are the same for both
datagroups and are called global parameters.}
\label{fig:terminology}
\end{figure}
Figure~\ref{fig:terminology} illustrates these definitions. In this example,
a dataset requires a model with $p$ parameters. There are simultaneous data
from $n$ detectors available which can be described by the same parameters.
These datasets define the datagroup A with $p$ free parameters. Another group
of data was recorded by $m$ detectors. These datasets define an individual
datagroup B with, again, $p$ free parameters. During the analysis of
both groups, however, it turns out that a specific parameter seems to be equal
for both data groups within the uncertainties. As a result, the two
individual values for this parameter are tied to each other, resulting in a
global parameter. That reduces the number of free parameters by one and the
remaining group parameters can be constrained better.
\subsection{Data- and analysis functions}
Since simultaneous fits can have large numbers of fit parameters connected
by a complicated logic, we provide a collection of all functions necessary
to initialize and perform simultaneous fits in ISIS\footnote{these functions
are available as part of the \texttt{isisscripts}, a collection of useful
functions, which can be downloaded at
http://www.sternwarte.uni-erlangen.de/git.public/?p=isisscripts}.
The initialization of a simultaneous fit is performed via
\begin{verbatim}
simfit = simultaneous_fit();
\end{verbatim}
where \verb|simultaneous_fit| returns a structure (\texttt{Struct\_Type}),
which has to be assigned to a variable, here \verb|simfit|. The structure
contains several
functions and fields to handle simultaneous fits. The documentation of each
function is available using the \verb|help|-qualifier. Some important functions
are described in the following.
\begin{verbatim}
simfit.add_data(filenames);
\end{verbatim}
This defines a datagroup and loads the spectra given by \verb|filenames|, which must
be an array of strings. The function also allows other data than spectra to be
loaded or defined.
\begin{verbatim}
simfit.fit_fun(model);
\end{verbatim}
The string \verb|model| defines the fit-function to be used for all
datasets. Here, the placeholder \% can be used instead of a component
instance. In this case individual group parameters are applied to each
datagroup automatically.
\begin{verbatim}
simfit.set_par_fun(parameter, function);
\end{verbatim}
This is probably one of the most useful functions. Like the ISIS intrinsic
function, the value of the \verb|parameter| is determined by the given
\verb|function|. The \%-placeholder can be used within the string
\verb|parameter| to apply the function to the corresponding parameter of
each data group. However, the function may contain other parameters or even a
single parameter name as well. In the latter case, if the function is also
applied to all datagroups (using the \%), the single parameter is treated as
global parameter from now on.\\
Because a simultaneous fit results in a large number of parameters, a single
call to a fit-routine (\verb|fit_counts|) will take a long time. In the
example of the previous section, the final model fitted to the data consists
of $(n+m)\times p$ parameters, where only $2 p - 1$ are free. To reduce the
runtime of a fit, three fit-routines are implemented within the simultaneous-
fit-structure.
\begin{verbatim}
simfit.fit_groups(groupID);
\end{verbatim}
Instead of perfoming a $\chi^2$-minimization of all parameters and datasets,
this function loops over all datagroups and fits only the associated parameters
(group parameters). If a group is specified by the optional \verb|groupID|,
then only the group parameters of this particular group are fitted.
\begin{verbatim}
simfit.fit_global();
\end{verbatim}
Instead of fitting the group parameters, this function fits the global
parameters only. Since all defined data groups have to be taken into
account, the fit usually takes longer than fitting the group parameters. \\
\subsection{Uncertainty calculations}\label{sec:uncertainties}
As already mentioned, the runtime of simultaneous fits is
increased compared to fitting a single dataset only. Thus, uncertainty
calculations of parameters, where a certain parameter's range has to be found
corresponding to a change in $\chi^2$, will be affected dramatically by the
high runtime. Furthermore, it is necessary to distinguish between group- and
global parameters. We recommend to compute the uncertainty intervals for each
parameter on a different machine by, e.g., using \citep{noble2006a} or
\texttt{mpi\_fit\_pars} and the \texttt{SLmpi}
module\footnote{http://www.sternwarte.uni-erlangen.de/\\git.public/?p=slmpi.git}.
We compared the runtime of a parallel uncertainty calculation in ISIS with a
serial approach in XSPEC. Estimating the uncertainties of 10 parameters in
parallel (i.e., on 10 cores) is faster by more than a factor of 3 (21\,ks vs. 60\,ks).
Additionally, the calculations in ISIS resulted in a better $\chi^2_\mathrm{red}$
because the parameter ranges being scanned are larger in the parallel approach.
Group parameters depend on a single datagroup only. As a consequence, all other
datagroups and therefore all other group parameters can be ignored during the
uncertainty calculation. Unfortunately, that is not the case for global
parameters. During the analysis of GRO~J1008--57\xspace (see Section~\ref{sec:gro1008}), the
uncertainties of the global parameters have been calculated by revealing the
$\chi^2$-landscape of each global parameter by individual fits. Afterwards,
every landscape has been interpolated to find the $\Delta \chi^2$-value of
interest (e.g., $\Delta \chi^2 = 2.71$ corresponding to the 90\%-confidence
level).
In this way the runtime of an uncertainty calculation of a single global
parameter could be reduced significantly. Note that depending on the model
and amount of data, such a computation can take up to several days.
\section{Applications}\label{sec:applications}
There are various applications of simultaneous fits and data analysis. Besides
determining specific parameters which seem to be constant over time by all
available data, more physical questions can be tackled. For instance, if a
physical property of the object of interest results in multiple
observables:
\begin{itemize}
\item the geometry of the accretion column in accreting neutron star X-ray
binaries affects the line shape of cyclotron resonant scattering features
(CRSF) \citep{schoenherr2007a} as well as the pulse profile shape (Falkner et
al., in prep.).
\end{itemize}
Furthermore, instead of deriving physical properties from parameters after fits
have been performed, these properties can be directly fitted to the data by
implementing the dependency on the model parameters:
\begin{itemize}
\item the components in radio maps of jets in active galactic
nuclei move with certain velocities. Assuming a constant velocity of
the jet components, the velocity itself could be a global fit parameter
\cite{grossberger}.
\item in the sub-critical accretion regime of neutron stars, the spectrum is
believed to harden with increasing luminosity \citep{becker2012a}.
Any possible dependency between power-law shape and luminosity could be
fitted simultaneously with multiple spectra.
\end{itemize}
\section{The Example GRO~J1008--57\xspace}\label{sec:gro1008}
As an example of a successful simultaneous fit we briefly summarize the results
of our analysis of \mbox{GRO~J1008--57\xspace} using almost all available X-ray spectra and
-lightcurves. This transient high-mass X-ray binary consists of a neutron star
orbiting a Be-type optical companion. For further details of the system as well
as for the results of the analysis see \cite{kuehnel2013a} and references therein.
Since sources are only visible for a small fraction of their full orbit, it is
challenging to obtain the orbital parameters of transient X-ray binaries by
analyzing, e.g., pulse arrival times \citep{deeter1981a,boynton1986a}.
Thus, an observed shift in the orbital phase with respect to initial
orbital parameters can be fitted with either a different orbital period or
time of periastron passage. This leads to a parameter degeneracy,
which can be visualized by a contour map of the $\chi^2$-landscape of these
parameters.
The resulting contour map shows that both parameters are degenerated
statistically (i.e., the ellipsoidal contour lines are tilted).
The outburst times of the source are, however, clearly connected to the
periastron passage. Performing a simultaneous fit of the pulse arrival times
and the outburst times reduces the parameter degeneracy and results in much
better constrained parameters (about a factor of 2-3) as seen by
recalculated contour map.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{param_evol_fit_talk}
\caption{A fit (black lines) of the power-law index $\Gamma$ and black body
flux $F_\text{BB}$ as functions of power-law flux $F_\text{PL}$. The different
colors correspond to different outbursts.}
\label{fig:paramfit}
\end{figure}
Initial fits of the spectra of three outbursts of GRO~J1008--57\xspace in 2005, 2007 and 2011
with an absorbed cutoff power-law and an additional black body component showed
that the folding energy $E_\text{fold}$, as well as the black body temperature
$kT$, are independent of time within uncertainties. In particular, it seems
that they do not change between different outbursts, i.e., these parameters are
constant properties of the source.
Thus, those parameters are set as global parameters using
\verb|simfit.set_par_fun| and their values are determined by all available
data. In addition, further parameters can be treated as global ones
\citep[see][for more details]{kuehnel2013a}. Finally, each observation
is described by 3 group parameters only ($\approx$ degrees of freedom for each
datagroup, the global parameters contribute marginally), which are the power-law
flux $F_\text{PL}$, the black body flux $F_\text{BB}$, and the photon index
$\Gamma$. The latter two strongly correlate with $F_\text{PL}$, but show
no dependency on the outburst time or -shape. This correlation can be fitted
to describe the spectrum of GRO~J1008--57\xspace by only one parameter: the power-law flux
$F_\text{PL}$. The fit is shown in Fig.~\ref{fig:paramfit} and its values
are given in Section 4.2 of \citep{kuehnel2013a}.
As already mentioned in Section~\ref{sec:uncertainties}, the runtime of
uncertainty calculations of the global parameters is increased dramatically.
In the case of this analysis, the $\chi^2$-landscape produced by taking all
68 spectra into account was interpolated to estimate the uncertainties. The
calculation of a single global parameter took $\sim\!7$\,days on 100~CPUs
(16320\,CPUh).
\section{Outlook}\label{sec:todo}
Although the simultaneous fits have already been applied successfully to
real data (see Section~\ref{sec:gro1008} and \cite{kuehnel2013a}), the
routines and functions are still under development. We recommend to pull the
\texttt{isisscripts}-GIT-repository\footnotemark[2] regularly to be up-to-date.
There are, however, some caveats according to Table~\ref{tab:procontra} which
one should be aware of (as with any routine, not just our ISIS implementation).
In particular, the runtime still has to be reduced.
One way to achieve this is by performing the fit on multiple CPUs, e.g., one
CPU handles one dataset or datagroup. This has not been implemented yet because the
dependencies of the datasets on each other require data exchange between
the processes on the different machines. Additionally, the question about
weighing the data is currently under discussion. The weight depends on the
number of datapoints available in each dataset (or -group) as well as their
uncertainties - but what does this mean for its importance, i.e., its effect on
the model parameters? These remaining issues have to be clarified and the
respective solutions will be published in the future.
\begin{acknowledgements}
M.~K\"uhnel was supported by the Bundesministerium f\"ur Wirtschaft und
Technologie under Deutsches Zentrum f\"ur Luft- und Raumfahrt grant
50OR1113. We thank John E. Davis for developing the \texttt{SLXfig} package,
which was used to create all figures shown in this paper.
\end{acknowledgements}
\renewcommand{\bibsep}{0pt}
|
\section{Introduction}
\label{sec:introduction}
Recently \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric systems have gained much attention. Such systems do feature a
special class of non-Hermitian Hamiltonians which exhibit special properties
such as a real eigenvalue spectrum \cite{Bender1999a}. An operator is considered
to be \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric if it commutes with the \texorpdfstring{$\mathcal{PT}$}{PT}-operator,
\begin{eqnarray}
[\mathcal{PT}, H] = 0,
\end{eqnarray}
where $\mathcal{P}$ is the parity operator ($\hat x \rightarrow -\hat x$ and
$\hat p \rightarrow -\hat p$), and $\mathcal{T}$ is the time-reversal operator
which changes $\hat p \rightarrow - \hat p$ and $\ii \rightarrow -\ii$. For a
system of which the Hamiltonian can be written as
\begin{eqnarray}
H = -\Delta + V(x),
\end{eqnarray}
\texorpdfstring{$\mathcal{PT}$}{PT}-symmetry implies the condition
\begin{eqnarray}
\label{eq:ptpot}
V(x) = V^*(-x)
\end{eqnarray}
for the (complex) potential. In such a system the imaginary part of the
potential represents the in- and out-coupling of particles into or from an
external reservoir.
Quantum systems fulfilling this property have been studied in, e.g.,
\cite{qm1,qm2,Mehri,Bender1999a}. However, the concept of \texorpdfstring{$\mathcal{PT}$}{PT}-symmetry is not¬
restricted to quantum mechanics. The first experimental realization of
\texorpdfstring{$\mathcal{PT}$}{PT}-symmetric systems was indeed achieved in optical wave guides where the
effects of \texorpdfstring{$\mathcal{PT}$}{PT}-symmetry and \texorpdfstring{$\mathcal{PT}$}{PT}-symmetry breaking were observed
\cite{Rueter10}. These first breakthroughs have increased the research effort
put into this field \cite{PhysRevA.88.053817, Deffner2015, Albeverio2015,
Mostafazadeh2013b}. \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric systems have also been studied in microwave
cavities \cite{Bittner2012a}, electronic devices
\cite{Schindler2011a,Schindler2012a}, and in optical \cite{0305-4470-38-9-L03,
opticPT1, ramezani10, musslimani08a, optic1, optic2, Makris2010a, makris08,
Chong2011, Peng2014} systems. Also in quantum mechanics the stationary
Schr\"odinger equation was solved for scattering solutions \cite{qm2} and bound
states \cite{Mehri}. The characteristic \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric properties are still found
when a many-particle description is used \cite{Dast2014}. In
\cite{Guthrlein2015} a \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric system was embedded as a subsystem into an
hermitian system, showing that the subsystem retained its \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric
properties.
In the first realizations of Bose-Einstein condensates
\cite{And95a,Bra95a,Dav95a} atoms without long-range interactions were used.
Here the dominating interactions can be described by s-wave scattering. However,
when other atoms with a large dipolar moment were investigated
\cite{Griesmaier2005,Santos2000,PhysRevA.77.061601, PhysRevLett.107.190401} new
effects emerged. Depending on the strength of the dipole-dipole interaction a
novel behaviour can be introduced into the condensates, such as an anisotropic
collapse of $^{52}$Cr atoms \cite{PhysRevLett.101.080401}.
Bose-Einstein condensates described by the Gross-Pitaevskii equation can be
placed in a \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric potential. While new effects of
condensates with long-range dipole-dipole interactions have been found
\cite{Raghavan1999, Zaman2009, Xiong2009} when placed in a double-well trap,
completely new structures arise if additional gain and loss terms are
introduced.
Different processes for the realisation of loss and gain in a condensate were
proposed. For the loss of particles the use of a focused electron beam
\cite{Gericke2008} at the loss site was examined. The gain might be realized by
feeding the condensate from a second condensate \cite{Robins2008} in a Raman
super-radiance-like pumping \cite{Schneble2004, Yoshikawa2004}. Also the coupling
to additional potential wells with a higher or lower potential base could be
used \cite{Kreibich2014}.
Since the \texorpdfstring{$\mathcal{PT}$}{PT}-symmetry of the
Gross-Pitaevskii equation depends on the \texorpdfstring{$\mathcal{PT}$}{PT}-symmetry of the wave function,
effects which change the geometry of the wave function can lead to additional
phenomena. Dipolar Bose-Einstein condensates exhibit such effects, e.g.\
structured ground states have been found \cite{PhysRevLett.98.030406}.
Therefore one would expect that the combination of dipolar Bose-Einstein
condensates with a \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric trap will lead to a new behaviour. In
\cite{Fortanier2014} a Bose-Einstein condensate with long-range dipole-dipole
interaction in a \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric double-well potential was examined. This
condensate shows a richer, much more elaborate bifurcation scenario with more
states involved than in the case of a condensate with only short-range
interactions. Some of these bifurcations include up to five states,
and therefore allow for the possibile existence of higher-order exceptional
points, that is points in the parameter space where the eigenvalues and
eigenfunctions coalesce \cite{Kato66, Holger2007, Heiss2008, Graefe08a , Dem12,
Heiss2012, Heiss13a}.
An analytical continuation of the Gross-Pitaevskii equation provides the
mathematical tool to examine bifurcations and exceptional points in detail. An
encircling of exceptional points in complex parameter space can reveal, through
the exchange behaviour of the participating states, the order of the exceptional
point \cite{Guthrlein2013}. Also additional states and bifurcations which only
exist in the analytically continued space are revealed. We can apply this
method to the system investigated in \cite{Fortanier2014} where bifurcations
with up to five states have been observed.
In this paper we will examine this bifurcation scenario in much more detail
using an analytically continued Gross-Pitaevskii equation, therefore allowing us
to find additional (mathematical) states which compose the bifurcations. This
mathematical tool also allows us to expose the exchange behaviour of the
exceptional points associated with the bifurcations. In \sref{sec:gpe}
we will first give a short recapitulation of the Gross-Pitaevskii equation and
the ansatz of the wave function. In \sref{sec:ac} we will give an introduction
to the analytical continuation with bicomplex numbers and the representation in
an idempotent basis. Finally in \sref{sec:res} we will present new features in
the bifurcation scenario and examine the properties of the associated
exceptional points.
\section{Gross-Pitaevskii equation and the time-dependent variational principle}
\label{sec:gpe}
In this section we recapitulate the theoretical description of a Bose-Einstein
condensate using an ansatz of coupled Gaussians for the wave function. We use
the following parameterization of the Gross-Pitaevskii equation
\begin{equation}
\label{eq:gpe}
\left[
- \Delta
+ c_{\rm sc} \left| \psi(\vec x, t) \right|^2
+ V_{\rm d}
+ V_{\rm ext}
\right] \psi(\vec x, t)
= \ii \frac{{\rm d}}{{\rm d} t} \psi(\vec x, t)
\end{equation}
with the scattering length $c_{\rm sc}$. The long-range dipole
interaction is described by the dipole potential
\begin{eqnarray}
V_{\rm d} = c_{\rm d} \int {\rm d}^3 x' \frac{1-3\cos^2 \theta}{|\vec x - \vec
x'|^3} \left| \psi(\vec {x'}, t) \right|^2
\end{eqnarray}
with the dipole strength $c_{\rm d} = 1$ and the angle $\theta$ between $\vec x
- \vec x'$ and the dipole alignment. A more complete overview of the
theoretical description and experimental realisation can be found in
\cite{Lahaye2009}.
We describe the external double-well
potential by two Gaussian beams \cite{Fortanier2014}
\begin{eqnarray}
V_{\rm ext} &= \sum_{i=1}^{2} c_{{\rm g}, i}
\exp\left( - (\vec x^T \mat A_{{\rm g},i} \vec x + \vec x^T \vec q_{{\rm g},i}
+ \gamma_{{\rm g},i}) \right),
\end{eqnarray}
where $c_{{\rm g},i}$ defines the amplitude, $\mat A$, $\vec q$ and $\gamma$
define the shape and location of the Gaussian beams. The two wells are located
at $\vec q = ( \pm x_{\rm pos}, 0, 0 )$ symmetrically with respect to the origin
and with the amplitude
\begin{eqnarray}
c_{{\rm g},1} = V_0 + \ii \Gamma
~{\rm and}~
c_{{\rm g},2} = V_0 - \ii \Gamma,
\end{eqnarray}
and therefore fullfilling the requirement \eref{eq:ptpot} for a \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric
potential. The parameter $V_0$ represents the strength of the real potential,
and $\Gamma$ determines the strength of the in- and outcoupling of particles
from an external reservoir. We are especially interested in how a change of the
in- and outcoupling strength $\Gamma$ effects the bifurcation behaviour.
To obtain the equations of motion we apply the time-dependent variational
principle by McLachlan \cite{McLachlan1964a} to \eref{eq:gpe}. To
fullfil the GPE as best as possible the difference betweeen the two sides of
the equation
\begin{eqnarray}
I = || \ii\psi - H \underbrace{\frac{\partial}{\partial t} \psi}_{=: \phi} ||^2
\end{eqnarray}
is minimized. For a given time $t$ the wave function $\psi$ is pinned and the
time-derivative $\phi$ is varied. Hence, if $I$ is a minimum for small $\delta\phi$
the change of $\delta I$ must vanish. We make an ansatz of coupled
Gaussians for the wave function
\begin{eqnarray}
\label{eq:ansatz}
\psi(\vec x, t)
= \sum_{k=1}^N g^k
= \sum_{k=1}^N \exp\left(-\left(
\vec x^T \mat A_k \vec x
+ \vec p_k^T \vec x
+ \gamma_k
\right)\right)
\end{eqnarray}
where the complex symmetric matrix $\mat
A$ defines the shape of the Gaussian, the complex momentum vector is given by
$\vec p$ and the complex number $\gamma$ defines the amplitude and phase of the
Gaussian. This ansatz does not allow for arbitrary matrices
$\mat A$, e.g. matrices which contain a negative gaussian widths are not
normalizable and are not accepted as solutions.
Because our trap exhibits a strong confinement in the $y$ direction
\cite{Fortanier2014} we can introduce further restrictions to the ansatz of our
wave-function. We do not consider all rotational and translational degrees of freedom in
the direction of the confinement ($A_{xy} = A_{yz} = p_{y} = 0$). With the help
of the time-dependent variational principle one can obtain the time derivatives
of the Gaussian parameters $\dot\mat A_k$, $\dot\vec p_k$ and $\dot \gamma_k$
\cite{Fortanier2014}. A numerical root search can be performed to
obtain the stationary states. It was shown \cite{Rau10c} that the
variational principle, while being numerically cheap, provides results with
extremely high quality. For all calculations in this paper, an ansatz
of $N=2$ coupled Gaussians was chosen, i.e.\ the wave function is approximated
in each well by one Gaussian function (but they are not restricted to a specific
location).
\section{Analytical continuation}
\label{sec:ac}
The non-analyticity of the Gross-Pitaevskii equation \eref{eq:gpe},
due to the square modulus of the wave-function, prevents us
from examining the exchange behaviour of exceptional points. Also since the
equation is not analytical the number of states can change from one side of the
bifurcation to the other \cite{Fortanier2014}. It has been shown for simpler
systems without long-range interactions that with an analytical continuation
\cite{Guthrlein2013,Dast2013} the whole bifurcation scenario of the equation is
revealed and the complete relation between the different stationary states can
be observed. But since the Gross-Piaevskii equation is non-analytical
and already in the complex domain, the analytical continuation has to go beyond
the complex domain. One way to perform the analytical continuation is
to separate all complex equations and complex parameters into twice the number
of real equations with twice the number of real parameters. Then these equations
can be analytically continued. An equivalent alternative is the introduction of
bicomplex numbers. In this approach the complex numbers $\mathbb{C}$ are
replaced by bicomplex numbers $\mathbb{BC}$ \cite{LUNA-ELIZARRARAS2012}. Both
methods are mathematically completely equivalent, however, the second approach
is much more elegant since it provides huge numerical benefits as we will see in
the following discussion. Note that there are different hypercomplex
numbers. But for the description we have in mind only bicomplex numbers have the
correct properties. Furthermore this choice ensures that while extending the
original model, all its properties and states are contained and no imformation
is lost. In other scenarios different hypercomplex numbers are used and
appropriate, e.g. quaternions and co-quaternions are discussed in
\cite{Brody2011a, brody2011b}.
\subsection{Bicomplex numbers}
We first recapitulate some basic properties of bicomplex numbers
\cite{LUNA-ELIZARRARAS2012}. In contrast to complex numbers bicomplex numbers
have three imaginary units with the following properties
\begin{equation}
\kk = \ii\jj ~,~~ \ii^2 = \jj^2 = -1 ~,~~ \kk^2 = 1.
\end{equation}
To analytically continue a complex equation we split the equation into its real
and imaginary part
\begin{equation}
z =
\underbrace{
\underbrace{z_{\rm r}}_{\in\mathbb{R}}
+ \ii \underbrace{z_{\rm i}}_{\in\mathbb{R}}
}_{\in \mathbb{C}}.
\end{equation}
These real and imaginary parts can be extended to complex numbers with the
imaginary unit $\jj$. Then $z$ becomes a bicomplex number
\begin{equation}
\label{eq:binumber}
z =
\underbrace{
\underbrace{(
\underbrace{z_0}_{\in\mathbb{R}}
+ \jj \underbrace{z_1}_{\in\mathbb{R}}
)}_{\in \mathbb{C}_\jj}
+ \ii \underbrace{(
\underbrace{z_2}_{\in\mathbb{R}}
+ \jj \underbrace{z_3}_{\in\mathbb{R}}
)}_{\in \mathbb{C}_\jj}
}_{\in \mathbb{BC}}
= z_0 + \jj z_1 + \ii z_2 + \kk z_3.
\end{equation}
The complex numbers $\mathbb{C}_\jj$ with the imaginary unit $\jj$ are
isomorphous to the complex numbers $\mathbb{C}_\ii$ with the imaginary unit
$\ii$.
\subsection{Representation of bicomplex numbers in the idempotent basis}
There exists a representation of bicomplex numbers in an idempotent basis. Let
us first consider the two idempotent elements
\begin{eqnarray}
\ensuremath{{\mathrm{e}^\oplus}} = \frac{1+\kk}{2}
~{\rm and}~
\ensuremath{{\mathrm{e}^\ominus}} = \frac{1-\kk}{2},
\end{eqnarray}
which fullfil the relations
\begin{eqnarray}
\ensuremath{{\mathrm{e}^\oplus}}^2 = \ensuremath{{\mathrm{e}^\oplus}}, ~\ensuremath{{\mathrm{e}^\ominus}}^2 = \ensuremath{{\mathrm{e}^\ominus}} ~{\rm and}~ \ensuremath{{\mathrm{e}^\oplus}} \ensuremath{{\mathrm{e}^\ominus}} = 0.
\end{eqnarray}
These properties allow us to decompose every bicomplex number of the form
\eref{eq:binumber} into
\begin{eqnarray}
z &= \left( (z_0 + z_3) + (z_2 - z_1) \ii \right) \ensuremath{{\mathrm{e}^\oplus}}
+ \left( (z_0 - z_3) + (z_2 + z_1) \ii \right) \ensuremath{{\mathrm{e}^\ominus}} \nonumber \\
&= \bpparam{z} \ensuremath{{\mathrm{e}^\oplus}} + \bmparam{z} \ensuremath{{\mathrm{e}^\ominus}}.
\end{eqnarray}
The coefficients $\bpparam{z}$ and $\bmparam{z}$ can be chosen to be either
elements of $\mathbb{C}_\ii$ or $\mathbb{C}_\jj$. In our calculations we choose
the $\mathbb{C}_\ii$ representation. We can now consider the implication the
idempotent properties have on the basic arithmetic operations. One finds that
\cite{LUNA-ELIZARRARAS2012}
\numparts
\begin{eqnarray}
a \pm b &=
(\bpparam{a}\pm\bpparam{b}) \ensuremath{{\mathrm{e}^\oplus}}
+ (\bmparam{a}\pm\bmparam{b}) \ensuremath{{\mathrm{e}^\ominus}},
\\
a \cdot b &=
(\bpparam{a}\cdot\bpparam{b}) \ensuremath{{\mathrm{e}^\oplus}}
+ (\bmparam{a}\cdot\bmparam{b}) \ensuremath{{\mathrm{e}^\ominus}},
\\
a / b &=
(\bpparam{a}/\bpparam{b}) \ensuremath{{\mathrm{e}^\oplus}}
+ (\bmparam{a}/\bmparam{b}) \ensuremath{{\mathrm{e}^\ominus}},
\end{eqnarray}
\endnumparts
viz.\ the operations can be performed independently for the plus and minus
components in the complex subspaces. These properties allow us to use a very
efficient numerical implementation (which uses already existing algorithms). In
particular for terms which contain the dipolar potential a highly optimized
algorithm for complex numbers is available. If the method without
bicomplex numbers would have been used other integrals had to be solved,
for which no algorithms exist. We will see in the next section that
special care must be taken if a calculation involves the complex conjugate.
\subsection{Complex conjugation}
There exist different kinds of complex conjugations for bicomplex numbers.
In \cite{Bagchi2015} possible conjugations and new \texorpdfstring{$\mathcal{PT}$}{PT}-symmetries
are discussed. However, for our application only one is relevant, since our
goal is that the extended model contains all states and properties of the
original model. In order to determine which one must be used we first
consider the complex conjugation which occurs in \eref{eq:gpe} without the
bicomplex continuation. We start with the complex conjugation of the real and
imaginary part of the complex number
\begin{eqnarray}
z^* &= z_{\rm r} - \ii z_{\rm i}.
\end{eqnarray}
Now the analytical continuation is applied by replacing the real and imaginary
parts with complex numbers with the imaginary unit $\jj$
\begin{eqnarray}
z^* &= (z_0 + \jj z_1) - \ii (z_2 + \jj z_3) = z_0 + \jj z_1 - \ii z_2 - \kk z_3.
\end{eqnarray}
This complex conjugation changes the sign of all $\ii$ and $\kk$ components.
Using this complex conjungation we can write the modulus squared of the
wave function in \eref{eq:gpe} as
\begin{eqnarray}
|\psi|^2 = \left<\psi \middle| \psi \right>_\mathbb{BC} = \psi^* \psi.
\end{eqnarray}
This complex conjugation also must be used in inner products. Note that in the
representation of the idempotent basis the complex conjugate reads
\begin{eqnarray}
\label{eq:cc}
z^* = \overline\bmparam{z} \ensuremath{{\mathrm{e}^\oplus}} + \overline\bpparam{z} \ensuremath{{\mathrm{e}^\ominus}},
\end{eqnarray}
where the bar denotes the complex conjugation with $\bpparam{z}, \bmparam{z}
\in \mathbb{C}_\ii$.
\subsection{Decomposition of the GPE using the idempotent basis}
With the properties shown so far we can examine the analytical continuation
of the Gross-Pitaevskii equation. A linear set of equations, such as the one
obtained by the time-dependent variational principle \cite{FortanierDoc}, can be
decomposed into two linear systems of equations using the idempotent basis
\numparts
\begin{eqnarray}
\bpparam{\mat K} \bpparam{\vec v} &= \bpparam{\vec r}, \\
\bmparam{\mat K} \bmparam{\vec v} &= \bmparam{\vec r}.
\end{eqnarray}
\endnumparts
The vectors $\bpparam{\vec r}$, $\bmparam{\vec r}$ and the matrices
$\bpparam{\mat K}$, $\bmparam{\mat K}$ contain both plus and minus coefficients
of the Gaussian parameters. The elements of the bicomplex matrices $\mat K$ and
vectors $\vec r$ are of the form \cite{FortanierDoc}
\begin{eqnarray}
\bbraketop{g_l}{\alpha^n \beta^m V}{g_k}
\end{eqnarray}
with $\alpha,\beta = x,y,z$ and $n, m \in \mathbb{N}$, and the Gaussians $g_l,
g_k$. If the bicomplex conjugate described in \eref{eq:cc} is used, one can
decompose the expression into
\begin{eqnarray}
\bbraketop{g_l}{\alpha^n \beta^m}{g_k}
=
\braketop{\bmparam{g_l}}{\alpha^n \beta^m \bpparam{V}}{\bpparam{g_k}} \ensuremath{{\mathrm{e}^\oplus}}
+
\braketop{\bpparam{g_l}}{\alpha^n \beta^m \bmparam{V}}{\bmparam{g_k}} \ensuremath{{\mathrm{e}^\ominus}}
.
\end{eqnarray}
This representation allows us to use an existing algorithm for complex numbers
to calculate the necessary integrals. The algorithm just must be called twice
for the appropriate parameter sets. For elliptic integrals the Carlson algorithm
\cite{Carlson} can be used. However, the largest numerical improvement is
achieved by using the highly optimized algorithm presented in
\cite{FortanierDoc} for the numerical integration of the dipolar term.
Finally we examine how the control parameters, which were real numbers before
the analytical continuation was applied, can be represented in the idempotent
basis. Originally the scattering length $c_{\rm sc}$ was real. After the complex
continuation it assumes the form
\begin{eqnarray}
c_{\rm sc} = c_{\rm sc}^0 + \jj c_{\rm sc}^1,
\end{eqnarray}
with $c_{\rm sc}^0, c_{\rm sc}^1 \in \mathbb{R}$. We can now represent this
parameter in the idempotent basis
\begin{eqnarray}
c_{\rm sc} &= c_{\rm sc}^0 + \jj c_{\rm sc}^1
= \underbrace{(c_{\rm sc}^0 - \ii c_{\rm sc}^1)}_{\bpparam{c_{\rm sc}}} \ensuremath{{\mathrm{e}^\oplus}} +
\underbrace{(c_{\rm sc}^0 + \ii c_{\rm sc}^1)}_{\bmparam{c_{\rm sc}}} \ensuremath{{\mathrm{e}^\ominus}}.
\end{eqnarray}
Since there are only two idependent real parameters the relation
$\overline{\bpparam{c}_{\rm sc}} = \bmparam{c}_{\rm sc}$ is fullfilled. The same
also applies for the in- and out-coupling parameter $\Gamma$.
Another special parameter type exists. If the bicomplex number $z$ only
contains a real part and an imaginary part for the complex unit $\ii$, but not
for $\jj$ or $\kk$, the number can be decomposed in the idempotent basis in such
a way that the two coefficients are equal, i.e.\ $\bpparam{z} = \bmparam{z}$.
\section{Results}
\label{sec:res}
\subsection{Bifurcation scenario of a BEC without long-range interaction}
To compare the bifurcation scenario between a BEC with and without dipolar
interaction, a short recapitulation of the bifurcation scenario of a BEC
without long-range interactions is given in this section. This scenario has been
discussed in detail in previous publications \cite{Graefe12,
Cartarius12b,Dast13,Haag14}. The bifurcation scenario of such a \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric
double well system can be described by the matrix equation
\begin{eqnarray}
\label{eq:mm}
\left(\begin{array}{cc}
-g |\psi_1|^2 - \ii \gamma & v \\
v & -g |\psi_2|^2 + \ii \gamma
\end{array}\right)
\left(\begin{array}{c}
\psi_1 \\ \psi_2
\end{array}\right)
= \mu
\left(\begin{array}{c}
\psi_1 \\ \psi_2
\end{array}\right)
\end{eqnarray}
with the coupling strength $v$, the nonlinearity $g$, and the in- and
outcoupling strength $\gamma$.
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig1.pdf}
\caption{
\label{fig:old}
Bifurcation scenarios of a Bose-Einstein condensate without long-range
interaction. The underlying equation \eref{eq:mm} was analytically continued
using bicomplex numbers. Solutions which exist in the complex plane are
represented by solid lines. Solutions which require the analytical
continuation are shown as dashed lines. States which are \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric are
denoted by the subscript s, while states which break this symmetry are
denoted by a. }
\end{figure}
In the linear case $g = 0$ there exist two parameter ranges for the in- and
out-coupling strength $\gamma$. In the regime below a critical parameter
$\gamma_c$ there are two \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric solutions, which merge in a tangent
bifurcation (O1 in \fref{fig:old}(b)) at the critical
parameter $\gamma_c$. For values greater than this critical strength, there
exist two \texorpdfstring{$\mathcal{PT}$}{PT}-broken solutions with complex chemical potentials. The tangent
bifurcation in this scenario is a second-order exceptional point. One can
observe the pitchfork bifurcation (O2 in \fref{fig:old}) which occurs on the
upper \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric branch for values of $-2 < g < 0$ (see \fref{fig:old}(a))
and switches to the lower \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric branch for values of $0 < g < 2$ (see
\fref{fig:old}(c)). This bifurcation vanishes for values $g < -2$ or $g > 2$
(see \fref{fig:old}(d)).
We will see that a bifurcation similar to bifurcation O1 appears in a dipolar
Bose-Einstein condensate. The behaviour of the bifurcation O2, which also
occurs, will be altered in certain parameter regions. The change in behaviour is
also dependent on new emerging bifurcations.
\subsection{Results for a BEC with long-range dipole interactions}
The calculations for the analytically continued GPE \eref{eq:gpe} reveales the
complete bifurcation structure of the system. In addition to the known
stationary states which exist in the complex domain new states which are truly
bicomplex are found. With these new states all branches of the bifurcations are
present. In addition it is possible to evaluate the exchange behaviour of the
exceptional points.
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig2.pdf}
\caption{
\label{fig:line0}
Real and imaginary parts of the plus and minus component of the bicomplex
mean-field energy in a dipolar Bose-Einstein condensate with scattering
length $c_{\rm sc} = -0.9$. The mean-field energy is composed as $E_{\rm mf}
= \bpparam{E_{\rm mf}} \ensuremath{{\mathrm{e}^\oplus}} + \bmparam{E_{\rm mf}} \ensuremath{{\mathrm{e}^\ominus}}$. States which
exist without the analytical continuation into bicomplex values must obey
$\bpparam{E_{\rm mf}} = \bmparam{E_{\rm mf}}$ and are shown as solid lines.
States which exist only in the bicomplex domain are shown as dashed lines.
The four bifurcations which occur between the states are marked with B1 to
B4. As discussed in \cite{Fortanier2014} states s1 and s2 obey \texorpdfstring{$\mathcal{PT}$}{PT}-symmetry
from $\Gamma = 0$ up to the bifurcation B1. Beyond this point the states are
\texorpdfstring{$\mathcal{PT}$}{PT}-broken. }
\end{figure}
In \fref{fig:line0} eigenvalues for states with $c_{\rm sc} = -0.9$ are shown.
The states which are purely complex (solid lines) are already known from
\cite{Fortanier2014}. Since these states are complex the relation
\begin{equation}
\bpparam{E_{\rm mf}} = \bmparam{E_{\rm mf}} \label{eq:rel_energy_complex}
\end{equation}
holds true. The real and imaginary parts of the mean-field energy
are
\begin{eqnarray}
\Re E_{\rm mf} = \frac{1}{2} \left( \bpparam{E_{\rm mf}} + \bmparam{E_{\rm mf}} \right)
~~{\rm and}~~
\Im E_{\rm mf} = \frac{\ii}{2} \left( \bpparam{E_{\rm mf}} - \bmparam{E_{\rm mf}} \right)
.
\end{eqnarray}
The imaginary part of the mean-field energy exists due to the loss or gain of
particles in the system. For the states present in the original model the real
and imaginary parts of the energy are real numbers. For new analytically
continued states these numbers are now complex. The relation
\eref{eq:rel_energy_complex} no longer holds true for states living only in the
bicomplex space (dashed lines). Since all branches of the
bifurcations are now present, there exist six states for the whole parameter
range of $\Gamma$. The chemical potential exhibits the same qualitative
behaviour as the mean-field energy. Since the qualitative
behaviour is the same in the plus and minus components only the plus component
of the mean-field energy is discussed in the following.
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig3.pdf}
\caption{
\label{fig:line1}
Mean-field energies $\bpparam{E_{\rm mf}}$ for different values of $c_{\rm
sc}$. The different bifurcation points are denoted by B1, B2, B3a and B3b.
When the bifurcations B3a and B3b coincide they are marked by B3. The
bifurcation B2 does not exist in (a). }
\end{figure}
In \fref{fig:line1} the real and imaginary parts of the plus component of the
mean-field energy are shown for different values of the scattering strength. For
small values of $\Gamma$ the states s1 and s2 show the typical behaviour also
known from BECs without long-range interactions. They merge in a tangent
bifurcation B1. Up to this point the states also exist without the analytical
continuation. For larger $\Gamma$ they become bicomplex. This qualitative
behaviour is independent of the scattering length, only the critical point is
slightly shifted.
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig4.pdf}
\caption{
\label{fig:line2}
For a scattering length of $c_{\rm sc,crit,1} = -0.93063$ the bifurcations
and EP2s B1 and B2 merge. This can be seen in subfigure a) and the resulting
bifurcation is marked as B12. In subfigure c) the EPs and bifurcations B2,
B3a and B3b merge for $c_{\rm sc,crit,2} = -0.93365$ into a higher order
exceptional point which is denoted by B23. Subfigure b) shows the
bifurcation scenario for an intermediate scattering length of $c_{\rm sc} =
-0.932$. }
\end{figure}
The bifurcation B2 undergoes multiple behaviour changes for different scattering
lengths. Also this bifurcation does not always exist (see \fref{fig:line1}(a)).
For smaller scattering lengths the bifurcation appears at $\Gamma = 0$. This is
a pitchfork bifurcation between the states s1, s3, and s6. The bifurcation moves
along the state s1 to larger $\Gamma$ as the scattering length is decreased (see
\fref{fig:line1}(b) and (c)). At a critical scattering length $c_{\rm sc, crit,
1} = -0.93063$ the bifurcation merges with the bifurcation B1 (see
\fref{fig:line2}(a)). For even smaller scattering lengths the bifurcation moves
back to $\Gamma = 0$ along the state s2. The pitchfork bifurcation is now formed
between the states s2, s3, and s6 (see \fref{fig:line2}(b)). However, before the
bifurcation reaches $\Gamma = 0$ another critical scattering length is reached.
At $c_{\rm sc, crit, 2} = -0.93365$ the bifurcation B2 merges with the
bifurcations B3a and B3b (see \fref{fig:line2}(c)). At this point the behaviour
of the bifurcation is altered. For larger scattering lengths at smaller
$\Gamma$ the participating states exist only for the bicomplex equations (see
\fref{fig:line2}(b)), but for larger $\Gamma$ the states exist in the complex
Gross-Pitaevskii equation (see \fref{fig:line2}(d)) without the bicomplex
extension. If the scattering length is smaller than $c_{\rm sc, crit, 2}$ this
behaviour is mirrored with respect to the $\Gamma$-axis. Until the bifurcation
vanishes for smaller scattering lengths at $\Gamma = 0$ it is composed of the
states s2, s4, and s5. If the bifurcation B2 is compared with the bifurcation O2
of the Bose-Einstein condensate without long-range interactions (see
\fref{fig:old}), the first change (the merger of B1 and B2) can also be
observed. However, the second change in behaviour is a new effect since the
bifurcations B3a and B3b do not exist without long-range interactions.
We have seen that the merger of the bifurcations B2, B3a, and B3b changes the
behaviour of bifurcation B2. The properties of the tangent bifurcations B3a and
B3b are altered. For scattering lengths greater than $c_{\rm sc, crit, 2}$ the
bifurcations B3a and B3b are bicomplex for larger $\Gamma$ (i.e., the states
have components of the imaginary units $\jj$ and $\kk$). However, for $\Gamma$
below the critical value only the real component and the component with complex
unit $\ii$ is nonzero (see \fref{fig:line2}(b)). For smaller values of the
scattering lengths the states in both $\Gamma$ regions are bicomplex, i.e., they
only exist as solutions of the analytically continued Gross-Pitaevskii equation
(see \fref{fig:line2}(d)).
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig5.pdf}
\caption{
\label{fig:scgamma}
Positions of the bifurcations and exceptional points in the $\Gamma$-$c_{\rm
sc}$-parameter space. In the inset the merger of multiple bifurcations can
be observed. Note that at the point P3 no merger occurs. There merely exist
two bifurcations between different states at the same point in parameter
space. }
\end{figure}
We have found that the critical scattering length parameters, which divide the
parameter region with different behaviours, are related with the merger of
multiple bifurcations. In \fref{fig:scgamma} the parameter pairs of the
scattering length $c_{\rm sc}$ and $\Gamma$ are shown at which the different
bifurcations occur. One observes three points at which the parameters of two
different bifurcations become identical. Point P3 is not special, the states
which are involved in the two bifurcations (B1 und B3) have different
eigenvalues and wave functions. Therefore just two independent bifurcations
occur at the same parameter pair. By contrast at P2 the bifurcations B1 and B2
are joined into one bifurcation. The bifurcation scenario is shown in
\fref{fig:line2}(a). Another bifurcation merger appears for the parameters at
point P1. The resulting bifurcation, which consits of B2, B3a, and B3b is shown
in \fref{fig:line2}(c). These merger points are also of special interest
because they have the prerequesites necessary that exceptional points of higher
order can appear.
\subsection{Exchange behaviour of the states around the exceptional points}
We now examine which signatures of exceptional points can be
observed. In \cite{Dem12,Guthrlein2013} it was discussed that a complex
encircling of a higher order exceptional point does not have to exhibit an
exchange of all states which are involved in the exceptional point. Using
different parameters to encircle the EP can show different exchange behaviour.
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig6.pdf}
\caption{
\label{fig:ep2}
Characteristic exchange of states when an exceptional
point is encircled in the complex parameter space. In this case the
bifurcation B1 (subfigure (a), $\Gamma_{\rm center} = 1.03904$, $r = 10^{-3}$),
bifurcation B2 (subfigure (b), $\Gamma_{\rm center} = 0.71967$, $r = 2\times10^{-3}$),
bifurcation B3a (subfigure (c), $\Gamma_{\rm center} = 1.10296$, $r = 10^{-4}$) and
bifurcation B3b (subfigure (d), $\Gamma_{\rm center} = 1.10296$, $r = 10^{-4}$)
are encircled on the parameter path
$\Gamma(x) = \Gamma_{\rm center} + r \ee^{\jj \phi}$ for $\phi \in [0,2\pi]$.
All plots were calculated for a scattering length of $c_{\rm sc} = -0.9$.
}
\end{figure}
The second-order exceptional points of the tangent bifurcations B1, B3a, and B3b
show the expected square root behaviour by exchanging each state with the other
(see \fref{fig:ep2}(a,c,d)) when the point is encircled in the complex parameter
space. On the other hand the third-order exceptional point at the bifurcation B2
only shows the exchange of two states. For the encircling in the complex
$\Gamma$ a cubic root exchange behaviour cannot be observed (see
\fref{fig:ep2}(b)).
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig7.pdf}
\caption{
\label{fig:ep4ab}
(a) and (c) show the exceptional point at the coalescence of bifurcation B1
and B2 at $c_{\rm sc,crit,1} = -0.93063$ and $\Gamma = 1.03531$. (b) and
(d) show the exceptional point for the merger of the bifurcations B2, B3a
and B3b at $c_{\rm sc,crit,2} = -0.93365$ and $\Gamma = 1.03215$. The
exceptional points are encircled in the complex parameter space of $\Gamma$
(subfigure (a) and (b)) with a radius of $r = 10^{-3}$. In (c) and (d) the
exceptional points are encircled in the complex parameter space of $c_{\rm
sc}$. The radius of the circle is $r = 4.65\times10^{-4}$ in (c) and $r =
2\times10^{-4}$ in (d). One exceptional point (subfigure (b) and (d))
exhibits the full exchange behaviour of a fourth-order exceptional point, in
which the original situation is not reached until the point is encircled
four times. By contrast the other exceptional point in (a) and (c) has two
separated groups of states which only exchange with each other, and
therefore the starting condition is already reached again after two full
orbits. }
\end{figure}
In \fref{fig:ep4ab}(a) we show the exchange behaviour of the states involved in
P2 where the bifurcations B1 and B2 merge. When encircling the exceptional point
in a complex $\Gamma$-parameter plane an exchange within pairs of states is
found, however, these two exchanges are separated and no exchange between all
four states can be observed. Therefore it is unclear whether these are two
second-order exceptional points or one fourth-order exceptional point. Also by
encircling the critical point in the complex scattering length plane does not
change the qualitative behaviour, again only two states exchange. To prove that
this must be an EP4, one must search for further complex perturbation
parameters. However, since for an exceptional point of order $n$, all $n$
eigenvalues and eigenstates must coalesce \cite{Kato66} we examine the wave
functions of the participating states and they all coalesce at the critical
point (which means that for the ansatz of coupled Gaussians all Gaussian
parameters must be the same, which is indeed the case).
The same examination can be performed for the merger of the bifurcations B2, B3a,
and B3b (point P1). At this critical point five eigenvalues coalesce. A circle
in the complex $\Gamma$ plane reveals the signature of four exchanging states
(see \fref{fig:ep4ab}(b)). Again the circle can be repeated in the complex
scattering plane (\fref{fig:ep4ab}(c)) resulting in the same exchange
behaviour. In this case the question arises whether this is an EP4 or an EP5.
The exchange behaviour proves that the order of the exceptional point must be at
least four. All wave functions of the participating states coalesce at the
critical point, however, to finally decide whether this is an EP5 further
perturbation parameter must be examined.
\begin{figure}
\includegraphics[width=0.98\textwidth]{fig8.pdf}
\caption{
\label{fig:ep5sym}
Mean-field energies $\bpparam{E_{\rm mf}}$ when the exceptional point where
the bifurcations B2, B3a, and B3b coalesce is encircled in the complex
symmetry-parameter space $s$ defined in \eref{eq:symparapath}. }
\end{figure}
We introduce a new asymmetry parameter in \eref{eq:gpe}, which lifts or lowers
the potential wells
\begin{eqnarray}
c_{{\rm g},1} = (V_0 + s) + \ii \Gamma
~{\rm and}~
c_{{\rm g},2} = (V_0 - s) + \ii \Gamma.
\end{eqnarray}
The new parameter $s$ allows us to break the remaining trap symmetry of the
system, e.g.\ one potential well is deepend, while the other is flattened. We
encircle the bifurcation for the scattering length $c_{\rm sc,crit,2} =
-0.93365$ and the coupling parameter $\Gamma = 1.03215$ on the path
\begin{eqnarray}
\label{eq:symparapath}
s = 5 \times 10^{-5} \ee^{\jj\phi} ~{\rm for}~
\phi \in [0, 2\pi]
\end{eqnarray}
and observe a permutation of all five states with each other (see
\fref{fig:ep5sym}). Thus we have proven the existence of an exceptional
point of order five in this system.
\section{Summary and conclusion}
We have shown how the Gross-Pitaevskii equation for dipolar Bose-Einstein
condensates can be analytically continued with an ansatz of coupled Gaussians
using bicomplex numbers. Especially the representation in the idempotent basis
of the bicomplex numbers can be used to separate the bicomplex equations into
twice the number of coupled complex equations. This allows for the reuse of
an algorithm developed for the integration of the complex equations.
By solving the analytically continued equations we have
shown that a dipolar Bose-Einstein condensate in a \texorpdfstring{$\mathcal{PT}$}{PT}-symmetric trap
exhibits a much richer bifurcation scenario than a
condensate without long-range interactions. Most of the properties examined in
this paper have revealed interesting mathematical relations in the bicomplex
parameter space, which is experimentally inaccessible. However, the
understanding of the bifurcation scenrio is important since bifurcations
influence crucially the stabiliy of a condensate \cite{Haag14}.
Furthermore the additional states, while not experimentally accessible,
allow us to examine the order of the exceptional points. Higher-order
exceptional points were found in other systems such as EP3s in the hydrogen
atom \cite{Holger2007} or in Bose-Einstein condensates \cite{Heiss13a}. Even
exeptional points of arbitrary order were shown to exist in a Bose-Hubbard
model \cite{Graefe08a}. However, no EPs with an order higher than three were
known for a BEC described by a mean-field model.
With the new (bicomplex) states we have demonstrated the properties of
the exceptional points associated with the bifurcations (\sref{sec:res}).
In particular the critical points where two or three bifurcation coalesce were
examined. By having shown that a parameter exists for which the
encircling of the critical value results in the permutation of all five states
participating in the exceptional point, we have indeed proven a fifth-order
exceptional point in dipolar Bose-Einstein condensates.
It was shown \cite{Brinker2015} that the signatures of exceptional points
influence the collapse dynamics of Bose-Einstein condensates. With the help of
a harmonic inversion analysis \cite{Fuchs2014} the order of degeneracies can be
translated into polynomial terms in the time-evolution. Further studies will
reveal the influence of such higher-order exceptional points on the collapse
dynamics.
\section*{References}
\bibliographystyle{unsrt}
|
\section{Introduction}
In recent years the \textit{Learning with Errors} (LWE) problem received a lot of attention in the cryptographic research community. As instance of the LWE problem we are given a uniformly distributed matrix $\textbf{\upshape A}\in\mathbb{Z}_q^{\lambda\times\kappa}$ and a noisy codeword $\textbf{\upshape y}=\textbf{\upshape Ax}+\textbf{\upshape e}\in\mathbb{Z}_q^{\lambda}$ with an error term $\textbf{\upshape e}\in\mathbb{Z}_q^{\lambda}$ chosen from a proper error distribution $\chi^{\lambda}$ and an unknown $\textbf{\upshape x}\in\mathbb{Z}_q^{\kappa}$. The task is to find the correct vector $\textbf{\upshape x}$. In the decisional version of this problem (DLWE problem) we are given $(\textbf{\upshape A}, \textbf{\upshape y})$ and have to decide whether $\textbf{\upshape y}=\textbf{\upshape Ax}+\textbf{\upshape e}$ or $\textbf{\upshape y}$ is a uniformly distributed vector in $\mathbb{Z}_q^{\lambda}$. In \cite{42} a search-to-decision reduction was provided to show that the two problems are essentially equivalent in the worst case and in \cite{43} a sample preserving search-to-decision reduction was provided for certain cases showing the equivalence in the average case. Moreover, in \cite{42} the average-case-hardness of the search problem was established by the construction of an efficient quantum algorithm for worst-case-lattice-problems using an efficient solver of the LWE problem if the error distribution $\chi$ is a \textit{discrete Gaussian} distribution. Accordingly, most cryptographic applications of the LWE problem used a discrete Gaussian error distribution for their constructions.\\
In this work we are concerned with the question whether the hardness of the LWE problem can be established for other discrete distributions, especially for reproducible distributions, i.e. distributions that are closed under convolution (- the discrete Gaussian distribution is not reproducible). This question is motivated by the following application. In \cite{2} the notion of \textit{Private Stream Aggregation} (PSA) was introduced. A PSA scheme is a cryptographic protocol between $n$ users and an untrusted aggregator. It enables each user to securely send encrypted time-series data to the aggregator. The aggregator is then able to decrypt the aggregate of all data in each time step, but cannot retrieve any further information about the individual data. In \cite{40} it was shown that a PSA scheme can be built upon any key-homomorphic weak pseudo-random function and some security guarantees were provided. In this paper we instantiate a concrete PSA scheme with a key-homomorphic weak pseudo-random function constructed from the DLWE problem.\\
A PSA scheme enables the users to output statistics over their data under \textit{differential privacy}. This notion was introduced in \cite{8} and is a measure for statistical disclosure of private data. Usually, $(\epsilon,\delta)$-differential privacy is preserved by using a mechanism that adds properly distributed noise to the statistics computed over a database with individual private data of users. In most of the works on differentialy privacy, these mechanisms are considered in the \textit{centralised} setting, where a trusted authority collects the individual data in the clear and performs the perturbation process. In contrast, a PSA scheme allows the users to perform differentially private data analysis in the \textit{distributed setting}, i.e. without the need of relying on a trusted authority. In light of that, the \textit{Skellam mechanism} was introduced in \cite{40} and shown to preserve differential privacy. The advantage of the Skellam mechanism over other mechanisms (like the Laplace \cite{8}, the Exponential \cite{9}, the Geometric \cite{52}, the Gaussian \cite{14} or the Binomial \cite{14}) mechanisms is that it is discrete (enabling cryptographic operations), maintains relatively high accuracy and is reproducible. This property allows all users to generate noise of small variance, that sums up to the value for the required level of differential privacy. Therefore the Skellam mechanism is well-suited for an execution through a PSA scheme.\\
We will take advantage of these properties of the Skellam distribution for our DLWE-based PSA scheme by using errors following the symmetric Skellam distribution $\text{\upshape Sk}_\mu$ with variance $\mu$ rather than the discrete Gaussian distribution. Therefore we need to show the average-case-hardness of the LWE problem with errors taken from the Skellam distribution. Now we can state the main theorem that will be shown in this work.
\begin{Thm}\label{lweskthm}
Let $\kappa$ be a security parameter
and let $\lambda=\lambda(\kappa)=\text{\upshape\sffamily poly}(\kappa)$ with $\lambda>3\kappa$. Let $q=q(\kappa)=\text{\upshape\sffamily poly}(\kappa)$ be a sufficiently large prime modulus and $\rho>0$ such that $\rho q\geq 2\lambda\sqrt{\kappa}$. If there exists a PPT-algorithm that solves the $\text{\upshape LWE}(\kappa,\lambda,q,{\text{\upshape Sk}_{(\rho q)^2/4}})$ problem with non-negligible probability, then there exists an efficient quantum-algorithm that approximates the decision-version of the shortest vector problem ($\text{\upshape GAPSVP}$) and the shortest independent vectors problem ($\text{\upshape SIVP}$) to within $\tilde{O}(\lambda\kappa/\rho)$ in the worst case.
\end{Thm}
Based on the same assumptions, the decisional version $\text{\upshape DLWE}(\kappa,\lambda,q,{\text{\upshape Sk}_{(\rho q)^2/4}})$ of the problem is also hard due to the search-to-decision reduction in \cite{43}. Hence, the error terms in our DLWE-based PSA scheme are used for two tasks: establishing the cryptographic security of the scheme and the distributed generation of noise for preserving differential privacy.\\
Our proof of Theorem \ref{lweskthm} is inspired by techniques used in \cite{39} where a variant of the LWE problem with uniform errors on a small support is shown to be hard. As in \cite{39}, we will construct a \textit{lossy code} for the error distribution $\text{\upshape Sk}_\mu$ from the LWE problem with discrete Gaussian errors. Variants of lossy codes were first used in \cite{53} and since then had applications in different hardness reductions, such as the reduction from the LWE problem to the Learning with Rounding problem in \cite{54}. Lossy codes are pseudo-random codes seeming to be good codes. However, encoding messages with a lossy code and adding certain errors annihilates any information on the message. On the other hand, encoding the same message using a truly random code and adding the same type of error preserves the message. We will conclude that recovering the message when encoding with a random code and adding noise must be computationally hard. If this was not the case, lossy codes could be efficiently distinguished from random codes, contradicting the pseudo-randomness-property of lossy codes. As in \cite{39} and opposed to the LWE problem with discrete Gaussian errors, our worst-to-average case reduction depends on the number $\lambda$ of LWE-samples. Thus, we will consider a $\lambda$-bounded LWE problem, where $\lambda$ has a fixed $\text{\upshape\sffamily poly}(\kappa)$ upper bound. This does not restrict our application to PSA and differential privacy, since we will identify $\lambda$ with the total number of queries processed during the execution of PSA.\\
Essential definitions and known facts about used distributions, the Learning with Errors problem and lossy codes are given in Sections \ref{nad} and \ref{fac} respectively. The proof of Theorem \ref{lweskthm} is provided in Section \ref{hardn} and in Section \ref{applic} we show how this result can be applied to the construction of a PSA scheme, thus yielding a prospective post-quantum cryptographic protocol for data analyses under differential privacy.
\begin{Rem} The result from $\cite{39}$ does not provide a proper error distribution for our DLWE-based PSA scheme since a differentially private mechanism with uniform noise provides no accuracy to statistical data analyses.
\end{Rem}
\begin{Rem} The original result from $\cite{42}$ states that the LWE problem is hard in the set $\mathbb{T}=\mathbb{R}/\mathbb{Z}$ when the noise is distributed according to the \mbox{\em continuous} Gaussian distribution (with a certain bound on the variance) modulo $1$. Although the continuous Gaussian distribution is reproducible as well, it does not seem to be a good idea to instantiate it instead of the Skellam distribution in the DLWE-based PSA scheme: For data processing reasons the values would have to be discretised. The resulting noise would follow a discrete Gaussian distribution which is not reproducible any more. The larger the number of users the more independent samples of discrete Gaussian noise would have to be added together. Moreover, if the total noise has to be invariant to the number of users (e.g. in order to keep the accuracy of the statistics), then the discretisation causes a stronger deviation from the discrete Guassian distribution the larger the number of users becomes. Therefore the deviation from a truely $(\epsilon,\delta)$-differentially private analysis would scale with the number of users.
\end{Rem}
\section{Preliminaries}\label{nad}
Let $q>2$ be a prime. We handle elements from $\mathbb{Z}_q$ as their central residue-class representation. This means that $x^\prime\in\mathbb{Z}_q$ is identified with $x\equiv x^\prime$ mod $q$ for $x\in\{-(q-1)/2,\ldots,(q-1)/2\}$ thereby lifting $x^\prime$ from $\mathbb{Z}_q$ to $\mathbb{Z}$.
\subsection{Distributions}
Let $X$ be a set. We denote by $\mathcal{U}(X)$ the uniform distribution on $X$. Let $\chi$ be a distribution (on $X$). We denote by $x\leftarrow\chi$ (or sometimes $x\leftarrow\chi(X)$) the sampling of $x$ (from $X$) according to $\chi$. If $A\leftarrow\chi^{a\times b}$ (or $A\leftarrow\chi(X^{a\times b})$) then $A$ is an $a\times b$-matrix constructed by picking every entry independently (from $X$) according to the distribution $\chi$.
\begin{Def}[Discrete Gaussian Distribution\cite{42}]
Let $q$ be an integer and let $\Phi_s$ denote the normal distribution with variance $s^2/(2\pi)$. Let $\overline{\Psi}_\alpha$ denote the discretised Gaussian distribution with variance $(\alpha q)^2/(2\pi)$, i.e. $\overline{\Psi}_\alpha$ is sampled by taking a sample from $\Phi_{\alpha q}$ and performing a randomised rounding $\cite{44}$. Let $D_\nu$ be the discretised Gaussian distribution with variance $\nu$, i.e. $D_\nu=\overline{\Psi}_{\sqrt{2\pi\nu}/q}$.
\end{Def}
\begin{Def}[Skellam Distribution \cite{29}]\label{skellam} Let $\mu_1$, $\mu_2> 0$. A discrete random variable $X$ is drawn according to the Skellam distribution with parameters $\mu_1,\mu_2$ (short: $X\leftarrow \text{\upshape Sk}_{\mu_1,\mu_2}$) if it has the following probability distribution function $\psi_{\mu_1,\mu_2}\colon\mathbb{Z}\mapsto\mathbb{R}$:
\[\psi_{\mu_1,\mu_2}(k)=\exp(-(\mu_1+\mu_2))\left(\frac{\mu_1}{\mu_2}\right)^{k/2}I_k(2\sqrt{\mu_1\mu_2}),\]
where $I_k$ is the modified Bessel function of the first kind (see pages $374$--$378$ in $\cite{28}$).
\end{Def}
A random variable $X\leftarrow \text{\upshape Sk}_{\mu_1,\mu_2}$ has variance $\mu_1+\mu_2$ and can be generated as the difference of two random variables drawn according to the Poisson distribution of mean $\mu_1$ and $\mu_2$, respectively \cite{29}. Note that the Skellam distribution is not generally symmetric. However, we only consider the particular case $\mu_1=\mu_2=\mu/2$ and refer to this symmetric distribution as $\text{Sk}_\mu = \text{Sk}_{\mu/2,\mu/2}$.
\subsection{Learning with Errors}\label{lwesubsec}
We will consider a $\lambda$-bounded LWE problem, where the adversary is given $\lambda(\kappa) = \text{\upshape\sffamily poly}(\kappa)$ samples (which we can write conveniently in matrix-form). As observed in \cite{39}, this consideration poses no restrictions to most cryptographic applications of the LWE problem, since they require only an a-priori fixed number of samples. In our application to differential privacy (Section \ref{applic}) we will identify $\lambda$ with the number of queries in a pre-defined time-series.\\
\noindent\textit{Problem} $1$. $\boldsymbol\lambda$\textbf{-bounded LWE Search-Problem, Average-Case Version.} Let $\kappa$ be a security parameter, let $\lambda = \lambda(\kappa) = \text{\upshape\sffamily poly}(\kappa)$ and $q = q(\kappa)\geq 2$ be integers and let $\chi$ be a distribution on $\mathbb{Z}_q$. Let $\textbf{x}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa)$, let $\textbf{A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ and let $\textbf{e}\leftarrow\chi^\lambda$. The goal of the $\text{\upshape LWE}(\kappa,\lambda,q,\chi)$ problem is, given $(\textbf{A}, \textbf{Ax} + \textbf{e})$, to find $\textbf{x}$.\\
\noindent\textit{Problem} $2$. $\boldsymbol\lambda$\textbf{-bounded LWE Distinguishing-Problem.} Let $\kappa$ be a security parameter, let $\lambda = \lambda(\kappa) = \text{\upshape\sffamily poly}(\kappa)$ and $q = q(\kappa)\geq 2$ be integers and let $\chi$ be a distribution on $\mathbb{Z}_q$. Let $\textbf{x}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa)$, let $\textbf{A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ and let $\textbf{e}\leftarrow\chi^\lambda$. The goal of the $\text{\upshape DLWE}(\kappa,\lambda,q,\chi)$ problem is, given $(\textbf{A}, \textbf{y})$, to decide whether $\textbf{y}=\textbf{Ax} + \textbf{e}$ or $\textbf{y}=\textbf{u}$ with $\textbf{u}\leftarrow\mathcal{U}(\mathbb{Z}_q^\lambda)$.\\
\subsection{Entropy and Lossy Codes}
We introduce the map-conditional entropy as starting point for our technical tools. It can be seen as a measure for ambiguity.
\begin{Def}[Map-conditional entropy]
Let $\chi$ be a probability distribution with finite support and let $X\leftarrow\chi$. Let $Supp(\chi)$ be the support of $\chi$, $\xi\in Supp(\chi)$ and let $f,g$ be two (possibly randomised) maps on the domain $Supp(\chi)$. The $(f,g,\xi)$-conditional entropy $H_{f,g,\xi} (X)$ of $X$ is defined as
\[H_{f,g,\xi}(X)=-\log_2(\Pr[X=\xi\,|\,f(X)=g(\xi)]).\]
\end{Def}
In the remainder of the paper we will see $f=f_{\textbf{A},\textbf{e}}$ and $g=g_{\textbf{A},\textbf{e}}$ as maps to the set of LWE instances, i.e.
\[f_{\textbf{A},\textbf{e}}(\textbf{y})=g_{\textbf{A},\textbf{e}}(\textbf{y})=\textbf{Ay} + \textbf{e}.\]
In this work we consider the $(f_{\textbf{A},\textbf{e}},f_{\textbf{A},\tilde{\textbf{e}}},\tilde{\textbf{x}})$-conditional entropy
\[H_{f_{\textbf{A},\textbf{e}},f_{\textbf{A},\tilde{\textbf{e}}},\tilde{\textbf{x}}}(\textbf{x})=-\log_2(\Pr[\textbf{x}=\tilde{\textbf{x}}\,|\,\textbf{Ax} + \textbf{e}=\textbf{A}\tilde{\textbf{x}} + \tilde{\textbf{e}}]),\]
of a random variable $\textbf{x}$, i.e. the entropy of $\textbf{\upshape x}$ given that a LWE instance generated with $(\textbf{\upshape A},\textbf{\upshape x},\textbf{e})$ is equal to another LWE instance generated with $(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{e}})$. Now we provide the notion of lossy codes, which is a main technical tool used in the proof of the hardness result.
\begin{Def}[Families of Lossy Codes \cite{39}]\label{lossydef}
Let $\kappa$ be a security parameter, let $\lambda = \lambda(\kappa) = \text{\upshape\sffamily poly}(\kappa)$ and let $q=q(\kappa)\geq 2$ be a modulus, $\Delta=\Delta(\kappa)$ and let $\chi$ be a distribution on $\mathbb{Z}_q$. Let $\{\mathcal{C}_{\kappa,\lambda,q}\}$ be a family of distributions where $\mathcal{C}_{\kappa,\lambda,q}$ is defined on $\mathbb{Z}_q^{\lambda\times\kappa}$. The distribution family $\{\mathcal{C}_{\kappa,\lambda,q}\}$ is $\Delta$-lossy for the error distribution $\chi$, if the following hold:
\begin{enumerate}
\item $\mathcal{C}_{\kappa,\lambda,q}$ is pseudo-random: It holds that $\mathcal{C}_{\kappa,\lambda,q}\approx_c\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$.
\item $\mathcal{C}_{\kappa,\lambda,q}$ is lossy: Let $f_{\textbf{\upshape B},\textbf{\upshape b}}(\textbf{\upshape y})=\textbf{\upshape B}\textbf{\upshape y} + \textbf{\upshape b}$. Let $\textbf{\upshape A}\leftarrow\mathcal{C}_{\kappa,\lambda,q}$, $\tilde{\textbf{\upshape x}}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa), \tilde{\textbf{\upshape e}}\leftarrow\chi^\lambda$, let $\textbf{\upshape x}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa)$ and $\textbf{\upshape e}\leftarrow\chi^\lambda$. Then it holds that
\[\Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}[H_{f_{\textbf{\upshape A},\textbf{\upshape e}},f_{\textbf{\upshape A},\tilde{\textbf{\upshape e}}},\tilde{\textbf{\upshape x}}}(\textbf{\upshape x})\geq\Delta]\geq 1-\text{\upshape\sffamily neg}(\kappa).\]
\item $\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ is non-lossy: Let $f_{\textbf{\upshape B},\textbf{\upshape b}}(\textbf{\upshape y})=\textbf{\upshape B}\textbf{\upshape y} + \textbf{\upshape b}$. Let $\textbf{\upshape A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$, $\tilde{\textbf{\upshape x}}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa), \tilde{\textbf{\upshape e}}\leftarrow\chi^\lambda$, let $\textbf{\upshape x}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa)$ and $\textbf{\upshape e}\leftarrow\chi^\lambda$. Then it holds that
\[\Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}[H_{f_{\textbf{\upshape A},\textbf{\upshape e}},f_{\textbf{\upshape A},\tilde{\textbf{\upshape e}}},\tilde{\textbf{\upshape x}}}(\textbf{\upshape x})=0]\geq 1-\text{\upshape\sffamily neg}(\kappa).\]
\end{enumerate}
\end{Def}
Note that our definition of lossy codes deviates from the definition of lossy codes provided in \cite{39}, since we use another type of entropy (which in general may be larger than the conditional min-entropy considered in \cite{39}). We will see that our notion of map-conditional entropy suffices for showing the hardness of the LWE problem.
\section{Basic Facts}\label{fac}
\subsection{Facts about the used Distributions}
We need to find a value such that a random variable distributed according to the discrete Gaussian distribution exceeds this value only with negligible probability.
\begin{Lem}[Bound on the discrete Gaussian distribution]\label{disgausbound}
Let $\kappa$ be a security parameter, let $s=s(\kappa)=\omega(\log(\kappa))$ and let $\nu=\nu(\kappa)=\text{\upshape\sffamily poly}(\kappa)$. Let $\zeta=\zeta(\kappa)>0$ be an integer. Let $g_1,\ldots,g_\zeta\leftarrow D_\nu$. Then \[\Pr\left[\left|g_1+\ldots +g_\zeta\right|>\sqrt{\zeta\nu s}\right]\leq\text{\upshape\sffamily neg}(\kappa).\]
\end{Lem}
\begin{proof}
Since $g_1,\ldots,g_\zeta$ are independent sub-gaussian random variables with sub-gaussian parameter $\sqrt{\nu}$, the result follows from an application of the Hoeffding-type inequality for sub-gaussian random variables (see Proposition $5.10$ in \cite{51}).
\end{proof}
We use the fact that the sum of independent Skellam random variables is a Skellam random variable, i.e. the Skellam distribution is reproducible.
\begin{Lem}[Reproducibility of $\text{\upshape Sk}_{\mu_1,\mu_2}$ \cite{29}]\label{sksum} Let $X\leftarrow \text{\upshape Sk}_{\mu_1,\mu_2}$ and $Y\leftarrow \text{\upshape Sk}_{\mu_3,\mu_4}$ be independent random variables. Then $Z:=X+Y$ is distributed according to $\text{\upshape Sk}_{\mu_1+\mu_3,\mu_2+\mu_4}$.
\end{Lem}
An induction step shows that the sum of $n$ i.i.d. symmetric Skellam random variables with variance $\mu$ is a symmetric Skellam random variable with variance $n\mu$. For our analysis, we will use the following bound on the ratio of modified Bessel functions of the first kind.
\begin{Lem}[Bound on $I_k(\mu)$ \cite{27}]\label{modbesrat} For real $k$, let $I_k(\mu)$ be the modified Bessel function of the first kind and order $k$. Then
\[1>\frac{I_k(\mu)}{I_{k-1}(\mu)}>\frac{-k+\sqrt{k^2+\mu^2}}{\mu},\, k\geq 0.\]
\end{Lem}
Moreover, we need a proper bound on the symmetric Skellam distribution that holds with probability exponentially close to $1$.
\begin{Lem}[Bound on the Skellam distribution]\label{skellbound}
Let $\kappa$ be a security parameter, let $s=s(\kappa)=\omega(\log(\kappa))$ and let $\mu=\mu(\kappa)=\text{\upshape\sffamily poly}(\kappa)$ with $\mu>s>0$. Let $X\leftarrow\text{\upshape Sk}_\mu$. Then \[\Pr[X>s\sqrt{\mu}]\leq\text{\upshape\sffamily neg}(\kappa).\]
\end{Lem}
\begin{proof}
Applying the Laplace transform and the Markov's inequality we obtain for any $t>0$,
\[\Pr[X>s\sqrt{\mu}]=\Pr[e^{tX}>e^{ts\sqrt{\mu}}]\leq\frac{\mathbb{E}[e^{tX}]}{e^{ts\sqrt{\mu}}}.\]
As shown in \cite{30}, the moment generating function of $X\sim \text{Sk}(\mu)$ is
\[\mathbb{E}[e^{tX}]=e^{-\mu(1-\cosh(t))},\]
where $\cosh(t)= (e^t + e^{-t})/2$. Hence, we have
\[\Pr[X>s\sqrt{\mu}]\leq e^{-\mu(1-\cosh(t))-ts\sqrt{\mu}}< e^{-s(1-\cosh(t))-ts^{3/2}}.\]
Fix $t=\operatorname{arsinh}(1/\sqrt{s})$. Then
\begin{align*} \Pr[X>s\sqrt{\mu}]< & e^{-s(1-\sqrt{1+1/s})-s^{3/2}\operatorname{arsinh}(1/\sqrt{s})}\\
= & e^{-s}\cdot e^{s\cdot(\sqrt{1+1/s}-\sqrt{s}\operatorname{arsinh}(1/\sqrt{s}))}\\
< & e^{-s}\cdot e^{2/3}\\
= & \text{\upshape\sffamily neg}(\kappa).
\end{align*}
To see the last inequality, observe that the function $f(s)=s\cdot(\sqrt{1+1/s}-\sqrt{s}\operatorname{arsinh}(1/\sqrt{s}))$ is monotonically increasing and its limit is $2/3$.
\end{proof}
\subsection{Facts about Learning with Errors}
In \cite{42}, Regev established worst-to-average-case connections between conjecturally hard lattice problems and the $\text{\upshape LWE}(\kappa,\lambda,q,D_\nu)$ problem.
\begin{Thm}[Worst-to-Average Case \cite{42}]\label{wtacr} Let $\kappa$ be a security parameter and let $q = q(\kappa)$ be a modulus, let $\alpha=\alpha(\kappa)\in(0,1)$ be such that $\alpha q> 2\sqrt{\kappa}$. If there exists a probabilistic polynomial-time algorithm solving the $\text{\upshape LWE}(\kappa,\lambda,q,D_{(\alpha q)^2/(2\pi)})$ problem with non-negligible probability,
then there exists an efficient quantum algorithm that approximates the decision-version of the shortest vector problem ($\text{\upshape GAPSVP}$) and the shortest independent vectors problem ($\text{\upshape SIVP}$) to within $\tilde{O}(\kappa/\alpha)$ in the worst case.
\end{Thm}
We use the search-to-decision reduction from \cite{43} basing the hardness of Problem $2$ on the hardness of Problem $1$ which works for any error distribution $\chi$ and is sample preserving.
\begin{Thm}[Search-to-Decision \cite{43}]\label{lwestod} Let $\kappa$ be a security parameter, $q = q(\kappa) = \text{\upshape\sffamily poly}(\kappa)$ a prime modulus and let $\chi$ be any distribution on $\mathbb{Z}_q$. Assume there exists a probabilistic polynomial-time distinguisher that solves the $\text{\upshape DLWE}(\kappa,\lambda,q,\chi)$ problem with non-negligible success-probability, then there exists a probabilistic polynomial-time adversary that solves the $\text{\upshape LWE}(\kappa,\lambda,q,\chi)$ problem with non-negligible success-probability.
\end{Thm}
Finally, we provide a matrix version of Problem $2$. The hardness of this version can be shown by using a hybrid argument as pointed out in \cite{39}.
\begin{Lem}[Matrix version of LWE]\label{lwematver}
Let $\kappa$ be a security parameter, $\lambda=\lambda(\kappa)= \text{\upshape\sffamily poly}(\kappa)$, $\kappa^\prime=\kappa^\prime(\kappa)=\text{\upshape\sffamily poly}(\kappa)$. Assume that the $\text{\upshape DLWE}(\kappa,\lambda,q,\chi)$ problem is hard. Then $(\textbf{\upshape A},\textbf{\upshape AX}+\textbf{\upshape E})$ is pseudo-random, where $\textbf{\upshape A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa}), \textbf{\upshape X}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\kappa\times\kappa^\prime})$ and $\textbf{\upshape E}\leftarrow\chi^{\lambda\times\kappa^\prime}$.
\end{Lem}
\subsection{Facts about Lossy Codes}
We will use the fact that the existence of a lossy code for an error distribution implies the hardness of the associated decoding problems. This means, solving the LWE problem is hard, even though with overwhelming probability the secret is information-theoretically unique. In \cite{39} it was shown that solving the LWE problem for $\chi$ is hard if there exists a lossy code for $\chi$ in the sense of \cite{39}. Here we prove this statement for our definition of lossy codes. The proof is very similar to the one in \cite{39}.
\begin{Thm}[Lossy code gives hard LWE]\label{lossythm
Let $\kappa$ be a security parameter, let $\lambda=\lambda(\kappa)=\text{\upshape\sffamily poly}(\kappa)$ and let $q=q(\kappa)$ be a modulus. Let the distribution $\chi$ on $\mathbb{Z}_q$ be efficiently samplable. Let $\Delta=\Delta(\kappa)=\omega(\log(\kappa))$. Then the $\text{\upshape LWE}(\kappa,\lambda,q,\chi)$ problem is hard, given that there exists a family $\{\mathcal{C}_{\kappa,\lambda,q}\}\subseteq\mathbb{Z}_q^{\lambda\times\kappa}$ of $\Delta$-lossy codes for the error distribution $\chi$.
\end{Thm}
\begin{proof}
Due to the non-lossiness of $\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ for $\chi$, instances of $\text{\upshape LWE}(\kappa,\lambda,q,\chi)$ have a unique solution with overwhelming probability. Now, let $\mathcal{T}$ be a probabilistic polynomial-time adversary solving the $\text{\upshape LWE}(\kappa,\lambda,q,\chi)$ problem with non-negligible probability $\sigma$. Using $\mathcal{T}$, we will construct a probabilistic polynomial-time distinguisher $\mathcal{D}_{\mbox{\scriptsize LWE}}$ distinguishing $\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ and $\{\mathcal{C}_{\kappa,\lambda,q}\}$ with non-negligible advantage.\\
Let $\mathcal{D}_{\mbox{\scriptsize LWE}}$ be given $\textbf{\upshape A}\in\mathbb{Z}^{\lambda\times\kappa}$ as input. It must decide, whether $\textbf{\upshape A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ or $\textbf{\upshape A}\leftarrow\{\mathcal{C}_{\kappa,\lambda,q}\}$. Therefore, $\mathcal{D}_{\mbox{\scriptsize LWE}}$ samples $\tilde{\textbf{\upshape x}}\leftarrow\mathcal{U}(\mathbb{Z}^\kappa)$ and $\tilde{\textbf{\upshape e}}\leftarrow\chi^\lambda$. It runs $\mathcal{T}$ on input $(\textbf{\upshape A},\textbf{\upshape A}\tilde{\textbf{\upshape x}}+\tilde{\textbf{\upshape e}})$. Then $\mathcal{T}$ outputs some $\textbf{\upshape x}\in\mathbb{Z}^\kappa$. If $\textbf{\upshape x}=\tilde{\textbf{\upshape x}}$, then $\mathcal{D}_{\mbox{\scriptsize LWE}}$ outputs $1$, otherwise it outputs $0$.\\
If $\textbf{\upshape A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$, then $\tilde{\textbf{\upshape x}}$ is unique and then $\textbf{\upshape x}=\tilde{\textbf{\upshape x}}$ with probability $\sigma$. Therefore
\[\Pr[\mathcal{D}_{\mbox{\scriptsize LWE}}(\textbf{\upshape A})=1\,|\,\textbf{\upshape A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})]=\sigma.\]
If $\textbf{\upshape A}\leftarrow\{\mathcal{C}_{\kappa,\lambda,q}\}$, then $\mathcal{T}$ outputs the correct value with probability \[\Pr[\textbf{x}=\tilde{\textbf{x}}\,|\,\textbf{Ax} + \textbf{e}=\textbf{A}\tilde{\textbf{x}} + \tilde{\textbf{e}}]=2^{-H_{f_{\textbf{\upshape A},\textbf{\upshape e}},f_{\textbf{\upshape A},\tilde{\textbf{\upshape e}}},\tilde{\textbf{\upshape x}}}(\textbf{\upshape x})}\leq 2^{-\Delta},\]
with $f_{\textbf{\upshape B},\textbf{\upshape b}}(\textbf{\upshape y})=\textbf{\upshape B}\textbf{\upshape y} + \textbf{\upshape b}$. This holds with overwhelming probability over the choice of $(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})$. This probability is negligible in $\kappa$, since $\Delta=\omega(\log(\kappa))$. Therefore \[\Pr[\mathcal{D}_{\mbox{\scriptsize LWE}}(\textbf{\upshape A})=1\,|\,\textbf{\upshape A}\leftarrow\{\mathcal{C}_{\kappa,\lambda,q}\}]=\text{\upshape\sffamily neg}(\kappa)\]
and in conclusion $\mathcal{D}_{\mbox{\scriptsize LWE}}$ distinguishes $\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ and $\{\mathcal{C}_{\kappa,\lambda,q}\}$ with probability at least $\sigma-\text{\upshape\sffamily neg}(\kappa)$, which is non-negligible.
\end{proof}
We will use the fact that $\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ is always non-lossy if the corresponding error distribution $\chi$ can be bounded.
\begin{Lem}[Non-lossiness of $\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$]\label{nonloss}
Let $\kappa$ be a security parameter and $\chi$ a probability distribution on $\mathbb{Z}$. Say, the support of $\chi$ can be bounded by $r=r(\kappa)=\text{\upshape\sffamily poly}(\kappa)$. Moreover, let $q>(4r+1)^{1+\tau}$ for a constant $\tau>0$ and $\lambda=\lambda(\kappa)>(1+2/\tau)\kappa$. Let $f_{\textbf{\upshape B},\textbf{\upshape b}}(\textbf{\upshape y})=\textbf{\upshape B}\textbf{\upshape y} + \textbf{\upshape b}$. Let $\textbf{\upshape A}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$, $\tilde{\textbf{\upshape x}}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa), \tilde{\textbf{\upshape e}}\leftarrow\chi^\lambda$, let $\textbf{\upshape x}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa)$ and $\textbf{\upshape e}\leftarrow\chi^\lambda$. Then
\[\Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}[H_{f_{\textbf{\upshape A},\textbf{\upshape e}},f_{\textbf{\upshape A},\tilde{\textbf{\upshape e}}},\tilde{\textbf{\upshape x}}}(\textbf{\upshape x})=0]\geq 1-\text{\upshape\sffamily neg}(\kappa).\]
\end{Lem}
The proof of Lemma \ref{nonloss} is identical to the proof of Lemma $3$ in \cite{39}: since $\Pr_{\textbf{\upshape A}}[||\textbf{\upshape A}\tilde{\textbf{\upshape x}}||_\infty\leq 2r]\leq\text{\upshape\sffamily neg}(\kappa)$ (as shown in \cite{39}) and $||\tilde{\textbf{\upshape e}}-\textbf{\upshape e}||_\infty\leq 2r$ for independent $\textbf{\upshape e},\tilde{\textbf{\upshape e}}\leftarrow\chi^\lambda$ (where the norm is computed in the central residue-class representation of the elements in $\mathbb{Z}_q$), with probability $1-\text{\upshape\sffamily neg}(\kappa)$ there cannot exist any $\textbf{\upshape x}\in\mathbb{Z}_q^\kappa$ with $\textbf{\upshape x}\neq\tilde{\textbf{\upshape x}}$ and $\textbf{\upshape A}\textbf{\upshape x}+\textbf{\upshape e}=\textbf{\upshape A}\tilde{\textbf{\upshape x}}+\tilde{\textbf{\upshape e}}$.
\section{Proof of the Hardness Result}\label{hardn}
Now we construct the lossy code for the Skellam distribution. It is essentially the same construction that was used as lossy code for the uniform error distribution in \cite{39}.
\begin{Ctn}[Lossy code for the symmetric Skellam distribution]\label{lossysk}
Let $\kappa$ be an even security parameter, let $\lambda=\lambda(\kappa)=\text{\upshape\sffamily poly}(\kappa)$, $\nu>0$ and let $q=q(\kappa)$ be a prime modulus. The distribution $\mathcal{C}_{\kappa,\lambda,q,\nu}$ defined on $\mathbb{Z}_q^{\lambda\times\kappa}$ is specified as follows.
Choose $\textbf{\upshape A}^\prime\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa/2})$, $\textbf{\upshape T}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\kappa/2\times\kappa/2})$ and $\textbf{\upshape G}\leftarrow D_\nu^{\lambda\times\kappa/2}$.
Output \[\textbf{\upshape A}=(\textbf{\upshape A}^\prime||\textbf{\upshape A}^\prime\textbf{\upshape T}+\textbf{\upshape G}).\]
\end{Ctn}
From Lemma \ref{lwematver} and Theorem \ref{lwestod} it is straightforward that $\mathcal{C}_{\kappa,\lambda,q,\nu}$ is pseudo-random in the sense of property $1$ of Definition \ref{lossydef} assuming the hardness of the $\text{\upshape LWE}(\kappa,\lambda,q,D_\nu)$ problem.
\begin{Lem}[Pseudo-randomness of Construction $\ref{lossysk}$ \cite{39}]\label{lossypr}
For Construction $\ref{lossysk}$ it holds that $\mathcal{C}_{\kappa,\lambda,q,\nu}\approx_c\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa})$ assuming the hardness of the $\text{\upshape LWE}(\kappa,\lambda,q,D_\nu)$ problem.
\end{Lem}
Let $\textbf{\upshape A}=(\textbf{\upshape A}^\prime||\textbf{\upshape A}^\prime\textbf{\upshape T}+\textbf{\upshape G})$ be the code as defined in Construction \ref{lossysk}. We show that in our further analysis we can restrict ourselves to considering only $\textbf{\upshape G}$ instead of $\textbf{\upshape A}$.
\begin{Lem}\label{ginsteadofa}
Let $\kappa$ be an even integer. Let $\textbf{\upshape A}=(\textbf{\upshape A}^\prime||\textbf{\upshape A}^\prime\textbf{\upshape T}+\textbf{\upshape G})$ with $\textbf{\upshape A}^\prime\in\mathbb{Z}_q^{\lambda\times\kappa/2}$, $\textbf{\upshape T}\in\mathbb{Z}_q^{\kappa/2\times\kappa/2}$ and $\textbf{\upshape G}\in\mathbb{Z}_q^{\lambda\times\kappa/2}$. Then for all $\textbf{\upshape x}\in\mathbb{Z}_q^{\kappa/2}$ there is a $\textbf{\upshape x}^\prime\in\mathbb{Z}_q^\kappa$ with $\textbf{\upshape Ax}^\prime=\textbf{\upshape Gx}$.
\end{Lem}
\begin{proof}
Define \[\widetilde{\textbf{\upshape T}}=\begin{pmatrix} \textbf{\upshape I} & \textbf{\upshape T}\\ \textbf{\upshape 0} & \textbf{\upshape I}\end{pmatrix}.\] Note that $\widetilde{\textbf{\upshape T}}\in\mathbb{Z}_q^{\kappa\times\kappa}$ is a regular matrix. For all $\textbf{\upshape x}\in\mathbb{Z}_q^{\kappa/2}$ set $\textbf{\upshape x}^\prime=\widetilde{\textbf{\upshape T}}^{-1}\cdot(\textbf{0}^{\text{\scriptsize tr}}||\textbf{\upshape x}^{\text{\scriptsize tr}})^{\text{\scriptsize tr}}$. Then
\begin{align*} \textbf{\upshape Ax}^\prime = & (\textbf{\upshape A}^\prime||\textbf{\upshape A}^\prime\textbf{\upshape T}+\textbf{\upshape G})\cdot\textbf{\upshape x}^\prime\\
= & (\textbf{\upshape A}^\prime||\textbf{\upshape G})\cdot\widetilde{\textbf{\upshape T}}\cdot\widetilde{\textbf{\upshape T}}^{-1}\cdot(\textbf{0}^{\text{\scriptsize tr}}||\textbf{\upshape x}^{\text{\scriptsize tr}})^{\text{\scriptsize tr}}\\
= & (\textbf{\upshape A}^\prime||\textbf{\upshape G})\cdot(\textbf{0}^{\text{\scriptsize tr}}||\textbf{\upshape x}^{\text{\scriptsize tr}})^{\text{\scriptsize tr}}\\
= & \textbf{\upshape Gx}.
\end{align*}
\end{proof}
To show that Construction \ref{lossysk} is a lossy code for the symmetric Skellam distribution we prove that the second property of Definition \ref{lossydef} is satisfied. We first prove two supporting claims and then show the lossiness of Construction \ref{lossysk}.
\begin{Lem}\label{techlem}
$-C+\sqrt{C^2+1}\geq\exp(-C)$ for all $C\geq 0$.
\end{Lem}
\begin{proof}
Let $f(C)=(-C+\sqrt{C^2+1})\exp(C)$. Then $f(C)$ is monotonically increasing and $f(0)=1\cdot(-0+\sqrt{0+1})=1$.
\end{proof}
\begin{Lem}\label{smallnormvec}
Let $\kappa$ be a security parameter, let $s=s(\kappa)=\omega(\log(\kappa))$ and let $\nu=\nu(\kappa)=\text{\upshape\sffamily poly}(\kappa)$. Let $\lambda=\lambda(\kappa), \zeta=\zeta(\kappa)$ be integers. Let $\textbf{\upshape G}\leftarrow D_\nu^{\lambda\times\zeta}$. Then for all $\textbf{\upshape z}\in\{0,1\}^\zeta$ the following hold:
\begin{enumerate}
\item $\Pr[||\textbf{\upshape Gz}||_1>\lambda\sqrt{\zeta\nu s}]\leq\text{\upshape\sffamily neg}(\kappa).$
\item $\Pr[||\textbf{\upshape Gz}||_2^2>\lambda\zeta\nu s]\leq\text{\upshape\sffamily neg}(\kappa).$
\end{enumerate}
\end{Lem}
\begin{proof}
The claims follow from Lemma \ref{disgausbound}.
\end{proof}
\begin{Lem}[Lossiness of Construction $\ref{lossysk}$]\label{lossysklem}
Let $\kappa$ be an even security parameter, $s=s(\kappa)=\omega(\log(\kappa))$ with $\kappa>s^3$, let $\nu=\nu(\kappa)$, let $q=\text{\upshape\sffamily poly}(\kappa)$ be a sufficiently large prime modulus, let $\lambda=\lambda(\kappa)> s$ and let $\Delta=\Delta(\kappa)=\omega(\log(\kappa))$. Let $\mu=\mu(\kappa)\geq\lambda^2\nu$. Let $f_{\textbf{\upshape B},\textbf{\upshape b}}(\textbf{\upshape y})=\textbf{\upshape B}\textbf{\upshape y} + \textbf{\upshape b}$. Let $\textbf{\upshape A}\leftarrow\{\mathcal{C}_{\kappa,\lambda,q,\nu}\}$, let $\tilde{\textbf{\upshape x}}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa), \tilde{\textbf{\upshape e}}\leftarrow\text{\upshape Sk}_\mu^\lambda$, let $\textbf{\upshape x}\leftarrow\mathcal{U}(\mathbb{Z}_q^\kappa)$ and $\textbf{\upshape e}\leftarrow\text{\upshape Sk}_\mu^\lambda$. Then
\[\Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}[H_{f_{\textbf{\upshape A},\textbf{\upshape e}},f_{\textbf{\upshape A},\tilde{\textbf{\upshape e}}},\tilde{\textbf{\upshape x}}}(\textbf{\upshape x})\geq\Delta]\geq 1-\text{\upshape\sffamily neg}(\kappa).\]
\end{Lem}
\begin{proof}
Let $(\textbf{\upshape Mz})_j$ denote the $j$-th entry of $\textbf{\upshape Mz}$ for a matrix $\textbf{\upshape M}$ and a vector $\textbf{\upshape z}$. Let $\textbf{\upshape A}=(\textbf{\upshape A}^\prime||\textbf{\upshape A}^\prime\textbf{\upshape T}+\textbf{\upshape G})$ be distributed according to $\{\mathcal{C}_{\kappa,\lambda,q,\nu}\}$ with $\textbf{\upshape A}^\prime\leftarrow\mathcal{U}(\mathbb{Z}_q^{\lambda\times\kappa/2})$, $\textbf{\upshape T}\leftarrow\mathcal{U}(\mathbb{Z}_q^{\kappa/2\times\kappa/2})$ and $\textbf{\upshape G}\leftarrow D_\nu^{\lambda\times\kappa/2}$. Let $\tilde{\textbf{\upshape e}}=(\tilde{e}_j)_{j=1,\ldots,\lambda}\leftarrow\text{\upshape Sk}_{\mu}^\lambda$. Then we have the following chain of (in)equations:
\begin{align*}
& \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}[H_{f_{\textbf{\upshape A},\textbf{\upshape e}},f_{\textbf{\upshape A},\tilde{\textbf{\upshape e}}},\tilde{\textbf{\upshape x}}}(\textbf{\upshape x})\geq\Delta]\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}\left[\Pr[\textbf{x}=\tilde{\textbf{x}}\,|\,\textbf{Ax} + \textbf{e}=\textbf{A}\tilde{\textbf{x}} + \tilde{\textbf{e}}]\leq 2^{-\Delta}\right]\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}\left[\Pr_{\textbf{\upshape e}}[\textbf{Ax} + \textbf{e}=\textbf{A}\tilde{\textbf{x}} + \tilde{\textbf{e}}\,|\,\textbf{\upshape x}=\tilde{\textbf{x}}]\cdot\frac{\Pr[\textbf{x}=\tilde{\textbf{x}}]}{\Pr_{(\textbf{x},\textbf{e})}[\textbf{\upshape A}\textbf{\upshape x}+\textbf{\upshape e}=\textbf{\upshape A}\tilde{\textbf{x}}+\tilde{\textbf{e}}]}\leq 2^{-\Delta}\right]\numberthis\label{bayeseq}\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}\left[\Pr_{\textbf{\upshape e}}[\textbf{Ax} + \textbf{e}=\textbf{A}\tilde{\textbf{x}} + \tilde{\textbf{e}}\,|\,\textbf{\upshape x}=\tilde{\textbf{x}}]\cdot\right.\\
& \mbox{\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,}\left.\cdot\frac{\Pr[\textbf{\upshape x}=\tilde{\textbf{x}}]}{\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\Pr_{\textbf{\upshape e}}[\textbf{\upshape A}\textbf{\upshape x}+\textbf{\upshape e}=\textbf{\upshape A}\tilde{\textbf{\upshape x}}+\tilde{\textbf{\upshape e}}\,|\,\textbf{\upshape x}=\textbf{\upshape z}]\cdot\Pr[\textbf{\upshape x}=\textbf{\upshape z}]}\leq 2^{-\Delta}\right]\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}\left[\Pr_{\textbf{\upshape e}}[\textbf{\upshape e}=\tilde{\textbf{\upshape e}}]\cdot\frac{\Pr[\textbf{\upshape x}=\tilde{\textbf{\upshape x}}]}{\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\Pr_{\textbf{\upshape e}}[\textbf{\upshape A}(\textbf{\upshape z}-\tilde{\textbf{\upshape x}})+\textbf{\upshape e}=\tilde{\textbf{\upshape e}}]\cdot\Pr[\textbf{\upshape x}=\textbf{\upshape z}]}\leq 2^{-\Delta}\right]\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}\left[\Pr_{\textbf{\upshape e}}[\textbf{\upshape e}=\tilde{\textbf{\upshape e}}]\cdot\frac{1}{\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\Pr_{\textbf{\upshape e}}[\textbf{\upshape A}(\textbf{\upshape z}-\tilde{\textbf{\upshape x}})+\textbf{\upshape e}=\tilde{\textbf{\upshape e}}]}\leq 2^{-\Delta}\right]\numberthis\label{applyuniform}\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\frac{\Pr_{\textbf{\upshape e}}[\textbf{\upshape e}=\tilde{\textbf{\upshape e}}]}{\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\Pr_{\textbf{\upshape e}}[\textbf{\upshape e}=\textbf{\upshape A}\textbf{\upshape z}+\tilde{\textbf{\upshape e}}]}\leq 2^{-\Delta}\right]\numberthis\label{eliminatevariable}\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\frac{\Pr_{\textbf{\upshape e}}[\textbf{\upshape e}=\textbf{\upshape A}\textbf{\upshape z}+\tilde{\textbf{\upshape e}}]}{\Pr_{\textbf{\upshape e}}[\textbf{\upshape e}=\tilde{\textbf{\upshape e}}]}\geq 2^{\Delta}\right]\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\frac{\prod_{j=1}^\lambda \exp(-\mu)\cdot I_{(\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j}(\mu)}{\prod_{j=1}^\lambda \exp(-\mu)\cdot I_{\tilde{e}_j}(\mu)}\geq 2^{\Delta}\right]\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\prod_{j=1}^\lambda\frac{I_{(\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j}(\mu)}{I_{\tilde{e}_j}(\mu)}\geq 2^{\Delta}\right]\\
\geq & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\prod_{j=1}^\lambda\prod_{k=1+\tilde{e}_j}^{(\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j}\frac{-k+\sqrt{k^2+\mu^2}}{\mu}\geq 2^{\Delta}\right]\numberthis\label{approxmodbes}\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\prod_{j=1}^\lambda\prod_{k=1}^{(\textbf{\upshape A}\textbf{\upshape z})_j}\left(\frac{-(k+\tilde{e}_j)}{\mu}+\sqrt{\left(\frac{k+\tilde{e}_j}{\mu}\right)^2+1}\right)\geq 2^{\Delta}\right]\\
\geq & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\prod_{j=1}^\lambda\left(\frac{-((\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j)}{\mu}+\sqrt{\left(\frac{(\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j}{\mu}\right)^2+1}\right)^{(\textbf{\upshape A}\textbf{\upshape z})_j}\geq 2^{\Delta}\right]\numberthis\label{mondeceq}\\
\geq & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\prod_{j=1}^\lambda\exp\left(-\frac{(\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j}{\mu}\right)^{(\textbf{\upshape A}\textbf{\upshape z})_j}\geq 2^{\Delta}\right]\numberthis\label{techlemeq}\\
= & \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape e}})}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\prod_{j=1}^\lambda\exp\left(-\frac{(\textbf{\upshape A}\textbf{\upshape z})_j^2+(\textbf{\upshape A}\textbf{\upshape z})_j\cdot\tilde{e}_j}{\mu}\right)\geq 2^{\Delta}\right]\\
\geq & \Pr_{\textbf{\upshape A}}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\prod_{j=1}^\lambda\exp\left(-\frac{(\textbf{\upshape A}\textbf{\upshape z})_j^2+(\textbf{\upshape A}\textbf{\upshape z})_j\cdot s\sqrt{\mu}}{\mu}\right)\geq 2^{\Delta}\right]-\text{\upshape\sffamily neg}(\kappa)\numberthis\label{skellamboundeq}\\
= & \Pr_{\textbf{\upshape A}}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\exp\left(-\frac{\sum_{j=1}^\lambda(\textbf{\upshape A}\textbf{\upshape z})_j^2+s\sqrt{\mu}\cdot\sum_{j=1}^\lambda(\textbf{\upshape A}\textbf{\upshape z})_j}{\mu}\right)\geq 2^{\Delta}\right]-\text{\upshape\sffamily neg}(\kappa)\\
= & \Pr_{\textbf{\upshape A}}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^\kappa}\exp\left(-\frac{||\textbf{\upshape A}\textbf{\upshape z}||_2^2+s\sqrt{\mu}\cdot||\textbf{\upshape A}\textbf{\upshape z}||_1}{\mu}\right)\geq 2^{\Delta}\right]-\text{\upshape\sffamily neg}(\kappa)\\
\geq & \Pr_{\textbf{\upshape G}}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^{\kappa/2}}\exp\left(-\frac{||\textbf{\upshape Gz}||_2^2+s\sqrt{\mu}\cdot||\textbf{\upshape Gz}||_1}{\mu}\right)\geq 2^{\Delta}\right]-\text{\upshape\sffamily neg}(\kappa).\numberthis\label{fromAtoG}
\end{align*}
Equation \ref{bayeseq} is an application of the Bayes rule and Equation \ref{applyuniform} applies, since $\textbf{\upshape x}$ is sampled according to a uniform distribution. Equation \ref{eliminatevariable} is valid since in the denominator we are summing over all possible $\textbf{\upshape z}\in\mathbb{Z}_q^\kappa$. Inequation \ref{approxmodbes} is an iterative application of Theorem \ref{modbesrat}. Note that the modified Bessel function of the first kind ist symmetric when considered over integer orders. Therefore, from this point of the chain of (in)equations (i.e. from Inequation \ref{approxmodbes}), we can assume that $\tilde{e}_j\geq 0$. Moreover, we can assume that $(\textbf{\upshape A}\textbf{\upshape z})_j\geq 0$, since otherwise $I_{(\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j}(\mu)>I_{-(\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j}(\mu)$. I.e. if $(\textbf{\upshape A}\textbf{\upshape z})_j< 0$, then we implicitly change the sign of the $j$th row in the original matrix $\textbf{\upshape A}$ while considering the particular $\textbf{\upshape z}$. In this way we are always considering the worst-case scenario for every $\textbf{\upshape z}$. Note that this step does not change the distribution of $\textbf{\upshape A}$, since $\{\mathcal{C}_{\kappa,\lambda,q,\nu}\}$ is symmetric. Inequation \ref{mondeceq} holds, since $f_\mu(k)=(-k+\sqrt{k^2+\mu^2})/\mu$ is a monotonically decreasing function. Inequation \ref{techlemeq} follows from Lemma \ref{techlem} by setting $C=((\textbf{\upshape A}\textbf{\upshape z})_j+\tilde{e}_j)/\mu$.
Inequation \ref{skellamboundeq} holds because of the bound in Lemma \ref{skellbound}. Inequation \ref{fromAtoG} follows from Lemma \ref{ginsteadofa}, since $\textbf{\upshape A}=(\textbf{\upshape A}^\prime||\textbf{\upshape A}^\prime\textbf{\upshape T}+\textbf{\upshape G})$.\\
Now consider the set $\mathcal{Z}\subset\{0,1\}^{\kappa/2}$ with each element in $\mathcal{Z}$ having hamming weight exactly $\kappa/4$. Then $|\mathcal{Z}|=\binom{\kappa/2}{\kappa/4}>2^{\kappa/4}$. Since $\mu=\lambda^2\nu$, from Lemma \ref{smallnormvec} it follows that
\[\Pr_{\textbf{\upshape G}}\left[\sum_{\textbf{\upshape z}\in\mathcal{Z}}\exp\left(-\frac{||\textbf{\upshape Gz}||_2^2+s\sqrt{\mu}\cdot||\textbf{\upshape Gz}||_1}{\mu}\right)\geq 2^{\kappa/4}\cdot\exp\left(-\frac{\kappa s}{4\lambda}-\frac{s\sqrt{\kappa s}}{2}\right)\right]\geq 1-\text{\upshape\sffamily neg}(\kappa),\]
where the norm is computed in the central residue-class representation of the elements in $\mathbb{Z}_q$. Moreover we have
\[2^{\kappa/4}\cdot\exp\left(-\frac{\kappa s}{4\lambda}-\frac{s\sqrt{\kappa s}}{2}\right)>C^\kappa\]
for some constant $C>1$, since $\kappa> s^3$ and $\lambda> s$. Therefore
\begin{align*}
& \Pr_{(\textbf{\upshape A},\tilde{\textbf{\upshape x}},\tilde{\textbf{\upshape e}})}[H_{f_{\textbf{\upshape A},\textbf{\upshape e}},f_{\textbf{\upshape A},\tilde{\textbf{\upshape e}}},\tilde{\textbf{\upshape x}}}(\textbf{\upshape x})\geq\Delta]\\
\geq & \Pr_{\textbf{\upshape G}}\left[\sum_{\textbf{\upshape z}\in\mathbb{Z}_q^{\kappa/2}}\exp\left(-\frac{||\textbf{\upshape Gz}||_2^2+s\sqrt{\mu}\cdot||\textbf{\upshape Gz}||_1}{\mu}\right)\geq 2^\Delta\right]-\text{\upshape\sffamily neg}(\kappa)\\
\geq & \Pr_{\textbf{\upshape G}}\left[\sum_{\textbf{\upshape z}\in\mathcal{Z}}\exp\left(-\frac{||\textbf{\upshape Gz}||_2^2+s\sqrt{\mu}\cdot||\textbf{\upshape Gz}||_1}{\mu}\right)\geq 2^\Delta\right]-\text{\upshape\sffamily neg}(\kappa)\\
\geq & 1-\text{\upshape\sffamily neg}(\kappa).
\end{align*}
\end{proof}
\noindent We put the previous results together in order to show the main theorem.\\
\noindent\textit{Proof of Theorem} \ref{lweskthm}. By Theorem \ref{wtacr} the $\text{\upshape LWE}(\kappa,\lambda,q,D_\nu)$ problem is hard for $\nu=(\alpha q)^2/(2\pi)>2\kappa/\pi$ if there exists no efficient quantum algorithm approximating the decision-version of the shortest vector problem ($\text{\upshape GAPSVP}$) and the shortest independent vectors problem ($\text{\upshape SIVP}$) to within $\tilde{O}(\kappa/\alpha)$ in the worst case. Let $q=q(\kappa)=\text{\upshape\sffamily poly}(\kappa)$, $s=s(\kappa)=\omega(\log(\kappa))$ and $\lambda>3\kappa$. Then for $\Delta=\omega(\log(\kappa))$, Lemma \ref{nonloss} (setting $r=s\sqrt{\mu}$), Lemma \ref{lossypr} and Lemma \ref{lossysklem} provide that Construction \ref{lossysk} gives us a family of $\Delta$-lossy codes for the symmetric Skellam distribution with variance $\mu\geq\lambda^2\nu$. By Theorem \ref{lossythm} this is sufficient for the hardness of the $\text{\upshape LWE}(\kappa,\lambda,q,\text{\upshape Sk}_\mu)$ problem. Setting $\rho=\alpha\lambda$ yields $(\rho q)^2>4\lambda^2\kappa$ and the claim follows.
\hfill$\Box$\\
\noindent By Theorem \ref{lwestod} we get the hardness of the DLWE problem as a corollary.
\begin{Cor}\label{dlweskcor}
Let $\kappa$ be a security parameter
and let $\lambda=\lambda(\kappa)=\text{\upshape\sffamily poly}(\kappa)$ with $\lambda>3\kappa$. Let $q=q(\kappa)=\text{\upshape\sffamily poly}(\kappa)$ be a sufficiently large prime modulus and $\rho>0$ such that $\rho q\geq 2\lambda\sqrt{\kappa}$. If there exists a PPT-algorithm that solves the $\text{\upshape DLWE}(\kappa,\lambda,q,{\text{\upshape Sk}_{(\rho q)^2/4}})$ problem with non-negligible probability,
then there exists an efficient quantum-algorithm that approximates the decision-version of the shortest vector problem ($\text{\upshape GAPSVP}$) and the shortest independent vectors problem ($\text{\upshape SIVP}$) to within $\tilde{O}(\lambda\kappa/\rho)$ in the worst case.
\end{Cor}
\section{Application to Differential Privacy}\label{applic}
We turn to showing how the previous result contributes to building prospective post-quantum secure protocols for differential privacy with a relatively high accuracy. In contrast to the $\text{\upshape LWE}(\kappa,\lambda,q,D_\nu)$ problem, note that for the hardness of the $\text{\upshape LWE}(\kappa,\lambda,q,\text{\upshape Sk}_\mu)$ problem we need the standard deviation $\sqrt{\mu}$ of the symmetric Skellam distribution to grow linearly in the number $\lambda$ of equations.
\subsection{Security}\label{secsubsec}
As mentioned in the introduction, the notion of Private Stream Aggregation (PSA) was introduced in \cite{2} and in \cite{40} it was shown that a PSA scheme can be built upon any key-homomorphic weak pseudo-random function and some security guarantees were provided. In the next theorem we recap the result from \cite{40} in a brief form.
\begin{Thm}[Weak PRF gives secure protocol \cite{40}]\label{PSATHEOREM}
Let $\kappa$ be a security parameter, and $m,n\in\mathbb{N}$ with $\log(m)=\text{poly}(\kappa),n=\text{poly}(\kappa)$. Let $(G,\cdot), (S,*)$ be finite groups and $G^\prime\subseteq G$. For some finite set $M$, let \[\mathcal{F}=\{\text{\upshape\sffamily F}_s\,|\,\text{\upshape\sffamily F}_s:M\to G^\prime\}_{s\in S}\] be a (possibly randomised) \textit{weak PRF family} and let \[\varphi:\{-mn,\ldots,mn\}\to G\] be a mapping. Let the algorithm \textbf{\mbox{\upshape \sffamily Setup}} be defined as follows.
\begin{description}
\item \textbf{\mbox{\upshape\sffamily Setup}}: $(\mbox{\upshape\sffamily pp},T,s_0,s_1,\ldots,s_n)\leftarrow \mbox{\upshape\sffamily Setup}(1^\kappa)$, where $\mbox{\upshape\sffamily pp}$ are parameters of $G,G^\prime,S,M,\mathcal{F},\varphi$. The keys are $s_i\leftarrow\mathcal{U}(S)$ for all $i\in[n]$ with $s_0=(\bigast_{i=1}^n s_i)^{-1}$ and $T\subset M$ such that all $t\in T$ are chosen uniformly at random from $M$.
\end{description}
Then for any ppt algorithm in the non-adaptive compromise model
the following algorithm generates ciphers indistinguishable under a chosen plaintext attack:
\begin{description}
\item\textbf{\mbox{\upshape\sffamily PSAEnc}}: $c^{(t)}_i\leftarrow \text{\upshape\sffamily F}_{s_i}(t)\cdot \varphi(x^{(t)}_i)$ for $x^{(t)}_i\in\{-m,\ldots,m\}, t\in T$.\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
\end{description}
Moreover, if $\mathcal{F}$ contains only deterministic functions that are \textit{homomorphic over $S$}
and if $\varphi$ is an $mn$-isomorphic embedding, then the following algorithm correctly decrypts $\sum_{i=1}^n x^{(t)}_i$ for all $t$:
\begin{description}
\item\textbf{\mbox{\upshape\sffamily PSADec}}: compute $\varphi\left(\sum_{i=1}^n x^{(t)}_i\right)=\text{\upshape\sffamily F}_{s_0}(t)\cdot c^{(t)}_1\cdot\ldots\cdot c^{(t)}_n$ and invert.\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
\end{description}
\end{Thm}
We can build an instantiation of Theorem \ref{PSATHEOREM} (without correct decryption) based on the $\text{\upshape DLWE}(\kappa,\lambda,q,\chi)$ problem as follows. Set $S=M=\mathbb{Z}_q^\kappa, G=\mathbb{Z}_q$, choose $s_i\leftarrow \mathcal{U}(S)$ for all $i=1,\ldots,n$ and $s_0=-\sum_{i=1}^n s_i$, set $\text{\upshape\sffamily F}_{s_i}(t)=\langle t,s_i\rangle +e^{(t)}_i$ (which is a so-called \textit{randomised} weak pseudo-random function as described in \cite{49} and \cite{44}), where $e^{(t)}_i\leftarrow\chi$ and let $\varphi$ be the identity function. Here the decryption function is defined by \[\text{\upshape PSADec}_{s_0}(c^{(t)}_1,\ldots,c^{(t)}_n)=\langle t,s_0\rangle+\sum_{i=1}^n c^{(t)}_i=\sum_{i=1}^n x^{(t)}_i+\sum_{i=1}^n e^{(t)}_i.\]
Thus, the decryption is not perfectly correct any more, but randomised and it gives us a noisy sum.
\begin{Exm}\label{discretegaussianexample}
Let $\chi=D_{\nu/n}$ with variance $\nu/n=2\kappa/\pi$, then the $\text{\upshape DLWE}(\kappa,\lambda,q,\chi)$ problem is hard due to Theorem $\ref{wtacr}$ and Theorem $\ref{lwestod}$. Thus, the above scheme is secure in the sense of Theorem $\ref{PSATHEOREM}$.
\end{Exm}
\begin{Exm}\label{skellamexample}
Let $\chi=\text{\upshape Sk}_{\mu/n}$ with variance $\mu/n=\lambda^2\kappa$ (where $\lambda=\lambda(\kappa)=\text{\upshape\sffamily poly}(\kappa)$ with $\lambda>3\kappa$), then the $\text{\upshape DLWE}(\kappa,\lambda,q,\chi)$ problem is hard due to Corollary $\ref{dlweskcor}$ and the above scheme is secure in the sense of Theorem $\ref{PSATHEOREM}$.
\end{Exm}
\subsection{Privacy}\label{privsubsec}
Moreover, the total noise $\sum_{i=1}^n e^{(t)}_i$ in Example \ref{skellamexample} is distributed according to $\text{\upshape Sk}_\mu$ due to Lemma \ref{sksum}. Thus, in contrast to the total noise in Example \ref{discretegaussianexample}, the total noise in Example \ref{skellamexample} preserves the distribution of the single noise and can be used for preserving differential privacy of the correct sum by splitting the task of perturbation among the users.\\
We provide the definition of $(\epsilon,\delta)$-differential privacy (\mbox{\upshape\sffamily DP}) and a bound on the variance $\mu$ of the symmetric Skellam distribution that is needed in order to preserve $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP}.\\
We recall that a randomised mechanism preserves differential privacy if its application on two adjacent databases, i.e. databases which differ in one entry only, leads to close distributions of the output.
\begin{Def}[Differential Privacy~\cite{8}]
Let $\mathcal{R}$ be a (possibly infinite) set and let $n\in\mathbb{N}$. A randomised mechanism $\mathcal{A}:\mathcal{D}^n\to\mathcal{R}$ preserves $(\epsilon,\delta)$-differential privacy, if for all adjacent databases $D_0, D_1\in\mathcal{D}^n$ and all $R\subseteq\mathcal{R}$:
\[\Pr[\mathcal{A}(D_0)\in R]\leq e^\epsilon\cdot \Pr[\mathcal{A}(D_1)\in R]+\delta.\]
The probability space is defined over the randomness of $\mathcal{A}$.
\end{Def}
Typically $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} is achieved by properly perturbing the correct statistics. The next theorem shows how to use the Skellam distribution for this task.
\begin{Thm}[Skellam mechanism preserves \mbox{\upshape\sffamily DP} \cite{40}]\label{skmech} Let $\epsilon>0$ and let $0<\delta<1$. For all databases $D\in\mathcal{D}^n$ the randomised mechanism \[\mathcal{A}_{Sk}(D):=f(D)+Y\]
preserves $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} with respect to any query $f$ of sensitivity $S(f)$, where $Y\leftarrow\text{\upshape Sk}_\mu$ with
\[\mu\geq\frac{\log(1/\delta)}{1-\cosh(\epsilon/S(f))+(\epsilon/S(f))\cdot\sinh(\epsilon/S(f))}.\]
\end{Thm}
\begin{Rem}\label{skdprem} The bound on $\mu$ from Theorem \text{\upshape \ref{skmech}} is smaller than $2\cdot (S(f)/\epsilon)^2\cdot\log(1/\delta)$, thus the standard deviation $\mu$ of $Y\leftarrow\text{\upshape Sk}_\mu$ may be assumed to be linear in $S(f)/\epsilon$ (for constant $\delta$).
\end{Rem}
Suppose that adding symmetric Skellam noise with variance $\mu$ preserves $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP}. We define $\mu_{user}=\mu/n$. Since the Skellam distribution is reproducible, the noise addition can be executed in a distributed manner: each (non-compromised) user simply adds (independent) symmetric Skellam noise with variance $\mu_{user}$ to her own value in order to preserve the privacy of the final output.
\subsection{Accuracy}\label{acsubsec}
\begin{Thm}[Accuracy of the Skellam mechanism \cite{40}]\label{erroranalysis} Let $\epsilon>0$ and $0<\delta<1$. Then for all $0<\beta<1$ the mechanism specified in Theorem $\ref{skmech}$ has $(\alpha,\beta)$-accuracy, where
\[\alpha=\frac{S(f)}{\epsilon}\cdot\left(\log\left(\frac{2}{\beta}\right)+\log\left(\frac{1}{\delta}\right)\right).\]
\end{Thm}
This means, the error does not exceed $\alpha$ with probability at least $1-\beta$. Theorem \ref{PSATHEOREM} indicates that the set $T$ contains all the time-frames where a query can be executed. For simplicity we assume that all queries are independent, i.e. the arguments of all queries are independent. As pointed out in section \ref{lwesubsec} we identify the number of queries with the number of equations in the instance of the $\lambda$-bounded LWE problem, thus $|T|=\lambda$. (A result in \cite{47} indicates that for an efficient and accurate mechanism this number cannot be substantially larger than $n^2$, where $n$ is the number of users in the network.) Due to sequential composition (see for instance Theorem $3$ of \cite{48}), in order to preserve $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} in all $\lambda$ queries together, the executed mechanism must preserve $(\epsilon/\lambda,\delta)$-\mbox{\upshape\sffamily DP} in every single query. Therefore the following holds: suppose $\text{\upshape Sk}_{\mu^\prime}$-noise is sufficient in order to preserve $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} in a single query. Then, due to Remark \ref{skdprem}, we must use $\text{\upshape Sk}_{\lambda^2\mu^\prime}$-noise in order to preserve $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} in all $\lambda$ queries.\\
With Theorem \ref{erroranalysis} we obtain $(\alpha,\beta)$-accuracy for every single query executed during $T$ with
\[\alpha=\frac{S(f)}{\epsilon/\lambda}\cdot\left(\log\left(\frac{2}{\beta}\right)+\log\left(\frac{1}{\delta}\right)\right)=O\left(\frac{S(f)\lambda}{\epsilon}\right),\]
which is optimal with respect to sequential composition.
\subsection{Combining Security, Privacy and Accuracy}
Set $S(f)=m$ and $\mu=2\cdot (m\lambda/\epsilon)^2\cdot\log(1/\delta)$. From the discussion from above it follows that if every user adds $\text{\upshape Sk}_{\mu/n}$-noise to her data in every time-step $t\in T$, then this is sufficient in order to preserve $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} in all $\lambda$ sum-queries that are executed during $T$, where for each time-step $t\in T$ the data of each user comes from $\{-m,\ldots,m\}$.\\
Furthermore, if for a security parameter $\kappa$ we have that $\mu/n=\lambda^2\kappa$, then we obtain a secure protocol for analysing sum-queries, where the security is based on prospectively hard lattice problems. As shown in \cite{40}, a combination of these two results provides \textit{computational} $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} in all $\lambda$ sum-queries. Assume that we want to find values for $\epsilon, \delta$ such that when every user adds $\text{\upshape Sk}_{\mu/n}$-noise to her data with $\mu=2\cdot (m\lambda/\epsilon)^2\cdot\log(1/\delta)$ to preserve $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} of the final statistics, then the same noise suffices for providing security. Therefore the following must be satisfied:
\[2\cdot (m\lambda/\epsilon)^2\cdot\log(1/\delta)\geq n\lambda^2\kappa.\]
This inequality holds for
\[\epsilon=\epsilon(\kappa)\leq\sqrt{\frac{2 m^2\cdot\log(1/\delta)}{\kappa\cdot n}},\numberthis\label{epsilonupperbound}\]
indicating that $\epsilon=\epsilon(\kappa)$ depends on $1/\kappa$. Note that this is consistent with the original definition of computational \mbox{\upshape\sffamily DP} in \cite{15}. Thus, in addition to a privacy/accuracy trade-off we also get a security/accuracy trade-off. More specifically, depending on $\kappa$ and $n$ we obtain a tight lower bound on the $(\alpha,\beta)$-accuracy for every single query executed during $T$:
\begin{align*}
\alpha= & \frac{m}{\epsilon/\lambda}\cdot\left(\log\left(\frac{2}{\beta}\right)+\log\left(\frac{1}{\delta}\right)\right)\\
\geq & \lambda\cdot\sqrt{\frac{\kappa\cdot n}{2\cdot\log(1/\delta)}}\cdot\left(\log\left(\frac{2}{\beta}\right)+\log\left(\frac{1}{\delta}\right)\right)\\
= & \Omega(\lambda\sqrt{\kappa\cdot n}).
\end{align*}
\begin{Exm}
There are $n=20,000$ users in a network with variable time-series data falling in an interval of $\{-1000,\ldots,1000\}$, i.e. $m=1000$. Note that the aggregated sum can not exceed $20,000,000$. The users want to preserve computational $(\epsilon,\delta)$-\mbox{\upshape\sffamily DP} with $\epsilon=1$ and $\delta=0.1$ while evaluating all $\lambda$ sum-queries over a time period $T$. They use the DLWE-based secure protocol for communicating with an untrusted aggregator and generate symmetric Skellam noise. For a hard DLWE problem, the security parameter is chosen to be $\kappa=200$. For these parameters, inequality $\ref{epsilonupperbound}$ is satisfied. For the $(\alpha,\beta)$-accuracy per query it holds that \[\alpha=\frac{m}{\epsilon/\lambda}\cdot\left(\log\left(\frac{2}{\beta}\right)+\log\left(\frac{1}{\delta}\right)\right)=1000\lambda\cdot\left(\log\left(\frac{2}{\beta}\right)+\log(10)\right).\]
\end{Exm}
\section{Conclusions}
In this work we provided a worst-to-average-case connection from conjecturally hard lattice problems to the LWE problem with errors following a symmetric Skellam distribution. Our proof relies on the notion of lossy codes from \cite{39}. An implication of this result is the construction of the first prospective post-quantum Private Stream Aggregation scheme for data analyses under differential privacy where the errors are used both for security of the scheme and for the distributed noise generation for preserving differential privacy. An interesting further direction is to reduce the size of the variance that is necessary for the hardness of the LWE problem with errors following a symmetric Skellam distribution, especially to abolish the dependence on the number of LWE-samples. Another problem to face is to show the hardness of the Ring LWE problem (a more efficient version of LWE introduced in \cite{45}) with errors following a symmetric Skellam distribution and to establish a corresponding search-to-decision reduction. Sufficient conditions on the error distribution for the existence of a search-to-decision reduction for the Ring LWE problem were provided in \cite{55}. The Skellam distribution does not seem to satisfy these conditions. Thus, we require a different proof than in \cite{55}.
\addcontentsline{toc}{chapter}{Literatur}
|
\section{Introduction}
A large number of existing models used in sparse signal processing and machine learning rely on $\ell_1${}-norm regularization in order to recover sparse signals or to identify sparse features for classification tasks. Sparse $\ell_1${}-norm regularization is also prominently used in the image-processing and computer vision domain, where it is used for segmentation, tracking, and background subtraction tasks.
In computer vision and image processing, we are often interested in regions that are not only sparse, but also {\em spatially smooth}, i.e., regions with contiguous support structure. In such situations, it is desirable to have regularizers that promote the selection of large, contiguous regions rather than merely sparse (and potentially isolated) pixels. In contrast, simple $\ell_1${}-norm regularization adopts an unstructured approach that induces sparsity wherein each variable is treated independently, disregarding correlation among neighboring variables.
For example, smooth support structure is relevant to compressive background subtraction \cite{
cevher2008compressive, cevher2009sparse} which detects contiguous regions of movement against a stationary background.
For imaging applications, $\ell_1${}-norm regularization may result in regions with spurious active (or isolated) pixels or non-smooth boundaries in the support set. This issue is addressed by the image-segmentation literature, where spatially correlated priors (such as total variation or normalized cuts) are used to enforce smooth support boundaries~\cite{boykov2001fast,felzenszwalb2004efficient,shi2000normalized,comaniciu2002mean}. An important hallmark of existing image-segmentation methods is that they are able to enforce spatially contiguous support. However, the concept of correlated support has yet to be ported to more complex reconstruction tasks, including (but not limited to) robust PCA and compressive background subtraction.
The development of such structured sparsity models has been an active research topic \cite{cevher2009sparse,baraniuk2010model,huang2011learning,jenatton2011structured,bach2011convex,jacob2009group}, with new models and applications still emerging \cite{jeni2014spatio,jenatton2009structured}.
In this paper, we develop a class of convex priors based on overlapping block/group sparsity, which are able to enforce sparsity of the support set and promote {\em spatial smoothness}.
\subsection{Relevant Previous Work}
Existing work on spatially-smooth support-set regularization can be divided into two main categories: (i) non-convex models that rely on graphs and trees, and (ii) convex models that rely on group-sparsity inducing norms.
Cevher \etal~\cite{cevher2009sparse} promote sparsity using Markov random fields (MRFs) in combination with compressive-sensing signal recovery, which is referred to as lattice matching pursuit (LaMP). LaMP recovers structured sparse signals using fewer noisy measurements than methods that ignore spatially correlated support sets. Baraniuk \etal \cite{baraniuk2010model} prove theoretical guarantees on robust recovery of structured sparse signals using a non-convex algorithm; their approach has been validated using wavelet-tree-based hierarchical group structure, as well as signals with non-overlapping blocks in the support set. Huang \etal \cite{huang2011learning} developed a theory of greedy approximation methods for general non-convex structured sparse models.
All these methods, however, are limited in that they are either non-convex, computationally expensive, or do not allow for overlapping (or not aligned) group structure.
Jenatton \etal \cite{jenatton2011structured} showed the possibility of coming up with a problem-specific optimal group-sparsity-inducing norm using prior knowledge of the underlying structure. While they consider a convex relaxation of the structured sparsity problem, it remains unclear how their proposed active-set algorithm for least squares regression can be generalized to a broader range of applications.
\subsection{Contributions}
Our work is inspired by the $\ell_1/\ell_2$-norm spatial coherence priors used in \cite{jenatton2011structured}, as well as group sparsity priors used in statistics (e.g., group lasso) \cite{leigsnering2014multipath,yuan2006model}. Our main contributions can be summarized as follows: (i)
We propose new regularizers for imaging and computer vision applications including compressive image recovery, sparse \& low rank decomposition, and a block-sparse generalization of total variation.
(ii) We develop computationally efficient global minimization algorithms that are suitable for overlapping pixel-cliques.
Existing methods for group sparsity use the alternating direction method of multipliers (ADMM), and have excessive memory requirements for large clique sizes. We therefore discuss a new approach using fast convolution algorithms to perform gradient descent with low memory requirements and a complexity that is independent of the clique size.
(iii) We propose the use of our regularizers within greedy pursuit methods for compressive reconstruction.
(iv) We demonstrate that our algorithms can be used to suppress artifacts and enhance the quality of sparse recovery methods when applied to a variety of imaging applications.
\subsection{Notation}
For any column vector $\mathbf{x} \in \mathbb{R}^n$, we define its $\ell_{\alpha}$-norm with $\alpha\geq1$ as $\|\mathbf{x}\|_{\alpha} = (\sum_{i=1}^n |x_i|^{\alpha})^{1/{\alpha}}$.
For $\mathbf{x} \in \mathbb{R}^n$, the vector $\mathbf{x}_c$ consists only of the entries associated to the index set $c$. The support set (i.e., the set of indices of non-zero entries) of a vector or vectorized image $\vecx$ is denoted by $\supp(\vecx).$
For a matrix $\mathbf{A} \in \mathbb{R}^{M\times N}$ with rank $r = \min\{M,N\}$ and singular values $\sigma_i$, the nuclear norm is defined by $\|\bA\|_* = \sum_{i=1}^r \sigma_i.$ We use $\|\bA\|_1 = \sum_{ij} |A_{ij}|$ to denote the element-wise $\ell_1${}-norm for $\mathbf{A}.$
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{cliques.pdf}
\vspace{0.1cm}
\caption{Illustration of cliques and overlapping cliques.}
\label{fig:cliques}
\end{figure}
\section{Problem Formulation}
Consider the measurement model $\vecy = \matPhi \vecx_0 + \vecz_0$,
where $\vecy \in \mathbb{R}^M$ is the observed signal, $\mathbf{x}_0 \in \mathbb{R}^N$ is the original sparse signal we wish to recover, $\mathbf{z}_0 \in \mathbb{R}^M$ is a non-sparse component of the signal (comprising both the background image and potential noise), $\mathbf{\Phi} \in \mathbb{R}^{M\times N}$ is the linear operator that models the signal acquisition process. Based on this model, we study signal recovery by solving convex optimization problems of the following general form:
\begin{align}
\{\hat\vecz,\hat\vecx\}= \argmin_{\mathbf{z} \in \mathbb{R}^{M}\!,\mathbf{x} \in \mathbb{R}^N} D(\vecx,\vecz\,|\,\vecy,\matPhi) + J(\vecx). \label{general}
\end{align}
Here, $D: \mathbb{R}^M\!\times \mathbb{R}^M \mathcal \to\, \mathbb{R}$ is a convex data-consistency term, and $J: \mathbb{R}^M \mathcal \to\, \mathbb{R}$ is a regularizer that enforces both sparsity and support smoothness on the vector $\vecx.$ The proposed regularizer is a hybrid $\ell_1/\ell_2$-norm penalty of the form
\eqn{reg}{ \textstyle J(\vecx) = \sum_{c \in \mathcal{C}} \|\mathbf{x}_c\|_2,}
where $\mathcal{C}$ is a set of cliques over the graph $\mathcal{G}$ defined over the pixels of $\vecx.$ This regularizer \eqref{reg} is a natural generalization of the {\em group (or block) sparsity} model that has been explored in the literature for a variety of purposes including statistics and radar \cite{jenatton2011structured,jacob2009group,jeni2014spatio,leigsnering2014multipath}. We focus on the case where the collection of sparse cliques consist of regularly-spaced groups of adjacent pixels.
For example, consider two types of cliques shown in Figure \ref{fig:cliques}(a) and \ref{fig:cliques}(b). Notice the (a) 2-clique and (b) 4-clique wherein all nodes are connected to each other. These cliques can be translated over the entire image graph to generate various overlapping clique geometries as shown in (c) and (d), respectively. In (c), eight overlapping cliques, each of size two, overlap at a central point. In the image processing literature this is referred to as an 8-connected neighborhood~\cite{cheng2009subband}. In contrast, Figure \ref{fig:cliques}(d) uses a higher-order connectivity model, which is obtained using four rectangular cliques of size four (each shown in a different color).
%
Overlapping group-sparsity models of the form depicted in Figure \ref{fig:cliques}(d) effectively enforce spatial coherence of the recovered support. When such an overlapping group-sparsity model is used, all pixels in a clique tend to be either zero or non-zero at the same time (see, e.g., \cite{bach2011convex}). Since each pixel shares multiple overlapping cliques with its neighbors, this regularizer suppresses ``rogue'' (or isolated) pixels from entering the support without their neighbors and hence, promotes smooth (contiguous) support boundaries.
\subsection{Applications}
The proposed regularizer \eqref{reg} can be used as a building block for various applications in computer vision, image processing, and compressive sensing. In what follows, we will focus on the following three imaging applications:
\textit{1) Compressive sensing signal recovery:} Consider a signal $\mathbf{x}\in\mathbb{R}^N$ that is $K$-sparse, i.e., only $K \ll N$ entries of~$\vecx$ are non-zero. In the CS literature, the signal is acquired via $M < N$ linear projections $\mathbf{y} = \mathbf{\Phi} \mathbf{x}$. The $K$-sparse signal $\mathbf{x}$ can then be recovered if, for example, the matrix $\mathbf{\Phi}$ satisfies the \textit{2K-RIP} or similar conditions~\cite{baraniuk2010model,candes2006robust}. The underlying recovery problem is usually formulated as follows:
\aln{
\vecx^\star =& \argmin_{\mathbf{x}\in \mathbb{R}^N} \|\mathbf{y} -\mathbf{\Phi} \mathbf{x}\|_2^2 \,\, \hbox{ \,\,subject to\,\, } \!\! \|\mathbf{x}\|_0 = K. \label{csrec}
}
When the sparse signals are images, simple sparse recovery may not exploit the entire image structure; this is particularly true for background-subtracted surveillance video. Background subtraction is used in applications where one is interested only in inferring foreground objects and activities. Background subtraction is easily achieved in the compressive domain by computing the difference between adjacent image data or by subtracting a long term signal mean (or median). Background-subtracted frames are generally more sparse than frames containing background information, and can thus be reconstructed from far fewer measurements $M$.
We propose to extend the problem in (\ref{csrec}) by adding a regularizer of the form \eqref{reg} to promote correlation in the support set of the foreground objects. The optimization problem defined in (\ref{csrec}) is non-convex and is commonly solved using greedy algorithms \cite{tropp2007signal, needell2009cosamp, cevher2009sparse}. We will show that the use of our prior~\eqref{reg} leads to faster signal recovery with a small number of measurements compared to existing methods.
\textit{2) Total-variation denoising:} Total variation (TV) denoising restores a noisy image $\vecy$ (e.g., vectorized image) by finding an image that lies close to $\vecy$ in an $\ell_2$-norm sense, while simultaneously having small total variation; this can be accomplished by solving
\aln{
\vecx^\star = \argmin_{\vecx\in\mathbb{R}^N} \textstyle \frac{1}{2} \|\vecx-\vecy\|^2 + \lambda\|\nabla_d \vecx\|_1, \label{rof}
}
where $\nabla_d: \mathbb{R}^N \to \mathbb{R}^{2N}$ is a discrete gradient operator that acts on an $N$-pixel image, and produces a stacked horizontal and vertical gradient vector containing all first-order differences between adjacent pixels. TV-based image processing assumes that images have a piecewise constant representation, i.e., the gradient is sparse and locally contiguous~\cite{rudin1992nonlinear,goldstein2009split}.
Numerous generalizations of TV exist, including the recently proposed vectorial TV for color images \cite{bresson2008fast,ono2014decorrelated}. Such regularizers are of the form of \eqref{rof} merely by changing the definition of the discrete gradient operator.
We propose to extend total variation by penalizing the gradient of cliques in order to enforce a greater degree of spatial coherence. In particular, we consider
\aln{\vecx^\star = \argmin_{\vecx\in\mathbb{R}^N} \textstyle \frac{1}{2} \|\vecx-\vecy\|^2 + J(\nabla_d \vecx), \label{rof2} }
where $J(\cdot)$ denotes the regularizer \eqref{reg}. Furthermore, we explore formulations where the discrete gradient operator is given by the decorrelated color TV operator described in \cite{ono2014decorrelated}. With our approach, we also show the application of proposed structured sparsity prior on 3-D blocks. Note that \cite{selesnick2013total,liu2015image} explores the use of \mbox{1-D} and \mbox{2-D} overlapping group sparsity for TV image denoising, but using a majorization-minimization algorithm combined with ADMM.
\textit{3) Robust PCA (RPCA):} Suppose $\mathbf{Y} = [\mathbf{y}_1, \dots, \mathbf{y}_L]$ is a matrix of $L$ measurement vectors, and $\mathbf{Y}$ is the sum of a low rank matrix $\mathbf{Z}$ and a sparse matrix $\mathbf{X}.$ For this case, Cand\`es \etal show that exact recovery of these components is possible using the following formulation \cite{candes2011robust}:
\begin{align} \begin{split}
\{\hat\bZ,\hat\bX\} = & \argmin_{\mathbf{Z},\mathbf{X}\in \mathbb{R}^{N \times L}} \|\mathbf{Z}\|_* + \lambda \|\mathbf{X}\|_1 \\
& \hbox{ \,\,subject to\,\, } \!\mathbf{Y} = \mathbf{Z} + \mathbf{X}. \label{eq4}
\end{split}
\end{align}
The nuclear-norm in \eqref{eq4} promotes a low rank solution for \matZ; the $\ell_1$-norm penalty on promotes sparsity in \matX. For this reason the solution to \eqref{eq4} is sometimes referred to as a sparse-plus-low-rank decomposition.
%
A well-known application of RPCA is background subtraction in videos with a stationary background. For such datasets, the shared background in the frames $\{y_i\}$ can be represented using a low-rank subspace. The moving foreground objects often have sparse support, and thus are absorbed into the sparse term \matX.
We propose to replace the $\ell_1${}-norm regularization prior on~$\bX$ in~(\ref{eq4}) with the proposed regularizer in \eqref{reg}; this enables us to promote spatial smoothness in the support set of the foreground objects.
Here, we build on the work of \cite{gao2012block}, where structured sparsity with non-overlapping blocks is used in RPCA for foreground detection, and \cite{yao2014foreground}, where a hybrid of ALM and network flow methods \cite{mairal2010network} are used to solve $\ell_1/\ell_\infty$ regularized RPCA problems.
\subsection{Optimization Algorithms}
We now develop efficient numerical methods for solving problems involving the regularizer \eqref{reg}.
A common approach to enforce group sparsity in the statistics literature is consensus ADMM \cite{deng2013group,boyd2011distributed}, which we will briefly discuss in Section \ref{sec:stuff}. For image processing and vision applications, where the datasets as well as the cliques tend to be large, the high memory requirements of ADMM render this approach unattractive. As a consequence, we propose an alternative method that uses fast convolution algorithms to perform gradient descent that exhibits low memory requirements and requires low computational complexity. In particular, our approach is capable of handling large-scale problems, such as those in video applications, which are out of the scope of memory-hungry ADMM algorithms.
We note that numerical methods for overlapping group sparsity have been studied in the context of statistical regression \cite{jacob2009group,yuan2011efficient,boyd2011distributed}, but for different purposes. Yuan \etal~\cite{yuan2011efficient} solves the regression variable selection problem using an accelerated gradient descent approach, whereas Deng \cite{deng2013group} and Boyd \cite{boyd2011distributed} use consensus ADMM, which does not scale to high-dimensional problems. Compared to these methods, our approach provides significant speedups (see Section~\ref{sec:experiments}).
\subsubsection{Proximal Minimization and ADMM}
\label{sec:stuff}
The simplest instance of the problem \eqref{general} is the proximal operator for the penalty term $J$ in \eqref{reg}, defined as follows:
\begin{equation}\label{prox}
\prox_{J}(\vecv,\lambda) = \argmin_\vecx \|\vecx-\vecv\|^2 + \lambda J(\vecx) .
\end{equation}
Proximal minimization is a key sub-step in a large number of numerical methods. For example, the ADMM for TV minimization \cite{rudin1992nonlinear,goldstein2009split} requires the computation of the proximal operator of the $\ell_1$-norm. For such methods, the regularizer~\eqref{reg} is easily incorporated into the numerical procedure by replacing this proximal minimization with \eqref{prox}.
In the simplest case where the cliques in $\mathcal{C}$ are small and no other regularizers are needed, the proximal minimization~\eqref{prox} can be computed using ADMM \cite{boyd2011distributed,goldstein2009split}. Similar approaches have been used for other applications of overlapping group sparsity \cite{deng2013group}.
It is key to realize that the regularization term in (\ref{prox}) can be reformulated as follows:
\begin{align}
\hat\vecx= \argmin_{\mathbf{x}\in \mathbb{R}^N} \, \|\mathbf{x}-\mathbf{v}\|_2^2 & + \lambda \sum_{i=1}^{s}\sum_{c\in \mathcal{C}_i} \|\mathbf{x}_c\|_2. \label{eq7}
\end{align}
Here, $\mathcal{C}_1, \dots, \mathcal{C}_s$ are clique subsets for which the cliques in $\mathcal{C}_i$ are {\em disjoint}. For example, consider the case where the set of cliques contains all $2\times 2$ image patches as shown in Figure \ref{fig:cliques}(d). For such a scenario, we need four subsets of disjoint cliques to represent every possible patch. The reformulated problem for the example graph will be of the form \eqref{eq7} with $s=4.$ In general, if cliques are formed by translating an $l\times l$ patch, $l^2$ subsets of cliques are required so that every subset contains only disjoint cliques.
To apply ADMM to this problem, we need to introduce~$s$ auxiliary variables $\mathbf{z}^1, \dots, \mathbf{z}^s$ each representing a copy of the original pixel values.
The resulting problem is
\begin{align} %
\begin{split}
\{\hat\vecx,\hat\vecz^i\,\forall i\} = & \argmin_{\mathbf{x}, \{\mathbf{z}^i\}_{i=1}^{s} } \|\mathbf{x}-\mathbf{v}\|_2^2 + \lambda \sum_{i=1}^{s}\sum_{c\in \mathcal{C}_i} \|\mathbf{z}^i_c\|_2 \\
&\!\! \hbox{ \,\,subject to\,\, } \mathbf{z}^i = \mathbf{x}, \,\, \forall i. \label{constrained}
\end{split}
\end{align}
This is an example of a consensus optimization problem, which can be solved using ADMM (see \cite{deng2013group} for more details).
An important property of this ADMM reformulation is that each vector $\vecz^i$ can be updated in closed form---an immediate result of the disjoint clique decomposition.
\subsubsection{Forward-Backward Splitting (FBS) with Fast Fourier Transforms}
The above discussed ADMM approach has several drawbacks. First, it is difficult to incorporate more regularizers (in addition to the support regularizer $J$) without the introduction of an excessive amount of additional auxiliary variables. Furthermore, the method becomes inefficient and memory intensive for large clique sizes and large data-sets (as it is the case for multiple images). For instance in RPCA, if the cliques are generated by $l\times l$ patches, $l^2$ variables $\{\vecz^i\}$ are required, each having the same dimensionality as original image data-set $NL$. Additionally, the dual variables for each equality constraints in \eqref{constrained} will require another $l^2NL$ storage entries.
As a consequence, for large values of $l,$ the memory requirements of ADMM become prohibitive.
We propose a new forward-backward splitting algorithm that exploits fast convolution operators and prevents the excessive memory overhead of ADMM-based methods. To this end, we propose to ``smoothen'' the objective via \emph{hyperbolic regularization} of the $\ell_2$-norm as
\begin{align} \label{smooth}
\|\mathbf{x}_c\|_2 \approx \|\mathbf{x}_c\|_{2,\epsilon} = \sqrt{x_1^2+\dots+x_n^2 + \epsilon^2}
\end{align}
for some small $\epsilon>0.$ For the sake of clarity, we describe the forward-backward splitting approach in the specific case of robust PCA. Note, however, that other regularizers are possible with only minor modifications.
Using the proposed support prior (\ref{eq4}), we write
\begin{align}
\{\hat\bZ,\!\hat\bX\}\!=\!\argmin_{\mathbf{Z},\mathbf{X}} \|\mathbf{Z}\|_* \!\! \textstyle+\! \lambda J_\epsilon(\bX) \!+\! \frac{\mu}{2} \|\mathbf{Y}\!-\!\mathbf{Z}\!-\!\mathbf{X}\|_F^2 \label{general2}
\end{align}
where
\begin{equation} \textstyle J_\epsilon(\bX) = \sum_{t=1}^L \sum_{c\in \mathcal{C}} \|\mathbf{X}_{t,c}\|_{2,\epsilon} \label{smoothed}\end{equation}
is the smoothed support regularizer, and $\mathbf{X}_{t,c}$ refers to the clique $c$ drawn from column $t$ of $\mathbf{X}.$ We note that this formulation differs from that in Liu \etal \cite{liu2013robust}, where the structured sparsity is induced across columns of $\mathbf{X}$ rather than blocks, and is solved using conventional ADMM.
The forward-backward splitting (or proximal gradient) method is a general framework for minimizing objective functions with two terms \cite{DBLP:journals/corr/GoldsteinSB14}. For the problem \eqref{general2}, the method alternates between gradient descent steps that only act on the smooth terms in \eqref{general2}, and a backward/proximal step that only acts on the nuclear norm term. The gradient of the (smoothed) proximal regularizer in \eqref{general2} is given column-wise (i.e., image-wise) by
\begin{equation}\textstyle \nabla J_\epsilon(\matX_t) = \sum_{\substack{c\in C }} \matX_{t,c} \|\matX_{t,c}\|_{2,\epsilon}^{-1}. \label{grad}\end{equation}
%
The gradient formula \eqref{grad} requires the computation of the sum \eqref{smoothed} for every clique $c,$ and then, a summation over the reciprocals of these sums; this is potentially expensive if done in a na\"ive way. Fortunately, every block sum can be computed simultaneously by squaring all of the entries in $\matX,$ and then convolving the result with a block filter. The result of this convolution contains the value of $\|\matX_{t,c}\|_{2,\epsilon}^2$ for all cliques $c.$ Each entry in the result is then raised to the $-\nicefrac{1}{2}$ power, and convolved again with a block filter to compute the entries in the gradient \eqref{grad}. Both of these two convolution operations can be computed quickly using fast Fourier transforms (FFTs), so that the computational complexity becomes independent of clique size.
Algorithm \ref{alg:fbs} shows the pseudocode for solving (\ref{general2}). In Steps 1 and 2, the values of \matX and \matZ are updated using gradient descent on \eqref{general2}, ignoring the nuclear norm regularizer. Step 3 accounts for the nuclear-norm term using its proximal mapping, which is given by
$$\prox_*(\bQ, \delta) = \matU (\text{sign}(\matS)\circ\max\{|\matS|-\delta,0\})\matV^T,$$
where $\bQ =\matU\matS\matV^T$ is a singular value decomposition of~$\bQ,$ $|\matS|$ denotes element-wise absolute value, and $\circ$ denotes element-wise multiplication.
\begin{algorithm}[t]
\caption{Forward-backward proximal minimization }
\label{alg:fbs}
\textbf{Input:} $\mathbf{Y}, \mu > 0, \lambda, \mathcal{C}_i , \alpha>0$\\
\textbf{Initialize:} $\mathbf{X}^{(0)} = \mathbf{0}$, $\mathbf{Z}^{(0)} = \mathbf{0}$\\
\textbf{Output:} $\mathbf{X}^{(n)}, \mathbf{Z}^{(n)} $
\begin{algorithmic}[1]
\WHILE {not converged}
\STATE \textbf{Step 1:} Forward gradient descent on $X$,
\STATE $\matX^{(n)}_k =$ \parbox[t]{.7\linewidth}{
$\matX^{(n-1)}_k - \alpha\lambda \nabla J_\epsilon(\matX)$ \\$+ \alpha\mu(\matY_k-\matZ^{(n-1)}_k-\matX^{(n-1)}_k)$}
\STATE \textbf{Step 2:} Forward gradient descent on $\matZ$,
\STATE $\,\,\, {\matZ}^{(n)}_k = \matZ^{(n-1)}_k + \alpha \mu(\matY_k-\matZ^{(n-1)}_k-\matX^{(n-1)}_k)$
\STATE \textbf{Step 3:} Backward gradient descent on $\matZ$,
\STATE $\quad \mathbf{\matZ}^{(n)} = \prox_*(\matZ^{(n)}, \alpha) $
\ENDWHILE
\end{algorithmic}
\end{algorithm}
The forward-backward splitting (FBS) procedure in Algorithm~\ref{alg:fbs} is known to converge for sufficiently small stepsizes $\alpha$ \cite{beck2009fast}. Practical implementations of FBS \ref{alg:fbs} include adaptive stepsize selection \cite{wright2009sparse}, backtracking line search, or acceleration~\cite{beck2009fast}. We use the FASTA solver from \cite{DBLP:journals/corr/GoldsteinSB14}, which combines such acceleration techniques.
We note that FBS \ref{alg:fbs} only requires a total of $4NL$ storage entries for $\mathbf{X}, \mathbf{Y}, \mathbf{Z}$ and gradient $\nabla J_\epsilon(\mathbf{X})$. However, in order to solve RPCA formulation using ADMM we require $2l^2NL$ storage entries for auxiliary variables (as discussed before) and $4NL$ storage entries for the variables $\mathbf{X}, \mathbf{Y}, \mathbf{Z}$ and dual variable of $\mathbf{Y} = \mathbf{X} + \mathbf{Z}$, leading to total of $(2l^2 + 4)NL$ storage entries. Since the memory usage and runtime of FBS is independent of the clique size, the advantage of FBS over ADMM is much greater for larger cliques.
\subsubsection{Matching Pursuit Algorithm}
For compressive-sensing problems involving large random matrices, {\em matching pursuit} algorithms (such as CoSaMP~\cite{needell2009cosamp}) are an important class of sparse recovery methods. When signals have structured support, model-based matching pursuit routines have been proposed that require non-convex minimizations over Markov random fields~\cite{cevher2009sparse}. In this section, we propose a model-based matching pursuit algorithm that achieves structured compressive signal recovery using {\em convex} sub-steps for which global minimizers are efficiently computable.
The proposed method, Convex Lattice Matching Pursuit (CoLaMP), is a greedy algorithm that attempts to solve
\begin{equation}
\begin{split}
\hat\vecx= & \argmin_\vecx \|\matPhi \vecx - \vecy\|^2_2 + \lambda J(\vecx)\\
& \!\!\!\hbox{ \,\,subject to\,\, } \!\! \|\vecx\|_0\le K. \label{colampeq}
\end{split}
\end{equation}
The complete method is listed in Algorithm \ref{alg:colamp}. In Step 1, CoLaMP proceeds like other matching pursuit algorithms; the unknown signal is estimated by multiplying the residual by the adjoint of the measurement operator. In Step 2, this estimate is refined by solving a support regularized problem of the form \eqref{prox}. We solve this problem either via ADMM or the FBS method in Algorithm \ref{alg:fbs}). In Step 3, a least-squares (LS) problem is solved to identify the signal that best matches the observed data, assuming the correct support was identified in Step 2. This LS problem is solved by a conjugate gradient method. Finally, in Step 4, the residual (the discrepancy between $\matPhi \vecx$ and the data vector \vecy) is calculated. The algorithm is terminated if the residual becomes sufficiently small or a maximum number of iterations is reached.
\begin{algorithm}[t]
\caption{CoLaMP - Convex Lattice Matching Pursuit}
\label{alg:colamp}
\textbf{Input:} $\mathbf{y}, \mathbf{\Phi}, K, \lambda, \epsilon$\\
\textbf{Initialize:} $\mathbf{x}^{(0)} = \mathbf{0}$, $\mathbf{s}^{(0)} = \mathbf{0}, \mathbf{r}^{(0)} = \mathbf{y}$ \\
\textbf{Output:} $\mathbf{x}^{(n)}$
\begin{algorithmic}[1]
\WHILE {$\, n \le$ {\em max\_iterations} and $\|\mathbf{r}^{(n)}\|_2 > \epsilon$}
\STATE \textbf{Step 1:} Form temporary target signal
\STATE $\quad \mathbf{v}^{(n)} \gets \Phi^T \mathbf{r}^{(n-1)} + \mathbf{x}^{(n-1)}$
\STATE \textbf{Step 2:} Refine signal support using convex prior\\
\STATE $\quad \vecx_r^{(n)} = \argmin_\vecx \|\vecx-\mathbf{v}^{(n)}\|_2^2+\lambda J(x),$
\STATE $\quad \vecs \gets \supp(\mathbf{x}_r^{(n)}) $
\STATE \textbf{Step 3:} Estimate target signal
\STATE $\quad$ Solve $ \matPhi_s^T \matPhi_s \vecx_s = \matPhi_s^T \vecy,$ with $\matPhi_s = \matPhi(:,\vecs)$
\STATE $\quad$ Set all but largest $K$ entries in $ \vecx_s$ to zero,
\STATE $\quad \vecx^{(n)}(s)= \vecx_s(s) $
\STATE \textbf{Step 4:} Calculate data residual
\STATE $\quad \mathbf{r}^{(n)} \gets \mathbf{y} - \mathbf{\Phi} \mathbf{x}^{(n)}$
\STATE $\quad n \gets n+1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
CoLaMP has several desirable properties. First, the support set regularization (Step 2) helps to prevent signal support from growing quickly, and thus minimizes the cost of the least-squares problem in Step 3. Secondly, the use of a convex prior guarantees that a global minimum is obtained for every subproblem in Step 2, regardless of the considered clique structure. This is in stark contrast to other model-based recovery algorithms, such as LaMP\footnote{It is possible to restrict LaMP to planar Ising models, in which case a global optimum is computable \cite{cevher2009sparse}.}, and model-based CoSaMP \cite{baraniuk2010model}, which requires the solution to non-convex optimization problems to enforce structured support set models.
\begin{figure*}[tbp]
\begin{center}
\includegraphics[width=\linewidth,height=2in]{highres-eps-converted-to.pdf}
\vspace{-.25cm}
\end{center}
\caption{\vspace{0.2cm}Compressed sensing recovery results for background subtracted images using $M=3K$.}
\label{fig:qualitative_bs}
\end{figure*}
\section{Numerical Experiments}
\label{sec:experiments}
We now apply the proposed regularizer to a range of datasets to demonstrate its efficacy for various applications. Unless stated otherwise, we showcase our algorithms using overlapping cliques of size $2\times2$ as shown in Figure~\ref{fig:cliques}(b). Note that the numerical algorithms need not be restricted to those discussed above as different schemes (such as primal-dual decomposition) are needed for different situations.
\subsection{Compressive Image Recovery}
We first consider the recovery of background-subtracted images from compressive measurements. We use the ``walking2" surveillance video data \cite{WuLimYang13} with frames of dimension $288\times384$. Test data is generated by choosing two frames from a video sequence and computing the pixel-wise difference between their intensities. We compare the output of our proposed CoLaMP algorithm to that of other state-of-the-art recovery algorithms, such as overlapping group lasso \cite{leigsnering2014multipath}, fixed-point continuation (FPC) \cite{hale2007fixed} and CoSaMP \cite{needell2009cosamp}. Note that CoSaMP defines the support set using the $2K$ largest components of the error signal. The group lasso algorithm is equivalent to minimizing the objective in \eqref{colampeq} using variational method. Unlike the CoLaMP algorithm, this method does not consider prescribed signal sparsity $K$. An example recovery using $M = 3K$ measurements is shown in Figure~\ref{fig:qualitative_bs}. The sparsity level $K$ is chosen such that the recovered images account for $97\%$ of the compressive signal energy. The average $K$ across datasets is $2800$ and we fix $\lambda = 2.$ Note that the spatially clustered pixels are recovered almost perfectly. Further, we randomly generated 50 such test images from the above dataset and compared the performance of the CoLaMP, group lasso, and FPC algorithms under varying numbers of measurements from $1K$ to $5K$. The performance is measured in terms of the magnitude of reconstruction error normalized by the original image magnitude. Results are shown in Figure \ref{fig:comparison} (left). We clearly see that
the proposed smooth sparsity prior significantly improves the reconstruction quality over FPC. Furthermore, our algorithm is $7\times$ faster than the group lasso algorithm. For $M/K = 3$, the average runtime is $215s$ for CoLaMP and $1510s$ for the group lasso algorithm.
\subsection{Robust Signal Recovery}
\begin{figure*}[!tbp]
\begin{center}
\vspace{-3mm}
\includegraphics[width=0.8\linewidth,height=2in]{qualitative_cs.pdf}
\vspace{-.25cm}
\end{center}
\caption{Robust recovery results for the phantom image from a noisy compressed signal.}
\label{fig:qualitative_cs}
\end{figure*}
\begin{figure*}[!tbp]
\vspace{-2mm}
\centering
\centering
\includegraphics[width=.33\linewidth]{quantitative_bs_3-eps-converted-to.pdf}
\includegraphics[width=.33\linewidth]{quantitative_cs_1-eps-converted-to.pdf}
\includegraphics[width=.33\linewidth]{quantitative_dn_1-eps-converted-to.pdf}
\caption{Quantitative Comparison: (left) Recovery performance of compressed sensing on background subtracted images; (center) Robust compressed sensing recovery error at various SNR; (right) Average denoising gain in PSNR (dB) for various values of $\kappa$}
\label{fig:comparison}
\end{figure*}
We next showcase the suitability of CoLaMP for signal recovery from noisy compressive measurements. We consider a $100\times100$ Shepp\textemdash Logan phantom image with a support size of $K = 2636.$ A Gaussian random measurement matrix was used to sample $M=2K$ measurements, and the measurements were corrupted with additive white Gaussian noise. The signal-to-noise ratio of the resulting measurements is $10$\,dB. Figure \ref{fig:qualitative_cs} shows the original and recovered images for various recovery algorithms. We also show the output from the first few iterates of the CoLaMP algorithm. The support of the target signal is almost exactly recovered within four iterations of CoSaMP and stabilizes by the end of 10 iterations. Figure \ref{fig:qualitative_cs} also shows the recovery times of various algorithms running on the same laptop computer. CoLaMP is approximately 40$\times$ faster than the CoSaMP algorithm and it is at least 2$\times$ faster than FPC.
To enable a fair comparison, we also show the output obtained with CoLaMP using the 8-connected pixel clique in Figure \ref{fig:cliques}(c), as well as the output of the group lasso algorithm \cite{yuan2006model}, where each clique is of size $2\times2$. All these algorithms and our proposed method are implemented using ADMM. Not surprisingly, while all these algorithms beat CoLaMP in terms of runtime, their recovered signals do not match CoLaMP in terms of perceived closeness to target signal as shown in Figure \ref{fig:qualitative_cs}. The CoLaMP results are regularized by $\lambda_0=16.$ We then used an increasing value of $\lambda_n=1.02^n\lambda_0$ where $n$ is the iteration number. In practice, we obtain better results if $\lambda$ increases over time as it will heavily penalize sparse, blocky noise. For all other algorithms, we used the implementations provided by the authors.
For detailed quantitative comparisons, we repeat the above experiment using 100 Gaussian random measurement matrices and record the average reconstruction error with SNR varying from $5$\,dB to $20$\,dB.
For each algorithm, $M$ is fixed to the minimal measurement number required to give close to perfect recovery in the presence of noise. For CoLaMP and overlapping group lasso, we set $M=2K,$ whereas for FPC and non-overlapping group lasso we set $M=3.5K.$ Figure \ref{fig:comparison} (center) illustrates that CoSaMP outperforms FPC at all SNRs even with $1.5K$ fewer measurements. Group lasso performs best at low SNR while its performance flattens out starting at $10$\,dB.
\vspace{-2mm}
\subsection{Color Image Denoising}
We now consider a variant of the denoising problem (\ref{rof2}) where the image gradient is defined over color images using the decorrelated vectorized TV (D-VTV) proposed in \cite{ono2014decorrelated}
\begin{align}
\hat\vecx = & \argmin_{\mathbf{x}\in \mathbb{R}^{3N}} \sum_{c\in\mathcal{C}} \lambda\|\nabla_d
\vecx^{\ell}_c\|_2 + \|\nabla_d \vecx^{ch}_c\|_2 \nonumber \\
& \!\! \hbox{ \,\,subject to\,\, } \|\mathbf{x}-\mathbf{y}\|_2 \leq \kappa m. \label{general3}
\end{align}
Here, $\nabla_d\vecx^{\ell}\in\mathbb{R}^{2N}$ and $\nabla_d\vecx^{ch}\in\mathbb{R}^{4N}$ represent the stacked gradients of luminance and chrominance channels of the input color image, the constant $m$ depends on the noise level, and $\kappa$ is a fidelity parameter. To solve this problem numerically, we use the primal-dual algorithm described in \cite{ono2014decorrelated}, but we replace the shrinkage operator with the proximal operator \eqref{prox} to adapt our clique-based regularizer.
\begin{figure*}[!tbp]
\begin{center}
\includegraphics[width=1.0\linewidth,height=2.4in]{qualitative_dn.pdf}
\vspace{-.6cm}
\end{center}
\caption{Restoration of noisy images using Block D-VTV and existing D-VTV (best viewed in color).}
\label{fig:qualitative_dn}
\end{figure*}
Following a protocol similar to D-VTV \cite{ono2014decorrelated}, we conduct experiments using 300 images from the Berkeley Segmentation Database \cite{MartinFTM01}. Noisy images with average PSNR 20\,dB are obtained by adding white Gaussian noise. The resulting denoised output of our method (Block D-VTV) is compared to D-VTV in Figure \ref{fig:qualitative_dn}. The zoomed-in version reveals that our method exhibits less uneven color artifacts and less pronounced staircasing artifacts than the D-VTV results. A quantitative comparison measured using average PSNR gain (in dB) is drawn in Figure \ref{fig:comparison} (right) for various values of $\kappa$. Our method outperforms D-VTV by $0.25$\,dB. Also note that our method, Block D-VTV, obtains relatively better PSNR gain than the state-of-the-art D-VTV method at smaller values of $\kappa$. This observed gain is significant because smaller $\kappa$ values lead to a tighter fidelity constraint and thus a smaller solution space around the noisy input. In such situations, Block D-VTV helps to improve image quality by leveraging input from neighboring pixels.
\subsection{Video Decomposition}
\begin{figure*}[t
\begin{center}
\includegraphics[width=1.0\linewidth,height=2.7in]{qualitative_rpca.pdf}
\vspace{-.6cm}
\end{center}
\caption{\vspace{1mm}Sparse-and-low-rank decomposition using original robust PCA and proposed approach.}
\label{fig:qualitative_rpca}
\end{figure*}
We finally consider the robust PCA (RPCA) problem for structured sparsity of size $10\times 10$ as formulated in (\ref{general2}) and using Algorithm 2. We consider the same airport surveillance video data \cite{li2004statistical} as in \cite{candes2011robust} with frames of dimension $144\times176$. For a clique formed from $l\times l$ patches, we observed that $\lambda = 1/(l\sqrt{n_1})$ works best for our experiments as opposed to $\lambda = 1/\sqrt{n_1}$ used in \cite{candes2011robust}. This is because each element of the matrix $\mathbf{X}$ is shared by $l^2$ sparsity inducing terms. The resulting low rank components (background) and foreground components of three such example video frames are shown in Figure \ref{fig:qualitative_rpca}. For all the approaches, the low rank components are nearly identical. We observe that the rank of the low-rank component remains the same. As highlighted with the green box, the noisy sparse edges appearing in the original RPCA disappear from the foreground component using our proposed method. We also display the foreground component obtained using smaller overlapping cliques of size $3\times 3$, but solved using ADMM as opposed to forward-backward splitting (Algorithm \ref{alg:fbs}).
We found that for clique size of $10\times 10$ the ADMM method becomes intractable because it requires approximately 50$\times$ more memory than the proposed forward-backward splitting method with fast convolutions (i.e., $204NL$ vs.\ $4NL$).
\section{Conclusions}
\vspace{-2mm}
We have proposed a novel structured support regularizer for convex sparse recovery. Our regularizer can be applied to a variety of problems, including sparse-and-low-rank decomposition and denoising. For compressive signal recovery using large unstructured matrices, our convex regularizer can be used to improve the recovery quality of existing matching-pursuit algorithms. Compared to existing algorithms for this task, our proposed approach enjoys the capability of fast signal reconstruction from fewer measurements while exhibiting superior robustness against spurious artifacts and noise. For color image denoising, the restored images reveal more homogeneous color effects. For robust PCA, we achieve improved foreground-background separation with far fewer artifacts.
We envision many more applications that could benefit of the proposed regularizer, including deblurring and inpainting. More sophisticated directions include using support regularization for structured dictionary learning \cite{zhang2013learning} and multitask classification.
\section*{Acknowledgments}
The work of S.~Shah and T.~Goldstein was supported in part by the US National Science Foundation (NSF) under grant CCF-1535902 and by the US Office of Naval Research under grant N00014-15-1-2676. The work of C.~Studer was supported in part by Xilinx Inc., and by the US NSF under grants ECCS-1408006 and CCF-1535897.
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
Magnetic vortices are ground state magnetization configurations that form spontaneously in magnetic elements such as (sub)micron disks \cite{Ha2003,Guslienko2004}. They consist of a curling magnetization that turns out-of-plane at the disk center over a nano-scale region known as the vortex core \cite{Cowburn1999-2,Shinjo2000,Wachowiak2002}. The vortex state has been extensively studied in the past decade, partly due to its rich excitation spectrum \cite{Ding2014,Kammerer2011,Aliev2009,Boust2004,Ivanov2005}. The lowest frequency excitation of a magnetic vortex is the gyrotropic mode, which corresponds to orbit-like motion of the vortex core about its equilibrium position \cite{Choe2004,Guslienko2002,Park2003}. This mode can be excited by the application of a time varying in-plane magnetic field, and its frequency, $\fg$, depends strongly on the vortex magnetization configuration \cite{Guslienko2002}. It is therefore not surprising that the application of an external magnetic field, which modifies the vortex spin structure through the addition of a Zeeman term, will alter $\fg$ \cite{deLoubens2009,Yoo2011,Buchanan2006,Yakata2013,Ivanov2002}. Indeed this field tune-ability has been exploited for several novel applications such as tunable electronic oscillators for radiofrequency signal generation \cite{Dussaux2010,Lebrun2014,Pribiag2007} and frequency based \cite{Braganca2010} magnetic field sensing \cite{Fried2016}.
All of the above applications that make use of the field tune-ability of $\fg$ rely on a detailed understanding of how the gyrotropic frequency changes with field amplitude. It has previously been demonstrated that low amplitude, spatially uniform magnetic fields perpendicular to the disk plane modify the vortex magnetization configuration so as to result in a linear change in $\fg$ with increasing field amplitude \cite{deLoubens2009,Yoo2011,Pribiag2007}. However, there can be a significant deviation from this linear frequency behavior when the applied field amplitude is close to the disk saturation field \cite{Dussaux2010,DussauxThesis}. In this work we carry out a detailed study of the phenomena underlying this non-linear field dependence. Such an analysis will no doubt be important for the further development of the many proposed technologies which exploit the field dependence of the gyrotropic frequency.
This paper is set out as follows. In Sec.~\ref{smethod} we give a brief description of the micromagnetic simulation technique. In Sec.~\ref{snonlinear} we present simulation results exhibiting a drop-off in the gyrotropic frequency for small disks in large out-of-plane fields. In Sec.~\ref{sdeformed} we look at the effect of an external out-of-plane field on vortex core deformation during gyrotropic motion. Here it is shown that in the case of a small disk in large out-of-plane fields, there is an increased deformation of the core magnetization profile due to the out-of-plane demagnetizing field created when the core is shifted laterally. Finally in Sec.~\ref{scalcs}, we relate this increased core deformation to the non-linear field dependence of the vortex gyrotropic resonance frequency by performing a detailed analysis of the field dependence of the vortex stiffness coefficient and gyroconstant. Here it is shown that for small disks in large out-of-plane fields, deformation of the vortex core driven by the demagnetizing field leads to a sharply decreasing exchange contribution to the vortex stiffness coefficient, resulting in a drop-off in the gyrotropic frequency.
\section{Micromagnetic Simulation}
\label{smethod}
We will concentrate primarily on simulations run in MuMax3 \cite{Vansteenkiste2014} for Permalloy-like disks with saturation magnetization $\ms=800\,$kA/m, exchange constant $A=$ 13$\,$pJ/m, magnetic damping parameter $\alpha=0.008$ and nil intrinsic anisotropy. A cell size of $3\times3\times3.75\,$nm$^{3}$ was used for the range of simulated disk geometries. For simplicity, in this work we will only consider the case of a vortex with a positive core polarity which is aligned with the external out-of-plane field. The equilibrium magnetization configuration was first found by initializing a vortex like spin structure, applying the desired static external field and evolving the magnetization (without precession) to reach the minimum energy state. To induce gyrotropic motion (and hence find $\fg$), vortex core dynamics were driven via the application of a transverse sinc pulse field: $A\sin(\omega(t-t_0))/(\omega(t-t_0))$. This excites modes up to a cut-off frequency of $\omega/2\pi=30\,$GHz. The resonant eigenfrequencies can then be determined by performing a Fourier analysis of the time dependent spatially averaged in-plane magnetization \cite{Baker2016}. We note that while this process also excites higher frequency spin waves \cite{Aliev2009,Park2005,Ivanov2005,Buess2004}, this work will focus on the lowest frequency excitation corresponding to the vortex gyrotropic mode.
\section{Non-linear frequency responses}
\label{snonlinear}
The simulated gyrotropic frequencies as a function of uniform out-of-plane field amplitude, are shown in Fig.~\ref{fdropoff}(a) for several disk diameters with thickness $L=30\,$nm. To be able to obtain comparable results across all disk geometries, $\fg$ has been normalized by the simulated frequency found in zero out-of-plane field for each disk size. Likewise the field amplitude has been normalized by the disk saturation field, $H_\mathrm{S}$, which was found by stepping through fields of increasing amplitude and plotting the equilibrium out-of-plane magnetization to determine when the vortex state is no longer present. The result of such a process is shown in Fig.~\ref{fdropoff}(b) for the 384 nm disk diameter where the saturation field is found to be $885\pm5\,$mT.
We see that while the gyrotropic frequency scales linearly in low field amplitudes, there is a drop-off (or mode `softening') in $\fg$ when the applied field is close to $\hs$. Moreover, this drop-off becomes less distinct as the disk radius is increased, with $\fg$ being almost linear up to the saturation field for the 768$\,$nm disk. This drop-off in $\fg$ has been previously observed via micromagnetic simulation and experiment \cite{Dussaux2010,Dussaux2011} in disks with a diameter of 170$\,$nm.
We finally note that we have confirmed the frequency behavior in Fig.~\ref{fdropoff}(a) for the 192 nm disk when decreasing the mesh size to $2\times 2 \times 3.75\,$nm$^3$. We have also compared these frequencies to results obtained using a similar time domain method in OOMMF\cite{oommf} ($3\times3\times3\,$nm$^{3}$ cell size) and an eigenmode method \cite{dAquino2009,Metaxas2014} in FinMag (derived from Nmag\cite{Fischbacher2007}) with good agreement being found across the range of applied fields [Fig.~\ref{fdropoff}(c)]. We used a non-uniform finite element mesh in FinMag with a characteristic internode length of 3$\,$nm at the disk center which smoothly transitioned to 5$\,$nm at the disk edge.
\begin{figure}[t]
\centering
\includegraphics[width=7.5cm]{Fig1}
\caption{(a) Simulated gyrotropic frequency as a function of uniform out-of-plane field amplitude for disks of thickness $L=30\,$nm and a range of diameters. $\fg$ has been normalized by the simulated frequency in zero out-of-plane field while the field amplitude has been normalized by the disk saturation field. (b) Out-of-plane magnetization for a slice through the disk center for a 384$\,$nm diameter disk in increasing field amplitudes (c) Comparison of the simulated values of $\fg$ found using MuMax3 (for two different cell sizes) with the frequencies found using OOMMF and FinMag (disk diameter is 192$\,$nm).}
\label{fdropoff}
\end{figure}
\section{Core dynamics and deformation}
\label{sdeformed}
To begin to understand the cause of the non-linear frequency behavior, it is necessary to look at how the vortex magnetization configuration changes \emph{during} gyrotropic motion. While this might initially seem unrelated to the observed drop-off in $\fg$, in later sections it will be shown that deformation of the vortex core magnetization profile during gyrotropic motion is the primary cause of the frequency drop-off.
\subsection{Differentiating static and dynamic core deformation}
Figure \ref{fdyncore} shows slices of the out-of-plane magnetization near the vortex core for a range of normalized field amplitudes. These one-dimensional slices of the thickness averaged magnetization have been taken so as to intersect the vortex core center and the disk center. For simplicity we will concentrate on results obtained for the 192 and 768$\,$nm disks. The solid red lines are for gyrotropically resonating (`dynamic') vortex cores whereas the dashed black lines are for `static' vortex cores which have been shifted to new equilibrium positions by static in-plane magnetic fields. All core displacements in Fig.~\ref{fdyncore} are $\approx$ 3$\,$nm in the positive $x$ direction to enable comparison between the different profiles.
At zero out-of-plane field, the static displaced cores are highly symmetric in both the small and large disk. While this symmetry is retained for the larger disk size when the out-of-plane field is increased, for the 192$\,$nm disk the static displaced cores become deformed with the minimum of the magnetostatic halo\cite{Gaididei2010} [labelled in Fig.~\ref{fdyncore}(a)] moving up from its position for a centered vortex (referenced by the solid horizontal lines) in the direction of core displacement. This is accompanied by a downward shift of the magnetostatic halo on the side of the core which is away from the direction of core displacement. Notably the magnitude of this asymmetry in the magnetization profile of statically displaced cores in the 192$\,$nm disk increases as the external field amplitude is increased.
\begin{figure}[t!]
\centering
\includegraphics[width=7.5cm]{Fig2}
\caption{Out-of-plane magnetization profile close to the vortex core for a 192 and 768$\,$nm disk diameter in several normalized field amplitudes. The solid red lines show the magnetization profile while the core is undergoing gyrotropic motion. The black dashed lines show the magnetization profile when the core has been shifted to a new equilibrium position by a static in-plane magnetic field. The horizontal black lines reference the minimum of the magnetostatic halo when the vortex core is stationary and at the disk center.}
\label{fdyncore}
\end{figure}
Dynamically displaced cores are however clearly asymmetric for both disk sizes in zero out-of-plane field. As the out-of-plane field is increased, the asymmetry of the statically and dynamically displaced core profiles become comparable. Indeed, for $H/\hs$ = 0.85, the two profiles are almost identical for both disks (i.e.~highly asymmetric in the 192$\,$nm disk and highly symmetric in the 768$\,$nm disk).
In the following section, this behavior will be shown to be due to two forms of core deformation. Namely for a disk in zero or low out-of-plane fields core deformation is predominantly due to the gyroforce resulting from vortex core motion and is thus only seen in the dynamically displaced profiles. However for large out-of-plane fields asymmetry in the core magnetization profile is driven by the demagnetizing field created when the core is shifted laterally and is thus observed in both statically and dynamically displaced vortex cores.
\subsection{Gyrofield driven core deformation}
Deformation of the magnetization profile of a dynamic vortex core (i.e.~one undergoing gyrotropic motion) has previously been observed in the absence of an out-of-plane field \cite{Yamada2007,Waeyenberge2006,Novosad2005} and was attributed to the gyroforce \cite{Guslienko2008b} resulting from the motion of the non-uniform magnetization configuration. The effect of the gyroforce can be described by an effective field, the gyrofield, the out-of-plane component of which causes core deformation and can be calculated by \cite{Guslienko2008b}:
\begin{equation}
h_z(r,t)=\frac{1}{\gamma}\frac{(\mathbf{m}\times\mathbf{\dot{m}})_z}{(m_z+p)^2}
\label{egfield}
\end{equation}
where $\gamma$ is the gyromagnetic ratio, $\mathbf{m}$ is the magnetization unit vector, $m_z$ is the out-of-plane magnetization, $p$ is the vortex core polarity and $\mathbf{\dot{m}}=-(\mathbf{\dot{X}}\cdot\mathbf{\nabla})\mathbf{m}$ where $\mathbf{\dot{X}}$ is the core velocity which can be determined from the vortex core position and the gyrotropic frequency.
\begin{figure}[t!]
\centering
\includegraphics[width=8.3cm]{Fig3}
\caption{The gyrofield close to the vortex core as calculated using Eq.~(\ref{egfield}) for disk diameters of 192 and 768$\,$nm and three different normalized out-of-plane field values.}
\label{fgyrofield}
\end{figure}
In Fig.~\ref{fgyrofield} we have numerically calculated the gyrofield for each core profile shown in Fig.~\ref{fdyncore}. For both disk diameters the gyrofield decreases as the out-of-plane field amplitude is increased. This is because increasing the field leads to a higher degree of out-of-plane canting of the curling magnetization which reduces the spatial magnetization gradient and thus the gyrofield (via $\mathbf{\dot{m}}$ which depends on $\nabla \mathbf{m}$). The gyrofield profile is consistent with the core asymmetry of dynamically displaced cores observed in Fig.~\ref{fdyncore}, i.e.~in the direction of core displacement the halo is shifted upward due to a positive gyrofield. On top of this, the weakening of the gyrofield with increasing out-of-plane fields explains why, as observed in Fig.~\ref{fdyncore}, the static and dynamic displaced core profiles become similar as the field amplitude is increased (due to the dynamic source of deformation, the gyrofield, being strongly reduced).
\begin{figure}[t!]
\centering
\includegraphics[width=7.5cm]{Fig4}
\caption{(a,b) Slices through the created vortex-like configuration for the (a) 192 and (b) 768$\,$nm disk diameter. The solid red lines show the magnetization profile for a core shifted 10$\,$nm in the positive $x$ direction. The black dashed lines show the magnetization profile for a centered vortex core. (c,d) Demagnetizing field profiles for a displaced and centered vortex core in the (c) 192 and (d) 768$\,$nm disk. (e,f) Demagnetizing field close to the vortex core for the (e) 192 and (f) 768$\,$nm disk. The solid horizontal lines reference the magnetostatic field 25$\,$nm either side of the core maximum for the case of a displaced vortex core. The dashed horizontal line marks the magnetostatic field 25$\,$nm either side of the core maximum for the case of a centered vortex core.}
\label{fmademag}
\end{figure}
\subsection{Demagnetizing field driven core deformation}
The deformation of dynamic and static displaced vortex cores for small disks in high out-of-plane fields can be shown to be a magnetostatic effect by looking at the demagnetizing field created by statically shifted, \emph{non-deformed} vortex cores. Note that it is necessary to look at the demagnetizing field created by symmetric, non-deformed core profiles to avoid confusion related to whether asymmetries in the demagnetizing field are the cause of the core deformation or a consequence of it. To do this we created simple magnetization configurations crudely approximating that of a shifted (by 10$\,$nm) and unshifted vortex in the presence of a high out-of-plane field [Fig.~\ref{fmademag}(a) and (b)]. The demagnetizing field profile (averaged across the disk thickness) created by these `artificial' magnetization configurations is shown Fig.~\ref{fmademag}(c) and (d).
As shown in Fig.~\ref{fmademag}(c) and (d), for both disk sizes the $z$ component of the magnetostatic field is strongest around the vortex core where the out-of-plane magnetization is highest. The field amplitude then drops-off close to the disk lateral boundary since here there is no adjacent moments to reinforce the demagnetizing field. For the smaller disk size, the reduced lateral dimensions mean that these edge effects result in a spatial gradient of the demagnetizing field close to the disk center [Fig.~\ref{fmademag}(c)]. This field gradient close to the vortex core in the smaller disk size leads to an asymmetry in the demagnetizing field created when the core is shifted laterally. Namely, in the direction of core displacement, the demagnetizing field a given distance from the core center will decrease in amplitude (become less negative) as it will be closer to the disk edge. Similarly, on the side away from the direction of core displacement the magnetostatic field will increase in amplitude (become more negative) as it will be farther away from the disk's lateral boundary. We would expect this effect to be significantly reduced for the larger disk diameter as here the influence of the disk edge on the created demagnetizing field becomes negligible close to the vortex core.
Such asymmetry in the demagnetizing field profile is seen in Fig.~\ref{fmademag}(e) which shows the same field profile as Fig.~\ref{fmademag}(c), however here we concentrate on the field 25$\,$nm either side of the core maximum. Note that the field profile of the non-displaced vortex has been translated 10$\,$nm in the positive $x$ direction to enable comparison between the two cases. For the smaller disk size there is a clear difference between the field amplitude 25$\,$nm either side of the core maximum (referenced by the solid horizontal lines) for the vortex with a displaced core. As expected this asymmetry is significantly reduced for the larger disk size [Fig.~\ref{fmademag}(f)]. Notably the asymmetry in the demagnetizing field profile is consistent with deformation of statically shifted vortex cores observed in Fig.~\ref{fdyncore}. Namely in the direction of core displacement the height of the magnetostatic halo is increased consistent with a reduced negative demagnetizing field.
We also note that when compared to the demagnetizing field created by a vortex with a centered core, the field increase to the right of the core is greater than the decrease to the left. This can be seen in Fig.~\ref{fmademag}(e) when comparing the distance between the solid and dashed horizontal lines: the dashed line references the magnetostatic field 25$\,$nm either side of the core for a centered vortex profile. The change in the demagnetizing field's `net' direction when the core is displaced is therefore aligned with the core polarity, which is in contrast to gyrofield driven deformation where the field opposing the core is always greater than that reinforcing it \cite{Guslienko2008b,Yamada2007,Waeyenberge2006}. Such behavior in the gyrofield profile can be identified in Fig.~\ref{fgyrofield} where, for the smaller disk in zero out-of-plane field, the maximum gyrofield opposing the vortex core is $\approx$ -4.80$\,$mT compared to $\approx$ 4.55$\,$mT reinforcing the core polarity.
\begin{figure}[t!]
\centering
\includegraphics[width=6cm]{Fig5}
\caption{Plots of the evolution of the spatially average out-of-plane magnetization as a function of time since gyrotropic motion was induced for a 192 and 768$\,$nm disk in various out-of-plane fields. The gyrotropic motion has been driven by a in-plane sinusoidal field with frequency equal to the $\fg$ values in Fig.~\ref{fdropoff} found via the sinc pulse excitation detailed in Sec.~II.}
\label{fdmz}
\end{figure}
\subsection{Time evolution of the out-of-plane magnetization}
The differing net directions of the field produced by the demagnetizing field and gyrofield driven core deformation results in different changes in the spatially averaged out-of-plane magnetization component ($\bar{m}_z$) when the vortex core is dynamically displaced. This effect is seen in Fig.~\ref{fdmz} where we have plotted $\bar{m}_z$ as a function of time after the start of gyrotropic motion for the two disk sizes in a range of field amplitudes. For both diameters, when in zero out-of-plane field, $\bar{m}_z$ decreases as the core oscillation amplitude increases with time. This is consistent with vortex core deformation at low out-of-plane fields being driven by the gyrofield which creates a net field that opposes the core polarity. For the larger disk size this effect decreases with increasing out-of-plane field since the amplitude of the gyrofield decreases (Fig.~\ref{fgyrofield}) and demagnetizing field driven deformation is minimal. Indeed at high out-of-plane fields the gyrofield has decreased to an extent where there is almost no vortex core deformation [Fig.~\ref{fdyncore}(f)] resulting in the change in $\bar{m}_z$ being $\approx 0$.
For the smaller disk diameter however, the demagnetizing field driven deformation is significant. As already shown, the change in the net direction of this field when the core is displaced aligns with the core polarity, thus leading to an increase in the average out-of-plane magnetization of the displaced vortex state. Indeed for the case of $H/\hs=0.85$ we see that $\bar{m}_z$ becomes positive as the core oscillation amplitude increases with time. This is consistent with core deformation predominantly coming from demagnetizing field effects.
\section{Gyroconstant and stiffness coefficient calculations}
\label{scalcs}
In the previous section it was shown that for disks of small lateral dimensions in large out-of-plane fields, the demagnetizing field created when the vortex core is shifted laterally leads to significant deformation of the magnetization profile. Moreover, this high field deformation was shown to result in an increase in the average out-of-plane magnetization of the displaced vortex state. In this section we will relate this deformation to the non-linear field dependence of the gyrotropic frequency observed in Fig.~\ref{fdropoff}(a). To do this we analyze the field dependence of the vortex stiffness coefficient, $\kappa$, and the gyroconstant, $G$ which together enable a calculation of $\fg$ according to the equation \cite{Guslienko2002}:
\begin{equation}
2\pi\fg=\frac{\kappa}{G}.
\label{ekG}
\end{equation}
The gyroconstant describes the lateral force acting on the core as a result of the motion of the non-uniform magnetization configuration \cite{Thiele1973,Huber1982}. $G$ can be calculated from the thickness averaged spin structure using the equation:
\begin{equation}
G=\frac{\ms L}{\gamma}\iint_{A}\mathbf{m}\cdot \left (\frac{d\mathbf{m}}{dx}\times\frac{d\mathbf{m}}{dy}\right )dx dy
\label{eG}
\end{equation}
where for the small amplitude oscillations considered here the integration should be done over the area of the vortex core \cite{Fried2016}.
In Eq.~(\ref{ekG}), the stiffness coefficient describes the resorting force acting on the vortex core due to an increase in magnetic energy when it is displaced. For small shifts of the vortex core one can assume parabolic scaling of the magnetic energy with core displacement \cite{Guslienko2002,Dussaux2013,Sukhostavets2013}:
\begin{equation}
W(X)=W(0)+\frac{1}{2}\kappa X^2+\mathcal{O}(X^4)
\label{ek}
\end{equation}
where $W$ is the total magnetic energy and $X=|\mathbf{X}|$ is the core displacement. Note that the Zeeman energy associated with the in-plane driving field should not be taken into account when determining $\kappa$. This is due to the fact that the in-plane field is only a mechanism to translate the core within the disk, thus enabling the confining potential of the vortex to be probed.
\subsection{Field dependence of the gyroconstant}
The influence of a spatially uniform out-of-plane field on $G$ has previously been analytically studied\cite{deLoubens2009}. There, the gyroconstant was calculated using the expression:
\begin{equation}
G(H)=G(0)(1-p\cos\theta)
\label{eGdeLoubens}
\end{equation}
where $\theta$ is the polar angle (i.e. down from the $z$ axis) in which the magnetization outside the core is tilted \cite{PigeauThesis}. The above equation shows that $G$ decreases as the curling magnetization tilts to align with the external field. Such behavior can be understood from the fact that out-of-plane canting of the curling spins decreases the spatial gradient of the magnetization within the vortex core leading to a smaller value of $G$.
As previously shown\cite{Fried2016}, we find good agreement between the gyroconstant calculated directly from Eq.~(\ref{eG}) and values of $G$ predicted by Eq.~(\ref{eGdeLoubens}) when $\theta$ is taken from the minimum of the magnetostatic halo. For example, in the case of the 192$\,$nm disk we observe a maximum discrepancy of $\approx$ 2\% across the range of applied fields. For this reason, in the following discussion all values of $G$ have been calculated using the simpler Eq.~(\ref{eGdeLoubens}) with $\theta$ extracted from the minimum of the magnetostatic halo.
\begin{figure}[t]
\centering
\includegraphics[width=7.5cm]{Fig6}
\caption{(a) Gyroconstant field dependence (normalised by its value in zero out-of-plane field). (b) Comparison of the static magnetization profiles of a centered vortex core for the 192 and 768 nm disks in various out-of-plane field amplitudes. The one dimensional slice has been taken across the disk center of the thickness averaged magnetization.}
\label{fG}
\end{figure}
Resultant calculated values of the gyroconstant are shown for a range of out-of-plane field amplitudes in Fig.~\ref{fG}(a). Consistent with the above discussion, $G$ decreases with increasing field amplitude. One will note that $G$ is consistently smaller for the larger disk diameter when in finite out-of-plane fields. To understand this behavior we compare the equilibrium out-of-plane magnetization configurations of the two disk sizes in a range of normalized field amplitudes [Fig.~\ref{fG}(b)]. Clearly there is a deepening of the magnetostatic halo for the smaller disk size when in a finite out-of-plane field. This is due to the nearby out-of-plane canting of the magnetization at the disk boundary which reinforces the negative demagnetizing field created by the vortex core (which is normally at the source of the halo \cite{Gaididei2010}). This halo deepening increases the spatial magnetization gradient within the vortex core resulting in a larger value of $G$ [as observed in Fig.~\ref{fG}(a)]. Nevertheless we note that for both disk sizes the gyroconstant continues to decrease with increasing field, a behavior which, ignoring field-dependencies of $\kappa$, would lead to a gyrotropic frequency which monotonically increases with out-of-plane field. Therefore to explain the drop-off in $\fg$ at high field amplitudes we must also to look at how $\kappa$ changes with increasing out-of-plane field.
\subsection{Field dependence of the stiffness coefficient}
The influence of a uniform out-of-plane field on the stiffness coefficient was also studied in Ref.~\onlinecite{deLoubens2009}. This was done assuming that the only contribution to $\kappa$ was from the magnetostatic energy of the dipole charges that are generated by the in-plane magnetization when the core is shifted laterally. This leads to the expression for the field dependent stiffness coefficient:
\begin{equation}
\kappa(H)=\kappa(0)\sin^2\theta
\label{ekdeLoubens}
\end{equation}
showing that $\kappa$ also decreases as the magnetization cants to align with the applied field. Qualitatively this behavior can be understood from the fact that, as the out-of-plane field is increased there is a reduced in-plane magnetization component, meaning a weaker in-plane demagnetizing field is created by the curling spins when the vortex core is displaced. This results in a smaller increase in the system's magnetostatic energy when the core is displaced, leading to a lower value of $\kappa$.
If one assumes that the canting of the curling magnetization varies linearly with field amplitude (i.e. $\cos\theta=H/\hs$), Eqs.~(\ref{eGdeLoubens}) and ({\ref{ekdeLoubens}) combine to give \cite{deLoubens2009}:
\begin{equation}
\fg(H)=\fg(0)\left (1+p\frac{H}{\hs}\right ).
\label{efH}
\end{equation}
This suggests that the gyrotropic frequency will increase linearly with field amplitude up to the disk saturation field (at which point the gyrotropic mode can no longer be excited).
To determine the stiffness coefficient from our simulations, the vortex core was driven by an in-plane sinusoidal field with amplitude 0.05$\,$mT and frequency equal to $\fg$ as found when analyzing the sinc pulse induced gyrotropic motion. This is necessary as the sinc pulse field used to determine $\fg$ induces very small amplitude oscillations in the core ($X<1\,$nm) which results in an increased numerical uncertainty in the extracted value of $\kappa$. The stiffness coefficient is then calculated by performing a parabolic fit to the total magnetic energy as a function of core displacement. It is again noted that the Zeeman energy associated with the in-plane driving field (which was calculated by multiplying the average in-plane magnetization by the driving field amplitude at that time), was not taken into account when performing such a fit.
We also note that driving large amplitude oscillations of the vortex core can lead to non-linear behavior in the gyrotropic mode dynamics which can manifest as resonant peak splitting \cite{Buchanan2007} and peak foldover \cite{Buchanan2007,Drews2012,Guslienko2010}. However we have confirmed that the described sinusoidal field is not large enough to induce such behavior. This was done by plotting the maximum core displacement versus the driving frequency and comparing this to values of $\fg$ determined from the sinc excitation field. As observed in Fig.~\ref{fkappa}(a) the frequency resulting in the largest maximum core displacement closely corresponds to the resonant gyrotropic frequency found when the core is driven by a low amplitude transverse sinc field (as referenced by the vertical black dashed lines).
\begin{figure}[t!]
\centering
\includegraphics[width=8.4cm]{Fig7}
\caption{(a) Maximum core displacement as a function of sinusoidal driving frequency for a range of disk geometries (in a field close to $\hs$). (b) Total stiffness coefficient normalized by its value in zero out-of-plane field as a function of field amplitude for the 192 and 768$\,$nm disks. (c) The contribution to $\kappa$ arising from the magnetostatic energy associated with the in-plane magnetization and (d) the exchange energy. Both of these contributions have been normalized by the total stiffness coefficient in zero out-of-plane field. (e) The change in exchange energy (relative to that for a centered vortex) as a function of core displacement for the 192$\,$nm disk in $H/\hs$ = 0 and 0.85.}
\label{fkappa}
\end{figure}
The resulting normalized values of $\kappa$ are shown for the two disk sizes in a range of field amplitudes in Fig.~\ref{fkappa}(b). We first note that, as expected from Eq.~(\ref{ekdeLoubens}), $\kappa$ decreases as the curling magnetization cants out-of-plane with increasing field amplitude. Such behavior in $\kappa$ acts to decrease $\fg$ (rather than increase it, as occurs in reality) showing that the linear frequency behavior observed for low out-of-plane fields is driven by changes in $G$.
At intermediate field amplitudes $\kappa$ is larger for the smaller disk diameter. However, this increase in $\kappa$ is compensated by a larger $G$ [as seen in Fig.~\ref{fG}(a)] resulting in similar normalized values of $\fg$ for the two disk sizes. As the field is further increased there is a sharper drop-off in $\kappa$ for the smaller disk size with the normalized stiffness coefficients of the two disk diameters becoming approximately equal close to $\hs$. This sharper drop-off in the stiffness coefficient of the smaller disk diameter is consistent with the non-linear frequency behavior observed in Fig.~\ref{fdropoff}(a). Namely this sharp drop-off leads to changes in $\kappa$ becoming comparable and then greater than changes in $G$, thus leading to a plateau and then drop-off in $\fg$ at high field amplitudes. In the next section we separate the individual contributions to the stiffness coefficient and demonstrate that this sharper drop-off in $\kappa$ is due to a decreased exchange contribution.
\subsection{Stiffness coefficient decomposition}
The field behavior of the total vortex stiffness can be decomposed into two contributions. The first of these results from the magnetostatic charges created by the in-plane magnetization when the vortex core is displaced and will be labeled $\kdip$. The second contribution, $\ke$, results from the change in exchange energy when the vortex core is displaced. $\ke$ and $\kdip$ were determined using an identical process to the total stiffness coefficient, however here only the appropriate energy contribution (as extracted from the simulations) is taken into account when performing the parabolic fit. It should be noted that we observed a non-zero contribution to $\kappa$ from the Zeeman energy associated with the out-of-plane magnetization, $\kz$, and the magnetostatic energy associated with the out-of-plane magnetization, $\kdoop$. However these two contributions cancel across the range of applied fields [$(\kdoop+\kz)/\kappa \leq 0.04$] and therefore have a negligible effect on $\fg$.
Fig.~\ref{fkappa}(c) shows the field dependence of $\kdip$. Note that $\kdip$ has been normailzed by the value of the total stiffness coefficient in zero out-of-plane field. Consistent with the fact that Eq.~(\ref{ekdeLoubens}) was derived considering only the magnetostatic charges created by the in-plane magnetization when the vortex core is displaced, $\kdip$ varies as $\sin^2\theta$ to within 2\% for both disk sizes. We also note a sharper decrease in $\kdip$ for the 768$\,$nm disk as the field amplitude is increased. This is due to the fact that, as shown in Fig.~\ref{fG}(b), for finite field amplitudes the curling magnetization around the vortex core is less in-plane for the larger disk size (i.e.~the magnetostatic halo is higher). This means a weaker demagnetizing field is created by the in-plane magnetization when the vortex core is displaced, leading to a sharper decrease in $\kdip$.
\begin{figure*}[t!]
\centering
\includegraphics[width=12cm]{Fig8}
\caption{(a) The gyrotropic frequency associated with the exchange contribution to the stiffness coefficient for a range of disk diameters. (b) The simulated values of $\fg$ [as in Fig.~\ref{fdropoff}(a)] minus the frequency associated with the exchange contribution to the stiffness coefficient for a range of disk diameters.}
\label{fminusexch}
\end{figure*}
The field dependence of $\ke$ is shown in Fig.~\ref{fkappa}(d). For zero out-of-plane field, $\ke$ is positive for both disk sizes. $\ke$ then decreases as the out-of-plane field amplitude is increased. Notably this decrease is sharper for the smaller disk size, with $\ke$ even becoming negative for large $\hs$ (inferring that here the exchange energy favors core displacement). This sharper decrease $\ke$ is consistent with the more pronounced drop-off in the total stiffness coefficient for the smaller disk size when in large out-of-plane fields [Fig.~\ref{fkappa}(b)].
We have confirmed that the non-zero $\ke$ in zero field is primarily a result of vortex core deformation. This dominates the contribution to $\ke$ resulting from changes in the curling magnetization configuration when the core is displaced \cite{Guslienko2002}. This was demonstrated by calculating an exchange-driven core stiffness for statically displaced cores $\kappa_{\mathrm{E,static}}$. For the 192$\,$nm disk in zero out-of-plane field [where there is no deformation of the statically shifted core, i.e.~see Fig.~\ref{fdyncore}(a)] we found that $\kappa_{\mathrm{E,static}}$ was two orders of magnitude lower than the corresponding value of $\ke$ (the exchange-driven stiffness coefficient of the \emph{dynamic} core). This confirms that the dominant contribution to $\ke$ originates from vortex core deformation.
The out-of-plane field dependence of $\ke$ can therefore be understood by looking at how gyrofield-, and demagnetizing field-driven core deformation change the vortex exchange energy. This is done in Fig.~\ref{fkappa}(e) where we plot the change in exchange energy as a function of (dynamic) core displacement. We have done this for the 192$\,$nm disk in $H/\hs=0$ (i.e.~where deformation is driven by the gyrofield) and $H/\hs=0.85$ (where deformation is predominately a demagnetizing field effect). When in zero out-of-plane field, we see that the exchange energy increases as the core is displaced. This is consistent with the fact that the gyrofield decreases the average out-of-plane magnetization (Fig.~\ref{fdmz}). This results in the displaced vortex having a higher exchange energy than the centered vortex, since the magnetization is now further away from being uniformly saturated out-of-plane. In contrast, for the case of $H/\hs$ = 0.85 we see that the exchange energy decreases as the core is displaced. This is consistent with the fact that here deformation is driven by the demagnetizing field which results in an increase in the average out-of-plane magnetization (Fig.~\ref{fdmz}). This leads to the displaced vortex having a lower exchange energy than the centered vortex, since the magnetization is now closer to being uniformly saturated out-of-plane.
The field behavior of $\ke$ can now be qualitatively explained as follows. When in zero out-of-plane field $\ke$ is positive for both disk sizes. This is consistent with the fact that here, the gyrofield (which generates a positive $\ke$) drives core deformation. For the larger disk size $\ke$ then decreases with increasing field amplitude due to a weakened gyrofield (Fig.~\ref{fgyrofield}). Indeed for large $H/\hs$, $\ke$ tends towards zero consistent with the fact that there is no strong core deformation [Fig.~\ref{fdyncore}(f)], due to a significantly reduced gyrofield. For the smaller disk size, $\ke$ also decreases with increasing field amplitude, however this decrease is sharper due to the presence of demagnetizing field driven core deformation (which generates a negative $\ke$). Indeed as previously noted, $\ke$ becomes negative at large out-of-plane fields, consistent with the fact that here core deformation is predominately a demagnetizing field effect.
In Fig.~\ref{fminusexch} we show the field dependence of $\ke$ is the cause of the non-linear frequency behavior observed in Fig.~\ref{fdropoff}(a) for small disks in large out-of-plane fields. In Fig.~\ref{fminusexch}(a) we plot the field dependence of the frequency associated with the exchange contribution to the stiffness coefficient (i.e.~$f_{\mathrm{G,E}}=\ke/G$) for the range of disk geometries considered in Fig.~\ref{fdropoff}(a). For the larger disk size, $f_{\mathrm{G,E}}$ is almost independent of the field amplitude up to $\hs$. However for smaller disks, $f_{\mathrm{G,E}}$ is only constant at low field amplitudes, after which there is a sharp drop-off in the frequency. Consistent with the simulated frequency behavior in Fig.~\ref{fdropoff}(a), as the disk diameter is reduced, the decrease in $f_{\mathrm{G,E}}$ becomes sharper and begins at a lower field amplitude. On top of this, when subtracting $f_{\mathrm{G,E}}$ away from the simulated values of $\fg$, we find that for all disk sizes the resulting frequencies scale (almost) linearly up to the disk saturation field [Fig.~\ref{fminusexch}(b)]. This result explicitly confirms that the non-linear frequency behavior observed for small disks in Fig.~\ref{fdropoff}(a) is a result of a strongly decreasing $\ke$, which even becomes negative, in large out-of-plane fields.
We finally note that it was found that the observed non-linear frequency behavior shows little dependence on the disk thickness. For example when comparing the normalized values of the gyrotropic frequency for disks of thickness $L$ = 15 and 30$\,$nm with a constant diameter of $2R$ = 192$\,$nm we see a maximum discrepancy of $\approx 2\%$ across the range of applied fields. This behavior is not surprising given that we have just shown that this non-linear behavior in $\fg$ for small disks in large out-of-plane fields is a result of a strongly decreasing $\ke$. This decrease in $\ke$ is a result of demagnetizing field driven vortex core deformation, which is present as a result of the close proximity of the vortex core to the disk's edge (and is therefore largely independent of the element thickness).
\section{Conclusion}
\label{sconc}
In conclusion we have shown that disks with small lateral dimensions exhibit an increasingly non-linear field dependence of the gyrotropic mode frequency in large out-of-plane spatially uniform magnetic fields. This non-linear frequency behavior was found to be associated with an increased deformation of the vortex core which arises for small disks in large out-of-plane fields due to asymmetries in the demagnetizing field profile when the core is shifted laterally. This form of deformation increases the average out-of-plane magnetization of the displaced vortex state and thus leads to a decrease in the exchange energy when the vortex core is shifted laterally. This results in a sharper decrease in the total vortex stiffness coefficient, which leads to the observed mode-softening or drop-off in $\fg$ for small magnetic disks in large out-of-plane fields.
\begin{acknowledgments}
This work was supported by the Australian Research Council's Discovery Early Career Researcher Award scheme (DE120100155), a research grant from the United States Air Force (Asian Office of Aerospace Research and Development, AOARD), the University of Western Australia's Early Career Researcher Fellowship Support and Research Collaboration Award schemes and by resources provided by the Pawsey Supercomputing Centre with funding from the Australian Government and the Government of Western Australia. The authors thank Maximilian Albert and Rebecca L.~Carey for assistance with the Finmag package.
\end{acknowledgments}
|
\section{Concluding Remarks}
The paper presents a novel convex optimization framework for transductive CGRL and a scalable algorithmic solution with guaranteed global optimum
and a time complexity that does not depend on the sizes of input graphs.
Our experiments on multi-graph data sets provide strong evidence for the superior power of the proposed approach in modeling cross-graph inference and large-scale optimization.
\section*{Acknowledgements}
We thank the reviewers for their helpful comments.
This work is supported in part by the
National Science Foundation (NSF) under grants IIS-1216282, 1350364, 1546329.
\section{The Proposed Method}
\begin{figure*}[ht]
\centering
\resizebox{1.3cm}{!}{$\mathscr{P}\bigg($}
\begin{minipage}[c]{0.11\linewidth}
\centering
$\underbrace{\includegraphics[width=0.9\linewidth]{figures/product_graphs_simple/g1.png}}_{\text{\large $G^{(1)}$}}$
\end{minipage
,
\begin{minipage}[c]{0.09\linewidth}
\centering
\vspace{0.35cm}
$\underbrace{\includegraphics[width=0.9\linewidth]{figures/product_graphs_simple/g2.png}}_{\text{\large $G^{(2)}$}}$
\end{minipage}
,
\begin{minipage}[c]{0.11\linewidth}
\centering
$\underbrace{\includegraphics[width=0.9\linewidth]{figures/product_graphs_simple/g3.png} }_{\text{\large $G^{(3)}$}}$
\end{minipage
\
\resizebox{1.25cm}{!}{$\bigg)=$}
\begin{minipage}[c]{0.375\linewidth}
\centering
\includegraphics[width=0.75\linewidth]{figures/product_graphs_simple/pg.png}
\end{minipage}
\caption{Graph product of $G^{(1)}$, $G^{(2)}$ and $G^{(3)}$.
Each vertex in the resulting graph $\mathscr{P}\big(G^{(1)}$, $G^{(2)}$, $G^{(3)}\big)$ corresponds to a multi-relation across the original graphs. E.g., vertex \texttt{3.II.B} in $\mathscr{P}$ corresponds to multi-relation \texttt{(3,II,B)} across $G^{(1)}$, $G^{(2)}$ and $G^{(3)}$.}
\label{fig:gp-example}
\end{figure*}
\label{sec:proposed}
We introduce our notation in \ref{sec:notations}
and the notion of graph product (GP) in \ref{sec:gp}.
We then narrow down to a specific GP family with desirable computational properties in \ref{sec:gp},
and finally propose our GP-based optimization objective in \ref{sec:obj}.
\subsection{Notations}
\label{sec:notations}
We are given $J$ heterogeneous graphs
where the $j$-th graph contains $n_j$ vertices and is associated with an adjacency matrix $G^{(j)} \in \mathbb{R}^{n_j \times n_j}$.
We use $i_j$ to index the $i_j$-th vertex of graph $j$,
and use a tuple $(i_1, \dots, i_J)$ to index each multi-relation across the $J$ graphs.
The system predictions over all possible $\prod_{j=1}^J n_j$ multi-relations is summarized in an order-$J$ tensor $f \in \mathbb{R}^{{n_1} \times \dots \times n_J}$,
where $f_{i_1,i_2,\dots,i_J}$ corresponds to the prediction about tuple $(i_1, \dots, i_J)$.
Denote by $\otimes$ the Kronecker (Tensor) product.
We use $\bigotimes_{j=1}^{J} x_j$ (or simply $\bigotimes_j x_j$) as the shorthand for $x_1 \otimes \dots \otimes x_J$.
Denote by $\times_j$ the $j$-mode product between tensors.
We refer the readers to \cite{kolda2009tensor} for a thorough introduction about tensor mode product.
\subsection{Graph Product}
\label{sec:gp}
In a nutshell, graph product (GP)
\footnote{
While traditional GP only applies to two graphs, we generalize it to the case of multiple graphs
(Section \ref{sec:sgp}).
}
is a mapping from each cross-graph multi-relation to each vertex in a new graph $\mathscr{P}$,
whose edges encode similarities among the multi-relations (illustrated in Fig.\ \ref{fig:gp-example}).
A desirable property of GP is
it provides a natural reduction from the original multi-relational learning problem over \emph{heterogeneous} information sources (Task \ref{task:1})
to an equivalent graph-based learning problem over a \emph{homogeneous} graph (Task \ref{task:2}).
\begin{task}
\label{task:1}
Given $J$ graphs $G^{(1)}, \dots, G^{(J)}$ with a small set of labeled multi-relations
$\mathcal{O} = \{(i_1, \dots, i_J)\}$,
predict labels of the unlabeled multi-relations.
\end{task}
\begin{task}
\label{task:2}
Given the product graph $\mathscr{P}\big(G^{(1)}, \dots, G^{(J)}\big)$ with a small set of labeled vertices
$\mathcal{O} = \{(i_1, \dots, i_J)\}$,
predict labels of its unlabeled vertices.
\end{task}
\subsection{Spectral Graph Product}
\label{sec:sgp}
We define a parametric family of GP operators named the spectral graph product (SGP),
which is of particular interest as
it subsumes the well-known Tensor GP and Cartesian GP (Table \ref{tab:example}),
is well behaved (Theorem \ref{thm:commutation})
and allows efficient optimization routines (Section \ref{sec:approximation}).
Let $\lambda^{(j)}_{i_j}$ and $v^{(j)}_{i_j}$ be the $i_j$-th eigenvalue and eigenvector for the graph $j$,
respectively.
We construct SGP by defining the eigensystem of its adjacency matrix
based on the provided $J$ heterogeneous eigensystems of $G^{(1)},\dots, G^{(J)}$.
\begin{definition}
\label{def:sgp}
The SGP of $G^{(1)}, \dots, G^{(J)}$ is a graph consisting of $\prod_j n_j$ vertices,
with its adjacency matrix $\mathscr{P}_\kappa := \mathscr{P}_\kappa\left( G^{(1)}, \dots, G^{(J)} \right)$ defined by the following eigensystem
\begin{align}
\Big\{
\kappa\big(\lambda^{(1)}_{i_1}, \dots, \lambda^{(J)}_{i_J} \big),
\bigotimes_j {v^{(j)}_{i_j}}
\Big\}_{i_1, \dots, i_J}
\end{align}
where $\kappa$ is a pre-specified nonnegative nondecreasing function over $\lambda_{i_j}^{(j)}, \forall j = 1,2,\dots, J$.
\end{definition}
In other words,
the $(i_1,\dots,i_J)$-th eigenvalue of $\mathscr{P}_\kappa$ is defined by
coupling the $\lambda^{(1)}_{i_1}, \dots, \lambda^{(J)}_{i_J}$ with function $\kappa$,
and the $(i_1,\dots,i_J)$-th eigenvector of $\mathscr{P}_\kappa$
is defined by coupling $v^{(1)}_{i_1}, \dots, v^{(J)}_{i_J}$
via tensor (outer) product.
\begin{remark}
If each individual $\big\{v^{(j)}_{i_j}\big\}_{{i_j}=1}^{n_j}$ forms an orthogonal basis
in $\mathbb{R}^{n_j}$, $\forall j \in 1,\dots,J$,
then $\big\{\bigotimes_j {v^{(j)}_{i_j}}\big\}_{i_1, \dots, i_J}$ forms an orthogonal basis in $\mathbb{R}^{\prod_{j=1}^J n_j}$.
\end{remark}
In the following example we introduce two special kinds of SGPs,
assuming $J=2$ for brevity.
Higher-order cases are later summarized in Table \ref{tab:example}.
\begin{example}
Tensor GP defines
$\kappa(\lambda_{i_1}, \lambda_{i_2}) = \lambda_{i_1} \lambda_{i_2}$,
and is equivalent to Kronecker product:
$
\mathscr{P}_{\text{Tensor}}\big(G^{(1)}, G^{(2)}\big)
= \sum_{i_1, i_2} (\lambda_{i_1} \lambda_{i_2}) \big(v^{(1)}_{i_1} \otimes v^{(2)}_{i_2}\big) \big(v^{(1)}_{i_1} \otimes v^{(2)}_{i_2}\big)^\top
\equiv G^{(1)} \otimes G^{(2)}
$.
Cartesian GP defines
$\kappa(\lambda_{i_1}, \lambda_{i_2}) = \lambda_{i_1} + \lambda_{i_2}$,
and is equivalent to the Kronecker sum:
$
\mathscr{P}_{\text{Cartesian}}\big(G^{(1)}, G^{(2)}\big)
= \sum_{i_1, i_2} (\lambda_{i_1} + \lambda_{i_2}) \big(v^{(1)}_{i_1} \otimes v^{(2)}_{i_2}\big) \big(v^{(1)}_{i_1} \otimes v^{(2)}_{i_2}\big)^\top
\equiv G^{(1)} \oplus G^{(2)}
$.
\begin{table}[ht]
\small
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{r|cc}
\toprule
SGP Type & $\kappa\big( \lambda_{i_1}^{(1)}, \cdots, \lambda_{i_J}^{(J)} \big)$ & $\left[\mathscr{P}_\kappa\right]_{(i_1, \cdots i_J), (i'_1, \cdots i'_J)}$ \\
\midrule
Tensor & $\prod_j \lambda_{i_j}^{(j)}$ & $\prod_j G^{(j)}_{i_j, i'_j}$ \\
Cartesian & $\sum_j \lambda_{i_j}^{(j)}$ & $\sum_j G^{(j)}_{i_j, i'_j} \prod_{j'\not=j} \delta_{i_{j'} = i'_{j'}}$ \\
\bottomrule
\end{tabular}
}
\caption{Tensor GP and Cartesian GP in higher-orders.}
\label{tab:example}
\end{table}
\end{example}
While Tensor GP and Cartesian GP provide mechanisms to associate multiple graphs
in a multiplicative/additive manner,
more complex cross-graph association patterns can be modeled by specifying $\kappa$.
E.g.,
$
\kappa\left( \lambda_{i_1}, \lambda_{i_2}, \lambda_{i_3} \right)
= \lambda_{i_1} \lambda_{i_2} + \lambda_{i_2} \lambda_{i_3} + \lambda_{i_3} \lambda_{i_1}
$
indicates pairwise associations are allowed among three graphs,
but no triple-wise association is allowed as term $\lambda_{i_1} \lambda_{i_2} \lambda_{i_3}$ is not involved.
Including higher order polynomials in $\kappa$
amounts to incorporating higher-order associations among the graphs,
which can be achieved by simply exponentiating $\kappa$.
Since what the product graph $\mathscr{P}$ offers is essentially a similarity measure among multi-relations,
shuffling the order of input graphs $G^{(1)}, \dots, G^{(J)}$ should not affect $\mathscr{P}$'s topological structure.
For SGP, this property is guaranteed by the following theorem:
\begin{theorem}[The Commutative Property]
\label{thm:commutation}
SGP is commutative (up to graph isomorphism) if $\kappa$ is commutative.
\end{theorem}
We omit the proof.
The theorem suggests the SGP family is well-behaved as long as $\kappa$ is commutative,
which is true for both Tensor and Cartesian GPs as both multiplication and addition operations are order-insensitive.
\subsection{Optimization Objective}
\label{sec:obj}
It is often more convenient to equivalently write tensor $f$ as a multi-linear map.
E.g., when $J = 2$,
tensor (matrix) $f \in \mathbb{R}^{n_1 \times n_2}$ defines a bilinear map
from $\mathbb{R}^{n_1} \times \mathbb{R}^{n_2}$ to $\mathbb{R}$ via $f(x_1, x_2) := x_1^\top f x_2$
and we have $f_{i_1,i_2} = f(e_{i_1}, e_{i_2})$.
Such equivalence is analogous to high-order cases
where $f$ defines a multi-linear map
from $\mathbb{R}^{n_1} \times \dots \times \mathbb{R}^{n_J}$ to $\mathbb{R}$.
To carry out transductive learning over $\mathscr{P}_\kappa$ (Task \ref{task:2}),
we inject the structure of the product graph into $f$ via a Gaussian random fields prior \cite{zhu2003semi}.
The negative log-likelihood of the prior $-\log p\left( f \mid \mathscr{P}_\kappa \right)$ is the same (up to constant) as the following squared semi-norm
\begin{align}
\label{eq:semi-norm}
\|f\|_{\mathscr{P}_\kappa}^2
&=
vec(f)^\top \mathscr{P}^{-1}_\kappa vec(f) \\
\label{eq:semi-norm-sum}
&=
\sum_{i_1, i_2, \dots, i_J}
\frac{f\big( v^{(1)}_{i_1}, \dots, v^{(J)}_{i_J} \big)^2}{\kappa\big( \lambda_{i_1}^{(1)}, \dots, \lambda_{i_J}^{(J)} \big)}
\end{align}
Our optimization objective is therefore defined as
\begin{equation} \label{eq:obj}
\min_{f \in \mathbb{R}^{n_1 \times \dots \times n_J}} \enskip \ell_{\mathcal{O}}\left(f\right)
+
\frac{\gamma}{2} \|f\|_{\mathscr{P}_\kappa}^2
\end{equation}
where $\ell_{\mathcal{O}}(\cdot)$ is a loss function to be defined later (Section \ref{sec:opt}),
$\mathcal{O}$ is the set of training tuples,
and $\gamma$ is a tuning parameter controlling the strength of graph regularization.
\section{Optimization}
\label{sec:opt}
Let $\left( x \right)_+ = \max\left( 0, 1-x \right)$ be the shorthand for hinge loss.
We define $\ell_\mathcal{O}(f)$ to be the ranking $\ell_2$-hinge loss
\begin{equation} \label{eq:loss}
\ell_\mathcal{O}(f)
=
\frac{\sum_{
\text{
\tiny
$
\begin{array}{c}
(i_1, \dots, i_J) \in \mathcal{O} \\
(i_1', \dots, i_J') \in \bar{\mathcal{O}}
\end{array}
$
}} \Big( f_{i_1 \dots i_J} - f_{i_1' \dots i_J'} \Big)^2_+}{|\mathcal{O}\times\bar{\mathcal{O}}|}
\end{equation}
where $\bar{\mathcal{O}}$ is the complement of $\mathcal{O}$ w.r.t.\ all possible multi-relations.
Eq.\ \eqref{eq:loss} encourages the valid tuples in our training set $\mathcal{O}$ to be ranked higher than those corrupted ones in $\bar{\mathcal{O}}$,
and is known to be a surrogate of AUC.
We use stochastic gradient descent for optimization as $|\mathcal{O}|$ is usually large.
In each iteration,
a random valid multirelation $(i_1, \dots, i_J)$ is uniformly drawn from $\mathcal{O}$,
a random corrupted multirelation $(i'_1, \dots, i'_J)$ is uniformly drawn from $\bar{\mathcal{O}}$.
The associated noisy gradient is computed as
\begin{align}
\nabla_\alpha &= \frac{\partial \ell_\mathcal{O}}{\partial f} \left(
\frac{\partial f_{i_1, \dots, i_J}}{\partial \alpha}
-
\frac{\partial f_{i'_1, \dots, i'_J}}{\partial \alpha}
\right)
+ \gamma \alpha \oslash \kappa \label{eq:sgd-alpha}
\end{align}
where we abuse the notation by defining $\kappa \in \mathbb{R}^{d_1 \times \dots \times d_J}$,
$\kappa_{k_1,\dots,k_J} := \kappa\big(\lambda^{(1)}_{k_1}, \dots, \lambda^{(J)}_{k_J}\big)$;
$\oslash$ is the element-wise division between tensors.
The gradient w.r.t.\ $\alpha$ in \eqref{eq:sgd-alpha} is
\begin{align}
\frac{\partial f_{i_1, \dots, i_J}}{\partial \alpha}
&= \frac{\partial \big(\alpha \times_1 V^{(1)}_{i_1} \times_2 \dots \times_J V^{(J)}_{i_J}\big)}{\partial \alpha} \\
&= \bigotimes_j V^{(j)}_{i_j} \quad \in \mathbb{R}^{d_1 \times \dots d_J} \label{eq:coupled}
\end{align}
Each SGD iteration costs $O\big( \prod_j d_j \big)$ flops,
which is independent from $n_1, n_2, \dots, n_J$.
After obtaining the solution $\hat{\alpha}(\kappa)$ of optimization \eqref{eq:tucker-obj} for any given SGP $\mathscr{P}_\kappa$,
our final predictions in $\hat{f}(\kappa)$ can be recovered via \eqref{eq:tucker}.
Following \cite{duchi2011adaptive}, we allow adaptive step sizes for each element in $\alpha$.
That is, in the $t$-th iteration we use
$
\eta^{(t)}_{k_1, \dots, k_J} = \eta_0 \Big/ \Big[\sum_{\tau=1}^t { {\nabla_\alpha}_{k_1, \dots, k_J}^{(\tau)}}^2\Big]^{\frac{1}{2}}
$
as the step size for $\alpha_{k_1, \dots, k_J}$,
where $\big\{{\nabla_\alpha}_{k_1, \dots, k_J}^{(\tau)}\big\}_{\tau=0}^t$ are historical gradients associated with $\alpha_{k_1, \dots, k_J}$
and $\eta_0$ is the initial step size (set to be $1$).
The strategy is particularly efficient with highly redundant gradients,
which is our case where the gradient is a regularized rank-2 tensor, according to \eqref{eq:sgd-alpha} and \eqref{eq:coupled}.
In practice (especially for large $J$),
the computation cost of tensor operations involving $\bigotimes_{j=1}^J V^{(j)}_{i_j} \in \mathbb{R}^{d_1, \dots, d_J}$
is not ignorable even if $d_1, d_2, \dots, d_J$ are small.
Fortunately, such medium-sized tensor operations in our algorithm are highly parallelable over GPU.
The pseudocode for our optimization algorithm is summarized in Alg.\ \ref{alg:code}.
\begin{algorithm}[t]
\label{alg:code}
\caption{\small{Transductive Learning over Product Graph (TOP)}}
\ForEach{$j \in 1, \dots, J$}
{
$\big\{ v^{(j)}_k, \lambda^{(j)}_k \big\}_{k=1}^{d_j} \gets \textsc{Approx\_Eigen}(G^{(j)})$\;
}
\ForEach{$\left( k_1, \dots, k_J \right) \in [d_1] \times \dots [d_J]$}
{
$\kappa_{k_1, \dots, k_J} \gets \kappa(\lambda^{(1)}_{k_1}, \dots, \lambda^{(J)}_{k_J})$\;
}
$\alpha \gets 0$, $Z \gets 0$\;
\While{not converge}
{
$(i_1, \dots, i_J) \stackrel{uni}{\sim} \mathcal{O}$,\enskip\,$(i'_1, \dots, i'_J) \stackrel{uni}{\sim} \bar{\mathcal{O}}$\;
$f_{i_1,\dots,i_J} \gets \alpha \times_1 V_{i_1}^{(1)} \times_2 \dots \times_J V_{i_J}^{(J)}$\;
$f_{i'_1,\dots,i'_J} \gets \alpha \times_1 V_{i'_1}^{(1)} \times_2 \dots \times_J V_{i'_J}^{(J)}$\;
$\delta = f_{i_1,\dots,i_J} - f_{i'_1,\dots,i'_J}$\;
\If{$\delta < 1$}
{
$\nabla_\alpha \gets 2(\delta-1) \Big(\bigotimes_{j} V^{(j)}_{i_j} - \bigotimes_{j} V^{(j)}_{i'_j}\Big) + \gamma \alpha \oslash \kappa$\;
}
\Else
{
$\nabla_\alpha \gets \gamma \alpha \oslash \kappa$\;
}
$Z \gets Z + \nabla_\alpha^{\odot 2}$\;
$\alpha \gets \alpha - \eta_0 Z^{\odot - \frac{1}{2}} \odot \nabla_\alpha$\;
}
\Return $\alpha$
\end{algorithm}
\section{Convex Approximation}
\label{sec:approximation}
\begin{figure*}
\centering
\includegraphics[width=0.85\linewidth]{figures/basis/basis.png}
\caption{
An illustration of the eigenvectors of $G^{(1)}$, $G^{(2)}$ and $\mathscr{P}\big( G^{(1)}, G^{(2)} \big)$.
We plot leading nontrivial eigenvectors of $G^{(1)}$ and $G^{(2)}$ in blue and red curves, respectively,
and plot the induced leading nontrivial eigenvectors of $\mathscr{P}\big( G^{(1)}, G^{(2)} \big)$ in 3D.
If $G^{(1)}$ and $G^{(2)}$ are symmetrically normalized,
their eigenvectors (corresponding to eigenvectors of the graph Laplacian) will be ordered by smoothness w.r.t.\ the graph structures.
As a result, eigenvectors of $\mathscr{P}\big( G^{(1)}, G^{(2)} \big)$ will also be ordered by smoothness.
}
\label{fig:smooth}
\end{figure*}
The computational bottleneck for optimization \eqref{eq:obj} lies in evaluating $\|f\|_{\mathscr{P}_\kappa}^2$ and its first-order derivative,
due to the extremely large size of $\mathscr{P}_\kappa$.
In section \ref{sec:exact},
we first identify the computation bottleneck of using the exact formulation,
based on which we propose our convex approximation scheme in \ref{sec:tucker}
that reduces the time complexity of evaluating the semi-norm $\|f\|^2_{\mathscr{P}_\kappa}$ from $O\left(\big(\sum_j n_j\big) \big(\prod_j n_j\big) \right)$ to $O\big(\prod_j d_j\big)$,
where $d_j \ll n_j$ for $j = 1,\dots,J$.
\subsection{Complexity of the Exact Formulation} \label{sec:exact}
The brute-force evaluation of $\|f\|^2_{\mathscr{P}_\kappa}$ according to \eqref{eq:semi-norm-sum} costs
$O\big( \big(\prod_j n_j\big)^2 \big)$,
as one has to evaluate $O\big( \prod_j n_j \big)$ terms inside the summation
where each term costs $O\big( \prod_j n_j \big)$.
However, redundancies exist and
the minimum complexity for the exact evaluation is given as follows
\begin{proposition}
The exact evaluation of semi-norm $\|f\|_{\mathscr{P}_\kappa}$ takes $O\big(\big(\sum_j n_j\big) \big(\prod_j n_j\big) \big)$ flops.
\end{proposition}
\begin{proof}
Notice that the collection of all numerators in \eqref{eq:semi-norm-sum}, namely $\big[f\big( v^{(1)}_{i_1}, \dots, v^{(J)}_{i_J} \big)\big]_{i_1, \cdots, i_J}$,
is a tensor in $\mathbb{R}^{n_1 \times \dots \times n_J}$ that can be precomputed via
\begin{equation} \label{eq:multiply}
\big(\big(f \times_1 V^{(1)}\big) \times_2 V^{(2)}\big) \dots \times_J V^{(J)}
\end{equation}
where $\times_j$ stands for the $j$-mode product between a tensor in $\mathbb{R}^{n_1 \times \dots \times n_j \times \dots \times n_J}$
and $V^{(j)} \in \mathbb{R}^{n_j \times n_j}$.
The conclusion follows as the $j$-th mode product in \eqref{eq:multiply} takes $O\big( n_j \prod_j n_j \big)$ flops,
and one has to do this for each $j = 1,\dots,J$.
When $J=2$,
\eqref{eq:multiply} reduces to the multiplication of three matrices ${V^{(1)}}^\top f V^{(2)}$ at the complexity of $O\left( (n_1 + n_2)n_1n_2 \right)$.
\end{proof}
\subsection{Approximation via Tucker Form}
\label{sec:tucker}
Equation \eqref{eq:multiply} implies the key for complexity reduction is to reduce the cost of the $j$-mode multiplications $\cdot \times_j V^{(j)}$.
Such multiplication costs $O\big( n_j \prod_j n_j \big)$ in general, but can be carried out more efficiently if $f$ is structured.
Our solution is twofold:
First, we include only the top-$d_j$ eigenvectors in $V^{(j)}$ for each graph $G^{(i)}$,
where $d_j \ll n_j$.
Hence each $V^{(j)}$ becomes a thin matrix in $\mathbb{R}^{n_j \times d_j}$.
Second,
we restrict tensor $f$ to be within the linear span of the top $\prod_{j=1}^J d_j$ eigenvectors of the product graph $\mathscr{P}_\kappa$
\begin{align}
f &= \sum_{k_1, \cdots, k_J = 1}^{d_1, \cdots, d_J} \alpha_{k_1, \cdots, k_J} \bigotimes_{j} v^{(j)}_{k_j} \label{eq:tucker} \\
&= \alpha \times_1 V^{(1)} \times_2 V^{(2)} \times_3 \dots \times_J V^{(J)}
\end{align}
The combination coefficients $\alpha \in \mathbb{R}^{d_1 \times \cdots \times d_J}$ is
known as the core tensor of Tucker decomposition.
In the case where $J=2$,
the above is equivalent to saying $f \in \mathbb{R}^{n_1 \times n_2}$
is a low-rank matrix parametrized by $\alpha \in \mathbb{R}^{d_1 \times d_2}$
such that $f = \sum_{k_1, k_2} \alpha_{k_1, k_2} v^{(1)}_{k_1} {v^{(2)}_{k_2}}^\top = V^{(1)} \alpha {V^{(2)}}^\top$.
Combining \eqref{eq:tucker} with the orthogonality property of eigenvectors
leads to the fact that $f\big( v_{k_1}^{(1)}, \dots, v_{k_J}^{(J)} \big) = \alpha_{k_1, \cdots, k_J}$.
To see this for $J=2$, notice
$f\big( v_{k_1}^{(1)}, v_{k_2}^{(2)} \big) = {v_{k_1}^{(1)}}^\top f v_{k_1}^{(2)} = {v_{k_1}^{(1)}}^\top V^{(1)} \alpha {V^{(2)}}^\top v_{k_1}^{(2)} = e_{k_1}^\top \alpha e_{k_2} = \alpha_{k_1,k_2}$.
Therefore the semi-norm in \eqref{eq:semi-norm} can be simplified as
\begin{equation}
\label{eq:reg}
\|f\|_{\mathscr{P}_\kappa}^2
=
\|\alpha\|_{\mathscr{P}_\kappa}^2
=
\sum_{k_1, \dots, k_J=1}^{d_1, \cdots, d_J}
\frac{\alpha^2_{k_1, \cdots, k_J}}{\kappa\big( \lambda_{k_1}^{(1)}, \dots, \lambda_{k_J}^{(J)} \big)}
\end{equation}
Comparing \eqref{eq:reg} with \eqref{eq:semi-norm-sum},
the number of inside-summation terms is reduced from $O\big( \prod_j n_j \big)$ to $O\big( \prod_j d_j \big)$ where $d_j \ll n_j$.
In addition, the cost for evaluating each term inside summation is reduced from $O\big( \prod_j n_j \big)$ to $O(1)$.
Denote by $V^{(j)}_{i_j} \in \mathbb{R}^{d_j}$ the $i_j$-th row of $V^{(j)}$,
we obtain the following optimization by replacing $f$ with $\alpha$ in \eqref{eq:obj}
\begin{equation} \label{eq:tucker-obj}
\begin{aligned}
\min_{\alpha \in \mathbb{R}^{d_1 \times \dots \times d_J}}
&\ell_\mathcal{O} \left(f\right)
+ \frac{\gamma}{2} \|\alpha\|_{\mathscr{P}_\kappa}^2 \\
\text{s.t.}\quad\enskip\ &f = \alpha \times_1 V^{(1)} \times_2 \dots \times_J V^{(J)}
\end{aligned}
\end{equation}
Optimization above has intuitive interpretations.
In principle,
it is natural to emphasis bases in $f$ that are ``smooth'' w.r.t.\ the manifold structure of $\mathscr{P}_\kappa$,
and de-emphasis those that are ``nonsmooth'' in order to obtain a parsimonious hypothesis with strong generalization ability.
We claim this is exactly the role of regularizer \eqref{eq:reg}.
To see this,
note any nonsmooth basis $\bigotimes_j v_{k_j}^{(j)}$ of $\mathscr{P}_\kappa$ is likely to be associated with small a eigenvalue $\kappa\big( \lambda^{(1)}_{k_1},\dots, \lambda^{(J)}_{k_J}\big)$
(illustrated in Fig.\ \ref{fig:smooth}).
The conclusion follows by noticing that
$\alpha_{k_1,\dots,k_J}$ is essentially the activation strength of $\bigotimes_j v_{k_j}^{(j)}$ in $f$ (implied by \eqref{eq:tucker}),
and that \eqref{eq:reg} is going to give any $\alpha_{k_1,\dots,k_J}$ associated with a small $\kappa\big( \lambda^{(1)}_{k_1},\dots, \lambda^{(J)}_{k_J}\big)$ a stronger penalty.
\eqref{eq:tucker-obj} is a convex optimization problem over $\alpha$ with any convex $\ell_\mathcal{O}(\cdot)$.
Spectral approximation techniques for graph-based learning has been found successful in standard classification tasks \cite{fergus2009semi},
which are special cases under our framework when $J=1$.
We introduce this technique for multi-relational learning,
which is particularly desirable as the complexity reduction will be much more significant for high-order cases $(J >= 2)$.
While $f$ in \eqref{eq:tucker} is assumed to be in the Tucker form,
other low-rank tensor representation schemes are potentially applicable.
E.g.,
the Candecomp/Parafac (CP) form that further restricts $\alpha$ to be diagonal,
which is more aggressive but substantially less expressive.
The Tensor-Train decomposition \cite{oseledets2011tensor}
offers an alternative representation scheme in the middle of Tucker and CP,
but the resulting optimization problem will suffer from non-convexity.
\section{Introduction}
\label{sec:intro}
Many important problems in multi-source relational learning could be cast as joint learning over multiple graphs about how heterogeneous types of objects interact with each other. In literature data analysis, for example, publication records provide rich information about how authors collaborate with each other in a co-authoring graph, how papers are linked in citation networks, how keywords are related via ontology, and so on. The challenging question is about how to combine such heterogeneous information in individual graphs for the labeling or scoring of the multi-relational associations in tuples like \texttt{(author,paper,keyword)}, given some observed instances of such tuples as the labeled training set. Automated labeling or scoring of unobserved tuples allows us to discover who have been active in the literature on what areas of research, and to predict who would become influential in which areas in the future.
In protein data analysis, as another example, a graph of proteins with pairwise sequence similarities is often jointly studied with a graph of chemical compounds with their structural similarities for the discovery of interesting patterns in \texttt{(compound,protein)} pairs. We call the prediction problem in both examples \textit{cross-graph learning of multi-relational associations}, or simply \textit{cross-graph relational learning} (CGRL), where the multi-relational associations are defined by the tuples of heterogeneous types of objects, and each object type has its own graph with type-specific relational structure as a part of the provided data. The task is to predict the labels or the scores of unobserved multi-relational tuples, conditioned on a relatively small set of labeled instances.
CGRL is an open challenge in machine learning for several reasons. Firstly, the number of multi-relational tuples grows combinatorially in the numbers of individual graphs and the number of nodes in each graph.
How to make cross-graph inference computationally tractable for large graphs is a tough challenge. Secondly, how to combine the internal structures or relations in individual graphs for joint inference in a principled manner is an open question. Thirdly, supervised information (labeled instances) is typically extremely sparse in CGRL due to the very large number of all possible combinations of heterogeneous objects in individual graphs. Consequently, the success of cross-graph learning crucially depends on effectively leveraging the massively available unlabeled tuples (and the latent relations among them) in addition to the labeled training data. In other words, how to make the learning transductive is crucial for the true success of CGRL. Research on transdcutive CGRL has been quite limited, to our knowledge.
Existing approaches in CGRL or CGRL-related areas can be outlined as those using tensors or graph-regularized tensors, and kernel machines that combine multiple kernels.
Tensor methods have been commonly used for combining multi-source evidence of the interactions among multiple types of objects \cite{nickel2011three, rendle2009learning, kolda2009tensor} as the combined evidence can be naturally represented as tuples. However, most of the tensor methods do not explicitly model the internal graph structure for each type of objects, although some of those methods implicitly leverage such information via graph-based regularization terms in their objective function that encourage similar objects within each graph to share similar latent factors
\cite{narita2012tensor, cai2011graph}.
A major weakness in such tensor methods is the lack of convexity in their models, which leads to ill-posed optimization problems particularly in high-order scenarios. It has also been observed that tensor factorization models suffer from label-sparsity issue, which is typically severe in CGRL.
Kernel machines have been widely studied for supervised classifiers,
where a kernel matrix corresponds to a similarity graph among a single type of objects.
Multiple kernels can be combined,
for example,
by taking the tensor product of each individual kernel matrix,
which results in a desired kernel matrix among cross-graph multi-relational tuples.
The idea has been explored in relational learning combined with SVMs \cite{ben2005kernel}, perceptions \cite{basilico2004unifying} or Gaussian process \cite{yu2008gaussian} for two types of objects and is generalizable to the multi-type scenario of CGRL.
Although being generic,
the complexity of such kernel-based methods grows exponentially in the number of individual kernels (graphs) and the size of each individual graph. As a result, kernel machines suffer from poor scalability in general.
In addition, kernel machines are purely supervised (not for transductive learning), i.e., they cannot leverage the massive number of available non-observed tuples induced from individual graphs and the latent connections among them. Those limitations make existing kernel methods less powerful for solving the CGRL problem in large scale and under severely data-sparse conditions.
In this paper, we propose a novel framework for CGRL which can be characterized as follows:
(i) It uses graph products to map heterogeneous sources of information and the link structures in individual graphs onto a single \emph{homogeneous} graph;
(ii) It provides a convex formulation and approximation of the CGRL problem that ensure robust optimization and efficient computation; and
(iii) It enables transductive learning in the form of label propagation over the induced homogeneous graph so that the massively available non-observed tuples and the latent connections among them can play an important role in effectively addressing the label-sparsity issue.
The proposed framework is most related to \cite{liu2015bipartite},
where the authors formulated graph products for learning the edges of a bipartite graph.
Our new framework is fundamentally different in two aspects.
First, our new formulation and algorithms allow the number of individual graphs to be greater than two,
while method in \cite{liu2015bipartite} is only applicable to two graphs.
Secondly, the
algorithms in \cite{liu2015bipartite} suffer from cubic complexity over the graphs sizes
(quadratic by using a non-convex approximation),
while our new algorithm enjoys both the convexity of the formulation and the low time complexity which is linear over the graph sizes.
The paper is organized as follows:
Section \ref{sec:proposed} shows how cross-graph multi-relations can be embedded into the vertex space of a homogeneous graph.
Section \ref{sec:approximation} describes how efficient label propagation among multi-relations can be carried out in such space with approximation.
We discuss our optimization algorithm in Section \ref{sec:opt}
and provide empirical evaluations over real-world datasets in Section \ref{sec:exp}.
\section{Adaptive Nonparametric SGP}
\label{sec:adaptive-sgp}
Let $\varphi\left(\alpha; \kappa\right)$ be the objection function of \eqref{eq:tucker-obj},
and denote by $\hat{\varphi}(\kappa) = \min_\alpha \ \varphi\left( \alpha; \kappa \right)$ and $\hat{\alpha}(\kappa) = \mathrm{argmin}_\alpha \ \varphi\left( \alpha; \kappa \right)$.
Denote by $\hat{\alpha}(\kappa)$ the solution of \eqref{eq:tucker-obj} for any fixed $\kappa$.
The derivative of \eqref{eq:tucker-obj}'s optimal criterion-function value w.r.t.\ $\kappa$ can be obtained via Danskin's theorem
\begin{equation}
\nabla_\kappa = - \frac{\gamma}{2} \hat{\alpha}(\kappa) \odot \kappa^{- \odot 2}
\end{equation}
We require $\kappa$ to satisfy the following affine constraints during adaptation
\begin{enumerate}[(a)]
\item Monotonicity: $\kappa_{\dots, k_{j-1}, k_j, k_{j+1}, \dots} \ge \kappa_{\dots, k_{j-1}, (k_j+1), k_{j+1}, \dots}\quad \forall j, k_j$
\item ``Simplexity''\footnote{The sum of $\kappa_{k_1, \dots, k_J}$ does not have to be one. Any constant will suffice.} : $\sum_{k_1, \dots, k_J = 1}^{d_1, \dots, d_J} \kappa_{k_1, \dots, k_J} = 1 \enskip \text{and} \enskip \kappa \ge 0$
\end{enumerate}
Projection (b) can be solved by \cite{duchi2008efficient}.
Projection (a) can be recast as an isotonic regression problem with partial-order constraints,
and be solved by alternating PAVA w.r.t.\ difference dimensions \cite{paul2004quantile}.
\section{Experiments}
\label{sec:exp}
\subsection{Datasets}
We evaluate our method on real-world data in two different domains:
the Enzyme dataset \cite{yamanishi2008prediction} for compound-protein interaction
and the DBLP dataset of scientific publication records.
Fig.\ \ref{fig:schema} illustrates their heterogeneous objects and relational structures.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.25\linewidth]{figures/drug-target.pdf}
\hspace{4em}
\includegraphics[width=0.32\linewidth]{figures/dblp-schema.pdf}
\caption{The heterogeneous types of objects (the circles) and the relational structures in the Enzyme (left) and DBLP (right) data sets.
The blue edges represent the within-graph relations and the red edges represent the cross-graph interactions. The corresponding tuples in Enzyme is in the form of \texttt{(Compound,Protein)},
and in DBLP is in the form of \texttt{(Author,Paper,Venue)}.}
\label{fig:schema}
\end{figure*}
The Enzyme dataset
has been used for modeling and predicting drug-target interactions,
which contains a graph of 445 chemical compounds (drugs) and a graph of 664 proteins (targets).
The prediction task is to label the unknown compound-protein interactions based on both the graph structures and a small set of 2,926 known interactions.
The graph of compounds is constructed based on the SIMCOMP score \cite{hattori2003heuristics},
and the graph of proteins is constructed based on the normalized SmithWaterman score \cite{smith1981identification}.
While both graphs are provided in the dense form,
we converted them into sparse $k$NN graphs where each vertex is connected with its top 1\% neighbors.
As for the DBLP dataset,
we use a subset of 34,340 DBLP publication records in the domain of Artificial Intelligence \cite{tang2008arnetminer}, from which 3 graphs are constructed as:
\begin{itemize}
\item For the author graph ($G^{(1)}$) we draw an edge between two authors if they have coauthored an overlapping set of papers, and remove the isolated authors using a DFS algorithm. We then obtain a symmetric $k$NN graph by connecting each author with her top $0.5\%$ nearest neighbors using the count of co-authored papers as the proximity measure. The resulting graph has 5,517 vertices with 17 links per vertex on average.
\item For the paper graph ($G^{(2)}$) we connect two papers if both of them cite another paper, or are cited by another paper.
Like $G^{(1)}$, we remove isolated papers using DFS and construct a symmetric 0.5\%-NN graph.
To measure the similarity of any given pair of papers,
we represent each paper as a bag-of-citations and compute their cosine similarity.
The resulted graph has 11,879 vertices and has an average degree of 50.
\item For the venue graph ($G^{(3)}$) we connect two venues if they share similar research focus.
The venue-venue similarity is measured by the total number of cross-citations in between,
normalized by the size of the two venues involved.
The symmetric venue graph has 22 vertices and an average degree of 7.
\end{itemize}
Tuples in the form of \texttt{(Author,Paper,Venue)} are extracted from the publication records,
and there are 15,514 tuples (cross-graph interactions) after preprocessing.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{figures/results/both-within.pdf}
\caption{Performance of TOP with different SGPs.}
\label{fig:TOP on Enzyme and DBLP}
\end{figure}
\subsection{Methods for Comparison}
\begin{itemize}
\item Transductive Learning over Product Graph (\textbf{TOP}). \\
The proposed method.
We explore the following
$\kappa$'s for parametrizing the spectral graph product. \\
{
\small
\\
\centering
\begin{tabular}{cll}
Name & $\kappa(x,y)$ $(J=2)$ & $\kappa(x,y,z)$ $(J=3)$ \\
\midrule
Tensor & $xy$ & $xyz$ \\
Cartesian & $x+y$ & $x+y+z$ \\
Exponential & $e^{x+y}$ & $e^{xy + yz + xz}$ \\
Flat & $1$ & 1 \\
\end{tabular}
}
\item Tensor Factorization (\textbf{TF}) and Graph-regularized TF (\textbf{GRTF}).
In TF we factorize $f \in \mathbb{R}^{n_1 \times \dots \times n_J}$
as a set of dimensionality-reduced latent factors
$C^{d_1, \times \dots \times d_J}$, $U_1^{n_1 \times d_1}, \dots, U_J \in \mathbb{R}^{n_J \times d_J}$.
In GRTF,
we further enhanced the traditional TF by adding graph regularizations to the objective function,
which enforce the model to be aware of the context information in $G^{(j)}$'s
\cite{narita2012tensor, cai2011graph};
\item One-class Nearest Neighbor (\textbf{NN}).
We score each tuple $\left(i_1, \dots, i_J\right)$ in the test set with
$
\hat{f}{(i_1, \dots, i_J)} = \mathrm{max}_{\left( i_1', \dots, i_J' \right) \in \mathcal{O}} \ \prod_{j=1}^J G_{i_j i'_j}
$.
That is,
we assume the tuple-tuple similarity can be factorized as
the product of vertex-level similarities across different graphs.
We experimented with several other similarity measures and empirically found the multiplicative similarity leads to the best overall performance.
Note it does not rely on the presence of any negative examples.
\item Ranking Support Vector Machines \cite{joachims2002optimizing} (\textbf{RSVM}).
For the task of completing the missing paper in \texttt{(Author,?,Venue)},
we use a Learning-to-Rank strategy by treating \texttt{(Author,Venue)} as the query
and \texttt{Paper} as the document to be retrieved.
The query feature is constructed by concatenating the eigen-features of $\texttt{Author}$ and $\texttt{Venue}$,
where we define the eigen-feature of vertex $i_j$ in graph $j$ as $V^{(j)}_{i_j} \in \mathbb{R}^{d_j}$.
The feature for each query-document pair is obtained by taking the tensor product of the query feature and document eigen-feature.
\item Low-rank Tensor Kernel Machines (\textbf{LTKM}).
While traditional tensor-based kernel construction methods for tuples suffer from poor scalability.
We propose to speedup by replacing each individual kernel with its low-rank approximation before tensor product,
leading to a low-rank kernel of tuples which allows more efficient optimization routines.
\end{itemize}
\begin{figure*}[ht]
\centering
\includegraphics[width=0.825\linewidth]{figures/results/enzyme.pdf}
\caption{Test-set performance of different methods on Enzyme.}
\label{fig:Methods on Enzyme}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[width=0.825\linewidth]{figures/results/dblp.pdf}
\caption{Test-set performance of different methods on DBLP.}
\label{fig:Methods on DBLP}
\end{figure*}
For fair comparison,
loss functions for TF, GRTF, RSVM and LTKM
are set to be exactly the same as that for TOP, i.e. E.q.\ \eqref{eq:loss}.
All algorithms are trained using a mini-batched stochastic gradient descent.
We use the same eigensystems (eigenvectors and eigenvalues) of the $G^{(j)}$'s
as the input for TOP, RSVM and LTKM.
The number of top-eigenvalues/eigenvectors $d_j$ for graph $j$ is chosen
such that $\lambda^{(j)}_1, \dots ,\lambda^{(j)}_{d_j}$
approximately cover $80\%$ of the total spectral energy of $G^{(j)}$.
With respect to this criterion,
we choose $d_1=1,281$, $d_2 = 2,170$, $d_3 = 6$ for DBLP,
and $d_1=150$, $d_2=159$ for Enzyme.
\subsection{Experiment Setups}
For both datasets, we randomly sample one third of known interactions
for training (denoted by $\mathcal{O}$), one third for validation and use the remaining ones for testing.
Known interactions in the test set,
denoted by $\mathcal{T}$, are treated as positive examples.
All tuples not in $\mathcal{T}$,
denoted by $\bar{\mathcal{T}}$, are treated as negative.
Tuples that are already in $\mathcal{O}$ are removed from $\bar{\mathcal{T}}$ to avoid misleading results \cite{bordes2013translating}.
We measure algorithm performance on Enzyme
based on the quality of inferred target proteins given each compound,
namely by the ability of completing $\texttt{(Compound,?)}$.
For DBLP,
the performance is measured by the quality of inferred papers given author and venue,
namely by the ability of completing $\texttt{(Author,?,Venue)}$.
We use Mean Average Prevision (MAP),
Area Under the Curve (AUC) and Hits at Top 5 (Hits@5) as our evaluation metrics.
\subsection{Results}
Fig.\ \ref{fig:TOP on Enzyme and DBLP} compares the results of TOP
with various parameterizations of the spectral graph product (SGP).
Among those, Exponential $\kappa$ works better on average.
Figs.\ \ref{fig:Methods on Enzyme} and \ref{fig:Methods on DBLP} show the main results,
comparing TOP (with Exponential $\kappa$) with other representative baselines.
Clearly, TOP outperforms all the other methods on both datasets in all the evaluation metrics of MAP \footnote{MAP scores for random guessing are 0.014 on Enzyme and 0.00072 on DBLP, respectively.}, AUC and Hit@5.
\begin{figure}[h]
\centering
\includegraphics[width=0.75\linewidth]{figures/results/scale.pdf}
\caption{Performance of TOP v.s.\ model size on Enzyme.}
\label{fig:eigen k}
\end{figure}
Fig.\ \ref{fig:eigen k} shows the performance curves of TOP on Enzyme over different model sizes (by varying the $d_j$'s).
With a relatively small model size compared with using the full spectrum, TOP's performance converges to the optimal point.
|
\subsection{Stochastic dynamic prediction}
\label{sec_intro_stochdyn}
Since the works of
\citet{Epstein} and
\citet{Tatarsky},
we know that accounting for the uncertainty in the initial forecast fields
can improve weather (and other geophysical) predictions.
Assigning a probability distribution for the truth at the start of the forecast
(instead of using deterministic initial data) and attempting to advance this distribution in time
according to the dynamic (forecast) model is called stochastic dynamic prediction.
The advantage of the stochastic dynamic prediction paradigm is twofold.
First, the resulting forecast probability distribution provides a valuable measure of the {\em uncertainty}
in the prediction, leading to probabilistic forecasting
and flow-dependent background-error statistics in data assimilation.
Second, for a nonlinear physical model, switching from the deterministic forecast
to the mean of the forecast probability distribution
improves the mean-square accuracy of the prediction, i.e.\ can improve the deterministic forecasting.
\subsection{Model errors}
\label{sec_intro_mem}
Since \citet[][]{Pitcher}, we realize that not only uncertainties in the initial data
(analysis errors) matter, forecast model (including boundary conditions) imperfections
also play an important role.
Simulation of model errors is the subject of this study, so we define them now.
Let the forecast model be of the form
\begin{linenomath*}
\begin {equation}
\label {m}
\frac {\d {\bf x}}{\d t} = {\bf F}({\bf x}),
\end {equation}
\end{linenomath*}
where $t$ is time, ${\bf x}$ is the vector that represents the (discretized) state of the system,
and ${\bf F}$ is the model (forecast) operator.
The imperfection of the model Eq.(\ref{m}) means that the (appropriately discretized)
truth does {\em not} exactly satisfy this equation. The discrepancy is called the model error \citep[e.g.][]{Orrell}:
\begin{linenomath*}
\begin {equation}
\label {me}
\boldsymbol\xi^t = {\bf F}({\bf x}^t) - \frac {\d {\bf x}^t}{\d t}.
\end {equation}
\end{linenomath*}
The true model error $\boldsymbol\xi^t$ is normally unknown.
In order to include model errors in the stochastic dynamic prediction paradigm,
one {\em models} $\boldsymbol\xi^t(t)$ as a {\em random process}, $\boldsymbol\xi(t)$, or,
in other words, as a spatio-temporal random field $\xi(t, {\bf s})$
(where ${\bf s}$ is the spatial vector).
The probability distribution of $\boldsymbol\xi(t)$
(in most cases, dependent on the flow)
is assumed to be known.
Rearranging the terms in Eq.(\ref{me}),
and replacing the unknown $\boldsymbol\xi^t$ with its stochastic counterpart
$\boldsymbol\xi$,
we realize that the resulting model of
truth is the stochastic dynamic equation
\begin{linenomath*}
\begin {equation}
\label {msd}
\frac {\d {\bf x}}{\d t}= {\bf F}({\bf x}) - \boldsymbol\xi.
\end {equation}
\end{linenomath*}
Thus, the extended stochastic dynamic prediction (or modeling) paradigm requires two input probability
distributions (that of initial errors and that of model errors) and aims to
transform them to the output (forecast) probability distribution.
\subsection{Ensemble prediction}
\label{sec_intro_ensm}
Stochastic dynamic modeling of complex geophysical systems
is hampered by their high dimensionality and non-linearity.
For realistic models, the output probability distribution is analytically
intractable. An affordable approximate solution is provided
by the Monte-Carlo method called in geosciences {\em ensemble prediction}.
In ensemble prediction, the input uncertainties (i.e.\ initial and model errors)
are represented by simulated {\em pseudo-random draws}
from the respective probability distributions.
A relatively small affordable number of these draws are fed to the forecast model
giving rise to an {\em ensemble} of predictions (forecasts).
{\em If initial and model errors are sampled from the correct respective distributions,
then the resulting forecast ensemble members are draws from the correct probability distribution
of the truth given all available external data (initial and boundary conditions).}
This mathematically justifies the ensemble prediction principle.
From the practical perspective, members of the forecast ensemble can be interpreted as ``potential
truths'' consistent with all available information.
In what follows, we concentrate on the model error field $\xi(t, {\bf s})$.
We briefly review existing models for $\xi(t, {\bf s})$ and then present our
stochastic pattern generator, whose goal is to simulate pseudo-random draws of $\xi(t, {\bf s})$
from a meaningful and flexible distribution.
\subsection{Practical model error modeling}
\label{sec_intro_pmem}
In meteorology, our knowledge of the actual model error probability distribution is scarce.
Justified stochastic model-error models are still to be devised and verified.
In the authors' opinion, the best way to stochastically represent spatio-temporal
forecast-model-error fields is to treat each error source separately,
so that, say, each physical parametrization is accompanied with a spatio-temporal
stochastic model of its uncertainty. Or, even better, to completely switch
from deterministic physical parameterizations to stochastic ones.
There is a growing number of such developments \citep[see][for a review]{Berner}, but the problem is
so complex that we cannot expect it to be solved in the near future.
Its solution is further hampered by the fact that the existing
meteorological observations are too scarce and too inaccurate
for model errors to be objectively identified by comparison with measurement data
with satisfactory accuracy \citep{TsyrulnikovGorin}.
As a result, in meteorology, {\em ad-hoc} model-error models are in wide use. The existing
approaches can be classified as either non-stochastic or stochastic.
Non-stochastic schemes can be multi-model (different ensemble members are generated using
different forecast models) or multi-parameterization (each ensemble member is generated using
the forecast model with a unique combination of different physical parameterization schemes or their
parameters).
These techniques are capable of introducing significant diversity in the ensemble \citep{Berner2011},
but the resulting ensemble members cannot be considered as independent and drawn from the
same probability distribution (an assumption normally made in using the ensembles).
Besides, there are not enough different models and not enough substantially different
physical parameterizations to generate large ensembles.
Finally, running many forecast models is a technologically very demanding task.
Stochastic approaches, on the contrary, offer the opportunity to generate as many
ensemble members taken from the same probability distribution as needed, while working with just one
forecast model and one set of physical parameterizations.
In atmospheric ensemble prediction and ensemble data assimilation,
the most widely used stochastic techniques are
SPPT (Stochastic Perturbations of Physical Tendencies, \citet[][]{Buizza}),
SKEB (Stochastic Kinetic Energy Backscatter scheme, \citet{Shutts}), and
SPP (Stochastically Perturbed Parameterizations, \citet{Christensen,Ollinaho}).
In the SPPT, multiplicative perturbations to the {\em tendencies} produced by the model's
physical parameterizations are introduced. The multiplier is a spatio-temporal random field
centered at 1.
In the SKEB, {\em additive} perturbations are computed by modulating a spatio-temporal random
field by the local kinetic energy dissipation rate.
In the SPP, selected {\em parameters} of the physical parameterization schemes are perturbed
again using a spatio-temporal field, which thus is seen to be needed in
all of the above stochastic model error representation schemes.
Stochastic parameterization schemes can also demand such fields \citep[e.g.][]{Bengtsson}.
\subsection{Generation of spatio-temporal random fields}
\label{sec_intro_gen}
The simplest non-constant pseudo-random field is the white noise, i.e.\
the uncorrelated in space and time random field. The white noise is the
default forcing in stochastic differential equations, e.g.\ \citet{Jazwinski} or \citet{Arnold}.
Its advantage is the complete absence of any spatio-temporal structure, it is a pristine
source of stochasticity. But in model-error modeling, this lack of structure
precludes its direct use as an additive or multiplicative perturbation field
because model errors are related to the weather pattern and so should be correlated
(dependent) both in space and time.
\citet{Tsyrulnikov2005} showed in a simulation study that model errors can exhibit complicated
spatio-temporal behavior.
A correlated pseudo-random spatio-temporal field can be easily computed
by generating independent random numbers at points of a {\em coarse} spatio-temporal grid
and then assigning each of them to all model grid points within the
respective coarse-grid cell \citep{Buizza}.
As a result, the model-grid field
becomes correlated in space and time. The decorrelation space and time scales are, obviously,
defined by the respective coarse grid spacings
(e.g.\ in \citet{Buizza} these were about 1000 km in space and 6 h in time).
This technique is extremely simple but it suffers from two flaws.
First, the resulting model-grid field appears to be discontinuous and inhomogeneous.
Second, the spatio-temporal structure of the
field is not scale dependent, that is, the resulting temporal
length scales do not depend on the respective spatial scales.
In reality, longer spatial scales ``live longer'' than shorter spatial scales,
which ``die out'' quicker. This `proportionality of scales' is widespread
in geophysical fields \citep[see][and references therein]{Tsyroulnikov2001}
and other media, \citep[e.g.][p.129]{Meunier},
so we believe this property should be represented by model-error models.
Note also that the ``proportionality of scales'' is a special case of
the {\em non-separability} of
spatio-temporal covariances. For a critique of simplistic separable space-time covariance models,
see \citet{CressieHuang}, \citet{Stein2005}, \citet{Gneiting}, and section \ref{sec_req} below.
Another popular space-time pseudo-random field generation technique
employs a spectral transform in space and then imposes independent
temporal auto-regressions for the coefficients of the spectral expansion
\citep{Berner2009,Palmer,Charron,Bouttier2012}.
This technique is more general and produces homogeneous fields, but the above
implementations use {\em the same time scale} for all spatial wavenumbers
so that there are still no space-time interactions in the generated spatio-temporal fields
(though \citet{Charron} noted that the decorrelation time scales can be made dependent on the
spatial scales and \citet{Palmer} allowed for this dependence in their SKEB pattern generator equations).
In this report, we propose and test a spatio-temporal
Stochastic (pseudo-random) Pattern Generator (SPG)
that accounts for the above ``proportionality of scales'' and
imposes meaningful space-time interactions.
The SPG operates on a limited-area domain.
It is based on a (spectral-space) solution to a stochastic partial differential equation,
more precisely, to a stochastic differential equation in time with a pseudo-differential spatial operator.
In what follows, we present the technique,
examine properties of the resulting spatio-temporal fields on 2D and 3D spatial domains,
describe the numerical scheme, and
explore the performance of the SPG.
The technique is implemented as a Fortran program freely available from
\url{https://github.com/gayfulin/SPG}.
\section {Model error fields: separability vs. ``proportionality of scales''}
\label{sec_req}
In this motivational section we show on a simple 1D (in space) example that
space-time interactions in the model error random field play a significant role.
Specifically, we demonstrate that these interactions determine whether the
spatial length scale of the resulting forecast error field grows, in a first
approximation, in time or remains constant.
We note that for small enough model error perturbations and small enough lead times,
the forecast error due to the accumulated model errors can be approximated by the so-called
model-error drift, that is, the time integrated model error:
$\bar\xi(t,s) =\int_0^t \xi(t,s) \,\d t$ \citep[][]{Orrell}.
Therefore, the methodology in this section is to take two fields,
one with separable spatio-temporal correlations and the other
with``proportional scales'', integrate them in time, and look at
the spatial length scales of the two time integrated random fields.
Theoretically, the time integration reduces (filters out) small-scale-in-time
components of the field. As a separable field has no space-time interactions, its time integral
should have exactly the same spatial length scale as $\xi(t,s)$.
For a proportional-scales field, smaller scales in time are associated with
smaller scales in space, so the amount of small spatial scales in the time integrated
field should decrease in time leading to an increase in the spatial length scale.
To verify these theoretical conclusions, we set up the following numerical experiment.
We considered a 1D domain of size 100 km and the time integration period of 3 h.
In this 2D spatio-temporal domain, we introduced a grid
with 100 points in space and 100 points in time.
On this grid, we simulated two random fields, both with unit variance and exactly the same
spatial and temporal exponential correlations.
The first field had separable correlations
$C_1(\Delta t,\Delta s)=\exp(-|\Delta s|/L) \cdot \exp(-|\Delta t|/T)$,
whereas the second field had non-separable correlations
$C_2(\Delta t,\Delta s)=\exp(-\sqrt{(\Delta s/L)^2 + (\Delta t/T)^2}$,
which can be shown to satisfy the ``proportionality of scales'' property.
The spatial length scale $L$ was selected in such a way that the spatial correlation
function intersects the 0.7 level at the distance of 50 km.
The temporal length scale was selected to be equal to $L/U$, where $U=20$ m/s was
taken as the characteristic flow velocity.
Note that both the separability and the exponential temporal correlation function are what the
scale-independent first-order auto-regressions
used in \citet{Berner2009}, \citet{Palmer}, \citet{Charron}, and \citet{Bouttier2012} imply.
Knowing the two correlation functions, we simulated pseudo-random realizations of the
two fields (by building the two covariance matrices, computing their square roots, and
applying the latter to vectors of independent $N(0,1)$ random variables), see Fig.\ref{Fig_ME}.
\begin{figure
\begin{center}
{ \scalebox{0.25}{ \includegraphics{ME_Sep100x100.pdf}}
\scalebox{0.25}{ \includegraphics{ME_PS100x100.pdf}}
}
\end{center}
\caption{Simulated spatio-temporal fields.
{\em Left}:: With separable space-time correlations.
{\em Right}: With non-separable proportional-scales correlations.
}
\label{Fig_ME}
\end{figure}
Comparing the two panels of Fig.\ref{Fig_ME}, one can see that the two fields look quite differently.
Visually, the most striking difference is the lack of isotropy in the separable case.
The proportional-scales field looks much more realistic than the separable one.
To get a more objective criterion, we computed the time integrated model error field $\bar\xi(t,s)$
(the model error drift, a proxy to the model-error induced forecast error,
see above in this section).
Figure \ref{Fig_MEFE} shows the spatial cross-sections of the
arbitrarily chosen realizations of the model error fields (left) and
the drift fields (right).
The realizations generated by the separable random field model are given in black
and the realizations of the proportional-scales field are represented by the red curves.
\begin{figure
\begin{center}
{ \scalebox{0.34}{\includegraphics{ME_1D_100x100.pdf}}
\scalebox{0.34}{\includegraphics{MEintegr_1D_100x100.pdf}}
}
\end{center}
\caption{Spatial cross-sections of the simulated fields.
{\em Left}:: Model error fields.
{\em Right}: Time integrated model error fields.
}
\label{Fig_MEFE}
\end{figure}
One can see that, indeed, the time integration did not change the spatial structure of the
{\em separable} field (compare the two black curves in Fig.\ref{Fig_MEFE}, left and right).
In contrast, the time integrated {\em proportional-scales} field becomes much smoother in space
(compare the two red curves in Fig.\ref{Fig_MEFE}, left and right).
Even more objectively, we estimated the {\em spatial micro-scale}
of the drift $\bar\xi(t,s)$.
The estimator was
$({ {\mathsf{Var}}\, \bar\xi} / {{\mathsf{Var}}\, \delta\bar\xi})^{1/2}\cdot h$,
where
$\delta\xi$ is the forward finite difference in space,
the variance ${\mathsf{Var}}\,$ was estimated by averaging over the space coordinate
and over an ensemble of 100 realizations), and
$h$ is the spatial mesh size.
The resulting spatial micro-scales for the two
fields in question are displayed in Fig.\ref{Fig_Lx}
as functions of time.
\begin{figure
\begin{center}
{ \scalebox{0.35}{ \includegraphics{LxFE_100x100.pdf}}
}
\end{center}
\caption{Spatial micro-scale as a function of integration time
for the two time integrated random fields (separable and proportional-scales).
}
\label{Fig_Lx}
\end{figure}
As expected, in the separable case, the spatial micro-scale did not change
as a result of the time integration (the flat black line), whereas in the non-separable
proportional-scale case, the spatial length scale of the model error drift
rapidly grew in time.
It is worth emphasizing that in data assimilation,
the spatial length scale is a very important
attribute of the forecast error field and thus needs to be correctly represented by a
forecast (background) ensemble.
Thus, we have shown that the specific type of the spatio-temporal interactions in a model (tendency) error
field has important consequences for the spatial structure of the
resulting practically relevant forecast error field.
We have no evidence on the actual model error spatio-temporal structure, but we know
that non-separability and, more specifically, proportionality of scales is ubiquitous
in geophysics \citep{Tsyroulnikov2001}, therefore,
we postulate that the SPG should produce proportional-scales fields.
\section {SPG: Requirements and approach}
\label{sec_req}
The general requirements are:
\begin{enumerate}
\item
\label{req_SPG_isotr}
The SPG should produce univariate stationary in time and
homogeneous (stationary) and isotropic in space Gaussian pseudo-random fields $\xi(t, {\bf s})$
in 3D and 2D spatial domains.
\item
The SPG should be fast enough so that it does not significantly slow down
the forecast model computations.
\item
The magnitude as well as the spatial and temporal length scales of $\xi(t, {\bf s})$ are to be tunable.
\end{enumerate}
\noindent We also impose more specific requirements:
\begin{enumerate}
\setcounter{enumi}{3}
\item
\label{req_SPG_cont}
The random field $\xi(t, {\bf s})$ should have
finite variance and continuous realizations (sample paths).
\item
\label{req_SPG_PS}
The spatio-temporal covariances should obey the ``proportionality of scales'' principle:
larger (shorter) spatial scales should be associated with larger (shorter) temporal scales
\citep{Tsyroulnikov2001}.
\item
\label{req_SPG_flex}
The SPG {\em ansatz} should be
flexible enough to allow
for practicable solutions in both physical space and spectral space.
\end{enumerate}
Two comments are in order.
Firstly, stationarity, homogeneity, isotropy, and Gaussianity imposed by requirement \ref{req_SPG_isotr}
are just the simplest natural properties of a spatio-temporal random field.
The SPG is intended to be used as a building block
in practical schemes like the above SPPT, SKEB, SPP, or others.
Its role is to be the source of meaningful and easily tunable spatio-temporal stochasticity,
whereas physical model error features (flow dependence, non-Gaussianity, etc.)
are to be provided by the specific model error modeling scheme on a point-by-point basis.
Secondly, requirement \ref{req_SPG_flex} demands an SPG equation to be solvable in physical space as well as in spectral space
for the following reasons. All the above mentioned existing pattern generators are
spectral space based because this is the simplest way to get a homogeneous and isotropic
field in physical space. So, following this path, we would like to have a spectral-space solver.
But we envision that a combination of a homogeneous and isotropic spatial structure
(provided by the SPG) and point-by-point flow dependent and/or non-Gaussian features
(provided by the specific model error modeling scheme)
can appear too restrictive in the near future. Specifically, this
combination approach cannot produce variable local spatial and temporal length scales
if used in schemes like the SKEB or SPP (say, we may wish to reduce the local
length scales in meteorologically active areas like cyclones or convective systems).
Therefore, we wish the SPG equation to allow for a physical-space
solver that would be capable of imposing variable in space and time structures.
As a starting point in the development of the SPG,
we select the general class of {\em linear evolutionary stochastic
partial differential equations} (SPDE).
This choice is motivated by the flexibility of this class of spatio-temporal models
\citep[e.g.][]{Lindgren2011}.
In particular, for an SPDE, it is relatively easy to introduce inhomogeneity
(non-stationarity) in space and time as well as
local anisotropy---either by changing coefficients of the spatial operator
or by changing local properties of the driving noise.
One can also produce non-Gaussian fields by making the random forcing
non-Gaussian \citep[e.g.][]{Aberg,Wallin}. Physical-space discretizations of
SPDEs lead to {\em sparse} matrices, which give rise to fast numerical algorithms.
If an SPDE has constant coefficients, then it can be efficiently solved
using spatial spectral-space expansions.
In this study, we develop the SPG that relies
on a spatio-temporal stochastic model with {\em constant coefficients} so that both physical-space
and spectral-space solvers can be employed.
To facilitate the spectral-space solution, the general strategy is to define the SPG model on a standardized spatial domain.
The operational pseudo-random fields are then produced by mapping of the generated fields
from the standardized domain to the forecast-model domain.
In 3D, the standardized spatial domain is chosen to be the unit cube with the periodic
boundary conditions in all three dimensions, in other words,
the three-dimensional (3D) unit torus.
In 2D, the standardized domain is the 2D unit torus.
The 3D and 2D cases are distinguished
by the dimensionality $d=2$ or $d=3$ in what follows.
To simplify the presentation, the default dimensionality will be $d=3$.
\section {Tentative first-order SPG model}
\label{sec_order1}
\subsection{Physical-space model}
\label{sec_order1_phy}
The random field in question $\xi(t,{\bf s})$ is a function of the time coordinate $t$
and the space vector ${\bf s}=(x,y,z)$, where $(x,y,z)$ are the three spatial coordinates.
Each of the spatial coordinates belongs to the the unit circle $\S^1$,
so that
${\bf s}$ is on the unit torus
$T^3=\S^1 \times \S^1 \times \S^1$ (${\mathbb T}^2$ in the 2D case).
We start with the simplest general form of the first-order Markov model:
\begin{linenomath*}
\begin {equation}
\label {spg_gen}
\frac {\partial \xi(t,{\bf s})}{\partial t} + A \, \xi(t,{\bf s}) = \alpha(t,{\bf s}),
\end {equation}
\end{linenomath*}
where $A$ is the spatial linear operator to be specified and $\alpha$ is the driving noise.
$\alpha$ is postulated to be white in space and time; this is done to facilitate a fast numerical solver
in physical space as demanded by requirement \ref{req_SPG_flex}
in section \ref{sec_req} because generation of the white noise
is computationally inexpensive (its values on a grid in space and time are just independent
Gaussian random variables).
The SPG is required to be fast, so we choose $A$ to be a {\em differential} operator
(because, as we noted, in this case a physical-space discretization of $A$ gives rise to
a very sparse matrix).
Further, since we wish $\xi(t,{\bf s})$ to be homogeneous and isotropic in space,
we define $A$ to be a polynomial of the negated spatial Laplacian:
\begin{linenomath*}
\begin {equation}
\label {ALapl}
A = P(-\Delta) = \sum^q_{j=0} c_j (-\Delta)^j,
\end {equation}
\end{linenomath*}
where $P(x)$ is the polynomial and $q$ its degree (a positive integer).
We will refer to $q$ as the spatial order of the SPG model.
Note that the negation of the Laplacian is convenient because $(-\Delta)$ is
a non-negative definite operator.
The model Eq.(\ref{ALapl}) appears to be too rich for the purposes
of the SPG at the moment, so
in what follows we employ an even more reduced (but still quite flexible) form
\begin{linenomath*}
\begin {equation}
\label {ALapl2}
A=P(-\Delta) = \mu (1 - \lambda^2\Delta)^q,
\end {equation}
\end{linenomath*}
where $\mu$ and $\lambda$ are positive real parameters.
So, we start with the following SPG equation:
\begin{linenomath*}
\begin {equation}
\label {bpg1}
\frac {\partial \xi(t,{\bf s})}{\partial t} + \mu (1 - \lambda^2\Delta)^q \,\xi(t,{\bf s}) = \alpha(t,{\bf s}).
\end {equation}
\end{linenomath*}
\subsection{Spectral-space model}
\label{sec_order1_spe}
On the torus ${\mathbb T}^\mathnormal{d}$, a Fourier series is an expansion in the basis functions
$\rm e^{\i ({\bf k}, {\bf s})} \equiv \rm e^{\i (\mathnormal{mx+ny+lz})}$,
where the wavevector ${\bf k}$ is, for $d=3$, the triple
of integer wavenumbers, ${\bf k}=(m,n,l)$.
We perform the Fourier decomposition for both $\alpha(t, {\bf s})$ and $\xi(t, {\bf s})$,
\begin{linenomath*}
\begin {equation}
\label {bpg7a}
\alpha(t, {\bf s})=\sum_{{\bf k} \in {\mathbb Z}^d} \tilde\alpha_{\bf k}(t) \rm e^{\i ({\bf k}, {\bf s})}
\end {equation}
\end{linenomath*}
and
\begin{linenomath*}
\begin {equation}
\label {bpg7}
\xi(t, {\bf s})=\sum_{{\bf k} \in {\mathbb Z}^d} \tilde\xi_{\bf k}(t) \rm e^{\i ({\bf k}, {\bf s})}
\end {equation}
\end{linenomath*}
(where ${\mathbb Z}$ denotes the set of integer numbers)
and substitute these expansions into Eq.(\ref{bpg1}).
From the orthogonality of the basis functions,
we obtain that Eq.(\ref{bpg1}) decouples into the set of ordinary stochastic differential
equations \citep[OSDE, e.g.][]{Jazwinski,Arnold} in time:
\begin{linenomath*}
\begin {equation}
\label {osde1}
\frac{\d \tilde\xi_{\bf k}} {\d t} + \mu (1 + \lambda^2 {k}^2)^q
\, \tilde\xi_{\bf k}(t)= \tilde\alpha_{\bf k}(t),
\end {equation}
\end{linenomath*}
where $k=|{\bf k}|=\sqrt{m^2+n^2+l^2}$.
The white driving noise $\alpha$ is stationary, hence the
spectral-space coefficients $\tilde\alpha_{\bf k}(t)$ are
probabilistically independent random processes.
This is well known for random fields on the $d$-dimensional real space ${\mathbb R}^d$ (where spectra are continuous),
see e.g.\ Chapter 2 in \citet{Adler} or section 8 in \citet{Yaglom},
and can be directly verified in our case of the fields on the torus (where spectra are discrete).
Therefore, for different wavevectors ${\bf k}$,
the resulting spectral-space equations
are probabilistically completely {\em independent} from each other.
This greatly simplifies the solution of the SPG equations because instead of handling
the complicated SPDE Eq.(\ref{bpg1}) we have to solve a number of independent simple OSDEs
Eq.(\ref{osde1}).
Further, from the postulated whiteness of the spatio-temporal random field
$\alpha(t,{\bf s})$, all $\tilde\alpha_{\bf k}(t)$
are white in time random processes with the same intensity $\sigma$,
(Appendix \ref{app_wnoise}):
\begin{linenomath*}
\begin {equation}
\label {swn1}
\tilde\alpha_{\bf k}(t) = \sigma \, \Omega_{\bf k}(t),
\end {equation}
\end{linenomath*}
where $\Omega_{\bf k}(t)$ are the independent {\em standard} white noises,
i.e.\ the derivatives of the independent standard Wiener processes $W_{\bf k}(t)$
such that
\begin{linenomath*}
\begin {equation}
\label {wp1}
\Omega_{\bf k}(t) \d t = \d W_{\bf k}(t).
\end {equation}
\end{linenomath*}
Thus, the first-order SPG model reduces to a series of OSDEs
\begin{linenomath*}
\begin {equation}
\label {osde1a}
\d \tilde\xi_{\bf k} + \mu (1 + \lambda^2 {k}^2)^q \, \tilde\xi_{\bf k} \, {\d t} =
\sigma \, \d W_{\bf k}.
\end {equation}
\end{linenomath*}
For practical purposes the series is truncated, so that ${\bf k}\equiv (m,n,l)$ is limited:
$|m| < m_{\rm max}$,
$|n| < n_{\rm max}$, and
$|l| < l_{\rm max}$,
where $m_{\rm max}$, $n_{\rm max}$, and $l_{\rm max}$ are the truncation limits.
If not otherwise stated, all the truncation limits are the same and denoted by $n_{\rm max}$.
\subsection{Stationary spectral-space statistics}
\label{sec_statio}
Equation (\ref{osde1a}) is a first-order OSDE with constant coefficients sometimes
called the Langevin equation
(e.g.\ \citet{Arnold} or \citet{Jazwinski}, Example 4.12). Its generic form is
\begin{linenomath*}
\begin {equation}
\label {osde2}
\d \eta + a \eta \,\d t= \sigma \d W,
\end {equation}
\end{linenomath*}
where $\eta(t)$ is the random process in question, $a$ and $\sigma$ are constants, and
$W(t)$ is the standard Wiener process.
The solution to Eq.(\ref{osde2}) is known as the Ornstein-Uhlenbeck random process,
whose stationary (steady-state) temporal covariance function is
\begin{linenomath*}
\begin {equation}
\label {osde4}
B_\eta(t)=\frac{\sigma^2}{2a} \, \rm e^\mathnormal{-a|t|}
\end {equation}
\end{linenomath*}
\citep[e.g.][Example 4.12]{Jazwinski}. From Eq.(\ref{osde4}), it is clear that $a$ has the
meaning of the inverse temporal length scale $\tau = 1/a$.
Now, consider
the stationary covariance function of the elementary random process $\tilde\xi_{\bf k}(t)$,
\begin{linenomath*}
\begin {equation}
\label {bt}
{\mathsf{E}\,} \tilde\xi_{\bf k}(t_0) \cdot \tilde\xi_{\bf k}(t_0 +t) = b_{\bf k} \cdot C_{\bf k}(t),
\end {equation}
\end{linenomath*}
where $b_{\bf k}$ is the variance and $C_{\bf k}(t)$ the correlation function.
According to Eq.(\ref{bpg7}), $\tilde\xi_{\bf k}$ is the spatial spectral component of the
random field in question $\xi(t,{\bf s})$. Therefore $b_{\bf k}={\mathsf{Var}}\,\tilde\xi_{\bf k}$
is called the {\em spatial spectrum} of $\xi(t,{\bf s})$.
From Eqs.(\ref{osde1a}) and (\ref{osde4}), we have
\begin{linenomath*}
\begin {equation}
\label {bk}
b_{\bf k} = \frac {\sigma^2} {2 \mu (1 + \lambda^2 {k}^2)^q}
\end {equation}
\end{linenomath*}
and $C_{\bf k}(t) = \exp( -{|t|} / {\tau_{\bf k}} )$, where
\begin{linenomath*}
\begin {equation}
\label {tauk}
\tau_{\bf k} = \frac {1} { \mu (1 + \lambda^2 {k}^2)^q}
\end {equation}
\end{linenomath*}
is the temporal length scale associated with the spatial wavevector ${\bf k}$.
Note that by the spectrum (e.g.\ $b_{\bf k}$) we always mean the {\em modal} spectrum,
i.e.\ the variance associated with a single basis function
(a single wavevector ${\bf k}$); the modal spectrum
is not to be confused with the variance (or energy) spectrum.
\subsection{Physical-space statistics}
In the stationary regime (i.e.\ after an initial transient period has passed),
the above independence of the spectral random processes $\tilde\xi_{\bf k}(t)$
(see section \ref{sec_order1_spe})
implies that the random field $\xi(t, {\bf s})$ is
spatio-temporally {\em homogeneous},
i.e.\ invariant under shifts in space and time:
\begin{linenomath*}
\begin {equation}
\label {hom}
{\mathsf{E}\,} \xi(t, {\bf s}) \cdot \overline {\xi(t + \Delta t, {\bf s} + \Delta {\bf s})} = B(\Delta t, \Delta {\bf s}),
\end {equation}
\end{linenomath*}
where ${\mathsf{E}\,}$ is the expectation operator and
\begin{linenomath*}
\begin {equation}
\label {Bts}
B(t, {\bf s}) = \sum_{\bf k} b_{\bf k}\, C_{\bf k}(t) \,\rm e^{\i ({\bf k}, {\bf s})}.
\end {equation}
\end{linenomath*}
In particular, the spatial covariance function is
\begin{linenomath*}
\begin {equation}
\label {Bs}
B({\bf s}) = B(t=0, {\bf s}) =
\sum_{\bf k} b_{\bf k} \,\rm e^{\i ({\bf k}, {\bf s})},
\end {equation}
\end{linenomath*}
where it is seen that the spatial spectrum $b_{\bf k}$ is the Fourier transform of the
spatial covariance function $B({\bf s})$.
The temporal covariance function is
\begin{linenomath*}
\begin {equation}
\label {B_t}
B({t}) = B(t, {\bf s=0}) =
\sum_{\bf k} b_{\bf k} \,C_{\bf k}(t).
\end {equation}
\end{linenomath*}
Finally, the variance is
\begin{linenomath*}
\begin {equation}
\label {Vxi}
{\mathsf{Var}}\,\xi = B(t=0, {\bf s=0}) = \sum_{\bf k} b_{\bf k}.
\end {equation}
\end{linenomath*}
\subsection{``Proportionality of scales'' requires that $q=\frac12$}
\label{sec_PSproperty}
The more precise formulation of the ``proportionality of scales'' requirement \ref{req_SPG_PS}
states that for large $k$,
the temporal length scale $\tau_{\bf k}$ should be inversely proportional to $k$:
\begin{linenomath*}
\begin {equation}
\label {cond_PS}
\tau_{\bf k} \sim \frac{1}{k} \quad \mbox{as} \quad k \to \infty.
\end {equation}
\end{linenomath*}
From Eq.(\ref{tauk}), this condition entails, importantly, that
\begin{linenomath*}
\begin {equation}
\label {cond_PS2}
q=\frac12.
\end {equation}
\end{linenomath*}
Below, we show that the choice $q=\frac12$ causes the generated spatio-temporal random fields
to possess, besides the ``proportionality of scales'', many other nice properties (sections
\ref{sec_isotr} and \ref{sec_covs}).
\subsection{The spatial operator of order $q=\frac12$}
\label{sec_Aq12}
The model's spatial operator $A$ becomes (see Eq.(\ref{ALapl2}))
\begin{linenomath*}
\begin {equation}
\label {ALapl12}
A = \mu ({1 - \lambda^2\Delta})^\frac12 \equiv \mu \sqrt{ 1 - \lambda^2\Delta }.
\end {equation}
\end{linenomath*}
This is a pseudo-differential operator \citep[e.g.][]{Shubin} with the {\em symbol}
\begin{linenomath*}
\begin {equation}
\label {symb}
a({k}) = \mu \sqrt{1 + \lambda^2{k}^2},
\end {equation}
\end{linenomath*}
so that the action of $A$
on the test function $\varphi({\bf s})$ is defined as follows.
First, we Fourier transform $\varphi({\bf s})$ getting $\{ \tilde\varphi_{\bf k} \}$.
Then, $\forall {\bf k} \in {\mathbb Z}^d$, we multiply $\tilde\varphi_{\bf k}$ by the symbol
$a({k})$.
Finally, we perform the backward Fourier transform of
$\{ a({k}) \tilde\varphi_{k} \}$
retrieving the result, the function $(A\varphi)({\bf s})$.
So, the action of the above fractional negated and shifted Laplacian on test functions in spectral space
is well defined.
Importantly, in physical space, the pseudo-differential operator $A$ can be
approximated by a discrete-in-space linear operator which is represented by a {\em very sparse matrix},
see Appendix \ref{app_symbol}.
So, in both spectral space and physical space, the resulting operator $A$ with the
fractional degree $q=\frac12$ is numerically tractable.
\subsection{The first-order model cannot satisfy the SPG requirements}
\label{sec_impose}
Let us compute ${\mathsf{Var}}\,\xi$ using Eqs.\ref{bk} and \ref{Vxi}.
Since $b_{\bf k}$ is a smooth function of the wavevector ${\bf k}$, we may approximate the sum in Eq.\ref{Vxi}
with the integral (where $b({\bf k})=b_{\bf k}$ for integer wavenumbers), getting
\begin {equation}
\label {fvd}
{\mathsf{Var}}\,\xi \propto \int _{{\mathbb R}^d} \frac {1} {\sqrt{1 + \lambda^2 {k}^2}} \,\d {\bf k} \propto
\int_{{\mathbb R}} \frac {k^{d-1}} {\sqrt{1 + \lambda^2 {k}^2}} \,\d k.
\end {equation}
To check the convergence of the latter integral in Eq.(\ref{fvd}),
we examine the $k \to \infty$ limit. For large $k$, the integrand is, obviously,
proportional to $k^{d-2}$.
As we know, the integral of this kind converges if the integrand decays faster than ${k^{-1-\epsilon}}$
with some $\epsilon >0$.
This implies that the integral in Eq.(\ref{fvd}) diverges for all $d \ge 1$.
In other words,
the spectrum Eq.(\ref{bk}) decays too slowly for ${\mathsf{Var}}\,\xi$ to be finite.
So, the SPG model Eq.(\ref{bpg1}) cannot simultaneously satisfy the proportional-scales requirement \ref{req_SPG_PS}
(which leads to $q=\frac12$) and the finite-variance requirement \ref{req_SPG_cont}.
Consequently, the SPG model is to be somehow changed.
The solution is to increase the temporal order of the model.
\section {Higher-order in time model}
\label{sec_phy}
\subsection{Formulation}
\label{sec_motiv}
The SPG model of higher temporal order is
\begin{linenomath*}
\begin {equation}
\label {bpg_p}
\left( \frac {\partial }{\partial t} + \mu \sqrt{1 - \lambda^2\Delta} \right)^p \xi(t,{\bf s}) =
\alpha(t,{\bf s}),
\end {equation}
\end{linenomath*}
where $p$ is the temporal order of the modified SPG model (a positive integer).
In spectral space, the model reads (cf. section \ref{sec_order1_spe})
\begin{linenomath*}
\begin {equation}
\label {bpg_ps}
\left(\frac{\d} {\d t} + \mu \sqrt{1 + \lambda^2 k^2}\right)^p \tilde\xi_{\bf k}(t) =
\sigma \, \Omega_{\bf k}(t).
\end {equation}
\end{linenomath*}
In this section, we explore the steady-state statistics of $\xi(t,{\bf s})$
and find out which values of the temporal order $p$ solve the above infinite variance problem.
\subsection{Stationary spectral-space statistics}
For each ${\bf k}$, Eq.(\ref{bpg_ps}) is a $p$th-order in time OSDE.
Using Table \ref{Tab_osde} in Appendix \ref{app_OSDE},
we can write down
the stationary variance $b_{\bf k}$ and the temporal correlation function $C_{\bf k}(t)$
of the solution to Eq.(\ref{bpg_ps}), the process $\tilde\xi_{\bf k}(t)$:
%
\begin{linenomath*}
\begin {equation}
\label {bk12_p}
b_{\bf k} \propto \frac {\sigma^2} { \mu^{2p-1} (1 + \lambda^2 {k}^2)^{p-\frac12} }
\end {equation}
\end{linenomath*}
(where the sign $\propto$ means proportional to)
and
%
\begin{linenomath*}
\begin {equation}
\label {Ck_p}
C_{\bf k}(t) = \left(1 + \frac{|t|} {\tau_{\bf k}} + r_2 \frac{|t|^2} {\tau_{\bf k}^2} + \dots +
r_{p-1} \frac{|t|^{p-1}} {\tau_{\bf k}^{p-1}} \right)
\,\rm e^{- \frac{|\mathnormal{t}|}{\tau_{\bf k}} }.
\end {equation}
\end{linenomath*}
Here $r_2,\dots, r_{p-1}$ are real numbers (given for $p=1,2,3$ in Table \ref{Tab_osde},
see Appendix \ref{app_OSDE})
and $\tau_{\bf k}$ are still defined by Eq.(\ref{tauk}).
Specifically, for the temporal order $p=3$, we have
%
\begin{linenomath*}
\begin {equation}
\label {bk_p3}
b_{\bf k}|_{p=3} = \frac {3\sigma^2} {16 \mu^5 (1 + \lambda^2 {k}^2)^\frac{5}{2} }
\end {equation}
\end{linenomath*}
and
%
\begin{linenomath*}
\begin {equation}
\label {Ck_p3}
C_{\bf k}(t)|_{p=3} = \left(1 + \frac{|t|} {\tau_{\bf k}} + \frac13 \,\frac{|t|^{2}} {\tau_{\bf k}^{2}} \right) \,\rm e^{- \frac{|\mathnormal{t}|}{\tau_{\bf k}} }.
\end {equation}
\end{linenomath*}
As Eq.(\ref{tauk}) is unchanged in the higher order model,
the ``proportionality of scales'' condition Eq.(\ref{cond_PS}) is still satisfied.
In order to achieve the desired dependency of $\tau_{\bf k}$ not only on $k$ (which we already have from Eq.(\ref{tauk})),
but also on $\lambda$ (the greater is $\lambda$ the greater should be $\tau_{\bf k}$),
we parameterize $\mu$ as
%
\begin{linenomath*}
\begin {equation}
\label {mu}
\mu = \frac{U} {\lambda},
\end {equation}
\end{linenomath*}
where $U>0$ is the velocity-dimensioned tuning parameter.
Note that $\lambda$ affects both the spatial length scale of $\xi$
(due to Eq.(\ref{bk12_p})) and the temporal length scale (thanks to Eq.(\ref{tauk})).
In contrast, $U$ affects only the temporal length scale.
\subsection{Finite-variance criterion}
\label{sec_finvar_p}
Substituting $b_k$ from Eq.(\ref{bk12_p}) into Eq.(\ref{Vxi}),
approximating the sum over the wavevectors by the integral,
and exploiting the isotropy of the integrand yields
\begin{linenomath*}
\begin {equation}
\label {vardp}
{\mathsf{Var}}\,\xi \approx {\mathsf{const}} \cdot \int_0^\infty \frac{\sigma^2} { (1 + \lambda^2 {k}^2)^{p-\frac12} } \, k^{d-1} \,\d k,
\end {equation}
\end{linenomath*}
so that we have ${\mathsf{Var}}\,\xi <\infty$ (requirement \ref{req_SPG_cont}) whenever
\begin{linenomath*}
\begin {equation}
\label {critdp}
p > \frac{d+1}{2}.
\end {equation}
\end{linenomath*}
\subsection{Isotropy}
\label{sec_isotr}
In this section, we show that, remarkably, $q=\frac12$ is the unique spatial order for which the field $\xi(t, {\bf s})$
appears to be isotropic in space-time. In particular, the shape of the correlation function
is the same in any spatial or temporal or any other direction in the spatio-temporal domain
${\mathbb T}^\mathnormal{d} \times {\mathbb R}$.
\subsubsection{Spatial isotropy}
\label{sec_spat_isotr}
We note that the spatial isotropy of the random field $\xi$
is the invariance of its covariance function $B({\bf s})$ under rotations.
If we were in ${\mathbb R}^d$ rather than on ${\mathbb T}^\mathnormal{d}$,
isotropy of $B({\bf s})=B(s)$, where $s = |{\bf s}|$ is the spatial distance, would be equivalent
to isotropy of its Fourier transform (spectrum) $ b({\bf k})$, so that the latter would be
dependent only $k=|{\bf k}|$.
On the torus, spectra are discrete, i.e. $m,n,l$ take only {\em integer} values,
so, strictly speaking, $b_{\bf k}$ cannot be isotropic there.
To avoid this technical difficulty, we resort (for the theoretical analysis only)
to the device used in sections \ref{sec_impose} and
\ref{sec_finvar_p}, the approximation of a sum over
the wavevectors by an integral.
Specifically, we assume that $b_{\bf k}$ is smooth enough
(which is tantamount to the assumption that $B({\bf s})$ decays on length scales much smaller than
the domain's extents) for the validity of the approximation
\begin{linenomath*}
\begin {equation}
\label {Bs_int}
B({\bf s}) = \sum_{\bf k \in {\mathbb Z}^d} b_{\bf k} \,\rm e^{\i ({\bf k}, {\bf s})} \approx
\int_{{\mathbb R}^d} b({\bf k}) \,\rm e^{\i ({\bf k}, {\bf s})} \d {\bf k},
\end {equation}
\end{linenomath*}
where $b({\bf k})$ is a smooth function of the real vector argument ${\bf k} \in {\mathbb R}^d$ such that
$\forall \bf k \in {\mathbb Z}^d, b({\bf k})=b_{\bf k}$.
The integral in Eq.(\ref{Bs_int}) with the isotropic $b({\bf k})$, see Eq.(\ref{bk12_p}),
can be easily shown to be invariant under rotations of ${\bf s}$.
This implies that $B({\bf s})$ and so the random field $\xi$ are
indeed approximately spatially isotropic.
In the theoretical analysis in this section, we will rely on the approximation Eq.(\ref{Bs_int})
and thus assume that the ``spectral grid'' is dense enough for the spatial spectra to be treated
as continuous ones.
\subsubsection{Isotropy in space-time}
\label{sec_spatemp_spec}
Consider the OSDE Eq.(\ref{bpg_ps}) in the stationary regime.
Following \citet[][section 8]{Yaglom}, the stationary random process can be spectrally
represented as the stochastic integral
\begin{linenomath*}
\begin {equation}
\label {timspe}
\tilde\xi_{\bf k}(t) = \int_{{\mathbb R}} \rm e^{\i \omega \mathnormal{t}} \, Z_{\bf k}(\d \omega),
\end {equation}
\end{linenomath*}
where $\omega$ is the angular frequency (temporal wavenumber) and $Z$ is the
orthogonal stochastic measure such that
\begin{linenomath*}
\begin {equation}
\label {b_k_omega}
{\mathsf{E}\,} |Z_{\bf k}(\d \omega)|^2 = b_{\bf k}(\omega) \, \d \omega,
\end {equation}
\end{linenomath*}
where $b_{\bf k}(\omega)$ is the spectral density of the process $\tilde\xi_{\bf k}(t)$
(i.e.\ the Fourier transform of its covariance function $b_{\bf k} C_{\bf k}(t)$,
see Eq.(\ref{bt})) and,
at the same time, the spatio-temporal spectrum of the field $\xi$.
In the spectral expansion of the driving white noise $\Omega_{\bf k}(t)$ (see Eq.(\ref{bpg_ps})),
\begin{linenomath*}
\begin {equation}
\label {wn_spe}
\Omega_{\bf k}(t) = \int_{{\mathbb R}} \rm e^{\i \omega \mathnormal{t}} \, Z_{\Omega_{\bf k}}(\d \omega),
\end {equation}
\end{linenomath*}
we have ${\mathsf{E}\,} |Z_{\Omega_{\bf k}}(\d \omega)|^2 ={\mathsf{const}} \cdot \d\omega$ because the white noise
has constant spectral density.
Next, we substitute Eqs.(\ref{timspe}) and (\ref{wn_spe}) into Eq.(\ref{bpg_ps}), getting
\begin{linenomath*}
\begin {equation}
\label {ZZ}
(\i\omega + \mu \sqrt{1 + \lambda^2 k^2})^p Z_{\bf k}(\d \omega) = Z_{\Omega_{\bf k}}(\d \omega).
\end {equation}
\end{linenomath*}
In this equation, taking expectation of the squared modulus of both sides,
recalling that $\mu=U/\lambda$,
and introducing the scaled angular frequency $\omega' = \omega /U$, we finally obtain
\begin{linenomath*}
\begin {equation}
\label {b_k_omega_p}
b_{\bf k}(\omega') \equiv b_{\bf K} \propto \frac{1}{ (\lambda^{-2} + (\omega')^2 + k^2)^p } =
\frac{1}{ (\lambda^{-2} + {\bf K}^2)^p },
\end {equation}
\end{linenomath*}
where
\begin{linenomath*}
\begin {equation}
\label {Kk}
{\bf K} = (\ \frac{\omega}{U}, {\bf k}) \equiv (\frac{\omega}{U}, m, n, l )
\end {equation}
\end{linenomath*}
is the spatio-temporal wavevector.
From Eq.(\ref{b_k_omega_p}), one can see that with the scaled frequency (note that the change
$\omega \to \omega / U$ corresponds to the change of the time coordinate
$t \to t \cdot U$), the spatio-temporal spectrum $b_{\bf k}(\omega') \equiv b_{\bf K}$ becomes
{\em isotropic in space-time}. This implies that the correlation function of $\xi$ is isotropic
in space-time as well (with the scaled time coordinate).
Note that this remarkable property can be achieved only with the spatial order $q=\frac12$.
It is worth noting that
the functional form of the spatio-temporal spectrum Eq.(\ref{b_k_omega_p})
together with the constraint Eq.(\ref{critdp}) imply that the conditions of Theorem 3.4.3 in
\citet[][]{Adler} are satisfied, so that spatio-temporal sample paths of the random field $\xi$ are almost surely continuous,
as we demanded in section \ref{sec_req}, see requirement \ref{req_SPG_cont}.
\subsection {Spatio-temporal covariances: the Mat\'ern class}
\label{sec_covs}
The spatio-temporal field satisfying the $p$-th order SPG model Eq.(\ref{bpg_p}) has the spatio-temporal
correlation function belonging to the so-called {\em Mat\'ern} class of
covariance functions \citep[e.g.][]{SteinBook,Guttorp}. To see this, we denote
%
\begin{linenomath*}
\begin {equation}
\label {nupd}
\nu = p - \frac{d+1}{2} >0,
\end {equation}
\end{linenomath*}
where the positivity follows from Eq.(\ref{critdp}).
Then Eq.(\ref{b_k_omega_p}) rewrites as
\begin{linenomath*}
\begin {equation}
\label {bk_p_nu}
b_{\bf K} \propto \frac {1} { (\lambda^{-2} + {\bf K}^2)^{\nu+\frac{d+1}{2}} }.
\end {equation}
\end{linenomath*}
Note that here $d+1$ is the dimensionality of space-time.
Equation (\ref{bk_p_nu}) indeed presents the spectrum of the
Mat\'ern family of correlation functions, see e.g.\ Eq.(32) in \citet{SteinBook}.
The respective isotropic correlation function is given by the equation that precedes Eq.(32)
in \citet{SteinBook} or by Eq.(1) in \citet[][]{Guttorp}:
\begin{linenomath*}
\begin {equation}
\label {matern}
B(r) \propto ({r}/{\lambda})^\nu K_\nu ({r}/{\lambda}),
\end {equation}
\end{linenomath*}
where $r=\sqrt{s^2 + (Ut)^2}$ is the distance
(in our case, the Euclidean distance in space-time with the coordinates $(x,y,z,Ut)$)
and $K_\nu$ is the MacDonald function (the modified Bessel function of the second kind).
The Mat\'ern family is often recommended
for use in spatial analysis due to its notable flexibility with only two
free parameters: $\nu$ and $\lambda$, see e.g.\ \citet{SteinBook} and \citet{Guttorp}.
Specifically, $\lambda$ controls the length scale, whereas
$\nu>0$ determines the degree of smoothness: the higher $\nu$, the
smoother the field. Note that the smoothness is
understood as the number of the mean-square derivatives of the random field in question.
The degree of smoothness
depends on the behavior of the correlation function at small distances
and manifests itself in field's realizations as the amount of small-scale noise
(for illustration see Appendix \ref{app_smoo}).
Table \ref{Tab_Matern} lists the resulting correlation functions (in any direction
in space-time) for
several combinations of $d$ and $p$ \citep[see][for details]{Guttorp}.
\begin{table
\caption{\bf Spatio-temporal correlation functions $B(r)$ for some plausible combinations of
the dimensionality $d$ and the temporal order $p$}
\begin{center}
\begin{tabular}{|c|c||c||c|}
\hline
$d$ & $p$ & $\nu=p - \frac{d+1}{2}$ & $B(r)$\\
\hline
2 & 2 & $\frac12$ & $\rm e^{-\frac{\mathnormal{r}}{\lambda}}$ \\
\hline
2 & 3 & $\frac32$ & $(1 + \frac{r}{\lambda})\,\rm e^{-\frac{\mathnormal{r}}{\lambda}}$ \\
\hline
2 & 4 & $\frac52$ & $(1 + \frac{r}{\lambda} + \frac13 \left(\frac{r}{\lambda}\right)^2)\,\rm e^{-\frac{\mathnormal{r}}{\lambda}}$ \\
\hline
3 & 3 & $1$ & $\frac{r}{\lambda} K_1 (\frac{r}{\lambda})$ \\
\hline
\end{tabular}
\end{center}
\label{Tab_Matern}
\end{table}
With the fixed $d$, the larger $p$ corresponds, according to Eq.(\ref{nupd}),
to the larger $\nu$ and so to the smoother in space and time field $\xi$.
This can be used to change the degree of smoothness of the generated field by changing
the temporal order of the SPG model.
From the constraint Eq.(\ref{critdp}), the minimal temporal order $p$ that can be used in both 2D and 3D is
equal to 3. {\em This value ${p=3}$ will be used by default in what follows and in the current SPG computer program}.
\subsection{Spatio-temporal correlation functions: illustrations}
\label{sec_spatim_illu}
Here we show spatial, temporal, and spatio-temporal correlation functions
computed using Eq.(\ref{matern}).
To make the plots more accessible, it is arbitrarily assumed that the
extent of the standardized spatial domain (the torus) in each dimension is 3000 km, so that
the distance is measured in kilometers.
The default SPG setup parameters are
$\lambda=125$ km and $U=20$ m/s.
\subsubsection{Spatial correlation functions}
\label{sec_spacrf}
Figure \ref{Fig_scrf} presents the {spatial} correlation functions
for different length scales in 2D and 3D.
One can notice, first, that the actual length scale is well controlled by the parameter $\lambda$.
Second, it is seen that in 2D (the left panel), where, according to Eq.(\ref{nupd}), $\nu= \frac{3}{2}$, the correlation functions
are somewhat {\em smoother} at the origin than in 3D (the right panel), where $\nu=1$.
This is consistent with the above statement that the greater $\nu$ the smoother the field.
But in general, the 2D and 3D spatial correlation functions are quite similar.
\begin{figure
\begin{center}
{ \scalebox{0.3}{ \includegraphics{Spacrf2D.pdf}}
\scalebox{0.3}{ \includegraphics{Spacrf3D.pdf}}
}
\end{center}
\caption{Spatial correlation functions for $p=3$ in 2D (the left panel) and 3D (the right panel)---for
the four spatial length
scales indicated in the legend.
}
\label{Fig_scrf}
\end{figure}
\subsubsection{Temporal correlation functions}
\label{sec_timcrf}
Equation (\ref{matern}) implies that the spatial and temporal correlations have the same {\em shapes}.
The latter feature is very nice because atmospheric spectra are known to be similar in the spatial and in the temporal domain,
e.g.\ the well-known ``-5/3'' spectral slope law
is observed both in space and time, see e.g.\ \citet[section 23]{Monin}.
So, the SPG does reproduce this observed in the atmosphere similarity of spatial and temporal correlations.
Figure \ref{Fig_tcrf_U} shows the temporal correlation functions
for different parameters $U$.
Comparing Fig.\ref{Fig_tcrf_U} with Fig.\ref{Fig_scrf}(right), one can observe
that the spatial and temporal correlations indeed have the same shape.
\begin{figure
\begin{center}
{ \scalebox{0.3}{ \includegraphics{Timcrf3D_U.pdf}}
}
\end{center}
\caption{Temporal correlation functions in 3D for the four values of $U$ indicated in the legend.
}
\label{Fig_tcrf_U}
\end{figure}
\subsubsection{Spatio-temporal correlations}
\label{sec_spatim}
Figure \ref{Fig_spatim1} presents the spatial correlation functions for different
time lags. Figure \ref{Fig_spatim2} displays the spatio-temporal correlation function.
In both Fig.\ref{Fig_spatim1} and Fig.\ref{Fig_spatim2},
another manifestation of the spatio-temporal ``proportionality of scales'' is seen:
the larger the time lag, the broader the spatial correlations.
Note that this is consistent with the behavior of the spatio-temporal covariances
found by \citet[Fig.8]{CressieHuang} in real-world wind speed data.
\begin{figure
\begin{center}
{ \scalebox{0.3}{ \includegraphics{SpacrfLagged.pdf}}
}
\end{center}
\caption{Spatial correlation functions in 3D for the four time lags indicated in the legend.
}
\label{Fig_spatim1}
\end{figure}
\begin{figure
\begin{center}
{ \scalebox{0.3}{ \includegraphics{Spatimcrf.pdf}}
}
\end{center}
\caption{Spatio-temporal SPG covariances.
}
\label{Fig_spatim2}
\end{figure}
\subsection{Introducing anisotropy in the vertical plane}
\label{sec_anis}
We have formulated the SPG model under the 3D isotropy assumption.
This implies that the ratio of the horizontal length scale to the horizontal domain size
is the same as the ratio of the vertical length scale to the vertical domain size.
This may be reasonable but, obviously, the independent specification of the horizontal
and vertical length scales would be much more flexible.
To get this capability, we can employ two equivalent modifications to the SPG model.
One approach is to change the radius of the ``vertical circle'' in the torus
from $1$ to the $\delta^{-1}$, where $\delta$ is a positive parameter.
Another approach is to replace the Laplacian
$\Delta = \frac{\partial^2}{\partial x^2} + \frac{\partial^2}{\partial y^2} + \frac{\partial^2}{\partial z^2}$
by its anisotropic version
$\Delta' = \frac{\partial^2}{\partial x^2} + \frac{\partial^2}{\partial y^2} + \delta^2\frac{\partial^2}{\partial z^2}$.
With both approaches, the vertical length scale increases by the factor of $\delta$.
\subsection{Preserving isotropy in the horizontal plane for non-square domains}
\label{sec_horz_isotr}
If the size of the domain in physical space in the $x$ direction, $D_x$, differs from the domain size
in the $y$ direction, $D_y$, then mapping from
a square SPG domain to the rectangular physical domain would result in an elliptic (also called geometric)
anisotropy in the horizontal plane.
This undesirable feature can be avoided by replacing $\Delta'$ defined in section \ref{sec_anis}
with
\begin{linenomath*}
\begin {equation}
\label {Lapl_anis}
\Delta_* = \frac{\partial^2}{\partial x^2} + \gamma^2 \frac{\partial^2}{\partial y^2} +
\delta^2\frac{\partial^2}{\partial z^2},
\end {equation}
\end{linenomath*}
where $\gamma={D_x} /{D_y}$.
The only change in all the above spectral equations is that the wavenumbers $n$ and $l$ are
to be multiplied by $\gamma$ and $\delta$, respectively. The total spatial
wavenumber squared $k^2$ is to be replaced everywhere by its scaled version
\begin{linenomath*}
\begin {equation}
\label {k2_anis}
k^2_*=m^2+ (\gamma n)^2+(\delta l)^2.
\end {equation}
\end{linenomath*}
More technically, in our implementation of the SPG, the number of grid points on the torus,
$n_x^T,n_y^T,n_z^T$,
differs from that on the physical-space domain, $n_x,n_y,n_z$, respectively, for two reasons.
Firstly, the grid on the torus is defined to have more grid points than in physical space,
let us denote them $n_x^+,n_y^+,n_z^+$.
This is done because on the cube with the periodic boundary conditions (i.e.\ on the torus) the correlations
between the opposite sides of the cube are close to 1 due to the periodicity.
In order to avoid these spurious correlations on the physical-space domain, we use only part
of the grid on the torus
(specifically, $n_x,n_y,n_z$ contiguous grid points) to map the field to the physical-space domain
grid point to grid point.
Having the user-defined grid sizes in the physical-space domain, $n_x,n_y,n_z$,
we specify the grid sizes on the torus, $n_x^+,n_y^+,n_z^+$, from the condition
that the resulting correlations between the opposite sides of the domain should be less than 0.2.
Secondly, we somewhat further increase the number of grid points
$n_x^+,n_y^+,n_z^+$ in order for the final grid sizes on the torus, $n_x^T,n_y^T,n_z^T$,
be multiples of 2,3,5 (as required by the fast Fourier transform software we use).
Then, we find $\gamma$ from the requirement that after the mapping from the torus to the physical-space domain,
the length scales of the horizontal function in the $x$ and $y$ directions be the same.
It is easy to see that this is the case if
\begin{linenomath*}
\begin {equation}
\label {Lapl_anis2gamma}
\gamma=\frac{n_x^T}{n_y^T}.
\end {equation}
\end{linenomath*}
This device indeed allows to preserve the horizontal isotropy and
to change the vertical length scale in a broad range (not shown).
We have refrained from introducing this feature to our basic SPG equations
for the sake of simplicity of presentation.
\subsection{The final formulation of the SPG model}
\label{sec_finSPG}
The temporal order of the SPG model is $\boxed{p=3}$.
The SPG model is
\begin {equation}
\label {bpg_3}
\boxed{ \left( \frac {\partial }{\partial t} + \frac{U}{\lambda} \sqrt{1 - \lambda^2\Delta_*} \right)^3 \xi(t,{\bf s}) = \alpha(t,{\bf s}) },
\end {equation}
where $\alpha(t,{\bf s})$ is the spatio-temporal white noise.
In spectral space, each spectral coefficient $\tilde\xi_{\bf k}(t)$ satisfies the equation
\begin {equation}
\label {bpg_3spe}
\boxed{ \left(\frac{\d} {\d t} + \frac{U}{\lambda} \sqrt{1 + \lambda^2 k_*^2}\right)^3 \tilde\xi_{\bf k}(t) =
\sigma \,\Omega_{\bf k}(t) },
\end {equation}
where $\Omega_{\bf k}(t)$ are mutually independent complex standard white noise processes.
The intensity of the spatio-temporal white noise $\alpha$ is $(2\pi)^{d/2} \,\sigma$.
\section {Time discrete solver for the third-order in time SPG model}
In {\em physical space}, our final evolutionary model Eq.(\ref{bpg_p}) with $p=3$ can be discretized using
the approximation of the operator $\sqrt{1 - \lambda^2\Delta}$ proposed in
Appendix \ref{app_symbol}.
The respective physical-space solver looks feasible but we do {\em not}
examine it in this study.
Below, we present our basic {\em spectral-space} technique.
From this point on, we will consider only the {\em spectral} SPG.
\subsection {The spectral solver}
To numerically integrate the SPG equations in spectral space,
we discretize Eq.(\ref{bpg_ps}) (with $p=3$) using an implicit scheme.
The model operator $(\frac{\d}{\d t} + a_{\bf k})^3$,
where $a_{\bf k}=\frac{U}{\lambda} \sqrt{1 + \lambda^2 k_*^2}$ and
$k_*^2$ is defined in Eq.(\ref{k2_anis}),
is discretized by replacing the
time derivative $\frac{\d}{\d t}$ with the backward finite difference $\frac{{\cal I - B}}{\Delta t}$,
where $\Delta t$ is the time step, ${\cal I}$ is the identity operator, and ${\cal B}$ is the backshift operator.
The white noise in the r.h.s.\ of Eq.(\ref{bpg_ps}) is discretized using Eq.(\ref{wp1}), where
$\d W_{\bf k}(t)$ is replaced with
$\Delta W_{\bf k}(t) = W_{\bf k}(t+\Delta t) - W_{\bf k}(t)$, and simulated as a zero-mean Gaussian random variable with the variance $\Delta t$.
As a result, we obtain the time discrete evolution equation
\begin{linenomath*}
\begin {equation}
\label {osde_num2}
\hat\xi_{\bf k}(i) = \frac{1} {\varkappa^3}
\left[ 3 \varkappa^2 \hat\xi_{\bf k}(i-1) -3\varkappa\hat\xi_{\bf k}(i-2)+\hat\xi_{\bf k}(i-3)+
\sigma {\Delta t}^{\frac52} \zeta_{{\bf k}t} \right],
\end {equation}
\end{linenomath*}
where $i=0, 1,2, \dots$ denotes the discrete time instance,
$\varkappa=1+ a_{\bf k} \Delta t$,
and $\zeta_{{\bf k}t} \sim CN(0,1)$ are independent complex standard Gaussian pseudo-random variables
(for their definition, see Appendix \ref{sec_dscr_wnoise}).
Note that the solution of the time-discrete Eq.(\ref{osde_num2}) is denoted by the hat, $\hat\xi_{\bf k}(i)$,
in order to distinguish it from the solution of the time-continuous Eq.(\ref{bpg_ps}), which is denoted by the tilde, $\tilde\xi_{\bf k}(t)$.
It can be shown that the numerical stability of the scheme Eq.(\ref{osde_num2})
is guaranteed whenever $\varkappa >1$, which is
always the case because $a_{\bf k} > 0 $ (see Eq.(\ref{symb})).
Note that the derivation of the numerical scheme for a higher-order (i.e.\ with $p>3$) SPG
model is straightforward: one should just raise the difference operator
$\frac{{\cal I - B}}{\Delta t}$ to a power higher than 3.
\subsection {Correction of spectral variances}
\label{sec_spec_correc}
Because of discretization errors,
the time discrete scheme Eq.(\ref{osde_num2}) gives rise to the steady-state spectral
variances $\hat b_{\bf k} = {\mathsf{Var}}\,\hat\xi_{\bf k}(i)$ that
are different from the ``theoretical'' ones, $b_{\bf k}$ (given in Eq.(\ref{bk_p3})).
The idea is to correct (multiply by a number) the solution $\hat\xi_{\bf k}(i)$
to Eq.(\ref{osde_num2})
so that the steady-state variance of the corrected $\hat\xi_{\bf k}(i)$ be equal to $b_{\bf k}$.
To this end, we derive $\hat b_{\bf k}$ from Eq.(\ref{osde_num2})
(using Eq.(\ref {sch_p3V}) in Appendix \ref{app_OSDE_dscr}) and
then, knowing the ``theoretical'' $b_{\bf k}$, we introduce the correction coefficients,
$\sqrt{b_{\bf k} / \hat b_{\bf k}}$, to be applied to $\hat\xi_{\bf k}(i)$.
This simple device ensures that for any time step, the spatial spectrum and thus the {\em spatial} covariances are perfect.
But the {\em temporal} correlations do depend on the time step, this aspect is discussed below in
section \ref{sec_gamma}.
\subsection {``Warm start'': ensuring stationarity from the beginning of the time integration}
To start the numerical integration of the third-order scheme Eq.(\ref{osde_num2}),
we obviously need three initial conditions.
If the integration is the continuation of a previous run, then we just take
values of $\hat\xi_{\bf k}(i)$ at the last three time instances $i$ from that previous run;
this ensures the continuity of the resulting
trajectory. If we start a new integration, we have to somehow generate values of ${\bf\hat \xi}_{\bf k}(i)$ at $i=1,2,3$,
let us denote them here as the vector
$\boldsymbol\xi^{ini} = ({\bf\hat \xi}_{\bf k}(1), {\bf\hat \xi}_{\bf k}(2), {\bf\hat \xi}_{\bf k}(3))^\top$.
Simplistic choices like specifying zero initial conditions give rise to a substantial initial transient period,
which distorts the statistics of the generated field in the short time range.
In order to have the steady-state regime right from the beginning of the time integration
and thus avoid the initial
transient period completely, we simulate $\boldsymbol\xi^{ini}$ as a pseudo-random draw from the
multivariate Gaussian distribution with zero mean and the steady-state covariance matrix of $\hat\xi_{\bf k}(i)$.
In Appendix \ref{app_OSDE_dscr}, we derive the components of this
$3\times 3$ matrix, namely, its diagonal elements
(all equal to the steady-state variance), see Eq.(\ref{sch_p3V}),
and the lag-1 and lag-2 covariances, see Eq.(\ref{sch_p3c12}).
\subsection {Computational efficiency}
\label{sec_effic}
In this subsection, we describe two techniques that allow us to significantly decrease
the computational cost of running the spectral SPG.
\subsubsection {Making the time step $\Delta t$ dependent on the spatial wavevector ${\bf k}$}
\label{sec_gamma}
For an ordinary differential equation, the accuracy of a finite-difference scheme
depends on the time step.
More precisely, it depends on the ratio of the time step $\Delta t$ to the temporal length
scale $\tau$ of the process in question.
For high accuracy, $\Delta t \ll \tau$ is needed.
In our problem, $\tau_{\bf k}$ decays with the total scaled wavenumber $k_*$,
see Eqs.(\ref{tauk}), (\ref{cond_PS}), and (\ref{k2_anis}).
This implies that for higher $k_*$, smaller time steps are needed.
To maintain the accuracy across the wavenumber spectrum, we choose the time step
to be a portion of the time scale:
\begin{linenomath*}
\begin {equation}
\label {osde_num4}
(\Delta t)_{\bf k} = \beta \tau_{\bf k}.
\end {equation}
\end{linenomath*}
The less $\beta$, the more accurate
and, at the same time, more time consuming the numerical integration scheme is.
We note that in atmospheric spectra, small scales have, normally,
much less variance (energy) than large scales.
But with the constant $\beta$, the computational time would be, on the contrary, spent predominantly on
high wavenumbers (because the latter require a smaller time step and are much more abundant in 3D or 2D).
So, to save computer time whilst ensuring reasonable overall (i.e.\ for the whole range of wavenumbers) accuracy,
we specify $\beta$ to be wavenumber dependent (growing with the wavenumber) in the following ad-hoc way:
\begin{linenomath*}
\begin {equation}
\label {gamma_k}
\beta_{\bf k} = \beta_{\rm min} + (\beta_{\rm max} - \beta_{\rm min}) \left( \frac{k_*} { \max k_*}\right)^2,
\end {equation}
\end{linenomath*}
where $\beta_{\rm min}$ and $\beta_{\rm max}$ are the tunable parameters.
The choice of the ``optimal'' $\beta_{\rm min}$ and $\beta_{\rm max}$ is discussed just below in section
\ref{sec_coarse}.
\subsubsection {Introduction of a coarse grid in spectral space}
\label{sec_coarse}
Here we propose another technique to reduce the computational cost of the spectral solver.
The technique exploits the {\em smoothness} of the SPG spectrum $b_{\bf k}$ Eq.(\ref{bk_p3}).
This smoothness allows us to introduce a {\em coarse grid in spectral space} and
save a lot of computer time by performing the integration of the time discrete
spectral OSDEs Eq.(\ref{osde_num2}) only for those wavevectors
that belong to the coarse grid. The spectral coefficients $\hat\xi_{\bf k}(i)$ are then
interpolated from the coarse grid to the dense (full) grid in spectral space.
The latter interpolation would introduce correlations between different spectral coefficients $\hat\xi_{\bf k}(i)$,
which would destroy the spatial homogeneity. In order to avoid this, we employ a device used to
generate so-called surrogate time series \citep[][section 2.4.1]{Theiler}.
At each $t$, we multiply the interpolated (i.e.\ dense-grid) $\hat\xi_{\bf k}(i)$
by $\rm e^{\i \theta_{\bf k}}$,
where
$\theta_{\bf k}$
are independent {\em random phases}, i.e.\ independent for different ${\bf k}$ random variables
uniformly distributed on the segment $[0,2\pi]$.
It can be easily seen that this multiplication
removes any correlation between the spectral coefficients.
Note also that the random phase rotation does not destroy the Gaussianity
because $\hat\xi_{\bf k}(i)$ are complex circularly-symmetric random variables
with uniformly distributed and independent of $|\hat\xi_{\bf k}(i)|$ arguments (phases)
\citep[e.g.][section A.1.3]{Tse}.
In order to preserve the temporal correlations, we keep the set of $\theta_{\bf k}$
constant during the SPG-model time integration.
The exact spectrum $b_{\bf k}$ after the trilinear (bilinear in 2D) interpolation
of $\hat\xi_{\bf k}(i)$ from the coarse to the full spectral grid
is imposed in a way similar to that described
in section \ref{sec_spec_correc} as follows.
At any time instance when we wish to compute the physical space field,
for each ${\bf k}$ on the full spectral grid, the linearly interpolated value $\check\xi_{\bf k}$
is the linear combination of the closest coarse-grid points ${\bf k}_j$:
\begin {equation}
\label {trilin}
\check\xi_{\bf k} = \sum_{j=1}^{2^d} w_j \, \hat\xi_{{\bf k}_j}
\end {equation}
where $\check{}$ denotes the interpolated value and
$w_j$ is the interpolation weight (note that the set of the closest coarse-grid points ${\bf k}_j$
depends, obviously, on ${\bf k}$).
In Eq.(\ref {trilin}), the coarse-grid variances
${\mathsf{Var}}\, \hat\xi_{{\bf k}_j}= {b}_{{\bf k}_j}$ are known for all ${\bf k}_j$
from the spectrum $\{b_{\bf k}\}$, see Eqs.(\ref{bk12_p}) or (\ref{bk_p3}). Therefore, we can find
${\mathsf{Var}}\, \check\xi_{\bf k}= \sum_j w_j^2 \, b_{{\bf k}_j}$.
Besides, we know which variance $\check\xi_{\bf k}$ {\em should} have on the fine grid, namely $b_{\bf k}$.
So, we normalize $\check\xi_{\bf k}$ by multiplying it by $\sqrt{b_{\bf k} / ({\mathsf{Var}}\, \check\xi_{\bf k})}$,
thus imposing the exact spatial spectrum for all ${\bf k}$.
Technically, the 3D coarse spectral grid is the direct product of three 1D grids.
Any of the (non-uniform) 1D coarse grids is specified as follows.
Its $j$th point is located at the fine-grid wavenumber $n_j$,
which equals $j$ for $|j| \le n_0$ (where $n_0$ is an integer) and equals the closest integer to
$n_0(1+\varepsilon)^{|j|-n_0}$ for $|j| > n_0$.
Here, $\varepsilon$ is a tunable small positive number.
In the below numerical experiments, the coarse-grid
parameters were $n_0=20$ and $\varepsilon=0.2$, which resulted in the following
positive 1D coarse-grid points:
0
1
2
3
\dots
19
20
24
29
35
42
50
60
72
86
103
124
150
(the 1D grid extent was 300 points and, correspondingly, the maximal wavenumber was 150).
\subsubsection {Numerical acceleration: results}
\label{sec_gamma_coarse}
As the two above acceleration techniques guarantee that the {\em spatial} spectrum is always precise,
we tested how these techniques impacted the {\em temporal} correlations
and what was the speedup.
We performed a numerical experiment with the 2D SPG on the grid
with $300 \times 300$ points, the mesh size $h=7$ km,
and the setup parameters
$\lambda=80$ km, $U=10$ m/s, and $\delta=\gamma=1$.
The time interval $\Delta t_{\rm FFT}$ between the successive backward Fourier transforms determines the effective
resolution of the generated field in time.
To make the temporal resolution consistent with the spatial resolution, we selected $\Delta t_{\rm FFT}$
close to $h/U$, namely, $\Delta t_{\rm FFT}=15$ min.
The computations were performed on a single CPU.
The results are presented in Table \ref{Tab_timings}.
We compared the non-accelerated scheme with the constant $\beta=0.1$ and without the sparse spectral grid (the second row)
and the accelerated scheme with $\beta_{\rm min}=0.15$, $\beta_{\rm max}=3$, and with the sparse spectral grid (the third row).
From column 2, it is seen that
the combined effect of the two numerical acceleration techniques
on the cost of the {\em spectral-space} computations (see column 2)
was dramatic:
the speedup was 66 times as compared to the non-accelerated scheme.
The contributions of the two above numerical acceleration techniques to the spectral-space speedup were comparable in magnitude (not shown).
Most importantly, this spectral-space speedup was achieved at the very little cost:
the temporal length scale $T_{0.5}$
(defined as the time shift at which the correlation function first intersects the 0.5 level) was distorted
by only 4 \% w.r.t.\ the theoretical model (column 6).
Note, however, that the cost of the interpolation from the sparse spectral grid (column 3) and of the discrete backward
Fourier transform (column 4) reduced the total speedup of the 2D SPG to 14 times (see column 5).
\begin{table
\caption{\bf CPU times of 2D SPG computations per 1 h of SPG model time and the relative error in the temporal length scale $T_{0.5}$.}
\begin{center}
\begin{tabular}{|c||c|c|c||c|c|}
\hline
Accelerators & CPU spec. & CPU interp. & CPU FFT & Speedup & Rel.err. $T_{0.5}$\\
\hline
NO & 0.66 & 0 & 0.027 & 1 & 3 \% \\
\hline
YES & 0.010 & 0.012 & 0.027 & 14 &4 \% \\
\hline
\end{tabular}
\end{center}
\small{\em Spec. stands for spectral-space, interp. means interpolation from the sparse spectral grid, and FFT is the fast Fourier transform.}
\label{Tab_timings}
\end{table}
In 3D, the SPG operating on the spatial grid with $300 \times 300 \times 64$ points, took
40-70 times more CPU time as compared to the above 2D case,
with the accuracy being similar to that indicated in Table \ref{Tab_timings} (not shown).
The total speedup was only 8 times due to an increased share of the Fourier transform.
\subsection {Examples of the SPG fields}
Figure \ref{Fig_xi_xy} shows a horizontal $x$-$y$ cross-section and
Fig.\ref{Fig_xi_xt} a spatio-temporal $x$-$t$ cross-section of the pseudo-random
field $\xi(t,x,y)$ simulated by the SPG with the setup parameters indicated in section \ref{sec_gamma_coarse}.
Note that with 300 grid points in each spatial direction, only 256 contiguous grid points are
shown in the Figs.\ref{Fig_xi_xy} and \ref{Fig_xi_xt} and are intended to be used in a mapping to a physical space domain.
This is done in the SPG for practical purposes in order to avoid
correlations between the opposite sides of the spatial domain, which would be spurious in
real-world applications.
\begin{figure
\begin{center}
{ \scalebox{0.5}{ \includegraphics{xy_cross.png}}
}
\end{center}
\caption{Horizontal ($x$-$y$) cross-section of an SPG field.}
\label{Fig_xi_xy}
\end{figure}
\begin{figure
\begin{center}
{ \scalebox{0.5}{ \includegraphics{xt_cross.png}}
}
\end{center}
\caption{Spatio-temporal ($x$-$t$) cross-section of an SPG field.}
\label{Fig_xi_xt}
\end{figure}
\section {Discussion}
\subsection {Physical-space or spectral-space SPG solver?}
In this study, we have investigated both the spectral-space and the physical-space approximations
of the SPG spatio-temporal model. We have found that both approaches can be used to build
a practical SPG scheme. We have selected the spectral-space technique.
Here, we briefly compare both approaches.
Advantages of the spectral-space technique are the following.
\begin {itemize}
\item
Simplicity of realization. If the SPG model has constant coefficients, then the complicated
SPG equation decouples into a series of simple OSDEs.
\item
Straightforward accommodation of non-local-in-physical-space spatial operators.
\end {itemize}
Advantages of the physical-space approach are:
\begin {itemize}
\item
The relative ease of introduction of inhomogeneous (non-stationary) and anisotropic
capabilities to the SPG.
\item
The SPG solver can be implemented in domains with complex boundaries.
\item
Better suitability for an efficient implementation on massively parallel computers.
\end{itemize}
\subsection {Extensions of the SPG}
The proposed SPG technique can be extended in the future along the following lines.
\begin {itemize}
\item
Development of a physical-space solver.
\item
Introduction of advection to the SPG model.
\item
Introduction of spatial inhomogeneity/anisotropy and non-stationarity.
\item
Introduction of non-Gaussianity. This can be done either by applying a nonlinear transform to the output SPG fields,
or by introducing a non-Gaussian driving noise \citep[as in][]{Aberg,Wallin}. The former approach is simpler
but the latter allows for much richer deviations from Gaussianity, including the multi-dimensional aspect.
\item
Going beyond additive and multiplicative perturbations for highly non-Gaussian variables like humidity,
cloud fields, or precipitation.
\item
Simulation of several mutually correlated pseudo-random fields.
\item
Making the temporal order $p$ a user defined variable. As noted above, the larger $p$ the smoother
the generated field.
\end {itemize}
\section {Conclusions}
\begin {itemize}
\item
The proposed Stochastic Pattern Generator (SPG) produces pseudo-random
spatio-temporal Gaussian fields on 2D and 3D limited area spatial domains
with the tunable variance, horizontal, vertical, and temporal length scales.
\item
The SPG model is defined on a standardized domain in space, specifically, on the unit
2D or 3D torus.
Fields on a limited-area geophysical domain in question are obtained by mapping from the standardized domain.
\item
The SPG is based on a linear third-order in time stochastic
model driven by the white in space and time Gaussian noise.
\item
The spatial operator of the stochastic model is built to ensure that solutions to the SPG model,
i.e.\ the generated pseudo-random fields, satisfy the ``proportionality of scales'' property:
large-scale (small-scale) in space field components have large (small) temporal length scales.
\item
Beyond the ``proportionality of scales'', the generated fields possess a number of other nice properties:
\begin{itemize}
\item
The spatio-temporal realizations are (almost surely) continuous.
\item
With the appropriately scaled time and vertical coordinates,
the spatio-temporal fields are isotropic in space-time.
\item
The correlation functions in space-time belong to the Mat\'ern class.
\item
The spatial and temporal correlations have the same shapes.
\end{itemize}
\item
The basic SPG solver is spectral-space based.
\item
Two techniques to accelerate the spectral-space computations are proposed and implemented.
The first technique selects the time step of the spectral-space numerical integration scheme to be
dependent on the wavenumber, so that the discretization error is smaller for more energetic larger
spatial scales and is allowed to be larger for less energetic smaller scales.
The second technique introduces a coarse grid in spectral space.
The combined speedup for spectral-space computations
from both techniques is as large as 40--60 times.
\item
Potential applications of the SPG include ensemble prediction and ensemble data assimilation
in meteorology, oceanography, hydrology, and other areas.
The SPG can be used to generate
spatio-temporal perturbations of the model fields (in the additive or multiplicative or other mode)
and of the boundary conditions.
\item
An application of the SPG
as a source of additive spatio-temporal model error perturbations
to the meteorological COSMO model \citep{Baldauf}
is described
in \citep[][]{TsyrulnikovGayfulin_MZ}.
\end {itemize}
\section*{Acknowledgements}
\addcontentsline{toc}{section}{\numberline{}Acknowledgements}%
The SPG has been developed as part of the Priority Project KENDA
(Kilometre scale Ensemble Data Assimilation) of COSMO.
We have used the discrete fast Fourier package \verb"fft991" developed by C.Temperton at ECMWF in 1978.
\section*{Appendices}
\addcontentsline{toc}{section}{\numberline{}Appendices}%
|
\section{Introduction}
\subsection{Motivation and relevant works}
The studies of the statistical properties of 2-dimensional hyperbolic
systems with singularities are motivated in large part by mathematical
billiards with chaotic behavior, introduced by Sinai in \cite{Sin} and
since then studied extensively by many authors \cite{BSC90, BSC91, Y98,Y99,CM}.
Statistical properties of chaotic dynamical systems are described by
the decay of correlations and by various limiting theorems. Let
$(\mathcal{M},\mathcal{F},\mu)$ be a \emph{dynamical system}, i.e., a measurable transformation
$\mathcal{F}\colon \mathcal{M}\to\mathcal{M}$ preserving a probability measure $\mu$ on the Borel sigma algebra of $\mathcal{M}$.
\nomenclature{$\mathcal{F}$}{The nonuniformly hyperbolic map}
For any real-valued functions $f$ and $g$ on $\mathcal{M}$ (often called
\emph{observables}) the {\em correlations} are defined by
\begin{equation}
C_n(f,g,\mathcal{F}) = \int_{\mathcal{M}} (f\circ \mathcal{F}^n)\, g\, d\mu -
\int_{\mathcal{M}} f\, d\mu \int_{\mathcal{M}} g\, d\mu
\label{Cn}
\end{equation}
Note that (\ref{Cn}) is well defined for all $f,g\in L^2_{\mu}(\mathcal{M})$.
It is a standard fact that $(\mathcal{F},\mu)$ is {\em mixing} if and only if
\begin{equation} \label{Cto0}
\lim_{n\to\infty} C_n(f,g,\mathcal{F}) = 0,
\qquad \forall f,g\in L^2_{\mu}(\mathcal{M})
\end{equation}
The statistical properties of the system $(\mathcal{M},\mathcal{F},\mu)$ are characterized by
the \emph{rate of decay of correlations}, i.e., by the speed of
convergence in (\ref{Cto0}) for ``good enough'' functions $f$ and $g$.
If $\mathcal{M}$ is a manifold and $\mathcal{F}$ is a smooth (or piecewise smooth) map,
then ``good enough'' usually means bounded and (piecewise) H\"older
continuous.
Generally, mixing dynamical systems (even very strongly mixing ones,
such as Bernoulli systems) may exhibit quite different statistical properties, depending on the rate of the
decay of correlations. If correlations decay exponentially fast (i.e.,
$|C_n| =\mathcal{O}( e^{-an})$ with $ a>0$), usually the classical Central Limit
Theorem (CLT) holds, as well as many other probabilistic limit laws,
such as Weak Invariance Principle (convergence to Brownian motion),
which play a crucial role in applications to statistical mechanics; we
refer the reader to the surveys in \cite{Ch95,CD,Den89,D04,Y98} and
\cite[Chapter~7]{CM}. Such strongly chaotic dynamical systems behave
very much like sequences of i.i.d.\ (independent identically
distributed) random variables in probability theory.
However, some other mixing and Bernoulli systems have slow rates of the
decay of correlations, such as $|C_n|=\mathcal{O}( n^{-a})$. Their
statistical properties are usually weak, they exhibit \emph{intermittent} behavior \cite{PM80}: intervals of chaotic motion
are followed by long periods of regular oscillations, etc. Such systems can help to understand the transition from regular to chaotic motion, and for that they have long attracted considerable interest in physics community \cite{FM88,Ma83,VCG}. We note that if\footnote{We say $A_n\sim B_n$ if there exist $0<c_1<c_2$ such that $c_1B_n\leq A_n\leq c_2 B_n$ for all $n\geq 1$.} $|C_n|\sim n^{-a}$ with $a\leq
1$, then even the classical CLT usually fails. In that case the system can be approximated by an unconventional Brownian motion in which the mean squared displacement grows faster than linearly in time. This may help to explain certain unusual physical phenomena, such as superconductivity and superdiffusion. In particular, for $a=1$ the mean squared displacement acquires an extra logarithmic factor \cite{GB,CDlorentz,SV}.
An challenging question to ask is ``What are the main reasons that have slowed down the decay rates of correlations for nonuniformly hyperbolic systems"? It has been a mathematically challenging problem to estimate the rates of the
decay of correlations for hyperbolic systems with singularities,
including chaotic billiards. The main difficulty is caused by
singularities and the resulting fragmentation of phase space during the
dynamics, which slows down the global expansion of unstable manifolds.
Moreover, the differential of the billiard map is unbounded and has
unbounded distortion near the singularities, which aggravates the
analysis of correlations: one has to subdivide the vicinity of singularities into countably many ``shells'' in which distortions can be effectively controlled.
Even for strongly chaotic billiards, exponential upper bounds on
correlations were proven only in 1998 when Young \cite{Y98} introduced
her tower construction as a universal tool for the description of
nonuniformly hyperbolic maps; see also \cite{C99}. Young also
sharpened her estimates on correlations by combining her tower
construction with a coupling technique borrowed from probability
\cite{Y99}. The coupling method was further developed by Bressaud and
Liverani in \cite{BL02} and then reformulated in pure dynamical terms
(without explicit tower construction) by Dolgopyat \cite{CD,D01} using standard
pairs; see also \cite[Chapter~7]{CM}. Dolgopyat's technique was proven
to be efficient in handling various types of strongly chaotic systems with singularities
\cite{CD,CDgalton,CDlorentz,CM,CZ09,D01,D04b}.
For weakly chaotic billiards, the first rigorous upper bounds on
correlations (based on Young's tower) were obtained in the mid-2000s
\cite{M04,CZ05a}. The results upper bounds were not optimal, as they included an
extra logarithmic factor, which was later removed in \cite{CZ08} by a
finer analysis of return time statistics. Besides billiards, the same
general scheme for bounding correlations has been applied to
linked-twist maps \cite{SS} and generalized baker's transformations
\cite{BM}, intermittent symplectic maps \cite{LM}, and solenoids \cite{AP}.
Surprisingly little progress has been made in obtaining \emph{lower}
bounds on correlations for hyperbolic systems, including billiards.
Among rare results in this direction are those for Bunimovich stadia
\cite{GB} and billiards with cusps \cite{BCD}, where a lower bound was
a byproduct of a non-classical CLT that forces correlations to be at
least of order $\mathcal{O}(n^{-1})$. For one dimensional non-uniformly expanding maps and for Markov maps,
lower bounds on correlations have been derived via the renewal methods
by Sarig \cite{Sr}, later improved by Gou\"{e}zel \cite{GO}. The
renewal techniques were then extended to more general non-uniformly
expanding maps and certain nonuniformly hyperbolic systems; see \cite{Hu,HV13,MT14,LT} and references therein. The main
scheme of the renewal methods relies on the construction of an induced
map, for which the corresponding transfer operator has a spectral gap
on a certain functional space. Actually the main reason why it is so
difficult to apply operator technique to billiards and related hyperbolic systems with singularities, is the lack of a suitable
functional space on which the transfer operator for the induced map
would have a spectral gap (and would be aperiodic).
For chaotic billiards
and their perturbations, a suitable Banach space of functions was
constructed in \cite{DZ11,DZ13}, and the spectral gap for
transfer operators was proven to exist. But it is still difficult to apply the renewal operator
methods on these systems because in order to take care of the unbounded differential of
the billiard map, the norms defined in \cite{DZ11,DZ13} cannot directly produce
the necessary estimates needed for the renewal technique \cite{Sr}. We
should however stress that similar difficulties are also encountered in
the studies of non-uniformly expanding (non-invertible) maps, as it
is pointed out in \cite{HV13}. Recently, some progress has been made in this direction for some hyperbolic maps, see \cite{LT,MT14}.
In this paper we are able to identify the main factors that affect the decay rates of correlations for rather general nonuniformly hyperbolic systems, in terms of the tail distribution of the return time function (used in the inducing scheme).
Since the singularities of the systems make it difficult to apply all existing methods, we revisit Dolgopyat's coupling method and the ideas of standard pairs \cite{CD,D01} for systems with exponential decay rates of correlations, see also \cite{Y99}, and develop a new coupling scheme for nonuniformly hyperbolic systems. Combining with the elegant Probability Renewal Theory, originated from Kolmogorov \cite{Kol}, we are able to obtain an optimal bound for the decay rates of correlations for general 2-dimensional hyperbolic systems.
Our formulas give a precise asymptotic, rather than upper or lower
bounds. To our knowledge, this is the first result of that type in the
context of nonuniformly hyperbolic billiards. And the results have greatly improved all existing results on decay rates of correlations for hyperbolic systems with slow decay rates of correlations. Moreover, our new method is more flexible, comparing to all other existing methods in this direction, as it can be applied to dynamical systems under small deterministic or random perturbations. In Section 10, we will describe several classes of billiards to which our results can be applied. Moreover we also obtain the optimal bound for an example of nonuniform link-twist map. Other nonuniform hyperbolic maps, some of which exhibiting attractors, have been investigated in another paper \cite{VZ}.
We conclude this Introduction by quoting forthcoming contributions and by addressing new questions. One of the main difficulty of the inducing approach is to control the rate of expansion and contraction along the invariant manifolds for the original map, being those rates exponential for the first return map. We deal with that issue in the present paper by a careful control of the set $C_{n,b}$, see equation (\ref{Cnb}), whose tails produces very often the (optimal) rate of decay for the lower bound of the correlation function (see our Theorem 2). It turns out that there are a few class of attractors for which the original map exhibit polynomial contraction/expansion along the invariant manifold in the induced region. We will present in a future article an application of our coupling technique to those systems: the advantage with respect the examples presented in this paper, is that the condition (\textbf{H2}), involving the set $C_{n,b}$ is now automatically satisfied, and this allows us to get easily lower bound for the correlation decay and even to improve previous results on upper bounds.
Another extension of our theory will be to consider higher dimensional systems, and this needs a corresponding generalization of the standard pair technique.
We finally quote the problem of lifting a few probabilistic limit theorems from the induced system to the original one; it is well known that problems arise for the induced observable which are not anymore essentially bounded and this requires the use of suitable functional spaces. It is worth mentioning that this not happens for the central limit theorem in our paper since we have a sufficient upper bound for the observables defined on the whole space.
We describe the structure of our work in more details below.
\subsection{Plan of exposition}
We first prove an upper bound on correlations under rather general
assumptions (compared to those in \cite{CZ05a, CZ08}), and then we obtain optimal estimates for the decay of correlations for certain
hyperbolic systems, including semi-dispersing billiards, billiards with cusps, Bunimovich stadia, etc. The key ingredient of the coupling scheme is the construction of a Markov tower, which we
will call {\it generalized Young tower} together with a new version of the coupling lemma for nonuniformly hyperbolic systems. Our
general scheme consists of three major steps:\\
(\textbf{a}) We first need to choose a subset $M \subset \mathcal{M}$ on which the induced (first return) map $F\colon M\to M$ is uniformly hyperbolic (with an exponential decay of correlations). We note that $F$ preserves the measure $\mu$ conditioned on $M$, which we denote by $\mu_M$; it will be an SRB measure. For the definition of SRB measure, see Section \ref{assumption}. \\
(\textbf{b}) Then we check that standard pairs and standard families (which were introduced in \cite{CD,D01,CM}), for the induced system $(F, M, \mu_M)$ satisfy the specific conditions of our earlier work \cite{CZ09}. These would imply a Coupling Lemma for the induced system, as well as an exponential decay of correlations. Based on the Coupling Lemma for the induced system \cite{CZ09}, a generalized Young tower for the induced system is constructed.\\
(\textbf{c}) In order to prove a Coupling Lemma for the original system $(\mathcal{M}, \mathcal{F}, \mu)$, a generalized Markov tower is carefully constructed based on that for the induced map. We decompose the original measure according to a stopping time function and perform coupling only at those stopping times. On the one hand, this procedure allows us to ``match" images of our measures efficiently when they both properly return to a reference set. On the other hand, the probability renewal theory enables us to keep track of points that have properly returned to the base of the tower sufficiently many times, but failed to couple for various reasons. Our procedure can be applied to 2D (nonuniformly) hyperbolic systems with or without singularities.\\
The quantitative part of our scheme involves the following estimates. For the set $M \subset \mathcal{M}$ selected in part (\textbf{a}), let $R \colon \mathcal{M}\to \mathbb{N}$ be the first hitting time to $M$ under iterations of $\mathcal{F}$; see precise definition in (\ref{defR}). The Poincar\'e recurrence theorem implies that $R<\infty$ almost everywhere on $\mathcal{M}$. We obtain the following upper bound on the decay of correlations in Theorem~\ref{main2} for any piecewise H\"older observables $f,g$ on $\mathcal{M}$, any $n>1$:
\begin{equation} \label{uppern}
|C_n(f,g,\mathcal{F})|\leq C_{f,g}\, \mu(R>n),
\end{equation}
where $C_{f,g}>0$ is a constant depending only on $f$ and $g$.
Moreover the following formula is derived in Theorem~\ref{TmMain} for a large class of systems:
\begin{equation}\label{firstorder}
\int_{\mathcal{M}} (f\circ \mathcal{F}^n)\, g\, d\mu -
\int_{\mathcal{M}} f\, d\mu \int_{\mathcal{M}} g\, d\mu= \mu(R>n)\mu(f)\mu(g)+o(\mu(R>n)),
\end{equation}
under the condition that ${\rm supp}(f)\subset M$ and ${\rm supp}(g)\subset M$. This is a novel result for billiard systems. Our asymptotic formula (\ref{firstorder}) implies that the tail bound for the return time function $\mu(R>n)$ is indeed an optimal estimate for the decay of correlations. Note that if one chooses certain observables satisfying conditions in Theorem 2 as well as $\mu(f)\mu(g)=0$, then we get a faster decay rate of correlations:
\begin{equation} \label{fastzero}
|C_n(f,g,\mathcal{F})|= o(\mu(R>n)),
\end{equation}
as it is proved in Theorem 2.
\section{Assumptions and main results}
\subsection{Assumptions}\label{assumption}
Let $\mathcal{M}$ be a 2-dimensional compact Riemannian manifold, possibly with
boundary. Let $\Omega \subset \mathcal{M}$ be an open subset and let $\mathcal{F}\colon
\Omega \to \mathcal{M}$ be a $C^{1+\gamma_0}$ diffeomorphism of $\Omega$ onto
$\mathcal{F}(\Omega)$ (here $\gamma_0\in (0,1]$). We assume that $\mathcal{S}_1 =
\mathcal{M}\setminus\Omega$ is a finite or countable union of smooth compact
curves. Similarly, $\mathcal{S}_{-1} = \mathcal{M}\setminus\mathcal{F}(\Omega)$ is a finite or
countable union of smooth compact curves. If $\mathcal{M}$ has boundary
$\partial \mathcal{M}$, it must be a subset of both $\mathcal{S}_1$ and $\mathcal{S}_{-1}$. We
call $\mathcal{S}_1$ and $\mathcal{S}_{-1}$ the \emph{singularity sets} for the maps $\mathcal{F}$ and
$\mathcal{F}^{-1}$, respectively. We denote by $\Omega_i$, $i\geq 1$, the connected components of
$\Omega$; then $\mathcal{F}(\Omega_i)$ are the connected components of
$\mathcal{F}(\Omega)$. We assume that $\mathcal{F}|_{\Omega}$ is \textbf{time-reversible}, and the restriction of the map $\mathcal{F}$ to any
component $\Omega_i$ can be extended by continuity to its boundary
$\partial \bar{\Omega}_i$, though the extensions to $\partial \bar{\Omega}_i \cap
\partial \bar{\Omega}_j$ for $i\neq j$ need not agree. Similarly, for each $i$ the
restriction of $\mathcal{F}^{-1}$ to any connected component $\mathcal{F}(\Omega_i)$
can be extended by continuity to its boundary $\partial \mathcal{F}(\bar{\Omega}_i)$.
Next we assume that the map $\mathcal{F}$ is (nonuniformly) hyperbolic, as defined by Katok and Strelcyn \cite{KS}. This means that $\mathcal{F}$ preserves
a probability measure $\mu$ such that $\mu$-a.e.\ point $x\in \mathcal{M}$ has
two non-zero Lyapunov exponents: one positive and one negative. Also,
the first and second derivatives of the maps $\mathcal{F}$ and $\mathcal{F}^{-1}$ do
not grow too rapidly near their singularity sets $\mathcal{S}_1$ and $\mathcal{S}_{-1}$,
respectively; and the $\varepsilon$-neighborhood of the singularity set has measure $\mathcal{O}(\varepsilon^{q_0})$ for some $q_0>0$. This is to ensure the existence and absolute continuity of stable
and unstable manifolds at $\mu$-a.e.\ point.
Let $$\mathcal{W}^{u}=\cap_{n\geq 0} \mathcal{F}^n (\mathcal{M}\setminus \mathcal{S}_1).$$ Obviously,
$\mathcal{W}^{u}$ is (mod 0) the union of all unstable manifolds, and we assume
that the partition $\mathcal{W}^{u}$ of $\mathcal{M}$ into unstable manifolds is measurable,
so that $\mu$ induces conditional distributions on $\mu$-almost all
unstable manifolds (see the definition and basic properties of
conditional measures in \cite[Appendix~A]{CM}). Most importantly, we
assume that $\mu$ is an
Sinai-Ruelle-Bowen (SRB) measure; i.e. the conditional distributions of $\mu$ on unstable
manifolds $W\subset \mathcal{W}^{u}$ are absolutely continuous with respect to
the Lebesgue measure on $W$. SRB measures are known to be the only physically observable measures, in the sense that
their basins of attraction have positive Lebesgue volume; see
\cite{Y02} and \cite[Sect.~5.9]{H02}. We also \textbf{assume that our SRB
measure $\mu$ is ergodic and mixing}.
Our work is devoted to statistical properties of $\mu$, so it is
natural for us to take its ergodicity and mixing for granted.
In
chaotic billiards, all the above assumptions are satisfied and are
usually easy to check. In particular, the invariant measure $\mu$ for billiards is smooth and has a
positive density on all of $\Omega$. In physics terms, this invariant measure $\mu$ is an
equilibrium state. Another important class of systems consists of small
perturbations of chaotic billiards (usually induced by external forces
or special boundary conditions) \cite{Ch01,Ch08}. Those systems model
electrical current \cite{CDlorentz,CELS}, heat conduction and viscous
flows \cite{BSp,CL95}, the motion under gravitation on the Galton board
\cite{CDgalton}, etc. For perturbed billiards all the above assumptions
are satisfied, too, but the measure $\mu$ is no longer absolutely
continuous: it is singular with respect to the Lebesgue measure on
$\mathcal{M}$ (though every open subset $U\subset \mathcal{M}$ still has a positive
$\mu$-measure). In physics, such a measure $\mu$ (for billiards under small perturbations) is called a nonequilibrium
steady state (NESS).
Even with the assumptions made thus far, the decay of correlations may be arbitrarily slow
\cite{CZ05a,CZ05b}. In order to ensure a specific rate for the decay of
correlations we introduce the inducing scheme that were used by Markarian \cite{M04} for Bunimovich Stadia, Chernov and Zhang \cite{CZ05a} for general hyperbolic systems with slow decay of correlations.
We first construct a subset $M\subset \mathcal{M}$, with $\mu(M)>0$, and assume that $\partial M\subset \mathcal{S}_1$. Let $R:\mathcal{M}\to\mathbb{N}$ be the first hitting time of $M$. By the Poincare Recurrence Theory, there exists $\hat \mathcal{M}\subset \mathcal{M}$, such that $\mu(\hat\mathcal{M})=1$ and for any $x\in \hat \mathcal{M}$, $R(x)<\infty$. For any $x\in M\cap\hat\mathcal{M}$, we define $Fx=\mathcal{F}^{R(x)}$. Then $F$ preserves a probability measure $\mu_M:=\mu|_{M}/\mu(M)$. For precise construction of the induced system, see Subsection \ref{induce}.
We next make more specific assumptions on the induced system $(M,F,\mu_M)$.
\bigskip
\noindent (\bH1) \textbf{Distribution of the first hitting time $R$.} We assume that there exist $C>0$, ${\alpha_0}>1$, such that the distribution of $R$ satisfies: \begin{equation}\label{R(x)=n} \mu(M_n)\leq \frac{C}{n^{1+{\alpha_0}} }\end{equation} where $M_n$ is the closure of the level set of $R$ restricted on $M$ which is $\{x\in M\,:\, R(x)=n\}$. Moreover, we assume that there exists $N_0\geq 1$ such that every level set $M_n$ contains at most $N_0$ connected components.
\bigskip
\noindent{\textbf{Remark.}}
\begin{itemize}
\item[(i)] \textit{Note that (\ref{R(x)=n}) implies that there exist $C_1>0$, such that $$ \mu(R>n)=\mu(M)\sum_{k\geq n}\mu_{M}(R>k)\leq \frac{C_1 }{ n^{{\alpha_0}-1}}$$
}
\item[(ii)] \textit{In assumption (\textbf{H1}), we can replace (\ref{R(x)=n}) by the following limit:
$$ \mu(M_n)\leq \frac{C}{L(n)n^{1+{\alpha_0}} },$$
where $L(n)$ is a slowly vary function at infinity. Our results on decay rates of correlations still hold by simply adjusting the order of the tail bound $\mu(R>n)$ by a slowly vary function in all estimations.}\end{itemize}
\medskip
For a large $b$ (whose precise value will be given in (\ref{dpbchi})), we denote
$\psi(n):=(b\ln n)^2$, and
define the set
\begin{equation}\label{Cnb}
C_{n,b}=\{x\in\mathcal{M} \,|\, \#_{1 \leq i\leq n} \{\mathcal{F}^i(x) \in M\}\in (1,\psi(n)]\}
\end{equation}
Clearly, $C_{n,b}$ contains those points in $(R< n)$ whose forward trajectory only returns to $M$ at most $\psi(n)$ times within $n$ collisions.\\
\noindent (\bH2) {\textbf{Measure of $C_{n,b}$.}} We assume that $$\mu(C_{n,b}) \leq C n^{1-{\alpha_0}};
\,\,\,\,\,\mu(C_{n,b}\cap M)\leq C n^{-{\alpha_0}}.$$
\medskip
Note that by the time-reversibility, the second condition implies that $$\mu(C_{n,b}\cap \mathcal{F}^{-n}M)\leq C n^{-{\alpha_0}},$$
i.e. in $M$, there are much less points in $C_{n,b}$.
Next we introduce
sufficient conditions for exponential decay rates of correlations, as well as for the coupling lemma, for the induced map. These assumptions are quite standard and have been made in many references \cite{C99,CD,CM,CZ09}.\\
\noindent (\bH3) {\textbf{Sufficient conditions for exponential decay of correlations of the induced map.}}
\begin{itemize}
\item[(\textbf{h1})] \textbf{Hyperbolicity\footnote{We have already assumed that Lyapunov exponents are not zero a.e., but our methods also use stable and unstable cones for the map $F$.} of $F$}. There exist two families of cones
$C^u_x$ (unstable) and $C^s_x$ (stable) in the tangent spaces
${\cal T}_x M$, for all $x\in M\setminus S_1$, and there exists a
constant $\Lambda>1$, with the following properties:
\begin{itemize}
\item[(1)] $D_x F (C^u_x)\subset C^u_{ F x}$ and $D_x F
(C^s_x)\supset C^s_{ F x}$, wherever $D_x F $ exists.
\item[(2)] $\|D_x F(v)\|\geq \Lambda \|v\|, \forall
v\in C_x^u; \quad\text{and}\quad \|D_xF^{-1}(v)\|\geq
\Lambda \|v\|, \forall v\in C_x^s$. \item[(3)] These
families of cones are continuous on $M$
and the angle between $C^u_x$ and $C^s_x$ is uniformly
bounded away from
zero.
\end{itemize}
We say that a smooth curve $W\subset M$ is an unstable (stable) \emph{curve} if at every point $x \in W$ the tangent line
$\mathcal{T}_x W$ belongs in the unstable (stable) cone $C^u_x$ ($C^s_x$).
Furthermore, a curve $W\subset M$ is an unstable (stable) \emph{manifold} if $F^{-n}(W)$ is an unstable (stable) curve for all $n \geq 0$ (resp. $\leq 0$).
\item[(\textbf{h2})] \textbf{Singularities.} The boundary $\partial M$ is transversal to both stable and
unstable cones. Every other smooth curve $W\subset S_1\setminus \partial M$ (resp.
$W\subset S_{-1}\setminus \partial M$ ) is a
stable (resp. unstable) curve. Every curve in $ S_1$
terminates either inside another curve of $ S_1$ or on
the boundary $\partial M$. A similar assumption is made for $S_{-1}$. Moreover, there exist $q_0, s_0\in (0,1]$ and $C>0$ such
that for any $x\in M\setminus S_1$ \begin{equation}\label{upper}
\|D_xF \|\leq C\, \text{\rm dist}(x, S_1)^{-s_0},\end{equation}
and for any $\varepsilon>0$,
\begin{equation}\label{epscs11}\mu\bigl(x\in M\colon \text{\rm dist}(x, S_1)<\varepsilon\bigr)<C\varepsilon^{q_0}.\end{equation}
Note that (\ref{epscs11}) implies that for $\mu$-a.e. $x\in M$, there exists a stable manifold $W^s(x)$ and an unstable manifold $W^{u}(x)$, such that $F^n W^s(x)$ and $F^{-n}W^u(x)$ never hit $ S_1$, for any $n\geq 0$.
\begin{defn}
For every $x, y \in M$, define $\mathbf{s}_+(x,y)$, the forward
\emph{separation time} of $x, y$, to be the smallest integer
$n\geq 0$ such that $x$ and $y$ belong to distinct elements of
$M\setminus S_n$.
Fix $\beta\in(0,1)$, then $d(x,y)=\beta^{\mathbf{s}_+(x,y)}$
defines a metric on $ M$.
Similarly we define the backward separation time $\mathbf{s}_-(x,y)$.
\end{defn}
\item[(\textbf{h3})] \textbf{Regularity of stable/unstable
manifolds}. We assume that the following families of stable/unstable curves, denoted by $\mathcal{W}^{s,u}_F$ are invariant under $F^{-1}$ (resp.$F$) and include all stable/unstable manifolds:
\begin{enumerate}
\item[(1)] \textbf{Bounded curvature.} There exist $B>0$ and $c_M>0$, such that the curvature
of any $W\in \mathcal{W}^{s,u}_F$ is uniformly bounded from above by $B$, and the length of the curve $|W|<c_M$.
\item[(2)] \textbf{Distortion bounds.} There exist $\gamma_0\in (0,1)$ and $C_{\mathbf{r}}>1$ such
that for any unstable curve $W\in \mathcal{W}^{u}_F$ and
any $x, y\in W$,
\begin{equation}
\left|\ln\mathcal{J}}\def\varepsilon{\varepsilon_W (F^{-1}x)-\ln \mathcal{J}}\def\varepsilon{\varepsilon_W (F^{-1}y)\right| \leq C_{\mathbf{r}}\, d (x, y)^{\gamma_0}, \label{distor10}
\end{equation}
where $$\mathcal{J}}\def\varepsilon{\varepsilon_W (F^{-1}x)=dm_{F^{-1}W}(F^{-1}x)/dm_W(x),$$ denotes the Jacobian
of $F^{-1}$ at $x\in W$ with respect to the Lebesgue measure $m_W$ on the unstable curve $W$.
\item[(3)] { \textbf{Absolute continuity.}}
Let $W_1,W_2\in \mathcal{W}^{u}_F$ be two unstable curves close to each other. Denote
\begin{equation*}
W_i'=\{x\in W_i\colon
W^s(x)\cap W_{3-i}\neq\emptyset\}, \hspace{.5cm} i=1,2.
\end{equation*} The map
$\mathbf{h}\colon W_1'\rightarrow W_2'$ defined by sliding along stable
manifolds
is called the \textit{holonomy} map. We assume $\mathbf{h}_*m_{W_1'}$ is absolutely continuous with respect to $m_{W_2'}$, i.e. $ \mathbf{h}_*m_{W_1'}\prec m_{W_2'}$; and furthermore, there exist uniform constants $C_{\mathbf{r}}>0$ and $\vartheta_0\in (0,1)$, such that the Jacobian of $\mathbf{h} $ satisfies
\begin{equation}\label{Jh}
|\ln\mathcal{J}}\def\varepsilon{\varepsilon\mathbf{h}(y)-\ln \mathcal{J}}\def\varepsilon{\varepsilon\mathbf{h}(x)| \leq C_{\mathbf{r}} \text{\rm dist}(W_1', W_2')^{\gamma_0}\,
\vartheta_0^{\mathbf{s}_+(x,y)},
\hspace{1cm}\forall x, y\in W_1' \end{equation}
where $\text{\rm dist}(W_1', W_2'):=\sup_{x\in W_1', y\in W_2'} d(x,y)$.
Similarly, for any $n\geq 1$ we can define the holonomy map
$$
\mathbf{h}_n=F^n\circ \mathbf{h} \circ F^{-n}:F^n W_1\to F^n W_2,
$$
and then (\ref{Jh}) and the uniform hyperbolicity (\textbf{h1}) imply
\begin{equation} \label{cJhn}
\ln \mathcal{J}}\def\varepsilon{\varepsilon\mathbf{h}_n(F^n \mathbf{x})\leq C_{\mathbf{r}} \text{\rm dist}(W_1',W_2')^{\gamma_0} \, \vartheta_0^{n}.
\end{equation}
\end{enumerate}
\item[(\textbf{h4})] {\textbf{ One-step expansion.}}
We have \begin{equation}
\liminf_{\delta\to 0}\
\sup_{W\colon |W|<\delta}\sum_{n: V_n\in FW\setminus S_{1}}
\left(\frac{|W|}{|V_{n
}|}\right)^{q_0}\cdot \frac{|F^{-1}V_{n}|}{|W|}<1,
\label{step1} \end{equation} where the supreme is taken over regular unstable curves $W\subset M$, $|W|$ denotes the length of $W$. \end{itemize}
Note that the boundary $\partial M$ is a part of the singular set $\mathcal{S}_1$, hence every stable manifold for $F$ is also a stable manifold for $\mathcal{F}$. Since we denote by $\mathcal{W}_F^s$ the collection of all stable manifolds for $F$, then the collection $\mathcal{W}^s$ of stable manifolds for $\mathcal{F}$ can be constructed by
\begin{equation}\label{def:cWs}
\mathcal{W}^{s}_*=\cup_{m=1}^{\infty}\cup_{k=0}^{m-1}\mathcal{F}^{k} \{W^s\in \mathcal{W}_F^{s}\colon W^s\subset M_m\}.
\end{equation}
On the other hand, every unstable manifold $W_F^u$ for $F$ is a (part of) an unstable manifold $W^u$ for $\mathcal{F}$, more precisely, $W^u_F = W^u\cap \Omega_i$ for some $\Omega_i \subset M$.
Since we denote by $\mathcal{W}_F^u$ the collection of all unstable manifolds for $F$, then it can be extended to the whole space $\mathcal{M}$ in a similar way:
\begin{equation}\label{def:cWu}
\mathcal{W}^{u}_{\ast}=\cup_{m=1}^{\infty}\cup_{k=0}^{m-1}\mathcal{F}^{-k} \{W^u\in \mathcal{W}_F^{u}\colon W^u\subset F M_m\}.
\end{equation}
Notice that this will not be exactly the collection $\mathcal{W}^u$ of unstable manifolds for $\mathcal{F}$, instead the latter would be obtained by concatenation of some curves from $\mathcal{W}^{u}_{\ast}$. Since the collection of curves in $\mathcal{W}^{u/s}\setminus \mathcal{W}^{u/s}_*$ is a null set, we will not make distinction between these two sets below. More precisely, we will identify $\mathcal{W}^{\sigma}=\mathcal{W}^{\sigma}_*$, for $\sigma\in\{u,s\}$.
\subsection{Statement of the main results}
For any $\gamma\in (0,1)$, we consider those bounded, real-valued functions $f\in L_{\infty}(\mathcal{M},\mu)$ such that, there exists a measurable foliation $\mathcal{W}_f^s$ of $\mathcal{M}$ into stable curves, with the property that for any $x$ and $y$ lying on one stable curve $W\in \mathcal{W}_f^s$,
\begin{equation} \label{sDHC-} |f(x) - f(y)| \leq \|f\|^-_{\gamma,\mathcal{W}^s_f} d (x,y)^{\gamma},\end{equation}
with
$$\|f\|^-_{\gamma,\mathcal{W}^s_f}\colon = \sup_{W\in \mathcal{W}^s_f,}\sup_{ x, y\in W}\frac{|f(x)-f(y)|}{ d (x,y)^{\gamma}}<\infty.$$
Note that there may exist several measurable partitions $\mathcal{W}^s_{f,\alpha}$ of $\mathcal{M}$ into stable curves, such that (\ref{sDHC-}) holds, for $\alpha$ belongs to an index set $\mathcal{A}_f$. We denote these partitions as $$\mathcal{W}^-_f:=\{\mathcal{W}^s_{f,\alpha},\alpha\in \mathcal{A}_f\}.$$
We also require that the stable foliation $\mathcal{W}^s$ of $\mathcal{M}$ belongs $\mathcal{W}^-_f$.
Now we define
$$\|f\|^-_{\gamma}\colon =\sup_{\alpha\in \mathcal{A}_f} \|f\|^-_{\gamma,\mathcal{W}^s_{f,\alpha}}$$
Let $\mathcal{H}^-(\gamma)$ be the collection of all such observables $f$, such that $\|f\|^-_{\gamma}<\infty$. Then for any $f\in \mathcal{H}^-(\gamma)$, any stable curve $W\in \mathcal{W}^-_f$, we have
\begin{equation} \label{DHC-} |f(x) - f(y)| \leq \|f\|^-_{\gamma} d (x,y)^{\gamma}.\end{equation}
Similarly, we define $\mathcal{H}^{+}(\gamma)$ as the set of all bounded, real-valued functions $g\in L_{\infty}(\mathcal{M},\mu)$ such that for any $g\in \mathcal{H}^+(\gamma)$, any unstable curve $W\in \mathcal{W}^+_g$, we have
\begin{equation} \label{DHC+} |g(x) - g(y)| \leq \|g\|^+_{\gamma} d (x,y)^{\gamma}.\end{equation}
Here $\mathcal{W}^+_g=\{\mathcal{W}^u_{g,\alpha},\, \alpha\in \mathcal{A}_g\}$ is the collection of measurable partitions $\mathcal{W}^u_{g,\alpha}$ of $\mathcal{M}$ into unstable curves $\mathcal{W}^u_g$, such that the unstable foliation $\mathcal{W}^u$ of $\mathcal{M}$ belongs $\mathcal{W}^+_g$. Moreover,
for any $x$ and $y$ lying on one unstable curve $W \in \mathcal{W}^{u}_{g,\alpha}$,
\begin{equation} \label{DHC+1}
|g(x) - g(y)| \leq \|g\|^+_{\gamma,\mathcal{W}^u_{g,\alpha}} d (x,y)^{\gamma},
\end{equation}
where
$$
\|g\|^+_{\gamma,\mathcal{W}^u_{g,\alpha}}\colon = \sup_{W\in \mathcal{W}^u_{g,\alpha}}\sup_{ x, y\in W}\frac{|g(x)-g(y)|}{ d (x,y)^{\gamma}}<\infty.
$$
Thus we have that
$$\|g\|^+_{\gamma}\colon =\sup_{\alpha\in \mathcal{A}_g} \|g\|^+_{\gamma,\mathcal{W}^u_{g,\alpha}}<\infty.$$
For every $f\in \mathcal{H}^{\pm}(\gamma)$ we define
\begin{equation} \label{defCgamma}
\|f\|^{\pm}_{C^{\gamma}}\colon=\|f\|_{\infty}+\|f\|^{\pm}_{\gamma}
\end{equation}
In particular, if $f$ is H\"{o}lder continuous on every component of $\mathcal{M}\setminus \mathcal{S}_k$, for some integer $k\in \mathbb{Z}$, with H\"{o}lder exponent $\gamma$, then one can check that $f\in \mathcal{H}^{\pm}(\gamma)$.
By using the coupling methods, we obtain the following upper bounds for the rate of decay of correlations.
\begin{theorem}\label{main2}
For systems satisfy (\bH1)-(\bH3), for any
observables $f\in \mathcal{H}^-(\gamma_f)$ and $g\in\mathcal{H}^+(\gamma_g)$ on $\mathcal{M}$, with $\gamma_f,\gamma_g>0$,
\begin{align*}
|\mu(f\circ \mathcal{F}^n\cdot g)-\mu(f)\mu(g)|&\leq C\|g\|^+_{_{C^{\gamma_g}}} \|f\|^-_{_{C^{\gamma_f}}}(\mu(C_{n/2,b}\cap{\rm supp}(g)\cap {\rm supp}(f\circ F^{n/2}))+\mu(R>n))\\
&\leq C\|g\|^+_{_{C^{\gamma_g}}} \|f\|^-_{_{C^{\gamma_f}}} n^{1-{\alpha_0}},\end{align*}
for $n\geq N$, with $\alpha_0>1$ was given in (\textbf{H1}), $N=N(g^1,g^2)\geq 1$, and $C=C(\gamma_f,\gamma_g)>0$ is a constant.
\end{theorem}
Theorem \ref{main2} implies that the upper bound for the decay rates is determined by both the measure of $(R>n)$ and that of $C_{n,b}\cap{\rm supp}(g)\cap {\rm supp}(f\circ F^n)$. Note that by assumption, $\mu(C_{n,b}\cap M)\leq C n^{-{\alpha_0}}$. This also gives us a hint that by choosing observables $f,g$ that only supported on $M$, one may make the measure of the second set to be of much smaller order than that of $(R>n)$. Next we indeed show that $\mu(R>n)$ characterizes the optimal bound for the decay rates of correlations for observables on $M$.
\begin{theorem} \label{TmMain}
Under conditions of Theorem 1, if we further assume both $f$ and $g$ are supported in $ M$ (note $M$ is the nice subset in $\mathcal{M}$), then
correlations decay as:
\begin{equation}\label{mainh2a}
\mu(f\circ \mathcal{F}^n\cdot g) -
\mu(f)\mu(g) = \mu(R>n)\mu(f)\mu(g)+ E(f,g,n),\end{equation}
for any $n\geq N$, with $$|E(f,g,n)|\leq C \|f\|^-_{C^{\gamma_f}}\|g\|^+_{C^{\gamma_g}}n^{-\beta_0}=o(\mu(R>n)),$$ and $\beta_0=\min\{{\alpha_0},2{\alpha_0}-2\},$ $C=C(\gamma_f,\gamma_g)>0$ is a constant.
\end{theorem}
For the case where $\mu(f)\mu(g)=0$ we also have a better bound $o(\mu(R>n))$ specified in the above Theorem 2. For example, for semi-dispersing billiards on a rectangle and Bunimovich Stadia, we have ${\alpha_0}=2$, so for general
observables, the correlations decay as $ O(n^{-1})$; see also
\cite{M04,CZ05a}. But for observables supported on
$M$ and such that $\mu(f)\mu(g)=0$, the above theorem implies
$ C_n(f,g,\mathcal{F})= O(n^{-2})$, which also implies the classical central limit theorem.
Next we will prove that for dynamical systems with slow decay rates of correlations, the class of H\"{o}lder observables $f$ with support on $M$ will satisfy the classical Central Limit Theorem.
\begin{theorem}\label{MCLXn0} Assume $\beta_0>1$ in Theorem 2. Let $f\in \mathcal{H}^-(\gamma)\cap\mathcal{H}^+(\gamma)$ with $\gamma>0$, ${\rm supp}(f)\subset M$ and $\mu(f)=0$. Assume $f$ is not a coboundary, i.e. there is no function $h$ such that $f=h-h\circ \mathcal{F}$. Then the following sequence converges:
\begin{equation}\label{ClT}\frac{f+\cdots+f\circ \mathcal{F}^n}{\sigma\sqrt{ n }}\xrightarrow{d} Z,\end{equation}
in distribution, as $n\to\infty$.
Here $$ \sigma^2 = \mu(f^2)+2\sum_{n=0}^{\infty}\mu(f\circ \mathcal{F}^n\cdot f)<\infty,$$ and $Z$ is a standard normal variable.
\end{theorem}
According to Theorem 3, for Bunimovich Stadia and billiards with cusps, even though the correlation for general variables usually decay at order $\mathcal{O}(n^{-1})$, if we pick a H\"{o}lder continuous function $f$ supported on $M$, with $f\in H^+(\gamma_0)\cap H^-(\gamma_0)$ and $\mu(f)=0$, then we still be able to get a classical Central limit theorem, instead of the abnormal Central limit theorem.
Another related observable we would like to discuss is $\hat f:=f-\mu(f)$, the centralized version of $f$, with ${\rm supp}(f)\subset M$. Note that we can express $\mu(f)=\mu(f)\mathbf{I}_{\mathcal{M}}$. Thus $\hat f=f-\mu(f)\mathbf{I}_{\mathcal{M}}$ is not supported on $M$ any more, so our Theorem 2 does not apply to it.
Moreover, to study the Central limit theorem for $\hat f\circ \mathcal{F}^n$, we need to define ${\tilde{f}}:=(f-\mu(f)R)\cdot\mathbf{I}_M$, which is the induced observable by $\hat f$. Thus the partial sums$$\tilde{S}_n:=\hat f+\hat f\circ \mathcal{F}+\cdots+\hat f\circ \mathcal{F}^n,$$ and $$S_n={\tilde{f}}+{\tilde{f}}\circ F+\cdots+{\tilde{f}}\circ F,$$ should have similar asymptotic behavior if we scale them by the reciprocal of their standard deviation. However, by the definition of ${\tilde{f}}$, we know that the variance of ${\tilde{f}}$ is dominated from below by that of $R$, which is unbounded for the case when $\mu(R>n)\sim n^{-1}$. This means that the classical Central limit fails, as well as the Green-Kubo formula, as $$\sum_{n=1}^{\infty} \mu(R>n)\sim\sum_{n=1}^{\infty} n^{-1}=\infty.$$
For the Central Limit Theorem of these general classes of observable for Bunimovich Stadia and billiards with cusps, see \cite{BG,BCD}.\\
\\
Throughout the paper we will use the following conventions: Positive and finite global constants whose value is unimportant, will be denoted by $c, c_1,c_2,\cdots$ or $C, C_1, C_2, \cdots$. These letters may denote different values in different equations throughout the paper. In Appendix, we also list some notations that we use throughout the paper.
\section{Standard families and the induced map}
In this paper, we will use the coupling method to prove our main theorems, which depends heavily on the concept of standard pairs proposed firstly by Dolgopyat in \cite{D01}, as well as the $\mathcal{Z}$ function by Chernov and Dolgopyat in \cite{CD,CM}.
\subsection{Construction of an induced system $(F,M,\mu_M)$.}\label{induce}
In this section, we carefully construct the induced system, and take care of the definition of singularity set for both systems.
Let $\mathcal{M}$ be a 2-dimensional compact Riemannian manifold, and $\Omega \subset \mathcal{M}$ be an open subset. $\mathcal{S}_1 =
\mathcal{M}\setminus\Omega$ and $\mathcal{S}_{-1} = \mathcal{M}\setminus\mathcal{F}(\Omega)$ are
the \emph{singularity sets} for the maps $\mathcal{F}$ and
$\mathcal{F}^{-1}$, respectively. Let $\Omega_i$, $i\geq 1$, be the connected components of
$\Omega$; then $\mathcal{F}(\Omega_i)$ are the connected components of
$\mathcal{F}(\Omega)$.
Let $\mathcal{D}=\cup_{i\in I}\Omega_{i}$ (card$\,I<\infty$) be a finite union
of some connected components of $\Omega$. For any $x\in \mathcal{M}$, let
\begin{equation} \label{defR}
R(x)=\min\{n\geq 1\colon \mathcal{F}^n(x) \in \mathcal{D},\ \
\mathcal{F}^m(x)\notin \mathcal{S}_1,\ m=1, \ldots, n-1\},
\end{equation}
be the first entrance time to the set $\mathcal{D}$. When restricted on $\mathcal{D}$, we also call $R$ the first return time. We denote by $\mathcal{N}_1\subset\mathcal{D}$ the set of
points that never return to $\mathcal{D}$ under forward iterations of $\mathcal{F}$; it consists of points of two types:
\begin{itemize}
\item[(i)] $\mathcal{F}^n(x) \in \mathcal{S}_1$ for some $n\geq 1$ and $\mathcal{F}^m(x) \notin \mathcal{D}$
for $m=1,\ldots,n-1$ (the orbit of $x$ hits a singularity before it comes back to $\mathcal{D}$);
\item[(ii)]
$\mathcal{F}^n(x) \in \Omega \setminus \mathcal{D}$ for all $n\geq 1$ (i.e., $x$ is a
wandering point).\end{itemize}
For each $n\geq 1$, the ``level'' set $$\mathcal{D}_n := \{x\in \mathcal{D}\,:\, R(x)=n\} \subset \mathcal{D}$$
is open, and if $\mathcal{D}_n \neq \emptyset$ then $\mathcal{F}^n$ is a diffeomorphism
of $\mathcal{D}_n$ onto $\mathcal{F}^n(\mathcal{D}_n) \subset \mathcal{D}$. We denote by $F$ the first
return map, i.e.,
$$
F(x) = \mathcal{F}^n(x)\qquad \forall x\in \mathcal{D}_n,\quad n\geq 1.
$$
It is easy to see that $F$ is a diffeomorphism of the open set
$\mathcal{D}^{+}=\cup_{n\geq 1} \mathcal{D}_n$ onto the open set $\mathcal{D}^-=\cup_{n\geq 1}
\mathcal{F}^n(\mathcal{D}_n)$. The inverse map $F^{-1}$ is defined on $\mathcal{D}^- \subset
\mathcal{D}$ and takes it back to $\mathcal{D}^+$. Let $M=\overline{\mathcal{D}}$ denote the
closure of $\mathcal{D}$, and for each $n\geq 1$ let $M_n = \overline{\mathcal{D}_n}$.
We put
$$
S_1 = M\setminus\mathcal{D}^+ = \mathcal{N}_1\cap\partial \mathcal{D}
$$
and
$$
S_{-1} = M\setminus\mathcal{D}^- = \mathcal{N}_{-1}\cap\partial \mathcal{D},
$$
where $\mathcal{N}_{-1}\subset \mathcal{D}$ denotes the set of points never coming back
to $\mathcal{D}$ under the iterations of $\mathcal{F}^{-1}$. We assume that both $S_1$
and $S_{-1}$ are finite or countable unions of smooth compact curves.
The sets $S_{\pm 1}$ play the role of singularities for the induced
maps $F^{\pm 1}$. We assume that the map $F$ restricted to any level
set $\mathcal{D}_n$ can be extended by continuity to its boundary $\partial
\mathcal{D}_n$, but the extensions to $\partial \mathcal{D}_n \cap \partial \mathcal{D}_m$ for
$n\neq m$ need not agree. A similar assumption is made for $F^{-1}$.
We assume that $\mu(\mathcal{D})>0$. It is easy to show \cite{CK} that the SRB
measure $\mu$ cannot be concentrated on curves, i.e., $\mu(W)=0$ for
any smooth curve $W\subset \mathcal{M}$. Thus all our singularity sets
$\mathcal{S}_{\pm 1}$, $S_{\pm 1}$, and their images under $\mathcal{F}^n$, $n\in{\mathbb Z}$, are
null sets. By the ergodicity of $\mu$ we have
\begin{equation} \label{cMMm}
\mathcal{M} = \bigcup_{n\geq 1}\bigcup_{m=0}^{n-1} \mathcal{F}^m \mathcal{D}_n=\bigcup_{m\geq 0}\bigcup_{n=m+1}^{\infty}\mathcal{F}^m \mathcal{D}_n=\bigcup_{m\geq 1}(R=m)\qquad {\rm (mod\ 0)},
\end{equation}
where $$(R=m)=\bigcup_{n=m}^{\infty}\mathcal{F}^{n-m} \mathcal{D}_{n} $$ is the $m$-th level set of $R$ in $\mathcal{M}$.
We note that $\int_M R\, d\mu=1$ by the Kac theorem. The first return
map $F$ preserves the measure $\mu$ conditioned on $M$; we denote it by $\mu_M$.
Clearly, this measure is ergodic. We also \textbf{assume that $\mu_M$ is mixing}. Unstable
manifolds $W\subset M$ are the unstable manifolds for the induced map
$F$ intersected with $M$, hence $\mu_M$ is an SRB measure.
For $n\geq 1$, let
$$
S_{n}=\bigcup_{i=0}^{n-1} F^{-i} S_1 \,\,\,\,\text{ and }\,\,\,\,\, S_{-n}=\bigcup_{i=0}^{n-1} F^i S_{-1},
$$
for each $n \geq 1$. Then the map $F^n\colon M\setminus S_{n}\to
M\setminus S_{-n}$ is a $C^{1+\gamma_0}$ diffeomorphism.\\
\subsection{Standard families}
For any unstable manifold $W\in \mathcal{W}^{u}$, let $\mu_{W}$ be the probability measure on $W$ determined by the unique probability density $\rho_{W}$ (with respect to the Lebesgue measure $m_W$) satisfying
\begin{equation}\label{dens}\frac{\rho_{W}(y)}{ \rho_{W}(x)} =
\lim_{n\rightarrow \infty}\frac{\mathcal{J}}\def\varepsilon{\varepsilon_{W}(F^{-n}y)} {\mathcal{J}}\def\varepsilon{\varepsilon_{W}(F^{-n}x)}.\end{equation}
Here $\rho_W=d\mu_W/d m_W$ is called {\it{the u-SRB
density on $W$}}, and the corresponding probability measure $\mu_{W}$
on ${W}$ is
called {\it{the u-SRB measure of $F$}}. The formula (\ref{dens}) is standard in ergodic
theory, see \cite{CM} page 105.
(\ref{dens}) implies that for any $m\geq 1$, and $x,y\in W$,
$$\frac{\rho_W(y) }{\rho_W(x)}=\frac{\rho_{F^{-m}W}(F^{-m}y)\mathcal{J}}\def\varepsilon{\varepsilon_W(F^{-m} y)}{\rho_{F^{-m}W}(F^{-m}x)\mathcal{J}}\def\varepsilon{\varepsilon_W(F^{-m}x)},$$
which is equivalent to
$$\frac{\rho_W(y) }{\rho_{F^{-m}W}(F^{-m}y)\mathcal{J}}\def\varepsilon{\varepsilon_W(\mathcal{F}^{-m} y)}=\frac{\rho_W(x)}{\rho_{F^{-m}W}(F^{-m}x)\mathcal{J}}\def\varepsilon{\varepsilon_W(F^{-m}x)}=c^{-1}(m,W,F),$$
for some constant $c=c(m,W,F)>0$. This implies that for any $x\in W$,
\begin{equation}\label{cmW}\rho_{F^{-m}W}(F^{-m}x)\mathcal{J}}\def\varepsilon{\varepsilon_W(F^{-m} x)=c(m,W,F)\rho_W(x). \end{equation}
To determine $c(m,W,F)$, we use the fact that both $\rho_W$ and $\rho_{F^{-m}W}$ are probability densities.
\begin{align*}
&c(m,W,F)=c(m,W,F)\int_W\rho_W(y) dm_W(y)\\
&=\int_{W}\rho_{F^{-m}W}(F^{-m}y) \mathcal{J}}\def\varepsilon{\varepsilon_W(F^{-m} y)\, dm_{W}( y)\\
&=\int_{F^{-m}W}\rho_{F^{-m}W}(x)\mathcal{J}}\def\varepsilon{\varepsilon_W(x)\frac{d m_{W}(F^m x)}{dm_{F^{-m}W}(x)}\, dm_{F^{-m}W}(x)\\
&=\int_{F^{-m}W}\rho_{F^{-m}W}(x)\, dm_{F^{-m}W}(x)=1,\end{align*}
which implies that $c(m,W,F)=1$.
Similarly, we have that for any $x\in W$, \begin{equation}\label{cmWF}\rho_{\mathcal{F}^{-m}W}(\mathcal{F}^{-m}x)\mathcal{J}}\def\varepsilon{\varepsilon_W(\mathcal{F}^{-m} x)=\rho_W(x),\,\,\,\,\,\,\,\rho_{F^{-m}W}(F^{-m}x)\mathcal{J}}\def\varepsilon{\varepsilon_W(F^{-m} x)=\rho_W(x). \end{equation}
Furthermore, it follows from the distortion bound (\ref{distor10}) that $\rho_W\sim |W|^{-1}$ on $W$. More precisely, we have
\begin{equation}\label{rhoWbd}
\frac{1}{|W|} e^{-C_{\mathbf{r}}|W|^{\gamma_0}}\leq \rho_W(x)\leq \frac{1}{|W|} e^{C_{\mathbf{r}}|W|^{\gamma_0}}.
\end{equation}
\begin{defn}[Standard pair]
A probability
measure $\nu$ supported on an unstable manifold $W$ is
called {\it{regular}}, if $\nu$ is absolutely continuous with respect to
the $u$-SRB measure $\mu_W$ on $W$, and the probability density function $g=d\nu/d \mu_W\in \mathcal{H}^+(\gamma_0)$ satisfies \begin{equation}\label{standardpair}
| \ln g(x) - \ln g(y)|
\leq C_{F} d(x,y)^{\gamma_0}, \end{equation} where $C_F$ is a fixed constant. In this case $(W, \nu)$ is
called a \emph{standard pair}. Moreover, if the probability density $g=d\nu/d \mu_W$ satisfies
\begin{equation}\label{sstandardpair}
| \ln g(x) - \ln g(y)|
\leq \Lambda^{\gamma_0} C_{F} d(x,y)^{\gamma_0}, \end{equation}
we call $(W,\nu)$ a pseudo-standard pair.
\end{defn}
Here we also need the concept of pseudo-standard pair, as in the proof of the Coupling Lemma, one needs to subtract a smooth function from the density of a standard pair. But the resulting conditional measure will only induce a pseudo-standard pair, as we will show in lemma \ref{defnN}.
\begin{defn}
Let $\mathcal{G}=\{(W_{\alpha}, \nu_\alpha), \alpha\in \mathcal{A}\}$ be a family of (pseudo-) standard pairs equipped with a factor measure $\lambda$ on the index set $\mathcal{A}$. We call $\mathcal{G}$ a
{(pseudo-) standard family} on $\mathcal{M}$, if it satisfies the following conditions:
\begin{enumerate}
\item[(i)] $\{W_{\alpha}, \alpha\in \mathcal{A}\}$ is a measurable partition of a measurable subset $\mathcal{W}\subset \mathcal{M}$ into unstable manifolds;
\item[(ii)] There is a Borel measure $\nu$ satisfying:
\begin{equation}\label{muMdisintegrate}
\nu(B)=\int_{\alpha\in\mathcal{A}} \nu_{\alpha}(B\cap
W_{\alpha})\,
\lambda(d\alpha),
\end{equation}
for any measurable set $B\subset \mathcal{M}$.
\end{enumerate}
For simplicity, we denote such a family by
$$\mathcal{G} = (\mathcal{W}, \nu)=\{(W_{\alpha},\nu_{\alpha}), \alpha\in \mathcal{A}, \lambda\}.$$
\end{defn}
For any $\gamma\in (0,1)$, any $f\in \mathcal{H}^-(\gamma)$, we define
\begin{equation}\label{cGphi}\mathcal{G}(f):=\int_{\alpha\in \mathcal{A}} \int_{W_{\alpha}} f(x)\, d\nu_{\alpha} \, \lambda(d\alpha).\end{equation}
Then one can check that any standard family $\mathcal{G}$ is a bounded linear functional (or generalized function) on the space of test functions $\cup_{0<\gamma<1}\mathcal{H}^-(\gamma)$.
To understand the distribution of short unstable manifolds in any standard family, we define a characteristic function $\mathcal{Z}$ on these standard families, such that
for any standard family $\mathcal{G}=(\mathcal{W},\nu)$,
\begin{equation}\label{cZ}
\mathcal{Z}(\mathcal{G})=\frac{1}{\nu(\mathcal{M})}\,\int_{\mathcal{A}}|W_{\alpha}|^{-q_0}\,\lambda(d\alpha), \end{equation}
where $q_0\in(0,1]$ was defined in (\ref{upper}).
We denote $\mathfrak{F}(\mathcal{M})$ ($\mathfrak{F}(M)$) as the collection of all standard families $\mathcal{G}$ on $\mathcal{M}$ (resp. $M$), such that $\mathcal{Z}(\mathcal{G})<\infty$. One can check that both sets are closed under positive scalar multiplications.
We fix a large number $C_{\rm q}>100 C_F$, whose value will be chosen in (\ref{delta0}). Given a standard family
$\mathcal{G}$, if $\mathcal{Z}(\mathcal{G})<C_{\rm q}$ we say $\mathcal{G}$ is a \emph{proper} (standard)
family.
Moreover, since a standard family can be viewed as a weighted sum of standard pairs, then for any sequence of nonnegative numbers $\{a_i, i\geq 1\}$ collection of standard families $\mathcal{G}_i=(\mathcal{W}_i,\nu_i)\in \mathfrak{F}(\mathcal{M})$, $i\geq 1$, the following sum is well-defined, $\sum_{i=1}^{\infty}\mathcal{G}_i:=(\mathcal{W},\nu)$, with
\begin{equation}\label{directsum}\mathcal{W}=\bigcup_{i=1}^{\infty} \mathcal{W}_i,\,\,\,\text{ and }\,\,\,\nu= \sum_{i=1}^{\infty}a_i \nu_i.\end{equation} It is also a standard family in $\mathfrak{F}(\mathcal{M})$, as long as $\{\mathcal{W}_1,\cdots,\mathcal{W}_n, \cdots\}$ are mutually disjoint, and $\sum_{i=1}^{\infty}a_i\nu_i(\mathcal{M})<\infty$.
In this paper, we always take the u-SRB measure $\mu_{\alpha}:=\mu_{W_{\alpha}}$ to be the reference measure on indexed unstable curve $W_{\alpha}$, sometimes we also denote it as $\mu_W$ for general unstable curve $W$. Next we will show that the set $\mathfrak{F}(\mathcal{M})$ is invariant under $\mathcal{F}$, and $\mathfrak{F}(M)$ is invariant under $F$.
\begin{comment}
Indeed since our standard families $(\mathcal{W},\nu)$ are defined only on unstable manifolds, so it is uniquely determined by its density $g\in \mathcal{H}^+(\gamma_0)$ (together with the measurable partition $\mathcal{W}^{u})$. More precisely, by comparing with the disintegration formula of measure $\mu$ using definition (\ref{muMdisintegrate}), we consider any standard family $(\mathcal{W},\nu)$ and the density $g:=d\nu/d\mu\in \mathcal{H}^+(\gamma_0)$, with $\mathcal{W}=\{W_{\alpha}, \alpha\in \mathcal{A}\}$. Since $\mathcal{A}\subset \mathcal{A}^u_{\mathcal{M}}$, so for any measurable set $A\subset \mathcal{M}$,
\begin{align}\label{gnustandard}\nu(A)&=\int_A g\,d\mu=\int_{\alpha\in \mathcal{A}^u_{\mathcal{M}}} \int_{W_{\alpha}\cap A} g\,d\mu_{\alpha}\,\lambda^u_{\mathcal{M}}(d\alpha)\nonumber\\
&=\int_{\alpha\in \mathcal{A}^u}\int_{W_{\alpha}\cap A} d\nu_{\alpha}\,\lambda(d\alpha)\end{align}
where $$d\nu_{\alpha}=g_{\alpha}\,d\mu_{\alpha},\,\,\,\,\,{ and }\,\,\,\,\, \lambda(d\alpha)=\mu_{\alpha}(g)\lambda^u_{\mathcal{M}}(d\alpha)$$ Here $g_{\alpha}$ is a function defined only on $W_{\alpha}$ such that for any $x\in W_{\alpha}$, $ g_{\alpha}(x)=g(x)/\mu_{\alpha}(g)$. Clearly, $\lambda(\mathcal{A})=\nu(\mathcal{M})=\mu(g)$.
Thus we have extended the family from $\mathcal{W}$ to $\mathcal{W}^{u}$ uniquely, by assigning zero factor measures on unstable manifold in $\mathcal{W}^{u}\setminus \mathcal{W}$. Consequently, we can say that the standard family $(\mathcal{W},\nu)$ is (almost surely) uniquely determined by the density function $g$, which can be decomposed along each unstable manifold, such that for any $x\in W_{\alpha}$:
\begin{equation}\label{galpha}
g(x)=g_{\alpha}(x)\mu_{\alpha}(g) \mathbf{I}_{W_{\alpha}}(x).\end{equation}
On the other hand, if $d\nu/d\mu=g$, then (\ref{gnustandard}) implies that there exists $\mathcal{A}\subset\mathcal{A}^u_{\mathcal{M}}$, such that for any $\alpha\in \mathcal{A}$, $\mu_{\alpha}(g)>0$; and any $\alpha\in \mathcal{A}^u_{\mathcal{M}}\setminus\mathcal{A}$, $\mu_{\alpha}(g)=0$. Thus we define $\mathcal{W}=\{W_{\alpha},\alpha\in \mathcal{A}\}$, then $(\mathcal{W},\nu)$ is the standard family defined by $\nu$ (or equivalently by density $g$). We call $g$ the {\it{associated density function of family $(\mathcal{W},\nu)$}}; and $(\mathcal{W},\nu)$ the {\it{ standard family of density function $g$.}}
In particular, for any $m\geq 1$, we define the index set $\mathcal{A}^u_m$ as all unstable manifolds of $F$ in the level set $\mathcal{D}_m$. More precisely,
\begin{equation}\label{DmW}\{W_{\alpha}\,:\,\alpha\in \mathcal{A}^u_m\}=\{W_{\alpha}\subset \mathcal{D}_m\,:\, \alpha\in \mathcal{A}^u_M\}.\end{equation}
One can check that $$\lambda^u_M(\mathcal{A}_m)=\mu_M(\mathcal{D}_m)=\mu(\mathcal{D}_m)/\mu(M).$$
We denote $(\mathcal{W}^u, \mu)|_{\mathcal{D}_m}$ as the standard family obtained by restricting $(\mathcal{W}^u,\mu)$ on $\mathcal{D}_m$. We will use it to show that
$\mathfrak{F}(\mathcal{M})$ is also invariant under $\mathcal{F}$.
\end{comment}
\begin{proposition}\label{SFinvariant}
For any $\mathcal{G}\in \mathfrak{F}(\mathcal{M})$, then $\mathcal{F}\mathcal{G}$ is a standard family. Similarly, for any $\mathcal{G}\in \mathfrak{F}(M)$, then $F\mathcal{G}$ is a standard family. Moreover, $(\mathcal{W}^u,\mu)\in\mathfrak{F}(\mathcal{M})$, and $(\mathcal{W}^u_F,\mu_M)\in\mathfrak{F}(M)$ are both proper families. \end{proposition}
To prove this proposition, we first need the following lemma.
\begin{lemma}\label{regular1} Let $(W, \nu)$ be a standard pair.
Then both
$\mathcal{F}^k(W, \nu)$ and $F^k(W,\nu)$ are standard families, for any $k\geq 1$. \end{lemma}
\begin{proof} Let $(W, \nu)$ be any standard pair, any $k\geq 1$. We first consider the iterations under $\mathcal{F}$. Assume $\mathcal{F}^k W=\{(V_{\alpha},\nu_{\alpha}),\alpha\in \mathcal{A}, \lambda\}$.
We denote $W_{\alpha}=\mathcal{F}^{-k}V_{\alpha}$, for any index $\alpha\in \mathcal{A}$.
Using (\ref{cmWF}) we know that for any $\alpha\in \mathcal{A}$, any $x\in V_{\alpha}$,
\begin{equation}\label{ccFkW}\frac {d\mathcal{F}_*^k \mu_{W_{\alpha}}(x)}{d \mu_{V_{\alpha}}(x)}=1.\end{equation}
We denote $g=d\nu/d\mu_W$, then for any $k\geq 1$, the density function $g_k$ of $\mathcal{F}^k_*\nu$ can be written as
\begin{align}\label{dcFknu}
g_k(x)&=\frac{d\mathcal{F}_*^k\nu_{W_{\alpha}}(x)}{d\mu_{V_{\alpha}}(x)}=
\frac{d\nu_{W_{\alpha}}(\mathcal{F}^{-k}x)}{d \mu_{W_{\alpha}}(\mathcal{F}^{-k} x)}\cdot \frac{d \mu_{W_{\alpha}}( \mathcal{F}^{-k}x)}{d \mu_{V_{\alpha}}( x)}\nonumber\\
& =\frac{d\nu_{W_{\alpha}}(\mathcal{F}^{-k}x)}{d \mu_{W_{\alpha}}(\mathcal{F}^{-k} x)}\cdot \frac{d \mathcal{F}^k_*\mu_{W_{\alpha}}(x)}{d \mu_{V_{\alpha}}( x)}=\frac{d\nu_{W_{\alpha}}(\mathcal{F}^{-k}x)}{d \mu_{W_{\alpha}}(\mathcal{F}^{-k} x)}\cdot \frac{d \mu_{V_{\alpha}}(x)}{d \mu_{V_{\alpha}}( x)}\nonumber\\
&= \frac{d\nu_{W}(\mathcal{F}^{-k}x)}{d \mu_{W}(\mathcal{F}^{-k} x)}\cdot \frac{\nu(W_{\alpha})}{\mu_W(W_{\alpha})}= g(\mathcal{F}^{-k}(x))\cdot \frac{\nu(W_{\alpha})}{\mu_W(W_{\alpha})},\end{align}
for all $ x\in V_{\alpha}$, here $\nu_{W_{\alpha}}$ is the conditional measure of $\nu$ restricted on $W_{\alpha}$.
Thus for any $x, y$ belong to one smooth component of
$\mathcal{F}^k(W)$, $k\geq 1$,
\begin{align*}|
\ln g_{k}(x)&- \ln g_{k}(y)|= |\ln g( \mathcal{F}^{-k}x)- \ln g(
\mathcal{F}^{-k}y)
\\
&\leq
C_F d( \mathcal{F}^{-k}x, \mathcal{F}^{-k}y)^{\gamma_0}\leq C_F d( x, y)^{\gamma_0}.
\end{align*}
Thus we can see that
$(\mathcal{F}^k W,\mathcal{F}^k_* \nu)$ is a standard family.
Similarly, one can show that $F(W,\nu)$ is a standard family.
\end{proof}
Next we return to the proof of Proposition \ref{SFinvariant}. For any standard family $\mathcal{G}=(\mathcal{W},\nu)$, since the $\mathcal{F}$ (resp. $F$) image of every standard pair in $\mathcal{G}$ is a standard family, thus $\mathcal{F}\mathcal{G}$ (resp. $F\mathcal{G}$) is a standard family. Moreover, the set $\mathcal{W}^{u}_{F}$ can be viewed as the measurable partition of $M$ into unstable manifolds. It was proved in \cite{CZ09} Lemma 12 that $(\mathcal{W}^{u}_{F},\mu_M)$ is indeed a proper standard family. Let $\mathcal{A}^u_M$ be the index set of unstable manifolds in $\mathcal{W}^u_F$, such that $$\mathcal{W}^u_F=\{W_{\alpha}\,:\, \alpha\in \mathcal{A}^u_M\}.$$
Then $\mu_M$ induces a factor measure $\lambda^u_{M}$ on the sigma algebra (induced by the Borel sigma algebra on $M$) of the index set $\mathcal{A}^u_{M}$, such that
\begin{equation}\label{standarcWM} (\mathcal{W}^u_F, \mu_M)=\{(W_{\alpha},\mu_{\alpha}), \alpha\in \mathcal{A}^u_M, \lambda^u_M\}.\end{equation}
Now it follows from the invariance of $\mu_M$ under $F$, that $(\mathcal{W}^u_F,\mu_M)\in \mathfrak{F}(M)$. Similarly, $(\mathcal{W}^u, \mu)\in \mathfrak{F}(\mathcal{M})$. The fact that $(\mathcal{W}^u, \mu)$ is a proper family follows from the fact that the singular set of $\mathcal{F}$ is a subset of that of $F$, thus (\ref{epscs11}) implies that\begin{equation}\label{epscs111}\mu\bigl(x\in \mathcal{M}\colon \text{\rm dist}(x, \mathcal{S}_1)<\varepsilon\bigr)<C\varepsilon^{q_0}.\end{equation}
\qed
\bigskip
It was shown above that $\mathfrak{F}(\mathcal{M})$ is invariant under $\mathcal{F}$.
More precisely, if $\mathcal{G}=(\mathcal{W}, \nu)$ is
a standard family with a factor measure $\lambda$, such that (\ref{muMdisintegrate}) holds, then $\mathcal{F}^n_*\nu$ induces a standard family with
$\mathcal{F}^n\mathcal{G}=\mathcal{F}^n(\mathcal{W},\nu):=(\mathcal{F}^n \mathcal{W}, \mathcal{F}^n_*\nu)$. Here, we denote \begin{equation}\label{FnG}\mathcal{F}^n (\mathcal{W},\nu)=\{(V_{\alpha},\nu_{\alpha})\,,\,\alpha\in \mathcal{A}^n, \lambda_n\}\end{equation} as the standard family with factor measure $\lambda_n$ on the index set $\mathcal{A}^n$ of unstable manifolds in $\mathcal{F}^n\mathcal{W}$.
We define \begin{equation}\label{vartheta1}\vartheta_1=\max\{\vartheta_0, \Lambda^{-\gamma_0}\}\end{equation}
where $\vartheta_0$ and $\Lambda$ was defined as in \textbf{(h3)}.
\begin{lemma}\label{densitybd} There exists $\delta_0>0$, such that for any standard pair $(W,\nu_W)$, with $|W|\leq 20\delta_0$, and $g(x)=d\nu_W/d\mu_W$, then the density $g$ satisfies:
$$|g(x)-1|\leq \varepsilon_{\mathbf{d}}$$
where $\varepsilon_{\mathbf{d}}<10^{-5}$, whose value will be chosen in (\ref{epsbdchoose}).
\end{lemma}
\begin{proof}
We choose $C_{\rm q}$ large enough, $\varepsilon_{\mathbf{d}}<10^{-5}$ and $\delta_0$ small enough to satisfy
\begin{equation}\label{delta0}
\frac{1}{C_{\rm q}}<\delta_0^{q_0},\,\,\,\,\,\,\text{ and }\,\,\,\,\, \frac{1}{C_{\rm q}}< (20\delta_0)^{\gamma_0}<\frac{\varepsilon_{\mathbf{d}}}{C_F}.
\end{equation}
Combining with (\ref{standardpair}), if one makes a standard pair
$(W,\nu_W)$ with density function $g\in\mathcal{H}^+(\gamma_0)$ with respect to the u-SRB measure $\mu_W$ on $W$, then (\ref{delta0}) implies that for any $x,y\in W$,
\begin{equation}\label{densitypair}
|\ln g(x)- \ln g(y)|\leq C_F(20\delta_0)^{\gamma_0}<\varepsilon_{\mathbf{d}}.
\end{equation}
Note that $\mu_W$ is a probability measure, with H\"{o}lder continuous probability density $\rho_W=d\mu_W/d m_W$. Thus there exists $x_0\in W$, such that $\rho_W(x_0)=|W|^{-1}$, which is the average value of $\rho_W$.
Since $g(x)=d\nu_W/d \mu_W$ is also a probability density, there exists $x_0\in W$, such that $g(x_0)=1$. This implies that for any $x\in W$,
\begin{equation}\label{densitypair1}| g(x)- 1|\leq \varepsilon_{\mathbf{d}}.\end{equation}
Thus for any standard pair with length
$|W|<20\delta_0$, the density function $g$ is bounded by $$1-\varepsilon_{\mathbf{d}}\leq g(x)\leq 1+\varepsilon_{\mathbf{d}}.$$
\end{proof}
For any $n\geq 1$, any $x\in M$, such that $W^u(x)$ exists, we define $r_n(x)$ as the minimal distance between $F^n x$ and the two end points of $W^u(F^n x)$. In particular, the following facts were proved in \cite{CZ09}. \begin{lemma}\label{properagain}
The following statements hold:\\
(1) There exists a uniform constant $\chi>0$, such that for any standard pair $(W,\nu)$, $F^n(W,\nu)$ is a proper family for any $n>\chi\ln |W|$;\\
(2) Let $\mathcal{G}=(\mathcal{W},\nu)$ be a standard family, with $\mathcal{Z}(\mathcal{G})<\infty$, then there exists $N>1$, such that $F^N\mathcal{G}$ is a proper family;\\
(3) For any $x\in M$, let $\mathbf{r}^{s/u}(x)$ be the minimal distance of $x$ to the boundary points of $W^{s/u}(x)$ measured along $W^{s/u}(x)$. Then any standard pair $(W,\nu)$ with length $|W|>\delta_0$ is proper; and there exists $C>0$ such that for any stable/unstable manifold $W^{s/u}$ with length $|W^{s/u}|>\delta_0$, we have
\begin{equation}\label{delta00}m_{W^{s}}(\mathbf{r}^u(x)<\varepsilon)<C \varepsilon^{q_0},\,\quad m_{W^u} (\mathbf{r}^s(x)<\varepsilon)<C \varepsilon^{q_0};
\end{equation}
\noindent(4) There exists $C>0$, such that for any standard family $\mathcal{G}=((W_{\alpha},\alpha\in \mathcal{A}),\nu)$, any $\varepsilon\in (0,1)$,
\begin{equation}\label{epsq}\nu(x\in W_{\alpha}\,:\, \mathbf{r}^u(x)<\varepsilon, \alpha\in \mathcal{A})<C\mathcal{Z}(\mathcal{G}) \varepsilon^{q_0},\,\,\,\,\,\,\,\,\nu(x\in W_{\alpha}\,:\, \mathbf{r}_n(x)<\varepsilon, \alpha\in \mathcal{A})<C\mathcal{Z}(F^n\mathcal{G}) \varepsilon^{q_0}. \end{equation}
\noindent(5) There exist constants $c>0$, $C_z>0$, and $\vartheta_3\in (0,1)$, such
that for any standard family $\mathcal{G}=(\mathcal{W},\nu)$ supported in $M$, if $\mathcal{Z}(\mathcal{G})<\infty$, then for any $n\geq 1$,
\begin{equation}\label{firstgrowth} \mathcal{Z}
(F^n \mathcal{G})\leq c\vartheta_3^{n}\mathcal{Z} (\mathcal{G})+C_z;
\end{equation}
\begin{equation}\label{firstgrowth1}
F^n_*\nu(r^{u/s}<\varepsilon)\leq c\,\vartheta_3^{n}
\nu(r^{u/s}<\varepsilon)+C_z\varepsilon^{q_0}.
\end{equation}
\end{lemma}
\noindent{\textbf{Remark.}} Note that the first equation in (\ref{delta00}) is the time-reversal version of the second equation, which follows from our assumption \textbf{(h2)} on dynamics near singularities.
\begin{comment}
\begin{defn}\label{ustandard}
For any standard family $\mathcal{G}=\{(W_{\alpha},\nu_{\alpha}), \alpha\in \mathcal{A},\lambda\}$, any $n\geq 0$, let $F^n\mathcal{G}$ be defined as in (\ref{FnG}) with the factor measure $\lambda_n$. We define a new standard family using the same factor measure $\lambda_n$, but changing the probability measure from $\nu_{\alpha}$ to the $u$-SRB measure $\mu_{\alpha}$ on each unstable curve $W_{\alpha}$: \begin{equation}\label{associated family}\mathcal{G}_{n}^u:=(F^n W, \mu_n):=((V_{\alpha},\mu_{\alpha})\,:\,\alpha\in \mathcal{A}_n, \lambda_n).\end{equation} We call $\mathcal{G}_n^u$ the {\it{ $\mu$-standard family associated to }}$F^n \mathcal{G}$.
\end{defn}
Next lemma discusses the relations between these two standard families, which implies that for $n$ large, essentially only the factor measure $\lambda_n$ matters.
\begin{lemma}\label{holdernorm} There exists $C>0$ such that for any standard family $\mathcal{G}:=(\mathcal{W},\nu)$ with $\mathcal{G}^u_n=( F^n \mathcal{W}, \mu_n)$ being the $\mu$-standard family associated to $ F^n \mathcal{G}$, for any $n\geq 1$, then \begin{equation}\label{nualphamu}| F^n_*\nu(A)-\mu_n(A)|\leq C\nu( F^{-n}A)\Lambda^{-\gamma_0 n},\end{equation} for any measurable collection of stable manifolds $A$.
In particular, if for any smooth component $V\in \mathcal{F}^n \mathcal{W}$, the length satisfies $|V|<20\delta_0$, where $\delta_0$ was chosen in Lemma \ref{densitybd}, then
\begin{equation}\label{nualphamudelta}| F^n_*\nu(A)-\mu_n(A)|\leq \varepsilon_{\mathbf{d}} \, \nu( F^{-n}A)\vartheta_1^{n+1},\end{equation}
\end{lemma}
\begin{proof}
We first consider the case when $\mathcal{G}=(W,\nu)$ is a standard pair with density $$g:=d\nu/d\mu_{W}\in\mathcal{H}^+(\gamma_0).$$ Let $$ F^n W=\{V_{\alpha}, \alpha\in \mathcal{A}_n\},$$ then $\mathcal{A}_n$ is at most countable. By the definition (\ref{standardpair}), we know that for any $x,y\in V_{\alpha}$:
\begin{equation}\label{holderlngalapha}|\ln g( F^{-n}x)-\ln g( F^{-n}y)|\leq C_F d( F^{-n}x, F^{-n}y)^{\gamma}\leq C_F| F^{-n}V_{\alpha}|^{\gamma_0}.\end{equation}
Or equivalently,
\begin{equation}\label{holderlngalaphan}|g( F^{-n}x)- g( F^{-n}y)|\leq C_F| F^{-n}V_{\alpha}|^{\gamma_0}g( F^{-n}y).\end{equation}
One can check that the push forward measure $ F^n_*\nu$ has density $g( F^{-n}x) $, with respect to the Lebesgure measure $m$. More precisely, for any measurable collection of stable manifolds $A\subset \mathcal{M}$,
\begin{align}\label{cFnnuAnug}
F^n_*\nu(A)&=\int \chi_A(x) \,d F^n\nu(x)
=\int \chi_A(x) \cdot g( F^{-n} x)\, d\mu(x)\nonumber\\
&=\sum_{\alpha\in \mathcal{A}_n}\int_{V_{\alpha}}\chi_A(x) g(\mathcal{F}^{-n}x) d\mu_{\alpha}\,\lambda^u(d\alpha)\\
&=\sum_{\alpha\in \mathcal{A}_n}\int_{V_{\alpha}}\chi_A(x) d\nu_{\alpha}\,\lambda_n(d\alpha),\end{align}
where $d\nu_{\alpha}= g_{\alpha}\,\,d \mu_{\alpha}(x)$ and $$\lambda_n(d\alpha)=\int_{V_{\alpha}} g( F^{-n} x)\,\,d \mu_{\alpha}(x) .$$ Here $g_{\alpha}$ is the density function defined only on $V_{\alpha}$ such that for any $x\in V_{\alpha}$, $$ g_{\alpha}(x)=\frac{g( F^{-n} x)}{\int_{V_{\alpha}} g( F^{-n} x) \, d\mu_{\alpha}}$$ Clearly, one can check that $$\lambda_n(\mathcal{A}_n)= F^n_*\nu(\mathcal{M})=\mu(g).$$
In particular, note that if we start from a standard pair $(W,\mu_W)$, then $ F^n(W,\mu_W)$ is a standard family on $ F^n W$ with measure $ F^n \mu_W$ satisfying
$$ F^n_* \mu_W(A)=\sum_{\alpha\in \mathcal{A}_n}\mu_{\alpha}(A\cap V_{\alpha})\, \lambda^u_{\mathcal{M}}(\alpha).$$
Thus we use (\ref{holderlngalaphan}) and divide $\mu_{\alpha}(g\circ F^{n})$ on every term to get
\begin{equation}\label{holderlngalaphang}|g_{\alpha}(x)- g_{\alpha}(y)|\leq C_F| F^{-n}V_{\alpha}|^{\gamma_0}g_{\alpha}(y).\end{equation}
Since the pBorelrobability density $g_{\alpha}$ is continuous on $V_{\alpha}$, and the average value is $1$, then there exists $y\in W_{\alpha}$, such that $g_{\alpha}(y)=1$. Then we get that for any $x\in V_{\alpha}$:
$$ |g_{\alpha}(x)-1|\leq C_F | F^{-n}V_{\alpha}|^{\gamma_0},$$
where the constant $C_F$ does not depend on $x$ and $y$.
Now for any Borel measurable set $A\subset \mathcal{M}$, we integrate the above inequality on $A$ with respect to $\mu_{\alpha}$ in order to get
\begin{equation}\label{ValphacapA}\left| \int_{V_{\alpha}\cap A} g_{\alpha}\,d \mu_{\alpha}-\mu_{\alpha}(V_{\alpha}\cap A)\right|\leq C_F| F^{-n}V_{\alpha}|^{\gamma_0}\mu_{\alpha}(V_{\alpha}\cap A).\end{equation}
Now we define a new standard family $$\mathcal{G}^u_n:=( F^n W, \mu_n)=((V_{\alpha}, \mu_{\alpha}), \alpha\in \mathcal{A}_n, \lambda_n),$$ using the same factor measure, such that
$$\mu_n(A)=\sum_{\alpha\in \mathcal{A}_n}\lambda_n(\alpha)\int_{V_{\alpha}} \chi_A\, d\mu_{\alpha}.$$
We integrate (\ref{ValphacapA}) with respect to the factor measure $\lambda_n$, then it follows that
$$| F^n_*\nu(A)-\mu_n(A)|\leq C_F F^n_*\nu(A)\cdot \sup_{\alpha\in \mathcal{A}_n, V_{\alpha}\cap A\neq \emptyset} \{| F^{-n}V_{\alpha}|^{\gamma_0}\}.$$
This implies (\ref{nualphamu}).
In general, we assume $$\mathcal{G}=(\mathcal{W},\nu)=\{(W_{\alpha}, \nu_{\alpha}), \alpha\in \mathcal{A}, \lambda\}$$ According to the above analysis, for any $n\geq 1$, and $\alpha\in \mathcal{A}$, $ F^n(W_{\alpha}, \nu_{\alpha})$ and its associated standard family, denoted as $$\mathcal{G}^u_{n,\alpha}=( F^n W_{\alpha}, \mu_{n,\alpha}),$$ are related by
$$| F^n_*\nu_{\alpha}(A)-\mu_{n,\alpha}(A)|\leq C_F\mu_{n,\alpha}(A)\cdot \sup_{\beta\in \mathcal{A}_{n,\alpha}, V_{\beta}\cap A\neq \emptyset} \{| F^{-n}V_{\beta}|^{\gamma_0}\}.$$
Now we integrate the above expression with respect to the factor measure, and using the fact that the associated family $\mathcal{G}^u_n$ of $ F^n \mathcal{G}$ can be written as $\mathcal{G}^u_n=( F^n\mathcal{W}, \mu_n)$, which satisfies
$$\mu_n(A)=\int_{\alpha\in \mathcal{A}}\mu_{n,\alpha}(A)\lambda (d\alpha)= F^n_*\nu(A).$$
Thus for any Borel measurable set $A\subset \mathcal{M}$, we have
\begin{align*}| F^n_*\nu(A)-\mu_{n}(A)|&=|\int_{\alpha\in \mathcal{A}}( F^n_*\nu_{\alpha}(A)-\mu_{n,\alpha}(A))\,\lambda(d\alpha)|\\
&\leq C_F\,\mu_n(A)\cdot \sup_{\beta\in \mathcal{A}_n, V_{\beta}\cap A\neq \emptyset} \{| F^{-n}V_{\beta}|^{\gamma_0}\}\leq C_F\sup_{\beta\in \mathcal{A}_n, V_{\beta}\cap A\neq \emptyset} |V_{\beta}| \Lambda^{-n\gamma_0}\end{align*}
here $\mathcal{A}_n=\{W_{\alpha}\in F^{-n} A\}$ is the index set of unstable foliation of $ F^{-n}A$, and $\mathcal{A}$ is the index set of $\mathcal{W}$; and we have used (\textbf{h1}), i.e., the uniform expanding property of unstable manifolds under $F$.
Now we consider the case $V_{\beta}|\leq 20\delta_0$, for any $\beta\in \mathcal{A}_n$. Now by the choice of $\delta_0$, see (\ref{delta0}), we obtain the claimed estimations:
$$|F^n_*\nu(A)-\mu_{n}(A)|\leq C_F (20\delta_0)^{\gamma_0}\Lambda^{-n\gamma_0}\leq \varepsilon_{\mathbf{d}} \vartheta_1^{n+1}.$$
where $\varepsilon_{\mathbf{d}}$ was defined in Lemma \ref{densitybd}.
\end{proof}
\end{comment}
\section{Coupling Lemma for the induced map}
It was proved in \cite{C99, CZ09} that Assumptions \textbf{(h1)-(h4)} imply exponential decay of correlation for the induced system $(M, F,\mu_M)$ and any observables $f\in \mathcal{H}^-(\gamma_1)$ and $g\in \mathcal{H}^+(\gamma_2)$, where $\gamma_1,\gamma_2>0$, with ${\rm supp}\, f \subset M$ and ${\rm supp}\, g\subset M$. More precisely, we have
\begin{equation}\label{expcz09}
\left|\int_{M} (f\circ F^n)\, g\, d\mu_M -
\int_{M} f\, d\mu_M \int_{M} g\, d\mu_M\right|\leq C \|f\|^-_{_{C^{\gamma_1}}}\|g\|^+_{_{C^{\gamma_2}}} \vartheta^n,
\end{equation}
for some uniform constants $\vartheta=\vartheta(\gamma_1,\gamma_2)\in (0,1)$ and $C>0$.
We will review in this section the
coupling method developed in \cite{D01,CD,CM,Y99} for the induced system, but we have to construct a special hyperbolic set.
\subsection{Construction of a hyperbolic set $\mathcal{R}^*$}
We first construct a hyperbolic
set $\mathcal{R}^*\subset M$ with positive measure, which will be used as the reference set for the coupling procedure.
\begin{defn}\label{defRsu} Let $\Gamma^{s}$ be a family of stable manifolds, and $\Gamma^u$ a family of unstable manifolds with length $\in( 10\delta_0, 20\delta_0)$. We say that $\mathcal{R}^*=\Gamma^u\cap\Gamma^s$ is a {\it{hyperbolic set with product structure}}, if it satisfies the following four conditions:\\
(i) There exist a family of stable manifolds $\hat\Gamma^{s}$, a family of unstable manifolds $\hat\Gamma^u$, and a region $\mathcal{U}^*$ bounded by two stable manifolds $W^s_i\in \hat\Gamma^s$ and two unstable manifolds $W^u_i\in \hat\Gamma^{u}$, for $i=1,2$;\\
(ii) Any stable manifold $W^s\in \hat\Gamma^s$ and any unstable manifold $W^u\in\hat\Gamma^u$ only intersect at exactly one point;\\
(iii) The two defining families $\Gamma^{s/u}$ are obtained by intersecting $\hat\Gamma^{s/u}$ with $\mathcal{U}^*$, such that
$$\Gamma^{s/u}:=\hat\Gamma^{u/s}\cap \mathcal{U}^*$$
\noindent(iv) Let $\nu^u=\mu|_{\Gamma^u}$ be obtained by restricting the SRB measure on $\Gamma^u$, then $(\Gamma^u, \nu^u)$ defines a standard family, and $\nu^u(\Gamma^s)>0$.
We say a stable or unstable curve $W$ {\it{properly across}} $\mathcal{U}^*$, if the two end points of the closure of $W\cap\mathcal{U}^*$ are contained in the boundary $\partial \mathcal{U}^*$. We say a set $A\subset \mathcal{R}^*$ is a
$u$-subset, if there exists a measurable collection of unstable manifolds $\Gamma^u_A\subset\Gamma^u$, such that $A=\Gamma^u_A\cap \Gamma^s$. Similarly a set $A\subset \mathcal{R}^*$ is called a
$s$-subset, if there exists a subset $\Gamma^s_A\subset\Gamma^s$, such that $A=\Gamma^s_A\cap \Gamma^u$.
\end{defn}
It follows from condition (iv) that we can define a factor measure $\lambda$ on the sigma algebra (induced by the Borel $\sigma$-algebra of $\mathcal{M}$) of the index set of $\Gamma^u=\{W_{\alpha},\alpha\in \mathcal{A}\}$, such that for any Borel set $A\subset \mathcal{U}^*$,
$$\nu^u(A)=\int_{\alpha\in\mathcal{A}} \mu_{\alpha}(W_{\alpha}\cap A)\,\lambda(d\alpha).$$
Hyperbolic product sets were constructed in several references, see for example \cite{D01,CD,CM,Y98, CZ09}. The next proposition is devoted for such a construction.
The construction of such hyperbolic set with property (i) - (iii) was done in details
in \cite{CM, CZ09}, so we will not repeat it here.
\begin{proposition}\label{firstproper}
There exist $\hat\delta_1$, a hyperbolic set with product structure $\mathcal{R}^*=\Gamma^s\cap\Gamma^u$ and the rectangle $\mathcal{U}^*$ containing $\mathcal{R}^*$ bounded by two stable manifolds and two unstable manifolds with length approximately $10\delta_0$, such that the following properties hold:\\
(i) $\mu(\mathcal{R}^*)>\hat\delta_1$ and for any unstable $W$ that fully crosses $\mathcal{U}^*$, $\mu_{W}(\mathcal{R}^*\cap W)>\hat\delta_1$;\\
(ii) There exists $n_0\geq 1$, such that for any $n\geq n_0$, $$\mu(F^n\mathcal{R}^*\cap\mathcal{R}^*)>0.$$For any $n\geq 1$, $$\mu(F^n\mathcal{R}^*\cap\mathcal{R}^*)>\hat\delta_1.$$
(iii) Moreover, if a point $x\in \mathcal{R}^*$ return to $\mathcal{R}^*$ under $\mathcal{F}^n$, for some $n\geq 1$, then $W^u(\mathcal{F}^n x)\subset \Gamma^u$.
\end{proposition}
\begin{comment}
\begin{proof}
The construction of such hyperbolic set with property (i) and (ii) was done in details in \cite{CM, CZ09}, so we will not repeat it here.
Next we will be devoted to prove property (iii). However, we need to reconstruct a hyperbolic set $\tilde\mathcal{R}^*$ satisfying properties (i) and (ii), and then by adding some more constructions, we get a new hyperbolic set $\mathcal{R}^*$ which inherits the property (i) and (ii), yet also enjoys (iii).
More precisely, using techniques described in \cite{CM, CZ09}, we first construct a family of stable/unstable manifolds $\tilde\Gamma^{s/u}$, a hyperbolic set $\tilde \mathcal{R}^*=\tilde\Gamma^s\cap\tilde\Gamma^u$, together with a rectangle $\tilde\mathcal{U}^*$ being the smallest region containing $\mathcal{R}^*$. Moreover they have the properties listed as in item (i) and (ii):\\
(i$'$) There exists $\tilde\delta_1>0$, such that $\mu(\tilde\mathcal{R}^*)>\tilde\delta_1$ and for any unstable $W$ that fully cross $\tilde\mathcal{U}^*$, $\mu_{W}(\tilde\mathcal{R}^*\cap W)>\tilde\delta_1$.
(ii$'$) There exists $n_0\geq1$, such that for any $n\geq n_0$, $\mu(F^n\tilde\mathcal{R}^*\cap\tilde\mathcal{R}^*)>\tilde\delta_1$.\\
Note that any stable manifold can only contains at most one periodic point. Thus using the dense property of periodic points in the support of the SRB measure $\mu$, see \cite{H02} page 206-209, and \cite{Katok}, one can in addition show that most of these periodic points have long stable and unstable manifolds. Mainly because a periodic orbit only consists of finite number of points, they can easily get away from the singular set $S_{\pm 1}$ which only consists of countably many smooth curves. Thus we can choose $\tilde\delta_1$ small enough, such that $\tilde\mathcal{R}^*$ contains plenty of periodic points. In particular, we can choose two stable manifolds in $\tilde\Gamma^s$, denotes as $W^s_i$, for $i=1,2$, with the following properties:\\
(I) $W^s_1$ contains a periodic point of period $p\geq 2$, and there exists $k_1\geq 2$ such that $F^{k_1} W^s_1\subset W^s_2$;\\
(II) $W^s_2$ contains a periodic point of period $p\geq 2$, and there exists $k_2\geq 2$ such that $F^{k_2} W^s_2\subset W^s_1$;\\
(iii) Let $A_s$ be the $s$-subset of $\tilde\mathcal{R}^*$ bounded by $W^s_1$ and $W^s_2$. Then whenever $F^nA_s$ intersects $A_s$, the intersection set is either a trivial set or a $u$-subset that fully crosses $A_s$.\\
In order to choose these two stable manifolds with properties (I)-(II), we only need to take one stable manifold $W^s_1$ which contains a periodic point of periodic $p\geq 2$; and let $W^s_2\in\tilde\Gamma^s$ be the stable manifold that contains $F^{k_1}W^s_1$ for some $k_1\in \{1,\cdots, p-1\}$. By the mixing property, there exists $n_1\geq 1$, such that $$\mu(\mathcal{F}^{n_1}(\tilde\mathcal{R}^*)\cap\tilde\mathcal{R}^*)\geq \tilde\delta_1.$$ Thus we can choose one stable manifold $W^s_1$ in $\tilde\Gamma^s$ such that $W^s_1$ contains a periodic point of periodic $p\geq 2$, and $\mathcal{F}^{n_1} W^s_1$ is contained in a stable manifold in $\tilde\Gamma^s$. Since $p$ is finite, we can find a stable manifold $W^s$ in $$A(W^s_1):=\{\mathcal{F}^k W^s_1, k=1,\cdots, p-1\}$$ such that $W^s=\mathcal{F}^{k_1} W^s_1$ belongs to a stable manifold $W^s_2\in \tilde\Gamma^s$, and there are no other stable manifolds in $A(W^s_1)$ lying between $W^s_1$ and $W^s_2$ in $\tilde\mathcal{U}^*$. Now we define $k_2=p-k_1$. Then we know that $W^s_2$ consists of a periodic point of period $p$ and $F^{k_2} W^s_2\subset W^s_1$.
Since the system is invertible, any unstable manifold can only contains at most one periodic point of $F^{-1}$. Thus using similar scheme, we can choose two unstable manifolds $W^u_i$, with $i=1,2$, such that:\\
(I$'$) $W^u_1$ contains a periodic point of period $p'\geq 2$, and there exists $k'_1\geq 2$ such that $F^{-k'_1} W^u_1\subset W^u_2$;\\
(II$'$) $W^u_2$ contains a periodic point of period $p'$, and there exists $k'_2\geq 2$ such that $$F^{-k'_2} W^u_2\subset W^u_1.$$
Next we define our hyperbolic set $$\mathcal{R}^*:=A_u\cap \tilde\mathcal{R}^*$$ and $$\Gamma^{s/u}=\tilde\Gamma^{s/u}\cap A_u.$$ Then $\mathcal{R}^*=\Gamma^s\cap\Gamma^u$ is a hyperbolic set. In addition, one can check that this new hyperbolic set has the properties given by item (i) and (ii) together with a constant $\delta_1<\tilde\delta_1$. Next we will verify item (iii) for this newly constructed hyperbolic set $\mathcal{R}^*$.
Let $W^u(x)\in \Gamma^u$, such that for some $n\geq 1$, $F^{n} x\in\mathcal{R}^*$ returns to $\mathcal{R}^*$. But we assume $W^u(F^n x)\cap \mathcal{R}^*$ is not fully stretched in $\mathcal{R}^*$. This implies that there exists an unstable manifold $W\in \Gamma^u$, such that $F^n x\in W$, but $W^u(F^n x)$ is only a subset of $ W$. Let $z$ be an end point of $W^u(F^n x)\cap W$. Note that an end point of an unstable manifold $z$ must belong to the singular set of $F^{-m}$ for some $m\geq 1$. More precisely, $F^{-m} z\in S_{-1}$. However, since $z\in W$ and $W$ is an unstable manifold, thus by the definition of unstable manifold, we know that $F^{-k} W$ is a smooth curve in $M\setminus S_{-1}$ for any $k\geq 1$. This contradicts to the fact that $ F^{-m} W\cap S_{-1}$ intersects at $z$. Thus such $z$ does not exist, which implies that $W^u(F^n x)=W$ since $F^n x\in W$. This verified item (iii)(a).
Next we show that for any $x\in \mathcal{R}^*$, $F^n x\in \mathcal{R}^*$ implies that $F^n W^u(x)$ must have properly returned to $\mathcal{U}^*$ too. According to (iii)(a), we know that $F^n W^u(x)$ must have properly returned to $\mathcal{U}^*$. So the only possibility that $F^n (W^u(x)\cap \mathcal{U}^*)$ does not properly return to $\mathcal{U}^*$ is that an end point, denoted by $z \in W^u(x)\cap \partial\mathcal{U}^*$, may have the property that $F^n z$ lies in the interior of $\mathcal{U}^*$. Since the stable manifold $W^s(z)$ is a stable boundary of $\mathcal{U}^*$, thus $F^n W^s(z)$ also has a piece contained in the interior of $\mathcal{U}^*$. However, this is a contradiction, as by our construction with property (I) and (II), we know that images of $W^s(z)$ can never touch the interior of $\mathcal{U}^*$, and it can at most touch the other stable boundary of $\partial\mathcal{U}^*$. This verifies item (iii)(b).
Using invertibility of the map, one can verify item (iii)(c) and (iii)(d) similarly, which we will omit it here.
Finally, for item (iv) and (v), it is enough to prove (iv) in details.
We prove it by way of contradiction.
Assume for some $n\geq 1$, the set $A_n:=F^n\mathcal{R}^*\cap\mathcal{R}^*$ is neither a null set nor a $u$-subset. Then there are only two possible cases:\\
\noindent Case I. $A_n$ is contained in a $u$-subset, but does not fully stretched in the unstable direction in $\mathcal{R}^*$. Item (iii) implies that the $s$-boundary of $A_n$ must come from either $F^nW^s_1$ or $F^nW^s_2$. But this is a contraction, since by our choice of $W^s_1$ and $W^s_2$, their forward images never come to the interior of $\mathcal{U}^*$.\\
\noindent Case II. $A_n$ is fully stretched in the unstable direction in $\mathcal{R}^*$, but it intersects either $W^u_1$ or $W^u_2$, which are the unstable boundaries of $\mathcal{U}^*$. This implies that $F^{-n}W^u_1$ or $F^{-n} W^u_2$ belongs to the interior of $\mathcal{U}^*$, which is again a contradiction to our construction of $W^u_i$, for $i=1,2$. \\
Thus we have shown that for any $s$-subset $R^s$ of $\mathcal{R}^*$, if $F^n R^s$ returns to $\mathcal{R}^*$ nontrivially for some $n\geq 1$, then $F^n R^s$ must be a $u$-subset. This finished the proof of the Proposition.
\end{proof}
\end{comment}
From now on, according to the construction in Proposition \ref{firstproper}, we will fix the hyperbolic set $\mathcal{R}^*$, as well as its defining families $\Gamma^s$ and $\Gamma^u$, with \begin{equation}\label{defnR}
\mathcal{R}^*=\Gamma^u\cap\Gamma^s.\end{equation}
\begin{comment}
Next we prove an important property for our hyperbolic set $\mathcal{R}^*$.
\begin{proposition}\label{firstproper}
There exist $\hat\delta_1$, a hyperbolic set with product structure $\mathcal{R}^*=\Gamma^s\cap\Gamma^u$ and the rectangle $\mathcal{U}^*$ containing $\mathcal{R}^*$ bounded by two stable manifolds and two unstable manifolds with length approximately $10\delta_0$, such that the following properties hold:\\
(i) $\mu(\mathcal{R}^*)>\hat\delta_1$ and for any stable / unstable $W$ that fully cross $\mathcal{U}^*$, $\mu_{W}(\mathcal{R}^*\cap W)>\hat\delta_1$;\\
(ii) There exists $n_0>1$, such that for any $n\geq n_0$, $$\mu(F^n\mathcal{R}^*\cap\mathcal{R}^*)>\hat\delta_1$$\\
(iii) Moreover, there is a set of full measure $\mathcal{R}'\subset \mathcal{R}^*$, such that for any $x\in \mathcal{R}'$, its any forward /backward return to $\mathcal{R}^*$ must be ``proper" in the following sense:
\begin{enumerate}
\item[(a)] if for some $n\geq 1$ , $F^n x\in\mathcal{R}^*$, then $W^u(F^n x)\cap \mathcal{R}^*$ must properly return to $\mathcal{R}^*$;
\item[(b)] if for some $n\geq 1$ , $F^n x\in\mathcal{R}^*$, then $F^n (W^u(x)\cap \mathcal{U}^*) \cap \mathcal{R}^*$ must properly return to $\mathcal{R}^*$;
\item[(c)] if for some $n\geq 1$, $F^{-n} x\in\mathcal{R}^*$, then $W^s(F^{-n} x)\cap \mathcal{R}^*$ must properly return to $\mathcal{R}^*$.
\item[(d)] if for some $n\geq 1$, then $F^{-n}(W^s(x)\cap\mathcal{U}^*)\cap \mathcal{R}^*$ must properly return to $\mathcal{R}^*$.
\end{enumerate}
(iv) For any $n\geq 1$, if $F^n \mathcal{R}^*\cap\mathcal{R}^*$ is not empty, then $F^n\mathcal{R}^*\cap\mathcal{R}^*$ is either a trivial set or a $u$-subset of $\mathcal{R}^*$. \\
(v) For any $n\geq 1$, if $F^{-n} \mathcal{R}^*\cap\mathcal{R}^*$ is not empty, then $F^{-n}\mathcal{R}^*\cap\mathcal{R}^*$ is either a trivial set or a $s$-subset of $\mathcal{R}^*$.
\end{proposition}
\begin{proof}
The construction of such hyperbolic set with property (i) and (ii) was done in details in \cite{CM, CZ09}, so we will not repeat here.
So below we will devoted to prove property (iii). However, we need to reconstruct a hyperbolic set $\tilde\mathcal{R}^*$ satisfying properties (i) and (ii), and then by adding some more constructions, we get a new hyperbolic set $\mathcal{R}^*$ which inherits the property (i) and (ii), yet also enjoys (iii).
More precisely, using techniques described as in \cite{CM, CZ09}, we first construct a family of stable/unstable manifolds $\tilde\gamma^{s/u}$, a hyperbolic set $\tilde \mathcal{R}^*=\tilde\Gamma^s\cap\tilde\Gamma^u$, together with a rectangle $\tilde\mathcal{U}^*$ being the smallest solid region containing $\mathcal{R}^*$; such they have the properties listed similar to item (i) and (ii):\\
(i') There exists $\tilde\delta_1>0$, such that $$\mu(\tilde\mathcal{R}^*)>\tilde\delta_1$$ and for any stable/unstable $W$ that fully cross $\tilde\mathcal{U}^*$, $$\mu_{W}(\tilde\mathcal{R}^*\cap W)>\tilde\delta_1$$
(ii') There exists $n_0>1$, such that for any $n\geq n_0$, $$\mu(F^n\tilde\mathcal{R}^*\cap\tilde\mathcal{R}^*)>\tilde\delta_1$$
Note that any stable manifold can only contains at most one periodic point. Thus using the dense property of periodic points in the support of the SRB measure $\mu$, see \cite{H02} page 206-209, and \cite{Katok}, one can in addition show that most of these periodic points have long stable and unstable manifolds. Mainly because a periodic point only has finite number of images, so they can easily get away from the singular set $S_{\pm 1}$ which only consists of countably many smooth curves. Thus we can choose $\tilde\delta_1$ small enough, such that $\tilde\mathcal{R}^*$ contains plenty of periodic points. In particular, we can choose two stable manifolds in $\tilde\Gamma^s$, denotes as $W^s_i$, for $i=1,2$, with the following properties:\\
(I) $W^s_1$ contains a periodic point of period $p\geq 2$, and there exists $k_1\geq 2$ such that $F^{k_1} W^s_1\subset W^s_2$;\\
(II) $W^s_2$ contains a periodic point of period $p\geq 2$, and there exists $k_2\geq 2$ such that $F^{k_2} W^s_2\subset W^s_1$;\\
Indeed to choose these two stable manifolds with properties (I)-(II), we only need to choose one stable manifold $W^s_1$ which contains a periodic point of periodic $p\geq 2$; and let $W^s_2\in\tilde\Gamma^s$ be the stable manifold that contains $F^{k_1}W^s_1$ for some $k_1\in \{1,\cdots, p-1\}$. By the mixing property, there exists $n_1\geq 1$, such that $$\mu(F^{n_1}(\tilde\mathcal{R}^*)\cap\tilde\mathcal{R}^*)\geq \tilde\delta_1$$ We choose one stable manifold $W^s_1$ in $\mathcal{R}^s_1$m such that $W^s_1$ contains a periodic point of periodic $p\geq 2$, and $\mathcal{F}^{n_1} W^s_1$ is contained in a stable manifold in $\tilde\Gamma^s$. Since $p$ is finite, we can find a stable manifold $W^s$ in $$A(W^s_1):=\{\mathcal{F}^k W^s_1, k=1,\cdots, p-1\}$$ such that $W^s=\mathcal{F}^{k_1} W^s_1$ belongs to a stable manifold $W^s_2\in \tilde\Gamma^s$, and there are no other stable manifolds in $A(W^s_1)$ lie between $W^s_1$ and $W^s_2$ in $\tilde\mathcal{U}^*$. Now we define $k_2=p-k_1$. Then we know that $W^s_2$ consists a periodic point of period $p$ and $F^{k_2} W^s_2\subset W^s_1$.
Since the system is invertible, note that any unstable manifold can only contains at most one periodic point of $F^{-1}$. Thus using similar scheme, we can choose two unstable manifolds $W^u_i$, with $i=1,2$, such that:\\
(I$'$) $W^u_1$ contains a periodic point of period $p'\geq 2$, and there exists $k'_1\geq 2$ such that $F^{-k'_1} W^u_1\subset W^u_2$;\\
(II$'$) $W^u_2$ contains a periodic point of period $p'$, and there exists $k'_2\geq 2$ such that $$F^{-k'_2} W^u_2\subset W^u_1$$
Next we define our hyperbolic set $$\mathcal{R}^*:=A_u\cap \tilde\mathcal{R}^*$$ and $$\Gamma^{s/u}=\tilde\Gamma^{s/u}\cap A_u$$ Then $\mathcal{R}^*=\Gamma^s\cap\Gamma^u$ is a hyperbolic set. In addition, one can check that this new hyperbolic set has the properties given by item (i) and (ii) together with a constant $\delta_1<\tilde\delta_1$. Next we will verify item (iii) for this newly constructed hyperbolic set $\mathcal{R}^*$.
Let $W^u(x)\in \Gamma^u$, such that for some $n\geq 1$, $F^{n} x\in\mathcal{R}^*$ return to $\mathcal{R}^*$. But we assume $W^u(F^n x)\cap \mathcal{R}^*$ is not fully stretched in $\mathcal{R}^*$. This implies that there exists an unstable manifold $W\in \Gamma^u$, such that $F^n x\in W$, but $W^u(F^n x)$ is only a subset of $ W$. Let $z$ be an end point of $W^u(F^n x)\cap W$. Note that an end point of an unstable manifold $z$ must belong to the singular set of $F^{-m}$ for some $m\geq 1$. More precisely, $F^{-m} z\in \mathcal{S}_{-1}$. However, since $z\in W$ and $W$ is an unstable manifold, thus by the definition of unstable manifold, we know that $F^{-k} W$ is a smooth curve in $M\setminus \mathcal{S}_{-1}$ for any $k\geq 1$. This contradicts to the fact that $ F^{-m} W\cap \mathcal{S}_{-1}$ intersects at $z$. Thus such $z$ does not exist, which implies that $W^u(F^n x)=W$ since $F^n x\in W$. This verified item (iii)(a).
Next we show that for any $x\in \mathcal{R}^*$, $F^n x\in \mathcal{R}^*$ implies that $F^n W^u(x)$ must properly return to $\mathcal{U}^*$ too. According to (iii)(a), we know that $F^n W^u(x)$ must have properly returned to $\mathcal{U}^*$. So the only possibility that $F^n (W^u(x)\cap \mathcal{U}^*)$ does not properly return to $\mathcal{U}^*$ is that, an end point, denoted by $z \in W^u(x)\cap \partial\mathcal{U}^*$, may has the property that $F^n z$ lies in the interior of $\mathcal{U}^*$. Since the stable manifold $W^s(z)$ is a stable boundary of $\mathcal{U}^*$, thus $F^n W^s(z)$ also has a piece contained in the interior of $\mathcal{U}^*$. However, this is contradiction, as by our construction with property (I) and (I), we know that images of $W^s(z)$ can never touch the interior of $\mathcal{U}^*$, and it can at most touch the other stable boundary of $\partial\mathcal{U}^*$. This verifies item (iii)(b).
Using invertibility of the map, one can verify item (iii)(c) and (iii)(d) similarly, which we will omit it here.
Finally, for item (iv) and (v), it is enough to prove (iv) in details.
We prove it by way of contradiction.
Assume for some $n\geq 1$, the set $A_n:=F^n\mathcal{R}^*\cap\mathcal{R}^*$ is neither trivial nor a $u$-subset. Then there are only two possible cases:\\
\noindent Case I. $A_n$ is contained in a $u$-subset, but does not stretch fully in the unstable direction in $\mathcal{R}^*$. Item (iii) implies that the $s$-boundary of $A_n$ must come from either $F^nW^s_1$ or $F^nW^s_2$. But this is a contraction, since by our choice of $W^s_1$ and $W^s_2$, their forward images never come to the interior of $\mathcal{U}^*$.\\
\noindent Case II. $A_n$ is fully stretched in the unstable direction in $\mathcal{R}^*$, but it intersects either $W^u_1$ or $W^u_2$, which are the unstable boundary of $\mathcal{U}^*$. This implies that $F^{-n}W^u_1$ or $F^{-n} W^u_2$ belongs to the interior of $\mathcal{U}^*$, which is again a contraction to our construction of $W^u_i$, for $i=1,2$. \\
Thus we have shown that for any $s$-subset $R^s$ of $\mathcal{R}^*$, if $F^n R^s$ returns to $\mathcal{R}^*$ nontrivially for some $n\geq 1$, then $F^n R^s$ must be a $u$-subset. This finished the proof of Proposition.
\end{proof}
\end{comment}
In the coupling scheme that will be described below, we will consider a standard pair $(W,\nu)$ by subtracting from its density a smooth function. Next lemma explains that after a few more iterations under $F$, the resulting measure also induces a standard pair.
\begin{lemma}\label{defnN} Let $(W,\nu)$ be a standard pair properly crosses $\mathcal{R}^*$, with $h=d\nu/d\mu_W$. Assume $ g\in \mathcal{H}(\gamma_0)$ is a function on $W$, with $\|g\|_{\gamma_0}\leq 1$, such that $(1-\mathbf{a})h/2<g<(1-\mathbf{a})h$, with $\mathbf{a}=\Lambda^{-\gamma_0}(1+\frac{1}{C_F})$. We denote $\eta$ as the measure with density $h_0=h-g$. Then $F(W,\eta/\eta(\mathcal{M}))$ is a standard pair, and $F (W, \eta)$ is a standard family.
\end{lemma}
\begin{proof} By the definition of standard pair, we know that
the positive density function $h=d\nu/d\mu_W$ satisfies (\ref{standardpair}):
\begin{equation}
| h(x) - h(y)|
\leq C_{F} d(x,y)^{\gamma_0}, \end{equation} where $\gamma_0\in (0,1)$ was given in (\ref{distor10}), and $C_F>C_{\mathbf{r}}$ is a fixed large constant.
Now for $ g\in \mathcal{H}(\gamma_0)$, with $\|g\|_{\gamma_0}\leq 1$ and $(1-\mathbf{a})h/2<g<h(1-\mathbf{a})$, we denote $$h_0=h-g$$ as the density of $\eta$. Note that
$$|g(x)-g(y)|\leq d(x,y)^{\gamma_0}$$
Then one can check that the new probability density $h':=h_0/\eta(M)$ satisfies
$$|\ln h'(x)-\ln h'(y)|\leq \frac{(C_F+1)}{\mathbf{a}}d(x,y)^{\gamma_0}\leq \Lambda^{\gamma_0} C_F d(x,y)^{\gamma_0}$$
which implies that $(W,\eta/\eta(M))$ is a psudo-standard pair.
We define $F^n W=\{V_{\alpha}, \alpha\in \mathcal{A}_n\}$, where $\mathcal{A}_n$ is a countable index set.
Note that by (\ref{dens}) and (\ref{cmW}), for any $x\in V_{\alpha}$, the $u$-SRB density $\rho_{V_{\alpha}}=d\mu_{V_{\alpha}}/dm_{V_{\alpha}}$ satisfies
$$\rho_{F^{-n}V_{\alpha}}(F^{-n}x)\mathcal{J}}\def\varepsilon{\varepsilon_{V_{\alpha}}(F^{-n} x)=\rho_{V_{\alpha}}(x) .$$
We denote for any $n\geq 1$, the density function $h_n$ of $F^n_*\eta$ as
\begin{align*}
h_n(x)&=\frac{d F_*^n\eta(x)}{d\mu_{V_{\alpha}}(x)}=
\frac{d\eta(F^{-n}x)}{d\mu_{F^{-n}V_{\alpha}}(F^{-n} x)}\cdot \frac {d\mu_{F^{-n}V_{\alpha}}(F^{-n} x)}{d m_{F^{-n}V_{\alpha}}( F^{-n}x)}\cdot \frac{d m_{F^{-n}V_{\alpha}}( F^{-n}x)}{d m_{V_{\alpha}( x)}}\cdot \frac{dm_{V_{\alpha}}( x)}{d\mu_{V_{\alpha}( x)}}\\
& = h_0(F^{-n}(x))\mathcal{J}}\def\varepsilon{\varepsilon_{V_{\alpha} (F^{-n}(x))}\cdot \frac{\rho_{F^{-n}V_{\alpha}}(F^{-n}x)}{\rho_{V_{\alpha}}(x)}\\
&=h_0(F^{-n}(x)),\end{align*}
for all $ x\in V_{\alpha}\subset F^nW$.
According to above analysis, and use the notation $\mu_{\alpha}=\mu_{V_{\alpha}}$, one can check that for any measurable set $A$,
\begin{align*}
F^n_*\eta(A)&=\int_W \chi_A(F^n x) h_0(x)\,d\mu_W(x)\\
&=\sum_{\alpha\in \mathcal{A}_n} \int_{V_{\alpha}} \chi_A(y) h_0(F^{-n}y) d \mu_{\alpha}(y)\\
&=\sum_{\alpha\in \mathcal{A}_n} \int_{V_{\alpha}} \frac{\chi_A(y) h_0(F^{-n}y)}{\mu_{\alpha}(h_0\circ F^{-n})}\, d\mu_{\alpha}(y) \cdot \mu_{\alpha}(h_0\circ F^{-n})\\
&=\sum_{\alpha\in \mathcal{A}_n}\int_{V_{\alpha}}\chi_A(x) d\nu_{\alpha}\,\cdot \lambda_n(\alpha),\end{align*}
where $$d\nu_{\alpha}= h_{\alpha}\,d\mu_{\alpha}$$ and \begin{equation}\label{lambdan}\lambda_n(\alpha)=\int_{V_{\alpha}} h_0(F^{-n} y)\, d\mu_{\alpha}(y)\end{equation} is the factor measure on index set $\mathcal{A}_n$. Here $h_{\alpha}$ is the probability density function defined only on $V_{\alpha}$ such that for any $x\in V_{\alpha}$, $$ h_{\alpha}(x)=\frac{h_0(F^{-n} x)}{\int_{V_{\alpha}}h_0(F^{-n} y)d\mu_{\alpha}(y)}.$$
Note that using the fact that $h/3\leq g\leq h/2$, we have for any $x, y\in V_{\alpha}\in F^n W$,
\begin{align*}
| \ln h_{\alpha}(x)- \ln h_{\alpha}(y)|&=|\ln h_0(F^{-n}x)-\ln h_0(F^{-n}y)|\leq \Lambda^{\gamma_0}C_F d(F^{-n}x, F^{-n}y)^{\gamma_0} \leq \Lambda^{\gamma_0} C_F \Lambda^{-n \gamma_0} d(x,y)^{\gamma_0}.\end{align*}
Thus we know that $F(W,\eta)$ is already a standard family.
If we define $\eta'=\eta/\eta(\mathcal{M})$, then one can check that $F(W,\eta')$ is indeed a standard pair.
\end{proof}
\begin{lemma}\label{t01} There exist $\hat\delta_0\in (0,\mu(\Gamma^s))$, $N_0>1$, such that for any proper family $\mathcal{G}=(\mathcal{W},\nu)$ with $\nu(M)=1$, then $F^{n}_*\nu$ has at least $\hat\delta_0$ portion of measure properly returned to $\mathcal{R}^*$, for any $n\geq N_0$.
\end{lemma}
\begin{proof}
Let $\hat\delta_1>0$ be defined as in Proposition \ref{firstproper}, such that $\mu(\mathcal{R}^*)>\hat\delta_1$.
By the uniform mixing property of the induced map $(F,M,\mu_M)$, and the fact that $\mathcal{G}$ is a proper family, (\ref{expcz09}) implies that for $n>1$,
$$|F^n_*\nu(\mathcal{R}^*)- \mu_M(\mathcal{R}^*)|\leq C \vartheta^n.$$ Moreover for any standard pair $(W_{\alpha},\nu_{\alpha})$, since $W_{\alpha}$ only has two end points, say $x_1, x_2$, so if $F^n W_{\alpha}$ intersects $\mathcal{R}^*$ at some $x\in \mathcal{R}^*$, then it must consist of the entire unstable manifold $W^u(x)$, unless $W^u(x)$ consists of one of points in $\{F^n x_1, F^n x_2\}$. Thus a majority of curves in $F^n \mathcal{W}$ must properly cross $\Gamma^s$.
Thus by taking a large $N_0$ and a small number $\hat\delta_0\in (0, \hat\delta_1)$, we have that for any $n\geq N_0$, $F^n_*\nu$ has at least $ \hat\delta_0$ portion of measure properly returned to $\mathcal{R}^*$.
Moreover, our choice of $N_0$ and $\hat\delta_0$ are uniform for all proper families.
\end{proof}
Let $N\geq 1$ be the integer chosen in Lemma \ref{defnN}. We define \begin{equation}\label{defn1}n_1=\max \{N,N_0\}\end{equation} Note that by Lemma \ref{defnN}, if we subtract a ``nice" function from the density of a proper standard pair $(W,\nu)$, then after $n_1$ iteration of $F$, the image $F^{n_1}(W,\eta)$ becomes a new proper family, where $\eta$ is the new conditional measure and has at least $\hat\delta_0$ portion of measure properly returned to $\mathcal{R}^*$.
\subsection{The Coupling Lemma for the induced map}
We first introduce a new concept called the {\it{generalized standard family}}.
\begin{defn}\label{pseodugs}
Let $(\mathcal{W},\nu)$ be a standard family, such that $\mathcal{W}\subset \Gamma^u$ is a measurable collection of unstable manifolds in $\Gamma^u$. Then we define $(\mathcal{W},\nu)|_{R^*}:=(\mathcal{W}\cap \mathcal{R}^*, \nu|_{\mathcal{R}^*})$. For any $n\geq 0$, we call $(\mathcal{W}_n,\nu_n):=\mathcal{F}^{-n} ((\mathcal{W},\nu)|_{R^*})$ as an $\mathcal{F}-$ generalized standard family with index $n$.
In addition, we call $(\hat\mathcal{W}_n,\hat\nu_n):=F^{-n} ((\mathcal{W},\nu)|_{R^*})$ as an $F-$ generalized standard family with index $n$. \end{defn}
Next we restate the Coupling Lemma \cite{CD,CM} for the induced
system $(F, \mu_M)$ using the concept of generalized standard families.
\begin{lemma}\label{coupling}
Under assumptions \textbf{(h1)-(h4)}. Let $\mathcal{G}^i=(\mathcal{W}^i, \nu^i), i=1,2$, be two proper standard families on $M$. For any $n\geq 1$,
there exist two sequences of $F-$ generalized standard families $\{(\mathcal{W}^i_n,\nu^i_n), n\geq 0\}$, such that \begin{equation}\label{decomposeGi}\mathcal{G}^i=\sum_{n=0}^{\infty}(\mathcal{W}^i_n,\nu^i_n):=\left(\bigcup_{n=0}^{\infty}\mathcal{W}^i_n,\sum_{n=0}^{\infty}\nu^i_n\right).
\end{equation} And they also have the following properties, for each $n\geq 0$:
\begin{itemize}
\item[(i)] \textbf{Proper returned to $\mathcal{R}^*$ at $n$.}\\
\,\,\,\,\, Both $(\mathcal{W}^1_n,\nu^1_n)$ and $(\mathcal{W}^2_n,\nu^2_n)$ are $F-$ generalized standard families of index $n$;
\item[(ii)] \textbf{Coupling $F^n_*\nu^1_n$ and $F^n_*\nu^2_n$ along stable manifolds in $\Gamma^s$.}\\
\,\,\,\,\,\,\,For any measurable collection of stable manifolds $A\subset \Gamma^s$, we have $$F^n_*\nu_n^1(A)=F^n_*\nu^2_n(A).$$
\item[(iii)] \textbf{Exponential tail bound for uncoupled measure at $n$.}\\
\,\,\,\,\, For any $n\geq 1$, we denote $\bar\nu_n^i:=\sum_{k\geq n}\nu^i_k$ as the uncoupled measure at $n$-th step, then
\begin{equation}\label{ctail}\bar\nu_n^i(M)<C\vartheta^n,\end{equation}
where $C>0$ and $\vartheta$ are uniform constants.
\end{itemize}
\end{lemma}
Note that the original Coupling Lemma was stated only for proper families, so to deal with a standard family which is not proper, we need to iterate $N$ times to make it proper, according to Lemma \ref{properagain}. Moreover, $\mathcal{W}^i_m$ and $\mathcal{W}^i_n$ may not be disjoint, for $m\neq n$, unless during the coupling process, one can couple the entire measure that properly returned to $\mathcal{R}^*$. Here $F^n_*\nu^i_n, i=1,2$, are the coupled components of $F^n_*\nu^i$. In practice, a
coupling procedure occurs at a sequence of times $0<t_1<t_2<\cdots<t_k<\infty$. In
particular, $\nu^i_j=0$, when $j\neq t_k$ for all $1\leq k$,
which means that $F^j_*\nu^i$ remains unchanged between successive
coupling times.
According to item (ii) of the above Lemma, for any bounded function $f$ that is constant on each stable manifold, we have $F^n_*\nu^1_k(f)-F^n_*\nu^2_k(f)=0$. This implies that:
\begin{align}\label{fconstant}|F^n_*\nu^1(f)-F^n_*\nu^2(f)|&\leq \sum_{k=1}^n |F^{n-k}_*(F^k_*\nu^1_k(f)-F^k_*\nu^2_k(f))|+|\bar\nu^1_n(f)-\bar\nu^2_n(f)|\nonumber\\
&=|\bar\nu^1_n(f)-\bar\nu^2_n(f)|\leq 2C \|f\|_{\infty}\vartheta^n.
\end{align}
Similarly, the above coupling lemma implies the exponential rates
for any bounded H\"{o}lder function $f\in \mathcal{H}^-(\gamma_f)$, any proper families $(\mathcal{W}^i,\nu^i)$, $i=1,2$, for the induced system $(F,M)$:
\begin{align}\label{fdecay1}
&\left|\int f \circ F^n \,d\nu^1 - \int f \circ F^n \,d\nu^2 \right|\nonumber\\
& \leq 2 \|f\|_\infty \bar\nu^1_n(M) +
\sum_{j \leq n} \|f\|^-_{\gamma_f} \Lambda^{-(n-j)\gamma_f} \nu^1_j(M)
\nonumber\\
& \leq 2C \|f\|^-_{_{C^{\gamma_f}}} \vartheta_2^{ n},
\end{align}
where $\vartheta_2=\min\{\Lambda^{-\gamma_f},\vartheta\}$.
Next we prove a lemma which directly follows from the above Coupling Lemma \ref{coupling}.
\begin{lemma}\label{extralemma1}
For any proper standard family $\mathcal{G}=(\mathcal{W},\nu)$ in $M$, there exists a sequence of $F-$ generalized standard families $\{(\mathcal{W}_n,\nu_n), n\geq 0\}$ with the following properties:
\begin{itemize}
\item[(i)] $(\mathcal{W}_n,\nu_n)$ is a $F-$ generalized standard families of index $n$;
\item[(ii)] $\mathcal{G}=\sum_{n\geq 0} (\mathcal{W}_n,\nu_n)$, with $\{\mathcal{W}_n, n\geq 0\}$ are disjoint sets, a.s.;
\item[(iii)] Furthermore for any $n\geq 1$,
\begin{equation}\label{ctail1}\sum_{k=n}^{\infty} \nu_k(\mathcal{M})<C\vartheta^n,\end{equation}
where $C>0$ and $\vartheta$ are the constants in (\ref{ctail}).\end{itemize}
\end{lemma}
\begin{proof}
For $i=1,2$, we take
$$\mathcal{G}^i:=(\mathcal{W}, \nu)$$ as two copies of the standard family. Then according to the Coupling Lemma \ref{coupling}, we can couple the entire measure that properly return to $\mathcal{R}^*$ at each step $n\geq 0$. This implies that there exists a sequence of $F-$ generalized standard families $\{(\mathcal{W}_n,\nu_n), n\geq 1\}$, such that (i)-(iii) hold (as stated in Lemma \ref{coupling}). In addition, $\{\mathcal{W}_n, n\geq 0\}$ are disjoint.
\end{proof}
Next we will show that there is a generalized Young tower on $M$, following the above Coupling Lemma. The proof of the existence of a generalized Young tower as a consequence of the Coupling Lemma was first derived in \cite{CZ09} implicitly, and also proved in \cite{WZZ}.
\begin{proposition}\label{Ytower}
The induced map $F$ defines a countable partition of $\mathcal{R}^*$ into $s$-subsets $$\mathcal{R}^*=\cup_{n\geq 1} \hat\mathcal{R}_n$$ with the following properties:
\begin{itemize}
\item[(a)] For any $n\geq 1$, if $ \hat\mathcal{R}_n$ is nontrivial, then $F^n\hat\mathcal{R}_n$ properly return to $\mathcal{R}^*$ for the first time. Moreover, there exists $\hat\delta_0>0$ such that $$\mu(\hat\mathcal{R}_{n_1})>\hat\delta_0$$ and \begin{equation}\label{ctail8}\sum_{k=n}^{\infty} \mu(\hat\mathcal{R}_k) < C\vartheta^n.\end{equation}
\item[(b)] For each $n\geq 1$, there exists at most countably many $s$-subset $\hat\mathcal{R}_{n,i}$, $i\geq 1$, such that $$\hat\mathcal{R}_n=\cup_{i\geq 1}\hat\mathcal{R}_{n,i},$$ where $F^n\hat\mathcal{R}_{n,i}$ properly return to $\mathcal{R}^*$ for the first time, for each $i\geq 1$.\end{itemize}
\end{proposition}
\begin{proof}
We first construct a partition of $\mathcal{R}^*$ into $s$-subsets. In the Coupling Lemma \ref{coupling}, for $i=1,2$, we take
$$\mathcal{E}:=(\Gamma^u,\mu|_{\Gamma^u})=\{(W,\mu_{W}), W\in \Gamma^u, \lambda^u\},$$ as the proper family induced by the SRB measure $\mu$ restricted on the family of all unstable manifolds $\Gamma^u$, with factor measure $\lambda^u$. Then Lemma \ref{extralemma1} implies that there exists a sequence of $F-$ generalized standard families $\{ (\mathcal{W}_n,\nu_n), n\geq 1\}$, such that $\mathcal{E}=\sum_{n\geq 1}(\mathcal{W}_n,\nu_n)$, with properties given by (i)-(iii); and $\mathcal{W}_n, n\geq 1$ are disjoint sets.
Because our singular set $S_{\pm nn_1}$ contains at most countably many smooth curves, it is possible that for an unstable manifold $W\in \Gamma^u$, its image $F^n W$ contains countably many smooth components that properly returned to $\mathcal{U}^*$. This implies that for each $n\geq 1$, there exists at most countably many solid minimal rectangles $\{U_{n,k}, k\geq 1\}$, such that for each $k\geq 1$, $U_{n,k}$ is the smallest rectangle, such that $F^n(U_{n,k}\cap \mathcal{W}_n)\cap \mathcal{R}^*$ is a $u$-subset of $\mathcal{R}^*$.
We define $\mathcal{V}_n\subset \mathcal{W}_n$, such that $\mathcal{V}_n=\cup_{k=1}^{\infty} U_{n,k}\cap \mathcal{W}_n\cap \mathcal{R}^*$. Then $\{\mathcal{V}_n, n\geq 1\}$ are disjoint sets, and $\cup_{n\geq 1} \mathcal{V}_n=\mathcal{R}^*$.
We define $$\hat\Gamma^s_{n,k}=\{W^s(x)\in \Gamma^s\,:\, x\in U_{n,k}\cap \mathcal{W}_n\cap \mathcal{R}^*\}.$$ i.e. $\hat\Gamma^s_{n,k}$ is the collection of stable manifolds in $U_{n,k}\cap\Gamma^s$. We denote $\hat\Gamma^s_n=\cup_{k\geq 1} \hat\Gamma^s_{n,k}$. Then we can check that $$\Gamma^s=\cup_{n=1}^{\infty}\hat\Gamma^s_n\,\,\,\,\,(\text{mod} \,0),$$ with the following properties:
\begin{itemize}
\item[(1)] $F^n\hat{ \Gamma}^s_n$ properly return to $\mathcal{R}^*$ for the first time under $F$, and $\{\hat\Gamma^s_n, n\geq 1\}$ are almost surely disjoint $s$-subsets of $\mathcal{U}^*$ in the following sense: $$\mu(\hat\Gamma^s_m\cap\hat\Gamma^s_n)=0,$$ for any $m\neq n$;
\item[(2)] Furthermore
\begin{equation}\label{ctail4}\sum_{k=n}^{\infty} \mu(\hat\Gamma^s_k)<C\vartheta^n,\end{equation}
where $C>0$ and $\vartheta$ is the constant in (\ref{ctail}).
\end{itemize}
The fact that a nonempty set $F^{ n}(\hat{ \Gamma}^s_n)$ properly return to $\mathcal{R}^*$ is guaranteed by Proposition \ref{firstproper}.
By taking $n_1$ large enough, Lemma \ref{t01} implies the existence of $\delta'\in (0,\hat\delta_0)$ such that $$\mu(\hat\Gamma^s_{n_1})>\delta'.$$
We define $$\hat\mathcal{R}_k=\mathcal{R}^*\cap\hat\Gamma^s_k,$$ and thus $$\mathcal{R}^*=\cup_{n\geq 1} \hat\mathcal{R}_k,$$ with property (a) and (b).
\end{proof}
Note that the partition $\mathcal{R}^*=\cup_{n\geq 1} \hat\mathcal{R}_n$ induces a first proper return time $\tau:\mathcal{R}^*\to \mathbb{N}$, such that each level set $(\tau=n)=\hat\mathcal{R}_n^s$.
We also define a first proper return map $T=F^{\tau}: \mathcal{R}^*\to \mathcal{R}^*$, such that for any $n\geq 0$, for $\mu$-almost every $x\in \hat\mathcal{R}_n$, we put \begin{equation}\label{defntau}T x:=F^{n}x.\end{equation} This stopping time $\tau$ is crucial in our coupling scheme. In addition according to our definition of $T$, we know that $T$ only sees proper returns to $\mathcal{R}^*$. This is important in proving the Coupling Lemma \ref{coupling1}.
Note that one can easily build up the generalized Young Tower based on the partition $\mathcal{R}^*=\cup_{n\geq 1}\hat\mathcal{R}_n$ in the spirit of \cite{Y98,Y99}. In addition, we have a partition of the phase space $M$:
\begin{equation}\label{Yytower}
M=\bigcup_{n\geq 1}\bigcup_{k=0}^{n-1} F^k \hat\mathcal{R}_n \,\,\,( \text{ mod }\, 0).
\end{equation}
One improvement here is that according to statement (b) in the above lemma, we allow the minimal $s$-rectangle containing $\hat\Gamma^s_n$ to consist of countably many minimal $s$-rectangles $\mathcal{U}_{n,i}$. This property is due to the fact that we allow the singular set of the system to contain countably many singular curves, since one unstable manifold may be cut into infinitely many small pieces, many of which may returned to the rectangle $\mathcal{U}^*$ simultaneously. To model general systems with countable singularities, a generalized Young tower with property (b) is indeed required.
Next we investigate the relation between the set $C_{n,b}$ defined in (\ref{Cnb}) and the reference set $\mathcal{R}^*$. According to the definition of $C_{n,b}$, we know that for any $x\in C_{n,b}$, its stable manifold $W^s(x)\in C_{n,b}$. Indeed by Assumption (\textbf{H2}), we know that $$\mu(M\cap C_{n,b})= \mathcal{O}(n^{-{\alpha_0}}).$$ We would like to see similar property for standard pairs that properly cross $\mathcal{R}^*$.
\begin{lemma}\label{nuwcb2}
There exists $ c_0>0$, such that for any standard pair $(W,\nu)$ properly cross $\mathcal{R}^*$, with $d\nu/d\mu_W=g\in \mathcal{H}^+(\gamma_0)$, then for any $n\geq 1$,
\begin{equation}\label{MUWCNB1}\nu(C_{n,b}\cap W\cap\mathcal{R}^*)\leq c_0n^{-{\alpha_0}}.\end{equation}
\end{lemma}
\begin{proof}
Since Lemma \ref{firstproper} implies that $\mu(\Gamma^s)>\hat\delta_1$, by Assumption (\textbf{H2}), for any $n\geq 1$,
$$\mu(\Gamma^s\cap C_{n,b})\leq \mu(C_{n,b}\cap M)\leq C n^{-{\alpha_0}},$$for some constant $C>0$.
One can check that for $x\in C_{n,b}$, then its stable manifold $W^s(x)\in C_{n,b}$, and
$$ \mu(W^s\in C_{n,b}\cap \Gamma^s )=\mu(\Gamma^s\cap C_{n,b})\leq C n^{-{\alpha_0}}.$$
Now we disintegrate the measure $\mu$ restricted on $\mathcal{U}^*$ along unstable leaves in $\Gamma^u=\{W_{\alpha},\alpha\in \mathcal{A}\}$, and let $\lambda$ be the factor measure on the index set $\mathcal{A}$, such that $\lambda(\mathcal{A})=\mu(\Gamma^s)$, and for any measurable set $A$,
$$\mu(\Gamma^s\cap A)=\int_{\alpha}\mu_{\alpha}(W_{\alpha}\cap A\cap \Gamma^s)\,\lambda(d\alpha).$$
Picking a curve $W_{\alpha_1}\in \Gamma^u$, by the absolute continuity of the stable holonomy map \textbf{(h3)}, there exist $0<c_1<c_2$, such that for any $\alpha\in \mathcal{A}$, any measurable set $A\subset \mathcal{M}$ satisfies
\begin{align*}c_1\mu_{\alpha_1}(W_{\alpha_1}\cap A\cap \Gamma^s)&\leq \mu_{\alpha}(W_{\alpha}\cap A\cap \Gamma^s)\leq c_2\mu_{\alpha_1}(W_{\alpha_1}\cap A\cap\Gamma^s).\end{align*}
This implies that
$$\mu_{\alpha_0}(W_{\alpha_1}\cap A\cap \Gamma^s)\lambda(\mathcal{A})\leq c_1^{-1}\mu(\Gamma^s\cap A).$$
Now we take any unstable manifold $W\in \Gamma^u$, and $A=C_{n,b}$, then we have proved that
$$\mu_{W}(W\cap C_{n,b}\cap \Gamma^s)\leq C c_1^{-1}\hat\delta_1^{-1} n^{-{\alpha_0}}.$$
Since unstable manifolds in $\Gamma^s$ has length $>10\delta_0$, Lemma \ref{properagain} implies that a standard pair $(W,\nu)$ is proper whenever $W$ cross $\Gamma^s$. So $(W,\nu)$ and $(W,\mu_W)$ are equivalent:
$$\nu(W\cap C_{n,b}\cap \Gamma^s)\leq C_1C c_1^{-1}\delta_1^{-1} \|g\|_{\infty}n^{-{\alpha_0}}\leq C_1C c_1^{-1}\delta_1^{-1} e^{\varepsilon_{\mathbf{d}}}n^{-{\alpha_0}},$$
for some constant $C_1$ depending on $C_F$ in (\ref{standardpair}), where we used Lemma \ref{densitybd} in the last step estimation. Now we take $$c_0=C_1C c_1^{-1}\delta_1^{-1} e^{\varepsilon_{\mathbf{d}}},$$ then (\ref{MUWCNB1}) has been proved.
\end{proof}
\section{Markov tower for the original map}
\subsection{Construction of the generalized Young tower for the original map}
In this subsection, we will construct a countable Markov partition of $\Gamma^s$ for the nonuniformly hyperbolic map and then use the return time to $\mathcal{U}^*$ to define a stopping time for our coupling scheme. To investigate the map $(\mathcal{F}, \mathcal{M},\mu)$ based on the induced system $(F,M,\mu_M)$, we know that $F$ and $\mathcal{F}$ share the same stable/unstable manifolds on $M$ almost surely. This allows us to use the same reference set $\mathcal{R}^*$ and $\mathcal{U}^*$, as well as the stable/unstable manifolds $\Gamma^{s/u}$ that defines $\mathcal{R}^*$.
First we extend the partition according to the original map by the following construction.
\begin{proposition}\label{YtowercM}
\begin{itemize}
\item[(i)] For any $n\geq 1$, the set $$\hat\mathcal{R}_n=\bigcup_{m\geq n} \mathcal{R}_{n,m}$$ has a decomposition into s-subsets $\mathcal{R}_{n,m}$, such that for any nontrivial $\mathcal{R}_{n,m}$, the set $\mathcal{F}^m\mathcal{R}_{n,m}$ properly returns to $\mathcal{R}^*$ for the first time under $\mathcal{F}$;
\item[(ii)] $\mathcal{R}^*$ has a partition into $s$-subsets $\mathcal{R}^*=\cup_{n\geq 1} \mathcal{R}_n$, such that for any nontrivial set $\mathcal{R}_n$, the set $\mathcal{F}^n \mathcal{R}_n$ properly return to $\mathcal{R}^*$ for the first time under iterations of $\mathcal{F}$;
\item[(iii)] There exist $\hat\delta_1\in (0,\hat\delta_0)$, $n_2\geq n_1$, such that $\mu(\mathcal{R}_{n_2})>\hat\delta_1$.
\item[(iv)] For any $n\geq 1$,
$\mu_M(\cup_{m\geq n}\mathcal{R}_m)\leq C n^{-\alpha_0}.$
\end{itemize}
\end{proposition}
\begin{proof}
By Proposition \ref{Ytower}, the set $\Gamma^s$ has a countable partition into $s$-subsets $$\Gamma^s=\cup_{n\geq 1} \hat\Gamma^s_n.$$
Inductively for any $n\geq 1$, and $m\geq n$, we can define $\Gamma_{n,m}$, which is the maximal $s$-subset of $\Gamma^s$, such that $\mathcal{F}^{m} \Gamma_{n,m}$ is a $u$-subset of $\mathcal{U}^*$. Moreover $\Gamma_{n,m}$ will also properly return to $\mathcal{U}^*$ under the induced map $F^n$. Then it follows that
$$\hat\Gamma^s_n=\bigcup_{m\geq n} \Gamma_{n,m}$$ is a disjoint decomposition of $\hat\Gamma^s_n$, for any $n\geq 1$.
Next we rearrange $\{\Gamma_{n,m}\}$ according to the index $m$. Note that \begin{align*}\Gamma^s&=\bigcup_{n\geq 1}\hat\Gamma^s_n=\bigcup_{n\geq 1}\bigcup_{m\geq n}\Gamma_{n,m}=\bigcup_{m=1}^{\infty}\left(\bigcup_{n=1}^{m}\Gamma_{n,m}\right)=\bigcup_{m=1}^{\infty}\Gamma^s_m,\end{align*} where $$\Gamma_m^s=\cup_{n=1}^{m} \Gamma_{n,m}.$$ Then by the definition of $\Gamma_{n,m}$, we know that $\mathcal{F}^m \Gamma^s_m$ returns properly to $\mathcal{R}^*$, and is nonempty.
Moreover, $$\mu(\hat\Gamma_{n_1})>\hat\delta_0$$ implies that there exists $n_2\geq n_1$, such that $$\mu(\Gamma^s_{n_2})>\hat\delta_1,$$ for some $\hat\delta_1\in (0,\hat\delta_0)$.
Now we define $\mathcal{R}_m=\Gamma^s_m\cap\mathcal{R}^*$, and $\mathcal{R}_{m,n}=\Gamma^s_{m,n}\cap\mathcal{R}^*$. Therefore, the above analysis verifies items (i)-(iii).
Note that \begin{align*}
\mu_M(\cup_{m=n}^{\infty}\Gamma^s_m)&\leq \mu_M(\cup_{m=n}^{\infty}\Gamma^s_m, C_{n,b})+\mu_M(\cup_{m=n}^{\infty}\Gamma^s_m, C_{n,b}^c) \\
&\leq C n^{-\alpha_0}+\mu_M(\cup_{m=n}^{\infty}\Gamma^s_m,C_{n,b}^c).
\end{align*}
We denote $$E_n:= \cup_{m=n}^{\infty}\Gamma^s_m \cap C_{n,b}^c,$$ which contains points that have returned to $M$ at least $(b\ln n)^2$ times within $n$ iterations under $\mathcal{F}$. We claim that \begin{equation}\label{muMEn}\mu_M(E_{n})=\mathcal{O}(n^{-1-\alpha_0}).\end{equation}
This implies that $$\mu_M(\cup_{m\geq n}\mathcal{R}_m)\leq \mu_M(\cup_{m=n}^{\infty}\Gamma^s_m)\leq C_1 n^{-\alpha_0}$$ as desired.
We prove claim (\ref{muMEn}) by considering $E_{n}$, and we divide the proof into two cases.
\quad(\textbf{a}). Let $I_n$ be all points $x\in E_{n}$, such that the forward orbit of $x$ hits $\mathcal{R}^*$ at most $b\ln n$ times within $n$ iterations.
Then there exists $k\in [1,n-b\ln n]$, such that $\mathcal{F}^kx\in M$ and the forward orbit of $\mathcal{F}^k x$ return to $M\setminus \mathcal{R}^*$ at least $b\ln n $ consecutive times under $\mathcal{F}$, but not hits $\mathcal{R}^*$.
According to Lemma \ref{extralemma1}, we know that the standard family $(\mathcal{W}^u_M,\mu_M)$ has a decomposition into $\{(\mathcal{W}_n,\mu_n), n\geq 1\}$, and by (\ref{ctail1}),
$\sum_{m=n}^{\infty}\mu_m(M)\leq C \vartheta^n$.
Thus $\mathcal{F}^k x$ belongs to the support of $\sum_{m\geq b\ln n} \mu_m$, which satisfies:
\begin{align*}
\mu_M(E_{n}\cap I_n)&\leq \sum_{m=b\ln n}^{\infty}\mu_m(M)\leq C\vartheta^{b\ln n}=\mathcal{O}(n^{-1-\alpha_0}),\end{align*}
where we have used (\ref{dpbchi}) in the last step, by choosing $b$ large.
\quad(\textbf{b}). Now we consider points in $E_{n}\setminus I_n$. Then we claim that for any $x\in (E_{n}\setminus I_n)$, iterations of $x$ hit $\mathcal{R}^*$ at least $b\ln n$ times within the $(b\ln n)^2$ returns to $M$. This is true because otherwise there must be an interval of length $b\ln n$ in these at least $(b\ln n)^2$ returns to $M$ such that iterates of $x$ never hit $\mathcal{R}^*$, and this contradicts the assumption that (\textbf{a}) does not hold. Thus there exists $k\in [1,n-b\ln n]$, such that $\mathcal{F}^kx\in \mathcal{R}^*$ and the forward trajectory of $\mathcal{F}^k x$ return to $\mathcal{R}^*$ at least $b\ln n $ times. Note that $E_{n}\subset \cup_{m=n}^{\infty}\Gamma^s$, which implies that the forward trajectory of $\mathcal{F}^k x$ never return to $\mathcal{R}^*$ properly within the next $n-k$ iterations. Thus again $\mathcal{F}^k x$ belongs to the support of $ \bigcup_{m\geq b\ln n}\mu_m.$ Similar to case (\textbf{a}), we know that
$$\mu_M(E_{n}\setminus I_n)\leq C n^{-1-\alpha_0}.$$
This finished the proof of our claim (\ref{muMEn}).
\end{proof}
Let $A_n$ be the set of all points in $\mathcal{R}^*$ that returned to $\mathcal{R}^*$ under $\mathcal{F}^n$. Then
$$A_n=\{x\in \mathcal{R}^*\,:\, \tau_0(x)+\cdots+\tau_0(L^{k-1} x)=n, \text{ for some } k=1,\cdots,n\},$$ as all points in $\mathcal{R}^*$ that will return to $\mathcal{R}^*$ after $n$-iterations.
On the other hand, \begin{equation}\label{Antaun}A_n=(\tau_0=n)\cup \left(\mathcal{F}^{-1}(\tau_0=n-1)\cap A_1\right)\cup\cdots \cup \left(\mathcal{F}^{-(n-1)}(\tau_0=1)\cap A_{n-1}\right).\end{equation}
Thus for any standard family $(\mathcal{W},\nu)$, we have for any $n\geq 1$,
\begin{align}\label{Annu}
\nu(A_n)&=\sum_{k=0}^{n-1} \nu(\mathcal{F}^{-k}(\tau_0=n-k)\cap A_k)\nonumber\\
&=\sum_{k=0}^{n-1} \mathcal{F}^k_*\nu((\tau_0=n-k)\cap \mathcal{F}^kA_k)\nonumber\\
&=\sum_{k=0}^{n-1} \mathcal{F}^k_*\nu(\mathcal{R}_{n-k}^s| \mathcal{F}^kA_k)\nu(A_k). \end{align}
\begin{proposition} Let $(\mathcal{W},\nu)$ be a standard family, assume there exist $c>0$ and $N\geq 1$, such that $\nu(A_N)\geq c$. Then $\nu(A_n)\geq c$, for any $n\geq N$.
In particular, if $(\mathcal{W},\nu)$ is a generalized standard family of index zero, then $\nu(A_n)\geq \hat\delta_1\nu(\mathcal{R}^*)$ for any $n\geq 1$. \end{proposition}
\begin{proof}
We use induction on $n=N, N+1, \cdots $. Clearly, by assumption, for $n=N$, $\nu(A_N)\geq c$. We assume $\nu(A_k)\geq c$, for $k=N,\cdots, N+n-1$, then by (\ref{Annu}), we have
\begin{align*}
\nu(A_{N+n})=\sum_{k=0}^{N+n-1} \mathcal{F}^k_*\nu(\mathcal{R}_{N+n-k}^s| \mathcal{F}^kA_k)\nu(A_k) \geq c \sum_{k=N}^{N+n-1} \mathcal{F}^k_*\nu(\mathcal{R}_{N+n-k}^s| \mathcal{F}^kA_k)\geq c.\end{align*}
This proves the first statement. Note that $A_1=\mathcal{R}_1$. If $(\mathcal{W},\nu)$ is a generalized standard family of index zero, then $\nu(A_1)\geq \hat\delta_1\nu(\mathcal{R}^*)$, according to Proposition \ref{YtowercM}. Thus by statement 1, we know that $\nu(A_n)\geq \hat\delta_1\nu(\mathcal{R}^*)$ for any $n\geq 1$.
\end{proof}
\begin{comment}
The following facts for renewal sequences were proved in a series of references, starting from Kolmogorov, \cite{Kol, Gel, Sto}, and mainly in \cite{Rog}.
\begin{lemma}\label{renewals} Let $\{p_k,k\geq 1\}$ be a nonnegative sequence such that $$\sum_{k=1}^{\infty}p_k=1.$$ Let $a_0=1$ and for any $n\geq 1$, we set $a_n=\sum_{k=1}^n p_k a_{n-k}$. If $$\lambda:=\sum_{k=1}^{\infty} k p_k<\infty,\,\,\,P_n:=\sum_{k\geq n} p_n=\mathcal{O}(n^{-\alpha})$$ and
$$\xi:=\max\{\eta\geq 1\,:\, \sum_{k=1}^{\infty} k^{\eta} p_k<\infty\}$$ with $\xi\geq 1$, then
$$a_n=\frac{1}{\lambda}+\frac{1}{\lambda^2}\sum_{k\geq n}\sum_{m=k+1}^{\infty} p_m +R_n(\xi),$$
where $R_n(\xi)=\mathcal{O}(n^{-\xi})$ if $\xi\geq 2$; and $R_n(\xi)=\mathcal{O}(n^{-2(\xi-1)})$ if $\xi\in (1,2)$.$$R_n=\left\{
\begin{array}{ll}
\mathcal{O}((\ln n)^2), & \text{ for }\, \alpha=\xi=1;\\
\mathcal{O}(n^{-2(\alpha-1)}), & \text{for } \, 1=\xi<\alpha<2; \\
\mathcal{O}(n^{-2}\ln n), & \text{for } \, \xi=1, \alpha=2;\\
\mathcal{O}(n^{-\alpha}), & \text{for } \,\xi>2, \alpha>2.
\end{array}
\right.$$
\end{lemma}
\vspace{0.3in}
We define $$p_k=\mu(\Gamma^s_k|\mathcal{R}^*)$$ for any $k\geq 1$. By Proposition \ref{YtowercM}, we know that $\sum_{k=n}^{\infty} p_k\leq C n^{-\alpha_0}$, for any $n\geq 1$. We let $\alpha=\alpha_0>1$, then one can check that $\mu(\tau^{\eta})<\infty$, for any $\eta<\alpha_0$. Thus $\xi<\alpha_0$. We first prove a lemma.
\begin{lemma} Let $(\mathcal{W},\nu)$ be a standard family with index zero, and $\nu(\mathcal{M})=1$. Then the total amount of measure of $(\mathcal{W},\nu)$ that properly returned to $\mathcal{R}^*$ after $\mathcal{F}^n$-iterations, which is denoted as $a'_n$, for any $n\geq 1$, satisfies
$a'_n\geq \nu(\mathcal{R}^*)/2$.
\end{lemma}
\begin{proof}
Let $a_0=1$ and for any $n\geq 1$, we define $$a_n=\sum_{k=1}^n p_k a_{n-k}.$$ Moreover, we denote $$\lambda:=\sum_{k=1}^{\infty} k p_k=1/\nu(\mathcal{R}^*)<\infty,$$ and by the assumption that $\nu(\tau^{\xi})<\infty$, we know that $$\sum_{k=1}^{\infty} k^{\xi} p_k<\nu(\tau^{\xi})<\infty.$$ Then the above renewal theory implies that
\begin{equation}\label{estan}a_n=\frac{1}{\lambda}+\frac{1}{\lambda^2}\sum_{k\geq n}\sum_{m=k+1}^{\infty} p_m +R_n(\xi),\end{equation}
where $R_n(\xi)=\mathcal{O}(n^{-\xi})$ if $\xi\geq 2$; and $R_n(\xi)=\mathcal{O}(n^{-2(\xi-1)})$ if $\xi\in [1,2)$.
We define $\mathcal{D}_{k,n}$ as the set of all points in $\mathcal{R}^*$ that return to $\mathcal{R}^*$ exactly $k$ times under $\mathcal{F}^n$, see (\ref{Dnm}). Then one can check that
$$\mathcal{D}_{k,n}=\{x\in \mathcal{R}^*\,:\, \tau(x)+\cdots+\tau(T^{k-1}x)=n\}.$$
Let $A_n$ be the set of all points that returned to $\mathcal{R}^*$ under $\mathcal{F}^n$. Then
$$A_n=\{x\in \mathcal{R}^*\,:\, \tau(x)+\cdots+\tau(T^{k-1} x)=n, \text{ for some } k=1,\cdots,n\}=\bigcup_{k=1}^{n}\mathcal{D}_{k,n},$$ as all points in $\mathcal{R}^*$ that will return to $\mathcal{R}^*$ after $n$-iterations.
We define the measure $\hat{\mu}:=\nu|_{\mathcal{R}^*}/\nu(\mathcal{R}^*)$. For any $k\geq 1$, we define the conditional measure on $\Gamma^s_k$ obtained by $\hat{\mu}$ as
$$\eta_k:=\frac{\hat\mu|_{\Gamma^s_k}}{\hat{\mu}(\Gamma^s_k)}.$$
Then
$$\mathcal{F}^k_*\eta_k(B)=\eta_k(\mathcal{F}^{-k}B)=\frac{\nu(\mathcal{F}^{-k}B\cap\Gamma^s_k)}{\nu(\Gamma^s_k)}$$
We denote $$\varepsilon_{n-k}:=\mathcal{F}^k_*\eta_k(A_{n-k})-\hat{\mu}(A_{n-k}).$$
Using the fact that the support of $\mathcal{F}^k_*\eta_k$ is properly return to $\mathcal{R}^*$, and Proposition \ref{estAnnn}, we get for any $k\leq n$,
\begin{equation}\label{cFkAn-k}|\varepsilon_{n-k}|\leq C_{\mathbf{d}}\hat{\mu}(A_{n-k}) \vartheta_1^{2}\end{equation}
We claim that
\begin{equation}\label{claimAn}A_n=(\Gamma^s_1\cap \mathcal{F}^{-1}A_{n-1})\cup (\Gamma^s_2\cap \mathcal{F}^{-2}A_{n-2})\cup\cdots\cup (\Gamma^s_{n-1}\cap \mathcal{F}^{-(n-1)}A_1)\cup \Gamma^s_n \end{equation}
For any $x\in A_n$, since $A_n\subset \mathcal{R}^*\subset \cup_{m\geq 1} \Gamma^s_m$, then there exists $m\geq 1$, such that $x\in A_n\cap \Gamma^s_m$. Note that $\mathcal{F}^m \Gamma^s_m$ is a $u$-subset of $\mathcal{R}^*$, so $\mathcal{F}^m x\in \mathcal{R}^*$. Using the definition of $A_n$, we know that $\mathcal{F}^n x\in \mathcal{R}^*$, which implies that $\mathcal{F}^{n-m}\mathcal{F}^m x\in \mathcal{R}^*$. Thus $\mathcal{F}^m x\in A_{n-m}$. This implies that
$$A_n\subset \cup_{m=1}^n(\Gamma^s_m\cap \mathcal{F}^{-m}A_{n-m}) $$
where we denote $A_0=\mathcal{R}^*$.
On the other hand, for any $m=1,\cdots, n$, any $x\in \Gamma^s_m$, if $\mathcal{F}^m x\in A_{n-m}$, then clearly $x\in A_n$.
Now it follows from (\ref{claimAn}) that
$$\hat{\mu}(A_n)=\sum_{m=1}^n \hat{\mu}(\Gamma^s_m\cap \mathcal{F}^{-m}A_{n-m})=\sum_{m=1}^n \hat{\mu}(\Gamma^s_m) \mathcal{F}^m_*\eta_m(A_{n-m})=\sum_{m=1}^n p_m\, \mathcal{F}^m_*\eta_m(A_{n-m})$$
Using (\ref{cFkAn-k}), we get
\begin{equation}\label{Anpnepsn}\hat{\mu}(A_n)=\sum_{m=1}^n p_m\, \hat{\mu}(A_{n-m})(1-\varepsilon_{n-m})\end{equation}
which implies that
$$\hat{\mu}(A_n)\geq \sum_{m=1}^n p_m\, \hat{\mu}(A_{n-m})(1-0.01\vartheta_1)$$
We define $a_n'=\hat{\mu}(A_n)$, then (\ref{Anpnepsn}) implies that
$$a_n'=\sum_{m=1}^n p_m\, a_{n-m}'+\sum_{m=1}^n p_m\, a_{n-m}'\,\varepsilon_{n-m}$$
Similar to \cite{Rog}, we denote
$\mathbf{R}_+(\mathcal{O}(n^{-\alpha}L(n)))$ (resp. $\mathbf{R}_+(o(n^{-\alpha}L(n)))$) as the set of all absolutely convergence series $x(z):=\sum_{n=0}^{\infty} x_n z^n$ for $|z|\leq 1$, such that $x_n=\mathcal{O}(n^{-\alpha}L(n))$ (resp. $x_n=o(n^{-\alpha}L(n))$); or in other words, $$\lim_{n\to\infty} |x_n| /(n^{-\alpha}L(n))<\infty,\, \text{(resp.} \lim_{n\to\infty} |x_n| /(n^{-\alpha}L(n))=0)$$ It was proved in \cite{Rog}~Lemma 2 that both sets are closed under addition, multiplication, and multiplication by a constant of
their generating functions. Moreover, if $x(z)\neq 0$ for $|z|\leq 1$, then we define $x(z)^{-1}=\lambda(z)=\sum_{n=0}^{\infty}\lambda(n)z^n$. It was shown in \cite{Rog}~Lemma 3 that
$$x(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha}L(n)))\Rightarrow \lambda(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha}L(n)));$$$$ x(z)\in \mathbf{R}_+(o(n^{-\alpha}L(n))) \Rightarrow \lambda(z)\in \mathbf{R}_+(o(n^{-\alpha}L(n)).$$
Here $\alpha\geq 1$, and for $\alpha=1$, $L(n)=\mathcal{O}(1)$.
For any complex number $z\in \mathbb{C}$, with $|z|\leq 1$, we define
$$A(z):=1+\sum_{n=1}^{\infty} a'_n z^n,\,\,\,\,\,\,p(z)=\sum_{n=1}^{\infty}p_n z^n,\,\,\,\,\varepsilon(z)=\sum_{n=1}^{\infty} \varepsilon_n z^n$$
Using (\ref{Anpnepsn}), we get
\begin{align}\label{Aza'}
A(z)&=\sum_{n=0}^{\infty}a'_n z^n\nonumber\\
&=1+\sum_{n=1}^{\infty}\sum_{m=1}^n p_m\, \hat{\mu}(A_{n-m})\,z^n+\sum_{n=1}^{\infty}\sum_{m=1}^n p_m\,\varepsilon_{n-m}\,z^n\nonumber\\
&=1+\sum_{m=1}^{\infty} \sum_{n=m}^{\infty} p_m\,z^m \hat{\mu}(A_{n-m})\,z^{n-m}+\sum_{m=1}^{\infty} \sum_{n=m}^{\infty} p_m\,z^m\,\hat{\mu}(A_{n-m})\varepsilon_{n-m}\,z^{n-m}\nonumber\\
&=1+\sum_{m=1}^{\infty} p_m\,z^m \sum_{k=0}^{\infty} \hat{\mu}(A_{k})\,z^{k}+\sum_{m=1}^{\infty} p_m\,z^m\, \sum_{k=0}^{\infty} \hat{\mu}(A_{k})\varepsilon_{k}\,z^{k}\nonumber\\
&=1+p(z) A(z)+p(z)A(z)\varepsilon(z)
\end{align}
We define
$$P(z)=\frac{1-p(z)}{1-z}=\sum_{n\geq 0} P_n z^n$$
Then one can check that $P_n=\sum_{k\geq n+1} p_k$, and $$P(1)=\sum_{n\geq 0} P_n =\sum_{n\geq 0}\sum_{k=n+1} p_k=\lambda$$
We claim that $P(z)\neq 0$ for any $|z|\leq 1$ with $z\neq 1$. Indeed if $|z|<1$, then
$$|p(z)|=|\sum_{n=1}^{\infty} p_n z^n|< \sum_{n=1}^{\infty} p_n=1$$
This implies that $|P(z)|>0$.
We assume $z=e^{i\theta}$, for $\theta\in [0,2\pi]$. Then
$$p(z)=\sum_{n=1}^{\infty} p_n \cos n\theta+i\sum_{n=1}^{\infty} p_n \sin n\theta $$
Note that $p(z)=0$ if and only if $\sum_{n=1}^{\infty} p_n \sin n\theta=0$ and $\sum_{n=1}^{\infty} p_n \cos n\theta=1$.
However, the second condition can be written as
\begin{equation}\label{secondII}\sum_{n=1}^{\infty} p_n \cos n\theta=\sum_{n=1}^{\infty} p_n \end{equation}
which implies that
$$\sum_{n=1}^{\infty} p_n (1-\cos n\theta)=0$$
Note that if $z=e^{i\theta} \neq 1$, then there exists $k\geq 1$, such that $|\cos k\theta|<1$, this implies that
$$\sum_{n=1}^{\infty} p_n (1-\cos n\theta)>p_k (1-\cos k\theta)$$
This contradicts the second condition (\ref{secondII}). Thus $p(z)=1$ for $|z|\leq 1$ if and only if $z=1$, as we have claimed.
By Proposition \ref{YtowercM}, we know that $\sum_{k=n}^{\infty} p_k\leq C n^{-\alpha_0}$, for any $n\geq 1$. This implies that
$P(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha_0}))$, which implies that $P(z)^{-1}\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha_0}))$.
(\ref{Aza'}) implies that
\begin{align*}
A(z)\cdot \frac{(1-p(z)(1+\varepsilon(z))}{1-z}&=\frac{1}{1-z}
\end{align*}
We multiple $P(z)^{-1}$ on both sides, to get for $|z|<1$,
\begin{align*}
A(z)&=\frac{1+p(z)\varepsilon(z)}{1-z}\cdot P(z)^{-1}=(1-p(z))^{-1}(1+p(z)\varepsilon(z))\\&\geq (1-p(z))^{-1}(1-\varepsilon_{\mathbf{d}})\end{align*}
According to \cite{Rog}~Theorem 1, the expression $(1-p(z))^{-1}$ can be represented as
$$(1-p(z))^{-1}=\sum_{n=0}^{\infty}\left(\frac{1}{\lambda}+\frac{1}{\lambda^2}\sum_{k\geq n} P_k+R_n\right)z^n,$$
where
$$R_n=\left\{
\begin{array}{ll}
\mathcal{O}(n^{-2(\alpha_0-1)}), & \text{for } \, 1=\xi<\alpha_0<2; \\
\mathcal{O}(n^{-2}\ln n), & \text{for } \, \xi=1, \alpha_0=2;\\
\mathcal{O}(n^{-\alpha_0}), & \text{for } \,\xi>2, \alpha_0>2.
\end{array}
\right.$$
Now using (\ref{Azeps}), we get for any $|z|\leq 1$,
$$A(z)=\sum_{n=0}^{\infty} a_n'z^n\geq (1-p(z))^{-1}(1-\varepsilon_{\mathbf{d}})=(1-\varepsilon_{\mathbf{d}})\sum_{n=0}^{\infty}\left(\frac{1}{\lambda}+\frac{1}{\lambda^2}\sum_{k\geq n} P_k+R_n\right)z^n$$
which implies that for any $n\geq 0$,
$$a_n'\geq (1-\varepsilon_{\mathbf{d}})\left(\frac{1}{\lambda}+\frac{1}{\lambda^2}\sum_{k\geq n} P_k+R_n\right)\geq \frac{1-\varepsilon_{\mathbf{d}}}{\lambda}$$
\end{comment}
\subsection{Proper returns to the base of the tower}
In the next sections, we will also consider the images of a standard family $(\mathcal{W},\nu)$ under iterations of the map $\mathcal{F}$. Indeed one can also show that $\mathcal{F}^n (\mathcal{W},\nu)$ essentially becomes a proper family as long as $n$ is large.
\begin{lemma}\label{propercF} Fix any $\delta\in(0,\mu(\mathcal{R}^*)/3)$. Let $\mathcal{G}=(\mathcal{W},\nu)$ be any standard family. Then there exists $N_1=N_1(\nu)\geq 1$, such that $\mathcal{F}^{n}\mathcal{G}$ has at least $\delta$ portion of the measure that fully returned to $\mathcal{R}^*$, for any $n\geq N_1$.
\end{lemma}
\begin{proof}
Since by the mixing property, we know that $\mathcal{F}^n_*\nu\to \mu$ weakly. Moreover, as $(\mathcal{W}^u,\mu)$ is a proper family, one can check that there exists $N_1\geq 1$, such that $\mathcal{F}^{n}\mathcal{G}$ is also proper, as $n\geq 1$. The second statement also follows from the mixing property and the fact that $\mu(\mathcal{R}^*)>0$.
\end{proof}
\begin{remark}\label{remark2}
By Proposition \ref{YtowercM}, we have $\mu(\Gamma^s_{n_2})>\hat\delta_1$; and by Proposition \ref{Ytower}, we have $\mu_M(\hat\Gamma^s_{n_1})>\hat\delta_0.$
\end{remark}
Similar to (\ref{Yytower}), we can decompose $\mathcal{M}$ as:
\begin{equation}\label{YytowercM}
\mathcal{M}=\bigcup_{n\geq 1}\bigcup_{k=0}^{n-1} \mathcal{F}^k \mathcal{R}_n \,\,\,( \text{ mod }\, 0).
\end{equation}
Now we define for any $n\geq 1$, the set \begin{equation}\label{levelset}
\mathcal{W}^u_n:=\bigcup_{m=n}^{\infty} \mathcal{F}^{m-n} \mathcal{R}_m.
\end{equation}
Then one can check that each set $\mathcal{W}^u_n$ has the property that $\mathcal{F}^n \mathcal{W}^u_n$ return to $\mathcal{R}^*$ properly for the first time. This also enable us to extend $\tau_0$ from $\mathcal{R}^*$ to the full phase space $\mathcal{M}$, almost surely, such that $\mathcal{W}^u_n$ is the $n$-th level set of $\tau_0$, i.e. $(\tau_0=n)=\mathcal{W}^u_n$.
\begin{lemma}\label{extralemma2}
Let $\mathcal{G}=(\mathcal{W},\nu)$ be a standard family. Then there exists a sequence of $\mathcal{F}-$ generalized standard families $\{(\mathcal{W}_n,\nu_n), n\geq 0\}$, such that $\mathcal{G}=\sum_{n=0}^{\infty}(\mathcal{W}_n,\nu_n)$; moreover, for any measurable collection of stable manifolds $A$, we have
$$\nu(A)=\sum_{n\geq 0} \nu_n(A).$$
In addition, $$\nu_n(\mathcal{M})\leq C \|g\|_{\infty}n^{-\alpha_0},$$ for some constant $C>0$.
\end{lemma}
\begin{proof}
We first extend the first return time function $\tau_0$ from $\mathcal{R}^*$ to $\mathcal{M}$, as in (\ref{YytowercM}). For any $n\geq 0$, let
$\mathcal{W}^u_n=(\tau_0=n)$ be its $n$-th level set, as in (\ref{levelset}). We define $\mathcal{E}=(\mathcal{W}^u,\mu)$ as the proper family generated by the invariant SRB measure $\mu$ on unstable manifolds $\mathcal{W}^u$. For any $n\geq 0$, let
$$\mathcal{W}^u_n=(\tau_0=n),\,\,\,\,\mu_n=\mu|_{(\tau_0=n)}.$$
Then one can check that $(\mathcal{W}^u_n,\mu_n)$ is a $\mathcal{F}-$ generalized standard family of index $n$, and
$$(\mathcal{W}^u,\mu)=\sum_{n\geq 0}(\mathcal{W}^u_n,\mu_n).$$
In addition, Item (iv) of Proposition \ref{YtowercM} implies that there exists a constant $C>0$ such that
$$\mu_n(\mathcal{M})\leq C n^{-\alpha_0}.$$
Next, we consider $\mathcal{G}=(\mathcal{W},\nu)$, which is a standard family, such that $\mathcal{W}\subset \mathcal{W}^u$ is a measurable partition of a Borel set $B\subset \mathcal{M}$ into unstable manifolds, and $\mu(B)>$, $g=d\nu/d\mu\in\mathcal{H}^+(\gamma_0)$. Then we define $\mathcal{W}_n:=\mathcal{W}^u_n\cap \mathcal{W}$, and $\nu_n=\nu|_{\mathcal{W}_n}$. Then $(\mathcal{W}_n,\nu_n)$ is a $\mathcal{F}-$ generalized standard family of index $n$.
Note that $\nu_n(\mathcal{M})\leq \|g\|_{\infty}\mu_n(\mathcal{M})$.
Using Proposition \ref{YtowercM}, we know that $$\nu_n(\mathcal{M})\leq C \|g\|_{\infty} n^{-\alpha_0}.$$
This verifies the two statements as claimed.
\end{proof}
\begin{comment}
By Lemma
\ref{extralemma1}, there exists a sequence of $F-$ generalized standard families $\mathcal{V}_{n,k}:=\{(V_{n,k}, \eta_{n,k}), k\geq 1\}$ associated with $\mathcal{E}_n$.
According to (\ref{ctail1}), we know that
\begin{equation}\label{etakn}\frac{\eta_{n,k}(\mathcal{M})}{\tilde\nu_n(\mathcal{M})}=\tilde\nu_n(V_{n,k}|\tilde\mathcal{W}_n)\leq C\vartheta^k\end{equation} for some constants $C>0$ and $\vartheta\in (0,1)$.
Moreover, we decompose $\{(V_{n,k}, \eta_{n,k}), k\geq 1\}$, according to its iterations under the original map $\mathcal{F}$, such that for any $m\geq k$, $$F^k(V_{n,k,m}, \nu_{n,k,m})=\mathcal{F}^m(V_{n,k,m},\nu_{n,k,m}).$$ Here $V_{n,k,m}$ is the support of $\nu_{n,k,m}$, and $F^k V_{n,k,m}=\mathcal{F}^m V_{n,k,m}$ is a $u$-subset of $\mathcal{U}^*$. We claim that:
\begin{equation}\label{vnkm}
\frac{\nu_{n,k,m}(\mathcal{M})}{\eta_{n,k}(\mathcal{M})}=F^k_*\eta_{n,k}(F^k V_{n,k,m}|F^kV_{n,k}))\leq C m^{-\alpha_0}
\end{equation}
Proof of the claim:
For convenience, we denote the probability measure $\tilde\eta_{n,k}:=F^k_*\eta_{n,k}(\,\cdot\, |F^kV_{n,k})$. Note that by definition, $ \tilde\eta_{n,k}$ is a probability measure fully support on $F^k V_{n,k}$ (that is fully across $\mathcal{R}^*$), thus $\tilde\eta_{n,k}(R>j)=0$, for $j$ large.
Note that
\begin{align*}
\sum_{m=j}^{\infty}\tilde\eta_{n,k}(V_{n,k,m})&\leq \tilde\eta_{n,k}(R\geq j)+ \tilde\eta_{n,k}(\cup_{m=j}^{\infty}V_{n,k,m}, R\leq j, C_{j,b})\\
&+\tilde\eta_{n,k}(\cup_{m=j}^{\infty}V_{n,k,m}, R\leq j, C_{j,b}^c) \\
&\leq \tilde\eta_{n,k}(\cup_{m=j}^{\infty}V_{n,k,m}, R\leq j, C_{j,b}^c)+\tilde\eta_{n,k}( R\leq j, C_{j,b}) \\
&\leq C_1 j^{-{\alpha_0}}+ C_2j^{-1-{\alpha_0}}\leq 2C_1 j^{-{\alpha_0}}
\end{align*}
Here in the last step of the above estimations, we have used the fact that the set $$E_j:=\{x\in M\,:\, \cup_{m=j}^{\infty}V_{n,k,m}, R<j, C_{j,b}^c\}$$ contains points that have returned to $M$ at least $b\ln j$ times within $j$ iterations, but have not returned to $\mathcal{R}^*$ yet. According to Lemma \ref{extralemma1}, we know that the standard family $(\mathcal{W}^u,\mu)$ has a decomposition into $\{(\mathcal{W}_j,\mu_j), j\geq 1\}$, and by (\ref{ctail1}),
$\sum_{m=j}^{\infty}\mu_m(M)\leq C \vartheta^j$.
Thus $E_j$ belongs to the support of $\sum_{m\geq j} \mu_m$, which satisfies:
\begin{align*}
\mu(\cup_{m=j}^{\infty}V_{n,k,m}, R<j, C_{j,b}^c)&=\sum_{m=b\ln j}^{\infty}\mu_m(R<j, C_{j,b}^c)\\
&\leq C\vartheta^{b\ln j}=\mathcal{O}(j^{-1-\alpha_0}).\end{align*}
Since $\tilde\eta_{n,k}$ is equivalent to $\mu$, the above estimation hold for $\tilde\eta_{n,k}$.
This proof the claim.
We denote $\mathcal{V}_{n,k,m}=(V_{n,k,m},\nu_{n,k,m})$, and define $$\tilde\mathcal{V}_{k,m+n}^n:=\mathcal{F}^{-n}\mathcal{V}_{n,k,m}=(\mathcal{F}^{-n}V_{n,k,m},\mathcal{F}^{-n}_*\nu_{n,k,m}).$$ Now
\begin{align*}
\mathcal{G}&=\sum_{n=0}^{\infty}\mathcal{G}_n=\sum_{n=0}^{\infty}\sum_{k=1}^{\infty}\sum_{m=k}^{\infty}\tilde\mathcal{V}_{k,m+n}^n\\
&=\sum_{n=0}^{\infty}\sum_{m=1}^{\infty}\sum_{k=1}^m \tilde\mathcal{V}_{k,m+n}^n =\sum_{n=0}^{\infty}\sum_{l=n+1}^{\infty}\sum_{k=1}^{l-n}\tilde\mathcal{V}_{k,l}^n\\
&=\sum_{l=1}^{\infty}\sum_{n=0}^{l}\sum_{k=1}^{l-n}\tilde\mathcal{V}_{k,l}^n=\sum_{l=0}^{\infty}(\hat\mathcal{W}_l,\hat\nu_l)
\end{align*}
where $$(\hat\mathcal{W}_l,\hat\nu_l)=\sum_{k=1}^l\sum_{n=1}^k \tilde\mathcal{V}_{k,l}^n=(\cup_{k=1}^l\cup_{n=1}^k\mathcal{F}^{-n}V_{n,k,l-n},\sum_{k=1}^l\sum_{n=1}^k\mathcal{F}^{-n}_*\nu_{n,k,l-n})$$ Then
by the definition of $\hat\nu_k$, we know that $\mathcal{F}^{k} (\hat\mathcal{W}_k)$ is also an (union of) $u$-subset that properly returned to $\mathcal{U}^*$.
In addition, note that
\begin{align*}
\hat\nu_l(\mathcal{M})&\leq \sum_{k=1}^l\sum_{n=1}^k\mathcal{F}^{-n}_*\nu_{n,k,l-n}(\mathcal{M})=\sum_{k=1}^l\sum_{n=1}^k\nu_{n,k,l-n}(\mathcal{M})\\
&\leq C\sum_{k=1}^l\sum_{n=1}^k \vartheta^k n^{-\alpha_0} (l-n)^{\alpha_0}\leq C_1 \sum_{k=1}^l\vartheta^k (l-k)^{-\alpha_0} \leq C_2l^{-\alpha_0},
\end{align*}
where we used (\ref{etakn}) and (\ref{vnkm}) in the above estimations.
\end{comment}
To make comparison, we need to construct a generalized standard family of index zero on $\mathcal{R}^*$. Let $\mu^*=\mu|_{\mathcal{R}^*}/\mu(\mathcal{R}^*)$ be the conditional measure of $\mu$ on $\mathcal{R}^*$. Then $(\mathcal{R}^*,\mu^*)$ can be viewed as a generalized standard family of index zero. Moreover, $(\mathcal{R}^*,\mu^*)$ can be decomposed into a sequence of $\mathcal{F}-$ generalized standard families, i.e.
$(\mathcal{R}^*,\mu^*)=\sum_{n=1}^{\infty} (\mathcal{R}_n, \mu^*_n)$
with $\mu^*_n=\mu^*|_{\mathcal{R}_n}$, and $(\mathcal{R}_n,\mu^*_n)$ is a generalized standard family of index $n$.\\
\begin{lemma}\label{Tmerateo} Let $\mathcal{G}^i=(\mathcal{W}^i,\nu^i)=((W_{\alpha},\nu_{\alpha}),\alpha\in \mathcal{A}^i, \lambda^i)$ , $i=1,2$, be two generalized families of index $0$, such that $\mathcal{W}^i=\mathcal{W}^i\cap \mathcal{R}^*$ properly cross $\mathcal{R}^*$. Assume that $\nu^1(\mathcal{R}^*)=\nu^2(\mathcal{R}^*)$.\\
(i) For any $m\geq n\geq 1$ we have
\begin{equation}\label{nu1gammanmm1}|\nu^1(\Gamma_{n,m})-\nu^2(\Gamma_{n,m})|\leq \varepsilon_{\mathbf{d}} \nu^2(\Gamma_{n,m}).\end{equation}
Here $\Gamma_{n,m}\subset \Gamma^s$, such that $F^n \Gamma_{n,m}=\mathcal{F}^m \Gamma_{n,m}$ is a $u$-subset of $\mathcal{R}^*$.\\
(ii) For any $n\geq 1$, we have
$$|\nu^1(\mathcal{R}_n)-\nu^2(\mathcal{R}_n)|\leq \varepsilon_{\mathbf{d}}\nu^2(\mathcal{R}_n).$$
\end{lemma}
\begin{proof}
Note that (i) implies (ii) according to Proposition \ref{YtowercM}, as $\nu^i(\mathcal{R}_n)=\sum_{k=1}^{n}\nu^i(\Gamma_{k,n})$, for $i=1,2$. Thus $$|\nu^1(\mathcal{R}_n)-\nu^2(\mathcal{R}_n)|\leq \varepsilon_{\mathbf{d}} \sum_{k=1}^n \nu^2(\Gamma_{k,n})\leq \varepsilon_{\mathbf{d}}\nu^2(\mathcal{R}_n).$$
Now it suffices to prove statement (i) .
Let $\mathcal{G}^i=(\mathcal{W}^i,\nu^i)=((W_{\alpha},\nu_{\alpha}),\alpha\in \mathcal{A}^i, \lambda^i),$ $i=1,2$, be two pseudo generalized families of index $0$.
For any $\alpha\in \mathcal{A}^i$, let $W^u_{\alpha}$ be the (unique) unstable manifold in $\Gamma^u$ that contains $W_{\alpha}$; and $\nu^u_{\alpha}$ a regular measure on $W_{\alpha}^u$, with $\nu^u_{\alpha}|_{\Gamma^s}=\nu_{\alpha}$. Then we can start from the standard family $$\hat{\mathcal{G}}^i:=((W^u_{\alpha},\nu^u_{\alpha}),\alpha\in \mathcal{A}^i, \lambda^i).$$ Now we can use the assumption (\ref{Jh}) on distortion bounds for the Jacobian of the stable holonomy map defined by $\Gamma^s$. More precisely, for any $\alpha\in \mathcal{A}^1$, $\beta\in \mathcal{A}^2$, we define $$\mathbf{h}_{\alpha,\beta}: W_{\alpha}\to W_{\beta}$$ as the stable holonomy map, with $$\mathbf{h}_{\alpha,\beta}(x)=d\nu^u_{\alpha}/d\nu^u_{\beta},$$ for any $x\in W_{\alpha}\cap \Gamma^s$. Then by the absolute continuity property of the holonomy map, especially (\ref{Jh}), as well as the fact that $\text{\rm dist}(W_{\alpha}, W_{\beta})\leq 20\delta_0$, for any $W_{\alpha},W_{\beta}\in \Gamma^u$; thus we have $$|\ln \mathbf{h}_{\alpha,\beta}|\leq C_{\mathbf{r}} \text{\rm dist}(W_{\alpha}, W_{\beta})^{\gamma_0}\leq C_{F} (20\delta_0)^{\gamma_0}<\varepsilon_{\mathbf{d}},$$ by the choice of $\delta_0$ as in Lemma \ref{densitybd}, as well as the definition of $\vartheta_1$ in (\ref{vartheta1}). Thus for any measurable collection of stable manifolds $A\subset \Gamma_{n,m}$, any standard pair $( W^u_{\alpha}, \nu^u_{\alpha})$ and $( ( W^u_{\beta}, \nu^u_{\beta}))$, we have
\begin{equation}\label{alphbeta1}|\nu^u_{\alpha}(A\cap W^u_{\alpha})-\nu^u_{\beta}(A\cap W^u_{\beta})|\leq \varepsilon_{\mathbf{d}}\,\nu^u_{\beta}(A\cap W^u_{\beta}).\end{equation}
Since $\nu^u_{\alpha}|_{\Gamma^s}=\nu_{\alpha}$ and $\nu^u_{\beta}|_{\Gamma^s}=\nu_{\beta}$, we indeed have
\begin{equation}\label{alphbeta}|\nu_{\alpha}(A\cap W^u_{\alpha})-\nu_{\beta}(A\cap W^u_{\beta})|\leq \varepsilon_{\mathbf{d}}\,\nu_{\beta}(A\cap W^u_{\beta}).\end{equation}
Since (\ref{alphbeta}) is true for all $\alpha\in \mathcal{A}^1$ and all $\beta\in \mathcal{A}^2$, using the fact that $\lambda^i(\mathcal{A}^i)=1$, we have,
$$|\nu^1(\Gamma_{n,m})- \nu^2(\Gamma_{n,m})|\leq \varepsilon_{\mathbf{d}}\nu^1(\Gamma_{n,m}).$$
\end{proof}
\begin{comment}
\begin{lemma}\label{Tmerateo1} Let $$\mathcal{G}^i=(\mathcal{W}^i,\nu^i)=((W_{\alpha},\nu_{\alpha}),\alpha\in \mathcal{A}^i, \lambda^i),$$ $i=1,2$, be two pseudo-generalized standard families of index zero, as defined in Definition \ref{pseodugs}, such that $\mathcal{W}^i$ properly cross $\mathcal{R}^*$. Assume that both $\nu^1$ and $\nu^2$ are probability measures.
For any $ n\geq 1$ and $k\geq n$, we have
\begin{equation}\label{nu1gammanmm}|\nu^1(\mathcal{D}_{n,k})-\nu^2(\mathcal{D}_{n,k})|\leq \varepsilon_{\mathbf{d}}\nu^2(\mathcal{D}_{n,k}) \vartheta_1^{n+1},\end{equation}
for some uniform constant $C_1>0$ and $\vartheta_1\in (0,1)$.
\end{lemma}
\begin{proof}
Let $\mathcal{G}^i=(\mathcal{W}^i,\nu^i)=((W_{\alpha},\nu_{\alpha}),\alpha\in \mathcal{A}^i, \lambda^i),$ $i=1,2$, be two pseudo generalized families of index $0$. Then it follows from the definition that $T\mathcal{G}^i$ are both generalized standard families of index $0$.
We fix any $n\geq 1$, for any $k\geq n$, let $U_{n,k}$ be the minimal rectangle containing $\mathcal{D}_{n,k}$, with $$T^n U_{n,k}=\mathcal{F}^k U_{n,k}$$ being a $u$- subset properly crosses $\mathcal{U}^*$.
For any $\alpha\in \mathcal{A}^i$, let $$W_{\alpha,k}=W_{\alpha}\cap U_{n,k},$$ and $$\mathcal{W}^i_{n,k}:=\{W_{\alpha,k}, \alpha\in \mathcal{A}^i\}.$$ Then we can start from the standard family $$\mathcal{G}^i_{n,k}:=( \mathcal{W}^i_{n,k}, \nu^i|_{_{U_{n,k}}}).$$
Note that $T^n\mathcal{D}_{n,k}$ fully crosses $\mathcal{R}^*$ and
\begin{align}\label{Gammanmksame}
\int_{\alpha\in \mathcal{A}^i}\nu_{\alpha}(\mathcal{D}_{n,k}) \,\lambda^i(d\alpha)&=\nu^i(\mathcal{D}_{n,k})=T^n_*\nu^i(T^n \mathcal{D}_{n,k})\nonumber\\
&= T^n_*(\nu^i|_{_{U_{n,k}}})(\mathcal{R}^*)=T^n_*\nu^i(T^n \mathcal{D}_{n,k})\nonumber\\
&=\int_{\alpha\in \mathcal{A}^i} T^n_*\nu_{\alpha}(T^n \mathcal{D}_{n,k})\,\lambda^i(d\alpha).\end{align} It follows that the standard family $\mathcal{G}^i_{n,k}$ and its image $T^n \mathcal{G}^i_{n,k}$ have the same factor measures -- mainly because unstable manifolds in $\mathcal{W}^i_{n,k}$ have not been cut by any singular curves under $T^n$.
(\ref{Gammanmksame}) implies that it is enough to compare $ T^n_*(\nu^1|_{_{U_{n,k}}})$ and $ T^n_*(\nu^2|_{_{U_{n,k}}})$, which are associated with two standard families $T^n\mathcal{G}^i_{n,k}$ that fully returned to $\mathcal{U}^*$. For any $n\geq 1$, and $k\geq n$, we also denote $$T^n\mathcal{G}^i_{n,k}=(\mathcal{V}_{n,k}^i, T^n_*(\nu^i|_{_{U_{n,k}}}))=\{(T^n W_{\alpha,k}, T^n_*\nu_{\alpha}),\alpha\in \mathcal{A}^i, k\geq 1,\lambda^i\}.$$
Moreover, we denote $$\mathcal{E}^i_{n,k}=\{(T^n W_{\alpha,k},\mu_{\alpha,k}),\alpha\in \mathcal{A}^i, \lambda^i\},$$
as the associated $\mu$- standard family of $T^n_*\nu_{\alpha}$ replaced by the $u$-SRB measure $$\mu_{\alpha,k}:=\mu_{T^n W_{\alpha,k}},$$ yet having the same index set as well as the same factor measure. We also denote
$$\mathcal{E}^i_{n,k}=(\mathcal{V}^i_{n,k}, \mu^i_{n,k})$$ for simplicity, such that for any measurable set $A\subset \mathcal{U}^*$,
\begin{equation}\label{newmui}\mu^i_{n,k}(A)=\int_{\alpha\in \mathcal{A}^i}\int_{x\in T^nW_{\alpha,k}} \chi_A(x)\, d\mu_{\alpha,k} \, \lambda^i(d\alpha).\end{equation}
Now Lemma \ref{holdernorm} implies that $T^n\mathcal{G}^i_{n,k}$ can be approximated by its associated $u$-standard family $\mathcal{E}^i_{n,k}$ in the following sense:
\begin{equation}\label{nualphamu11}|T^n_*(\nu^i|_{U_{n,k}})(\mathcal{R}^*)-\mu^i_{n,k}(\mathcal{R}^*)|\leq C_F\mu^i_{n,k}(\mathcal{R}^*) \Lambda^{-n\gamma_0}\leq 0.1(1-\vartheta_1)\vartheta_1^{n+1}\mu^i_{n,k}(\mathcal{R}^*) .\end{equation}
Or we can write
$$T^n_*(\nu^i|_{_{U_{n,k}}})(\mathcal{R}^*)=\mu^i_{n,k}(\mathcal{R}^*) (1+\mathcal{O}(\Lambda^{-n\gamma_0})),$$
where $\mu^i_{n,k}$ was defined as in (\ref{newmui}), and
$$ |\mathcal{O}(\Lambda^{-n\gamma_0})| \leq \varepsilon_{\mathbf{d}}\Lambda^{-n\gamma_0}.$$
For any $n\geq 1$, we denote $\mathcal{B}_n=\mathcal{B}_n(W)$ as the index set of $T^n W$. According to (\ref{cFnnuAnug}), we know that the pushforward image of a standard pair $(W,\mu_W)$ must be of the form $(T^n W, T^n_* \mu_W)$, where $$T^n W=\{V_{\beta}, \beta\in \mathcal{B}_n\}$$ and for any measurable set $A\subset \mathcal{M}$,
$$T^n_* \mu_W(A)=\int_{\beta\in \mathcal{B}_n}\mu_{\beta}(A\cap V_{\beta})\, \lambda^u_{\mathcal{M}}(d\alpha).$$
Thus the standard family $\mathcal{E}^1_{n,k}$ must come from the $n$-th iterations of $((W_{\alpha,k}, \mu_{W_{\alpha,k}}), \mathcal{A}^i, \lambda^i)$. This implies that
\begin{equation}\label{munmki}\mu^i_{n,k}(\mathcal{R}^*)=\int_{\alpha\in \mathcal{A}^i}\mu_{{\alpha,k}}( \mathcal{R}^*)\, \lambda^i(d\alpha).\end{equation}
Moreover, note that $$\lambda^i(\mathcal{A}^i)=1$$ as $\nu^i$ is a probability measure. So in order to compare $\mu^1_{n,k}(\mathcal{R}^*)$ and $ \mu^2_{n,k}(\mathcal{R}^*)$, it is enough to compare $\mu_{{\alpha,k}}( \mathcal{R}^*)$ for $\alpha\in \mathcal{A}^1$ and $\alpha\in \mathcal{A}^2$.
Since for each $k\geq 1$, the stable boundaries of $U_{n,k}$ shrink exponentially fast under the iteration of $T^n$, thus the standard pairs in the family $\mathcal{E}^1_{n,k}=(\mathcal{V}^1_{n,k}, \mu^1_{n,k}) $ should be very ``close" to those in $\mathcal{E}^2_{n,k}=(\mathcal{V}^2_{n,k}, \mu^2_{n,k})$. Indeed note that unstable manifolds in $\mathcal{V}^1_{n,k}$ and those in $\mathcal{V}^2_{n,k}$ are at most $0.1\Lambda^{-n}$ far apart . Now we can use the assumption (\ref{Jh}) on distortion bounds for the Jacobian of the stable holonomy map defined by $\Gamma^s$. More precisely, for any $\alpha\in \mathcal{A}^1$, $\beta\in \mathcal{A}^2$, we define $$\mathbf{h}: W_{\alpha}\to W_{\beta}$$ as the stable holonomy map, with $$\mathbf{h}(x)=d\mu_{\alpha}/d\mu_{\beta},$$ for any $x\in W_{\alpha}\cap \Gamma^s$ and $$\mathbf{h}_n: T^n (W_{\alpha}\cap \Gamma_{n}) \to T^n (W_{\beta}\cap \Gamma_{n}).$$ Then by the absolute continuity property of the holonomy map, especially (\ref{Jh}) and (\ref{cJhn}), as well as the fact that $\text{\rm dist}(W_{\alpha}, W_{\beta})\leq 20\delta_0$, for any $W_{\alpha},W_{\beta}\in \Gamma^u$; thus we have $$|\ln \mathbf{h}_n|\leq C_{\mathbf{r}} \text{\rm dist}(W_{\alpha}, W_{\beta})^{\gamma_0}\vartheta^n_0\leq C_{\mathbf{r}} (20\delta_0)^{\gamma_0}\vartheta^n_0<\varepsilon_{\mathbf{d}}\vartheta^n_0<\varepsilon_{\mathbf{d}}^{n+1}$$ as points in $\Gamma_{n}$ have returned to $M$ at least $n$ times under the induced map $T$, also by the choice of $\delta_0$ as in (\ref{holdernorm}), as well as the definition of $\theta_1$ in (\ref{vartheta1}). Thus for any measurable collection of stable manifolds $A\subset \Gamma^s$, any standard pair $(T^n W_{\alpha,k}, \mu_{\alpha,k})\in \mathcal{E}^1_{n,k}$ and $(T^n W_{\beta,k},\mu_{\beta,k})\in \mathcal{E}^2_{n,k}$, we have
\begin{equation}\label{alphbeta}|\mu_{\alpha,k}(A\cap T^n W_{\alpha,k})-\mu_{\beta,k}(A\cap T^n W_{\beta,k})|\leq \varepsilon_{\mathbf{d}}^{n+1}\,\mu_{\beta,k}(A\cap T^n W_{\beta,k}).\end{equation}
Since (\ref{alphbeta}) is true for all $\alpha\in \mathcal{A}^1$ and all $\beta\in \mathcal{A}^2$, using the fact that $\lambda^i(\mathcal{A}^i)=1$, we have by (\ref{munmki}),
$$|\mu^1_{n,k}(\mathcal{R}^*)- \mu^2_{n,k}(\mathcal{R}^*)|\leq \varepsilon_{\mathbf{d}}\mu^1_{n,k}(\mathcal{R}^*)\,\vartheta_1^{n+1}.$$
Now combining with (\ref{nualphamu11}) and (\ref{Gammanmksame}) , we get
$$|\nu^1(\mathcal{D}_{n,k})-\nu^2(\mathcal{D}_{n,k})|\leq \varepsilon_{\mathbf{d}}\nu^2(\mathcal{D}_{n,k}) \vartheta_1^{n+1},$$ where $\vartheta\in (0,1)$ was defined in (\ref{vartheta1}).
\end{proof}
\end{comment}
\begin{comment}\begin{proposition}\label{estAnnn}
Let $$\mathcal{G}^i=(\mathcal{W}^i,\nu^i)=((W_{\alpha},\nu_{\alpha}),\alpha\in \mathcal{A}^i, \lambda^i),$$ $i=1,2$, be two pseudo-generalized standard families of index zero, as defined in Definition \ref{pseodugs}, such that $\mathcal{W}^i$ properly cross $\mathcal{R}^*$. Assume that both $\nu^1$ and $\nu^2$ are probability measures.
For any $ n\geq 1$, we have
\begin{equation}\label{nu1Ank}|\nu^1(A_{n})-\nu^2(A_n)|\leq \varepsilon_{\mathbf{d}}\nu^1(A_{n}).\end{equation}\end{proposition}
\begin{proof}
Using the definition of $A_n$, as well as Lemma \ref{Tmerateo}, we have
\begin{align*}
|\nu^1(A_n)-\nu^2(A_n)|&\leq \sum_{k=1}^{n}|\nu^1(\mathcal{D}_{k,n})-\nu^2(\mathcal{D}_{k,n})|\\
&\leq \sum_{k=1}^{n}\varepsilon_{\mathbf{d}}\nu^1(\mathcal{D}_{k,n})
\end{align*}
\end{proof}
\end{comment}
\begin{comment}
\subsection{Proof of Lemma \ref{renewal1}}\label{subsectionrenewal}
The following facts for renewal sequences were proved in a series of references, starting from Kolmogorov, \cite{Kol, Gel, Sto}, and mainly in \cite{Rog}.
\begin{lemma}\label{renewals} Let $\{p_k,k\geq 1\}$ be a nonnegative sequence such that $$\sum_{k=1}^{\infty}p_k=1.$$ Let $a_0=1$ and for any $n\geq 1$, we set $a_n=\sum_{k=1}^n p_k a_{n-k}$. If $$\lambda:=\sum_{k=1}^{\infty} k p_k<\infty,\,\,\,P_n:=\sum_{k\geq n} p_n=\mathcal{O}(n^{-\alpha})$$ and
$$\xi:=\max\{\eta\geq 1\,:\, \sum_{k=1}^{\infty} k^{\eta} p_k<\infty\}$$ with $\xi\geq 1$, then
$$a_n=\frac{1}{\lambda}+\frac{1}{\lambda^2}\sum_{k\geq n}\sum_{m=k+1}^{\infty} p_m +R_n(\xi),$$
where $R_n(\xi)=\mathcal{O}(n^{-\xi})$ if $\xi\geq 2$; and $R_n(\xi)=\mathcal{O}(n^{-2(\xi-1)})$ if $\xi\in (1,2)$.$$R_n=\left\{
\begin{array}{ll}
\mathcal{O}((\ln n)^2), & \text{ for }\, \alpha=\xi=1;\\
\mathcal{O}(n^{-2(\alpha-1)}), & \text{for } \, 1=\xi<\alpha<2; \\
\mathcal{O}(n^{-2}\ln n), & \text{for } \, \xi=1, \alpha=2;\\
\mathcal{O}(n^{-\alpha}), & \text{for } \,\xi>2, \alpha>2.
\end{array}
\right.$$
\end{lemma}
\vspace{0.3in}
We define $$p_k=\mu(\Gamma^s_k|\mathcal{R}^*)$$ for any $k\geq 1$. By Proposition \ref{YtowercM}, we know that $p_n=\mathcal{O}(n^{-1-\alpha_0})$ and $\sum_{k=n}^{\infty} p_k\leq C n^{-\alpha_0}$, for any $n\geq 1$. We let $\alpha=\alpha_0>1$, then one can check that $\mu(\tau^{\eta})<\infty$, for any $\eta<\alpha_0$. Thus $\xi<\alpha_0$. We first prove a lemma.
\begin{lemma} Let $(\mathcal{W}^i,\nu^i)$ be two proper standard families, and $\nu^i(\mathcal{M})=1$. Then the total amount of remaining measure of $(\mathcal{W}^i,\nu^i)$ that properly returned to $\mathcal{R}^*$ after $n$-th step coupling, which is denoted as $s^i_n$, for any $n\geq 1$, satisfies
$$s^i_n\geq (1-\varepsilon_{\mathbf{d}})\sum_{j\geq n} \sum_{k\geq j+1} p_k$$
\end{lemma}
\begin{proof}
By Lemma \ref{properttau0}, any proper family has a decomposition into generalized standard families:
$$\mathcal{G}^i_0:=(\mathcal{W}^i,\nu^i)=\sum_{m=0}^{\infty} \mathcal{G}^i_{0,m}=\sum_{m=0}^{\infty}(\mathcal{W}^i_{0,m},\nu^i_{0,m}),$$ according to Lemma \ref{properttau0}, where $\mathcal{G}^i_{0,m}$ is a generalized standard family with index $m$, such that $\mathcal{F}^{m}\mathcal{W}^i_{0,m}$ properly returns $\mathcal{R}^*$.
We define $a^i_n=\nu^i_{0,n}(\mathcal{M})$ as the amount of measure that will arrive $\mathcal{R}^*$ (properly) under $\mathcal{F}^n$ for the first time. By Lemma \ref{couplescheme}, we denote $d^i_n$ as the total portion of measure that coupled at the $n$-th step (out of the total returned amount of measure $s^i_n$) .
Let $e_n:=s^i_n d^i_n$ be the amount of measure coupled at $n$-th step.
We denote $\hat s^i_k=s^i_k-e_k$, $\hat a_k^i=a_k^i-e_k$, and
$$p_n=\mu(\Gamma^s_n|\mathcal{R}^*),\,\,\,\,\,\,\, p^i_n=\nu^i(\Gamma^s_n|\mathcal{R}^*),\,\,\,\,\,\,\,p_{j, k-j}^i=\mathcal{F}^{jt_0}_*\nu^i_{j}(\Gamma^s_{k-j}|\mathcal{F}^{t_0 j}A_{j})$$
Let $a_0=1$ and for any $n\geq 1$, we define $$a_n=\sum_{k=1}^n p_k a_{n-k}.$$ Moreover, we denote $$\lambda:=\sum_{k=1}^{\infty} k p_k=1/\nu(\mathcal{R}^*)<\infty,$$ and by the assumption that $\nu(\tau^{\xi})<\infty$, we know that $$\sum_{k=1}^{\infty} k^{\xi} p_k<\nu(\tau^{\xi})<\infty.$$ Then the above renewal theory implies that
\begin{equation}\label{estan}a_n=\frac{1}{\lambda}+\frac{1}{\lambda^2}\sum_{k\geq n}\sum_{m=k+1}^{\infty} p_m +R_n(\xi),\end{equation}
where $R_n(\xi)=\mathcal{O}(n^{-\xi})$ if $\xi\geq 2$; and $R_n(\xi)=\mathcal{O}(n^{-2(\xi-1)})$ if $\xi\in [1,2)$.
We define $\mathcal{D}_{k,n}$ as the set of all points in $\mathcal{R}^*$ that return to $\mathcal{R}^*$ exactly $k$ times under $\mathcal{F}^{t_0n}$, see (\ref{Dnm}). Then one can check that
$$\mathcal{D}_{k,n}=\{x\in \mathcal{R}^*\,:\, \tau_0(x)+\cdots+\tau_0(L^{k-1}x)=n\}.$$
Let $A_n$ be the set of all points that returned to $\mathcal{R}^*$ under $\mathcal{F}^{t_0 n}$. Then
$$A_n=\{x\in \mathcal{R}^*\,:\, \tau_0(x)+\cdots+\tau_0(L^{k-1} x)=n, \text{ for some } k=1,\cdots,n\}=\bigcup_{k=1}^{n}\mathcal{D}_{k,n},$$ as all points in $\mathcal{R}^*$ that will return to $\mathcal{R}^*$ after $n$-steps.
We claim that
\begin{equation}\label{claimAn}A_n=(\Gamma^s_1\cap \mathcal{F}^{-1}A_{n-1})\cup (\Gamma^s_2\cap \mathcal{F}^{-2}A_{n-2})\cup\cdots\cup (\Gamma^s_{n-1}\cap \mathcal{F}^{-(n-1)}A_1)\cup \Gamma^s_n \end{equation}
For any $x\in A_n$, since $A_n\subset \mathcal{R}^*\subset \cup_{m\geq 1} \Gamma^s_m$, then there exists $m\geq 1$, such that $x\in A_n\cap \Gamma^s_m$. Note that $\mathcal{F}^m \Gamma^s_m$ is a $u$-subset of $\mathcal{R}^*$, so $\mathcal{F}^m x\in \mathcal{R}^*$. Using the definition of $A_n$, we know that $\mathcal{F}^n x\in \mathcal{R}^*$, which implies that $\mathcal{F}^{n-m}\mathcal{F}^m x\in \mathcal{R}^*$. Thus $\mathcal{F}^m x\in A_{n-m}$. This implies that
$$A_n\subset \cup_{m=1}^n(\Gamma^s_m\cap \mathcal{F}^{-m}A_{n-m}) $$
where we denote $A_0=\mathcal{R}^*$.
On the other hand, for any $m=1,\cdots, n$, any $x\in \Gamma^s_m$, if $\mathcal{F}^m x\in A_{n-m}$, then clearly $x\in A_n$.
Now it follows from (\ref{claimAn}) that
$$\nu^i(A_n)=\sum_{m=1}^n \nu^i(\Gamma^s_m\cap \mathcal{F}^{-m}A_{n-m})=\sum_{m=1}^n \nu^i(\Gamma^s_m) \mathcal{F}^m_*\eta_m(A_{n-m})=\sum_{m=1}^n p_m\, \mathcal{F}^m_*\eta_m(A_{n-m})$$
Using (\ref{cFkAn-k}), we get
\begin{equation}\label{Anpnepsn}\hat{\mu}(A_n)=\sum_{m=1}^n p_m\, \hat{\mu}(A_{n-m})(1-\varepsilon_{n-m})\end{equation}
which implies that
$$\hat{\mu}(A_n)\geq \sum_{m=1}^n p_m\, \hat{\mu}(A_{n-m})(1-\varepsilon_{\mathbf{d}})$$
We define $a_n'=\hat{\mu}(A_n)$, then (\ref{Anpnepsn}) implies that
$$a_n'=\sum_{m=1}^n p_m\, a_{n-m}'+\sum_{m=1}^n p_m\, a_{n-m}'\,\varepsilon_{n-m}$$
\begin{equation}\label{hatsik}
\delta^i_k=a_k^i+\alpha\sum_{l=1}^{k}p_{l}^i \delta^i_{k-l}\end{equation}
where $\alpha=(1-\varepsilon_{\mathbf{d}})/2\in (0,1/2)$
Similar to \cite{Rog}, we denote
$\mathbf{R}_+(\mathcal{O}(n^{-\alpha}L(n)))$ (resp. $\mathbf{R}_+(o(n^{-\alpha}L(n)))$) as the set of all absolutely convergence series $x(z):=\sum_{n=0}^{\infty} x_n z^n$ for $|z|\leq 1$, such that $x_n=\mathcal{O}(n^{-\alpha}L(n))$ (resp. $x_n=o(n^{-\alpha}L(n))$); or in other words, $$\lim_{n\to\infty} |x_n| /(n^{-\alpha}L(n))<\infty,\, \text{(resp.} \lim_{n\to\infty} |x_n| /(n^{-\alpha}L(n))=0)$$ It was proved in \cite{Rog}~Lemma 2 that both sets are closed under addition, multiplication, and multiplication by a constant of
their generating functions. Moreover, if $x(z)\neq 0$ for $|z|\leq 1$, then we define $x(z)^{-1}=\lambda(z)=\sum_{n=0}^{\infty}\lambda(n)z^n$. It was shown in \cite{Rog}~Lemma 3 that
$$x(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha}L(n)))\Rightarrow \lambda(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha}L(n)));$$$$ x(z)\in \mathbf{R}_+(o(n^{-\alpha}L(n))) \Rightarrow \lambda(z)\in \mathbf{R}_+(o(n^{-\alpha}L(n)).$$
Here $\alpha\geq 1$, and for $\alpha=1$, $L(n)=\mathcal{O}(1)$.
For any complex number $z\in \mathbb{C}$, with $|z|\leq 1$, we define
$$\delta^i(z):=\delta^i_0+\sum_{n=1}^{\infty} \delta^i_n z^n,\,\,\,\,\,\,p(z)=\sum_{n=1}^{\infty}p_n z^n,\,\,\,\,A(z)=a^i_0+\sum_{n=1}^{\infty} a^i_n z^n$$
Note that $ a^i_0=\delta^i_0$.
Clearly, Proposition \ref{YtowercM} implies that $p(z)$ belongs to $\mathbf{R}_+(\mathcal{O}(n^{-1-\alpha_0}))$. Moreover, $A^i(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha_0}))$.??
Moreover, we claim that $p(z)=1$ for $|z|\leq 1$ if and only if $z=1$.
Using (\ref{hatsik}), we get
$$\delta^i(z)(1-\alpha_0 p(z))=A^i(z)$$
Indeed if $|z|<1$, then
$$|p(z)|=|\sum_{n=1}^{\infty} p_n z^n|< \sum_{n=1}^{\infty} p_n=1$$
We assume $z=e^{i\theta}$, for $\theta\in [0,2\pi]$. Then
$$p(z)=\sum_{n=1}^{\infty} p_n \cos n\theta+i\sum_{n=1}^{\infty} p_n \sin n\theta $$
Note that $p(z)=0$ if and only if $\sum_{n=1}^{\infty} p_n \sin n\theta=0$ and $\sum_{n=1}^{\infty} p_n \cos n\theta=1$.
However, the second condition can be written as
\begin{equation}\label{secondII}\sum_{n=1}^{\infty} p_n \cos n\theta=\sum_{n=1}^{\infty} p_n \end{equation}
which implies that
$$\sum_{n=1}^{\infty} p_n (1-\cos n\theta)=0$$
Note that if $z=e^{i\theta} \neq 1$, then there exists $k\geq 1$, such that $|\cos k\theta|<1$, this implies that
$$\sum_{n=1}^{\infty} p_n (1-\cos n\theta)>p_k (1-\cos k\theta)$$
This contradicts the second condition (\ref{secondII}). Thus $p(z)=1$ for $|z|\leq 1$ if and only if $z=1$, as we have claimed.
By assumption, we have for $m=1,\cdots, n$, $\hat{\mu}(A_{n-m})\geq d(1+\vartheta_1^{n-m})$. Thus we have
\begin{align*}
\mu(A_n)&\geq \sum_{m=1}^n p_m\, \hat{\mu}(A_{n-m})(1- \varepsilon_{\mathbf{d}}^2)\\
&\geq d \sum_{m=1}^n p_m\, (1+\vartheta_1^{n-m})(1- \varepsilon_{\mathbf{d}}^2)\\
&\geq d (1+cp_n)(1- \varepsilon_{\mathbf{d}}^2)\\
&\geq d(1- \varepsilon_{\mathbf{d}}^2)+d(1- \varepsilon_{\mathbf{d}}^2) \sum_{m=1}^n p_m\, \vartheta_1^{n-m}\\
&\geq d(1- \varepsilon_{\mathbf{d}}^2)+cd(1- \varepsilon_{\mathbf{d}}^2) p_n\\
&=d+d(cp_n- \varepsilon_{\mathbf{d}}^2- \varepsilon_{\mathbf{d}}^2 p_n)\end{align*}
Let $a'_0=1$, and $a'_n=\mu(A_n|\mathcal{R}^*)$ as the total measure that will return to $\mathcal{R}^*$ properly. More precisely, we have
$$
a'_n=\frac{\mu(A_n\cap \mathcal{R}^*)}{\mu(\mathcal{R}^*)} =\frac{\mu(A_n)}{\mu(\mathcal{R}^*)}
$$
We define the measure $\hat{\mu}:=\mu|_{\mathcal{R}^*}/\mu(\mathcal{R}^*)$; clearly it is invariant under the return map to $\mathcal{R}^*$. For any $k\geq 1$, we define the conditional measure on $\Gamma^s_k$ obtained by $\hat{\mu}$ as
$$\eta_k:=\frac{\hat\mu|_{\Gamma^s_k}}{\hat{\mu}(\Gamma^s_k)}.$$
Then
$$\mathcal{F}^k_*\eta_k(B)=\eta_k(\mathcal{F}^{-k}B)=\frac{\mu(\mathcal{F}^{-k}B\cap\Gamma^s_k)}{\mu(\Gamma^s_k)}$$
Using the fact that the support of $\mathcal{F}^k_*\eta_k$ is properly return to $\mathcal{R}^*$, and Proposition \ref{estAnnn}, we get for any $k\leq n$,
\begin{equation}\label{cFkAn-k}|\mathcal{F}^k_*\eta_k(A_{n-k})-\hat{\mu}(A_{n-k})|\leq \varepsilon_{\mathbf{d}}\sum_{i=1}^{n-k} \hat{\mu}(\mathcal{D}_{i,n-k}) \vartheta_1^{i+1}\end{equation}
Then we can check that
$$a'_1=\hat{\mu}(A_1)=\hat{\mu}(\Gamma^s_1)=a_1=p_1 a_0.$$
Furthermore,
\begin{align*}
a'_2=\hat{\mu}(A_2)=\hat{\mu}(\mathcal{D}_{2,2})+ \hat{\mu}(\mathcal{D}_{1,2}) =\frac{\hat{\mu}(\Gamma_1^s\cap\mathcal{F}^{-1}\Gamma^s_1)}{\hat{\mu}(\Gamma_1^s)}p_1+p_2=\mathcal{F}_*\eta_1(A_1)
p_1+p_2 a_0. \end{align*}
Using (\ref{cFkAn-k}), we get
We can check that
$$\hat{\mu}(\mathcal{F}^{-1}\mathcal{D}_{1,1}|\Gamma^s_1)=F_*\eta_1(\mathcal{D}_{1,1}).
$$ Since both $\mathcal{F}_*\eta_1$ and $\hat{\mu}$ are probability measures with support properly cross $\mathcal{R}^*$, we can apply Lemma \ref{Tmerateo1}, to get
$$| \mathcal{F}_*\eta_1(\mathcal{D}_{1,1})-\hat{\mu}(\mathcal{D}_{1,1})|\leq \varepsilon_{\mathbf{d}}\hat{\mu}(\mathcal{D}_{1,1})\vartheta_1^2,$$
which implies that
\begin{equation}\label{eta1}
|\mathcal{F}_*\eta_1(A_1)-\hat{\mu}(A_1)|\leq \varepsilon_{\mathbf{d}}\hat{\mu}(A_1)\vartheta_1^2
\end{equation}
This is equivalent as
$$|\hat{\mu}(\mathcal{D}_{2,2})-a_1p_1|\leq \varepsilon_{\mathbf{d}} a_1p_1\vartheta_1^2.$$
Here $\vartheta_1\in (0,1)$ were given in (\ref{vartheta1}), and $\varepsilon_{\mathbf{d}}=0.1(1-\vartheta_1)$ was defined in Lemma \ref{densitybd}.
\begin{equation}\label{a1'}|a'_2-a_2|\leq \varepsilon_{\mathbf{d}} a_1p_1\vartheta_1^2.\end{equation}
$$a_2=a_0p_2+a_1p_1=p_2+p_1^2$$
$$a_3=a_0p_3+a_1p_2+a_2p_1=p_3+2p_1p_2+p_1^3$$
$$a_4=a_0p_4+a_1p_3+a_2p_2+a_3p_1=p_4+p_1p_3+p_2^2+p_2p_1^2+p_3p_1+2p_1^2p_2=p_4+2p_1p_3+p_2^2+$$
This also implies that
\begin{equation}\label{a3a2p}|a_3-(p_1a_2'+a_1'p_2+p_3)|=|a_2'-a_2|p_1\leq \varepsilon_{\mathbf{d}} a_1p_1^2\vartheta_1^2.\end{equation}
Note that \begin{align*}
a'_3&=\hat{\mu}(A_3)
=\mathcal{F}_*\eta_1(A_{2}) p_1+\mathcal{F}^2_*\eta_2(A_{1}) p_2+p_3.\end{align*}
Thus we have
\begin{align*}
|a_3'&-a_2'p_1-a_1'p_2-p_3|\leq |\mathcal{F}_*\eta_1(A_2)-a_2'|p_1+|\mathcal{F}_*^2\eta_2(A_1)-a_1'|p_2\\
&\leq \varepsilon_{\mathbf{d}}p_1\sum_{i=1}^{2} \hat{\mu}(\mathcal{D}_{i,2}) \vartheta_1^{i+1}+\varepsilon_{\mathbf{d}}p_2 \hat{\mu}(\mathcal{D}_{1,1}) \vartheta_1^{2}\\
&\leq \varepsilon_{\mathbf{d}}^2(\hat{\mu}(\mathcal{D}_{1,2})p_1+\hat{\mu}(\mathcal{D}_{1,1})p_2)+p_1\hat{\mu}(\mathcal{D}_{2,2})\varepsilon_{\mathbf{d}}^2\vartheta_2^4\\
&\leq \varepsilon_{\mathbf{d}}^2(\hat{\mu}(\mathcal{D}_{1,2})p_1+\hat{\mu}(\mathcal{D}_{1,1})p_2)+p_1^2a_1\varepsilon_{\mathbf{d}}^2\vartheta_2^4+p_1^2a_1\varepsilon_{\mathbf{d}}^3\vartheta_2^6\\
\end{align*}
Combining with (\ref{a3a2p}), we get
$$|a_3'-a_3|\leq \varepsilon_{\mathbf{d}}(p_2p_1+p_1p_2+a_1p_1^2)\vartheta_1^2+p_1^3\varepsilon_{\mathbf{d}}^2\vartheta_2^4+p_1^3\varepsilon_{\mathbf{d}}^3\vartheta_2^6$$
Note that
\begin{align*}
\hat{\mu}(\mathcal{D}_{3,3})& =\frac{\hat{\mu}(\Gamma^s_1\cap \mathcal{F}^{-1}\mathcal{D}_{2,2})}{\hat{\mu}(\Gamma^s_1)}\cdot \hat{\mu}(\Gamma^s_1)=\frac{\hat{\mu}|_{_{\Gamma^s_1}}(\mathcal{F}^{-1}\mathcal{D}_{2,2})}{\hat{\mu}(\Gamma^s_1)} \cdot \hat{\mu}(\Gamma^s_1)=F_*\eta_1(\mathcal{D}_{2,2})p_1
\end{align*}
Note that $F_*\eta_1$ is a probability measure with support properly return to $\mathcal{R}^*$.
Using Lemma \ref{Tmerateo1}, we get
$$|F_*\eta_1(\mathcal{D}_{2,2}) -\hat{\mu}(\mathcal{D}_{2,2})|\leq \varepsilon_{\mathbf{d}}\,\hat{\mu}(\mathcal{D}_{2,2})\vartheta_1^3= \varepsilon_{\mathbf{d}} \vartheta_1^3(\,a_1 p_1+ \varepsilon_{\mathbf{d}}a_1p_1\vartheta_1^2 ).$$
Thus
\begin{align*}
|\hat{\mu}(\mathcal{D}_{3,3})-p_1\hat{\mu}(\mathcal{D}_{2,2})|&\leq \varepsilon_{\mathbf{d}} \,a_1 p_1^2\vartheta_1^3(1+ \varepsilon_{\mathbf{d}} \vartheta_1^2) \end{align*}
This implies that
\begin{equation}\label{D33}
|\hat{\mu}(\mathcal{D}_{3,3})-a_1p^2|\leq \varepsilon_{\mathbf{d}} p_1^3\vartheta_1^2(1+\vartheta_1+\varepsilon_{\mathbf{d}}^2)
\end{equation}
In addition,
\begin{align*}
\hat{\mu}(\mathcal{D}_{2,3})& =\hat{\mu}(\Gamma^s_1\cap \mathcal{F}^{-1}\Gamma_2^s)+\hat{\mu}(\mathcal{F}^{-2}\Gamma^s_1\cap \Gamma_2^s)\\
&=\hat{\mu}(\Gamma^s_1\cap F^{-1}\Gamma_2^s)+\hat{\mu}(F^{-1}\Gamma^s_1\cap \Gamma_2^s)\\
&=\frac{\hat{\mu}(F^{-1}\Gamma_2^s| \Gamma_1^s)}{\hat{\mu}(\Gamma_1^s)}\cdot \hat{\mu}(\Gamma_1^s) +\frac{\hat{\mu}(F^{-1}\Gamma^s_1|
\Gamma_2^s)}{\hat{\mu}(\Gamma^s_2)}\cdot \hat{\mu}(\Gamma^s_2)\\
&=\frac{\hat{\mu}|_{_{\Gamma^s_1}}(\mathcal{F}^{-1}\Gamma^s_2)}{\hat{\mu}(\Gamma^s_1)} \cdot \hat{\mu}(\Gamma^s_1)+\frac{\hat{\mu}|_{_{\Gamma^s_2}}(\mathcal{F}^{-1}\Gamma^s_1)}{\hat{\mu}(\Gamma^s_2)} \cdot \hat{\mu}(\Gamma^s_2)=F_*\eta_1(\Gamma^s_2) p_1+F_*\eta_2(\Gamma^s_1) p_2
\end{align*}
Using Lemma \ref{Tmerateo1}, we get
$$|F_*\eta_1(\Gamma^s_2) -\hat{\mu}(\Gamma^s_2)|\leq \varepsilon_{\mathbf{d}}\,\hat{\mu}(\Gamma^s_2)\vartheta_1^2= \varepsilon_{\mathbf{d}}\,p_2\vartheta_1^2,$$
$$|F_*\eta_2(\Gamma^s_1) -\hat{\mu}(\Gamma^s_1)|\leq \varepsilon_{\mathbf{d}}\,\hat{\mu}(\Gamma^s_1)\vartheta_1^2= \varepsilon_{\mathbf{d}}\,p_1\vartheta_1^2$$
Thus
\begin{align*}
|\hat{\mu}(\mathcal{D}_{2,3})-\hat{\mu}(\mathcal{D}_{1,2})p_1-\hat{\mu}(\mathcal{D}_{1,1})p_2|&\leq \,2p_1p_2\varepsilon_{\mathbf{d}}^2\\
\end{align*}
Using the fact that $a_3'=\hat{\mu}(\mathcal{D}_{3,3}+\mathcal{D}_{2,3}+\mathcal{D}_{1,3})$, and combining above estimations, we get
\begin{align*}
|a_3'-(p_3 a_0+a_1p_2+p_1 a_2)|&\leq 2\varepsilon_{\mathbf{d}} \vartheta_1^2 a_1p_2+ \varepsilon_{\mathbf{d}} \,a_1 p_1^2\vartheta_1^3(1+ \varepsilon_{\mathbf{d}} \vartheta_1^2)
\end{align*}
Using the estimation for $a_1'$ and $a_2'$, see (\ref{a1'}). We get
\begin{align*}\label{a3'}
|a_3-a_3'|&\leq p_1 \varepsilon_{\mathbf{d}}\,a_1 p_1\vartheta_1^2+ \varepsilon_{\mathbf{d}} \,a_1 p_1^2\vartheta_1^3(1+ \varepsilon_{\mathbf{d}} \vartheta_1^2) +2a_1p_2\varepsilon_{\mathbf{d}}^2\\
&\leq \varepsilon_{\mathbf{d}} \vartheta_1^2(p_1^3(1+\vartheta_1+\varepsilon_{\mathbf{d}}^3)+2p_2p_1)\\
&\leq \varepsilon_{\mathbf{d}} \frac{a_3\vartheta_1^2}{1-\vartheta_1}=\frac{0.1(1-\vartheta_1)\vartheta_1^2 a_3}{1-\vartheta_1}<\varepsilon_3:=0.1\, a_3\vartheta_1^2
\end{align*}
\begin{align*}
|a_3'-a_3|&\leq \varepsilon_{\mathbf{d}}^2(a_2p_1+a_1p_2+a_1p_1^2+a_1p_1^2\varepsilon_{\mathbf{d}}\vartheta_2^2)\\
&\leq \varepsilon_{\mathbf{d}}^2(a_3+p_1^3\varepsilon_{\mathbf{d}}\vartheta_2^2)\leq a_3 \varepsilon_{\mathbf{d}}^2(1+\varepsilon_{\mathbf{d}}\vartheta_2^2) \\
\end{align*}
Note that \begin{align*}
a'_4&=\hat{\mu}(A_4)\\
&=\mathcal{F}_*\eta_1(A_{3}) p_1+\mathcal{F}^2_*\eta_2(A_{2}) p_2+\mathcal{F}_*^3\eta_{3}(A_1)p_3+p_4 a_0.\end{align*}
Using (\ref{cFkAn-k}) we get
\begin{align*}
|a_4'&-a_3'p_1-a_2'p_2-a_1'p_3-p_4|
\leq \varepsilon_{\mathbf{d}}\sum_{k=1}^3\sum_{i=1}^{3} p_{4-k}\hat{\mu}(\mathcal{D}_{i,k}) \vartheta_1^{i+1}\\
&=\varepsilon_{\mathbf{d}}\sum_{i=1}^{3}(\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{i,k})) \vartheta_1^{i+1}\\
&= \sum_{k=2}^3 \varepsilon_{\mathbf{d}}|a_k'-a_k|p_{n-k}\vartheta_1^2+\sum_{k=2}^3 \varepsilon_{\mathbf{d}}a_kp_{n-k}\vartheta_1^2
\end{align*}
Moreover,
\begin{equation}\label{a4a3p}|a_4-(a_3'p_1+a_2'p_2+a_1'p_3+p_4|=|(a_3-a_3')p_1+(a_2-a_2')p_2|\leq \varepsilon_{\mathbf{d}}^2(p_2p_1^2+p_1p_2p_1+p_1^4+ a_1p_1p_2+p_1^2\hat{\mu}(\mathcal{D}_{2,2})\varepsilon_{\mathbf{d}}\vartheta_2^2) \end{equation}
Thus combining with (\ref{a4a3p}), we have
\begin{align*}
&|a_4'-a_4|\leq |\mathcal{F}^1_*\eta_1(A_3)-a'_3|p_1+|\mathcal{F}^2_*\eta_2(A_2)-a'_2|p_2+|\mathcal{F}^3_*\eta_3(A_1)-a'_1|p_3+\sum_{j=2}^3 |a_j'-a_j|p_{4-j}\\
&\leq \varepsilon_{\mathbf{d}}\sum_{i=1}^{3}(\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{i,k})) \vartheta_1^{i+1}+\varepsilon_{\mathbf{d}}^2(p_2p_1^2+p_1p_2p_1+p_1^4+ a_1p_1p_2+p_1^2\hat{\mu}(\mathcal{D}_{2,2})\varepsilon_{\mathbf{d}}\vartheta_2^2) \\
&= \varepsilon_{\mathbf{d}}\left(\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{1,k})\vartheta_1^2+\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{2,k}) \vartheta_1^3+\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{3,k}) \vartheta_1^4\right)\\
&+\varepsilon_{\mathbf{d}}^2(p_2p_1^2+p_1p_2p_1+p_1^4+ a_1p_1p_2+p_1^2\hat{\mu}(\mathcal{D}_{2,2})\varepsilon_{\mathbf{d}}\vartheta_2^2) \\
&= \varepsilon_{\mathbf{d}}(\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{1,k})+p_2p_1^2+p_1p_2p_1+p_1^4+ a_1p_1p_2)\vartheta_1^2\\
&+\varepsilon_{\mathbf{d}}(\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{2,k})\vartheta_1^3+\sum_{k=1}^3p_{4-k}\hat{\mu}(\mathcal{D}_{3,k}) ) \vartheta_1^4+\varepsilon_{\mathbf{d}}^2p_1^2\hat{\mu}(\mathcal{D}_{2,2})\vartheta_1^4\\
&\leq
a_4\varepsilon_{\mathbf{d}}^2+
\end{align*}
$$|a_3'-a_3|\leq \varepsilon_{\mathbf{d}}(a_2p_1+a_1p_2+a_1p_1^2)\vartheta_1^2+a_1p_1^2\varepsilon_{\mathbf{d}}^2\vartheta_2^4\leq a_3 \varepsilon_{\mathbf{d}}^2 +p_1^3\varepsilon_{\mathbf{d}}\vartheta_2^2(1+\varepsilon_{\mathbf{d}}^2)$$
\begin{equation}\label{a1'}|a'_2-a_2|\leq \varepsilon_{\mathbf{d}} p_1^2\vartheta_1^2\leq \varepsilon_2:= 0.1\,a_2 \vartheta_1^2.\end{equation}
Note that $a_4'=\hat{\mu}(\mathcal{D}_{4,4}+\mathcal{D}_{3,4}+\mathcal{D}_{2,4}+\mathcal{D}_{1,4})$.
Moreover,
\begin{align*}
\hat{\mu}(\mathcal{D}_{4,4})&=\hat{\mu}(\mathcal{F}^{-1}\Gamma^s_1\cap\mathcal{F}^{-2}\Gamma^s_1\cap \cap\mathcal{F}^{-3}\Gamma^s_1|\Gamma^s_1) p_1\\
&= F_*\eta_1(\mathcal{D}_{3,3}) p_1 \end{align*}
Using Lemma \ref{Tmerateo1} to get
$$|F_*\eta_1(\mathcal{D}_{3,3}) -\hat{\mu}(\mathcal{D}_{3,3})|\leq \varepsilon_{\mathbf{d}}\hat{\mu}(\mathcal{D}_{3,3})\vartheta_1^{4}.$$
Note that $$a_1=p_1, a_2=p_1a_1+p_2 a_0=p_1^2+p_2, a_3=a_0p_3+a_1p_2+a_2p_1=p_1^3+p_2p_1+p_1p_2+p_3$$
Thus
\begin{align*}
|\hat{\mu}(\mathcal{D}_{4,4})-p_1^4|&\leq \varepsilon_{\mathbf{d}} p_1^4\vartheta_1^2(1+\vartheta_1+\varepsilon_{\mathbf{d}}^2)+\varepsilon_{\mathbf{d}}p_1^2\vartheta_1^4\left(p^2+ \varepsilon_{\mathbf{d}} p_1^2\vartheta_1^2(1+\vartheta_1+\varepsilon_{\mathbf{d}}^2)\right)\\
&= \varepsilon_{\mathbf{d}} \vartheta_1^2 p_1^4(1+\vartheta_1+\vartheta_1^2+\vartheta_1^3+\vartheta_1^4+\vartheta_1^5+\vartheta_1^6)\\
&\leq 0.1\,\vartheta_1^2 p_1^4\end{align*}
In addition,
\begin{align*}
\hat{\mu}(\mathcal{D}_{3,4})&=\hat{\mu}(\mathcal{F}^{-1}\Gamma_1\cap \mathcal{F}^{-2}\Gamma^s_2|\Gamma^s_1)p_1+\hat{\mu}(\mathcal{F}^{-1}\Gamma_2\cap \mathcal{F}^{-3}\Gamma^s_1|\Gamma^s_1)p_1+\hat{\mu}(\mathcal{F}^{-1}\Gamma^s_1\cap\mathcal{F}^{-2}\Gamma_1|\Gamma^s_2)p_2\\
&=F_*\eta_1(\mathcal{D}_{2,3}) p_1+F_*\eta_1(\mathcal{D}_{2,2}) p_1\end{align*}
Thus
$$|\hat{\mu}(\mathcal{D}_{3,4})-3p_1^2p_2|\leq 3p_1^2p_2\varepsilon_{\mathbf{d}}(\vartheta_1+\vartheta_1^2+\vartheta_1^4)< 3p_1^2p_2\varepsilon_{\mathbf{d}}(\vartheta_1+\vartheta_1^2+\vartheta_1^3)\leq 0.3p_1^2p_2\vartheta_1^2
$$
\begin{align*}
\hat{\mu}(\mathcal{D}_{2,4})&=\hat{\mu}(\mathcal{F}^{-1}\Gamma_3|\Gamma^s_1)p_1+\hat{\mu}( \mathcal{F}^{-3}\Gamma^s_1|\Gamma^s_3)p_3+\hat{\mu}( \mathcal{F}^{-2}\Gamma^s_2|\Gamma^s_2)p_2\\
&=F_*\eta_1(\mathcal{D}_{1,3}) p_1+F_*\eta_3(\mathcal{D}_{1,1}) p_3+F_*\eta_2(\mathcal{D}_{1,2})p_2\\
&\leq \varepsilon_{\mathbf{d}} p_1(p_3+p_3 \vartheta_1^2)+p_3(p_1+p_1\vartheta_1^2)+p_2(p_2+p_2\vartheta_1^2)
\end{align*}
Thus
$$|\hat{\mu}(\mathcal{D}_{2,4})-2p_1p_3-p_2^2|\leq \varepsilon_{\mathbf{d}}(2p_1p_3 +p_2^2)\vartheta_1^2$$
Combining the above facts, as well as $a_4=p_4+2p_3p_2+3p_1^2p_2+p_2^2+p_1^4$,
we get an estimation for $a_4'$:
\begin{align*}
|a_4'-a_4|\leq \varepsilon_{\mathbf{d}}\frac{a_4\vartheta^2_1}{1-\vartheta_1}\leq \varepsilon_4:=0.1 \,a_4\vartheta_1^2
\end{align*}
Inductively, assume for any $k\leq n-1$, we have
$$|a'_{k}-a_{k}|\leq \varepsilon_{k}\leq 0.1\,a_k\vartheta_1^2$$
Next we estimate $a'_{n}$. Note that
\begin{align*}
a'_n&=\hat{\mu}(A_n)\\
&=\mathcal{F}_*\eta_1(A_{n-1}) p_1+\mathcal{F}^2_*\eta_2(A_{n-2}) p_2+\cdots+p_{n-1}\mathcal{F}^{n-1}_*\eta_{n-1}(A_1)+p_n a_0.\end{align*}
Using Lemma \ref{Tmerateo1}, we get for any $k=1,\cdots, n-1$,
$$|\mathcal{F}^k_*\eta_k(A_{n-k}) -a_{n-k}|\leq 0.1\,a_{n-k}\vartheta_1^{2}.$$
Combing above estimates as well as our assumptions, we get
\begin{equation}\label{a'n}|a'_n-a_n|\leq \sum_{k=2}^n\varepsilon_{k}p_{n-k}\leq 0.1a_n\vartheta_1^2,\end{equation}
where again, we denote $p_0=1$, and the error term $\varepsilon_n$ satisfies
$$|\varepsilon_{n}|\leq 0.1a_n\vartheta_1^2
$$
\medskip
Next we consider any any pseudo-generalized standard family of index zero, $(\mathcal{W},\nu)$, with probability measure $\nu(\mathcal{R}^*)=1$.
Then by Lemma \ref{Tmerateo1}, we can check that
for any $k\geq 1$,
$$|\nu(\Gamma^s_k)-\hat{\mu}(\Gamma_k^s)|\leq \varepsilon_{\mathbf{d}}\hat{\mu}(\Gamma^k_s),$$
which implies that
\begin{equation}\label{pk}
|\nu(\Gamma^s_k)-p_k|\leq \varepsilon_{\mathbf{d}} p_k\vartheta_1.
\end{equation}
In particular, the total amount of measure in $\mathcal{F} ((\mathcal{W},\nu)$ that properly returned to $\mathcal{R}^*$ satisfies
$$r_1:=\nu(\Gamma^s_1)=\nu(\Gamma^s_1)= p_1+\varepsilon_0 =a_1+\varepsilon_0,$$
where $|\varepsilon_0|<C p_1$.
We also define $$r_0:=1.$$
Let $r_n$ be the total amount of measure in $\mathcal{F}^n (\mathcal{W},\nu)$ that properly returned to $\mathcal{R}^*$. Using Lemma \ref{Tmerateo1}, we first check $r_2$ satisfies
\begin{align*}r_2&= \nu(\cup_{k=1}^2 \mathcal{D}_{k,2})=\sum_{k=1}^2\nu(\mathcal{D}_{k,2})\\
&\leq \sum_{k=1}^2\hat{\mu}(\mathcal{D}_{k,2})(1+0.1 \,\vartheta_1^{2})\\
&=a'_2+0.1\vartheta_1^2 \,\sum_{k=1}^2 p_k a_{2-k} \leq a'_2(1+0.1\vartheta_1^2),
\end{align*}.
Inductively we can check that the total amount of measure in $\mathcal{F}^n (\mathcal{W},\nu)$ that properly returned to $\mathcal{R}^*$ satisfies
\begin{align*}
r_n &=\sum_{k=1}^{n} \nu(\mathcal{D}_{k,n})\leq a'_n(1+0.1\vartheta_1^2)
\end{align*}
Combining with (\ref{a'n}), and using the fact that $a'_n\leq 1$ is bounded, we get
\begin{align*}
|r_n-a_n|&\leq 0.2\vartheta_1^2 \,\sum_{m=2}^n p_{n-m}\sum_{j=1}^{m} a_{m-j} p_{j}\\
&=0.2\vartheta_1^2 \,\sum_{j=1}^m\left(\sum_{m=j}^n p_{n-m} a_{m-j}\right)p_j
\leq 0.2\, a_n\vartheta_1^2.
\end{align*}
\end{proof}
Next we consider the general case.
\begin{lemma}\label{lemma 23} Let $\mathcal{G}_m=(\mathcal{W}_m,\nu_m)$ be a generalized standard family of index $m$, for $m=0,\cdots, t_0$. We denote $$(\mathcal{W},\nu)=\sum_{m=0}^{t_0}\mathcal{G}_m,$$ and assume $\nu$ is a probability measure. Then the total amount of measure of $(\mathcal{W},\nu)$ that properly returned to $\mathcal{R}^*$ after $\mathcal{F}^{t_0}$-iterations, which is denoted as $s_{t_0}$, satisfies
$$s_{t_0}\geq \nu(\mathcal{R}^*)/2.$$
\end{lemma}
\begin{proof}
Let $$(\mathcal{W},\nu)=\sum_{m=0}^{t_0}\mathcal{G}_m$$ be the sum of $t_0$ generalized standard families, with probability measure $\nu$.
We define $$q_m=\mathcal{F}^m_*\nu_m(\mathcal{R}^*).$$ Clearly, $q_m\leq C m^{\alpha_1+1}$. Note that once $\mathcal{F}^m\mathcal{G}_m$ arrives at $\mathcal{R}^*$, its further iterations will return to $\mathcal{R}^*$ according to the Markov decomposition $$\Gamma^s=\cup_{k\geq 1} \Gamma^s_k.$$
Note that $(\mathcal{W}_0,\nu_0)$ has index $0$, we denote $a_0^0=q_0$ and $p_0=0$.
Then we can apply the above lemma to get that the total amount of measure in $\mathcal{F}^n (\mathcal{W}_0,\nu_0)$ that properly returned to $\mathcal{R}^*$ satisfies
$a_n^0\geq q_0/(2\lambda)$.
Inductively, one can show that in the generalized family $(\mathcal{W}_k,\nu_k)$, for $k=1,\cdots,n$, the total amount of measure in $\mathcal{F}^n (\mathcal{W}_k,\nu_k)$ that properly returned to $\mathcal{R}^*$ satisfies
$a_n^k\geq q_k/(2\lambda).$
Thus the total amount of measure of $(\mathcal{W},\nu)$ that properly returned to $\mathcal{R}^*$ after $\mathcal{F}^n$-iterations, which is denoted as $s_n$, for any $n\geq 0$, satisfies
\begin{align*}
s_n&=a_n^0+a_n^1+\cdots+a_n^n\\
&\geq \frac{1}{2\lambda}\sum_{k=0}^n q_k.
\end{align*}
Since we can chose $t_0$ large, such that
\begin{align*}
s_{t_0}&\geq \mu(\mathcal{R}^*)/3
\end{align*}We have proved the claimed results.
\end{proof}
\end{comment}
\section{Coupling Lemma for the original system}
In this section, we will prove the Coupling Lemma for the original nonuniformly hyperbolic map, which is new to our knowledge, as the construction is significantly different from that for systems with uniformly hyperbolicity. This will enable us to define the coupling decompositions of probability measures on $\mathcal{M}$, which will be used to investigate the rate of decay of correlations of iterations of those measures.
\subsection{Statement of the Coupling Lemma}
We now state the coupling lemma for the original nonuniformly hyperbolic system $(\mathcal{F},\mu)$.
\begin{lemma}\label{coupling1} Let $\mathcal{G}^i:=(\mathcal{W}^i,\nu^i)$ be two standard families, with $\mathcal{W}^i\subset \mathcal{W}^u$ being a measurable partition of a Borel set $B^i\subset \mathcal{M}$ in to unstable manifolds, and $\mu(B^i)>0$, and $d\nu^i=g^id\mu$, $i=1,2$, where $g^i\in\mathcal{H}^+(\gamma_0)$ are probability density functions. \\
(\textbf{C1}) There exist $N=N(\nu_1,\nu_2)\geq 1$, and some
uniform constant $C_0>0$, such that for any $n\geq 1$, there is a decomposition $$\mathcal{F}^{N}\mathcal{G}^i=\sum_{k=1}^{n}(\hat\mathcal{W}_k^i,\eta_k^i)+(\bar\mathcal{W}^i_n,\bar\nu^i_n),$$ for $i=1,2$, with the following properties for any $k=1,\cdots, n$:
\begin{itemize}\item[(i)] $(\hat\mathcal{W}^i_k, \eta^i_k)$ is a generalized standard family with index $k$;
\item[(ii)] For any measurable function $f$ that is constant on each $W^s\in \Gamma^s$, we have $\mathcal{F}^{k}_*\eta^1_k(f)=\mathcal{F}^{k}_*\eta^2_k(f)$;
\item[(iii)] For any bounded function $f\in L_{\infty}(\mathcal{M})$, the uncoupled measure $\bar\nu^i_k$ satisfies
$|\bar\nu^i_k(f)|\leq C_0\|f\|_{\infty}k^{1-{\alpha_0}}.$
\end{itemize}
\noindent(\textbf{C2}) Moreover, there exists $C_1=C_1(\gamma_0)>0$, such that the portion of measure coupled at $k$-th step satisfies:
$$\mathcal{F}^k_*\eta^i_k(\mathcal{R}^*)\leq C_1\|g^i\|_{\infty} k^{-{\alpha_0}}.$$
\noindent(\textbf{C3}) The uncoupled measure $\bar\nu_n^i(\mathcal{M})$ is dominated by $\bar\nu_n^i(C_{n,b})$ and $\nu^i(R>n)$:
\begin{align*}
\bar\nu^i_n(\mathcal{M})&= \nu^i(R>n)+\bar\nu_n^i((R\leq n)\cap C_{n,b})+ \mathcal{O}(n^{-1-\alpha_0})\\
\end{align*}
\end{lemma}
The proof of this Coupling Lemma can be found in Subsection \ref{proofcoupling}, after we describe in detail the coupling procedure for measures that properly returned to $\mathcal{R}^*$ in next subsection. \\
We begin by considering a special situation, i.e. $\mathcal{G}^i=(\mathcal{W}^i,\nu^i)$, $i=1,2$, are two generalized standard families with index $0$.
We first prove a lemma describing the coupling process which will be used in our proof of Lemma \ref{coupling1}.
\begin{lemma}\label{couplescheme} Assume that for $i=1,2$, $\mathcal{G}^i=(\mathcal{W}^i,\nu^i)$ are generalized standard families with index $0$, and $$\min\{\nu^1(\Gamma^s), \nu^2(\Gamma^s)\}>0.$$ Then there exist a generalized standard family $\mathcal{E}^i=(\mathcal{W}^i,\eta^i)$ with index $0$, and $$\mathcal{K}^i:=\mathcal{G}^i-\mathcal{E}^i=(\mathcal{W}^i,\xi^i),$$ with the following properties:\\
(\textbf{a}) $\mathcal{E}^1$ and $\mathcal{E}^2$ are coupled in the following sense:
(\textbf{a1}) For any bounded function $f$ that is constant on each $W^s\in\Gamma^s$, we have $\eta^1(f)=\eta^2(f)$;
(\textbf{a2}) The total coupled measure satisfies $$\eta^i(\Gamma^s)=c_0\min\{\nu^1(\Gamma^s), \nu^2(\Gamma^s)\},$$ where $c_{0}\in [(1-\mathbf{a})/2,1-\mathbf{a}]$, $\mathbf{a}$ was defined in Lemma \ref{defnN}. And we denote $d^i=\eta^i(\Gamma^s)/\nu^i(\Gamma^s)$.\\
(\textbf{b}) The remaining uncoupled family $\mathcal{K}^i$ is a pseudo-generalized family with index $0$, it has the property that $\mathcal{F}^{n}(\mathcal{K}^i|_{\Gamma^s_n})$ becomes a generalized standard family of index $0$, for any $n\geq 1$.\\
(\textbf{c}) For any measurable collection of unstable manifolds $A\subset \Gamma^s$, the remaining uncoupled measure can be calculated as:
$$|\nu^1(A)-\nu^2(A)|=|\xi^1(A)-\xi^2(A)|\,\,\,\,\,\,\text{ \em{and} }\,\,\,\, \,\,\,
\xi^i(\Gamma^s)=(1-d^i)\nu^i(\Gamma^s).$$
\end{lemma}
\begin{proof}
Since for $i=1,2$, $\mathcal{G}^i$ is a generalized standard family of index $0$, by definition, it has density function $g^i\in \mathcal{H}^+(\gamma_0)$. Then we denote $$\mathcal{G}^i=\{(W_{\alpha},\nu_{\alpha})\,:\,\alpha\in \mathcal{A}^i, \lambda^i\},$$ such that for any measurable set $A\subset \mathcal{M}$,
$$\nu^i(A\cap\Gamma^s)=\int_{\alpha\in \mathcal{A}^i}\int_{W_{\alpha}\cap A} g^i_{\alpha}\,d\mu_{\alpha}\,\lambda^i(d\alpha),$$
where $$g^i_{\alpha}=g^i/\mu_{\alpha}(g^i)$$ with $$d\nu^i_{\alpha}=g^i_{\alpha}\,d\mu_{\alpha},$$ and $$\lambda^i(d\alpha)=\mu_{\alpha}(g^i)\lambda^u(d\alpha).$$ Clearly, $$\lambda^i(\mathcal{A}^i)=\nu^i(\Gamma^s)=\mu(g^i|_{\mathcal{R}^*}).$$
For any $\alpha\in \mathcal{A}^i$, as the standard pair $(W_{\alpha},\nu^i_{\alpha})$ properly crosses $\Gamma^s$, by Lemma \ref{densitybd}, we know that the density function satisfies $$g^i_{\alpha}\geq e^{-\varepsilon_{\mathbf{d}}}.$$ Thus one should be able to match at least a positive portion of measures from both families along stable manifolds in $\Gamma^s$. Now we follow the coupling scheme as described by Chernov and Markarian in the book \cite{CM} -- page 200-202, on choosing a function $\rho^i_{\alpha}\in \mathcal{H}(\gamma_0)$ on $W_{\alpha}$, with the following properties. More precisely, we take $c_0\in [(1-\mathbf{a})/2,(1-\mathbf{a})]$, as described in \cite{CM}, we have the flexibility to choose a function $\rho^i_{\alpha}\in \mathcal{H}(\gamma_0)$, such that its norm is bounded by $1$, with $\|\rho^i_{\alpha}\|_{\gamma_0}\leq 1$, and $$\rho^i_{\alpha}\in ( g^i_{\alpha}(1-\mathbf{a})/2, (1-\mathbf{a})g^i_{\alpha}),$$ for any $\alpha\in \mathcal{A}^i$, with $\mathbf{a}$ defined as in Lemma \ref{defnN}, such that
\begin{equation}\label{rhoialpha}\int_{\alpha\in \mathcal{A}^i}\mu_{\alpha}(\rho^i_{\alpha}\cdot\mathbf{I}_{\mathcal{R}^*\cap W_{\alpha}})\,d\lambda^i(\alpha)=c_0\min\{\nu^1(\Gamma^s), \nu^2(\Gamma^s)\},\quad \forall i=1,2.\end{equation}
Now we are ready to define the coupled families $\mathcal{E}^i$ for $i=1,2$.
(\ref{rhoialpha}) implies that for any $\alpha\in \mathcal{A}^i$, one can define a standard pair corresponding to the measure defined by the density function $\rho^i_{\alpha}$ on $W_{\alpha}$, denoted as $(W_{\alpha},\eta_{\alpha})$. Let $$\mathcal{E}^i:=(\mathcal{W}^i,\eta^i)=((W_{\alpha}, \eta_{\alpha})|_{\Gamma^s}, \lambda^i)$$ be the corresponding generalized standard family with index zero. One could choose $\rho^i_{\alpha}$ and $c_0$ carefully to make sure that for any measurable collection $A$ of stable manifolds in $\Gamma^s$, we have $$\eta^1(A)=\eta^2(A),$$
and
$$\eta^i( \Gamma^s\cap\mathcal{W}^i)=c_0\min\{\nu^1(\Gamma^s), \nu^2(\Gamma^s)\}.$$ Let $d^i=\eta^i(\Gamma^s)/\nu^i(\Gamma^s)$. This verifies items \textbf{(a1)-(a2)}.
Next, we define the remaining uncoupled family $\mathcal{K}^i$ by subtracting the density of $\eta^i$ from $\nu^i$. More precisely, for any $\alpha\in \mathcal{A}^i$, we subtract $\rho^i_{\alpha}$ from the density function $g^i_{\alpha}$. The remaining family in $\mathcal{G}^i$ is denoted as $\mathcal{K}^i$, which may not have the required regularity of being a generalized standard family. We apply Lemma \ref{defnN}, which states that $F\mathcal{K}^i$ is already a standard family. It follows that restricted on $\Gamma^s_n$, the family $\mathcal{F}^{n}(\mathcal{K}^i|_{\Gamma^s_n})$ becomes a generalized standard family with index $0$, for any $n\geq 1$.
Note that (\textbf{a2}) implies that, for any measurable collection of stable manifolds $A\subset \Gamma^s$,
\begin{equation}\label{coupleA}\nu^1(A)-\nu^2(A)=\eta^1(A)-\eta^2(A)+\xi^1(A)-\xi^2(A)=\xi^1(A)-\xi^2(A).\end{equation}
Combining above facts, we get $$\mathcal{G}^i=\mathcal{E}^i+\mathcal{K}^i$$ satisfying \textbf{(a)-(c)}, as claimed.
\end{proof}
Next we prove another lemma that will be used in the estimations of measures that proper return to $\mathcal{R}^*$ at any step $n\geq 1$.
Similar to \cite{Rog}, we denote
$\mathbf{R}_+(\mathcal{O}(n^{-\alpha_0}))$ (resp. $\mathbf{R}_+(o(n^{-\alpha_0}))$) as the set of all absolutely convergence series $x(z):=\sum_{n=0}^{\infty} x_n z^n$ for $|z|\leq 1$, such that $x_n=\mathcal{O}(n^{-\alpha_0})$ (resp. $x_n=o(n^{-\alpha_0})$); or in other words, $$\overline{\lim}_{n\to\infty} |x_n| /(n^{-\alpha_0})<\infty,\, \text{(resp.} \lim_{n\to\infty} |x_n| /(n^{-\alpha_0})=0).$$ It was proved in \cite{Rog}~Lemma 2 that both sets are closed under addition, multiplication, and multiplication by a constant of
their generating functions. Moreover, if $x(z)\neq 0$ for $|z|\leq 1$, then we define $x(z)^{-1}=\lambda(z)=\sum_{n=0}^{\infty}\lambda(n)z^n$. It was shown in \cite{Rog}~Lemma 3 that
$$x(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha_0}))\Rightarrow \lambda(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha_0}));$$$$ x(z)\in \mathbf{R}_+(o(n^{-\alpha_0})) \Rightarrow \lambda(z)\in \mathbf{R}_+(o(n^{-\alpha_0})).$$
\begin{lemma}\label{lemmarenewal}
Assume there exist $\{\delta_n\}$, $\{a_n\}$, $\{p_{n}\}$, such that for $n\geq 1$, \begin{equation}\label{hatsik1}
\delta_n=a_n+\alpha\sum_{l=1}^{n}p_{l} \delta_{n-l},\end{equation}
where $\alpha\in (0,1)$. For any complex number $z\in \mathbb{C}$, with $|z|\leq 1$, we define
$$\delta (z):=\delta _0+\sum_{n=1}^{\infty} \delta _n z^n,\,\,\,\,\,\,p(z)=\sum_{n=1}^{\infty}p_n z^n,\,\,\,\,A(z)=a _0+\sum_{n=1}^{\infty} a _n z^n,$$
where $ a_0=\delta_0$.
We assume that $p(z)\in\mathbf{R}_+(\mathcal{O}(n^{-1-\alpha_0}))$, $A (z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha_0}))$, with $p(1)=1$ and $A(1)=1$.
Then there exist constants $c_2>c_1>0$, such that for any $n\geq 1$,
$$c_1\alpha n^{-\alpha_0}\leq \delta_n\leq c_2 n^{-\alpha_0}.$$
\end{lemma}
\begin{proof} By assumption, for any $n\geq 1$, we have
$\delta _n=a_n +\alpha\sum_{l=1}^{n}p_{l} \delta _{n-l}$,
where $\alpha\in (0,1)$.
Moreover, we claim that $p(z)=1$ for $|z|\leq 1$ if and only if $z=1$.
Using the definition of $\delta(z)$, $p(z)$ and $A(z)$, we can check that (\ref{hatsik1}) implies that $$\delta (z)(1-\alpha p(z))=A (z).$$
Indeed if $|z|<1$, then
$$|p(z)|=|\sum_{n=1}^{\infty} p_n z^n|< \sum_{n=1}^{\infty} p_n=1.$$
We assume $z=e^{i\theta}$, for $\theta\in [0,2\pi]$. Then
$$p(z)=\sum_{n=1}^{\infty} p_n \cos n\theta+i\sum_{n=1}^{\infty} p_n \sin n\theta. $$
Note that $p(z)=1$ if and only if $\sum_{n=1}^{\infty} p_n \sin n\theta=0$ and $\sum_{n=1}^{\infty} p_n \cos n\theta=1$.
However, the second condition can be written as
\begin{equation}\label{secondII}\sum_{n=1}^{\infty} p_n \cos n\theta=1=\sum_{n=1}^{\infty} p_n, \end{equation}
which implies that
$$\sum_{n=1}^{\infty} p_n (1-\cos n\theta)=0.$$
Note that if $z=e^{i\theta} \neq 1$, then there exists $k\geq 1$, such that $|\cos k\theta|<1$, this implies that
$$\sum_{n=1}^{\infty} p_n (1-\cos n\theta)>p_k (1-\cos k\theta).$$
This contradicts the second condition (\ref{secondII}). Thus $p(z)=1$ for $|z|\leq 1$ if and only if $z=1$, as we have claimed.
Thus $0<1-\alpha p(z)\leq 1-\alpha$ for $|z|\leq 1$, since $\alpha\in (0,1)$. Moreover, $p(z)\in \mathbf{R}_+(\mathcal{O}(n^{-1-\alpha_0}))$ implies that $1-\alpha p(z)\in \mathbf{R}_+(\mathcal{O}(n^{-1-\alpha_0}))$. It follows from \cite{Rog}~Lemma 3 that
$$(1-\alpha p(z))^{-1}\in \mathbf{R}_+(\mathcal{O}(n^{-1-\alpha_0})).$$
Note that
\begin{align*}
\delta(z)&=(1-\alpha p(z))^{-1}A(z)\in \mathbf{R}_+(\mathcal{O}(n^{-\alpha_0})).
\end{align*}
We denote
\begin{equation}\label{1-alpha}q(z):=(1-\alpha p(z))^{-1}=\sum_{n=0}^{\infty}q_nz^n,\end{equation}
where $q_0=1$.
We use the fact that $q_0=1$ and
$$ q(z)(1-\alpha p(z))=1,$$
which implies that for $n\geq 1$,
$$q_{n}=\alpha\sum_{k=0}^{n-1} p_{n-k}q_{k}.$$
Using the Taylor expansion
$$q(z)\frac{1}{1-\alpha p(z)}=1+\alpha p(z)+\alpha^2 p(z)^2+ \cdots+\alpha^n(p(z))^n+\cdots=1+\alpha p(z)+\mathcal{O}(\alpha^2 p(z)^2)$$
Thus there exists a constant $C_1$, such that for any $n\geq 1$,
$$\alpha p_{n}\leq q_{n}\leq C_1 n^{-1-\alpha_0}.$$
Note that
\begin{align*}
\delta(z)&=(1-\alpha p(z))^{-1}A(z)\\
&=(\sum_{k=0}^{\infty} q_kz^k)(a_0+\sum_{k=1}^{\infty} a_kz^k)\\
&= a_0\sum_{n=0}^{\infty}q_nz^n+\sum_{n=1}^{\infty}\left(\sum_{k=1}^{n}q_{n-k} a_{k}\right) z^n\\
&=a_0+\sum_{n=1}^{\infty}\left(q_n+\sum_{k=1}^{n}q_{n-k} a_{k}\right) z^n\\
&=a_0+\sum_{n=1}^{\infty}\left(\sum_{k=0}^{n}q_{n-k} a_{k}\right) z^n.
\end{align*}
This implies that for any $n\geq 1$,
$$\alpha p_n+\alpha\sum_{k=1}^{n}p_{n-k} a_{k} \leq \delta_n=q_n+\sum_{k=1}^{n}q_{n-k} a_{k}.$$
Using the fact that $q_n\leq C_1 n^{-1-\alpha_0}$, and $a_n\leq C n^{-\alpha_0}$, and the left hand side is an averaging of $\{a_1,\cdots, a_n\}$, this implies that there exist $c_2>c_1>0$, such that for any $n\geq 1$,
$$c_1\alpha n^{-\alpha_0} \leq \delta_n=\sum_{k=0}^{n}q_{n-k} a_{k} \leq c_2 n^{-\alpha_0}.$$
\end{proof}
We consider another first proper return map $L :\mathcal{R}^*\to \mathcal{R}^*$, and a first proper return time $\tau_0:\mathcal{R}^*\to\mathbb{N}$, such that each level set $(\tau_0=n)=\mathcal{R}_n$ is a $s$-subset of $\mathcal{R}^*$, and $\mathcal{F}^{n } (\tau_0=n)$ properly return to $\mathcal{R}^*$ for the first time. Moreover, for $\mu$-almost every $x\in (\tau_0=n)$, we define $L x=\mathcal{F}^{n }x$.
As defined in (\ref{Antaun}), we denote $A_0=\mathcal{R}^*$, and for any $n\geq 1$, $$A_n=\{x\in \mathcal{R}^*\,:\, \tau_0(x)+\cdots+\tau_0(L^{k-1} x)=n, \text{ for some } k=1,\cdots,n\}.$$
Note that $A_n$ contains points in $\mathcal{R}^*$ that properly return to $\mathcal{R}^*$ under $\mathcal{F}^{n }$. In particular, it can be decomposed as:$$A_{n}=\bigcup_{k=0}^{n-1} \left(\mathcal{F}^{-k}(\tau_0=n-k)\cap A_k\right).$$
This implies that for any standard family $(\mathcal{W},\nu)$,
$$\nu(A_n)=\sum_{k=0}^{n-1}\mathcal{F}^{k}_*\nu(\tau_0=n-k|\mathcal{F}^k A_k) \nu(A_k).$$
Note that $\mathcal{F}^k A_k$ is a $u$ subset of $\mathcal{R}^*$ that properly cross $\Gamma^s$. We denote \begin{equation}\label{defnpk} p_k=\mu(\tau_0=k|\mathcal{R}^*).\end{equation} Then by Lemma \ref{Tmerateo}, we have
\begin{equation}\label{edtau0}
p_{n-k}(1-\varepsilon_{\mathbf{d}})\leq \mathcal{F}^{k}_*\nu(\tau_0=n-k|\mathcal{F}^k A_k) \leq p_{n-k}(1+\varepsilon_{\mathbf{d}}).
\end{equation}
This relation is crucial in the below estimations of the Coupling Lemma.
From now on, we choose $\varepsilon_{\mathbf{d}}<10^{-5}$, such that
\begin{equation}\label{epsbdchoose}
\varepsilon_{\mathbf{d}}< \frac{c_0 c_1(1-\varepsilon_{\mathbf{d}})}{4c_2 }
\end{equation}
\begin{comment}
\subsection{Outline of the proof of the Coupling Lemma}
Let $\mathcal{G}^i:=(\mathcal{W}^i,\nu^i)$ be two standard families for $i=1,2$. We first describe an outline of the coupling scheme using above lemmas. The main difficulty of the coupling process is to guarantee frequently, simultaneous returns of certain portion of the conditional measure from both families. This fact is not straightforward because of the nonuniformly hyperbolicity. We overcome this difficulty by using the Markov structure of the generalized Young Tower based on $\mathcal{R}^*=\cup_{m\geq 1}\Gamma^s_m$ as introduced by Proposition \ref{YtowercM}, as well as the renewal theory proved by \cite{Gel, Kol, Rog}. More precisely, we follow the 3 steps below:
\begin{itemize}
\item[(I)] We first define $N_1\geq 1$, such that both $\mathcal{F}^{N_1}_*\nu^1$ and $\mathcal{F}^{N_1}_*\nu^2$ have at least $\delta$ portion of measure properly returned to $\mathcal{R}^*$. We further decompose $$\mathcal{F}^{N_1}\mathcal{G}^i=\sum_{m=0}^{\infty}\mathcal{G}^i_{0,m}, $$ such that $\mathcal{F}^{m}\mathcal{G}^i_{0,m}$ is a generalized standard family of index $0$.
\item[(II)] Now we have that both $\mathcal{G}^1_{0,0}$ and $\mathcal{G}^2_{0,0}$ are generalized standard family of index $0$. By applying Lemma \ref{couplescheme} we can couple at least $\delta/3$ of measures from both families, to get $$\mathcal{G}^i_{0,0}=\mathcal{E}^i_0+\mathcal{K}^i_0,$$ where $\mathcal{E}^i_0$ is the coupled family at this step. Since $\mathcal{K}^i_0$ is a pseudo-generalized standard family of index $0$, we apply Lemma \ref{renewal1}, which guaranteed that at least $\delta$ portion of measure in this family will return to $\mathcal{R}^*$ after another $t_0$ iterations.
Thus the remaining families are split into two parts, $\hat\mathcal{K}^i_1$ and $$\bar\mathcal{G}^i_1=\sum_{m\geq t_0+1}\mathcal{G}^i_{0,m},$$ where $\hat\mathcal{K}^i_1$ consists of the remaining family $\mathcal{K}^i_1$ as well as $\mathcal{G}^i_{0,m}$ for $m=1,\cdots, t_0$.
\item[(III)] Now we are in the similar situation as the beginning of step (II). We evolve both families under $\mathcal{F}^{t_0}$, and consider the higher iteration map $\mathcal{F}^{t_0}$.
Now we consider another first return map $L :\mathcal{R}^*\to \mathcal{R}^*$, and a first return time $\tau_0:\mathcal{R}^*\to\mathbb{N}$, such that each level set $(\tau_0=n)$ is a $s$-subset of $\mathcal{R}^*$, and $\mathcal{F}^{n t_0} (\tau_0=n)$ properly return to $\mathcal{R}^*$ for the first time. Moreover, for $\mu$-almost every $x\in (\tau_0=n)$, we define $L x=\mathcal{F}^{n t_0}x$.
By applying Lemma \ref{renewal1}, it is guaranteed that $\mathcal{F}^{t_0}\hat\mathcal{K}^i_1$ has at least $\delta>0$ portion of the measure properly returned to $\mathcal{R}^*$. We again apply Lemma \ref{couplescheme} to couple at least $\delta/3$ portion of part of the remaining measure from each family.
The above algorithm is recurrent, thus by induction, at least the same portion $\delta/3$ of remaining probability measures could be coupled by applying Lemma \ref{couplescheme} under every iterations $t_0$.
\end{itemize}
An intuitive way of visualizing this coupling scheme is that, although both families $\nu^1$ and $\nu^2$ may have totally different first return distributions to $\mathcal{R}^*$, once their supports properly crosses $\mathcal{R}^*$ for the first time, the returning measures will have almost ``uniform" mixing, according to the Markov structure of $\mathcal{R}^*$. The processes can be approximated by a renewal process associated with an ergodic Markov chain with countably many states.
\end{comment}
\subsection{Proof of the Coupling Lemma}\label{proofcoupling}
\noindent{\textbf{Step 0. Decompose both families into sums of generalized standard families.}}\\
We first iterate both families for $N=N(\nu^1,\nu^2)$ iterations, such that
both families $\mathcal{F}^N\mathcal{G}^i$ are proper standard families, when restricting on $M$, and have at least $\delta>0$ portion of measure properly return to $\mathcal{R}^*$, with $\delta(1-\mathbf{a})/2>10\varepsilon_{\mathbf{d}}$.
For $i=1,2$, let $\mathcal{G}^i_0=(\mathcal{W}^i_0,\nu^i_0):=\mathcal{F}^N\mathcal{G}^i$. By Lemma \ref{extralemma2}, there exists a sequence of $\mathcal{F}-$ generalized standard families $\{(\mathcal{W}^i_{0,n},\nu^i_{0,n}), n\geq 0\}$, such that $$\mathcal{G}^i_0=\sum_{m=0}^{\infty} \mathcal{G}^i_{0,m}=\sum_{m=0}^{\infty}(\mathcal{W}^i_{0,m},\nu^i_{0,m}),$$ where $\mathcal{G}^i_{0,m}$ is a generalized standard family with index $m$, such that $\mathcal{F}^{m}\mathcal{W}^i_{0,m}$ properly returns $\mathcal{R}^*$. In addition, $$\nu^i_{0,n}(\mathcal{M})\leq C \|g^i\|_{\infty}n^{-\alpha_0},$$ for some constant $C>0$.
Note that \begin{equation}\label{delta000}s^i_0:=\nu^i_{0,0}(\Gamma^s)>\delta,\end{equation} Also for convenience, we denote $a_n^i=\nu^i_{0,n}(\mathcal{M})$. Clearly, $a_0^i=s^i_0$.
\\
\noindent{\textbf{Step 1. Capture and then coupling along $\Gamma^s$.}}\\
Since the family $\mathcal{G}^i_{0,0}$ is a generalized standard family with index $0$, we can apply Lemma \ref{couplescheme}, to get a decomposition $$\mathcal{G}^i_{0,0}= \mathcal{E}^i_{0}+ \mathcal{K}^i_{0},$$ with $$\mathcal{E}^i_{0}=(\mathcal{W}^i_{0,0}, \eta^i_{0}),$$ and $$\mathcal{K}^i_{0}=(\mathcal{W}^i_{0, 0 }, \xi^i_{0 }),$$ where $\mathcal{E}^i_0$ is a generalized standard family of index $0$. Note that $\mathcal{E}^1_{0}$ and $\mathcal{E}^2_{0}$ are coupled along stable manifolds in $\Gamma^s$. More precisely,
for any $f\in\mathcal{H}^-(\gamma_0)$ that is constant on each $W^s\in\Gamma^s$, \begin{equation}\label{couple2step1}\eta^1_{0}(f)=\eta^2_{0 }(f).\end{equation} Moreover, $$\eta^i_{0 }(\Gamma^s)=c_{0}\min\{s^1_0, s^2_0\}\geq c_0\delta,$$ where $c_{0}$ is chosen in $ [(1-\mathbf{a})/2,1-\mathbf{a}]$. Thus by (\ref{delta000}), we have
$$d^i_{0}:=\frac{\eta^i_0(\Gamma^s)}{s^i_0}\geq \frac{c_{0}\min\{s^1_0, s^2_0\}}{\max\{s^1_0, s^2_0\}}\in ( \delta(1-\mathbf{a})/2, 1-\mathbf{a}).$$
This implies that we can ``match" at least $(1-\mathbf{a})/2$ portion of each measure from $\nu^1_{0, 0 }$ and $\nu^2_{0, 0}$, as they both have proper returned to $\mathcal{R}^*$.
In terms of measures, note that
$\mathcal{G}^i_0=\sum_{m\geq 0}\mathcal{G}^i_{0,m}.$
After the first step of coupling, we know that the remaining family can be denoted as:
$$\mathcal{G}^i_1=(\mathcal{W}^i_1,\nu^i_1):=\sum_{m\geq 1}\mathcal{G}^i_{0,m}+\mathcal{K}^i_0=\sum_{m\geq 1}(\mathcal{W}^i_{0,m},\nu^i_{0,m})+(\mathcal{W}^i_{0,0}, \xi^i_0),$$
with $$\xi^i_0(\mathcal{M})=(1-d^i_0)\nu^i_{0,0}(\mathcal{M}).$$
Next, we need to estimate the total amount of points (in $\nu^i_1$) that properly returned to $\mathcal{R}^*$ under $\mathcal{F}$. We denote $A_0=\mathcal{R}^*$, and for any $n\geq 1$, $$A_n=\{x\in \mathcal{R}^*\,:\, \tau_0(x)+\cdots+\tau_0(L^{k-1} x)=n, \text{ for some } k=1,\cdots,n\}.$$
Then one can check that
$A_{1}=(\tau_0=1)$.
Thus the total amount of points (in $\nu^i_1$) that properly returned to $\mathcal{R}^*$ under $\mathcal{F}$ satisfies
\begin{align*}s^i_1&:=\nu^i_{0,1}(\mathcal{M})+\xi^i_{0}(A_{1})=\nu^i_{0,1}(\mathcal{M})+\xi^i_{0}(\tau_0=1|A_0)\xi^i_0(A_0)\\
&=\nu^i_{0,1}(\mathcal{M})+\nu^i_{1}(\tau_0=1|A_0)\xi^i_0(\mathcal{M})\\
&=\nu^i_{0,1}(\mathcal{M})+(1-d_0^i)\nu_1^i(\tau_0=1|A_0) s^i_0,\end{align*}
where we used the fact that $s^i_0=\nu^i_{0}(A_0)=\nu^i_{0,0}(\mathcal{M})$.
We put $$\hat\mathcal{K}^i_1=(\hat\mathcal{W}^i_1,\hat\nu^i_1):=\mathcal{K}^i_{0}|_{A_1}+\mathcal{G}^i_{0,1}.$$
Now we have shown that the total uncoupled portion after the first step can be represented as:
\begin{equation}\label{cGi100}\mathcal{G}^i_0- \mathcal{E}^i_0= \hat\mathcal{K}^i_1+\mathcal{K}^i_{0}|_{A_1^c}+\sum_{m\geq 2} \mathcal{G}^i_{0,m}.\end{equation}
To summarize, the remaining families contain three parts:
\begin{itemize}
\item[(1a)]$\sum_{m\geq 2} \mathcal{G}^i_{0,m}$ is the part that has not yet reached $\mathcal{R}^*$ after two steps;
\item[(1b)]$\hat\mathcal{K}^i_1$ corresponds to the portion that has reached $\mathcal{R}^*$ at the second step;
\item[(1c)] $\mathcal{K}^i_{0}|_{A_0^c}$ corresponds to the portion in $\mathcal{K}^i_0$ that will not return to $\mathcal{R}^*$ at the second step.
\end{itemize}
Here $s^i_1$ is the total amount of measure that have returned to $\mathcal{R}^*$ at the second step, for $i=1,2$. Moreover using (\ref{edtau0}), we have
\begin{equation}\label{si1}
s^i_1=\nu^i_{0,1}(\mathcal{M})+(1-d_0^i)\nu_0^i(\tau_0=1|A_0) s^i_0\geq \delta_1^i:=\nu^i_{0,1}(\mathcal{M})+\delta p_1(1-\varepsilon_{\mathbf{d}})/2\end{equation}
using the fact that $\delta<s^i_0=a^i_0$. In addition, we have
$$s^i_1\leq a^i_1+ p_1(1+\varepsilon_{\mathbf{d}})(1-d_0)a^i_0\leq \hat\delta_1^i:=a_1^i+\theta p_1a^i_0,
$$
where $d_0:=\delta(1-\mathbf{a})/2$, using the fact that $\varepsilon_{\mathbf{d}}\ll \delta(1-\mathbf{a})/2$, we can choose $\theta\in (0,1)$, such that $(1+\varepsilon_{\mathbf{d}})(1-d_0)\in (0,\theta)$.
Thus \begin{equation}\label{d1}d^i_1:=\frac{c_0\min\{\delta_1^1,\delta_1^2\}}{\hat\delta_1^i}\end{equation}
\begin{comment}
\noindent{\textbf{Step 2. Release and recapture.}}\\
Note that $\mathcal{F}\hat\mathcal{K}^i_1$ has properly returned to $\mathcal{R}^*$, for $i=1,2$. It follows from (\ref{si1}) that
$\mathcal{F}\hat\mathcal{K}^i_1(\mathcal{R}^*)=s^i_1\geq \delta_1$, for $i=1,2$.
Since we are in a similar situation as in step 1, we can apply Lemma \ref{couplescheme}, to get a decomposition
$$\mathcal{F}\hat\mathcal{K}^i_1= \mathcal{F}\mathcal{E}^i_{1}+ \mathcal{F}\mathcal{K}^i_{1},$$ with $$\mathcal{E}^i_{1}=(\hat{\mathcal{W}}^i_{1}, \eta^i_{1}),\,\,\,\,\,\mathcal{K}^i_{1}=(\hat{\mathcal{W}}^i_{1}, \xi^i_{1}).$$ Note that $\mathcal{F}\mathcal{E}^1_{1}$ and $\mathcal{F}\mathcal{E}^2_{1}$ are coupled along stable manifolds in $\Gamma^s$. More precisely, for any measurable collection of stable manifolds we have:
$A\subset \Gamma^s$, \begin{equation}\label{couple2step2}\mathcal{F}_*\eta^1_{1}(A)=\mathcal{F}_*\eta^2_{1}(A)\end{equation} Moreover, by (\ref{d1}), we know that $$d^i_{1}:=\mathcal{F}\eta^i_{1}(\mathcal{R}^*)/\mathcal{F}_*\hat\nu^i_1(\mathcal{R}^*)\in (d_1, 1/2).$$
We apply the first statement of
Lemma \ref{renewal1} to $\mathcal{F}\mathcal{K}^1_1$ and $\mathcal{F}\mathcal{K}^2_1$, as they are pseudo -generalized standard families of index zero. Thus both $\mathcal{F}^{2}\mathcal{K}^1_1$ and $\mathcal{F}^{2}\mathcal{K}^2_1$ have at least $\delta$ portion of measures properly returned to $\mathcal{R}^*$.
After the second step of coupling, we know that the remaining family can be denoted as:
$$\mathcal{G}^i_1=(\mathcal{W}^i_2,\nu^i_2):=\sum_{m\geq 2}\mathcal{G}^i_{0,m}+\mathcal{K}^i_1+\mathcal{K}^i_{0}|_{A_1^c}$$
We need to estimate the total amount of points (in $\nu^i_2$) that properly returned to $\mathcal{R}^*$ under $\mathcal{F}^{2}$.
Then one can check that
$A_{2}=(\tau_0=2)\cup (\mathcal{F}^{-1}(\tau_0=1)\cap A_1)$.
Thus the the total amount of points that properly returned to $\mathcal{R}^*$ under $\mathcal{F}^2$ satisfies
\begin{align*}s^i_2&:=\nu^i_{0,2}(\mathcal{M})+\nu^i_{2}(A_{2})\\
&=\nu^i_{0,2}(\mathcal{M})+\nu^i_{2}(\tau_0=2)+\nu^i_2 (\mathcal{F}^{-1}(\tau_0=1)\cap A_1)\\
&=\nu^i_{0,2}(\mathcal{M})+\xi^i_0(\tau_0=2)+\xi^i_1 (\mathcal{F}^{-1}(\tau_0=1)\cap A_1)\\
&=\nu^i_{0,2}(\mathcal{M})+(1-d_0^i)\xi^i_{0}(\tau_0=2|A_0)\nu^i_0(A_0)+(1-d_1^i)\mathcal{F}_*\xi^i_1 (\tau_0=1|\mathcal{F} A_1)\nu^i_1(A_1)\\
&=\nu^i_{0,2}(\mathcal{M})+(1-d_0^i)\xi^i_{0}(\tau_0=2|A_0)s^i_0+(1-d_1^i)\mathcal{F}_*\xi^i_1 (\tau_0=1|\mathcal{F} A_1)s^i_1\end{align*}
We define $$\hat\mathcal{K}^i_2=(\hat\mathcal{W}^i_2,\hat\nu^i_2):=\mathcal{K}^i_{1}|_{A_2}+\mathcal{K}^i_{0}|_{A_2/A_1}+\mathcal{G}^i_{0,2}.$$
Thus the total remaining family can be represented as $\sum_{m\geq 3} \mathcal{G}^i_{0,m}.$
Here $s^i_2$ is the total amount of measure that have returned to $\mathcal{R}^*$ within the first three steps, for $i=1,2$. Moreover
\begin{align}\label{si12}
s^i_2&=\nu^i_{0,2}(\mathcal{M})+(1-d_0^i)\xi^i_{0}(\tau_0=2|A_0)s^i_0+(1-d_1^i)\mathcal{F}_*\xi^i_1 (\tau_0=1|\mathcal{F} A_1)s^i_1\nonumber\\
&\geq \nu^i_{0,2}(\mathcal{M})+\nu^i_{0}(\tau_0=2|A_0)\delta/2+\mathcal{F}_*\nu^i_1 (\tau_0=1|\mathcal{F} A_1)\delta^i_1/2\nonumber\\
&\geq \delta_2:=\nu^i_{0,2}(\mathcal{M})+(1-\varepsilon_{\mathbf{d}})\left(p_2\delta+p_1\delta_1\right)/2
\end{align}
Similarly, we get
\begin{align*}
s^i_2&\leq a^i_2+(1+\varepsilon_{\mathbf{d}})(1-d_0)a_0^ip_2+(1+\varepsilon_{\mathbf{d}})(1-d_1^i)p_1
\hat\delta^i_1\\
&\leq \hat\delta^i_2:=a_2^i+\theta (p_2 a_0^i+p_1 \hat\delta_1^i)
\end{align*}
where we need $\varepsilon_{\mathbf{d}}<d_1^i$, and $(1+\varepsilon_{\mathbf{d}})(1-d_1)\in (0,\theta)$.
Thus \begin{equation}\label{d1}d^i_2:=\frac{c_0\min\{\delta_2^1,\delta_2^2\}}{\hat\delta_2^i}=\frac{c_0\min\{\tilde A_2^1,\tilde A_2^2\}}{\hat\delta_2^i}\end{equation}
This implies that
both $\mathcal{F}\nu_2$ and $\mathcal{F}\nu_2$ have at least $\delta_2$ portion of measures properly returned to $\mathcal{R}^*$.
Thus we are in similar situation as in the beginning of step 2.
Combining with (\ref{cGi100}), we know that the total uncoupled portion at the second step can be represented as:
\begin{equation}\label{firststepFtau2''}\mathcal{G}^i_0-\mathcal{E}^i_0-\mathcal{E}_1^i=\hat\mathcal{K}^i_{2}+ \mathcal{K}^i_1|_{A_2^c}+\mathcal{K}^i_{0}|_{(A_2\cup A_1)^c}+\sum_{m\geq 3} \mathcal{G}^i_{0,m}.\end{equation}
We need to estimate the total amount of points (in $\nu^i_2$) that properly returned to $\mathcal{R}^*$ under $\mathcal{F}^{2}$.
Then one can check that
$A_{3}=(\tau_0=3)\cup (\mathcal{F}^{-1}(\tau_0=2)\cap A_1)\cup (\mathcal{F}^{-2 }(\tau_0=1)\cap A_2)$.
Thus the the total amount of points that properly returned to $\mathcal{R}^*$ under $\mathcal{F}^2$ satisfies
\begin{align*}s^i_3&:=\nu^i_{0,3}(\mathcal{M})+\nu^i_{3}(A_{3})\\
&=\nu^i_{0,3}(\mathcal{M})+\nu^i_{3}(\tau_0=3)+\nu^i_3 (\mathcal{F}^{-1}(\tau_0=2)\cap A_1)+\nu^i_3 (\mathcal{F}^{-2 }(\tau_0=1)\cap A_2)\\
&=\nu^i_{0,3}(\mathcal{M})+\xi^i_0(\tau_0=3)+\xi^i_1 (\mathcal{F}^{-1}(\tau_0=2)\cap A_1)+\xi^i_2 (\mathcal{F}^{-2 }(\tau_0=1)\cap A_2)\\
&=\nu^i_{0,3}(\mathcal{M})+(1-d_0^i)\xi^i_{0}(\tau_0=3|A_0)\nu^i_0(A_0)+(1-d_1^i)\mathcal{F}_*\xi^i_1 (\tau_0=2|\mathcal{F} A_1)\nu^i_1(A_1)+(1-d_2^i)\mathcal{F}_*\xi^i_2 (\tau_0=1|\mathcal{F} A_2)\nu^i_2(A_2)\\
&=\nu^i_{0,3}(\mathcal{M})+(1-d_0^i)\xi^i_{0}(\tau_0=3|A_0)s^i_0+(1-d_1^i)\mathcal{F}_*\xi^i_1 (\tau_0=2|\mathcal{F} A_1)s^i_1+(1-d_2^i)\mathcal{F}^2_*\xi^i_2 (\tau_0=1|\mathcal{F} A_2)s^i_2\\
&\geq \nu^i_{0,3}(\mathcal{M})+\nu^i_{0}(\tau_0=3|A_0)\delta/2+\mathcal{F}_*\xi^i_1 (\tau_0=2|\mathcal{F} A_1)\delta^i_1/2+\mathcal{F}^2_*\xi^i_2 (\tau_0=1|\mathcal{F}^2 A_2)\delta^i_2/2\nonumber\\
&\geq \delta_3:=\nu^i_{0,3}(\mathcal{M})+(1-\varepsilon_{\mathbf{d}})\left(p_3\delta+p_2\delta_1+p_1\delta_2\right)/2
\end{align*}
\end{comment}
\noindent{\textbf{Step 3. Coupling at the repeated proper returns.}}\\
Next we consider higher iterations by induction, with $k\geq 2$.
We assume $s^i_j$ is well defined, for $j=1,\cdots, k-1$, and for $k\geq 1$,
\begin{equation}\label{k-1inductG}
\mathcal{G}^i_0=\mathcal{E}^i_0+\mathcal{E}_1^i+\mathcal{E}^i_2+\cdots+\mathcal{E}^i_{k-1}+\hat\mathcal{K}^i_{k}+\mathcal{K}^i_{k-1}|_{A_{k}^c}+\mathcal{K}^i_{k-2}|_{A_{k}^c\cup A_{k-1}^c}+\mathcal{K}^i_1|_{\cup_{j=2}^{k-1} A_j^c}+\mathcal{K}^i_{0}|_{\cup_{j=1}^{k-1} A_j^c}+\sum_{m\geq k+1} \mathcal{G}^i_{0,m}.
\end{equation}
Here for any $n\leq k-1$, $\mathcal{E}^i_{n}=(\hat{\mathcal{W}}^i_n,\eta^i_n)$ is a generalized standard family with index $n$, such that $\mathcal{E}^1_n$ and $\mathcal{E}^2_n$ are coupled along stable manifolds $\Gamma^s$, and satisfying
$$\eta^1_n(A)=\eta^2_n(A),$$ for any measurable collection $A$ of stable manifolds in $\Gamma^s$. In addition, $\eta^i_n(\mathcal{M})=d_n^i s^i_n$.
Moreover,
$$\hat\mathcal{K}^i_{k}=(\hat{\mathcal{W}}^i_{k}, \hat{\nu}^i_{k})$$ is a generalized standard family with index $k$, i.e. $\mathcal{F}^{k}\hat\mathcal{K}^i_{k}$ properly return to $\mathcal{R}^*$. The family $\mathcal{K}^i_n=(\mathcal{W}^i_n,\xi^i_n)$ is the uncoupled family in $\hat\mathcal{K}^i_{n}$, i.e. $\hat\mathcal{K}^i_{n}=\mathcal{E}^i_n+\mathcal{K}^i_n$.
We need to estimate the total amount of points (in $\nu^i_{k}$) that properly returned to $\mathcal{R}^*$ under $\mathcal{F}^{k}$.
Thus the the total amount of points that properly returned to $\mathcal{R}^*$ under $\mathcal{F}^{k}$ satisfies
\begin{align*}s^i_k&:=\nu^i_{0,k}(\mathcal{M})+\nu^i_{k}(A_{k})\\
&=\nu^i_{0,k}(\mathcal{M})+\sum_{j=0}^{k-1}\nu^i_k (\mathcal{F}^{-j}(\tau_0=k-j)\cap A_j)\\
&=\nu^i_{0,k}(\mathcal{M})+\sum_{j=0}^{k-1}\xi^i_j (\mathcal{F}^{-j}(\tau_0=k-j)\cap A_j)\\
&=\nu^i_{0,k}(\mathcal{M})+\sum_{j=0}^{k-1}(1-d_{j}^i)\mathcal{F}^{j}_*\xi^i_{j}(\tau_0=k-j|\mathcal{F}^{j}A_j)\nu^i_j(A_j)\\
&=\nu^i_{0,k}(\mathcal{M})+\sum_{j=0}^{k-1}(1-d_{j}^i)\mathcal{F}^{j}_*\xi^i_{j}(\tau_0=k-j|\mathcal{F}^{ j}A_j)s^i_j.\end{align*}
Next, we can estimate the lower and upper bound of $s^i_k=\hat{\nu}^i_{k}(\mathcal{M})$.
\begin{align*}
s^i_k&\geq \nu^i_{0,k}(\mathcal{M})+\sum_{j=0}^{k-1}(1-d_{j}^i)p_{k-j}(1-\varepsilon_{\mathbf{d}})\delta^i_j\\
&\geq \delta^i_k:=a^i_k+\frac{1-\varepsilon_{\mathbf{d}}}{2}\sum_{j=0}^{k-1}p_{k-j}\delta^i_j.\end{align*}
We apply Lemma \ref{lemmarenewal}, to get that there exist $c_1<c_2$, such that
\begin{equation}\label{lowersi}
\frac{c_1(1-\varepsilon_{\mathbf{d}})}{2} k^{-\alpha_0}\leq \delta^i_k\leq c_2 k^{-\alpha_0}.
\end{equation}
Next we consider the upper bound.
We assume, for $l=1,\cdots, k-1$,
\begin{align}\label{sk}
s^i_l&\leq \hat\delta^i_l:=a_l^i+\theta \sum_{j=0}^{l-1}p_{k-j}\hat\delta^i_j,\end{align}
here we assume that $(1+\varepsilon_{\mathbf{d}})(1-d^i_l)\in (0,\theta)$, for $l=1,\cdots k-1$. According to Lemma \ref{extralemma2}, we know that
\begin{equation}\label{aki}
a_l^i\leq C l^{-\alpha_0}.
\end{equation}
Next, we claim that $s_k^i$ has similar property. We define
\begin{equation}\label{deltatheta}
\hat\delta^i_{k}:=a_k^i+\theta \sum_{j=0}^{k-1}p_{k-j}\hat\delta^i_j.\end{equation}
We need to estimate the upper bound of $\hat\delta^i_{k}$, as well as the lower bound of $d^i_k$. Note that by Lemma \ref{lemmarenewal}, we introduce some new auxiliary series, as the following:
$$\Delta^i_k=\left\{
\begin{array}{ll}
\hat \delta^i_l, & {l=1,\cdots, k;} \\
0, & \text{otherwise.}
\end{array}
\right. \,\,\,\,\,\,\,\,\,\,\,\,\text{ and }\,\,\,\,\,\,\,\,\,\hat a^i_l=\left\{
\begin{array}{ll}
a^i_l, & {l=1,\cdots, k;} \\
0, & \text{otherwise.}
\end{array}
\right.
$$
Now (\ref{deltatheta}) implies that
\begin{equation}\label{deltatheta2}
\Delta^i_l:=\hat a_l^i+\theta \sum_{j=0}^{l-1}p_{l-j}\Delta^i_j,\,\,\,\,\,l\geq 1\end{equation}
Thus by Lemma \ref{lemmarenewal}, there exists constant $c_2>0$, such that for $n\leq k$,
$$ \hat\delta^i_{n} \leq c_2 n^{-\alpha_0}.$$
This implies that
\begin{equation}\label{sik}\frac{c_1(1-\varepsilon_{\mathbf{d}})}{2}k^{-\alpha_0}\leq s^i_k\leq c_2 k^{-\alpha_0}.\end{equation}
Note that \begin{align*}
d_{k}^i:&=\frac{c_0\min\{s_{k}^1,s_{k}^2\}}{s_{k}^i}\geq \frac{c_0 c_1(1-\varepsilon_{\mathbf{d}})}{2c_2 }.\end{align*}
Indeed the above estimation is true for any $n\leq k$, i.e.
\begin{equation}\label{defnd}d_n^i \geq d:=\frac{c_0 c_1(1-\varepsilon_{\mathbf{d}})}{2c_2 }.\end{equation}
By assumption (\ref{epsbdchoose}), we have $\varepsilon_{\mathbf{d}}< \frac{c_0 c_1(1-\varepsilon_{\mathbf{d}})}{4c_2 }$, thus we can proved inductively that $(1-d_k^i)(1-\varepsilon_{\mathbf{d}})<\theta
$, for any $k\geq 1$.
This implies that at every step, we can couple $d^i_k\geq d$ portion of the remaining measure that properly return to $\mathcal{R}^*$ under $\mathcal{F}^{k }$.
\begin{comment}
Combining with (\ref{k-1inductG}), we have shown that
\begin{equation}\label{firststepFtauk+1''}
L^{k}\mathcal{G}^i_1=L^{k}\mathcal{E}_1^i+s_1^iL^{k-1}\mathcal{E}^i_2+\cdots+s^i_{k}L\mathcal{E}^i_{k+1}+s_1^iL^{k-1}\mathcal{U}^i_2+\cdots+s^i_{k}L\mathcal{U}^i_{k+1}+s^i_{k+1}\hat\mathcal{K}^i_{k+1}+ \sum_{m\geq k+2} \mathcal{G}^i_{0,m}.\end{equation}
Inductively, we can also get a formula for $s^i_{k+1}$:
\begin{align*}
s^i_{k+1}&=Q^i_{k+1}+(1-d^i_{k+1})Q^i_k+(1-d^i_{k+1})(1-d^i_k)Q^i_{k-1}+\cdots+\Pi_{j=0}^{k+1}(1-d_{j}^i)Q^i_0\\
&\leq \sum_{m=0}^{k+1}Q^i_{k+1-m}\zeta^m\leq CQ^i_{k+1},
\end{align*}
where $\zeta=1-\delta/3$.
Thus the total uncoupled measure at $k+1$-th step is:
\begin{align}\label{tauk}
\hat\nu^i_{k+1}(\mathcal{M})+\tilde\nu^i_{k+1}(\mathcal{M})&=s^i_{k+1}+\tilde\nu^i_{k+1}(\mathcal{M})=\tilde\nu^i_{k+1}(\mathcal{M})+\mathcal{O}(Q^i_{k+1})\\
&=\nu_1^i(\tau>kt_0)+\mathcal{O}(\nu_1^i(kt_0\leq \tau\leq (k+1) )).\end{align}
\bigskip
\noindent\textbf{Step 4. Rearrange the coupled measures by the real iteration time under $\mathcal{F}$.}\\
\end{comment}
Now we rearrange the above coupled and uncoupled families according to the real iteration time under $\mathcal{F}$. Then
we have shown that for any $n\geq 1$, there is a decomposition
$$\mathcal{G}^i_0=\mathcal{E}^i_0+\mathcal{E}_1^i+\mathcal{E}^i_2+\cdots+\mathcal{E}^i_{n}+\bar\mathcal{G}^i_n,$$ for $i=1,2$, where $\bar\mathcal{G}^i_n=(\bar\mathcal{W}^i_n,\bar\nu^i_n)$ is defined as the remaining uncoupled family after step $n$, such that
$$\bar\mathcal{G}^i_n=\hat\mathcal{K}^i_{n+1}+\mathcal{K}^i_{n}|_{A_{n+1}^c}+\mathcal{K}^i_{n-1}|_{A_{n}^c\cup A_{n}^c}+\cdots+\mathcal{K}^i_1|_{\cup_{j=2}^{n+1} A_j^c}+\mathcal{K}^i_{0}|_{\cup_{j=1}^{n+1} A_j^c}+\sum_{m\geq n+1} \mathcal{G}^i_{0,m}.$$
Inductively, one can show that, by the mixing property, that $\mathcal{G}^i_0$ has a decomposition into $\{\mathcal{E}^i_k, k\geq 0\}$, i.e we can also denote this decomposition as $$\mathcal{G}^i_0=\sum_{k=0}^{\infty}\mathcal{E}^i_k.$$
Moreover, the leftover at $n$-th step, $\bar\mathcal{G}^i_n$ can be represented as
$$\bar\mathcal{G}^i_n=\sum_{k=n+1}^{\infty}\mathcal{E}^i_k=\sum_{k=n+1}^{\infty}(\hat{\mathcal{W}}^i_k,\eta^i_n).$$
Note that $\eta^i_n(\mathcal{M})= d_n^i s_n^i$, which implies that the coupled measure at the $n$-th step is $$\frac{dc_1(1-\varepsilon_{\mathbf{d}})}{2}n^{-\alpha_0}\leq d s_n^i\leq \eta^i_n(\mathcal{M})= d_n^i s_n^i\leq s_n^i\leq c_2 n^{-\alpha_0}.$$
Using the estimation (\ref{sik}),
so
the remaining uncoupled measure at the $n$-th step is
\begin{equation}\label{barnucM}\bar\nu^i_n(\mathcal{M})=\sum_{k=n+1}^{\infty} d_n^i s_n^i\leq \sum_{k=n+1}^{\infty} s_n^i\leq C n^{1-\alpha_0}.\end{equation}
Note that for any $n\geq 0$, $\mathcal{E}^i_{n}=(\hat{\mathcal{W}}^i_n,\eta^i_n)$ is a generalized standard family with index $n$, such that $\mathcal{F}^n\mathcal{E}^1_n$ and $\mathcal{F}^n\mathcal{E}^2_n$ are coupled along stable manifolds $\Gamma^s$, and satisfying
$$\mathcal{F}^n_*\eta^1_n(A)=\mathcal{F}^n_*\eta^2_n(A),$$ for any measurable collection $A$ of stable manifolds in $\Gamma^s$.
Thus for any measurable function $f$ that is constant on each $W^s\in \Gamma^s$, we have $$\mathcal{F}^{n}\eta^1_n(f)=\mathcal{F}^n\eta_n^2(f).$$ This verifies the items (\textbf{C1})(i)-(ii) in the Coupling Lemma \ref{coupling1}
An important fact following from (\ref{defnd}) is the exponential decay of correlations under the induced map $F$. Note that (\ref{defnd}) implies that one can couple at least $d$ portion of the proper return measure under each iteration $\mathcal{F}$, which implies that one can also couple at least $d$ portion of the proper return measure under each iteration of the induced map $F$. Thus we obtain by the Coupling Lemma \ref{coupling}, that for the two proper standard family $\mathcal{G}^i_0:=\mathcal{F}^N(\mathcal{W}^i, \nu^i)$, for any $n\geq 1$,
there exist two sequences of $F-$ generalized standard families $\{(\hat{\mathcal{W}}^i_n,\hat{\nu}^i_n), n\geq 0\}$, such that \begin{equation}\label{decomposeGi1}\mathcal{G}^i_0=\sum_{n=0}^{\infty}(\hat{\mathcal{W}}^i_n,\hat{\nu}^i_n):=\left(\bigcup_{n=0}^{\infty}\hat{\mathcal{W}}^i_n,\sum_{n=0}^{\infty}\hat{\nu}^i_n\right).
\end{equation} And for each $n\geq 0$, both $(\hat{\mathcal{W}}^1_n,\hat{\nu}^1_n)$ and $(\hat{\mathcal{W}}^2_n,\hat{\nu}^2_n)$ are $F-$ generalized standard families of index $n$; $F^n_*\hat{\nu}^1_n$ and $F^n_*\hat{\nu}^2_n$ are coupled along stable manifolds in $\Gamma^s$.
For any $n\geq 1$, the uncoupled measure satisfies
\begin{equation}\label{ctail2}\sum_{k\geq n}\hat{\nu}^i_k(M)<C\vartheta^n,\end{equation}
where $C>0$ and $\vartheta$ are uniform constants.
From now on, we choose the large constant $b=b(\gamma_0,\Lambda,\vartheta,{\alpha_0})>1$ such that
\begin{equation}\label{dpbchi}
\Lambda^{-\gamma_0 b\ln n}<n^{-1-{\alpha_0}},\,\,\,\,\,\,\,\,\vartheta^{ b\ln n}<n^{-1-{\alpha_0}},
\end{equation}
where $\vartheta\in (0,1)$ is given by (\ref{ctail}).
\\
Note that (\ref{barnucM}) implies that for any $f\in L_{\infty}(\mathcal{M})$, we have
$$|\bar\nu^i_n(f)|\leq \|f\|_{\infty} \bar\nu^i_n(\mathcal{M})\leq C\|f\|_{\infty} n^{1-\alpha_0}.$$
\qed
Next we claim that
\begin{equation}\label{cnbc}
\bar\nu^i_n(C_{n,b}^c\cap (R\leq n)) \leq C n^{-1-{\alpha_0}}.
\end{equation}
To see this, note that points in $C_{n,b}^c\cap (R\leq n)$ will
mostly visit cells with small indices and return to $M$ at least $\psi$ times within $n$ iterations. We prove this claim by considering two cases.
\quad(\textbf{a}). Let $I_n$ be all points $x\in C_{n,b}^c\cap (R\leq n)$, such that the iterations of $x$ hits $\mathcal{R}^*$ at most $b\ln n$ times within $n$ iterations.
Then there exists $k\in [1,n-b\ln n]$, such that $\mathcal{F}^kx\in M$ and the forward trajectory of $\mathcal{F}^k x$ return to $M\setminus \mathcal{R}^*$ at least $b\ln n $ consecutive times under $F$.
According to Lemma \ref{extralemma1}, we know that the standard family $(\mathcal{W}^u_M,\mu_M)$ has a decomposition into $\{(\mathcal{W}_n,\mu_n), n\geq 1\}$, and by (\ref{ctail1}),
$\sum_{m=n}^{\infty}\mu_m(M)\leq C \vartheta^n$.
Thus $\mathcal{F}^k x$ belongs to the support of $\sum_{m\geq b\ln n} \mu_m$, which satisfies:
\begin{align*}
\mu_M(C_{n,b}^c\cap (R\leq n)\cap I_n)&\leq \sum_{m=b\ln n}^{\infty}\mu_m(M)\\
&\leq C\vartheta^{b\ln n}=\mathcal{O}(n^{-1-\alpha_0}),\end{align*}
where we have used (\ref{dpbchi}) in the last step.
\quad(\textbf{b}). Now we consider points in $C_{n,b}^c\cap (R\leq n)\setminus I_n$. Then we claim that for any $x\in (C_{n,b}^c\cap (R\leq n)\setminus I_n)$, iterations of $x$ hit $\mathcal{R}^*$ at least $b\ln n$ times within the $(b\ln n)^2$ returns to $M$, but have not been coupled. This is true because otherwise there must be an interval of length $b\ln n$ in these at least $(b\ln n)^2$ returns to $M$ such that iterates of $x$ never hit $\mathcal{R}^*$, and this contradicts the assumption that (\textbf{a}) does not hold. Thus there exists $k\in [1,n-b\ln n]$, such that $\mathcal{F}^kx\in \mathcal{R}^*$ and the forward trajectory of $\mathcal{F}^k x$ return to $\mathcal{R}^*$ at least $b\ln n $ times. Thus $\mathcal{F}^k x$ belongs to the support of $\sum_{k\geq b\ln n}\hat{\nu}^i_k$. Using (\ref{ctail2}), we know that $$\sum_{k\geq b\ln n}\hat{\nu}^i_k(M)<C\vartheta^{b\ln n}=\mathcal{O}(n^{-1-\alpha_0}).$$
This implies that $$\bar\nu_n^i(C_{n,b}^c\cap (R\leq n)\setminus I_n)\leq C n^{-1-\alpha_0}.$$
This claim implies the last statement (\textbf{C3}):
\begin{align*}
\bar\nu_n^i(\mathcal{M})&=\nu^i(R>n)+\bar\nu_n((R\leq n)\cap C_{n,b})+\bar\nu_n^i((R\leq n)\cap C_{n,b}^c)\\
&= \nu^i(R>n)+\bar\nu_n^i((R\leq n)\cap C_{n,b})+ \mathcal{O}(n^{-1-\alpha_0}).\end{align*}
\begin{comment}Thus the amount of uncoupled measure is essentially dominated by $ C_{n,b}\cup(R>n)$, which contains points that returned
to $M$ fewer than $\psi=(b\ln n)^2$ times within $n$ iterations. Thus by Assumption (\textbf{H2}), we know that $$\mu(C_{n,b})\leq C n^{1-{\alpha_0}}.$$ Therefore, $$\nu^i(C_{n,b})\leq C\|g^i\|_{\infty} n^{1-{\alpha_0}},$$ where $d\nu^i=g^id\mu$.\\
Combining the above facts, we have shown that for any $n\geq 1$,
\begin{align*}
|\mathcal{F}^n_*\bar\nu^i_n(f)|&\leq C \,\|f\|_{\infty} \mu(C_{n,b}^c\cap (R<n))+|\bar\nu^i_n(f\circ \mathcal{F}^n \cdot\mathbf{I}_{C_{n,b}\cup(R>n)})| \\
&\leq C \,\|f\|_{\infty} n^{-{\alpha_0}}+\|f\|_{\infty}\|g^i\|_{\infty} \mu((C_{n,b}\cup(R>n))) \\
&\leq C\|g^i\|_{\infty} \|f\|_{\infty}n^{1-{\alpha_0}}.\end{align*}
This verifies \textbf{(C1)} (iii).\\
It follows from (\ref{sik}), \textbf{(C2)} also holds for $n\geq 1$.
\end{comment}
\section{Proof of Theorem 1.}
Now it is time to investigate the rates of decay of correlations using the above Coupling lemma. We first prove a lemma that will be used later.
We consider any two standard families $\mathcal{G}^i=(\mathcal{W}^i, \nu^i)$ with probability density $g^i=d\nu^i/d\mu\in\mathcal{H}^+(\gamma_0)$ for $i=1,2$, and $\mathcal{W}^i$ is a measurable partition of Borel set $B^i\subset \mathcal{M}$ into unstable manifolds. And we consider any $f\in \mathcal{H}^-(\gamma_1)$, with $\gamma_1 >0$. According to the Coupling Lemma \ref{coupling1}, for any $n\geq 1$, there exists a decompositions
$$\mathcal{F}^N_*\nu^i=\sum_{m=1}^{n}\nu_m^i+\bar\nu^i_n,$$ for $n\geq 1$, where $N=N(\nu_1,\nu_2)$, $\mathcal{F}^m_*\nu^1_m$ is coupled with $\mathcal{F}^m_*\nu^2_m$. This implies that for any $m\geq 1$, for any measurable function $h\in\mathcal{H}^-(\gamma_1)$ that is constant on any stable manifold $W^s\in\Gamma^s$, we have
\begin{equation}\label{coupling mapf1} \mathcal{F}^m_*\nu^1_m(h)=\mathcal{F}^m_*\nu^2_m(h).\end{equation}
\begin{lemma}\label{decayinflemma}
There exists $C_1=C_1(\gamma_1)>0$, which does not depend on $f$ and $g^i$, $i=1,2$, such that for any $n\geq 1$,
\begin{align*}
&\left|\sum_{m=1}^{n}(\nu_m^i(f)-\nu^2_m(f))\right| \leq C_1\|f\|^-_{C^{\gamma_1}} n^{-{\alpha_0}}.\end{align*}\end{lemma}
\begin{proof}
For any $n\geq 1$, by definition of $C_{n,b}^c$ and the choice of $b$, for $f\in\mathcal{H}^-(\gamma_1)$ there exists $C>0$ such that for $x\in C_{n,b}^c$ and
$y\in W^s(x)$,
\begin{equation}\label{holderbarx}|f(\mathcal{F}^n(x))-f(\mathcal{F}^n(y))|\leq C\|f\|^-_{\gamma_1}\Lambda^{-\gamma_1 b\ln n}\leq C\|f\|^-_{\gamma_1} n^{-{\alpha_0}}.\end{equation}
Now for any $x\in W^s\subset \Gamma^s$, we choose $\bar x\in W^s(x)$, such that $f(\bar x)=\max_{y \in W^s(x)} f(y)$ be the maximum value of $f$ along stable manifold $W^s(x)$. Then (\ref{holderbarx}) implies that for $x\in C_{n-m,b}^c\cap \Gamma^s$,
$$|f\circ \mathcal{F}^{n-m}(x)-f\circ\mathcal{F}^{n-m}(\bar x)|\leq C\|f\|_{\gamma_1}^-(n-m)^{-{\alpha_0}}.$$
Thus there exists $C_1=C_1(\gamma_1)>0$, such that for $i=1,2$,
\begin{align*}I_n^i:&=\sum_{m=1}^{n-1}\int_{\Gamma^s\cap C_{n-m,b}^c}\biggl| f \circ \mathcal{F}^{n-m}(x)-f\circ\mathcal{F}^{n-m}(\bar x)\biggr| d\mathcal{F}^{m}_*\nu_m^i(x)\\
&\leq C\|f\|_{\gamma_1}^-\sum_{m=1}^{n-1}\mathcal{F}^m_*\nu_m^i(\Gamma^s)(n-m)^{-{\alpha_0}}\\
&\leq C\|f\|_{\gamma_1}^-\left( \sum_{m=1}^{n/2}\mathcal{F}^m_*\nu_m^i(\Gamma^s)(n-m)^{-{\alpha_0}}+\sum_{m=n/2}^{n-1}\mathcal{F}^m_*\nu_m^i(\Gamma^s)(n-m)^{-{\alpha_0}}\right) \\&\leq C_1\|f\|_{\gamma_1}^- n^{-{\alpha_0}},\end{align*}
where we have used \textbf{(C2)} in the Coupling Lemma \ref{coupling1} in the last estimate, as well as the following estimate:
\begin{equation}\label{n-xtl}\int_{1}^{n-1} \frac{1}{x^l} \cdot \frac{1}{(n-x)^t} \, dx
\leq C_2 n^{-t}+ C_3 \frac{\ln n}{n^{l+t-1}}\leq C_1 n^{-t},\end{equation} for any $l\geq t\geq 1$.
Now we consider for $i=1,2$,
\begin{align*} I\!I^i_n\colon&=\sum_{m=1}^n\biggl|\int_{\Gamma^s\cap C_{n-m,b}} f \circ \mathcal{F}^{n-m}(x) d\mathcal{F}^{m}_*\nu_m^i(x)\biggr|\\
&\leq C\|f\|_{\infty}\sum_{m=1}^n \mathcal{F}^m_*\nu^i_m ( C_{(n-m,b)})\\
&\leq C\|f\|_{\infty}\sum_{m=1}^n(n-m)^{-{\alpha_0}} \mathcal{F}^m_*\nu^i_m(M)\\
&\leq C_2\|f\|_{\infty} n^{-{\alpha_0}},
\end{align*}
where we have used Lemma \ref{nuwcb2}, and $C_2=C_2(\gamma_1)>0$.
Combining the above estimates, we have \begin{align*}
I^1_n+I^2_n+I\!I^1_n+I\!I^2_n&\leq C\|f\|^-_{C^{\gamma_1}} n^{-{\alpha_0}},\end{align*}
for some constant $C=C(\gamma_1)>0$.
This implies that
\begin{align*}
&\sum_{m=1}^n\biggl|\int_{\mathcal{W}_m^1} f\circ \mathcal{F}^n d\nu_m^1-\int_{\mathcal{W}_m^2} f\circ \mathcal{F}^n d\nu_m^2\biggr|\\
&\leq \sum_{m=1}^n\biggl|\int_{\mathcal{F}^m\mathcal{W}_m^1} f \circ \mathcal{F}^{n-m} -f\circ\mathcal{F}^{n-m}(\bar x) d\mathcal{F}^{m}_*\nu_m^1-\int_{\mathcal{F}^m\mathcal{W}_m^2} f\circ\mathcal{F}^{n-m} -f\circ\mathcal{F}^{n-m}(\bar x) d\mathcal{F}^{m}_*\nu_m^2\biggl|\\
&\,\,\,+\sum_{m=1}^n\biggl|\int_{\mathcal{F}^m\mathcal{W}_m^1} f\circ\mathcal{F}^{n-m}(\bar x) d\mathcal{F}^{m}_*\nu_m^1-\int_{\mathcal{F}^m\mathcal{W}_m^2} f\circ\mathcal{F}^{n-m}(\bar x) d\mathcal{F}^{m}_*\nu_m^2\biggr|\\
&\leq (I^1_n+I^2_n+I\!I^1_n+I\!I^2_n)+\sum_{m=1}^n\biggl|\int_{\mathcal{F}^m\mathcal{W}_m^1} f\circ\mathcal{F}^{n-m}(\bar x) d\mathcal{F}^{m}_*\nu_m^1(x)-\int_{\mathcal{F}^m\mathcal{W}_m^2} f\circ\mathcal{F}^{n-m}(\bar x) d\mathcal{F}^{m}_*\nu_m^2(x)\biggr|\\
&=I^1_n+I^2_n+I\!I^1_n+I\!I^2_n
\leq C_1\|f\|^-_{C^{\gamma_1}} n^{-{\alpha_0}}.\end{align*}
\end{proof}
Using this lemma we can estimate the following decay rates of correlations.
\begin{lemma}\label{decayinf}
For any two standard families $\mathcal{G}^i=(\mathcal{W}^i, \nu^i)$ with $g^i=d\nu^i/d\mu\in\mathcal{H}^+(\gamma_0)$ for $i=1,2$, for any $f\in \mathcal{H}^-(\gamma_1)$, with $\gamma_1>0$, $\nu^1(\mathcal{M})=\nu^2(\mathcal{M})=1$, we have
\begin{align*}
&\left|\int_{\mathcal{M}} f\circ \mathcal{F}^n d\nu^1 -\int_{\mathcal{M}} f\circ \mathcal{F}^n d\nu^2\right| \leq C\|f\|^-_{C^{\gamma_1}}\max\{\|g^1\|_{\infty},\|g^2\|_{\infty}\} n^{1-\alpha_0},\end{align*}
for any $n\geq N$, where $N=N(\nu^1,\nu^2)\geq 1$, and $C=C(\gamma_1)>0$. \end{lemma}
\begin{proof}
By Lemma \ref{coupling1} for any $n\geq 1$, there exist $N=N(\nu^1,\nu^2)$, and a decomposition
$$\mathcal{F}^N_*\nu^i=\sum_{m=1}^{n}\eta_m^i+\bar\nu^i_n,$$ where $\mathcal{F}^m_*\nu^1_m$ is coupled with $\mathcal{F}^m\nu^2_m$ such that for any $m\geq 1$, for any measurable function $f\in\mathcal{H}^-(\gamma_1)$ that is constant on any stable manifold $W^s\in\Gamma^s$,
\begin{equation}\label{coupling mapf} \mathcal{F}^m_*\nu^1_m(f)=\mathcal{F}^m_*\nu^2_m(f).\end{equation}
Using the above Lemma, we get
\begin{align*}
&\sum_{m=1}^n\biggl|\int_{\hat{\mathcal{W}}_m^1} f\circ \mathcal{F}^n d\eta_m^1-\int_{\hat\mathcal{W}_m^2} f\circ \mathcal{F}^n d\eta_m^2\biggr|\\
& \leq C_1\|f\|^-_{C^{\gamma_1}} n^{-{\alpha_0}},\end{align*}
where $C_1=C_1(\gamma_1)>0$.
Now combining the above facts together with the Coupling Lemma \ref{coupling1}, for both $i=1,2$,
\begin{equation}\label{realbd}\int_{\mathcal{M}} f\circ \mathcal{F}^{n+N} d\nu^1-\int_{\mathcal{M}} f\circ \mathcal{F}^{n+N}\, d\nu^2=\bar\nu^1_n(f\circ \mathcal{F}^n)-\bar\nu^2_n(f\circ \mathcal{F}^n)+\sum_{m=1}^n\bigg(\int_{\hat{\mathcal{W}}_m^1} f\circ \mathcal{F}^n d\nu_m^1-\int_{\hat{\mathcal{W}}_m^2} f\circ \mathcal{F}^n d\nu_m^2\biggr). \end{equation}
Our analysis implies that the second term is of order $n^{-{\alpha_0}}$, thus the decay rate is essentially dominated by $\bar\nu^1_n(f\circ \mathcal{F}^n)$ for general observable $f$.
Note that
\begin{align}\label{suppfg}
|\bar\nu^1_n(f\circ \mathcal{F}^n)|&\leq C_2\|f\|_{\infty}n^{1-\alpha_0},
\end{align}
for some constant $C_2>0$.
Thus we have for $n> 1$,
\begin{align*}
&\left|\int_{\mathcal{M}} f\circ \mathcal{F}^{n+N} d\nu^1-\int_{\mathcal{M}} f\circ \mathcal{F}^{n+N}\, d\nu^2\right|
\\&\leq \sum_{i=1}^{2} C_1\|f\|^-_{C^{\gamma_1}}\|g^i\|_{\infty}
\mu(\mathcal{F}^{-n}{\rm supp}(f)\cap {\rm supp}(g^i)\cap C_{n,b})\\
&\leq 2\|f\|^-_{C^{\gamma_1}}\max\{\|g^1\|_{\infty},\|g^2\|_{\infty}\}n^{1-\alpha_0} ,\end{align*}
where we have used Lemma \ref{coupling1} for the last step, and $C=C(\gamma_1)>0$. This leads to the desired estimate as we have claimed. \end{proof}
Next we consider the case when we do not have standard families, but only H\"older observables.
\begin{lemma}\label{Lemma9}
For any piecewise H\"older continuous functions $f\in\mathcal{H}^-(\gamma_1), g^i\in\mathcal{H}^+(\gamma_2)$, $i=1,2$, with $\mu(g^1)=\mu(g^2)$, $\gamma_i>0$, there exists $N=N(g^1, g^2)\geq 1$, such that for any $n\geq N$,
\begin{align*}
\Delta&(f\circ \mathcal{F}^n,g^1,g^2):=|\mu(f\circ \mathcal{F}^n\cdot g^1)-\mu(f\circ \mathcal{F}^n\cdot g^2)|\\&
\leq C\|f\|_{C_{\gamma_1}}\left(\|g^1\|_{_{C^{\gamma_2}}}(\mu({\rm supp}(g^1)\cap C_{n/2,b})+n^{-\alpha_0})+\|g^2\|_{_{C^{\gamma_2}}}(\mu({\rm supp}(g^2)\cap C_{n/2,b})+n^{-\alpha_0})\right),
\end{align*}
where $C=C(\gamma_1,\gamma_2)>0$.
\end{lemma}
\begin{proof}
We first consider the case when $g^i\geq 0$. Since $\mu$ can be disintegrated on the measurable family of unstable manifolds $\mathcal{W}^u=\{W_{\alpha},\alpha\in \mathcal{A}^u\}$, as a standard family $((W_{\alpha},\mu_{\alpha}),\mathcal{A}^u,\lambda^u)$, such that for any measurable set $A\subset \mathcal{M}$,
$$\mu(A)=\int_{\alpha\in \mathcal{A}^u}\mu_{\alpha}(W_{\alpha}\cap A)\,\lambda^u(d\alpha).$$
Since $g^i\in \mathcal{H}^+(\gamma_2)$ and $\mathcal{F}^{[n/2]}{\rm supp}(g^i)\subset \mathcal{M}$, there exists $\mathcal{A}^i\subset \mathcal{A}^u$, such that $\mathcal{W}=\{W_{\alpha}\,:\,\alpha\in \mathcal{A}^i\}$ is a measurable foliation of $\mathcal{F}^{n/2}{\rm supp}(g^i)$ into unstable manifolds, with factor measure satisfies $\lambda^i(\mathcal{A}^i)=\mu({\rm supp}(g^i))$, and for any measurable set $A\subset M$, we define a probability measure $\nu^i$ such that
$$\nu^i(A):=\lambda^i(\mathcal{A}^i)^{-1}\int_{\alpha\in \mathcal{A}^i}\int_{W_{\alpha}\cap A} g^i_{\alpha}\,d\mu_{\alpha}\,\lambda^i(d\alpha),$$
with $g^i_{\alpha}=g^i\circ \mathcal{F}^{-n/2}/\bar g_{\alpha}^i$ and $d\lambda^i(\alpha)=\bar g_{\alpha}^i\,d\lambda^u(\alpha)$.
Here
$$\bar g_{\alpha}^i=\int_{W_{\alpha}} g^i\circ\mathcal{F}^{-n/2}\,d\mu_{\alpha}.$$
We know that
\begin{align}\label{I1g}
\int_{\mathcal{A}^i}&\int_{W_{\alpha}\cap \mathcal{F}^{n/2}C_{n/2,b}} f\circ\mathcal{F}^{n/2}\cdot g^i\circ \mathcal{F}^{-n/2}\,d\mu_{\alpha}\,\lambda^u(d\alpha)\nonumber\\&\leq \|f\|_{\infty}\|g^i\|_{\infty}\mu
\left(\mathcal{F}^{-n/2}{\rm supp}(f)\cap {\rm supp}(g^i)\cap C_{n/2,b}\right).
\end{align}
On the other hand, if $W_{\alpha}$ belongs to $\mathcal{F}^{n/2}C_{n/2,b}^c$, then $\mathcal{F}^{-n/2}W_{\alpha}$ returns to $M$ at least $(b\ln (n/2))^2>b\ln n$ times.
We define $\bar g^i(x)= \bar g^i_{\alpha}$, for any $x\in W_{\alpha}$, any $\alpha\in \mathcal{A}$, then $\bar g^i\in \mathcal{H}^+(1)$. By the H\"older continuity of $g^i$, we have
$$\sup_{\alpha\in \mathcal{A}^u}\sup_{x\in W_{\alpha}\cap \mathcal{F}^{n/2}C_{n/2,b}^c} |g^i\circ \mathcal{F}^{-n/2} -\bar g^i_{\alpha}|\leq C(\gamma_2) \|g^i\|_{{\gamma_2}}\vartheta^{\gamma_2 (b\ln \frac{n}{2})^2},$$
where $C(\gamma_2)>0$ is a constant.
Thus
\begin{align*
\Delta(f\circ \mathcal{F}^n,g^1,g^2)&=\Delta(f\circ \mathcal{F}^{n/2},\bar g^1,\bar g^2)+\mu\left(f\circ \mathcal{F}^{n/2}\cdot \left((g^1\circ\mathcal{F}^{-n/2}-\bar g^1)-(g^2\circ\mathcal{F}^{-n/2}-\bar g^2)\right)\right)\nonumber\\
&\leq \Delta(f\circ \mathcal{F}^{n/2},\bar g^1,\bar g^2) +C_1\|f\|_{_{C^{\gamma_1}}}(\|g^1\|_{{\gamma^2}}+\|g^2\|_{\gamma^2}) \vartheta^{\gamma_2 (b\ln \frac{n}{2})^2} \nonumber\\
&+C_2\|f\|_{_{C^{\gamma_1}}}\|g^1\|_{_{C^{\gamma_2}}} \mu(\mathcal{F}^{-n/2}{\rm supp}(f)\cap {\rm supp}(g^1)\cap C_{n/2,b})\\
&+C_3\|f\|_{_{C^{\gamma_1}}}\|g^2\|_{_{C^{\gamma_2}}} \mu(\mathcal{F}^{-n/2}{\rm supp}(f)\cap {\rm supp}(g^2)\cap C_{n/2,b}).
\end{align*}
We let $\mathcal{G}=(\mathcal{W}^i, \nu^i)$, such that $d\nu^i=\bar g^i/\mu(g^i) d\mu$. Then by Lemma \ref{decayinf}, there exists $N=N(g^1,g^2)\geq 1$ such that
for any $f\in \mathcal{H}^-(\gamma_1)$, any $n/2>N$:
\begin{align*}
\biggl|\nu^1(f\circ \mathcal{F}^{n/2}) -\nu^2(f\circ \mathcal{F}^{n/2}) \biggr|
&\leq C_1(\gamma_1)\|f\|_{_{C^{\gamma_1}}}\|g^1\|_{_{C^{\gamma_2}}}\mu({\rm supp}(g^1)\cap C_{n/2,b})\\
&+C_2(\gamma_1)\|f\|_{_{C^{\gamma_1}}}\|g^2\|_{_{C^{\gamma_2}}}\mu({\rm supp}(g^2)\cap C_{n/2,b}\cap \mathcal{F}^{-n/2}{\rm supp}(f)),\end{align*}
where $C_1(\gamma_2)>0$ and $C_2(\gamma_2)>0$ are constants.
Combining the above estimates, we get
that for $g^i\geq 0$,
\begin{align*}
\Delta(f\circ \mathcal{F}^n, g^1, g^2)&\leq C_1\|f\|_{_{C^{\gamma_1}}}\|g^1\|_{_{C^{\gamma_2}}} \left(\mu({\rm supp}(g^1)\cap C_{n/2,b})+n^{-\alpha_0}\right)\\
&+C_2\|f\|_{_{C^{\gamma_1}}}\|g^2\|_{_{C^{\gamma_2}}} \left(\mu({\rm supp}(g^2)\cap C_{n/2,b})+n^{-\alpha_0}\right),
\end{align*} where $C_1=C_1(\gamma_1,\gamma_2)>0$ and $C_2=C_2(\gamma_1,\gamma_2)>0$ are constants.
For the general case when $g^i$ is not nonnegative, we decompose $g^i=g^i_+-g^i_-$ into its positive and negative parts. Since $g^i_{\pm}\in \mathcal{H}^+(\gamma_2)$, a similar statement can be proved.
\end{proof}
Note that Theorem 1 directly follows from the above lemma.
\section{Proof of Theorem \ref{TmMain} and Theorem 3.}
\label{secPMT}
\subsection{Proof of Theorem 2}
By (\ref{cMMm}), we denote $$B_n=\bigcup_{m>n}\bigcup_{k=1}^{m-n}\mathcal{F}^k M_m=(R>n)\setminus M,$$ as the set of points in $M^c$ that take at least $n$-iterations under $\mathcal{F}$ before they come back to $M$. We first prove the following lemma that will be used to prove Theorem \ref{TmMain}.
\begin{lemma}\label{defnBn} For any large $n$, we define
$\mu^n=\frac{\mu|_{B_n^c}}{\mu(B_n^c)}$. Then for any probability measure $\nu$ with support contained in $M$, we have for any bounded function $f$ on $M$,
\begin{equation}\label{decay11}\mathcal{F}^n_* \nu(f)-\mu(f)-\mu(f)\mu(R>n)=\mathcal{F}^n_*\nu(f)-\mathcal{F}^n_* \mu^n(f)+ \mathcal{O}(n^{-\beta}),\end{equation}
with $\beta=\min\{{\alpha_0}, 2{\alpha_0}-2\}$.
\end{lemma}
\begin{proof}
First note that $$B_n=\bigcup_{m>n}\bigcup_{k=1}^{m-n}\mathcal{F}^k M_m=(R>n)\setminus (\cup_{m\geq n}M_m),$$ which implies that \begin{equation}\label{BnRn}\mu(B_n)=\mu(R>n)+\mathcal{O}(n^{-{\alpha_0}}).\end{equation}
Since the support of the initial measure $\nu$ is contained in $M$, we have that after
$n$ iterations the push forward measure $\mathcal{F}^n_*\nu$ can never reach the
region $\mathcal{F}^n B_n$. Thus we can ignore the measure $\mu$ restricted on $B_n$ within first $n$-iterations. This fact implies that for
each $n\geq 1$ the measure $\mu$ is a linear combination of two
probability measures, $$\mu=\mu(B_n)\cdot\frac{\mu|_{B_n}}{\mu(B_n)}+\mu(B_n^c)\cdot\frac{\mu|_{B_n^c}}{\mu(B_n^c)}.$$
We denote
$\mu^n=\frac{\mu|_{B_n^c}}{\mu(B_n^c)}$. Note that for
any bounded function $f$ supported on $M$, we have $$\mu(f)=\mu^n
(f)\mu(B_n^c)=\mu^n(f)-\mu^n(f)\mu(B_n).$$
Notice also that (\ref{BnRn}) implies that $\mu(B_n)= O( n^{1-{\alpha_0}})$. Thus we have
\begin{align*}
\mu^n(f)=\frac{\mu (f\cdot \mathbf{I}_{B^c_n})}{\mu(B_n^c)}=\frac{\mu(f)}{1-\mu(B_n)}=\mu(f)(1+ \mu(B_n)+O(\mu(B_n)^2)).
\end{align*}
We apply $\mathcal{F}^n_*$ on both sides of the equation, and using the fact that $\mathcal{F}^n_*\mu=\mu$ to get
$$\mathcal{F}^n_* \mu^n(f)=\mu(f)(1+ \mu(B_n)+O(\mu(B_n)^2)).$$
This also implies that
$$\mu(f)(1+\mu(R>n))=\mathcal{F}^n_* \mu^n(f)+\mu(f)\mathcal{O}(n^{-\beta_0})$$
where $\beta_0=\min\{\alpha_0, 2\alpha_0-2\}$.
Here we have used the fact given by (\ref{BnRn}).
Combining above facts, we have
\begin{align*}\mathcal{F}^n_* \nu(f)-\mu(f)-\mu(f)\mu(R>n)&=\mathcal{F}^n_*\nu(f)-\mathcal{F}^n_* \mu^n(f)+\mu(f)\mathcal{O}(n^{-\beta_0}).\end{align*}
\end{proof}
To prove Theorem \ref{TmMain}, we first assume the observable $g$ defines a probability measure on $M$, with $d\nu=g d\mu$. By Lemma \ref{defnBn}, we have
\begin{equation}\label{decay101}\mathcal{F}^n_* \nu(f)-\mu(f)-\mu(f)\mu(R>n)=\mathcal{F}^n_*\nu(f)-\mathcal{F}^n_* \mu^n(f)+O(n^{-\beta_0}).\end{equation}
Our goal is to show that for systems satisfy (\bH2) the correlation $\mathcal{F}^n_*\nu(f)-\mathcal{F}^n_* \mu^n(f)=o(\mu(R>n))$.
We use assumption \textbf{(H2)}, which states that $$\mu(C_{n,b}\cap \mathcal{F}^{-n} M)=\mathcal{O}(n^{-{\alpha_0}}).$$
Next we start to prove Theorem 2. Let $\gamma_1>0$. We consider any $f\in \mathcal{H}^-(\gamma_1)$ supported on $M$, and
any proper family $\mathcal{G}=(\mathcal{W}, \nu)$ with $g=d\nu/d\mu\in\mathcal{H}^+(\gamma_0)$ supported on $M$. For any large $n$, we define $\mathcal{G}^1=\mathcal{G}$ and $\mathcal{G}^2=(\mathcal{W}^u,\mu^n)$, then they are both proper families. We denote $\nu^1=\nu$ and $\nu^2=\mu^n$.
First note that since both measures are essentially supported on $(R\leq n)$, thus we have
$$\nu^i(R>n)=\mathcal{O}(n^{-\alpha_0}).$$
According to Lemma \ref{decayinflemma} we know that for any $k=1,\cdots, n$,
\begin{align*}
&\left|\int_{\mathcal{M}} f\circ \mathcal{F}^k d\nu^1 -\int_{\mathcal{M}} f\circ \mathcal{F}^k d \nu^2\right|\\&=\left|\bar\nu^1_k(f\circ \mathcal{F}^k)-\bar\nu^2_k(f\circ \mathcal{F}^k)+\sum_{m=1}^k\bigg(\int_{\mathcal{W}_m^1} f\circ \mathcal{F}^k d\nu_m^1-\int_{\mathcal{W}_m^2} f\circ \mathcal{F}^k d\nu_m^2\biggr)\right|\\
&\leq \left|\bar\nu^1_k(f\circ \mathcal{F}^k)-\bar\nu^2_k(f\circ \mathcal{F}^k)\right|+C(\gamma_1)\|f\|_{\gamma_1}\|g\|_{\infty} k^{-{\alpha_0}}.\end{align*}
According to the Coupling Lemma \ref{coupling1} and (\ref{suppfg}), we know that the uncoupled measure $\bar\nu^i_k(\mathcal{M})$ is dominated by $C_{k,b}$, while $\bar\nu^i_k(C_{k,b}^c\cap(R\leq k))=\mathcal{O}( k^{-{\alpha_0}}).$ Thus
\begin{align*}
\bar\nu^i_k(f\circ \mathcal{F}^k)&=
\bar\nu^i_k(f\circ \mathcal{F}^k\cdot \mathbf{I}_{C_{k,b}})+
\bar\nu^i_k(f\circ \mathcal{F}^k\cdot \mathbf{I}_{R>n})+\bar\nu^i_k(f\circ \mathcal{F}^k\cdot \mathbf{I}_{C_{k,b}^c})= \bar\nu^1_k(f\circ \mathcal{F}^k\cdot \mathbf{I}_{C_{k,b}})+\mathcal{O}( k^{-{\alpha_0}})\nonumber\\
& \leq \|f\|_{\infty}\|g^i\|_{\infty}\left(\mu(C_{k,b}\cap \mathcal{F}^{-k}{\rm supp}(f)\cap {\rm supp}(g^i))\right)+\mathcal{O}( k^{-{\alpha_0}})\\
&=\|f\|_{\infty}\|g^i\|_{\infty}\left(\mu(C_{k,b}\cap \mathcal{F}^{-k}M\cap M)\right)+\mathcal{O}( k^{-{\alpha_0}})=\mathcal{O}(k^{-{\alpha_0}}),
\end{align*}
where we have used Assumption (\textbf{H2}) in the last estimate.
Combining above facts, we take $k=n$, then
$$|\int_{\mathcal{M}} f\circ \mathcal{F}^n d\nu^1 -\int_{\mathcal{M}} f\circ \mathcal{F}^n d\nu^2|\leq C_1(\gamma_1)\|f\|^-_{C^{\gamma_1}}\|g\|_{\infty} n^{-{\alpha_0}}.$$
One can check that (\textbf{H1}) implies that $\mu(R>n|R>n)=1+\mathcal{O}(n^{-1})$. Combining this with (\ref{decay101}), then for any $n>1$, we get
\begin{align*}
&|\mathcal{F}^n_* \nu(f)-\mu(f\circ \mathcal{F}^n)-\mu(f)\mu(B_n)|\leq |\mathcal{F}^n_*\nu(f)-\mu^n(f\circ \mathcal{F}^n)|+ C_1(\gamma_1)\|f\|^-_{C^{\gamma_1}}\|g\|_{\infty}n^{-\beta_0}\\
&\leq C(\gamma_1) \|f\|^-_{C^{\gamma_1}}\|g\|_{\infty}n^{-{\alpha_0}}+ C_1(\gamma_1)\|f\|^-_{C^{\gamma_1}}\|g\|_{\infty}n^{-\beta_0},
\end{align*}
here $\beta_0=\min\{2{\alpha_0}-2,{\alpha_0}\}$.
For general H\"older observable $g$, we denote $g=g^+-g^-$. It is enough to consider the case when $\mu(g^{\pm})\neq 0$, then we consider $g^+$ and $g^-$ separately as in the proof of Theorem 1, to get (\ref{mainh2a}) in Theorem 2.
\subsection{Prove of Theorem 3.}
One interesting application of Theorem 2 is that one gets classical Central limiting theorem for stochastic processes generated by certain observables for dynamical systems with decay rates of order $\mathcal{O}(1/n)$.
We consider an observable $f\in H^-(\gamma)\cap H^+(\gamma)$ with ${\rm supp}(f)\subset M$ supported in $M$, and $\gamma>0$. We assume $f$ is not a coboundary. For any $n\geq 1$, we consider the two partial sums $\tilde{S}_n(f):=f+f\circ \mathcal{F}+\cdots+f\circ \mathcal{F}^n$ and $S_n(f)=f+f\circ F+\cdots+ f\circ F^n$. We assume $\mu(f)=0$.
For the induced map $(F,M,\mu_M)$, it follows from Theorem 7.52 in \cite{CM} and \cite{CZZ}, that conditions \textbf{(h1)-(h4)} implies that
\begin{equation}\label{CLT}\frac{S_n}{\sigma\sqrt{n}}\to N(0,1),\end{equation} in distribution, where $N(0,1)$ is the standard normal variable, and by the Green-Kubo formula,
\begin{equation}\label{sigmaF}\sigma^2=\mu_M(f^2)+2\sum_{n=1}^{\infty}\mu_M(f\circ F^n\cdot f).\end{equation}
Note that $$\int f\circ F^n\cdot f\,\,d\mu_M\leq C\|f\|_{C^{\gamma}} \vartheta^n,$$ as the induced map enjoys exponential decay of correlations.
In \cite{BCD}, the partial sum $S_n$ was associated with the so-called induced observable, $\tilde{S}_n{f}(x)=\sum_{m=0}^{R(x)-1} f(\mathcal{F}^m x)$. However, since our observable $f$ has support contained in $M$, thus $\tilde{S}_n(f)=S_n(f)$ coincide with the induced observable.
Next we review the relation between CLT of $S_n$ and $\tilde{S}_n$, see a detailed proof in \cite{BCD}.
\begin{lemma}\label{MCLXn4} For any $n\geq 1$, if $S_n$ satisfies the CLT (\ref{CLT}), then $\tilde{S}_n$ also satisfy a CLT:
\begin{equation}\label{ClTZ}\frac{\tilde{S}_n}{\tilde{\sigma} \sqrt{n}}\to N(0,1),\end{equation}
in distribution.
Here $\tilde{\sigma}^2 =\sigma^2\mu(M)$.
\end{lemma}
Thus (\ref{ClTZ}) implies that $\tilde{S}_n$ satisfies the classical CLT with variance $\sigma\sqrt{\mu(M)}$. Moreover, again by the Green-Kubo formula we get
\begin{align*}
\sigma^2\mu(M)&=\mu(f^2)+2\sum_{n=1}^{\infty}\mu(f\circ \mathcal{F}^n\cdot f).
\end{align*}
Comparing with (\ref{sigmaF}), and use the definition $d\mu_M=d\mu/\mu(M)$, we get the interesting relation
$$\sum_{n=1}^{\infty}\mu(f\circ \mathcal{F}^n\cdot f)=\sum_{n=1}^{\infty}\mu(f\circ F^n\cdot f).$$
This finishes the proof of Theorem 3.\\
\section{Sufficient conditions for (\textbf{H2}).}
In this section, we introduce two sufficient conditions to guarantee assumption (\textbf{H2}).\\
\noindent\textbf{Condition \textbf{(H2)(a)} }. We assume there exist $ \zeta\in(0,1), \eta_0 \in(0,1)$, $C, C_1>0$, $N>1$, such that for any sufficiently large $n>N$, there exist $ \varepsilon_n\in(0,1), \gamma_n \in(0,1)$, and for each $m=1, \ldots, (b\ln n)^2$,
$$\mu\bigl(\{x\in M\colon R(F^m(x))>\varepsilon_n^m n\}|M_n\bigr)<C \eta_0^m +C_1 n^{-\zeta}.$$
\\
\noindent\textbf{Condition \textbf{(H2)(b)}}. We assume there exist $q \in(0,1)$, $C>0, p>0$, and $N>1$ such that for any sufficiently large $n>N$, and for each $m=1, \ldots, (b\ln n)^2$:
$$\mu\bigl(\{x\in M\colon R(F^m(x))> n^{1-q}\}|M_n\bigr)<C n^{-p}.$$
We will prove the following two lemmas.
\begin{lemma}\label{tmH3H4} Condition (\bH2)(a) implies that there exists $\varepsilon_1\in (0,1)$ such that \textbf{(H2)} holds.
\end{lemma}
\begin{proof}
For any sufficiently large $n$, any $x \in C_{n,b}$, we define $$k_0=k_0(x)=\sum_{m=0}^{n-1}\mathbf{I}_{M}(\mathcal{F}^m(x))$$ as the number of returns to $M$ within $n$ iterations under $\mathcal{F}$ along the forward trajectory of $x$. By the definition of $C_{n,b}\subset (R<n)$, we know that any $x \in C_{n,b}$ must returns to $M$ at least once, with $$2\leq k_0<(b\ln n)^2.$$ Let $x_n\in M$ be the last enter to $M$ within $n$ iterations of $x$, and we define
$$m_{1}(x):=\max_{1\leq k\leq k_0}\{R(F^{-k}(x_n))\},$$ to be the largest
return time function value of $R$ along $n$ iterations of $x$ under $\mathcal{F}$. i.e.\ there exists $k_1=k_1(x)\in\{1,\cdots, n\}$, such that $$x_1:=\mathcal{F}^{-k_1}x_n\in M_{m_1}.$$ Moreover we define indices: $$m_1(x)=R(x_1), m_2(x):=R(F(x_1)), \cdots, m_{k_2}(x):=R(F^{k_2-1}(x_1)),$$ with $F^{k_2-1}(x_1)=F^{-1}x_n$ being the second last return to $M$ along $n$ iterations of $x$.
Without loss of generality, we assume $$\sum_{k=1}^{k_2} m_k\geq n/2,$$ i.e. the largest index $m_1$ occurs within the first $n/2$ iterations of $x$ under $\mathcal{F}$. Then by assumption, since $m_1(x)$ is the largest index within $k_0$ returns to $M$ along forward $n$ iterations of $x$, we have
$$
n/2\leq m_1+\cdots +m_{k_2}\leq n.
$$
Since points $x\in C_{n,b}$ return to $M$ at most $\psi$ times during
the first $n$ iterates of $\mathcal{F}$, there are $\leq \psi$ number of
intervals between successive returns to $M$, and hence the longest
interval has length $m_1(x)\geq n/\psi$. Let $c_0<1/2$ be a constant. We split $ C_{n,b}=C_{n,b}'\cup C_{n,b}''$ into two disjoint parts.\\
(I) $C_{n,b}'$ is a `good part' of
$C_{n,b}$ such that for any $y\in C_{n,b}'$,
\begin{equation}\label{epsk}
m_k(y)<\varepsilon_n^k m_1(y),
\end{equation}
for all
$$
\frac{c_0n}{m_1(y)} \leq k \leq k_0\leq \psi(n)=(b\ln n)^2,
$$where $\varepsilon_n\in (0,1)$ was given in (\textbf{H2(a)}). More precisely, for $y\in C'_{n,b}$, there is a sequence of returns to $M_{m_k}$'s, with index decreasing exponentially in $k$, within $n$-iterations.
For `good' points $y\in C_{n,b}'$ we have
$$
\frac{n}{2} \leq \sum_{k=1}^{\psi(n)} m_k \leq
m_1\cdot \frac{c_0n}{m_1} + m_1\sum_{k=\frac{c_0n}{m_1}}^{\psi(n)}\varepsilon_n^k \leq c_0 n+m_1\varepsilon_n^{\frac{c_0 n}{m_1}} \leq c_0 n+m_1.
$$
Since $c_0<1/2$,
we conclude $m_1 \geq cn$ for a positive constant $c:=\frac{1}{2}-c_0>0$. This implies that for points in $C_{n,b}'$, the largest index $m_1$ within $n$ iterations must be approximately of order $n$.
Accordingly, $$ C'_{n,b}\cap\mathcal{F}^{-n}M\subset \bigcup_{m_1\geq cn} \bigcup_{k=1}^{\psi(n)} \mathcal{F}^{-n}(F^{k}M_{m_1}).$$ Note that for any $m_1$, for any $1\leq k_1<k_2\leq \psi(n)$, the sets $\mathcal{F}^{-n}(F^{k_1}M_{m_1})\cap C'_{n,b}$ and $\mathcal{F}^{-n}(F^{k_2}M_{m_1})\cap C'_{n,b}$ are disjoint. Thus
the measure of good points in $C_{n,b}$ is bounded by
$$\mu(C_{n,b}'\cap M)\leq \sum_{m_1\geq cn}\mu(M_{m_1})\leq C n^{-\alpha_0}
\text{ and }
\mu(C_{n,b}')\leq \sum_{m_1\geq cn}m_1\mu(M_{m_1})\leq C n^{1-\alpha_0}.$$
(II) On the other hand $C_{n,b}''$ consists of `bad' points, such that (\ref{epsk}) fails:
\begin{equation}\label{epsk1}
m_k(y)>\varepsilon_n^k m_1(y),
\end{equation}
for some
$
\frac{c_0n}{m_1(y)} \leq k \leq b_0\ln n.
$
We divide $C_{n,b}''$ according to the maximal index $m_1$ of its points $C_{n,b}''=\bigcup_{m_1>\frac{n}{\psi}} C_{n,b,m_1}$
such that $C_{n,b,m_1}$ contains all points in $C_{n,b}''$ with the largest return time $m_1$ within $n$ iterations.
The contribution of $M_{m_1}$ to these `bad' points in $C_{n,b}''$ will be estimated according to (\bH2)(\textbf{a}) as following:
$$
\mu(C_{n,b, m_1}) \leq \mu(M_{m_1})\sum_{k=\tfrac{c_0n}{m_1}}^{\psi}
(\eta_0^k+ O(m_1^{-\zeta})).
$$
By assumption $m_1\geq n/\psi$, so the total measure of
$C_{n,b}''$ is
\begin{align}\label{estC''nb}
\mu(C_{n,b}''\cap M)&\leq\sum_{m_1=n/\psi}^{n}
\mu(M_{m_1})\sum_{k=\tfrac{c_0n}{m_1}}^{\psi} (\eta_0^k+ O(m_1^{-\zeta}))\nonumber\\
&\leq \sum_{k=1}^{\psi}\left( \eta_0^k
\sum_{m_1=\tfrac{c_0n}{k}}^{\infty}\mu(M_{m_1})+ \sum_{m_1=\tfrac{c_0n}{k}}^{\infty}\mu(M_{m_1}) m_1^{-\zeta}\right)\nonumber\\
&\leq \sum_{k=1}^{\psi}
\left(\frac{k}{c_0n}\right)^{\alpha_0}\, \eta_0^k+C \frac{\psi^{\alpha_0+1-\zeta}}{n^{\alpha_0+\zeta}}
\nonumber\\
&\leq C n^{-\alpha_0}\sum_{k=1}^{\psi}
k^{\alpha_0}\,\eta_0^k+C \frac{\psi^{\alpha_0+1-\zeta}}{n^{\alpha_0+\zeta}}
\leq C_1 n^{-\alpha_0},
\end{align}
where we have used the assumption on $\psi=(b\ln n)^2$.
Combining the above estimates we get
\begin{align*} \mu(C_{n,b}\cap M)= O(n^{-\alpha_0}),\,\,\,\,\, \mu(C_{n,b})= O(n^{1-\alpha_0}).
\end{align*}
\end{proof}
\begin{lemma}\label{CnbBnc} Under assumption \textbf{(H2)(b)}, \textbf{(H2)} holds.
\end{lemma}
\begin{proof}
For any sufficiently large $n$, any $x \in C_{n,b}$, we define $$k_0=k_0(x)=\sum_{m=0}^{n-1}\mathbf{I}_{M}(\mathcal{F}^m(x)),$$ as the number of returns to $M$ within $n$ iterations under $\mathcal{F}$ along the forward trajectory of $x$. By the definition of $C_{n,b}\subset (R\leq n)$, we know that any $x \in C_{n,b}$ must return to $M$ at least once, with $1\leq k_0<(b\ln n)^2$. Let $x_n\in M$ be the last entrance to $M$ within $n$-iterations, and we define
$$m_{1}(x):=\max_{1\leq k\leq k_0-1}\{R(F^{-k}(x_n))\}$$ to be the largest
return time function value of $R$ along $n$ iterations of $x$ under $\mathcal{F}$. i.e.\ there exists $k_1=k_1(x)\in\{1,\cdots, n\}$, such that $x_1:=\mathcal{F}^{-k_1}x_n\in M_{m_1}$, and $k_1+m_1\leq n$. Moreover we define indices: $$m_1(x)=R(x_1), m_2(x):=R(F(x_1)), \cdots, m_{k_2}(x):=R(F^{k_2-1}(x_1)),$$ with $F^{k_2-1}(x_1)=F^{-1}x_n$ being the second last return to $M$ along $n$ iterations of $x$.
Without loss of generality, we assume $$\sum_{k=1}^{k_2} m_k\geq n/2.$$ i.e. the largest index $m_1$ occurs within the first $n/2$ iterations of $x$ under $\mathcal{F}$. Then by assumption, since $m_1(x)$ is the largest index within $k_0$ returns to $M$ along forward $n$ iterations of $x$, we have
$$
n/2\leq m_1+\cdots +m_{k_2}\leq n.
$$
Since points $x\in C_{n,b}$ return to $M$ at most $\psi$ times during
the first $n$ iterates of $\mathcal{F}$, there are $\leq \psi$ number of
intervals between successive returns to $M$, and hence the longest
interval has length $m_1(x)\geq n/\psi$. Let $c_0<1/2$ be a constant. We split $ C_{n,b}=C_{n,b}'\cup C_{n,b}''$ into two disjoint parts.\\
(I) $C_{n,b}'$ is a `good part' of
$C_{n,b}\cap \mathcal{F}^{-n} M$ such that for any $y\in C_{n,b}'$,
\begin{equation}\label{epsk2}
m_k(y)<m_1(y)^{1-q},
\end{equation}
for all
$$
\frac{c_0n}{m_1(y)} \leq k \leq k_0\leq \psi(n)=(b\ln n)^2.
$$ More precisely, for $y\in C'_{n,b}$, there is a sequence of returns to $M_{m_k}$'s, with index decreasing exponentially in $k$, within $n$-iterations.
For `good' points $y\in C_{n,b}'$ we have
$$
\frac{n}{2} \leq \sum_{k=1}^{\psi(n)} m_k \leq
m_1\cdot \frac{c_0n}{m_1} + \sum_{k=\frac{c_0n}{m_1}}^{\psi(n)}m_1^{1-q} \leq c_0 n+m_1\leq c_0 n+m_1.
$$
Since $c_0<1/2$,
we conclude $m_1 \geq cn$ for a positive constant $c:=\frac{1}{2}-c_0>0$. This implies that for points in $C_{n,b}'$, the largest index $m_1$ within $n$ iterations must be approximately of order $n$.
Accordingly, the measure of good points in $C_{n,b}$ is bounded by
\begin{equation}\label{Cnb'H2b}\mu(C'_{n,b}\cap M)\leq \sum_{m_1\geq cn}\mu(M_{m_1})\leq C n^{-{\alpha_0}}.\end{equation}
And
\begin{equation}\label{Cnb'H2bM}\mu(C'_{n,b})\leq m \sum_{m_1\geq cn}\mu(M_{m_1})\leq C n^{1-{\alpha_0}}.\end{equation}
(II) On the other hand $C_{n,b}''$ consists of `bad' points $y\in C_{n,b}$, such that (\ref{epsk2}) fails. i.e.,
\begin{equation}\label{epsk3}
m_k(y)>m_1(y)^{1-q},
\end{equation}
for some
$$
\frac{c_0n}{m_1(y)} \leq k \leq \psi(n).
$$
We divide $C_{n,b}''$ according to the maximal index $m_1$ of its points $C_{n,b}''=\bigcup_{m_1>\frac{n}{\psi}} C_{n,b,m_1}$
such that $C_{n,b,m_1}$ contains all points in $C_{n,b}''$ with the largest return time $m_1$ within $n$ iterations.
The contribution of $M_{m_1}$ to these `bad' points in $C_{n,b}''$ will be estimated according to (\bH2)(\textbf{b}) as following:
$$
\mu(C_{n,b, m_1}) \leq \mu(M_{m_1})\sum_{k=\tfrac{c_0n}{m_1}}^{\psi}
m_1^{-p}.
$$
By assumption $m_1\geq n/\psi$, so the total measure of
$C_{n,b}''$ can be estimated as in (\ref{estC''nb}): $
\mu(C_{n,b}''\cap M)\leq C_1 n^{-\alpha_0}.
$ This implies $\mu(C_{n,b}\cap M)= O(n^{-\alpha_0})
$.
\end{proof}
\section{Applications to hyperbolic systems}
To illustrate our method, we apply it to several classes of dynamical systems.
Since the induced maps for most of these examples were studied in \cite{CZ09},
we only remind some basic facts here.
First we recall standard definitions, see \cite{BSC90,BSC91,C99}. A
2-D flat billiard is a dynamical system where a point moves freely at
unit speed in a domain $Q\subset \mathbb{R}^2$ and bounces off its
boundary $\partial Q$ by the laws of elastic reflection. We assume that
$\partial Q=\cup_{i}\Gamma_i$ is a finite union of piecewise smooth curves,
such that each smooth component $\Gamma_i\subset\partial Q$ is either convex
inward (dispersing), or flat, or convex outward (focusing).
Following Bunimovich, see \cite{Bu74,Bu79} and \cite[Chapter~8]{CM}, we
assume that every focusing component $\Gamma_i$ is an arc of a circle
such that there are no points of $\partial Q$ on that circle or inside it,
other than the arc $\Gamma_i$ itself. Under these assumptions the
billiard dynamics is hyperbolic.
Let $ \mathcal{M}=\partial Q\times [-\pi/2,\pi/2]$ be the {\em collision space}, which
is a standard cross-section of the billiard flow. Canonical coordinates
in $\mathcal{M}$ are $r$ and $\varphi$, where $r$ is the arc length parameter
on $\partial Q$ and $\varphi\in [-\pi/2,\pi/2]$ is the angle of reflection.
The collision map $\mathcal{F} \colon \mathcal{M} \to \mathcal{M}$ takes an inward unit vector
at $\partial Q$ to the unit vector after the next collision, and
preserves smooth measure $d\hat{\mu} = c\cdot \cos \varphi\, dr\, d\varphi$
on $\mathcal{M}$, where $c$ is a normalization constant. Furthermore, $\partial
\mathcal{M}\cup\mathcal{F}^{-1}(\partial \mathcal{M})$ is the singular set of $\mathcal{F}$.
For billiards with focusing boundary components, the expansion and
contraction (per collision) may be weak during long series of
successive reflections along certain trajectories. To study the
mixing rates, one needs to find and remove the spots in the phase
space where expansion (contraction) slows down. Such spots come in
several types and are easy to identify, for example, see \cite{CZ05a}
and \cite[Chapter~8]{CM}. Traditionally, we denote
\begin{equation*}
\partial Q=\partial^0 Q \cup \partial^{\pm} Q,
\end{equation*}
where $\partial Q^0$ is the union of flat boundaries, $\partial Q^-$
contains focusing boundaries and $\partial Q^+$ contains dispersing
boundaries. The collision space can be naturally divided into
focusing, dispersing and neutral parts:
\begin{equation*}
\mathcal{M}_{0}=\{(r, \varphi)\colon r\in \partial^0 Q\},
\quad\quad
\mathcal{M}_{\pm}=\{(r, \varphi)\colon r\in \partial^{\pm} Q\}.
\end{equation*}
Now we define the induced phase space:
\begin{equation} \label{M3}
M= \{x\in \mathcal{M}_-\colon\, \pi(x)\in \Gamma_i,\,
\pi(\mathcal{F}^{-1}x)\in \Gamma_j, \,j\neq i\}\cup \mathcal{M}_+.
\end{equation}
Note that $M$ only contains all collisions on dispersing boundaries and
the \emph{first} collisions with each focusing arc (the collisions with
straight lines are skipped altogether). The map $F$ preserves the
measure $\mu$ conditioned on $M$, which we denote by $ \mu =
[\hat{\mu}(M)]^{-1} \hat{\mu}$.
Furthermore, $F$ has a larger singular set than
the original map. Let $ S_0=\partial M$, then $ S_1:= S_0\cup
F^{-1} S_0$ is the singular set of $F$.
Let $R\colon \mathcal{M}\to \mathbb{N}$ be the first return time function, such
that for any $x\in \mathcal{M}$, $\mathcal{F}^{R(x)}x$ returns to $M$ for the first time. We define $M_m=R^{-1}\{m\}\cap M$
as the level set of the return time function.
To be more specific, we consider billiard systems that have been studied
in \cite{CZ05a,CZ05b,CZ07,CZ08,CM07}, which include semi-dispersing
billiards on a rectangle, Bunimovich billiard, Bunimovich Stadia, skewed
stadia, billiards with flat points, billiards with cusps. It was proved in
these references that the induced system $(F,M,\mu)$ satisfies the
condition \textbf{(H1)} and \textbf{ (H3)} and enjoys exponential decay of
correlations. It is enough to check (\bH2). We first introduce some new conditions that imply \textbf{(H2)} and which are easier to check.\\
\subsection{Billiards with property \textbf{(H2)(a)}.}
\noindent\textbf{New condition (A1).} Assume for any $n$ large, there exist $\hat M_n\subset M_n$, $\zeta_1\in (0,1)$, such that $$\mu(M_n\setminus \hat M_n| M_n)<C /n^{\zeta_1},$$ and for any $x\in \hat M_n$, there exist $d\in (0,1)$, $C>0$ and a large $b>0$, such that for $m=1, \cdots, (b\ln n)^2$, we have the following condition:
$$\mathbb{E}\left[R(F^m(x))|x\in\hat M_n\right]\leq d^m n +\mathcal{O}(n^{-1-\zeta_1}).$$
\begin{proposition} \label{prh1h2H3} Condition (\textbf{A1}) implies \textbf{(H2)(a)}. \end{proposition}
\begin{proof} Let $\psi(n)=(b\ln n)^2$.
For any $n$ large, any $k=1,\cdots, \psi(n)$, consider the quotient
$$\xi_{n,k}(x)\colon=R(F^kx)/R(F^{k-1}x),$$ for any $x\in \hat M_n$.
Let $$S_{n,k}= \ln \xi_{n,k}(x)+\cdots+\ln \xi_{n,1}(x).$$
Then the moment generating function of $S_k$ at $1$ satisfies:
$$\mathbb{M}_k(1):=\mu_n(e^{S_{n,k}})=\mu_n(\prod_{i=1}^k \xi_{n,i})\leq d^k +\mathcal{O}(n^{-\zeta_1}),$$
where we have used assumption (\textbf{A1}) in the last step, and $\mu_n:=\mu|_{ M_n}/\mu( M_n)$. Now by the Markov inequality,
for any $\rho>0$
\begin{align*}
\mu_n(e^{S_{n,k}}>e^{-\rho k})&\leq e^{\rho k} \mu_n(e^{ S_{n,k}})\leq (d e^{\rho})^k +\mathcal{O}(n^{-\zeta_1}) e^{\rho k}.
\end{align*}
We choose $q\in (0,1)$ such that $0<\zeta_1-q<1$, and define $$\rho_n=\frac{q}{b\ln n},$$ then one can check that there exists $N>1$, such that $$\eta_0:=\sup_{n\geq N} d e^{ \rho_n}<1,$$ and $ e^{\rho_n k}<n^q$, for any $k=1,\cdots, \psi(n)$. Now we have
$$\mu_n(R(F^k(x))>e^{-\rho_n k}n)\leq \eta_0^k+ \mathcal{O}(n^{-\zeta_1+q}).$$
Let $\varepsilon_n=e^{-\rho_n}$, then above inequality is equivalent to
$$\mu_n(R(F^k(x))>\varepsilon_n^k n)<\eta_0^k +\mathcal{O}(\frac{1}{n^{\zeta_1-q}}).$$
Thus \textbf{(H2)(a)} holds, with $n\varepsilon_n^k \geq n^{1-q}$, for $k=1,\cdots, (b\ln n)^2$, and $\zeta:=\zeta_1-q\in (0,1)$.\end{proof}
The stadium billiard table, introduced by Bunimovich in \cite{Bu74},
is comprised of two equal
semicircles which are connected by two parallel lines.
Dynamics on the stadium have been shown to be
non-uniformly hyperbolic, ergodic, and mixing; for some discussion of
these facts see
\cite{Bu74, Bu79, CM07}. In \cite{M04} Markarian proved that
correlations in
Stadia decay as $\mathcal{O}((\ln n)^2/n)$.
Chernov and Zhang later improved in \cite{CZ08} the decay rate to $\mathcal{O}(1/n)$.
It was shown in \cite{Bu79, CM} that if $x \in M_m$, then $F x \in M_k$ for
$k\in B_m=[a_m,b_m]$ with $a_m\asymp m/\beta$, $b_m\asymp \beta m$, where $\beta=3$. Here $a_m\asymp
\sqrt{m}$ means that there exist $c_1> c_2>0$, such that $c_2\leq
\frac{a_m}{\sqrt{m}}\leq c_1$, for any $m\geq 1$.
The transition probability between cells is given by
$$\mu_M(F x \in M_k | x \in M_m) = \frac{3m}{8k^2} +
\mathcal{O}\left(\frac{1}{m^2}\right).$$
It was also shown in \cite{CM, CZ05a,CZ09} that for any $x\in M_n$, $$\bar\eta:=\mathbb{E}(\ln(\tfrac{R(F)}{R})|R(x)=m)=\sum_{k\in B_m}\ln\frac{k}{m} \cdot \frac{3m}{8k^2}<1-\frac{5}{4}\ln 3<0.$$
In \cite{CZ08}, condition (3.5)-(3.8) were checked for both Bunimovich Stadia and Skewed Stadia. Moreover, the remarks before Lemma 3.3 in \cite{CZ08} shows that there exists $\hat M_n\subset M_n$, with
$$\mu(M_n\setminus \hat M_n| M_n)\leq C n^{-1/2},$$ and for any $x\in \hat M_n$,
$$\ln (R(F^m)/R(x))\leq \eta_1+\cdots+\eta_m,$$
where $\{\eta_1,\cdots,\eta_m\}$ are independent random variables with the same distribution as $\eta$. Here $\eta$ is a random variable supported on $[-\ln \beta,\ln\beta+1]$, and
having a negative mean value $\mathbb{E}(\eta)=\bar\eta<0$. We define $d:=e^{\bar\eta}<1$.
Thus one can check that for any $x\in M_n$, for $t=1,\cdots, (b\ln n)^2$,
$$\mathbb{E}(\tfrac{R(F^t)}{R}| R(x)=n)\leq
\mathbb{E}(e^{\eta_1+\cdots+\eta_t}| R(x)=n)+\mathcal{O}(n^{-\frac{1}{2}})\leq d^t+\mathcal{O}(n^{-\frac{1}{2}}).$$
This implies condition (\textbf{A1}), with $\zeta_1=1/2$.
Now by Proposition \ref{prh1h2H3}, we know that condition (\textbf{A1}) implies \textbf{(H2)(a)}.
The case for skewed stadia is very similar to above analysis, so it satisfies \textbf{(H2)(a)}, which we will omit here.
\subsection{Billiards with property \textbf{(H2)(b)}.}
Assume that each cell $M_n$ has dimension $\asymp n^{-a}$ in the unstable direction, dimension $\asymp n^{-\beta}$ in the stable direction, and measure $\mu(M_n)\asymp n^{-d}$, with $d\geq a+\beta>2$. We first foliate $M_n$ with unstable curves $W_{{\alpha}}\subset M_n$ (where $\alpha$ runs through an index set $\mathcal{A}$). These curves have length $|W_{\alpha}|\asymp n^{-\beta}$. Let $\nu_n:=\frac{1}{\mu(M_n)}\mu|_{M_n}$ be the conditional measure of $\mu$ restricted on $M_n$. Let $\mathcal{W}=\cup_{\alpha\in \mathcal{A}} W_{\alpha}$ be the collection of all unstable curves, which foliate the cell $M_n$. Then we can disintegrate the measure $\nu$ along the leaves $W_{\alpha}$. More precisely, in this way we can obtain a standard family $\mathcal{G}_n=(\mathcal{W},\nu_n)$, such that
for any measurable set $A\subset M_n$,
$$\nu_n(A)=\int_{\mathcal{A}}\nu_{\alpha}(W_{\alpha}\cap A) \, d\lambda(\alpha),$$
where $(W_{\alpha},\nu_{\alpha})$ is a standard pair, and $\lambda$ is the probability factor measure on $\mathcal{A}$. For some $k\leq n$, let $\mathcal{R}_k=\bigcup_{i>k}M_i$, which contains all the cells with index greater than $k$. For each unstable curve $W_{\alpha}\in\mathcal{W}$, if $F^n W_{\alpha}$ crosses $\mathcal{R}_k$, then $F^m W_{\alpha}$ is cut into pieces by the boundary of cells in $\mathcal{R}_k$. Moreover, the largest length of these pieces is $\sim k^{-\beta}$. According to the growth lemma (\ref{firstgrowth}), there exists $\theta_0\in(0,1)$, such that we have
\begin{equation}\label{growthest}F^m_*\nu_n(\mathcal{R}_k)\leq c\vartheta_0^m F_*\nu_n(\mathcal{R}_k)+C_z k^{-\beta q_0}.\end{equation}
\bigskip
\noindent{\textbf{Case I. Billiards with cusps.}}\\
This class of billiards were first studied by Machta \cite{Ma83}. It is
known that the billiard maps on these
tables are hyperbolic and ergodic. However, the hyperbolicity is
non-uniform. As a result, correlations
decay with order $ \mathcal{O}(n^{-1})$, see \cite{CZ05a,CZ08, CM07}. Moreover, it was showed that it satisfies the One-Step Expansion \textbf{(h4)} with $q_0=1$.
In \cite{CM07} Chernov and Markarian showed that
the induced map $F$ on a subset $M\subset \mathcal{M}$ has exponential decay
of correlations.
Dynamics of $F$ on billiards with cusps are remarkably different than those
on a stadium when it comes to points travelling
between $m$-cells: if $x \in M_m$ and $F x \in M_k$, then $k\in
B_m=[a_m,b_m]$, with $a_m\asymp \sqrt{m}, b_m\asymp m^2$.
And the transition probability from the $m$-cell to the $k$-cell is
$$\mu_M(F x \in M_k | x \in M_m):=
\frac{\mu(\{x\in M_m: F x \in M_k\})}{\mu(M_m)}
\asymp \frac{m^{2/3}}{k^{7/3}},$$ with $k\in [\sqrt{m}, m^2]$.
Each cell $M_m$
has length approximately $ m^{-7/3}$
in the unstable direction and length
approximately $ m^{-2/3}$
in the stable direction. Its measure is $\mu(M_m)\sim m^{-3}$.
Moreover, it was checked in \cite{CZ08} at the end of section 5 that this class of billiards satisfies for any small enough $e\in (0,1/4)$:
$$\mu(R(F(x))>n^{\frac{1}{2}+e} |R(x)=n)\leq C n^{-\frac{1}{2e}},$$
for some uniform constant $C>0$.
Since each cell $M_n$ has length approximately $ n^{-7/3}$
in the unstable direction, we take $\beta=7/3$, and $k=n^{\frac{1}{2}+e}$. Then we have
$$F_*\nu_n(\mathcal{R}_{n^{\frac{1}{2}+e}})\leq C n^{-\frac{1}{2e}}.$$
Now we apply (\ref{growthest}) to get for any $i=1,\cdots, (b\ln n)^2$,
\begin{align*}\nu_n(R(F^i(x))&>n^{\frac{1}{2}+e})=F^i_*\nu_n(\mathcal{R}_k)\\
&\leq c\vartheta_0^i F_*\nu_n(\mathcal{R}_k)+C_z k^{-\beta}\\
&\leq C\vartheta_0^i n^{-\frac{1}{2e}}+C_z n^{-\frac{7}{3}(\frac{1}{2}+e)}.\end{align*}
This verifies \textbf{(H2)}(b) with $q=1/2-e$, $p=\frac{1}{2e}$.\\
\noindent{\textbf{Case II. Semi-dispersing billiards.}}
Billiards in a square with a finite number of fixed, disjoint circular obstacles removed are
known as semi-dispersing billiards.
Chernov and Zhang proved \cite{CZ08} that this system has a decay of
correlations bounded by
$\text{const} \cdot n^{-1}$. Here the reduced phase space $M$ is made up
only of collisions with the
circular obstacles. The induced map $F: M \rightarrow M$ is then equivalent
to the well studied Lorentz gas
billiard map without horizon \cite{CZ05a}, which is known to have
exponential decay of correlations
(see \cite{CM07}, for instance). The structure of the $m$-cells $M_m =
\{x\in M : R(x) = m\}$ is
examined thoroughly in \cite{BSC90,BSC91,CM07}. We will use some of the
facts presented in
those references.
Many properties of the $m$-cells and of the induced billiard map in the
semi-dispersing case
are quite similar to those in billiards with cusps. In particular, the
measure of each $m$-cell is again
$\mu_M(M_m) \asymp m^{-3}$, with $u$-dimension approximately $m^{-2}$. Thus we take $\beta=2$. Moreover it satisfies the One-Step Expansion Estimate \textbf{(h4)} with $q_0=1$.
It is also know that for a point $x\in M_m$, $Fx\in M_k$
where $k\in B_m=[a_m,b_m]$, with $$a_m\asymp \sqrt{m}, b_m\asymp
m^2,$$
as in billiards with cusps. One major change is the transition
probabilities between cells. For semi-dispersing
billiards, we have (for admissible $k$) that
$$\mu_M(F x \in M_k | x \in M_m) \asymp \frac{m+k}{k^3}.$$
Moreover, it was checked in \cite{CZ08} at Section 5 that this class of billiards satisfies for any small enough $e\in (0,1/4)$:
$$\mu(R(F(x))>n^{\frac{1}{2}+e} |R(x)=n)\leq C n^{-\frac{1}{2e}},$$
for some uniform constant $C>0$.
We take $k=n^{\frac{1}{2}+e}$. Then we have
$$F_*\nu_n(\mathcal{R}_{n^{\frac{1}{2}+e}})\leq C n^{-\frac{1}{2e}}.$$
Now we apply (\ref{growthest}) to get for any $i=1,\cdots, (b\ln n)^2$,
\begin{align*}\nu_n(R(F^i(x))&>n^{\frac{1}{2}+e})=F^i_*\nu_n(\mathcal{R}_k)\\
&\leq c F_*\nu_n(\mathcal{R}_k)+C_z k^{-\beta}\\
&\leq C n^{-\frac{1}{2e}}+C_z n^{-2(\frac{1}{2}+e)}.\end{align*}
This verifies \textbf{(H2)}(b) with $q=1/2-e$, $p=\frac{1}{2e}$.\\
This implies that the semi-dispersing billiards on a rectangle and dispersing billiards with cusps have optimal bounds of decay rates of correlations given by Theorem
\ref{TmMain}.
\subsection{Application to linked-twist maps}
In this section, we apply our main theorem to the linked-twist map
studied in \cite{SS}. We claim that this map satisfy the new condition (\textbf{A1})
We consider the two-dimensional torus $\mathbb{T}^2=[0,2)\times [0,2)$ with coordinates $(x, y)$ (mod 2). On this torus we define two overlapping annuli $P, Q$ by
$P = [0,2]\times [0,1]$, $Q = [0,1]\times [0,2]$.
We denote the union of the annuli by $R = P\cup Q$ and the intersection by
$M = P\cap Q$. The annuli $P$ and $Q$ are vertical and horizontal
strips in the torus.
In order to define a linked twist map on the torus we first define a twist map
on each annulus. A twist map is simply a map in which the orbits move along
parallel lines, but with a uniform shear. In particular, we define
$F : R\to R$, such that
$$F(x,y)=\left\{
\begin{array}{ll}
(x + 2y, y), & \text{ if } (x, y)\in P; \\
(x,y), & \text{ if } (x,y)\in R\setminus P.
\end{array}
\right.$$
Note that $F$ leaves points in $R\setminus P$ unchanged, and any horizontal line in $P$ is invariant. We define the map $G$ similarly:
$$G(x,y)=\left\{
\begin{array}{ll}
(x , y+2x), & \text{ if } (x, y)\in Q; \\
(x,y), & \text{ if } (x,y)\in R\setminus Q.
\end{array}
\right.$$Now the linked twist map $H$ is defined by $H:=G\circ F$, which maps from $R$ to $R$. By calculating the differential $DH$, one can easily check that $\det DH=1$, which implies that $H$ preserves the Lebesgue measure $m$ on $M$.
We will first define a reduced map which enjoys the exponential decay of correlations. More precisely, we define $F_M:M\to M$, to be the return map with respect to $F$, such that for any $(x,y)\in M$, $F_M(x,y)=F^{n}(x,y)$, where $n=R_F(x,y)$ is the first return time of $(x,y)$ to $M$ under iterations of $F$. Similarly, we define $G_M:M\to M$, such that $G_M(x,y)=G^n(x,y)$, where $n=R_G(x,y)$ is the first return of $(x,y)$ to $M$ under iterations of $G$. We define the reduced map as $T:=G_M\circ T$. Then $T$ is the first return map obtained from $H$ onto $M$. Note that as $G$ is an Anosov diffeomorphism restricted on $M$, so by the uniformly hyperbolicity of $G$ on $M$, there exists $N=N(G)>1$ such that $G^N M\subset M$. Let $m_M$ be the conditional Lebesgure measure on $M$, then $T$ preserves $m_M$.
Let $ S_{\pm 1}$ be the singular set of the reduced map $H^{\pm
1}\colon =H_S^{\pm 1}$. In \cite{SS}, Figure 2 shows the structure of
$ S_1$ while Figure 5 shows the image of $ S_{-1}$. Using the
notation of that paper, we label by $\{\Sigma_n\}$ the connected
regions near $(1,0)$ in $ S_1$, as shown in Figure 6 of \cite{SS}, on which the
return time function is $n$. We know from Appendix A of \cite{SS} that
the cell $\Sigma_n$ has length of order $1/n$ and width of order
$1/n^2$. Similarly, we denote by $\{M_n\}$ the level set (or called cells) in $ S_{-1}$ with
backward return time $n$. As it was shown in Lemma 5.3 of \cite{SS},
unstable manifolds have slope $1+\sqrt{2}$, thus we know that the
longer boundary curves of $M_n$ all have slope approximately
$1+\sqrt{2}$, and these cells converge to the fixed point $(1,0)$ as
$n \to \infty$. In addition, one can show that $M_n$ has length of
order $ O(n^{-1})$ and width of order $ O(n^{-2})$. In the proof of
Lemma 5.4 of \cite{SS}, it was shown that when an unstable manifold $W$
intersects $ \Sigma_n$ for some $n$ large enough, it only crosses those
$\Sigma_m$ with $m\in [n, (3+2\sqrt{2})n]$. If we redefine $n$, then we
can say that $W$ intersects only cells $\Sigma_m$, with $m\in I_n =
[n/\beta+c_1, \beta n+c_1]$, where $\beta=1+\sqrt{2}$, for some
constants $c_1, c_2$.
In terms of the singular set $ S_{-1}$, this implies that the image of
$\partial M_n\subset S_{-1}$ will only intersect $\Sigma_m$, for $m\in I_n$,
i.e. $ M_n\subset \cup_{m\in I_n} \Sigma_m$. Thus we take an unstable
manifold $W$ that completely stretches across $M_n$, then its image
$HW$ will be cut into pieces such that each piece is stretched
completely across $M_m$, for $m\in I_n$.
Note that for large $n$, the region $M_n\cap \Sigma_n$ is nearly a
rectangle with dimension $ O(m^{-2}) \times O(n^{-2})$. Now Lemma
5.2 in \cite{SS} implies that the expansion factor of unstable manifolds
in $\Sigma_m$ is $ O(m)$, thus $TM_n\cap \Sigma_m$ is a strip in $M_m$
that completely stretched in the unstable direction and has width $
O(\frac{1}{ mn^2})$.
Thus one can now check that the transition probability of moving from
$\Sigma_n$ to $\Sigma_m$ is
$$\frac{\mu(\Sigma_m
\cap T M_n)}{\mu(TM_n)}=\frac{c_0 \frac{1}{m^2 n^2}}{\frac{1}{n^3}}=c_0\frac{n}{m^2},$$
where $c_0=\beta-\beta^{-1}$ is the normalizing constant, such that
$$\sum_{m\in I_n} \mu(\Sigma_m
\cap T M_n)=\mu(TM_n).$$ More precisely, $c_0$ solves $$\sum_{m=n/\beta}^{\beta n} c_0 \frac{1}{m^2 n^2} = \frac{1}{n^3}.$$
Thus we have shown that this class of maps satisfy (\textbf{A1}). By Proposition \ref{prh1h2H3}, the map satisfies condition
(\textbf{H2}(a)). Thus Theorem 1-3 hold for this map.
\begin{comment}
\section{Appendix. List of major notations}
Here we list notations in the paper that have been used at least twice.
\begin{itemize}
\item $(\mathcal{F},\mathcal{M},\mu)$ --- the original nonuniform hyperbolic system;
\item $(F,M,\mu_M)$ --- the induced uniformly hyperbolic system;
\item $\mathcal{W}^u$ (resp. $\mathcal{W}^s$) --- the collection of all unstable (resp. stable) manifolds for $\mathcal{F}$;
\item $\mathcal{W}^{u}_F$ ( resp. $\mathcal{W}^s_F$) --- the collection of all unstable (resp. stable) manifolds for $F$;
\item $R:\mathcal{M} \to \mathbb{N}$ --- the first hitting time to $M$. It is an extension of the first return time function on $M$;
\item $M_m$ --- the $m$-th level set of $R$ in $M$, i.e. $M_m=\{x\in M\,:\, R(x)=m\}$, it is the closure of the open set $\mathcal{D}_m$, for $m\geq 1$;
\item $\mathcal{S}_1$ ( resp. $\mathcal{S}_{-1}$) --- the singular set of $\mathcal{F}$ ( resp. $\mathcal{F}^{-1}$);
\item $S_1$ ( resp. $S_{-1}$) --- the singular set of $F$ ( resp. $F^{-1}$);
\item $\rho_W$ --- the density function of the $u$-SRB measure $\mu_W$;
\item $(W,\nu)$ --- a standard pair;
\item $(\mathcal{W},\nu)$ --- a standard family, also denoted as $\mathcal{G}=(\mathcal{W},\nu)$, equipped with a factor measure $\lambda$ on the index set $\mathcal{A}$ of $\mathcal{W}$;
\item $F^n (\mathcal{W},\nu)$ --- a standard family defined as $(F^n\mathcal{W}, F^n_*\nu)$ with factor measure $\lambda_n$ on the index set $\mathcal{A}^n$;
\item $\{(W_{\alpha},\mu_{\alpha}),\alpha\in \mathcal{A}^u_M, \lambda^u_M\}$ --- the standard family $(\mathcal{W}^u_F,\mu_M)$;
\item $\{(W_{\alpha},\mu_{\alpha}),\alpha\in \mathcal{A}^u_{\mathcal{M}}, \lambda^u_{\mathcal{M}}\}$ --- the standard family $(\mathcal{W}^u,\mu)$;
\item $\mathfrak{F}(\mathcal{M})$ --- the collection of all standard families on $\mathcal{M}$;
\item $C_{n,b}$ --- the set of all points in $\mathcal{M}$ whose forward orbits have returned to $M$ at least once, but at most $(b\ln n)^2$ times within $n$ iterations, defined as in (\ref{Cnb});
\item $\gamma_0\in (0,1)$ -- is the constant defined in the distortion bound (\ref{distor10});
\item $\varepsilon_1$ --- a constant defined in assumption (\textbf{H2});
\item $\mathcal{R}^*=\Gamma^u\cap \Gamma^s$ --- the hyperbolic product set defined as in Definition \ref{defRsu};
\item $N_1$ --- the number of iterations needed for a standard family to become proper, defined as in lemma \ref{propercF}, $N_1$ depends on the choice of standard family;
\item $T: \mathcal{M} \to \mathcal{R}^*$ --- the first hitting time to $\mathcal{R}^*$, as defined in (\ref{defntau}) such that $T=\mathcal{F}^{\tau}$;
\item $\mathcal{D}_{n,m}$ --- an $s$-subset in $\gamma^s$, such that $T^n\mathcal{D}_{n,m}=\mathcal{F}^m\mathcal{D}_{n,m}$, as defined in (\ref{Dnm});
\item $\alpha_0>1$ --- the constant in (\textbf{H1}), defining the order of $\mu(R>n)\leq C n^{1-\alpha_0}$;
\item $\gamma_i\geq \gamma_0$, $i=1,2$ --- the H\"{o}lder exponent of functions in $\mathcal{H}^{\pm}(\gamma_i)$;
\item $C_{\mathbf{r}}>0$ --- the distortion constant defined in (\textbf{h3});
\item $C_F>C_{\mathbf{r}}$ --- the constant in the definition of standard pair (\ref{standardpair});
\item $C_q>100 C_F$ --- the constant defining a proper family $\mathcal{G}$, i.e. $\mathcal{Z}(\mathcal{G})<C_q$;
\item $\delta_0$ --- a constant defined in (\ref{delta0}), such that a standard pair $(W,\nu)$ is called a proper standard pair, if $|W|<20\delta_0$;
\item $\hat\delta_1\in (0,1)$ --- constants defined in Proposition \ref{firstproper};
\item $\hat\delta_0\in(0,\hat\delta_1)$ --- a constant defined in Lemma \ref{t01};
\item $\delta>0$ --- a constant defined in Lemma \ref{propercF}, and also used in Lemma \ref{renewal1};
\item $\Lambda>1$ --- the minimal expansion factor given in assumption (\textbf{h1});
\item $\vartheta_0\in (0,1)$ --- a constant defined in the absolute continuity property (\textbf{h3});
\item $\vartheta\in (0,1)$ --- a constant chosen in the coupling lemma for the induced map, i.e. Lemma \ref{coupling};
\item $\vartheta_1=\max\{\vartheta_0,\Lambda^{-\gamma_0}\}$ --- a constant in $(0,1)$ as chosen in Lemma \ref{Tmerateo1};
\item $\vartheta_2=\max\{\vartheta, \lambda^{-\gamma_1}\}$ --- a constant for the exponential decay rates of $\{f\circ F^n\}$ as in (\ref{fdecay1});
\item $\vartheta_3\in (0,1)$ --- the constant defined in the Growth Lemma \ref{properagain}.
\end{itemize}
\end{comment}
\medskip
\noindent\textbf{Acknowledgement}.
SV was supported by the ANR-
Project Perturbations, by
the PICS05968 with U. Houston, and with a CNRS support to the Centro de Modelamiento Matem\`atico,
UMI2807, in Santiago de Chile. S.V. would like to thank UMass Amherst for the kind hospitality during the completion of this work. H.-K.Z.\ was partially supported by NSF grant DMS-1151762, by a grant from the Simons
Foundation (337646, HZ); and by the French CNRS with a {\em poste d'accueil} position at the Center of Theoretical Physics in Luminy and by the University of Toulon.
|
\section{Introduction}
\subsection{Setup and main results}
We work over an algebraically closed field $\Bbbk$ of characteristic~$0$.
The purpose of this paper is to survey the results on Hilbert schemes of lines and conics and
automorphism groups of Fano threefolds of Picard rank~$1$. These are usually known to experts,
but sometimes are scattered in the literature or even
in the mathematical folklore.
Let $X$ be a Fano threefold with at worst canonical Gorenstein singularities. In this case, the number
\begin{equation*}
\operatorname{g}(X)=-\frac{1}{2}K_X^3+1
\end{equation*}
is called the \textit{genus} of $X$.
By Riemann--Roch theorem and Kawamata--Viehweg vanishing one has
\begin{equation*}
\dim |-K_X|=\operatorname{g}(X)+1
\end{equation*}
(see e.\,g. \cite[2.1.14]{Iskovskikh-Prokhorov-1999}).
In particular, $\operatorname{g}(X)$ is an integer, and $\operatorname{g}(X)\geqslant 2$.
Recall that~\mbox{$\mathrm{Pic}(X)$} is a finitely generated torsion free abelian group, so that
\begin{equation*}
\mathrm{Pic}(X)\cong \mathbb{Z}^{\rho(X)}
\end{equation*}
(this holds even for Fano varieties with log terminal singularities, see e.g.~\mbox{\cite[Proposition 2.1.2]{Iskovskikh-Prokhorov-1999}}).
The integer $\rho(X)$ is called the \emph{Picard rank} of~$X$.
The maximal number $\iota=\iota(X)$
such that $-K_X$ is divisible by $\iota$ in~\mbox{$\mathrm{Pic}(X)$}
is called the \textit{Fano index}, or just~\emph{index}, of~$X$.
Let $H$ be a divisor class such that
\begin{equation*}
-K_X\sim\iota(X) H.
\end{equation*}
The class $H$ is unique since $\mathrm{Pic}(X)$ is torsion free.
Define the \textit{degree} of $X$ as
\begin{equation*}
\operatorname{d}(X)=H^3.
\end{equation*}
In this paper we concentrate on smooth Fano threefolds of Picard rank 1.
Their classification can be found in~\mbox{\cite[\S12.2]{Iskovskikh-Prokhorov-1999}} (see also \cite{mukai1989biregular}).
We recall it in Tables~\xref{table:Fanos-i-ge-2} and~\xref{table:Fanos-i-1}
which contain the lists of Fanos with index at least two and index one, respectively.
For our purposes it will be important to know for each type of Fano threefolds the minimal integer~$m_0$ such that~\mbox{$m_0H$} is very ample. We list these $m_0$ in the last columns of the tables.
\begin{table}[h]
\caption{Smooth Fano threefolds with $\rho=1$ and $\iota\geqslant 2$}\label{table:Fanos-i-ge-2}
\begin{tabular}{|c|c|c|p{11.5cm}|c|}
\hline $\iota$ & $\operatorname{d}$ & $h^{1,2}$ & Brief description & $m_0$ \\
\hline
\hline $4$ & $1$ & 0 & $\mathbb{P}^{3}$ & 1 \\
\hline
\hline $3$ & $2$ & 0 & quadric hypersurface in $\mathbb{P}^{4}$ & 1 \\
\hline
\hline $2$ & $1$ & 21 & hypersurface in $\mathbb{P}(1,1,1,2,3)$ of degree $6$ & 3 \\
\hline $2$ & $2$ & 10 & double cover of $\mathbb{P}^3$ branched in a quartic surface & 2 \\
\hline $2$ & $3$ & 5 & cubic hypersurface in $\mathbb{P}^{4}$ & 1 \\
\hline $2$ & $4$ & 2 & complete intersection of two quadrics in $\mathbb{P}^{5}$ & 1 \\
\hline $2$ & $5$ & 0 & section of $\mathrm{Gr}(2,5)\subset\mathbb{P}^9$ by a linear subspace of codimension $3$ & 1 \\
\hline
\end{tabular}
\end{table}
\begin{table}[h]
\caption{Smooth Fano threefolds with $\rho=1$ and $\iota=1$}\label{table:Fanos-i-1}
\begin{tabular}{|c|c|c|p{11.5cm}|c|}
\hline $\operatorname{g}$ & $\operatorname{d}$ & $h^{1,2}$ & Brief description & $m_0$ \\
\hline
\hline $2$ & $2$ & $52$ & double cover of $\mathbb{P}^3$ branched in a sextic surface & 3 \\
\hline $3$ & $4$ & $30$ & (a) quartic hypersurface in $\mathbb{P}^4$, or & 1 \\ &&&
(b) double cover of a smooth quadric in
$\mathbb{P}^{4}$ branched in an intersection with a quartic & 2 \\
\hline $4$ & $6$ & $20$ & complete intersection of a quadric and a cubic in $\mathbb{P}^{5}$ & 1 \\
\hline $5$ & $8$ & $14$ & complete intersection of three quadrics $\mathbb{P}^{6}$ & 1 \\
\hline $6$ & $10$ & $10$ & (a) section of $\mathrm{Gr}(2,5)\subset\mathbb{P}^9$ by a linear subspace of codimension~$2$ and a quadric, or
\newline
(b) double cover of the Fano threefold $Y$ with $\iota(Y)=2$ and $\operatorname{d}(Y)=5$ branched in an anticanonical divisor & 1 \\
\hline $7$ & $12$ & $7$ & section of a connected component of the orthogonal Lagrangian Grassmannian $\mathrm{OGr}_{+}(5,10)\subset\mathbb{P}^{15}$ by a linear subspace of codimension~$7$ & 1 \\
\hline $8$ & $14$ & $5$ & section of $\mathrm{Gr}(2,6)\subset\mathbb{P}^{14}$ by a linear subspace of codimension $5$ & 1 \\
\hline $9$ & $16$ & $3$ & section of the symplectic Lagrangian Grassmannian \mbox{$\mathop{\mathrm{LGr}}(3,6)\subset\mathbb{P}^{13}$} by a linear subspace of codimension~$3$ & 1 \\
\hline $10$ & $18$ & $2$ & section of the homogeneous space $\mathrm{G}_2/P\subset\mathbb{P}^{13}$ by a linear subspace of codimension~$2$ & 1 \\
\hline $12$ & $22$ & $0$ & zero locus of three sections of the rank $3$ vector bundle $\Lambda^2{\mathscr{U}}^\vee$, where ${\mathscr{U}}$ is the universal subbundle on $\mathrm{Gr}(3,7)$ & 1 \\
\hline
\end{tabular}
\end{table}
Note that although in some cases (for $\rho = \iota = 1$ and $\operatorname{g} = 3$ or $\operatorname{g} = 6$) there are two types of Fano threefolds, they belong to the same deformation family.
The first main result of this paper is an explicit description of the Hilbert schemes
of lines $\Sigma(Y)$ on Fano threefolds $Y$ of Picard rank 1, index 2 and degree $\operatorname{d}(Y) \geqslant 3$
and the Hilbert schemes of conics $S(X)$ on Fano threefolds $X$ of Picard rank~1, index~1
and genus~\mbox{$\operatorname{g}(X) \geqslant 7$}
(by lines and conics we mean lines and conics in the embedding given by the linear system~$|H|$).
We collect the results we have in the following theorem.
\begin{theorem}
\label{theo:main1}
Let $Y$ be a smooth Fano threefold with $\rho(Y)=1$, $\iota(Y) = 2$, and~\mbox{$\operatorname{d}(Y)\geqslant 3$}.
Then the Hilbert scheme of lines $\Sigma(Y)$ is a smooth irreducible surface and moreover:
\begin{itemize}
\item[(2.3)]
if $\operatorname{d}(Y)=3$, then $\Sigma(Y)$ is a minimal surface of general type
with irregularity $5$, geometric genus $10$ and canonical degree~\mbox{$K_{\Sigma(Y)}^2=45$};
\item[(2.4)]
if $\operatorname{d}(Y)=4$, then $\Sigma(Y)$ is an abelian surface;
\item[(2.5)]
if $\operatorname{d}(Y)=5$, then $\Sigma(Y)\cong\mathbb{P}^2$.
\end{itemize}
Let $X$ be a smooth Fano threefold $\rho(X)=1$, $\iota(X)=1$, and $\operatorname{g}(X)\geqslant 7$.
Then the Hilbert scheme of conics $S(X)$ is a smooth irreducible surface and moreover:
\begin{itemize}
\item[(1.7)]
if $\operatorname{g}(X)=7$, then $S(X)$ is the symmetric square of a smooth curve of genus~$7$;
\item[(1.8)]
if $\operatorname{g}(X)=8$, then $S(X)$ is a minimal surface of general type
with irregularity $5$, geometric genus $10$ and canonical degree~\mbox{$K_{S(X)}^2=45$};
\item[(1.9)]
if $\operatorname{g}(X)=9$, then $S(X)$ is a ruled surface isomorphic to the projectivization of a simple rank $2$ vector bundle on a smooth curve of genus $3$;
\item[(1.10)]
if $\operatorname{g}(X)=10$, then $S(X)$ is an abelian surface;
\item[(1.12)]
if $\operatorname{g}(X)=12$, then $S(X)\cong\mathbb{P}^2$.
\end{itemize}
\end{theorem}
To be honest, most of the information provided by Theorem~\xref{theo:main1}
can be found in the literature
(see \cite{Altman1977}, \cite{DesaleRamanan}, \cite{Furushima1989a},
\cite{Tennison1974}, \cite{Iskovskikh-1980-Anticanonical},
\cite{Puts1982}, \cite{Markushevich1981},
\cite{Iliev-Fano-3folds-genus9}, \cite{iliev2007manivel}, \cite{BrambillaFaenzi}, etc).
Our goal was, in a sense, in collecting all the results together, and cleaning things a bit.
One new improvement here is the case $\iota(X) = 1$ and $\operatorname{g}(X) = 10$, where originally
in~\cite[Proposition 3]{iliev2007manivel} a description of $S(X)$
was known for general $X$ only.
Another improvement is the case $\iota(X) = 1$ and $\operatorname{g}(X) = 9$ where
it was previously known that $S(X)$ is a projectivization of a vector bundle
over a curve of genus~$3$, but simplicity of the vector bundle was known only for
a general threefold~$X$ (see~\cite[\S3]{BrambillaFaenzi}).
Also, our proof for the even genus cases, i.e.\ $\iota(X) = 1$ and \mbox{$\operatorname{g}(X) \in \{8,10,12\}$},
emphasizes the relation between Fano threefolds of index~1 and~2.
We show that if $Y$ is a Fano threefold of index 2 and degree
\begin{equation*}
\operatorname{d}(Y) = \frac{\operatorname{g}(X)}{2} - 1
\end{equation*}
associated to $X$ by~\cite{kuznetsov2009derived}
(see also Appendix~\ref{section-Lines-and-conics}) then $S(X) \cong \Sigma(Y)$.
For small degrees and genera the situation with the Hilbert schemes of lines and conics is much more complicated.
For instance, in the case $\iota(Y) = 2$ and $\operatorname{d}(Y) = 2$ the scheme~\mbox{$\Sigma(Y)$} may be singular and in the case $\iota(X) = 1$ and $\operatorname{g}(X) = 6$ the scheme $S(X)$ may be even reducible.
Furthermore, for small values of $\operatorname{g}(X)$ it is quite hard to get a satisfactory explicit description of~$S(X)$.
Say, for~\mbox{$\operatorname{g}(X) = 2$} the only more or less explicit description of~$S(X)$ we are aware of is as a $240$-to-$1$ branched cover of $\mathbb{P}^2$, which is not very much useful.
So, it seems that our result is a kind of optimal in that direction.
A description of the Hilbert schemes of lines and conics allows to produce some results on the automorphism
groups of the corresponding varieties.
The automorphism groups act on the Hilbert schemes and we prove that the action is faithful in all cases listed in Theorem~\xref{theo:main1}.
In fact, we deduce faithfulness from a more general result (Theorem~\xref{theorem:faithful}).
In particular, it applies to all $X$ with $\operatorname{g}(X) \geqslant 4$ and to some~$X$ with~\mbox{$\operatorname{g}(X) = 3$},
and shows that the action on an irreducible component of a Hilbert scheme is faithful unless $X$ is a double cover of a smooth Fano threefold~$Y$
with $\rho(Y) = 1$, $\iota(Y) = 2$, and~$\operatorname{d}(Y) = \operatorname{g}(X) - 1$, in which case there is an irreducible component of
the Hilbert scheme of conics on $X$
on which the Galois involution of the double cover acts trivially.
We note that in most of the cases listed in Theorem~\xref{theo:main1},
the maximal linear algebraic subgroup
of the automorphism group of the surface $\Sigma(Y)$ and $S(X)$
is at most finite; the only exceptions are cases~\mbox{$(2.5)$} and~\mbox{$(1.12)$}. Due to the faithfulness result, this proves
that the automorphism groups of the threefolds
listed in Theorem~\xref{theo:main1}, except the threefolds~\mbox{$(2.5)$} and~\mbox{$(1.12)$},
are finite as well.
Our second main result is an extension of this observation to the following general statement
describing all possible infinite automorphism groups of Fano threefolds of Picard rank~$1$.
\begin{theorem}
\label{theorem:Prokhorov}
Let $X$ be a smooth
Fano threefold with $\rho(X)=1$. Then the group~\mbox{$\mathrm{Aut}(X)$} is finite
unless one of the following cases occurs:
\begin{itemize}
\item
$\iota(X) = 4$ so that $X\cong \mathbb{P}^3$; then $\mathrm{Aut}(X) \cong \mathrm{PGL}_4(\Bbbk)$;
\item
$\iota(X) = 3$ so that $X$ is a quadric in $\mathbb{P}^4$; then $\mathrm{Aut}(X) \cong \mathrm{PSO}_5(\Bbbk)$;
\item
$\iota(X)=2$, $\operatorname{d}(X)=5$; then $\mathrm{Aut}(X)\cong\mathrm{PGL}_2(\Bbbk)$;
\item\label{theorem:Prokhorov-i=1}
$\iota(X)=1$, $\operatorname{g}(X)=12$, and $X$ is one of the following
\begin{enumerate}
\item\label{theorem:Prokhorov-i=1-mu}
$X=X_{22}^{\mathrm{MU}}$ is the Mukai--Umemura threefold; then $\mathrm{Aut}(X)\cong \mathrm{PGL}_2(\Bbbk)$;
\item\label{theorem:Prokhorov-i=1-a}
$X=X_{22}^{\mathrm a}$ is the threefold of Example~\textup{\xref{example-V22-a}}; then $\mathrm{Aut}(X)\cong \mathbb{G}_{\mathrm a} \rtimes \boldsymbol{\mu}_4$;
\item\label{theorem:Prokhorov-i=1-m}
$X=X_{22}^{\mathrm m}(u)$ is a threefold from the one-dimensional family
of Example~\textup{\xref{example-V22-m}}; then $\mathrm{Aut}(X)\cong \mathbb{G}_{\mathrm m} \rtimes \boldsymbol{\mu}_2$.
\end{enumerate}
\end{itemize}
\end{theorem}
Note that a Fano threefold with $\rho(X) = 1$, $\iota(X) = 2$, and $\operatorname{d}(X) = 5$ is unique (up to isomorphism),
see~\cite[Theorem~II.1.1]{Iskovskikh-1980-Anticanonical} or~\cite[3.3.1--3.3.2]{Iskovskikh-Prokhorov-1999}).
Again, we should say that almost all results of Theorem~\ref{theorem:Prokhorov} were already known,
see~\cite{Prokhorov-1990c}.
The new results here is the explicit description of $\mathrm{Aut}(X_{22}^{\mathrm{a}})$ and $\mathrm{Aut}(X_{22}^{\mathrm{m}}(u))$.
Using the classification of Fano threefolds of Picard rank~$1$
(see~\cite[\S12.2]{Iskovskikh-Prokhorov-1999},
or Tables~\xref{table:Fanos-i-ge-2} and~\xref{table:Fanos-i-1})
we conclude that Theorem~\xref{theorem:Prokhorov} implies the following.
\begin{corollary}
\label{corollary:Jac}
Let $X$ be a smooth
Fano threefold with $\rho(X)=1$. If the group $\mathrm{Aut}(X)$ is infinite, then $h^{1,2}(X)=0$.
\end{corollary}
Our proof of Theorem~\xref{theorem:Prokhorov} relies on a classification
of smooth Fano threefolds. It would be interesting to find a proof
of Corollary~\xref{corollary:Jac} that does not depend on a classification,
and use it to obtain an alternative proof of Theorem~\xref{theorem:Prokhorov}.
Note that \cite[Theorem~1]{Tolman10} can be considered
as a symplectic counterpart of Corollary~\xref{corollary:Jac},
and to some extent can be used to recover it; namely, the results of~\cite{Tolman10}
imply that an automorphism group of a smooth Fano threefold $X$
of Picard rank~$1$ can contain a subgroup isomorphic to~$\mathbb{G}_{\mathrm m}$ only if $h^{1,2}(X)=0$.
\subsection{Applications and future directions}
One of the motivations for writing this paper was the problem of classification of finite subgroups of the Cremona group of rank~3 (cf.~\cite{Prokhorov2009e}, \cite{Prokhorov2011a},
\cite{Prokhorov-2-elementary}, \cite{Prokhorov-Shramov-p-groups}).
This classification problem reduces to investigation of finite automorphism groups of Fano threefolds of Picard number 1 with terminal singularities and Mori fiber spaces.
In particular, it includes classification of finite group of automorphisms of smooth Fano threefolds of Picard number 1.
Notice that here it is important to consider \textit{all} possible Fano threefolds in their deformation classes,
while restricting to \textit{general} Fano threefolds (as it is practiced by many authors) does not work.
This is why we try to push our arguments forward in full generality.
Our results allow to give explicit upper bounds on some parameters of automorphism groups,
which may be useful for further applications to studying birational automorphisms
(see~\cite{Prokhorov-Shramov-JCr3}, and
cf.~\cite{Prokhorov-Shramov-J}, \cite{ProkhorovShramov-RC},
\cite{ProkhorovShramov-dim3}, \cite{Yasinsky2016a}).
It would be interesting to understand which results
of this paper can be generalized to the case of singular Fano threefolds, and
what kind of conclusions one can make about their automorphism groups
(cf.~\cite{Prokhorov-planes}, \cite{Prokhorov-e-QFano7}, \cite{Prokhorov-factorial-Fano-e},
\cite{PrzyjalkowskiShramov2016}).
It would be also nice to extend the results of this paper to higher dimensions.
Naturally, the questions we discuss here become much more complicated.
Besides other things, no classification of higher-dimensional Fano varieties is available
(though, there are some partial results, see, e.g.,
\cite{Kuchle,Kuznetsov-Kuchle,Kuznetsov-Kuchle-c5,Kuznetsov-2018}).
\subsection{Outline of the paper}
The plan of our paper is as follows.
In \S\xref{section:hilbert-schemes} we collect the necessary results about Hilbert schemes of lines and conics on Fano threefolds.
In~\S\ref{subsection:lines-and-conics} we discuss general properties of Hilbert schemes,
while in~\S\ref{subsection:lines-index-2} and~\S\ref{subsection:conics-index-1} we concentrate on Hilbert scheme of lines and conics respectively on Fano threefolds.
Some rather technical parts of the material were moved to Appendix~\xref{appendix:new} for the readers convenience.
Another part of the arguments that uses derived categories perspective and technique is collected in Appendix~\xref{section-Lines-and-conics}.
The main result of this section is a proof of Theorem~\xref{theo:main1}.
In \S\xref{section:automorphisms} we recall various general results
about automorphism groups of algebraic varieties, including actions
on invariant linear systems and some well-known finiteness assertions.
In \S\xref{section:finiteness-for-Fanos} we
prove finiteness of automorphism groups for all Fano threefolds of Picard rank~1 except those listed in Theorem~\xref{theorem:Prokhorov}.
We start in~\S\xref{subsection:faithfulness-general} by proving a general faithfulness result (Theorem~\xref{theorem:faithful}) for an algebraic group action
on irreducible components of Hilbert schemes of (anticanonical) conics on Fano varieties of arbitrary dimension.
In~\S\xref{subsection:act-on-lines} we apply it to the action of the automorphism group of a Fano threefold of index $2$
and degree at least~$3$ on (an irreducible component of) the Hilbert scheme of lines, and
in~\S\xref{subsection:act-on-conics} we apply it to the action of the automorphism group of a Fano threefold of index $1$
and genus at least~$3$ on (an irreducible component of) the Hilbert scheme of conics.
For~\mbox{$\operatorname{d} \geqslant 3$} and~\mbox{$\operatorname{g} \geqslant 7$} we prove faithfulness of these actions and combining it with the description
of Hilbert schemes provided by Theorem~\xref{theo:main1}, deduce finiteness of the automorphism group.
Finally,
in~\S\xref{subsection:finite-Fano} we prove finiteness of the automorphism groups in the remaining (easy)
cases in a more straightforward way.
In \S\xref{section:infinite} we study Fano threefolds of index $1$ and
genus $12$ with infinite automorphism groups via a double projection method
and complete our proof of Theorem~\xref{theorem:Prokhorov}.
In~\S\ref{subsection:v5} we discuss geometry of the Fano threefold $Y$ of index~2 and degree~5 and give an explicit description of its Hilbert scheme of lines.
In~\S\ref{subsection:v22} we explain the double projection method and describe the relation
between the Hilbert scheme of lines on a Fano threefold $X$ of index~1 and genus~12 and the Hilbert scheme of lines on $Y$.
In~\S\ref{subsection:v22-special} we explain the construction of threefolds with infinite automorphism groups, and
in~\S\ref{subsection:aut-explicit} we describe explicitly their Hilbert schemes of lines and automorphisms groups.
In Appendix~\xref{appendix:new} we collect some well-known facts about conics.
First, we remind a classification of surfaces whose Hilbert scheme of conics is at least two-dimensional.
After that we remind a description of normal bundles of reducible and non-reduced conics.
In Appendix~\xref{section-Lines-and-conics} we prove Theorem~\xref{theorem:S-vs-Sigma} relating the Hilbert schemes of conics on Fano threefolds of index 1 and genera $8$, $10$, and $12$
to the Hilbert schemes of lines on Fano threefolds of index 2 and degrees $3$, $4$, and $5$, respectively.
The proof is based on the relation between derived categories of these threefolds established in~\cite{kuznetsov2009derived}.
We remind this approach, discuss some details of the relation, and then prove Theorem~\xref{theorem:S-vs-Sigma}.
We also write down proofs for some well-known results of Mukai concerning vector bundles on Fano threefolds that we could not find in the literature.
\medskip
\textbf{Notation and conventions.}
As we already mentioned,
we work over an algebraically closed field $\Bbbk$ of characteristic $0$.
We assume that the Fano varieties appearing in the paper are smooth unless
the converse is mentioned explicitly.
We remind about the smoothness assumption
only at the most important points of our exposition.
We use the following notation throughout the paper.
By $\mathrm{Pic}(X)$ and $\mathrm{Cl}(X)$ we denote the Picard group and the class group of Weil divisors on the variety $X$, respectively.
Linear equivalence of divisors is denoted by~$\sim$.
For a Fano threefold $X$ we keep the notation $\rho(X)$, $\iota(X)$, $\operatorname{d}(X)$, and $\operatorname{g}(X)$
for the Picard rank, the Fano index, the degree, and the genus of $X$, respectively.
If $\rho(X)=1$, we always denote by $H$ or $H_X$ the ample generator of~\mbox{$\mathrm{Pic}(X)\cong\mathbb{Z}$}.
If $Z$ is a subscheme in $X$, we denote by $[Z]$ the point corresponding to $Z$ in the appropriate Hilbert scheme,
and by~\mbox{$\mathrm{Aut}(X;Z)$} the group of automorphisms of $X$ that preserve~$Z$.
Similarly, if $[D]$ is a divisor class in $\mathrm{Pic}(X)$ or~$\mathrm{Cl}(X)$,
we denote by $\mathrm{Aut}(X;[D])$ the group of automorphisms of $X$ that
preserve the class~$[D]$.
By $\mathrm{Gr}(k,n)$ we denote the Grassmannian of vector subspaces of dimension~$k$ in a vector space of dimension~$n$;
similarly, by $\mathrm{Gr}(k,W)$ we denote the Grassmannian of vector subspaces of dimension~$k$ in a vector space~$W$.
By a linear section of a Grassmannian we always mean
its linear section in the Pl\"ucker embedding, i.e. in the embedding
defined by the ample generator of its Picard group.
By $v_2\colon\mathbb{P}(V)\to\mathbb{P}(\mathrm{Sym}^2 V)$ we denote the second Veronese
embedding.
We denote by $\boldsymbol{\mu}_m$ the group of~\mbox{$m$-th} roots of unity (isomorphic to a cyclic group of order $m$).
\medskip
\textbf{Acknowledgements.}
We are grateful to Olivier Debarre, Francesco Russo, Richard Thomas, and Fyodor Zak for useful discussions.
We are also grateful to the referee for reading our paper.
\section{Hilbert schemes of lines and conics}\label{section:hilbert-schemes}
In this section we discuss general properties of Hilbert schemes of lines and conics
on Fano threefolds and give an explicit description for some of them.
\subsection{General properties of Hilbert schemes}
\label{subsection:lines-and-conics}
Let $X$ be a projective variety with a fixed ample divisor class $H$. Recall that
a \emph{line} (or an \emph{$H$-line} to be more precise) on~$X$ is a subscheme $L \subset X$ with Hilbert polynomial
\begin{equation*}
p_L(t) = 1 + t.
\end{equation*}
Similarly,
a \emph{conic} (or an \emph{$H$-conic}) on $X$ is a subscheme $C \subset X$ with Hilbert polynomial
\begin{equation*}
p_C(t) = 1 + 2t.
\end{equation*}
We denote by
\begin{equation*}
\Sigma(X) = \mathop{\mathsf{Hilb}}\nolimits^{p(t) = 1 + t}(X;H)
\end{equation*}
the Hilbert scheme of lines on $X$,
and by
\begin{equation*}
S(X) = \mathop{\mathsf{Hilb}}\nolimits^{p(t) = 1 + 2t}(X;H)
\end{equation*}
the Hilbert scheme of conics on $X$.
\begin{lemma}\label{lemma:lines-conics}
Let $X$ be a projective variety with an ample divisor class $H$.
\begin{itemize}
\item[(i)] If $2H$ is very ample and $L \subset X$ is an $H$-line
then $L \cong {\mathbb{P}}^1$ and ${\mathscr{O}}_X(H)\vert_L \cong {\mathscr{O}}_L(1)$.
\item[(ii)]
If $H$ is very ample and $C \subset X$ is an $H$-conic
then $C$ is purely
one-dimensional and
\begin{itemize}
\item[$\bullet$] either $C$ is a \emph{smooth conic}, i.e.\ $C \cong {\mathbb{P}}^1$ and ${\mathscr{O}}_X(H)\vert_C \cong {\mathscr{O}}_C(2)$,
\item[$\bullet$] or $C$ is a \emph{reducible conic}, i.e.\ $C = L_1 \cup L_2$ for two distinct lines $L_1$ and $L_2$ on $X$ intersecting transversally at a point,
\item[$\bullet$] or $C$ is a \emph{non-reduced conic},
i.e. a non-reduced subscheme $C \subset X$ such that $C_{\operatorname{red}} = L$ is a line
and $I_{L}/I_C \cong {\mathscr{O}}_{L}(-1)$.
\end{itemize}
\end{itemize}
\end{lemma}
\begin{proof}
First, assume that $H$ is very ample. Then we may assume that $X = {\mathbb{P}}^n$,
and $H$ is the class of a hyperplane.
If $p_L(t) = 1 + t$ then all irreducible components of $L$ have dimension at most 1,
and the sum of the degrees (with multiplicities) of all one-dimensional components is 1.
Let~$L_0$ be the purely one-dimensional part of $L$ and let $\ell$ be the sum of the lengths
of all zero-dimensional components (including embedded ones). Then by the above
observation~$L_0$ is integral of degree 1, hence $L_0$ is ${\mathbb{P}}^1$ linearly embedded into ${\mathbb{P}}^n$.
In particular, one has~\mbox{$p_{L_0}(t) = 1 + t$}, hence~\mbox{$p_L(t) = 1 + \ell + t$}, which means $\ell = 0$
and so $L = L_0$.
Analogously, let $p_C(t) = 1 + 2t$. Then all irreducible components of $C$ have dimension at most 1,
and the sum of the degrees (with multiplicities) of all one-dimensional components is 2.
Let $C_0$ be the purely one-dimensional part of $C$ and let $\ell$ be the sum of the lengths
of all zero-dimensional components (including embedded ones). If $C_0$ is integral, then
it is contained in the linear span of any triple of its points. Thus $C_0$ is a divisor
of degree~2 on~${\mathbb{P}}^2$, so $C_0 \cong {\mathbb{P}}^1$ and ${\mathscr{O}}_X(H)\vert_{C_0} \cong {\mathscr{O}}_{C_0}(2)$.
Furthermore, we have~\mbox{$p_{C_0}(t) = 1 + 2t$}, hence~\mbox{$p_C(t) = 1 + \ell + 2t$}
which means $\ell = 0$ and $C = C_0$.
If $C_0$ is not integral, then either it has two different irreducible components $L_1$ and~$L_2$ of degree~1,
or one irreducible component $L$ of degree~1 with multiplicity~2. In the first case,~$L_1$ and~$L_2$ are lines,
hence their scheme-theoretic intersection has length~\mbox{$\delta = 0$} or~\mbox{$\delta = 1$}.
It follows that
\begin{equation*}
p_C(t) = \ell + (1 + t) + (1 + t) - \delta,
\end{equation*}
which means
that $\ell = 0$ and $\delta = 1$. In other words, $L_1$ and $L_2$ meet at a point and~\mbox{$C = L_1 \cup L_2$}.
In the second case we have a canonical epimorphism ${\mathscr{O}}_{C_0} \to {\mathscr{O}}_L$
and its kernel is a line bundle on $L$, hence is isomorphic to ${\mathscr{O}}_L(k)$ for some $k \in \mathbb{Z}$.
Then
\begin{equation*}
p_C(t) = \ell + (1+t) + (1+k+t),
\end{equation*}
which implies $k = - (1 + \ell)$.
On the other hand, it is easy to see that ${\mathscr{O}}_L(k)$ is a quotient of $I_L/I_L^2$ which is the conormal
bundle of $L$ in ${\mathbb{P}}^n$, hence is isomorphic to~\mbox{${\mathscr{O}}_L(-1)^{\oplus (n-1)}$}.
Therefore $k \geqslant -1$, so comparing with the previous observation, we see that $k = -1$ and~$\ell = 0$.
In other words, $C_\mathrm{red} = L$ and $I_L/I_C \cong {\mathscr{O}}_L(-1)$.
Finally, assume that $H$ is not ample, but $2H$ is very ample and $p_L(t) = 1 + t$.
Then with respect to $2H$ the Hilbert polynomial of $L$ is $1 + 2t$, hence in the embedding
of $X$ given by the linear system~$2H$ it is a conic. But it can be neither reducible, nor non-reduced conic,
since $X$ contains no curves which have degree 1 with respect to $2H$. Thus $L \cong {\mathbb{P}}^1$
and ${\mathscr{O}}_X(2H)\vert_L \cong {\mathscr{O}}_L(2)$, which implies the claim.
\end{proof}
\begin{remark}\label{remark:double-lines}
If for a line $L \subset X$ there is a non-reduced conic $C$ with $C_\mathrm{red} = L$, we will say that $L$ \emph{admits a structure of a non-reduced conic}.
It is worth noting that in contrast to the case of a projective space,
not every line admits such a structure.
Indeed, as we have seen in the proof of Lemma~\xref{lemma:lines-conics} above, a line $L$ admits a structure of a non-reduced conic if and only if there is an epimorphism
${\mathscr{N}}^\vee_L = I_L/I_L^2 \to {\mathscr{O}}_L(-1)$.
In Remark~\xref{remark:dl-special} below we discuss for which lines on Fano threefolds this holds.
\end{remark}
\begin{remark}
It is easy to see that one cannot have the same results as in Lemma~\xref{lemma:lines-conics} under weaker assumptions.
Indeed, assume we consider $\Sigma(X)$ and only the divisor~$3H$, but not~$2H$, is very ample.
Then we can realize $\Sigma(X)$ as a subscheme in $\mathop{\mathsf{Hilb}}\nolimits^{p(t) = 1 + 3t}({\mathbb{P}}^n; 3H)$.
The latter, however, has two irreducible components: one parameterizing normal rational cubic curves, and the other parameterizing plane cubics
plus a point (possibly an embedded one). Therefore the same is true in general for $\Sigma(X)$.
Similarly, assume we consider~\mbox{$S(X)$} and only~$2H$, but not~$H$, is very ample.
Then we can realize~$S(X)$ as a subscheme in~\mbox{$\mathop{\mathsf{Hilb}}\nolimits^{p(t)=1+4t}({\mathbb{P}}^n; 2H)$}.
The latter Hilbert scheme also has several irreducible components,
some of which parameterize curves of other types
than those listed in Lemma~\xref{lemma:lines-conics}(ii).
\end{remark}
From now on we consider the Hilbert schemes of lines and conics on
Fano threefolds of Picard rank~1 and index~1 or~2
(with respect to the ample generator $H$ of the Picard group). We note that the Hilbert schemes $\Sigma(X)$ and $S(X)$ are
nonempty by \cite{Shokurov1979a} (see also \cite{Reid-lines}). As it was explained above to avoid pathologies
when considering $\Sigma(X)$ we should restrict to the case when $2H$ is very ample, i.e., to Fano threefolds of index~1 and genus $\operatorname{g} \geqslant 3$,
as well as Fano threefolds of index~2 with~\mbox{$\operatorname{d} \geqslant 2$}.
Similarly, when considering~$S(X)$ we should restrict to the case when $H$ is very ample:
in the index~1 case
this means that either $\operatorname{g} \geqslant 4$, or $\operatorname{g} = 3$ and $X$ is a quartic threefold,
while in the index~2 case
this means $\operatorname{d} \geqslant 3$.
Under our assumptions, by Lemma~\xref{lemma:lines-conics} both lines and conics are locally complete intersections, hence their conormal and normal sheaves are locally free.
We will need some facts about them. The first is quite standard.
\begin{lemma}[{\cite[Propositions~III.1.3(ii) and III.2.1(i), Lemma III.3.2]{Iskovskikh-1980-Anticanonical}}]
\label{lemma:normal-bundles}
If $L$ is a line and $C$ is a smooth conic on a Fano threefold $X$ of index $1$ then
\begin{equation*}
{\mathscr{N}}_{L/X} \cong {\mathscr{O}}_L(a) \oplus {\mathscr{O}}_L(-1-a)
\qquad\text{and}\qquad
{\mathscr{N}}_{C/X} \cong {\mathscr{O}}_C(a) \oplus {\mathscr{O}}_C(-a)
\end{equation*}
for some $a \geqslant 0$.
If $L$ is a line and $C$ is a smooth conic on a Fano threefold $Y$ of index $2$ then
\begin{equation*}
{\mathscr{N}}_{L/Y} \cong {\mathscr{O}}_L(a) \oplus {\mathscr{O}}_L(-a)
\qquad\text{and}\qquad
{\mathscr{N}}_{C/Y} \cong {\mathscr{O}}_C(1+a) \oplus {\mathscr{O}}_C(1-a)
\end{equation*}
for some $a \geqslant 0$.
\end{lemma}
It is a bit harder to deal with the normal bundle of a reducible or non-reduced conic~$C$
(see, however,~\S\xref{subsection:degenerate-conics}).
Recall that by~\cite{FGA} or \cite[Theorem I.2.8]{Kollar-1996-RC} the tangent space to the Hilbert scheme at a point corresponding to a locally complete intersection
subscheme $Z \subset X$ is~\mbox{$H^0(Z,{\mathscr{N}}_{Z/X})$} and the obstruction space is $H^1(Z,{\mathscr{N}}_{Z/X})$.
Therefore, the dimension of any irreducible
component of the Hilbert scheme is bounded from below by
the Euler characteristic~$\chi({\mathscr{N}}_{Z/X})$ of the normal bundle.
By Lemma~\xref{lemma:normal-bundles} and Corollaries~\xref{corollary:normal-reducible-conic-3fold}, and~\xref{corollary:normal-nonreduced-conic-3fold}
in the cases that are most relevant for us this gives.
\begin{corollary}\label{proposition:dimensions-hilbert-schemes}
The following assertions hold.
\begin{enumerate}
\item[(i)]
If $Y$ is a Fano threefold of index $2$, then the dimension of any component of $\Sigma(Y)$ is at least $2$.
\item[(ii)]
If $X$ is a Fano threefold of index $1$, then the dimension of any component of $\Sigma(X)$ is at least $1$.
\item[(iii)]
If $X$ is a Fano threefold of index $1$, then the dimension of any component of $S(X)$ is at least $2$.
\end{enumerate}
\end{corollary}
A bit later we will see that in all the cases listed in Corollary~\xref{proposition:dimensions-hilbert-schemes},
the Hilbert schemes are equidimensional of dimensions 2, 1, and 2 respectively (see Lemmas~\xref{lemma-hilb-x-1}, \xref{lemma-hilb-y-1}, \xref{lemma:exotic-component}, and~\xref{lemma-hilb-x-2}).
In what follows we will say that a line or a smooth conic is \emph{ordinary}, if in the notation of Lemma~\xref{lemma:normal-bundles} we have $a = 0$, and \emph{special}, if $a \geqslant 1$.
Furthermore, if $a=1$ we will say that the corresponding line (or conic) is \emph{$1$-special},
and if $a \geqslant 2$ we will say that it is $2$-special.
\begin{corollary}\label{corollary:hilbert-smoothness}
The following assertions hold.
\begin{enumerate}
\item[(i)] If $Y$ is a Fano threefold of index $2$, then the Hilbert scheme $\Sigma(Y)$ is smooth of dimension $2$ at points
corresponding to ordinary lines or $1$-special lines and singular at points corresponding to $2$-special lines.
\item[(ii)] If $X$ is a Fano threefold of index $1$, then the Hilbert scheme $\Sigma(X)$ is smooth of dimension $1$ at points
corresponding to ordinary lines,
and is singular at points corresponding to special lines.
\item[(iii)] If $X$ is a Fano threefold of index $1$, then
the Hilbert scheme $S(X)$ is smooth of dimension~$2$ at points
corresponding to smooth ordinary or smooth $1$-special conics,
and singular at points corresponding to smooth $2$-special conics.
\end{enumerate}
\end{corollary}
\begin{proof}
By Lemma~\xref{lemma:normal-bundles} in the cases claimed to be corresponding to smooth points the obstruction space $H^1(Z,{\mathscr{N}}_{Z/X})$ vanishes,
and in the cases claimed to be corresponding to singular points the tangent space $H^0(Z,{\mathscr{N}}_{Z/X})$ jumps.
\end{proof}
\begin{remark}\label{remark:dl-special}
Note that according to Remark~\xref{remark:double-lines},
only $1$-special lines on Fano threefolds (both of index 1 and 2)
admit a structure of a non-reduced conic,
and this structure is unique.
In particular, if $X$ is a Fano threefold of index 1 such that $\Sigma(X)$ is smooth, then~$X$ has no non-reduced conics.
\end{remark}
As we will see below, it is useful to know that Fano threefolds do not contain some special surfaces.
We check this in the next lemma.
\begin{lemma}\label{lemma:cones}
Let $X$ be a Fano threefold with $\rho(X)=1$ and $\iota(X)=1$,
and suppose that~\mbox{$-K_X$} is very ample.
Then the following assertions hold.
\begin{enumerate}
\item [(i)]
The threefold $X$ contains neither
the Veronese surface $v_2(\mathbb{P}^2)$, nor any of its linear projections.
\item [(ii)]
If $X$ contains a two-dimensional cone $Z$
then $X$ is a quartic in~${\mathbb{P}}^4$ and the base of~$Z$ is a smooth plane quartic curve.
\end{enumerate}
\end{lemma}
\begin{proof}
Assume that $Z \subset X$ is one of the surfaces listed in assertion~(i), so that in
particular~\mbox{$H^2\cdot Z \leqslant 4$}.
Since $\rho(X) = 1$, we have~\mbox{$Z \sim rH$} for some positive integer $r$, hence
\begin{equation*}
4\geqslant H^2\cdot Z = rH^3 = r(2\operatorname{g}(X) - 2)
\end{equation*}
Since $-K_X$ is very ample, we have $\operatorname{g}(X) \geqslant 3$.
Hence the only possible case is when~$X$ is a quartic in~$\mathbb{P}^4$, $r = 1$, and $H^2\cdot Z=4$, so that $Z$ is a regular projection of the Veronese surface.
Moreover, we see that $Z$ is a hyperplane section of the smooth hypersurface~\mbox{$X\subset\mathbb{P}^4$},
so that $Z$ has at worst isolated singularities, and $Z$ is contained in~$\mathbb{P}^3$.
But the latter is impossible for a regular projection of a Veronese surface.
This gives assertion~(i).
If $Z \subset X$ is a two-dimensional cone with vertex at a point $P \in Z$ then $Z$ is contained
in the embedded tangent space to $X$ at $P$. Since $X$ is smooth, the embedded tangent space to $X$ at $P$ is ${\mathbb{P}}^3$, so $Z$
is an irreducible component of a hyperplane section of $X$. But since~\mbox{$\mathrm{Pic}(X) =\mathbb{Z}\cdot H$},
it follows that $Z$ is a hyperplane section, so~\mbox{$\operatorname{g}(X) = 3$} and hence~$X$ is a quartic threefold.
The base of $Z$ is a quartic curve; it is smooth since $Z$, being a hyperplane section of a smooth hypersurface $X \subset \mathbb{P}^4$, can have at worst isolated singularities.
This gives assertion~(ii).
\end{proof}
An easy parameter count shows that a general quartic threefold in ${\mathbb{P}}^4$ does not contain cones.
However, there are examples of quartic threefolds with cones.
\begin{example}[see \cite{Tennison1974}]\label{example:quartic-cone}
Consider the Fermat quartic threefold
\begin{equation*}
X = \{x_0^4 + x_1^4 + x_2^4 + x_3^4 + x_4^4 = 0\} \subset {\mathbb{P}}^4.
\end{equation*}
Let $P \in X$ be a point with the last three coordinates equal to zero (there are four such points) and consider the plane $\Pi = \{x_0 = x_1 = 0\}$.
Consider the hyperplane~\mbox{$H(P,\Pi) \subset {\mathbb{P}}^4$}
spanned by the point $P$ and the plane $\Pi$.
Then $X \cap H(P,\Pi)$ is the cone with vertex~$P$ and the base being the plane Fermat quartic $\Pi \cap X$.
Using the action of the automorphism group~\cite{Shioda-Fermat-curves}
\begin{equation*}
\mathrm{Aut}(X)\cong\boldsymbol{\mu}_4^4 \rtimes \mathfrak{S}_5,
\end{equation*}
we can construct $4\cdot 10 = 40$ such cones.
\end{example}
\subsection{Hilbert schemes of lines}
\label{subsection:lines-index-2}
Let $X$ be a smooth Fano threefold. Let $\Sigma_0$ be an irreducible component of the Hilbert scheme
of lines $\Sigma(X)$, and consider the reduced scheme
structure on~$\Sigma_0$.
Restricting to $\Sigma_0$ the universal family of lines,
we obtain a diagram
\begin{equation}\label{diagram-universal-line}
\vcenter{\xymatrix{
& {\mathscr{L}}_0(X) \ar[dl]_q \ar[dr]^p \\
\Sigma_0 && X
}}
\end{equation}
The map $q\colon{\mathscr{L}}_0(X) \to \Sigma_0$ is a $\mathbb{P}^1$-bundle.
Let $L \subset X$ be a line corresponding to a point~$[L]$ in the component $\Sigma_0$ of the Hilbert scheme.
The fiber $q^{-1}([L])$ is identified by the map~$p$ with the line $L$. Note that the normal bundle of $L$ in ${\mathscr{L}}_0(X)$
is the trivial bundle of rank equal to the dimension of the tangent space to $\Sigma_0$ at $[L]$.
So,
the differential of $p$ is the map
\begin{equation}\label{equation-dp-lines}
dp \colon
{\mathscr{N}}_{L/{\mathscr{L}}_0(X)} = T_{[L]}\Sigma_0 \otimes {\mathscr{O}}_L \hookrightarrow
T_{[L]} \Sigma(X) \otimes {\mathscr{O}}_L =
H^0(L,{\mathscr{N}}_{L/X}) \otimes {\mathscr{O}}_L \to {\mathscr{N}}_{L/X}
\end{equation}
with the last map being given by evaluation.
This is very useful for understanding the infinitesimal structure of the map $p$ along $L$.
\begin{lemma}
\label{lemma-hilb-x-1}
If $X$ is a Fano threefold with $\rho(X)=1$, $\iota(X)=1$, and very ample~$-K_X$, then
every irreducible component of the Hilbert scheme $\Sigma(X)$ of lines on $X$ is one-dimensional.
\end{lemma}
\begin{proof}
Let $\Sigma_0 \subset \Sigma(X)$ be an irreducible component of dimension $k \geqslant 2$,
and consider the reduced scheme structure on~$\Sigma_0$.
Consider the map~\eqref{equation-dp-lines}.
Its source is a trivial vector bundle,
and by Lemma~\xref{lemma:normal-bundles} its target is
\begin{equation*}
{\mathscr{N}}_{L/X} \cong {\mathscr{O}}_L(a) \oplus {\mathscr{O}}_L(-1-a)
\end{equation*}
with $a \geqslant 0$.
Since the second summand has no global sections, the image of $dp$ is contained in the first summand, hence the rank of $dp$ does not exceed 1.
Moreover, since $\Sigma_0$ is reduced, so is~${\mathscr{L}}_0(X)$, and hence so is the general fiber of the map $p$.
This means that the map
\begin{equation*}
p\colon{\mathscr{L}}_0(X) \to X
\end{equation*}
has fibers of dimension at least $k-1$,
hence the image $Z = p({\mathscr{L}}_0(X))$ has dimension at most
$k + 1 - (k - 1) = 2$.
Therefore, $Z \subset X$ is a surface with $\dim\Sigma(Z) \geqslant k \geqslant 2$.
By Corollary~\xref{corollary:lines-conics-plane-1} the surface $Z$ is a plane,
but by Lemma~\xref{lemma:cones}(i) the threefold $X$ contains no planes,
which is a contradiction.
\end{proof}
\begin{remark}
\label{remark:nonreduced-sigma}
On most Fano threefolds $X$ with $\rho(X)=1$ and $\iota(X)=1$ a general point of every irreducible component of $\Sigma(X)$ is an ordinary line.
However, there are some exceptions.
First, if~$X$ is the Mukai--Umemura threefold of genus 12 (see \cite[\S 6]{Mukai-Umemura-1983} or Theorem~\xref{theorem-mu-22} below)
then all lines on $X$ are special, and in fact $\Sigma(X)$ is everywhere non-reduced with~\mbox{$\Sigma(X)_{\operatorname{red}}\cong{\mathbb{P}}^1$}
(see Proposition~\ref{proposition:lines-special-v22}).
In the opposite direction not that much is known. What we know is that the
Mukai--Umemura threefold is the only one with everywhere non-reduced $\Sigma(X)$ in genus~12,
and that in genus~10 and~9 there are no threefolds with~\mbox{$\Sigma(X)$} everywhere non-reduced
\cite{Prokhorov-1990b}, \cite{Gruson-Laytimi-Nagaraj}.
Another interesting example is a quartic $X$ in ${\mathbb{P}}^4$ containing a cone (see Example~\xref{example:quartic-cone}) so that $\operatorname{g}(X) = 3$.
In this case each line $L$ on the cone has a structure of a non-reduced conic (obtained by
intersecting the cone with its tangent plane along $L$),
hence by Remark~\xref{remark:dl-special} each such~$L$ is 1-special, hence
the corresponding irreducible component of~$\Sigma(X)$ is everywhere non-reduced with the underlying reduced scheme being a smooth plane quartic.
For instance, if $X$ is the Fermat quartic of Example~\ref{example:quartic-cone},
then $\Sigma(X)$ is the union of 40 such non-reduced components, (see \cite[Example in \S 2]{Tennison1974}).
\end{remark}
\begin{remark}
Suppose that $X$ is a Fano threefold with $\rho(X)=1$ and $\iota(X)=1$.
If $X$ is general in the corresponding deformation family,
then $\Sigma(X)$ is a smooth curve, and its genus can be
computed in every case, see~\cite[Theorem~4.2.7]{Iskovskikh-Prokhorov-1999}.
\end{remark}
Now consider the Hilbert scheme of lines on threefolds $Y$ of index 2.
\begin{lemma}\label{lemma-hilb-y-1}
Let $Y$ be a Fano threefold with $\rho(Y)=1$ and $\iota(Y)=2$.
Suppose that the divisor $2H$ is very ample, i.\,e. that $\operatorname{d}(Y)\geqslant 2$.
Then every irreducible component of the Hilbert scheme $\Sigma(Y)$ of lines on $Y$ is two-dimensional and its general point corresponds to an ordinary line.
In particular, every irreducible component of $\Sigma(Y)$ is generically smooth.
Moreover,
the map $p\colon{\mathscr{L}}_0(Y) \to Y$ is surjective, generically finite, does not contract divisors, and is not birational.
\end{lemma}
\begin{proof}
Let $\Sigma_0 \subset \Sigma(Y)$ be an irreducible component, and consider the reduced scheme structure
on~$\Sigma_0$.
Assume that a line corresponding to a general point of $\Sigma_0$ is special.
By the argument of Lemma~\xref{lemma-hilb-x-1} the rank of the map $dp$ does not exceed 1, and the map
\begin{equation*}
p\colon{\mathscr{L}}_0(Y) \to Y
\end{equation*}
has fibers of dimension at least $k-1$, where $k = \dim \Sigma_0$. Therefore, the image
\mbox{$Z = p({\mathscr{L}}_0(Y))$}
has dimension
at most $k + 1 - (k - 1) = 2$. Thus $Z \subset Y$ is a surface,
and by Corollary~\xref{proposition:dimensions-hilbert-schemes} one has $\dim\Sigma(Z) \geqslant \dim \Sigma_0 = k \geqslant 2$.
By our assumption the divisor~$2H$ is very ample, hence by Corollary~\xref{corollary:lines-conics-plane-1} the surface $Z$ is a plane.
But $Y$ cannot contain a plane by adjunction, which gives a contradiction.
Therefore a general point of~$\Sigma_0$ corresponds to an ordinary line $L$,
hence $\dim \Sigma_0 = 2$ by Corollary~\xref{corollary:hilbert-smoothness}.
Moreover, for such $L$ all the maps in~\eqref{equation-dp-lines} are isomorphisms, so
the map $dp$ is an isomorphism on $L$, hence the map~$p$ is dominant and unramified along $L$. Since $p$ is also proper, it is surjective.
Moreover, since~\mbox{$\dim {\mathscr{L}}_0(Y) =3= \dim Y$}, the map $p$ is generically finite.
Now consider the ramification locus $R(p) \subset {\mathscr{L}}_0(Y)$ of the map $p$.
Let $L$ be a line corresponding to an arbitrary point of $\Sigma_0$.
If $L$ is an ordinary line then we have already seen that~$p$ is unramified along $L$.
If, however, $L$ is special, the map~$dp$ is degenerate at all points of $L$.
Therefore the ramification locus $R(p) \subset {\mathscr{L}}_0(Y)$ is just the preimage under~$q$ of the locus of special lines in $\Sigma_0$.
Assume that $D \subset {\mathscr{L}}_0(Y)$ is an irreducible divisor contracted by $p$.
Then $D \subset R(p)$, hence $D$ is a union of fibers of $q$.
Therefore, $p(D)$ is a union of lines. Since $D$ is irreducible and~\mbox{$\dim p(D) < \dim D = 2$},
it is just one line $L$. But then $D \subset q^{-1}([L])$ is not a divisor.
Assume that $p$ is birational. Since $\rho({\mathscr{L}}_0(Y)) \geqslant 2$ and $\rho(Y) = 1$,
the morphism
$p$ cannot be an isomorphism. Since $Y$ is smooth, the exceptional locus of $p$
should be a divisor contracted by $p$ (see {\cite[\S2.3, Theorem 2]{Shafarevich1994a}}),
which contradicts the above conclusions.
\end{proof}
\begin{remark}
If $Y$ is a Fano threefold with $\rho(Y) = 1$, $\iota(Y) = 2$ and
$\operatorname{d}(Y)=4$ or $5$ then the map $p$ is finite of degree $4$ and $3$, respectively.
This is no longer true in the cases~\mbox{$\operatorname{d}(Y)=3$} and~\mbox{$\operatorname{d}(Y)=2$}.
For a cubic threefold $Y\subset {\mathbb{P}}^4$ the map $p$
has one-dimensional fibers exactly when~$Y$ contains \emph{generalized Eckardt points},
i.e. points $P$ such that the embedded tangent space at~$P$
cuts out a cone on~$Y$.
For example, the Fermat cubic contains 30 generalized Eckardt points.
Similarly, there are examples of double covers of ${\mathbb{P}}^3$ branched in quartic surfaces
(for example, over Fermat quartic surfaces) that
contain points over which $p$ is not finite.
\end{remark}
The following result is well known
(see e.\,g.~\cite[Proposition~III.1.3(iii)]{Iskovskikh-1980-Anticanonical}).
\begin{proposition}\label{hilb-lines-smooth}
Let $Y$ be a Fano threefold with $\rho(Y) = 1$, $\iota(Y) = 2$, and $\operatorname{d}(Y) \geqslant 3$.
Then the Hilbert scheme of lines $\Sigma(Y)$ is a smooth surface.
\end{proposition}
\begin{proof}
By Corollary~\xref{corollary:hilbert-smoothness} it is enough to show that there are no $2$-special lines.
Since~\mbox{$\operatorname{d}(Y) \geqslant 3$},
the class $H$ is very ample and defines an embedding~\mbox{$Y \hookrightarrow {\mathbb{P}}^n$}.
Consider the standard exact sequence
\begin{equation*}
0 \to {\mathscr{N}}_{L/Y} \to {\mathscr{N}}_{L/{\mathbb{P}}^n} \to {\mathscr{N}}_{Y/{\mathbb{P}}^n}\vert_L \to 0.
\end{equation*}
Note that ${\mathscr{N}}_{L/{\mathbb{P}}^n} \cong {\mathscr{O}}_L(1)^{\oplus (n-1)}$. Thus ${\mathscr{N}}_{L/Y}$ is a subbundle in
the direct sum of $n-1$ copies of ${\mathscr{O}}_L(1)$. This means that ${\mathscr{N}}_{L/Y}$ cannot have a summand isomorphic to ${\mathscr{O}}_L(a)$ with~\mbox{$a \geqslant 2$}.
Hence $L$ cannot be $2$-special.
\end{proof}
\begin{remark}
If $\operatorname{d}(Y) = 2$, so that
$f\colon Y \to \mathbb{P}^3$ is a double cover of $\mathbb{P}^3$ branched in a smooth quartic surface $S$, the Hilbert scheme of lines $\Sigma(Y)$ is,
in fact, a double cover of the subscheme of $\mathrm{Gr}(2,4)$ parameterizing bitangent lines to the surface $S$,
branched in a finite number of points corresponding to lines contained in $S$.
Moreover, if $L_0 \subset S$ is such a line,
and $L= f^{-1}(L_0)_{\mathrm{red}}$, then $L$ is a $2$-special line on $Y$ and hence $\Sigma(Y)$ is singular at~$L$
(cf. \cite[Remark to Proposition III.1.3]{Iskovskikh-1980-Anticanonical}).
\end{remark}
For Fano threefolds of index 2 and degree at least 3 one can describe $\Sigma(Y)$ explicitly.
\begin{proposition}
\label{hilb-lines-explicit}
Let $Y$ be a \textup(smooth\textup) Fano threefold with $\rho(Y) = 1$, $\iota(Y) = 2$, and~\mbox{$\operatorname{d}(Y)\geqslant 3$}.
Then $\Sigma(Y)$ is smooth and irreducible. Moreover
\begin{itemize}
\item[(i)] if $\operatorname{d}(Y)=3$, then $\Sigma(Y)$ is a minimal surface of general type
with irregularity $5$, geometric genus $10$, and canonical degree~\mbox{$K_{\Sigma(Y)}^2=45$};
\item[(ii)] if $\operatorname{d}(Y)=4$, then $\Sigma(Y)$ is an abelian surface;
\item[(iii)] if $\operatorname{d}(Y)=5$, then $\Sigma(Y)\cong\mathbb{P}^2$.
\end{itemize}
\end{proposition}
\begin{proof}
If $\operatorname{d}(Y)=3$, then $Y$ is a cubic hypersurface in $\mathbb{P}^4$,
and the assertion holds by~\mbox{\cite[\S1]{Altman1977}}.
If $\operatorname{d}(Y)=4$, then $Y$ is a complete intersection of two quadrics in~$\mathbb{P}^5$, and the assertion holds
by~\cite[Theorem~5]{NarasimhanRamanan1969} (see also \cite[Theorem 4.8]{Reid1972}, \cite[Theorem~2]{DesaleRamanan}, \cite[\S6.3]{Griffiths-Harris-1978}).
If $\operatorname{d}(Y)=5$, then $Y$ is isomorphic
to a linear section $\mathrm{Gr}(2,5)\cap\mathbb{P}^6$
of the Grassmannian $\mathrm{Gr}(2,5)\subset\mathbb{P}^9$,
and the assertion holds by~\mbox{\cite[Proposition~III.1.6]{Iskovskikh-1980-Anticanonical}}
or \cite{Furushima1989a}
(see also~\S\ref{subsection:v5} for an explicit description of lines and
Proposition~\ref{proposition-v22-v5} for an alternative approach).
\end{proof}
\begin{remark}
\label{remark:hyperelliptic-curve}
The abelian surface $\Sigma(Y)$ associated to a Fano threefold $Y$ with $\rho(Y) = 1$, $\iota(Y) = 2$, and~\mbox{$\operatorname{d}(Y) = 4$} can be described as follows. Recall that such $Y$
is an intersection of two quadrics in $\mathbb{P}^5$.
The corresponding pencil contains precisely 6 degenerate quadrics (\cite[Proposition~2.1]{Reid1972}),
so one can consider the double cover $B(Y) \to \mathbb{P}^1$ branched in these six points.
This is a curve of genus 2. It can be regarded as a curve parameterizing the families of
planes in the quadrics of our pencil.
One can show that $\Sigma(Y)$ is isomorphic to the Jacobian of the curve~$B(Y)$,
see~\cite[Theorem~5]{NarasimhanRamanan1969}.
Moreover, the surface~$\Sigma(Y)$ is isomorphic to
the intermediate Jacobian of~$Y$ (as an abstract variety), see \cite[\S6.4]{Griffiths-Harris-1978}.
\end{remark}
\subsection{Hilbert schemes of conics}
\label{subsection:conics-index-1}
In this section we restrict to the case of smooth Fano threefolds $X$ with $\rho(X)=1$ and
$\iota(X)=1$ and their Hilbert schemes of conics $S(X)$.
Let~$S_0$ be an irreducible component of $S(X)$, and consider the reduced scheme
structure on~$S_0$.
Restricting to $S_0$ the universal family of conics, we obtain a diagram
\begin{equation}\label{diagram-universal-conic}
\vcenter{
\xymatrix{
& \mathscr{C}_0(X) \ar[dl]_q \ar[dr]^p \\
S_0 && X
}
}
\end{equation}
The map $q\colon\mathscr{C}_0(X) \to S_0$ is a conic bundle.
Let $C \subset X$ be a conic corresponding to a point $[C]$ in the component $S_0$ of the Hilbert scheme.
The fiber $q^{-1}([C])$ is identified by the map $p$ with the conic $C$. The normal bundle of $C$ in $\mathscr{C}_0(X)$
is the trivial bundle of rank equal to the dimension of the tangent space to $[C]$ at $S_0$.
Like in~\eqref{equation-dp-lines},
the differential
of $p$ is the map
\begin{equation}\label{equation-dp-conics}
dp \colon
{\mathscr{N}}_{C/\mathscr{C}_0(X)} = T_{[C]} S_0 \otimes {\mathscr{O}}_C \hookrightarrow
T_{[C]} S(X) \otimes {\mathscr{O}}_C =
H^0(C,{\mathscr{N}}_{C/X}) \otimes {\mathscr{O}}_C \to {\mathscr{N}}_{C/X}
\end{equation}
with the last map being given by evaluation.
We will call an irreducible component of $S(X)$ \emph{exotic}
if it does not contain smooth conics.
The next lemma shows that exotic components appear only for quartics with cones and describes them explicitly.
\begin{lemma}\label{lemma:exotic-component}
Let $X$ be a Fano threefold with $\rho(X)=1$ and $\iota(X)=1$,
and suppose that~\mbox{$-K_X$} is very ample.
Let $S_0 \subset S(X)$ be an exotic component.
Then $X$ is a quartic with a cone, and $S_0 \cong \mathop{\mathsf{Hilb}}\nolimits^2(B)$, where $B$ is a smooth curve, which is the base of the cone.
In particular, one has~\mbox{$\dim S_0 = 2$}.
Moreover, the irreducible component of $S(X)$ underlying $S_0$ is everywhere non-reduced.
\end{lemma}
\begin{proof}
By Corollary~\xref{proposition:dimensions-hilbert-schemes} we have $\dim S_0 \geqslant 2$.
On the other hand, by Lemma~\xref{lemma-hilb-x-1} every irreducible component of $\Sigma(X)$ is one-dimensional.
Since a line $L$ on $X$ admits at most one structure of a non-reduced conic (see Remark~\xref{remark:dl-special}),
it follows that a conic corresponding to a general point of~$S_0$ is a union of two distinct lines.
Since $X$ does not contain projections of the Veronese surface by Lemma~\xref{lemma:cones}(i)
(in particular, $X$ does not contain smooth quadric surfaces),
we deduce from Lemma~\xref{lemma:many-reducible-conics} that~$X$ contains a two-dimensional cone with base $B$
such that $S_0$ is the set of conics formed by unions of rulings of the cone.
In other words, one has
\begin{equation*}
S_0 = \{ L_{b_1} \cup L_{b_2} \mid (b_1,b_2) \in \mathrm{Sym}^2(B) \}.
\end{equation*}
Moreover, $X$ is a quartic threefold and $B$ is a smooth curve by Lemma~\xref{lemma:cones}(ii).
In particular, one has $\mathrm{Sym}^2(B) \cong \mathop{\mathsf{Hilb}}\nolimits^2(B)$.
As we already mentioned in Remark~\ref{remark:nonreduced-sigma} the component of $\Sigma(X)$ underlying $B$ is everywhere non-reduced.
A similar argument shows that the component of $S(X)$ underlying~$\mathop{\mathsf{Hilb}}\nolimits^2(B)$ is also everywhere non-reduced.
Indeed, if $C = L_{b_1} \cap L_{b_2}$ spans a plane $\Pi$, the corresponding double plane provides $C$ with a non-reduced structure
corresponding to a surjective map ${\mathscr{N}}^\vee_{C/X} \to {\mathscr{O}}_X(-1)\vert_C$.
By duality this gives an embedding ${\mathscr{O}}_X(1)\vert_C \to {\mathscr{N}}_{C/X}$, hence $\dim H^0(C,{\mathscr{N}}_{C/X}) \geqslant \dim H^0(C,{\mathscr{O}}_X(1)\vert_C) = 3$.
\end{proof}
\begin{lemma}\label{lemma-hilb-x-2}
If $X$ is a Fano threefold with $\rho(X)=1$, $\iota(X)=1$ and $-K_X$ very ample then
every irreducible component $S_0$ of the Hilbert scheme $S(X)$ of conics on $X$ is two-dimensional.
If $S_0$ is not exotic, then the map $p\colon\mathscr{C}_0(X) \to X$ is surjective, generically finite,
does not contract divisors, and is not birational;
moreover, the natural scheme structure on $S_0$ is generically reduced.
If $S_0$ is exotic, then~$X$ is a quartic and~\mbox{$p(\mathscr{C}_0(X))$} is a cone over a
smooth curve.
\end{lemma}
\begin{proof}
First, we note that $p$ is surjective unless $S_0$ is exotic.
Indeed, if the image of~$p$ is a surface $Z \subset X$ then $\dim S(Z) \geqslant \dim S_0 \geqslant 2$ by Corollary~\xref{proposition:dimensions-hilbert-schemes},
hence by Lemma~\xref{lemma:lines-conics-plane} the surface~$Z$
is a linear projection of the Veronese surface, which contradicts Lemma~\xref{lemma:cones}(i),
or~$Z$ is a cone. In the latter case clearly $S_0$ is an exotic component.
Assume that $\dim S_0=k \geqslant 3$.
By Lemma~\xref{lemma:exotic-component} the component $S_0$ is not exotic, and by Corollary~\xref{corollary:hilbert-smoothness}
a general point of~$S_0$ corresponds to a smooth special (and even 2-special) conic.
The differential $dp$ on such a conic has rank at most 1 everywhere,
therefore the fibers of the map $p$ have dimension~\mbox{$k-1$}, hence the image $p(\mathscr{C}_0(X))$ has dimension~\mbox{$k + 1 - (k-1) = 2$}.
In particular, $p$ is not surjective, which contradicts the above conclusions.
Thus we have $\dim S_0 = 2$.
Assume that $S_0$ is not exotic and $\dim S_0 = 2$.
Then $p$ is surjective, and since $\dim\mathscr{C}_0(X) =3= \dim X$, the morphism~$p$ is generically finite.
Consider the ramification locus $R(p)$ of the map~$p$.
Let $C$ be a conic corresponding to a smooth point of $S_0$.
Then both~${\mathscr{N}}_{C/\mathscr{C}_0(X)}$ and~${\mathscr{N}}_{C/X}$ are vector bundles on $C$ of rank 2 and Euler characteristic~2,
see Lemma~\xref{lemma:normal-bundles} and Corollaries~\xref{corollary:normal-reducible-conic-3fold}, and~\xref{corollary:normal-nonreduced-conic-3fold}.
Hence the kernel and the cokernel of the map $dp$ have the same rank and Euler characteristic. If the rank
of the cokernel is $0$, then so is the rank of the kernel. Since ${\mathscr{N}}_{C/\mathscr{C}_0(X)}$ is a trivial vector bundle,
it is torsion free, so it follows that the kernel is zero. But then the cokernel is zero as well.
This means that either the cokernel of the map $dp$ is zero, hence the map $p$ is unramified along $C$,
or the support of the cokernel is either $C$, or if $C = L_1 \cup L_2$ is reducible, one of the lines $L_i$.
This shows that away of the $q$-preimage of the singular locus of $S(X)$
the ramification locus $R(p)$ is the union of (irreducible components of) fibers of~$q$.
Arguing as in Lemma~\xref{lemma-hilb-y-1} (with obvious modifications), we conclude that $p$ cannot contract divisors and cannot be birational.
The above arguments also show that a generic point of $S_0$ corresponds to a smooth ordinary conic $C$.
Therefore, the tangent space to $S(X)$ at $C$ is 2-dimensional, hence~$S(X)$ is generically reduced along $S_0$.
Finally, if $S_0$ is an exotic component, a description of Lemma~\xref{lemma:exotic-component} shows that $X$ is a quartic,
and $p(\mathscr{C}_0(X))$ is a cone over a smooth curve~$B$.
\end{proof}
Our next goal, as before, is a proof of smoothness of $S(X)$ and its explicit description for some $X$.
The direct proof of smoothness is very complicated, since there are three types of conics
and it is much more difficult to analyze the tangent space to the Hilbert scheme
at a point corresponding to a reducible or non-reduced conic, and the corresponding obstruction
space (see Appendix~\xref{subsection:degenerate-conics}, or~\cite[\S3.2]{iliev2011fano}).
Typically such considerations work only for general Fano threefolds.
So, instead of using the above straightforward approach, we use the ideas of~\cite{kuznetsov2009derived},
where it was argued that the geometry of Fano threefolds of index 1 and even genus $\operatorname{g}$ is related to the geometry
of Fano threefolds of index 2 and degree $\operatorname{d} = \operatorname{g}/2 - 1$. The reason for this is a similarity between the structure
of their derived categories. Using this idea we will prove the following result.
\begin{theorem}\label{theorem:S-vs-Sigma}
Let $X$ be a \textup(smooth\textup) Fano threefold with $\rho(X) = 1$, $\iota(X) = 1$, and
\begin{equation*}
\operatorname{g}(X) \in \{8,10,12\}.
\end{equation*}
Then there is a smooth Fano threefold $Y$ with $\rho(Y) = 1$, $\iota(Y) = 2$, and
\begin{equation*}
\operatorname{d}(Y) = \frac{\operatorname{g}(X)}{2}-1
\end{equation*}
such that $S(X) \cong \Sigma(Y)$.
\end{theorem}
Since the proof of this result uses a completely different technique, we moved it to Appendix~\xref{section-Lines-and-conics}.
Actually, for $\operatorname{g}(X) = 10$ and $\operatorname{g}(X) = 12$ we identify the Hilbert schemes explicitly and show that these identifications match up,
while
for~\mbox{$\operatorname{g}(X)=8$} we construct a direct isomorphism of $S(X)$ and $\Sigma(Y)$.
Combining Theorem~\xref{theorem:S-vs-Sigma} with a description of Hilbert schemes of lines of index 2 threefolds and
with some other results, we can state now the following proposition.
Recall that a vector bundle ${\mathscr{E}}$ is called \emph{simple}, if $\mathop{\mathsf{Hom}}\nolimits({\mathscr{E}},{\mathscr{E}}) = \Bbbk$.
\begin{proposition}\label{proposition:conics}
Let $X$ be a \textup(smooth\textup) Fano threefold of index $1$ and
genus $\operatorname{g}(X)\geqslant 7$.
Then $S(X)$ is a smooth irreducible surface and
\begin{itemize}
\item[(i)] if $\operatorname{g}(X)=7$, then $S(X)$ is the symmetric square of a
smooth curve of genus $7$;
\item[(ii)] if $\operatorname{g}(X)=8$, then $S(X)$ is a minimal surface of general type
with irregularity $5$, geometric genus $10$, and canonical
degree~\mbox{$K_{S(X)}^2=45$};
\item[(iii)] if $\operatorname{g}(X)=9$, then $S(X)$ is a ruled surface
that is a projectivization of a simple rank~$2$ vector bundle on a smooth curve of genus $3$;
\item[(iv)] if $\operatorname{g}(X)=10$, then $S(X)$ is an abelian surface;
\item[(v)] if $\operatorname{g}(X)=12$, then $S(X)\cong\mathbb{P}^2$.
\end{itemize}
\end{proposition}
\begin{proof}
If $\operatorname{g}(X)=7$, the assertion holds by~\cite[Theorem~6.3]{Kuznetsov-V12}.
If $\operatorname{g}(X)=8$, the assertion holds by Theorem~\xref{theorem:S-vs-Sigma} and Proposition~\xref{hilb-lines-explicit}(i).
If $\operatorname{g}(X)=9$, the surface $S(X)$ is ruled by~\cite[Proposition~3.10]{BrambillaFaenzi},
and the simplicity of the corresponding vector bundle is proved in Lemma~\xref{lemma:v-simple}.
If $\operatorname{g}(X)=10$, the assertion holds by Proposition~\ref{proposition:S-Pic-iso}, or equivalently by
Theorem~\xref{theorem:S-vs-Sigma} and Proposition~\xref{hilb-lines-explicit}(ii).
If $\operatorname{g}(X)=12$, the assertion holds by~\mbox{\cite[Theorem~2.4]{Kollar2004b}}
(alternatively, one can apply Proposition~\ref{proposition-v22-v5}, or equivalently
Theorem~\xref{theorem:S-vs-Sigma} and Proposition~\xref{hilb-lines-explicit}(iii)).
Smoothness of~$S(X)$ is clear from the above case-by-case analysis.
\end{proof}
Propositions~\xref{hilb-lines-explicit} and~\xref{proposition:conics}
together give Theorem~\xref{theo:main1}.
\begin{remark}
\label{remark:hilb-conics-reducible}
Note that if $X$ is a Fano
threefold of genus~\mbox{$\operatorname{g}(X)\leqslant 6$}, then
the surface~$S(X)$ may
be singular and even reducible.
For example, let $\pi\colon X\to Y$ be a double cover of a smooth Fano variety $Y$
with $\rho(Y)=1$ and $\iota(Y)=2$ branched in a smooth anticanonical divisor.
Then $X$ is a smooth Fano threefold with $\rho(X) = 1$, $\iota(X) = 1$, and $\operatorname{g}(X) = \operatorname{d}(Y) + 1$ (see Lemma~\xref{lemma:trivial-action-conics} below), and
$S(X)$ is a union of two irreducible components; one of them is identified with
the Hilbert scheme of lines in~$Y$,
and the other is a double cover of the subvariety of the Hilbert scheme of conics $S(Y)$ bitangent
to the branch divisor (see \cite[Proposition 2.1.2]{Iliev1994a} for the case~\mbox{$\operatorname{g}(X)=6$}).
\end{remark}
For the sake of completeness, we conclude this section by a discussion of Hilbert
schemes of conics on some Fano threefolds of Picard rank $1$ and index~$2$.
These results, of course, are well known to experts, however, we do not know a good reference for them
except for the case~\mbox{$\operatorname{d}(Y)=5$}.
According to our conventions, we consider Hilbert schemes of conics only on those Fano threefolds whose ample generator of the Picard group is very ample.
In the case of index 2 this means that~\mbox{$\operatorname{d}(Y) \geqslant 3$}.
Recall the description of the Hilbert scheme of lines~$\Sigma(Y)$ for these threefolds from Proposition~\xref{hilb-lines-explicit}.
Also, recall that there is a curve $B(Y)$ of genus 2 associated to a Fano threefold $Y$ of index 2 and degree 4,
see Remark~\xref{remark:hyperelliptic-curve}.
\begin{proposition}
Let $Y$ be a \textup(smooth\textup) Fano threefold with $\rho(Y)=1$, $\iota(Y)=2$, and~\mbox{$\operatorname{d}(Y) \geqslant 3$}.
Then $S(Y)$ is a smooth fourfold, and
\begin{itemize}
\item[(i)] if $\operatorname{d}(Y)=3$, then $S(Y)$ is a ${\mathbb{P}}^2$-bundle over the surface $\Sigma(Y)$.
\item[(ii)] if $\operatorname{d}(Y)=4$, then $S(Y)$ is a ${\mathbb{P}}^3$-bundle over the curve $B(Y)$.
\item[(iii)] if $\operatorname{d}(Y)=5$, then $S(Y)\cong\mathbb{P}^4$.
\end{itemize}
\end{proposition}
\begin{proof}
Let us start with assertion~(i).
Let $Y \subset \mathbb{P}^4$ be a smooth cubic hypersurface.
The linear span of a conic $C \subset Y$ is a plane~\mbox{$\langle C \rangle \cong \mathbb{P}^2$}.
This plane is not contained in $Y$, because the Picard group of $Y$ is generated by a hyperplane section by Lefschetz theorem.
Hence the intersection~\mbox{$\langle C \rangle \cap Y$} is a plane cubic curve, containing the conic $C$.
This means that
\begin{equation*}
\langle C \rangle \cap Y = C \cup L(C),
\end{equation*}
where $L(C)$ is a line (usually called the residual line of $C$).
It is easy to see that the association $C \mapsto L(C)$ defines a regular map $S(Y) \to \Sigma(Y)$.
The fiber of the map over a point $[L] \in \Sigma(Y)$ is the space of all planes in $\mathbb{P}^4$ containing~$L$ (hence is isomorphic to $\mathbb{P}^2$).
Indeed, if $\Pi$ is such a plane then $\Pi \cap Y = L \cup C(\Pi)$ with~$C(\Pi)$ a conic, and conversely, every conic whose residual line is $L$ spans a plane containing $L$.
Altogether, this shows that
$$
S(Y) \cong \Sigma(Y) \times_{\mathrm{Gr}(2,5)} \mathrm{Fl}(2,3;5),
$$
where $\mathrm{Fl}(2,3;5)$ is the flag variety. Thus, we proved assertion~(i).
The idea of the proof of assertion~(ii) is similar to that for assertion~(i).
Let $Y \subset \mathbb{P}^5$ be a complete intersection of two quadrics.
Given a conic~\mbox{$C \subset Y$}, we consider its linear span~\mbox{$\langle C \rangle \subset \mathbb{P}^5$}.
The restriction to $\langle C \rangle$ of the pencil of quadrics defining $Y$ is a pencil of conics containing $C$.
This means that there is a unique quadric $Q(C)$ in the pencil defining~$Y$ that contains $\langle C \rangle$
(again, because the plane $\langle C \rangle \cong \mathbb{P}^2$ is not contained in $Y$).
In other words, the association
$C \mapsto (Q(C),\langle C\rangle)$
defines a regular map
\begin{equation*}
S(Y) \to \mathop{\mathsf{Hilb}}\nolimits^{p(t) = (1+t)(2+t)/2}({\mathscr{Q}}/\mathbb{P}^1)
\end{equation*}
into the relative Hilbert scheme of
planes in the divisor~\mbox{${\mathscr{Q}} \subset \mathbb{P}^5\times\mathbb{P}^1$}
(defined by the pencil of quadrics) over $\mathbb{P}^1$.
For a smooth quadric $Q \subset \mathbb{P}^5$ the Hilbert scheme of planes~\mbox{$\mathop{\mathsf{Hilb}}\nolimits^{p(t) = (1+t)(2+t)/2}(Q)$} is isomorphic to a union of two copies of~$\mathbb{P}^3$,
while for a cone over a smooth quadric in $\mathbb{P}^4$, it is isomorphic to~$\mathbb{P}^3$.
Altogether, this means that the Stein factorization for the canonical map
$\mathop{\mathsf{Hilb}}\nolimits^{p(t) = (1+t)(2+t)/2}({\mathscr{Q}}/\mathbb{P}^1) \to \mathbb{P}^1$ is a composition of a $\mathbb{P}^3$-bundle with a double cover
$B(Y) \to \mathbb{P}^1$, branched in the points of $\mathbb{P}^1$ corresponding to singular quadrics in the pencil.
This proves assertion~(ii).
For assertion~(iii) see \cite[Proposition~2.32]{Sanna} or~\cite[Proposition~1.2.2]{Iliev1994a}.
\end{proof}
\section{Automorphism groups}
\label{section:automorphisms}
In this section we remind some general results on automorphism groups of projective varieties,
in particular showing that under appropriate conditions they are linear algebraic groups.
We also discuss some general approaches to finiteness of automorphism groups.
Throughout the section we work
under rather general assumptions.
\subsection{Actions on linear systems}
\label{subsection:actions}
Let $X$ be a normal projective variety and let $A$ be a Weil divisor on $X$.
If the linear system $|A|$ is not empty, we denote by
\begin{equation*}
\varphi_{|A|}\colon X \dashrightarrow \mathbb{P}\big(H^0(X, {\mathscr{O}}_X(A))^{\vee}\big)
\end{equation*}
the corresponding rational map.
If the class $[A]$ of $A$ in $\mathrm{Cl}(X)$ is invariant with respect to a subgroup $\Gamma \subset \mathrm{Aut}(X)$,
then there is a natural action of $\Gamma$ on ${\mathbb{P}}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$ and the map $\varphi_{|A|}$ is $\Gamma$-equivariant.
Note also that the $\Gamma$-action on ${\mathbb{P}}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$ is induced by an action on $H^0(X,{\mathscr{O}}_X(A))^\vee$ of a central extension
\begin{equation*}
1 \to \boldsymbol{\mu}_m \to \widetilde\Gamma \to \Gamma \to 1,
\end{equation*}
where $m = \dim H^0(X,{\mathscr{O}}_X(A))$. Indeed, the above exact sequence is the pullback of
\begin{equation*}
1 \to \boldsymbol{\mu}_m \to \mathrm{SL}(H^0(X, {\mathscr{O}}_X(A))^{\vee}) \to \mathrm{PGL}(H^0(X, {\mathscr{O}}_X(A))^{\vee}) \to 1
\end{equation*}
via the map $\Gamma \to \mathrm{Aut}({\mathbb{P}}(H^0(X, {\mathscr{O}}_X(A))^{\vee})) \cong \mathrm{PGL}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$.
The induced map~\mbox{$\widetilde\Gamma \to \mathrm{SL}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$} gives a $\widetilde\Gamma$-action on $H^0(X, {\mathscr{O}}_X(A))^{\vee}$.
\begin{remark}\label{remark:need-extension}
Note that the action of $\Gamma$ on ${\mathbb{P}}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$ may not be induced by the action on $H^0(X, {\mathscr{O}}_X(A))^\vee$
of $\Gamma$ itself, i.\,e. passing to a central extension above is indeed necessary.
On the other hand, if the sheaf ${\mathscr{O}}_X(A)$ admits a $\Gamma$-linearization
(that is, if the action of $\Gamma$ on $X$ lifts to its action on~${\mathscr{O}}_X(A)$),
then the map~\mbox{$\Gamma \to \mathrm{PGL}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$} lifts to a map~\mbox{$\Gamma \to \mathrm{GL}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$}.
\end{remark}
The following lemma is easy and well known.
\begin{lemma}\label{lemma:anticanonical-ring0}
Let $X$ be a normal projective variety and $A$ be a Weil divisor on $X$.
Let $\mathrm{Aut}(X; [A])$ be the stabilizer of the class $[A]\in\mathrm{Cl}(X)$ in $\mathrm{Aut}(X)$.
If the map $\varphi_{|A|}$
is birational onto its image then the action of $\mathrm{Aut}(X; [A])$ on ${\mathbb{P}}(H^0(X, {\mathscr{O}}_X(A))^{\vee})$ is faithful.
In particular, in this case $\mathrm{Aut}(X; [A])$ is a linear algebraic group.
\end{lemma}
\begin{proof}
If some element $g \in\mathrm{Aut}(X; [A])$ acts trivially on ${\mathbb{P}}((H^0(X, {\mathscr{O}}_X(A))^{\vee}))$, then by assumption it also acts trivially
on an open dense subset of $X$, hence on the whole $X$.
\end{proof}
Note that any multiple of the canonical class is invariant under the automorphism group~\mbox{$\mathrm{Aut}(X)$}
and, moreover, has a natural $\mathrm{Aut}(X)$-linearization. Applying Lemma~\xref{lemma:anticanonical-ring0}
and taking into account Remark~\xref{remark:need-extension},
we obtain the following result.
\begin{corollary}\label{corollary:anticanonical-ring}
Let $X$ be a normal projective variety.
Suppose that for some $m\in \mathbb{Z}$ \textup(either positive or negative\textup)
the map~\mbox{$\varphi_{|mK_X|}$} is birational onto its image. Then the action of the group $\mathrm{Aut}(X)$ on
$\mathbb{P}\big(H^0(X,\mathscr{O}_X(mK_X))^{\vee}\big)$
is faithful and lifts to an embedding
\begin{equation*}
\mathrm{Aut}(X) \hookrightarrow \mathrm{GL}(H^0(X,{\mathscr{O}}_X(mK_X))^\vee).
\end{equation*}
In particular, $\mathrm{Aut}(X)$ is a linear algebraic group.
\end{corollary}
\begin{corollary}\label{corollary:hypersurface}
Let $X\subset\mathbb{P}^N$ be a normal complete intersection
of dimension~\mbox{$\dim X\geqslant 3$} that is not contained in a hyperplane in~$\mathbb{P}^N$.
Then there is a natural embedding~\mbox{$\mathrm{Aut}(X)\hookrightarrow\mathrm{PGL}_{N+1}(\Bbbk)$}.
\end{corollary}
\begin{proof}
By Lefschetz theorem one has $\mathrm{Pic}(X)=\mathbb{Z}\cdot H$, where $H$ is a hyperplane
section (see e.\,g. \cite[Corollary~IV.3.2]{Hartshorne1970}).
Thus, the embedding $X \hookrightarrow \mathbb{P}^N$ is given by an invariant linear system
$|H|$, so the assertion follows from Lemma~\xref{lemma:anticanonical-ring0}.
\end{proof}
\subsection{Finiteness results}
\label{subsection:known-finiteness}
Let us recall several easy finiteness results for
automorphism groups of algebraic varieties.
\begin{lemma}
\label{lemma:non-uniruled}
If a linear algebraic group $G$ acts faithfully on a variety $X$ which is not ruled, then $G$ is finite.
\end{lemma}
\begin{proof}
If $G$ is not finite, it contains a subgroup isomorphic to $\mathbb G_{\mathrm{m}}$ or $\mathbb G_{\mathrm{a}}$.
An open subset of $X$ is covered by one-dimensional orbits of this subgroup, hence $X$ is ruled, which is a contradiction.
\end{proof}
\begin{corollary}\label{corollary:nef-K}
Let $X$ be a variety of Kodaira dimension $\kappa(X)\geqslant 0$.
Suppose that a linear algebraic group $G$ acts faithfully on $X$.
Then $G$ is finite.
\end{corollary}
\begin{proof}
Since the linear system $|nK_X|$ is not empty for some $n>0$,
the variety $X$ is not uniruled (see \cite[Theorem 1]{MiyaokaMori}).
Thus we can apply Lemma~\xref{lemma:non-uniruled}.
\end{proof}
\begin{corollary}\label{corollary:general-type}
Let $X$ be a variety of general type. Then the group $\mathrm{Aut}(X)$ is finite.
\end{corollary}
\begin{proof}
Apply Corollaries~\xref{corollary:anticanonical-ring}
and~\xref{corollary:nef-K}.
\end{proof}
\begin{remark}
Actually, even the group of birational selfmaps of a variety of
general type is finite, since it coincides with the automorphism group of its canonical model.
\end{remark}
Another collection of finiteness results concerns hypersurfaces and complete intersections.
\begin{theorem}[{see \cite{MatsumuraMonsky}}]
\label{theorem:MatsumuraMonsky}
Let $X$ be a smooth hypersurface of degree $d\geqslant 3$ in $\mathbb{P}^N$, where
$N\geqslant 2$. Then the automorphism group of $X$ is finite
unless either $N=2$ and $d=3$, or $N=3$ and $d=4$.
\end{theorem}
There are many classification results
on automorphism groups of hypersurfaces of small degree,
in particular, cubic hypersurfaces (see \cite{Hosoh97},
\cite{Hosoh02}, \cite[\S9.5]{Dolgachev-book},
\cite{Adler78}, \cite{GAL11}, \cite{OY15}).
Also, Theorem~\xref{theorem:MatsumuraMonsky} has the following
recent generalization.
\begin{theorem}[{\cite[Theorem~3.1]{Benoist2013}}, see also \cite{ChenPanZhang-2015}]
\label{theorem:Benoist}
If $X \subset {\mathbb{P}}^N$ is a smooth complete intersection of
dimension $\dim X\geqslant 3$
and codimension $\operatorname{codim}(X) \geqslant 2$ not contained in a hyperplane
in~$\mathbb{P}^N$, then $\mathrm{Aut}(X)$ is finite.
\end{theorem}
Another well-known finiteness result that we will need is as follows.
Recall that for any morphism $\phi\colon Y\to X$ there is a subgroup
$\mathrm{Aut}(Y/X)\subset \mathrm{Aut}(Y)$ that consists of all automorphisms whose action is fiberwise with respect to $\phi$;
we will refer to this group as \emph{the group of automorphisms of $Y$ over $X$}.
\begin{lemma}
\label{lemma:projectivization-stable-bundle}
Let ${\mathscr{E}}$ be a simple vector bundle on a projective scheme~$X$. Then the group~\mbox{$\mathrm{Aut}({\mathbb{P}}_X({\mathscr{E}})/X)$}
of the automorphisms of the projectivization ${\mathbb{P}}_X({\mathscr{E}})$ over $X$ is finite.
\end{lemma}
\begin{proof}
This is Corollary to Proposition~2
in~\cite{Grothendieck1958geometrie}
(note also that the group denoted by~$\Gamma$ in~\cite{Grothendieck1958geometrie} is a subgroup in the $2$-torsion subgroup of $\mathrm{Pic}(X)$, hence is finite).
\end{proof}
\begin{corollary}
If ${\mathscr{E}}$ is a simple vector bundle on a smooth curve $C$ of genus $g > 1$, then the
group $\mathrm{Aut}({\mathbb{P}}_C({\mathscr{E}}))$ is finite.
\end{corollary}
\begin{proof}
Indeed, the morphism ${\mathbb{P}}_C({\mathscr{E}}) \to C$ is canonical, hence there is an exact sequence
\begin{equation*}
1 \to \mathrm{Aut}({\mathbb{P}}_C({\mathscr{E}})/C) \to \mathrm{Aut}({\mathbb{P}}_C({\mathscr{E}})) \to \mathrm{Aut}(C).
\end{equation*}
The term on the left is finite by Lemma~\xref{lemma:projectivization-stable-bundle}, and the term on the right is finite since~\mbox{$g>1$}, see Corollary~\xref{corollary:general-type}.
Therefore $\mathrm{Aut}({\mathbb{P}}_C({\mathscr{E}}))$ is finite.
\end{proof}
\section{Finiteness for Fano threefolds}
\label{section:finiteness-for-Fanos}
In this section we prove finiteness of automorphism groups for most of smooth Fano threefolds of Picard rank 1.
\subsection{Faithfulness of action on a family of curves}
\label{subsection:faithfulness-general}
In this subsection we prove a general result on faithfulness of an automorphism group action on a Hilbert scheme of
curves of degree 2 with respect to the anticanonical class.
In the next subsections we apply it to Hilbert schemes of lines on Fano threefolds of index 2 and Hilbert schemes of conics on Fano threefolds of index 1.
Let $X$ be a smooth projective variety (of any dimension).
Let $S$ be an irreducible and reduced projective subscheme in a Hilbert scheme of curves on $X$, let
$\mathscr{C} \subset S \times X$ be the corresponding family of curves, and let
\begin{equation*}
\xymatrix{
& \mathscr{C} \ar[dl]_q \ar[dr]^p \\
S && X
}
\end{equation*}
be the corresponding diagram of projections.
\begin{theorem}\label{theorem:faithful}
Assume that $X$ is a smooth Fano variety of any dimension greater than~$1$ with~\mbox{$\mathrm{Pic}(X)\cong\mathbb{Z}$} and~$-K_X$ very ample.
Assume that
\begin{itemize}
\item
for general $s \in S$ the fiber $\mathscr{C}_s$ of $q$ is a smooth rational curve and $\mathscr{C}_s\cdot(-K_X) = 2$;
\item
the morphism
$p$ is dominant, not birational, and does not contract divisors.
\end{itemize}
Let $G \subset \mathrm{Aut}(X)$ be a nontrivial algebraic
subgroup that acts trivially on $S$.
Then the group $G$ is cyclic of order~$2$,
and a curve $C$ corresponding to a general point of $S$ is the preimage of a curve $C' \subset X' = X/G$ under the quotient map $\pi \colon X \to X'$.
Furthermore, if $\dim X$ is odd then $X'$ is smooth, and
if $\dim X$ is even, then $X'$ is either smooth or has one singular point of type $\frac12(1,\ldots,1)$.
In both cases the branch locus of $\pi$ in $X'$ is the union of a smooth anticanonical divisor $B$ and $\mathrm{Sing}\,(X')$.
Finally, $X'$ is a Fano variety with $\mathrm{Pic}(X') \cong \mathbb{Z}$, and the divisor $K_{X'}$ is divisible by $2$ in
the class group~\mbox{$\mathrm{Cl}(X')$}.
\end{theorem}
\begin{proof}
First let us show that $G$ is finite.
Since $G$ is an algebraic group, it is enough to show that the connected component $G^0$ of identity in $G$ is trivial.
Since $\mathscr{C}_s$ is smooth for general $s \in S$ and $p$ is generically finite, a general point $x \in X$ does not lie on a reducible curve from the family.
Since $p$ is dominant and not birational, the fiber $p^{-1}(x)$ over a general point $x \in X$ consists of more than one point.
Thus for a general $x \in X$ there are two distinct irreducible curves $C_1$ and $C_2$ in the family $\mathscr{C}$ that pass through~$x$.
Since the action of~$G^0$ on $S$ is trivial, both curves $C_i$ are $G^0$-invariant.
Hence~\mbox{$G^0\cdot x \subset C_1 \cap C_2$}.
Since~$C_1$ and~$C_2$ are distinct and irreducible, it follows that $G^0\cdot x$ is finite.
But $G^0$ is connected, hence~\mbox{$G^0\cdot x = x$}.
Thus, a general point of $X$ is fixed by $G^0 \subset \mathrm{Aut}(X)$, hence~$G^0$ is trivial.
This means that the group~$G$ is finite.
Now let $G_0 \subset G$ be a cyclic subgroup of order~\mbox{$n>1$}.
The action of~$G_0$ on~$\mathscr{C}$ is fiberwise over~$S$.
Therefore, it has two fixed points on a general smooth
fiber of $q$, so
the fixed locus of $G_0$ in $\mathscr{C}$ contains a divisor which intersects a general smooth fiber of $q$ at two distinct points.
Since the morphism $p \colon \mathscr{C} \to X$ contracts no divisors, the fixed locus of~$G_0$ in~$X$ contains a divisor $F$ which intersects a general smooth curve from $\mathscr{C}$ at two points.
Since~$\mathrm{Pic} (X)\cong \mathbb{Z}$, this means that $F \sim -K_X$. Put
\begin{equation*}
V = H^0(X,{\mathscr{O}}(-K_X))^\vee,
\end{equation*}
so that $X \subset \mathbb{P}(V)$ is the anticanonical embedding.
The action of $G_0$ on $X$ induces an action on $V$ by Corollary~\xref{corollary:anticanonical-ring}.
The fixed divisor $F$ generates a hyperplane $V_0 \subset V$ and we have a direct sum decomposition
\begin{equation*}
V = V_0 \oplus V_1,
\end{equation*}
where $V_0$ and $V_1$ are eigenspaces for (a generator of) $G_0\cong\boldsymbol{\mu}_n$, and
$V_1$ is one-dimensional.
It follows that the fixed locus of $G_0$ on $\mathbb{P}(V)$ is $\mathbb{P}(V_0) \sqcup \mathbb{P}(V_1)$, and its fixed locus on $X$
is either $F = X \cap \mathbb{P}(V_0)$, or the union of $F$ with the point $P=\mathbb{P}(V_1) \in \mathbb{P}(V)$ corresponding to the one-dimensional
eigenspace $V_1 \subset V$ (if the point $P$ lies on $X$).
Let $X' = X/G_0$ be the quotient with $\pi\colon X \to X'$ being the projection.
If $P \in X$ then~\mbox{$P' = \pi(P)$} is a quotient singularity of type $\frac1n(1,\ldots,1)$ on $X'$ and $X'_0 = X' \setminus P'$ is smooth;
otherwise $X'$ is smooth and we set $X'_0 = X'$.
Put $X_0 = \pi^{-1}(X'_0)$ and $\pi_0 = \pi\vert_{X_0}$.
Since $\pi$ is a finite morphism, for any Weil divisor $R$ on $X'$ the pull-back $\pi^*R$ is a
well-defined $G_0$-invariant Weil divisor
(the closure of $\pi_0^{-1}(R\vert_{X'_0})$).
Furthermore, one has (see, e.g. \cite[1.7.5]{Fulton-Intersection-theory})
\begin{equation}
\label{equation-faithful-action-conics}
\mathrm{Cl}(X') = \mathrm{Cl}(X'_0) = \mathrm{Pic}(X'_0),\qquad \mathrm{Cl}(X')\otimes \mathbb{Q}=(\mathrm{Cl}(X)\otimes \mathbb{Q})^{G_0}.
\end{equation}
Since $\pi_0\colon X_0 \to X'_0$ is a cyclic degree $n$ cover with ramification divisor $F$,
the class of the branch divisor~\mbox{$B=\pi_0(F) \subset X'_0$} is divisible by $n$ in $\mathrm{Pic}(X'_0)$,
see e.g. \cite[Theorem~1.2]{Wavrik1968}, so that $B \sim nD$ for some~\mbox{$D \in \mathrm{Pic}(X'_0)$} with $F \sim \pi_0^*D$.
Let $C$ be a smooth curve corresponding to a general point of $\mathscr{C}$.
Then $C$ does not pass through $P$, since otherwise~\mbox{$p^{-1}(P) \subset \mathscr{C}$} would be a divisor contracted by $p$.
Thus~\mbox{$C = \pi_0^*(\pi_0(C))$} and hence
\begin{equation}
\label{equation-faithful-action-conics-LD}
2 = C \cdot (-K_X) = C\cdot F = \pi_0^*(\pi_0(C)) \cdot \pi_0^*(D) = n\pi_0(C)\cdot D,
\end{equation}
so $n$ divides 2. Since $n>1$ by our assumption, we have $n = 2$ and $G_0 \cong \boldsymbol{\mu}_2$.
Furthermore, since $n = 2$ by Hurwitz formula we have
\begin{equation*}
\pi_0^*K_{X'_0} \sim K_{X_0} - F \sim -2F.
\end{equation*}
Applying $\pi_{0*}$ we obtain
\begin{equation}
\label{eq:k-divisible}
K_{X'_0} \sim -2D.
\end{equation}
Thus the divisor $K_{X'}$ is divisible by~2 in $\mathrm{Cl}(X')$.
If $\dim X$ is odd and~\mbox{$P \in X$}, the image of $K_{X'}$ in the local class group
$\mathrm{Cl}(X',P') \cong \mathbb{Z}/2\mathbb{Z}$ is the generator.
This gives a contradiction with~\eqref{eq:k-divisible} and thus shows that $P \not\in X$ when~$\dim X$ is odd, and hence $X' = X'_0$ is smooth.
Since $\pi\colon X \to X'$ is a double cover and $\pi^*K_{X'} \sim 2K_X$,
it follows that $-K_{X'}$ is ample, i.e.~$X'$ is a Fano variety.
By \eqref{equation-faithful-action-conics} we have $\rho(X') = 1$.
Finally, it remains to show that $G = G_0 \cong \boldsymbol{\mu}_2$.
We have already shown that any
nontrivial element in $G_0$ has order $2$ and acts on $V$ as an involution with eigenspaces of dimension $1$ and \mbox{$\dim V - 1$}.
It follows that
$G\cong \boldsymbol{\mu}_2^{r}$ for some $r$, and it remains to show that~\mbox{$r=1$}.
Suppose that $r>1$. Take two different involutions $\tau_1,\tau_2\in G$.
The action of the abelian group~\mbox{$G\cong \boldsymbol{\mu}_2^{r}$} on $V = H^0(X,{\mathscr{O}}(-K_X))^\vee$ is diagonalizable.
Thus we may assume that the action of $\tau_1$ (respectively,~$\tau_2$) on $V$ is given by $\operatorname{\operatorname{diag}}(-1,1,\ldots,1)$
(respectively,~\mbox{$\operatorname{\operatorname{diag}}(1,-1,1,\ldots,1)$}).
Then the action of $\tau_1\circ \tau_2$ is given by $\operatorname{\operatorname{diag}}(-1,-1,1,\ldots,1)$.
If~$\dim V > 3$ the dimension of both eigenspaces is greater than 1, which contradicts the above observation.
On the other hand, the case $\dim V = 3$ is impossible, since then~$X \cong \mathbb{P}^2$ and has no curves of degree $2$ with respect to the anticanonical class.
\end{proof}
\subsection{Action on lines}
\label{subsection:act-on-lines}
We start with studying an action of the automorphism group of a smooth Fano threefold $Y$
of Picard rank $1$ and index~$2$ on its Hilbert scheme of lines.
Recall the notation introduced in \S\xref{subsection:lines-and-conics}. In particular, for an irreducible
component~\mbox{$\Sigma_0 \subset \Sigma(Y)$} of the Hilbert
scheme of lines on~$Y$ we denote by~${\mathscr{L}}_0(Y)$ the corresponding component of the universal line, so that we have the diagram~\eqref{diagram-universal-line}
Note that every component of~$\Sigma(Y)$ is generically reduced by Lemma~\ref{lemma-hilb-y-1}.
As we explained earlier, we are interested in proving faithfulness of the action of the automorphism group $\mathrm{Aut}(Y)$ on the Hilbert scheme $\Sigma(Y)$.
In the next lemma we consider an irreducible component $\Sigma_0$ of $\Sigma(Y)$ and the subgroup $\mathrm{Aut}_{\Sigma_0}(Y) \subset \mathrm{Aut}(Y)$ stabilizing~it.
Although for $\operatorname{d}(Y) \geqslant 3$ the Hilbert scheme $\Sigma(Y)$ is irreducible by Proposition~\xref{hilb-lines-explicit},
for~\mbox{$\operatorname{d}(Y) = 2$} the Hilbert scheme $\Sigma(Y)$ might a priori have several components. This is why we formulate the lemma in this form.
\begin{lemma}\label{lemma:Fano-lines}
Let $Y$ be a Fano threefold with $\rho(Y) = 1$, $\iota(Y) = 2$, and
assume that~$-K_Y$ is very ample, i.e.\ $2 \leqslant \operatorname{d}(Y) \leqslant 5$.
Then the action of $\mathrm{Aut}_{\Sigma_0}(Y)$ on an irreducible component $\Sigma_0$ of the Hilbert scheme of lines on $Y$ is faithful.
In particular, the action of~\mbox{$\mathrm{Aut}(Y)$} on $\Sigma(Y)$ is faithful.
\end{lemma}
\begin{proof}
Let~\mbox{$G \subset \mathrm{Aut}_{\Sigma_0}(Y)$} be the kernel of the action of the group $\mathrm{Aut}_{\Sigma_0}(Y)$ on $\Sigma_0$.
Suppose that $G$ is nontrivial.
Since $-K_Y$ is ample, the group $G$ is a linear algebraic group by Corollary~\xref{corollary:anticanonical-ring}.
By Lemma~\xref{lemma-hilb-y-1} the conditions of Theorem~\xref{theorem:faithful} are satisfied.
Since~\mbox{$\dim Y = 3$} is odd, we conclude that there is a double cover $\pi \colon Y \to Y'$ over a smooth Fano variety $Y'$ with $\rho(Y') = 1$.
Since the branch divisor $B \subset Y'$ is anticanonical, it follows that $K_Y \sim\frac12\pi^*K_{Y'}$, hence
\begin{equation*}
K_Y^3 = \frac18 \cdot 2\cdot K_{Y'}^3 = \frac14 K_{Y'}^3.
\end{equation*}
In particular,
one has
\begin{equation*}
-K_{Y'}^3 = -4K_Y^3 = 32\operatorname{d}(Y) \geqslant 64.
\end{equation*}
This is only possible if $Y' \cong \mathbb{P}^3$ (see Table~\xref{table:Fanos-i-ge-2}).
On the other hand, by Theorem~\xref{theorem:faithful} for a line $L$ corresponding to a general point of $\Sigma_0$ we have $L = \pi^*L'$
for a curve $L' \subset Y'$, hence
\begin{equation}\label{eq:2-KY-L}
2 = -K_Y \cdot L = -\frac12 \pi^*K_{Y'} \cdot \pi^*L' = -K_{Y'} \cdot L'.
\end{equation}
The right side of~\eqref{eq:2-KY-L} is divisible by $\iota(Y') = \iota(\mathbb{P}^3) = 4$.
This contradiction shows that~$G$ is trivial, hence the action of $\mathrm{Aut}_{\Sigma_0}(Y)$ on $\Sigma_0$ is faithful.
\end{proof}
\begin{corollary}\label{corollary-aut-index-2}
Let $Y$ be a \textup(smooth\textup) Fano threefold with $\rho(Y) = 1$, $\iota(Y) = 2$, and~\mbox{$\operatorname{d}(Y)\geqslant 3$}.
The following assertions hold:
\begin{itemize}
\item[(i)]
if $\operatorname{d}(Y) = 3$ then
$\mathrm{Aut}(Y)\subset\mathrm{Aut}(S)$
for a smooth minimal surface $S$ of general type with irregularity $5$, geometric genus $10$, and canonical degree~\mbox{$K_{\Sigma(Y)}^2=45$};
\item[(ii)]
if $\operatorname{d}(Y) = 4$ then
$\mathrm{Aut}(Y)\subset\mathrm{Aut}(S)$
for an abelian surface $S$;
\item[(iii)]
if $\operatorname{d}(Y) = 5$ then
$\mathrm{Aut}(Y) \subset \mathrm{Aut}({\mathbb{P}}^2) \cong\mathrm{PGL}_3(\Bbbk)$.
\end{itemize}
In particular, for $\operatorname{d}(Y) = 3$ and $\operatorname{d}(Y) = 4$, the group $\mathrm{Aut}(Y)$ is finite.
\end{corollary}
\begin{proof}
Assertions (i), (ii) and (iii) follow from Proposition~\xref{hilb-lines-explicit}
and Lemma~\xref{lemma:Fano-lines}.
Finiteness for $\operatorname{d}(Y)=3$
follows from assertion~(i) and Corollary~\xref{corollary:general-type},
while finiteness for~\mbox{$\operatorname{d}(Y)=4$} follows from assertion~(ii) and
Corollaries~\xref{corollary:anticanonical-ring}
and~\xref{corollary:nef-K}.
\end{proof}
An alternative proof of finiteness of the automorphism
group $\mathrm{Aut}(Y)$ for $\operatorname{d}(Y)=3$ and~\mbox{$\operatorname{d}(Y)=4$}
is by applying Theorems~\xref{theorem:MatsumuraMonsky} and~\xref{theorem:Benoist}.
For~\mbox{$\operatorname{d}(Y)=5$} one actually has
\begin{equation*}
\mathrm{Aut}(Y)\cong\mathrm{PGL}_2(\Bbbk),
\end{equation*}
see Theorem~\xref{theorem-mu-5} below.
\begin{remark}
Besides its action on $\Sigma(Y)$, the automorphism group $\mathrm{Aut}(Y)$ also acts on the intermediate Jacobian $J(Y)$ of $Y$.
For $\operatorname{d}(Y) = 3$ one can check that this action is faithful.
Indeed, by~\cite[\S5]{Beauville1982singularities} the intermediate Jacobian $J(Y)$ contains an $\mathrm{Aut}(Y)$-invariant theta divisor $\Theta \subset J(Y)$
which is equal to the image of the canonical Abel--Jacobi map
\begin{equation*}
\Sigma(Y) \times \Sigma(Y) \to J(Y),
\qquad
(L_1,L_2) \mapsto [L_1] - [L_2].
\end{equation*}
Moreover, $\Theta$ has a unique singular point $P \in \Theta$ (the image of the diagonal in $\Sigma(Y) \times \Sigma(Y)$)
and the exceptional divisor of the blow up of $\Theta$ at $P$ is $\mathrm{Aut}(Y)$-equivariantly isomorphic to $Y$.
The faithfulness of the action on $J(Y)$ follows immediately.
For $\operatorname{d}(Y) = 4$ one can show that the group $\mathrm{Aut}(Y)$ contains a subgroup~$\Gamma$ of order $32$ that acts trivially on the corresponding pencil
of quadrics, and $\Gamma$ contains a subgroup~$\Gamma_0$ of order $16$
that acts trivially on the associated curve $B(Y)$ of genus $2$
mentioned in Remark~\xref{remark:hyperelliptic-curve}, and also on the intermediate Jacobian
of~$Y$.
Finally, for $\operatorname{d}(Y) = 5$ the intermediate Jacobian of $Y$ is zero (see for instance~\cite[\S12.2]{Iskovskikh-Prokhorov-1999}).
\end{remark}
\subsection{Action on conics}
\label{subsection:act-on-conics}
Now we will analyze the action of the automorphism group of a smooth Fano threefold
$X$ of Picard rank $1$ and index $1$ on the Hilbert scheme $S(X)$ of conics on~$X$. We will
assume that $H \sim -K_X$ is very ample; in particular, this means that~\mbox{$\operatorname{g}(X) \geqslant 3$}.
As in the case of lines, we are interested in proving faithfulness of the action of the automorphism group $\mathrm{Aut}(X)$ on $S(X)$,
but by the same reason as in \S\xref{subsection:act-on-lines} we consider the action of the subgroup $\mathrm{Aut}_{S_0}(X)$ stabilizing an irreducible component $S_0$ of~$S(X)$.
Note that we know irreducibility of $S(X)$ for $\operatorname{g}(X) \geqslant 7$ (see Proposition~\xref{proposition:conics}),
but already for~$\operatorname{g}(X) = 6$ the scheme $S(X)$ can be reducible (see Remark~\xref{remark:hilb-conics-reducible}).
We start by discussing some cases
when the action of the subgroup $\mathrm{Aut}_{S_0}(X) \subset \mathrm{Aut}(X)$ on an irreducible component $S_0$ of $S(X)$
is not faithful.
\begin{lemma}\label{lemma:trivial-action-conics}
Let $Y$ be a Fano threefold with $\rho(Y) = 1$, $\iota(Y) = 2$, and~\mbox{$\operatorname{d}(Y) \geqslant 2$}.
Let~\mbox{$\pi\colon X \to Y$} be a double cover branched in a smooth anticanonical divisor $B \subset Y$.
Then~$X$ is a smooth Fano threefold with $\rho(X)=1$, $\iota(X) = 1$, \mbox{$\operatorname{g}(X) = \operatorname{d}(Y) + 1$},
and~$S(X)$ has an irreducible component $S_0$ such that the action of $\mathrm{Aut}_{S_0}(X)$ on it is not faithful.
\end{lemma}
\begin{proof}
By the Hurwitz formula one has
\begin{equation}\label{equationHurwitz}
-K_X\sim \pi^* H_Y,
\end{equation}
where $H_Y$ is the ample generator of the Picard group of $Y$ (so that $-K_Y\sim 2H_Y$).
Hence~$X$ is a (smooth) Fano threefold.
Furthermore, by \cite{Cornalba1981}
the pullback morphism \mbox{$\pi^* \colon H^2(Y,\mathbb{Z}) \to H^2(X,\mathbb{Z})$} is an isomorphism, hence $\rho(X) = 1$ and $\mathrm{Pic}(X)$ is generated by $K_X$, i.e., $\iota(X) = 1$.
It follows easily that $\operatorname{g}(X) = \operatorname{d}(Y) + 1$.
For every line $L \subset Y$ its preimage~\mbox{$\pi^{-1}(L) \subset X$} is a conic.
This defines a morphism $\Sigma(Y) \to S(X)$, whose image is a union of components of $S(X)$.
The Galois involution of the double cover is an automorphism of $X$ which acts trivially on
any such component $S_0$ of $S(X)$,
hence is contained in the kernel of $\mathrm{Aut}_{S_0}(X)$-action on $S_0$.
\end{proof}
\begin{example}\label{example:trivial-action-exotic-component}
Assume that $X$ is a quartic threefold with cones and $S_0 \subset S(X)$ is an exotic component (see Lemma~\xref{lemma:exotic-component}).
Then there may be a nontrivial subgroup in $\mathrm{Aut}(X)$ acting trivially on $S_0$. For example, assume that $X \subset {\mathbb{P}}^4$ is the Fermat quartic,
consider a cone on $X$ described in Example~\xref{example:quartic-cone} and consider the action of the group $\boldsymbol{\mu}_4$ on~${\mathbb{P}}^4$
by the primitive character on the first two coordinates, and trivial on the last three coordinates. The equation of $X$ is preserved
by this action, hence $\boldsymbol{\mu}_4$ acts (faithfully) on~$X$. On the other hand, its action on the base $B$ of the cone is trivial,
hence so is its action on~\mbox{$S_0=\mathop{\mathsf{Hilb}}\nolimits^2(B)$}.
However, it may act nontrivially on nilpotents (recall that the scheme structure on the component of $S(X)$ underlying $S_0$ is everywhere non-reduced).
\end{example}
In the next lemma we show that the situations of
Lemma~\xref{lemma:trivial-action-conics} and Example~\xref{example:trivial-action-exotic-component} are the only ones
when the action of the automorphism group $\mathrm{Aut}_{S_0}(X)$ on an irreducible component $S_0$ of $S(X)$ is not faithful,
at least in the case of a very ample~$-K_X$.
\begin{lemma}\label{lemma:Fano-conics}
Let $X$ be a Fano threefold with $\rho(X)=1$ and $\iota(X)=1$ such that~\mbox{$-K_X$} is very ample,
i.e. either $\operatorname{g}(X)\geqslant 4$, or $X$ is a quartic threefold.
If the action of $\mathrm{Aut}_{S_0}(X)$ on a non-exotic irreducible component $S_0$ of $S(X)$ is not faithful,
then~$X$ is a double cover of a smooth Fano threefold $Y$ with $\rho(Y) = 1$, $\iota(Y) = 2$,
and~\mbox{$\operatorname{d}(Y)=\operatorname{g}(X)-1\geqslant 2$},
the irreducible component $S_0$ comes from $\Sigma(Y)$ as in Lemma~\xref{lemma:trivial-action-conics},
and the kernel of the action of $\mathrm{Aut}_{S_0}(X)$ on $S_0$ is generated by the Galois involution of the double cover.
In particular, for $\operatorname{g}(X) \geqslant 7$ the action of $\mathrm{Aut}(X)$ on $S(X)$ is faithful.
\end{lemma}
\begin{proof}
Let~\mbox{$G \subset \mathrm{Aut}_{S_0}(X)$} be the kernel of the action of the group $\mathrm{Aut}_{S_0}(X)$ on $S_0$.
Suppose that $G$ is nontrivial.
The group $G$ is a linear algebraic group by Corollary~\xref{corollary:anticanonical-ring}.
If~$S_0$ is not exotic then by Lemma~\xref{lemma-hilb-x-2} the conditions of Theorem~\xref{theorem:faithful} are satisfied.
Since~\mbox{$\dim X = 3$} is odd, we conclude that there is a double cover $\pi \colon X \to Y$ over a smooth Fano
variety~$Y$ with $\rho(Y) = 1$ (which corresponds to the variety $X'$ of Theorem~\xref{theorem:faithful}).
The index of $Y$ is even and $Y \not\cong \mathbb{P}^3$ (by the same reason as in the proof of Lemma~\xref{lemma:Fano-lines},
where we had a similar situation with the threefold $Y'$ instead of~$Y$),
hence~\mbox{$\iota(Y) = 2$}.
Since the branch divisor $B \subset Y$ is anticanonical, it follows
that~\mbox{$-K_X \sim \pi^*H_Y$}, where~$H_Y$ is the ample generator of the Picard group of $Y$ (so that $-K_Y\sim 2H_Y$).
Hence
\begin{equation*}
K_X^3 = \frac18 \cdot 2\cdot K_Y^3 = \frac14 K_Y^3,\quad \operatorname{d}(Y) = -\frac18K_Y^3 = -\frac12K_X^3 = \operatorname{g}(X) - 1.
\end{equation*}
Since $\operatorname{g}(X) \geqslant 3$ it follows that $\operatorname{d}(Y) \geqslant 2$.
On the other hand, since $\operatorname{d}(Y) \leqslant 5$, we have~\mbox{$\operatorname{g}(X) \leqslant 6$}, so for $\operatorname{g}(X) \geqslant 7$ the action is faithful.
\end{proof}
We think that for $\operatorname{g}(X) \leqslant 6$ the action of $\mathrm{Aut}(X)$ on $S(X)$ is still faithful.
\begin{corollary}\label{corollary:high-genus-Fanos-Aut}
Let $X$ be a \textup(smooth\textup) Fano threefold with $\rho(X) = 1$, $\iota(X) = 1$, and
genus $\operatorname{g}(X)\geqslant 7$.
The following assertions hold:
\begin{itemize}
\item[(i)]
if $\operatorname{g}(X) = 7$ then $\mathrm{Aut}(X) \subset \mathrm{Aut}(C)$ for a smooth irreducible curve $C$ of genus~$7$;
\item[(ii)]
if $\operatorname{g}(X) = 8$ then
$\mathrm{Aut}(X)\subset\mathrm{Aut}(S)$ for a minimal surface of general type
with irregularity $5$, geometric genus $10$, and canonical
degree~\mbox{$K_{S(X)}^2=45$};
\item[(iii)]
if $\operatorname{g}(X) = 9$ then
$\mathrm{Aut}(X) \subset \mathrm{Aut}(S)$ for a surface $S$ isomorphic to a projectivization of a simple rank $2$ vector bundle on a smooth irreducible curve of genus~$3$;
\item[(iv)]
if $\operatorname{g}(X) = 10$ then
$\mathrm{Aut}(X)\subset\mathrm{Aut}(S)$ for an abelian surface $S$;
\item[(v)]
if $\operatorname{g}(X) = 12$ then
$\mathrm{Aut}(X) \subset \mathrm{Aut}({\mathbb{P}}^2) \cong \mathrm{PGL}_3(\Bbbk)$.
\end{itemize}
In particular, if $7 \leqslant \operatorname{g}(X) \leqslant 10$, then $\mathrm{Aut}(X)$ is finite.
\end{corollary}
\begin{proof}
By Lemma~\xref{lemma:Fano-conics} the action of $\mathrm{Aut}(X)$ on the Hilbert scheme
$S(X)$ of conics on~$X$ is faithful. Thus assertions~(ii), (iii), (iv) and~(v)
are implied by assertions~(ii), (iii), (iv) and~(v) of Proposition~\xref{proposition:conics},
respectively. In case of assertion~(i) we also take into account an isomorphism
$\mathrm{Aut}\big(\mathrm{Sym}^2(C)\big)\cong\mathrm{Aut}(C)$,
see~\cite{Ran1986}.
Keeping in mind Corollary~\xref{corollary:general-type}, we see that finiteness
for $\operatorname{g}(X)=7$ and~$8$ is implied by assertions~(i) and~(ii)
respectively. If $\operatorname{g}(X)=9$, finiteness is implied by Lemma~\xref{lemma:projectivization-stable-bundle}
and assertion~(iii).
Finally, keeping in mind Corollary~\xref{corollary:anticanonical-ring} and
Lemma~\xref{lemma:non-uniruled} we see that finiteness for $\operatorname{g}(X)=10$ is implied by assertion~(iv).
\end{proof}
As we will see in~\S\xref{section:infinite} some Fano threefolds of index 1 with
$\operatorname{g}(X) = 12$ actually have an infinite automorphism group.
\subsection{Small degree and genus}
\label{subsection:finite-Fano}
As we have shown in Corollaries~\xref{corollary-aut-index-2}
and~\xref{corollary:high-genus-Fanos-Aut}, for Fano threefolds of index 2 and degree $3 \leqslant \operatorname{d}(Y) \leqslant 4$, and
for Fano threefolds of index 1 and genus~\mbox{$7 \leqslant \operatorname{g}(X) \leqslant 10$} the automorphism groups are always finite.
In this subsection we show the same for smaller values of degree and genus.
We start with the cases when $H$ is not very ample.
\begin{lemma}\label{lemma:double-cover-finite}
Let $X$ be a Fano threefold of index $1$ or $2$ with $\rho(X) = 1$. If the ample generator~$H$ of~\mbox{$\mathrm{Pic}(X)$}
is not very ample then the group $\mathrm{Aut}(X)$ is finite.
\end{lemma}
\begin{proof}
According to
Tables~\xref{table:Fanos-i-ge-2}
and~\xref{table:Fanos-i-1} all such varieties are double covers
\begin{equation*}
\varphi \colon X \to X',
\end{equation*}
where $X'$ is a Fano threefold with $\mathrm{Cl}(X')=\mathbb{Z}\cdot H'$ for some ample divisor $H'$ on $X'$.
In Table~\xref{table:double-covers} we list all possible situations.
The first column of Table~\xref{table:double-covers} lists the invariants of~$X$.
The second column describes $X'$; here
$\mathbb{P}(1,1,1,2)$ is the weighted projective space and~$Q$ is a smooth three-dimensional quadric.
The third column specifies the class in~\mbox{$\mathrm{Pic}(X')$} of the branch divisor $B \subset X'$ of the double cover $\varphi \colon X \to X'$
(note that in the case~\mbox{$X' = \mathbb{P}(1,1,1,2)$} the double cover is also branched over the singular point of $X'$).
\begin{table}[h]
\caption{Double covers}\label{table:double-covers}
\begin{tabular}{|c|c|c|}
\hline
invariants of $X$ & $X'$ & $B$ \\
\hline
$\iota(X) = 2,\ \operatorname{d}(X) = 1$ &
$\mathbb{P}(1,1,1,2)$ & $6H'$ \\
\hline
$\iota(X) = 2,\ \operatorname{d}(X) = 2$ & ${\mathbb{P}}^3$ & $4H'$ \\
\hline
$\iota(X) = 1,\ \operatorname{g}(X) = 2$ & ${\mathbb{P}}^3$ & $6H'$ \\
\hline
$\iota(X) = 1,\ \operatorname{g}(X) = 3$ & $Q$ & $4H'$ \\
\hline
\end{tabular}
\end{table}
The map $\varphi$ is anticanonical in all cases except when $\iota(X)=2$ and $\operatorname{d}(X)=2$;
in the latter case it is defined by the complete linear system of the divisor $H$ such that~$2H\sim -K_X$.
In particular, $\varphi$ is equivariant with respect to the whole automorphism group $\mathrm{Aut}(X)$, hence we have a natural map
\begin{equation}\label{eq:AutXB}
\mathrm{Aut}(X) \to \mathrm{Aut}(X'; B)
\end{equation}
into the group of automorphisms of $X'$ preserving $B$.
Moreover, we have
\begin{equation*}
X \cong \operatorname{Spec}_{X'}\left({\mathscr{O}}_{X'} \oplus {\mathscr{O}}_{X'}\left(-\textstyle{\frac12}B\right)\right),
\end{equation*}
where ${\mathscr{O}}_{X'}\left(-\textstyle{\frac12}B\right)$ is the reflexive sheaf corresponding to the Weil divisor class $-\textstyle{\frac12}B$,
and the algebra structure is determined by the composition
\begin{equation*}
{\mathscr{O}}_{X'}\left(-\textstyle{\frac12}B\right) \otimes {\mathscr{O}}_{X'}\left(-\textstyle{\frac12}B\right) \xrightarrow{\ \ \ }
{\mathscr{O}}_{X'}(-B) \xrightarrow{\ B\ } {\mathscr{O}}_{X'}
\end{equation*}
with the canonical first map and with the second map given by the divisor $B$.
In particular, every automorphism of $X'$ that fixes $B$ induces an automorphism of $X$, hence
the map~\eqref{eq:AutXB}
is surjective.
Its kernel is clearly generated by the Galois involution of the double cover~$\varphi$,
hence is isomorphic to $\boldsymbol{\mu}_2$.
On the other hand, it is clear from Table~\ref{table:double-covers} that the divisor $2H'$ is very ample in all cases, so that
$X' \subset \mathbb{P}(V)$, where $V = H^0(X,{\mathscr{O}}_{X'}(2H'))^\vee$, and $\mathrm{Aut}(X') \subset \mathrm{PGL}(V)$.
Furthermore,
$B$ is not contained in a hyperplane in $\mathbb{P}(V)$, hence
the natural map
\begin{equation*}
\mathrm{Aut}(X'; B) \to \mathrm{Aut}(B; [2H'])
\end{equation*}
into the group of automorphisms of $B$ preserving (the class in $\mathrm{Pic}(B)$ of) the restriction of $2H'$ to $B$
is injective.
Thus we have an exact sequence
\begin{equation*}
1 \to \boldsymbol{\mu}_2 \to \mathrm{Aut}(X) \to \mathrm{Aut}(B; [2H']).
\end{equation*}
It remains to notice that $B$ is smooth (as the fixed locus of an involution on a smooth variety $X$) and its canonical bundle
is nef by adjunction formula, hence $\mathrm{Aut}(B; [2H'])$ is finite
by Lemma~\xref{lemma:anticanonical-ring0} and Corollary~\xref{corollary:nef-K}.
\end{proof}
Now we will combine the above results with Theorems~\xref{theorem:MatsumuraMonsky} and Theorem~\xref{theorem:Benoist}.
\begin{proposition}\label{proposition:nearly-done}
If $Y$ is a \textup(smooth\textup) Fano threefold with $\rho(Y) = 1$, $\iota(Y) = 2$, and~\mbox{$\operatorname{d}(Y) \leqslant 4$}
then the group $\mathrm{Aut}(Y)$ is finite.
If $X$ is a \textup(smooth\textup) Fano threefold with~$\rho(X) = 1$, $\iota(X) = 1$, and~\mbox{$\operatorname{g}(X) \leqslant 10$}
then the group $\mathrm{Aut}(X)$ is finite.
\end{proposition}
\begin{proof}
First, let $Y$ be a Fano threefold with $\rho(Y) = 1$ and $\iota(Y) = 2$.
If~\mbox{$\operatorname{d}(Y) \in \{1,2\}$}, then
the ample generator of $\mathrm{Pic}(Y)$ is not very ample, and
we apply Lemma~\xref{lemma:double-cover-finite}.
If~\mbox{$\operatorname{d}(Y) \in \{3,4\}$}, then we apply Corollary~\xref{corollary-aut-index-2}.
Second, let $X$ be a Fano threefold with $\rho(X) = 1$ and $\iota(X) = 1$.
If $\operatorname{g}(X) = 2$ or~\mbox{$\operatorname{g}(X) = 3$} and~$-K_X$ is not very ample,
we apply Lemma~\xref{lemma:double-cover-finite}.
If $\operatorname{g}(X) = 3$ and~$-K_X$ is very ample then $X$ is a quartic in $\mathbb{P}^4$
and we apply Theorem~\xref{theorem:MatsumuraMonsky}.
If~\mbox{$\operatorname{g}(X)=4$} or~\mbox{$\operatorname{g}(X)=5$}, then $X$ is a complete intersection in a projective space
of multidegree~\mbox{$(2,3)$} and~\mbox{$(2,2,2)$} respectively,
and we apply Theorem~\xref{theorem:Benoist}.
If $\operatorname{g}(X) = 6$ we refer to~\cite[Proposition~3.21(c)]{Debarre-Kuznetsov2015}.
Finally, if~\mbox{$7 \leqslant \operatorname{g}(X) \leqslant 10$}, we apply Corollary~\xref{corollary:high-genus-Fanos-Aut}.
\end{proof}
To complete the proof of Theorem~\xref{theorem:Prokhorov}, we need to describe the automorphism groups of Fano threefolds of index 2 and degree 5,
and of index 1 and genus 12. This is done in the next section.
\section{Infinite automorphism groups}
\label{section:infinite}
We already know from Proposition~\xref{proposition:nearly-done} that the only Fano threefolds of Picard rank~$1$ and index 1 or 2
which can have infinite automorphism groups are the threefold $Y$ with~$\iota(Y) = 2$ and $\operatorname{d}(Y) = 5$
(such threefold is actually unique up to isomorphism, see \cite[Theorem~II.1.1]{Iskovskikh-1980-Anticanonical} or \cite[3.3.1--3.3.2]{Iskovskikh-Prokhorov-1999}),
and some of the threefolds $X$ with~$\iota(X) = 1$ and $\operatorname{g}(X) = 12$.
\subsection{Fano threefolds of index 2 and degree 5}
\label{subsection:v5}
We start with a detailed description of the Fano threefold $Y$ with $\iota(Y) = 2$ and $\operatorname{d}(Y) = 5$.
From the classification it is known that $Y$ is isomorphic to a linear section of the Grassmannian $\mathrm{Gr}(2,5)\subset \mathbb{P}^9$
by a subspace~\mbox{$\mathbb{P}^6\subset\mathbb{P}^9$} (see Table~\xref{table:Fanos-i-ge-2}).
For our purposes, however, the description of $Y$ suggested by Mukai and Umemura~\cite{Mukai-Umemura-1983} is more convenient.
Let
\begin{equation*}
M_d=\mathrm{Sym}^d (\Bbbk^{2})^\vee
\end{equation*}
be the space of binary forms of degree $d$.
We denote by $x$ and $y$ the elements of the standard basis of the vector space~\mbox{$(\Bbbk^2)^\vee$},
so that elements of $M_d$ are polynomials of degree~$d$ in~$x$ and~$y$.
The group $\mathrm{GL}_2(\Bbbk)$ acts naturally on the space $M_d$ by the rule
\begin{equation*}
\left(\begin{smallmatrix} a & b \\ c & d \end{smallmatrix}\right) \colon
x \mapsto ax + cy,
\qquad
y \mapsto bx + dy,
\end{equation*}
and induces an action of $\mathrm{PGL}_2(\Bbbk)$ on the projective space $\mathbb{P}(M_d)$. Consider the form
\begin{equation*}
\phi_6(x,y)=xy(x^4-y^4)\in M_6
\end{equation*}
and the corresponding point $[\phi_6]\in\mathbb{P}(M_6)\cong\mathbb{P}^6$.
\begin{theorem}[{\cite{Mukai-Umemura-1983}, see also~\cite[Proposition~7.1.10]{CheltsovShramov2016}}]
\label{theorem-mu-5}
The stabilizer of $[\phi_6]$ is the octahedral group
$$\mathrm{Oct}\cong\mathfrak{S}_4 \subset \mathrm{PGL}_2(\Bbbk),$$
and the closure of its orbit
\begin{equation*}
Y= \overline{\mathrm{PGL}_2(\Bbbk)\cdot [\phi_6]}\subset\mathbb{P}^6
\end{equation*}
is the smooth Fano threefold with $\rho(Y)=1$, $\iota(Y)=2$, and $\operatorname{d}(Y)=5$
embedded by the ample generator of $\mathrm{Pic}(Y)$.
The automorphism group of $Y$ is $\mathrm{Aut}(Y) \cong \mathrm{PGL}_2(\Bbbk)$.
\end{theorem}
We will need a description of the $\mathrm{PGL}_2(\Bbbk)$-orbits on $Y$ (see \cite[Lemma 1.5]{Mukai-Umemura-1983}).
For this we need to introduce notation for the standard connected subgroups in $\mathrm{PGL}_2(\Bbbk)$.
We denote by
\begin{itemize}
\item $B_2 \subset \mathrm{PGL}_2(\Bbbk)$ the standard Borel subgroup (upper triangular matrices),
\item $U_2 \subset \mathrm{PGL}_2(\Bbbk)$ the standard unipotent subgroup (upper triangular matrices with units on the diagonal), and
\item $T_2 \subset \mathrm{PGL}_2(\Bbbk)$ the standard torus (diagonal matrices).
\end{itemize}
The orbit decomposition of $Y$ is
\begin{equation*}
Y = \Orb_3(Y) \sqcup \Orb_2(Y) \sqcup \Orb_1(Y)
\end{equation*}
with $\Orb_k(Y)$ standing for the unique $\mathrm{PGL}_2(\Bbbk)$-orbit of dimension $k$; explicitly
\begin{alignat*}{2}
\Orb_3(Y) & = \mathrm{PGL}_2(\Bbbk)\cdot [\phi_6] && \cong \mathrm{PGL}_2(\Bbbk)/\mathrm{Oct},\\
\Orb_2(Y) & = \mathrm{PGL}_2(\Bbbk)\cdot [xy^5] && \cong \mathrm{PGL}_2(\Bbbk)/T_2,\\
\Orb_1(Y) & = \mathrm{PGL}_2(\Bbbk)\cdot [x^6] && \cong \mathrm{PGL}_2(\Bbbk)/B_2.
\end{alignat*}
It is clear from this description that
$\Orb_{1}(Y)$ is a normal rational sextic curve, and that
\begin{equation*}
\oOrb_2(Y) = \Orb_2(Y) \sqcup \Orb_1(Y)
\end{equation*}
is the image of ${\mathbb{P}}^1 \times {\mathbb{P}}^1 = {\mathbb{P}}(M_1) \times {\mathbb{P}}(M_1)$ under the map
\begin{equation*}
\nu \colon {\mathbb{P}}(M_1) \times {\mathbb{P}}(M_1) \xrightarrow{\qquad} {\mathbb{P}}(M_6),
\qquad (f,g) \mapsto f^5g.
\end{equation*}
Geometrically, $\oOrb_2(Y)$ is the tangential scroll of $\Orb_1(Y)$, i.e., the surface swept by tangent lines to the twisted sextic curve $\Orb_1(Y)$,
and $\nu$ is its normalization morphism (for more details, see for instance~\cite[Lemma~7.2.2]{CheltsovShramov2016}).
\begin{remark}
\label{remark:curves-oorb2}
It follows from the above description that any irreducible curve of degree at most~5 contained in $\oOrb_2(Y)$ is
either a line~$\nu(\{f\} \times {\mathbb{P}}(M_1))$ (as we show below these are special lines on~$Y$, as defined in~\S\xref{subsection:lines-and-conics}),
or is the image of the normal rational quintic
\begin{equation}\label{eq:z:mu}
Z_{\mathrm{MU}} = \nu({\mathbb{P}}(M_1) \times \{x\}) = \overline{B_2 \cdot [xy^5] } \subset Y
\end{equation}
under the $\mathrm{PGL}_2(\Bbbk)$-action.
The reason for the notation in~\eqref{eq:z:mu} will become clear later.
And meanwhile, just note that $Z_{\mathrm{MU}}$ is preserved by the Borel subgroup $B_2 \subset \mathrm{PGL}_2(\Bbbk)$.
\end{remark}
Recall that by Proposition~\ref{hilb-lines-explicit}(iii) the Hilbert scheme of lines $\Sigma(Y)$ is isomorphic to~$\mathbb{P}^2$.
In fact, we have a $\mathrm{PGL}_2(\Bbbk)$-equivariant isomorphism
\begin{equation*}
\Sigma(Y) \cong \mathbb{P}(M_2)
\end{equation*}%
see~\cite[Theorem~I]{Furushima1989a} or \cite[Proposition~2.20]{Sanna}.
Below we describe explicitly lines on~$Y$ corresponding to points of $\mathbb{P}(M_2)$.
Note that any pair of points $f,g \in \mathbb{P}(M_1)$ gives a point $fg \in \mathbb{P}(M_2)$,
and, if $f \ne g$, a point $fg(f^4 - g^4) \in \Orb_3(Y) \subset Y$.
\begin{lemma}
\label{lemma:lines-v5-explicit}
Every line on $Y$ can be written in one of the following two forms:
\begin{equation*}
\begin{aligned}
L_{fg} &= \{ fg(s_1f^4 - s_2g^4) \in Y \}_{(s_1:s_2) \in \mathbb{P}^1}, \qquad&& \text{for $f,g\in\mathbb{P}(M_1)$, $f \ne g$},\\
L_{f^2} &= \{ f^5(s_1x + s_2y) \}_{(s_1:s_2) \in \mathbb{P}^1}, && \text{for $f\in\mathbb{P}(M_1)$}.
\end{aligned}
\end{equation*}
\end{lemma}
\begin{proof}
It is clear that both $L_{fg}$ and $L_{f^2}$ are lines on $Y$.
The first intersects $\oOrb_2(Y)$ at two points $fg^5$ and $f^5g$, while the second is contained in $\oOrb_2(Y)$.
Thus, they correspond to points of different $\mathrm{PGL}_2(\Bbbk)$-orbits in $\Sigma(Y)$.
It remains to recall that $\Sigma(Y) \cong \mathbb{P}(M_2)$
contains just two $\mathrm{PGL}_2(\Bbbk)$-orbits, and to notice that the images of the lines $L_{fg}$ and $L_{f^2}$
under the action of $\mathrm{PGL}_2(\Bbbk)$ are lines of the same form.
\end{proof}
From the description of Lemma~\ref{lemma:lines-v5-explicit} it is easy to obtain the following result.
Recall that ${\mathscr{L}}(Y)$ denotes the universal line on $Y$, and $q \colon {\mathscr{L}}(Y) \to \Sigma(Y)$, $p \colon {\mathscr{L}}(Y) \to Y$ denote its natural projections.
\begin{corollary}[{cf. \cite[1.2.1(3)]{Iliev1994a} and \cite[Corollary~2.24]{Sanna}}]
\label{corollary:V5-lines-description}
The set $q(p^{-1}([\varphi]))$ of lines on $Y$ passing through a point $[\varphi] \in Y$
can be described as follows:
\begin{equation*}
q(p^{-1}([\varphi])) =
\begin{cases}
\{[fg], [f^2 - g^2], [f^2 + g^2]\}, & \text{if $\varphi = fg(f^4 - g^4) \in \Orb_3(Y)$,}\\
\{[fg], [f^2]\}, & \text{if $\varphi = f^5g \in \Orb_2(Y)$,}\\
\{[f^2]\}, & \text{if $\varphi = f^6 \in \Orb_1(Y)$,}\\
\end{cases}
\end{equation*}
The ramification divisor of the map $p \colon {\mathscr{L}}(Y) \to Y$ is the union of lines $L_{f^2}$ for $f \in \mathbb{P}(M_1)$.
In other words, any line $L_{fg}$ with $f \ne g$ is ordinary and any line $L_{f^2}$ is special.
\end{corollary}
The three points $[fg]$, $[f^2 - g^2]$, and $[f^2 + g^2]$ parameterizing three lines through a general point of $Y$
correspond to the three axes of an octahedron.
\subsection{Fano threefolds of index 1 and genus 12}
\label{subsection:v22}
There is a similar example of a Fano threefold $X$ with $\rho(X) = 1$, $\iota(X) = 1$, and~\mbox{$\operatorname{g}(X) = 12$},
that was also found by Mukai and Umemura.
Consider the form
\begin{equation*}
\phi_{12}(x,y)=xy(x^{10} +11 x^5y^5+y^{10}) \in M_{12}
\end{equation*}
and the point
\begin{equation*}
\upsilon= (\phi_{12}, 1) \in {\mathbb{P}}(M_{12}\oplus M_0) \cong {\mathbb{P}}^{13}.
\end{equation*}
\begin{theorem}[\cite{Mukai-Umemura-1983}]
\label{theorem-mu-22}
The stabilizer of $[\upsilon]$ is the icosahedral group
$$\mathrm{Icos}\cong\mathfrak{A}_5 \subset \mathrm{PGL}_2(\Bbbk),$$
and the closure of its orbit
\begin{equation*}
X^{\mathrm{MU}} = \overline{\mathrm{PGL}_2(\Bbbk)\cdot [\upsilon]}\subset \mathbb{P}^{13}
\end{equation*}
is a smooth anticanonically embedded Fano threefold with $\rho(X^{\mathrm{MU}})=1$, $\iota(X^{\mathrm{MU}})=1$, and~\mbox{$\operatorname{g}(X^{\mathrm{MU}})=12$}.
The automorphism group of $X^{\mathrm{MU}}$ is $\mathrm{Aut}(X^{\mathrm{MU}}) \cong \mathrm{PGL}_2(\Bbbk)$.
\end{theorem}
Note, however, that $X^{\mathrm{MU}}$ is just a single variety from a six-dimensional family of Fano threefolds of this type.
One of descriptions of other Fano threefolds of index 1 and genus~12 is based on the double projection method.
\begin{theorem}[{\cite{Iskovskikh1989}}, {\cite{Prokhorov-1992b}}, {\cite[Theorem 4.3.7]{Iskovskikh-Prokhorov-1999}}]
\label{theorem:double-projection}
The following assertions hold:
\begin{itemize}
\item[(i)]
Let $X$ be a smooth Fano threefold with $\rho(X) = 1$, $\iota(X) = 1$, and $\operatorname{g}(X) = 12$, and let $L \subset X$ be a line.
Then the linear system $|H_X-2L|$, where $H_X$ is the ample generator of $\mathrm{Pic}(X)$, defines a birational map of $X$
onto the smooth Fano threefold~$Y$ with $\rho(Y)=1$, $\iota(Y)=2$, and $\operatorname{d}(Y)=5$.
\item[(ii)]
Let $Y$ be the smooth Fano threefold with $\rho(Y) = 1$, $\iota(Y)=2$, and $\operatorname{d}(Y)=5$, and let $Z\subset Y \subset \mathbb{P}^6$ be a normal rational quintic curve.
Then the linear system~\mbox{$|3H_Y-2Z|$}, where $H_Y$ is the ample generator of $\mathrm{Pic}(Y)$, defines a
birational map of~$Y$ onto a smooth Fano threefold $X$ with $\rho(X)=1$, $\iota(X)=1$, and~\mbox{$\operatorname{g}(X)=12$}.
\end{itemize}
The constructions of~{\rm(i)} and~{\rm(ii)} are mutually inverse, and the corresponding birational transformation between $X$ and $Y$ can be described by a diagram
\begin{equation}\label{equation(1)}
\vcenter{
\xymatrix{
&&
X' \ar[dl]_{\sigma_X} \ar@{-->}[rr]^-{\upchi} && Y' \ar[dr]^{\sigma_Y}
\\
L \ar@{^{(}->}[r] &
X \ar@{-->}[rrrr]^-\xi &&&& Y &
Z \ar@{_{(}->}[l]
}
}
\end{equation}
where the morphism $\sigma_X$ is the blow up of $L$, the morphism $\sigma_Y$ is the blow up of $Z$,
and the upper dashed arrow $\upchi$ is a flop.
\end{theorem}
\begin{remark}\label{remark:contracted-divisors}
In the above diagram, the map $\xi \colon X\dashrightarrow Y$ contracts a divisor which is a unique member of the linear system $|H_X-3L|$.
Similarly, the map $\xi^{-1} \colon Y\dashrightarrow X$ contracts a divisor which is a unique member of the linear system $|H_Y-Z|$.
\end{remark}
We denote by $E_L \subset X'$ and $E_Z \subset Y'$ the exceptional divisors of the blowups $\sigma_X$ and~$\sigma_Y$.
\begin{lemma}
\label{lemma:flopping-locus}
The flopping locus of the map $\upchi$ is the union of strict transforms of lines on $X$ intersecting $L$,
and of the exceptional section of the divisor $E_L$ if the line $L$ is special.
The flopping locus of the map $\upchi^{-1}$ is the union of strict transforms of bisecants of $Z$ on $Y$.
\end{lemma}
\begin{proof}
The first assertion can be found in~\cite[Proposition~4.3.1]{Iskovskikh-Prokhorov-1999}.
For the second assume that $C \subset Y'$ is a flopping curve of $\upchi^{-1}$ and let $C_X \subset X'$ be the corresponding flopped curve.
Then either $\sigma_X(C_X)$ is a line meeting $L$ or $C_X$ is the exceptional section of $E_L$.
Therefore, one has
\begin{equation*}
(\sigma_X^* H_X-2 E_L)\cdot C_X= -1.
\end{equation*}
By the construction of flops \cite{Kollar-flops} we have $\sigma_Y^* H_Y \cdot C = 1$.
Therefore, $\sigma_Y(C)$ is a line on~$Y$.
Since $K_{Y'}\cdot C = 0$, it is a bisecant of $Z$.
\end{proof}
One can also show that the flopping curves of the map $\upchi$ are disjoint and have normal bundles
of the form $\mathscr{O}_{\mathbb{P}^1}(-1)\oplus\mathscr{O}_{\mathbb{P}^1}(-1)$ or $\mathscr{O}_{\mathbb{P}^1}\oplus\mathscr{O}_{\mathbb{P}^1}(-2)$, see \cite{Cutkosky1989},
and therefore, near each flopping curve, the flop $\upchi$ is given by Reid's pagoda \cite{Reid1983}.
\begin{remark}
\label{remark:functoriality}
The construction of Theorem~\ref{theorem:double-projection} is functorial: an isomorphism between pairs $(X_1,L_1)$ and $(X_2,L_2)$ induces
an isomorphism of the associated diagrams~\eqref{equation(1)}, and hence an isomorphism of the corresponding pairs $(Y,Z_1)$ and $(Y,Z_2)$.
Conversely, an isomorphism of pairs $(Y,Z_1)$ and $(Y,Z_2)$ induces in the same way an isomorphism of pairs $(X_1,L_1)$ and $(X_2,L_2)$ associated with them.
In particular, if the pair $(Y,Z)$ corresponds to a pair $(X,L)$ then
\begin{equation}
\label{eq:aut-xl-yz}
\mathrm{Aut}(X; L) \cong \mathrm{Aut}(Y; Z),
\end{equation}
where $\mathrm{Aut}(X; L) \subset \mathrm{Aut}(X)$ and $\mathrm{Aut}(Y; Z) \subset \mathrm{Aut}(Y)$ are the subgroups preserving $L$ and~$Z$ respectively.
In particular, if $G \subset \mathrm{Aut}(X)$ and $L$ is $G$-invariant, then $G \subset \mathrm{PGL}_2(\Bbbk)$ and $Z$ is $G$-invariant.
Conversely, if $Z$ is stabilized by a subgroup $G \subset \mathrm{PGL}_2(\Bbbk)$ then $G$ acts faithfully on $X$ and preserves the line $L$.
\end{remark}
Denote by $\Sigma^0_L(X) \subset \Sigma(X)$ the open subscheme of the Hilbert scheme of lines on $X$ that parameterizes lines
which intersect neither $L$, nor any other line intersecting $L$.
Similarly, denote by
\begin{equation*}
\Sigma_Z(Y) = q(p^{-1}(Z)) \subset \Sigma(Y)
\end{equation*}
the closed subscheme of $\Sigma(Y)$ parameterizing lines intersecting a normal rational quintic~$Z$, and let
$\Sigma^0_Z(Y) \subset \Sigma_Z(Y)$ be its open subscheme that parameterizes lines which are
neither bisecants of $Z$, nor intersect any bisecant of $Z$.
\begin{lemma}
\label{lemma:sigma-0-iso}
The scheme $\Sigma^0_L(X)$ is an open dense subscheme of $\Sigma(X)$, and
the map $L' \mapsto \xi(L')$ is a rational map $\Sigma(X) \dashrightarrow \Sigma_Z(Y)$
inducing an isomorphism $\Sigma^0_L(X) \cong \Sigma^0_Z(Y)$.
\end{lemma}
\begin{proof}
By Lemma~\ref{lemma:flopping-locus}
any line intersecting a given line $L_0$ on $X$ is a flopping line for the double projection from $L_0$,
hence the number of such lines is finite.
Since any component of $\Sigma(X)$ is one-dimensional (see Lemma~\ref{lemma-hilb-x-1}),
it follows that $\Sigma^0_L(X) \subset \Sigma(X)$ is dense.
If $L'$ corresponds to a point of $\Sigma^0_L(X)$, the map $\xi$ is regular on $L'$.
Since
\begin{equation*}
H_Y \cdot \xi(L') = (H_X - 2E_L) \cdot L' = 1
\qquad\text{and}\qquad
E_Z \cdot \xi(L') = (H_X - 3E_L) \cdot L' = 1,
\end{equation*}
it follows that $\xi(L')$ is a line, and intersects $Z$ at one point.
Moreover, $\xi(L')$ does not intersect bisecants of~$Z$, since $L'$ does not intersect flopping lines.
Hence $\xi(L')$ corresponds to a point of $\Sigma^0_Z(Y)$.
Thus, the map $\xi$ is well defined on an open subscheme $\Sigma^0_L(X)$ as a map $\Sigma^0_L(X) \to \Sigma^0_Z(Y)$.
Conversely, if $L'$ corresponds to a point of $\Sigma^0_Z(Y)$, the map $\xi^{-1}$ is regular on $L'$,
and a computation similar to the above shows that $\xi^{-1}(L')$ is a line on $X$.
This defines a morphism $\Sigma^0_Z(Y) \to \Sigma_L^0(X)$.
The two morphisms are evidently mutually inverse.
\end{proof}
In the above lemma we do not claim that $\Sigma^0_Z(Y)$ is dense in $\Sigma_Z(Y)$.
In fact, $\Sigma_Z(Y)$ can have a component consisting of lines
meeting both $Z$ and a bisecant of $Z$ (cf.\ Lemma~\ref{lemma:sigma-z-y})
and then $\Sigma^0_Z(Y)$ is contained in the complement of this component.
\begin{corollary}
\label{corollary:sigma-x-gorenstein}
The Hilbert scheme of lines $\Sigma(X)$
is a Gorenstein curve.
\end{corollary}
\begin{proof}
As it was mentioned in the proof of Lemma~\ref{lemma:sigma-0-iso}, the number of lines in $X$ intersecting a given line $L$ is finite.
This means that the open subschemes $\Sigma^0_L(X)$ form an open covering of $\Sigma(X)$.
So, by Lemma~\ref{lemma:sigma-0-iso} it is enough to prove that $\Sigma^0_Z(Y)$ is Gorenstein.
Since $Z$ is a smooth curve, it is a locally complete intersection in $Y$.
Since the map $p\colon {\mathscr{L}}(Y) \to Y$ is finite, the scheme $p^{-1}(Z) \subset {\mathscr{L}}(Y)$ is also a locally complete intersection,
and since ${\mathscr{L}}(Y)$ is smooth, we conclude that $p^{-1}(Z)$ is a Gorenstein curve.
It remains to notice that the map $q\colon p^{-1}(Z) \to \Sigma_Z(Y)$ is an isomorphism over $\Sigma^0_Z(Y)$, hence the latter is also Gorenstein.
\end{proof}
The above argument also shows that the curve $\Sigma(X)$ has only planar singularities.
\subsection{Special Fano threefolds of genus $12$}
\label{subsection:v22-special}
In this section we construct some examples of Fano threefolds $X$ of genus 12 with infinite automorphism groups, and after that
we show that all $X$ with infinite automorphism groups are covered be these examples.
By Remark~\ref{remark:functoriality} to produce an example of such $X$,
it is enough to find a normal rational quintic $Z$ stabilized by an infinite subgroup of $\mathrm{Aut}(Y) \cong \mathrm{PGL}_2(\Bbbk)$.
Recall the notation for subgroups $B_2$, $U_2$, and $T_2$ of $\mathrm{PGL}_2(\Bbbk)$ introduced in~\S\ref{subsection:v5}.
\begin{example}
\label{example-MU}
Let $Z = Z_{\mathrm{MU}} \subset Y$ be the quintic of Remark~\ref{remark:curves-oorb2}.
The corresponding Fano threefold of index 1 and genus 12 has a faithful action of the subgroup $B_2 \subset \mathrm{PGL}_2(\Bbbk)$.
In Theorem~\ref{theorem:Prokhorov-g-12-finite} we prove it is the Mukai--Umemura threefold $X^{\mathrm{MU}}$ of Theorem~\ref{theorem-mu-22}.
\end{example}
\begin{example}[{\cite{Prokhorov-1990c}}] \label{example-V22-a}
The curve
\begin{equation}\label{eq:z:a}
Z_{\mathrm{a}} = \overline{ U_2 \cdot [\phi_6]}\subset Y\subset\mathbb{P}(M_6)
\end{equation}
is a normal rational quintic curve preserved by the subgroup $U_2 \subset \mathrm{PGL}_2(\Bbbk)$.
We have
\begin{equation*}
Z_{\mathrm{a}} \cap \Orb_3(Y) = U_2 \cdot [\phi_6] \cong \mathbb{A}^1,
\qquad
Z_{\mathrm{a}} \cap \Orb_2(Y) = \varnothing,
\qquad
Z_{\mathrm{a}} \cap \Orb_1(Y) = [x^6].
\end{equation*}
We denote by $X^{\mathrm{a}}$ the Fano threefold of index 1 and genus 12
corresponding to the quintic~$Z_{\mathrm{a}}$ via the construction of Theorem~\xref{theorem:double-projection}.
\end{example}
\begin{example}[{\cite{Prokhorov-1990c}}]\label{example-V22-m}
For every parameter $u \in \Bbbk$ put
\begin{equation}
\label{eq:normal-quintic}
\phi_{6,u}(x,y) =
\left(\begin{smallmatrix}
1 & u \\ 0 & 1
\end{smallmatrix}\right)
\cdot \phi_6 =
x(ux+y)(x^4 - (ux+y)^4).
\end{equation}
Clearly, one has
\begin{equation*}
[\phi_{6,u}] \in U_2 \cdot[\phi_6] \subset \Orb_3(Y) \subset Y.
\end{equation*}
Expanding the right side of~\eqref{eq:normal-quintic} we get
\begin{equation*}
\phi_{6,u}(x,y) = u(1-u^4)x^6 + (1 - 5u^4)x^5y - 10u^3x^4y^2 - 10u^2x^3y^3 - 5ux^2y^4 - xy^5.
\end{equation*}
If all the coefficients of this polynomial are non-zero, i.e., if
\begin{equation}\label{eq:uu15u1}
u(u^4-1)(5u^4-1) \ne 0,
\end{equation}
the closure of the $T_2$-orbit of $\phi_{6,u}$
\begin{equation}\label{eq:z:m}
Z_{\mathrm{m}}(u) = \overline{T_2 \cdot [\phi_{6,u}]}
\end{equation}
is a normal rational quintic curve preserved by the subgroup $T_2 \subset \mathrm{PGL}_2(\Bbbk)$.
If~\eqref{eq:uu15u1} fails the orbit closure is either the line $L_{xy}$ (if $u = 0$),
or a normal rational quartic curve (if $u^4 = 1$), or a singular rational quintic curve (if $u^4 = 1/5$).
We have
\begin{align*}
Z_{\mathrm{m}}(u) \cap \Orb_3(Y) &= T_2 \cdot [\phi_{6,u}] \cong \mathbb{A}^1 \setminus \{0\},\\
Z_{\mathrm{m}}(u) \cap \Orb_2(Y) &= [xy^5],\\
Z_{\mathrm{m}}(u) \cap \Orb_1(Y) &= [x^6].
\end{align*}%
We denote by $X^{\mathrm{m}}(u)$ the Fano threefold of index 1 and genus 12 corresponding to
the quintic $Z_{\mathrm m}(u)$ via the construction of Theorem~\xref{theorem:double-projection}.
\end{example}
In what follows we refer to varieties $X^{\mathrm{MU}}$, $X^{\mathrm a}$, and $X^{\mathrm m}(u)$ defined in Theorem~\ref{theorem-mu-22}
and Examples~\ref{example-V22-a} and~\ref{example-V22-m} as \emph{special} Fano threefolds of genus $12$.
According to Remark~\ref{remark:functoriality} and in view of the construction of the curves $Z_{\mathrm{MU}}$, $Z_{\mathrm{a}}$, and $Z_{\mathrm{m}}(u)$
we have $B_2 \subset \mathrm{Aut}(X^{\mathrm{MU}})$, $U_2 \subset \mathrm{Aut}(X^{\mathrm{a}})$, and $T_2 \subset \mathrm{Aut}(X^{\mathrm{m}}(u))$.
We already know that $\mathrm{Aut}(X^{\mathrm{MU}})$ is actually much bigger.
In~\S\ref{subsection:aut-explicit} we will show that the other two groups are slightly bigger as well.
\begin{remark}
The construction of varieties $X^{\mathrm{MU}}$ and $X^{\mathrm{a}}$ does not depend on any parameter, so these are single varieties.
On the contrary, the construction of $X^{\mathrm{m}}$ depends on the parameter~$u$.
In fact, this is not quite precise.
Indeed, let $\upzeta$ be a primitive fourth root of unity.
It is easy to see that the polynomials $\phi_{6,u}(x,y)$ and $\phi_{6,\upzeta u}(x,\upzeta y)$ are proportional,
hence the $T_2$-orbits of $[\phi_{6,u}]$ and $[\phi_{6,\upzeta u}]$ coincide.
Thus $Z_{\mathrm{m}}(u) = Z_{\mathrm{m}}(\upzeta u)$ and~\mbox{$X^{\mathrm{m}}(u) \cong X^{\mathrm{m}}(\upzeta u)$}.
So the space of parameters for the family of varieties $X^{\mathrm{m}}(u)$ is
\begin{equation*}
\Big({\mathbb{P}}_u^1 \setminus \{0, \sqrt[4]{1}, \sqrt[4]{1/5}, \infty\} \Big)/\boldsymbol{\mu}_4 =
{\mathbb{P}}_{u^4}^1 \setminus \{0,1,1/5,\infty\},
\end{equation*}
with $u^4$ being a coordinate.
\end{remark}
The next lemma shows that the quintics $Z_{\mathrm{MU}}$, $Z_{\mathrm{a}}$, and $Z_{\mathrm{m}}(u)$
exhaust all rational normal quintic curves in $Y$ with an infinite stabilizer inside $\mathrm{PGL}_2(\Bbbk)$.
\begin{lemma}
\label{lemma:z:classification}
Assume that $Z \subset Y$ is a rational normal quintic curve, invariant with respect to a non-trivial connected solvable algebraic group $B \subset \mathrm{PGL}_2(\Bbbk)$.
Then $Z$ is conjugate under the action of $\mathrm{Aut}(Y) = \mathrm{PGL}_2(\Bbbk)$ to one of the curves $Z_{\mathrm{MU}}$, $Z_{\mathrm{a}}$, and~$Z_{\mathrm{m}}(u)$
described by~\eqref{eq:z:mu}, \eqref{eq:z:a}, or~\eqref{eq:z:m}.
\end{lemma}
\begin{proof}
Since the subgroup $B \subset \mathrm{PGL}_2(\Bbbk)$ is conjugate to one of the subgroups $B_2$, $T_2$, or $U_2$
discussed in~\S\xref{subsection:v5},
we can assume without loss of generality that $B$ is one of these subgroups. Let us consider these cases one-by-one.
First, assume that $B = B_2$. Since $Z \cong {\mathbb{P}}^1$, every point of $Z$
has a nontrivial one-dimensional stabilizer in $B$, hence $Z \subset \oOrb_2(Y)$.
By Remark~\ref{remark:curves-oorb2} $Z$ is conjugate to~$Z_{\mathrm{MU}}$.
Moreover, the quintics conjugate to $Z_{\mathrm{MU}}$ are the only smooth rational quintics contained in~\mbox{$\oOrb_2(Y)$},
so from now on we may assume that $Z \not\subset \oOrb_2(Y)$.
An arbitrary point of the open orbit~\mbox{$\Orb_3(U)$} can be written as $[\varphi]$, where
\begin{equation*}
\varphi = fg(f^4 - g^4),
\end{equation*}
and $f$, $g$ are linear forms, so we may assume that $Z$ is the closure of the $B$-orbit
of~$[\varphi]$.
Now, assume that $B = U_2$. For general $f$ and $g$ the closure of the $U_2$-orbit of $[\varphi]$ is a curve of degree $6$.
For it to have degree 5, it is necessary for $\varphi$ to be divisible by $x$.
Conjugating by an element of $\mathrm{Aut}(Y)_\varphi \cong \mathrm{Oct}$, we may assume that $f = x$
(up to a scalar multiple). But then (again up to a scalar multiple) $\varphi$ should be equal to
\begin{equation*}
\phi_{6,u,v}(x,y) = x(ux+vy)(x^4 - (ux+vy)^4)
= \left(\begin{smallmatrix}
1 & u \\ 0 & v
\end{smallmatrix}\right)
\cdot \phi_6,
\qquad u \in \Bbbk,\ v \in \Bbbk^\times.
\end{equation*}
Such point is obtained from $[\phi_6]$ by a $B_2$-action. But the group $B_2$ normalizes the
subgroup~\mbox{$U_2 \subset \mathrm{PGL}_2(\Bbbk)$},
hence $Z$ is conjugate to the closure of the $U_2$-orbit of $[\phi_6]$, i.e.
to~$Z_{\mathrm{a}}$.
Finally, assume that $B = T_2$.
Again, for general $f$ and $g$ the closure of the $T_2$-orbit of $[\varphi]$ has degree 6, and the degree is smaller if and only if $\varphi$ is divisible by $x$ or $y$.
Conjugating by an element of $\mathrm{Aut}(Y)_\varphi \cong \mathrm{Oct}$ we again may assume that $f= x$, i.e.\ $\varphi = \phi_{6,u,v}$.
But this point is in the $T_2$-orbit of $[\phi_{6,u}]$, hence $Z$ is conjugate to~$Z_{\mathrm{m}}(u)$.
\end{proof}
Now finally, we can classify all Fano threefolds of index 1 and genus 12 with infinite automorphism groups,
which is the first main result of this section.
\begin{theorem}[{see \cite{Prokhorov-1990c}}]
\label{theorem:Prokhorov-g-12-finite}
Let $X$ be a \textup(smooth\textup) Fano threefold with $\rho(X) = 1$, $\iota(X) = 1$, and $\operatorname{g}(X) = 12$.
Then the automorphism group of $X$ is finite unless $X$ is a special Fano threefold of genus~$12$.
More precisely, if $X$ admits a faithful action of the group $B_2$, then $X \cong X^{\mathrm{MU}}$ is
the Mukai--Umemura threefold described in Theorem~\xref{theorem-mu-22} and mentioned in Example~\xref{example-MU}.
Otherwise, either $X \cong X^{\mathrm{a}}$ or $X \cong X^{\mathrm{m}}(u)$, see Examples~\xref{example-V22-a} and~\xref{example-V22-m}.
\end{theorem}
\begin{proof}
Let $B$ denote a maximal solvable (Borel) subgroup of the connected component~\mbox{$\mathrm{Aut}^0(X)$} of identity in the group~\mbox{$\mathrm{Aut}(X)$}.
The group $\mathrm{Aut}(X)$ is finite if and only if $B$ is trivial.
The group $B$ acts on the Hilbert scheme $\Sigma(X)$ of lines on $X$, and by the fixed-point theorem~\mbox{\cite[Theorem~VIII.21.2]{Humphreys1975}}
there exists a $B$-invariant line~\mbox{$L\subset X$}.
Then by Remark~\ref{remark:functoriality}
the associated curve~$Z$ is $B$-invariant and by Lemma~\xref{lemma:z:classification} the curve~$Z$ is $\mathrm{Aut}(Y)$-conjugate
to one of the curves $Z_{\mathrm{MU}}$, $Z_{\mathrm{a}}$, or $Z_{\mathrm{m}}(u)$, hence $X$ is isomorphic
to one of the special threefolds of Example~\ref{example-MU}, \ref{example-V22-a}, or~\ref{example-V22-m}.
Moreover, if $X$ is the Mukai--Umemura threefold of Theorem~\ref{theorem-mu-22} then $B = B_2$,
hence the corresponding quintic is $B_2$-invariant, hence is conjugate to $Z_{\mathrm{MU}}$.
Therefore, the threefold of Example~\ref{example-MU} is the Mukai--Umemura threefold.
\end{proof}
To complete the proof of Theorem~\ref{theorem:Prokhorov} it remains to describe explicitly
the automorphism groups of the threefolds $X^{\mathrm{a}}$ and $X^{\mathrm{m}}(u)$.
We do this in the next subsection.
\begin{remark}
One can also use the approach of Theorem~\xref{theorem:Prokhorov-g-12-finite} to establish finiteness of the automorphism group
of an arbitrary smooth Fano threefold $X$ with $\rho(X)=1$, $\iota(X)=1$, and $7\leqslant \operatorname{g}(X)\leqslant 10$,
see \cite{Prokhorov-1990c} for details.
\end{remark}
\subsection{Explicit automorphisms groups}
\label{subsection:aut-explicit}
The main ingredient in the explicit description of the automorphisms groups of $X = X^{\mathrm{a}}$ and $X = X^{\mathrm{m}}(u)$
is the description of the Hilbert scheme $\Sigma(X)$ of lines on $X$.
For this
we use Lemma~\ref{lemma:sigma-0-iso} relating it to $\Sigma_Z(Y)$,
where $Z$ is the corresponding quintic curve.
Accordingly, we start by describing $\Sigma_Z(Y)$.
We include the case $Z = Z_{\mathrm{MU}}$ for completeness.
\begin{lemma}
\label{lemma:sigma-z-y}
If $Z = Z_{\mathrm{MU}}$, $Z = Z_{\mathrm{a}}$, or $Z = Z_{\mathrm{m}}(u)$,
then the curve $\Sigma_Z(Y)$ is a plane quintic curve which can be described by the following picture:
\begin{equation*}
\arraycolsep = 3em
\begin{array}{ccc}
\begin{tikzpicture}[xscale = .7, yscale = .8]
\draw[very thick, double] (0,0) ellipse (3em and 10ex);
\draw[thick] (-1.7,1.8) -- (0,1.8) node [above] {$P$} --(1.7,1.8) node [right] {$\ell$};
\draw[fill] (0,1.8) circle (.3em);
\end{tikzpicture}
&
\begin{tikzpicture}[xscale = .7, yscale = .8]
\draw[thick] (0,0) ellipse (3em and 10ex);
\draw[thick] (0,.52) ellipse (2em and 7ex);
\draw[thick] (-1.7,1.8) -- (0,1.8) node [above] {$P$} --(1.7,1.8) node [right] {$\ell$};
\draw[fill] (0,1.8) circle (.2em);
\end{tikzpicture}
&
\begin{tikzpicture}[xscale = .7, yscale = .8]
\draw[thick] (0,0) ellipse (3em and 10ex);
\draw[thick] (0,0) ellipse (2em and 10ex);
\draw[thick] (-1.7,1.8) -- (0,1.8) node [above] {$P$} --(1.7,1.8) node [right] {$\ell$};
\draw[fill] (0,1.8) circle (.2em);
\end{tikzpicture}
\\
\Sigma_{Z_{\mathrm{MU}}}(Y) &
\Sigma_{Z_{\mathrm{a}}}(Y) &
\Sigma_{Z_{\mathrm{m}}(u)}(Y)
\end{array}
\end{equation*}
In other words, $\Sigma_Z(Y)$ is the union of a line $\ell$ and two conics \textup{(}or a double conic, in the Mukai--Umemura case\textup{)}
tangent to $\ell$ at a certain point $P \in \Sigma(Y)$.
\end{lemma}
\begin{proof}
To describe $\Sigma_Z(Y)$ we use consecutively Corollary~\ref{corollary:V5-lines-description}.
First, put $Z = Z_{\mathrm{a}} = \overline{U_2 \cdot [\phi_6]}$.
Clearly, one has
\begin{equation*}
q(p^{-1}([\phi_6])) = \{ [xy], [x^2-y^2], [x^2+y^2] \},
\end{equation*}
Hence
\begin{align*}
q(p^{-1}(U_2 \cdot [\phi_6])) &= (U_2 \cdot [xy]) \cup (U_2 \cdot [x^2 - y^2]) \cup (U_2 \cdot [x^2 + y^2]) =
\\
&= \{x(sx + y)\} \cup \{ x^2 - (sx + y)^2\} \cup \{ x^2 + (sx + y)^2\},
\end{align*}
where $s \in \Bbbk$.
The point at the boundary of $Z_{\mathrm{a}}$ is $[x^6]$ and $q(p^{-1}([x^6])) = [x^2]$.
We see that~$\Sigma_{Z_{\mathrm{a}}}(Y) = q(p^{-1}(Z_{\mathrm{a}}))$ is the union of a line
\begin{equation*}
\ell = \{ x(s_1x + s_2 y) \},
\end{equation*}
and two conics
\begin{equation*}
\gamma'_{\mathrm{a}} = \{ (s_1^2 - s_2^2)x^2 - 2s_1s_2xy - s_1^2y^2 \},
\qquad
\gamma''_{\mathrm{a}} = \{ (s_1^2 + s_2^2)x^2 + 2s_1s_2xy + s_1^2y^2 \},
\end{equation*}
tangent to it (and tangent to each other with multiplicity~4) at the point $P = [x^2]$.
If $Z = Z_{\mathrm{m}}(u) = \overline{T_2 \cdot [\phi_{6,u}]}$, then
\begin{equation*}
q(p^{-1}([\phi_{6,u}])) = \{ [x(ux + y)], [x^2-(ux + y)^2], [x^2+(ux + y)^2] \},
\end{equation*}
hence
\begin{align*}
q(p^{-1}(T_2 \cdot [\phi_{6,u}])) &= (T_2 \cdot [x(ux + y)]) \cup (T_2 \cdot [x^2 - (ux + y)^2]) \cup (T_2 \cdot [(x^2 + (ux + y)^2])=
\\
&= \{x(ux + ty)\} \cup \{ x^2 - (ux + ty)^2\} \cup \{ x^2 + (ux + ty)^2\},
\end{align*}
where $t \in \Bbbk^\times$.
The points at the boundary of $Z_{\mathrm{m}}(u)$ are $[x^6]$ and $[xy^5]$, and we have~\mbox{$q(p^{-1}([x^6])) = [x^2]$} and $q(p^{-1}([xy^5])) = \{[xy], [y^2]\}$.
Thus, $\Sigma_{Z_{\mathrm{m}}}(Y) = q(p^{-1}(Z_{\mathrm{m}}(u)))$ is the union of the line $\ell$ (the same line as in the case of $Z = Z_{\mathrm{a}}$)
and two conics
\begin{equation*}
\gamma'_{\mathrm{m}}(u) = \{ s_1^2(1 - u^2)x^2 - 2s_1s_2uxy - s_2^2y^2 \},
\quad
\gamma''_{\mathrm{m}}(u) = \{ s_1^2(1 + u^2)x^2 + 2s_1s_2uxy + s_2^2y^2 \},
\end{equation*}
tangent to~$\ell$ at the point $P = [x^2]$, and also tangent to each other with multiplicity~2 at the points~$[x^2]$ and~$[y^2]$ respectively.
Finally, if $Z = Z_{\mathrm{MU}} = \{x(s_1x + s_2y)^5\}_{(s_1:s_2) \in \mathbb{P}^1}$, then
\begin{equation*}
\Sigma_{Z_{\mathrm{MU}}(Y)} = q(p^{-1}(Z_{\mathrm{MU}})) = \{ x(s_1x + s_2y) \} \cup \{ (s_1x + s_2y)^2 \}.
\end{equation*}
Thus, $\Sigma_{Z_{\mathrm{MU}}}(Y)$ is the union of the line $\ell$ (the same line again) and of the conic
\begin{equation*}
\gamma_{\mathrm{MU}} = \{ (s_1x + s_2y)^2 \}
\end{equation*}
tangent to~$\ell$ at the point $P = [x^2]$.
Since the lines parameterized by this conic are special, its preimage $q^{-1}(\gamma_{\mathrm{MU}})$ is the ramification divisor of $p \colon {\mathscr{L}}(Y) \to Y$,
hence the component of $\Sigma_{Z_{\mathrm{MU}}}(Y)$ underlying the conic $\gamma_{\mathrm{MU}}$ is everywhere non-reduced.
Summarizing, we can write
\begin{equation*}
\Sigma_{Z}(Y) =
\begin{cases}
\ell \cup 2\gamma_{\mathrm{MU}}, & \text{if $Z = Z_{\mathrm{MU}}$},\\
\ell \cup \gamma'_{\mathrm{a}} \cup \gamma''_{\mathrm{a}}, & \text{if $Z = Z_{\mathrm{a}}$},\\
\ell \cup \gamma'_{\mathrm{m}}(u) \cup \gamma''_{\mathrm{m}}(u), & \text{if $Z = Z_{\mathrm{m}}(u)$}.
\end{cases}
\end{equation*}
This completes the proof of the lemma.
\end{proof}
Another observation that we need is the following.
Denote by $\langle Z \rangle$ the linear span of the quintic $Z$.
It is a hyperplane in $\mathbb{P}(M_6) = \mathbb{P}^6$.
\begin{lemma}
\label{lemma:bisecant}
Let $Z = Z_{\mathrm{MU}}$, $Z = Z_{\mathrm{a}}$, or $Z = Z_{\mathrm{m}}(u)$ and
\begin{equation*}
F = Y \cap \langle Z \rangle.
\end{equation*}
Then $F$ is a non-normal quintic surface whose normalization is the Hirzebruch surface~$\mathrm{F}_3$.
The normalization map $\mathrm{F}_3 \to F$
glues the exceptional section with one fiber of $\mathrm{F}_3$ into the line $L_{x^2} = \mathrm{Sing}\,(F)$.
The line $L_{x^2}$ is the unique bisecant of $Z$ and corresponds to the distinguished point $P \in \Sigma_Z(Y)$.
Any line on $Y$ intersecting both $Z$ and $L_{x^2}$ is the image of the fiber of $\mathrm{F}_3$; these lines are parameterized by the component $\ell \subset \Sigma_Z(Y)$.
\end{lemma}
\begin{proof}
In all three cases the linear span $\langle Z \rangle$ is the hyperplane
\begin{equation*}
\langle x^6, x^5y, x^4y^2,x^3y^3, x^2y^4, xy^5 \rangle \subset \mathbb{P}(M_6),
\end{equation*}
so~$F$ is the corresponding hyperplane section of $Y$.
In particular, it is a quintic surface.
The line $L_{x^2}$ is contained in the hyperplane, hence also in $F$, and the same is true for any line~$L_{x(s_1x + s_2y)}$ parameterized by $\ell \subset \Sigma_Z(Y)$.
The lines $L_{x^2}$ and $L_{x(s_1x + x_2y)}$ meet at the point~$[x^5(s_1x+s_2y)]$, so the surface $F$ is swept out by secants of $L_{x^2}$.
Therefore, applying the main result of~\cite{FurushimaTada1989} we conclude that $F$ is non-normal and its normalization is a Hirzebruch surface.
Moreover, since the line~$L_{x^2}$ is special (see Corollary~\ref{corollary:V5-lines-description}), the normalization of~$F$ is~$\mathrm{F}_3$ and the normalization map
glues the exceptional section of $\mathrm{F}_3$ with a fiber.
Let $s$ denote the class of the exceptional section of $\mathrm{F}_3$, and $f$ the class of a fiber.
Since the fibers of $\mathrm{F}_3$ are mapped to lines on $Y$, and since the image of $\mathrm{F}_3$ is a quintic surface, the map $\mathrm{F}_3 \to F \to \langle Z \rangle \cong \mathbb{P}^5$
is given by an incomplete linear subsystem in $|s + 4f|$.
Let us check that the curve $Z$ is also the image of a member of the same linear system.
Indeed, $Z$ is a smooth quintic curve, and $|s + 4f|$ is the only linear system that contains integral curves of degree 5 with respect to $s + 4f$.
Now, we can check the last two assertions of the lemma.
Any bisecant of $Z$ is contained in the linear span $\langle Z \rangle$, hence in the surface $F$.
Therefore, it is the image of a fiber of $\mathrm{F}_3$.
Since $(s+4f) \cdot f = 1$, the image of a fiber intersects $Z$ in a single point and the intersection is transversal,
unless this is the fiber that is glued with the exceptional section.
This shows that $L_{x^2}$ is the unique bisecant of $Z$.
Finally, any line intersecting both $Z$ and $L_{x^2}$ is also contained in $\langle Z \rangle$, hence lies on $F$, hence is the image of a fiber of $\mathrm{F}_3$.
And as we have seen above, these lines are parameterized by $\ell$.
\end{proof}
Lemma~\ref{lemma:bisecant} allows to describe the flopping locus of the birational transformation $\upchi$
in the diagram~\eqref{equation(1)}.
\begin{proposition}\label{proposition:flopping-loci}
Let $X = X^{\mathrm{MU}}$, $X = X^{\mathrm{a}}$, or $X = X^{\mathrm{m}}(u)$ and $Z \subset Y$ is the corresponding quintic curve.
The flopping loci of the birational transformation $\upchi$
in the diagram~\eqref{equation(1)} is the exceptional section of the exceptional divisor $E_L \subset X'$,
and the strict transform of the unique bisecant $L_{x^2}$ of $Z$ in $Y'$.
In particular, the line $L \subset X$ is special and does not intersect any other line on $X$.
\end{proposition}
\begin{proof}
By Lemma~\ref{lemma:flopping-locus} the flopping locus in $Y'$ consists of strict transforms of bisecants of~$Z$.
So, by Lemma~\ref{lemma:bisecant} there is a unique flopping curve for $\upchi^{-1}$.
Consequently, the same is true for the map~$\upchi$.
On the other hand, the surface $F$ is the image of the exceptional divisor $E_L$.
Since its normalization is isomorphic to $\mathrm{F}_3$, it follows that $L$ is a special line on $X$.
Indeed, otherwise $E_L\cong \mathrm{F}_1$ and the map $\sigma_Y \circ \upchi$ is regular near a general fiber
of $E_L$. Then the image of fibers must be irreducible conics on $F$.
On the other hand, $F$ does not contain irreducible conics.
Hence, the exceptional section of the exceptional divisor is in the flopping locus.
But as we have already shown, the flopping locus consists of a single curve.
This means that no other line on $X$ intersects $L$.
\end{proof}
Combining the above assertions we obtain the following
\begin{proposition}
\label{proposition:lines-special-v22}
The Hilbert scheme of lines on a special Fano threefold $X$ of genus~$12$ has the following description:
\begin{itemize}
\item If $X = X^{\mathrm{MU}}$ then $\Sigma(X)$ is a smooth rational curve with an non-reduced scheme structure.
\item If $X = X^{\mathrm{a}}$ then $\Sigma(X)$ is the union of two rational curves glued at a point $P$, such that $\mathrm{Sing}\,(\Sigma(X)) = P$.
\item If $X = X^{\mathrm{m}}(u)$ then $\Sigma(X)$ is the union of two smooth rational curves glued at two simple tangency points $P$ and $P'$.
\end{itemize}
\end{proposition}
\begin{warning}
One can actually prove that in the case $X = X^{\mathrm{a}}$ the components of~$\Sigma(X)$ are smooth and the point $P$ is their tangency point of multiplicity 4.
Moreover, in fact in all cases considered in Proposition~\ref{proposition:lines-special-v22} the Hilbert scheme $\Sigma(X)$
has a natural structure of a plane quartic (see also Remark~\ref{remark-V22-VSP}),
which is either a double conic, or a union of two conics with a single common point, or a union of two conics with two tangency points,
so the picture below is adequate.
However, we do not need all these facts, and the proof that we have in mind requires going in too much details, so we skip it.
\end{warning}
The next picture shows how $\Sigma(X)$ looks:
\begin{equation*}
\arraycolsep = 4em
\begin{array}{ccc}
\begin{tikzpicture}[xscale = .7, yscale = .8]
\draw[very thick, double] (0,0) ellipse (3em and 10ex);
\draw (0,1.8) node [below] {$P$};
\draw (0,-1.8) node [above] {$P'$};
\draw[fill] (0,1.8) circle (.3em);
\draw[fill] (0,-1.8) circle (.3em);
\end{tikzpicture}
&
\begin{tikzpicture}[xscale = .7, yscale = .8]
\draw[thick] (0,0) ellipse (3em and 10ex);
\draw[thick] (0,.52) ellipse (2em and 7ex);
\draw (0,1.8) node [below] {$P$};
\draw[fill] (0,1.8) circle (.2em);
\end{tikzpicture}
&
\begin{tikzpicture}[xscale = .7, yscale = .8]
\draw[thick] (0,0) ellipse (3em and 10ex);
\draw[thick] (0,0) ellipse (2em and 10ex);
\draw (0,1.8) node [below] {$P$};
\draw (0,-1.8) node [above] {$P'$};
\draw[fill] (0,1.8) circle (.2em);
\draw[fill] (0,-1.8) circle (.2em);
\end{tikzpicture}
\\
\Sigma(X^{\mathrm{MU}}) &
\Sigma(X^{\mathrm{a}}) &
\Sigma(X^{\mathrm{m}}(u))
\end{array}
\end{equation*}
\begin{proof}
Let $L$ be the line on $X$ obtained from the pair $(Y,Z)$ by the construction of Theorem~\ref{theorem:double-projection}(ii).
By Proposition~\ref{proposition:flopping-loci} and Lemmas~\ref{lemma:sigma-0-iso} and~\ref{lemma:bisecant} we have an isomorphism
\begin{equation*}
\Sigma(X) \setminus \{[L]\} = \Sigma_L^0(X) \cong \Sigma_Z^0(Y) = \Sigma_Z(Y) \setminus \ell.
\end{equation*}
Thus, $\Sigma(X)$ is a one-point compactification of $\Sigma_Z(Y) \setminus \ell$.
Using the description of~$\Sigma_Z(Y)$ given in Lemma~\ref{lemma:sigma-z-y} we deduce all assertions of the proposition,
except for the local description of $\Sigma(X)$ at the point $P$ (corresponding to the line $L$) in the Mukai--Umemura and multiplicative cases.
For this we can argue as follows.
First, replace the line $L$ with the line $L'$ corresponding to the point $P'$ (any other point in the Mukai--Umemura case,
and the other singular point in the multiplicative case) and consider the quintic curve~$Z' \subset Y$ associated with the pair $(X,L')$.
By Lemma~\ref{lemma:z:classification} we conclude that~$Z'$ is conjugate to~$Z_{\mathrm{m}}(u')$,
for some $u'$ possibly different from $u$, with respect to the $\mathrm{Aut}(Y)$-action,
so it follows that
the local behavior of $\Sigma(X)$ at $P$ is the same as at $P'$.
\end{proof}
Now we are ready to prove the second main result of this section.
\begin{proposition}
\label{proposition:automorphisms-special-v22}
The automorphism groups of special Fano threefolds of genus~$12$ are the following:
\begin{equation*}
\mathrm{Aut}(X^{\mathrm{MU}}) \cong \mathrm{PGL}_2(\Bbbk),
\qquad
\mathrm{Aut}(X^{\mathrm{a}}) \cong \mathbb{G}_{\mathrm a} \rtimes \boldsymbol{\mu}_4,
\qquad
\mathrm{Aut}(X^{\mathrm{m}}(u)) \cong \mathbb{G}_{\mathrm m} \rtimes \boldsymbol{\mu}_2.
\end{equation*}
\end{proposition}
\begin{proof}
The first isomorphism is given by Theorem~\ref{theorem-mu-22}, so we concentrate on the other two.
The group~$\mathrm{Aut}(X)$ acts on the Hilbert scheme $\Sigma(X)$, and as a consequence on the set~$\mathrm{Sing}\,(\Sigma(X))$,
which by Proposition~\ref{proposition:lines-special-v22} is a single point in the case $X = X^{\mathrm a}$ and a two-point set in the case~$X = X^{\mathrm m}(u)$.
Let $L \subset X$ be the line corresponding to the singular point~$P$ of $\Sigma(X)$,
and let $\mathrm{Aut}(X; L) \subset \mathrm{Aut}(X)$ be the subgroup that preserves~$L$.
Then we have an equality
\begin{equation*}
\mathrm{Aut}(X^{\mathrm a}) = \mathrm{Aut}(X^{\mathrm a}; L),
\end{equation*}
and an exact sequence
\begin{equation}
\label{eq:aut-xm-sequence}
1 \to \mathrm{Aut}(X^{\mathrm m}(u); L) \to \mathrm{Aut}(X^{\mathrm m}(u)) \to \boldsymbol{\mu}_2,
\end{equation}
where the group $\boldsymbol{\mu}_2$ is considered as the group of permutations of the set $\{P,P'\}$.
On the other hand, we have an isomorphism~\eqref{eq:aut-xl-yz}.
A simple computation shows that
\begin{equation*}
\mathrm{Aut}(X^{\mathrm a}; L) \cong {} \mathrm{Aut}(Y; Z_{\mathrm a}) = U_2 \rtimes \boldsymbol{\mu}_4 \cong \mathbb{G}_{\mathrm a} \rtimes \boldsymbol{\mu}_4.
\end{equation*}
Indeed, if $g \in \mathrm{Aut}(Y)$ preserves $Z_{\mathrm a} = \overline{U_2 \cdot [\phi_6]}$,
then $g([\phi_6]) =[\phi_{6,u}]$ for some $u \in U_2$,
hence~$u^{-1}g$ is an element of the stabilizer $\mathrm{Oct}$ of the point $[\phi_6]$ that preserves $U_2$.
Therefore, we have $u^{-1}g \in \boldsymbol{\mu}_4$, where $\boldsymbol{\mu}_4$ is the subgroup of the octahedral group fixing one of the octahedron axes.
It is generated by the element
\begin{equation*}
\tau=
\left(\begin{smallmatrix}
\upzeta & 0\\ 0&1
\end{smallmatrix}\right)
\in \mathrm{PGL}_2(\Bbbk),
\end{equation*}
where $\upzeta$ is a fourth root of unity, and we finally get the required description of $\mathrm{Aut}(X^{\mathrm{a}})$.
Similarly,
\begin{equation*}
\mathrm{Aut}(X^{\mathrm m}(u); L) \cong {} \mathrm{Aut}(Y; Z_{\mathrm m}(u)) = T_2 \cong \mathbb{G}_{\mathrm m}.
\end{equation*}
Indeed, if $g \in \mathrm{Aut}(Y)$ preserves $Z_{\mathrm m}(u) = \overline{T_2 \cdot [\phi_{6,u}]}$, then $g([\phi_{6,u}]) = t([\phi_{6,u}])$ for some~\mbox{$t \in T_2$},
hence $t^{-1}g$ is an element of the stabilizer $u \cdot \mathrm{Oct} \cdot u^{-1}$ of the point $[\phi_{6,u}]$ (where we consider $u$ as an element of $U_2$) that preserves $T_2$.
In other words,
\begin{equation*}
t^{-1} g \in (u \cdot \mathrm{Oct} \cdot u^{-1}) \cap (T_2 \cup w(T_2)),
\end{equation*}
where $w$ is the nontrivial element of the Weyl group ${\mathfrak S}_2$ of $\mathrm{PGL}_2(\Bbbk)$.
But for $u$ satisfying the inequality of Example~\ref{example-V22-m} the intersection on the
right hand side is trivial, and we finally see that $g \in T_2$.
To conclude
we note that by~\cite[Proposition~5.1]{Dinew-Kapustka-Kapustka-2015}
the group $\mathrm{Aut}(X^{\mathrm{m}}(u))$ contains an extra involution
hence~\mbox{$\mathrm{Aut}(X^{\mathrm{m}}(u)) \ne \mathrm{Aut}(X^{\mathrm{m}}(u); L)$},
so the second map in~\eqref{eq:aut-xm-sequence} is surjective, and we get the required description of $\mathrm{Aut}(X^{\mathrm{m}}(u))$.
\end{proof}
\begin{remark}
\label{remark-V22-VSP}
According to S.\,Mukai \cite{mukai1989biregular} (see also \cite{Mukai-1992}, \cite{Schreyer2001}) any
Fano threefold $X$ with $\rho(X)=1$, $\iota(X)=1$, and $\operatorname{g}(X)=12$
can be realized as a variety of sums of powers that parameterizes
polar hexagons of a plane quartic curve $\mathfrak{C}$ (see \cite[\S5]{Mukai-1992} for a definition and~\cite{Dinew-Kapustka-Kapustka-2015} for some details),
and the Hilbert scheme $\Sigma(X)$ is the \emph{Scorza transform} of
$\mathfrak{C}$, (see~\cite[\S7]{DolgachevKanev} for a definition).
Unfortunately, a complete proof of these facts is not yet published,
while the construction of the extra involution in~\cite{Dinew-Kapustka-Kapustka-2015} relies on them.
To establish the existence of an involution in $\mathrm{Aut}(X^{\mathrm{m}}(u))\setminus \mathbb{G}_{\mathrm m}$,
independent of the above Mukai's results,
one can use another equivariant Sarkisov link similar to~\eqref{equation(1)},
see~\cite{Kuznetsov-Prokhorov-Gm} for details.
\end{remark}
Propositions~\xref{proposition:nearly-done} and~\xref{proposition:automorphisms-special-v22}
together with Theorem~\ref{theorem:Prokhorov-g-12-finite} (and a classification of smooth Fano threefolds of Picard rank $1$, see~\cite[\S12.2]{Iskovskikh-Prokhorov-1999})
give a proof of Theorem~\xref{theorem:Prokhorov}.
|
\chapter{#1}}
\newcommand{\sossub}[1]{\section{#1}}
\newcommand{\et}[1]{{\large\begin{flushleft} \color{blue}\textbf{#1} \end{flushleft}}}
\newcommand{\sto}[1]{\begin{center} \textit{#1} \end{center}}
\newcommand{\rf}[1]{{\color{blue}[\textit{#1}]}}
\newcommand{\el}[1]{\label{#1}}
\newcommand{\er}[1]{\eqref{#1}}
\newcommand{\df}[1]{\textbf{#1}}
\newcommand{\mdf}[1]{\pmb{#1}}
\newcommand{\ft}[1]{\footnote{#1}}
\newcommand{\n}[1]{$#1$}
\newcommand{\fals}[1]{$^\times$ #1}
\newcommand{{\color{red}$^{NEW}$ }}{{\color{red}$^{NEW}$ }}
\newcommand{\ci}[1]{}
\newcommand{\de}[1]{{\color{green}\underline{#1}}}
\newcommand{\rangle}{\rangle}
\newcommand{\langle}{\langle}
\newcommand{\left(}{\left(}
\newcommand{\right)}{\right)}
\newcommand{\right.}{\right.}
\newcommand{\left[}{\left[}
\newcommand{\right]}{\right]}
\newcommand{\Big(}{\Big(}
\newcommand{\Big)}{\Big)}
\newcommand{\nonumber \\}{\nonumber \\}
\newcommand{\partial}{\partial}
\newcommand{\pd}[1]{\frac {\partial} {\partial #1}}
\newcommand{\nabla}{\nabla}
\newcommand{$>$}{$>$}
\newcommand{\begin{eqnarray}}{\begin{eqnarray}}
\newcommand{\end{eqnarray}}{\end{eqnarray}}
\newcommand{\begin{equation}}{\begin{equation}}
\newcommand{\end{equation}}{\end{equation}}
\newcommand{\bay}[1]{\left(\begin{array}{#1}}
\newcommand{\end{array}\right)}{\end{array}\right)}
\newcommand{\textit{e.g.} }{\textit{e.g.} }
\newcommand{\textit{i.e.}, }{\textit{i.e.}, }
\newcommand{\iv}[1]{{#1}^{-1}}
\newcommand{\st}[1]{|#1\rangle}
\newcommand{\at}[1]{{\Big|}_{#1}}
\newcommand{\zt}[1]{\rm{#1}}
\newcommand{\zi}[1]{\textit{#1}}
\def{\alpha}{{\alpha}}
\def{\Alpha}{{\Alpha}}
\def{\beta}{{\beta}}
\def{\Beta}{{\Beta}}
\def{\delta}{{\delta}}
\def{\Delta}{{\Delta}}
\def{\epsilon}{{\epsilon}}
\def{\Epsilon}{{\Epsilon}}
\def{\varepsilon}{{\varepsilon}}
\def{\gamma}{{\gamma}}
\def{\Gamma}{{\Gamma}}
\def{\kappa}{{\kappa}}
\def{\Kappa}{{\Kappa}}
\def{\lambda}{{\lambda}}
\def{\Lambda}{{\Lambda}}
\def{\omega}{{\omega}}
\def{\Omega}{{\Omega}}
\def{\varphi}{{\varphi}}
\def{\sigma}{{\sigma}}
\def{\Sigma}{{\Sigma}}
\def{\theta}{{\theta}}
\def{\Theta}{{\Theta}}
\def {\rm Tr} {{\rm Tr}}
\def{\cal A}{{\cal A}}
\def{\cal C}{{\cal C}}
\def{\cal D}{{\cal D}}
\def{\cal E}{{\cal E}}
\def{\cal F}{{\cal F}}
\def{\cal H}{{\cal H}}
\def{\cal J}{{\cal J}}
\def{\cal K}{{\cal K}}
\def{\cal L}{{\cal L}}
\def{\cal M}{{\cal M}}
\def{\cal N}{{\cal N}}
\def{\cal O}{{\cal O}}
\def{\cal P}{{\cal P}}
\def{\cal Q}{{\cal Q}}
\def{\cal R}{{\cal R}}
\def{\cal S}{{\cal S}}
\def{\cal T}{{\cal T}}
\def{\cal V}{{\cal V}}
\def{\cal W}{{\cal W}}
\def{\cal Y}{{\cal Y}}
\def{\cal Z}{{\cal Z}}
\def\mathbb{C}{\mathbb{C}}
\def\mathbb{R}{\mathbb{R}}
\def\mathbb{Z}{\mathbb{Z}}
\def\mathscr{A}{\mathscr{A}}
\def\mathscr{B}{\mathscr{B}}
\def\mathscr{D}{\mathscr{D}}
\def\mathscr{E}{\mathscr{E}}
\def\mathscr{F}{\mathscr{F}}
\def\mathscr{G}{\mathscr{G}}
\def\mathscr{H}{\mathscr{H}}
\def\mathscr{J}{\mathscr{J}}
\def\mathscr{L}{\mathscr{L}}
\def\mathscr{M}{\mathscr{M}}
\def\mathscr{N}{\mathscr{N}}
\def\mathscr{O}{\mathscr{O}}
\def\mathscr{P}{\mathscr{P}}
\def\mathscr{R}{\mathscr{R}}
\def\mathscr{Q}{\mathscr{Q}}
\def\mathscr{X}{\mathscr{X}}
\defSL(2,\BZ){SL(2,\mathbb{Z})}
\def$SL(2,R)\times SL(2,R)$ {$SL(2,R)\times SL(2,R)$ }
\def${AdS}_5\times {S}^5$ {${AdS}_5\times {S}^5$ }
\def${AdS}_3$ {${AdS}_3$ }
\defSU(N){SU(N)}
\def\ad#1#2{{\frac \delta {\delta\sigma^{#1}} (#2)}}
\def\bar Q_{\bar f}{\bar Q_{\bar f}}
\defN_f{N_f}
\defSU(N_f){SU(N_f)}
\def{^\circ_\circ}{{^\circ_\circ}}
\def{B_\nu{}^\nu}{{B_\nu{}^\nu}}
\def{B_{yy}}{{B_{yy}}}
\def{\zt{sgn}(y)}{{\zt{sgn}(y)}}
\usepackage[usenames,dvipsnames]{xcolor}
\definecolor{orange}{cmyk}{0,0.5,1,0}
\definecolor{rossoCP3}{cmyk}{0,.88,.77,.40}
\definecolor{graa}{rgb}{0.8,0.8,0.8}
\definecolor{blaa}{rgb}{0.2,0.2,0.6}
\begin{document}
\title{\PRE{\vspace*{0.9in}} \color{rossoCP3}{Constraints on 750~GeV
colorless $\mathcal Q$-onia from running couplings
}}
\author{{\bf Luis A. Anchordoqui}}
\affiliation{Department of Physics and Astronomy,\\ Lehman College, City University of
New York, NY 10468, USA
\PRE{\vspace*{.05in}}
}
\affiliation{Department of Physics,\\
Graduate Center, City University
of New York, NY 10016, USA
\PRE{\vspace*{.05in}}
}
\affiliation{Department of Astrophysics,\\
American Museum of Natural History, NY
10024, USA
\PRE{\vspace*{.05in}}
}
\author{{\bf Haim \nolinebreak Goldberg}}
\affiliation{Department of Physics,\\
Northeastern University, Boston, MA 02115, USA
\PRE{\vspace*{.05in}}
}
\author{{\bf Xing Huang}}
\affiliation{Department of Physics, \\
National Taiwan Normal University, Taipei, 116, Taiwan
\PRE{\vspace*{.05in}}
}
\PRE{\vspace*{.15in}}
\begin{abstract}\vskip 2mm
\noindent We present yet another composite model explaining the
relatively broad peak in the diphoton invariant mass distribution
around 750 GeV recently observed at the LHC experiments. We consider
the excess originates in bound states of vector-like fermions $\cal
Q$ transforming under the electroweak group (but not QCD) of the
standard model and which are also charged under a new $SU(N_{\rm
TC})$ QCD-like confining force. Since the new uncolored fields
transform as $SU(2)$ doublets they can acquire a mass through Yukawa
interactions with the electroweak Higgs doublet, as quarks and
leptons. We parametrize the ${\mathcal Q}$-onium bound state using
the Coulomb approximation and give a numerical fit to the diphoton
excess consistent with no strong coupling effects up to about
30~TeV, which is the energy scale for next generation colliders.
The range of allowed hypercharge $Y_{\mathcal Q}$ is given by $2.26
\leq Y_{\mathcal Q} \leq 2.53$ for $N_{\rm TC} =2$ and $2.17 \leq
Y_{\mathcal Q} \leq 2.31$ for $N_{\rm TC} = 3$. The smoking gun for
the model would be the discovery of $\mathcal{Q}$-onia decaying into
Higgs and $Z$-boson.
\end{abstract}
\maketitle
\section{Introduction}
In December 2015, ATLAS~\cite{ATLAS} and CMS~\cite{CMS:2015dxe}
famously announced the observation of a peak in the diphoton mass
distribution around 750~GeV, using (respectively) $3.2~{\rm fb}^{-1}$
and $2.6~{\rm fb}^{-1}$ of data recorded at a center-of-mass energy
$\sqrt{s} = 13~{\rm TeV}$. The diphoton excesses could be interpreted
as the decay products of a new massive particle $X$, with spin 0, 2,
or higher~\cite{Landau:1948kw}. Assuming a narrow width approximation
ATLAS gives a local significance of $3.6\sigma$, or else a global
significance of $2.0\sigma$ when the look-elsewhere-effect in the mass
range $M_X/{\rm GeV} \in [200 - 2000]$ is accounted for. Signal-plus-background fits were also implemented for a broad
signal component with a large decay width. The largest
deviation from the background-only hypothesis corresponds to $M_X \sim
750~{\rm GeV}$ with a total width \mbox{$\Gamma_{\rm total} \sim
45~{\rm GeV}$.} The local and global significances evaluated for the
broad resonance fit are roughly 0.3 higher than that for the fit using
the narrow width approximation, corresponding to $3.9\sigma$ and
$2.3\sigma$, respectively. The CMS data gives a local significance of
$2.6\sigma$ and a global significance smaller than
$1.2\sigma$. Fitting the LHC data at $\sqrt{s} = 13~{\rm TeV}$ to a
resonant structure leads to a cross section times branching fraction of
\begin{equation}
\sigma_{\rm LHC13} (pp \to X +
{\rm
anything}) \times {\cal B} (X \to \gamma \gamma) \sim
\left\{\begin{array}{cl}
(10 \pm 3)~{\rm fb} &~~~ {\rm ATLAS} \\
\phantom{0}(6 \pm 3)~{\rm fb} &~~~ {\rm CMS} \end{array}
\right. \,,
\end{equation}
at $1\sigma$~\cite{Franceschini:2015kwy}. However, the LHC data at
$\sqrt{s} = 8~{\rm TeV}$ show no excess over standard model processes
(SM)~\cite{Aad:2015mna,Khachatryan:2015qba}, placing a critical upper
bound on the cross section times branching fraction: $\sigma_{\rm LHC8}
(pp \to X + {\rm anything}) \times {\cal B} (X \to \gamma
\gamma) < 2.00~{\rm fb}$ at 95\% CL~\cite{Khachatryan:2015qba}.
Altogether, the LHC8 data would be compatible with the LHC13 data within
$\sim 2 \sigma$, if the diphoton cross section grows by more than
about a factor of 3 or 3.5.
Quite recently, ATLAS and CMS updated their diphoton resonance
searches~\cite{Delmastro,Musella,CMS:2016owr}. ATLAS reanalyzed the
$3.2~{\rm fb}^{-1}$ of data, targeting separately spin-0 and spin-2
resonances. For spin-0, the most significant deviation from the
background-only hypothesis corresponds to $M_X \sim 750~{\rm GeV}$ and
$\Gamma_{\rm total} \sim 45~{\rm GeV}$. The local significance is now
increased to $3.9\sigma$ but the global significance remains at the $2
\sigma$ level. For the spin-2 resonance, both the local and global
significances are reduced down to $3.6\sigma$ and $1.8\sigma$,
respectively. The new CMS analysis includes additional data (recorded
in 2015 while the magnet was not operated) for a total of 3.3~${\rm
fb}^{-1}$. The largest excess is observed for $M_X = 760~{\rm GeV}$
and $\Gamma_{\rm total} \approx 11~{\rm GeV}$, and has a local
significance of $2.8\sigma$ for spin-0 and $2.9\sigma$ spin-2
hypothesis. After taking into account the effect of searching for
several signal hypotheses, the significance of the excess is reduced
to $< 1 \sigma$. CMS also communicated a combined search with data
recorded at $\sqrt{s} = 13~{\rm TeV}$ and $\sqrt{s} = 8~{\rm
TeV}$. For the combined analysis, the largest excess is observed at
$M_X = 750~{\rm GeV}$ and $\Gamma_{\rm total} = 0.1~{\rm GeV}$. The
local and global significances are $\approx 3.4\sigma$ and
$1.6\sigma$, respectively.
Among many possible interpretations~\cite{Staub:2016dxq}, one of the
most attractive type of models is that of pairs of new heavy fermions
which produce resonant signals of their near-threshold bound
states. It seems natural to classify the various models in accordance
with the fermion properties. The simplest explanation for the
structure underlying the bump in the diphoton spectrum could be that
of a QCD bound state of a heavy vector-like quark $Q$, with a mass around
375~GeV~\cite{Barger:1987xg,Kats:2012ym,Luo:2015yio,Zhang:2016xei,Han:2016pab,Kats:2016kuz,Hamaguchi:2016umx}.
For this model, the signal rate can be explained by a scalar $Q \bar
Q$ bound state $\eta_Q$, with $Q$ transforming as $(3, 1, -4/3)$ under
the $SU(3) \otimes SU(2) \otimes U(1)_Y$ gauge groups of the
SM~\cite{Han:2016pab,Kats:2016kuz}. Actually, in analogy to $\eta_c$
and $J/\psi$, the QCD bound states would contain both scalars and
vectors. We may note in passing that the vector $J/\psi$-like state
will not only decay into $\gamma \gamma$ but also into dilepton
topologies. The latter is severely constrained by
ATLAS~\cite{Aad:2014cka} and CMS~\cite{Khachatryan:2014fba} direct
searches; however, since the $J/\psi$-like state can only be produced
through $q \bar q$ annihilation, its production cross section at the
LHC is significantly smaller than that of the $Q \bar Q$ scalar state.
A somewhat related model interprets the excess of diphoton events as
arising also from the decay of a $\mathpzc Q$-onium state $\eta
_{\mathpzc Q}$, but with the colored fermions ${\mathpzc Q}$ bound by
a hidden confining
$SU(N)$~\cite{Harigaya:2015ezk,Nakai:2015ptz,Curtin:2015jcv,Bian:2015kjt,Craig:2015lra,Redi:2016kip,Kamenik:2016izk,Ko:2016sht,Bai:2016czm}. One
particularly interesting possibility to describe the hidden gauge
dynamics is given by $SU(N_{\rm TC})$, with $N_{\rm TF}$
techni-flavors. This is because the dynamics of this theory is
renowned from QCD and can be also understood in the large $N_{\rm TC}$
limit: the gauge theory is asymptotically free provided it satisfies
the familiar bound on the number of techini-flavors and confines at a
scale $\Lambda_{\rm TC}$. To avoid the strict constraints common to
old techni-color models the heavy-fermions must be in a vectorial
representation of the SM and in the fundamental $\bm{N}_{\rm TC}$ of
$SU(N_{\rm TC})$; namely
\begin{equation}
{\mathpzc Q} = \sum_{i =1}^{N_s} {\mathpzc Q}_{\ i}, \quad {\rm with}
\quad
{\mathpzc Q}_{\ i} =
(\bm{N}_{\rm TC}, \bm{R}_i) \oplus (\bm{\bar{ N}}_{\rm TC}, \bm{\bar{R}}_i) \,,
\label{vec-rep}
\end{equation}
where $\bm{R}_i$ denotes a generic SM representation and $N_s$ is the
number of species with mass below the confinement
scale~\cite{Redi:2016kip}. Yet a third dynamics is possible if
the heavy fermions $\mathcal{Q}$ are colorless particles, bound by a
confining $SU(N_{\rm TC})$~\cite{Iwamoto:2016ral}.
At this point it is important to stress two fundamental differences
between these three models. Firstly, while $\eta_Q$ and
$\eta_{\mathpzc Q}$ could be produced either via gluon or photon
fusion in LHC collisions, the production of $\eta_\mathcal{Q}$ could
only proceed via photon fusion. Secondly, if the heavy particles
belong to the $SU(2)$ doublet and feel the strong $SU(3)$ color
interactions, then the 125 GeV~Higgs signal strength in the $gg \to H
\to \gamma \gamma$ channel would be significantly modified.
ATLAS~\cite{Aad:2015gba} and CMS~\cite{Khachatryan:2014jba} data would
then place severe constraints on model parameters unless the coupling
of $Q$ or ${\mathpzc Q}$ to the Higgs is small. However,
$\mathcal Q$ transforms as an $SU(3)$ singlet and so it does not
contribute to $gg \to H \to \gamma \gamma$. Therefore, the $\mathcal Q$'s could
get mass like quarks and leptons, as $SU(2)$ doublets and singlets can form a
mass term after being Higgsed. This is precisely what motivates our
study.
In this paper we investigate the phenomenology of $\mathcal Q$-onia
states of a hidden $SU(N_{\rm TC})$. We combine the requirements to
reproduce the LHC diphoton signal with those arising from the
renormalization group (RG) equations to constrain the parameter space
of $\mathcal{Q}$ fermions transforming under the electroweak group
(but not QCD) of the SM. For related RG studies in which the 750~GeV
excitation is not a bound state,
see~\cite{Gu:2015lxj,Son:2015vfl,Goertz:2015nkp,Dev:2015vjd,Gross:2016ioi}. Before
proceeding we note that the LHC phenomenology of $\mathcal Q$-onia
states, with constituents that carry only hypercharge, but no SM
$SU(3)$ or $SU(2)$ quantum numbers has been presented
in~\cite{Iwamoto:2016ral}. Though at first sight our study seems quite
similar to the analysis in~\cite{Iwamoto:2016ral}, it differs in a
fundamental aspect: herein we consider colorless fermions that
transform as $SU(2)$ doublets and thus the heavy $\mathcal Q$
particles can obtain a mass, as all other SM fermions, through the
Higgs mechanism. The layout of the paper is as follows. In
Sec.~\ref{secII} we introduce the Lagrangian of the model with a new
confining gauge interaction and derive the RG equations. In
Sec.~\ref{secIII} we discuss the LHC phenomenology of our set up. We
parametrize the $\mathcal{Q}$-onium bound state using the Coulomb
approximation~\cite{Kats:2012ym} and give a numerical fit to the
diphoton excess consistent with the running couplings. We request no
strong coupling effects up to about 30~TeV, which is the energy scale
for next generation colliders~\cite{Gershtein:2013iqa}. In
Sec.~\ref{secIV} we verify consistency with early universe
cosmology. Our conclusions are collected in Sec.~\ref{secV}.
\section{Lagrangian and Renormalization Group equations}
\label{secII}
To develop our program in the simplest way, we will work within the construct of a minimal model in which we consider one generation of heavy colorless
fermions, which contains two $SU(2)$ doublets (one left- and one
right-handed to make the representation vectorial, i.e. symmetric
under parity) and four chiral fermions that transform as $SU(2)$
singlets (two of them are left-handed and the other two
right-handed). We label the four flavors as club,
diamond, heart, and spade, and so we named the $\mathcal Q$ particles
``qards'' accordingly.\footnote{Note that qards are quirks
transforming under the electroweak group (but not QCD) of the
SM~\cite{Kang:2008ea,Kribs:2009fy,Martin:2010kk,Harnik:2011mv,Fok:2011yc}.}
The Lagrangian for the technicolor qards is
given by
\begin{eqnarray} \nonumber
\mathscr{L}_{\mathcal Q} &=& \Big( i\overline {\bm{\mathcal Q}_{L}} \gamma
_{\mu}D^{\mu} \bm{\mathcal{Q}}_{L} + i\overline {\mathcal Q^{_\clubsuit}_{R}}
\gamma _{\mu} D^{\mu} \mathcal Q^{_\clubsuit}_R
+i\overline {\mathcal Q^{_\diamondsuit}_R} \gamma
_{\mu}D^{\mu} \mathcal Q^{_\diamondsuit}_R +\\
&+ & i\overline {\bm{{\cal Q}_{R}}} \gamma
_{\mu}D^{\mu} \bm{{\cal Q}}_{R} + i\overline
{\mathcal Q^{_\heartsuit} _{L}}
\gamma _{\mu}D^{\mu} \mathcal Q^{_\heartsuit}_L
+i\overline {\mathcal Q^{_\spadesuit}_L} \gamma
_{\mu}D^{\mu} \mathcal Q^{_\spadesuit}_L \Big) +\zt{h.c.}\, ,
\end{eqnarray}
where the covariant derivative for the $SU(2)$ doublets $\bm{{\cal Q}}_{L,R}$,
\begin{equation}
\bm{{\cal Q}}_L = \bay{c}
\mathcal Q^{_\clubsuit}_L \\ \mathcal Q^{_\diamondsuit}_L \end{array}\right) \quad
{\rm and} \quad \bm{{\cal Q}}_R = \bay{c}
\mathcal Q^{_\heartsuit}_R \\ \mathcal Q^{_\spadesuit}_R \end{array}\right) \,,
\end{equation}
reads
\begin{equation}
{D}_\mu = \partial_\mu - i g_2 \tau^a A^a_\mu - i g_{\zt{TC}} \tilde T^a
\tilde G^a_\mu- i g_Y Y_{\cal Q} B_\mu\,,
\end{equation}
while those for SU(2) singlets $\mathcal Q^{_\heartsuit}_{L}$, $\mathcal
Q^{_\spadesuit}_{L}$, $\mathcal Q^{_\clubsuit}_{R}$, $\mathcal Q^{_\diamondsuit}_{R}$ are (plus
sign for $\mathcal Q^{_\clubsuit}_R$, $\mathcal Q^{_\heartsuit}_L$)
\begin{equation}
{D}_\mu = \partial_\mu - i g_{\zt{TC}} \tilde T^a
\tilde G^a_\mu- i g_Y \left(Y_{\cal Q} \pm \tfrac 1 2 \right)B_\mu\,.
\end{equation}
The Yukawa interactions that provide the masses to the qards are:
\begin{equation}
\mathscr{L}_{\cal Y} = -{\cal Y}_{_{\! \diamondsuit}}\overline {\bm{{\cal Q}}_{L}} \cdot H \mathcal Q^{_\diamondsuit}_{R}
- {\cal Y}_{_{\! \clubsuit}}\overline {\bm{{\cal Q}}_{L}} \cdot
\widetilde H \mathcal Q^{_\clubsuit}_{R}
-{\cal Y}_{_{\!\spadesuit}}\overline {\mathcal Q^{_\spadesuit}_{L}} \bm{{\cal Q}}_{R}\cdot H
-{\cal Y}_{_{\! \heartsuit}}\overline {\mathcal Q^{_\heartsuit}_{L}} \bm{{\cal Q}}_{R}\cdot \widetilde H +\zt{h.c.}\label{L_Yukawa}
\end{equation}
where $\widetilde H=i\sigma^2 H^*$. The couplings of the technicolor qards
with the Higgs are the same as those with SM fermions and hence they get masses
in the same way. Substituting in (\ref{L_Yukawa}) the Higgs vacuum expectation value,
\begin{equation}
\langle H \rangle = \frac 1 {\sqrt 2} \bay{c}
0 \\ v \end{array}\right) ,
\end{equation}
we have the mass terms
\begin{equation}
\mathscr{L}_m = -\frac 1 {\sqrt 2} {\cal Y}_{_{\!\diamondsuit}}\overline {\mathcal
Q^{_\diamondsuit}_{L}} \mathcal Q^{_\diamondsuit}_{R}v
-\frac 1 {\sqrt 2} {\cal Y}_{_{\! \clubsuit}}\overline {\mathcal
Q^{_\clubsuit}_{L}} \mathcal Q^{_\clubsuit}_{R}v
-\frac 1 {\sqrt 2} {\cal Y}_{_ {\! \spadesuit}}\overline {\mathcal Q^{_\spadesuit}_{L}}
\mathcal Q^{_\spadesuit}_{R}v
-\frac 1 {\sqrt 2} {\cal Y}_{_ {\! \heartsuit}}\overline {\mathcal Q^{_\heartsuit}_{L}}
\mathcal Q^{_\heartsuit}_{R}v +\zt{h.c.}\label{L_Yukawa_m}
\end{equation}
The RG equation for the coupling of an
$SU(N_c)$ gauge field with $N_f$ fundamental matter reads
\begin{equation}
\frac {d }{d t}g(t)=-\frac{g^3 }{(4 \pi )^2}\left(\frac{11
N_c}{3}-\frac{2 N_f}{3}\right)\,.
\end{equation}
Throughout we take both the left and right-handed technicolor $SU(N_{\zt{TC}})$ qards to be in the representation of
\begin{equation} \el{t-rep}(\bm{ N}_{\zt{TC}},\df 1, \df 2 \oplus \df 1 \oplus \df 1)_Y \end{equation}
under $SU(N_{\zt{TC}}) \otimes SU(3) \otimes SU(2) \times U(1)_Y$,
satisfying a relation akin to (\ref{vec-rep}),
\begin{equation}
{\mathcal Q} =
(\bm{N}_{\rm TC}, \bm{1},\bm{2} \oplus \bm{1} \oplus \bm{1} )_Y \oplus (\bm{\bar{N}}_{\rm TC}, \bm{1},
\bm{2} \oplus \bm{1} \oplus \bm{1})_{-Y} \, .
\end{equation}
Thus, the RG equation become
\begin{equation}
\frac{d}{d t} g_{\zt{TC}} = \frac{g_{\rm TC}^3}{16\pi ^2}\left[ -\frac {11} 3 N_{\zt{TC}}+\frac{2}{3} \times
4 \right] = \frac{g_{\rm TC}^3}{16\pi ^2}\left( -\frac {11} 3 N_{\zt{TC}}+\frac
8 3\right)\, ,
\end{equation}
For the $SU(3)$ and $SU(2)$ gauge couplings $g_3$ and $g_2$, the RG
equations are
\begin{equation}
\frac{d}{dt}g_3 = \frac{g_3^3}{16\pi ^2}\left[ -11+\frac{2}{3}(2 \times
3)\right] = \frac{g_3^3}{16\pi ^2}\left( -7\right)\, ,\el{rg-g3}
\end{equation}
and
\begin{equation}
\frac{d}{dt}g_2 = \frac{g_2^3}{16\pi ^2}\left[ -\frac{22}{3}+\frac{2}{3}
\left(\frac
3 2 \times 3 + \frac 1 2 \times 3 + N_{\zt{TC}} \right)+\frac{1}{6}\right] =
\frac{g_2^3}{16\pi ^2}\left(-\frac{19}{6}+\frac 2 3 N_{\zt{TC}}\right)
\, .\el{rg-g2}
\end{equation}
The RG running of the hypercharge coupling can be worked out in a
similar fashion,
\begin{equation} \frac{d}{dt}g_Y = \frac{1}{16\pi
^2}\left\{\frac{2}{3} \sum _f Y_f^2 + \frac{1}{3}\sum _s Y_s^2 \right\}
g_Y^3 \, ,
\label{rgy}
\end{equation} where the sum is over the hypercharges of chiral
fermions $f$ and complex scalars $s$. For the case at hand (\ref{rgy}) can be
rewritten as
\begin{equation}
\frac{dg_Y}{dt} = \frac{g_Y^3}{16 \pi^2}
\left\{ \frac{2}{3}
\times 2 \times \left[2N_{\zt{TC}} Y_{\mathcal
Q}^2+N_{\zt{TC}} \left(Y_{\mathcal
Q}-\frac{1}{2}\right)^2+N_{\zt{TC}}
\left(Y_{\mathcal{Q}}+\frac{1}{2}\right)^2\right]+\frac {41}
6\right\} \!,
\end{equation}
where $Y_{\mathcal Q}$ is the hypercharge of the technicolor qards
in the $SU(2)$ doublet. Notice that the qards in the doublet of $SU(2)$
have charge $Y_{\mathcal Q}$ while those in the singlet have charges
$Y_{\mathcal Q} \pm
\frac 1 2$ in order to allow Yukawa coupling. The extra factor of $2$
follows from the fact that we have doublets for both left- and
right-handed qards.
We denote the Yukawa coupling of the qard with hypercharge
$Y_{\mathcal Q}\pm\frac 1 2$ as ${\cal Y}_\pm$. The RG equation for
${\cal Y}_\pm$ is given by
\begin{equation}\label{RGE_yuk_top}
\frac{d}{d t} {\cal Y}_\pm = \frac{{\cal Y}_\pm}{16\pi
^2}\left[\frac{2 N_{\zt{TC}}+3}{2} {\cal Y}_\pm^2+\frac 3 2
{\cal Y}_\mp^2 -\frac {3(N_{\zt{TC}}^2-1)} {N_{\zt{TC}}} g_{\zt{TC}}^2-
\frac{9}{4}g_2^2-\left(\frac{3}{4} \pm 3 Y_Q +6 Y_Q^2 \right) g_Y^2 \right]\, .
\end{equation}
The $SU(2)$ quantum number is the same
as in SM and hence so is the contribution from $SU(2)$ gauge fields. The only
difference is in the hypercharges. It can be easily checked that for
$Y_{\mathcal Q} = 1 /6$, the $g_Y$ contribution ($3/4+3 Y_{\mathcal Q} +6
Y_{\mathcal Q}^2$) gives $17/12$, which is the value for the SM.
\section{Playing qards at the LHC}
\label{secIII}
For simplicity, we further assume the initial value of ${\cal Y}_+ = {\cal Y}_-$, and so there are four possible ${\cal Q}$-onium bound states that can contribute the diphoton excess
\begin{equation}
\bar{\mathcal Q}^{_\heartsuit} \mathcal Q^{_\heartsuit},\quad \bar
{\mathcal Q}^{_\spadesuit} \mathcal Q^{_\spadesuit},\quad \bar{\mathcal
Q}^{_\clubsuit} \mathcal Q^{_\clubsuit},\quad \bar{\mathcal
Q}^{\diamond} \mathcal Q^{\diamond}\,,
\end{equation}
where we have already combined the left and right techni-qards to form
Dirac spinors. All of them are spin-0 ($\eta_{\cal Q}$) and have equal
masses $M_X \sim 2 m_{\cal Q}$.
Following~\cite{Barger:1987xg,Kats:2012ym} we describe the $SU(N_{\rm TC})$ binding potential in the Coulomb
approximation. The radial wave function $R^{(n,l)}(r)$ follows
from that of the hydrogen atom
\begin{equation}
\left(\frac{|R^{(n,l)}(0)|^2}{M_X^3}\right)_{\rm Coul} = \frac 1 {16
n^3} (C_N {\bar \alpha}_{\rm TC})^3 \,,
\el{wavefunction}
\end{equation}
where $n$ is the principal quantum number, $l$ is the orbital angular
momentum, and $\bar \alpha_{\rm TC} \equiv \alpha_{\rm TC} (a_0^{-1})$
is the techni-color gauge coupling in the MS scheme, with $a_0$ the
Bohr radius of the bound state~\cite{Iwamoto:2016ral}. For fermions in the
fundamental representation, $C_N = (N_{\rm TC}^2 -1)/(2N_{\rm TC})$.
The Coulomb approximation is reliable when the non-perturbative effect
is small. In what follows, we demand that the inverse Bohr radius, which is the characteristic
scale of the bound state dynamics, is above the confinement scale:
\begin{equation} \label{Coul-range}
a_0^{-1} \sim {\alpha}_{\rm TC}(a_0^{-1}) M_X > \Lambda_{\rm TC} \end{equation} where the
confinement scale ${\Lambda}_{\rm TC}$ reads \begin{equation} \Lambda_{\rm
TC}\sim M_X \exp\left[-\frac {6\pi}{(11 N_{\rm TC}-2N_f)
\alpha_{\rm TC}(M_X)}\right]\,. \end{equation}
Intuitively, (\ref{Coul-range}) states that the perturbative
treatment (like the Coulomb approximation) breaks down below the
confinement scale. The
value of $a_0^{-1}$ ranges from 58~GeV to about 320~GeV depending on
the value of $N_{\rm TC}$ and $Y_{\cal Q}$.
The four resonance states can be produced via photon fusion~\cite{Fichet:2015vvy,Csaki:2015vek,Anchordoqui:2015jxc,Csaki:2016raa,Harland-Lang:2016qjy,Martin:2016byf,Anchordoqui:2016rve}.
The total photo-production cross section at LHC13 can be parametrized by
\begin{equation} \sigma_{\rm LHC13} (\gamma \gamma
\to \eta_\mathcal{Q} \to \gamma \gamma) =
4.1~\text{pb}~\left(\frac{\Gamma_{\zt{total}}}{45~\text{GeV}} \right)
{\cal B}^2(\eta_{\mathcal Q} \rightarrow \gamma\gamma)
\label{shl-1}
\end{equation}
and the one at LHC8 by
\begin{equation} \sigma_{\rm LHC8} (\gamma \gamma \to \eta_{\mathcal Q} \to \gamma \gamma) =
1.4~\text{pb}~\left(\frac{\Gamma_{\zt{total}}}{45~{\rm GeV}} \right)
{\cal B}^2(\eta_{\mathcal Q} \rightarrow \gamma\gamma)\, , \end{equation}
showing consistency with the 95\% CL upper limit~\cite{Khachatryan:2015qba}. Actually, the ratio of
the LHC13/LHC8 partonic luminosity is largely dominated by
systematic uncertainties driven by the parton distribution functions.
The luminosity ratio is~\cite{Harland-Lang:2016qjy}
\begin{equation}
\frac{{\cal L}_{\gamma \gamma} (\sqrt{s} = 13~{\rm TeV})}{{\cal L}_{\gamma \gamma} (\sqrt{s} =
8~{\rm TeV})} = 3^{+0.1}_{-0.2}, \ 2.65\pm 0.15, \ 2.1 \pm0.4,
\end{equation}
for CT14QED~\cite{Schmidt:2015zda}, MRST2004~\cite{Martin:2004dh}, and
NNPDF2.3~\cite{Ball:2013hta}; respectively. We note that the predictions
of NNPDF2.3 are only marginally compatible with LHC8
data~\cite{Khachatryan:2015qba}.
The decay width of $\eta_{\cal Q}$ to diphotons is given
by~\cite{Iwamoto:2016ral} \begin{equation} \frac{\Gamma(\eta_{\mathcal Q} \to {\gamma}
{\gamma})}{M_X} = 4 N_{\rm TC} \left( Y_{\mathcal Q}+\frac 1 2 \right) ^4 \alpha^2
\frac{|R(0)|^2}{M_X^3}\,, \end{equation} where $R(0) \equiv R^{(1,0)}(0)$. The
total width is dominated by two major channels. One of them is to techni-gluons
$\Gamma(\eta_{\mathcal Q} \to g_{\zt{TC}} g_{\zt{TC}})$, with
\begin{equation}
\frac {\Gamma(\eta_{\mathcal Q} \to g_{\zt{TC}}
g_{\zt{TC}})}{\Gamma(\eta_{\mathcal Q} \to
\gamma\gamma)} = \frac{ \left(N_{\rm
TC}^2-1\right)}{4 N_{\rm TC}^2}\frac{ \alpha_{\zt{TC}}^2} {\left(Y_{\mathcal
Q}+\frac{1}{2}\right)^4\alpha^2} \,. \end{equation}
and the other is $\Gamma(\eta_{\mathcal Q} \to Z H)$ \cite{Barger:1987xg}
\begin{equation}
\frac {\Gamma(\eta_{\mathcal Q} \to Z H)}{\Gamma(\eta_{\mathcal Q} \to
\gamma\gamma)} = \frac{M_X^4} {4 {M_Z}^4}\frac{ a^2 {\alpha}_Z^2 \left[(1-R_H-R_H)^2-4 R_H R_Z\right]^{3/2}} {4\left(Y_{\mathcal
Q}+\frac{1}{2}\right)^4\alpha^2} \,, \end{equation}
where $R_i = (M_i/M_X)^2$ and $M_i$ ($i=H,Z$) are the masses of Higgs and $Z$-boson. Also ${\alpha}_Z = {\alpha}/(\sin^2 {\theta}_W \cos^2 {\theta}_W)$ and $a = \frac 1 2 (I_{3L} - I_{3R})$, where $I_{3L}$ ($I_{3R}$) is the isospin for the left- and right-handed techni-qards. Notice that the decay width into $ZH$ is enhanced by a factor of $\frac{M_X^4} {{M_Z}^4}$ due to the longitudinal mode of the $Z$-boson. The decay width into $HH$ would receive similar enhancement had it not been forbidden by CP symmetry ($HH$ has $J^{\zt{PC}} = 0^{++}$ while $\eta_{\cal Q}$ has $J^{\zt{PC}} = 0^{-+}$). We will determine ${\alpha}_{\zt{TC}}$ (as a function of $N_{\rm
TC},Y_{\mathcal Q}$) by fitting the predicted production cross
section times branching to the observed value. Before proceeding we
note that the ATLAS excess is quite broad and probably with a large
uncertainty. The CMS excess, however, is smaller and has no clear
preference for a large width. This seems to indicate that the ATLAS
excess could be a real signal combined with a large fluctuation,
making the excess appear larger and wider than the underlying physical
signal. Throughout we assume the resonance needs to have a
signal~\cite{Kats:2016kuz}
\begin{equation}
\sigma_{\rm LHC13} (pp \to X + {\rm
anything}) \times {\cal B} (X \to \gamma \gamma) \approx 3 -
6~{\rm fb}\, .
\label{mattS}
\end{equation}
Note that the $2$ bound states of electric charge $Y_{\cal Q} + \frac 1 2$ contribute equally to the total cross section and dwarf the contributions from the other two states of charge $Y_{\cal Q} - \frac 1 2$ because ${\sigma}$ is proportional to the 8th power of the electric charge.
Our results are encapsulated in Fig.~\ref{fig:1}, where we show the
region of the parameter space that can explain the diphoton signal and satisfy the bound $\Gamma(\eta_{\mathcal Q} \to Z H) < 10~\Gamma(\eta_{\mathcal Q} \to {\gamma} {\gamma})$. The (blue) banana-shape region in the top panel is obtained by
demanding both that the Landau pole of the Yukawa coupling is above
30~TeV and that the $\eta_{\cal Q}$ production cross section times
branching into diphotons (\ref{shl-1}) is 5~fb. The (red) cross-hatched tail is excluded because the Coulomb approximation
fails. The (orange) region on the top of the figure is where the contribution to the decay width from the channel into diphotons is no longer negligible. The parameter space of the
$Y_{\cal Q} - N_{\rm TC}$ plane is further constrained by the perturbativity condition $\alpha_{\rm TC}
(a_0^{-1}) < 1$ and by the requirement $\Gamma(\eta_{\mathcal Q} \to Z
H) < 10~\Gamma(\eta_{\mathcal Q} \to {\gamma} {\gamma})$. Note that within the
allowed region of the parameter space we always have $\Gamma_{\rm
total} \leq 10~{\rm GeV}$ for each bound state. In fact, the
constraint $\alpha_{\rm TC} (a_0^{-1})<1$ itself requires $\Gamma_{\rm total} < 11~{\rm GeV}$. The allowed region of the parameter space is significantly bounded; the hypercharge needs to lie between $2.26 \leq Y_{\mathcal Q} \leq 2.53$ for $N_{\rm TC} =2$ and be within the range $2.17 \leq Y_{\mathcal Q} \leq 2.31$ for $N_{\rm TC} = 3$. Inside this region, the partial decay widths into $W^+ W^-$ and into $Z Z$ are negligible compared to $\Gamma(\eta_{\mathcal Q} \to {\gamma} {\gamma})$ and hence the corresponding bounds are satisfied trivially.
\begin{figure}[tpb]
\begin{minipage}[t]{0.80\textwidth}
\postscript{RG_Q-onium3}{0.78}
\end{minipage}
\begin{minipage}[t]{0.80\textwidth}
\postscript{RG_Q-onium_10GeV}{0.78}
\end{minipage}
\caption{{\bf Top.} The (blue) banana-shape region is obtained by
demanding that the Landau pole of the Yukawa coupling is above
30~TeV and that the production cross section of $\eta_{\cal Q}$ times
its branching into diphotons is about 5~fb. The (red) cross-hatched tail is excluded because the Coulomb approximation
fails. The (orange) band is where the contribution from decay into diphotons accounts for more than 10\% of the total width. {\bf
Bottom.} Allowed region of the parameter space for $\Gamma(\eta_{\mathcal Q} \to Z H) < 10~\Gamma(\eta_{\mathcal Q} \to {\gamma}
{\gamma})$.}
\label{fig:1}
\end{figure}
\section{Constraints from cosmology}
\label{secIV}
Primordial nucleosynthesis provides the earliest observationally
verified landmark and constraints from big bang nucleosynthesis (BBN)
can bound the parameters of beyond SM physics models~\cite{Sarkar:1995dd}. Of particular
interest here, the techni-gluons from the decay of $\mathcal{Q}$-onium
would hadronize to techni-glueballs $G_{\zt{TC}}$ of mass $M_G \sim 7
{\Lambda}_{\rm TC}$. We must then
verify that the dominant techni-glueball decay, $G_{\zt{TC}} \to \gamma \gamma$,
does not drastically alter
any of the light elemental abundances synthesized during
BBN.
Note that if the $G_{\rm TC}$ decay takes place before BBN, then the photons injected
into the plasma would rapidly redistribute their energy through
scattering off background photons and through inverse
Compton scattering. The thermalization process will be particularly efficient at
plasma temperatures above 1~MeV, which is the threshold for background
$e^+e^-$ pair annihilation, and which, incidentally, coincides with time of
about 1~second. For $N_{\rm TC}=3$, the techni-glueball decay
width is given by \begin{equation} \Gamma \left(G_{\zt{TC}} \to \gamma \gamma
\right) \approx \frac{(Y_{\cal Q} + 1/2)^4 \alpha^2}{64 \pi^3}
\frac{M_{G}^3}{m_{\cal Q}^2} \left( \frac{3 M_{G}^3}{60 m_{\cal Q}^3} \right)^2
\, ; \end{equation} the values for different $N_{\rm TC}$ are of the same
order~\cite{Chen:2005mg,Juknevich:2009ji,Juknevich:2009gg}. A straightforward substitution shows that in our model $G_{\rm
TC}$
decay would not not alter BBN as the techni-glueball lifetime $<
10^{-28}~{\rm s}$.
\section{Looking ahead}
\label{secV}
In this work we have attempted to associate the possible event excess
in the diphoton invariant mass spectrum around 750~GeV, as indicated
in LHC13 data, with bound states of a new asymptotically free gauge
theory. In particular, we have constructed a minimal model with one
generation of uncolored fermions $\mathcal Q$, in which the $\mathcal
Q$ fields transform as $SU(2)$ doublets and singlets, and are
invariant under the $SU(N_{\rm TC})$ transformation of a hidden strong
gauge interaction to be explored during the LHC Run II data taking
period. Since the new colorless fields transform as $SU(2)$ doublets
they can acquire a mass through Yukawa interactions with the
electroweak Higgs doublet, as quarks and leptons. We parametrized the
${\mathcal Q}$-onium bound state using the Coulomb approximation and
gave a numerical fit to the diphoton excess consistent with no strong
coupling effects up to about 30~TeV. We have shown that allowed
hypercharges lie within the range of $2.26 \leq Y_{\mathcal Q} \leq
2.53$ for $N_{\rm TC} =2$ and $2.17 \leq Y_{\mathcal Q} \leq 2.31$ for
$N_{\rm TC} = 3$. The smoking gun for the model would be the
discovery of $\eta_\mathcal{Q} \to ZH$.
In closing, we note that the $\mathcal{Q}$ Yukawa couplings drive the
quartic Higgs coupling to negative values in the ultraviolet and the
SM scalar effective potential develops an instability above about
100~TeV. As noted elsewhere~\cite{Anchordoqui:2012fq} (see also~\cite{Kadastik:2011aa,EliasMiro:2012ay}) the potential instability of
the electroweak vacuum can be evaded if the scalar sector contains a
hidden heavy scalar singlet (with a large vacuum expectation value),
which mixes with the SM Higgs doublet. The quartic interaction
between the heavy scalar singlet and the Higgs doublet leads to a
positive tree-level threshold correction for the Higgs quartic
coupling, which is very effective in stabilizing the potential. In
addition, the hidden scalar singlet could deliver mass terms for the
vector-like hidden quarks $Q$ and/or ${\mathpzc Q}$~\cite{inprepa}.
\acknowledgments{We would like to thank Xerxes Tata, Vernon Barger,
and Chaehyun Yu for valuable discussions. L.A.A. is supported by
U.S. National Science Foundation (NSF) CAREER Award PHY1053663 and
by the National Aeronautics and Space Administration (NASA) Grant
No. NNX13AH52G; he thanks the Center for Cosmology and Particle
Physics at New York University for its hospitality. X.H. is
supported by the MOST Grant 103-2811-M-003-024; he thanks the
Institute of Modern Physics at Northwest University for its
hospitality. Any opinions, findings, and conclusions or
recommendations expressed in this material are those of the authors
and do not necessarily reflect the views of the National Science
Foundation.}
|
\section{Introduction}
The search for topological superconductors, which host Majorana zero modes (MZMs),
has becomes an active pursuit in condensed matter physics \cite{Reich,Brouwer_Science,Wilczek2012,AliceaRev}.
Such exotic modes are predicted to obey non-Abelian braiding statistics\cite{Moore1991,Nayak1996,ReadGreen},
and have potential application in topological quantum computations \cite{kitaev,TQCreview}.
Many theoretical proposals for realizing topological superconductors in the laboratory have been
put forward recently~\cite{Fu&Kane08,Fu&Kane09,Sau10,Alicea10,LutchynPRL10,1DwiresOreg,MajoranaTInanowires,SauNature2012,Nadj-Perge13}, and, more excitingly, devices for detecting MZMs were successfully fabricated in the laboratory and the preliminary signatures of MZMs were
observed~\cite{Mourik2012,Das2012,Deng2012,Fink2012,Churchill2013,Nadj-Perge14,Deng2014,Higginbotham15,Albrecht16,HaoZhang16}.
Most research activity has focused on the topological superconductors belonging to class D ( {\it i.e.},
SCs with broken time-reversal-symmetry) and supporting an odd number of MZMs at a topological defect~\cite{Altland'97, TIClassification, Kitaev2009}.
However, Majorana zero modes can also appear in pairs in time-reversal invariant topological superconductors (TRITOPS)
belonging to class ${\rm DIII}$ \cite{ TIClassification, Kitaev2009, Teo&Kane10}. Those MZM pairs are referred to as
``Majorana Kramers pairs'' (MKPs), and their stability is protected by the time-reversal (TR) symmetry and the quasiparticle excitation gap. Recently, several theoretical proposals were put forward to realize TRITOPS~\cite{wong&Law12,dengPRL12,zhangPRL13,NakosaiPRL13,KeselmanPRL13,Gaidamauskas14,Klinovaja14,SchradePRL15}.
Transport signatures of MKPs and their detection schemes using a QPC were also recently investigated in a quantum spin Hall system~\cite{Li15}.
Most previous works on MKPs considered non-interacting (or effectively non-interacting) models. It is well-known, however,
that interactions in one-dimensional systems are very important~\cite{Gangadharaiah11, Lobos'12, Fidkowski'12, Affleck'13} and
in some cases may even modify the classification of non-interacting systems~\cite{Fidkowski'11}.
For non-interacting systems, the presence of a MKP leads to a quantized conductance of $4e^2/h$ due to
perfect Andreev reflection at the junction. This quantization of the conductance is due to the constraints imposed by TR symmetry which leads to complete decoupling of MKP in the non-interacting models. The situation is different, however, in the presence of interparticle interactions, and the fate of the perfect Andreev reflection fixed point is unclear. In this paper,
we study the stability of MKPs with respect to electron-electron interactions and consider two generic systems - a) MKP
coupled to an interacting Luttinger liquid (see Fig.~\ref{fig:device} a)); b) MKP coupled to an interacting quantum dot (see Fig.~\ref{fig:device} b)).
\begin{figure}[t]
\includegraphics[width=9cm]{device.pdf}
\caption{(Color online) a) Schematic setup consisting of a) a junction between the Luttinger liquid and TRITOPS,
and b) a quantum dot coupled to a normal lead and a TRITOPS. Here, $x=0$ denotes the point in the lead which couples to the MKPs or quantum dots.}
\label{fig:device}
\end{figure}
We first consider a spinful Luttinger liquid lead with $\mathbb{SU}(2)$ spin symmetry coupled to a TRITOPS with a single MKP per end. In this case the boundary problem has an additional $\mathbb{U}(1)$ symmetry. We find that for weak repulsive interactions, $1>K_{\rho}\gtrsim 1/3$ with $K_\rho$ being the Luttinger parameter, the Andreev reflection fixed point ($\mathbb{A\times A}$) is stable
and the normal reflection fixed point ($\mathbb{N\times N}$) is unstable. For intermediate interaction strength $1/4<K_{\rho}\lesssim 1/3$,
the phase diagram depends on microscopic details, i.e. on the strength of four-fermion interactions allowed by TR symmetry,
which causes a Berezinsky-Kosterlitz-Thouless~(BKT)
transition between $\mathbb{A\times A}$ and $\mathbb{N\times N}$. Finally, for sufficiently strong repulsive interactions $K_{\rho}<1/4$,
the two electron backscattering term becomes relevant, and drives the system to a stable normal reflection fixed point.
In the presence of spin-orbit coupling, the corresponding boundary problem may break $\mathbb{U}(1)$ symmetry. In this case, allowed processes involve spin-preserving and spin-flip Andreev scattering which drive the system to different boundary conditions for lead electrons: spin-preserving Andreev boundary ($\mathbb{A\times A}$) condition corresponds to $\psi_\sigma(0)=-\psi^\dag_\sigma(0)$ and spin-flip Andreev boundary ($\mathbb{SFA}$) condition corresponds to $\psi_\sigma(0)= -i\psi^\dag_{-\sigma}(0)$.
Thus, the corresponding phase diagram depends on
the relative strength of the corresponding Andreev scattering amplitudes. We find that the boundary conditions in this case are similar to those in a spin-triplet superconductor-Luttinger liquid junction and are stable with respect to weak repulsive interactions. In this sense, the physics is fundamentally different from an s-wave superconductor-Luttinger junction where weak repulsive interactions destabilize Andreev reflection fixed point~\cite{Fidkowski'12}.
In this paper, we also study the effect of local repulsive interactions by considering a MKP coupled to a quantum dot (QD) and
an $\mathbb{SU}(2)$-invariant normal lead (NL). In the limit of a large Coulomb interaction in the QD and single-electron occupancy, we investigate the competition between Kondo and Majorana correlations. When the coupling to the MKP is absent ($\lambda=0$), the system flows to the Kondo fixed point with the corresponding boundary conditions for NL electrons $\psi_{R\sigma}(0)=-\psi_{L\sigma}(0)$ where $R/L$ denote right and left movers. As we increase the coupling constant $\lambda$, the system exhibits a crossover
from the Kondo dominated regime to a Majorana dominated regime where the QD spin builds up a strong correlation with the MKP. The latter is characterized by $\mathbb{A\times A}$ boundary conditions $\psi_{R\sigma}(0)=-\psi^\dag_{L\sigma}(0)$. Thus, the problem at hand represents a new class of boundary impurity problems where spin in the dot is coupled to the fermion parity of a topological superconductor.
In order to understand thermodynamic and transport properties of this Majorana fixed point,
we have developed a slave-boson mean-field theory (please refer to Refs. \cite{Coleman84,BickersRMP87} for Anderson impurity models)
for this system. We show that the Majorana dominated regime corresponds to a new (i.e. different from Kondo) saddle-point solution.
We have analyzed the stability of this mean-field solution with respect to Gaussian fluctuations (in the spirit of Refs. \cite{Read&Newns83, Coleman'87}) finding that the mean field theory is stable (in the quasi-long range order sense) and can be used to calculate different observable quantities. We use this approach to calculate differential tunneling conductance as a function of applied voltage bias.
The paper is organized as follows. In Secs.~\ref{sec:SU2NoRashba} and \ref{sec:RashbaEffect}, we introduce the model of a MKP - Luttinger liquid junction, and consider the boundary problem with and without (e.g., due to Rashba spin-orbit coupling in the lead) $\mathbb{U}(1)$ symmetry. In Sec. \ref{sec:QD}, we study the signatures of a MKP in a QD-NL junction using both the renormalization group (RG) analysis
and the slave-boson mean-field theory. We also consider the Gaussian fluctuations around the mean-field solution, and analyze the stability
of the slave-boson mean-field solution. Finally, we conclude in Sec. \ref{sec:conclusion}.
\section{Majorana Kramers pair - Luttinger liquid junction}\label{sec:MKPLL}
In this section we consider the setup shown in Fig. \ref{fig:device} a) consisting of a semi-infinite spinful Luttinger liquid coupled weakly to a TRITOPS. We assume that the topological gap of the superconductor is much larger than the other relevant energy scales
(i.e. tunneling amplitudes $t_\sigma$, $\Delta$ and $\Delta_{\rm AN}$ see the text below Eq. \eqref{eq:model0} for definitions). Thus, in the low-energy approximation the superconductor
Hamiltonian consists of only the MKPs localized at its opposite ends. In this section, we will use $\psi_{\sigma}(0)$ to describe the operators at the boundary
$x=0$, and use $t(l_0)$ (similarly for $\tilde{t}$, $\Delta$ and $\tilde{\Delta}$) as the initial value in RG flow with the initial length cutoff $l_0$.
\subsection{Majorana Kramers pair coupled to $\mathbb{SU}(2)$-invariant Luttinger liquid}\label{sec:SU2NoRashba}
\subsubsection{Theoretical Model}
We first consider an $\mathbb{SU}(2)$-invariant interacting nanowire coupled to a MKP. The Hamiltonian for the 1D lead can be written as the spinful Luttinger model
\begin{equation}
H_{\rm{lead}} = \sum_{j=\rho,\sigma}\frac{v_{j}}{2\pi}\int_{0}^{\infty}dx\left(K_{j}(\partial_{x}\theta_{j})^{2}+\frac{(\partial_{x}\phi_{j})^{2}}{K_{j}}\right)
\label{eq:Hlead}
\end{equation}
where $v_{\rho/\sigma}$ and $K_{\rho/\sigma}$are velocity and
Luttinger parameter for charge and spin modes, respectively.
The bosonic fields satisfy the commutation relation
$[\phi_{\alpha}(x),\theta_{\beta}(x')]=i \pi K_{\alpha} \delta_{\alpha\beta}\text{sign}(x-x')$. We use here the following convention for the Abelian bosonization procedure~\cite{giamarchi}:
\begin{equation}
\psi_{R/L,s}(x)= \frac{\Gamma_{R/L,s}}{\sqrt{2\pi a}}e^{i\frac{1}{\sqrt{2}}\left\{ \pm[\phi_{\rho}(x)+s\phi_{\sigma}(x)]+\theta_{\rho}(x)+s\theta_{\sigma}(x)\right\} }
\end{equation}
where $R/L$ represents right/left moving modes, $a$ is an ultraviolet cutoff length scale, $s=\uparrow/\downarrow$
denotes fermion spin, and $\Gamma_{R/L,s}$ is the Klein factor.
The total Hamiltonian is given as
\begin{align}\label{eq:totalH}
H=H_{\rm lead}+H_B.
\end{align}
where $H_B$ the coupling between the Luttinger liquid lead and the MKP. We neglect here the ground-state degeneracy splitting energy. The most general form of the TR invariant boundary Hamiltonian describing the coupling between the MKP and Luttinger liquid and including only two and four-fermion operators reads
\begin{align}
H_B &=i\, t_{\uparrow}\gamma_{\uparrow}\left(\psi_{\uparrow}(0)+\psi_{\uparrow}^{\dagger}(0)\right)
-i\, t_{\downarrow}\gamma_{\downarrow}\left(\psi_{\downarrow}(0)+\psi_{\downarrow}^{\dagger}(0)\right) \nonumber\\
& -\Delta i\gamma_{\uparrow}\gamma_{\downarrow}\left(-i\psi_{\uparrow}^{\dagger}(0)\psi_{\downarrow}(0)+i\psi_{\downarrow}^{\dagger}(0)\psi_{\uparrow}(0)\right) \nonumber\\
&-\Delta_{\rm AN} i\gamma_{\uparrow}\gamma_{\downarrow}\left(-i\psi_{\uparrow}^{\dagger}(0)\psi_{\downarrow}^{\dagger}(0)+i\psi_{\downarrow}(0)\psi_{\uparrow}(0)\right)
\label{eq:model0}
\end{align}
where $\gamma_{\uparrow/\downarrow}$ are Majorana operators with $\{\gamma_\sigma,\gamma_{\sigma'}\}=2\delta_{\sigma,\sigma'}$. Here $t_{\uparrow/\downarrow}$, $\Delta$
and $\Delta_{\text{AN}}$ are set to be real. The first two terms represent tunneling between the lead and the MKP with the
amplitudes $t_{\uparrow/\downarrow}$. Under TR symmetry $T$, field operators $\psi$ transform as
\begin{eqnarray}
&T \psi_{R/L\uparrow} = \psi_{L/R\downarrow},\\
&T \psi_{R/L\downarrow} = -\psi_{L/R\uparrow},
\end{eqnarray}
(i.e. $T^2=-1$) and the coupling constants in the Hamiltonian need to complex conjugated. TR symmetry requires that $t_{\uparrow}=t_{\downarrow}=t$ with $t$ being real.
Assuming the spin-quantization axis is fixed in the whole system, the overall Hamiltonian $H$
has $\mathbb{U}(1)$ spin-rotation symmetry, leaving it invariant under the unitary transformation:
\begin{align}
(\psi_{\uparrow}, \psi_{\downarrow})^T &\rightarrow R(\theta) (\psi_{\uparrow}, \psi_{\downarrow})^T \\
(\gamma_{\uparrow},\gamma_{\downarrow})^T & \rightarrow R(-\theta) (\gamma_{\uparrow},\gamma_{\downarrow})^T.
\label{eq:U1symmetry}
\end{align}
Here $R(\theta)=e^{i\frac{\theta}{2}\sigma_y}$ represents a $\mathbb{U}(1)$ spin-rotation matrix by an angle $\theta$. Thus, electron tunneling between Luttinger liquid and topological superconductor preserves the spin.
The last two terms $\Delta$ and $\Delta_{\rm AN}$
represent normal, and anomalous backscattering terms,
which, in fact, will also be generated by the tunneling terms in the RG flow in the presence of interactions in the Luttinger liquid.
\subsubsection{Weak coupling RG analysis near normal reflection fixed point}
We now study the stability of the weak coupling normal reflection fixed point using perturbative RG analysis.
In the ultraviolet, the boundary conditions for lead electrons at $x=0$ are given by $\psi_{R\sigma}(0)=\psi_{L\sigma}(0)$ (i.e. perfect normal reflection). In terms of bosonization language, this boundary condition corresponds to $\Gamma_{L,s}=\Gamma_{R,s}$ and pinning $\phi_{\rho,\sigma}(0)$.
Once we turn on the boundary couplings $t$, $\Delta$ and $\Delta_{\rm AN}$, boundary conditions for lead electrons may change depending on the strength of interaction in the lead.
Let us study now the stability of this normal reflection fixed point. After integrating out the fields away from $x=0$, the corresponding imaginary-time partition function becomes
\begin{equation}
\mathcal{Z}={\displaystyle \int}D[\theta_{\rho}]D[\theta_{\sigma}]\; e^{-(S_{0}+S_{T})},
\end{equation}
with
\begin{equation}
S_{0}=\sum_{j=\rho,\sigma}\frac{K_{j}}{2\pi}\int\frac{d\omega}{2\pi}|\omega||\theta_{j}(\omega)|^{2},
\label{eq:ActionBC_Normal}
\end{equation}
and the boundary coupling term reads
\begin{align}
S_{T}&=\int \frac{d\tau}{2\pi a}\Bigg[t\big( i\gamma_{\uparrow}\Gamma_{\uparrow}\cos\frac{\theta_{\rho}+
\theta_{\sigma}}{\sqrt{2}}-i\gamma_{\downarrow}\Gamma_{\downarrow}\cos\frac{\theta_{\rho}-\theta_{\sigma}}{\sqrt{2}}\big)\nonumber\\
& -\Delta\gamma_{\uparrow}\gamma_{\downarrow}\Gamma_{\uparrow}\Gamma_{\downarrow}\cos\sqrt{2}\theta_{\sigma}
-\Delta_{\rm AN}\gamma_{\uparrow}\gamma_{\downarrow}\Gamma_{\uparrow}\Gamma_{\downarrow}\cos\sqrt{2}\theta_{\rho}\Bigg],
\end{align}
where $a$ is the ultraviolet cutoff. Here we used short-hand notation $\theta_{j}(\tau)$ denoting the fields at $x=0$.
We now perform a perturbative RG procedure by separating the bosonic fields $\theta_{j}$ into slow, and fast modes and integrating out the fast modes. After some manipulations, the new effective action can be calculated using the
cumulant expansion:
\begin{equation}
S_{{\rm eff}}[\theta_{j}^{<}]=S_{{\rm 0}}[\theta_{j}^{<}]+\langle S_{T}\rangle-\frac{1}{2}\left(\langle S_{T}^{2}\rangle-\langle S_{T}\rangle^{2}\right),
\end{equation}
where the average $\langle\cdots\rangle$ describes an integration
over the fast modes. The details of this calculation are presented in the Appendix \ref{app:RG2order_SU2}, and we simply summarize the RG equations here
\begin{eqnarray}
\frac{dt}{dl} & = & \left(1-\frac{1}{4K_{\rho}}-\frac{1}{4K_{\sigma}}\right)t -\frac{\Delta t }{4\pi v K_{\sigma}} -\frac{\Delta_{\rm AN} t}{4\pi v K_{\rho}},\,\,\,\,\,\,\,\,\label{eq:U1RG_t} \\
\frac{d\Delta}{dl} & = & \left(1-\frac{1}{K_{\sigma}}\right)\Delta-\left(\frac{1}{K_{\rho}}-\frac{1}{K_{\sigma}}\right)\frac{t^2}{4\pi v} , \label{eq:U1RG_Delta} \\
\frac{d\Delta_{\rm AN}}{dl} & = & \left(1-\frac{1}{K_{\rho}}\right)\Delta_{\rm AN}+\left(\frac{1}{K_{\rho}}-\frac{1}{K_{\sigma}}\right)\frac{t^2}{4\pi v}. \label{eq:U1RG_delta}
\end{eqnarray}
Here $dl=d\ln b$ where $b$ is the ratio of the cutoff change from $\Lambda$ to $\Lambda/b$ with $\Lambda =v/a$.
One can notice that $t$ is a relevant perturbation and grows under RG. Therefore, in the non-interacting case when $\Delta, \Delta_{\rm AN}=0$,
the system will flow to the perfect Andreev reflection fixed point ($\mathbb{A\times A}$) corresponding to the boundary condition $\psi_{L,s}^{\dagger}(0)=-\psi_{R,s}(0)$~\cite{Fidkowski'12} and quantized differential conductance $G=\frac{4e^2}{h}$ at zero temperature.
\begin{figure}
\includegraphics[scale=0.55]{SU2symRGD.pdf}
\caption{Illustration of the RG flow diagram near the normal reflection fixed point $\mathbb{N\times N}$ for the case with $U(1)$ spin symmetry.
The red line indicates the regime where $\mathbb{N\times N}$ is stable. Here, we set $K_\rho=1/3$}
\label{fig:SU2symRGD}
\end{figure}
Let us now try to understand the effects of interactions. In this section, we will focus on an $\mathbb{SU}(2)$ spin-invariant lead ($K_{\sigma}=1$) and repulsive interactions in the nanowire $K_{\rho}<1$. In this case,
the coupling $\Delta_{\rm AN}$ is irrelevant and can be neglected, and RG equations simplify to
\begin{eqnarray}
\frac{dt}{dl} & = & \left(\frac{3}{4}-\frac{1}{4K_{\rho}}\right)t-\frac{\Delta t}{4\pi v} ,\\
\frac{d\Delta}{dl} & = & - \left(\frac{1}{K_{\rho}}-1\right) \frac{ t^2}{4\pi v}.
\end{eqnarray}
The coupling $t$ is relevant for not too strong repulsive interactions. It becomes marginal, however, if initial value of $\Delta(l_0)$ is equal to the special value $\Delta^* = \pi v (3-\frac{1}{K_\rho})$. Indeed, then above RG equations (after a slight redefinition of variables) are identical to the anisotropic Kondo model~\cite{giamarchi}, the solution of which is well-known. If the initial value of $\Delta (l_0)$ is zero, and the parameter $\Delta^*>0$ (i.e. $K_{\rho}>1/3$), the system will flow to strong coupling $\mathbb{A\times A}$ fixed point
whereas for $K_{\rho}\lesssim 1/3$, the system will flow to $\mathbb{N\times N}$ fixed point for small $t(l_0)$
and flow to strong coupling $\mathbb{A\times A}$ for larger $t(l_0)$. The perturbative RG flow is summarized in Fig. \ref{fig:SU2symRGD}.
\subsubsection{Weak coupling RG analysis near perfect Andreev reflection fixed point}
As shown in the previous section, the normal reflection fixed point is unstable for weak repulsive interactions and the system flows to the perfect Andreev fixed point corresponding to the boundary conditions $\psi_{L,s}^{\dagger}(0)=-\psi_{R,s}(0)$ which, in bosonic variables corresponds to pinning $\theta_{\rho}$ and $\theta_{\sigma}$ fields at $x=0$. Thus, the fluctuating degrees of freedom are the fields $\phi_{\rho}$ and $\phi_{\sigma}$ and the corresponding boundary action reads
\begin{equation}
S_{0}=\sum_{j=\rho,\sigma}\frac{1}{2\pi K_{j}}\int\frac{d\omega}{2\pi}|\omega||\phi_{j}(\omega)|^{2}.
\label{eq:ActionBC_Andreev}
\end{equation}
We now consider perturbations near the Andreev fixed point which are consistent with time-reversal and the spin-$\mathbb{SU}(2)$ symmetry of the Luttinger liquid lead. The only fermion bilinear boundary perturbation preserving aforementioned symmetries is
\begin{align}
H_{1B}&=\lambda_1 (\psi_{R\uparrow}^{\dagger}(0)\psi_{L\uparrow}(0)+\psi_{R\downarrow}^{\dagger}(0)\psi_{L\downarrow}(0))+h.c.\nonumber\\
&=\frac{\lambda_1}{2\pi a}\cos\left(\sqrt{2}\phi_{\rho}\right)\cos\left(\sqrt{2}\phi_{\sigma}\right).
\label{eq:NormalBackscattering}
\end{align}
In addition, one has to also consider the following four-fermion perturbation consistent with the above symmetries:
\begin{align}
H_{2B}=&\lambda_2 \psi_{L\uparrow}^{\dagger}(0)\psi_{R\uparrow}(0)\psi_{L\downarrow}^{\dagger}(0)\psi_{R\downarrow}(0)+h.c. \nonumber \\
&= \frac{\lambda_2}{(2\pi a)^2}\sin(2\sqrt{2}\phi_{\rho}),
\label{eq:TwoBackscattering}
\end{align}
which corresponds to two-electron backscattering. The perturbative RG equations for $\lambda_1$ and $\lambda_2$ are given by
\begin{align}
\frac{d\lambda_1}{dl}&=(1-K_{\rho}-K_{\sigma})\lambda_1\\
\frac{d\lambda_2}{dl}&=(1-4K_{\rho})\lambda_2
\end{align}
One can see that the first term $\lambda_1$ is irrelevant since $K_\sigma=1$ whereas the second coupling becomes relevant for $K_{\rho}<1/4$ indicating that $\mathbb{A\times A}$ fixed point becomes unstable for strong repulsive interactions.
Taking into account the perturbative RG analysis near both $\mathbb{N\times N}$ and $\mathbb{A\times A}$ fixed points,
we conjecture the qualitative phase diagrams shown in Fig. \ref{fig:SU2symFlowD}.
\begin{figure}
\includegraphics[scale=0.45]{SU2symFlowD.pdf}
\caption{Illustration of the flow between the normal reflection fixed point $\mathbb{N\times N}$
and the Andreev reflection fixed point $\mathbb{A\times A}$ for the case with $U(1)$ spin symmetry.
For the analysis at $\mathbb{N\times N}$, the boundary perturbation $\Delta$ bends the transition
line, i.e. the brown dashed line connecting $K_{\rho}=1/3$ at $\mathbb{N\times N}$ and $K_{\rho}=1/4$ at $\mathbb{A\times A}$.
We also assume the initial value $\Delta(l_0)$ in RG is zero.}
\label{fig:SU2symFlowD}
\end{figure}
\subsubsection{Differential tunneling conductance}
We now discuss transport signatures of MKPs. The simplest experiment to detect the presence of a MKP is the differential conductance measurement.
We focus on the case with an $SU(2)$ symmetric wire $K_{\sigma}=1$ and calculate $G=dI/dV$ at zero voltage bias as a function of temperature. The RG flow between the normal $\mathbb{N\times N}$ and Andreev reflection $\mathbb{A\times A}$ defines a crossover temperature $T^*$, which roughly corresponds to the width of the zero bias peak.
Although the conductance for the whole crossover regime requires involved calculations, the conductance around $\mathbb{N\times N}$
and $\mathbb{A\times A}$ fixed points can be obtained using perturbation theory, see, e.g., Ref.~[\onlinecite{Lutchyn2013}].
First of all, we consider the case $1/3<K_{\rho}<1$, where $\mathbb{A\times A}$ fixed point is stable.
In the ultraviolet (i.e. near the unstable normal reflection fixed point), the leading relevant perturbation is the coupling to the MKP, $t$, which has scaling dimension $\frac{3}{4}-\frac{1}{4K_{\rho}}$. Near the stable Andreev reflection fixed point (i.e. in the infrared),
the deviation from the quantized value comes from the leading irrelevant operators which cause backscatterings,
i.e. single-electron backscattering shown in Eq. (\ref{eq:NormalBackscattering}) with scaling dimension $-K_{\rho}$ and
two-electron backscattering shown in Eq. (\ref{eq:TwoBackscattering}) with scaling dimension $1-4K_{\rho}$.
Here, for $K_{\rho}>1/3$, the single-electron backscattering shown in Eq. (\ref{eq:NormalBackscattering}) is the leading irrelevant operator. We can now obtain scaling of the conductance with temperature at zero bias (assuming the initial value of $\Delta$ coupling is zero, i.e. $\Delta(l_0)=0$):
\begin{equation}
\frac{G}{4e^2/h} \Bigg |_{K_{\rho}>\frac{1}{3}} = \begin{cases}
c_{1,T}(K_\rho)\left(\frac{T}{T^*}\right)^{2(\frac{1}{4K_{\rho}}-\frac{3}{4})}, & \quad T\gg T^* \\
1 - c_{2,T} (K_{\rho}) \left(\frac{T}{T^*}\right)^{2K_{\rho}}, & \quad T\ll T^* \\
\end{cases},
\end{equation}
where $c_{1/2,T}(K_{\rho})$ are numerical coefficients of the order one. Similarly, one can obtain voltage corrections to the conductance at zero temperature. Interestingly, the analogous coefficient $c_{1,V}(K_{\rho})$ vanishes in the non-interacting limit and, therefore, the scaling of the conductance with voltage and temperature is different at $K_{\rho}=1$, see Ref.~[\onlinecite{Lutchyn2013}] for details.
Next, we consider $K_{\rho}<1/4$, where $\mathbb{N\times N}$ is stable in the infrared.
In this case, we start near the high energy unstable fixed point $\mathbb{A\times A}$ and calculate the conductance by perturbing with the two-electron backscattering operator which is the leading relevant operator in this regime.
Thus, we obtain
\begin{equation}
\frac{G}{4e^2/h} \Bigg |_{K_{\rho}\lesssim\frac{1}{3}} \sim \begin{cases}
1- c_{3,T}(K_{\rho}) \left(\frac{T}{T^*}\right)^{2(4K_{\rho}-1)}, & \quad T\gg T^* \\
c_{4,T}(K_\rho)\left(\frac{T}{T^*}\right)^{2(\frac{1}{4K_{\rho}}-\frac{3}{4})}, & \quad T\ll T^* \\
\end{cases},
\end{equation}
where $c_{3/4,T}(K_{\rho})$ are $\mathcal{O}(1)$ numerical coefficients.
The calculation of the conductance in the regime $1/4<K_{\rho}\lesssim 1/3$ depends on microscopic details (i.e. strength of $t(l_0)$), and is outside the scope of the paper.
\subsection{The effect of the Rashba spin-orbit coupling in the lead}\label{sec:RashbaEffect}
\subsubsection{Theoretical Model}\label{sec:IIB1}
In this section, we consider the effect of Rashba spin-orbit coupling (SOC) in the nanowire. When coupling to MKP, the spin eigenstates of the MKP do not have to be the same as the
the spin eigenstates of the nanowire. Therefore, tunneling between the lead and the TRITOPS will have both spin-preserving and spin-flip components. In order to see how the spin flip tunneling
is generated, we consider the direction of the Rashba coupling which has an angle $\theta$ rotation
compared to that of the MKP. The corresponding tight binding model can written as
\begin{align}
H &= H_{\rm lead} + H_{\rm T}\\
H_{\rm lead} &= -t \sum_{j=1}^{N}\sum_{s} \left(c_{j+1, s}^{\dagger}c_{j, s} +h.c. \right)
+ \mu \sum_{j s} c_{j, s}^{\dagger}c_{j, s} \nonumber\\
& +\sum_{js s'} (-i) \alpha_R c_{j+1,s}^{\dagger} \left( \cos\theta \sigma_z +\sin\theta \sigma_y \right)_{ss'} c_{j,s'} +h.c.,\nonumber\\
H_{\rm T}&= i t_0 \left[ \gamma_{\uparrow} ( c_{N\uparrow}+c_{N\uparrow}^{\dagger} )
-\gamma_{\downarrow} ( c_{N\downarrow}+c_{N\downarrow}^{\dagger} )\right].
\end{align}
One can see that the above Hamiltonian respects TR symmetry. We apply the following unitary transformation
\begin{equation}
\left( \begin{array}{c}
d_{i\uparrow} \\
d_{i\downarrow}
\end{array}
\right) = e^{-i\frac{\theta}{2}\sigma_x}
\left( \begin{array}{c}
c_{i\uparrow} \\
c_{i\downarrow}
\end{array}
\right),
\end{equation}
and then the bulk and boundary Hamiltonians become
\begin{eqnarray}
H_{\rm lead} &=& \mu \sum_{js}d_{j, s}^{\dagger}d_{j, s} + \sum_j \Big[ (-t-i\alpha_R)d_{j+1, \uparrow}^{\dagger}d_{j, \uparrow} \nonumber\\
&& + (-t+i\alpha_R)d_{j+1, \downarrow}^{\dagger}d_{j, \downarrow} +h.c.\Big] ,
\end{eqnarray}
and
\begin{eqnarray}
H_{\rm T} &= & i t \sum_{s=\uparrow,\downarrow} s \gamma_{s} (d_{N,s}+d_{N,s}^{\dagger})\nonumber\\
& & + \tilde{t} \sum_{s} s \gamma_{s} (d_{N,-s}^{\dagger}-d_{N,-s}),
\end{eqnarray}
where $t=t_0 \cos\theta$ and $\tilde{t}=t_0 \sin\theta$, and $s=1(-1)$ for spin-$\uparrow(\downarrow)$.
Therefore, the spin-flip tunneling is non-zero for any $\theta\neq 0$, i.e. due to the presence of SOC.
One can simply check that, in the presence of both $t$ and $\tilde t$, the $U(1)$ symmetry shown in Eq. (\ref{eq:U1symmetry}) is broken.
In this case, the boundary condition at the Andreev reflection fixed point is determined by the relative magnitude of $t$ and $\tilde t$. For the discussion of boundary condition and bosonization procedure in the normal reflection fixed point, please refer to Appendix \ref{app:soc_boundary}.
It is instructive to analyze the boundary conditions in the non-interacting case using the scattering matrix approach. The unitary scattering matrix
is defined as (see, e.g., Ref.~\cite{Nilsson08})
\begin{equation}\label{eq:Sw}
S(\omega) = \hat{I} + 2\pi i \hat{W}^{\dagger} \left( H_{MK} -\omega - i\pi \hat{W}\hat{W}^{\dagger} \right)^{-1} \hat{W},
\end{equation}
where $H_{MK}$ is the Hamiltonian for the MKP (2 by 2 matrix) which vanishes in the limit $L \gg \xi$ with $L$ and $\xi$ being respectively the length and coherence length of the superconductor. Note that the local term $i \delta E \gamma_{\uparrow} \gamma_{\downarrow}$ is not allowed by TR symmetry. The matrix $\hat{W}$
describes the coupling between the MKP ${\gamma_{\uparrow},\gamma_{\downarrow}}$ and the lead degrees of
freedom in the basis $({\psi_{\uparrow},\psi_{\downarrow},\psi_{\uparrow}^{\dagger},\psi_{\downarrow}^{\dagger}})$:
\begin{equation}
\hat{W} =
\begin{pmatrix}
it & \tilde{t} & it & -\tilde{t} \\
-\tilde{t} & -it & \tilde{t} & -it
\end{pmatrix}.
\end{equation}
Note that we assume that lead Hamiltonian is diagonal here. Therefore, $\psi_\sigma$ represent helicity eigenstates in the case of a Rashba model. Using Eq. \eqref{eq:Sw}, we can represent the scattering matrix at $\omega=0$ as
\begin{equation}
S(0)=
\begin{pmatrix}
S^{ee}(0) & S^{eh}(0) \\
S^{he}(0) & S^{hh}(0)
\end{pmatrix}.
\end{equation}
The components $S^{ee}(0)$ and $S^{eh}(0)$ describe normal and Andreev reflection, respectively. As pointed out in Ref. \cite{Li15}, the normal part $S^{ee}(0)$ is zero so we focus on the non-diagonal components:
\begin{align}
S^{eh}(0) &=
\begin{pmatrix}
\frac{\tilde{t}^2-t^2}{t^2+\tilde{t}^2} & -\frac{2i \tilde{t} t}{t^2+\tilde{t}^2} \\
-\frac{2i \tilde{t} t}{t^2+\tilde{t}^2} & \frac{\tilde{t}^2-t^2}{t^2+\tilde{t}^2}
\end{pmatrix},\nonumber\\
&=-\cos 2\theta-i\sigma_x\sin 2\theta
\end{align}
where the diagonal term is the coefficient of the same-spin Andreev reflection $\psi_{\uparrow}\rightarrow \psi_{\uparrow}^{\dagger}$,
and the off-diagonal term is the coefficient of the spin-flip Andreev reflection $\psi_{\uparrow}\rightarrow\psi_{\downarrow}^{\dagger}$.
As we change the angle of SOC, $\theta$, from $0\,(\tilde{t}=0)$ to $\pi/4\,(t=\tilde{t})$, the Andreev reflection boundary condition
changes continuously from $\psi_{L,s}(0)=-\psi^\dagger_{R,s}(0)(\mathbb{A\times A})$ with $s=\uparrow,\downarrow$
(i.e. $t\neq 0$ and $\tilde{t}=0$)
to $\psi_{L\uparrow}(0)=-i\psi^\dagger_{R\downarrow}(0)$ and $\psi_{L\downarrow}(0)=-i\psi^\dagger_{R\uparrow}(0)$.
We denote this boundary condition for $t=\tilde{t}$ as \textit{spin flip Andreev reflection boundary condition} $(\mathbb{SFA})$,
which describes an Andreev reflection with spin-flip processes.
Upon increasing $\theta$ to $\pi/2$, the boundary condition becomes $\psi_{L,s}(0)=\psi^\dagger_{R,s}(0) (\mathbb{\tilde{A}\times \tilde{A}})$
(i.e. $t= 0$ and $\tilde{t}\neq 0$).
\begin{figure}
\centering
\includegraphics[scale=0.50]{RGflownew.pdf}
\caption{RG flow diagram for the junction without $U(1)$ symmetry: (a) for $1/3<K_{\rho}<1$ and $K_{\sigma}=1$,
(b) for $K_{\rho}<1/4$ and $K_{\sigma}=1$, and (c) for $1/4<K_{\rho}\lesssim 1/3$ and $K_{\sigma}=1$, and the green line indicates
the conjectured BKT phase transition. The inset table summarizes
the important time-reversal invariant boundary perturbations near $\mathbb{SFA}$, $\mathbb{A\times A}$, and $\mathbb{\tilde{A}\times \tilde{A}}$ fixed points.
For $K_{\sigma}=1$, along each line of the RG flow, the phase diagram as a function of $K_{\rho}$ is similar to the that shown in Fig.~\ref{fig:SU2symRGD}.}
\label{fig:RGflow}
\end{figure}
Here we would like to emphasize that the $\mathbb{SFA}$ boundary condition is different from the Andreev boundary condition in s-wave
spin-singlet superconducting junction where $\psi_{L\uparrow}(0)=\mp i\psi^\dagger_{R\downarrow}(0)$ and $\psi_{L\downarrow}(0)=\pm i\psi^\dagger_{R\uparrow}(0)$ (see, e.g., Ref.\onlinecite{maslovPRB96}). Notice different signs in this case for spin-up and spin-down components. The $\mathbb{SFA}$ boundary condition in our case corresponds to spin-triplet Andreev reflection which typically is realized at junctions between a normal lead and a spin-triplet p-wave superconductor. Indeed, if we denote spin-triplet pair potential as $\Delta(p) \propto (\overrightarrow{d}(p)\cdot \overrightarrow{\sigma})i\sigma_{y}$, then different orientations of the $\overrightarrow{d}$-vector correspond to $\mathbb{SFA}$ ($\overrightarrow{d}\propto (0,0,1)$) and $\mathbb{A \times A}$ ($\overrightarrow{d} \propto (0,\pm 1,0)$) boundary conditions. This difference between conventional (s-wave) spin-singlet Andreev boundary conditions and $\mathbb{SFA}$ boundary conditions considered here becomes very important later when we consider allowed boundary perturbations.
\subsubsection{RG analysis near normal reflection fixed point $\mathbb{N\times N}$}
Let's now analyze the interaction effects in the lead. In the absence of U(1) spin-rotation symmetry, we can have additional terms in the boundary action:
\begin{eqnarray}
S_{\rm T} & = & \int d\tau \Bigg[ i\, t\gamma_{\uparrow}\left(\psi_{\uparrow}(0)+\psi_{\uparrow}^{\dagger}(0)\right)
-i\, t\gamma_{\downarrow}\left(\psi_{\downarrow}(0)+\psi_{\downarrow}^{\dagger}(0)\right) \nonumber\\
& & +\tilde{t}\gamma_{\uparrow}\left(\psi_{\downarrow}(0)-\psi_{\downarrow}^{\dagger}(0)\right)-\tilde{t}\gamma_{\downarrow}\left(\psi_{\uparrow}(0)-\psi_{\uparrow}^{\dagger}(0)\right)\nonumber\\
& & -\Delta i\gamma_{\uparrow}\gamma_{\downarrow}\left(-i\psi_{\uparrow}^{\dagger}(0)\psi_{\downarrow}(0)+i\psi_{\downarrow}^{\dagger}(0)\psi_{\uparrow}(0)\right) \nonumber\\
& & +\tilde{\Delta}i\gamma_{\uparrow}\gamma_{\downarrow}\left(\psi_{\uparrow}^{\dagger}(0)\psi_{\uparrow}(0)-\psi_{\downarrow}^{\dagger}(0)\psi_{\downarrow}(0)\right) \Big].
\end{eqnarray}
We have omitted here the irrelevant terms, e.g. $\Delta_{\text{AN}}$, analogous to those considered in Sec. \ref{sec:SU2NoRashba}.
After the bosonization, the boundary action reads
\begin{eqnarray}
S_{T}&=&\int d\tau \Bigg[ \frac{t}{2\pi a} \left( i\gamma_{\uparrow}\Gamma_{\uparrow}\cos\frac{\theta_{\rho}+
\theta_{\sigma}}{\sqrt{2}}-i\gamma_{\downarrow}\Gamma_{\downarrow}\cos\frac{\theta_{\rho}-\theta_{\sigma}}{\sqrt{2}}\right) \nonumber\\
&& + \frac{\tilde{t}}{2\pi a}\left(i\gamma_{\downarrow}\Gamma_{\uparrow}\sin\frac{\theta_{\rho}+\theta_{\sigma}}{\sqrt{2}}
- i\gamma_{\uparrow}\Gamma_{\downarrow}\sin\frac{\theta_{\rho}-\theta_{\sigma}}{\sqrt{2}}\right) \nonumber\\
& & -\frac{\Delta}{2\pi a}\gamma_{\uparrow}\gamma_{\downarrow}\Gamma_{\uparrow}\Gamma_{\downarrow}\cos\sqrt{2}\theta_{\sigma}
+\frac{\tilde{\Delta}}{2\pi v}i\gamma_{\uparrow}\gamma_{\downarrow} \frac{i \partial_{\tau}\theta_{\sigma}}{\sqrt{2}} \Bigg].
\end{eqnarray}
Note the appearance of the new marginal term described by coupling constant $\tilde{\Delta}$.
We now perform a perturbative RG analysis up to the second-order in coupling coefficients. The details of the calculations are presented in Appendix \ref{app:RG2order_SU2_BreakingBoundary}. Here we summarize our results for $K_{\sigma}=1$:
\begin{eqnarray}\label{eq:RG2}
\frac{dt}{dl} & = & \left(\frac{3}{4}-\frac{1}{4K_{\rho}}-\frac{\Delta}{4\pi v } \right)t -\frac{\tilde{\Delta} \tilde{t}}{2\pi v } , \label{eq:U1B_RG_t} \\
\frac{d\tilde{t}}{dl} & = & \left(\frac{3}{4}-\frac{1}{4K_{\rho}}+\frac{\Delta}{4\pi v }\right)\tilde{t} -\frac{ \tilde{\Delta} t }{2\pi v } , \label{eq:U1B_RG_tt}\\
\frac{d\Delta}{dl} & = & -\left(\frac{1}{K_{\rho}}-1\right) \frac{t^{2}-\tilde{t}^{2}}{4 \pi v},\label{eq:U1B_RG_Delta} \\
\frac{d\tilde{\Delta}}{dl} & = & - B(K_{\rho}) \frac{t \tilde{t}}{4 \pi v} .\label{eq:U1B_RG_tDelta}
\end{eqnarray}
The generation of the $\Delta$ term (proportional to $t^2-\tilde{t}^2$) originates from the processes involving two different spin channels of the lead whereas
the generation of the $\tilde{\Delta}$ term (proportional to $t\tilde{t}$ ) comes from processes within the same spin channel. Both of these terms can be generated {\it only} in the presence of the interaction in the lead. This fact follows from the definition of the
function $B(K_{\rho})$
\begin{equation}
B(K_{\rho}) = \frac{ C(1/2K_\rho-1/2) }{C(1/2)\, C(1/2K_\rho)} \left(\frac{1}{K_{\rho}}+1 \right)>0.
\end{equation}
Here the function $C(\nu)$ is defined as
\begin{align}\label{eq:Cnu}
C(\nu)=\lim_{\delta \rightarrow 0^+} \int_0^{\infty} \frac{e^{-\delta z}\cos(z)}{(z+1)^\nu}dz,
\end{align}
and originates from the integration over relative coordinate, $\tau-\tau'$ during the RG procedure, see Appendix \ref{app:RG2order_SU2_BreakingBoundary}.
In the non-interacting limit, $K_\rho\rightarrow 1$, $C(\nu\rightarrow 0^{+})\propto \nu$, and thus, the RG equation for $\tilde\Delta$ becomes
\begin{equation} \label{eq:RGeqforDelta}
\frac{d\tilde{\Delta}}{dl} \approx -\frac{c_5}{4\pi v}\left(\frac{1}{K_{\rho}}-1\right) t \widetilde{t},
\end{equation}
where numerical constant $c_5\approx 11.5$.
As mentioned, both $\Delta$ and $\tilde{\Delta}$ cannot be generated in the RG in the absence of interactions in the lead (i.e. $K_\rho=1$).
Using Eqs.\eqref{eq:RG2} it is instructive to analyze first the flow in the non-interacting limit, in which case $\Delta=\tilde{\Delta}=0$. Both $t$ and $\tilde{t}$ are relevant and flow to strong coupling. As follows from the discussion in the previous section,
the exact boundary condition at the Andreev reflection fixed point is determined by the initial values of $t$ and $\tilde{t}$ and
we can identify the corresponding limits by looking at the scattering matrix, i.e. $t \gg \tilde t $ corresponds
to $\psi_{\sigma}(0)=-\psi^\dag_{\sigma}(0)$, $t \ll \tilde t $ corresponds to $\psi_{\sigma}(0)=\psi^\dag_{\sigma}(0)$ and finally
$t = \tilde t $ corresponds to $\psi_{\sigma}(0)=-i\psi^\dag_{-\sigma}(0)$, see Fig. \ref{fig:RGflow}a.
We now analyze the RG flow for not-too-strong repulsive interactions $1/3\lesssim K_\rho<1$.
First of all, one can notice that even if we start with initial conditions $\Delta(l_0)=0$, $\tilde{\Delta}(l_0)=0$,
the corresponding four-fermion terms are going to be generated by the RG procedure. Here $l_0$ is initial length cutoff.
Since the couplings $\Delta$ and $\tilde{\Delta}$ affect the RG flow differently, we now have 4-parameter phase diagram.
Based on the perturbative RG equations, one can see that both $t$ and $\tilde{t}(l)$ will grow under RG, see Fig. ~\ref{fig:RGflow}. Thus, normal reflection fixed point is unstable in this parameter regime.
\subsubsection{RG analysis near spin-flip Andreev reflection fixed point $\mathbb{SFA}$}
We now analyze the stability of the spin-flip Andreev reflection fixed point $\mathbb{SFA}$ which corresponds to the following boundary conditions:
\begin{eqnarray}\label{eq:SFA_BC}
\psi_{L\uparrow}(0) &=& -i\psi^\dagger_{R\downarrow}(0), \\
\psi_{L\downarrow}(0) &=& -i\psi^\dagger_{R\uparrow}(0).
\end{eqnarray}
In the bosonization language, the bosonic fields $\phi_{\sigma}(0)=0$ and $\theta_{\rho}(0)=-\pi/(2\sqrt{2})$ are pinned,
and the Klein factors have the relation $\Gamma_{\uparrow L}=\Gamma_{\downarrow R}$ and
$\Gamma_{\downarrow L}=\Gamma_{\uparrow R}$. Now, let us study all the fermion bilinear perturbations at the boundary allowed by TR symmetry. First of all, one can show that the normal backscattering is vanishing in this case, in agreement with the scattering calculation in Sec.~\ref{sec:IIB1}.
Indeed, using the boundary conditions~\eqref{eq:SFA_BC} one can show that
\begin{eqnarray}
&&\psi_{L\uparrow}^{\dagger}(0)\psi_{\uparrow,R}(0)+ \psi_{R\downarrow}^{\dagger}(0)\psi_{L\downarrow}(0) + h.c.\nonumber \\
&&=-i\psi_{L\uparrow}^{\dagger}(0)\psi_{L\downarrow}^{\dagger}(0) + i\psi_{L\uparrow}(0)\psi_{L\downarrow}(0) +h.c. = 0
\end{eqnarray}
Note that for s-wave spin-singlet superconductor the boundary conditions are different:
$\psi_{L\uparrow}(0)=\mp i\psi^\dagger_{R\downarrow}(0)$ and $\psi_{L\downarrow}(0)=\pm i\psi^\dagger_{R\uparrow}(0)$, and the backscattering term
$\sim \sin \sqrt{2}\phi_{\rho}$ does not vanish. Since this term is relevant for $K_\rho<1$, the Andreev reflection fixed point is unstable in an s-wave superconductor-LL junction.
Let's now consider allowed Andreev reflection bilinear processes. Among those, the only allowed bilinear term is spin-conserving Andreev reflection:
\begin{eqnarray}
H_{1B}^{SFA}&=&\lambda_1^{SFA}(\psi_{L\uparrow}^{\dagger}\psi_{\uparrow,R}^{\dagger} + \psi_{R\downarrow}^{\dagger}\psi_{L\downarrow}^{\dagger} + h.c.) \nonumber\\
&=& \lambda_1^{SFA}(i\psi_{L\uparrow}^{\dagger}\psi_{L\downarrow} + i\psi_{R\downarrow}^{\dagger}\psi_{\uparrow,R} + h.c.) \nonumber\\
&=& 2\frac{\lambda_1^{SFA}}{2\pi a} \left(i\Gamma_{L\uparrow}\Gamma_{L\downarrow}+i\Gamma_{R\downarrow}\Gamma_{R\uparrow}\right)\cos\sqrt{2}\theta_{\sigma}.
\end{eqnarray}
Additionally, we also consider the following four-fermion term
\begin{eqnarray}
H_{2B}^{SFA}&=&\frac{\lambda_2^{SFA}}{(2\pi a)^2}(\psi_{L\uparrow}^{\dagger}\psi_{R\uparrow}\psi_{L\downarrow}^{\dagger}\psi_{R\downarrow}+h.c.)\nonumber\\
&=& 2\lambda_2^{SFA} \;\Gamma_{L\uparrow}\Gamma_{R\uparrow}\Gamma_{L\uparrow}\Gamma_{L\downarrow}\Gamma_{R\downarrow}\cos2\sqrt{2}\phi_{\rho},
\end{eqnarray}
which corresponds to two-electron backscattering. The leading order perturbative RG equations for $\lambda_1^{SFA}$
and $\lambda_2^{SFA}$ are give by
\begin{eqnarray}
\frac{d\lambda_1^{SFA}}{dl}&=& \left(1-\frac{1}{K_{\sigma}}\right)\lambda_1^{SFA}, \\
\frac{d\lambda_2^{SFA}}{dl}&=& (1-4K_{\rho})\lambda_2^{SFA}.
\end{eqnarray}
One can see that the first term $\lambda_1^{SFA}$ is marginal for $\mathbb{SU}(2)$ symmetric Luttinger liquid lead $K_{\sigma}=1$,
whereas the second coupling becomes relevant for $K_{\rho}<1/4$ indicating that the $\mathbb{SFA}$ fixed point becomes unstable
for strong repulsive interactions. If the $SU(2)$ spin symmetry is broken in the lead, the $\mathbb{SFA}$ fixed point
becomes unstable for $K_{\sigma}>1$, and the system will flow towards the $\mathbb{A\times A}$ fixed point.
On the other hand, the $\mathbb{SFA}$ is stable for $K_{\sigma}<1$.
\subsubsection{RG analysis near spin-conserving Andreev fixed point $\mathbb{A \times A}$}
As shown in Sec.~\ref{sec:IIB1}, the boundary conditions near $\mathbb{A\times A}$ fixed point are $\psi_{L,s}(0) = e^{i\alpha}\psi_{R,s}^{\dagger}(0)$ with $\alpha=0$ or $\pi$. Thus, the boson fields are $\theta_{\rho}=\pm\pi/\sqrt{2}$ and $\theta_{\sigma}=0$ are pinned at the boundary, and the Klein factors satisfy the relations
$\Gamma_{L,s}=\Gamma_{R,s}=\Gamma_{R}$. In the $U(1)$-conserving case, there are only
irrelevant perturbations for $1/3<K_{\rho}<1$ such as two-electron backscattering
\begin{align}
H_{2B}^{A\times A}=&\lambda_2^{A\times A} \psi_{L\uparrow}^{\dagger}(0)\psi_{R\uparrow}(0)\psi_{L\downarrow}^{\dagger}(0)\psi_{R\downarrow}(0)+h.c. \nonumber \\
&= \frac{\lambda_2}{(2\pi a)^2}\sin(2\sqrt{2}\phi_{\rho}).
\end{align}
Additionally, if $U(1)$ symmetry is broken, the spin-flip Andreev
reflection processes are allowed
\begin{eqnarray}
H_{1B}^{A\times A}&=& \lambda_1^{A\times A}(\psi_{R,\uparrow}^{\dagger}(0)\psi_{L,\downarrow}^{\dagger}(0)-\psi_{R,\downarrow}^{\dagger}(0)\psi_{L,\uparrow}^{\dagger}(0))+h.c.\nonumber\\
&=& 4 i \Gamma_{\uparrow} \Gamma_{\downarrow} \sin \sqrt{2} \phi_{\sigma}.
\end{eqnarray}
The leading order perturbative RG equations for $\lambda_1^{A\times A}$
and $\lambda_2^{A\times A}$ are give by
\begin{eqnarray}
\frac{d\lambda_1^{A\times A}}{dl}&=& (1-K_{\sigma})\lambda_1^{A\times A}, \\
\frac{d\lambda_2^{A\times A}}{dl}&=& (1-4K_{\rho})\lambda_2^{A\times A}.
\end{eqnarray}
One can see that the first term $\lambda_1^{A\times A}$ is marginal for $SU(2)$ symmetric Luttinger liquid lead $K_{\sigma}=1$,
whereas the second coupling becomes relevant for $K_{\rho}<1/4$ indicating that $\mathbb{A\times A}$ fixed point becomes unstable
for strong repulsive interactions. If the $SU(2)$ spin symmetry is broken in the lead, the $\mathbb{A\times A}$ fixed point
becomes unstable for $K_{\sigma}<1$, and the system will flow towards the $\mathbb{SFA}$ fixed point.
On the other hand, the $\mathbb{A \times A}$ is stable for $K_{\sigma}>1$. Exactly at $K_\sigma=1$, both $\lambda_1^{A\times A}$ and $\lambda_1^{SFA}$ terms are marginal and compete with each other. Thus, generically both spin-conserving and spin-flip Andreev reflection processes will be present and their relative strength depends on microscopic details. This conclusion is consistent with the non-interacting results ($K_{\rho}=1$) discussed in Sec.\ref{sec:IIB1}. Our main results are summarized in Fig. \ref{fig:RGflow}.
\section{Majorana Kramers pair - Quantum dot - Normal lead junction}\label{sec:QD}
\subsection{Theoretical model}\label{sec:QDmodel}
In this section we study effect of local electron-electron interactions and consider the system consisting of a QD with a single spin-degenerate level coupled to MKP $\gamma_{\uparrow,\downarrow}$, localized at the end of a TRITOPS, and a NL. The schematic plot of the device is shown in Fig. \ref{fig:device} b). Assuming that TR symmetry and $\mathbb{U}(1)$-spin rotation symmetry are preserved and the induced gap in the topological superconductor is sufficiently larger than other energy scales of the problem, the low-energy effective Hamiltonian of the system can be written as
\begin{align}
H&=\sum_\sigma\epsilon d^\dagger_\sigma d_\sigma + U n_\uparrow n_\downarrow + V +H_{NL}\\
V&=\sum_\sigma [i\lambda_\sigma\gamma_\sigma (d_\sigma + d^\dagger_\sigma)+t_\sigma(d^\dagger_\sigma \psi_\sigma(0)+\text{h.c.})]
\end{align}
where $d^\dagger_\sigma$ and $d_\sigma$ are creation and annihilation operators on the QD, $n_\sigma=d^\dagger_\sigma d_\sigma$, $\epsilon$ is the chemical potential of the QD, $U$ is the strength of the electron-electron interaction on the QD, $\psi^\dagger_\sigma$ and $\psi_\sigma$ are fermion creation and annihilation operators in the NL, and $t(\lambda_\sigma)$ is the tunneling coefficient between the NL(MKP) and the QD. For the perturbative RG analysis, we adopted the same Hamiltonian for NL as Eq. \eqref{eq:Hlead} with $K=1$. For slave-boson mean-field theory analysis, we assumed quadratic dispersion $\xi_k$ for the NL. We set $t_\sigma$ and $\lambda_\sigma$ to be real. Time-reversal symmetry requires $t_\uparrow=t_\downarrow=t$ and $\lambda_\uparrow=-\lambda_\downarrow=\lambda$. The Hamiltonian $H_{NL}$ represents semi-infinite NL ($x\geq0$) with hopping $t_0$. We are interested in the limit where $\epsilon<0$, $U+\epsilon>0$ such that the QD favors single occupation, and weak coupling regime $|t|,|\lambda|\ll \min(-\epsilon,U-\epsilon)$. We also consider the non-interacting limit for NL. In this limit, one can simplify the effective Hamiltonian by projecting it onto single-occupation subspace~\cite{wolff}, see Appendix \ref{app:SW} for details. The effective Hamiltonian becomes $H=H_{\text{NL}}+H_b$ with the boundary Hamiltonian $H_b$ being
\begin{widetext}
\begin{align}
H_b&=\xi_+\bigg[\frac{t^2}{2}\vec{S}\cdot\vec{s}(0)+\frac{i\lambda^2}{2}\gamma_\uparrow\gamma_\downarrow S_y +\frac{i\lambda t}{2} \bigg(\gamma_\uparrow(\psi_\uparrow+\psi^\dagger_\uparrow)S_z+ \gamma_\downarrow(\psi_\downarrow+\psi^\dagger_\downarrow)S_z +\gamma_\uparrow(\psi_\downarrow S^- +\psi^\dagger_\downarrow S^+)-\gamma_\downarrow(\psi_\uparrow S^+ +\psi^\dagger_\uparrow S^-) \bigg) \bigg]\nonumber\\
&+\xi_-\left[\frac{i\lambda t}{2}\left(\gamma_\uparrow(\psi_\uparrow+\psi^\dagger_\uparrow)-\gamma_\downarrow(\psi_\downarrow+\psi^\dagger_\downarrow)\right)\right],
\label{eq:Hb}
\end{align}
\end{widetext}
where $\vec{S}=d^\dagger_\alpha \boldsymbol{\sigma}_{\alpha\beta} d_\beta$, $\vec{s(0)}=\psi^\dagger_\alpha(0) \boldsymbol{\sigma}_{\alpha\beta} \psi_\beta(0)$, $S^+=S_x+iS_y$, $S^-=S_x-iS_y$ and the coefficients $\xi_\pm$ are defined as
\begin{equation}
\xi_\pm=\frac{1}{|\epsilon|}\pm\frac{1}{U-|\epsilon|}.
\end{equation}
In the limit $\lambda \rightarrow0$, the first term $\sim t^2$ drives the system to the Kondo fixed point where a spin in QD and a spin in the lead form a spin-singlet state. In the presence of the Majorana coupling $\lambda$, additional terms appear in the Hamiltonian. These Majorana-induced couplings favor the strong-correlation between QD spin and MKP, and, therefore, compete with Kondo coupling.
The critical difference between the present Hamiltonian~\eqref{eq:Hb} and that of time-reversal broken case with single Majorana mode, for example in Ref. \onlinecite{Kondo-Majorana}, is the presence of the second term $\sim\lambda^2$. This time-reversal preserving interaction term between QD and MKP replaces the Zeeman-like coupling in the single Majorana mode case. While the Zeeman-like coupling becomes zero at the particle-hole symmetric point in the previous study\cite{Kondo-Majorana}, this interaction term is proportional to $\xi_+$ and is always non-zero for any position of the level $\epsilon$ in the dot. {Therefore, one cannot apply the same method as in Ref.~[\onlinecite{Kondo-Majorana}] to find the exact solution at the particle-hole symmetric point.} To understand low-energy properties of the system, we present below the results from two complementary calculations: perturbative RG analysis and slave-boson mean-field theory in the limit of an infinite on-site repulsion.
\subsection{Weak coupling perturbative RG analysis}\label{sec:QDRG}
In order to understand the effect of Majorana induced couplings on the infrared(IR) fixed point the system flows to, we study RG flow of the boundary couplings in the weak-coupling limit.
First, we introduce the following rescaled couplings: $M(l_0)=\xi_+\lambda^2$, $T_1(l_0)=\lambda t\xi_-$, $T_2(l_0)=\lambda t \xi_+$ and $J(l_0)=t^2\xi_+$. After performing standard bosonization procedure and rescaling the parameters, we obtain the following effective action at the boundary:
\begin{align}
S_{b}&=\int \frac{d\tau}{2\pi a} \bigg\{ iM\gamma_\uparrow\gamma_\downarrow S_y -\frac{iaJ^zS_z}{\sqrt{2} v}\partial_\tau\theta_\sigma \nonumber\\
+&iT_1 \left[ \gamma_\uparrow\Gamma_\uparrow\cos\left(\frac{\theta_\rho+\theta_\sigma}{\sqrt{2}}\right)- \gamma_\downarrow\Gamma_\downarrow\cos\left(\frac{\theta_\rho-\theta_\sigma}{\sqrt{2}}\right) \right] \nonumber\\
+&iT_2^z S_z \left[ \gamma_\uparrow\Gamma_\uparrow \cos\left(\frac{\theta_\rho+\theta_\sigma}{\sqrt{2}}\right)+\gamma_\downarrow\Gamma_\downarrow \cos\left(\frac{\theta_\rho-\theta_\sigma}{\sqrt{2}}\right) \right] \nonumber\\
+&iT_2^\bot \bigg[ \gamma_\uparrow\Gamma_\downarrow\left(S_x\cos \left(\frac{\theta_\rho-\theta_\sigma}{\sqrt{2}}\right)+S_y\sin\left(\frac{\theta_\rho-\theta_\sigma}{\sqrt{2}}\right)\right) \nonumber\\
-&\gamma_\downarrow\Gamma_\uparrow\left( S_x \cos\left(\frac{\theta_\rho+\theta_\sigma}{\sqrt{2}}\right)-S_y\sin\left(\frac{\theta_\rho+\theta_\sigma}{\sqrt{2}}\right)\right)\bigg]\nonumber\\
-&iJ^\bot \Gamma_\uparrow\Gamma_\downarrow \left(S_x\sin{\sqrt{2}\theta_\sigma}+S_y\cos\sqrt{2}\theta_\sigma \right) \bigg\}
\label{eq:Sb0}
\end{align}
Here we have introduced couplings $T_2^{z,\bot}$ and $J^{z,\bot}$ for convenience. We will recover spin-rotation symmetry by setting $T_2^z=T_2^\bot$ and $J^z=J^\bot$ at the end of the calculation. We will focus on the non-interacting limit for NL, but adding small repulsive interaction in NL does not change our conclusion.
Let us now perform perturbative RG analysis up to the second order in couplings near ultraviolet normal reflection fixed point. The procedure of the calculations is similar to the one presented in Appendix \ref{app:RG2order_SU2} and \ref{app:RG2order_SU2_BreakingBoundary}. The RG equations for the couplings read
\begin{eqnarray}
\frac{dM}{dl}&=&M+\frac{T_2^2}{\pi v}\\
\frac{dT_1}{dl}&=&\frac{T_1}{2}\\
\frac{dT_2}{dl}&=&\frac{T_2}{2}+\frac{T_2J}{\pi v}\\
\frac{dJ}{dl}&=&\frac{J^2}{\pi v}
\end{eqnarray}
From these RG equations, we can see that $M$ is the most relevant coupling while the Kondo coupling is only marginally relevant. Thus, the system generically flows to the Majorana strong coupling fixed point.
If initially $M(l_0) \ll J(l_0)$, the system can still reach the Kondo strong coupling fixed point.
One can estimate the crossover scale, $\lambda_c$, by solving $M(l^*)=J(l^*)\sim 1$ ($l^*$ is the crossover length scale) which leads to the following estimate for the critical coupling
\begin{equation}
\lambda_c \sim \frac{1}{\xi_+} \exp \left(-\frac{\pi v}{2 \xi^+ t^2}\right),
\label{eq:lambdac}
\end{equation}
which defines a crossover between the two regimes. In deriving this estimate, we have ignored the second order contributions from $T_2^2$ term assuming that it is small.
Around the Kondo strong coupling fixed point, the spin operator in QD is absorbed by NL and, therefore, acquires scaling dimension one. As a result, $T_2$ and $M$ terms become irrelevant and marginal, respectively. However, the term proportional to $T_1$ is still relevant and drives the system to Andreev reflection strong coupling fixed point with $\mathbb{A\times A}$ boundary condition and $G(0)=4e^2/h$, see Sec.~\ref{sec:SU2NoRashba}.
Let us now study the nature of Majorana strong coupling fixed point defined by $M(l^*)\sim 1$ and $J(l^*) \ll 1$. The two degenerate (Kramers) states that minimize $iM\gamma_\uparrow\gamma_\downarrow S_y$ term are
\begin{eqnarray}
\ket{\psi_1}&=&\ket{i\gamma_\uparrow\gamma_\downarrow=-1, S_y=1}\nonumber\\\,\\
\ket{\psi_2}&=&\ket{i\gamma_\uparrow\gamma_\downarrow=1, S_y=-1}.\nonumber\\\nonumber
\end{eqnarray}
Assuming that $M$ is large, one can project the rest of the boundary terms on to this low-energy manifold and simplify the boundary problem. Since the ground state is an eigenstate of $S_y$ and $i\gamma_\uparrow\gamma_\downarrow$, the terms that are proportional to $\gamma\otimes I$ and $\gamma\otimes S_y$ will be projected to zero. The remaining boundary terms at particle-hole symmetric point (i.e. $T_1=0$) are
\begin{align}\label{eq:Hmb}
\!\!H_{\rm mb}
&=i T_2(l^*)\left[ \beta_\uparrow(\psi_\uparrow+\psi_\uparrow^\dagger)-\beta_\downarrow(\psi_\downarrow+\psi_\downarrow^\dagger)\right] \nonumber\\
&+\frac{iJ(l^*)}{2} \beta_\uparrow\beta_\downarrow(-i\psi_\uparrow^\dagger\psi_\downarrow+i\psi_\downarrow^\dagger\psi_\uparrow),
\end{align}
where introduced generalized Majorana operators $\beta_\uparrow=(\gamma_\uparrow S_z -\gamma_\downarrow S_x)/2$ and $\beta_\downarrow=-(\gamma_\downarrow S_z +\gamma_\uparrow S_x)/2$. In the ground state manifold with $i\gamma_\uparrow\gamma_\downarrow S_y=-1$, these new operators behave as MKP, $\{\beta_\sigma,\beta_{\sigma'}\}=2\delta_{\sigma,\sigma'}$. Now one can notice that the effective Hamiltonian~\eqref{eq:Hmb} is exactly the same as that in Eq.~\eqref{eq:model0} with $K_\rho=1$ and $\Delta_{\text{AN}}=0$. Therefore, using the results from the previous section and the condition $T_2(l^*)\gg J(l^*)$, we can immediately conclude that the system will flow to the strong coupling fixed point under RG and will be governed by the $\mathbb{A\times A}$ boundary condition $\psi_\sigma(0)=-\psi_\sigma^\dagger(0)$.
As follows from the aforementioned analysis, the coupling of QD to MKP leads to a non-trivial many-body ground-state where the spin on the QD gets entangled with the fermion parity of the MKP. Due to the change in the boundary conditions for lead electrons, the zero-bias tunneling conductance is $G=4 e^2/h$ due to perfect Andreev reflection phenomenon. Further insight about the physical properties of the system can be obtained using a complementary approach - slave-boson mean field theory.
\subsection{Slave-boson mean field theory}\label{sec:QDSB}
In this section, we develop a slave-boson mean-field approach for an infinite repulsive interaction in QD (i.e. $U \rightarrow \infty$). In this case one can completely exclude double occupancy state from the Hilbert space, see, for example, Ref. \cite{Nagaosa_book}. Next, one can represent the creation and annihilation operators for the QD as $d^\dagger_\sigma\rightarrow f^\dagger_\sigma b$ and $d_\sigma \rightarrow f_\sigma b^\dagger$ with an additional constraint $b^\dagger b+\sum_\sigma f^\dagger_\sigma f_\sigma=1$ where $b$ is a boson operator representing an empty state. Thus, the effective action of the system in terms of new fields variables reads
\begin{widetext}
\begin{align}
S_{\text{sb}}=&\int d\tau \sum_\sigma \bigg[\sum_k\psi^*_{k,\sigma}(\partial_\tau+\xi_k)\psi_{k,\sigma}+f^*_\sigma(\partial_\tau+\epsilon)f_\sigma + i\lambda_\sigma\gamma^1_\sigma(f_\sigma b^* + f^*_\sigma b)+\sum_k t(f^*_\sigma\psi_{k,\sigma} b+\psi^*_{k,\sigma}f_\sigma b^*)\nonumber\\
&+\frac{1}{2}\sum_{i=1,2}\gamma^i_\sigma\partial_\tau \gamma^i_\sigma +i\delta_{1\sigma}\gamma^1_\sigma\gamma^2_\sigma+i\delta_2 \gamma^1_\sigma\gamma^2_{-\sigma} +\eta\left(\frac{b^* b-1}{2}+f^*_\sigma f_\sigma\right)\bigg]
\label{eq:mfe0},
\end{align}
\end{widetext}
where $\eta$ is the Lagrange multiplier, $\gamma^1$ and $\gamma^2$ correspond to the Majorana modes at the end of the TSC near the QD and at the opposite end. $\delta_{1\uparrow}=-\delta_{1\downarrow}=\delta_1$ and $\delta_2$ mix the Majorana modes at the opposite ends for finite size system.
\subsubsection{Mean-field solution}
We now develop self-consistent mean-field theory for the problem. We first calculate mean-field solution for the action~\eqref{eq:mfe0} and replace boson fields with their mean-field value $\langle b \rangle= \langle b^* \rangle=b$ and solve for $b$ and $\eta$. Here, without loss of generality, we assumed that $b$ is real since the phase can be gauged away. In the next section, we will study effect fluctuations around the mean-field saddle point and specify precisely the meaning of the mean-field solution for this low-dimensional model.
The mean-field equations can be obtained by minimizing the action~\eqref{eq:mfe0}:
\begin{eqnarray}
\frac{\partial S}{\partial \eta}&=&b^2+\sum_{\sigma}\langle f^*_\sigma f_\sigma \rangle -1=0 \label{eq:mfe1}\\
\frac{\partial S}{\partial b}&=&2b\eta+t\sum_{k,\sigma}(\langle f^*_\sigma\psi_{k,\sigma}\rangle + \langle \psi^*_{k,\sigma} f_\sigma \rangle)\nonumber\\
&&+i \sum_\sigma \lambda_\sigma \langle \gamma^1_\sigma(f^*_\sigma + f_\sigma)\rangle =0 \label{eq:mfe2}
\end{eqnarray}
The details of the calculation of the correlation functions are presented in the Appendix \ref{app:SB1}. We first consider the limit $T, \delta_1,\delta_2 \rightarrow 0$ and assume that $|\epsilon|\gg |\lambda|,\Gamma$ where $\Gamma=\pi \nu_F|t|^2$ such that the probability for empty state in QD $b^2$ is small. In this limit, the first equation becomes
\begin{equation}
\epsilon+\eta\approx \frac{\pi}{2}\Gamma b^4,
\end{equation}
Substituting $\eta\approx -\epsilon$ back into Eq.\eqref{eq:mfe2} and neglecting smaller terms, one finds
\begin{equation}
\eta-\frac{2\Gamma}{\pi}\ln\frac{\Lambda}{\Gamma b^2}-\frac{|\lambda|}{\sqrt{2}b}\approx 0\label{eq:sbr}
\end{equation}
where we assumed $\lambda\gg\Gamma b$.
For $\lambda\rightarrow 0$ we recover the solution for the Kondo-dominated regime:
\begin{eqnarray}
T_K&\equiv&\Gamma b^2=\Lambda e^{-\frac{\pi|\epsilon|}{2\Gamma}}.
\end{eqnarray}
If Majorana coupling $\lambda \gg \lambda_c$, $b$ is determined by the last term in Eq. \eqref{eq:sbr}:
\begin{equation}
b\approx\frac{|\lambda|}{\sqrt{2}|\epsilon|}.
\end{equation}
The crossover between two regimes occurs at
\begin{equation}
\lambda_c \approx \sqrt{\frac{2\Lambda}{\Gamma}}|\epsilon|e^{-\frac{\pi|\epsilon|}{4\Gamma}}
\end{equation}
which qualitatively agrees with the estimate for $\lambda_c$ from the RG analysis, see Eq.~\eqref{eq:lambdac}.
In the presence of the Majorana splitting $\delta_1$ and $\delta_2$ and for arbitrary value of $\lambda/\Gamma b$, we can solve the mean-field equations numerically. In terms of $\delta_1^2+\delta_2^2\equiv\delta^2$, the second mean-field equation\eqref{eq:mfe2} now becomes
\begin{equation}
\frac{|\epsilon|}{\Gamma}-\frac{2}{\pi}\ln\frac{\Lambda}{\Gamma b^2}-2I(b,\tilde\lambda,\tilde\delta)=0,\label{eq:sbr2}
\end{equation}
where
\begin{align}
I(b,&\tilde\lambda,\tilde\delta)=\frac{b^2\tilde\lambda^2}{\pi}\times\\
&\int_0^\infty dx \frac{x(x-\tilde\delta^2-b^2\tilde\lambda^2)}{(x+b^4)(x(x-\tilde\delta^2-2b^2\tilde\lambda^2)^2+b^4(x-\tilde\delta^2)^2)} \nonumber
\end{align}
One can numerically solve the Eq. \eqref{eq:sbr2} for self-consistent solution $b$ as a function of $\tilde\lambda=\lambda/\Gamma$ and $\tilde\delta=\delta/\Gamma$, see Fig. \ref{fig:mfs2} for results. One can see that if we increase the splitting for Majoranas $\delta$, the Kondo correlations become more important and eventually start to dominate. As a result, the magnitude of $\lambda_c$ defining the crossover between two different fixed points is increased.
\begin{figure}[t]
\centering
\includegraphics[scale=0.45]{mfs2.pdf}
\caption{The solution $b$ of the mean-field equation as a function of $\tilde\lambda$ and $\tilde\delta$. We set $\epsilon=-6\Gamma$ and $\Lambda=50\Gamma$.}
\label{fig:mfs2}
\end{figure}
\subsubsection{Gaussian fluctuations around mean-field solution}\label{sec:gaussianfluctuation}
In the previous section, we found mean-field solution for $b$ and $\eta$ of Eq.~\eqref{eq:mfe0}. We now analyze the stability of the mean-field solution with respect to fluctuations. This issue is rather subtle, and has been discussed extensively in the context of the Kondo problem~\cite{Coleman'87}. Indeed, one can check that the action~\eqref{eq:mfe0} is invariant with respect to local gauge transformations $b \rightarrow b e^{i\theta}$ and $f\rightarrow e^{i\theta}f$. The mean-field solution appears to break this $\mathbb{U}(1)$ symmetry. However, as we will show below, the fluctuations will restore this symmetry.
We now make a transformation to the ``radial coordinates" and rewrite $b(\tau)=s(\tau)e^{i\theta(\tau)}$. One can check that the action~\eqref{eq:mfe0} is invariant with respect to local gauge transformations $b \rightarrow b e^{i\theta}$ and $f\rightarrow e^{i\theta}f$ and $\eta(\tau)\rightarrow \eta+i\partial_\tau\theta$. Therefore, we can expand the action in terms of fluctuations $\delta s(\tau)$ and $\partial_\tau \theta(\tau)$
\begin{equation}
s(\tau)=\bar{s}+\delta s(\tau) , \,\,\,\,\,\, \eta(\tau)=\bar\eta+i\partial_\tau\theta(\tau).
\end{equation}
around the corresponding saddle point. Here $\bar{s}$ is the mean-field solution for $b$, defined in the previous section.
After integrating out fermions, the effective action reads
\begin{equation}
S_{\text{eff}}=-\Tr\ln\left[ \mathcal{G}^{-1}(s,\eta) \right]+\int d\tau \left[\eta(s^2-1)+s\partial_\tau s \right]\label{eq:effS}
\end{equation}
where
\begin{equation}
\mathcal{G}^{-1}(s,\eta)=\left(
\begin{array}{cc}
-G_{f}^{-1} & sG_{\gamma} s\\
sG_{\gamma}s & -\tilde{G}_{f}^{-1}
\end{array}
\right)
\end{equation}
and
\begin{align}
G_{f}&=-\frac{1}{\partial_\tau+\epsilon+\eta+s(G_\psi+G_{\gamma})s}, \\
\tilde{G}_{f}&=-\frac{1}{\partial_\tau-\epsilon-\eta+s(\tilde{G}_\psi+G_{\gamma})s}, \\
G_\psi&=-\sum_k\frac{t^2}{\partial_\tau+\xi_k}, \,\,\,\,\, \tilde{G}_\psi=-\sum_k\frac{t^2}{\partial_\tau-\xi_k}, \\
G_{\gamma}&=-\frac{\lambda^2}{\partial_\tau+\delta}.
\end{align}
We now expand the fields $s$ and $\eta$ around their mean-field values and collect the quadratic terms in $\delta s$ and $\partial_\tau \theta$. In Matsubara frequency domain, the effective action for the Gaussian fluctuation can be written as
\begin{equation}
S^{(2)}_{\text{eff}}=\frac{1}{2\beta}\sum_\nu (\dot\theta_{-\nu} \,\, \delta s_{-\nu}) \left(\begin{array}{cc}
\Gamma^{\dot\theta \dot\theta}_\nu & \Gamma^{\dot\theta s}_\nu \\
\Gamma^{\dot\theta s}_\nu & \Gamma^{ss}_\nu
\end{array}\right) \left(\begin{array}{c}
\dot\theta_\nu \\
\delta s_\nu
\end{array}\right)
\end{equation}
where we introduced Fourier transform
\begin{equation}
\delta s(\tau) =\frac{1}{\beta}\sum_\nu \delta s_\nu e^{-i\omega_\nu \tau}, \,\,\,\,\, \dot\theta(\tau) =\frac{1}{\beta}\sum_\nu \dot\theta_\nu e^{-i\omega_\nu \tau}
\end{equation}
with bosonic Matsubara frequency $\omega_\nu=2\pi\nu/\beta$. The details of the calculation of the above matrix elements $\Gamma^{ij}$ are given in Appendix \ref{app:SB2}. Here we simply highlight the main results. We first note that $\Gamma^{\dot\theta s}_\nu\approx 2i\bar{s} $ near the mean-field solution $\bar{\eta} \approx -\epsilon$. Diagonal element $\Gamma^{ss}_{\nu}$ and $\Gamma^{\theta \theta}_{\nu}$ can be obtained using the analytic continuation of fermionic Matsubara frequency $i\omega_n\rightarrow \omega$ and integrating around the two branch cuts $\text{Im}[\omega]=0$ and $\text{Im}[\omega]=-\omega_\nu$. The correlation function of $\delta s$ and $\dot\theta$ is given by
\begin{eqnarray}
D_{\dot\theta \dot\theta}(i\omega_{\nu})&=&\frac{\Gamma^{ss}_\nu}{\Gamma^{\dot\theta \dot\theta}_\nu\Gamma^{ss}_\nu+4\bar{s}^2},\\
D_{ss}(i\omega_\nu)&=&\frac{\Gamma^{\dot\theta \dot\theta}_\nu}{\Gamma^{\dot\theta \dot\theta}_\nu\Gamma^{ss}_\nu+4\bar{s}^2},
\end{eqnarray}
and govern the dynamics of the fluctuating fields $\delta s(\tau)$ and $ \dot\theta(\tau)$. We can now address the question regarding the restoration of the broken $\mathbb{U}(1)$ symmetry.
Let us consider the correlation function $\langle b(\tau)b^*(0)\rangle$. The mean-field solution assumes that $\langle b(\tau)b^*(0)\rangle \rightarrow \bar{s}^2$ for $\tau \rightarrow \infty$. It has been shown, however, in Ref.~\cite{Coleman'87} that above correlation function for the generalized Anderson model decays as a power-law $\langle b(\tau)b^*(0)\rangle\propto |\tau|^{-\alpha}$ with some non-universal exponent.
\begin{figure}[t]
\includegraphics[width=3.5in]{ImGss1.pdf}\\
\includegraphics[width=3.5in]{ImGss2.pdf}
\caption{The function $\text{Im}[D_{\dot\theta\dot\theta}(\Omega)]$ for different values of $\lambda$ and $\delta$. Here panels a) and b) correspond to $\delta=0.01$ and $\lambda=0.2$; we used $\Gamma=1$, $\epsilon=-5$, $\Lambda=50$ here. }
\label{fig:ImGss}
\end{figure}
We now perform a similar analysis for QD-MKP problem at hand. Since $\langle s(\tau)s(0)\rangle\sim \bar{s}^2$ in the long time limit, one can decouple amplitude and phase fluctuations
\begin{eqnarray} \label{eq:theta_corr}
\langle b(\tau)b^*(0)\rangle &\approx& \bar{s}^2 \langle e^{i(\theta(\tau)-\theta(0))}\rangle\nonumber\\
&=&\bar{s}^2 \exp(-\frac{1}{2}\langle[\theta(\tau)-\theta(0)]^2\rangle).
\end{eqnarray}
We can evaluate the exponent, following Ref.\cite{Coleman'87}, as
\begin{align}
\frac{1}{2}\langle[\theta(\tau)&-\theta(0)]^2\rangle=\frac{1}{\beta}\sum_{\nu\neq0} \frac{D_{\dot\theta\dot\theta}}{\omega_\nu^2}(1-e^{-i\omega_\nu\tau})\label{eq:bcontour}\\
&=-\oint\frac{d\Omega}{2\pi i}\frac{1-e^{-\Omega\tau}}{1-e^{-\beta \Omega}}\frac{D_{\dot\theta\dot\theta}(\Omega)}{\Omega^2}\\\nonumber
&\stackrel{T\rightarrow 0 }{=}-\int^\infty_0 \frac{d\Omega}{\pi}\frac{1-e^{-\Omega\tau}}{\Omega^2}\text{Im}[\lim_{\xi\rightarrow 0^+}D_{\dot\theta\dot\theta}(\Omega+i\xi)].\nonumber
\end{align}
Here Matsubara sum was evaluated by integrating along the branch cut $\text{Im}[\Omega]=0$ using the analytic continuation for bosonic Matsubara frequency $i\omega_\nu\rightarrow \Omega$. We find that $\text{Im}[D_{\dot\theta \dot\theta}(\Omega+i\epsilon)] \propto -\alpha\Omega$ in low frequency limit, see Fig. \ref{fig:ImGss}.
Here we eventually take $\epsilon \rightarrow 0$. Thus, the correlation function
\begin{equation}
\langle b(\tau)b^*(0)\rangle \propto \tau^{-\alpha}
\end{equation}
decays as a power law, which is a key result of this section. In this sense, the situation is analogous to slave-boson theory for the Kondo problem.
The expression for $\alpha$ as a function of $\lambda_0=\lambda/\Gamma\bar{s}$ in the limit of zero splitting for MKP, $\delta\rightarrow 0$, is given by
\begin{equation}
\alpha=\frac{1}{2}\frac{1}{(h(\lambda_0)+\bar{s}^2\pi^2/4)}
\end{equation}
\begin{equation}
h(\lambda)\!=\!\begin{cases}
-\frac{\ln[4\lambda^4]}{8\lambda^2}\!-\!\frac{1-4\lambda^2}{8\lambda^2\sqrt{1\!-\!8\lambda^2}}\ln\left[\frac{1\!-\!4\lambda^2+\sqrt{1-8\lambda^2}}{1-4\lambda^2-\sqrt{1-8\lambda^2}} \right], &\lambda<\frac{1}{2}\\
-\frac{\ln[4\lambda^4]}{8\lambda^2}\!-\!\frac{1-4\lambda^2}{4\lambda^2\sqrt{1\!-\!8\lambda^2}}\left(\frac{\pi}{2}\!-\!\tan^{-1}\frac{1-4\lambda^2}{\sqrt{1-8\lambda^2}}\right), & \lambda \geq \frac{1}{2}
\end{cases} \nonumber
\end{equation}
Using the corresponding mean-field solution of Eq. \eqref{eq:mfe2}, one can evaluate the exponent $\alpha$, see Fig. \ref{fig:exponent}. We find that the exponent $\alpha$ moderately increases with $\lambda$. When the Majorana splitting energy $\delta$ becomes larger, $\alpha$ decreases and eventually approaches the value in the Kondo limit $\alpha =\frac{1}{2}+O(\bar{s}^2)$.
\begin{figure}[t]
\includegraphics[width=3.4in]{exponent.pdf}
\caption{The exponent $\alpha$ as a function of Majorana coupling strength $\lambda$.}
\label{fig:exponent}
\end{figure}
Overall, we find that the correlation function~\eqref{eq:theta_corr} decays as a power law in the long-time limit which is qualitatively similar to phase fluctuations in the Kondo problem~\cite{Coleman'87}. This is the main result of this section showing that fluctuations ultimately restore $U(1)$ symmetry, in agreement with the Mermin-Wagner theorem, but the correlation function decays slowly in comparison with the ``disordered" high-temperature limit. The situation is reminiscent of quasi-long range order where the fluctuations ultimately restore broken symmetry but, at the same time, there is a well-defined mean-field amplitude of fluctuations (i.e. $\bar{s}\neq 0$) which opens up a gap in the spectrum.
\subsection{Differential tunneling conductance}
Using the mean-field theory developed in the previous sections, one can now calculate transport properties of the NL-QD-TSC junction. To compute the differential conductance $G$, one needs to compute scattering matrix of the system within the mean-field approximation. The slave-boson mean-field Hamiltonian can be written as
\begin{eqnarray}
H_{sb}&=&H_{NL}+\sum_\sigma\bigg[\sum_{k}tb(f^\dagger_\sigma \psi_{k,\sigma}+\psi^\dagger_{k,\sigma}f_\sigma)+\tilde\epsilon f^\dagger_\sigma f_\sigma\nonumber\\
&& +i\lambda_\sigma b\gamma^1_\sigma(f^\dagger_\sigma+f_\sigma)+i\delta_{1\sigma}\gamma^1_\sigma\gamma^2_\sigma+i\delta_2\gamma^1_\sigma\gamma^2_{-\sigma} \bigg].
\end{eqnarray}
The scattering matrix for electrons close to the Fermi level is given by
\begin{equation}
S(E)=1+2\pi i\hat{W}^\dagger(H_{local}-E-\pi i \hat{W}\hat{W}^\dagger)^{-1}\hat{W},
\end{equation}
where $H_{local}$ is the Hamiltonian describing the ``local impurity" and $\hat{W}\propto tb$ is the matrix of coupling constants between local degrees of freedom and lead electrons.
Using the scattering matrix one can compute the probability for Andreev reflection and ultimately obtain differential conductance $G(V)$. In agreement with the analysis in Sec. \ref{sec:QDRG}, we find that zero-bias differential conductance is quantized $G(0)=4e^2/h$. In the limit of small bias voltage and zero splitting $\delta\rightarrow0$, the differential conductance $G(V)$ reads
\begin{equation}
G(V)\approx\frac{4e^2}{h}\,\frac{\Gamma_{\rm eff}^2}{\Gamma_{\rm eff}^2+(eV)^2}
\end{equation}
with the width of the zero-bias peak changing from $\Gamma_{\text{eff}}\approx\text{min}\{ T_K, \frac{2\lambda^2}{\Gamma} \}$ in Kondo-dominated to $\Gamma_{\text{eff}}\approx\frac{\Gamma\lambda^2}{2\epsilon^2}$ in the Majorana-dominted regime.
In addition to the differential conductance, the signatures of MKP should be observable in shot noise and full counting statistics measurements as have been discussed in the context of a quantum dot coupled to a single Majorana zero mode, see, e.g., Refs.~\onlinecite{QDShotNoise,QDFCS}.
\section{Conclusions}\label{sec:conclusion}
We study two new boundary impurity problems involving MKPs: Luttinger liquid - MKP and NL - QD - MKP junctions, see Fig. \ref{fig:device}. The presence of MKPs in these systems leads to a drastic change of their physical properties. A well-known example is the change of the differential tunneling conductance through such a junction from essentially zero (perfect normal reflection) to $4e^2/h$ (perfect Andreev reflection)~\cite{wong&Law12, Li15}. This result, however, was obtained for non-interacting systems, and we extend the analysis to interacting systems.
For the first example (i.e. Luttinger liquid - MKP junction) we consider electron-electron interactions in the bulk of the wire. We find that perfect Andreev reflection fixed point is stable with respect to weak repulsive interactions in the lead. This result should be contrasted with the conventional Luttinger liquid - s-wave superconductor junction where weak repulsive interactions destabilize Andreev reflection fixed point and drive the system back to the normal reflection fixed point~\cite{Fidkowski'12}. The reason for such a difference is the relative sign change in Andreev boundary conditions indicating that Luttinger liquid - MKP junction is similar, in this sense, to Luttinger liquid coupled to spin-triplet p-wave superconductor rather than spin-singlet s-wave superconductor. We perform perturbative RG analysis near perfect normal reflection and perfect Andreev reflection fixed points and propose a phase diagram, see Fig.~\ref{fig:RGflow}.
Next we investigate effect of local repulsive interactions in the normal lead-quantum dot- TRI topological superconductor junction. We show that the system flows to a new fixed point which is characterized by a strong entanglement of a QD spin with a MKP. These correlations ultimately lead to the change of boundary conditions for lead electrons: from Kondo to perfect Andreev boundary conditions. Using a combination of a perturbative RG analysis and slave-boson mean-field theory we identify the ground-state of the system and calculate tunneling conductance through the junction, demonstrating that zero-temperature differential tunneling conductance is $4e^2/h$.
As we increase Majorana coupling $\lambda$, the width of the zero-bias peak exhibits a crossover from the Kondo temperature $T_K$ to $\Gamma \lambda^2/\epsilon^2$ in the Majorana-dominated regime.
We have also studied effect of quantum fluctuations near the slave-boson mean-field saddle point and demonstrated that the mean-field solution is well-defined (in the quasi-long range order sense) and thus can be used to calculate the spectrum in the QD as well as other observables.
Our work represents the first step in understanding the signatures of Majorana Kramers pairs (class DIII superconductor) when coupled to an interacting lead. As we have shown, the phase diagram in this case is much richer than in the case of a single Majorana mode (class D superconductor), see Ref.~\cite{Fidkowski'12}. Our perturbative approach does not capture the crossover regime between perfect normal reflection and perfect Andreev reflection. It would be interesting to understand the phase diagram for $1/4 <K_{\rho}\lesssim 1/3$ using numerical methods and check our conjectured phase diagram.
In this paper we also analyzed Gaussian fluctuations around the slave-boson mean-field solution, see Sec. \ref{sec:gaussianfluctuation}.
We confirmed the stability of the low-energy mean-field solution by calculating the correlation function of the slave bosons. It would be interesting to study other physical quantities numerically such as impurity spectral
function and magnetic susceptibility of the impurity.
{\it Note added}. While this manuscript was in preparation, we became aware of related independent work on this subject~\cite{Pikulin2016} which has some overlap with Sec.~\ref{sec:MKPLL}.
\section{Acknowledgements}
We are grateful to Dmitry Pikulin, Jian Li and Andrei Bernevig for stimulating discussions. YK was supported by the Samsung Scholarship. RL wishes to acknowledge the hospitality of the Aspen Center for Physics and the support under NSF Grant \#1066293.
EG, JP and KF acknowledges the support from the Danish Council for Independent Research $|$ Natural Sciences.
\newpage
|
\section{Introduction}
Sequential data can take many forms. Written text, video data, language, and many other forms of information are naturally sequential. Designing models for predicting sequential data, or otherwise extracting information from a sequence is an important problem in machine learning. Often, recurrent neural networks are used for this task. Unlike a non-recurrent network, a recurrent network's input at a given time-step consists of any new information along with the output of the network at the previous time-step. Since the network receives both new input as well as its previous output, it can be said to have a ``memory'', since its previous activations will affect the current output. Training a RNN model with gradient descent, or similar gradient-based optimization algorithms is subject to the usual problem of vanishing gradients \cite{bengio1994learning}. At each time step, the gradient diminishes and eventually disappears. Consequently, if the RNN needs information from more than a few time-steps ago to compute the correct output at the current time step, it will be incapable of making an accurate prediction. The model ``Long Short-Term Memory'', \cite{hochreiter1997long} greatly mitigated this problem. LSTM incorporates ``gates'', which are neurons that use a sigmoid activation, and are multiplied with the output of other neurons. Using gates, the LSTM can adaptively ignore certain inputs. LSTM also maintains a set of values that are protected by the gates and that do not get passed through an activation function.
In this work we develop a modification to LSTM that aims to make better use of the existing LSTM structure while using a small number of extra parameters. We claim that there are three issues with LSTM with forget gates. First, forget gates impose an exponential decay on the memory, which may not be appropriate in some cases. Second, the memory cells cannot communicate or exchange information without opening the input and output gates, which also control the flow of information outside the memory cells. Third, the hyperbolic tangent function is not ideal since LSTM memory values can grow large, but the the hyperbolic tangent has a very small gradient when its input value is large.
In our modified version, the forget gate is replaced with a functional layer that resides between the input and output gates. We call this modification LSTM with working memory, henceforth abbreviated LSTWM. LSTWM incorporates an extra layer that operates directly on the stored data. Rather than multiplying previous cell values by a forget gate, it uses a convex combination of the current memory cell value and the output of this extra layer whose input consists of previous memory cell values. In addition, we find that using a logarithm-based activation function improves performance with both the LSTM architecture and our modified variant. We evaluate this method on the Hutter challenge dataset \cite{hutter} as well as a task designed using the MNIST \cite{lecun2010mnist} handwritten digit dataset.
\section{Architecture and Training}
We begin with a review of LSTM. Since nearly all modern implementations use forget gates, we will call LSTM with forget gates standard LSTM or just LSTM. The term ``memory cell'' will be used to refer to the inner recurrence with a weight of one and ``LSTM cell'' to refer to an entire, individual LSTM unit.
\subsection{Background on LSTM}
We introduce notation to be used throughout the rest of the paper: $x_t$ is the input vector at time $t$, $y_t$ is the network layer output at time $t$, $\sigma=\frac{1.0}{1.0+e^{-x}}$ and $f$ is an activation function. The standard LSTM formulation includes an input gate, output gate, and usually a forget gate (introduced in \cite{gers2000learning}). The output of an LSTM unit is computed as follows:
\begin{align}
a &= f(W\cdot [x_t;y_{t-1}]+b) \\
g^{(i)} &= \sigma(W_{g_i}\cdot [x_t;y_{t-1}] + b_{g_i}) \\
g^{(o)}&= \sigma(W_{g_o}\cdot [x_t;y_{t-1}] + b_{g_o}) \\
g^{(f)} &= \sigma(W_{g_s}\cdot [x_t;y_{t-1}] + b_{g_s}) \\
c_t &= g^{(i)}\odot a + g^{(f)}\odot c_{t-1} \\
y_t &= g^{(o)}\odot f(c_t)
\end{align}
\begin{figure}[t!]
\centering
\includegraphics[width=.4\textwidth]{lstm_figure}
\caption{LSTM with forget gates. The orange circles are multiplicative gate units, the white circle is the addition operator. Double arrows represent the output of a neuron with multiple inputs, while single arrows represent the path of a single value.}
\label{fig:lstmfig}
\end{figure}
It should be noted that other variants of LSTM exist, such as LSTM with peepholes \cite{gers2003learning} and more recently Associative LSTM \cite{danihelka2016associative}. We chose LSTM with forget gates as it is a simple yet commonly used LSTM configuration.
The input and output gate values, $g^{(i)}$ and $g^{(o)}$, serve to regulate incoming and outgoing data, protecting the memory cell value. The forget gate can be used to clear the memory cell when its data is no longer needed. If the input/output gates are closed, and the forget gate is open, then there is no transformation on the memory cell values. This is an extreme example, since the gate cells will likely take different values at any given timestep and can never be fully open or closed. However, it illustrates the main point of LSTM, which is that it can store data for an arbitrary amount of time, in principle.
LSTM with forget gates does not share the same mathematical properties of the original LSTM design. The memory-cell value decays exponentially due to the forget gate. This is important, since the network has a finite memory capacity and the most recent information is often more relevant than the older information. It may, however, be useful to decay that information in a more intelligent manner. Another point is that the information in the memory cells cannot be used without releasing it into the outer recurrence. This exposes them to the downstream network. A memory cell must also, to some degree, accept information from the upstream network to perform computation on the recurrent information.
Another related architecture in this domain is the Gated Recurrent Unit, or GRU \cite{cho2014learning}, which uses a convex combination instead of an input gate. It does not use an output gate. This makes the architecture easier to implement, since there is only one set of recurrent values per layer. This also does not share the theoretical properties of the original LSTM, since the memory cell values are replaced by their input rather than summed with their input.
Empirical studies \cite{jozefowicz2015empirical} \cite{chung2014empirical} comparing the performance of LSTM and GRU are somewhat inconclusive. Despite this, we hypothesize that a potential weak-point of the GRU is that its cells cannot accumulate large values as evidence of the strong presence of a certain feature. Suppose a GRU sees the same feature three times in a row, and its input gate is open. The cell will (roughly speaking) be replaced three times, but there is comparatively little evidence that the GRU saw this feature more than once. On the other hand, the LSTM memory cell has been increased three times, and it can arguably retain this large-magnitude value for a longer period of time. For this reason, we feel that the LSTM is a stronger base upon which to build.
\begin{figure}[t!]
\centering
\includegraphics[width=.4\textwidth]{lstwm_figure}
\caption{LSTWM. The orange square represents a convex combination of the two inputs on the right.}
\label{fig:lstwmfig}
\end{figure}
\subsection{LSTWM}
For a regular LSTM to operate on the values stored in its memory cell and store them again, it must open both the output gate and the input gate. This releases the value to the downstream network, and new input from the upstream network must also, to some degree, be allowed to enter the cell. We propose that a network may make better use of its input and output gates if they weren't serving the two, not necessarily related, purposes of regulating recurrent data as well as regulating inputs and outputs to and from other parts of the network. Additionally, the activation function is applied to the memory cell values before they are multiplied by the output gate. Typically, the hyperbolic tangent function is used, and if the memory cell value already has a high magnitude, the errors entering the memory cell will be very small. We argue that a non-saturating activation function will confer a significant performance increase. The output of a LSTWM cell is computed in the following manner:
\begin{align}
a &= f(W\cdot [x_t;y_{t-1}]+b) \\
g^{(i)} &= \sigma(W_{g_i}\cdot [x_t;y_{t-1}] + b_{g_i}) \\
g^{(o)}&= \sigma(W_{g_o}\cdot [x_t;y_{t-1}] + b_{g_o}) \\
g^{(s)} &= \sigma(W_{g_s}\cdot [x_t;y_{t-1}] + b_{g_s}) \\
c^{(r_l)} &= \rho(c_{t-1},-1) \\
c^{(r_r)} &= \rho(c_{t-1},1) \\
i_t &= f (w_{v_1}\odot c_{t-1} + w_{v_2} \odot c^{(r_l)} + w_{v_3} \odot c^{(r_r)} + b_{v_1}) \\
r_t &= g^{(s)} \odot c_{t-1} + (1.0-g^{(s)}) \odot i_t \\
c_t &= g^{(i)}\odot a + r_t \\
y_t &= g^{(o)}\odot f(c_t)
\end{align}
Where $f$ is an activation function, either
\[ f(x) =
\begin{cases}
-ln(-x+1) & \text{if } x \text{< 0} \\
ln(x+1) & \text{otherwise}
\end{cases}
\]
or
\[ f(x) =
tanh(x)\\
\]
This logarithm-based activation function does not saturate and can better handle larger inputs than tanh. Fig. ~\ref{fig:activation} illustrates both functions. The recurrent cell values may grow quite large, causing the hyperbolic tangent function to quickly saturate and gradients to disappear. To obtain good performance from our design, we wanted to develop a non-saturating function that can still squash its input. Earlier works have used logarithm-based activations, and the function we used appears in \cite{isa2010suitable} and originally in \cite{bilski2000backpropagation}.
While the rectified linear unit \cite{glorot2011deep} (ReLU) is a common choice for non-recurrent architectures, they are usually not suitable for LSTM cells since they only have positive outputs (although they have been used with some success in LSTM and RNNs, see \cite{le2015simple} \cite{krueger2015regularizing}), and exploding feedback loops are more likely when the activation function does not apply some ``squashing effect''. However, tanh is less suited for large inputs due to the small values of its derivative outside the small interval around zero.
\begin{figure}[t!]
\centering
\includegraphics[width=.5\textwidth]{figure_1
\caption{Wikipedia Text Prediction Task. Two layer networks.}
\label{fig:text_training_2l}
\end{figure}
The forget gate output has been renamed $g^{(s)}$, and now controls the convex combination of the previous memory cell value $c_{t-1}$ with the output of the inner layer $i_t$. The weight vectors $w_{v_2}$ and $w_{v_3}$ only connect a neuron to its left and right neighbors. In software, we did not implement this as a matrix multiplication, but as an element-wise roll of the input which gets multiplied by weight vectors and then has bias vectors added. The function $\rho(x,y)$ is an element-wise roll where $x$ is the input vector and $y$ is the number of times to roll left or right. For example $\rho([1,2,3],1) = [3,1,2]$ and $\rho([1,2,3],-1) = [2,3,1]$. So the inter-neuron connections within the memory cells are sparse and local. The reason for choosing a sparse layer is that dense layers grow quadratically with respect to layer width. We wanted to compare equal-width networks, since a LSTM's memory capacity is determined by the number of individual memory cells. By setting near-zero weights and biases for the inner layer, the network can also achieve a traditional forget gate. Since this layer does not use many extra parameters, we can compare equal width networks. $w_{v_1}$, $w_{v_2}$, $w _{v_3}$ and $b_{v_1}$ are initialized with zeros, so the network starts out as a standard LSTM. In summary, an LSTWM network can modify its memory in a more complex manner without necessarily accepting new values or exposing its current values. Since the inner layer only uses the previous memory cell values, it can be computed in parallel with any downstream network and does not present a computation bottleneck if implemented in a parallel manner.
This architecture was adapted from a design in a previous version \cite{pulver2016lstm} of this work, which used normal forget gates and included an extra layer and extra gate after the memory cell update. We found that the previous four-gate architecture did not perform as well on tasks that required a precise memory. Likely, having three different memory operations at each timestep resulted in excessive changes in to the memory. Setting appropriate initial bias values helped the situation in some cases, however, we found better designs that did not require as much hand-tuning. Our first attempt at resolving the issue was removing the forget gate. Removing the forget gate from our earlier design did not yield good result by itself.
Figures ~\ref{fig:lstmfig} and ~\ref{fig:lstwmfig} illustrate LSTM and LSTWM, respectively. The $i$ subscript indicates that this is the $i^{th}$ unit in a layer. The white double arrows indicate an input that determines a gate value (as opposed to a value that gets multiplied by a gate).
\begin{figure}[t!]
\centering
\includegraphics[width=.5\textwidth]{figure_2
\caption{Wikipedia Text Prediction Task. Single layer networks.}
\label{fig:text_training_1l}
\end{figure}
\subsubsection{Training, Regularization and Other Details}
Given that the memory cells are not necessarily decaying in an exponential manner at each timestep, it is important to control their magnitude. Rather than just using a regularization term on the network weights themselves, we also regularize the overall magnitude of the memory-cells at each timestep.
When training an ANN of any type, a regularization term is usually included in the cost function. Often, the L2-norms of the weight matrices are added together and multiplied by a very small constant. This keeps the weight magnitudes in check. For our architecture to train quickly, it is important to use an additional regularization term. Given a batch of training data, we take the squared mean of the absolute value plus the mean absolute value of the memory cell magnitudes at each timestep for every element in the batch. In other words, $\eta\cdot(mean(|cells|)^2 + mean(|cells|))$ for some small constant $\eta$. We found that using $\eta\approx 10^{-2}$ or $\eta\approx 10^{-3}$ worked well. Similar regularization terms are discussed in \cite{krueger2015regularizing}, although they differ in that they are applied to the change in memory cell values from one timestep to the next, rather than their magnitude. Using direct connections between the inner cells can amplify gradients and produce quickly exploding values. By adding this regularization term, we penalize weight configurations that encourage uncontrollable memory cell values. Since the cells values get squashed by the activation function before being multiplied by the output gate, regularizing these values shapes the optimization landscape in a way that encourages faster learning. This is also true to some extent for regular LSTM, which benefits from this type of regularization as well. We also applied this regularization function to the weights themselves. LSTWM can learn effectively without extra regularization, but it will learn slowly in some cases.
Note that we use the square-of-abs-mean not the mean-squared. Using the square-of-abs-mean allows some values to grow large, and only encourages most of the values to be small. This is important, especially for the memory cells, because the ability of a neural network to generalize depends on its ability to abstract away details and ignore small changes in input. Indeed, this is the entire purpose of using sigmoid-shaped squashing functions. If the goal were to keep every single memory cell within tanh's ``gradient-zone'', one could use the hyperbolic cosine as a regularization function since it grows very large outside this range.
We used Python and Theano \cite{2016arXiv160502688short} to implement our network and experiments.
\begin{figure}[t!]
\centering
\includegraphics[width=.5\textwidth]{transfer_functions}
\caption{Comparison of tanh and log-based activation function}
\label{fig:activation}
\end{figure}
\section{Experiments}
We have two experimental tasks to test our network on: text prediction and a combination digit-recognition and addition task. The networks were trained using ADAM \cite{kingma2014adam}, an optimization algorithm based on gradient descent, with the following settings $\alpha=.001$ $\beta_1=.9$ $\beta_2=.999$.
\section{Text Prediction}
\begin{figure}[t]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Architecture & BPC on test set after training \\
\hline
LSTM-256-tanh & 1.893 \\
\hline
LSTWM-256-tanh & 1.892\\
\hline
LSTM-256-log & 1.880 \\
\hline
LSTWM-256-log & 1.880 \\
\hline
LSTM-256-256-tanh & 1.742 \\
\hline
LSTWM-256-256-tanh & 1.733 \\
\hline
LSTM-256-256-log & 1.730 \\
\hline
LSTWM-256-256-log & 1.725 \\
\hline
\end{tabular}
\caption{Performance on text prediction task.}
\label{fig:text_results}
\end{figure}
Like many other works e.g. \cite{danihelka2016associative}, we use the hutter challenge \cite{hutter} dataset as a performance test. This is a dataset of text and XML from Wikipedia. The objective is to predict the next character in the sequence. (Note that we only use the dataset itself, and this benchmark is unrelated to the Hutter compression challenge) The first 95\% of the data was used as a training set and the last 5\% for testing. Error is measured in bits-per-character, BPC, which is identical to cross entropy error, except that the base-2 logarithm is used instead of the natural log. We used a batch size of 32 and no gradient clipping. To reduce the amount of training time needed, we used length-200 sequences for the first epoch, and then used length-2000 sequences for the remaining five epochs.
Figures ~\ref{fig:text_training_2l} and ~\ref{fig:text_training_1l} show a running average of the training error and ~\ref{fig:text_results} shows the BPC on the test set after training. The results are close for this particular task, with LSTWM taking a slight advantage. Notice that for this test, the logarithm based activation does carry some benefit, and the best performing network was indeed LSTWM with the logarithmic activation.
\subsection{Training Information}
Given the popularity of this task as a benchmark, a quick training phase is desirable. Input vectors are traditionally given in a one-hot format. However, the network layers can be quite wide, and each cell has many connections. We found that using a slightly larger nonzero value in the vectors resulted in quicker training. Instead of using a value of 1.0 for the single non-zero element in the input vectors, we used $log(n)+1.0$ where $n$ is the number of input symbols. In this case, $n=205$, since there are 205 distinct characters in this dataset. On tasks where there are many symbols, the mean magnitude of the elements of the input vector is not as small, which accelerated training in our experiments.
Another method we used to speed up training was using a pre-train epoch with shorter sequences. The shorter sequences mean that there are more iterations in this epoch and the iterations are performed more quickly. This rapid pre-training phase moves the parameters to a better starting point more quickly. This method could also be used in any task where the input is an arbitrary-length sequence. Longer sequences result in better quality updates, but they take proportionally longer to compute. We also added a very small amount of Gaussian noise to the inputs during the pre-train epoch, but not during the main training phase. Fast and noisy updates work well early in the training process and reduce the amount of time necessary to train networks.
\section{Digit recognition and addition combined task}
\begin{figure}[t]
\centering
\includegraphics[width=.3\textwidth]{input1}
\\
\vspace{.1cm}
\includegraphics[width=.3\textwidth]{input2}
\\
\vspace{.1cm}
\includegraphics[width=.3\textwidth]{input3}
\caption{Example input sequences for combination digit task}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.4\textwidth]{separated}
\caption{Inputs colored and separated into columns to show individual input columns}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.5\textwidth]{4digit}
\caption{Digit Combo task training error.}
\label{fig:combo_training}
\end{figure}
\begin{figure}[t!]
\centering
\begin{tabular}{|c|c|}
\hline
Architecture & \pbox{20cm}{ \# of correct sum outputs \\ \centering{(out of 10000)} } \\
\hline
LSTM-128-128-tanh & 8864 \\
\hline
LSTWM-128-128-tanh & 8820 \\
\hline
LSTM-128-128-log & 8972 \\
\hline
LSTWM-128-128-log & 9015 \\
\hline
\end{tabular}
\caption{Performance on digit combo task.}
\label{fig:combo_results}
\end{figure}
This task is a combination of digit recognition and addition. Other works (e.g. \cite{ha2016hypernetworks}, \cite{kalchbrenner2015grid}) use the MNIST dataset as a performance test for artificial neural networks. Addition is also a common task for evaluating the viability of new RNN designs, and usually the numbers to be added are explicitly encoded by a single input vector at each timestep. The difference with this task is that we horizontally concatenate a number of images from MNIST and train on the sum of those digits. This large concatenated image is presented to the network in column vectors. After the network has seen the entire image, it should output the sum of the digits in the input. No error is propagated until the end of the image. This makes the task a useful benchmark for memory, generalization ability, and vision. The training targets are placeholders until the end, and then two (more if needed) size-10 vectors corresponding to the two digits of the sum. A similar but non-sequential task appears in \cite{ba2014multiple}.
For this task, the error coming from the end of the sequence must be back-propagated all the way to the first digit image. If this does not occur, the model will not correctly learn. Unlike text prediction, this requires a robust memory and the ability to back-propagate the error an arbitrary number of time-steps. We suspect that text prediction is a task where memory values are often replaced. For text-prediction, at the end of training, the forget gate biases are often < 0, which indicates low forget-gate activity, i.e. it tends to forget things quickly. This task was developed as a way to test memory functionality, since text-recognition does not stress this capability.
Another interesting property of this task is that it demonstrates the ability of an architecture to form a generalized algorithm. Increasing the length by concatenating more digit images will remove non-algorithmic minimums from the search space. In other words, the network will need to learn an addition algorithm to correctly process a longer sequence. Simply ``memorizing'' the correct outputs will not work when the network does not have enough parameters, to work on longer sequences, it must learn an addition algorithm. This can also be applied without the strict memory requirements of back-propagating to the first digit. In this case, one would simply train to output the running sum at frequent intervals rather than just at the end of the sequence. In the boundary case where the number of digit images is one, it becomes a simple digit-recognition task.
We trained in a similar manner to the text-prediction experiment. Each training input consists of randomly selected digits along with their sum as a target. The same ADAM parameters were used as in the text-prediction task. We trained on sequences with four concatenated digit images. Therefore, each input sequence was 112 input vectors with three more placeholders at the end. As a preprocessing step, the digit images were added with Gaussian (scaled down $\times 10^{-5}$) noise, blurred (3x3 kernel, Gaussian), mean-subtracted and squashed with the log-activation function in that order. Better performance can be obtained by omitting the noise and blur, however it was difficult to obtain consistent results without some smoothing of the inputs.
The LSTWM architecture learned more quickly, and achieved the best performance, although only using the logarithm-based activation. Figures ~\ref{fig:combo_training} and ~\ref{fig:combo_results} show the training error curve and peak test set results respectively.
\subsection{Plain Digit Recognition}
We also tested the case for the task above where there is only one digit image, for a variety of different network sizes. Since the layers are narrow, we decided to slightly increase the width of LSTM so that the LSTM networks have more parameters. Since this is a much simpler task, it did not require as much training.
\begin{figure}[t!]
\centering
\includegraphics[width=.5\textwidth]{digi_single}
\caption{Plain Digit Recognition}
\label{fig:digit_training}
\end{figure}
\begin{figure}[t!]
\centering
\begin{tabular}{|c|c|}
\hline
Architecture & \pbox{20cm}{ Best test-set performance \\ \centering{(out of 10000)} } \\
\hline
LSTM-32-33-tanh & 9600 \\
\hline
LSTWM-32-32-tanh & 9620 \\
\hline
LSTM-32-33-log & 9657 \\
\hline
LSTWM-32-32-log & 9655 \\
\hline
LSTM-32-32-33-33-tanh & 9676 \\
\hline
LSTWM-32-32-32-32-tanh & 9690 \\
\hline
LSTM-32-32-33-33-log & 9690 \\
\hline
LSTWM-32-32-32-32-log & 9712 \\
\hline
\end{tabular}
\caption{Performance on digit recognition. (out of 10,000 test sequences)}
\label{fig:digit_results}
\end{figure}
\subsection{Results}
For text-prediction, the performance is roughly equal, although LSTWM with the logarithm activation slightly outperforms the other networks. Notably, the logarithmic activation function increases performance consistently on this task. For the digit-combo task, we observed better performance with LSTWM, and likewise on the simple digit recognition task. The number of extra parameters used is linear, not quadratic, with respect to layer width, so we consider it a good improvement given the small size increase. We claim that using the inner layer, along with using a logarithm-based activation will offer a modest but significant performance benefit. Another notable feature of LSTWM is that the training error decreases more quickly compared to standard LSTM.
\section{Conclusion}
We discussed several interesting properties of LSTM and introduced a modified LSTM architecture that outperforms LSTM in several cases using few additional parameters. We observe that a logarithm-based activation function works well in LSTM and our modified variant. Finally, we presented a benchmark for recurrent networks based on digit-recognition.
\printbibliography
\end{document}
|
\section{Introduction}
Since the first experimental verification of the Bose-Einstein condensation (BEC) \cite{ak,anderson,wwcch}, occurred more then seven decades after its theoretical prediction \cite{bose,eins}, a great effort in the theoretical and experimental viewpoint has been made in the study of this quantum many body physical phenomenon \cite{Hulet,Dalfovo,l01,Bagnato,cw,Piza,Bloch,Carusotto}. Looking for new applications an atom-molecule Bose-Einstein condensates was experimentally produced applying magnetic field near a Feshbach resonance in an atomic BEC \cite{Donley,Zoller}. The magnetic field pulses produces a coherent interconversion dynamics of atoms and molecules. Another techniques used to produce an atom-molecule Bose-Einstein condensate are the photoassociation \cite{Souza} and the two-photons Raman transition \cite{Wynar}. These experimental opened the possibility of the cold and ultracold quantum chemistry \cite{Wynar,Richter,Hutson,Regal,Ospelkaus,Heinzen,Goral,Carr,Kevin,Makrides,Dulieu,Pillet,Amelink,Pillet2,Durr,GSantos09 }. I am considering here an atom-molecule Bose-Einstein condensate model, used to study this coherent interconversion dynamics of atoms and homonuclear diatomic molecules \cite{GSantos06a,GSantos10}. This model is integrable in the sense that it can be solved by the quantum inverse scattering method (QISM) \cite{GSantos11a,GSantos11b,jlletter,jlreview,jinter,ATonel,jlsigma,Angela2,Angela,GSantos06b,GSantosBVS-PLB,jdensity,jbroots,GSantos13}. In this work I will discuss the symmetries and the interconversion currents algebra of the model and use they to study the quantum dynamics of the currents. This method was applied in the study of the Josephson tunnelling phenomenon in a two-site Bose-Hubbard model to get the quantum dynamics of the tunnelling currents of atoms as a function of the parameters of the Hamiltonian \cite{GSantosCA1,GSantosCA2}. We can also apply this method to a model recently used to study the tunnelling between two atom-molecule BECs coupled by Josephson tunnelling \cite{Motohashi1,Motohashi2,Motohashi3,Motohashi4,Dukelsky}. It is also worth to note that we can considerer the atom-molecule BEC as a two mode system, an atomic mode and the another one molecular. This bipartite system is entangled by the interconversion dynamic of atoms and molecules and we can study the quantum phase transition of the system using tools of the quantum information \cite{GSantos10,Angela2}.
The dynamics of interconversion of atoms and molecules as an open system, with particles losses was studied in \cite{Cui}. We are considering a closed system with the total number of atoms conserved\cite{GSantosCA1}. The dynamic of interconversion in an atom-molecule Bose-Einstein condensate is described by the Hamiltonian
\begin{eqnarray}
\hat{H} & = & U_a\hat{N}_a^2 + U_b\hat{N}_b^2 + U_{ab}\hat{N}_a\hat{N}_b + \mu_a\hat{N}_a +\mu_b\hat{N}_b - \Omega(\hat{a}^{\dag}\hat{a}^{\dag}\hat{b} + \hat{b}^{\dag}\hat{a}\hat{a}),
\label{ham}
\end{eqnarray}
\noindent where $\hat{a}^{\dagger}$ and $\hat{a}$ are the creation and annihilation boson operators of an unbound atom while $\hat{b}^{\dagger}$ and $\hat{b}$ are the creation and annihilation boson operators of a molecule. The number boson operators $\hat{N}_a=\hat{a}^{\dagger}\hat{a}$ and $\hat{N}_b=\hat{b}^{\dagger}\hat{b}$ are the respective number of unbound atoms and number of molecules operators. These bosons operators satisfies the canonical commutation relations
\begin{equation}
[\hat{a},\hat{a}^\dagger] = [\hat{b},\hat{b}^\dagger] = \hat{I},
\end{equation}
\noindent with the condition that the unbound atom and molecule boson operators commutes
\begin{equation}
[\hat{a}^\dagger,\hat{b}] = [\hat{a},\hat{b}] = 0,
\end{equation}
\noindent and
\begin{equation}
[\hat{N}_a,\hat{a}] = - \hat{a}, \qquad [\hat{N}_a,\hat{a}^\dagger] = +\hat{a}^\dagger,
\end{equation}
\begin{equation}
[\hat{N}_b,\hat{b}] = - \hat{b}, \qquad [\hat{N}_b,\hat{b}^\dagger] = +\hat{b}^\dagger,
\end{equation}
\noindent where $\hat{I}$ is the identity operator.
The parameter $U_{a}$ is the atom-atom interaction, $U_{ab}$ is the atom-molecule interaction and $U_{b}$ is the molecule-molecule interaction. The strength of these interactions are proportional to the scattering length and usually is enough to consider only the low energy $s$-wave scattering length.
The parameter $\mu_a$ is the external potential for the unbound atoms and $\mu_b$ is the external potential for the molecules. The parameter $\Omega$ is the amplitude for interconversion of atoms and molecules.
In the limit $U_{a}=U_{ab}=U_{b}=0$, the Hamiltonian (\ref{ham}) has been studied using a
variety of methods \cite{Vardi,Hines,Clare,Shen,Graefe}. However in the experimental context, the $s$-wave scattering interactions play a significant role. It will be seen below that for the general model (\ref{ham}) the inclusion of these scattering terms has a non-trivial consequence. The $s$-wave scattering length for the atom-atom interaction $U_{a}$ can be determined precisely by photoassociation spectroscopy of one and two photons \cite{Amelink,AnnualReview}. We mention that generally the values for the molecule-molecule $U_b$ and the atom-molecule $U_{ab}$ interactions are unknown \cite{Wynar}, although some estimates exist in the case of $^{85}Rb$ \cite{Cusack}. For an up to dated discussion of theoretical and experimental data see \cite{AnnualReview}.
The paper is organized as follows. In the section 2, I present the symmetries of the Hamiltonian. In the section 3, I present the interconversion currents algebra and the Casimir operators. In the section 4, I calculate the quantum dynamics of the currents. In the section 5, I summarize the results.
\section{Symmetries}
The Hamiltonian (\ref{ham}) is invariant under the global $U(1)$ gauge transformation $\hat{a} \rightarrow e^{i\alpha}\hat{a}$, $\hat{b} \rightarrow e^{2i\alpha}\hat{b}$, where $\alpha$ is an arbitrary $c$-number and $\hat{a}^{\dagger}\rightarrow e^{-i\alpha}\hat{a}^{\dagger}$, $\hat{b}^{\dagger}\rightarrow e^{-2i\alpha}\hat{b}^{\dagger}$. The global $U(1)$ gauge invariance is associated with the conservation of the total number of atoms (unbound atoms plus bound atoms) $\hat{N} = \hat{N}_a + 2\hat{N}_b$. The Hamiltonian (\ref{ham}) is not invariant under the exchange of atoms and molecules because the interconversion dynamics is not symmetric under this exchange.
The Hamiltonian (\ref{ham}) is invariant under the $\mathbb{Z}_2$ mirror transformation of the unbound atom operators $\hat{a} \rightarrow -\hat{a}, \hat{a}^{\dagger} \rightarrow -\hat{a}^{\dagger}$, but it is not invariant under the $\mathbb{Z}_2$ mirror transformation of the molecule operators $\hat{b} \rightarrow -\hat{b}, \hat{b}^{\dagger} \rightarrow -\hat{b}^{\dagger}$. The $\mathbb{Z}_2$ symmetry is associated with the parity of the wave function by the relation
\begin{equation}
\hat{P}_a \; |\Psi\rangle = (-1)^{N_a} |\Psi\rangle,
\end{equation}
\noindent where $\hat{P}_a$ is the parity operator acting only on $\hat{a}$ and $\hat{a}^{\dagger}$. Because this symmetry we have $[\hat{H},\hat{P}_a] = 0$. When we consider the parity operator $\hat{P}_b$ acting only on the molecule operators $\hat{b}$ and $\hat{b}^{\dagger}$, the parity of the wave function is not well defined and we have $[\hat{H},\hat{P}_b] \neq 0$.
If $N_a$ is even, $N$ is even and the dimension of the state space is $D=N/2 +1$. For $N_a$ even we can write a general state of the Hamiltonian (\ref{ham}) as
\begin{equation}
|\Psi\rangle = \sum_{n_b = 0}^{\frac{N}{2}} \; C_{N-2n_b,n_b} \frac{(\hat{a}^{\dagger})^{N-2n_b}}{\sqrt{(N-2n_b)!}} \frac{(\hat{b}^{\dagger})^{n_b}}{\sqrt{n_b!}} |0,0\rangle,
\end{equation}
\noindent and
\begin{equation}
\hat{P}_a \; |\Psi\rangle = |\Psi\rangle.
\end{equation}
\noindent If $N_a$ is odd, $N$ is odd and the dimension of the state space is $D=(N+1)/2$. For $N_a$ odd we can write a general state of the Hamiltonian (\ref{ham}) as
\begin{equation}
|\Psi\rangle = \sum_{n_b = 0}^{\frac{N - 1}{2}} \; C_{N-2n_b,n_b} \frac{(\hat{a}^{\dagger})^{N-2n_b}}{\sqrt{(N-2n_b)!}} \frac{(\hat{b}^{\dagger})^{n_b}}{\sqrt{n_b!}} |0,0\rangle,
\end{equation}
\noindent and
\begin{equation}
\hat{P}_a \; |\Psi\rangle = - |\Psi\rangle.
\end{equation}
The symmetries of the Hamiltonian (\ref{ham}) imply degeneracy. For the conservancy of $\hat{N}$ we have that all wave function of the Hamiltonian (\ref{ham}) are degenerated eigenfunctions of $\hat{N}$ with the same eigenvalue $N$. For the parity operator $\hat{P}_a$ all wave function of the Hamiltonian (\ref{ham}) are even or odd depending if $N$ is even or odd. All wave functions are degenerated eigenfunctions of $\hat{P}_a$ with the same eigenvalue $\lambda = +1$ if $N$ is even or they are degenerated eigenfunctions of $\hat{P}_a$ with the same eigenvalue $\lambda = -1$ if $N$ is odd.
\section{Interconversion Currents Algebra}
The quantum dynamics of any operator $\hat{O}$ in the Heisenberg picture is determined by the Heisenberg equation of motion
\begin{equation}
\frac{d\hat{O}}{dt} = \frac{i}{\hbar}[\hat{H},\hat{O}].
\label{dOdtH}
\end{equation}
The boson operator total number of atoms, $\hat{N} = \hat{N}_a + 2\hat{N}_b$, is a conserved quantity, $[\hat{H},\hat{N}]=0$, and it is commutable compatible operator (CCO) with the number of unbound atoms and number of molecules boson operators, $[\hat{N},\hat{N}_a]=[\hat{N},\hat{N}_b]= [\hat{N}_a,\hat{N}_b]=0$. The boson operators number of unbound atoms and the number of molecules don't commute with the Hamiltonian and their time evolution is dictated by the interconversion current operator,
\begin{equation}
\hat{\mathcal{J}} = \frac{1}{4i} (\hat{a}^\dagger \hat{a}^\dagger\hat{b} - \hat{b}^\dagger\hat{a}\hat{a}),
\label{cj}
\end{equation}
\noindent in coherent opposite phases because of the conservancy of $\hat{N}$, with
\begin{equation}
[\hat{H},\hat{N}_a]= + 8 i\Omega\hat{\mathcal{J}}, ~~~ [\hat{H},\hat{N}_b]= -4i\Omega\hat{\mathcal{J}},
\end{equation}
\noindent and
\begin{equation}
\frac{d\hat{N}_a}{dt} = - 8\frac{\Omega}{\hbar}\hat{\mathcal{J}}, \label{CDN1}
\end{equation}
\begin{equation}
\frac{d\hat{N}_b}{dt} = + 4 \frac{\Omega}{\hbar}\hat{\mathcal{J}}. \label{CDN2}
\end{equation}
Integrating the Eqs. (\ref{CDN1}) and (\ref{CDN2}) we get
\begin{eqnarray}
\hat{N}_a(t) &=& \hat{N}_a(0) - 8\frac{\Omega}{\hbar}\int_0^t\hat{\mathcal{J}}(\tau) \; d\tau, \label{SCDN1} \\
\hat{N}_b(t) &=& \hat{N}_b(0) + 4\frac{\Omega}{\hbar}\int_0^t\hat{\mathcal{J}}(\tau) \; d\tau. \label{SCDN2}
\end{eqnarray}
The operator $\hat{\mathcal{J}}$ together with the imbalance current operator $\hat{\mathcal{I}}$,
\begin{equation}
\hat{\mathcal{I}} = \frac{1}{4}(\hat{N}_a - 2\hat{N}_b),
\label{ci}
\end{equation}
\noindent and the coherent correlation interconversion current operator $\hat{\mathcal{T}}$,
\begin{equation}
\hat{\mathcal{T}} = \frac{1}{4}(\hat{a}^\dagger\hat{a}^\dagger\hat{b} + \hat{b}^\dagger\hat{a}\hat{a}),
\label{ct}
\end{equation}
\noindent generates the currents algebra
\begin{equation}
[\hat{\mathcal{T}},\hat{\mathcal{J}}]= +if(\hat{\mathcal{I}},\hat{N}), ~~~ [\hat{\mathcal{T}},\hat{\mathcal{I}}]= -i\hat{\mathcal{J}}, ~~~ [\hat{\mathcal{J}},\hat{\mathcal{I}}]= +i\hat{\mathcal{T}} \nonumber,
\label{calg}
\end{equation}
\noindent with
\begin{equation}
f(\hat{{\cal I}},\hat{N})=\frac{3}{2}\hat{{\cal I}}^{2}+\frac{1}{4}\hat{N}\hat{{\cal I}}-\hat{\mathcal{N}},
\label{calgf}
\end{equation}
\noindent and
\begin{equation}
\hat{\mathcal{N}}=\frac{\hat{N}}{8}\left(\frac{\hat{N}}{4}+1\right).
\end{equation}
\noindent With the identification $\hat{L}_x \equiv \hat{\mathcal{T}}$, $\hat{L}_y \equiv \hat{\mathcal{J}}$, and $\hat{L}_z \equiv \hat{\mathcal{I}}$ we can write (\ref{calg}) as the deformed momentum angular algebra
\begin{equation}
[\hat{L}_k,\hat{L}_l] = i\varepsilon_{klx}\hat{L}_x + i\varepsilon_{kly}\hat{L}_y + i\varepsilon_{klz} f(\hat{L}_z,\hat{N}),
\end{equation}
\noindent where $\varepsilon_{klm}$ is the antisymmetric Levi-Civita tensor, with $k,l=x,y,z$ and $\varepsilon_{xyz} = +1$.
We have two Casimir operators for that current algebra using the deformed momentum angular realization. One of them is the total number of atoms $\hat{N}$, related to the global $U(1)$ gauge symmetry, $\hat{\mathcal{\mathfrak{C}}}_{1} = \hat{N}$, and the another one is related to the deformed momentum angular algebra with the broken $O(3)$ symmetry
\begin{eqnarray}
\hat{\mathcal{\mathfrak{C}}}_{2} &=& \hat{L}_{x}^{2}+\hat{L}_{y}^{2}+\hat{L}_{z}^{3}+\frac{\hat{N}}{4}\hat{L}_{z}^{2}+\frac{1}{2}\left(1-\frac{\hat{N}}{2} -\frac{\hat{N}^2}{8}\right)\hat{L}_{z}. \nonumber\\ \label{CQ2}
\end{eqnarray}
\noindent We can show that $\hat{\mathcal{\mathfrak{C}}}_{2}$ is just a function of $\hat{\mathcal{\mathfrak{C}}}_{1}$
\begin{equation}
\hat{\mathcal{\mathfrak{C}}}_{2} = \frac{\hat{\mathcal{\mathfrak{C}}}^2_{1}}{16}\left(\frac{\hat{\mathcal{\mathfrak{C}}}_{1}}{4} + 1\right).
\end{equation}
\noindent The Casimir operators $\hat{\mathcal{\mathfrak{C}}}_{1}$ and $\hat{\mathcal{\mathfrak{C}}}_{2}$, the boson number of unbound atoms $\hat{N}_a$, the boson number of molecules $\hat{N}_b$, and the imbalance current operator, $\hat{\mathcal{I}}$, are CCO and so they have the same set of eigenfunctions and can simultaneous have well defined values
\begin{eqnarray}
\hat{\mathcal{\mathfrak{C}}}_{2} |n_a,n_b\rangle &=& \frac{N^2}{16}\left(\frac{N}{4} + 1\right) |n_a,n_b\rangle, \\
\hat{\mathcal{I}} |n_a,n_b\rangle &=& \frac{1}{4} \left(n_a - 2 n_b\right) |n_a,n_b\rangle.
\end{eqnarray}
We also can use the realization of a deformed $SU(2)$ algebra \cite{Graefe,Korsch,Bonatsos,Watanabe}
\begin{equation}
\hat{\mathcal{L}}_{\pm} = \hat{L}_{x} \pm i\hat{L}_{y}, \qquad
\hat{\mathcal{L}}_{z} = \hat{L}_{z},
\end{equation}
\noindent with the commutation relations
\begin{equation}
[\hat{\mathcal{L}}_{z}, \hat{\mathcal{L}}_{\pm}] = \pm \hat{\mathcal{L}}_{+}, \qquad
[\hat{\mathcal{L}}_{+}, \hat{\mathcal{L}}_{-}] = 2 f(\hat{\mathcal{L}}_{z},\hat{N}),\nonumber
\end{equation}
\noindent that we can write as
\begin{eqnarray}
[\hat{\mathcal{L}}_{k},\hat{\mathcal{L}}_{l}] &=& \varepsilon_{kl-}\hat{\mathcal{L}}_{+} + \varepsilon_{kl+}\hat{\mathcal{L}}_{-} + 2 \varepsilon_{zkl}f(\hat{\mathcal{L}}_{z},\hat{N}),
\end{eqnarray}
\noindent with $k,l=z,+,-$ and $\varepsilon_{z+-} = +1$.
This deformed $SU(2)$ algebra has three Casimr operators, $\hat{\mathcal{\mathfrak{C}}}_{1}$,
\begin{equation}
\hat{\mathcal{\mathfrak{C}}}_{3} = \hat{\mathcal{L}}_{+}\hat{\mathcal{L}}_{-} + \hat{\mathcal{L}}_{z}^{3} + \frac{1}{2}\left(\frac{\hat{N}}{2}-3\right)\hat{\mathcal{L}}_{z}^{2} + \frac{1}{2}\left[1-\hat{N}\left(\frac{\hat{N}}{8}+1\right)\right]\hat{\mathcal{L}}_{z}
\label{CQ3}
\end{equation}
\noindent and
\begin{equation}
\hat{\mathcal{\mathfrak{C}}}_{4} = \hat{\mathcal{L}}_{-}\hat{\mathcal{L}}_{+} + \hat{\mathcal{L}}_{z}^{3} + \frac{1}{2}\left(\frac{\hat{N}}{2}+3\right)\hat{\mathcal{L}}_{z}^{2} + \frac{1}{2}\left(1-\frac{\hat{N}^{2}}{8}\right)\hat{\mathcal{L}}_{z}.
\label{CQ4}
\end{equation}
That we can write, using the $\hat{\mathcal{\mathfrak{C}}}_{1}$ Casimir operator, as
\begin{equation}
\hat{\mathcal{\mathfrak{C}}}_{3}=\frac{1}{64}(\hat{\mathcal{\mathfrak{C}}}_{1}^{3}+2\hat{\mathcal{\mathfrak{C}}}_{1}^{2}-8\hat{\mathcal{\mathfrak{C}}}_{1}),
\end{equation}
\begin{equation}
\hat{\mathcal{\mathfrak{C}}}_{4}=\frac{1}{64}(\hat{\mathcal{\mathfrak{C}}}_{1}^{3}+6\hat{\mathcal{\mathfrak{C}}}_{1}^{2}+8\hat{\mathcal{\mathfrak{C}}}_{1}).
\end{equation}
They are related by the equation,
\begin{eqnarray}
\hat{\mathcal{\mathfrak{C}}}_{4} & = & \hat{\mathcal{\mathfrak{C}}}_{3}+\frac{\hat{\mathcal{\mathfrak{C}}}_{1}}{4}\left(\frac{\hat{\mathcal{\mathfrak{C}}}_{1}}{4}+1\right).
\end{eqnarray}
Using the identities
\begin{equation}
\hat{\mathcal{L}}_{+}\hat{\mathcal{L}}_{-} = \hat{\mathcal{L}}_{x}^{2} + \hat{\mathcal{L}}_{y}^{2} + f(\hat{\mathcal{L}}_{z},\hat{N}), \qquad
\hat{\mathcal{L}}_{-}\hat{\mathcal{L}}_{+} = \hat{\mathcal{L}}_{x}^{2} + \hat{\mathcal{L}}_{y}^{2} - f(\hat{\mathcal{L}}_{z},\hat{N}), \label{ident1}
\end{equation}
\noindent we can write the Cassimir operators (\ref{CQ3}) and (\ref{CQ4}) as
\begin{eqnarray}
\hat{\mathcal{\mathfrak{C}}}_{3} & = & \hat{\mathcal{L}}_{x}^{2}+\hat{\mathcal{L}}_{y}^{2}+\hat{\mathcal{L}}_{z}^{3}+ \frac{\hat{N}}{4}\hat{\mathcal{L}}_{z}^{2}+\frac{1}{2}\left(1-\frac{\hat{N}}{2}-\frac{\hat{N}^{2}}{8}\right)\hat{\mathcal{L}}_{z}-\hat{\mathcal{N}},
\end{eqnarray}
\begin{eqnarray}
\hat{\mathcal{\mathfrak{C}}}_{4} & = & \hat{\mathcal{L}}_{x}^{2}+\hat{\mathcal{L}}_{y}^{2}+\hat{\mathcal{L}}_{z}^{3}+\frac{\hat{N}}{4}\hat{\mathcal{L}}_{z}^{2}+\frac{1}{2}\left(1-\frac{\hat{N}}{2}-\frac{\hat{N}^{2}}{8}\right)\hat{\mathcal{L}}_{z}+\hat{\mathcal{N}}.
\end{eqnarray}
\noindent It is easy to show that $\hat{\mathcal{\mathfrak{C}}}_{4} - \hat{\mathcal{N}} = \hat{\mathcal{\mathfrak{C}}}_{3} + \hat{\mathcal{N}} = \hat{\mathcal{\mathfrak{C}}}_{2}$ and $\hat{\mathcal{\mathfrak{C}}}_{4} - \hat{\mathcal{\mathfrak{C}}}_{3} = 2\hat{\mathcal{N}}$. Therefore, if we take the average value as in \cite{Graefe,Korsch,Watanabe} we see that the Casimir operators $\hat{\mathcal{\mathfrak{C}}}_{2}$, $\hat{\mathcal{\mathfrak{C}}}_{3} + \hat{\mathcal{N}}$ and $\hat{\mathcal{\mathfrak{C}}}_{4} - \hat{\mathcal{N}}$ describe the same surfaces. Some of these Casimir surfaces has been denoted as Kummer shapes \cite{Graefe,Korsch,Holm1,Holm2,Kummer1,Kummer2,Kummer3,Kummer4}. We plot these surfaces in the Fig. (\ref{CSF}) using the following parametrization
\begin{eqnarray}
X(u,v) &=& f(v)\; cos\;u, \\
Y(u,v) &=& f(v)\; sin\;u, \\
Z(u,v) &=& b\;v,
\end{eqnarray}
\noindent where
\begin{equation}
f(v) = \sqrt{C_2 - v^3 - \frac{N}{4}v^2 - \frac{1}{2}\left( 1 - \frac{N}{2} - \frac{N^2}{8} \right)v},
\end{equation}
\noindent with $ C_2 = \frac{N^2}{16}\left(\frac{N}{4} +1\right)$ the eigenvalue of the Casimir operator $\hat{\mathcal{\mathfrak{C}}}_{2}$, $0\leq u \leq 2\pi$ and $-\frac{N}{4} \leq v \leq \frac{N}{4}$. The parameter $b$ changes the scale in the $Z$-axis direction. The radii in the boundaries of the $Z$-axis direction are $f(-\frac{N}{4}) = \sqrt{\frac{N}{8}}$ and $f(\frac{N}{4}) = \sqrt{\frac{N}{8}\left(N - 1\right)}$. If we extend the physical limit of $v$ we get the surfaces in Fig. (\ref{CSF2}).
\begin{figure}
\begin{center}
\begin{tabular}{cc}
$(a)$ & $(b)$ \\
\includegraphics[scale=0.5]{Casimir-ManifoldN10b10gbw.eps} & \includegraphics[scale=0.5]{Casimir-ManifoldN100b10gbw.eps} \\
$(c)$ & $(d)$ \\
\includegraphics[scale=0.5]{Casimir-ManifoldN1000b20gbw.eps} & \includegraphics[scale=0.5]{Casimir-ManifoldN15000b200gbw.eps} \\
\end{tabular}
\caption{Casimir surfaces for $(a)$ $N = 10$ and $b=10$, $(b)$ $N = 100$ and $b=10$, $(c)$ $N = 1000$ and $b=20$, $(d)$ $N = 15000$ and $b=200$.}
\label{CSF}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
$(a)$ & $(b)$ \\
\includegraphics[scale=0.5]{Casimir-ManifoldN10b10Nextbw.eps} & \includegraphics[scale=0.5]{Casimir-ManifoldN100b10Nextbw.eps} \\
$(c)$ & $(d)$ \\
\includegraphics[scale=0.5]{Casimir-ManifoldN1000b20Nextbw.eps} & \includegraphics[scale=0.5]{Casimir-ManifoldN15000b100Nextbw.eps} \\
\end{tabular}
\caption{Casimir surfaces for $v\in[-0.55N,0.55N]$ and $(a)$ $N = 10$ and $b=10$, $(b)$ $N = 100$ and $b=10$, $(c)$ $N = 1000$ and $b=20$, $(d)$ $N = 15000$ and $b=100$.}
\label{CSF2}
\end{center}
\end{figure}
Using the commutation relations of the currents (\ref{calg}) it is easy to calculate the anticommutators
\begin{eqnarray}
[\hat{\mathcal{T}},\hat{\mathcal{I}}]_{+} & = & 2 \hat{\mathcal{I}} \hat{\mathcal{T}} - i\hat{\mathcal{J}}, \\ \label{AC1a}
[\hat{\mathcal{T}},\hat{\mathcal{J}}]_{+} & = & 2 \hat{\mathcal{J}} \hat{\mathcal{T}} + i f(\hat{\mathcal{I}},\hat{N}), \\ \label{AC2a}
[\hat{\mathcal{J}},\hat{\mathcal{I}}]_{+} & = & 2 \hat{\mathcal{I}} \hat{\mathcal{J}} + i\hat{\mathcal{T}}.\label{AC3a}
\end{eqnarray}
\noindent We will use these anticommutators together with the commutators (\ref{calg}) in the calculus of the currents quantum dynamics.
\section{Interconversion Currents Quantum Dynamics}
We can rewrite the Hamiltonian (\ref{ham}) using the currents operators (\ref{cj}), (\ref{ci}) and (\ref{ct}) as
\begin{eqnarray}
\hat{H} &=& \alpha\hat{\mathcal{I}}^2 + \hat{\mathbf{\zeta}}\hat{\mathcal{I}} - 4\Omega\hat{\mathcal{T}} + \rho\hat{\mathcal{\mathfrak{C}}}_{1}^2 + \xi\hat{\mathcal{\mathfrak{C}}}_{1},
\label{ham2}
\end{eqnarray}
\noindent with the following identification for the parameters
\begin{eqnarray}
\alpha & = & 4U_a + U_b - 2U_{ab}, \nonumber \\
\beta & = & \left(2U_a - \frac{U_b}{2} \right),\nonumber \\
\gamma & = & 2\mu_a - \mu_b, \nonumber \\
\rho & = & \left(\frac{U_a}{4} + \frac{U_b}{16} + \frac{U_{ab}}{8} \right), \nonumber \\
\xi & = & \left( \frac{\mu_a}{2} + \frac{\mu_b}{4} \right), \nonumber
\end{eqnarray}
\noindent and with the definition of the Casimir operator $\hat{\mathbf{\zeta}} = \beta\hat{\mathcal{\mathfrak{C}}}_{1} + \gamma$.
The quantum dynamic of the currents are determined by the currents algebra (\ref{calg}), their commutation relations with the Hamiltonian and the parameters. Following \cite{GSantosCA1} we can write the second time derivative of any operator $\hat{O}$ in the Heisenberg picture as
\begin{equation}
\frac{d^2\hat{O}}{dt^2} = \left(\frac{i}{\hbar}\right)^2 [\hat{H},[\hat{H},\hat{O}]],
\label{dOdt0}
\end{equation}
\noindent or as
\begin{equation}
\frac{d^2\hat{O}}{dt^2} = \frac{i}{\hbar}[\hat{H},\frac{d\hat{O}}{dt}].
\label{dOdt}
\end{equation}
We can see from the Hamiltonian (\ref{ham2}) and Eq. (\ref{dOdt0}) that all Casimir operators are conserved quantities, $[\hat{H},\hat{\mathcal{\mathfrak{C}}}_{1}]=0$. Because the operator $\hat{N}$ is a Cassimir operator and a conserved quantity all wave function of the Hamiltonians (\ref{ham}) or (\ref{ham2}) are eigenfunctions of that operator with degenerated eigenvalue $N$ \cite{GSantosBVS-PLB}. As the another Casimir operators are functions of $\hat{N}$ we have also that all wave function of the Hamiltonians are degenerated eigenfunctions of these operators with the respective eigenvalues as functions of $N$. Therefore, we will consider the Casimir operators $\hat{\mathbf{\zeta}}$ and $\hat{N}$ as the respective $c$-numbers $N$ and $\zeta = \beta N + \gamma$ in the calculus of the currents quantum dynamics. So we also will write $\hat{\mathcal{N}}$ and $f(\hat{{\cal I}},\hat{N})$ as the respective functions of $N$,
\begin{equation}
\mathcal{N} = \frac{N}{8}\left(\frac{N}{4}+1\right), \qquad f(\hat{{\cal I}},N) = \frac{3}{2}\hat{{\cal I}}^{2}+\frac{N}{4}\hat{{\cal I}}-\mathcal{N}.
\label{c-n}
\end{equation}
Using the Eq. (\ref{ham2}) and (\ref{dOdt0}) or (\ref{dOdt}) we found the following equations for the quantum dynamics of the three currents
\begin{eqnarray}
\frac{d^{2}\hat{\mathcal{I}}}{dt^{2}}+i\frac{\alpha}{\hbar}\frac{d\hat{\mathcal{I}}}{dt}+24\frac{\Omega^{2}}{\hbar^{2}}\hat{{\cal I}}^{2}+4\frac{\Omega^{2}}{\hbar^{2}}N\hat{{\cal I}} & = & - 8\frac{\Omega\alpha}{\hbar^{2}}\hat{\mathcal{I}}\hat{\mathcal{T}} - 4\frac{\Omega\zeta}{\hbar^{2}}\hat{\mathcal{T}} + 16\frac{\Omega^{2}}{\hbar^{2}}\mathcal{N},
\label{eq1:wideeq}
\end{eqnarray}
\begin{eqnarray}
\frac{d^{2}\hat{\mathcal{J}}}{dt^{2}}+\frac{1}{\hbar^{2}}\left[\alpha^{2}+4\Omega^{2}N+\zeta^{2}\right]\hat{\mathcal{J}} & = & -4\frac{\alpha^{2}}{\hbar^{2}}\hat{\mathcal{I}}^{2}\hat{\mathcal{J}} - 4i\frac{\alpha^{2}}{\hbar^{2}}\hat{\mathcal{I}}\hat{\mathcal{T}} - 8\frac{\Omega\alpha}{\hbar^{2}}\hat{\mathcal{J}}\hat{\mathcal{T}} \nonumber\\
& - & \frac{4}{\hbar^{2}}\left(\alpha\zeta+12\Omega^{2}\right)\hat{{\cal I}}\hat{\mathcal{J}}-\frac{2i}{\hbar^{2}}\left(\alpha\zeta+12\Omega^{2}\right)\hat{\mathcal{T}} \nonumber\\
& - & 6i\frac{\Omega\alpha}{\hbar^{2}}\hat{{\cal I}}^{2} - i\frac{\Omega\alpha}{\hbar^{2}}N\hat{{\cal I}} + 4i\frac{\Omega\alpha}{\hbar^{2}}\mathcal{N},
\label{eq2:wideeq}
\end{eqnarray}
\begin{eqnarray}
\frac{d^{2}\hat{\mathcal{T}}}{dt^{2}}+\frac{1}{\hbar^{2}}\left[\alpha^{2}+\zeta^{2}\right]\hat{\mathcal{T}} & = & -4\frac{\alpha^{2}}{\hbar^{2}}\hat{\mathcal{I}}^{2}\hat{\mathcal{T}}+4i\frac{\alpha^{2}}{\hbar^{2}}\hat{\mathcal{I}}\hat{\mathcal{J}}-4\frac{\alpha\zeta}{\hbar^{2}}\hat{\mathcal{I}}\hat{\mathcal{T}}+2i\frac{\alpha\zeta}{\hbar^{2}}\hat{\mathcal{J}} \nonumber\\
& + & 8\frac{\Omega\alpha}{\hbar^{2}}\hat{\mathcal{J}}^{2} - 8\frac{\Omega\alpha}{\hbar^{2}}\hat{\mathcal{I}}f(\hat{{\cal I}},N) - 4\frac{\Omega\zeta}{\hbar^{2}}f(\hat{{\cal I}},N).
\label{eq3:wideeq}
\end{eqnarray}
\noindent Different choices of the parameters gives us different dynamics for the currents. The parameters $\rho$ and $\xi$ don't change the dynamics of the currents because they are coupling the Cassimir operator $\hat{\mathcal{\mathfrak{C}}}_{1}$. For $\alpha = \zeta = 0$, the current $\hat{\mathcal{T}}$ is a conserved quantity, $[\hat{\mathcal{H}},\hat{\mathcal{T}}]=0$, but this don't means that we don't have interconversion dynamics. We can see from Eqs. (\ref{SCDN1}) and (\ref{SCDN2}) that the quantum dynamic of $\hat{N}_a$, $\hat{N}_b$, and $\hat{\mathcal{I}}$ only depend of the interconversion current $\hat{\mathcal{J}}$ and the interconversion parameter $\Omega$. We only have dynamics if $\Omega \neq 0$. We get the following relations between the interactions and external potentials parameters
\begin{equation}
U_{ab} = 2U_a + \frac{U_b}{2}, \qquad \frac{\mu_b - 2\mu_a}{N} = 2U_a - \frac{U_b}{2}.
\label{cpar1}
\end{equation}
\noindent The simplest choice is the non interaction limit with all $U$'s equal to zero and with all external potentials $\mu$'s equal to zero, but the equations for the currents quantum dynamics are the same since that the parameters obey the relations (\ref{cpar1}). For $\alpha = 0$ and $\zeta \neq 0$, we have yet the external potentials $\mu$'s and the atom-atom $U_a$, molecule-molecule $U_b$ and atom-molecule $U_{ab}$ interactions, but the interactions are constrained by the first equation in (\ref{cpar1}). We can also choose $2\mu_a = \mu_b$ or $U_a = U_b/4$ and we will get $\zeta \neq 0$.
\section{Summary}
I have discussed the symmetries of the model and calculated the Casimir operators for the $O(3)$ and $SU(2)$ deformed algebras and showed that the surfaces corresponding to these Casimir operators are the same. I have showed that an interconversion currents algebra appears when I calculate the quantum dynamics of the unbound atoms $\hat{N}_a$ and molecules $\hat{N}_b$ boson number operators. I have used the Heisenberg equation of motion to write the second time derivative of the currents operators. Then I calculated the quantum dynamics of these currents and showed that different dynamics appear when I consider different choices of the parameters of the Hamiltonian.
\section*{Acknowledgments}
The author acknowledge CAPES/FAPERJ (Coordena\c{c}\~ao de Aperfei\c{c}oamento de Pessoal de N\'{\i}vel Superior/Funda\c{c}\~ao de Amparo \`a Pesquisa do Estado do Rio de Janeiro) for the financial support.
|
\section{Introduction}
Sparse Signal recovery (SSR) refers to algorithms which seek sparse solutions to underdetermined systems of equations \cite{eladbook}, which occur naturally when one seeks a representation of a given signal under an overcomplete dictionary. Overcomplete dictionaries have gained popularity in a wide range of applications because they are much more flexible than their undercomplete counterparts and lead to unique solutions under certain constraints, when sparsity has been enforced \cite{donoho2006stable}. Constraining the solution of underdetermined problems to be sparse represents prior knowledge about the solution and makes finding it tractable. In certain applications, structured sparsity, such as block sparsity, has been enforced on the desired coefficient vector, i.e., a small number of blocks of the solution are non-zero \cite{eldar2010block}.
SSR has become a very active research area in recent times because of its wide range of engineering applications. For example, in several popular computer vision problems, such as face recognition \cite{wright2009robust}, motion segmentation \cite{elhamifar2009sparse}, and activity recognition \cite{yang2008distributed}, signals lie in low-dimensional subspaces of a high dimensional ambient
space. An important class of methods to deal with this depends on exploiting the notion of sparsity. Following this path, Sparse Representation based Classification (SRC) \cite{wright2009robust} was proposed and produced state of the art results in a face recognition (FR) task.
In many applications, we often encounter outliers in measurements, which leads traditional SSR algorithms to fail and necessitates the development of an outlier robust SSR algorithm. The need for outlier resistant SSR algorithms motivates our present work, in which we develop a robust SSR algorithm and extend it to recover simultaneous block sparse signals. To show the efficacy of our approach, we focus on FR, which refers to identifying a subject's face given a labeled database of faces. The pioneering work of Wright et al. \cite{wright2009robust} on SRC showed that a face classifier can be devised by using the downsampled images from the training database as a dictionary and considering the sparse representation of a given image under that dictionary as the "identity" of the person. In this scenario, it is intuitive to assume that the dictionary is broken up into blocks corresponding to each specific person and constraining the encoding of the image to be block-sparse leads to performance gains \cite{elhamifar2011robust}.
One significant challenge in the FR problem is dealing with occlusions. Occlusions are outliers within the SRC model because it assumes that the dictionary spans the space of all possible observations. A popular way to incorporate robustness to outliers into the SSR model is to assume that the outliers themselves have a sparse representation
\cite{wright2009robust}, which has been shown to yield improved resilience to various forms of face occlusion and corruption \cite{wright2009robust, yang2011robust, wagner2012toward, elhamifar2011robust, li2013robust}. In certain cases, such as when the entire face is occluded or lighting conditions are extremely poor, FR within the SRC framework can yield unsatisfactory results because it is difficult to solve the single measurement vector (SMV) SSR problem. When possible, it is advantageous to acquire multiple measurements of the same source and instead solve the multiple measurement vector (MMV) problem. The MMV problem assumes that the support of the non-zero coefficients that encode each measurement does not change, while the actual values of the coefficients can vary. It is well known that the MMV problem yields much better recovery results than the SMV problem \cite{cotter2005sparse}.
In this work, we extend the SRC framework to the MMV case and consider performing FR when multiple images of the same subject, corrupted by non-stationary occlusions, are presented to the classifier. Our work is motivated, in part, by the person re-identification problem \cite{hirzer2011person, lisanti2015person, karanam2015sparse}. Srikrishna et al. \cite{karanam2015sparse} addressed the re-identification problem by applying SSR to each individual image of the subject and aggregating the results to form a global classifier. As such, \cite{karanam2015sparse} did not address the MMV nature of the problem. The main motivation behind our work is to enforce the prior knowledge that the input images correspond to the same person within the SSR process, while still maintaining resilience to time-varying occlusions.
Our SSR framework builds upon the hierarchical Bayesian framework discussed in \cite{tipping2001sparse,wipf2007empirical,zhang2011sparse}, known as Sparse Bayesian Learning (SBL).
This choice is motivated by the superior recovery results obtained for the standard SSR problem \cite{wipf2007empirical, giri2015type} and the Bayesian framework is convenient for extensions to problems with structure \cite{zhang2013extension}. In this work, we extend the SBL framework to the MMV block-sparse case and explicitly model time-varying occlusions, referring to our method as robust SBL (Ro-SBL).
\subsection{Contributions}
\begin{itemize}
\item We introduce a novel hierarchical Bayesian Robust SSR algorithm, Ro-SBL, for solving the MMV block-sparse problem with time-varying outliers. This work has connections to \cite{li2013robust}, where a Robust Block Sparse Bayesian Learning (BSBL) method was proposed. In contrast with our work, BSBL only considered the SMV problem and did not harness the ability of the SBL framework to capture non-stationary outliers.
\item We validate our proposed method with synthetic data results and also apply our method to a robust simultaneous FR task. Unlike \cite{karanam2015sparse}, our proposed approach exploits the prior knowledge that the input images correspond to the same person within the SSR process.
\end{itemize}
\vspace{-1.5em}
\section{Ro-SBL for Simultaneous Block Sparse Recovery}
The signal model for simultaneous block sparse recovery is given by
\begin{align}\label{eq:model}
\vY = \vA \vX + \vE + \mathbf{V}
\end{align}
where, $\vY \in \mathbb{R}^{n \times L}$ is the matrix of $L$ measurements, $\vect{A} \in \mathbb{R}^{n \times m}$ is the dictionary, $\mathbf{V} \in \mathbb{R}^{n \times L}$ is the independent and identically distributed (IID) Gaussian noise term with mean zero and variance $\sigma^2$, $\vX \in \mathbb{R}^{m \times L} $ is the encoding of the measurements under $\vA$, and $\vE \in \mathbb{R}^{n \times L}$ is the matrix containing the outliers in the measurements.
The key assumption in the MMV problem is that, if a given column of $\vect{A}$ is activated (i.e. its corresponding coefficient in $\vX$ is non-zero) for one of the measurements, then it will be activated for all of the measurements \cite{cotter2005sparse}. This means that the same set of basis vectors have been used to generate all of the measurements, which is reflected in the encoding matrix $\vX$ in the form of joint sparsity, i.e. $\lbrace \vx_{(:,i)} \rbrace_{i=1}^L$ share the same support, where $\vx_{(:,i)}$ is the $i$'th column of $\vX$ \cite{cotter2005sparse}. Within a Bayesian framework, the joint sparsity assumption translates to placing a prior on the \textit{rows } of $\vX$. In the context of our work, we build upon the extension of the SBL framework to the MMV problem in \cite{wipf2007empirical} and adopt a hierarchical prior, namely a Gaussian Scale Mixture (GSM), over the rows of $\vX$:
\begin{equation}
p\left(\mathbf{x}_{(j,:)} | \gamma_j \right) = \mathsf{N} \left(\mathbf{x}_{(j,:)};0, \gamma_j \vI_L \right)
\end{equation}
where, $\mathbf{x}_{(j,:)}$ denotes the $j$'th row of $\vX$ and $\gamma_j$ is the unknown variance hyperparameter. In addition, we also consider block sparsity in each $\vx_{(:,i)}$, where the block structure is shared among all of the encoding vectors. Assuming that the support of $\vx_{(:,i)}$ is separated into disjoint sets $\mathscr{G}_g, 1 \leq g \leq G$, which are known a-priori and shared across all $i$, $p(\vX)$ is amended to reflect that each of the rows in a given group $\mathscr{G}_g$ share the same $\vg_g$ \cite{zhang2013extension}:
\begin{equation}
p(\vX) = \prod_{g=1}^G \prod_{j \in \mathscr{G}_g} p\left(\mathbf{x}_{(j,:)} | \gamma_g\right)
\end{equation}
Although the joint block sparsity constraint is a valid one for the encoding matrix $\vX$, it does not hold for the outlier matrix $\vE$ since the outliers could be non-stationary, i.e., time varying. Therefore, we will treat each $\ve_{(:,i)}$ independently and not constrain the outliers to share the same support across all measurements. As such, we adopt a sparsity enforcing GSM prior on $\ve_{(:,i)}$, which induces the following prior on $\vE$:
\begin{equation}
p (\vE) = \prod_{j=1}^n \prod_{i=1}^L p(\epsilon_{ji}| \delta_{ji}) = \prod_{j=1}^n \prod_{i=1}^L \mathsf{N}(\epsilon_{ji}; 0, \delta_{ji})
\end{equation}
This set of assumptions is unique to this work and is motivated by the FR task.
\vspace{-1em}
\subsection{Incorporating Robustness to Outliers}
\vspace{-0.2em}
For an SMV problem, \cite{li2013robust}\cite{jin2010algorithms} showed that sparse (under the standard basis) outliers can be incorporated into the well known SBL framework by introducing a simple modification to the dictionary $\vA$. In the present work, we extend this idea to the MMV case, which results in the following modification to the signal model in \eqref{eq:model}:
\begin{align}\label{eq:zhilin}
\vY = \underbrace{\begin{bmatrix}
\vA & \vI_n
\end{bmatrix}}_{\boldsymbol{\tilde{A}}} \underbrace{\begin{bmatrix}
\vX \\ \vE
\end{bmatrix}}_{\vXt} + \vV.
\end{align}
Note that \eqref{eq:zhilin} and \eqref{eq:model} are equivalent, but, as will be shown next, the signal model in \eqref{eq:zhilin} lends itself much more nicely to a closed form inference procedure.
\vspace{-1em}
\subsection{Ro-SBL Inference Procedure}
The goal of the inference procedure is to estimate the hyperparameters $\vg = \begin{bmatrix}
\gamma_1 & \cdots & \gamma_G
\end{bmatrix}^T$ and $\vDel = \begin{bmatrix}
\vd_{(:,1)} & \cdots & \vd_{(:,L)}
\end{bmatrix}$, where $\vd_{(:,i)} = \begin{bmatrix}
\delta_{1i} &\cdots& \delta_{ni}
\end{bmatrix}^T$. As in \cite{tipping2001sparse}\cite{wipf2007empirical}, we adopt an Expectation Maximization (EM) procedure where we treat $\vXt$ as the hidden data. In the E-step, we seek the expectation of the complete data $(\vY,\vXt,\vg,\vDel)$ log likelihood under the posterior $p(\vXt | \vY, \vg^t,\vDel^t,\sigma^2)$, where $t$ denotes the iteration index. Because $\lbrace \vxt_{(:,i)} \rbrace_{i=1}^L$ are conditionally independent given $\vY$, $\vg$, and $\vDel$, the E-step reduces to
\begin{align}\label{eq:Q}
\begin{split}
& Q(\vg,\vDel, \sigma^2, \vg^t,\vDel^t) = \\
& \sum_{i=1}^L E_{\vxt_{(:,i)} | \vy_{(:,i)},\vg^t,\vd_{(:,i)},\sigma^2} \left[ \log p\left(\vy_{(:,i)},\vxt_{(:,i)},\vg,\vd_{(:,i)},\sigma^2 \right)\right].
\end{split}
\end{align}
The posterior needed to compute \eqref{eq:Q} is given by $\mathsf{N}\left( \vxt_{(:,i)} ; \vmu^i, \vS^i \right)$, where $\vmu^i$ and $\vS^i$ are given by \cite{tipping2001sparse}\cite{wipf2007empirical}
\begin{align*}
\begin{split}
\vmu^i &= \vP^i \vAtilde^T \left(\sigma^2 \boldsymbol{I}_{m+n} + \vAtilde \vP^i \vAtilde^T\right)\inv \vy_{(:,i)} \\
\vS^i &= \vP^i - \vP^i \vAtilde^T \left(\sigma^2 \boldsymbol{I}_{m+n} + \vAtilde \vP^i \vAtilde^T \right)\inv \vAtilde \vP^i
\end{split}
\end{align*}
where $\vP^i$ is a diagonal matrix containing $\begin{bmatrix}
\vgtilde^T & \vd_{(:,i)}^T
\end{bmatrix}$ on the diagonal and $\vgtilde \in \mathbb{R}^m$ is set to $\tilde{\gamma}_j = \gamma_g$ for $j \in \mathscr{G}_g$. Unlike \cite{wipf2007empirical}, where the covariance of the posterior is shared for all $i$, the covariance is a function of $i$ here because each $\vxt_{(:,i)}$ consists of $\vx_{(:,i)}$, whose support does not vary with $i$, and $\ve_{(:,i)}$, whose support does vary.
In the M-step, $Q(\vg,\vDel,\sigma^2, \vg^t,\vDel^t)$ is maximized with respect to $(\vg,\vDel, \sigma^2)$, leading to the update rules:
\begin{align}\label{eq:update rules mmv}
\begin{split}
&\gamma_g = \sum_{i=1}^L\sum_{j \in \mathscr{G}_g} \frac{\vS_{jj}^i + \left(\vmu_j^i\right)^2}{\vert \mathscr{G}_g \vert L} \\
&\delta_{ji} = \vS_{\tilde{j} \tilde{j}}^i + \left(\vmu_{\tilde{j}}^i\right)^2 \; ,\; 1 \leq j \leq n \; , \; \tilde{j} = j + m \\
&\sigma^2 = \\
&\sum_{i=1}^L \frac{\left\Vert \vy_{(:,i)} \right\Vert^2 -2\vy_{(:,i)}^T \vAt \vmu^i + \mathsf{tr}\left( \vAt^T \vAt \left( \vS^i + \vmu^i \left(\vmu^i\right)^T\right)\right) }{L n}
\end{split}
\end{align}
where $\mathsf{tr}(\cdot)$ refers to the trace operator.
Upon convergence of the EM algorithm to the estimates $\hat{\vg}$ and $\hat{\vDel}$, $\vX$ and $\vE$ can be estimated using the maximum a-posteriori (MAP) estimator:
\begin{align*}
\begin{bmatrix}
\hat{\vx}_{(:,i)}^T & \hat{\ve}_{(:,i)}^T
\end{bmatrix}^T = \argmax_{\vxt_{(:,i)}} p\left( \vxt_{(:,i)} | \vy_{(:,i)},\hat{\vg},\hat{\vd}_{(:,i)},\sigma^2 \right) = \hat{\vmu}^i
\end{align*}
\vspace{-1.5em}
\section{Results}
\vspace{-0.5em}
\subsection{Synthetic Data Results}
To validate the proposed method, we conducted SSR experiments on synthetic data. To generate the synthetic data, we begin by randomly selecting $s$ sets from $\lbrace \mathscr{G}_g \rbrace_{g=1}^G$ and generate $\vx_{(:,i)}$ such that the non-zero elements are indexed by one of the selected sets. We use equally sized blocks of length $8$. The non-zero elements of $\vx_{(:,i)}$ are drawn from the $\mathsf{N}(0,1)$ distribution. We generate $\vA \in \mathbb{R}^{80 \times 160}$ by drawing its elements from the $\mathsf{N}(0,1)$ distribution and normalizing the columns to have unit $\ell_2$ norm. Finally, we use the robust modeling strategy and replace
$\vA$ by $\tilde{\vA} = \begin{bmatrix}
\vA & \vI_n \end{bmatrix}$. In order to simulate a noisy SSR scenario, we generate $\vv_{(:,i)}$ by drawing its elements from the $\mathsf{N}(0,1)$ distribution and $\ve_{(:,i)}$ by drawing its elements form the student-t distribution with one degree of freedom. Finally, we generate observations $\vy_{(:,i)}$ according to \eqref{eq:model} after scaling $\vv_{(:,i)}$ and $\ve_{(:,i)}$ to achieve a specified Signal-to-Gaussian noise ratio (SGNR) and Signal-to-Outlier noise ratio (SONR).
Let $\hat{\vX}$ denote the approximation to $\vX$ generated by the SSR algorithm. We measure the quality of the recovery using the relative $\ell_2$ error: $\frac{1}{L}\sum_{i=1}^L \frac{\Vert \vx_{(:,i)} - \boldsymbol{\hat{x}}_{(:,i)} \Vert^2}{\Vert \vx_{(:,i)} \Vert^2}$. We performed the synthetic data experiment 500 times and report the average performance results.
We compare the performance of the proposed method to several standard SSR algorithms. As a baseline, we use the $\ell_1$ SSR approach and the block sparse extension of the $\ell_1$ approach, the $\ell_2-\ell_1$ block SSR algorithm (also known as Group LASSO \cite{bakin1999adaptive}), which seeks
\begin{align}\label{eq:l2-l1}
\boldsymbol{\hat{x}} &= \argmin_{\vx} \Vert \vA \vx - \vy\Vert_2^2 + \lambda \sum_{g=1}^G \Vert \vx_{\mathscr{G}_g} \Vert_2.
\end{align}
Note that \eqref{eq:l2-l1} reduces to the $\ell_1$ SSR objective function when each element of $\vx$ is a separate group. We use the SLEP \cite{liu2009slep} software package to solve the $\ell_1$ and $\ell_2-\ell_1$ problems. For comparison purposes we naively extend the $\ell_1$ and $\ell_2-\ell_1$ approaches to the MMV case by solving each MMV problem as $L$ independent SMV problems.
We also compare our approach with Block-SBL (BSBL) \cite{zhang2013extension, li2013robust}, which is a hierarchical Bayesian framework for solving the SMV block-sparse recovery problem. We naively extend BSBL to the MMV case by assuming that the outliers have stationary support, denoting the resulting algorithm as M-BSBL. In the context of the signal model in \eqref{eq:model}, M-BSBL corresponds to assuming \textit{row sparsity} on $\vE$.
Simulation results for a 5 measurement SSR problem with 40 dB SGNR and 5 dB SONR are shown in Fig. \ref{fig:tvo}. M-BSBL and Ro-SBL drastically outperform the $\ell_1$ and $\ell_2-\ell_1$ SSR approaches, which shows that hierachical Bayesian approaches outperform deterministic methods even in the challenging SSR setup considered. In addition, since the Bayesian approaches explicitly model the MMV nature of the problem, this result shows that significant improvements can be achieved by incorporating the prior knowledge that the support of $\vx_{(:,i)}$ does not change with $i$ in the SSR algorithm. We observe a $25\%-51\%$ improvement in relative $\ell_2$ error from Ro-SBL compared to M-BSBL for $s \leq 6$. This suggests that Ro-SBL is better able to capture outliers due to its superior model.
For illustrative purposes, we conducted the same synthetic data experiment, with the exception that the outliers were constrained to be the same for all $i$. The results are shown in Fig. \ref{fig:so}. As expected, the performance gains of the hierachical Bayesian approaches over the deterministic approaches carries over into the stationary outlier scenario. It is important to note that M-BSBL slightly outperforms Ro-SBL because, in this case, the M-BSBL signal model is better fitted to estimate the outliers since M-BSBL assumes that the outliers are stationary and enjoys the advantages of MMV modeling, even for outliers.
\begin{table*}[t]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
& \multicolumn{2}{c|}{$10\%$} & \multicolumn{2}{c|}{$20\%$} & \multicolumn{2}{c|}{$30\%$} & \multicolumn{2}{c|}{$40\%$} & \multicolumn{2}{c|}{$50\%$} \\ \hline
& $L = 1$ & $L = 5$ &$L = 1$ &$L = 5$ &$L = 1$ &$L = 5$ &$L = 1$ &$L = 5$ &$L = 1$ &$L = 5$ \\ \hline
SRC \cite{wright2009robust} & 89.72 & 100 & 83.61 & 100 & 71.29 & 97.81 & 54.24 & 92.54 & 35.97 & 67.98 \\ \hline
$P_{\ell_2-\ell_1}$ \cite{elhamifar2011robust} & 91.60 & 100& 84.50 & 100& 69.90& 96.93&54.08 & 93.42 & 39.07 & 72.37\\ \hline
M-BSBL \cite{li2013robust} & --- & 100 & --- & 100 & --- & 99.12& ---& 94.30 & --- & 76.75\\ \hline
Ro-SBL (proposed) & \textbf{91.92} & \textbf{100} & \textbf{91.11} & \textbf{100}& \textbf{83.52}& \textbf{100}& \textbf{65.58}& \textbf{100} & \textbf{41.44} & \textbf{91.22} \\ \hline
\end{tabular}
\vspace{1em}
\caption{Face classification accuracy results for $L = 1$ and $L = 5$ for various occlusion rates}
\label{table:classification}
\end{table*}
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering\includegraphics[height=15em, width = \linewidth]{l2_error_snr40_sor5_TVO_2.png}
\caption{Time-Varying outliers}
\label{fig:tvo}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\centering\includegraphics[height=15em, width = \linewidth]{l2_error_snr40_sor5_SO_2.png}
\caption{Stationary outliers}
\label{fig:so}
\end{subfigure}
\caption{Comparison of SSR algorithms on synthetic data}
\vspace{-0.5em}
\end{figure}
\begin{figure*}
\centering
\begin{multicols}{2}
\centering
\begin{tikzpicture}
\node[anchor=north west,inner sep=0] at (0,0) (c1) {\includegraphics[scale=0.25]{face1.png}};
\node[anchor=north west,inner sep=0] at (6,0) (c2) {\includegraphics[scale=0.25]{coefficients1.png}};
\node[anchor=north west,inner sep=0] at (13,0) (c3) {\includegraphics[scale=0.25]{cleanface1.png}};
\node[anchor=south west,inner sep=0] at (0,0) (c4) {\includegraphics[scale=0.25]{face2.png}};
\node[anchor=south west,inner sep=0] at (6,0) (c5) {\includegraphics[scale=0.25]{coefficients2.png}};
\node[anchor=south west,inner sep=0] at (13,0) (c6) {\includegraphics[scale=0.25]{cleanface2.png}};
\draw[draw=yellow!60!black,line width=10pt,-stealth] (c1) -- (c2)
node[midway,text=black,font=\footnotesize\bfseries]{SSR};
\draw[draw=yellow!60!black,line width=10pt,-stealth] (c2) -- (c3)
node[midway,text=black,font=\footnotesize\bfseries]{Reconstruction};
\draw[draw=yellow!60!black,line width=10pt,-stealth] (c4) -- (c5)
node[midway,text=black,font=\footnotesize\bfseries]{SSR};
\draw[draw=yellow!60!black,line width=10pt,-stealth] (c5) -- (c6)
node[midway,text=black,font=\footnotesize\bfseries]{Reconstruction};
\end{tikzpicture}
\end{multicols}
\caption{Example of face reconstruction using Ro-SBL. The red bar denotes coefficients corresponding to the true subject class.}
\label{fig:reconstruction}
\end{figure*}
\vspace{-1em}
\subsection{FR Results}
\vspace{-0.2em}
In this section, we present results demonstrating the efficacy of the proposed method in a FR task. We use the Extended Yale B Database \cite{georghiades2001few}, which consists of 2441 images of 38 subjects under various illumination conditions. Each $192 \times 168$ image is a frontal perspective of the subject's face and has been cropped such that only the face can be seen. We randomly split the database into training and test sets. Following the SRC framework, we downsample the images in the training set by $\frac{1}{12}$, vectorize the result, and concatenate the vectors to form the dictionary $\vA$:
\begin{align*}
\vA = \begin{bmatrix}
\underbrace{\cdots \vA^1 \cdots}_{Block 1} | & \underbrace{\cdots \vA^2 \cdots}_{Block 2} |& \cdots \cdots |& \underbrace{\vA^{38}}_{Block 38}
\end{bmatrix}
\end{align*}
where $\vA^k$ consists of training images from the $k$'th subject. Note that this automatically introduces a block structure in $\vA$. As before, we replace $\vA$ by $\boldsymbol{\tilde{A}}$.
In the testing phase, we seek to identify a given subject from $L$ images of that subject's face. To simulate time-varying occlusions, we occlude each image by one of 10 animal images, choosing the location of the occluding image randomly. Given $L$ observations, we use one of the SSR algorithms to estimate $\hat{\vX}$ and $\hat{\vE}$. Similar to the person re-identification classifier presented in \cite{karanam2015sparse}, we label the test images using
\begin{align*}
k^* = \argmin_k \sum_{i = 1}^L \left\Vert \vy_{(:,i)} - \boldsymbol{{A}} \left(\boldsymbol{\phi}^k \odot \boldsymbol{\hat{x}}_{(:,i)}\right) - \boldsymbol{\hat{\epsilon}}_{(:,i)}\right\Vert^2
\end{align*}
where $\boldsymbol{\phi_{\mathscr{G}_k}}^k = 1$ and $\boldsymbol{\phi_{\mathscr{G}_c}}^k = 0, \forall c \neq k$.
For each test subject, around $30$ test images were available and the classification experiment was run $\left \lfloor{\frac{30}{L}}\right \rfloor$ times. We report averaged classification results in Table \ref{table:classification} for two cases: $L=1$, i.e. an SMV FR problem considered in \cite{yang2011robust, wright2009robust, elhamifar2011robust}, and $L=5$, i.e. an MMV problem considered in person re-identification \cite{karanam2015sparse}. Note that our proposed algorithm becomes equivalent to M-BSBL for $L=1$, hence we do not report M-BSBL results for this case in Table \ref{table:classification}. It is evident from the results that, for every algorithm, the $L=5$ case leads to much better classification accuracy compared to $L=1$, which corroborates the well known result that MMV modeling is superior to SMV in harsh conditions.
In all cases, Ro-SBL performs better than all other competing algorithms. For low occlusion rates ($10\%$), all of the competing algorithms perform comparably, despite the fact that only Ro-SBL explicitly models non-stationary occlusions. This result can be explained by the fact that $10\%$ occlusion represents a minor fraction of the overall image area and a significant amount of facial features remain un-occluded. On the other hand, Ro-SBL drastically outperforms the other algorithms at high occlusion rates ($50 \%$) with $L=5$, which can be attributed to the fact that a large portion of the facial features in each image are occluded and the SSR algorithm is forced to use all 5 measurements to jointly recover $\hat{\vX}$. Since Ro-SBL models outliers more accurately than the other SSR methods, it is better able to approximate the identity of the occluded subject.
Finally, as a visual example of how the proposed method performs face classification, we show that the occlusion can be removed from the test image by considering $\vA\vx_{(:,i)}$ as the estimate of the original, un-occluded test image. The face reconstruction result is shown in Fig. \ref{fig:reconstruction} (results are generated using $L = 5$ and a downsampling factor of $\frac{1}{6}$ was used for visualization purposes), which shows that the proposed method removes much of the occluding image and provides a relatively good reconstruction of the original face. Moreover, the coefficient plots show that the dominant coefficients all reside in the block corresponding to the test subject's index in $\vA$.
\vspace{-0.9em}
\section{Conclusion}
\vspace{-0.9em}
In this article, we have proposed a novel robust sparse recovery algorithm based on the well known SBL framework to recover simultaneous block sparse signals in presence of time varying outliers. Along with validating our method on synthetic data, we show the efficacy of our approach in simultaneous FR in the presence of time varying outliers.
|
\section{Introduction}
The recent discovery of unconventional superconductivity (SC) in Fe pnictides has led to an intense research effort aimed towards understanding their fundamental properties and the underlying mechanisms that lead to strong correlations between the lattice, charge and magnetic degrees of freedom \cite{Johnston10,Canfield10,Stewart11}. Similar to the layered cuprate superconductors, the SC in FeAs-based compounds seems to arise close to an antiferromagnetic (AFM) phase, suggesting that magnetism and SC are closely intertwined in these systems \cite{Johnston97,Maple98,Lee06,Lumsden10,Dai12}. However, the parent cuprates are insulators with strongly-correlated localized magnetic moments while the parent FeAs-based superconductors are metals with itinerant moments \cite{Johnston97,Johnston10,Si08}. MnAs-based systems form a bridge between the high-{$T_{\textrm c}$} cuprates and the FeAs-based materials, such as in tetragonal {BaMn$_{2}$As$_{2}$} that orders with a G-type AFM structure and shares the same $I4/mmm$ crystal structure as many of the Fe pnictides but manifests an insulating ground state with localized moments similar to cuprates \cite{Singh09,Johnston11}. Interestingly, unlike {BaMn$_{2}$As$_{2}$}, the insulator {SrMn$_{2}$As$_{2}$} crystallizes in a trigonal unit cell (space group $P\bar{3}m1$) with a corrugated honeycomb structure \cite{Brechtel78,Wang11,Sangeetha16} which yields a possibly frustrated Mn spin-system \cite{Rastelli79,Mazin13,McNally15}. This kind of system attracts a lot of attention because depending upon the strength and nature of the spin interactions, different magnetic structures, from a collinear N\'{e}el AFM phase to a magnetic spiral and even to a stripe phase with alternating ferromagnetic (FM) stripes are possible \cite{Rastelli79,Mazin13}.
\begin{figure}
\centering
\includegraphics{Fig1final.eps}
\caption{\label{Fig1}
(Color online) Temperature dependence of magnetic susceptibility $\chi(T)$ for $\textbf{\textit{H}}\parallel \textbf{\textit{c}}$ and $\textbf{\textit{H}}\perp \textbf{\textit{c}}$ with $H$=1000 Oe. Two views of the magnetic structure are shown as insets. Right inset: A side view of the collinear AFM alignment of the Mn atoms at two different heights along the $\textbf{\textit{c}}$-axis marked by blue (Mn1) and red (Mn2) planes respectively, with two antiparallel spin directions. Left inset: The ordered Mn atoms as viewed along the $\textbf{\textit{c}}$-axis.}
\end{figure}
\begin{figure*}
\centering
\includegraphics{Fig2.eps}
\caption{\label{Fig2}
Neutron diffraction measurements at $T = 5$ K to search for magnetic Bragg peaks. (a) $l$-scans along $(h0l)$ with $h = 0$, 1, and 2. (b) $h$-scans along $(h0l)$ with $l = 0$, and 2. (c) $hh$-scans along $(hhl)$ with $l = 0$, and 2. The plots are offset along the vertical axis. Additional peaks scattered from the aluminum (Al) sample holder are observed in (c).}
\end{figure*}
Here we report single-crystal neutron diffraction studies on {SrMn$_{2}$As$_{2}$} which orders in a collinear N\'{e}el AFM phase below {$T_{\textrm N}$}~$=118(2)$~K showing that the dominant spin interactions are not frustrated. The magnetic Mn moments are antiferromagnetically aligned in the basal plane of the trigonal unit cell with an ordered moment of $3.6(2)$ $\mu_{\textrm B}$/Mn ion at $T=5$~K. The two Mn moments within the unit cell form a bilayer with antiparallel spins and can be viewed as a corrugated honeycomb lattice. High-resolution x-ray diffraction measurements were also performed to study the influence of magnetoelastic coupling in this system but no distortion of the lattice was observed down to a base temperature of 6 K within the resolution limit.
Single crystals of {SrMn$_{2}$As$_{2}$} were grown out of Sn flux using conventional high-temperature solution growth techniques \cite{Canfield01,Sangeetha16}. Both magnetic susceptibility $\chi(T)$ and neutron diffraction measurements were carried out on the same single crystal of mass 44 mg, with dimensions $4.0 \times 3.0 \times 1.0$~mm$^3$. The crystals grow as flat plates with the $\textbf{\textit{c}}$-axis perpendicular to them. A Quantum Design, Inc., superconducting quantum interference device magnetic properties measurement system was used for the $\chi(T)$ measurements. Energy-dispersive x-ray measurements on the same single crystal confirmed the composition to be a pure 122 phase. Single-crystal neutron diffraction measurements were performed at the thermal triple-axis spectrometer, TRIAX, at the University of Missouri Research Reactor. Measurements were carried out with an incident energy of 14.7 meV, using S\"{o}ller collimations of $60^\prime$-$40^\prime$-sample-$40^\prime$-$80^\prime$. Pyrolytic graphite filters were placed both before and after the sample to reduce higher-order wavelengths. The sample was mounted on the cold finger of a closed-cycle helium cryostat to reach temperatures of $5$ K $\le T \le 300$ K. The lattice parameters were measured to be $a = b = 4.29(1)~\text{\AA}$ and $c = 7.24(1)~\text{\AA}$ at 5~K. Rocking scans performed through the Bragg peaks showed a full width at half maximum of $\approx$~$0.3\degree$ confirming the good mosaicity of the sample.
High-energy x-ray diffraction measurements were performed on a 1 mg single crystal at the 6-ID-D station at the Advanced Photon Source using an x-ray wavelength of $\lambda = 0.123712 ~\text{\AA}$ and a beam size of $100 \times 100 ~\mu \text{m}^2$. The sample was cooled down using a closed-cycle He cryostat. Two Be domes were placed over the sample and evacuated, and a small amount of He exchange gas was subsequently added to the inner dome for thermal equilibrium. An aluminized-Kapton heat shield also surrounded the sample and inner Be dome. The cryostat was mounted on a 6-circle diffractometer and a MAR345 image plate was used to measure the diffracted x-rays \cite{Kreyssig07}.
The $T$-dependence of the magnetic susceptibilities $\chi(T)$ with applied magnetic field ($\textbf{\textit{H}}$) along the $\textbf{\textit{c}}$-axis ($\chi_c$, $\textbf{\textit{H}}\parallel \textbf{\textit{c}}$) and perpendicular to it ($\chi_{ab}$, $\textbf{\textit{H}}\perp \textbf{\textit{c}}$) are shown in Fig.~\ref{Fig1}. For $\textbf{\textit{H}}\perp \textbf{\textit{c}}$, $\chi_{ab}$ starts to decrease rapidly below {$T_{\textrm N}$} $=118(2)$ K with a distinct change in the slope while $\chi_{c}$ remains almost constant suggesting an antiferromagnetic (AFM) transition with the ordered moments aligned within the $\textbf{\textit{ab}}$ plane.
Theoretical calculations predict that frustration in a corrugated honeycomb lattice can result in a spiral or an incommensurate magnetic ground state \cite{Rastelli79,Mazin13}. Therefore, neutron diffraction measurements were performed to determine the magnetic structure. Experiments were carried out in two configurations -- first in the $(h0l)$ horizontal scattering plane followed by measurements in the $(hhl)$ scattering plane. Extensive measurements along the three principal directions $[00l]$, $[h00]$, and $[hh0]$ found no additional magnetic scattering beyond the magnetic contributions at the Bragg peaks corresponding to the chemical crystal structure as shown in Fig.~\ref{Fig2}. Additional diffraction peaks observed in Fig.~\ref{Fig2}(c) correspond to the scattering from the aluminum (Al) sample holder which was confirmed by rocking scans through the Al peak positions with no significant observed variation in peak intensities.
\begin{figure}
\centering
\includegraphics{Fig3.eps}
\caption{\label{Fig3}
(Color online) Integrated intensities of the (002), (100) and (112) Bragg peaks as a function of temperature $T$ are shown in (a)--(c) respectively. The solid line is a power law fit given by, $I_{\textrm M}=I_0(1-T/T_{\textrm N})^{2\beta}$, for $T\ge 90$ K. (d)--(f) Rocking scans at specific temperatures [as mentioned in panel (e)] around (002), (100) and (112) peaks, respectively. The data at 120 K and 140 K overlap with each other.}
\end{figure}
As shown in Fig.~\ref{Fig3}, magnetic intensity develops only at the nuclear Bragg peak positions below {$T_{\textrm N}$} suggesting that the magnetic unit cell is same as that of the chemical unit cell. Therefore, the magnetic structure is built by the two magnetic Mn atoms within the same unit cell only. Symmetry analysis using the program SARAh-Representational Analysis \cite{Sarah} provided distinct magnetic structures for this system which includes FM or AFM alignments with the moments aligned either along the $\textbf{\textit{c}}$-axis or in the $\textbf{\textit{ab}}$-plane. No canting of the moments is allowed by symmetry for a second-order magnetic phase transition. Since we observe magnetic intensity in $(00l)$-type Bragg peaks in our neutron diffraction measurements, we can confirm that the spins are not aligned along the $\textbf{\textit{c}}$-axis. This is due to the fact that the neutron scattering cross-section is sensitive only to the component of the ordered magnetic moment perpendicular to the scattering vector. Moreover, no indication of a FM signal was found in magnetization measurements \cite{Sangeetha16}. The only remaining possibility is a collinear N\'{e}el AFM phase with the Mn moments in the $\textbf{\textit{ab}}$-plane with $180^\circ$ AFM alignments between a moment and all nearest neighbor moments as shown in the insets to Fig.~\ref{Fig1}.
All accessible peaks in the $(h0l)$ and $(hhl)$ scattering planes have been measured and analyzed with a total of 14 independent peaks, excluding peaks with high non-magnetic intensities. Figures~\ref{Fig3}(a)--(c) show the temperature dependence of the integrated intensities of the (002), (100) and (112) peaks. Rocking scans at specific temperatures [as mentioned in panel (e)] around the (002), (100) and (112) peaks are shown in Figs.~\ref{Fig3}(d)--(f), respectively. There are no changes in the peak intensities between 120~K and 140~K but as the temperature is lowered below {$T_{\textrm N}$}=118(2) K, the magnetic contribution sets in and the peak intensities increase.
The magnetic moment can be calculated from the integrated intensities measured by rocking scans on a series of peaks in both the $(h0l)$ and $(hhl)$ configurations. The magnetic intensity is given by \cite{Shirane}
\begin{equation}
\label{eq1}
I_\textrm{M} = N_\textrm{M} \frac{(2\pi)^3}{v_\textrm{M}}\sum_{\textbf{\textit{G}}_\textrm{M}} \delta(\textbf{\textit{Q}} - \textbf{\textit{G}}_\textrm{M})\left | \textbf{\textit{F}}_\textrm{M} (\textbf{\textit{G}}_\textrm{M})\right |^2,
\end{equation}
where $v_\textrm{M}$ is the magnetic unit cell volume, $N_\textrm{M}$ is the number of such cells in the sample, $\textbf{\textit{Q}}$ is the scattering vector and $\textbf{\textit{G}}_\textrm{M}$ is the magnetic reciprocal-lattice vector. $\textbf{\textit{F}}_\textrm{M}$ is the magnetic structure factor and is given by
\begin{equation}
\label{eq2}
\textbf{\textit{F}}_\textrm{M} (\textbf{\textit{G}}_\textrm{M}) = \frac{\gamma r_0}{2}\sum_{j}gf_j(\textbf{\textit{G}}_\textrm{M})\textbf{\textit{S}}_{\perp j}e^{i\textbf{\textit{G}}_\textrm{M}\cdot \textbf{\textit{d}}_j} e^{-\textbf{\textit{W}}_j},
\end{equation}
where $\gamma = 1.193$ is the magnetic dipole moment of the neutron in units of nuclear Bohr magneton, $r_0 = 2.818 \times 10^{-15}$ m and $g$ is the spectroscopic splitting factor. $f(\textbf{\textit{G}}_\textrm{M})$ and $e^{-\textbf{\textit{W}}}$ are the magnetic form factor and the Debye-Waller factor for the magnetic Mn ions, respectively. The index $j$ is a sum over all $j^{\textrm{th}}$ Mn ions in the unit cell. The quantity $\textbf{\textit{S}}_{\perp}$ is the magnetic interaction vector and is given by $\left | \textbf{\textit{S}}_{\perp} \right |^2 = \sum_{\alpha\beta}(\delta_{\alpha \beta} - \hat{Q}_\alpha \hat{Q}_\beta) S^{\ast}_{\alpha}S_{\beta}$. $\hat{Q}$ is a unit vector parallel to the scattering vector $\textbf{\textit{Q}}$ and the indices $\alpha$ and $\beta$ represent the $x$, $y$ and $z$ components in the summation. This reflects the fact that only the component of $\textbf{\textit{S}}$ perpendicular to $\textbf{\textit{Q}}$ contributes to the magnetic scattering amplitude. The magnetic intensities were obtained from the difference in integrated intensities between 5 K and 140 K ($>$ {$T_{\textrm N}$}) data sets. The difference is normalized with the corresponding integrated intensities at 140 K which are purely of nuclear origin. The temperature dependence of the nuclear intensity is almost constant above {$T_{\textrm N}$} as evident from Figs.~\ref{Fig3}(a)--(c). Thus any excess contribution of nuclear intensity at 5 K due to the Debye-Waller factor can be neglected.
Though the ordered collinear Mn moments lie in the $\textbf{\textit{ab}}$ plane of the trigonal lattice, it is not possible to uniquely determine their orientation in this plane due to the symmetry of the crystal structure. Moreover, one has to take into account the possible domain orientations in calculating the ordered moment. Considering equal domain populations and the magnetic moments aligned along the high symmetry directions, there are six possible spin directions (or domains) for the Mn moment: $\textbf{\textit{S}} = \pm [100], \pm[010]$ and $\pm[1\bar{1}0]$. The second Mn moment within the magnetic unit cell is aligned in the corresponding antiparallel direction. If $\eta_i$ is the angle between $\textbf{\textit{S}}_i$ and $\textbf{\textit{Q}}$ for the $i^{\textrm{th}}$ domain, then the average magnetic interaction vector contributing to the magnetic intensity for that particular peak is given by $\left \langle \left | \textbf{\textit{S}}_{\perp} \right |^2 \right \rangle = S^2(1-\left \langle \textrm{cos}^2\eta_i \right \rangle)$. The ordered moment is obtained by $\mu = gS\mu_{\textrm B}$ and was determined to be $3.6(2)$ $\mu_{\textrm B}$/Mn at 5 K. The quoted error includes the uncertainty in the domain populations. The ordered moment suggests local moment AFM behavior as observed in {BaMn$_{2}$As$_{2}$} (Ref.\citenum{Singh09}). It is somewhat lower than the nominal $5.0$ $\mu_{\textrm B}$/Mn expected for the high-spin state of Mn$^{2+}$ but is comparable to other Mn-122 compounds like {BaMn$_{2}$As$_{2}$} (Ref.\citenum{Singh09}), {BaMn$_{2}$P$_{2}$} (Ref.\citenum{Brock94}), {SrMn$_{2}$P$_{2}$} (Ref.\citenum{Brock94}), {CaMn$_{2}$Bi$_{2}$} (Ref.\citenum{Gibson15}), and {CaMn$_{2}$Sb$_{2}$} (Refs.\citenum{McNally15,Ratcliff09,Bridges09}). The reduced moment can be attributed to the strong spin-dependent hybridization between the Mn $3d$ and the As $4p$ orbitals as shown by density functional calculations \cite{An09} and to the expected quasi-two-dimensionality of the Mn--Mn spin interactions (see below). This is, however, different from the Fe-122 compounds where the Fe moment is greatly reduced due to the itinerant nature of the magnetism \cite{Johnston10,Stewart11,Si08}.
Based on the proposed domain configuration, we now calculate the expected ratio between $\chi_{ab}$ and $\chi_c$. If $\theta$ is the angle between the spin direction and the applied field, the susceptibility is given by $\chi_\theta = \chi_\parallel \mathrm{cos}^2\theta + \chi_\perp \mathrm{sin}^2\theta$, where $\chi_\parallel$ and $\chi_\perp$ are the susceptibilities parallel and perpendicular to the applied field \cite{Buschow}. For $\textbf{\textit{H}}\parallel \textbf{\textit{c}}$, $\theta = 90\degree$ for all the domains and thus $\chi_c = \chi_\perp$. To simplify the calculation for $\chi_{ab}$, we further assume $\textbf{\textit{H}}\parallel [100]$ for the in-plane susceptibility measurement. Then equal domain populations with $\theta = 0\degree$, $60\degree$ and $120\degree$ result in $\chi_{ab}=\chi_\perp/2$ at $T=0$ ($\chi_\parallel=0$ at $T=0$ for a collinear AFM). This result holds for any direction of applied field in the plane. Therefore, $\chi_{ab}/\chi_c=0.5$ at $T=0$ which is close to the experimental value of $0.58(1)$ determined from $\chi_{ab}$ and $\chi_c$ at $T = 2$~K in Fig.~\ref{Fig1}. The minor difference is probably due to a deviation from our assumption of equal domain populations.
In the following, we analyze the detailed temperature dependence of the magnetic contribution ($I_{\textrm M}$) to the Bragg peaks which is proportional to the square of the ordered moment. For $T$ close to $T_{\textrm N}$, it is predicted to have a power law behavior with a critical exponent $2\beta$ given by $I_{\textrm M}=I_0(1-T/T_{\textrm N})^{2\beta}$ (Ref. \citenum{Blundell}). Figures~\ref{Fig3}(a)--(c) show the power-law fit of the measured magnetic intensity for $T\ge 90$ K. The critical exponent $\beta$ was determined for six independent peaks as shown for the three peaks in Figs.~\ref{Fig3}(a)--(c) and was consistently found to be $0.21(3)$ which lies between the values expected for a three-dimensional (3D) Heisenberg spin system ($\beta_\textrm{3D, Heissenberg} = 0.36$) and that for a purely 2D Ising/XY system which predict critical exponents of $\beta_\textrm{2D, Ising}=0.125$ and $\beta_\textrm{2D, XY}=0.13$, respectively \cite{Blundell}. Thus {SrMn$_{2}$As$_{2}$} appears to behave like a quasi-2D AFM system which is also consistent with high-$T$ magnetic susceptibility measurements [$\chi (T \ge T_{\textrm N})$]\cite{Sangeetha16}.
This quasi-2D can be explained qualitatively from the magnetic and lattice structure. Within the unit cell, the two antiparallel Mn spins lie in different planes perpendicular to $\textbf{\textit{c}}$, forming a corrugated honeycomb layer (Fig.~\ref{Fig1} insets) with a nearest-neighbor distance of $3.06(1) ~\text{\AA}$. The distance between the nearest Mn atoms between the two layers is $6.05(1) ~\text{\AA}$ which is twice the nearest Mn-Mn distance within the layer. This difference is expected to result in a weaker interplanar exchange interaction compared to intraplanar interactions leading to a quasi-2D spin system. It has been rigorously shown by Mermin and Wagner \cite{Mermin66} that a 2D arrangement of spins cannot form a long-range magnetically-ordered state except at $T=0$. However, based on recent theoretical models including the 2D quantum Heisenberg model \cite{Chakravarty89} and a combination of numerical and renormalization group arguments \cite{Taroni08}, one obtains the window for the critical exponent $\beta$ for 2D/quasi-2D systems as $\sim0.1 \le \beta \le 0.25$. This has been successful in explaining a range of systems including {La$_2$CoO$_{4}$} with a {$T_{\textrm N}$} = 274.7(6) K and a critical exponent $\beta = 0.20(2)$ showing a crossover regime from 3D to 2D behavior \cite{Yamada89} similar to the system studied here.
Additionally, we have performed high-energy x-ray diffraction measurements to study the strength of the magnetoelastic coupling in {SrMn$_{2}$As$_{2}$} which would result in the distortion of the lattice as the system enters the ordered magnetic state and breaks the trigonal symmetry. Magnetically-induced lattice distortions have been observed in other pnictide systems like in {Ba(Fe$_{1-x}$Co$_{x}$)$_{2}$As$_{2}$} (Ref. \citenum{Nandi10}), {Ba(Fe$_{0.961}$Rh$_{0.039}$)$_{2}$As$_{2}$} (Ref. \citenum{Kreyssig10}) and many others. However, no lattice distortion was observed in {SrMn$_{2}$As$_{2}$} down to 6 K and an upper limit on any possible relative lattice distortion is $3 \times 10^{-4}$ as calculated from the experimental resolution.
In conclusion, we have shown that {SrMn$_{2}$As$_{2}$} orders in a collinear N\'{e}el AFM phase with the Mn spins aligned along the basal plane of the trigonal unit cell having a net ordered moment of $3.6(2)$ $\mu_{\textrm B}$/Mn which is smaller than the full high-spin value of $5.0$ $\mu_{\textrm B}$/Mn. The magnetic interaction is quasi-two-dimensional with a strong in-plane exchange interaction within the corrugated honeycomb Mn layer compared to weak interplanar interaction between the layers. Lastly, the magnetoelastic coupling does not cause a measurable lattice distortion in the ordered state.
Work at the Ames Laboratory was supported by the Department of Energy, Basic Energy Sciences, Division of Materials Sciences \& Engineering, under Contract No. DE-AC02-07CH11358. This research used resources at the University of Missouri Research Reactor. We are grateful to D. S. Robinson for support during the x-ray experiments. This research used resources of the Advanced Photon Source, a US Department of Energy (DOE) Office of Science User Facility operated for the DOE Office of Science by Argonne National Laboratory under Contract
No. DE-AC02-06CH11357.
|
\section{Introduction}
In the theory of noncommutative topology or geometry \cite{Conn}, a
generally noncommutative C*-algebra $\mathcal{A}$ is viewed as the algebra
C\left( X_{q}\right) $ of continuous functions on a virtual spatial object
X_{q}$, called a quantum space. Many interesting examples of quantum spaces
have been constructed with a topological or geometrical motivation, and
analyzed in comparison with their classical counterpart. Different
topological or geometrical viewpoints of the same object may give rise to
different quantum versions of quantum spaces. For example, quantum
odd-dimensional spheres and associated complex projective spaces have been
introduced and studied by Soibelman, Vaksman, Meyer, and others \cit
{VaSo,Me} as $\mathbb{S}_{q}^{2n+1}$ and $\mathbb{C}P_{q}^{n}$ via quantum
universal enveloping algebra approach, and by Hajac and his collaborators
including Baum, Kaygun, Matthes, Pask, Sims, Szyma\'{n}ski, Zieli\'{n}ski,
and others \cite{BaHaMaSz,HaMaSz,HaKaZi,HaPaSiZi} as $\mathbb{S}_{H}^{2n+1}$
and $\mathbb{P}^{n}\left( \mathcal{T}\right) $ via a multi-pullback and
Toeplitz algebra approach.
Recall that the concept of a vector bundle $E$ over a compact space $X$ can
be reformulated in the noncommutative context as a finitely generated
projective left module $\Gamma\left( E_{q}\right) $ over $C\left(
X_{q}\right) $, viewed as the space of continuous cross-sections of some
virtual noncommutative or quantum vector bundle $E_{q}$ over $X_{q}$, as
suggested by Swan's work \cite{Swan}. Based on the strong connection
approach to quantum principal bundles \cite{Haja:sc} for compact quantum
groups \cite{Woro,Wo:cm}, Hajac and his collaborators introduced quantum
line bundles $L_{k}$ of degree $k$ over $\mathbb{P}^{n}\left( \mathcal{T
\right) $ as some rank-one projective modules realized as spectral subspaces
$C\left( \mathbb{S}_{H}^{2n+1}\right) _{k}$ of $C\left( \mathbb{S
_{H}^{2n+1}\right) $ under a $U\left( 1\right) $-action, and analyzed them
via pairing of cyclic cohomology and K-theory \cite{HaMaSz,HaPaSiZi}. In
particular, it was found that $L_{k}$ is not stably free unless $k=0$,
revealing some information about the $K_{0}$-group of $C\left( \mathbb{P
^{n}\left( \mathcal{T}\right) \right) $. On the other hand, even for the
most crucial case of $n=1$, the $K_{0}$-group was not fully determined
despite the significant progress made in the 2003 paper \cite{HaMaSz}.
Going beyond the $K$-theoretic study of C*-algebras that classifies finitely
generated projective modules only up to stable isomorphism, some successes
have been achieved in the study of cancellation problem, made popular by
Rieffel \cite{Ri:dsr,Ri:ct}, that classifies finitely generated projective
modules up to isomorphism, for some quantum algebras \cit
{Ri:ct,Ri:pm,Sh:ct,Bach,Pete}. It is of interest and a natural question to
classify finitely generated projective modules over $C\left( \mathbb{P
^{n}\left( \mathcal{T}\right) \right) $ and identify the line bundles $L_{k}$
among them, beside finding the $K$-groups of $C\left( \mathbb{P}^{n}\left(
\mathcal{T}\right) \right) $.
In this paper, we use the powerful groupoid approach to C*-algebras
initiated by Renault \cite{Rena} and popularized by Curto, Muhly, and
Renault \cite{CuMu,MuRe} to realize $C\left( \mathbb{P}^{n}\left( \mathcal{T
\right) \right) $ as a concrete groupoid C*-algebra \cite{Rena}. Focusing on
the quantum complex line $\mathbb{P}^{1}\left( \mathcal{T}\right) $, we get
the C*-algebra structure of $C\left( \mathbb{P}^{1}\left( \mathcal{T}\right)
\right) $ analyzed and its $K$-groups computed. Furthermore, we get the
finitely generated left projective modules over $C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $ classified up to isomorphism by
classifying the projections over $C\left( \mathbb{P}^{1}\left( \mathcal{T
\right) \right) $, i.e. in $M_{\infty}\left( C\left( \mathbb{P}^{1}\left(
\mathcal{T}\right) \right) \right) $, up to unitary equivalence, and
explicitly identify the quantum line bundles $L_{k}$ among the classified
projections, showing that these modules $L_{k}$ do exhaust all rank-one
projections over $C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) $.
(After posting this work at ArXiv, a revised version of \cite{HaPaSiZi}
appeared at arXiv containing a computation of $K$-groups of $C\left( \mathbb
P}^{n}\left( \mathcal{T}\right) \right) $ for all $n$ \cite{HaNePaSiZi}.)
\section{Quantum projective spaces as groupoid C*-algebras}
Taking the groupoid approach to C*-algebras initiated by Renault \cite{Rena}
and popularized by the work of Curto, Muhly, and Renault \cite{CuMu,MuRe},
we give a description of the C*-algebras $C\left( \mathbb{S
_{H}^{2n-1}\right) $ and $C\left( \mathbb{P}^{n-1}\left( \mathcal{T}\right)
\right) $ of \cite{HaPaSiZi} as some concrete groupoid C*-algebras. We refer
to \cite{Rena,MuRe} for the concepts and theory of groupoid C*-algebras used
freely in the following discussion.
Let $\left. \left( \mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}}^{n}\right)
\right\vert _{\overline{\mathbb{Z}_{\geq}}^{n}}$ with $n>1$ be the
transformation group groupoid $\mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}}^{n}
$ restricted to the positive \textquotedblleft cone\textquotedblright\
\overline{\mathbb{Z}_{\geq}}^{n}$ where $\overline{\mathbb{Z}}:=\mathbb{Z
\cup\left\{ \infty\right\} $ carries the standard topology and $\mathbb{Z
^{n}$ acts on $\overline{\mathbb{Z}}^{n}$ componentwise in the canonical
way. From the invariant open subset $\mathbb{Z}_{\geq}^{n}$ of the unit
space $\overline{\mathbb{Z}_{\geq}}^{n}$ of $\left. \left( \mathbb{Z
^{n}\ltimes\overline{\mathbb{Z}}^{n}\right) \right\vert _{\overline{\mathbb{
}_{\geq}}^{n}}$, we get the short exact sequence of C*-algebra
\begin{equation*}
0\rightarrow C^{\ast}\left( \left. \left( \mathbb{Z}^{n}\ltimes \overline
\mathbb{Z}}^{n}\right) \right\vert _{\mathbb{Z}_{\geq}^{n}}\right)
\rightarrow C^{\ast}\left( \left. \left( \mathbb{Z}^{n}\ltimes \overline
\mathbb{Z}}^{n}\right) \right\vert _{\overline{\mathbb{Z}_{\geq}
^{n}}\right) \rightarrow C^{\ast}\left( \left. \left( \mathbb{Z}^{n}\ltime
\overline{\mathbb{Z}}^{n}\right) \right\vert _{\overline {\mathbb{Z}_{\geq}
^{n}\backslash\mathbb{Z}_{\geq}^{n}}\right) \rightarrow0
\end{equation*}
and furthermore since the open invariant set $\mathbb{Z}_{\geq}^{n}$ is
dense in the unit space, it induces a faithful representation $\pi$ of
C^{\ast }\left( \left. \left( \mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}
^{n}\right) \right\vert _{\overline{\mathbb{Z}_{\geq}}^{n}}\right) $ on
\ell^{2}\left( \mathbb{Z}_{\geq}^{n}\right) $. From the groupoid isomorphism
\begin{equation*}
\left. \left( \mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}}^{n}\right)
\right\vert _{\overline{\mathbb{Z}_{\geq}}^{n}}\ \cong\ \times^{n}\left.
\left( \mathbb{Z}\ltimes\overline{\mathbb{Z}}\right) \right\vert _{\overline
\mathbb{Z}_{\geq}}}
\end{equation*}
and the C*-algebra isomorphism $C^{\ast}\left( \left. \left( \mathbb{Z
\ltimes\overline{\mathbb{Z}}\right) \right\vert _{\overline{\mathbb{Z}_{\geq
}}}\right) \cong\mathcal{T}$ for the Toeplitz C*-algebra $\mathcal{T}$ with
C^{\ast}\left( \left. \left( \mathbb{Z}\ltimes\overline{\mathbb{Z}}\right)
\right\vert _{\mathbb{Z}_{\geq}}\right) \cong\mathcal{K}\left(
\ell^{2}\left( \mathbb{Z}_{\geq}\right) \right) $, we get
\begin{equation*}
C^{\ast}\left( \left. \left( \mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}
^{n}\right) \right\vert _{\overline{\mathbb{Z}_{\geq}}^{n}}\right)
\cong\otimes^{n}\mathcal{T}
\end{equation*}
with $C^{\ast}\left( \left. \left( \mathbb{Z}^{n}\ltimes\overline {\mathbb{Z
}^{n}\right) \right\vert _{\mathbb{Z}_{\geq}^{n}}\right) \cong\otimes^{n
\mathcal{K}\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right) \right) $.
Since $C\left( \mathbb{S}_{H}^{2n-1}\right) \cong\left( \otimes ^{n}\mathcal
T}\right) /\left( \otimes^{n}\mathcal{K}\right) $ by (A.2) of \cite{HaPaSiZi
, we have
\begin{equation*}
C\left( \mathbb{S}_{H}^{2n-1}\right) \cong C^{\ast}\left( \left. \left(
\mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}}^{n}\right) \right\vert _
\overline{\mathbb{Z}_{\geq}}^{n}}\right) \diagup C^{\ast}\left( \left.
\left( \mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}}^{n}\right) \right\vert _
\mathbb{Z}_{\geq}^{n}}\right) \cong C^{\ast}\left( \mathfrak{G}_{n}\right)
\end{equation*}
realized as the groupoid C*-algebra of the concrete groupoid
\begin{equation*}
\mathfrak{G}_{n}:=\left. \left( \mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}
^{n}\right) \right\vert _{\overline{\mathbb{Z}_{\geq}}^{n}\backslash \mathbb
Z}_{\geq}^{n}}\ \ .
\end{equation*}
Next we note that the $U\left( 1\right) $-action on $C\left( \mathbb{S
_{H}^{2n-1}\right) \cong C^{\ast}\left( \mathfrak{G}_{n}\right) $ considered
in \cite{HaPaSiZi} is implemented by the multiplication operator
\begin{equation*}
U_{\zeta}:f\in C_{c}\left( \mathfrak{G}_{n}\right) \mapsto h_{\zeta}f\in
C_{c}\left( \mathfrak{G}_{n}\right)
\end{equation*}
for $\zeta\in U\left( 1\right) \equiv\mathbb{T}$ where
\begin{equation*}
h_{\zeta}:\left( m,p\right) \in\mathfrak{G}_{n}\subset\mathbb{Z}^{n}\ltime
\overline{\mathbb{Z}}^{n}\mapsto\zeta^{\Sigma m}\in\mathbb{T}\text{\ \ with
\Sigma m:=\sum_{i=1}^{n}m_{i}
\end{equation*}
is a groupoid character of $\mathfrak{G}_{n}$ and hence $U_{\zeta}$ is an
automorphism of $C^{\ast}\left( \mathfrak{G}_{n}\right) $, since $h_{\zeta
}h_{\zeta^{\prime}}=h_{\zeta\zeta^{\prime}}$ and $U_{\zeta}\left(
w_{i}\right) =\zeta w_{i}$ for the generators $w_{i}$ of $C\left( \mathbb{S
_{H}^{2n-1}\right) $ \cite{HaPaSiZi} identified with the characteristic
function $\chi_{A_{i}}\in C_{c}\left( \mathfrak{G}_{n}\right) $ of
\begin{equation*}
A_{i}:=\left\{ \left( e_{i},p\right) :p\in\overline{\mathbb{Z}_{\geq}
^{n}\backslash\mathbb{Z}_{\geq}^{n}\right\} \subset\mathfrak{G}_{n}\subse
\mathbb{Z}^{n}\ltimes\overline{\mathbb{Z}}^{n}.
\end{equation*}
The C*-algebra $C\left( \mathbb{P}^{n-1}\left( \mathcal{T}\right) \right) $
of the quantum complex projective space studied in \cite{HaKaZi,HaPaSiZi} is
then isomorphic to the $U\left( 1\right) $-invariant C*-subalgebra $C^{\ast
}\left( \mathfrak{G}_{n}\right) ^{U\left( 1\right) }$ of $C^{\ast}\left(
\mathfrak{G}_{n}\right) $, which can be realized as the groupoid C*-algebra
C^{\ast}\left( \left( \mathfrak{G}_{n}\right) _{0}\right) $ of the
subgroupoid $\left( \mathfrak{G}_{n}\right) _{0}$ of $\mathfrak{G}_{n}$,
where
\begin{equation*}
\left( \mathfrak{G}_{n}\right) _{k}:=\left\{ \left( m,p\right) \in\mathfrak{
}_{n}:\Sigma m=k\right\}
\end{equation*}
for $k\in\mathbb{Z}$. Furthermore, $C^{\ast}\left( \mathfrak{G}_{n}\right) $
becomes a graded algebra $\oplus_{k\in\mathbb{Z}}\overline{C_{c}\left(
\left( \mathfrak{G}_{n}\right) _{k}\right) }$ with the component $\overline
C_{c}\left( \left( \mathfrak{G}_{n}\right) _{k}\right) }$ being the quantum
line bundle $C\left( \mathbb{S}_{H}^{2n-1}\right) _{k}$ \cit
{HaKaZi,HaPaSiZi} of degree $k$ over the quantum space $\mathbb{P
^{n-1}\left( \mathcal{T}\right) $.
As shown in \cite{HaPaSiZi}, the case of $\mathbb{P}^{n-1}\left( \mathcal{T
\right) $ with $n=2$ plays a crucially important role in the study of the
quantum line bundle $C\left( \mathbb{S}_{H}^{2n-1}\right) _{k}$ in general,
so we focus on the case of $\mathbb{P}^{1}\left( \mathcal{T}\right) $ in the
remaining part of this paper, while leaving the higher-dimensional cases to
a subsequent paper.
\section{$K$-groups of quantum projective line}
In the case of $n=2$, the groupoid $\mathcal{G}:=\mathfrak{G
_{2}\equiv\left. \left( \mathbb{Z}^{2}\ltimes\overline{\mathbb{Z}
^{2}\right) \right\vert _{\overline{\mathbb{Z}_{\geq}}^{2}\backslash\mathbb{
}_{\geq}^{2}}$ has the unit space
\begin{equation*}
\mathcal{G}^{\left( 0\right) }=\overline{\mathbb{Z}_{\geq}}^{2}\backslas
\mathbb{Z}_{\geq}^{2}=\left( \overline{\mathbb{Z}_{\geq}}\times\left\{
\infty\right\} \right) \cup\left( \left\{ \infty\right\} \times\overline
\mathbb{Z}_{\geq}}\right)
\end{equation*}
and consists of points $\left( m,l,p,q\right) $ with $\left( m,l\right) \i
\mathbb{Z}^{2}$ and $\left( p,q\right) \in\mathcal{G}^{\left( 0\right) }$
such that $\left( m+p,l+q\right) \in\mathcal{G}^{\left( 0\right) }$ where
m+\infty=\infty$ for any $m\in\mathbb{Z}$ is understood.
The previous discussion realizes the C*-algebra $C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $ of the quantum projective line
\mathbb{P}^{1}\left( \mathcal{T}\right) $ as a groupoid C*-algebra
C^{\ast}\left( \mathcal{G}_{0}\right) $ where the subgroupoid
\begin{equation*}
\mathcal{G}_{0}:=\left\{ \left( n,-n,p,q\right) :n\in\mathbb{Z}\text{ such
that }\left( p,q\right) ,\left( p+n,q-n\right) \in\mathcal{G}^{\left(
0\right) }\right\} \subset\mathcal{G}
\end{equation*}
shares the same unit space $\left( \mathcal{G}_{0}\right) ^{\left( 0\right)
}=\mathcal{G}^{\left( 0\right) }$ with $\mathcal{G}$.
Note that the open dense invariant subset $U:=\left( \mathbb{Z
_{\geq}\times\left\{ \infty\right\} \right) \sqcup\left( \left\{
\infty\right\} \times\mathbb{Z}_{\geq}\right) $ of $\mathcal{G}^{\left(
0\right) }$ consists of two disjoint free orbits $\mathbb{Z
_{\geq}\times\left\{ \infty\right\} $ and $\left\{ \infty\right\} \time
\mathbb{Z}_{\geq}$ of $\mathcal{G}_{0}$, from which we get a faithful
representation $\pi$ of $C^{\ast}\left( \mathcal{G}_{0}\right) \equiv
C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) $ on the Hilbert
space
\begin{equation*}
\mathcal{H}:=\ell^{2}\left( U\right) \cong\ell^{2}\left( \mathbb{Z}_{\geq
}\right) \oplus\ell^{2}\left( \mathbb{Z}_{\geq}\right)
\end{equation*}
such that $\pi\left( \delta_{\left( n,-n,p,q\right) }\right) $ for any
\left( n,-n,p,q\right) \in\mathcal{G}_{0}$ with $\left( p,q\right) \in U$ is
the partial isometry sending $\delta_{\left( p,q\right) }\in\ell ^{2}\left(
U\right) $ to $\delta_{\left( p+n,q-n\right) }\in\ell ^{2}\left( U\right) $
and all other $\delta_{\left( p^{\prime},q^{\prime }\right)
}\in\ell^{2}\left( U\right) $ to $0$.
The open subgroupoid
\begin{equation*}
\mathcal{G}_{0}|_{U}=\left\{ \left( n,-n,p,\infty\right) ,\left(
n,-n,\infty,q\right) :n\in\mathbb{Z}\text{ such that }p,q,p+n,q-n\in \mathbb
Z}_{\geq}\right\}
\end{equation*}
of $\mathcal{G}_{0}$ is isomorphic to the disjoint union $\mathfrak{K
_{+}\sqcup\mathfrak{K}_{-}$ of two copies of the groupoid $\mathfrak{K
:=\left. \left( \mathbb{Z}\ltimes\mathbb{Z}\right) \right\vert _{\mathbb{Z
_{\geq}}$ under the map $\left( n,-n,p,\infty\right) \mapsto\left(
n,p\right) \in\mathfrak{K}_{+}$ and $\left( n,-n,\infty ,q\right)
\mapsto\left( -n,q\right) \in\mathfrak{K}_{-}$. Thus
\begin{equation*}
C^{\ast}\left( \mathcal{G}_{0}|_{U}\right) \cong C^{\ast}\left( \mathfrak{K
_{+}\right) \oplus C^{\ast}\left( \mathfrak{K}_{-}\right) \cong\mathcal{K
\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right) \right) \oplus\mathcal{K
\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right) \right) .
\end{equation*}
With $\mathcal{G}^{\left( 0\right) }\backslash U=\left\{ \left(
\infty,\infty\right) \right\} $ and $\mathcal{G}_{0}|_{\left\{ \left(
\infty,\infty\right) \right\} }=\left\{ \left( n,-n,\infty,\infty\right)
:n\in\mathbb{Z}\right\} $ isomorphic to the group $\mathbb{Z}$, we have the
short exact sequence
\begin{equation*}
0\rightarrow C^{\ast}\left( \left. \mathcal{G}_{0}\right\vert _{U}\right)
\cong\mathcal{K}\oplus\mathcal{K}\overset{\iota}{\rightarrow}C^{\ast}\left(
\mathcal{G}_{0}\right) \overset{\sigma}{\rightarrow}C^{\ast}\left( \left.
\mathcal{G}_{0}\right\vert _{\left\{ \left( \infty,\infty\right) \right\}
}\right) \cong C\left( \mathbb{T}\right) \rightarrow0,
\end{equation*}
where $C^{\ast}\left( \left. \mathcal{G}_{0}\right\vert _{U}\right) \con
\mathcal{K}\oplus\mathcal{K}$ under the representation $\pi$ with $\mathcal{
}\equiv\mathcal{K}\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right) \right) $,
and $\delta_{\left( 1,-1,\infty,\infty\right) }\in C^{\ast }\left( \left.
\mathcal{G}_{0}\right\vert _{\left\{ \left( \infty ,\infty\right) \right\}
}\right) $ is identified with
\begin{equation*}
z:=\func{id}_{\mathbb{T}}\in C\left( \mathbb{T}\right) .
\end{equation*}
In the induced 6-term exact sequenc
\begin{equation*}
\begin{array}{rcccccl}
\mathbb{Z}\left[ e_{11}\right] \oplus\mathbb{Z}\left[ e_{11}\right] \cong &
K_{0}\left( \mathcal{K}\oplus\mathcal{K}\right) & \overset{K_{0}\left(
\iota\right) }{\rightarrow} & K_{0}\left( C^{\ast}\left( \mathcal{G
_{0}\right) \right) & \overset{K_{0}\left( \sigma\right) }{\rightarrow} &
K_{0}\left( C\left( \mathbb{T}\right) \right) & \cong\mathbb{Z} \\
& \uparrow\eta & & & & \downarrow\varepsilon & \\
\mathbb{Z}\left[ z\right] \cong & K_{1}\left( C\left( \mathbb{T}\right)
\right) & \overset{K_{1}\left( \sigma\right) }{\leftarrow} & K_{1}\left(
C^{\ast}\left( \mathcal{G}_{0}\right) \right) & \overset{K_{1}\left(
\iota\right) }{\leftarrow} & K_{1}\left( \mathcal{K}\oplus\mathcal{K}\right)
& =0
\end{array}
\end{equation*}
the homomorphism $K_{0}\left( \sigma\right) $ is clearly surjective, and we
claim that the index homomorphism $\eta$ sends $\left[ z\right] $ to $\left(
-\left[ e_{11}\right] \right) \oplus\left[ e_{11}\right] $, where $e_{11}$
is the standard matrix unit and $z\equiv\func{id}_{\mathbb{T}}\in
GL_{1}\left( C\left( \mathbb{T}\right) \right) $. Indeed $z$ lifts via
\sigma$ to the characteristic function $\chi_{W}\in C^{\ast }\left( \mathcal
G}_{0}\right) $ of the set
\begin{equation*}
W:=\left\{ \left( 1,-1,p,\infty\right) :p\geq0\right\} \cup\left\{ \left(
1,-1,\infty,q\right) :q\geq1\right\} ,
\end{equation*}
and $\pi\left( \chi_{W}\right) =\mathcal{S}\oplus\mathcal{S}^{\ast}$ a
partial isometry with kernel projection $0\oplus e_{11}$ and cokernel
projection $e_{11}\oplus0$, where $\mathcal{S}$ is the (forward) unilateral
shift. Hence
\begin{equation*}
\eta\left( \left[ z\right] \right) =\left[ 0\oplus e_{11}\right] -\left[
e_{11}\oplus0\right] \in K_{0}\left( \mathcal{K}\oplus \mathcal{K}\right) .
\end{equation*}
(It is understood that the index homomorphism $\eta$ used here may be
different by a $\pm$-sign from the one used by other authors.)
Thus we get $K_{0}\left( \iota\right) \left( \left[ 0\oplus e_{11}\right]
\left[ e_{11}\oplus0\right] \right) =0$ in $K_{0}\left( C^{\ast}\left(
\mathcal{G}_{0}\right) \right) $, and hence $\left[ e_{11}\oplus0\right]
\left[ 0\oplus e_{11}\right] $ in $K_{0}\left( C^{\ast}\left( \mathcal{G
_{0}\right) \right) $. A simple diagram chase concludes that $K_{0}\left(
C^{\ast}\left( \mathcal{G}_{0}\right) \right) \cong\mathbb{Z}\left[
e_{11}\oplus0\right] \oplus\mathbb{Z}\left[ \tilde {I}\right] $ for the
identity element $\tilde{I}$ of $C^{\ast}\left( \mathcal{G}_{0}\right) $,
while $K_{1}\left( C^{\ast}\left( \mathcal{G}_{0}\right) \right) =0$.
Furthermore
\begin{align*}
K_{0}\left( \iota\right) & :m\left[ e_{11}\right] \oplus l\left[ e_{11
\right] \in K_{0}\left( \mathcal{K}\oplus\mathcal{K}\right) \cong\mathbb{Z
\oplus\mathbb{Z} \\
& \mapsto\left( m+l\right) \left[ e_{11}\oplus0\right] \oplus0\left[ \tilde{
}\right] \in K_{0}\left( C^{\ast}\left( \mathcal{G}_{0}\right) \right) \con
\mathbb{Z}\oplus\mathbb{Z}.
\end{align*}
We summarize as follows.
\textbf{Theorem 1}. \textit{For the quantum complex projective space }
\mathbb{P}^{1}\left( \mathcal{T}\right) $\textit{, there is a short exact
sequence of C*-algebras decomposing its algebra }$C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $\textit{\ as
\begin{equation*}
0\rightarrow\mathcal{K}\oplus\mathcal{K}\overset{\iota}{\rightarrow}C\left(
\mathbb{P}^{1}\left( \mathcal{T}\right) \right) \overset{\sigma }{\rightarro
}C\left( \mathbb{T}\right) \rightarrow0,
\end{equation*}
\textit{and its }$K$\textit{-groups coincide with those of its classical
counterpart, i.e.
\begin{equation*}
K_{0}\left( C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) \right)
\cong\mathbb{Z}\oplus\mathbb{Z}\text{\ \ \textit{and}\ \ }K_{1}\left(
C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) \right) =0.
\end{equation*}
\section{Classification of projections over quantum projective line}
In the following, we denote by $M_{\infty}\left( \mathcal{A}\right) $ the
direct limit (or the union as sets) of the increasing sequence of matrix
algebras $M_{n}\left( \mathcal{A}\right) $ over $\mathcal{A}$ with the
canonical inclusion $M_{n}\left( \mathcal{A}\right) \subset M_{n+1}\left(
\mathcal{A}\right) $ identifying $x\in M_{n}\left( \mathcal{A}\right) $ with
$x\boxplus0\in M_{n+1}\left( \mathcal{A}\right) $ for any algebra $\mathcal{
}$, where $\boxplus$ denotes the standard diagonal concatenation of two
matrices. So the size of an element in $M_{\infty}\left( \mathcal{A}\right) $
can be taken arbitrarily large. We also use $U_{\infty}\left( \mathcal{A
\right) $ to denote the direct limit of the unitary groups $U_{n}\left(
\mathcal{A}\right) \subset M_{n}\left( \mathcal{A}\right) $ for a unital
C*-algebra $\mathcal{A}$ with $U_{n}\left( \mathcal{A}\right) $ embedded in
U_{n+1}\left( \mathcal{A}\right) $ by identifying $x\in U_{n}\left( \mathcal
A}\right) $ with $x\boxplus1\in U_{n+1}\left( \mathcal{A}\right) $.
Before proceeding with the classification problem, we briefly recall the
relation between projections over a C*-algebra $\mathcal{A}$ and finitely
generated left projective modules over $\mathcal{A}$, and between them and $K
$-theory.
By a projection over a unital C*-algebra $\mathcal{A}$, we mean a
self-adjoint idempotent in $M_{\infty}\left( \mathcal{A}\right) $. Two
projections $P,Q\in M_{n}\left( \mathcal{A}\right) $ are called unitarily
equivalent if there exists a unitary $U\in M_{N}\left( \mathcal{A}\right) $
with $N\geq n$ such that $UPU^{-1}=Q$. Each projection $P\in M_{n}\left(
\mathcal{A}\right) $ over $\mathcal{A}$ defines a finitely generated left
projective module $\mathcal{A}^{n}P$ over $\mathcal{A}$ where elements of
\mathcal{A}^{n}$ are viewed as row vectors. The mapping $P\mapsto\mathcal{A
^{n}P$ induces a bijective correspondence between the unitary equivalence
classes of projections over $\mathcal{A}$ and the isomorphism classes of
finitely generated left projective modules over $\mathcal{A}$ \cite{Blac}.
Two finitely generated projective left modules $E,F$ over $\mathcal{A}$ are
called stably isomorphic if they become isomorphic after being augmented by
the same finitely generated free $\mathcal{A}$-module, i.e. $E\oplus
\mathcal{A}^{k}\cong F\oplus\mathcal{A}^{k}$ for some $k\geq0$.
Correspondingly, two projections $P$ and $Q$ are called stably equivalent if
$P\boxplus I_{k}$ and $Q\boxplus I_{k}$ are unitarily equivalent for some
identity matrix $I_{k}$. The $K_{0}$-group $K_{0}\left( \mathcal{A}\right) $
classifies projections over $\mathcal{A}$ up to stable equivalence. The
classification of projections over a C*-algebra up to unitary equivalence,
appearing as the cancellation problem, was popularized by Rieffel's
pioneering work \cite{Ri:dsr,Ri:ct} and is in general an interesting but
difficult question.
The set of all unitary equivalence classes of projections over a C*-algebra
\mathcal{A}$ is an abelian monoid $\mathfrak{P}\left( \mathcal{A}\right) $
with its binary operation provided by the diagonal sum $\boxplus$ of
projections. The image of the canonical homomorphism from $\mathfrak{P
\left( \mathcal{A}\right) $ into $K_{0}\left( \mathcal{A}\right) $ is the
so-called positive cone of $K_{0}\left( \mathcal{A}\right) $.
In the following, we use $\tilde{I}$ to denote the multiplicative unit of
the unital C*-algebra $\left( \mathcal{K}\oplus\mathcal{K}\right)
^{+}\subset C^{\ast}\left( \mathcal{G}_{0}\right) $ where $\mathcal{A}^{+}$
denotes the unitization of $\mathcal{A}$, and $\tilde{I}_{n}$ to denote the
identity matrix in $M_{n}\left( \left( \mathcal{K}\oplus\mathcal{K}\right)
^{+}\right) $, while
\begin{equation*}
P_{m}:=\sum_{i=1}^{m}e_{ii}\in M_{m}\left( \mathbb{C}\right) \subset
\mathcal{K}
\end{equation*}
denotes the standard $m\times m$ identity matrix in $M_{m}\left( \mathbb{C
\right) \subset\mathcal{K}$ for any integer $m\geq0$ (with $M_{0}\left(
\mathbb{C}\right) \equiv\left\{ 0\right\} $ and $P_{0}\equiv0$ understood).
We also use the notation
\begin{equation*}
P_{-m}:=I-P_{m}\in\mathcal{K}^{+}
\end{equation*}
for integers $m>0$, where $I$ is the identity operator canonically contained
in $\mathcal{K}^{+}$, and symbolically adopt the notation
\begin{equation*}
P_{-0}\equiv I-P_{0}=I\neq P_{0}.
\end{equation*}
Furthermore by abuse of notation, we take
\begin{equation*}
P_{-m}\oplus P_{-l}:=\tilde{I}-\left( P_{m}\oplus P_{l}\right) \in\left(
\mathcal{K}\oplus\mathcal{K}\right) ^{+}
\end{equation*}
if $m,l\geq0$. Note that $P_{m}\oplus P_{l}\notin\left( \mathcal{K}\oplu
\mathcal{K}\right) ^{+}$ if $m$ and $l$ are of strictly opposite $\pm$-sings.
Let $\alpha\in M_{\infty}\left( C^{\ast}\left( \mathcal{G}_{0}\right)
\right) $ be a projection. Since projections in $M_{\infty}\left( C\left(
\mathbb{T}\right) \right) $ are classified up to unitary equivalence as the
constant functions on $\mathbb{T}$ with an identity matrix $I_{n}\in
M_{n}\left( \mathbb{C}\right) $ as the value for some $n\in\mathbb{Z}_{\geq }
$ (and hence $K_{0}\left( C\left( \mathbb{T}\right) \right) =\mathbb{Z}$),
\alpha$ is unitarily equivalent over $C^{\ast}\left( \mathcal{G}_{0}\right) $
to some projection $\beta\in M_{N}\left( C^{\ast}\left( \mathcal{G
_{0}\right) \right) $ with $\sigma\left( \beta\right) =I_{n}$ for some
n\geq0$ and a suitably large size $N\geq n$. It is easy to see that $n$
depends only on $\alpha$, and we call $n$ the rank of $\alpha$.
So in the following, we concentrate on classifying projections $\alpha$ over
$C^{\ast}\left( \mathcal{G}_{0}\right) $ with $\sigma\left( \alpha\right)
=I_{n}$ and $\alpha\in M_{N}\left( C^{\ast}\left( \mathcal{G}_{0}\right)
\right) $ for some $N\geq n$.
Now since $\sigma\left( \alpha-\tilde{I}_{n}\right) =I_{n}-I_{n}=0$,
\begin{equation*}
\alpha-\tilde{I}_{n}\in M_{N}\left( \mathcal{K}\oplus\mathcal{K}\right)
\equiv M_{N}\left( \mathcal{K}\right) \oplus M_{N}\left( \mathcal{K}\right)
\end{equation*}
which can be approximated by elements in $M_{N}\left( M_{k}\left( \mathbb{C
\right) \right) \oplus M_{N}\left( M_{k}\left( \mathbb{C}\right) \right) $.
So we can replace $\alpha$ by a unitarily equivalent projection $\tilde{I
_{n}+x$ for some $x$ in
\begin{equation*}
M_{N}\left( M_{k}\left( \mathbb{C}\right) \right) \oplus M_{N}\left(
M_{k}\left( \mathbb{C}\right) \right) \subset M_{N}\left( \pi\left(
C_{c}\left( \left. \mathcal{G}_{0}\right\vert _{U}\right) \right) \right)
\subset M_{n}\left( \left( \mathcal{K}\oplus\mathcal{K}\right) ^{+}\right)
\end{equation*}
with a suitably large $k$. Let $I_{n}^{\prime}$ be the identity element of
\begin{equation*}
M_{n}\left( M_{k}\left( \mathbb{C}\right) \right) \oplus M_{n}\left(
M_{k}\left( \mathbb{C}\right) \right) \subset M_{n}\left( \left( \mathcal{K
\oplus\mathcal{K}\right) ^{+}\right) .
\end{equation*}
Then since $I_{n}^{\prime}+x\in M_{N}\left( M_{k}\left( \mathbb{C}\right)
\right) \oplus M_{N}\left( M_{k}\left( \mathbb{C}\right) \right) $ is
unitarily equivalent over $M_{k}\left( \mathbb{C}\right) \oplus M_{k}\left(
\mathbb{C}\right) $ to $P_{m}\oplus P_{l}$ for some $0\leq m,l\leq Nk$ where
$P_{m}\oplus P_{l}$ is the identity element of
\begin{equation*}
M_{m}\left( \mathbb{C}\right) \oplus M_{l}\left( \mathbb{C}\right) \subset
M_{Nk}\left( \mathbb{C}\right) \oplus M_{Nk}\left( \mathbb{C}\right) \equiv
M_{N}\left( M_{k}\left( \mathbb{C}\right) \right) \oplus M_{N}\left(
M_{k}\left( \mathbb{C}\right) \right) ,
\end{equation*}
we have $\tilde{I}_{n}+x$ unitarily equivalent over $\left( M_{k}\left(
\mathbb{C}\right) \oplus M_{k}\left( \mathbb{C}\right) \right)
^{+}\subset\pi\left( C_{c}\left( \left. \mathcal{G}_{0}\right\vert
_{U}\right) \right) ^{+}$ to
\begin{equation*}
\left( \tilde{I}_{n}-I_{n}^{\prime}\right) +\left( P_{m}\oplus P_{l}\right)
\in M_{N}\left( \left( M_{k}\left( \mathbb{C}\right) \oplus M_{k}\left(
\mathbb{C}\right) \right) ^{+}\right) \subset M_{N}\left( \left( \mathcal{K
\oplus\mathcal{K}\right) ^{+}\right)
\end{equation*}
by the canonical embedding of $U_{\infty}\left( \mathcal{A}\right) $ in
U_{\infty}\left( \mathcal{A}^{+}\right) $ for any unital C*-algebra
\mathcal{A}$. Note that $\left( \tilde{I}_{n}-I_{n}^{\prime}\right) +\left(
P_{m}\oplus P_{l}\right) $ can be expressed in the form
\begin{equation*}
\text{(*)\ \ }\left( P_{m_{1}}\oplus P_{l_{1}}\right) \boxplus\cdots
\boxplus\left( P_{m_{N}}\oplus P_{l_{N}}\right) \in M_{N}\left( \left(
M_{k}\left( \mathbb{C}\right) \oplus M_{k}\left( \mathbb{C}\right) \right)
^{+}\right)
\end{equation*}
for some $m_{i},l_{i}\in\mathbb{Z}$ with $\left\vert m_{i}\right\vert
,\left\vert l_{i}\right\vert \leq k$, and since $\sigma\left( \alpha\right)
=\sigma\left( \tilde{I}_{n}\right) =I_{n}$, we have $m_{i},l_{i}\leq0$
(viewing $P_{m_{i}}\oplus P_{l_{1}}$ as $\tilde{I}-\left( P_{\left\vert
m_{i}\right\vert }\oplus P_{\left\vert l_{1}\right\vert }\right) $) for
i\leq n$ and $m_{i},l_{i}\geq0$ for $i>n$.
It remains to classify projections $\alpha\in M_{N}\left( C^{\ast}\left(
\mathcal{G}_{0}\right) \right) $ of the form (*) up to unitary equivalence
over $C^{\ast}\left( \mathcal{G}_{0}\right) $.
When the rank $n$ is $0$, we have $m_{i},l_{i}\geq0$ for all $i$ in (*), and
$P_{m_{i}},P_{l_{i}}\in M_{k}\left( \mathbb{C}\right) $. With
P_{m_{i}},P_{l_{i}}$ viewed as elements in $M_{Nk}\left( \mathbb{C}\right)
\supset M_{k}\left( \mathbb{C}\right) $, the projections $P_{m_{1}}\boxplus
\cdots\boxplus P_{m_{N}}$ and $P_{l_{1}}\boxplus\cdots\boxplus P_{l_{N}}$
lying in $M_{N}\left( M_{Nk}\left( \mathbb{C}\right) \right) $ with ranks
bounded by $Nk$ are unitarily equivalent over $M_{Nk}\left( \mathbb{C
\right) $ to $P_{m}\boxplus0\boxplus\cdots\boxplus0$ and
P_{l}\boxplus0\boxplus\cdots\boxplus0$ in $M_{N}\left( M_{Nk}\left( \mathbb{
}\right) \right) $ respectively, where $m:=\sum_{i}m_{i}\leq Nk$ and
l:=\sum_{i}l_{i}\leq Nk$. Hence $\left( P_{m_{1}}\oplus P_{l_{1}}\right)
\boxplus\cdots\boxplus\left( P_{m_{N}}\oplus P_{l_{N}}\right) $ is unitarily
equivalent over $\pi\left( C_{c}\left( \left. \mathcal{G}_{0}\right\vert
_{U}\right) \right) ^{+}$ to $P_{m}\oplus P_{l}\in M_{1}\left( \left(
\mathcal{K}\oplus\mathcal{K}\right) ^{+}\right) \equiv\left( \mathcal{K
\oplus\mathcal{K}\right) ^{+}$.
On the other hand, if such projections $P_{m}\oplus P_{l}$ and
P_{m^{\prime}}\oplus P_{l^{\prime}}$ with $m,l,m^{\prime},l^{\prime}\i
\mathbb{Z}_{\geq}$ are unitarily equivalent over $C^{\ast}\left( \mathcal{G
_{0}\right) \subset\mathcal{B}\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right)
\right) \oplus\mathcal{B}\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right)
\right) $, then their ranks must coincide, i.e. $m=m^{\prime}$ and
l=l^{\prime}$. Thus for the case of $n=0$, we get unitary equivalence
classes of projections over $C^{\ast}\left( \mathcal{G}_{0}\right) $
classified by $\left( m,l\right) \in\mathbb{Z}_{\geq}\times\mathbb{Z}_{\geq}$
as $P_{m}\oplus P_{l}$.
When the rank $n$ is strictly positive, we claim that the projections
\tilde{I}_{n}$, $\tilde{I}_{n}\boxplus\left( P_{m}\oplus P_{0}\right) $, and
\begin{equation*}
\tilde{I}_{n-1}\boxplus\left( P_{-m}\oplus P_{-0}\right) \equiv\tilde {I
_{n-1}\boxplus\left( \tilde{I}-\left( P_{m}\oplus0\right) \right)
\end{equation*}
with $m\in\mathbb{N}\equiv\mathbb{Z}_{>}$, give a complete list of unitary
equivalence classes of projections $\alpha$ with $\sigma\left( \alpha\right)
=I_{n}$.
First we observe that for $m,l,n\geq0$.
\begin{equation*}
\left[ \tilde{I}_{n}\boxplus\left( P_{m}\oplus P_{l}\right) \right] =\left[
\tilde{I}_{n}\right] +K_{0}\left( \iota\right) \left( m\left[ e_{11}\right]
\oplus l\left[ e_{11}\right] \right) =\left( m+l\right) \left[ e_{11}\oplus
\right] \oplus n\left[ \tilde{I}\right]
\end{equation*}
and
\begin{equation*}
\left[ \tilde{I}_{n}\boxplus\left( P_{-m}\oplus P_{-l}\right) \right] =\left[
\tilde{I}_{n}\boxplus\left( \tilde{I}-\left( P_{m}\oplus P_{l}\right)
\right) \right] =\left[ \tilde{I}_{n}\right] +\left[ \tilde{I}-\left(
P_{m}\oplus P_{l}\right) \right]
\end{equation*
\begin{equation*}
=\left[ \tilde{I}_{n}\right] +\left[ \tilde{I}\right] -\left[ P_{m}\oplus
P_{l}\right] =\left( n+1\right) \left[ \tilde{I}\right] -K_{0}\left(
\iota\right) \left( m\left[ e_{11}\right] \oplus l\left[ e_{11}\right]
\right)
\end{equation*
\begin{equation*}
=-\left( m+l\right) \left[ e_{11}\oplus0\right] \oplus\left( n+1\right)
\left[ \tilde{I}\right]
\end{equation*}
in $K_{0}\left( C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right)
\right) $. So the stable equivalence class over $C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $ of a projection of the form $\tilde{
}_{n}\boxplus\left( P_{m}\oplus P_{l}\right) $ with $n,ml\geq0$ (so $m,l$
are integers not of opposite $\pm$-signs) is determined exactly by $m+l\i
\mathbb{Z}$ and its rank ($n$ or $n+1$). In particular, for $n>0$, the
projections $\tilde{I}_{n}$, $\tilde{I}_{n}\boxplus\left( P_{m}\oplus
P_{0}\right) $, and $\tilde{I}_{n-1}\boxplus\left( P_{-m}\oplus
P_{-0}\right) $ with $m>0$ are mutually stably and hence unitarily
inequivalent.
It remains to show that any projection $\alpha$ of the form (*) with $n>0$
is unitarily equivalent to one of $\tilde{I}_{n}$, $\tilde{I
_{n}\boxplus\left( P_{m}\oplus P_{0}\right) $, and $\tilde{I
_{n-1}\boxplus\left( P_{-m}\oplus P_{-0}\right) $ with $m\in\mathbb{N}$.
Recall that $\sigma\left( \alpha\right) =I_{n}$ implies that
m_{i},l_{i}\leq0$ for $i\leq n$ and $m_{i},l_{i}\geq0$ for $i>n$. Since
\mathcal{K}\oplus\mathcal{K}\subset C\left( \mathbb{P}^{1}\left( \mathcal{T
\right) \right) $, using some (unitary) finite permutation matrices, we can
convert $\alpha$ to a unitarily equivalent projection $\beta$ of the for
\begin{equation*}
\beta=\left( \boxplus^{n-1}\tilde{I}\right) \boxplus\left( P_{m^{\prime
\prime}}\oplus P_{l^{\prime\prime}}\right) \boxplus\left(
P_{m^{\prime}}\oplus P_{l^{\prime}}\right) \boxplus\left(
\boxplus^{N-n-1}0\right) \in M_{N}\left( \left( \mathcal{K}\oplus\mathcal{K
\right) ^{+}\right)
\end{equation*}
with $m^{\prime\prime}=\sum_{i=1}^{n}m_{i}\leq0$, $l^{\prime\prime}=\sum
_{i=1}^{n}l_{i}\leq0$, $m^{\prime}=\sum_{i=n+1}^{N}m_{i}\geq0$, $l^{\prime
}=\sum_{i=n+1}^{N}l_{i}\geq0$, or for short
\begin{equation*}
\beta=\tilde{I}_{n-1}\boxplus\left( P_{m^{\prime\prime}}\oplus P_{l^{\prime
\prime}}\right) \boxplus\left( P_{m^{\prime}}\oplus P_{l^{\prime}}\right)
\in M_{n+1}\left( \left( \mathcal{K}\oplus\mathcal{K}\right) ^{+}\right) ,
\end{equation*}
by swapping the largest (finite) identity diagonal blocks in $P_{m_{i}}$ and
$P_{l_{i}}$ for $i>n+1$ with suitable disjoint diagonal zero blocks of
P_{m_{n+1}}$ and $P_{l_{n+1}}$ respectively, and by swapping the largest
(finite) diagonal zero blocks in $P_{m_{i}}$ and $P_{l_{i}}$ for $i<n$ with
suitable disjoint diagonal identity blocks of $P_{m_{n}}$ and $P_{l_{n}}$
respectively. Here it is understood that $m^{\prime\prime}$ and $l^{\prime
\prime}$ carry a negative sign and hence $P_{m^{\prime\prime}}\oplus
P_{l^{\prime\prime}}=\tilde{I}-\left( P_{\left\vert m^{\prime\prime
}\right\vert }\oplus P_{\left\vert l^{\prime\prime}\right\vert }\right) $.
By swapping a suitable (finite) diagonal zero block of $P_{l^{\prime\prime}}$
with a suitable (finite) identity block of $P_{l^{\prime}}$, we get $\beta$
unitarily equivalent to either
\begin{equation*}
\tilde{I}_{n-1}\boxplus\left( P_{m^{\prime\prime}}\oplus P_{l^{\prime\prime
}+l^{\prime}}\right) \boxplus\left( P_{m^{\prime}}\oplus0\right)
\end{equation*}
if $l^{\prime\prime}+l^{\prime}<0$, or to
\begin{equation*}
\tilde{I}_{n-1}\boxplus\left( P_{m^{\prime\prime}}\oplus P_{-0}\right)
\boxplus\left( P_{m^{\prime}}\oplus P_{l^{\prime}+l^{\prime\prime}}\right)
\end{equation*}
if $l^{\prime\prime}+l^{\prime}\geq0$.
With $\pi\left( \chi_{W}\right) =\mathcal{S}\oplus\mathcal{S}^{\ast}$ as
discussed earlier, conjugating $\tilde{I}_{n-1}\boxplus\left( P_{m^{\prime
\prime}}\oplus P_{l^{\prime\prime}+l^{\prime}}\right) \boxplus\left(
P_{m^{\prime}}\oplus0\right) $ or $\tilde{I}_{n-1}\boxplus\left(
P_{m^{\prime\prime}}\oplus P_{-0}\right) \boxplus\left( P_{m^{\prime}}\oplus
P_{l^{\prime}+l^{\prime\prime}}\right) $ by the unitary
\begin{equation*}
\tilde{I}_{n-1}\boxplus\left(
\begin{array}{cc}
\pi\left( \chi_{W}\right) ^{\left\vert l^{\prime\prime}+l^{\prime
}\right\vert } & \tilde{I}-\pi\left( \chi_{W}\right) ^{\left\vert
l^{\prime\prime}+l^{\prime}\right\vert }\left( \pi\left( \chi_{W}\right)
^{\ast}\right) ^{\left\vert l^{\prime\prime}+l^{\prime}\right\vert } \\
\tilde{I}-\left( \pi\left( \chi_{W}\right) ^{\ast}\right) ^{\left\vert
l^{\prime\prime}+l^{\prime}\right\vert }\pi\left( \chi_{W}\right)
^{\left\vert l^{\prime\prime}+l^{\prime}\right\vert } & \left( \pi\left(
\chi_{W}\right) ^{\ast}\right) ^{\left\vert l^{\prime\prime}+l^{\prime
}\right\vert
\end{array}
\right) \in M_{n+1}\left( C\left( \mathbb{P}^{1}\left( \mathcal{T}\right)
\right) \right)
\end{equation*}
or its adjoint respectively converts each to the form
\begin{equation*}
\gamma:=\tilde{I}_{n-1}\boxplus\left( P_{-j}\oplus P_{-0}\right)
\boxplus\left( P_{k}\oplus0\right)
\end{equation*}
for some $j,k\geq0$ (up to swap of finite diagonal blocks in the first
\oplus$-summand).
Finally by swapping a suitable (finite) diagonal zero block of $P_{-j}$ with
a suitable (finite) identity block of $P_{k}$, we get $\gamma$ unitarily
equivalent to either $\tilde{I}_{n-1}\boxplus\left( P_{k-j}\oplus
P_{-0}\right) \boxplus0$ if $k-j<0$, or $\tilde{I}_{n}\boxplus\left(
P_{k-j}\oplus0\right) $ if $k-j\geq0$. Thus $\alpha$ is unitarily equivalent
over $C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) $ to either
\tilde{I}_{n-1}\boxplus\left( \tilde{I}-\left( P_{j-k}\oplus0\right) \right)
$ if $k-j<0$, or $\tilde{I}_{n}\boxplus\left( P_{k-j}\oplus0\right) $ if
k-j\geq0$ as wanted.
Now we summarize what we have found.
\textbf{Theorem 2}. \textit{The abelian monoid }$\mathfrak{P}\left( C\left(
\mathbb{P}^{1}\left( \mathcal{T}\right) \right) \right) $\textit{\ of
unitary equivalence classes of projections over }$C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $\textit{\ consists of (the
representatives) }$P_{m}\oplus P_{l}$\textit{, }$\tilde{I}_{n}\boxplus\left(
P_{j}\oplus0\right) $\textit{, and }$\tilde{I}_{n-1}\boxplus\left( \tilde {I
-\left( P_{k}\oplus0\right) \right) $\textit{\ for }$m,l,j\in \mathbb{Z
_{\geq}$\textit{\ and }$n,k\in\mathbb{Z}_{>}$\textit{, where }$\tilde{I}
\textit{\ is the identity of }$\left( \mathcal{K}\oplus \mathcal{K}\right)
^{+}\subset C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) $\textit
\ and }$P_{k}$\textit{\ is the identity element of }$M_{k}\left( \mathbb{C
\right) \subset\mathcal{K}$\textit{, with its binary operation }$\cdot
\textit{\ specified by
\begin{equation*}
\left\{
\begin{array}{lll}
\left( P_{m}\oplus P_{l}\right) \cdot\left( \tilde{I}_{n}\boxplus\left(
P_{j}\oplus0\right) \right) =\tilde{I}_{n}\boxplus\left(
P_{m+l+j}\oplus0\right) , & & \\
\left( \tilde{I}_{n}\boxplus\left( P_{j}\oplus0\right) \right) \cdot\left(
\tilde{I}_{n^{\prime}-1}\boxplus\left( \tilde{I}-\left( P_{k}\oplus0\right)
\right) \right) =\tilde{I}_{n+n^{\prime}}\boxplus\left( P_{j-k}\oplus0\right)
& \text{\textit{if }} & j\geq k, \\
\left( \tilde{I}_{n}\boxplus\left( P_{j}\oplus0\right) \right) \cdot\left(
\tilde{I}_{n^{\prime}-1}\boxplus\left( \tilde{I}-\left( P_{k}\oplus0\right)
\right) \right) =\tilde{I}_{n+n^{\prime}-1}\boxplus\left( \tilde{I}-\left(
P_{k-j}\oplus0\right) \right) & \text{\textit{if }} & j<k, \\
\left( P_{m}\oplus P_{l}\right) \cdot\left( \tilde{I}_{n-1}\boxplus\left(
\tilde{I}-\left( P_{k}\oplus0\right) \right) \right) =\tilde{I
_{n}\boxplus\left( P_{m+l-k}\oplus0\right) & \text{\textit{if }} & m+l\geq k,
\\
\left( P_{m}\oplus P_{l}\right) \cdot\left( \tilde{I}_{n-1}\boxplus\left(
\tilde{I}-\left( P_{k}\oplus0\right) \right) \right) =\tilde{I
_{n-1}\boxplus\left( \tilde{I}-\left( P_{k-m-l}\oplus0\right) \right) &
\text{\textit{if }} & m+l<k
\end{array}
\right.
\end{equation*}
\textit{for representatives of different types and by adding up
corresponding indices }$m,l,j,n,k$ involved\textit{\ for representatives of
the same type}.
\textbf{Corollary 1}. \textit{The cancellation law holds for projections }
\alpha$\textit{\ over }$C\left( \mathbb{P}^{1}\left( \mathcal{T}\right)
\right) $\textit{\ of rank }$n\geq1$\textit{\ where }$n$\textit{\ is the
rank of the projection }$\sigma\left( \alpha\right) \in M_{\infty}\left(
C\left( \mathbb{T}\right) \right) $\textit{\ at any point of }$T$\textit{,
but fails for projections }$\alpha$\textit{\ over }$C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $\textit{\ of rank} $0$.
We also get the following details about the positive cone of $K_{0}\left(
C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) \right) $, extending
the information provided by Corollary 3.4 of \cite{HaMaSz}.
\textbf{Corollary 2}. \textit{The positive cone of }$K_{0}\left( C\left(
\mathbb{P}^{1}\left( \mathcal{T}\right) \right) \right) \cong\mathbb{Z}\left[
e_{11}\oplus0\right] \oplus\mathbb{Z}\left[ \tilde {I}\right] $\textit{\ is
\begin{equation*}
\left( \mathbb{Z}_{\geq}\left[ e_{11}\oplus0\right] \oplus0\left[ \tilde{I
\right] \right) \cup\left( \mathbb{Z}\left[ e_{11}\oplus0\right] \oplu
\mathbb{Z}_{>}\left[ \tilde{I}\right] \right) .
\end{equation*}
\textit{The canonical homomorphism from the monoid }$\mathfrak{P}\left(
C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) \right) $\textit{\
to }$K_{0}\left( C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right)
\right) $\textit{\ sends
\begin{equation*}
\left\{
\begin{array}{lll}
P_{m}\oplus P_{l} & \mapsto & \left( m+l\right) \left[ e_{11}\oplus0\right]
\\
\tilde{I}_{n}\boxplus\left( P_{j}\oplus0\right) & \mapsto & j\left[
e_{11}\oplus0\right] \oplus n\left[ \tilde{I}\right] \\
\tilde{I}_{n-1}\boxplus\left( \tilde{I}-\left( P_{k}\oplus0\right) \right) &
\mapsto & -k\left[ e_{11}\oplus0\right] \oplus n\left[ \tilde{I}\right
\end{array}
\right.
\end{equation*}
\textit{for }$m,l,j\in\mathbb{Z}_{\geq}$\textit{\ and} $n,k\in\mathbb{Z}_{>}
.
We briefly compare the quantum complex projective space $\mathbb{P
^{1}\left( \mathcal{T}\right) $ with the Podle\'{s} quantum sphere $\mathbb{
}_{\mu c}^{2}$ for $\mu\in\left( -1,1\right) $ and $c>0$, using the groupoid
approach.
By the description of the structure of $C\left( \mathbb{S}_{\mu
c}^{2}\right) $ in \cite{Sh:qpsu}, $C\left( \mathbb{S}_{\mu c}^{2}\right) $
can be realized as the groupoid C*-algebra $C^{\ast}\left( \mathcal{F
\right) $ of the subgroupoid
\begin{equation*}
\mathcal{F}:=\left\{ \left( n,n,p,q\right) :n\in\mathbb{Z}\text{ such that
\ \left( p,q\right) ,\left( p+n,q+n\right) \in\mathcal{G}^{\left( 0\right)
}\right\}
\end{equation*}
of $\mathcal{G}$, sharing the same unit space $\mathcal{F}^{\left( 0\right)
}=\mathcal{G}^{\left( 0\right) }$ with $\mathcal{G}$.
The open subgroupoid
\begin{equation*}
\mathcal{F}|_{U}=\left\{ \left( n,n,p,\infty\right) ,\left( n,n,\infty
,q\right) :n\in\mathbb{Z}\text{ such that }p,q,p+n,q+n\in\mathbb{Z}_{\geq
}\right\}
\end{equation*}
of $\mathcal{F}$ is isomorphic to the disjoint union $\mathfrak{K}_{+}\sqcu
\mathfrak{K}_{-}$ of two copies of the groupoid $\mathfrak{K}:=\left. \left(
\mathbb{Z}\ltimes\mathbb{Z}\right) \right\vert _{\mathbb{Z}_{\geq}}$ under
the map $\left( n,n,p,\infty\right) \mapsto\left( n,p\right) \in\mathfrak{K
_{+}$ and $\left( n,n,\infty,q\right) \mapsto\left( n,q\right) \in\mathfrak{
}_{-}$. Thus
\begin{equation*}
C^{\ast}\left( \mathcal{F}|_{U}\right) \cong C^{\ast}\left( \mathfrak{K
_{+}\right) \oplus C^{\ast}\left( \mathfrak{K}_{-}\right) \cong\mathcal{K
\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right) \right) \oplus\mathcal{K
\left( \ell^{2}\left( \mathbb{Z}_{\geq}\right) \right) .
\end{equation*}
With $\mathcal{F}^{\left( 0\right) }\backslash U=\left\{ \left(
\infty,\infty\right) \right\} $ and $\mathcal{F}|_{\left\{ \left(
\infty,\infty\right) \right\} }=\left\{ \left( n,n,\infty,\infty\right) :n\i
\mathbb{Z}\right\} $ isomorphic to the group $\mathbb{Z}$, we get the short
exact sequence
\begin{equation*}
0\rightarrow C^{\ast}\left( \left. \mathcal{F}\right\vert _{U}\right) \con
\mathcal{K}\oplus\mathcal{K}\overset{\iota}{\rightarrow}C^{\ast}\left(
\mathcal{F}\right) \cong C\left( \mathbb{S}_{\mu c}^{2}\right) \overset
\sigma}{\rightarrow}C^{\ast}\left( \left. \mathcal{F}\right\vert _{\left\{
\left( \infty,\infty\right) \right\} }\right) \cong C\left( \mathbb{T
\right) \rightarrow0,
\end{equation*}
where $\delta_{\left( 1,1,\infty,\infty\right) }\in C^{\ast}\left( \left.
\mathcal{F}\right\vert _{\left\{ \left( \infty,\infty\right) \right\}
}\right) $ is identified with $z:=\func{id}_{\mathbb{T}}\in C\left( \mathbb{
}\right) $ under $\sigma$. In the induced 6-term exact sequence of $K
-groups, the index homomorphism $\eta:K_{1}\left( C\left( \mathbb{T}\right)
\right) \rightarrow K_{0}\left( \mathcal{K}\oplus\mathcal{K}\right) $ sends
\left[ z\right] $ to $\left( -\left[ e_{11}\right] \right) \oplus\left(
\left[ e_{11}\right] \right) $, and hence $K_{0}\left( \iota\right) \left(
\left[ e_{11}\oplus0\right] \right) =-K_{0}\left( \iota\right) \left( \left[
0\oplus e_{11}\right] \right) $, leading to $K_{0}\left( \iota\right) \left(
\left[ P_{m}\oplus P_{l}\right] \right) =\left( m-l\right) \left[
e_{11}\oplus0\right] $ in $K_{0}\left( C\left( \mathbb{S}_{\mu c}^{2}\right)
\right) \cong\mathbb{Z}\left[ e_{11}\oplus0\right] \oplus\mathbb{Z}\left[
\tilde {I}\right] $.
By the same kind of analysis carried out above for $C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $, we get the following results.
\textbf{Theorem 3}. \textit{The abelian monoid }$\mathfrak{P}\left( C\left(
\mathbb{S}_{\mu c}^{2}\right) \right) $\textit{\ of unitary equivalence
classes of projections over }$C\left( \mathbb{S}_{\mu c}^{2}\right) $\textit
\ consists of (the representatives) }$P_{m}\oplus P_{l}$\textit{, }$\tilde{I
_{n}\boxplus\left( P_{j}\oplus0\right) $\textit{, and }$\tilde {I
_{n-1}\boxplus\left( \tilde{I}-\left( P_{k}\oplus0\right) \right) $\textit{\
(or equivalently, }$\tilde{I}_{n}\boxplus\left( 0\oplus P_{k}\right)
\textit{) for }$m,l,j\in\mathbb{Z}_{\geq}$\textit{\ and }$n,k\in\mathbb{Z
_{>}$\textit{, where }$\tilde{I}$\textit{\ is the identity of }$\left(
\mathcal{K}\oplus\mathcal{K}\right) ^{+}\subset C\left( \mathbb{S}_{\mu
c}^{2}\right) $\textit{\ and }$P_{k}$\textit{\ is the identity of }
M_{k}\left( \mathbb{C}\right) \subset\mathcal{K}$\textit{, with its binary
operation }$\cdot$\textit{\ specified by
\begin{equation*}
\left\{
\begin{array}{lll}
\left( P_{m}\oplus P_{l}\right) \cdot\left( \tilde{I}_{n}\boxplus\left(
P_{j}\oplus0\right) \right) =\tilde{I}_{n}\boxplus\left(
P_{m+j-l}\oplus0\right) & \text{\textit{if }} & m+j\geq l, \\
\left( P_{m}\oplus P_{l}\right) \cdot\left( \tilde{I}_{n}\boxplus\left(
P_{j}\oplus0\right) \right) =\tilde{I}_{n-1}\boxplus\left( \tilde {I}-\left(
P_{l-m-j}\oplus0\right) \right) & \text{\textit{if }} & m+j<l, \\
\left( \tilde{I}_{n}\boxplus\left( P_{j}\oplus0\right) \right) \cdot\left(
\tilde{I}_{n^{\prime}-1}\boxplus\left( \tilde{I}-\left( P_{k}\oplus0\right)
\right) \right) =\tilde{I}_{n+n^{\prime}}\boxplus\left( P_{j-k}\oplus0\right)
& \text{\textit{if }} & j\geq k, \\
\left( \tilde{I}_{n}\boxplus\left( P_{j}\oplus0\right) \right) \cdot\left(
\tilde{I}_{n^{\prime}-1}\boxplus\left( \tilde{I}-\left( P_{k}\oplus0\right)
\right) \right) =\tilde{I}_{n+n^{\prime}-1}\boxplus\left( \tilde{I}-\left(
P_{k-j}\oplus0\right) \right) & \text{\textit{if }} & j<k, \\
\left( P_{m}\oplus P_{l}\right) \cdot\left( \tilde{I}_{n-1}\boxplus\left(
\tilde{I}-\left( P_{k}\oplus0\right) \right) \right) =\tilde{I
_{n}\boxplus\left( P_{m-k-l}\oplus0\right) & \text{\textit{if }} & m\geq k+l,
\\
\left( P_{m}\oplus P_{l}\right) \cdot\left( \tilde{I}_{n-1}\boxplus\left(
\tilde{I}-\left( P_{k}\oplus0\right) \right) \right) =\tilde{I
_{n-1}\boxplus\left( \tilde{I}-\left( P_{k+l-m}\oplus0\right) \right) &
\text{\textit{if }} & m<k+l
\end{array}
\right.
\end{equation*}
\textit{for representatives of different types and by adding up
corresponding indices }$m,l,j,n,k$ involved\textit{\ for representatives of
the same type}.
\textbf{Corollary 3}. \textit{The cancellation law holds for projections }
\alpha$\textit{\ over }$C\left( \mathbb{S}_{\mu c}^{2}\right) $\textit{\ of
rank }$n\geq1$\textit{\ where }$n$\textit{\ is the rank of the projection }
\sigma\left( \alpha\right) \in M_{\infty}\left( C\left( \mathbb{T}\right)
\right) $\textit{\ at any point of }$\mathbb{T}$\textit{, but fails for
projections }$\alpha$\textit{\ over }$C\left( C\left( \mathbb{S}_{\mu
c}^{2}\right) \right) $\textit{\ of rank} $0$.
The following details about the positive cone of $K_{0}\left( C\left(
\mathbb{S}_{\mu c}^{2}\right) \right) $ extend the information provided by
Corollary 4.4 of \cite{HaMaSz}.
\textbf{Corollary 4}. \textit{The positive cone of }$K_{0}\left( C\left(
\mathbb{S}_{\mu c}^{2}\right) \right) \cong\mathbb{Z}\left[ e_{11}\oplus
\right] \oplus\mathbb{Z}\left[ \tilde{I}\right] $\textit{\ is
\begin{equation*}
\left( \mathbb{Z}\left[ e_{11}\oplus0\right] \oplus0\left[ \tilde {I}\right]
\right) \cup\left( \mathbb{Z}\left[ e_{11}\oplus0\right] \oplus\mathbb{Z}_{>
\left[ \tilde{I}\right] \right) .
\end{equation*}
\textit{The canonical homomorphism from the monoid }$\mathfrak{P}\left(
C\left( \mathbb{S}_{\mu c}^{2}\right) \right) $\textit{\ to }$K_{0}\left(
C\left( \mathbb{S}_{\mu c}^{2}\right) \right) $\textit{\ sends
\begin{equation*}
\left\{
\begin{array}{lll}
P_{m}\oplus P_{l} & \mapsto & \left( m-l\right) \left[ e_{11}\oplus0\right]
\\
\tilde{I}_{n}\boxplus\left( P_{j}\oplus0\right) & \mapsto & j\left[
e_{11}\oplus0\right] \oplus n\left[ \tilde{I}\right] \\
\tilde{I}_{n-1}\boxplus\left( \tilde{I}-\left( P_{k}\oplus0\right) \right) &
\mapsto & -k\left[ e_{11}\oplus0\right] \oplus n\left[ \tilde{I}\right
\end{array}
\right.
\end{equation*}
\textit{for }$m,l,j\in\mathbb{Z}_{\geq}$\textit{\ and} $n,k\in\mathbb{Z}_{>}
.
Comparing the above results, we see that quantum complex projective lines
\mathbb{P}^{1}\left( \mathcal{T}\right) $ and $\mathbb{S}_{\mu c}^{2}$ are
distinguished apart by the monoid structures of $\mathfrak{P}\left( C\left(
\mathbb{P}^{1}\left( \mathcal{T}\right) \right) \right) $ and $\mathfrak{P
\left( C\left( \mathbb{S}_{\mu c}^{2}\right) \right) $, and also by the
positive cone of their $K_{0}$-groups.
\section{Line bundles over quantum projective line}
In this section, we identify the quantum line bundles $L_{k}:=C\left(
\mathbb{S}_{H}^{3}\right) _{k}$ of degree $k$ over $C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $ with a concrete (unitary equivalence
class of) projection classified in the previous section.
To distinguish between ordinary function product and convolution product, we
denote the groupoid C*-algebraic multiplication of elements in $C^{\ast
}\left( \mathcal{G}\right) \supset C_{c}\left( \mathcal{G}\right) $ by $\ast
, while omitting $\ast$ when the elements are presented as operators or when
they are multiplied together pointwise as functions.
Recall from earlier section that $L_{k}=\overline{C_{c}\left( \mathcal{G
_{k}\right) }\subset C\left( \mathbb{S}_{H}^{3}\right) $ where
\begin{equation*}
\mathcal{G}_{k}:=\left\{ \left( n+k,-n,p,q\right) :n\in\mathbb{Z}\text{ such
that }\ \left( p,q\right) ,\left( p+n+k,q-n\right) \in\mathcal{G}^{\left(
0\right) }\right\} \subset\mathcal{G},
\end{equation*}
and $L_{0}=\overline{C_{c}\left( \mathcal{G}_{0}\right) }=C\left( \mathbb{P
^{1}\left( \mathcal{T}\right) \right) $. Furthermore the groupoid C*-algebra
$C^{\ast}\left( \mathcal{G}\right) \cong C\left( \mathbb{S}_{H}^{3}\right)
\overline{\oplus_{k\in\mathbb{Z}}L_{k}}$ is a $\mathbb{Z}$-graded algebra.
Let $k>0$. We identify below separately $L_{k}$ and $L_{-k}$ with a
representative of projections over $C\left( \mathbb{P}^{1}\left( \mathcal{T
\right) \right) $ classified earlier.
The characteristic function $\chi_{A}\in C_{c}\left( \mathcal{G}_{k}\right)
\subset C\left( \mathbb{S}_{H}^{3}\right) $ of the compact set
\begin{equation*}
A:=\left\{ \left( k,0,p,q\right) :\left( p,q\right) ,\left( p+k,q\right) \i
\mathcal{G}^{\left( 0\right) }\right\} \subset\mathcal{G}
\end{equation*}
is an isometry with $\chi_{A}^{\ast}\ast\chi_{A}=\chi_{\mathcal{G}^{\left(
0\right) }}\equiv1\in C^{\ast}\left( \mathcal{G}_{0}\right) $ and
\begin{equation*}
\chi_{A}\ast\chi_{A}^{\ast}=\chi_{B}=\tilde{I}-\left( P_{k}\oplus0\right)
\end{equation*}
a projection in $C^{\ast}\left( \mathcal{G}_{0}\right) $ $\equiv C\left(
\mathbb{P}^{1}\left( \mathcal{T}\right) \right) $for the set
\begin{equation*}
B:=\left\{ \left( 0,0,p^{\prime},q^{\prime}\right) \in\mathcal{G}:\
p^{\prime}\geq k,\ q^{\prime}\geq0\right\} .
\end{equation*}
So with $\chi_{B}\ast\chi_{A}=\chi_{A}\in C_{c}\left( \mathcal{G}_{k}\right)
$, we get a left $C^{\ast}\left( \mathcal{G}_{0}\right) $-module homomorphis
\begin{equation*}
x\in C^{\ast}\left( \mathcal{G}_{0}\right) \ast\chi_{B}\mapsto x\ast\chi
_{A}\in\overline{C_{c}\left( \mathcal{G}_{k}\right) }\equiv L_{k}
\end{equation*}
with well-defined inverse map
\begin{equation*}
y\in\overline{C_{c}\left( \mathcal{G}_{k}\right) }\mapsto y\ast\chi
_{A}^{\ast}=y\ast\chi_{A}^{\ast}\ast\chi_{B}\in C^{\ast}\left( \mathcal{G
_{0}\right) \ast\chi_{B}
\end{equation*}
since $\chi_{A}^{\ast}\in C_{c}\left( \mathcal{G}_{-k}\right) $ and hence
C_{c}\left( \mathcal{G}_{k}\right) \ast\chi_{A}^{\ast}\subset C_{c}\left(
\mathcal{G}_{0}\right) $. Now $L_{k}$ being isomorphic to the left $C^{\ast
}\left( \mathcal{G}_{0}\right) $-module $C^{\ast}\left( \mathcal{G
_{0}\right) \left( \tilde{I}-\left( P_{k}\oplus0\right) \right) $ is
identified with the rank-one projection $\tilde{I}-\left(
P_{k}\oplus0\right) $.
Next we show that in the left $C^{\ast}\left( \mathcal{G}_{0}\right)
-module decompositio
\begin{equation*}
L_{-k}=L_{-k}\ast\chi_{B}\oplus L_{-k}\ast\left( 1-\chi_{B}\right) \equiv
L_{-k}\left( \tilde{I}-\left( P_{k}\oplus0\right) \right) \oplus
L_{-k}\left( P_{k}\oplus0\right)
\end{equation*}
by the projection $\chi_{B}$, the first component $L_{-k}\ast\chi_{B}$ can
be identified with the projection $\tilde{I}$ and the second component
L_{-k}\ast\left( 1-\chi_{B}\right) $ can be identified with the projection
P_{k}\oplus0$, leading to the conclusion that $L_{-k}$ can be identified
with the projection $\tilde{I}\boxplus\left( P_{k}\oplus0\right) $.
Indeed since $\chi_{A}\ast\chi_{A}^{\ast}=\chi_{B}$ and $\chi_{A}^{\ast}\as
\chi_{A}=1\equiv\chi_{\mathcal{G}^{\left( 0\right) }}$ with $\chi
_{A}^{\ast}\in C_{c}\left( \mathcal{G}_{-k}\right) $, the ma
\begin{equation*}
x\in C^{\ast}\left( \mathcal{G}_{0}\right) \mapsto x\ast\chi_{A}^{\ast
}=x\ast\chi_{A}^{\ast}\ast\chi_{B}\in\overline{C_{c}\left( \mathcal{G
_{-k}\right) }\ast\chi_{B}\equiv L_{-k}\ast\chi_{B}
\end{equation*}
is a $C^{\ast}\left( \mathcal{G}_{0}\right) $-module isomorphism with
inverse $y\mapsto y\ast\chi_{A}$ and hence $L_{-k}\ast\chi_{B}$ is
identified with the projection $\tilde{I}$.
On the other hand, comparing
\begin{align*}
C_{c}\left( \mathcal{G}_{-k}\right) \ast\left( 1-\chi_{B}\right) &
=C_{c}\left( \left\{ \left( n-k,-n,p,\infty\right) :0\leq p<k\text{ and}\
p+n\geq k\right\} \right) \\
& =C_{c}\left( \left\{ \left( n,-n-k,p,\infty\right) :0\leq p<k\text{ and}\
p+n\geq0\right\} \right)
\end{align*}
an
\begin{equation*}
C_{c}\left( \mathcal{G}_{0}\right) \ast\left( 1-\chi_{B}\right) =C_{c}\left(
\left\{ \left( n,-n,p,\infty\right) :0\leq p<k\text{\ and}\ p+n\geq0\right\}
\right)
\end{equation*}
where with the last coordinate being $\infty$, the second coordinate becomes
irrelevant, we get a $\left( C_{c}\left( \mathcal{G}_{0}\right) ,\ast\right)
$-module isomorphis
\begin{equation*}
f\in C_{c}\left( \mathcal{G}_{0}\right) \ast\left( 1-\chi_{B}\right) \mapsto
f\circ\tau\in C_{c}\left( \mathcal{G}_{-k}\right) \ast\left(
1-\chi_{B}\right)
\end{equation*}
where $\tau\left( n,-n-k,p,\infty\right) :=\left( n,-n,p,\infty\right) $,
which extends to a $C^{\ast}\left( \mathcal{G}_{0}\right) $-module
isomorphism
\begin{equation*}
C^{\ast}\left( \mathcal{G}_{0}\right) \ast\left( 1-\chi_{B}\right) \equiv
C^{\ast}\left( \mathcal{G}_{0}\right) \left( P_{k}\oplus0\right) \rightarro
\overline{C_{c}\left( \mathcal{G}_{-k}\right) }\ast\left( 1-\chi_{B}\right)
\equiv L_{-k}\ast\left( 1-\chi_{B}\right) .
\end{equation*}
So the $C^{\ast}\left( \mathcal{G}_{0}\right) $-module $L_{-k}\ast\left(
1-\chi_{B}\right) $ is identified with the projection $P_{k}\oplus0$.
We summarize as follows.
\textbf{Theorem 4}. \textit{The quantum line bundle }$L_{k}\equiv C\left(
\mathbb{S}_{H}^{3}\right) _{k}$\textit{\ of degree }$k\in\mathbb{Z}$\textit
\ over }$C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) $\textit{\
is isomorphic to the finitely generated projective left module over }
C\left( \mathbb{P}^{1}\left( \mathcal{T}\right) \right) $\textit{\
determined by the projection }$\tilde{I}-\left( P_{k}\oplus0\right) $\textit
\ if }$k\geq 0$\textit{, and the projection }$\tilde{I}\boxplus\left(
P_{-k}\oplus0\right) $\textit{\ if} $k<0$.
\textbf{Corollary 5}. \textit{The quantum line bundles }$L_{k}\equiv C\left(
\mathbb{S}_{H}^{3}\right) _{k}$\textit{\ with }$k\in\mathbb{Z}$\textit{\
provide a complete list of mutually non-isomorphic rank-one finitely
generated left projective modules over} $C\left( \mathbb{P}^{1}\left(
\mathcal{T}\right) \right) $.
It is interesting to note that in the case of quantum teardrops
WP_{q}\left( k,l\right) $, the quantum principal $U\left( 1\right) $-bundles
$\mathcal{L}\left( k\right) $ of degree $k$ over $C\left( WP_{q}\left(
k,l\right) \right) $ introduced by Brzezi\'{n}ski and Fairfax \cite{BrzeFair}
do not exhaust all rank-one finitely generated projective modules over
C\left( WP_{q}\left( k,l\right) \right) $ by the result of \cite{Sh:qt}.
|
\section*{Abstract}
\input{sec_abstract.tex}
\IEEEpeerreviewmaketitle
\section{Introduction}
\input{sec_introduction.tex}
\section{Materials and methods}
\input{sec_materials.tex}
\section{Procedure}
\input{sec_estimation.tex}
\section{Results and discussion}
\input{sec_results.tex}
\section{CONCLUSION}
\input{sec_conclusion.tex}
\setcounter{secnumdepth}{0}
\section{Appendix}
\input{sec_appendix1.tex}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\input{sec_references.tex}
\end{document}
\subsection{Data acquisition}
While participants were performing the test we recorded the videos with a Pointgrey Flea3 high frame rate USB camera. We stored the video for future work. Simultaneously, EEG signals were captured using Mitsar-EEG 201 amplifier and accompanying WinEEG software. The electrodes were placed according to the international “10-20 system” standards of electrode placement.
Electro-gel was injected into electrodes hollow in order to decrease the electrode-skin resistance. The focus of current work is to analyze the statistical characteristics of blinks. Nevertheless, we record EEG from all the electrodes for further research.
\begin{figure}[!htbp]
\includegraphics[width=3.5in]{img/setup.eps}
\caption{Setup of the experiment}
\label{fig2}
\end{figure}
\subsection{Testing procedure}
Fifty one male and female subjects aged from 19 to 25 years, were recruited for the experiment. Everyone provided their written consent. The subjects had no history of psychiatric illness, and they had not been affected by any significant medical, neurological or ophthalmological illness. We prepared a questionnaire testifying that subjects had none of these conditions.
The experiment software was developed in JavaScript with jQuery (Fig. \ref{fig5a}). The software was designed in a way that no intervention of participants or an experiment supervisor is required during the procedure. \\
The whole testing session took 30 minutes and consisted of 5 minute resting, followed by a 10 minute IQ testing, and a 5 minute rest. After the second resting period, a passage about Ethiopia was presented, followed by a quiz to invoke memory recall activity. The first 5 minute resting period was included to calm the participants down. Participants were asked to try to keep their mind calm as much as possible. The following IQ test is not discussed in this paper as more investigation is needed. The latter 5 minute resting stage is included in this study and the comparison between the two resting sessions is provided in later sections. The passage is presented after the second resting interval for five minutes. Participants were asked to read the passage carefully with the purpose of memorizing the facts from it. The passage is 40 sentences long, and contains basic facts about Ethiopia. None of the participants were familiar with the topic of the text. The passage followed by the 5 minute quiz to test the memory retention, we will discuss it in our next study. \\
Since EEG is prone to noise due to movement, we made sure that subjects had been informed about staying still and their movement was minimized. During both resting stages, information with instruction about staying still with countdown progress bar, were displayed. To prevent head movements, text was displayed on a single screen with no need to scroll. Due to heavy noise caused by the subjects falling asleep, adjusting the cap or constant head movements, the total of 40 subjects' data were dropped.\\
\begin{figure}[!htbp]
\includegraphics[trim=0.0cm 1.5cm 0.0cm 2.5cm, width=3.5in]{img/software.eps}
\caption{Question delivery software}
\label{fig5a}
\end{figure}
|
\section{Introduction}
QKD allows two distant and trusted parties, Alice and Bob, to create a secret key through the use of quantum and classical communications, and can be implemented using discrete-variable (DV) or CV quantum states. In DV-QKD protocols, key information is encoded on the properties of single photons, and in CV-QKD protocols key information is encoded on the quadrature variables of light. In the former technology detection is realized by single-photon detectors e.g., \cite{DV1, DV2, DV3}, while in the latter technology detection is realised by more efficient and faster detectors, such as homodyne (or heterodyne) detectors e.g., \cite{CV1, CV2, thesis, Weedbrook2012, Weedbrook2013}.
QKD is mostly implemented experimentally in a prepare-and-measure (PM) type scheme e.g., \cite{exp-PM}, where Alice prepares the quantum states which are then transmitted over an insecure quantum channel towards Bob, who measures the received quantum states, resulting in creation of correlated data between the two trusted parties. Each PM scheme can be represented in an equivalent entanglement-based (EB) scheme, where Alice generates a two-mode entangled state, with one mode being kept by Alice and the other mode being transmitted over an insecure quantum channel. Alice and Bob then proceed to invoke a QKD protocol by measuring their own modes to create correlated data. In both the PM scheme and EB scheme, after generation of the correlated data, Alice and Bob proceed with classical postprocessing including sifting, parameter estimation, reconciliation and privacy amplification over a public (but authenticated) classical channel, to generate a secret key even in the presence of Eve (a potential eavesdropper).
CV-QKD protocols using Gaussian resources have been well analysed in theory and experimentally implemented e.g., \cite{thesis, Weedbrook2012, Weedbrook2013, exp-PM}. In the Gaussian PM scheme \cite{thesis, Weedbrook2012, exp-PM}, the prepared quantum states are either coherent states or squeezed states which are modulated by Gaussian distributions. Each Gaussian PM scheme can be represented by an equivalent EB scheme \cite{thesis, Weedbrook2012, Weedbrook2013} in which Alice generates a two-mode squeezed vacuum (TMSV) state. Although Gaussian quantum states are a well-established resource from both a theoretical and an experimental perspective (for review see\cite{Weedbrook2012}), the use of non-Gaussian states in the implementation of CV-QKD protocols has also garnered interest, e.g. \cite{nG-modulation, nG1, nG2, nG-coherent}.
Such studies can be motivated in part by the fact that non-Gaussian operations such as photon subtraction \cite{1st_PSS, 2, telep, 3, 9, Oxford, Oxford2013, 7} and photon addition \cite{1st_PAS, telep, Oxford, 7, added} on an incoming TMSV state can lead to higher levels of entanglement \cite{Oxford}.
It is widely anticipated that the use of satellites will assist in the deployment of QKD over global scales. Thus, it is important to analyse the effectiveness of CV-QKD protocols over atmospheric channels towards (and from) LEO satellites. Such channels are highly fading in nature since the transmissivity of the channel fluctuates due to atmospheric effects. CV-QKD protocols using Gaussian resources have been studied over atmospheric channels in \cite{Usenko, Heim, Neda1, Neda2}. However, works on using non-Gaussian resources in CV-QKD protocols have focussed on fixed-attenuation channels. Thus, it remains unclear whether non-Gaussian entangled states can be effective in the implementation of CV-QKD protocols over atmospheric fading channels (in terms of the quantum key generation rates). The entanglement-generation rate produced by non-Gaussian entangled states passing through atmospheric channels has been recently studied by us \cite{Neda3}. In this present work we extend such studies to consider bounds on the actual quantum key rates generated by such non-Gaussian transfer. We will be particularly interested in comparing such quantum key rates with those arising from Gaussian transfer over atmospheric channels, and Gaussian and non-Gaussian transfer over fixed-attenuation channels.
We will focus on transmission fading caused by the beam wander \cite{Usenko, fso, beamwander}, which is expected to dominate photon losses in Earth-to-satellite channels. To explore non-Gaussian key generation we will utilize non-Gaussian entangled states which are created just-in-time via photonic subtraction from incoming Gaussian states. In our security analysis, we will include the effects of channel fading, the probabilistic production of the non-Gaussian states, and the most likely imperfections in a CV-QKD protocol (excess noise, inefficient and noisy detectors, and non-perfect reconciliation algorithms). A schematic illustration of the CV-QKD protocol we adopt in a fading uplink configuration is shown in Fig.~\ref{scheme1}.
\begin{figure}[!t]
\begin{center}
{\includegraphics[width=3.4 in, height=2.5 in]{scheme1.eps}}
\caption{An example of quantum communication scheme in the ground-to-LEO configuration for implementation of our CV-QKD protocol.}\label{scheme1}
\end{center}
\end{figure}
The structure of the remainder of this paper is as follows. In Section II, our CV-QKD protocol using a non-Gaussian resource is described, and its security analysis is discussed in detail. In Section III, the performance of the protocol over the atmospheric channels is presented and compared to the corresponding protocol over the fixed-attenuation channels. Finally, concluding results are provided in Section IV.
\section{system model and quantum key rate}
We now describe the implementation of a CV-QKD protocol using non-Gaussian states and outline how to determine the quantum key rates of this protocol for the atmospheric channel.
\subsection{CV-QKD Protocol using non-Gaussian states}
In this work, we will focus on the implementation of a specific CV-QKD protocol in the EB scheme, in which a two-mode non-Gaussian entangled state is shared between Alice and Bob. Initially, we focus on the case where the non-Gaussian state is created just-in-time via photonic subtraction from an incoming Gaussian state - a non-Gaussian operation which has previously been experimentally demonstrated \cite{exp1, exp2}.
The protocol begins with Alice possessing a TMSV state, $\rho_{AB}$, described in the Fock basis as ${\left| {TMSV} \right\rangle _{AB}} = \sum\limits_{n = 0}^\infty {{q_n}} {\left| n \right\rangle _A}{\left| n \right\rangle _B}$, where ${q_n} = {\lambda ^n}\sqrt {1 - {\lambda ^2}} $, and where $\lambda = \tanh r $ with $r$ being the squeezing parameter (indices $A$ and $B$ indicate the two modes). The two-mode squeezing in dB is given by $ - 10{\log _{10}}\left( {\exp ( - 2r)} \right)$. Since the TMSV state is a Gaussian state, it can be completely described by its first moment, which is zero, and its covariance matrix (CM) given by
\begin{eqnarray}\label{TMSV-CM}
M_{AB}^{in}= \left( {\begin{array}{*{20}{c}}
{v\,I}&{\sqrt {{v^2} - 1} \,Z}\\
{\sqrt {{v^2} - 1} \,Z}&{v\,I}
\end{array}} \right) ,
\end{eqnarray}
where $I$ is a $2 \times 2$ identity matrix, $Z = diag\left( {1, - 1} \right)$, and $v = \cosh \left( {2r} \right)$ is the quadrature variance of each mode.
While mode~$A$ is held by Alice, the other mode undergoes the photon-subtraction operation. To invoke this operation, Mode~$B$ of Alice's TMSV state interacts with a vacuum mode in a beam splitter BS1 with transmissivity $T$ and reflectivity $1-T$, with one of the outputs of the beam splitter feeding a single-photon detector. When the detector registers one photon, a pure photon-subtracted squeezed (PSS) state is heralded in mode~$B_1$, the non-measured output of the beam splitter, defined as ${\rho _{A{B_1}}}$, which is a non-Gaussian entangled state. The normalized state arising from this process and its creation probability $P_{ss}$ are given by ${\left| {PSS} \right\rangle _{A{B_1}}} = \sum\limits_{n = 0}^\infty {{q_n}} {\left| {n + 1} \right\rangle _A}{\left| n \right\rangle _{B_1}}$, where \cite{Neda3, Oxford}
\begin{eqnarray}\label{PSS}
\begin{array}{l}
{q_n} = \left( {1 - {\lambda ^2}T} \right){\left( {\lambda \sqrt T } \right)^n}\sqrt {n + 1}, \\
\\
{P_{ss}} = \frac{{{\lambda ^2}\left( {1 - {\lambda ^2}} \right)\left( {1 - T} \right)}}{{{{\left( {1 - {\lambda ^2}T} \right)}^2}}}.
\end{array}
\end{eqnarray}
After applying the photon subtraction, the output of the beam splitter, mode~$B_{1}$ is transmitted towards Bob over an unsecured quantum channel with transmissivity $\tau$. In any real-world implementation of this CV-QKD protocol, the quantum state preparation at Alice's side will produce some additional noise. In the following we refer to this excess noise as $\varepsilon$.
The QKD protocol then proceeds with Bob making a homodyne detection of the amplitude or phase quadrature of the received mode $B_2$, and Alice measures both quadratures using a heterodyne detection on mode~$A$ to create correlated data.
Bob's realistic homodyne detection will have an efficiency $\mu$ and an electronic noise $\nu_{el}$ \cite{thesis, inefficient_homodyne}. The efficiency can be \emph{modeled} by placing a beam splitter BS2 of transmissivity $\mu$ before an ideal homodyne detector. This detector's electronic noise can be modeled by an EPR state, $\rho_{{H_0}G}$, of quadrature variance $\nu _d$, where ${\nu _{el}} = \left( {1 - \mu } \right)\left( {{\nu _d} - 1} \right)$. One input port of the beam splitter BS2 is the received mode $B_2$, and the second input port is fed by one half of the EPR state, mode~$H_0$, while the output ports are mode~$B_{3}$ (which is measured by the ideal homodyne detector) and mode $H$. Our CV-QKD protocol using the PSS state is shown in Fig.~\ref{scheme2}.
\begin{figure}[!t]
\begin{center}
{\includegraphics[width=3.4 in, height=2 in]{scheme4.eps}}
\caption{The entanglement-based CV-QKD protocol over fading channels using the PSS state which is created just-in-time from the incoming TMSV state.}\label{scheme2}
\end{center}
\end{figure}
Since non-Gaussian states are not completely described by the first and second moments of the quadrature operators, we are not able to quantify the evolution of our PSS state solely through the CM. Thus, we will employ the Kraus representation \cite{Kraus} in order to analyze the evolution of our PSS state through the channel. Considering a quantum state with density operator ${\rho _{in}}$ as the input of a trace-preserving completely positive channel, the output density operator of the channel can be described in an operator-sum representation of the form
${\rho _{out}} = \sum\limits_{\ell = 0}^\infty {{G_\ell }{\rho _{in}}\,G_\ell ^\dag }$,
where the Kraus operators ${G_\ell }$ satisfy $\sum\limits_{\ell = 0}^\infty {{G_\ell }\,G_\ell ^\dag } = I$, with $I$ being the identity operator.
The Kraus operators of a fixed-attenuation channel with transmissivity $\tau $ can be written as \cite{Kraus},
\begin{eqnarray}\label{Atten}
{G_\ell }\left( \tau \right) = \sum\limits_{m = 0}^\infty {\sqrt {{}^{m + \ell }{C_\ell }} {{\left( {\sqrt {1 - \tau } } \right)}^\ell }{{\left( {\sqrt \tau } \right)}^m}\left| m \right\rangle \left\langle {m + \ell } \right|,}
\end{eqnarray}
where ${}^{m + \ell }{C_\ell }$ is the binomial coefficient. From these operators it can then be shown that the elementary density operator $\left| m \right\rangle \left\langle n \right|$ after the evolution through the fixed-attenuation channel can be written as \cite{Kraus},
\begin{equation}\
\begin{array}{l}
\left| m \right\rangle \left\langle n \right| \to \sum\limits_{\ell = 0}^\infty {{G_\ell }\left( \tau \right)\left| m \right\rangle \left\langle n \right|\,G_\ell ^\dag } \left( \tau \right) = \\
\\
\sum\limits_{\ell = 0}^{\min \{ m,n\} } {\sqrt {{}^m{C_\ell }{}^n{C_\ell }} } {\left( {1 - \tau } \right)^\ell }{\left( {\sqrt \tau } \right)^{\left( {m + n - 2\ell } \right)}}\left| {m - \ell } \right\rangle \left\langle {n - \ell } \right|.
\end{array}\label{Evol}
\end{equation}
Considering the initial density operator of our PSS state as the following
\begin{eqnarray}\label{initial}
{\rho _{A{B_1}}} = \sum\limits_{m = 0}^\infty {\sum\limits_{n = 0}^\infty {{q_m}} {q_n}{{\left| {m + 1} \right\rangle }_A}{{\left\langle {n + 1} \right|}_A} \otimes {{\left| m \right\rangle }_{{B_1}}}{{\left\langle n \right|}_{{B_1}}}},
\end{eqnarray}
the density operator of the output mixed state ${\rho _{A{B_2}}}$ can be calculated in the Fock basis through the use of Eq.~\eqref{Evol}, giving
\begin{equation}\label{output}
{\rho _{A{B_2}}} = \sum\limits_{a' = 1}^\infty {\sum\limits_{b' = 0}^\infty {\sum\limits_{c' = 1}^\infty {\sum\limits_{d' = 0}^\infty {{\rho _{a'b'c'd'}}{{\left| {a'} \right\rangle }_A}{{\left\langle {c'} \right|}_A} \otimes {{\left| {b'} \right\rangle }_{{B_2}}}{{\left\langle {d'} \right|}_{{B_2}}}} } } } ,
\end{equation}
with
\begin{equation}\label{elements}
\begin{array}{l}
{\rho _{a'b'c'd'}} = {q_{a' - 1}}\,{q_{c' - 1}} \times \\
\\
\,\,\,\,\,\,\,\,\,\,\,\,\,\,\sqrt {{}^{a' - 1}{C_{a' - 1 - b'}}{}^{c' - 1}{C_{c' - 1 - d'}}} {\left( {1 - \tau } \right)^{a' - 1 - b'}}{\left( {\sqrt \tau } \right)^{b' + d'}}
\end{array}
\end{equation}
if $a' - b' = c' - d' $ and $a' - 1 - b' \ge 0$, otherwise $\rho _{a'b'c'd'} = 0$.
We can compute the lower bound on the actual key rate, $K$, of the protocol by replacing the shared non-Gaussian state between Alice and Bob $\rho_{A{B_2}}$ with a Gaussian state $\rho _{A{B_2}}^G$ having the same mean value and CM, resulting in $K(\rho _{A{B_2}}^G) \le K({\rho _{A{B_2}}})$. The output mixed state ${\rho _{A{B_2}}}$ has zero mean and CM in the following form
\begin{equation}\label{AB2}
{M_{A{B_2}}} = \left( {\begin{array}{*{20}{c}}
{x'I}&{z'Z}\\
{z'Z}&{y'I}
\end{array}} \right),
\end{equation}
where
\begin{equation}\label{x'y'z'}
\begin{array}{l}
x' = \sum\limits_{a' = 1}^\infty {\sum\limits_{b' = 0}^\infty {\sum\limits_{c' = 1}^\infty {\sum\limits_{d' = 0}^\infty {(2a' + 1){\rho _{a'b'c'd'}}} } } } \left| {_{a' = c'}} \right.,\\
\\
y' = \sum\limits_{a' = 1}^\infty {\sum\limits_{b' = 0}^\infty {\sum\limits_{c' = 1}^\infty {\sum\limits_{d' = 0}^\infty {(2b' + 1){\rho _{a'b'c'd'}}} } } } \left| {_{a' = c'}} \right.,\\
\\
z' = \sum\limits_{a' = 1}^\infty {\sum\limits_{b' = 0}^\infty {\sum\limits_{c' = 1}^\infty {\sum\limits_{d' = 0}^\infty {\left( {\sqrt {a'b'} {\rho _{a'b'c'd'}}\left| {_{a' = c' + 1}} \right.} \right.} } } } \\
\\
\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\left. { + \sqrt {(a' + 1)(b' + 1)} {\rho _{a'b'c'd'}}\left| {_{a' = c' - 1}} \right.} \right).
\end{array}
\end{equation}
Since the output density operator ${\rho _{A{B_2}}}$ possesses an infinite number of elements, we are required to deploy a numerical method to approximate ${\rho _{A{B_2}}}$ by limiting its size, i.e. creating a truncated ${\rho _{A{B_2}}}$. Here, a very similar approach to that described in our previous work \cite{Neda3} can be used to estimate ${\rho _{A{B_2}}}$, and then calculate the elements of the CM ${M_{A{B_2}}}$ through the use of Eq.~\eqref{x'y'z'}. However, to determine the true resultant key rates (not just lower bounds) for the non-Gaussian state, Eq.~\eqref{output} must be utilized.
There is also an analytical approach to obtain the CM ${M _{A{B_2}}}$ \cite{nG1, nG2}, in which we first need to calculate the CM of the initial PSS state ${\rho _{A{B_1}}}$ given by
\begin{equation}\label{AB1}
\begin{array}{l}
{M_{A{B_1}}} = \left( {\begin{array}{*{20}{c}}
{xI}&{zZ}\\
{zZ}&{yI}
\end{array}} \right),\,\rm where\\
\\
x = 2v' + 1,\,\,\,\,y = 2v' - 1,\,\,\,\,z = 2\sqrt {{{v'}^2} - 1\,} ,
\end{array}
\end{equation}
and where $v' = \frac{{1 + \left( {1 + T} \right){{\sinh }^2}r}}{{1 + \left( {1 - T} \right){{\sinh }^2}r}}$. Note the PSS state ${\rho _{A{B_1}}}$ has zero mean.
After transmission of mode~$B_1$ through a quantum channel with transmissivity $\tau$ and excess noise $\varepsilon$, the CM of the resulting mixed state ${\rho _{A{B_2}}}$ is given by
\begin{equation}\label{AB2-CM}
{M_{A{B_2}}} = \left( {\begin{array}{*{20}{c}}
{x\,I}&{\sqrt {\tau \,} z\,Z}\\
{\sqrt \tau \,z\,Z}&{\left( {\tau (y + {\chi _c})} \right)\,I}
\end{array}} \right),
\end{equation}
where ${\chi _c} = \varepsilon + \frac{{1 - \tau }}{\tau }$. Note also the numerical method (truncated Eq.~\eqref{AB2}) gives effectively the same output CM as the CM in Eq.~\eqref{AB2-CM} for $\varepsilon=0$, provided the output density matrix ${\rho _{A{B_2}}}$ is appropriately estimated \cite{Neda3}. After mode~$B_2$ undergoes the beam splitter BS2, the CM of the resulting mixed state ${\rho _{A{B_3}}}$ is given by
\begin{equation}\label{AB3-CM}
{M_{A{B_3}}} = \left( {\begin{array}{*{20}{c}}
{x\,I}&{\sqrt {\mu \tau \,} z\,Z}\\
{\sqrt {\mu \tau \,} z\,Z}&{\left( {\mu \tau (y + \chi )} \right)\,I}
\end{array}} \right),
\end{equation}
where $\chi = {\chi _c} + \frac{{{\chi _d}}}{\tau }$, and where ${\chi _d} = \frac{{1 - \mu }}{\mu }{\nu _d}$. Having the CMs ${M_{A{B_2}}}$ in Eq.~\eqref{AB2-CM} and ${M_{A{B_3}}}$, we are then able to compute a key rate, which will be a lower bound on the actual key rate of our CV-QKD protocol.
If the original TMSV state $\rho_{AB}$ is exploited as the source of the CV-QKD protocol, the output state at the end of the channel will still be a Gaussian state, whose CM is again given by Eq.~\eqref{AB2-CM}, but where now $x$, $y$ and $z$ are given by
\begin{equation}\label{TMSV-output-CM}
x = y = v,\,\,\,z = \sqrt {{v^2} - 1}.
\end{equation}
\subsection{Computation of the Secret Key Rate}
Here, we summarize some known results in computing a key rate based on the CM of a shared entangled state between the two trusted parties \cite{thesis, inefficient_homodyne}. It is known in the typical point-to-point CV QKD, the reverse reconciliation (RR) scenario always leads to higher key rates than the direct reconciliation (DR) scenario \cite{RR_2002}. Hence, we will only consider the RR scenario where Bob's data is the reference of reconciliation.
For a realistic reconciliation algorithm, the asymptotic key rate against collective attacks for RR is given by $K = \xi {I_{a{b_3}}} - I_E$, where $0<\xi<1$ is the reconciliation efficiency, $I_{a{b_3}}$ is the mutual information between Alice and Bob's measurements after accounting for additional detector noise and efficiencies.
$I_{E}$ is the Holevo quantity, which gives an upper bound on the quantum information stolen by Eve.
We will assume Bob makes the homodyne detection of ${{\hat q}_{{B_3}}}$, the amplitude quadrature of mode~$B_3$. On the other side, Alice invokes heterodyne detection by combining mode~$A$ with a vacuum mode on a balanced beam splitter, and applies homodyne detection on conjugate quadratures of the two output modes.
Due to the symmetry between the two conjugate quadratures, we only consider the measurement of the amplitude quadrature of mode~$A^{he}$ (one output mode of the balanced beam splitter), which is given by ${{\hat q}_{{A^{he}}}} = \frac{1}{{\sqrt 2 }}\left( {{{\hat q}_A} + {{\hat q}_0}} \right)$, where ${{{\hat q}_A}}$ is the amplitude quadrature of mode~$A$ prior to the heterodyne detection and ${{{\hat q}_0}}$ is the amplitude quadrature of the vacuum input.
Then
$I_{a{b_3}} = \frac{1}{2}{\log _2}\left( {\frac{{{V_{{B_3}}}}}{{{V_{{B_3}\left| {{A^{he}}} \right.}}}}} \right)$, where $V$ denotes the variance, and
\begin{equation}\label{mutual1}
{V_{{B_3}}} = \left\langle {\hat q_{{B_3}}^2} \right\rangle ,\,\,{V_{{B_3}\left| {{A^{he}}} \right.}} = \left\langle {\hat q_{{B_3}}^2} \right\rangle - \frac{{{{\left\langle {{{\hat q}_{{B_3}}},{{\hat q}_{{A^{he}}}}} \right\rangle }^2}}}{{\left\langle {\hat q_{{A^{he}}}^2} \right\rangle }},
\end{equation}
where $\left\langle . \right\rangle $ denotes the expectation value. According to the CM ${M_{A{B_3}}}$ in Eq.~\eqref{AB3-CM} we will have
\begin{equation}\label{mutual2}
\begin{array}{l}
\left\langle {\hat q_{{B_3}}^2} \right\rangle = \mu \tau \left( {y + \chi } \right),\\
\\
\left\langle {{{\hat q}_{{B_3}}},{{\hat q}_{{A^{he}}}}} \right\rangle = \frac{1}{{\sqrt 2 }}\left\langle {{{\hat q}_{{B_3}}},{{\hat q}_A}} \right\rangle = \frac{1}{{\sqrt 2 }}\sqrt {\mu \tau } z,\\
\\
\left\langle {\hat q_{{A^{he}}}^2} \right\rangle = \frac{1}{2}\left( {\left\langle {\hat q_A^2} \right\rangle + 1} \right) = \,\frac{1}{2}\left( {x + 1} \right).
\end{array}
\end{equation}
Thus, we find ${V_{{B_3}\left| {{A^{he}}} \right.}} = \mu \tau \left( {y - \frac{{{z^2}}}{{x + 1}} + \chi } \right)$, which according to the values of $x$, $y$ and $z$ of the PSS state in Eq.~\eqref{AB1} is given by ${V_{{B_3}\left| {{A^{he}}} \right.}} =\mu \tau (1 + \chi ) $. Hence, the mutual information for the CV-QKD protocol using the PSS states is given by $I_{a{b_3}}=\frac{1}{2}{\log _2}\left( {\frac{{y + \chi }}{{1 + \chi }}} \right)$
Eve's information on Bob's measurement outcome, $b_3$, is given by ${I_E} = S\left( {{\rho _E}} \right) - S\left( {{\rho _{E\left| {{b_3}} \right.}}} \right)$, where $S(\rho )$ is the von Neumann entropy of the state $\rho$. To determine $ S\left( {{\rho _E}} \right)$ we assume Eve's system $\rho _E$ purifies $\rho _{AB_2}$, that is, $S({\rho _E}) = S({\rho _{A{B_2}}})$.
The entropy $S({\rho _{A{B_2}}})$ can be calculated through the symplectic eigenvalues ${\nu _{1,2}}$ of ${M_{A{B_2}}}$ given by Eq.~\eqref{AB2-CM}. This leads to $S({\rho _{A{B_2}}}) = f({\nu _1}) + f({\nu _2})$, where $f(x) = \frac{{x + 1}}{2}{\log _2}\left( {\frac{{x + 1}}{2}} \right) - \frac{{x - 1}}{2}{\log _2}\left( {\frac{{x - 1}}{2}} \right)$, and $\nu _{1,2}^2 = \left( {\Delta \pm \sqrt {{\Delta ^2} - 4\Omega } } \right)/2$, with
\begin{equation}\label{compare}
\begin{array}{l}
\Delta = {x^2} + {\tau ^2}{\left( {y + {\chi _c}} \right)^2} - 2\tau {z^2},\\
\\
\Omega = {\left( {x\tau \left( {y + {\chi _c}} \right) - \tau {z^2}} \right)^2}.
\end{array}
\end{equation}
The second entropy we require in order to determine $I_E$ can be written as
$S({\rho _{E\left| {b_3} \right.}}) = S\left( {{\rho _{AHG\left| {b_3} \right.}}} \right)$,
since $\rho_{AEHG}$ is pure.
The conditional entropy $S\left( {{\rho _{AHG\left| {b_3} \right.}}} \right)$ can be calculated as $S\left( {{\rho _{AHG\left| {b_3} \right.}}} \right) = f({\nu _3}) + f({\nu _4})$, where
\begin{equation}\label{compare}
\begin{array}{l}
\nu _{3,4}^2 = \left( {\Delta ' \pm \sqrt {{{\Delta '}^2} - 4\Omega '} } \right)/2, \ \ {\rm with}\\
\\
\Delta ' = \frac{1}{{\tau \left( {y + \chi } \right)}}\left( {\tau \left( {y + {\chi _c}} \right) + x\sqrt \Omega + {\chi _d}\Delta } \right),\\
\\
\Omega ' = \frac{{\sqrt \Omega }}{{\tau \left( {y + \chi } \right)}}\left( {x + {\chi _d}\sqrt \Omega } \right).
\end{array}
\end{equation}
\subsection{CV-QKD Protocol over Atmospheric Fading Channels}
In atmospheric channels the transmissivity $\eta_{tran}$ fluctuates due to several effects. Such fading channels can be characterized by a distribution of values $\eta $, with a probability density distribution $p\left( \eta \right)$, where $\eta = \sqrt {{\eta _{tran}}} $. Consistent with other recent studies \cite{fso, beamwander, Usenko}, we will assume that fading arising from the atmosphere is due only to beam wander. Assuming the beam spatially fluctuates around the center of the receiver's aperture, the probability density distribution $p\left( \eta \right)$ can be described by the log-negative Weibull distribution \cite{beamwander},
\begin{equation}\
p\left( \eta \right) = \frac{{2{L^2}}}{{\sigma _b^2\gamma_s \eta }}{\left( {2\ln \frac{{{\eta _0}}}{\eta }} \right)^{\left( {\frac{2}{\gamma_s }}- 1 \right) }}\exp \left( { - \frac{{{L^2}}}{{2\sigma _b^2}}{{\left( {2\ln \frac{{{\eta _0}}}{\eta }} \right)}^{\left( {\frac{2}{\gamma_s }} \right)}}} \right)
\label{f1}
\end{equation}
for $\eta \in \left[ {0,\,{\eta _0}} \right]$, with $p\left( \eta \right) = 0$ otherwise.
Here, $\sigma _b^2$ is the beam wander variance,
$\gamma_s$ is the shape parameter, $L$ is the scale parameter, and ${\eta _0}$ is the maximum value of $\eta$. The latter three parameters are given by
\begin{equation}\label{f2}
\begin{array}{l}
\gamma_s = 8h\frac{{\exp \left( { - 4h} \right){I_1}\left[ {4h} \right]}}{{1 - \exp \left( { - 4h} \right){I_0}\left[ {4h} \right]}}{\left[ {\ln \left( {\frac{{2\eta _0^2}}{{1 - \exp \left( { - 4h} \right){I_0}\left[ {4h} \right]}}} \right)} \right]^{ - 1}},\\
\\
L = \beta{\left[ {\ln \left( {\frac{{2\eta _0^2}}{{1 - \exp \left( { - 4h} \right){I_0}\left[ {4h} \right]}}} \right)} \right]^{ - \left( {{1 \mathord{\left/
{\vphantom {1 \gamma_s }} \right.
\kern-\nulldelimiterspace} \gamma_s }} \right)}},\,\,\eta _0^2 = 1 - \exp \left( { - 2h} \right) ,
\end{array}
\end{equation}
where ${I_0}\left[ . \right]$ and ${I_1}\left[ . \right]$ are the modified Bessel functions, and where $h = {\left( {{\beta \mathord{\left/
{\vphantom {a W}} \right.
\kern-\nulldelimiterspace} W}} \right)^2}$, with $\beta$ being the receiver aperture radius and $W$ the beam-spot radius. In our subsequent calculations we will adopt $W=\beta$, and let the mean fading loss be controlled only by adjustments to the value of $\sigma _b$.
Since depolarization is very weak in the atmospheric channel, dephasing will also be weak and thus we will ignore it \cite{sem}. We will also assume the transmissivity of the channel can be measured in real-time at the receiver by passing a local oscillator through the channel in an orthogonal polarized mode to the signal.
In the implementation of our CV-QKD protocol over the atmospheric fading channels towards (and from) a LEO satellite, we will assume one mode of the PSS state remains at the ground station (satellite), while the other mode (photon-subtracted mode) is transmitted to the satellite (ground station) over the fading uplink (downlink) with probability density distribution $p\left( \eta \right)$. Let us assume Alice is the sender and Bob is receiver. After each realization of $\eta $, using the CM in Eqs.~\eqref{AB2-CM} and \eqref{AB3-CM}, where the transmissivity $\tau$ needs to be replaced by $\eta^2 $, we can compute the lower bound on the key rate given by ${K(\eta )}$. Since $\eta$ is a random variable, the elements of the final key rate is computed by averaging ${K(\eta )}$ over all possible transmission factors of the fading channel giving $K = \int_0^{{\eta _0}} {K(\eta )} p(\eta )\,d\eta $ in units of bits per pulse.
Note the key rate of the protocol must be computed by including the creation probability $P_c$ of the initial entangled state as $P_cK$. In our protocol, when the PSS state is created just-in-time from the incoming TMSV state, and then used for the CV-QKD protocol $P_c=P_{ss}$. Here, we have assumed the creation probability of the original TMSV state is one.
The range of losses we consider covers a wide range of anticipated scenarios for LEO satellite-based communications with losses in the range 0-10 dB (downlink) to 20-30 dB (uplink). Our results will also be applicable to direct line-of-sight terrestrial communications through air.
\section{Simulation Results}
We now consider the performance of our CV-QKD protocol.
In Fig.~\ref{high} (top), we plot the key rate $P_cK$ resulting from the PSS states as well as the original TMSV states over the fading channels as a function of channel loss for different values of squeezing of the original TMSV state; 5dB, 10dB and 16dB.\footnote{Note for the photon subtraction operation, the transmissivity $T$ of the exploited beam splitter BS1 can be chosen arbitrarily from 0 to 1, leading to the change of the creation probability $P_{ss}$ and also the CM in Eqs.~\eqref{AB2-CM} and \eqref{AB3-CM}. Thus, for each value of loss, the key rate $P_cK$ varies with different $T$. Here, in our all simulations we have chosen an optimal value of $T$ for each value of loss to maximise the key rate $P_cK$.} As discussed earlier, in this protocol Alice and Bob apply heterodyne detection and homodyne detection, respectively, to their own modes, followed by a RR process. First, we focus on a high-efficiency reconciliation, i.e., $\xi=0.95$. Our all simulations show the key rate in the presence of several noise sources; the input excess noise of $\varepsilon = 0.01$, and the efficiency and electronic noise of the homodyne detector $\mu = 0.526$ and $\nu_{el}=0.04361$, respectively \cite{nG1}.
The abscissa of Fig.~\ref{high} (top) corresponds to $ - 10{\log _{10}}(\int_0^{{\eta _0}} {{\eta ^2}p(\eta )\,d} \eta )$ and represents the mean fading loss in the fading channel under different conditions (different $\sigma _{b}$). The impact of a low-efficiency reconciliation is investigated in Fig.~\ref{low} (top), where we have adopted $\xi=0.8$. According to the top plots of Fig.~\ref{high} and Fig.~\ref{low}, the computed key rate resulting from the PSS state is always lower than its original TMSV state over the atmospheric fading channels.
We also simulate the performance of our CV-QKD protocol over fixed-attenuation channels.
In order to make a valid comparison, we will assume the loss in the fixed-attenuation channel is the same as the mean fading loss in the corresponding fading channel, i.e., $\tau = \int_0^{{\eta _0}} {{\eta ^2}p(\eta )\,d} \eta $. The bottom plots of Fig.~\ref{high} and Fig.~\ref{low} show the key rate over the fixed-attenuation channels as a function of channel loss, i.e., $ - 10{\log _{10}}({\tau })$ with all the settings and parameters being the same as the corresponding protocols in the top figures.
The key rate over the fading channels is always higher than that over the corresponding fixed-attenuation channels. This fact (which comes from the probabilistic nature of the fading channels) is most evident from Fig.~\ref{low} where we see the fixed-attenuation channel is not able to generate positive key rates (at high losses) for squeezing of 10dB and 16dB. This is not the case for the corresponding fading channels. From the bottom plots of Fig.~\ref{high} and Fig.~\ref{low}, we can also see that for the high squeezing regime (here 16dB) which is not experimentally achievable at the moment, the PSS state can be more effective in terms of the key rates relative to the original TMSV state for some range of losses (see also \cite{nG1, nG2}).
\begin{figure}[!t]
\begin{center}
{\includegraphics[width=2.7 in, height=3.8 in]{Beta95_fading_fixed.eps}}
\caption{The key rate (in bits per pulse) over the fading channels (top) and the corresponding fixed-attenuation channels (bottom) using the PSS state and the original TMSV state for different values of initial squeezing with $\xi=0.95$.}\label{high}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
{\includegraphics[width=2.7 in, height=3.8 in]{Beta8_fading_fixed.eps}}
\caption{The key rate (in bits per pulse) over the fading channels (top) and the corresponding fixed-attenuation channels (bottom) using the PSS state, PSS2 state, and the original TMSV state for different values of initial squeezing with $\xi=0.8$.}\label{low}
\end{center}
\end{figure}
We have also simulated the performance of the CV-QKD protocol using a state created from the incoming TMSV state by subtracting one photon from \emph{each} mode (using beam splitters with the same transmissivity $T$). Such a non-Gaussian state and its creation probability is given by (e.g., \cite{Oxford, Neda3})
\begin{equation}\label{PSS2}
\begin{array}{l}
\left| {PSS2} \right\rangle = \sum\limits_{n = 0}^\infty {{q_n}} {\left| n \right\rangle _1}{\left| n \right\rangle _2},\,\rm where\\
\\
{q_n} = \sqrt {\frac{{{{\left( {1 - {\lambda ^2}{T^2}} \right)}^3}}}{{1 + {\lambda ^2}{T^2}}}} {\left( {\lambda T} \right)^n}(n + 1),\\
\\
{P_{sb}} = \frac{{{\lambda ^2}\left( {1 - {\lambda ^2}} \right)\left( {1 + {\lambda ^2}{T^2}} \right){{\left( {1 - T} \right)}^2}}}{{{{\left( {1 - {\lambda ^2}{T^2}} \right)}^3}}}.
\end{array}
\end{equation}
The CM of the evolved $\left| {PSS2} \right\rangle$ state is again given by Eq.~\eqref{AB2-CM} but where now $x$, $y$ and $z$ are given by
\begin{equation}\label{xyz-PSS2}
x = y = \frac{{1 + 3{\lambda ^4}{T^4} + 8{\lambda ^2}{T^2}}}{{1 - {\lambda ^4}{T^4}}},\,\,\,z = \frac{{4\lambda T\left( {1 + 2{\lambda ^2}{T^2}} \right)}}{{1 - {\lambda ^4}{T^4}}}.
\end{equation}
For $\left| {PSS2} \right\rangle$ states over fading channels, we find similar trends to our previous results but with lower key rates (at higher squeezing levels) relative to the single-photon subtracted states. However, the point remains that these new bounds for the $\left| {PSS2} \right\rangle$ states are still less than the TMSV rates. Fig.~\ref{low} (top) illustrates this point. Over fixed-attenuation channels the $\left| {PSS2} \right\rangle$ states showed similar trends (not plotted due to closeness of curves).
Finally, we investigate the impact of the imperfect reconciliation process on the resulting key rates (it will suffice to use the TMSV state for this purpose).
Fig.~\ref{3d} (top) shows the key rate $K$ of the CV-QKD protocol achieved by the TMSV state with squeezing 10dB (the state-of-the-art squeezing) as a function of the reconciliation efficiency $\xi$ and as a function of the fading channel loss. This figure clearly illustrates the negative impact of the imperfect reconciliation process on the resulting key rate. The results shown in the Fig.~\ref{3d} (bottom) illustrate the key rate $K$ with reconciliation efficiency $\xi=0.8$ as a function of the initial squeezing of the TMSV state and as a function of the fading channel loss.
According to Fig.~\ref{3d} (bottom), it is evident that an increase in fading loss reduces the key rate, while increasing the initial squeezing is able to partly compensate the fading's negative effect. However, Fig.~\ref{3d} (bottom) also shows that a rise in the initial squeezing is not always able to increase the key rate. In fact, for each value of the fading loss there is an optimal value of squeezing which maximizes the key rate. This optimal squeezing is reduced by increasing the fading loss. For instance, for a mean fading loss of $0.8$dB, the optimal squeezing is about $15$dB, while for higher loss at $11.2$dB, the optimal value is around $10$dB. Note that in case of perfect reconciliation, i.e. $\xi=1$, the key rate always grows with the initial squeezing.
Not shown in Fig.~\ref{3d} is the impact of other real-world issues such as the effect of time delays caused by the reconciliation process (and other such effects). However, the impact of the reconciliation efficiency is in most circumstances the dominant effect in reducing the quantum key rates. State-of-the-art CV reconciliation efficiencies are close to values of 0.95 \cite{eff}. At these efficiencies, achievable squeezing levels will produce key rates of $10^{-3}$ bits per pulse at mean fading loss of $25$dB.
\begin{figure}[!t]
\begin{center}
{\includegraphics[width=2.7 in, height=3.8 in]{3D.eps}}
\caption{The key rate (in bits per pulse) over the fading channels using the TMSV state as a function of the mean fading loss and the reconciliation efficiency $\xi$ for squeezing 10dB (top), and also as a function of the mean fading loss and initial squeezing for $\xi=0.8$ (bottom).}\label{3d}
\end{center}
\end{figure}
\section{Conclusion}
In this work we have explored the usefulness of Gaussian and non-Gaussian entangled states over atmospheric channels towards LEO satellites. Focussing on the Gaussian component of both types of states as they evolve over the atmospheric channels, we have computed the quantum key rate for the Gaussian case and a lower bound for the non-Gaussian case.
We have also investigated how the presence of noise and imperfections of the protocol affect the key rates. Relative to the corresponding fixed-attenuation channels (such as those found in fiber), we find the fading channels (due to their stochastic nature) can lead to higher key rates. This somewhat counter-intuitive result is most noticeable for the very high-loss channels. We have also found that in some circumstances the non-Gaussian states could result in enhanced quantum key rates over the fixed-attenuation channels compared to the Gaussian states. This is the opposite of what we find in the atmospheric fading channels, once again highlighting the impact the atmosphere can have on the quantum key outcomes. We caution that our rates for the non-Gaussian states are lower bounds - the true key rate for such states remains an open problem. If the bounds shown are here not tight the relative performance levels discussed here may vary.
The role played by the efficiency of the classical reconciliation process, as a function of the mean fading loss and the initial squeezing of the Gaussian states, was also investigated. For state-of-the-art squeezing levels (10dB) and state-of-the-art reconciliation efficiencies (0.95) we found ground-to-satellite quantum key rates of roughly $10^{-3}$ bits per pulse are viable. This channel will be the limiting channel in a ground-satellite-ground relay system. Due to the reduced beam wandering losses incurred from space to Earth, the key rates of the satellite-to-ground channels will be roughly an order of magnitude larger.
The calculations presented here should be of value in the quantitative assessment of the CV-QKD protocols using Gaussian and non-Gaussian resources for future quantum communications over atmospheric channels.
|
\section{Introduction \label{introduction}}
One of the biggest mysteries of the Fe-based superconductors is the striking difference between the magnetic phase of the pnictide and chalcogenide compounds. The parent compounds of the Fe-pnictide superconductors show a magnetic state with a $(\pi,0)$ spin density wave (SDW) vector which gives way to superconductivity upon doping and/or application of pressure~\cite{RevModPhys.83.1589}. The Fermi surfaces in both the pnictides and chalcogenides in general show common features such as hole pockets at the $\Gamma$ point and electron pockets at the X and Y points of the Brillouine Zone~\cite{PhysRevLett.100.237003,PhysRevLett.101.026403,PhysRevLett.101.057003,PhysRevLett.101.177005,PhysRevB.79.054517,0295-5075-83-4-47001,PhysRevLett.101.216402}. These pockets are well-nested at the wave vector of $(\pi,0)$ or $(0,\pi)$. This is also the wave vector of the SDW, and hence band nesting is commonly accepted as the mechanism of the SDW in the pnictide compounds~\cite{PhysRevB.81.024511}.
However the chalcogenides show a completely different magnetic structure, with a SDW wave vector of $(\pi/2,\pi/2)$ for zero or low doping, forming the so-called double-striped (DS) phase and hence the origin of the SDW in the chalcogenides cannot be explained by the nesting of the hole and electron bands~\cite{PhysRevB.79.054503}. The origin of the SDW in the chalcogenides is still under hot debate. Among the chalcogenides, the compound Fe$_{1+y}$Te poses a particular mystery. Apart from the fact that the SDW is different from the pnictides, it has been pointed out that a generic commensurate $(\pi/2,\pi/2)$ SDW is a superposition of ${\bf Q}_{1/2}=(\pi/2,\pm\pi/2)$
wave vectors~\cite{PhysRevLett.109.157206}. It was shown in the latter reference that quantum fluctuations in a localized model stabilize the so-called orthogonal double stripe (ODS) phase which contains both these wave vectors.
Experimentally, it was shown that Fe$_{1+y}$Te tends to order in the ODS state based on the neutron scattering structure factor~\cite{PhysRevLett.107.216403}. Theoretical studies such as exact diagonalization~\cite{1742-6596-200-2-022058} and mean field studies of the $t$-$J$ model~\cite{PhysRevB.86.134512} also favor the ODS state.
A recent spin-polarized scanning tunnelling microscopy (STM) study confirmed the DS phase at low excess iron, but suggested the ODS phase at higher $y$ values~\cite{Enayat08082014}. The magnetic order becomes even more complex upon increasing the excess iron.
In particular, recent neutron diffraction results show an incommensurate SDW state~\cite{PhysRevLett.102.247001}, and also possibly a helical state~\cite{PhysRevLett.115.177203}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.3,angle=0]{./Figures/Fig1_abc.jpg}\\
\vspace*{-1.cm}
\hspace{0.45cm}
\includegraphics[scale=0.5,angle=0]{./Figures/Fig1_de.pdf}
\end{center}
\vspace{-1.6cm}
\caption{The magnetic structure with $(\pi/2,\pi/2)$ SDW wave vector with (a) double striped structure ($\phi=\pi/4$), and (b) extended-stripe phase ($\phi=0$).
(c) Shows the ODS phase with both $(\pi/2,\pi/2)$ and $(\pi/2,-\pi/2)$ ordering wave-vectors.
(d) The magnetic Brillouin zone (red rectangle) for an SDW with ${\bf Q}_1=(\pi/2,\pi/2)$ and (e) the magnetic Brillouin zone (red square) for an
SDW with ${\bf Q}_{1/2}=(\pi/2,\pm\pi/2)$. The black square is the fulll Brillouin zone for the paramagnetic case.}
\label{fbzmbzc2c4}
\end{figure}
Due to the close proximity of magnetism and superconductivity in the iron-compounds, the superconductivity is believed to be mediated by spin fluctuations rather than lattice vibrations~\cite{0034-4885-74-12-124508}. Hence understanding the magnetic phase is important to understand the superconducting state. The investigations of the quasiparticle excitation can give valuable information on the nature of superconductivity since they are intimately connected to the superconducting gap, $\Delta_{\bf k}$. These quasiparticles can be detected by momentum-sensitive techniques such as ARPES to find the quasiparticle spectrum and the quasiparticle density of states. Among phase-sensitive techniques such as Josephson tunnelling, SQUID interferometry, etc., the one that directly investigates quasiparticle effects is the quasiparticle interference (QPI) spectroscopy based on scanning tunneling microscopy (for a review see~\cite{0034-4885-74-12-124513}). In this method, interference of quasiparticles due to random impurities in the sample is detected through the spatial modulation of the local quasiparticle density of states at a constant energy bias. This density of states corresponds to the differential conductance which is measured by STM, and the Fourier transform of these real space data then gives the wave vectors at which the dominant scatterings occur. QPI is a powerful tool because it simultaneously yields energy-dependent real-space and momentum-space information on the quasiparticle wave-functions, scattering processes and coherence factors. This information can be used to distinguish between different superconducting
order parameters~\cite{0034-4885-74-12-124513,PhysRevLett.71.3363,0295-5075-85-3-37005,Hanaguri23042010,Hanaguri13022009,Allan:2012aa,Chi:2014aa,Akbari:2010aa,PhysRevLett.104.257001,Yamakawa:2015aa,Hirschfeld:2015aa,Sykora:2011aa,Akbari:2011aa,0295-5075-100-3-37004,Akbari:2013aa,0295-5075-103-2-27004,0295-5075-100-3-37002,Akbari:2014aa,doi:10.7566/JPSJ.83.061015}. In addition, QPI can also access the momentum space structure of the unoccupied states that are inaccessible to photoemission~\cite{McElroy1}.
Given the complicated scenario regarding the magnetic phases of Fe$_{1+y}$Te, it is natural to ask if one can apply a non-magnetic method to detect these phases. In this article we explore the possibility of using non-magnetic-QPI as a possible tool to detect the magnetic phases of this compound. QPI can be a valuable tool to characterize the different possible magnetic structures since the magnetic states will influence the band structure which will in turn influence the QPI features. To illustrate the possibility, we undertake a theoretical investigation of QPI for three different kinds of magnetic structures involving the ${\bf Q}_1$ and ${\bf Q}_2$ wave vectors. For low $y$, only one of either ${\bf Q}_1$ or ${\bf Q}_2$ is present with a phase-angle $\phi=\pi/4$ which results in the DS phase shown in Fig.~\ref{fbzmbzc2c4}(a). This is the first phase that we consider. For the same wave vector, one can also theoretically change the magnetic structure by varying the phase angle $\phi$, which thus offers a continuous parameter which determines the magnetic structure. The second magnetic phase we consider is with ${\bf Q}_1$ and $\phi=0$. The third magnetic phase we consider is the ODS phase discussed in~\cite{PhysRevLett.109.157206}. We consider the five-orbital tight-binding model of Ducatman, et. al\cite{PhysRevB.90.165123}. This model is meant for the Fe$_{1+y}$Te compounds and explicitly considers the effect of the additional interstitial Fe atoms. \\
{\it Spin density wave with ${\bf Q}_1$ wave vector:}
\label{secone}
In the following, we present the mean field frame work for general SDW with ${\bf Q}_1$.
For this, the mean-field ansatz for the electronic density in orbital $\gamma$ and site $i$ can be written as
\begin{equation}
\langle n_{i\gamma\sigma} \rangle = \frac{n_\gamma}{2}+\sigma\frac{m_\gamma}{2}\cos\left(
{\bf Q}_1\cdot{\bf r}_i + \phi\right),
\label{ek}
\end{equation}
where $\phi$ is a general phase angle, and
$n_\gamma$ is the total (spin-up + spin-down)
number of electrons in orbital $\gamma$ at lattice site $i$. The magnetisation is assumed to point along the $z$-direction, so that
$
2\langle S_{i\gamma z} \rangle = m_\gamma\cos\left(
{\bf Q}_1\cdot{\bf r}_i + \phi\right).
$
Our starting point for finding the mean field solution for the magnetic phase is the Hamiltonian,
$
H=H_{TB}+H_{int},
$
where $H_{TB}$ is the kinetic energy (tight-binding) part and $H_{int}$ is the interaction. The latter is given by
\begin{eqnarray}
\begin{aligned}
H_{int} =
& U \sum_{i\gamma} n_{i\gamma\uparrow} n_{i\gamma\downarrow}
+U'
\sum_{i\sigma{\tilde \sigma} \gamma>\beta} n_{i\gamma\sigma}
n_{i\beta{\tilde \sigma}}
\\
&- 2J \sum_{i\gamma>\beta} {\bf S}_{i\gamma}\cdot {\bf S}_{i\beta},
\end{aligned}
\label{tri}
\end{eqnarray}
where $\gamma$ is an orbital index and $\sigma$ refers to the spin. Here $U$ and $U'$ are the intra- and inter-orbital Coulomb repulsions, respectively, and $J$ is the Hund's coupling.
Using Eq. (\ref{ek}), the total mean field Hamiltonian reduces to
\begin{equation}
H^{MF}_{int} = H_U + H_{U'} + H_J,
\label{chatur1}
\end{equation}
where
\begin{eqnarray}
\begin{aligned}
&H_U = -\frac{U}{6}\sum_{\gamma\sigma}\sigma m_\gamma (n_{{\bf Q}\gamma\sigma}e^{i\phi}+n_{-{\bf Q}\gamma\sigma}e^{-i\phi}),
\\
&H_{U'} = \left( U'-\frac{J}{2} \right)\sum_{{\bf k}\sigma,\gamma\neq\beta}n_\beta a^\dagger_{{\bf k}\gamma\sigma}a_{{\bf k}\gamma\sigma},
\\
&
H_J = -\frac{J}{4}\sum_{\sigma,\gamma\neq\beta}\sigma m_\beta (n_{{\bf Q}\gamma\sigma}e^{i\phi} + n_{-{\bf Q}\gamma\sigma}e^{-i\phi}),
\label{panch}
\end{aligned}
\end{eqnarray}
and $n_{{\bf Q}\gamma\sigma} = \sum_{\bf k} a^\dagger_{{\bf k}\gamma\sigma}a_{{\bf k}+{\bf Q},\gamma\sigma}$. The magnetic Brillouin zone (MBZ) for the SDW modulated by the ${\bf Q}_1$- wave vector is shown by the red line in Fig.~\ref{fbzmbzc2c4}(d). \\
The procedure for solving the mean-field Hamiltonian self consistently has been described in~\cite{PhysRevB.79.104510} and consists of the following steps. Initial trial values of $n_\gamma$ and $m_\gamma$ are chosen and the Hamiltonian is diagonalized for all points in the MBZ for these values. The Fermi energy corresponding to the given electronic filling is then determined. Using the eigenvectors, eigenvalues, and the Fermi energy so obtained, new values of $n_\gamma$ and $m_\gamma$ are calculated.
These are fed as input into the next iteration and the process is repeated until self-consistency is reached for a given temperature. \\
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{./Figures/Fig2.pdf}
\end{center}
\vspace{-0.5cm}
\caption{(a) The Fermi surface for $n=6.64$ electrons and $U=2.125$ eV in the full Brillouin zone for the double stripe ($\phi=\pi/4$) magnetic order, showing the electron pocket at the $\Gamma$ point. There are additional electron pockets identified by the wave vectors ${\bf q}_1$, ${\bf q}_1'$ and ${\bf q}_2'$, which can be identified with the scattering wave vectors in the QPI patterns of Fig.~\ref{perkins_qpidisp}. (b) Fermi surfaces for the extended-stripe phase ($\phi=0$) case, and (c) for the orthogonal double stripe (ODS) case.
}
\label{fig2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\linewidth, angle=0]{./Figures/Fig3_abcd.pdf}
\includegraphics[width=1\linewidth, angle=0]{./Figures/Fig3_efgh.pdf}
\caption{
(a-d) Spectral intensity maps for energies -50, 0, 50 and 100meV for the double stripe magnetic order.
The electron pockets associated with ${\bf q}_1$, ${\bf q}_1'$ and ${\bf q}_2'$ (shown in Figs.~\ref{fig2} and~\ref{perkins_qpidisp})
appear around -50 meV and increase in size as the energy increases. These wave vectors correspond to the electron-like dispersions in Fig.~\ref{perkins_qpidisp}.
%
(e-h) Corresponding QPI maps for energies -50, 0, 50 and 100meV for the double stripe magnetic order.}
\label{qpiakwds}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{./Figures/Fig4.pdf}
\end{center}
\vspace*{-0.6cm}
\caption{QPI dispersion along the $(0,0)\rightarrow (-\pi,\pi)$, and $(0,0)\rightarrow (\pi,\pi)$ directions for the double stripe magnetic order. The dominant scattering wave vectors ${\bf q}_1$, ${\bf q}_1'$ and ${\bf q}_2'$ are shown and correspond to the wave vectors connecting the Fermi surfaces in Fig.~\ref{fig2}.
}
\label{perkins_qpidisp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\linewidth, angle=0]{./Figures/Fig5_abcd.pdf}
\includegraphics[width=1\linewidth, angle=0]{./Figures/Fig5_efgh.pdf}
\caption{
(a-d) Spectral intensity maps for four different energies in the full BZ for the extended-stripe ($\phi=0$) case. The electron pocket with the flat cusp is reflected by the sudden appearance of the electron pocket at -70meV.
%
(e-h) QPI maps in the full BZ for four different energies for the extended-stripe case.}
\label{fig5}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth,angle=0]{./Figures/Fig6.pdf}
\end{center}
\vspace*{-0.6cm}
\caption{The QPI dispersion along the (a) $(-\pi,\pi)$ direction and (b) along the $(\pi,\pi)$ direction for $\phi=0$.
}
\label{qpidisp_phi0}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth, angle=0]{./Figures/Fig7_abcd.pdf}
\includegraphics[width=1\linewidth, angle=0]{./Figures/Fig7_efgh.pdf}
\caption{
(a-d) Spectral intensities at the energies indicated in the BZ double the size of the MBZ.
%
(e-h) Corresponding QPI maps for four different energies in the ODS phase in the full BZ.}
%
\label{ods_akw}
\label{ods_qpi}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\linewidth,angle=0]{./Figures/Fig8.pdf}
\end{center}
\vspace*{-0.6cm}
\caption{QPI dispersion along the (a) $(\pi,0)$ direction and (b) $(\pi,\pi)$ direction for the ODS phase. For (a), there is a hole feature (${\bf q}_1$) starting from about 70 meV, an electron feature starting from about 0 meV, a feature at -60 meV, and an electron feature at -100 meV (${\bf q}_2$). There is also a faint hole feature at about -25 meV. For (b), there is a hole feature at 70 meV (${\bf q}_1'$), an electron feature at 0 meV, a feature at -60 meV and an electron feature at -100 meV (${\bf q}_2'$).
}
\label{ods_qpidisp_both}
\end{figure}
{\it Theoretical formulation of QPI}:
The scattering off a magnetic impurity is modelled by the Hamiltonian
\begin{equation}
H_{imp}=\sum_{{\bf k}{\bf k'}\mu\nu\sigma\sigma'}c^{\dagger}_{\mu{\bf k}\sigma}(J^{\mu\nu}_{\sigma\sigma'}{\bf S}\cdot {\mbox{{\boldmath$\sigma$}}}_{\sigma\sigma'})c_{\nu{\bf k'}\sigma'},
\end{equation}
where ${\bf S}$ is the impurity spin. We take impurity spin along $z$-direction and assume purely localized impurity scattering.
Introducing the matrices
${\hat U} = \sigma_z \otimes {\hat J}$,
and
${\hat T}
=[1-{\hat U}\sum_{\bf k}G_0({\bf k},\omega)]^{-1}{\hat U}
$,
the QPI under the $T$-matrix approximation is given by~\cite{Balatsky:2006aa}
\begin{eqnarray}
\begin{aligned} %
{\rm QPI}
\sim
{\rm Im}
\sum_{{\bf k}\sigma}
{\rm Tr}
\Big[
G_{0\sigma}({\bf k},\omega){\hat T}_\sigma G_{0\sigma}({\bf k}+{\bf q},\omega)
\Big],
\end{aligned}
\end{eqnarray}
where $\otimes$ is the direct product and ${\hat J}$ is a matrix in the purely orbital basis: ${\hat J}_{\mu\nu}=J^{\mu\nu}_{\uparrow\uparrow}=J^{\mu\nu}_{\downarrow\downarrow}$.\\
{\it Numerical Results for the double stripe phase:}
This particular magnetic structure is exhibited by the iron-chalcogenides and is obtained by putting $\phi=\pi/4$ in Eq.~(\ref{ek}). In this section we present numerical results for Fe$_{1+y}$Te in the mean-field formulation. We employ the 5-orbital tight binding model of~\cite{PhysRevB.90.165123} and consider Fe$_{1.08}$Te which, assuming that each excess Fe atom contributes eight electrons corresponds to an electronic density of $n=6.64$.
We should emphasise that the chemical potential shifts by about 0.4 eV in the $y=0.08$ case compared to $y=0$
and that the hole pocket at $\Gamma$ is replaced by an electron pocket~\cite{PhysRevB.90.165123}.
This latter feature has been observed by ARPES in other electron doped Fe compounds as well~\cite{PhysRevB.84.020509}.
For the mean field calculation we fix $J=0.25U$ and carry out the self-consistent calculation for the magnetization.
The magnetic moment in Fe$_{1+y}$Te is around $2.1\mu_B$~\cite{PhysRevB.79.054503}. Further, from LDA and ARPES, the band structure in the magnetic phase of Fe$_{1+y}$Te is known to have an electron pocket at the $\Gamma$ point~\cite{PhysRevLett.111.217002}. From our mean-field analysis, a magnetic moment of $2.1 \mu_B$ corresponds to about $2.2$~eV. There is a hole like feature at the $\Gamma$ point for this value of $U$, which is replaced by an electron like feature at lower values.
%
Thus in order to maintain consistency of our results with LDA and ARPES~\cite{PhysRevLett.111.217002}, we choose $U=2.125$~eV. The magentization has a somewhat lower value of $1.02\mu_B$ for this interaction, but the Fermi surface so obtained indeed shows an electron pocket at $\Gamma$ (see Fig.~\ref{fig2}(a)).
The spectral functions for four different energies and the corresponding QPI maps are shown in Fig.~\ref{qpiakwds}.
One can trace the main momentum scattering vectors, shown by the ${\bf q}_i$ in Fig.~\ref{fig2}(a), in the QPI maps of Fig.~\ref{qpiakwds}.
In Fig.~\ref{perkins_qpidisp} we show the ``QPI dispersion'', i.e. the QPI intensities as a function of frequency and momentum, along the $(\pi,-\pi)$- and $(\pi,\pi)$-directions in the full Brillouin zone (BZ).
In Fig.~\ref{perkins_qpidisp} we see an electron like feature starting from about -0.05 eV along the $(-\pi,\pi)$ direction. We have labeled the main momentum characterizing this by ${\bf q}_1$. We also have an electron like feature starting from the same energy, but at lower momenta, ${\bf q}_1'$, along the $(\pi,\pi)$ direction. Along the latter direction there is also an electron-like feature starting from the same energy but at higher momenta, ${\bf q}_2'$. The origin of these features is clear from Fig.~\ref{fig2} (a) where we have labeled these momenta on the Fermi surface. Thus these features can be related to the ``twin'' electron pockets connected by ${\bf q}_2'$. These ``twin'' pockets indeed start at an energy about -0.05 eV as can be seen from Fig.~\ref{qpiakwds}.\\
{\it Numerical Results for the extended-stripe phase:}
The extended-stripe phase is obtained by setting $\phi=0$ in Eqs.~(\ref{ek}-\ref{panch}). The corresponding magnetic structure is shown in Fig.~\ref{fbzmbzc2c4}(b) and consists of two interpenetrating antiferromagnetic sublattices and a sublattice with zero magnetic moment. We repeat the self-consistent calculation for $n=6.64$ electrons. To have a magnetization similar to the $\phi=\pi/4$ case, we choose in this case $U=2$ eV, for which the magnetization equals $0.91\mu_B$. The corresponding Fermi surface is shown in Fig.~\ref{fig2}(b). The Fermi surface consists of a hole pocket at the $\Gamma$ point, and electron-pockets with a relatively flat cusp at ${\bf Q}_2$.
Fig.~\ref{fig5} shows the spectral intensities at four different energies in the full BZ. The electron pocket with the flat cusp makes a sudden appearance at around -70 meV around the $(-\pi/2,\pi/2)$ point (see Fig.~\ref{fig5}(b)). Furthermore Fig.~\ref{fig5} shows the corresponding QPI maps for these energies in the full BZ. At -70 meV there is an abrupt qualitative shift in the QPI map around $\Gamma$, coincident with the appearance of the electron pocket with the flat cusp. In Fig.~\ref{qpidisp_phi0} we have shown the QPI dispersion along the $(\pi,\pi)$ and $(-\pi,\pi)$ directions. Along both directions we see an electron like feature with a heavy mass which corresponds to the electron pocket with the broad cusp in the quasiparticle dispersion.
This flat pocket is also present in the DS phase, however there it is located at a much lower energy than that considered in the QPI dispersion here and hence does not show up for that case.
{\it The orthogonal double stripe phase:}
This phase is shown in Fig.~\ref{fbzmbzc2c4}(c) and has been discussed in Ref.~\cite{PhysRevLett.109.157206}. The MBZ is the small diamond in the center in Fig.~\ref{fbzmbzc2c4}(e). Unlike the other two phases which had $C_2$ symmetry this phase possesses $C_4$-tetragonal symmetry. It can be described by the following mean-field ansatz:
\begin{equation}
\langle n_{i\gamma\sigma} \rangle = \frac{n_\gamma}{2}+\sigma\frac{m_\gamma}{2}
[
\sin\left(
{\bf Q}_1\cdot{\bf r}_i\right) + \cos \left(
{\bf Q}_2\cdot{\bf r}_i\right)
],
\label{ek_ods}
\end{equation}
In this case the magnetization at site $i$ is given by
$
\langle S_{i\gamma z} \rangle = m_\gamma \left( \sin {\bf Q}_1\cdot{\bf r}_i + \cos {\bf Q}_2\cdot{\bf r}_i\right)/2
$.
Under this ansatz the mean field Hamiltonian for the interaction is given by Eq.~(\ref{chatur1}) with
\begin{eqnarray}
\begin{aligned}
H_U
&=
\frac{iU}{6}
\sum_{\sigma \gamma}
\sigma m_{\gamma}
(n_{{\bf Q}_1\gamma\sigma}-n_{{\bf Q}_2\gamma\sigma}) +h.c.,
\\
H_J & = \frac{J}{4}
\sum_{\sigma,\gamma\neq\beta}
\sigma m_\beta(i n_{{\bf Q}_1\gamma\sigma}-n_{{\bf Q}_2\gamma\sigma})+h.c.,
\end{aligned}
\end{eqnarray}
and $H_{U'}$ being the same as earlier.
Based on the self-consistent results for the magnetization as a function of $U$, with $n=6$ and $J=0.25U$, we choose $U=2$ eV which corresponds to a magnetization of 2.54$\mu_B$. The Fermi surfaces are shown in Fig.~\ref{fig2}(c) in a region twice the size of the MBZ and rotated by 45$^{\rm o}$. The Fermi surface consists of a hole pocket at $\Gamma$. We show the spectral intensities at four different energies in Fig.~\ref{ods_akw}. The corresponding QPI maps for these energies are also shown in Fig.~\ref{ods_qpi}, while the QPI dispersion is shown in Fig.~\ref{ods_qpidisp_both}. Figure~\ref{ods_qpidisp_both}(a) shows the cut along the $(\pi,0)$ direction and Fig.~\ref{ods_qpidisp_both}(b) along the $(\pi,\pi)$ direction.
There are several features in the spectral intensities which give rise to a more complex QPI pattern and QPI dispersion compared to the DS and extended-stripe cases.
In Fig.~\ref{ods_qpidisp_both}(a) there is a hole feature starting from about 70meV, an electron feature starting from about 0meV, a feature at -60meV, and an electron feature at -100meV. There is also a faint hole feature at about -25meV. All of these can be understood in terms of the spectral intensities of Fig.~\ref{ods_akw}.
%
As an example, the hole dispersion marked by ${\bf q}_1$ and the electron dispersion marked by ${\bf q}_2$ in Fig.~\ref{ods_qpidisp_both}(a) can be associated with ${\bf q}_1$ and ${\bf q}_2$ in Fig.~\ref{ods_akw}(d) and (a) respectively. Similarly ${\bf q}_1'$ and ${\bf q}_2'$ in Fig.~\ref{ods_qpidisp_both}(b) can be associated with ${\bf q}_1'$ and ${\bf q}_2'$ in Fig.~\ref{ods_akw}(d) and (a) respectively.
In this article, using QPI as a possible tool to detect the magnetic structure based on scattering of quasipartciles from nonmagnetic impurities, we theoretically investigated three different possible magnetic phases for the parent compounds of the iron-chalcogenide superconductors.
Most importantly, we have shown that different magnetic phases realized in iron chalocgenides can be easily identified by means of QPI from non-magnetic impurities due to very different electronic structure reconstruction in all these phases.
The first phase considered was the experimentally observed DS phase, which has a SDW wave vector ${\bf Q}_1$ and a phase angle of $\phi=\pi/4$. The second phase was the extended stripe phase which has $\phi=0$, and the third phase was the ODS phase which consists of both ${\bf Q}_1$ and ${\bf Q}_2$ wave vectors. We used a tight binding model developed for Fe$_{1+y}$Te and a multi-orbital interaction term to obtain the magnetic states self-consistently. For the first two states we considered an electronic filling of 6.64 electrons and a magnetic moment of about 1$\mu_B$ to match the electron pocket experimentally seen for the DS phase. For the ODS phase we chose 6 electrons and a magnetic moment of 2.5$\mu_B$. The QPI calculations were carried out using the $T$-matrix approximation.
In the DS phase, the Fermi surfaces consist of an electron pocket at $\Gamma$ and twin-electron pockets at the corners of the magnetic BZ. The QPI features are mainly dominated by these twin pockets as seen in the QPI dispersion plots which predominantly show an electron-type dispersion. For the extended stripe phase, the Fermi surface consists of a hole pocket at $\Gamma$, and an electron type dispersion with a flat cusp in the $(-\pi,\pi)$ direction. It is predominantly this feature which is reflected in the QPI behavior. The scattering across this electron-pocket dominates the QPI intensity map and results in the electron-feature in the QPI dispersion.
The ODS phase was relatively more complicated than the other two phases considered. The scatterings between various contours of constant energy resulted in the features seen in the QPI dispersion.\\
{\it Acknowledgments:}
We thank P. Wahl, A. Yaresko, P. Thalmeier, M.N. Gastiasoro, B.M. Andersen, and M. Enayat for helpful comments and for useful discussions. We are grateful to the Max Planck Institute for the Physics of Complex Systems (MPI-PKS) for the use of computer facilities.
B.K. and A.A. wish to acknowledge the Korea Ministry of Education, Science and Technology, Gyeongsangbuk-Do and Pohang City for Independent Junior Research Groups at the Asia Pacific Center for Theoretical Physics.
The work by B.K. and A.A. was supported through NRF funded by MSIP of Korea (2015R1C1A1A01052411). A.A. acknowledges support by Max Planck POSTECH / KOREA Research Initiative (No. 2011-0031558) programs through NRF funded by MSIP of Korea.
The work of IE was supported by the Focus Program 1458
Eisen-Pniktide of the DFG. IE acknowledges support by the Ministry of
Education and Science of the Russian Federation in the
framework of Increase Competitiveness Program of NUST
MISiS (N 2-2014-015).
|
\section{Introduction}\label{sec:intro}
Iterated function systems are often used to construct fractals. Given a metric space, a classical \emph{iterated function system} is a finite set of functions
$$\left\{f_i:X\rightarrow X\ \big|\ i=1,2,\dots,N\right\}\quad (N\in\mathbb{N})$$
such that each $f_i$ is a contraction. In \cite{Barnsley:2011dy}, Barnsley and Vince showed that the IFSs of noncontractive type (i.e. composed of maps that are not contractions with respect to any topologically equivalent metric in X) can yield attractor. These arise naturally in projective spaces, though classical irrational rotation on the circle can be adapted too.
In the present paper, we will construct a converging geometric IFS for which the functions are not contraction maps under the natural metric.
We study the limit behavior of iteratedly dividing a triangle in a natural way. Let $\mathcal{T}$ be the moduli space of hyperbolic/Euclidean triangles modulo (ordered) similarities. A hyperbolic triangle is given in Figure \ref{fig:hyperbolic_triangle}.
\begin{figure}[htbp]
\begin{center}
\includegraphics{hyperbolic_triangle}
\end{center}
\begin{minipage}{0.9\textwidth}
\caption{\label{fig:hyperbolic_triangle}
{\bf A hyperbolic triangle in the upper plane model.}}
\end{minipage}
\end{figure}
An ordered triangle is parametrized by the (ordered) triple of its three angles. $\mathcal{T}$ is naturally identified with the tetrahedron $\left\{(x,y,z)\in\mathbb{R}^3\ \big|\ x,y,z>0, x+y+z\leqslant\pi\right\}$ in $\mathbb{R}^3$. Given a hyperbolic triangle $\triangle_{ABC}$, the three \emph{mid-lines} joining the midpoints of its edges divide it into four smaller triangles.
\begin{figure}[htp]
\center{\includegraphics[width=400pt]{defining_triangle}}\\
\begin{minipage}{0.9\textwidth}
\caption{\label{fig:defining_triangle}
{\bf The defining functions} The shadowed triangles are $f_A(\triangle_{ABC})$ and $f_M(\triangle_{ABC})$ respectively. }
\end{minipage}
\end{figure}
Let $f_A, f_B, f_C$ and $f_M$ be functions on $\mathcal{T}$ that maps $\triangle_{ABC}$ to one of the smaller triangles, with vertices ordered in the natural way so that they are identical for Euclidean ones, as indicated in Figure \ref{fig:defining_triangle}. Our main result is the following
\begin{thm}\label{thm:main}
The iterated function system $\{f_A,f_B,f_C,f_M\}$ on $\mathcal{T}$ is converging and the limit functions are continuous.
\end{thm}
The functions can be determined as follows. Let $a, b, c$ be the length of edges $BC$, $CA$ and $AB$ respectively and define $a^\prime$, $b^\prime$ and $c^\prime$ to be the length of $f_M(\triangle_{ABC})$ in the similar way. Then $a^\prime, b^\prime$ and $c^\prime$ are related to $A$, $B$ and $C$ by
\begin{align}\label{eqn:edge_angle_relation}
\begin{split}
\cos(A)=\frac{\cosh(b^\prime)\cosh(c^\prime)-\cosh(a^\prime)}{\sinh(b^\prime)\sinh(c^\prime)},\\
\cos(B)=\frac{\cosh(c^\prime)\cosh(a^\prime)-\cosh(b^\prime)}{\sinh(c^\prime)\sinh(a^\prime)},\\
\cos(C)=\frac{\cosh(a^\prime)\cosh(b^\prime)-\cosh(c^\prime)}{\sinh(a^\prime)\sinh(b^\prime)}.
\end{split}
\end{align}
And the edge lengths $a$, $b$ and $c$ are related to $a^\prime$, $b^\prime$ and $c^\prime$ by
\begin{equation}\label{eqn:edge_length_iteration}
\cosh(a^\prime) =\cosh\left(\frac{a}{2}\right)\cdot\mu,\quad \cosh(b^\prime) =\cosh\left(\frac{b}{2}\right)\cdot\mu,\quad \cosh(c^\prime) =\cosh\left(\frac{c}{2}\right)\cdot\mu,
\end{equation}
where
$$\mu:=\frac{1-\displaystyle{\tanh\frac{a+b+c}{4} \tanh\frac{a+b-c}{4} \tanh\frac{c+a-b}{4} \tanh\frac{b+c-a}{4}}}
{1+\displaystyle{\tanh\frac{a+b+c}{4} \tanh\frac{a+b-c}{4} \tanh\frac{c+a-b}{4} \tanh\frac{b+c-a}{4}}}.$$
We use Equation \ref{eqn:edge_angle_relation} to compute the edge lengths of $f_M(\triangle_{ABC})$. Then the edge lengths of the image triangle $\triangle_*$ of $\triangle_{ABC}$ under $f_A$, $f_B$, $f_C$ or $f_M$ is known. Now Equation \ref{eqn:edge_length_iteration} determines the edge lengths of the triangle $f_M(\triangle_*)$ and Equation \ref{eqn:edge_angle_relation} then compute the angles of $\triangle_*$ from the edge lengths of $f_M(\triangle_*)$.
We remark that the corresponding result holds equally well for spherical triangles. We may also ask the following questions. Are the limit functions smooth? Does the limit function give a half-line bundle structure over $\mathcal{T}$?
The papers is organized as follows. In Section 2, we study the iterated behaviours of the edge lengths, areas and angles. In Section 3, we show the continuity of the limit functions.
\section{Iterated division}
We refer to \cite{Casey:1889} for formulas in hyperbolic geometry.
Let $\{\triangle_n\}_{n=0}^\infty$ be a sequence of iterated action of the functions $f_A,f_B,f_C$ and $f_M$ on the triangle $\triangle_0=\triangle_{A_0B_0C_0}$, as illustrated in Figure \ref{fig:iteration_triangle}.
\begin{figure}[htbp]
\center{\includegraphics[width=400pt]{iteration_triangle}}\\
\begin{minipage}{0.9\textwidth}
\caption{\label{fig:iteration_triangle}
{\bf The inductive definition.} The shadowed triangle is $\triangle_n$. }
\end{minipage}
\end{figure}
We use $A_n$ to denote the angle $\angle B_nA_nC_n$, and $a_n$ the edge $B_nC_n$ (or its length). We define $B_n$, $C_n$, $b_n$ and $c_n$ similarly. Let $S_n$ be the (hyperbolic) area of $\triangle_n$. We have $A_n+B_n+C_n=1-S_n$.
The total angle will converge to $\pi$ since the area will converge to zero. It is then prone to think that each angle will increase and share the angle defect. However, this is not the case. For example, for an isosceles triangle $\triangle_*$ with edges lengths $4$, $4$ and $7$, the apex angle will increase while the bottom angles will decrease for $f_M$. So it is not trivial that the sequence will automatically converge to a nondegenerate Euclidean triangle. The distance of $\triangle_*$ from the fixed point $\left(\displaystyle{\frac{\pi}3,\frac\pi3,\frac\pi3}\right)$ will increase. Hence $f_M$ is NOT a contraction map.
The following lemma estimate the edge length changing under the IFS.
\begin{lemma}\label{lemma:edge_estimate}
For the sequence $\left\{\triangle_n\right\}_{n=0}^\infty$, we have
$$\sinh\frac{a_{n+1}}2<\frac12\sinh\frac{a_{n}}2.$$
Under the assumptions $\max\left\{\displaystyle{\sinh\frac{a_0}2}, \displaystyle{\sinh\frac{b_0}2}, \displaystyle{\sinh\frac{c_0}2}\right\}<\sigma$, we have
$$\left(\frac1{e\sqrt{e}}\right)^\sigma\cdot\frac{1}{2^n}\sinh\frac{a_0}2<\sinh\frac{a_{n}}2<\frac1{2^n}\sinh\frac{a_{0}}2.$$
\end{lemma}
\begin{proof} In the triangle $\triangle_n$, let $\ell$ be the line passing through the midpoints $C_{n+1}$ on $A_nB_n$ and $B_{n+1}$ on $A_nC_n$. Draw lines through $B_n$ and the midpoint $A_{n+1}$ of $B_nC_n$ to $\ell$. Let $E$ and $G$ be the intersection points respectively. Then we get the \emph{Lambert quadrilateral} as in Figure \ref{fig:lambert_quadrilateral}.
\begin{figure}[htbp]
\begin{center}
\includegraphics{lambert_quadrilateral}
\end{center}
\begin{minipage}{0.9\textwidth}
\caption{\label{fig:lambert_quadrilateral}
{\bf The Lambert quadrilateral.}}
\end{minipage}
\end{figure}
The segment $A_{n+1}G$ is also perpendicular to $B_nA_{n+1}$. By congruence of triangles, the length of $EG$ is the same as the length of the mid-line $B_{n+1}C_{n+1}$. We have
$$\sinh\left(\frac{a_n}2\right)=\sinh(a_{n+1})\cosh(l_n)=2\sinh\frac{a_{n+1}}2\cosh\frac{a_{n+1}}2\cosh{l_n}$$
Therefore
$$\sinh\left(\frac{a_{n+1}}2\right)\left/\sinh\left(\frac{a_n}2\right)\right.=\frac{1}{2\cosh(a_{n+1}/2)\cosh(l_n)}<\frac12$$
When $\displaystyle{a_{n+1}=\frac12a_n}$, the inequality comes from the convexity of the function $\sinh(x)$.
Therefore, we may suppose $\displaystyle{\sinh\frac{a_0}2<1}$, $\displaystyle{\sinh\frac{b_0}2<1}$ and $\displaystyle{\sinh\frac{c_0}2<1}$. Then we have
$$\sinh\frac{a_n}2<\frac{1}{2^n},\quad\sinh\frac{b_n}2<\frac1{2^n},\quad\sinh\frac{c_n}2<\frac1{2^n}$$
When $a_n$ joins the midpoint, the Lambert quadrilateral tells that $l_n<\displaystyle{\frac{b_n}2}$, and hence we have
we have
$$\sinh^2\frac{a_n}2=\frac{1}{4\cosh^2(a_n/2)\cosh^2(l_{n-1})}\cdot\sinh^2\frac{a_{n-1}}2>\frac{1}{4\cosh^2(a_n/2)\cosh^2(b_{n-1}/2)}\cdot\sinh^2\frac{a_{n-1}}2$$
When $a_n=\displaystyle{\frac12a_{n-1}}$, we have
$$\sinh^2\frac{a_n}2=\frac{1}{4\cosh^2(a_n/2)}\cdot\sinh^2\frac{a_{n-1}}2>\frac{1}{4\cosh^2(a_n/2)\cosh^2(b_{n-1}/2)}\cdot\sinh^2\frac{a_{n-1}}2$$
In all cases, we have
{\allowdisplaybreaks\begin{align*}
\sinh^2\frac{a_n}2>&\frac{1}{4\cosh^2(a_n/2)\cosh^2(b_{n-1}/2)}\cdot\sinh^2\frac{a_{n-1}}2\\
>&\cdots>\frac{1}{4^n}\sinh^2\frac{a_0}2\left/\left(\prod_{k=1}^n\big(1+\sinh^2(a_k/2)\big)\big(1+\sinh^2(b_{k-1}/2)\big)\right)\right.\\
>&\frac{1}{4^n}\sinh^2\frac{a_0}2\left/\exp\left(\sum_{k=1}^\infty\ln\big(1+\sinh^2(a_k/2)\big)+\sum_{k=0}^\infty\ln\big(1+\sinh^2(b_{k-1}/2)\big)\right)\right.\\
>&\frac{1}{4^n}\sinh^2\frac{a_0}2\left/\exp\left(\sum_{k=1}^\infty\ln\big(1+\frac1{2^k}\big)+\sum_{k=0}^\infty\ln\big(1+\frac{1}{2^k}\big)\right)\right.\\
>&\frac{1}{4^n}\sinh^2\frac{a_0}2\left/\exp\left(\sum_{k=1}^\infty\frac1{2^k}+\sum_{k=0}^\infty\frac{1}{2^k}\right)\right.=\frac1{4^ne^3}\sinh^2\frac{a_0}2\\
\end{align*}}
So
$$\sinh\frac{a_n}2>\frac{1}{e\sqrt{e}}\cdot\frac{1}{2^n}\sinh\frac{a_0}2.$$
\end{proof}
\begin{prop}\label{prop:converging_nondegenerate}
For any hyperbolic triangle $\triangle_0$, the sequence $\left\{\triangle_n\right\}_{n=0}^\infty$ converges to a nondegenerate Euclidean triangle.
\end{prop}
\begin{proof} Let $l_{a_n}$ be the length of the lambert qudrilateral on $a_n$, as in figure 2. Then we have
\begin{equation}\label{eqn:lambert_identity}
\sinh\frac{a_n}2=\sinh\frac{a_{n+1}}2\cosh l_{a_n}
\end{equation}
Now we compare $\sin\beta$ with $\sin A_n$. By the law of sines, we have
$$\sin\beta=\frac{\displaystyle{\sinh\frac{a_n}{2}\sin B_n}}{\sinh b_{n+1}},\quad \sin A_n=\frac{\sinh a_n\sin B_n}{\sinh b_n}$$
Hence
$$\frac{\sin\beta}{\sin A_n}=\frac{\displaystyle{\sinh(a_n/2)\sinh b_n}}{\sinh b_{n+1}\sinh a_n}=\frac{\displaystyle{\sinh(a_n/2)\sinh b_n\cosh l_{b_n}}}{\displaystyle{\sinh\frac {b_n}2\sinh a_n}}=\frac{\cosh(l_{b_n})}{\cosh(a_n/2)}\cosh(b_n/2)$$
In the corresponding right-angle triangle, we get $\displaystyle{l_{b_n}<\frac{a_n}2}$, hence we obtain
\begin{equation}
\frac{\cosh(b_n/2)}{\cosh(a_n/2)}<\frac{\sin\beta}{\sin A_n}<\cosh(b_n/2)
\end{equation}
Similarly, we get
\begin{equation}
\frac{\cosh(c_n/2)}{\cosh(a_n/2)}<\frac{\sin\gamma}{\sin A_n}<\cosh(c_n/2)
\end{equation}
Now we compare $\sin\alpha$ and $\sin A_n$. By the Cagnoli formula, we have
\begin{equation}
\sin\frac{S_n}2=\frac{\sinh(b_{n}/2)\sinh(c_n/2)\sin A_n}{\cosh(a_n/2)}
\end{equation}
By the Keogh formula, $S_n$ can also be computed by $b_{n+1},c_{n+1}$ and $\alpha$ as follows
\begin{equation}
\sin\frac{S_n}2=\sinh(b_{n+1})\sinh(c_{n+1})\sin(\alpha)
\end{equation}
So we get
\begin{align*}
\frac{\sin\alpha}{\sin A_n}=& \frac{\sinh(b_{n}/2)\sinh(c_n/2)}{\sinh(b_{n+1})\sinh(c_{n+1})}\cdot\frac{1}{\cosh(a_n/2)}\\
=& \frac{\cosh(l_{b_n})\cosh(l_{c_n})}{\cosh(a_n/2)}\quad (\text{Apply }(\ref{eqn:lambert_identity}))
\end{align*}
Apply $l_{b_n}<\frac{c_n}2$ and $l_{c_n}<\frac{b_n}2$ to get the inequality
$$\frac1{\cosh(a_n/2)}<\frac{\sin \alpha}{\sin A_n}<\cosh(b_n/2)\cosh(c_n/2)$$
So in all cases, we get
\begin{equation}
\frac{1}{\cosh(a_n/2)}<\frac{\sin A_{n+1}}{\sin A_n}<\cosh(b_n/2)\cosh(c_n/2)
\end{equation}
Let $\rho_n=\ln\sin A_n$, then we have
$$-\sum_{i=n}^{n+k-1}\ln\cosh\left(\frac{a_i}2\right)<\rho_{n+k}-\rho_n<\sum_{i=n}^{n+k-1}\left(\ln \cosh\left(\frac{b_i}2\right)+\ln\cosh\left(\frac{c_i}2\right)\right)$$
Hence
\begin{align*}
\abs{\rho_{n+k}-\rho_n}<&\sum_{i=n}^{n+k-1}\left(\ln \cosh\frac{a_i}2+\ln \cosh \frac{b_i}2 +\ln\cosh\frac{c_i}2\right)\\
=&\frac12\sum_{i=n}^{n+k-1}\left(\ln\left(1+\sinh^2\frac{a_i}2\right)+\ln\left(1+\sinh^2\frac{b_i}2\right)+\ln\left(1+\sinh^2\frac{c_i}2\right)\right)\\
<&\frac12\sum_{i=n}^{n+k-1}\left(\sinh^2\frac{a_i}2+\sinh^2\frac{b_i}2+\sinh^2\frac{c_i}2\right)\\
<&\frac12\sum_{i=n}^{\infty}\frac{1}{2^i}\left(\sinh^2\frac{a_0}2+\sinh^2\frac{b_0}2+\sinh^2\frac{c_0}2\right)\\
=&\frac{1}{2^n}\left(\sinh^2\frac{a_0}2+\sinh^2\frac{b_0}2+\sinh^2\frac{c_0}2\right)
\end{align*}
Therefore $\left\{\ln\sin A_n\right\}$ is a Cauchy sequence and converges. Similarly, $\{\ln\sinh B_n\}$ and $\{\ln\sin C_n\}$ converge. The proposition follows.
\end{proof}
For iteration of $f_M$, we compute the areas of the triangles as follows
\begin{prop} Let $\triangle_n=f^{n}_M(\triangle_0)$. Suppose that $\displaystyle{\sinh\frac{a_0}2<1}$, $\displaystyle{\sinh\frac{b_0}2<1}$ and $\displaystyle{\sinh\frac{c_0}2<1}$, then
$$\frac1{\sqrt{e}}\cdot\frac{1}{4^n}\cdot\sinh\frac{S_0}2\leq \sinh\frac{S_n}2\leq \frac{1}{4^n}\sinh\frac{S_0}2,\quad\forall\ n.$$
\end{prop}
\begin{proof} For any $n$, let
$$x=2\cosh a_{n},\quad y=2\cosh b_{n},\quad z=2\cosh c_{n}$$
Then by the trace identity, we have
$$\left(2\cos\frac{S_{n-1}}2\right)^2=x^2+y^2+z^2-xyz$$
and
$$\left(\cos\frac{S_{n}}2\right)^2=\frac{(x+y+z+2)^2}{(x+2)(y+2)(z+2)}$$
Therefore we have
{\allowdisplaybreaks\begin{align*}
\left(\sin\frac{S_{n}}2\right)^2=&1-\left(\cos\frac{S_{n}}2\right)^2=1-\frac{(x+y+z+2)^2}{(x+2)(y+2)(z+2)}\\
=&\frac{(x+2)(y+2)(z+2)-(x+y+z+2)^2}{(x+2)(y+2)(z+2)}\\
=&\frac{4-(x^2+y^2+z^2-xyz)}{(x+2)(y+2)(z+2)}=\frac{4-4\left(\cos\displaystyle{\frac{S_{n-1}}2}\right)^2}{(x+2)(y+2)(z+2)}\\
=&\frac{4}{(2\cosh a_{n}+2)(2\cosh b_{n}+2)(2\cosh c_{n}+2)}\left(\sin\frac{S_{n-1}}2\right)^2\\
=&\frac{4}{(4\sinh^2 \frac{a_{n}}2+4)(4\sinh^2 \frac{b_{n}}2+4)(4\sinh^2 \frac{c_{n}}2+4)}\left(\sin\frac{S_{n-1}}2\right)^2\\
=&\left(\frac14\sin\frac{S_{n-1}}2\right)^2\left/\left[(1+\sinh^2 \frac{a_{n}}2)(1+\sinh^2 \frac{b_{n}}2)(1+\sinh^2 \frac{c_{n}}2)\right]\right.
\end{align*}}
Without loss of generality, we may assume $\sinh\displaystyle{\frac{a_0}2}<1$. By Lemma, we have $\sinh\displaystyle{\frac{a_{n}}2}<\frac12 \sinh\frac{a_{n-1}}2$, then $\displaystyle{\sinh\frac{a_n}2<\frac1{2^n}}$, we get
\begin{align*}
\left(\sin\frac{S_{n}}2\right)^2\geq&\left(\frac14\sin\frac{S_{n-1}}2\right)^2\left/\left(1+\frac1{4^n}\right)^3\right.\geq\cdots\geq\frac{1}{16^n}\left(\sin\frac{S_0}2\right)^2\left/\prod_{i=1}^n\left(1+\frac1{4^i}\right)^3\right.\\
=&\frac{1}{16^n}\left(\sin\frac{S_0}2\right)^2\left/\exp\left\{3\sum_{i=1}^n\ln\left(1+\frac1{4^i}\right)\right\}\right.\\
\geq&\frac{1}{16^n}\left(\sin\frac{S_0}2\right)^2\left/\exp\left\{3\sum_{i=1}^\infty\frac1{4^i}\right\}\right.=\frac1{16^n}\cdot\frac1e\cdot\left(\sin\frac{S_0}2\right)^2
\end{align*}
\end{proof}
\begin{cor} For the sequence $\left\{\triangle_n=f_M^n(\triangle_0)\right\}$, let $S_n$ be the area of $\triangle_n$, then the sequence $\displaystyle{\left\{\frac{\sin(S_n/2)}{\sin(S_0/2)}\cdot4^n\right\}}$ converges and its limit belongs to $\displaystyle{\left(\frac{1}{\sqrt{e}},\sqrt{e}\right)}$.
\end{cor}
So the sine of the area of the middle triangle is shrinking to approximately $\frac14$ the sine of area of the previous triangle, with accumulated error of ratio less than $\sqrt{e}$.
\begin{comment}
\begin{prop} For any hyperbolic triangle $\triangle_0$, the sequence $\left\{\triangle_n\right\}_{n=0}^\infty$ converges to a nondegenerate Euclidean triangle.
\end{prop}
\begin{proof} Without loss of generality, we may assume $A\leq B\leq C$. Let $c_n$ be the corresponding edge of $\triangle(A_n, B_n, C_n)$ and $h_n$ be the height on $c_n$. Then we have
$$\sin\frac{S_{n-1}}2=\sinh(c_n)\sinh(h_n)$$
We get
\begin{align*}
\frac1{\sqrt{e}}\cdot\frac1{4^{n-1}}\cdot\sin\frac{S_0}2\leq&\sin\frac{S_{n-1}}2=\sinh(c_n)\sinh(h_n)=\frac{\sinh(h_n)}{\sinh(c_n)}\Big(\sinh(c_n)\Big)^2\\
<&\frac{\sinh(h_n)}{\sinh(c_n)}\cdot\left(\frac{\sinh(c_0)}{2^n}\right)^2=\frac{\sinh(h_n)}{\sinh(c_n)}\cdot\frac{\sinh^2(c_0)}{4^n}
\end{align*}
Therefore
$$\frac{\sinh(h_n)}{\sinh(c_n)}>\frac{4}{\sqrt{e}}\cdot\frac{\sin({S_0}/2)}{\sinh^2(c_0)}$$
\end{proof}
\end{comment}
\section{Continuity of the limit functions}
Let $\mathcal{T}_E$ be the Techm\"uller space of nontrivial Euclidean triangles modulo similarities and $\mathcal{T}_H$ be the Techm\"uller space of nontrivial hyperbolic triangles. $\mathcal{T}_E$ is naturally identified with the interior of the triangle in $\mathbb{R}^3$ with vertices $(1,0,0)$, $(0,1,0)$ and $(0,0,1)$. $\mathcal{T}_H$ is naturally identified with the interior of the tetrahedron in $\mathbb{R}^3$ with vertices $(0,0,0)$, $(1,0,0)$, $(0,1,0)$ and $(0,0,1)$. $\mathcal{T}_E$ and $\mathcal{T}_H$ inherit metrics as subspaces of $\mathbb{R}^3$. We have $\mathcal{T}=\mathcal{T}_E\cup\mathcal{T}_H$.
Let $\mathcal{S}$ be the set of infinite sequences in four letters $A$, $B$, $C$ and $M$. A sequence $\mathfrak{s}$ is called {\it rational} if exactly one of the three letters $A$, $B$ and $C$ appears infinite times, and {\it irrational} otherwise. Fix a Eulidean or hyperbolic triangle $\triangle=\triangle_{ABC}$. A sequence in $s\in\mathcal{S}$ defines nested triangles $\{\triangle_n\}$ in $\triangle$ via $f_A$, $f_B$, $f_C$ and $f_M$. The nested triangles $\{\triangle_n\}$ have a unique intersection point, denoted by $\phi(s)$. It is not hard to see that $\phi$ defines a surjective map from $\mathcal{S}$ to $\triangle$. Endow $\mathcal{S}$ with the smallest topology $\mathcal{T}$ such that $\phi$ is continuous. The topology does not depend on the choice of the Euclidean triangle $\triangle$.
An easy investigation of $\phi$ gives the following
\begin{prop}\label{prop:sequence_map}
For any two distinct sequences $\mathfrak{s}$ and ${\mathfrak{t}}$, we have $\phi(\mathfrak{s})=\phi({\mathfrak{t}})$ if and only if there exists
\begin{itemize}
\item a finite sequence $\tau_1,\cdots,\tau_{n}$ for some $n\geqslant0$,
\item a permutation $\sigma:\{A,B,C\}\rightarrow\{A,B,C\}$, and
\item a finite sequence $\zeta:\{1,\cdots,m\}\rightarrow\{x,y\}$ in two indeterminants $x$ and $y$ with $m\geqslant0$,
\end{itemize}
such that $\mathfrak{s}$ and ${\mathfrak{t}}$ are of the following six forms
\begin{tabular}{c c c c c c c c c c c c c c}
$(1)$&$\tau_1$&$\tau_2$&$\cdots$&$\tau_{n}$&$\sigma(A)$&$\alpha_1$&$\alpha_2$&$\cdots$&$\alpha_m$&$M$&$\sigma(A)$&$\sigma(A)$&$\cdots$\\
$(2)$&$\tau_1$&$\tau_2$&$\cdots$&$\tau_{n}$&$\sigma(A)$&$\alpha_1$&$\alpha_2$&$\cdots$&$\alpha_m$&$\sigma(B)$&$\sigma(C)$&$\sigma(C)$&$\cdots$\\
$(3)$&$\tau_1$&$\tau_2$&$\cdots$&$\tau_{n}$&$\sigma(A)$&$\alpha_1$&$\alpha_2$&$\cdots$&$\alpha_m$&$\sigma(C)$&$\sigma(B)$&$\sigma(B)$&$\cdots$\\
$(4)$&$\tau_1$&$\tau_2$&$\cdots$&$\tau_{n}$&$M$&$\beta_1$&$\beta_2$&$\cdots$&$\beta_m$&$M$&$\sigma(A)$&$\sigma(A)$&$\cdots$\\
$(5)$&$\tau_1$&$\tau_2$&$\cdots$&$\tau_{n}$&$M$&$\beta_1$&$\beta_2$&$\cdots$&$\beta_m$&$\sigma(B)$&$\sigma(C)$&$\sigma(C)$&$\cdots$\\
$(6)$&$\tau_1$&$\tau_2$&$\cdots$&$\tau_{n}$&$M$&$\beta_1$&$\beta_2$&$\cdots$&$\beta_m$&$\sigma(C)$&$\sigma(B)$&$\sigma(B)$&$\cdots$
\end{tabular}
\noindent where $\alpha_i=\zeta_i(\sigma(B),\sigma(C))$ and $\beta_i=\zeta_i(\sigma(C),\sigma(B))$.
\end{prop}
Two distinct sequences with the same image under $\phi$ are rational. An irrational sequence does not have the same image under $\phi$ with another sequence. Proposition \ref{prop:sequence_map} implies that the space $(\mathcal{S}, \mathcal{T})$ is not Hausdorff.
Given a sequence $\mathfrak{s}\in\mathcal{S}$ and a hyperbolic triangle $\triangle_0=\triangle_{A_0B_0C_0}\in\mathcal{T}_H$. Inductively define $\triangle_n=f_{s_n}(\triangle_{n-1})$. By Proposition \ref{prop:converging_nondegenerate}, the sequence $\{\triangle_n\}$ converge to a nondegenerate Euclidean triangle. Hence we have a well-defined map $\Phi:\mathcal{S}\times \mathcal{T}_H\rightarrow \mathcal{T}_E$.
\begin{prop}\label{prop:continuity}
The map $\Phi:\mathcal{S}\times \mathcal{T}_H\rightarrow \mathcal{T}_E$ satisfies the following continuity properties
\begin{enumerate}
\item For any $\mathfrak{s}\in\mathcal{S}$, $\Phi_\mathfrak{s}=\Phi(\mathfrak{s},\cdot):\mathcal{T}_H\rightarrow \mathcal{T}_E$ is continous and surjective.
\item For any $\triangle\in \mathcal{T}_H$, $\Phi_\triangle=\Phi(\cdot,\triangle):\mathcal{S}\rightarrow \mathcal{T}_E$ is continous at irrational points.
\end{enumerate}
\end{prop}
\begin{proof}
The functions $f_A$, $f_B$, $f_C$ and $f_M$ are smooth. For the continuity, it suffices to show that the series $\displaystyle{\sum_{n=1}^\infty(\ln\sin A_n-\ln\sin A_{n-1})}$ is uniformly convergent, which directly follows from the last inequality in the proof of Proposition \ref{prop:converging_nondegenerate}.
Let $p:\mathcal{T}_H\rightarrow\mathcal{T}_E$ be the projection
$$(x,y,z)\longmapsto \left(\frac{x}{x+y+z}\cdot\pi,\frac{z}{x+y+z}\cdot\pi,\frac{z}{x+y+z}\cdot\pi\right).$$
If $\Phi_\mathfrak{s}$ is not surjective, take $v\in\mathcal{T}_E\setminus\phi_\mathfrak{s}(\mathcal{T}_H)$. There exists some $\varepsilon>0$ such that $\mathcal{T}_E$ contains the circle $C_1$ with center $v$ and radius $\varepsilon$.
By the proof of Proposition \ref{prop:converging_nondegenerate}, there exists some sufficiently small $s>0$ such that
$$\abs{\Phi_\mathfrak{s}(x)-p(x)}< \frac{\varepsilon}2$$
for any $x$ in the circle $C_2$, which is the intersection of $p^{-1}(C_1)$ and $x+y+z=\pi-s$. Then the singular loop $\Phi_\mathfrak{s}(C_2)$ is free homotopic to $C_1$ in $\mathcal{T}_E\setminus\{v\}$ by straight-line homotopy, and hence not null homotopic. However, $C_2$ bounds a disk in $x+y+z=\pi-s$ whose image gives a null homotopy of $\Phi_\mathfrak{s}(C_2)$. The contradiction implies the surjectivity of $\Phi_s$.
Given a hyperbolic triangle $\triangle=\triangle_{ABC}$. Use $\triangle$ to define the topology on $\mathcal{S}$. Let $\mathfrak{s}=(s_1,s_2,\cdots)$ be an irrational sequence. Let $\{\triangle_n\}$ be the sequence to define $\Phi(\mathfrak{s},\triangle)$. For any $n$, the set $U_n$ of sequences starting with $(s_1,s_2,\cdots,s_n)$ is a neighborhood of $\mathfrak{s}$. Given any $\varepsilon\in(0,1)$, let
$$\mu=\min\left\{\ln\left(\frac\varepsilon{14\pi}+1\right), \ln\left(\frac{1}{1-\varepsilon/(14\pi)}\right)\right\}$$
By Proposition \ref{prop:converging_nondegenerate}, there exist $N$ such that for any ${\mathfrak{t}}\in U_N$, the ratio of the angles of $\Phi({\mathfrak{t}},\triangle)$ and the angles $(A_N,B_N,C_N)$ of $\triangle_N$ is within $(e^{-\mu}, e^\mu)$. For any ${\mathfrak{t}}\in U_N$, suppose $\Phi_\triangle({\mathfrak{t}})=(\alpha,\beta,\gamma)$ and $\Phi_\triangle(\mathfrak{s})=(\alpha_0,\beta_0,\gamma_0)$, then
\begin{align*}
d(\Phi_\triangle({\mathfrak{t}}),\Phi_\triangle(\mathfrak{s}))\leq&\abs{\alpha-\alpha_0}+\abs{\beta-\beta_0}+\abs{\gamma-\gamma_0}\\
\leq& \abs{\alpha-A_N} +\abs{A_N-\alpha_0}+\abs{\beta-B_N}+\abs{B_N-\beta_0}+\abs{\gamma-C_N}+\abs{C_N-\gamma_0}\\
\leq& 2(A_N+B_N+C_N)\cdot \big((1-e^{-\mu})+(e^\mu-1)\big)\\
\leq& 6\pi\cdot \left(\frac\varepsilon{14\pi}+\frac\varepsilon{14\pi}\right)=\frac67\cdot\varepsilon<\varepsilon.
\end{align*}
Hence $\Phi_\triangle$ is continuous at $\triangle$.
\end{proof}
We remark that in generral $\Phi_\triangle$ can be discontinuous at an rational sequence.
\begin{proof}[Proof of Theorem \ref{thm:main}]
The theorem evidently follows from Proposition \ref{prop:converging_nondegenerate} and Proposition \ref{prop:continuity}.
\end{proof}
\subsubsection*{Acknowledgement} The first named author is partially supported by NSFC 11425102.
|
\section{Introduction}
Time series is a ubiquitous datatype in our life, ranging from finance, medicine, geology, etc.
It is clear that different problems depend on different interpretation and processing of the observed time series. In some situations, the information can be easily read from the signal, for example, the cardiac arrest could be easily read from the electrocardiogram (ECG) signal; in others, it is less accessible, for example, the heart rate variability (HRV) hidden inside the ECG signal; in yet others, the information might be masked and cannot be read directly from the observed time series. %
This comes from the fact that while the time series encodes the temporal dynamics of the system under observation, most of {the} time the dynamical information we could perceive is masked or deformed due to the observation process and the nature of the physiology.
When the information is masked or deformed but exists in the observed time series, we might need more sophisticated approaches to extract the information relevant to the situation we have interest {in}. In general, inferring the dynamical information from the time series is challenging.
We could view the challenge in two parts. First, we need to choose a model to quantify the recorded signal, which captures the features or information about the underlying dynamical system we have interest {in}. This model could come from the field background knowledge, or in some cases it could be relatively blind. Second, we need to design an associated algorithm to extract the desired features from the recorded signal. With the acquired features, we could proceed to study the dynamical problem we have interest {in}.
We take physiological signals to illustrate the challenge of the modeling issue. Note that the procedure could be applied to other suitable fields. It is well known that how the signal oscillates contains plenty of information about a person's health condition.
Based on the oscillatory behavior and the widely studied Fourier analysis, common features we could discuss are the frequency, which represents how fast the signal oscillates, and the amplitude, which represents how strongly the signal oscillates at that frequency. However, these features have been found limited when the signal is not stationary, which is a property shared by most physiological signals. Indeed, { these signals mostly oscillate} with time-varying frequency and amplitude. To capture this property, we could consider the {\em adaptive harmonic model} encoding the features {\em instantaneous frequency (IF)} and {\em amplitude modulation (AM)} \cite{Daubechies_Lu_Wu:2011,Chen_Cheng_Wu:2014}; that is, the signal is modeled as
\begin{equation}\label{Introduction:AHM}
f_0(t)=A(t)\cos(2\pi\phi(t)),
\end{equation}
where $A$ is a smooth positive function and $\phi$ is a smooth monotonically increasing function. In other words, at time $t$, the signal $f$ repeats itself as a sinusoidal function within about $1/\phi'(t)$ seconds, and the oscillation is modulated by the AM function $A(t)$. These features have been proved useful and could well represent the physiological dynamics and health status, and have been applied to different problems \cite{Lin_Hseu_Yien_Tsao:2011,Lin_Wu_Tsao_Yien_Hseu:2014,Wu_Talmon_Lo:2015}.
There are actually more detailed features embedded in the oscillatory signals that cannot be captured by (\ref{Introduction:AHM}). One particular feature is the non-sinusoidal oscillatory pattern.
For example, respiratory flow signals usually do not oscillate like the sinusoidal function, since the inspiration is normally shorter than the expiration, and this difference is intrinsic to the respiratory system \cite{Benchetrit:2000}.
These observations lead us to consider the following model \cite{Wu:2013,Yang:2014,Hou_Shi:2016},
\begin{equation}\label{Introduction:ANHM0}
f_1(t)=A(t)s(\phi(t)),
\end{equation}
where $A(t)$ and $\phi(t)$ are the same as those of (\ref{Introduction:AHM}), and $s$ is a real 1-periodic function {with the unitary $L^2$ norm}, that is $s(t+1)=s(t)$ for all $t$, so that the first Fourier coefficient $\hat{s}(1)\neq 0$, which could be different from the cosine function.
We call the periodic signal $s(t)$ the {\em wave-shape function}, $\phi(t)$ the phase function, the derivative $\phi'(t)$ the IF, and $A(t)$ the AM of $f_1(t)$. Note that when $s$ is smooth enough, (\ref{Introduction:ANHM0}) could be expanded pointwisely by the Fourier series as
\begin{equation}\label{Introduction:ANHM1}
f_1(t)=\sum_{k=0}^\infty A(t)a_k\cos(2\pi k\phi(t)+\alpha_k),
\end{equation}
where $a_k\geq0$, $k\in\mathbb{N}\cup\{0\}$ are associated with the Fourier coefficients of $s$, $\alpha_0=0$ and $\alpha_k\in[0,2\pi)$, $k\in\mathbb{N}${, and $a_0^2+2\sum_{k=1}^\infty a_k^2=1$}. Note that we could have two different aspects of the same signal $f_1$. First, we could view it as an oscillatory signal with one oscillatory component with non-sinusoidal oscillation (\ref{Introduction:ANHM0}). Second, we also could view it as an oscillatory signal with multiple oscillatory components with the cosine oscillatory pattern (\ref{Introduction:ANHM1}); in this case, we call the first oscillatory component $A(t)a_1\cos(2\pi \phi(t)+\alpha_1)$ the {\em fundamental component} and $A(t)a_k\cos(2\pi k\phi(t)+\alpha_k)$, $k\geq 2$, the {\em $k$-th multiple} of the fundamental component. Clearly, the IF of the $k$-th multiple is $k$-times that of the fundamental component. Note that $A(t)a_0$ could be viewed as the {trend coming from the DC (direct current) or zero-frequency} term of the wave-shape function. While the second viewpoint (\ref{Introduction:ANHM1}) is better for the theoretical analysis, the first viewpoint (\ref{Introduction:ANHM0}) is more physical in several applications.
{ Let us take} the ECG signal as an example, where the IF, AM and the wave-shape function have their own physiological meanings.
The oscillatory morphology of the ECG signal, the wave-shape function, reflects not only the electrical pathway inside the heart and how the sensor detects the electrophysiological dynamics, but also the respiration as well as the heart anatomy. Several clinical diseases are diagnosed by reading the oscillatory morphology. With these physiological understanding, it is better to consider model (\ref{Introduction:ANHM0}) to study the ECG signal and view IF, AM and wave-shape function as separate features.
As for IF and AM, it is well known that while the rate of the pacemaker is constant, the heart rate generally is not constant. The discrepancy comes from neural and neuro-chemical influences on the pathway from the pacemaker to the ventricle. This non-constant heart beat rate could be modeled as the IF of the ECG signal.
The AM of the ECG signal is directly related to the respiration via the variation of thoracic impedance. Indeed, when the lung is full of air, the thoracic impedance increases and hence and ECG amplitude decreases, and vice versa. Note that IF and AM could be captured by both (\ref{Introduction:ANHM0}) and (\ref{Introduction:ANHM1}).
Several algorithms were proposed to extract IF and AM from a given oscillatory signal in the past decade, like empirical mode decomposition \cite{Huang_Shen_Long_Wu_Shih_Zheng_Yen_Tung_Liu:1998}, reassignment method (RM) \cite{Auger_Flandrin:1995}, synchrosqueezing transform (SST) \cite{Daubechies_Lu_Wu:2011}, concentration of frequency and time \cite{Daubechies_Wang_Wu:2016}, Blaschke decomposition \cite{Coifman_Steinerberger:2015}, iterative filtering \cite{Cicone_Liu_Zhou:2014}, sparsification approach \cite{Hou_Shi:2013a}, approximation approach \cite{Chui_Mhaskar:2016}, convex optimization \cite{Kowalski_Meynard_Wu:2015}, Gabor transform based on different selection criteria \cite{Balazs_Dorfler_Jaillet_Holighaus_Velasco:2011,Ricaud_Stempfel_Torresani:2014}, etc. In general, we could view these methods as a nonlinear time-frequency (TF) analysis.
However, to capture the wave-shape function, an extra step is needed -- we could fit a non-sinusoidal periodic function to the signal after/while extracting the IF by, for example, applying the functional regression \cite[Section 4.7]{Chui_Lin_Wu:2015}, designing a dictionary \cite{Hou_Shi:2016} or unwrapping the phase \cite{Yang:2014}. The obtained features have been used to study field problems, such as the sleep stage prediction \cite{Wu_Talmon_Lo:2015}, the blood pressure analysis \cite{Wu_Chang_Wu_Wang_Yang_Wu:2015}. See \cite{Daubechies_Wang_Wu:2016} for a review of the applications.
As useful as the above-mentioned model and algorithms { are} to extract dynamical features from time series, there are{, however,} several unsolved limitations.
First, for most physiological signals, the wave-shape function {varies from time to} time. The time-varying wave-shape function might { prevent} the current available methods from extracting the wave-shape function. We will provide physiological details in Section \ref{Section:Limitation}.
Second, there might be more than one oscillatory component in a signal, and each oscillatory component has its own wave-shape function. See Figure \ref{fig:Introduction:Example0sig} for an photoplethysmogram signal (PPG) as an example. In this PPG signal, there are two oscillatory components, hemodynamic rhythm and respiratory rhythm.
Third, although we could obtain reasonable information about IF and AM from the above-mentioned approaches, when the signal has multiple oscillatory components with non-sinusoidal waves, these methods are limited. In particular, the multiples of different fundamental components will interfere with each other.
Furthermore, an automatic determination of the number of oscillatory components becomes more difficult when each component oscillates with a non-sinusoidal wave.
Hence, modifications are needed.
\begin{figure}[h!]
\begin{centering}
\includegraphics[width= \textwidth]{Intro1sig.png}
\end{centering}
\caption{\label{fig:Introduction:Example0sig} A photoplethysmogram signal. It is visually clear that there are two rhythms inside the signal. The faster one is associated with the heart rate, which beats about 100 times per minute; the slower one is associated with the respiration, which is about 18 times per minute.}
\end{figure}
In this paper, we resolve these limitations. We introduce the adaptive non-harmonic model to model oscillatory signals with multiple components and time-varying wave-shape functions. Motivated by cepstrum, we introduce an algorithm called {\em de-shape SST} to alleviate the influence caused by non-sinusoidal wave-shape functions in the TF analysis. Hence, we provide an enhanced TF representation in the following sense -- the de-shape SST would provide a TF representation with only IF and AM information without the influence of non-sinusoidal wave-shape functions.
We illustrate the effectiveness of de-shape SST by showing results on a simulated signal. In this example, the clean signal $f(t)$ is composed of two oscillatory components $f_{1}$ and $f_{2}$, where $f_{1}(t)=A_1(t)(\sum_{k\in\mathbb{Z}}\delta_k\star h)(t)\chi_{[0,60]}(t)$, $A_{1}(t)=1.5e^{-(\frac{t-20}{100})^2}$, $h(t)=e^{-18t^2}$, $\delta_k$ is the Dirac delta measure supported at $k$, $\chi_{I}$ is the indicator function supported on $I\subset \mathbb{R}$ and $f_{2}(t)=A_{2}(t)\text{mod}(\phi_{2}(t),1)$, where $A_{2}(t)>0$ and $\phi_{2}'(t)>0$ are two non-constant smooth function and $\text{mod}(x,1):=x-\lfloor x\rfloor$ for $x\in\mathbb{R}$ and $\lfloor x\rfloor$ means the largest integer less than or equal to $x$. Clearly, $f_{1}$ oscillates at the fixed frequency $\phi'_1(t)=1$ with a non-sinusoidal wave-shape function -- the wave-shape function of $f_1$ looks like a Gaussian function; $f_{2}$ oscillates with a time-varying frequency with the non-sinusoidal wave, which behaves like a sawtooth wave. This signal is sampled at rate $100$Hz, from $t=0$ to $t=100$ seconds.
Figure \ref{fig:Introduction:Example1sig} shows the two constituents of the total signal $f(t)=f_{1}(t)+f_{2}(t)$, as well as $A_{2}(t)$ and $\phi_{2}'(t)$. Note that $f_1$ ``lives'' during only part of the full time observation time interval.
The panels in Figure \ref{fig:Introduction:Example1} show the results of short-time Fourier transform of $f(t)$, the SST of $f(t)$ and the de-shape SST of $f(t)$. It is clear that compared with the TF representation provided by STFT or SST, the TF representation provided by the de-shape SST contains only the fundamental frequency information of the two oscillatory components, even when the wave-shape function is far from the sinusoidal wave. More discussions will be provided in Section \ref{Section:Numerics}, including how $A_{2}(t)$ and $\phi_{2}(t)$ are generated.
\begin{figure}[h!]
\begin{centering}
\includegraphics[width= \textwidth]{Main0sig.png}
\end{centering}
\caption{\label{fig:Introduction:Example1sig}Top panel: $f_1(t)$; second panel: $f_2(t)$; third panel: the $A_2(t)$ (dashed curve) and $\phi_2'(t)$ (solid curve) of $f_2(t)$; {bottom panel: $f(t)$.} To enhance the visibility, we only show the signal from the 25-th second to the 65-th second.}
\end{figure}
\begin{figure}[h!]
\begin{centering}
\includegraphics[width= \textwidth]{Main0.png}
\end{centering}
\caption{\label{fig:Introduction:Example1}Upper left: the short time Fourier transform (STFT) of $f(t)$; upper right: the synchrosqueezed STFT of $f(t)$; lower left: the de-shape SST of $f(t)$. It is clear that the de-shape SST provides only the fundamental frequency information of the two oscillatory components, even when the wave-shape function is far from the sinusoidal wave; lower right: the de-shape SST of $f(t)$ superimposed with the ground truth IF's of both components in red. To enhance the visibility, we show the de-shape SST only up to 6 Hz in the frequency axis.}
\end{figure}
The paper is organized in the following way. In Section \ref{Section:Model}, we discuss the limitation of model (\ref{Introduction:ANHM0}), and provide a modified model, the adaptive non-harmonic model. In Section \ref{Section:Cepstrum}, the existing cepstrum algorithms in the engineering field are reviewed, and the new algorithm de-shape SST is introduced. The theoretical justification of the de-shape SST is postponed to Appendix \ref{Appendix:Proof}. Section \ref{Section:Numerics} shows the numerical results of de-shape SST on several different simulated, medical, musical, and biological signals. Section \ref{Section:NumericalIssues} discusses numerical issues of the de-shape SST algorithm. Section \ref{Section:Conclusions} summarizes the paper.
\section{Adaptive Non-harmonic Model}\label{Section:Model}
In this section we first review the phenomenological model based on the wave-shape function (\ref{Introduction:ANHM0}) fixed over time. Then, we discuss the relationship between the wave-shape function and several commonly encountered physiological signals, and discuss limitations. This discussion leads us to introduce the adaptive non-harmonic model.
We start from introducing some notations. The Schwartz space is denoted as $\mathcal{S}$; the tempered distribution space, which is the dual space of the Schwartz space, is denoted as $\mathcal{S}'$; $\ell^p$, where $p>0${,} indicates the sequence space including all sequences $x:\mathbb{N}\to \mathbb{R}$ so that $\sum_{n\in\mathbb{N}} |x(n)|^p<\infty$, where $x(n)$ is the $n$-th element of the sequence $x$. For each $k\in\mathbb{N}\cup\{\infty\}$, $C^k$ indicates the space of continuous functions with all the derivatives continuous, up to the $k$-th derivates, and $C_c^k$ indicates the space of compactly supported continuous functions with all the derivatives continuous, up to the $k$-th derivates. For each $p\in\mathbb{N}$, $L^p$ includes all measurable functions $f$ so that $\int_{-\infty}^\infty|f(x)|^pdx<\infty$; $L^\infty$ includes all measurable functions which are bounded almost surely. For $f\in\mathcal{S}'$ and $g\in \mathcal{E}'$, where $\mathcal{E}'$ is the set of compactly supported distribution{s}, denote $f\star g$ to be the convolution. We will interchangeably use $\mathcal{F}f$ or $\hat{f}$ to denote the Fourier transform of the function $f\in \mathcal{S}'$. When $f\in L^1(\mathbb{R})$, the Fourier transform is equally defined as $f(\xi)=\int_{-\infty}^\infty f(t)e^{-i2\pi \xi t}dt$; when $f\in \mathcal{E}'$, we know that $\hat{f}(\xi)=\langle f,e^{-i2\pi \xi\cdot}\rangle$, where $\langle\cdot,\cdot\rangle$ indicates the evaluation of the {distributions} $f$ at the $C^\infty$ function $e^{-i2\pi \xi t}$.
For a periodic function $s$, denote $\hat{s}(k)$, $k\in\mathbb{Z}$, to be its Fourier series coefficients. For each $N\in\mathbb{N}$, denote the Dirichlet kernel $D_N(x):=\sum_{\ell=-N}^N e^{i2\pi \ell x}$.
\subsection{Review of the wave-shape function}
We continue the discussion of the model (\ref{Introduction:ANHM0})
\begin{equation}\label{model:linearRelationship}
f(t)=A(t)s(\phi(t)),
\end{equation}
where $A\in C^1(\mathbb{R})$ is strictly positive, $\phi\in C^2(\mathbb{R})$ is strictly monotonically increasing, and $s\in C^{1,\alpha}$, $\alpha>1/2$, is a 1-periodic function {with the unitary $L^2$ norm} so that its Fourier series coefficients satisfy $|\hat{s}(1)|>0$, $|\hat{s}(k)|\leq \delta |\hat{s}(1)|$ for some $\delta\geq0$ and $\sum_{k=N+1}^\infty |k\hat{s}(k)|\leq \theta$ for some $\theta\geq 0$ and $N\in\mathbb{N}$.
We need more conditions for {the} analysis. Take $0\leq \epsilon\ll1$, we require
$|\phi''(t)|\leq\epsilon \phi'(t)$ and $|A'(t)|\leq \epsilon \phi'(t)$ for all $t$. This means that we allow the IF and AM to vary in time, as long as the variations are slight from one period to the next.
\subsection{Limitations in modeling physiological signals}\label{Section:Limitation}
While many physiological signals are oscillatory and have ``similar'' patterns, at first glance they could be well modeled by (\ref{model:linearRelationship}) and the analysis could proceed. However, it is not always possible to do so. In this section we provide examples to discuss limitations.
\subsubsection{Electrocardiographic signal} \label{Section:ECGdiscussion}
{The ECG signal, which provides information of the electrical activity of the heart, is ubiquitous in healthcare setting now.}
It not only contains a wealth of information {regarding} the cardiac/cardiovascular health but also provides a unique non-invasive portal to physiological dynamical states of the human body, via for example the HRV assessment. While the HRV{, the non-constant heart rate,} could be studied by evaluating the IF of the ECG signal and well estimated by the ``R peak detection algorithm'', in several cases a modern TF analysis could help improve the estimation accuracy \cite{Herry_Frasch_Wu:2015}.
We now discuss the limitation of modeling the ECG signal by (\ref{model:linearRelationship}).
Take the relationship between the RR and QT intervals of the lead II ECG signal $E_{\texttt{II}}(t)$ as an example\footnote{{The P, Q, R, S, and T are significant landmarks of the ECG signal. The P wave represents atrial depolarization. The Q wave is any downward deflection after the P wave. The R wave follows as an upward deflection, which is spiky, and the S wave is any downward deflection after the R wave. The Q wave, R wave, and S wave form the QRS complex, which corresponds to the ventricular depolarization. The T wave follows the S wave, which represents the ventricular repolarization.} The QT interval (respectively RR interval) is the length of the time interval between the start of the Q wave and the end of the T wave of one heart beat (respectively two R landmarks of two consecutive heart beats). We could view the R peak as a surrogate of the cardiac cycle, and hence the RR interval could be viewed as a surrogate of the inverse of the heart rate. {See Figure \ref{fig:ECG} for an example of the P, Q, R, S, and T landmarks and the RR and QT intervals. For more information about ECG signal, we refer the readers to \cite{ecg}.}}.
The nonlinearity relationship between the QT interval and the RR interval has been well accepted -- for example, the Fridericia's formula (QT interval is proportional to the cubic root of RR interval) \cite{Fridericia:1920} or a fully nonlinear depiction {\cite[Figure 3]{Fossa_Zhou:2010}}.
If we model $E_{\texttt{II}}(t)$ by the model (\ref{model:linearRelationship}), and have
\begin{align}
E_{\texttt{II}}(t)&=A_{\texttt{II}}(t)s_{\texttt{II}}(\phi_{\text{II}}(t))=\sum_{\ell=1}^\infty A_{\texttt{II}}(t)c_{\texttt{II}}(\ell)\cos(2\pi \ell\phi_{\text{II}}(t)+\alpha_{\texttt{II},\ell}),\label{model:ECG0}
\end{align}
where $\alpha_{\texttt{II},0}=0$, $\alpha_{\texttt{II},\ell}\in [0,2\pi)$ when $\ell\in\mathbb{N}$, $c_{\texttt{II}}(1)>0$ and $c_{\texttt{II}}(\ell)\geq 0$ for $\ell\neq 1$ are related to Fourier series coefficients of the wave-shape function $s_{\texttt{II}}$. Here, $s_{\texttt{II}}$ models the oscillation in the lead II ECG signal, which is non-sinusoidal.
Note that under this model, the QT interval has to be ``almost'' linearly related to the RR interval. To see this, suppose there is a 1-periodic function $s_{\text{II}}$ for the lead II ECG signal, where the R peak happens at time $0$, and a monotonically increasing function $\phi_{\text{II}}(t)$ so that the ECG signal could be modeled as $s_{\text{II}}(\phi_{\text{II}}(t))$; that is, the wave-shape function is fixed all the time. Suppose that the $k$-th R peak happens at time $t_k$, where $k\in\mathbb{Z}$; that is, $t_k=\phi_{\text{II}}^{-1}(k)$. By the mean value theorem, in this model we have the following relationship for the ECG signal at time $t\in[t_k,t_{k+1}]$ :
\begin{align}
&s_{\text{II}}(\phi_{\text{II}}(t)){=\sum_{k\in\mathbb{Z}}s_{\text{II}}(\phi_{\text{II}}(t))\chi_{[t_k,t_{k+1})}(t)}\nonumber\\
=&\,\sum_{k\in\mathbb{Z}}s_{\text{II}}(\phi_{\text{II}}(t_k)+(t-t_k)\phi'_{\text{II}}(\tilde{t}_k))\chi_{[t_k,t_{k+1})}(t)=\sum_{k\in\mathbb{Z}}s_{\text{II}}\left(\frac{t-t_k}{1/\phi'_{\text{II}}(\tilde{t}_k)}\right)\chi_{[t_k,t_{k+1})}(t),\label{Model:LinearRelationship}
\end{align}
where $\tilde{t}_k\in[t_k,t_{k+1}]$, $\chi_{[t_k,t_{k+1})}$ is a indicator function defined on $[t_k,t_{k+1})$, and the second equality holds since $\phi_{\text{II}}(t_k)=k$ and $s$ is $1$-periodic. While the RR interval between the $k$-th and the $(k+1)$-th R peaks is proportional to $1/\phi'_{\text{II}}(\tilde{t}_k)$ up to order $O(\epsilon)$ by the slowly varying IF assumption of $\phi_{\text{II}}$, we know that the wave-shape function is approximately linearly dilated according to $1/\phi'_{\text{II}}(\tilde{t}_k)$. If the the wave-shape function is linearly dilated according to the RR interval, then the QT interval should be linearly related with the RR interval and hence the claim. Clearly, this model contradicts the physiological finding that the QT interval should be nonlinearly related to the RR interval, so we need a modified model to better quantify the ECG signal.
Furthermore, note that since the cardiac axis varies from time to time due to respiration, physical activity and so on, even if the RR interval is fixed all the time and we focus on the lead II ECG signal, we cannot find a fixed wave-shape function to exactly model the ECG signal. Note that the wave-shape function variation caused by respiration could be applied to extract the respiratory information from the ECG signal \cite{Chui_Lin_Wu:2015}.
\subsubsection{Respiratory signal}
Oscillation is a typical pattern in breathing in normal subjects. It is well known that there is a rhythmic controller in the Pre-B\"otzinger complex in the brain stem which regularly oscillates. In a normal subject the respiratory period is about 5 seconds per cycle. Note that when we are awake, we could also control our respiration by our will, but to simplify the discussion, we do not take this into account.
The existence of breathing pattern variability has been well known \cite{Benchetrit:2000}. For example, the period of each respiratory cycle for a normal subject under normal status varies according to time. The ratio between the length of inspiration period and the length of expiration period is not linearly related to the instantaneous respiratory rate, and its variability also contains plenty of physiological information \cite{Benchetrit:2000}. In other words, the wave-shape function associated with the respiration is not fixed all the time.
By the same argument as that for the ECG signal, this nonlinear relationship between the instantaneous respiratory rate and the wave-shape function could not be fully captured by (\ref{model:linearRelationship}).
The same argument holds for the other physiological signals, like the photoplethysmography signal that reflects the hemodynamics information, the capnogram signal that monitors inhaled and exhaled concentration or partial pressure of carbon dioxide and is a surrogate of the oscillatory dynamics of the respiratory system, and so on.
\subsubsection{Natural vibration of stiff strings}\label{Section:StiffStringModel}
In this section we discuss the signal commonly encountered in music, in particular the sound generated by the string musical instrument. The acoustic signal generated by the string musical instrument could be well modeled by the transversal vibration behavior of an ideal string. For an ideal string of length $L>0$ placed on $[0,L]$ with both ends fixed ideally, when the string has {\em stiffness}; that is, there is a restoring force proportional to the displacement (or more generally the bending angle), we could consider the following differential equation for $y\in \mathbb{R}^+\times [0,L]$ satisfying \cite{fletcher1964normal,fletcher2010physics}
\begin{equation}\label{Equation:Music:Stiffness}
\mu \frac{\partial^2 y}{\partial t^2} = T \frac{\partial^2 y}{\partial x^2} - ESK^2 \frac{\partial^4 y}{\partial x^4}\,,
\end{equation}
where $\mu>0$ is mass per unit length, $T\geq0$ is tension, $E\geq0$ is Young's modulus of the string, $S\geq0$ is the cross-sectional areas of the string, and $K\geq0$ is the radius of gyration, with the initial condition $y(0,x)=0$ for all $x\in[0,L]$ and the boundary condition $y(t,0)=0$ and $y(t,L)=0$ for all time $t\geq 0$.
Consider the case of a {\em pinned string}, that is, $y(t,0)=y(t,L)=0$ and $\frac{\partial^2 y}{\partial x^2}(t,0)=\frac{\partial^2 y}{\partial x^2}(t,L)=0$ for all $t$. The solution $y(t,x)$ is the transversal displacement of the string point $x$ at time $t$ \cite{fletcher1964normal,fletcher2010physics}, which is the linear combination of the {\em normal modes} represented by $y_n(t,x)=\sin(2\pi k_nx)\sin(2\pi {\xi_n} t)$ with $k_n=\frac{n}{2L}$, $\xi_1=\frac{1}{2L}\sqrt{\frac{T}{\mu}}$ and
\begin{equation}
\xi_n = n\xi_1\sqrt{1+\beta n^2}\,,
\end{equation}
where $\beta=\frac{\pi^2ESK^2}{TL^2}$; that is, the $n$-th component with the $n$-th lowest frequency is deviated from $n\xi_1$ in a nonlinear way.
In other words, the sound associated with the solution oscillates with a non-sinusoidal wave and the fundamental frequency is $\frac{1}{2L}\sqrt{\frac{T}{\mu}}$ with several multiples. Clearly, when $E=0$, (\ref{Equation:Music:Stiffness}) is reduced to the wave equation, and the solution is well known.
In music signal processing, this phenomenon is well known as {\em inharmonicity}, which appears in instruments, like piano and guitar. In these instruments, natural vibration appears after the excitation (i.e., plucking or pressing the keyboard) of the modes. For the piano, $\beta$ is in the ranges from around $10^{-4}$ to $10^{-3}$. Obviously, the sound with inharmonicity does not well fit (\ref{model:linearRelationship}).
\subsection{Time-varying wave-shape function}
The above discussions indicate that we need a model with time-varying wave-shape functions.
Thus, we wish to generalize (\ref{model:linearRelationship}). To achieve this goal, we will directly generalize the equivalent expression (\ref{Introduction:ANHM1}) to capture an oscillatory signal with the ``time-varying wave-shape function''.
\begin{defn}[Adaptive non-harmonic function]\label{Definition:ANHFunction}
Take $\epsilon { > 0}$, a non-negative { $\ell^1$} sequence $c =\{c(\ell)\}_{\ell=0}^\infty$, $0<C<\infty$, and $N\in\mathbb{N}$. The set $\mathcal{D}_{\epsilon}^{c,C,N}\subset C^1(\mathbb{R})\cap L^\infty(\mathbb{R})$ of {\it adaptive non-harmonic (ANH) functions} is defined as the set consisting of functions
\begin{align}
f(t)=\frac{1}{2}B_0(t)+\sum_{\ell=1}^\infty B_\ell(t)\cos(2\pi \phi_\ell(t))\label{model:nonlinearRelationship}
\end{align}
satisfying the following conditions:
\begin{itemize}
\item the {\em regularity condition} :
\begin{align}
&B_\ell\in C^1(\mathbb{R})\cap L^\infty(\mathbb{R}),\, { \mbox{ for } \ell = 0 ,\ldots \infty,}\\
&\phi_\ell\in C^2(\mathbb{R}), \, { \mbox{ for } \ell= {1} ,\ldots \infty.}
\end{align}
For all $t\in\mathbb{R}$, $B_\ell(t)\geq 0$ for all {$\ell=0,1,2,\ldots,\infty$} and $\phi'_\ell(t)>0$ for all $\ell=1,\ldots,\infty$.
\item the {\em time-varying wave-shape} condition: for all $t\in\mathbb{R}$,
\begin{equation}\label{Condition:ANH:phi_ell}
\left|\phi'_\ell(t)-\ell\phi'_1(t)\right|\leq \epsilon \phi'_1(t)
\end{equation}
for all $\ell=1,\ldots,\infty$,
\begin{equation}\label{Condition:ANH:B_ell}
{B_\ell(t)\leq c(\ell)B_1(t)}
\end{equation}
for all {$\ell=0,1,\ldots,\infty$,
\begin{equation}\label{Condition:ANH:B_elltail1}
\sum_{\ell=N+1}^\infty B_\ell(t)\leq \epsilon \sqrt{\frac{1}{4}B_0(t)^2+\frac{1}{2}\sum_{\ell=1}^\infty B_\ell(t)^2},
\end{equation}
and
\begin{equation}\label{Condition:ANH:B_elltail2}
\sum_{\ell=1}^\infty \ell B_\ell(t)\leq C \sqrt{\frac{1}{4}B_0(t)^2+\frac{1}{2}\sum_{\ell=1}^\infty B_\ell(t)^2}.
\end{equation}}
\item the {\em slowly varying} condition: for all $t\in\mathbb{R}$,
\begin{align}
\label{def:slow_varying}
&|B_\ell'(t)|\leq \epsilon c(\ell)\phi_1'(t),\, { \mbox{ for } \ell = 0 ,\ldots \infty,}\\
&|\phi_\ell''(t)|\leq \epsilon \ell\phi_1'(t),\, { \mbox{ for } \ell= {1} ,\ldots \infty,}
\end{align}
and $\|\phi_1'(t)\|_{L^\infty}<\infty$.
\end{itemize}
\end{defn}
The adjective {\em adaptive} in ANH function indicates that the frequency and amplitude are time-varying, and the adjective {\em non-harmonic} indicates that the oscillation might be non-sinusoidal. When $\frac{B_{\ell}(t)}{B_1(t)}$ are constants for all $\ell=0,1,\ldots,\infty$ and $\phi'_\ell(t)=\ell\phi_1'(t)+\alpha_\ell$ for some $\alpha_\ell\in\mathbb{R}$ for all $\ell=1,\ldots,\infty$, (\ref{model:nonlinearRelationship}) is reduced to (\ref{Introduction:ANHM1});
when the other conditions for the wave-shape function in (\ref{model:linearRelationship}) are further satisfied, (\ref{model:nonlinearRelationship}) is reduced to (\ref{model:linearRelationship}).
Thus, (\ref{model:nonlinearRelationship}) is a direct generalization of (\ref{Introduction:ANHM1}) by allowing $c_\ell$ and $\alpha_\ell$ in (\ref{Introduction:ANHM1}) to vary, which quantifies the time-varying wave-shape function.
We call $B_1(t)\cos(2\pi \phi_1(t))$ the {\it fundamental component} and $\phi'_1$ the {\it fundamental IF} (or {\em pitch} in the music signal analysis) of the signal $f(t)$. {Note that the condition $|\phi_1''(t)|\leq \epsilon \phi_1'(t)$ says that locally the fundamental IF is nearly constant, but it does not imply that the fundamental IF is nearly constant globally.} By a slight abuse of terminology, for $\ell>1$, we call $B_\ell\cos(2\pi\phi_\ell(t))$ the {\it $\ell$-th multiple}, although $\phi_\ell$ might not be proportional to $\phi_1$. {Note that we can ``view'' $\sqrt{\frac{1}{4}B_0(t)^2+\frac{1}{2}\sum_{\ell=1}^\infty B_\ell(t)^2}$ as the AM of $f(t)$. This comes from the fact that in (\ref{Introduction:ANHM1}), $A(t)^2a_0^2+2\sum_{\ell=1}^\infty A(t)^2a^2_\ell=A^2(t)$, and $B_k(t)$ is the the generalization of $A(t)a_k$ in (\ref{Introduction:ANHM1}) for $k=0,1,\ldots,\infty$. In this definition, however, we do not control how large $\sqrt{\frac{1}{4}B_0(t)^2+\frac{1}{2}\sum_{\ell=1}^\infty B_\ell(t)^2}$ should be. Also note that the series $\big(\frac{B_0(t)/2}{\sqrt{\frac{1}{4}B_{0}(t)^2+\frac{1}{2}\sum_{\ell=1}^\infty B_{\ell}(t)^2}},\,\frac{B_1(t)/\sqrt{2}}{\sqrt{\frac{1}{4}B_{0}(t)^2+\frac{1}{2}\sum_{\ell=1}^\infty B_{\ell}(t)^2}},\ldots,\big)$ has the unitary $\ell^2$ norm, which is a generalization of the assumption that the wave-shape function has the unitary $L^2$ norm. The condition (\ref{Condition:ANH:B_elltail1}) says that only the first $N$ multiples are significant. The condition (\ref{Condition:ANH:B_elltail2}) is a direct generalization of the $C^{1,\alpha}$ condition of the wave-shape function in (\ref{model:linearRelationship}).}
To see how the wave-shape function varies according to time, denote {$t_k:=\phi_1^{-1}(k)$}. Clearly, for signals in $I_k:=[t_k,t_{k+1})$, we could not find a single 1-periodic function $s(t)$ so that $\frac{f}{B_1(t)}|_{I_k}$ is the composition of $s$ and $\phi_1(t)$.
Thus, we could view the model (\ref{model:nonlinearRelationship}) either as an adaptive non-harmonic model with one oscillatory component with the {\it time-varying wave-shape function}, or as an adaptive harmonic model with many oscillatory components with the sinusoidal wave pattern.
\begin{defn}[Adaptive non-harmonic model]\label{DefBClassMultipleTimeSeries}
Take $\epsilon { > 0}$ and $d>0$.
The set $\mathcal{D}_{\epsilon,d}\subset C^1(\mathbb{R})\cap L^\infty(\mathbb{R})$ consists of \textit{superposition of ANH functions}, that is
\begin{equation}
f(t)=\sum_{k=1}^{K}f_{k}(t)
\end{equation}
for some finite $K>0$ and \
\begin{equation*}
f_{k}(t)={\frac{1}{2}B_{k,0}(t) +}\sum_{\ell={1}}^\infty B_{k,\ell}(t)\cos(2\pi \phi_{k,\ell}(t))\in \mathcal{D}_{\epsilon_k}^{c_{k},C_k,N_k}
\end{equation*}
for some $0\leq \epsilon_k\leq \epsilon$, non-negative sequence $c_{k}=\{c_k(\ell)\}_{\ell={0}}^\infty$, $0<C_k<\infty$ and $N_k\in\mathbb{N}$, where {for all $t\in\mathbb{R}$,} the fundamental IF's of all ANH functions satisfy
\begin{itemize}
\item the {\em frequency separation} condition:
\begin{equation}\label{condition_Cepsilon_d}
\phi'_{k,1}(t)-\phi'_{k-1,1}(t)\geq d
\end{equation}
for $k=2,\ldots,K$
\item the {\em non-multiple} condition: for each $k=2,\ldots,K$, $\phi'_{k,1}(t)/\phi'_{\ell,1}(t)$ is not an integer for $\ell=1,\ldots,k-1$.
\end{itemize}
\end{defn}
\section{De-shape SST}\label{Section:Cepstrum}
In this section, we propose an algorithm, de-shape SST, to study a given oscillatory signal. De-shape SST provides a TF representation which contains essentially the IF and AM information of the fundamental component of each ANH function and removes the influence caused by the non-trivial wave-shape function. In Section \ref{Section:ReviewCepstrum}, we provide a review of how cepstrum is applied in engineering.
{In Section \ref{Section:TFcepstrum}, the short time cepstral transform (STCT) is introduced with a theoretical justification in Theorem \ref{Theorem:TimeVaryingCepstrum} to generalize cepstrum to the time-frequency analysis.}
The proof of the theorem is postponed to the Appendix.
{In Section \ref{Section:deShape}, we introduce the inverse STCT that will be used in the de-shape algorithm. }
In Sections {\ref{Section:deshapeSTFT}-\ref{Section:Synchrosqueezing}}, the de-shape STFT and de-shape SST are discussed.
\subsection{A quick review of cepstrum}\label{Section:ReviewCepstrum}
Cepstrum is a commonly applied signal processing technique \cite{Oppenheim_Schafer:2009}.
One motivation of introducing cepstrum is the {\em pitch detection} problem in music (recall that pitch means the fundamental frequency).
It is closely related to the {\em homomorphic signal processing}, which aims at converting signals structured by complicated algebraic systems into simple ones.
Since its invention in 1963 \cite{bogert1963quefrency}, the cepstrum has been applied in various discrete-time signal processing problems, such as detecting the echo delay, deconvolution, feature representations for speech recognition like the Mel-Frequency Cepstral Coefficients (MFCC), and estimating the pitch of an audio signal. A thorough review of the cepstrum can be found in \cite{oppenheim2004frequency,Oppenheim_Schafer:2009}.
We start from recalling the {\em complex cepstrum}. For a suitable chosen signal $f(t)\in\mathbb{R}$, the {\em cepstrum}, denoted as $\tilde{f}^{C}(q)$, where $q\in\mathbb{R}$ is called {\em quefrency}\footnote{The term ``cepstrum'' is invented by interchanging the consonants of the first part of the word ``spectrum'' in order to signify their difference. Similarly, the word ``{quefrency}'' is the inversion of the first part of ``frequency''. By definition, the quefrency has the same unit as time.}, is defined as the inverse Fourier transform of the logarithm of the Fourier transform \cite{Oppenheim_Schafer:2009}:
\begin{equation}
\tilde{f}^{C}(q) := \int \log \hat{f}(\xi) e^{2\pi iq\xi } d\xi \,,\label{eq:cepstrum_basic}
\end{equation}
whenever the inverse Fourier transform of $\log \hat{f}(\xi)$ makes sense, where $\log$ is defined on a chosen branch. We call the domain of $\tilde{f}^C$ the {\em quefrency domain}. Numerically, since the computation of the complex cepstrum requires the phase unwrapping process, it causes instability. Therefore, we could also consider the {\em real cepstrum}, denoted as $\tilde{f}^{R}$, which is represented as
\begin{equation}
\tilde{f}^{R}(q) = \int \log |\hat{f}(\xi)| e^{-2\pi iq\xi }d\xi\,,\label{eq:real_cepstrum}
\end{equation}
whenever the Fourier transform of $\log |\hat{f}(\xi)|$ makes sense. Note that there is no difference to take the Fourier transform or inverse Fourier transform since the signal is in general real, so we take the Fourier transform instead of the inverse Fourier transform.
In audio signal analysis, the logarithm operation on the magnitude spectrum can be interpreted to be an approximation of the perceptual scale of sound intensity, thus it is conventionally measured in dB.
Intuitively, the cepstrum measures ``the rate of the harmonic peaks per Hz'', namely the {\em period} of the signal, where the period is the inverse of the frequency; that is, the prominent peaks in the cepstrum indicate the periods and their multiples in the signal.
Besides periodicity detection, this method has also been used in a wide variety of fields which requires deconvolution of a {\em source-filter model}.
The main idea behind cepstrum is to find ``the spectral distribution of the spectrum'', which contains the period information of the signal. It is effective since it could transform the ``slow-varying envelope'' of the spectrum to the low-quefrency range, separated from the fast-varying counterpart of the spectrum, which is transformed to the high-quefrency range and represents the period information of the signal.
\begin{example}
We consider an acoustic signal to demonstrate how the overall idea beyond cepstrum or homomorphic signal processing could help in signal processing when the signal comes from a complicated combination of two components. A human voice $f\in C^\infty$ could be modeled by the glottal vibration, which is a pulse sequence $g=p\sum_{k\in\mathbb{Z}}\delta_{T_0k}\in \mathcal{S}'$, where $p\in \mathcal{S}$ and $T_0>0$, convolved with the impulse response of the vocal tract $h\in\mathcal{S}$ so that $\hat{h}$ is a non-negative function, i.e., $f(t)=(g\star h)(t)$. A mission of common interest is to separate these two components.
First, the Fourier transform converts the convolution into multiplication in the frequency domain $\hat{f}(\xi)=\hat{h}(\xi)\hat{g}(\xi)$, where $\hat{g}=\frac{1}{T_0}\hat{p}\star \sum_{k\in\mathbb{Z}}\delta_{k/T_0}\in C^\infty$ by the Poisson summation formula. Second, the logarithm converts multiplication into addition, but we have to be careful when we take the logarithm. To simplify the discussion, we assume that supports of both $\hat{g}$ and $\hat{h}$ are positive-valued. Thus, $\log(\hat{f}(\xi)) = \log(\hat{g}(\xi))+\log(\hat{h}(\xi))$. Thus, the convolution operator in the time domain becomes the addition operator. Although under our simplified assumption, $\hat{g}\in C^\infty\cap L^\infty$ and $\hat{h}\in\mathcal{S}$, after taking logarithm we might not be able to define the Fourier transform. So we further assume that $\log(\hat{g}(\xi)),\log(\hat{h}(\xi))\in\mathcal{S}'$ so that we could apply the Fourier transform. For example, if $h$ is a Gaussian function, $\log(\hat{h}(\xi))$ is a quadratic polynomial function. We call the domain where $\mathcal{F}\log(\hat{f})$ is defined the quefrency domain.
In summary, the periodic glottal excitation is modeled as a series of harmonic peaks in the frequency domain by the Poisson summation formula (contributing to pitch), while the frequency response of the vocal tract, $\hat{h}(\xi)$, contributes to the amplitude of the spectrum. { Let us further assume} that after taking Fourier transform on $\log(\hat{f}(\xi))$ the glottal excitation lies in the high quefrency region while the vocal tract in the low quefrency region\footnote{In the music processing, the high-quefrency part in the cepstrum is related to the pitch while the low-quefrency part to timbre (i.e., sound color). }, then a simple high pass filtering, which is called the {\em liftering} (again an interchange of the consonants of ``filtering'') process, can separate the two components. One simple example of $h$ is that when $h$ is a Gaussian function, the Fourier transform of $\log(\hat{h}(\xi))$ is proportional to the second distributional derivative of the Dirac measure supported at $0$. These two components could then be reconstructed by reversing the procedure -- apply the Fourier transform, take the exponential and apply the inverse Fourier transform. The whole process is called the {\em homomorphic deconvolution}.
\end{example}
Although the real cepstrum avoids phase unwrapping, it is still limited by evaluating the logarithm, which is prone to numerical instability either in synthetic data or real-world data. To address this issue, it has been proposed in the literature to replace the logarithm by the {\em generalized logarithm function} \cite{kobayashi1984spectral,tokuda1994mel,taxt1997comparison},
\begin{equation}
L_\gamma(x):=\frac{|x|^\gamma-1}{\gamma},
\end{equation}
where $\gamma>0$, or the {\em root function} \cite{lim1979spectral,alexandre1993root,taxt1997comparison}, defined as
\begin{equation}
g_\gamma(x):=|x|^\gamma,
\end{equation}
where $\gamma>0$. Note that $L_\gamma$ approximates the logarithm function as $\gamma\to 0$. As $g_\gamma$ and $L_\gamma$ are related by a constant and a dilation, there is no practical difference which relaxation we choose. Thus, although we could also consider the generalized logarithm function $L_\gamma$ \cite{kobayashi1984spectral,tokuda1994mel}, to simplify the discussion, in this paper we relax the real cepstrum by the root function $g_\gamma$, and we call the resulting ``cepstrum'' the {\em $\gamma$-generalized cepstrum} (In the literature it is also called the root cepstrum):
\begin{equation}
\tilde{f}_\gamma(q) := \int g_\gamma(\hat{f}(\xi)) e^{{-}2\pi iq\xi }d\xi\,.\label{eq:root_cepstrum}
\end{equation}
There are several proposals for the choice of $\gamma$.
First, when $\gamma=2$, the formulation is equivalent to the autocorrelation function of $f$, which is a basic feature for {\em single} pitch detection but has been found unfeasible for multipitch estimation (MPE). To deal with the issue of multipitch, we should consider $\gamma<2$ \cite{kobayashi1984spectral,tolonen2000computationally}.
$\gamma=0.67$ is suggested by the nonlinear relationship between the sound intensity and perceived loudness determined by experiment, known as a case of Stevens' power law, which states that the sound intensity $x$ and the perceived loudness $y$ are related by $y\propto|x|^{0.67}$ \cite{stevens1957psychophysical,hermansky1990perceptual,tolonen2000computationally}.
Previous researches also suggested $\gamma$ to be $0.6$ \cite{kraft2014polyphonic}, $0.25$ \cite{indefrey1985design} and $0.1$ \cite{klapuri2008multipitch}. In short, the $\gamma$-generalized cepstrum has been shown more robust to noise than the real or complex cepstrum in the literature of speech processing \cite{lim1979spectral,alexandre1993root}.
In addition to the robustness, the $\gamma$-generalized cepstrum has been found useful in various problems like speech recognition \cite{hermansky1990perceptual}, speaker identification \cite{zhao2013analyzing}, especially in multiple pitch estimation \cite{indefrey1985design,tolonen2000computationally,klapuri2008multipitch,kraft2014polyphonic,su2015combining,su2016exploiting}. Due to its usefulness and for the sake of simplification, in the paper we focus on the $\gamma$-generalized cepstrum.
\subsection{Combining cepstrum and time-frequency analysis -- short time cepstral transform (STCT)}\label{Section:TFcepstrum}
As useful as the Fourier transform is for many practical problem, however, it has been well known that when the IF or AM is not constant, Fourier transform might not perform correctly. Indeed, for the ANH functions, since IF and AM are time-varying, the momentary behavior of oscillation is mixed up by the Fourier transform, and hence the cepstrum approach discussed in the previous section fails. To study this kind of dynamical signal, we need a replacement for the Fourier transform. A lot of efforts have been made in the past few decades to achieve this goal. TF analysis based on different principals \cite{Flandrin:1999} has attracted a lot of attention in the field and many variations are available. Well known examples include short time Fourier transform (STFT), continuous wavelet transform (CWT), Wigner-Ville distribution (WVD), etc. We refer the reader to \cite{Daubechies_Wang_Wu:2016} for a summary of the current progress of TF analysis. In this paper, we consider STFT, since it is a direct and intuitive generalization of the Fourier transform. A generalization of cepstrum to other TF analyses will be studied in {future works}.
Recall the definition of STFT. For a chosen window function $h\in\mathcal{S}$, the STFT of $f\in \mathcal{S}'$ is defined by
\begin{equation}
V^{(h)}_f(t, \xi) = \int f(\tau) h(\tau-t)e^{-i2\pi \xi (\tau-t)} \,\mathrm{d} \tau\,,\label{eq: stft1}
\end{equation}
where $t\in\mathbb{R}$ indicates time and $\xi\in\mathbb{R}$ indicates frequency\footnote{The phase factor $e^{i2\pi\xi t}$ in this definition is not always present in the literature, leading to the name {\em modified STFT} for this particular form. To slightly abuse the notation, we still call it STFT.}. We call $V^{(h)}_f(t, \xi)$ the TF representation of the signal $f$.
Since STFT could capture the spectrum or local oscillatory behavior of a signal, we could combine the ideas of STFT and cepstrum, which leads to the {\em short time cepstral transform (STCT)}:
\begin{defn}
Fix $\gamma>0$. For $f\in\mathcal{S}'$ and $h\in\mathcal{S}$, we have the {\em short time cepstral transform (STCT)}:
\begin{equation}
C^{(h,\gamma)}_f(t, q) := \int g_\gamma(V^{(h)}_f(t, \xi)) e^{-i2\pi q \xi} \,\mathrm{d} \xi,
\label{eq: rceps1}
\end{equation}
where $q\in\mathbb{R}$.
\end{defn}
$q$ in $C^{(h,\gamma)}_f(t, q)$ is called the {quefrency}, and its unit is second or any feasible unit in the time domain. Clearly, $C^{(h,\gamma)}_f(t, \cdot)$ is the $\gamma$-generalized cepstrum of the signal $f(\cdot) h(\cdot-t)$ and in general $C^{(h,\gamma)}_f(t, q)$ is not positive. To show the well-definedness of STCT, note that while $f\in\mathcal{S}'$ and $h\in\mathcal{S}$, $V^{(h)}_f(t, \xi)\in C^\infty$ is smooth and slowly increasing on both time and frequency axes. By a slowly increasing $C^\infty$ function $f$, we mean that $f$ and all its derivatives have at most polynomial growth at infinity. Thus, we know that $g_\gamma(V^{(h)}_f(t, \cdot))$ is continuous and slowly increasing. Hence its Fourier transform can be well-defined in the distribution sense since a continuous slowly increasing function is a tempered distribution. In the special case that $f\in C^\infty\cap L^\infty$, $g_\gamma(V^{(h)}_f(t, \xi))$ is a continuous function vanishing at infinity faster than any power of $|\xi|$, and hence $C^{(h,\gamma)}_f(t, q)$ is a well-defined continuous function in the {quefrency} axis.
As discussed above, since the cepstrum provides the information about periodicity, we call $C^{(h,\gamma)}_f(t,q)$ the {\em time-periodicity} (TP) representation of the signal $f$. Before proceeding, we consider the following example to demonstrate how the STCT works.
\begin{example}
Consider the Dirac comb $f(t)=\frac{1}{\xi_0}\sum_{k\in\mathbb{Z}}\delta_{k/\xi_0}$, where $\xi_0>0$. This is the typical periodic distribution, and we could view it as an ANH function with $K=1$, the delta measure as the shape function, the constant fundamental frequency $\xi_0$ Hz and the constant fundamental period $1/\xi_0$, although the wave-shape function is more general than { what} we consider in the ANH model; {it is more general than the ANH model since with the delta measure $f(t)$ does not satisfy the ANH model.} By the Poisson's summation formula, $f(t)=\sum_{k\in\mathbb{Z}}e^{i2\pi k\xi_0 t}$, where the summation holds in the distribution sense.
Choose a smooth window function $h\in\mathcal{S}$ so that $\hat{h}$ is supported on $[-\Delta,\Delta]$, where $0<\Delta<\xi_0/2$. By a direct calculation, the STFT of $f$ is
\begin{equation}
V_f^{(h)}(t,\xi)=\sum_{k\in\mathbb{Z}}\hat{h}(\xi-k\xi_0),
\end{equation}
and since $\Delta<\xi_0/2$,
\begin{equation}
|V_f^{(h)}(t,\xi)|=\sum_{k\in\mathbb{Z}}|\hat{h}(\xi-k\xi_0)|.
\end{equation}
To evaluate the STCT, where $\gamma>0$, we need to evaluate $|V_f^{(h)}(t,\xi)|^\gamma$. Under our assumption, it is trivial and we have
\begin{equation}
|V_f^{(h)}(t,\xi)|^\gamma=\sum_{\ell\in\mathbb{Z}}|\hat{h}(\xi-\ell\xi_0)|^\gamma= (\sum_{\ell\in\mathbb{Z}}\delta_{\ell\xi_0}\star |\hat{h}|^\gamma)(\xi),
\end{equation}
where $|\hat{h}|^\gamma\in C_c^0(\mathbb{R})$. Note that the convolution is well-defined since $\sum_{\ell\in\mathbb{Z}}\delta_{\ell\xi_0}$ is a tempered distribution and $|\hat{h}|^\gamma$ is a compactly supported distribution. By taking Fourier transform of $|V_f^{(h)}(t,\xi)|^\gamma$ and applying the Poisson summation formula, the STCT of $f$ is
\begin{equation}
C_f^{(h,\gamma)}(t,q)=\frac{\widehat{|\hat{h}|^\gamma}(q)}{\xi_0}\sum_{\ell\in\mathbb{Z}}\delta_{\ell/\xi_0}{(q)}\,,
\end{equation}
which provides the period information.
\end{example}
This example indicates the overall behavior of STCT when there is only one periodic function with a non-sinusoidal wave. In general, when there are more than one oscillatory functions with non-sinusoidal waves and different fundamental frequencies, the calculation is no longer direct since the multiples of different oscillatory functions may collide. Moreover, since the frequency and amplitude are time-varying, the calculation is more intricate. For the signals in { the set $\mathcal{D}_{\epsilon,d}$ defined in Definition \ref{DefBClassMultipleTimeSeries}}, however, we have the following Theorem showing how STCT works.
Before stating the theorem, we make the following general assumption about the window function.
\begin{assumption}\label{Assumption:GeneralAssumption}
Fix $\epsilon {>0}$ and $d>0$. Take $f(t)=\sum_{k=1}^Kf_k(t)\in \mathcal{D}_{\epsilon,d}$ for some $K\geq 1$. Suppose the fundamental frequency satisfies
\begin{equation}
\inf_{t\in\mathbb{R}} \phi_{1,1}'(t) > 0.
\end{equation}
Fix a window function $h\in \mathcal{S}$, which is chosen so that $\hat{h}$ is compactly supported and $\text{supp}(\hat{h})\subset[-\Delta,\Delta]$, where $\Delta>0$. Also assume that $\Delta$ is small enough so that
\begin{equation}
0<\Delta<\min\{\inf_{t\in\mathbb{R}}\phi'_{1,1}(t)/4,\,d/4\}.
\end{equation}
\end{assumption}
For a chosen window $h\in \mathcal{S}$, denote
\begin{equation}
I_k:=\int |h(x)||x|^kdx,
\end{equation}
where $k\in\{0\}\cup \mathbb{N}$.
We mention that a more general window could be considered with more error terms showing up in the proof. Since these extra efforts do not provide more insight about the theory, we choose to work with this setup.
\begin{defn} {Let $\phi_{k,\ell}(t)$ and $B_{k,\ell}(t)$ for $k=1,\cdots, K$ and $\ell \in \mathbb{N}$ be defined as in Definition \ref{DefBClassMultipleTimeSeries}.}
Under Assumption \ref{Assumption:GeneralAssumption}, define
\begin{equation*}
\phi_{k,-\ell}(t):=-\phi_{k,\ell}(t)\mbox{ and }B_{k,-\ell}(t):=B_{k,\ell}(t)
\end{equation*}
for $\ell\in \mathbb{N}$,
\[
\phi_{k,0}=0\,,
\]
and define a set of intervals
\begin{align}
Z_{k,\ell}(t)&=[\ell\phi'_{k,1}(t)-\Delta,\ell\phi'_{k,1}(t)+\Delta]\subset \mathbb{R}\,,\label{Definition:Zkl}
\end{align}
associated with $f$, where $k\in\{1,\ldots,K\}$ and $\ell\in\mathbb{Z}$.
\end{defn}
The following Theorem describes the behavior of STCT when the signal is in $\mathcal{D}_{\epsilon,d}$. The proof of Theorem \ref{Theorem:TimeVaryingCepstrum} is postponed to Appendix \ref{Appendix:Proof}.
\begin{thm}\label{Theorem:TimeVaryingCepstrum}
Suppose Assumption \ref{Assumption:GeneralAssumption} holds.
The STFT of $f$ at time $t\in\mathbb{R}$ is
\begin{align}
V^{(h)}_f(t,\xi)=
\frac{1}{2}\sum_{k=1}^K\sum_{\ell=-N_k}^{N_k} B_{k,\ell}(t)\hat{h}(\xi-\phi'_{k,\ell}(t))e^{i2\pi\phi_{k,\ell}(t)}+\epsilon(t,\xi),\label{MainTheorem:STFTExpansion}
\end{align}
where $\xi\in\mathbb{R}$ and $\epsilon(t,\xi)$ is defined in (\ref{Lemma:Proof:Definition:epsilon}). Furthermore, $\epsilon(t,\xi)$ is of order $\epsilon$ and decays at the rate of $|\xi|^{-1}$ as $|\xi|\to \infty$.
Take $0<\gamma\leq 1$.
For each $k\in\{1,\ldots,K\}$, denote a series $b_k\in \ell^1$, where $b_k(j)=0$ for all $|j|>N_k$ and $b_k(j)=B^\gamma_{k,j}(t)$ for all $|j|\leq N_k$. Then, for $q>0$, we have
\begin{align}
C^{(h,\gamma)}_f(t,q)=\frac{\widehat{|\hat{h}|^\gamma}(q) }{2^\gamma}\sum_{k=1}^K \hat{b}_k(q) +E_1+ E_2,\label{Theorem:STCT:MainStatement}
\end{align}
where $\hat{b}_k$ is {is the discrete-time Fourier transform of $b_k$}, $E_1$ is defined in (\ref{Theorem:Statement:STCT:E1}), {which is the Fourier transform of $\delta_3$ defined in (\ref{Proof:Lemma:Definition:delta3part2}),} and $E_2$ is defined in (\ref{Theorem:Statement:STCT:E2}), which {is the Fourier transform of $\epsilon_3$ defined in (\ref{Theorem:Statement:STCT:epsilon4}) and} in general is a distribution. {When $K=1$, $E_1=0$, and when $K>1$ it} satisfies
\begin{equation*}
|E_1|\leq {2}\Delta I_0^\gamma \sum_{k=2}^K B^\gamma_{k,1}(t)\|c_k^\gamma\|_{\ell^\infty}N_k\sum_{\ell=1}^{k-1}\Big[\frac{4\Delta}{\phi'_{\ell,1}(t)}+E^{(\ell)}(N_k)\Big],
\end{equation*}
where $E^{(\ell)}(N_k) { =o(N_k)}$ is defined in (\ref{Theorem:Statement:STCT:Eell}). $E_2$ satisfies $|E_2(\psi)|\leq \|{\epsilon_3}(t,\cdot)\|_{L^\infty} \|\hat{\psi}\|_{L^1}$ for all $\psi\in \mathcal{S}$, {and $\epsilon_3$} is of order {$\epsilon^\gamma$}.
\end{thm}
{
The equation (\ref{Theorem:STCT:MainStatement}) does not indicate the relationship between the relationship among $\frac{\widehat{|\hat{h}|^\gamma}(q) }{2^\gamma}\sum_{k=1}^K \hat{b}_k(q)$, $E_1$, and $E_2$, so we could not conclude that we could obtain the inverse of the IF from the STCT. We need more conditions to obtain what we are after. The following corollary is immediate from Theorem \ref{Theorem:TimeVaryingCepstrum}.
\begin{cor}\label{Corollary:TimeVaryingCepstrum}
Fix $\epsilon>0$ and $d>0$. Take $f(t)=\sum_{k=1}^Kf_k(t)\in \mathcal{D}_{\epsilon,d}$ for some $K\geq 1$. In addition to Assumption \ref{Assumption:GeneralAssumption}, suppose
\begin{equation}\label{Assumption:ANH:B_ell}
\frac{B_{k,\ell}(t)}{\sqrt{\frac{1}{4}B^2_{k,0}(t)+\frac{1}{2}\sum_{\ell=1}^\infty B^2_{k,\ell}(t)}} > \epsilon^{1/2}
\end{equation}
for all $k=1,\ldots,K$ and $\ell=0,1,\ldots,N_k$.
Then, when $\epsilon<1$ is sufficiently small, $\Delta N_k$ is sufficiently small and $\sqrt{\frac{1}{4}B^2_{k,0}(t)+\frac{1}{2}\sum_{\ell=1}^\infty B^2_{k,\ell}(t)}$ is sufficiently large for $k=1,\ldots,K$, the term $\frac{\widehat{|\hat{h}|^\gamma}(q) }{2^\gamma}\sum_{k=1}^K \hat{b}_k(q)$ in (\ref{Theorem:STCT:MainStatement}) dominates and $\hat{b}_k$ is a real, continuous and periodic function of period $1/\phi_{k,1}'(t)$ for $k=1,\ldots,K$.
\end{cor}
The assumption that $\sqrt{\frac{1}{4}B^2_{k,0}(t)+\frac{1}{2}\sum_{\ell=1}^\infty B^2_{k,\ell}(t)}$ is sufficiently large for $k=1,\ldots,K$ means that the ANH functions we have interest in have large enough AMs.
Condition (\ref{Condition:ANH:B_elltail1}) and Assumption (\ref{Assumption:ANH:B_ell}) together mean that the first $N_k$ multiples of the fundamental component of the $k$-th ANH function are strong enough, while the remaining multiples are not significant. When $\gamma$ is chosen small enough, this assumption leads to the fact that $b_k(j)=B_{k,j}^\gamma(t)$ is close to $1$ for $|j|\leq N_k$, and ``small'' otherwise. Thus, the Fourier transform of the $\ell^1$ series $b_k$ reflects faithfully the inverse of the IF. We could call $1/\phi_{k,1}'(t)$ the {\em instantaneous period (IP)} of the $k$-th ANH function, which is the inverse of its fundamental frequency.
Note that the assumption (\ref{Assumption:ANH:B_ell}) can be generalized, but more conditions are needed to guarantee that we obtain the IP. For example, if the condition (\ref{Assumption:ANH:B_ell}) is failed for $\ell=2j$ so that $B_{k,2j}(t)=0$ for $j=1,\ldots,\lfloor N_k/2\rfloor$, then the $\ell^1$ series $b_k$ has an oscillation of period $2\phi'(t)$, and hence its Fourier transform is dominant in $\frac{1}{2\phi'(t)}$ instead of $\frac{1}{\phi'(t)}$. This will lead to an incorrect conclusion about the IF in the end. In this paper, to simplify the discussion, we focus on this assumption. See more discussions in the Discussion section.
The bounds for $E_1$ and $E_2$ need some discussions.}
\begin{enumerate}
\item The {bound for} $E_1$ comes from controlling the possible overlaps between the multiples of different ANH components in the STFT $V^{(h)}_f(t,\xi)$. When $K=1$, there is no danger of overlapping, so $E_1=0$. When $K>1$, the term $N_k\sum_{\ell=1}^{k-1}[\frac{4\Delta}{\phi'_{\ell,1}(t)}+E^{(\ell)}(N_k)]$ is the upper bound of all possible overlaps between the $k$-th component and all $\ell$-th component, where $\ell\in\{1,\ldots,k-1\}$. {
The origin of this upper bound is the fundamental Erd\"{o}s-Tur\'{a}n inequality, which gives a quantitative form of Weyl's criterion for equidistribution, and the convergence rate of $E^{(\ell)}(N_k)\to 0$ when $N_k\to\infty$ depends on the algebraic nature of the ratio $\phi_{k,1}'(t)/\phi_{\ell,1}'(t)$. Note that even when the IF's of all oscillatory components are constant, if $K>1$, the $E_1$ term still exists due to the fundamental equidistribution property.
\item When $K>1$, the bound for $E_1$ is the worst bound. Since we could not control the locations of the overlaps between those multiples of different ANH components in the STFT, when we evaluate the STCT by the Fourier transform, the discrepancy caused by the overlaps, denoted as $\delta_3$ in (\ref{Proof:Lemma:Definition:delta3part2}), is bounded simply by the Riemann-Lebesgue theorem. The bound is shown in (\ref{Proof:Lemma:Definition:delta3Fourier}). See Remark \ref{Remark:BoundE1} for more discussions.
The constant could be improved. However, since the focus here is showing how the result is influenced by the fundamental limitation of the number of overlapped multiples, no effort has been made to optimize it.
\item Note that the bound of $E_1$ blows up when $N_k\to \infty$. Thus, the bound of $E_1$ is not useful when $N_k$ is ``huge''. In practice, however, most non-sinusoidal oscillatory signals have $N_k$ less than 20. The most extreme case we have encountered up to now is the ECG signal, which has $N_k$ about 40. Thus, in practice, we could choose a small $\Delta$ so that $E_1$ is well controlled for a ``reasonable'' $N_k$, and this is the condition ``$\Delta N_k$ is sufficiently small'' in Corollary \ref{Corollary:TimeVaryingCepstrum}.
However, $\Delta$ cannot be chosen arbitrarily small. Note that the smaller the $\Delta$ is, the longer the window will be, and the larger the absolute moments $I_k$ will be. Thus, the smaller the $\Delta$ is, the worse the bound of $E_2$ is.
In sum, when $K>1$, except for special non-sinusoidal oscillations with huge $N_k$, the bound for $E_1$ could be well controlled for practical applications. }
\item The {term} $E_2$ comes from the non-constant AM and IF of each ANH component. When the IF and AM are constant, this term becomes zero. Note that when $\gamma$ is chosen small, $b_k$ becomes more like a constant sequence and $\hat{b}_k(q)$ behaves more like a Dirichlet kernel. On the other hand, $E_2$ becomes large when $\gamma$ is small.
\end{enumerate}
{The theorem and the corollary say} that the STCT encodes the IF information in the format of IP via a periodic function.
To better understand periodic functions $\hat{b}_k$, we take a look at the following example.
\begin{example}\label{Example:WithNontrivialShape}
Consider a signal $f(t)=s(\xi_0 t)$, where $\xi_0>0$ and $s$ is real, smooth and 1-periodic. This special case has only one oscillatory component, $K=1$, with the fixed wave-shape function and a constant IF. Thus we do not worry about the error terms $E_1$ and $E_2$ in Theorem \ref{Theorem:TimeVaryingCepstrum}. By a direct expansion, $f(t)=\sum_{k=0}^Nc(k)\cos(2\pi k\xi_0t+\alpha_k)$, where $N\in\mathbb{N}\cup \{\infty\}$, $\alpha_0=0$, $c(1)>0$ and $\alpha_k\in[0,2\pi)$ and $c(k)\geq0$ for all $k\neq 1$. To simplify the calculation, we choose a smooth window function $h$ so that $\hat{h}$ is supported on $[-\Delta,\Delta]$, where $0<\Delta<\xi_0/2$. By the Plancherel identity and a direct calculation, we have
\begin{equation}
V_f^{(h)}(t,\xi)=\sum_{k=-N}^Nc(k)\hat{h}(\xi-k\xi_0)e^{i(2\pi k\xi_0 t+\alpha_k)},
\end{equation}
where we denote $c(-k)=c(k)$ for all $k\in\mathbb{N}$. Since $\Delta<\xi_0/2$, for $0<\gamma\ll 1$, we have
\begin{equation}
|V_f^{(h)}(t,\xi)|^\gamma=\sum_{k=-N}^Nc(k)^\gamma|\hat{h}(\xi-k\xi_0)|^\gamma=\big[\sum_{k=-N}^Nc(k)^\gamma\delta_{k\xi_0}\star |\hat{h}|^\gamma\big](\xi).
\end{equation}
The evaluation of the Fourier transform of $[\sum_{k=-N}^Nc(k)^\gamma\delta_{k\xi_0}\star |\hat{h}|^\gamma](\xi)$ is straightforward and we have for $q>0$,
\begin{align}
C_f^{(h,\gamma)}(t,q)\,= \mathcal{F}[|V_f^{(h)}(t,\cdot)|^\gamma](q)=\widehat{|\hat{h}|^\gamma}(q)\sum_{k=-N}^Nc(k)^\gamma e^{-i2\pi k\xi_0q}\,=\widehat{|\hat{h}|^\gamma}(q)S^{{(\gamma)}}_{1/\xi_0}(q),
\end{align}
where $S^{{(\gamma)}}_{1/\xi_0}(q)$ is a periodic distribution with the period of length $1/\xi_0$ so that $\widehat{S^{{(\gamma)}}_{1/\xi_0}}(k)=c(k)^\gamma$ for $k\in\{-N,-N+1,\ldots,N-1,N\}$ and $\widehat{S^{{(\gamma)}}_{1/\xi_0}}(k)=0$ otherwise.
We could take a look at a special case to have a better picture of what we get eventually. Suppose $N$ is finite and $c(k)=1$ for $k\in\{-N,\ldots,N\}$. In this case, we have
\begin{equation*}
C_f^{(h,\gamma)}(t,q)\,=\widehat{|\hat{h}|^\gamma}(q)D_N(\xi_0q),
\end{equation*}
where $D_N(\xi_0q)$ is the Dirichlet kernel, which is periodic with the period $1/\xi_0$ since $D_N(\xi_0q)=\frac{\sin(\pi(2N+1)\xi_0q)}{\sin(\pi\xi_0q)}$. Also, it becomes more and more spiky at $\ell/\xi_0$ and eventually the Delta comb supported on $\ell/\xi_0$, $\ell\in\mathbb{Z}$, when $N\to\infty$. On the other hand, when $N$ is finite and small, the STCT could be oscillatory but still contains information we need. For example, when $N=1$, $D_1(\xi_0q)=\frac{\sin(\pi 3\xi_0q)}{\sin(\pi\xi_0q)}$ and $D_1(\xi_0q)$ still has dominant values at $q=\ell/\xi_0$ for $\ell\in\mathbb{Z}$ .
\end{example}
\subsection{inverse STCT}\label{Section:deShape}
Based on Theorem \ref{Theorem:TimeVaryingCepstrum} and a careful observation, we see that to determine the fundamental frequency for an ANH signal $f(t)$, a candidate frequency should have the saliency of its multiples in the TF representation $V_f^{(h)}(t,\xi)$, and the associated period and its multiples in the TP representation $C^{(h,\gamma)}_f(t,q)$. In \cite{su2015combining,su2016exploiting}, this observation is summarized as a practical {principle} called the {\em constraint of harmonicity}, which is described as follows: at a specific time $t_0$, a pitch candidate, $\xi_1>0$, is determined to be the true pitch when there exists $M_v, M_u \in \mathbb{N}$ such that there are
\begin{enumerate}
\item A sequence of ``peaks'' found around $V_f^{(h)}(t_0,\xi_1)$, $V_f^{(h)}(t_0,2\xi_1)$, $\ldots$, $V_f^{(h)}(t_0,M_v \xi_1)$;
\item A sequence of ``peaks'' found around $C_f^{(h,\gamma)}(t_0,q_1)$, $C_f^{(h,\gamma)}(t_0,2q_1)$, $\ldots$, $C_f^{(h,\gamma)}(t_0,M_u q_1)$;
\item $\xi_1=1/q_1$.
\end{enumerate}
The sequence $\{\xi_1,2\xi_1,\ldots,M_v\xi_1\}$ is commonly called {\em harmonic series} associated with multiples of the pitch $\xi_1$.
The constraint of harmonicity {principle} leads to the following consideration. If we ``invert'' the quefrency axis of the TP representation by the operator $\mathcal{I}$,
\begin{equation}
\mathcal{I}:q\mapsto 1/q,
\end{equation}
when $q>0$, then by the relationship that the period is the inverse of the frequency, we could obtain information about the frequency in $C_f^{(h,\gamma)}(t,\mathcal{I}q)$. Note that $\mathcal{I}$ is open from $(0,\infty)$ to $(0,\infty)$ and the differentiation of $\mathcal{I}$ is surjective on $(0,\infty)$, so for a distribution $T$ defined on $(0,\infty)$, we could well-define the composition $T\circ \mathcal{I}$, or the pull-back of $T$ via $\mathcal{I}$ \cite[Theorem 6.1.2]{Hormander:1990}. Since in general $C_f^{(h,\gamma)}(t,\cdot)$ is a tempered distribution, we could consider the following definition to extract the frequency information for $f$:
\begin{defn}
For a function $f\in\mathcal{S}'$, window $h\in\mathcal{S}$ and $\gamma>0$, the {\em inverse short time cepstral transform (iSTCT)} is defined on $\mathbb{R}\times \mathbb{R}^+$ as
\begin{equation}
U_f^{(h,\gamma)}(t,\xi):=C_f^{(h,\gamma)}(t,\mathcal{I}\xi),
\end{equation}
where $\xi>0$ and $U_f^{(h,\gamma)}(t,\cdot)$ is in general a distribution.
\end{defn}
The unit of $\xi$ in $U_f^{(h,\gamma)}(t,\xi)$ is Hz or any feasible unit in the frequency domain. We mention that in the special case that $f\in C^\infty\cap L^\infty$, $U^{(h,\gamma)}_f(t, \xi)$ is a well-defined continuous function in the frequency axis. Also, if $C_f^{(h,\gamma)}(t,\cdot)$ is integrable and we want to preserve the integrability, we could weight $C_f^{(h,\gamma)}$ by the Jacobian of $\mathcal{I}$. However, since the integrability is not the main interest here, we do not consider it.
We view $U_f^{(h,\gamma)}(t,\xi)$ as a TF representation determined by a nonlinear transform composed of several transforms.
While this operator looks natural at the first glance, it is actually not stable.
See the following example for the source of the instability.
\begin{example}
{ Let us continue} the discussion of Example \ref{Example:WithNontrivialShape}.
Suppose $N$ is finite and hence $C_f^{(h,\gamma)}(t,q)=\widehat{|\hat{h}|^\gamma}(q)D_N(\xi_0q)$ for $q>0$.
%
Thus, by inverting the axis by $\xi\mapsto 1/\xi$ when $\xi\neq 0$, the iSTCT becomes $U^{(h,\gamma)}_f(t,\xi)= \widehat{|\hat{h}|^\gamma}(1/\xi) D_N(\xi_0/\xi)$, where $\xi>0$.
Clearly, due to the oscillatory nature of the Dirichlet kernel, the non-zero region of $D_N(\xi_0q)$ around $q=0$ would be flipped to the high frequency region, which amplifies the unwanted information in the low frequency and represents it in the high frequency region. To be more precise, since $D_N(\xi_0q)$ decays monotonically from $2N$ to about $-0.43N$ as $q$ goes from $0$ to $x_1\in (\frac{1}{(2N+1)\xi_0},\frac{2}{(2N+1)\xi_0})$, where $x_1$ is the local extremal point, in iSTCT, $U_f^{(h,\gamma)}(t,\xi)$ increases from about $-0.43N$ to $2N$ as $\eta$ goes from $1/x_1$ to $\infty$. This indicates that $|U_f^{(h,\gamma)}(t,\xi)|>N$ for all $\xi>\Xi$ for some $\Xi>1/x_1$.
\end{example}
Motivated by the above example, in practice, we need to apply a filtering process on the STCT to stabilize the algorithm. Here is the main idea. Since our interest is to capture the IF's of the signal, we have to effectively remove components unrelated to IF's in the STCT. In practice, the irrelevant components lie in the low quefrency region. Therefore, we need to apply a {\em long-pass lifter} on $U^{(h,\gamma)}_f(t,\xi)$, where the {\em lifter} refers to a ``filter'' processed in the cepstral domain, again by inverting the first four letters of ``filter'', to distinguish it from the {\em filter} processed in the spectral domain \cite{bogert1963quefrency,oppenheim2004frequency}. Moreover, since the quefrency is measured in the unit of time, a lifter is identified as a {\em short-pass} or {\em long-pass} one rather than a low-pass or a high-pass one \cite{bogert1963quefrency,oppenheim2004frequency}. In short, a long-pass lifter passes mainly the component of high quefrency (long period) while rejects mainly the component of low quefrency (short period).
\subsection{de-shape STFT}\label{Section:deshapeSTFT}
Take the music signal as an example to examine the iSTCT. The constraint of harmonicity {principle} tells us that while at a fixed time $t$ we could find a harmonic series associated with multiples of the pitch $\xi_0$ in the TF representation $V^{(h)}_f(t,\xi)$, we should find a sequence of peaks in the TF representation $U^{(h,\gamma)}_f(t,\xi)$, denoted as $\{\xi_1, \xi_1/2, \ldots, \xi_1/M_u\}$ and this sequence is called the {\em sub-harmonic series} associated with the fundamental frequency $\xi_1$ in the literature.
This observation motivates a combination of the STFT and iSTCT to extract the pitch information; that is, we consider the following combination of the TF representation and TP representation via the iSTCT, which we coined the name {\em de-shape STFT}:
\begin{defn}
For a function $f\in\mathcal{S}'$, window $h\in\mathcal{S}$ and $\gamma>0$, the de-shape STFT is defined on $\mathbb{R}\times \mathbb{R}^+$ as
\begin{equation}
\label{eq:W}
W^{(h,\gamma)}_f(t, \xi) := V^{(h)}_f(t,\xi)U^{(h,\gamma)}_f(t, \xi),
\end{equation}
where $\xi>0$ is interpreted as frequency.
\end{defn}
In general, since $V^{(h)}_f(t,\xi)$ is a $C^\infty$ function in the frequency axis and $U^{(h,\gamma)}_f(t, \xi)$ is a distribution in the frequency axis, the de-shape STFT is well-defined as a distribution. Again, in the special case that $f\in C^\infty\cap L^\infty$, $W^{(h,\gamma)}_f(t, \xi)$ is a well-defined continuous function in the frequency axis.
The motivation beyond the nomination ``de-shape'' is intuitive -- since the harmonic series associated with multiples of the fundamental frequency $\xi_0$ in $V^{(h)}_f(t,\xi)$ overlaps with the sub-harmonic series associated with multiples of the fundamental frequency $\xi_0$ in $U^{(h,\gamma)}_f(t, \xi)$ only at $\xi_0$, by multiplying $V^{(h)}_f(t,\xi)$ and $U^{(h,\gamma)}_f(t, \xi)$, only the information associated with the pitch is left in the result.
Thus, the influence caused by the non-trivial wave-shape function in the TF representation is removed, and hence we could view the de-shape process as an {\em adaptive and nonlinear filtering technique} for the STFT.
Since $\xi>0$ in $W^{(h)}_f(t, \xi)$ is interpreted as frequency, the de-shape STFT provides a TF representation.
We mention that in the music field, a similar idea called the {\em combined temporal and spectral representations} has been applied to the single pitch detection problem \cite{peeters2006music,emiya2007parametric}. With our notation, the proposed idea of detecting the pitch at time $t$, denoted as $\xi_0(t)$, is simply by $\xi_0 (t)=\operatornamewithlimits{argmax}_{\xi>0} |W^{(h,\gamma)}_f(t,\xi)|$ \cite{peeters2006music,emiya2007parametric}.
In the last section of \cite{peeters2006music}, the authors showed a figure of polyphonic music and slightly addressed the ``potential'' of this idea in multiple pitch estimation problems. But this idea was not noticed until \cite{su2015combining,su2016exploiting}, which gives an explicit methodology, systematic investigation, and evaluation of using this idea in multiple pitch estimation.
\subsection{Sharpen de-shape STFT by the synchrosqueezing transform -- de-shape SST}\label{Section:Synchrosqueezing}
While the de-shape STFT could alleviate the influence of the wave-shape function, it again suffers from the Heisenberg-Gabor uncertainty principle and tends to be blurred in the TF representation \cite{Flandrin:1999}.
One approach to sharpen a TF representation is by applying the SST, and we propose to combine SST to obtain a sharp TF representation without the influence of the wave-shape function. SST is a nonlinear TF analysis technique, which is special case of the more general RM method \cite{Auger_Flandrin:1995}. In summary, it aims at moving the spectral-leakage terms caused by Heisenberg-Gabor uncertainty principle to the correct location, and therefore sharpens the TF representation with high concentration \cite{Daubechies_Lu_Wu:2011,Chen_Cheng_Wu:2014,Thakur:2015,Oberlin_Meignen_Perrier:2015}. The main step in SST is estimating the {\em frequency reassignment vector}, which guides how the TF representation should be nonlinearly deformed. The resulting TF representation has been applied to several fields. For example, in the physiological signal processing, SST leads to a a better estimation of IF and AM, which is applied to study sleep dynamics \cite{Wu_Talmon_Lo:2015}, coupling \cite{Iatsenko_Bernjak_Stankovski_Shiogai_Owen_Clarkson_McClintock_Stefanovska:2013} and others, or a better spectral analysis, which is applied to study the noxious stimulation problem \cite{Lin_Wu:2016};
in the mechanical engineering, it has been applied to estimate speed of rotating machinery \cite{Xi_Cao_Chen_Zhang_Jin:2015} and others;
in finance, it is applied to detect the non-stationary dynamics in the financial system \cite{Guharay_Thakur_Goodman_Rosen_Houser:2013};
in the music processing, such an approach can better discriminate closely-located components, and applications have be found in chord recognition \cite{khadkevich2011time}, sinusoidal synthesis \cite{peeters1999sinola} and others.
The frequency reassignment vector associated with a function $f\in\mathcal{S}'$ is determined by
\begin{equation}
\Omega^{(h,\upsilon)}_f(t,\xi):=
\left\{
\begin{array}{ll}
-\Im\frac{V_f^{(\mathcal{D}h)}(t,\xi)}{2\pi V_f^{(h)}(t,\xi)}&\mbox{ when }|V_f^{(h)}(t,\xi)|> \upsilon\\
-\infty&\mbox{ when }|V_f^{(h)}(t,\xi)|\leq \upsilon
\end{array}
\right.
\,,\label{RM:omega}
\end{equation}
where $\mathcal{D}h(t)$ is the derivative of the chosen window function $h\in\mathcal{S}$, $\Im$ means the imaginary part and $\upsilon>0$ gives a threshold so as to avoid instability in computation when $|V^{(h)}_f(t,\xi)|$ is small. The theoretical analysis of the frequency reassignment vector has been studied in several papers \cite{Daubechies_Lu_Wu:2011,Wu:2013,Chen_Cheng_Wu:2014}, and we refer the reader with interest to these papers. {In general, we could consider variations of the reassignment vectors for different purposes. For example, the reassignment vectors used in the second order SST \cite{Oberlin_Meignen_Perrier:2015}. To keep the discussion simple, we focus on the original SST.}
The SST of $V^{(h)}_f(t,\xi)$ is therefore defined as
\begin{equation}\label{definition:SSTV}
SV^{(h,\upsilon)}_{f}(t,\xi)=\int_{\mathfrak{N}_\upsilon(t)} V^{(h)}_f(t,\eta) \frac{1}{\alpha}g\left(\frac{|\xi-\Omega^{(h,\upsilon)}_f(t,\eta)|}{\alpha}\right)\,\mathrm{d} \eta.
\end{equation}
where $\xi\geq0$, $\alpha>0$, $g\in\mathcal{S}$ and $\frac{1}{\alpha}g(\frac{\cdot}{\alpha})$ converges weakly to the Dirac measure supported at $0$ when $\alpha\to 0$, $\mathfrak{N}_\upsilon(t):=\{\xi\geq 0|\, |V^{(h)}_f(t,\xi)|>\upsilon\}$; similarly, we have the {\em de-shape SST} defined as
\begin{equation}\label{definition:SSTW}
SW^{(h,\gamma,\upsilon)}_{f}(t,\xi)=\int_{\mathfrak{N}_\upsilon(t)} W^{(h,\gamma)}_f(t,\eta) \frac{1}{\alpha}g\left(\frac{|\xi-\Omega^{(h,\upsilon)}_f(t,\eta)|}{\alpha}\right)\,\mathrm{d} \eta\,,
\end{equation}
where $\xi\geq0$. Numerically, $g$ could be chosen to be the Gaussian function with $\alpha>0$ or as a direct discretization of the Dirac measure when $\alpha\ll 1$. For numerical implementation details and the stability results of SST, we refer the reader with interest to \cite{Chen_Cheng_Wu:2014}.
With the de-shape STFT, the wave-shape information is decoupled from the IF and AM in the TF representation; with the de-shape SST, the TF representation is further sharpened. We could continue to do the analysis to, for example, carry out the wave-shape reconstruction, count the oscillatory components, etc. Furthermore, we could combine the de-shape SST information and current wave-shape analysis algorithms, including the functional regression \cite[Section 4.7]{Chui_Lin_Wu:2015}, designing a dictionary \cite{Hou_Shi:2016} or unwrapping the phase \cite{Yang:2014}, to study the oscillatory signal with time-varying wave-shape function. The work of estimating the time-varying wave-shape function with applications will be explored systematically in {a} coming work.
\section{Numerical Results}\label{Section:Numerics}
In this section we demonstrate how the de-shape SST performs in various kinds of signals with multiple ANH components with non-trivial time-varying shape function. We consider a wide range of physiological, biological, audio and mechanical signals, which are generated in different dynamical system and recorded by different sensors. The signals are: (1) abdominal fetal ECG signal, (2) different photoplethysmography signals under different challenges -- respiratory and heartbeat, motion and heartbeat, and non-contact PPG signal, (3) music and bioacoustic signals including the violin sonata, choir and wolves sound. The code of SST and de-shape SST and test datasets are available via request.
For a fair comparison, the parameters for computing the de-shape SST are set to be the same for all signals throughout the paper: $\gamma=0.3$ for the STCT and $\upsilon=10^{-4}$\% of the root mean square energy of the signal under analysis for the de-shape SST.
\subsection{Simulated signal}
We continue the example shown in the Introduction section, make clear how $f_{2}$ is generated, and consider a more complicated example.
Take $W$ to be the standard Brownian motion defined on $\mathbb{R}$ and define random processes $\Phi_{A,\sigma,a}:=\frac{(|W|+1)\star K_\sigma}{\|(|W|+1)\star K_\sigma\|_{L^\infty[0,100]}}+a$ and $\Phi_{\phi,\sigma,b,c}:=\frac{b|W|}{\|W\|_{L^\infty[0,100]}}\star K_{\sigma}+c$, where $a,b,c\in \mathbb{R}$, $K_{\sigma}$ is the Gaussian function with the standard deviation $\sigma>0$. $A_2(t)$ is a realization of $\Phi_{A,10,0.9}$, $A_3(t)$ is a realization of $\Phi_{A,10,0.9}$, $\phi_2(t)$ is a realization of $\Phi_{\phi,5,2,\pi/2}$, and $A_3$ is a realization of $\Phi_{\phi,5,1,4}$ on $[0,100]$. Here all realizations are independent.
The signal $f_2$ is generated by $A_2(t)\text{mod}(\phi_2,1)$. To generate $f_3$, denote $t_{k}=\phi_3^{-1}(k)$. The signal $f_3$ is $A_3(t)\chi_{[30,100]}(\sum_{k}\delta_{t_k}\star \chi_{[-3/100,3/100]})$.
Consider a clean signal $f(t)=f_1(t)+f_2(t)+f_3(t)$ from $t=0$ to $t=100$ sampled at 100Hz. Clearly, while $f_1$, $f_2$ and $f_3$ are oscillatory, the wave-shape functions are all non-trivial and the wave-shape functions of $f_2$ and $f_3$ are time-varying, and $f_2$ and $f_3$ exist for only part of the full time observation time. To further challenge the algorithm, we add a white noise $\xi(t)$ to $g$ by considering $Y(t)=f(t)+\xi(t)$, where for all $t$, $\xi(t)$ is a student t4 random variable with the standard deviation $0.5$. The signal-to-noise (SNR) ratio of $Y(t)$ is 1.8dB, where SNR is defined as $20\log\frac{\text{std}(f)}{\text{std}(\xi)}$ and std means the standard deviation.
The signal $f_1$ and $f_2$ are shown in Figure \ref{fig:Introduction:Example1sig}, and the signal of $f_3(t)$, $g(t)$ and $Y(t)$ are shown in Figure \ref{fig:simulation}. The results of STCT, iSTCT, de-shape STFT and de-shape SST of $g(t)$ and $Y(t)$ are shown in Figure \ref{fig:simulationResult}. The ground truths are superimposed for the comparison.
There are several findings.
Note that even when the signal is clean, we could see several interferences in either STFT or SST. For example, we could see the ``bubbling pattern'' in these TF representations around 2Hz from 0 to 60 seconds (indicated by red arrows), which comes from the interference of the 2nd-multiple of $f_1$ and the fundamental component of $f_2$. These interferences are eliminated in de-shape SST, since the wave-shape is ``decomposed'' in the analysis. Second, when the signal is clean, we could see a ``curve'' starting from about 3.4Hz at 0 second and climbing up to 4Hz at 40s in STFT and SST (indicated by green arrows). Certainly this is not a true component but an artifact, which comes from the incidental appearance of different multiples of different ANH functions. This might mislead us and conclude that there is an extra component. Note that this possible artifact is eliminated in the de-shape SST.
Third, around 85s, the IF's of $f_2$ and $f_3$ cross over (indicated by blue arrows). How to directly decouple signals with this kind of cross-over IF's with TF analysis technique is still an open question.
Last but not the least, while the SNR is low, the de-shape SST could still be able to provide a reasonable IF information regarding the components.
This comes from the robustness of the frequency reassignment vector, which has been discussed in \cite{Chen_Cheng_Wu:2014}.
We mention that we could further stabilize the TF representation determined by the de-shape SST by the currently proposed multi-taper technique called concentration of frequency and time (ConceFT) \cite{Daubechies_Wang_Wu:2016}. We refer the reader with interest to \cite{Daubechies_Wang_Wu:2016} for a detailed discussion of ConceFT.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main1sig.png}\\
\caption{Top: the simulated signal $f_3$; top middle: $f=f_1+f_2+f_3$; top bottom: $\xi(t)$, bottom: $Y=f(t)+\xi(t)$. To enhance visibility, we only show the signal over $[25,65]$.}
\label{fig:simulation}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Main1.png}\\
\includegraphics[width=\columnwidth]{Main1part2.png}\\
\caption{Top left: the STFT of the clean signal $f$, $|V^{(h)}_f(t,\xi)|$; top middle: the SST of $f$, $|SV^{(h)}_f(t,\xi)|$. The colored arrows indicates three findings mentioned in the main context; top right: the STCT of $f$, $|C^{(h)}_f(t,\xi)|$; middle left: the inverse STCT of $f$, $|U^{(h,\gamma)}_f(t,\xi)|$; middle middle: de-shape STFT of $f$, $|W^{(h,\gamma)}_f(t,\xi)|$; middle right: the de-shape SST of $f$, $|SW^{(h,\gamma)}_f(t,\xi)|$;
bottom left: the STFT of the noisy signal $Y$, $|V^{(h)}_Y(t,\xi)|$;
bottom middle: the de-shape SST of the noisy signal $Y$, $|SW^{(h,\gamma)}_Y(t,\xi)|$; bottom right: the de-shape SST of the clean signal $f$, $|SW^{(h,\gamma)}_f(t,\xi)|$, superimposed with $\phi'_1(t)$, $\phi'_2(t)$ and $\phi'_3(t)$ in red.}
\label{fig:simulationResult}
\end{figure}
\subsection{ECG signal}
As discussed in Section \ref{Section:ECGdiscussion}, we need the modified wave-shape function to better capture the features in the ECG signal. We now show that by the de-shape SST, we could obtain a TF representation without the influence of the time-varying wave-shape function. For the ECG signal, we follow the standard median filter technique to remove the baseline wandering \cite{Clifford_Azuaje_McSharry:2006}, and the sliding window is chosen to be 0.1 second.
\subsubsection{Normal ECG signal}
The lead II ECG signal $f(t)$ is recorded from a normal subject for 85 seconds, which is sampled at 1000Hz. The average heart rate of the subject is about 70 times per minute; that is, the IF is about 1.2 Hz. By reading Figure \ref{fig:ECG}, it is clear that the ECG signal is oscillatory with a non-trivial wave-shape function, and the wave-shape function is time varying, as is discussed in Section \ref{Section:ECGdiscussion}.
Figure \ref{fig:ECG} shows the analysis result. We could see a dominant curve in the STCT, which shows the period information of the oscillation and it is about 0.9 second per wave. The iSTCT flips the period information back to frequency information, and hence we see a dominant curve around 1.2 Hz. Eventually, the multiples associated with the ECG wave-form are well eliminated by the de-shape STFT $SW^{(h)}_f$ and the TF representation is sharp.
Thus we conclude that the de-shape SST provides a more faithful TF representation and decouples the IF, AM and the wave-shape function information. Moreover, the dominant curve around 1.2 Hz fits the ground truth instantaneous heart rate (IHR), which indicates the potential of the de-shape SST in the ECG signal analysis.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main3sigv2.png}\\
\includegraphics[width=\columnwidth]{Main3v2.png}\\
\caption{Top: the ECG signal $f$ recorded for 85 seconds from a normal subject.
{The basic landmarks of the ECG signal, P, Q, R, S, and T, and the QT and RR intervals are shown. Note that the QT interval (respectively RR interval) is the length of the time interval between the start of the Q wave and the end of the T wave of one heart beat (respectively two R landmarks of two consecutive heart beats).}
To enhance the visibility, we only show the first {10} seconds. Second row, left panel: $|V^{(h)}_f(t,\xi)|$; middle panel: $|SV^{(h)}_f(t,\xi)|$; right panel: $|SW^{(h,\gamma)}_f(t,\xi)|$. Third row, left panel: $|C^{(h,\gamma)}_f(t,\xi)|$; middle panel: $|U^{(h,\gamma)}_f(t,\xi)|$; right panel: $|SW^{(h,\gamma)}_f(t,\xi)|$ superimposed with the instantaneous heart rate. To enhance the visibility, we show $|C^{(h,\gamma)}_f(t,\xi)|$, $|U^{(h,\gamma)}_f(t,\xi)|$ and $|SW^{(h,\gamma)}_f(t,\xi)|$ only up to 4Hz in the frequency axis.}
\label{fig:ECG}
\end{figure}
\subsubsection{Abdominal fetal ECG}
The fetal ECG could provide critical information for physicians to make clinical decision. While several methods are available to obtain the fetal ECG, the abdominal fetal ECG signal is probably the most convenient and cheap one. We take the abdominal fetal ECG signal with the annotation provided by a group of cardiologists from PhysioNet \cite{Goldberger_Amaral_Glass_Hausdorff_Ivanov_Mark_Mietus_Moody_Peng_Stanley:2000}. In this database, four electrodes are placed around the navel, a reference electrode is placed above the pubic symphysis and a common mode reference electrode is placed on the left leg, which leads to four channels of abdominal ECG signal. The signal is recorded at 1000Hz for 300 seconds. In this example we show the result with the third abdominal ECG signal. Note that while the signal is carefully collected, the signal to noise ratio of the abdominal fetal ECG is relatively low.
We refer the reader with interest to \url{https://www.physionet.org/physiobank/database/adfecgdb/} for more details.
The results of different TF analyses, including de-shape SST, are shown in Figure \ref{fig:fetalECG}. In the STFT and SST, we could see a light curve around 2Hz, which coincides with the fetal IHR we have interest {in}. However, this information is masked by the multiples of the maternal ECG signal. In the de-shape SST, the wave-shape influence is removed and the fetal IHR is better extracted, and the estimated fetal IHR coincides well with the annotation provided by the physician. The curve around 1.5Hz is the IHR associated with the maternal heart beats. The potential of applying de-shape SST to study fetal ECG will be explored and reported in {future works}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main10sig.png}\\
\includegraphics[width=\columnwidth]{Main10.png}\\
\caption{Top: the abdominal fetal ECG signal $f$ recorded for 300 seconds from a recorded in labor, between 38 to 41 weeks gestation. To enhance the visibility, we only show the signal of 30 seconds long. The annotation of the fetal heart beats are marked in red. Middle left panel: $|V^{(h)}_f(t,\xi)|$; middle right panel: $|SV^{(h)}_f(t,\xi)|$; bottom left panel: $|SW^{(h,\gamma)}_f(t,\xi)|$; bottom right panel: $|SW^{(h,\gamma)}_f(t,\xi)|$ superimposed with the fetal instantaneous heart rate (IHR) determined by a group of cardiologists. {To enhance the visibility, we show $|SW^{(h,\gamma)}_f(t,\xi)|$} only up to 4Hz in the frequency axis. The curve around 1.5Hz is the IHR associated with the maternal heart beats.}
\label{fig:fetalECG}
\end{figure}
\subsection{PPG signal}
Pulse waves represent the hemodynamics, and it can be monitored via plethysmographic technologies in different regions of the body. These technologies often use photo sensors usually placed on the earlobe or finger, by illuminating the tissue and simultaneously measuring the transmitted or the reflected light using a specific wavelength. More recently, noncontact techniques such as video signals (e.g., PhysioCam \cite{Davila:2012Thesis}) have been used to monitor the pulse wave from the face at a distance. Collectively, the application of photosensors to monitor pulse wave are known as photoplethysmography (PPG). See, for example, \cite{Davila:2012Thesis} for a review of the PPG technique. In addition to acquire the hemodynamical information, it also contains the respiration information. Indeed, mechanically, inspiration leads to a reduction in tissue blood volume, which leads to a lower amplitude of the PPG signal. Since none of the pulse wave or the respiration-induce variation oscillates like a sinusoidal wave, the signal should be modeled by the ANH model.
\subsubsection{PPG signal with respiration}
Figure \ref{fig:PPG2comp} shows a PPG signal from the Capnobase dataset\footnote{\url{http://www.capnobase.org}} and its analysis result with the de-shape STFT. The PPG signal, the capnogram signal and the ECG signal are simulateneously recorded from a subject without any motion at 300 Hz for 480 seconds. By a visual inspection, it is clear that there are two oscillations inside the PPG signal -- the faster (respectively slower) oscillations are associated with the heartbeat (respectively respiration). Clearly, the non-sinusoidal oscillatory waves complicate STFT $V^{(h)}_f$ and SST $SV^{(h)}_f$, while these multiples are elliminated in the de-shape STFT and de-shape SST. Also, we could see that the estimated IHR and instantaneous respiratory rate (IRR) estimated from the PPG signal fit the IHR and IRR derived directly from the ECG signal and the capnogram signal. This indicates the potential of simultaneously obtaining IHR and IRR from the PPG signal.
We mention that when $\gamma$ is chosen to be $2$, the heartbeat component is missed (the result is not shown). This coincides with the general knowledge that $\gamma=2$ is not a good periodicity detector when there exists multiple periodicity in the signal.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main7sig.png}\\
\includegraphics[width=\columnwidth]{Main7.png}\\
\caption{Top row: the photoplethysmography signal $f$ recorded from a normal subject for 480 seconds. To enhance the visibility, we only show the segment between the 150-th second and 250-th second.
Second row: left: $|V^{(h)}_f(t,\xi)|$; middle: $|SV^{(h)}_f(t,\xi)|$; right: the de-shape SST of the capnogram signal superimposed with the instantaneous respiratory rate (IRR) in red, which is estimated from the PPG signal. Here, only up to 2Hz in the frequency axis is shown to enhance the visibility.
Bottom row: left: $|SW^{(h,\gamma)}_f(t,\xi)|$. To enhance the visibility, we show $|SW^{(h,\gamma)}_f(t,\xi)|$ only up to 4Hz in the frequency axis. Clearly, the multiples of each component are eliminated; middle: $|SW^{(h,\gamma)}_f(t,\xi)|$ superimposed with the estimated instantaneous heart rate (IHR) and IRR. The red curve around 0.3Hz is associated with the IRR and the red curve around 1.6 Hz is associated with the IHR. {The blue curve is five times the IRR curve}, which indicates that the component with the higher frequency is not a multiple of the component with lower frequency; right: the de-shape SST of the electrocardiographic signal superimposed with the IHR in red, which is estimated from the PPG signal. Here, only up to 2Hz in the frequency axis is shown to enhance the visibility.
}
\label{fig:PPG2comp}
\end{figure}
\subsubsection{PPG signal with motion}
Figure \ref{fig:MotionPPG} shows the result of one PPG sample used in the training dataset of ICASSP 2015 signal processing cup\footnote{\url{http://www.zhilinzhang.com/spcup2015/}}. The sample is a 5-minute PPG signal sampled at 125Hz when the subject runs with changing speeds, scheduled as: rest (30s) $\to$ 8km/h (1min) $\to$ 15km/h (1min) $\to$ 8km/h (1min) $\to$ 15km/h (1min) $\to$ rest (30s). {From the recorded signal} it is not easy to see how the motion and heartbeat vary. The heartbeat component starts from around 1.7 Hz at 50 seconds, to 2.2 Hz from 150 to 170 seconds, when the subject has just finished the 15km/h running section. Then, the heartbeat goes lower in the 8km/h section and higher in the final 15km/h section.
Note that the IF of the heartbeats (marked by the red arrow) lies between two other components, supposedly contributed by motion. The higher frequency component associated with motion has IF about twice the IF of the lower one.
We conjecture that the higher one is contributed by the movement of body while the lower is contributed by the movement of arms and legs. The body finishes a period by just one step, while the leg finished a period by two steps (one leg needs to finish a forward and backward movement). This is very similar to the ``octave'' detection problem in music signal processing (see Section \ref{Section:Numerical:ViolinSonata}) and it is quite natural to catch two {components} here as they are indeed (at least) two different oscillatory signals, where the one has IF almost twice from the other one. An extensive study of this signal is needed to fully understand how the body motion influences the physiological signal and will be reported in {a} future work.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main8sig.png}\\
\includegraphics[width=0.9\columnwidth]{Main8sigZoom.png}\\
\includegraphics[width=\columnwidth]{Main8.png}\\
\caption{Top row: the photoplethysmography signal $f$ recorded from a normal subject, who is scheduled to run at different speeds. Second row: the first 100 seconds photoplethysmography signal $f$. It is clear that the signal is composed of several components with complicated dynamics. Third row: $|V^{(h)}_f(t,\xi)|$ is shown on the left and $|SV^{(h)}_f(t,\xi)|$ is shown on the right. Bottom row: $|U^{(h,\gamma)}_f(t,\xi)|$ is shown on the left and $|SW^{(h,\gamma)}_f(t,\xi)|$ is shown on the right. The heartbeat component is marked by red arrows. To enhance the visibility, we shown $|U^{(h,\gamma)}_f(t,\xi)|$ and $|SW^{(h,\gamma)}_f(t,\xi)|$ only up to 6Hz in the frequency axis.}
\label{fig:MotionPPG}
\end{figure}
\subsubsection{Non-contact PPG signal}
Figure \ref{fig:NonContactPPG3} shows the non-contact PPG signal recorded from a normal subject when he is walking on the treadmill at 0.6 Hz. The sampling rate is 100Hz. The non-contact PPG is collected with the PhysioCam technology, and we refer the reader with interest to \cite{Davila:2012Thesis} for details. The ECG signal is simultaneously recorded from the subject at the sampling rate 1000Hz, so we have the true IHR for comparison. Clearly the signal is noisy and contains the walking rhythm; that is, the non-contact PPG signal is composed of two oscillatory signals -- one is associated with the hemodynamics and one is associated with the walking rhythm. Despite the heavy corruption terms in the low frequency, which comes from the ``trend'' inside the signal, we could see that the de-shape STFT successfully extracts the walking rhythm around 0.6 Hz and the IF around 2Hz, which coincides with the IHR determined from the ECG signal. A systematic study of this kind of signal, including the associated de-trend technique, is critical for practical applications and will be reported in {a} future work.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main9sig.png}\\
\includegraphics[width=\columnwidth]{Main9.png}\\
\caption{Top: the non-contact PPG signal $f$ recorded from a normal subject walking on the treadmill at a fixed speed. Middle left: $|V^{(h)}_f(t,\xi)|$; middle right: $|SV^{(h,\gamma)}_f(t,\xi)|$; bottom left: $|SW^{(h,\gamma)}_f(t,\xi)|$; bottom right: $|SW^{(h,\gamma)}_f(t,\xi)|$ superimposed with the instantaneous heart rate. To enhance the visibility, we show $|U^{(h,\gamma)}_f(t,\xi)|$ and $|SW^{(h,\gamma)}_f(t,\xi)|$ only up to 5Hz in the frequency axis.}
\label{fig:NonContactPPG3}
\end{figure}
\subsection{Music and bioacoustic sounds}
The idea of de-shape STFT has been applied in the task called {\em automatic music transcription} (AMT) \cite{peeters2006music,emiya2007parametric,su2015combining,su2016exploiting}, and this approach has been shown competitive in comparison to the state-of-the-art AMT methods in the MIREX-MF0 challenge, an annual competition in the field of music information retrieval (MIR).\footnote{\url{http://www.music-ir.org/mirex/wiki/MIREX_HOME}} AMT is still a technology under active development by now, where one big challenge is how to correctly identify the pitches of the notes played at the same time. In this subsection, we show the potential of applying the de-shape SST to the AMT problem.
\subsubsection{Violin sonata}\label{Section:Numerical:ViolinSonata}
Figure \ref{fig:Mozart} shows a 6-second segment from {\em Mozart's Violin Sonata in E minor, K.304}, where the annotations are provided by musicians. The sampling rate of the signal is 44.1 kHz. This segment contains the sounds of two instruments, violin (melody) and piano (accompaniment). The number of concurrent pitches of this signal at every timestamp varies from 1 to 4, where the violin is played in single pitch and the piano in multiple pitches. The patterns of the two instruments are different, which can be seen from reading the TF representations of STFT and SST. The violin sound exhibits a clear {\em vibrato} (i.e., periodic variation of the IF) together with a strong and frequency-dependent AM effect. See the red arrows in Figure \ref{fig:Mozart} for an example. It is to say that the spectral envelope of the sound varies strongly during one cycle of vibrato \cite{fletcher2010physics}. On the other hand, piano notes have stable IF's, strong attack and long decay of AMs, and, as mentioned in Section \ref{Section:StiffStringModel}, the {\em inharmonicity} makes the high-order harmonic peaks deviate from the integral multiple of the fundamental frequency $f_1$.
The notes of this segment are with pitches ranging from E2 (the fundamental frequency is $82.4$ Hz) to G5 (the fundamental frequency is $784.0$ Hz), and they are shown in the red lines in Figure \ref{fig:Mozart}. The resolution of the labels formatted in Musical Instrument Digital Interface (MIDI) is one semitone.
We indicate one specific tricky problem commonly encountered in this kind of signal. Take the signal from 0.76 to 1.14 seconds as an example. The highest note of piano, B3 (the fundamental frequency is $246.9$ Hz), is just one half of the violin note, B4 (the fundamental frequency is $493.88$ Hz). It is to say, all multiples of violin note are (nearly) overlapped with the piano note, thereby violates the frequency separation condition in Definition \ref{DefBClassMultipleTimeSeries}. The problem of detecting these ``overlaps'' is commonly understood as the {\em octave detection} \cite{su2014resolving}. A systematic study of this specific problem is out of the scope of this paper, and it will be discussed in {a} future work.
From the result of the de-shape SST, we see that the multiples are distinguished from the IF's and are eliminated. All the notes of both violin and piano are well captured. For violin we can even obtain the vibrato rate and vibrato depth of the notes, which are not recorded in the MIDI ground truth. We could also see that the octave problem mentioned above is well resolved.
However, we can still see some false detections in the ``inner part'' of the music. For example, there is a component appearing at around 330 Hz from 1.46 to 1.8 seconds, but there is no note played here. To explain this, notice that the fake component has frequency twice of a piano note while at the same time one half of the violin note. This causes an issue called the {\em stacked harmonics} ambiguity, which is caused by double or even more octave ambiguities. This open problem has also been raised in \cite{su2015combining,su2016exploiting}. Again, a systematic study of this specific problem is out of the scope of this paper, and it will be discussed in future {works}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main4sig.png}\\
\includegraphics[width=\columnwidth]{Main4.png}\\
\caption{Top: the Mozart violin sonata signal $f$, which is zoomed in to the period from 4.2 second to 4.7 second to enhance the visualization. Since there are several oscillatory components with complicated wave-shape function, it is not clear what information is hidden inside, even it is hard to identify oscillations. Middle left: $|V^{(h)}_f(t,\xi)|$; middle right: $|SV^{(h)}_f(t,\xi)|$; bottom left: $|SW^{(h,\gamma)}_f(t,\xi)|$, where the red arrows indicate the violin sound with vibrato; bottom right: $|SW^{(h,\gamma)}_f(t,\xi)|$ with the annotations superimposed in red. To enhance the visibility, we show $|SW^{(h,\gamma)}_f(t,\xi)|$ only up to 1000Hz in the frequency axis.}
\label{fig:Mozart}
\end{figure}
\subsubsection{Choir}
Figure \ref{fig:Choir} shows the analysis result of a recorded choir music with the annotation provided by experts. Similar to the above example, the choir music also has multiple components and usually in consonant intervals. Moreover, in the choir music, every perceived individual note is typically sung {\em in unison} by more than one performer. However, since there is always some small and independent variation of the IF among performers, the resulting sound would have wider mainlobe in the STFT than the other music sung by a single performer. Such a phenomenon, called {\em pitch scattering} \cite{ternstrom1993perceptual}, usually appears in choir and symphony music, as a challenge in correctly estimate the pitch of every note.
This example is a 3-part choir (first soprano, second soprano and alto), with pitches ranging from B3 (the fundamental frequency is $246.9$ Hz) to E5 (the fundamental frequency is $659.3$ Hz). We could see in Figure \ref{fig:Choir} that the pitch scattering issue can be partially addressed by the SST. However, we can still find some intertwined components, like the component at around 920 Hz from 2.2 to 3.5 seconds, which might be contributed by more than two notes with different vibrato behaviors. By using the de-shaped SST, this wide-spread terms are correctly identified as the multiples and removed. All labeled notes are captured and there are few false alarm terms.
Although we have shown the usefulness of de-shape SST in both physiological and musical signals, we need to emphasize some differences between them. In comparison to physiological data, musical signals can have a much larger number of components (e.g., more than 10 components in a symphony), which complicate the patterns of the multiples. Besides, most of the musical works are composed following the {\em theory of harmony}, which holds a principle that a sound is {\em consonant} when the ratio of the IF's are in simple ratios. This implies that the spectra of the components are highly overlapped. Moreover, the octave is very often seen in music composition. Therefore, musical signals usually violate Definition \ref{DefBClassMultipleTimeSeries} and make the problem of AMT ill-posed. To reduce the ambiguities of octaves and other consonant intervals, we may impose more strict constraints when we analyze the signal, like the constraint of harmonicity discussed in Section \ref{Section:deShape}.
For more information of this approach in AMT, readers could refer to \cite{su2015combining,su2016exploiting}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main6sig.png}\\
\includegraphics[width=\columnwidth]{Main6.png}\\
\caption{Top: the choir signal $f$, which is zoomed in to the period from 4.2 second to 4.7 second to enhance the visualization. Middle left: $|V^{(h)}_f(t,\xi)|$; middle right: $|SV^{(h)}_f(t,\xi)|$; bottom left: $|SW^{(h,\gamma)}_f(t,\xi)|$; bottom right: $|SW^{(h,\gamma)}_f(t,\xi)|$ with the annotations superimposed in red. To enhance the visibility, we show $|SW^{(h,\gamma)}_f(t,\xi)|$ only up to 800Hz in the frequency axis.}
\label{fig:Choir}
\end{figure}
\subsubsection{Wolf howling}
An important topic in conservation biology is monitoring the number of wolves in the field \cite{Passilongo_Mattioli_Bassi_Szabo_Apollonio:2015}. Analyzing the wolf howling signal is an efficient approach to evaluate how many wolves are there in the field under survey. In this final example we show the analysis result with a field signal recorded The sound is downloaded from Wolf Park website\footnote{\url{http://www.wolfpark.org/Images/Resources/Howls/Chorus_1.wav}}. The signal is sampled at 11.025 kHz for 40 seconds. In Figure \ref{fig:Wolf} we could directly see that while TF representations provided by STFT and SST are complicated by the multiples caused by the non-trivial wave-shape, the TF representation provided by de-shape SST contains only the fundamental components. By reading the de-shape SST, we could suggest that there are at least three wolves in the field, since during the recording period, there are at most three dominant curves at a fixed time. However, the ground truth for this database is not provided, and identifying each single wolf needs field experts, so this conclusion is not confirmed, and a further collaborative exploration with biologists is needed. To sum up, this suggests that the de-shape SST {has} potential to provide {an} audio visualization for this kind of application.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Main5sig.png}\\
\includegraphics[width=\columnwidth]{Main5.png}\\
\caption{Top: the wolf howling signal $f$. Since each component inside the signal oscillates at the frequency at least 400Hz, the oscillation could not be visualized except the overall amplitude modulation. Middle left: $|V^{(h)}_f(t,\xi)|$; middle right: $|SV^{(h)}_f(t,\xi)|$; bottom left: $|SW^{(h,\gamma)}_f(t,\xi)|$; bottom right: the zoom-in $|SW^{(h,\gamma)}_f(t,\xi)|$ only up to 1000Hz in the frequency axis. The three red arrows at around 18 seconds indicate that there are at least three wolves.}
\label{fig:Wolf}
\end{figure}
\section{Numerical issues}\label{Section:NumericalIssues}
While the numerical implementation of STCT, iSTCT, de-shape STFT and de-shape SST are straightforward, we should pay an attention to evaluate {iSTCT}. In particular, the map from $C_f^{(h,\gamma)}(t,\eta)$ to $U_f^{(h,\gamma)}(t,\xi)$ depends on the inverse map $\mathcal{I}$, which is numerically unstable. To stabilize it, there are two critical process: (1) long-pass lifter; (2) discretize $\mathcal{I}$ by a suitable weighting, for example, by the Jacobian of $\mathcal{I}$, so that the iSTCT is defined on the uniform frequency grid. Let the sampling frequency of the signal $f(t)$ be $\zeta>0$ and we sample $N\in\mathbb{N}$ points from $f$. Then, for the $N$-point STFT, the frequency axis is discretized into $\eta_n=n\zeta/N$, where $n = 0,1,\ldots,N-1$, $\eta_n$ is the $n$-th index in the frequency axis, and the frequency resolution is $\Delta\zeta:=\zeta/N$. Similarly, the quefrency axis in STCT is discretized into $q_n=n/\zeta$, where $n=0,1,\ldots, N-1$, $q_n$ is the $n$-th index in the quefrency axis and the quefrency resolution is $\Delta q:=1/\zeta$. We discretize the frequency axis of iSTCT in the same way as that of STFT; that is, the frequency axis of iSTCT is discretized into $\eta_n$, where $n = 0,1,\ldots,N-1$, and the frequency resolution is $\Delta\zeta$.
To implement the long-pass lifter mentioned in Section \ref{Section:deShape}, we consider a simple but effective hard threshold approach by choosing a cutoff quefrency $q_c$, where $c\in\mathbb{N}$ is chosen by the user; that is, all entries with index less than $c$ are set to zero and the other entries are not changed. While it depends on the characteristic of the signal, in practice we suggest to choose the cutoff quefrency in the range of $10\leq c\leq 20$ and numerically it performs well.
One main issue of the mapping $\mathcal{I}$ is that it maps uniform grid to a non-uniform grid and hence there are insufficient low-quefrency elements in $C_f^{(h,\gamma)}(t,\mathcal{I}\eta)$, which could be directed implemented by inverting the quefrency axis index of $C_f^{(h,\gamma)}(t,\eta)$, to represent the high-frequency content in $U_f^{(h,\gamma)}(t,\xi)$. For example, we have only about $\lfloor 0.1/\Delta q\rfloor=\lfloor 0.1\zeta\rfloor$ entries on the quefrency interval $[0.1,0.2]$ in $C_f^{(h,\gamma)}(t,\eta)$, while we have $\lfloor\frac{5N}{\zeta}\rfloor$ entries on the frequency interval $[5,10]$ in $U_f^{(h,\gamma)}(t,\xi)$. On the other hand, there are too many high-quefrency elements to represent the low-frequency content. Therefore, we suggest to do interpolation over the {quefrency} axis in the STCT to alleviate this issue. Denote the finer grid in the quefrency axis as $\tilde{q}_j$, $j=1,\ldots,M$ and $M>N$. Further, if we want to preserve the integrability of the function after the mapping $\mathcal{I}$, we should weight the entries by the Jacobian of $\mathcal{I}$. To sum up, after obtaining $C_f^{(h,\gamma)}$ with a finer resolution in the {quefrency} axis, the elements in $C_f^{(h,\gamma)}$ are weighted and summed up to the closest frequency bin corresponding to it; that is, we implement iSTCT by
\begin{equation}
U_f^{(h,\gamma)}(t,\eta_n)=\sum_{j \in \mathfrak{P}(\eta_n)}C^{(h,\gamma)}_f\left(t, \tilde{q}_j\right)\tilde{q}_j,
\end{equation}
where $\mathfrak{P}(\eta_n):=\{j: 1/(\eta_n+0.5\Delta\zeta)<\tilde{q}_j\leq 1/(\eta_n-0.5\Delta\zeta)\}$ for each $n=0,1,\ldots,N-1$.
\section{Conclusions}\label{Section:Conclusions}
To handle oscillatory signals in the real world, we provide a model capturing oscillatory features, including IF, AM and time-varying wave-shape function. To alleviate the limitation of TF analysis caused by the existence of non-trivial wave-shape function, we consider the idea of cepstrum and introduce the STCT, de-shape STFT and de-shape SST. A theoretical proof is provided to study how STCT works. When {the STCT and its theoretical proof is} combined with the previous study of SST, we have a theoretical understanding of the efficiency of de-shape SST. In addition to the simulated signal, several real datasets are studied and confirm the potential of the proposed algorithms. The proposed algorithm could be easily combined with several other algorithms to study a given database. For example, we could apply ConceFT \cite{Daubechies_Wang_Wu:2016} to stabilize the influence of the noise, the RM technique \cite{Auger_Flandrin:1995} could be applied to further sharpen the TF representation if causality is not an issue, we could apply the adaptive local iterative filtering \cite{Cicone_Liu_Zhou:2014} to reconstruct each oscillatory component, we could consider the template fitting scheme by designing a good dictionary based on the available information from the de-shape SST \cite{Hou_Shi:2013a}, to name but a few.
{However,} there are several problems left unanswered in this paper. We summarize them below.
To facilitate the discussion, we could call the sequence $\{B_{k,\ell}(t)\}_{\ell=-N_k}^{N_k}$ in (\ref{MainTheorem:STFTExpansion}) the {\em spectral envelope} of the $k$-th ANH model. The assumption in Theorem \ref{Theorem:TimeVaryingCepstrum} says that the spectral envelope of an ANH function should be ``far away'' from $0$.
In the ideal case, we would expect that the spectral envelope is ``slow-varying'' in comparison to the harmonic series in the spectrum, so that the cepstrum can well extract the periodicity-related elements from the filter-like elements. This ideal case is satisfied by the assumption in Theorem \ref{Theorem:TimeVaryingCepstrum} in the sense that the IP information is recovered in the STCT. However, this is not always true for real-world signals; in some challenging cases we could see non-trivial patterns in the spectral envelope, which breaks the assumptions in Theorem \ref{Theorem:TimeVaryingCepstrum}. This contaminates the information associated with the IP information we have interest {in}, and hence causes fake detection of periodicity.
Here we discuss two real scenarios when the spectral envelope has a non-trivial pattern.
The first scenario could be observed in the ECG signal with the fundamental frequency around $\xi_1>0$. For example, in some cases, we could find relatively stronger peaks around $3\xi_1$, $6\xi_1$ and $9\xi_1$ in comparison to other peaks in the spectrum. Therefore, in the cepstrum we can find not only a prominent peak at $q_1=1/\xi_1$ but also a small bump around $q_1/3$. To take a closer look at this phenomenon, we recall that it has been well known that the 12-lead ECG signals, denoted as $E(t)\in \mathbb{R}^{12}$, are the projection of the representative dipole current, denoted as $d(t)\in\mathbb{R}^3$, where $t\in\mathbb{R}$, of the electrophysiological cardiac activity on different directions. Physiologically, for a normal subject $d(t)$ is oscillatory with the period about $1$ second. If we could record $d(t)$, the recorded signal is called the vectocardiogram signal. For the $\ell$-th ECG channel, where $\ell=1,\ldots,12$, there is an associated projection direction $v_\ell\in \mathbb{R}^3$. The $\ell$-th ECG channel is thus the projection of $d(t)$ on $v_\ell$; that is, $E_\ell(t)=v_\ell^Td(t)$ or $E(t)=v^Td(t)$, where $v=[v_1 v_2 \ldots v_{12}]\in \mathbb{R}^{3\times 12}$. In general, $v$ changes according to time due to the cardiac axis deviation caused by the respiratory activity and other physical movements. To simplify the discussion, we ignore this facts. Thus, since $d(t)$ is oscillatory, it is clear that $E_\ell(t)$ is also oscillatory. In some cases, this complicated procedure leads to an {oscillation} in the spectral envelop, and hence the first scenario.
{The second example is the sound of clarinets.}
Clarinet is one kind of woodwind instrument which makes air resonating in a cylindrical tube with one ended closed. Because of such a physical structure, the even-numbered harmonics including $2\xi_1$ and $4\xi_1$ are highly suppressed\footnote{The absence of even harmonics is (part of) what is responsible for the ``warm'' or ``dark'' sound of a clarinet compared to the ``bright'' sound of a saxophone.}, which breaks the assumption of Corollary \ref{Corollary:TimeVaryingCepstrum}{, and is discussed after Corollary \ref{Corollary:TimeVaryingCepstrum}}. {But, in many real cases,} the cepstrum of the clarinet note do have a peak at {$q_1$} and its multiples {because the even-number harmonics are not totally eliminated}.
In several real examples, including the {Clarinet and ECG} examples, the unwanted terms in the above situations can be simply eliminated by hard-thresholding{; however,} it is not that easy {to achieve this naive idea}, and a systematic study of this challenge is needed, where we might incorporate more background knowledge into the analysis.
Next, we discuss another scenario when the proposed method works on multi-component signals. Consider {an {\em octave} signal mentioned in Section \ref{Section:Numerical:ViolinSonata}, where one of the two components has the fundamental frequency $\xi_{1,1}$ and the one of the other is higher than it by one octave, thereby with the fundamental frequency $2\xi_{1,1}$. Suppose the phases of these two components match in a way so that} the spectrum of the multi-component signal has stronger peaks at even-order harmonics, especially {$\xi_{1,2}=2\xi_{1,1}$, $\xi_{1,4}=4\xi_{1,1}$, $\xi_{1,6}=6\xi_{1,1}$, $\xi_{1,8}=8\xi_{1,1}$, etc. In this special case,} the spectral envelope oscillates and we may recall the IF's of both components in the de-shape SST. { Note that this special case contradicts the assumption of the ANH model so that we could not model it as a composition of two ANH functions, and the proposed method may or may not work. In the real-world music, the condition of phase matching does not always happen, and it makes the octave detection problem even harder, as is discussed in Section \ref{Section:Numerical:ViolinSonata}. We mention that while it is a difficult job in signal processing, human beings could identify the difference via learning the oscillatory pattern of the signal or the ``timbre''. For example, the timbre of the note C4, which has the fundamental frequency 262 Hz, is different from that of the combined note (or called ``an interval'' in music) C4+C5, where the fundamental frequency of C5 is 524 Hz. By learning the timbre, we could tell the difference.} The above scenarios all have their own interest but {are} out of the scope of this paper. We will report a systematic study in {a} future work.
We mention that there are several challenging cases in processing real-world multi-pitch signals, like {\em missing fundamental} or {\em stacked harmonics}, both of which have been discussed in \cite{su2015combining,su2016exploiting}. These could be treated as exceptional cases of the proposed model and a modification of the model and algorithm is needed to better handle these signals.
{Last but not the least, from the data analysis viewpoint, in general we cannot decide the model parameters, like the $\epsilon$, and the sequence $c$, a priori. This is an estimation problem in nature, and has been open for a while. However, for most problems we face in practice, we have some background knowledge that could guide us to ``guess the model''. For example, for the fetal ECG extraction problem, the heart rates of the mother and the fetus have a well-known range guided by the physiological background, and this is the information we could use to determine the parameters. But for a randomly given dataset without any background knowledge, at this moment, there is still no ideal way to determine the model parameters directly from the data itself. This fundamental estimation problem will be explored in a future work.}
|
\section{Introduction}
\section{Introduction}
Modern high-speed wireless networks heavily depend on reliable and efficient transmission of large data packets through the use of coding and information theory. The advent of machine-to-machine (M2M), vehicular-to-vehicular (V2V), and various streaming systems have spawned a renewed interest in developing information theoretical bounds and codes for communication of short packets \cite{Popovski2014}\cite{Durisi2015}\cite{FiveDisruptive}. Additionally, these applications often have tight reliability and latency constraints compared to typical wireless systems today. Communication at shorter blocklengths introduces several new challenges which are not present when considering communication of larger data packets. For example, the overhead caused by control signals and header data is insignificant if large data packets are sent, and hence, this overhead is often neglected in the analysis of protocols. However, more stringent latency requirements lead to shortened blocklengths for transmission such that
the size of control information may approach, or even exceed, the size of the data part in the packet. This is especially true for multiuser systems such as broadcast channels, two-way channels, or multiple access channels, where the control information must include information about the packet structure, security, and user address information for identification purposes.
The fundamentals of communication of short packets have been addressed by Strassen and, recently, Polyanskiy \emph{et al.} in \cite{Strassen2009} and \cite{Polyanskiy2010b}. It was shown that the maximum coding rate of a fixed-length code with $n$ channel uses and maximum error probability $\varepsilon$ over a discrete-time AWGN point-to-point channel has an asymptotic expansion given by
\begin{IEEEeqnarray}{rCl}
R^*(n,\varepsilon) = C - \sqrt{\frac{V}{n}}Q^{-1}(\varepsilon) + \frac{1}{2 n}\log_2 n +\mathcal{O}\left(\frac{1}{n}\right)\label{eq:fbl}
\end{IEEEeqnarray}
as $n\rightarrow \infty$.
Here, $C$ is the Shannon capacity, $V$ is the channel dispersion, and $Q^{-1}(\cdot)$ denotes the inverse $Q$-function. In addition to the asymptotic expansion in \eqref{eq:fbl}, \cite{Polyanskiy2010b} used nonasymptotic bounds to numerically demonstrate that $R^*(n,\varepsilon)$ is tightly approximated by the first three terms of \eqref{eq:fbl}.
The approximation \eqref{eq:fbl} and similar ones are important in the design of communication systems because the specifics of code selection can be neglected in the optimization of protocol parameters. For example, such approximations have been applied in the optimization of packet scheduling problems \cite{Xu2016}, hybrid ARQ protocols \cite{Makki2014}, and cloud radio access networks \cite{Khalili2015}.
In this paper, we consider downlink transmission with a discrete-time AWGN broadcast channel that consists of a transmitter and $K$ users. Downlink transmissions are organized in frames, whose structure is the main topic of this paper. In each frame, there is a message from the transmitter to the $k$-th user with a certain probability $1-q$. If, in a given frame, there is a message for user $k$, then this user is said to be \emph{active} in that frame. The size of the message to user $k$ is denoted by $D_k$ and is a random variable itself. Hence, the transmitter needs to convey information about which users are active, the structure of the transmission, and sizes of the messages. As a result, the \emph{frame duration}, which corresponds to the total transmission time, and the total power consumption at the users are also random variables. An important observation from \eqref{eq:fbl} is that larger data packets are encoded more efficiently. This introduces an interesting trade-off with two extremes: (a) in a broadcast setting one can either encode all messages in one large packet, or (b) one can encode each message separately, which is the norm in modern wireless protocols. In (a), the average frame duration is minimized, which implies that the average \emph{latency} across the users is minimized. However, the downside of (a) is that all users need to receive for the whole period of transmission to be able to decode their messages, which is undesirable for devices that are power-constrained. The latter approach (b), depicted in Fig.~\ref{fig:protocol_conventional}, uses codes which are less efficient, and thus the average frame duration is larger. On the other hand, each user only needs to decode the information intended for that user. The key point, however, is that these design considerations enlarge the space of feasible protocols and enable the protocol designer to seek a trade-off between frame duration (latency) and power consumption at the users. Despite this trade-off, practically all wireless systems solely use the extreme approach (b).
\subsubsection*{Contribution}
The purpose of this paper is to revisit the way a downlink frame is designed when it contains short packets. Specifically, it aims at exploring the trade-off between the average frame duration and the average power consumption at the users.
Instead of using a traditional frame structure, we enlarge the design space for a frame by doing the following: the users are divided into groups that may depend on the realization of the message sizes and the messages of each group are jointly encoded using optimal channel codes. We analyze the problem using asymptotic expansions similar to \eqref{eq:fbl}, and we find a lower bound for the trade-off curve. Next, we introduce three protocols: (a) a genie-aided protocol with performance close to the lower bound, (b) protocol with a fixed message that works for the case in which each message has either the size $0$ or $\alpha\in\mathbb{N}$ bits, and (c) a protocol with variable message sizes, where the message sizes are distributed according to a probability mass function $P_D$ with finite and nonnegative integer support. The protocols (b) and (c) both convey enough control information to make them practically usable. Our numerical results demonstrate trade-offs which are particularly interesting when the message sizes are small.
\subsubsection*{Organization} Section~\ref{sec:fbl} introduces the finite blocklength approximations and bounds for optimal channel codes while the system model is introduced in Section~\ref{sec:system_model}. Section~\ref{sec:lower_bnd} presents a lower bound for the average power at each user expressed as a function of the average frame duration. Section~\ref{sec:protocol_design} provides some concrete protocol designs, which are subsequently compared with the lower bound. Finally, numerical examples are presented in Section~\ref{sec:numerical} and Section~\ref{sec:conclusions} concludes the paper.
\subsubsection*{Notation}
Vectors are denoted by boldface letters (e.g., $\vect{x}$) while their entries are denoted by roman letters (e.g., $x_i$). We denote the $n$-dimensional all-zero vector and all-one vector by $\vect{0}_n$ and $\vect{1}_n$, respectively. We denote by $\vect{\bar 0}^n_i(x)$ the $n$-dimensional vector with $x$ in the $i$-th entry and zeroes in the rest.
We let $\oplus$ denote the concatenation of two bit string, e.g., for $\vect{a} \in\{0,1\}^n$ and $\vect{b}\in\{0,1\}^m$, $\vect{a}\oplus \vect{b}$ is the concatenated bit string.
Throughout the paper, the index $k$ belongs always to the set $\mathcal{K}\triangleq \{1,\cdots,K\}$, although this is sometimes not explicitly mentioned. We define the upper concave envelope of a function $f: \mathbb{R}_+ \mapsto \mathbb{R}_+$ as $\text{uce}(f) \triangleq \inf_g\{g \geq f \text{ and } g \text{ is concave}\}$. Similarly, the lower convex envelope is defined by $\text{lce}(f) \triangleq \sup_g\{g \leq f \text{ and } g \text{ is convex}\}$. Finally, $\mathbb{N}$ denotes the set of positive integers, $\mathbb{Z}_+ \triangleq \mathbb{N}\cup \{0\}$, and the symbol $\mathbb{R}$ indicate the set of real numbers.
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{protocol_conventional.pdf}
\caption{Conventional approach to downlink broadcasting. We denote by $\mathrm{M}_2,\mathrm{M}_6,\cdots,\mathrm{M}_{20}$ the messages of varying size (in bits) destined to the active users. An initial packet contains control information that defines the structure of the remaining part of the transmission. Each message is encoded separately.}
\label{fig:protocol_conventional}
\end{figure}
\section{Finite Blocklength Bounds and Approximations}
\label{sec:fbl}
In our analysis, we apply results from finite blocklength information theory.
For the (real) AWGN channel under a short-term power constraint $P$, \cite{Strassen2009} and \cite{Polyanskiy2010b} showed that the maximum coding rate $R^*(n,\varepsilon)$ of a code with fixed blocklength $n$ and error probability $\varepsilon\in(0,1)$ has the asymptotic expansion given by \eqref{eq:fbl}, where the channel capacity $C$ and the channel dispersion $V$ are given by
\begin{IEEEeqnarray}{rCl}
C &\triangleq& \frac{1}{2} \log_2(1+P)\label{eq:capacity}
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}
V &\triangleq& \frac{P(P+2)}{2(P+1)^2} \log_2(\exp(1))^2\label{eq:dispersion}
\end{IEEEeqnarray}
respectively. One can obtain tight nonasymptotic upper and lower bounds for $R^*(n,\varepsilon)$ using the achievability and converse bounds in \cite{Polyanskiy2010b}, and it was numerically demonstrated that the first three terms of the right-hand side of \eqref{eq:fbl} provide a tight approximation of $R^*(n,\varepsilon)$.
\begin{figure}[!t]
\centering
\begin{tikzpicture}[thick,scale=0.95, every node/.style={scale=0.95}]
\begin{axis}[
grid=both,
tick align=center,
width=9.1cm,
height=7.1cm,
minor tick num=4,
every minor grid/.style={opacity=0.2},
xmin=0,
xtick={0,500,1000,1500,2000},
xticklabels={{ 0},{ 500},{1000}, {1500},{2000}},
xmax=2000,
ymin=1.2,
ymax=1.8,
legend entries ={ {}{$N_a(k,\varepsilon)/k$},{}{$\bar N(k,\varepsilon)/k$},{}{$N(k,\varepsilon)/k$},{}{$N_c(k,\varepsilon)/k$},{}{$1/C$}},
xlabel={Information bits $k$},
ylabel={Ch. uses/information bit}
]
\addplot [
name path=ach,
color=magenta,
solid,
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{Nach.dat};
\addplot [
name path=approx2,
color=green,
solid,
line width=1.0pt
] table [x index = {0}, y index = {3}, col sep=comma]{Napproximations.dat};
\addplot [
name path=approx,
color=red,
solid,
line width=1.0pt
] table [x index = {0}, y index = {2}, col sep=comma]{Napproximations.dat};
\addplot [
name path=converse,
color=blue,
dashed,
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{Nconverse.dat};
\addplot [
name path=cap,
color=black,
solid,
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{Napproximations.dat};
\end{axis}
\end{tikzpicture}
\caption{Bounds and approximations for $N^*(k,\varepsilon)$ plotted for $\varepsilon=10^{-3}$ and $P=0\ \text{dB}$. The converse $N_c(k,\varepsilon)$ and achievability bound $N_a(k,\varepsilon)$ are plotted using the SPECTRE toolbox.}
\label{fig:fbl}
\end{figure}
We define $N^*(k, \varepsilon) \triangleq \minn{n \geq 0: n R^*(n,\varepsilon) \geq k}$ for $k\geq 1$ and $N^*(0,\varepsilon) \triangleq 0$ which is the smallest number of channel uses that allows the encoding of $k$ bits with error probability $\varepsilon$.
We obtain the following asymptotic approximation of $N^*(k, \varepsilon)$ as $k\rightarrow \infty$:
\begin{IEEEeqnarray}{rCl}
N^*(k, \varepsilon) &=& \frac{k}{C} + \sqrt{\frac{V k}{C^3}}Q^{-1}(\varepsilon) - \frac{1}{2 C} \log_2\frac{k}{C} + \mathcal{O}(1).\IEEEeqnarraynumspace\label{eq:Nasymptotic}
\end{IEEEeqnarray}
This can be verified by setting $\bar n$ equal to RHS of \eqref{eq:Nasymptotic} and by computing $\bar n R^*(\bar n,\varepsilon)$. Then, one finds that $\bar n R^*(\bar n,\varepsilon) = k +\mathcal{O}(1)$ from which \eqref{eq:Nasymptotic} follows.
We define the approximation
\begin{IEEEeqnarray}{rCl}
N(k,\varepsilon) \triangleq \text{uce}\mathopen{}\left(\frac{k}{C} + \sqrt{\frac{V k}{C^3}}Q^{-1}(\varepsilon) - \frac{1}{2 C} \log_2\frac{k}{C} \right)\label{eq:Napprox_uce}
\end{IEEEeqnarray}
where $\text{uce}(\cdot)$ stands for the upper concave envelope. It can be shown that the approximation of $N^*(k,\varepsilon)$ inside $\text{uce}(\cdot)$ in \eqref{eq:Napprox_uce} is concave for $k \geq \frac{4 C}{Q^{-1}(\varepsilon)^2 V \log_e(2)^2}$, implying that $N(k,\varepsilon) = N^*(k,\varepsilon) + \mathcal{O}(1)$. Additionally, in all numerical examples in this paper, we have $\frac{4 C}{Q^{-1}(\varepsilon)^2 V \log_e(2)^2} < 1$, and hence the upper concave envelope does not affect our numerical results. In Fig.~\ref{fig:fbl}, we have plotted the $\kappa\beta$-achievability bound $N_a(k,\varepsilon)$ and the meta-converse bound $N_c(k,\varepsilon)$ from \cite{Polyanskiy2010b} along with the approximation \eqref{eq:Napprox_uce}, and $\bar N(k,\varepsilon)=n$, where $n$ is a solution to:
\begin{IEEEeqnarray}{rCl}
n C - \sqrt{n V}Q^{-1}(\varepsilon) + \frac{1}{2} \log_2 n = k.
\end{IEEEeqnarray}
We observe that $N(k,\varepsilon)$ provides an approximation of $N^*(k,\varepsilon)$ that matches the converse bound closely.
In the remaining part of this paper, when referring to the blocklength of an optimal code conveying $k$ bits with a probability of error not exceeding $\varepsilon$, we consistently use the approximation $N(k,\varepsilon)$ in place of $N^*(k,\varepsilon)$ in all computations and derivations.
\begin{comment}
\begin{IEEEproof}
We set
\begin{IEEEeqnarray}{rCl}
\bar n = \frac{k}{C} + \sqrt{\frac{V k}{C^3}}Q^{-1}(\epsilon) - \frac{1}{2 C} \log_2\frac{k}{C} + \frac{d}{C}
\end{IEEEeqnarray}
and compute $\bar n R^*(\bar n,\epsilon)$:
\begin{IEEEeqnarray}{rCl}
\bar n R^*(\bar n,\epsilon) &=& k + \sqrt{\frac{V k}{C}}Q^{-1}(\epsilon)- \frac{1}{2} \log_2\frac{k}{C} + d \nonumber\\
&& {}-\sqrt{\frac{V k}{C}}Q^{-1}(\epsilon) +\frac{1}{2} \log_2\farg{\frac{k}{C}}+ \mathcal{O}(1).
\end{IEEEeqnarray}
Here, we have used that $\sqrt{Vk/C + \mathcal{O}(\sqrt{k})} = \sqrt{V k/C}+\mathcal{O}(1)$ and that $\log_2(k/C + \mathcal{O}(\sqrt{k})) = \log_2(k/C)+\mathcal{O}(1/\sqrt{k})$. This completes the proof
\end{IEEEproof}
\end{comment}
\section{System model}\label{sec:system_model}
We consider an AWGN broadcast channel with one transmitter and $K$ users. In the $t$-th time slot, the $k$-th user receive
\begin{IEEEeqnarray}{rCl}
Y_{k,t} \triangleq \sqrt{\gamma_k} X_t + Z_{k,t}.\label{eq:system_model}
\end{IEEEeqnarray}
where $Z_{k,t}\sim \mathcal{N}(0,1)$ and $X_t\in\mathbb{R}$ is the channel input. Throughout the paper, we assume that $\gamma_k = 1$. The assumption of equal channel conditions can, to some extend, be justified as follows. Consider a downlink broadcast scenario with many users with varying channel conditions. A viable communication strategy is to first divide the users into several CSI-groups such that the users assigned to a certain CSI-group have similar channel conditions. Then, the transmitter serves each CSI-group sequentially, and our system model in \eqref{eq:system_model} models a single CSI-group. A satellite-based broadcast system with line-of-sight to all users and predictable channel conditions constitute a practical example of our system model. If, however, CSI-grouping is not performed, then the transmitter needs to protect a packet destined to multiple users with a code that is strong enough to ensure that even the worst-channel user can decode. The assumption of nonfading channels is mainly introduced for simplicity, but we note that there are results in finite blocklength information theory for fading channels \cite{Yang2014}.
The message $\mathrm{M}_k$ destined to the $k$-th user is nonempty with probability $1-q\in(0,1)$, and we say that the $k$-th user is \emph{active} if there is a message destined to that user. We assume that the size of the message $\mathrm{M}_k$ (in bits) is given by $D_k\in\mathbb{Z}_+$ which is a discrete random variable distributed independently according to the probability mass function
\begin{IEEEeqnarray}{rCl}
P_D(d) \triangleq \left\{ \begin{array}{ll}
q & \text{if } d=0 \\
(1-q)p_i & \text{if } d=\alpha_i \text{ for } i \in\{1,\cdots,S\}.
\end{array}
\right.
\end{IEEEeqnarray}
The message $\mathrm{M}_k$ is drawn uniformly randomly from the set $\{0,1\}^{D_k}$. We use $\vect{\alpha}=(\alpha_1,\alpha_2, \ldots, \alpha_S)$ to denote a $S$-dimensional vector of distinct ordered positive integers ($\alpha_i<\alpha_s$ if $i<s$) that correspond to the possible message sizes.
The frame duration $T$ is a random variable that depends on the message sizes $\{D_k\}$. The transmitter encodes the message $\{\mathrm{M}_k\}$ into a sequence of channel inputs using the encoder function $f_t(\cdot)$ such that
\begin{align}
X_t \triangleq f_t(\{D_k\},\{\mathrm{M}_k\})
\end{align}
for $t\in\{1,\cdots,T\}$ and $X_t = 0$ for $t\in\{T+1,\cdots\}$. Additionally, we require that
\begin{IEEEeqnarray}{rCl}
\E{\frac{1}{T}\sum_{i=1}^T X_t} \leq P.
\end{IEEEeqnarray}
We define the ON-OFF function $g_{k,t}: (\mathbb{R} \cup \{\mathrm{e}\})^{t-1}\rightarrow \{0,1\}$ that defines the receiver activity for user $k$:
\begin{align}
\bar Y_{k,t} \triangleq \left\{ \begin{array}{ll}
Y_{k,t}, & g_{k,t}(\bar Y_{k}^{t-1}) = 1, \textrm{receiver is ON} \\
\mathrm{e}, & \text{receiver is OFF}
\end{array}
\right..
\end{align}
\begin{sloppypar}
The ON-OFF function replaces the $t$-th channel output with an erasure if the user is OFF at that time. The \emph{stopping time} $T_k$ represents the time index of the last nonerased channel output in the sequence $\bar Y_{k,t}$; after $T_k$ the receiver $k$ is OFF until the end of the frame. Formally, $T_k \triangleq \inff{n \geq 1:\forall t > n, g_{k,t}(\bar Y_{k}^{t-1}) = 0}$ for which we require $T_k < \infty$. Considering that a user can only use the channel outputs for which it is ON, we define the decoding function $h_{k,t}(\bar Y_k^t)$ to estimate the message $\mathrm{M}_k$ based on $\bar Y_{k}^t$. The ON-OFF functions are causal in the sense that the decision of whether the users are ON at time $t$ depends on previous channel outputs, $\bar Y_{k}^{t-1}$. Unless an error occurs during decoding, the stopping times $T_k$ are less than or equal $T$ for any practical applications of this model. We merely define $T_k$ to emphasize that $T$ is a random variable which is not known by the users, and hence the users need to obtain this information through the sequence $\bar Y_{k,t}$. In a conventional approach to downlink broadcast, as depicted in Fig.~\ref{fig:protocol_conventional}, control information in the initial packet defines the structure of the remaining transmission. Hence, after successfully decoding the control information in the initial packet, the $k$-th user knows $T_k$ and when to be ON and OFF to receive the message intended for that user.
\end{sloppypar}
The average power consumption of the $k$-th user is given by
\begin{IEEEeqnarray}{rCl}
P_k \triangleq \E{\sum_{i=1}^{T_k} \indi{g_{k,i}(\bar Y_{k}^{i-1} )=1}}\label{eq:power_def}
\end{IEEEeqnarray}
where $\indi{\cdot}$ is the indicator function, and is determined by the ON-OFF function. Note that $\E{P_1} = \E{P_k}$, for $k\in\{1,\cdots,K\}$, since the message sizes $D_k$ are distributed identically.
Finally, the active users need to decode their messages with reliability larger than or equal $1-\epsilon$ such that
\begin{align}
\pr{h_{k,T_k}(\bar Y_k^{T_k}) \not= \mathrm{M}_k | D_k > 0} \leq \epsilon\label{eq:reliability}
\end{align}
for $k\in\{1,\cdots,K\}$ and $\epsilon\in(0,1)$.
The above system model provides a general framework for the problem of downlink broadcast framing. For tractability, we constrain ourselves to an important and practical class of protocols described as follows. The transmitter forms $L\in\mathbb{Z}_+$ packets which are encoded using optimal codes with error probabilities $\{\bar \epsilon_l\}_{l\in\{1,\cdots,L\}}$. Here, $L$ and $\{\bar \epsilon_l\}$ are random variables that depend only on $\{D_k\}$. Let $L_{\text{max}}$ be a constant that denotes the maximum number of packets that the transmitter can send, defined as the smallest integer such that $L_{\text{max}}\geq L$ for all realizations of $\{D_k\}$. Let $\{\mathrm{M}^{(C)}_l\}_{l=1}^{L_{\text{max}}}$ denote the control information that needs to be conveyed in order to describe how the data for different users is conveyed (see the example below). Let
$\{D^{(C)}_l\}_{l=1}^{L_{\text{max}}}$ denote the sizes (in bits) of $\{\mathrm{M}^{(C)}_l\}_{l=1}^{L_{\text{max}}}$, i.e., $\mathrm{M}^{(C)}_l\in\{0,1\}^{D^{(C)}_l}$ and $D_l^{(C)} = 0$ for $l> L$. Let $\{\mathcal{U}_l\}_{l=1}^{L_{\text{max}}}$ denote disjoint random sets that depend only on $\{D_k\}$ such that $\bigcup_{i=1}^{L_{\text{max}}} \mathcal{U}_l = \mathcal{K}$ and such that $\mathcal{U}_l = \emptyset$ for $l> L$. The $l$-th packet then consists of the information bits $\mathrm{M}^{(C)}_l \oplus \bigoplus_{k\in\mathcal{U}_l} \mathrm{M}_k$ which are encoded by an optimal code with reliability $\bar \epsilon_l$ using $N\farg{D^{(C)}_l+\sum_{k\in\mathcal{U}_l} D_k,\bar \epsilon_l}$ channel uses. The encoder function $f_t(\cdot,\cdot)$ is defined by sequentially transmitting the $L$ encoded packets. The frame duration $T$ is given by $\sum_{l=1}^L N\farg{D^{(C)}_l+\sum_{k\in\mathcal{U}_l} D_k,\bar \epsilon_l}$.
We assume that the optimal code has the following property: If $j$ bits are encoded into $n$ channel uses by an optimal code with error probability $\varepsilon$, then the user needs to receive all $n$ channel uses so as to decode any of the $j$ bits with error probability $\varepsilon$.
As an illustration, we describe how the general framework is instantiated to describe a
conventional downlink frame from Fig.~\ref{fig:protocol_conventional}. Suppose $S = 3$ such that $D_k \in \{0,\alpha_1,\alpha_2,\alpha_3\}$. As there are four possible lengths, the control information about $\{D_k\}$ can be represented by at most $2K$ information bits which are conveyed in the first packet, commonly referred to as the \emph{header}.
We let $D^{(C)}_1 = 2K$ and let $\mathrm{M}_1^{(C)}$ be the bitstring of length $2K$ representing $\{D_k\}$.
Since there is a header packet and at most $K$ other packets, we set $L_{\text{max}} = L = K+1$.
We also set $\bar \epsilon_1 = \varepsilon_1$ and $\bar \epsilon_l = \varepsilon_2$ for $l\in\{2,\cdots,L_{\text{max}}\}$ where $(\varepsilon_1,\varepsilon_2)\in[0,1]^2$ are such that $\epsilon = 1-(1-\varepsilon_1)(1-\varepsilon_2)$. Since all control information is concentrated in the frame header, we have $D_l^{(C)}=0$ for $l\geq 2$.
The sets $\{\mathcal{U}_l\}_{l=1}^{L_{\text{max}}}$ are defined such that the header has no user data and
$\mathcal{U}_1 = \emptyset$, while $\mathcal{U}_l = \{l-1\}$ for $l\in\{2,\cdots,L_{\text{max}}\}$.
User $k$ is ON during the transmission of the first packet which it decodes with probability $1-\varepsilon_1$. If user $k$ successfully decodes the first packet, it learns $\{D_k\}$, and thereby it obtains a pointer to the location of the $(k+1)$-th packet, which contains the desired message $\mathrm{M}_k$. After decoding the header, the $k$-th user is OFF for the remaining time except when the $(k+1)$-th packet is transmitted. The $(k+1)$-th packet is successfully decoded with probability $1-\varepsilon_2$. The overall probability of error for the protocol from the viewpoint of a single user is given by $1-(1-\varepsilon_1)(1-\varepsilon_2) = \epsilon$ as desired.
For large message sizes $\alpha_s \gg 1$ we get the lower bounds:
\begin{IEEEeqnarray}{rCl}
\E{T} \geq \frac{K \E{D_1} }{C}\label{eq:Tbound} \\
\E{P_1} \geq \frac{\E{D_1}}{C} .\label{eq:Pbound}
\end{IEEEeqnarray}
When $\alpha_s \gg 1$, the control information becomes negligible, and hence for the conventional approach both $\E{T}$ and $\E{P_1}$ simultaneously approach the lower bounds in \eqref{eq:Tbound} and \eqref{eq:Pbound}.
Our objective is to explore trade-offs between the competing goals of minimizing $\E{T}$ and $\E{P_1}$.
\section{Lower bound}\label{sec:lower_bnd}
We establish a lower bound by assuming that the users are provided with control information from a genie, i.e., $\{D_k\}$ are known at all users. In that case, the transmitter and all users can agree on a protocol that only conveys the messages $\{M_k\}$, i.e., $D^{(C)}_l = 0$ for $l\in\{1,\cdots,L_{\text{max}}\}$. Hence, the transmitter may encode the messages $\{\mathrm{M}_k\}$ into at most $K$ separate packets such that each message is encoded in exactly one of these packets. Each packet may contain either no messages at all, a single message, or multiple concatenated messages, and they are encoded using optimal codes with error probabilities that do not exceed $\epsilon$ upon decoding; recall that all users experience the same error probability since $\gamma=1$.
Any genie-aided protocol can be characterized using $K$ random nonnegative integer vectors $\vect{N}_l\in \mathbb{Z}_+^S$, for $l\in\{1,\cdots,K\}$, that depend only on $\{D_k\}$. The content of the $l$-th packet is described by $\vect{N}_l$; the packet encodes $N_{l,1}$ messages of length $\alpha_1$, it encodes $N_{l,2}$ messages of length $\alpha_2$, etc. Note that the integer vectors $\{\vect{N}_l\}$ do not uniquely describe which messages are encoded in which packets. For a genie-aided protocol defined by a set of vectors $\{\vect{N}_l\}$, we compute the frame duration and average power as follows
\begin{IEEEeqnarray}{rCl}
T &=& \sum_{l=1}^K N(\vect{\alpha}\tr \vect{N}_l, \epsilon)\label{eq:prot_T} \\
\frac{1}{K}\sum_{k=1}^K P_i &=& \frac{1}{K}\sum_{l=1}^K \vect{1}_S\tr \vect{N}_l N(\vect{\alpha}\tr \vect{N}_l, \epsilon).\label{eq:prot_P}
\end{IEEEeqnarray}
Here, $T$ and $\frac{1}{K}\sum_{k=1}^K P_i$ are random variables that depend only on the realization of $\{D_k\}$. We aim to lower bound $\E{T} + \beta \E{P_1}$ for any $\beta>0$ and thereby obtain a lower bound on the average power consumption $\E{P_1}$ as a function of average frame duration $\E{T}$.
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=2000,
xmin=8500,
ymax=9000,
xmax=9200,
xlabel={Frame duration [ch. uses]},
ylabel={Power [ch. uses]},
legend entries = { {}{\eqref{eq:ex4} for all $\beta>0$}, {}{\eqref{eq:ex4} with $\beta = 0.035$}, {}{\eqref{eq:ex4} with $\beta = 0.095$}, {}{\eqref{eq:ex4} with $\beta = 0.245$}}
]
\addplot [
color=black,
line width=1.0pt] table [x index = {0}, y index = {1}, col sep=comma]{low_bnd_ex.dat};
\addplot [
color=green,
line width=1.0pt] table [x index = {0}, y index = {2}, col sep=comma]{low_bnd_ex.dat};
\addplot [
color=red,
line width=1.0pt] table [x index = {0}, y index = {3}, col sep=comma]{low_bnd_ex.dat};
\addplot [
color=blue,
line width=1.0pt] table [x index = {0}, y index = {4}, col sep=comma]{low_bnd_ex.dat};
\addplot [
only marks,
mark = *,
line width=1.0pt] table [x index = {0}, y index = {1}, col sep=comma]{low_bnd_ex_points.dat};
\end{axis}
\end{tikzpicture}
\caption{Depicts the lower bound in \eqref{eq:ex4} for three different values of $\beta$ for $P=0\ \text{dB}$, $\epsilon = 10^{-4}$, and $\alpha_1 = 1000$. The black curve is obtained by evaluating \eqref{eq:ex4} for all $\beta>0$ and by combining the resulting lower bounds. The dots correspond to five genie-aided protocols described by $(N_1,\cdots,N_4) \in [(4,0,0,0), (3, 1,0,0),(2,2,0,0),(2,1,1,0),(1,1,1,1)]$ (enumerated from top-left corner to bottom-right corner).}
\label{fig:low_bnd_ex}
\end{figure}
Before stating the lower bound, we introduce the technique through an example. Suppose $K=4$, $q=0$, $S=1$ such that $D_1 = \cdots= D_4 = \alpha_1$ and the frame duration and average power are deterministic. Since the users know $\{D_k\}_{k=1}^4$ and each of the four messages belongs to one encoded packet, any genie-aided protocol can be described through the four nonnegative integers $N_1,N_2, N_3$, and $N_4$ satisfying $N_1 +\cdots+ N_4 = 4$. These integers represent the number of messages encoded in the first, second, third, and fourth packet, respectively. For fixed $\beta>0$, our objective is to minimize $T + \frac{\beta}{4} \sum_{l=1}^4 P_k$ with respect to $N_1, \cdots,N_4\in \{0,\cdots,4\}$ subject to $N_1+\cdots+N_4 = 4$. For this particular example, one can easily solve the resulting integer optimization problem. However, we can also find a lower bound on $T + \frac{1}{4}\beta \sum_{l=1}^4 P_k$ through the following steps
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{T + \frac{\beta}{4} \sum_{l=1}^4 P_k}\nonumber\\
\quad &=& \sum_{l=1}^4 N(\alpha_1 N_l, \epsilon)+ \frac{\beta}{4}\sum_{l=1}^4 N_l N(\alpha_1 N_l, \epsilon)\\
&\geq& \min_{\substack{n_1,\cdots,n_4\in \{0,\cdots,4\}:\\ n_1+\cdots+n_4=4 }}\sum_{l=1}^4 \Phi_\beta(n_l)\label{eq:ex1}\\
&\geq& \min_{\substack{n_1,\cdots,n_4\in \{0,\cdots,4\}:\\ n_1+\cdots+n_4=4 }}\sum_{l=1}^4 \breve{\Phi}_\beta(n_l)\label{eq:ex2}\\
&\geq& \min_{\substack{n_1,\cdots,n_4\in \{0,\cdots,4\}:\\ n_1+\cdots+n_4=4 }} 4\breve{\Phi}_\beta\farg{\sum_{l=1}^4 \frac{n_l}{4}}\label{eq:ex3}\\
&=& 4\breve{\Phi}_\beta\farg{1}.\label{eq:ex4}
\end{IEEEeqnarray}
Here, \eqref{eq:ex1} follows by defining
\begin{IEEEeqnarray}{rCl}
\Phi_\beta(x) \triangleq N(\alpha_1 x, \epsilon)\left(1+ \frac{\beta x}{4}\right),
\end{IEEEeqnarray}
and by a minimization with respect to $n_1,\cdots,n_4$, \eqref{eq:ex2} follows by defining $\breve{\Phi}_\beta(\cdot)$ as the lower convex envelope of $\Phi_\beta(\cdot)$, and \eqref{eq:ex3} is by convexity of the lower convex envelope of $\Phi_\beta(\cdot)$. Interestingly, the bound in \eqref{eq:ex4} is fairly tight and simple to compute. We illustrate the bound \eqref{eq:ex4} in Fig.~\ref{fig:low_bnd_ex}, confirming the intuition that when $S=1$, one should attempt to have an equal number of non-empty messages in each packet, which for this example is $1$, $2$, or $4$ messages in each packet.
For the general setting with arbitrary $S\geq 1$ and $K$, we apply the above ideas in the following proposition which enables us to compute a lower bound on $\E{P_1}$ for certain $\E{T}$.
\begin{proposition}
\label{prop:prop_converse}
For every $\beta> 0$, we have
\begin{IEEEeqnarray}{rCl}
\E{T} + \beta \E{P_1} \geq \E{ \vect{1}_S\tr \vect{L}_{1:S} \breve{\phi}_{\beta}\farg{ \frac{\vect{L}_{1:S}}{ \vect{1}_S\tr \vect{L}_{1:S}} } }
\end{IEEEeqnarray}
where $\vect{L}\in\mathbb{Z}_+^{S+1}$ is multinomial distributed with $S+1$ categories, $K$ trials, and event probabilities $[(1-q)p_1, \cdots, (1-q)p_S,q]$, $\vect{L}_{1:S}$ denotes the first $S$ entries of $\vect{L}$, and $\breve{\phi}_{\beta}: \mathbb{R}_+^{S} \mapsto \mathbb{R}_+$ is the lower convex envelope of the function
\begin{IEEEeqnarray}{rCl}
\phi_\beta(\vect{x})\triangleq N\farg{\vect{\alpha}\tr \vect{x},\epsilon} \left(1+\frac{\beta \vect{1}_S\tr \vect{x}}{K}\right)\IEEEeqnarraynumspace\label{eq:non_phi_def}
\end{IEEEeqnarray}
defined for $\vect{x}\in\mathbb{R}_+^{S}$.
\end{proposition}
\begin{IEEEproof}
Fix $\beta> 0$. The users $\mathcal{K}$ can be decomposed into $S+1$ disjoint subsets $\{\mathcal{U}^{(s)}\}_{s\in\{0,\cdots,S\}}$ such that $\mathcal{U}^{(s)}\triangleq \{k \in \mathcal{K}: D_k=\alpha_s\}$ for $s\in\{0,\cdots,S\}$, where we let $\alpha_0 \triangleq 0$ for notational convenience. We denote the (random) set of active users by $\mathcal{U}=\bigcup_{s=1}^S \mathcal{U}^{(s)}$.
Fix a genie-aided protocol. Then, since we assume that the users are provided with control information by a genie, the protocol must decompose the set of active users $\mathcal{U}$ into at most $K$ (possibly empty) disjoint subsets $\{\mathcal{U}_l\}_{l\in\{1,\cdots,K\}}$. Note that these subsets are random, depend only on $\{D_k\}$, and are induced by the protocol. Define the random integer vectors $\vect{N}_l\in\mathbb{Z}_+^S$, for $l\in\{1,\cdots,K\}$ and $s\in\{1,\cdots,S\}$, as follows:
\begin{IEEEeqnarray}{rCl}
N_{l,s} \triangleq \sum_{k\in\mathcal{K}} \indi{k \in\mathcal{U}_l \text{ and } D_k = \alpha_s}.
\end{IEEEeqnarray}
The average frame duration and the average power for the genie-aided protocol in terms of $\{\vect{N}_i\}$ are now given by
\begin{IEEEeqnarray}{rCl}
\E{T} &=& \E{\sum_{l=1}^K N(\vect{\alpha}\tr \vect{N}_l,\epsilon)}\label{eq:low_bnd_ET}\\
\E{P_1} &=& \E{\frac{1}{K} \sum_{l=1}^K \vect{1}_S \tr \vect{N}_l N(\vect{\alpha}\tr \vect{N}_l,\epsilon)}.\label{eq:low_bnd_EP}
\end{IEEEeqnarray}
Now, we compute a lower bound on $\E{T}+\beta \E{P_1}$ based on \eqref{eq:low_bnd_ET} and \eqref{eq:low_bnd_EP}:
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\E{T}+\beta \E{P_1}} \nonumber\\\quad
&=& \E{\sum_{l=1}^K N(\vect{\alpha}\tr \vect{N}_l,\epsilon) +\frac{\beta}{K} \sum_{l=1}^K \vect{1}_S \tr \vect{N}_l N(\vect{\alpha}\tr \vect{N}_l,\epsilon)}\label{eq:non_UK0}\\
&=& \E{\sum_{l=1}^{K} \phi_\beta(\vect{N}_{l})}\label{eq:non_UK1}\\
&\geq& \E{\min_{\substack{\vect{n}_{1},\cdots,\vect{n}_{|\mathcal{U}|}\in\mathbb{Z}^S_+:\\\sum_{l=1}^{|\mathcal{U}|} n_{l,s}=|\mathcal{U}^{(s)}|}} \sum_{l=1}^{|\mathcal{U}|} \phi_\beta(\vect{n}_{l})}\label{eq:non_UK2}
\end{IEEEeqnarray}
where \eqref{eq:non_UK1} is by the definition of $\phi_\beta(\cdot)$ in \eqref{eq:non_phi_def}. In \eqref{eq:non_UK2}, the expectation is only with respect to the random variables $|\mathcal{U}^{(1)}|, \cdots,|\mathcal{U}^{(S)}|$ and $|\mathcal{U}|$.
Next, \eqref{eq:non_UK2} is lower-bounded by using the lower convex envelope of $\phi_{\beta}(\cdot)$ and its convexity:
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\E{T}+\beta \E{P_1}}\nonumber\\\quad
&\geq& \E{\min_{\substack{\vect{n}_{1},\cdots,\vect{n}_{|\mathcal{U}|}\in\mathbb{Z}_+^S:\\\sum_{l=1}^{|\mathcal{U}|} n_{l,s}=|\mathcal{U}^{(s)}|}} \sum_{l=1}^{|\mathcal{U}|} \breve{\phi}_\beta(\vect{n}_i)}\label{eq:non_UK5}\\
&\geq& \E{\min_{\substack{\vect{n}_{1},\cdots,\vect{n}_{|\mathcal{U}|}\in\mathbb{Z}_+^S:\\\sum_{l=1}^{|\mathcal{U}|} n_{l,s}=|\mathcal{U}^{(s)}|}} |\mathcal{U}| \breve{\phi}_\beta\mathopen{}\Bigg(\frac{1}{|\mathcal{U}|}\sum_{l=1}^{|\mathcal{U}|}\vect{n}_{l}\Bigg)}\label{eq:non_UK6}\\
&=& \E{|\mathcal{U}|\breve{\phi}_{\beta}\farg{ \left[\frac{|\mathcal{U}^{(1)}|}{|\mathcal{U}|},\cdots,\frac{|\mathcal{U}^{(S)}|}{|\mathcal{U}|}\right] }}.\label{eq:non_UK7}
\end{IEEEeqnarray}
Here, \eqref{eq:non_UK5} follows because the lower convex envelope $\breve{\phi}_{\beta}(\cdot)$ of $\phi_{\beta}(\cdot)$ is smaller than or equal $\phi_{\beta}(\cdot)$ and \eqref{eq:non_UK6} follows from convexity of $\breve{\phi}_{\beta}(\cdot)$. The result follows by noting that the random vector $\left[|\mathcal{U}^{(1)}|,\cdots,|\mathcal{U}^{(S)}|,|\mathcal{U}^{(0)}|\right]$ is multinomial distributed with $S+1$ categories, $K$ trials, and event probabilities $[(1-q)p_1, \cdots, (1-q)p_S,q]$.
\end{IEEEproof}
The following lemma shows that we can use the concavity of $N(\cdot,\epsilon)$ to simplify the computation of $\breve{\phi}_\beta(\cdot)$.
\begin{lemma}
\label{lem:breve_phi}
For every $\beta > 0$, we have
\begin{IEEEeqnarray}{rCl}
\breve{\phi}_\beta(\vect{x}) = \min_{\substack{\vect{\zeta}\in\mathbb{R}^S:\\ \vect{1}_S\tr \vect{\zeta} = 1 \\\forall s: \zeta_s >0 }} \sum_{s=1}^S \zeta_s \breve{\phi}^{(s)}_\beta(x_s/\zeta_s)\label{eq:breve_simple_comp_claim}
\end{IEEEeqnarray}
where we have defined
\begin{IEEEeqnarray}{rCl}
\phi_\beta^{(s)}(x) \triangleq N(\alpha_s x,\epsilon)(1+\beta x/K)\label{eq:phis_def}
\end{IEEEeqnarray}
for $x\geq 0$ and $s\in\{1,\cdots,S\}$.
Additionally, the optimization problem in \eqref{eq:breve_simple_comp_claim} is convex.
\end{lemma}
\begin{IEEEproof}
See Appendix~\ref{app:lem_proof}
\end{IEEEproof}
For the case with fixed message sizes, i.e., when $S=1$, Proposition~\ref{prop:prop_converse} reduces to the following corollary.
\begin{corollary}
For every $\beta\geq 0$, we have
\begin{IEEEeqnarray}{rCl}
\E{T} + \beta \E{P_1} \geq (1-q) K \breve{\phi}^{(1)}_{\beta}(1)
\end{IEEEeqnarray}
where $\phi^{(1)}_\beta(\cdot)$ is defined in \eqref{eq:phis_def}.
\end{corollary}
This readily follows from $\vect{L}_{1:S}/(\vect{1}_S\tr \vect{L}_{1:S}) = 1$ and because $\vect{1}_S\tr \vect{L}_{1:S}$ is Binomial distributed with parameters $K$ and $1-q$, and hence $\vect{1}_S\tr \vect{L}_{1:S} = (1-q)K$.
\begin{comment}
\begin{proposition}
Fix $K\in\mathbb{N}$ and $L \in\{1,\cdots,K\}$. Define
\begin{IEEEeqnarray}{rCl}
T_{L} &=& \E{ \lfloor |\mathcal{U}|/L\rfloor} N(L\alpha,\epsilon)\\
P_{L} &=& \E{\frac{\lfloor|\mathcal{U}|/L\rfloor}{\lfloor K/L\rfloor}} N(L\alpha,\epsilon).
\end{IEEEeqnarray}
Then, there does not exist a protocol that simultanously satisfies
\begin{IEEEeqnarray}{rCl}
\E{T} &<& T_{L}\\
\text{and } \E{P_1} &<& P_L
\end{IEEEeqnarray}
for $k\in\{1,\cdots,K\}$.
\end{proposition}
\begin{IEEEproof}
For notational convenience, we will use $N(\cdot) \triangleq N(\cdot,\epsilon)$. For any $\beta> 0$ and $M\in\mathbb{N}$, let $\{\mathcal{\hat U}_i^{(M)}\}_{i\in\{1,\cdots,M\}}$ be disjoint subsets such that $\bigcup_{i=1}^M \mathcal{\hat U}_i^{(M)} = \{1,\cdots,M\}$ and such that they solve the optimization problem
\begin{IEEEeqnarray}{rCl}
U_M = \min_{\{\mathcal{U}_i^{(M)}\}} \sum_{i=1}^M N(|\mathcal{U}^{(M)}_i| \alpha) + \frac{\beta}{M} \sum_{i=1}^M |\mathcal{U}^{(M)}_i| N(|\mathcal{U}^{(M)}_i| \alpha).\IEEEeqnarraynumspace\label{eq:Us}
\end{IEEEeqnarray}
For convenience, we also define $\mathcal{U}_{M+1}^{(M)} \triangleq \emptyset$.
Assume, without loss of generality, that $|\mathcal{\hat U}_1^{(M)}| \geq |\mathcal{\hat U}_2^{(M)}| \geq \cdots \geq |\mathcal{\hat U}_M^{(M)}|$ and that $M\in\mathcal{\hat U}_1^{(M)}$. Define
\begin{IEEEeqnarray}{rCl}
L &\triangleq& \maxx{M\in\mathbb{N}: |\mathcal{\hat U}_2^{(M)}| = 0}.
\end{IEEEeqnarray}
It follows that, for any $v\in\mathbb{N}$, we have $|\mathcal{\hat U}_1^{(v L)}| = \cdots = |\mathcal{\hat U}_v^{(v L)}| = L$ and $|\mathcal{\hat U}_{i}^{(v L)}|=0$ for $i\geq v+1$
We will also need the property $U_{\tilde K+1}\geq U_{\tilde K}$ for $\tilde K\in\mathbb{N}$. Let $\mathcal{U}^{(\tilde K)}_i = \mathcal{\hat U}_i^{(\tilde K+1)}$ for $i\in\{2,\cdots,\tilde K+1\}$ and let $\mathcal{U}^{(\tilde K)}_1 = \mathcal{\hat U}_1^{(\tilde K+1)}\setminus \{\tilde K+1\}$. Then, $\{\mathcal{U}^{(\tilde K)}_i\}_{i\in\{1,\cdots,\tilde K+1\}}$ consists of at most $\tilde K$ nonempty disjoint sets satisfying $\bigcup_{i=1}^{(\tilde K+1)}\mathcal{U}^{(\tilde K)}_i = \{1,\cdots,\tilde K\}$. Now, the property $U_{\tilde K+1}\geq U_{\tilde K}$ follows from the inequalities
\begin{IEEEeqnarray}{rCl}
\sum_{i=1}^{\tilde K+1} N(|\mathcal{\hat U}^{(\tilde K+1)}_i| \alpha) \geq \sum_{i=1}^{\tilde K+1} N(|\mathcal{U}^{(\tilde K)}_i| \alpha)
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\frac{1}{\tilde K+1}\sum_{i=1}^{\tilde K+1} |\mathcal{\hat U}^{(\tilde K+1)}_i| N(|\mathcal{\hat U}^{(\tilde K+1)}_i| \alpha)}\nonumber\\
&\geq& \frac{1}{\tilde K}\Bigg( \left(|\mathcal{\hat U}^{(\tilde K+1)}_1|-1\right) N(|\mathcal{\hat U}^{(\tilde K+1)}_1| \alpha)\nonumber\\
&&\qquad {}+\sum_{i=2}^{\tilde K+1} |\mathcal{\hat U}^{(\tilde K+1)}_i| N(|\mathcal{\hat U}^{(\tilde K+1)}_i| \alpha)\Bigg) \label{eq:largest}\\
&\geq&\frac{1}{\tilde K}\sum_{i=1}^{\tilde K+1} |\mathcal{U}_i^{(\tilde K)}| N(|\mathcal{U}^{(\tilde K)}_i| \alpha).\label{eq:largest2}
\end{IEEEeqnarray}
Here, \eqref{eq:largest} follows because $N(|\mathcal{\hat U}^{(\tilde K+1)}_1| \alpha) \geq N(|\mathcal{\hat U}^{(\tilde K+1)}_i| \alpha)$ for $i\geq 2$ and \eqref{eq:largest2} is by $|\mathcal{\hat U}_1^{(\tilde K+1)}| = |\mathcal{U}_1^{(\tilde K+1)}|+1$ and $N(|\mathcal{\hat U}^{(\tilde K+1)}_i| \alpha) \geq N(|\mathcal{U}^{(\tilde K)}_i| \alpha) $ for $i\in\{1,\cdots,\tilde K+1\}$.
By defining $\bar K\triangleq L\lfloor K/L\rfloor$, we have
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\E{T} + \beta \E{P_k}\geq U_K}\nonumber\\
&\geq& U_{\bar K}\\
&=& \E{\min \sum_{i=1}^{\bar K} N(|\mathcal{U}_i| \alpha) + \frac{\beta}{\bar K} \sum_{i=1}^{\bar K} |\mathcal{U}_i| N(|\mathcal{U}_i| \alpha) } \\
&\geq&\E{ \lfloor |\mathcal{U}|/L\rfloor N(L \alpha) + \frac{\beta}{L \lfloor K/L\rfloor } \lfloor |\mathcal{U}|/L\rfloor L N(L \alpha)} \label{eq:proof_lower_bound1}\\
&=&\E{ \lfloor |\mathcal{U}|/L\rfloor +\beta\frac{ \lfloor |\mathcal{U}| /L\rfloor }{\lfloor K/L\rfloor} } N(L \alpha).\label{eq:proof_lower_bound}
\end{IEEEeqnarray}
Here, \eqref{eq:proof_lower_bound1} follows because $\bar K$ is a multiple of $L$.
For each $L$, there exists $\beta_L>0$ such that \eqref{eq:proof_lower_bound} is satisfied. Hence,
\begin{IEEEeqnarray}{rCl}
\E{T} + \beta_L \E{P} &\geq&\E{ \lfloor |\mathcal{U}|/L\rfloor +\beta_L\frac{ \lfloor |\mathcal{U}| /L\rfloor }{\lfloor K/L\rfloor} } N(L \alpha)\\
&=& T_L + \beta_L P_L.\label{eq:contradiction}
\end{IEEEeqnarray}
Suppose that $\E{T} < T_L$ and that $\E{P_k}< P_L$, then we have $\E{T} + \beta_L \E{P_k} < T_L + \beta_L P_L$ which contradicts \eqref{eq:contradiction}. This argument completes the proof.
\end{IEEEproof}
\end{comment}
\begin{comment}
\begin{proposition}
For every $\beta\geq 0$, we have
\begin{IEEEeqnarray}{rCl}
\E{T} + \beta \E{P_1} \geq (1-q) K \breve{\phi}_{\beta}(\alpha)
\end{IEEEeqnarray}
where $\breve{\phi}_{\beta}(x)$ is the lower convex envelope of the function
\begin{IEEEeqnarray}{rCl}
\phi_{\beta}(x)\triangleq N(x,\epsilon) (1+\beta x/(K \alpha)).\label{eq:phi_def}
\end{IEEEeqnarray}
\end{proposition}
\begin{IEEEproof}
For notational convenience, we will use $N(\cdot) \triangleq N(\cdot,\epsilon)$ throughout the proof. Fix $\beta> 0$ and $K\in\mathbb{N}$. Since we assume that the users are provided with control information by a genie, the converse bound reduces to optimally distributing the active users $\mathcal{U}$ into at most $K$ disjoint subsets $\{\mathcal{U}_i\}_{i\in\{1,\cdots,K\}}$. Thus, we have the following lower bound
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\E{T}+\beta \E{P_1}=\E{T}+\frac{\beta}{K} \sum_{k=1}^K\E{P_k}} \nonumber\\
&\geq& \EE{\mathcal{U}}{\min_{\substack{\{\mathcal{U}_i\} \text{ disjoint}:\\ \bigcup_i \mathcal{U}_i = \mathcal{U}}}\sum_{i=1}^K N(\alpha|\mathcal{U}_i| )+ \frac{\beta}{K} \sum_{i=1}^K |\mathcal{U}_i| N(\alpha |\mathcal{U}_i|) }\IEEEeqnarraynumspace\\
&=& \EE{\mathcal{U}}{\min_{\substack{n_1,\cdots,n_{|\mathcal{U}|}\in\mathbb{Z}_+:\\\sum_{i=1}^{|\mathcal{U}|} n_i=|\mathcal{U}|}} \sum_{i=1}^{K} \phi_\beta(\alpha n_i)}\label{eq:UK2}
\end{IEEEeqnarray}
where $\phi(\cdot)$ is defined in \eqref{eq:phi_def}.
We can lower-bound \eqref{eq:UK2} as follows:
\begin{IEEEeqnarray}{rCl}
\E{T}+\beta \E{P_1} &\geq& \EE{\mathcal{U}}{\min_{\substack{n_1,\cdots,n_{|\mathcal{U}|}\in\mathbb{Z}_+:\\\sum_{i=1}^{|\mathcal{U}|}n_i=|\mathcal{U}|}} \sum_{i=1}^{|\mathcal{U}|} \breve{\phi}_{\beta}(\alpha n_i)}\label{eq:UK5}\\
&\geq& \EE{\mathcal{U}}{\min_{\substack{c_1,\cdots,c_{|\mathcal{U}|}\in\mathbb{R}_+:\\\sum_{i=1}^{|\mathcal{U}|} c_i=|\mathcal{U}|}} \sum_{i=1}^{|\mathcal{U}|} \breve{\phi}_\beta(\alpha c_i)}\label{eq:UK4}\\
&=& \EE{\mathcal{U}}{|\mathcal{U}|\breve{\phi}_{\beta}\farg{\alpha}}\label{eq:UK6}\\
&=& (1-q) K \breve{\phi}_{\beta}\farg{\alpha}.
\end{IEEEeqnarray}
Here, \eqref{eq:UK5} follows because the lower convex envelope $\breve{\phi}_{\beta}(\cdot)$ of $\phi_{\beta}(\cdot)$ is smaller than $\phi_{\beta}(\cdot)$, \eqref{eq:UK4} follows by extending the domain of optimization from the nonnegative integers to the nonnegative real numbers, and finally, \eqref{eq:UK6} follows from convexity of $\breve{\phi}_{\beta}(\cdot)$.
\end{IEEEproof}
\end{comment}
\subsection{Genie-aided protocol}\label{sec:genie}
We put forth a genie-aided protocol that uses the intuition obtained through Proposition~\ref{prop:prop_converse} and Lemma~\ref{lem:breve_phi}. Here, ``genie-aided'' refers to the fact that the protocol assumes that the knowledge about $\{D_k\}$ is available at all users. Lemma~2 suggests that one should group messages of the same sizes together rather than grouping messages of mixed message sizes. The purpose of introducing a genie-aided protocol is to show that it achieves a trade-off close to that of the lower bound. Moreover, we can compare the non-genie-aided protocols, introduced in Section~\ref{sec:protocol_design}, to the genie-aided protocol to show the impact of control information. Such comparisons are provided in Section~\ref{sec:numerical}.
First, for a set of users $\mathcal{\bar U}\subseteq \mathcal{K}$, $V\in\mathbb{N}$, and $\varepsilon\in(0,1)$, we define a $(\mathcal{\bar U}, V, \varepsilon)$-protocol as follows. The users $\mathcal{\bar U}$ are divided into $G\triangleq\lceil |\mathcal{\bar U}|/V \rceil$ disjoint sets $\{\mathcal{\bar U}_l\}_{l\in\{1,\cdots,G\}}$ such that
\begin{IEEEeqnarray}{rCl}
|\mathcal{\bar U}_l|&\triangleq& \left\{\begin{array}{ll}
\lfloor |\mathcal{\bar U}|/ G \rfloor+1, & l\in\{1,\cdots,\text{mod}(|\mathcal{\bar U}|, G)\}\\
\lfloor |\mathcal{\bar U}|/ G \rfloor, & \text{otherwise}.
\end{array}\right.
\end{IEEEeqnarray}
One can verify that $\sum_{l=1}^G |\mathcal{\bar U}_l| = |\mathcal{\bar U}|$.
Sequentially, for $l\in\{1,\cdots,G\}$, the transmitter encodes and conveys a packet containing $\bigoplus_{k\in \mathcal{\bar U}_l}\mathrm{M}_k$ with error probability $\varepsilon$ using $N(\sum_{k\in\mathcal{\bar U}_l} D_k,\varepsilon)$ channel uses. Here, $\oplus$ denotes the concatenation of messages. While the number of channel uses spend at the transmitter is given by $\sum_{i=1}^{G} N\farg{\sum_{k\in\mathcal{\bar U}_i} D_k,\epsilon}$, each user only needs to receive and decode one of the $G$ packets. We also note that a $(\mathcal{\bar U}, V, \varepsilon)$-protocol assumes control information at all the users $\mathcal{\bar U}$, i.e., the users needs to know $\mathcal{\bar U}$, $\{D_k\}_{k\in\mathcal{\bar U}}$, $V$, and $\varepsilon$.
For our genie-aided protocol, we define $\mathcal{U}^{(s)} \triangleq \{k\in \mathcal{K}: D_k = \alpha_s\}$, for $s\in\{0,\cdots, S\}$, and fix a vector $\vect{V}\in\mathcal{K}^S$. Now, sequentially for each $s\in\{1,\cdots,S\}$, the transmitter delivers the messages of the users $\mathcal{U}^{(s)}$ using a $(\mathcal{U}^{(s)}, V_s,\epsilon)$-protocol. We denote the average frame duration $\E{T}$ and the average power $\E{P_1}$ by $\bar T_{\text{genie}}^{(\vect{V},\epsilon)}$ and $\bar P_{\text{genie}}^{(\vect{V},\epsilon)}$, respectively.
The vector $\vect{V}$ is left to be specified. We can trace of optimal trade-off between average frame duration and average power by solving the integer optimization problem for all $\beta \geq 0$:
\begin{IEEEeqnarray}{rCl}
\min_{\vect{V} \in\mathcal{K}^S} \bar T_{\text{genie}}^{(\vect{V},\epsilon)} + \beta \bar P_{\text{genie}}^{(\vect{V},\epsilon)}
\end{IEEEeqnarray}
\section{Protocol Design}\label{sec:protocol_design}
In the following, we devise actual protocols that trade-off between average frame duration and average power consumption at the users. In contrast to the genie-aided protocol in Section~\ref{sec:genie}, these protocols need to convey control information.
\subsection{Fixed message size}
We initiate our discussion of protocol design with the case of fixed message size, i.e., $S=1$. In this case, the control information only consists of which users are active.
We divide the set of users $\mathcal{K}$ into $B \triangleq \lceil K/W\rceil$ disjoint subsets $\mathcal{K}_1,\cdots,\mathcal{K}_{B}$ such that $\bigcup_{i=1}^{B}\mathcal{K}_i = \mathcal{K}$ and such that
\begin{IEEEeqnarray}{rCl}
|\mathcal{K}_i| &=& \left\{\begin{array}{ll}
\lfloor K/B\rfloor + 1, & i\in\{1,\cdots,\text{mod}(K, B)\} \\
\lfloor K/B\rfloor, & \text{otherwise}.
\end{array}
\right.
\end{IEEEeqnarray}
Here, $W\in\mathbb{N}$ is a protocol parameter to be set.
The subsets $\{\mathcal{K}_i\}$ of $\mathcal{K}$ are termed user groups (UG). The transmitter forms a packet that contains only the number of active users in each UG, i.e., the packet encodes the vector $\left[|\mathcal{U}\cap \mathcal{K}_1|, |\mathcal{U}\cap \mathcal{K}_2|,\cdots,|\mathcal{U}\cap \mathcal{K}_{B}|\right]$. This vector constitutes a first layer of control information and can be uniquely represented by at most $k_1 = \lceil\lceil K/W\rceil\log_2 W\rceil$ bits. We encode the control information by an optimal channel code with error probability not exceeding $\epsilon_1\in(0,1)$ which can be achieved by approximately $N(k_1,\epsilon_1)$ channel uses.
After successfully decoding the first packet, the users know the number of users in each UG, and thereby the structure of the remaining part of the transmission. The second layer encodes control information and messages associated with each UG. Specifically, for the $i$-th UG, the transmitter needs to inform the users of the $i$-th UG about which $|\mathcal{U}\cap \mathcal{K}_i|$ users of $\mathcal{K}_i$ are active. Hence, the control information for the $i$-th UG, can be represented by $k_{2,i} \triangleq\Big\lceil\log_2 {{|\mathcal{K}_i|} \choose {|\mathcal{U}\cap \mathcal{K}_i|}}\Big\rceil$ bits and is conveyed by using an optimal code with error probability not exceeding $\epsilon_2\in(0,1)$, which requires approximately $N(k_{2,i},\epsilon_2)$ channel uses. Now, the messages of the active users in the $i$-th UG $\mathcal{U}_i \triangleq \mathcal{U}\cap \mathcal{K}_i$ are conveyed with error probability not exceeding $\epsilon_3\in(0,1)$ using an $(\mathcal{U}_i, V, \epsilon_3)$-protocol, where $V\in\mathcal{K}$ is another protocol parameter to be set. We emphasize that we can use an $(\mathcal{U}_i, V, \epsilon_3)$-protocol because the set of active users $\mathcal{U}_i$ knows $\mathcal{U}_i$ from the the control information provided that the first two packets are successfully decoded. Based on the description of the protocol above, one can compute $\E{T}$ and $\E{P_1}$ which we denote by $\bar T_{\text{fixed}}^{(V,W,\vect{\epsilon})}$ and $\bar P_{\text{fixed}}^{(V,W,\vect{\epsilon})}$, respectively. Here, $\vect{\epsilon}$ is the vector $[\epsilon_1,\epsilon_2,\epsilon_3]$.
The parameters $V$, $W$, and $\vect{\epsilon}$ are left to be specified. We can trace the optimal achievable trade-off of the proposed protocol by solving the following optimization problem for all $\beta \geq 0$:
\begin{IEEEeqnarray}{rCl}
\min_{(V,W)\in\mathcal{K}^2} \min_{\substack{\vect{\epsilon}\in[0,1]^3:\\\prod_{k=1}^3 (1-\epsilon_k) \geq 1-\epsilon}} \bar T_{\text{fixed}}^{(V,W,\vect{\epsilon})} + \beta\bar P_{\text{fixed}}^{(V,W,\vect{\epsilon})}.\label{eq:general_prot_opt}
\end{IEEEeqnarray}
While the outer minimization is an integer optimization problem which can only be solved using exhaustive search, the inner minimization is convex and can be solved using standard convex optimization algorithms. This is shown in the following lemma.
\begin{lemma}
The inner optimization problem in \eqref{eq:general_prot_opt} is convex in $\vect{\epsilon}$.
\end{lemma}
\begin{IEEEproof}
Note that, for fixed $V$ and $W$, the objective function in \eqref{eq:general_prot_opt} depends only on $\vect{\epsilon}$ through a nonnegative linear combination of $Q$-functions of $\epsilon_1$, $\epsilon_2$, and $\epsilon_3$, i.e., there exist nonnegative constants $a_1, a_2,$ and $a_3$ such that
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\bar T_{\text{fixed}}^{(V,W,\vect{\epsilon})} + \beta\bar P_{\text{fixed}}^{(V,W,\vect{\epsilon})}} \nonumber\\
&= a_{1} Q^{-1}(\epsilon_1) +a_{2} Q^{-1}(\epsilon_2) + a_{3} Q^{-1}(\epsilon_3). \label{eq:obj_func}
\end{IEEEeqnarray}
This is because $\bar T_{\text{fixed}}^{(V,W,\vect{\epsilon})}$ and $\beta\bar P_{\text{fixed}}^{(V,W,\vect{\epsilon})}$ are evaluated using $N(k,\varepsilon)$.
To show convexity of the optimization problem \eqref{eq:general_prot_opt}, we use the substitution $\epsilon_i = 1-\ee{u_i}$ for $u_i \leq 0$ and $i\in\{1,2,3\}$, which yields the equivalent constraint $u_1+u_2+u_3=\log\left(\prod_{k=1}^3 (1-\epsilon_k)\right) \leq \log(1-\epsilon)$ which is linear. Consequently, it is sufficient to show that $Q^{-1}(1-\ee{u_i})$ is convex for $i\in\{1,2,3\}$. This follows because the logarithm of the cumulative distribution function of the Gaussian distribution $f(x)\triangleq \log(1-Q(x))$ is concave and increasing. Thus, its inverse function $f^{-1}(x) = Q^{-1}(1-\ee{x})$ is convex and increasing.
\end{IEEEproof}
At this point, we have not discussed the possibility of undetected errors. Approximations like \eqref{eq:fbl} do not give any guarentee for the probability of detecting an error. Using CRCs, the probability of undetected error can be made arbitrarily small, but it is always positive and less than or equal $\epsilon$. Suppose that decoding of the first packet, containing control information, fails for the $k$-th user. In this case, the subsequent behavior is random, and the $k$-th user will (with high probability) not correctly decode the following packets. However, since the packet sizes are limited by $\alpha_S$, we can compute the worst-case power consumption at the users, say $P_{\text{worst}}$. We then cope with the problem of undetected errors simply by adding, to the power consumption at each user, the term $\epsilon P_{\text{worst}}$, which corresponds to the worst-case contribution to the power consumption.
\subsection{Variable message size}
\label{sec:varying}
\begin{figure}[!t]
\centering
\includegraphics[width=3.4in]{protocol_new.pdf}
\caption{An example of the protocol in Section~\ref{sec:varying} with $S= 2$, $K=40$, $W=10$, and $\vect{V} = [3,2]$. Packets surrounded by black separators corresponds to an encoded packet. Grey separators means ``encoded jointly'', e.g., the messages $\mathrm{M}_{12}$ and $\mathrm{M}_{15}$ are jointly encoded in one packet. The red shaded parts of the protocol depicts the packets that the users $12$ and $15$ needs to decode.}
\label{fig:protocol}
\end{figure}
Next, we consider the case $S \geq 2$.
The users are grouped into $B\triangleq \lceil K/W\rceil$ UGs in the same way as for the fixed message size protocol. The UGs are encoded sequentially after the control information of the first layer. The control information of the first layer consists of pointers to the time indices of the beginning of each UG. Thus, based on the control information of the first layer, each user can identify the location of its UG. Note that we need only $B-1$ pointer because the first UG is transmitted immediately after the control information. Each pointer is encoded separately in a packet using an optimal code with an error probability not exceeding $\epsilon_1$. Observe that one can compute the maximum length (in channel uses) of each UG and thereby the number of bits required for each pointer.
\begin{comment}The following sections describe how the UGs are encoded for the two protocols.
\subsubsection{Protocol 1: small amount of control information}
The next packet jointly encodes which users belong to each category, represented using $\Big\lceil \log_2 { {|\mathcal{K}_i|} \choose {\vect{a}_0, \vect{a}_1, \cdots, \vect{a}_S}} \Big\rceil$ bits, along with the messages of all users in $\mathcal{K}_i$, i.e., $\left\{\bigcup_{k\in\mathcal{K}_i} \mathrm{M}_k\right\}$. Note that some of the messages are empty. This packet is encoded using an optimal code with error probability not exceeding $\epsilon_3$. An example of the protocol is depicted in Fig.~\ref{fig:protocol}.
We denote $\E{T}$ and $\E{P_1}$ by $\bar T_{\text{prot1}}^{W,\vect{\epsilon}}$ and $\bar P_{\text{prot1}}^{W,\vect{\epsilon}}$, respectively, and optimize the parameters of the protocol using the optimization problem
\begin{IEEEeqnarray}{rCl}
\min_{W\in\mathcal{K}} \min_{\substack{\vect{\epsilon}\in[0,1]^3:\\ \prod_{k=1}^3 (1-\epsilon_k) \geq 1-\epsilon}} \bar T_{\text{prot1}}^{W,\vect{\epsilon}} + \beta \bar P_{\text{prot1}}^{W,\vect{\epsilon}}
\end{IEEEeqnarray}
for all $\beta\geq 0$.
\end{comment}
The control information of the second layer for the $i$-th UG consists of $\{D_k\}_{k\in\mathcal{K}_i}$, represented by $\lceil |\mathcal{K}_i| \log_2 (S+1) \rceil$ bits. These bits are transmitted using an optimal code with error probability not exceeding $\epsilon_2$.
\begin{comment}
of a vector with the number of users in $\mathcal{K}_i$ having $D_k=0$, $D_k=\alpha_1, \cdots, D_k = \alpha_S$, i.e.,
\begin{multline}
\vect{a}\triangleq \big[ \mathcal{K}_i\cap \{k: D_k = 0\}, \mathcal{K}_i\cap \{k: D_k = \alpha_1\}\\,\cdots,\mathcal{K}_i\cap \{k: D_k = \alpha_S\} \big].\label{eq:vecta}
\end{multline}
Here, we denote the entries of $\vect{a}$ by $\vect{a}_0,\cdots,\vect{a}_S$.
This vector is multinomial distributed with $S+1$ categories, $|\mathcal{K}_i|$ trials, and event probabilities $[q, (1-q)p_1, \cdots,(1-q)p_S]$. This vector can take $\multiset{|\mathcal{K}_i|}{S+1}$ distinct values, where $\multiset{n}{k}$ denotes the number of multisets of cardinality $k$ that can be drawn from a set of cardinality $n$. This vector is conveyed using an optimal code with error probability not exceeding $\epsilon_2$.
\end{comment}
Finally, sequentially for each $s\in\{1,\cdots,S\}$, the transmitter encodes the messages of the users $\mathcal{U}_i^{(s)}$ using an $(\mathcal{U}_i^{(s)},V_s,\epsilon_3)$-protocol, where $\vect{V} = [V_1,\cdots,V_S]$ are protocol parameters to be specified. The protocol is illustrated in Fig.~\ref{fig:protocol}.
We denote $\E{T}$ and $\E{P_1}$ by $\bar T_{\text{variable}}^{(\vect{V},W,\vect{\epsilon})}$ and $\bar P_{\text{variable}}^{(\vect{V},W,\vect{\epsilon})}$, respectively, and optimize the parameters of the protocol using the optimization problem
\begin{IEEEeqnarray}{rCl}
\min_{(\vect{V},W)\in\mathcal{K}^{S+1}} \min_{\substack{\vect{\epsilon}\in[0,1]^3:\\ \prod_{k=1}^3 (1-\epsilon_k) \geq 1- \epsilon} } \bar T_{\text{variable}}^{(\vect{V}, W,\vect{\epsilon})} +\beta \bar P_{\text{variable}}^{(\vect{V}, W,\vect{\epsilon})}.
\end{IEEEeqnarray}
As for the fixed message size protocol, the inner minimization is convex.
\section{Numerical Results}\label{sec:numerical}
In this section, we plot the lower bound along with the optimal achievable trade-offs for the proposed protocols. All results are for $\epsilon=10^{-4}$, $P=0\ \text{dB}$, $q=0.5$.
We first present results for the case with fixed message size. Fig.~\ref{fig:trade_off_16_100} and Fig.~\ref{fig:trade_off_128_100} show the trade-offs for $\alpha_1=100$ and $S=1$ for $K=16$ and $K=128$, respectively. We plot the lower bound given by Proposition~\ref{prop:prop_converse} and the trade-offs achievable by the genie-aided protocol and the fixed message size protocol. For the fixed message size protocol, we also plot the trade-off for the case where the inner minimization in \eqref{eq:general_prot_opt} is not performed and $\epsilon_1, \epsilon_2, \epsilon_3$ are set equally to $1-(1-\epsilon)^{1/3}$. For the protocols, we plot the lower convex envelopes and note that any point on them can be achieved by time-sharing between two sets of protocol parameters. We observe, as expected, that differences between the genie-aided protocols and the lower bounds are negligible. Optimizing over $\vect{\epsilon}$ also improves the trade-off slightly. This happens because the control information which is destined to many users needs better protection compared to a group of messages destined only to a group of users. Finally, we observe a significant gap between the genie-aided protocol and the fixed message size protocol which reflects the significance of control information for broadcast of small messages. Fig.~\ref{fig:trade_off_16_1000} and Fig.~\ref{fig:trade_off_128_1000} shows the trade-offs for $\alpha=1000$. In this case, we see that the gap between the genie-aided protocol and the fixed message size protocol becomes less significant.
Finally, in Fig.~\ref{fig:trade_off_two_mess_16_100} and Fig.~\ref{fig:trade_off_two_mess_16_1000}, we depict the trade-offs for $K=16$, $\vect{p}=[0.5,0.5]$ and with $\vect{\alpha} = [50,150]$ and $\vect{\alpha}=[500,1500]$, respectively. Our observations are similar to those for the fixed message size protocol.
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=1,
xmin=1800,
ymax=1100,
xmax=2450,
xlabel={Average frame duration [ch. uses]},
ylabel={Average power [ch. uses]},
legend entries = { {}{Protocol (no $\vect{\epsilon}$ opt.)}, {}{Protocol}, {}{Genie-aided protocol}, {}{Lower bound}}
]
\addplot [
color=green,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha100_fixedconcat.dat};
\addplot [
color=magenta,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha100_fixedopt.dat};
\addplot [
color=red, mark=*, line width=1.0pt,mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha100_genie.dat};
\addplot [line width=1.0pt,color=black, solid] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha100_lower.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average transmission time and average power consumption for the case $K=16$, $P=1$, $q=0.5$, $\alpha = 100$, $S=1$, and $\epsilon=10^{-4}$. Here, ``Protocol'' refers to the fixed message size protocol, while ``Protocol (no $\vect{\epsilon}$ opt.)'' refers to the fixed message size protocol with $\epsilon_1=\epsilon_2 = \epsilon_3 = 1-(1-\epsilon)^{1/3})$.}
\label{fig:trade_off_16_100}
\end{figure}
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=0,
xmin=13500,
ymax=4000,
xmax=19500,
xlabel={Average frame duration [ch. uses]},
ylabel={Average power [ch. uses]},
legend entries = { {}{Protocol (no $\vect{\epsilon}$ opt.)}, {}{Protocol}, {}{Genie-aided protocol}, {}{Lower bound}}
]
\addplot [
color=green,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha100_fixedconcat.dat};
\addplot [
color=magenta,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha100_fixedopt.dat};
\addplot [
color=red, mark=*, line width=1.0pt,mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha100_genie.dat};
\addplot [line width=1.0pt,color=black, solid] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha100_lower.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average transmission time and average power consumption for the case $K=128$, $P=1$, $q=0.5$, $\alpha = 100$, $S=1$, and $\epsilon=10^{-4}$. Here, ``Protocol'' refers to the fixed message size protocol, while ``Protocol (no $\vect{\epsilon}$ opt.)'' refers to the fixed message size protocol with $\epsilon_1=\epsilon_2 = \epsilon_3 = 1-(1-\epsilon)^{1/3})$.}
\label{fig:trade_off_128_100}
\end{figure}
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=0,
xmin=16700,
ymax=9000,
xmax=18500,
xlabel={Average frame duration [ch. uses]},
ylabel={Average power [ch. uses]},
legend entries = { {}{Protocol (no $\vect{\epsilon}$ opt.)}, {}{Protocol}, {}{Genie-aided protocol}, {}{Lower bound}}
]
\addplot [
color=green,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha1000_fixedconcat.dat};
\addplot [
color=magenta,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha1000_fixedopt.dat};
\addplot [
color=red, mark=*, line width=1.0pt,mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha1000_genie.dat};
\addplot [line width=1.0pt,color=black, solid] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K16_P1.0_q0.50_alpha1000_lower.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average transmission time and average power consumption for the case $K=16$, $P=1$, $q=0.5$, $\alpha = 1000$, $S=1$, and $\epsilon=10^{-4}$. Here, ``Protocol'' refers to the fixed message size protocol, while ``Protocol (no $\vect{\epsilon}$ opt.)'' refers to the fixed message size protocol with $\epsilon_1=\epsilon_2 = \epsilon_3 = 1-(1-\epsilon)^{1/3})$.}
\label{fig:trade_off_16_1000}
\end{figure}
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=0,
xmin=130000,
ymax=35000,
xmax=150000,
xlabel={Average frame duration [ch. uses]},
ylabel={Average power [ch. uses]},
legend entries = { {}{Protocol (no $\vect{\epsilon}$ opt.)}, {}{Protocol}, {}{Genie-aided protocol}, {}{Lower bound}}
]
\addplot [
color=green,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha1000_fixedconcat.dat};
\addplot [
color=magenta,
mark=*,
line width=1.0pt,
mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha1000_fixedopt.dat};
\addplot [
color=red, mark=*, line width=1.0pt,mark size=1pt] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha1000_genie.dat};
\addplot [line width=1.0pt,color=black, solid] table [x index = {0}, y index = {1}, col sep=comma]{fixed_plot_K128_P1.0_q0.50_alpha1000_lower.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average transmission time and average power consumption for the case $K=128$, $P=1$, $q=0.5$, $\alpha = 1000$, $S=1$, and $\epsilon=10^{-4}$. Here, ``Protocol'' refers to the fixed message size protocol, while ``Protocol (no $\vect{\epsilon}$ opt.)'' refers to the fixed message size protocol with $\epsilon_1=\epsilon_2 = \epsilon_3 = 1-(1-\epsilon)^{1/3})$.}
\label{fig:trade_off_128_1000}
\end{figure}
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=100,
xmin=1850,
ymax=600,
xmax=2620,
xlabel={Average frame duration [ch. uses]},
ylabel={Average power [ch. uses]},
legend entries = { {}{Protocol ($S\geq 2$)}, {}{Genie-aided protocol}, {}{Lower bound}}
]
\addplot [color=magenta, mark=*, line width=1.0pt, mark size=1pt]
table [x index = {0}, y index = {1}, col sep=comma]{two_mess_plot_K16_P1.0_q0.50_alpha50_150_opt.dat};
\addplot [color=red, mark=*, line width=1.0pt,mark size=1pt]
table [x index = {0}, y index = {1}, col sep=comma]{two_mess_plot_K16_P1.0_q0.50_alpha50_150_genie.dat};
\addplot [line width=1.0pt,color=black, solid]
table [x index = {0}, y index = {1}, col sep=comma]{two_mess_plot_K16_P1.0_q0.50_alpha50_150_lower.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average transmission time and average power consumption for the case $K=16$, $P=1$, $q=0.5$, $\vect{p}=[0.5,0.5]$, $\vect{\alpha} = [50,150]$, $S=2$, and $\epsilon=10^{-4}$.}
\label{fig:trade_off_two_mess_16_100}
\end{figure}
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=1000,
xmin=17000,
ymax=5100,
xmax=18600,
xlabel={Average frame duration [ch. uses]},
ylabel={Average power [ch. uses]},
legend entries = { {}{Protocol ($S\geq 2$)}, {}{Genie-aided protocol}, {}{Lower bound}}
]
\addplot [color=magenta, mark=*, line width=1.0pt, mark size=1pt]
table [x index = {0}, y index = {1}, col sep=comma]{two_mess_plot_K16_P1.0_q0.50_alpha500_1500_opt.dat};
\addplot [color=red, mark=*, line width=1.0pt,mark size=1pt]
table [x index = {0}, y index = {1}, col sep=comma]{two_mess_plot_K16_P1.0_q0.50_alpha500_1500_genie.dat};
\addplot [line width=1.0pt,color=black, solid]
table [x index = {0}, y index = {1}, col sep=comma]{two_mess_plot_K16_P1.0_q0.50_alpha500_1500_lower.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average transmission time and average power consumption for the case $K=16$, $P=1$, $q=0.5$, $\vect{p}=[0.5,0.5]$, $\vect{\alpha} = [500,1500]$, $S=2$, and $\epsilon=10^{-4}$.}
\label{fig:trade_off_two_mess_16_1000}
\end{figure}
\section{Conclusion}\label{sec:conclusions}
In this paper, we considered the AWGN broadcast channel with $K$ users with symmetric channel conditions. The downlink transmission is organized in frames. In each frame, a message of random size (in bits) is destined to each of the users in such a way that the message sizes are unknown to the users. The message can also be of size zero, which means the user should not receive data in that frame. A user, however, still needs to decode a certain amount of information from the frame in order to learn that there is no data destined to her in this particular frame. Hence, in addition to the messages, a protocol needs to convey control information that describes the structure of the transmission and the sizes of the messages. We used approximations of the maximum coding rate for the AWGN channel from finite blocklength information theory to show that jointly encoding different groupings of the messages enable the protocol designer to trade-off between average frame duration and the average power consumption at the users. Specifically, we derived a lower bound for the trade-off curve which assumed that control information was available at the users, a genie-aided protocol, and two practical protocols. Our numerical results showed that the genie-aided protocol achieved a trade-off curve that closely matched the lower bound. For both of our practical protocols, the control information led to a significantly worse trade-off curves when the messages were small and when compared to the genie-aided protocol. There are several directions for future research:
\begin{enumerate}
\item In Section~\ref{sec:system_model}, we significantly restricted our general system model to a space of practical and tractable protocols. A rigorous information-theoretic treatment of our general system model might lead to improved protocols and lower bounds.
\item The system model has two obvious extensions: one can extend the system model to include fading, and one can introduce asymmetric channel conditions using results from \cite{Yang2014}.
\item While we are able to quantify the suboptimality of our protocols by comparison to the lower bound, our protocols are still heuristic. One interesting idea for future research is to systematically investigate the design of good protocols that include control information.
\end{enumerate}
\begin{comment}
\section{Protocol design}\label{sec:protocol_design}
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{protocol.pdf}
\caption{Proposed protocol for $K=30$, $L_B=10$, and $L = 2$. In this case $\{11,12,15,16\}\subseteq \mathcal{U}$ are among the active users. Grey separators means that data on both sides are encoded jointly. The red shaded regions correspond to the packets that the users $15$ and $16$ need to decode.}
\label{fig:protocol}
\end{figure}
There are various ways in which the messages $\{M_k\}$ can be conveyed to the respective users. Our approach is to design a protocol in which the transmitter forms multiple packets which are encoded separately. For each of these packets, we apply the finite blocklength approximation in \eqref{eq:fbl} to find the optimal rate at which they can be encoded.
We assume that the users are not provided with any control information such as the active users and $\{D_k\}$. Thus, the transmitter needs to encode packets about which users are active, the packet sizes of $\mathrm{M}_k$, $D_k$, and the structure of the transmission. Clearly, this leaves us with a large space of feasible protocols. Here we introduce one class of protocols.
We first discuss what information, the transmitter needs to convey:
\begin{enumerate}
\item \emph{Messages} $\{\mathrm{M}_k\}$: The message $\mathrm{M}_k$ needs only to be received by the $k$-th user, but as discussed previously, messages can be grouped and encoded jointly.
\item \emph{Message sizes} $\{D_k\}$: The $k$-th user needs to know the message size $D_k$ before attempting to decode the actual message $\mathrm{M}_k$ (otherwise, the user does not know how many channel uses the message $\mathrm{M}_k$ takes).
\item \emph{Reciever activity} $\mathcal{U}$: It is necessary to convey whether the $k$-th user is active. In total, it requires $K$ information bits to convey this information to all users.\footnote{For the case $q\not= 1/2$, one can apply compression to reduce the number of information bits. This is, however, left for future work.} As $K$ information bits may represent a significant overhead, it may be beneficial to encode user activity bits in multiple packets such that each user needs only to decode one such packet.
\end{enumerate}
In the proposed protocol, depicted in Fig.~\ref{fig:protocol}, users are grouped into $\lceil K/L_B\rceil$ \emph{user groups} with at most $L_B$ users in each user group. User acitivity, messages, and message sizes associated to each of these user groups are conveyed sequentially in \emph{user group frames} (UGF).
A transmission is initiated by a packet that jointly encodes the frame duration (equivalent to the an end of transmission pointer) along with the $\lceil K/L_B\rceil-1$ time indices that points to the time indices where the $2$-th, $3$-th, ..., and $\lceil K/L_B\rceil$-th UGF begin. This packet is transmitted with a reliability $1-\epsilon_4$. The first UGF trivially begins after the initiating packet.
Let $\mathcal{K}\triangleq\{1,\cdots,K\}$ and let the users in the $u$-th user group be $\mathcal{K}_u \subseteq \mathcal{K}$. Then, the UGF for the $u$-th user group is constructed as follows. Initially, the transmitter divides the active users $\mathcal{ U}_u\subseteq \mathcal{K}_u$ of the $u$-th user group into subgroups $\mathcal{U}_{u,i}\subseteq \mathcal{ U}_u$, $i\in\{1,\cdots,\lceil |\mathcal{ U}_u|/L\rceil\}$ of at most $L$ users. The transmitter and users can agree on how to partition the users into subgroups for every set $\mathcal{U}_u$. The set of users $\mathcal{U}_{u,i}\subseteq \mathcal{ U}_u$ is referred to as the $i$-th subgroup of the $u$-th user group.
The main idea of our protocol is to jointly encode each of the subgroups.
A UGF consists of the following types of packets
\begin{enumerate}
\item \emph{Bit field packet}: A bit field, encoding the the information $\{\indi{D_k = 0}\}_{k\in\mathcal{K}_{u}}$. Hence, the packet consists of $|\mathcal{K}_u|$ information bits which are encoded with reliability $\epsilon_{1}$.
\item \emph{Size packets}: After grouping the active users of the $u$-th user group, $\mathcal{U}_u$, into $\lceil |\mathcal{U}_u|/L\rceil$ subgroups, the transmitter constructs a packet for each subgroup. For the $i$-th subgroup, the transmitter conveys a packet consisting of $\sum_{k\in\mathcal{U}_{u,i}} D_k$ along with a pointer to the packet that jointly encodes $\{\mathrm{M}_k\}_{k\in\mathcal{U}_{u,i}}$. Since $\sum_{k\in\mathcal{U}_{u,i}} D_k$ can take at most $L(S-1)+1$ distinct values, the size packet for the $i$-th subgroup needs to convey $ \lceil \log_2 (L (S-1) +1) \rceil + \text{ptr}$ information bits which are encoded with reliablity $\epsilon_2$. Here, $\text{ptr}$ denotes the number of bits needed to convey a pointer to a time index. The size packets are transmitted sequentially.
\item \emph{Message packets}: Next, the transmitter encodes the messages of each subgroup, $\{\mathrm{M}_k\}_{k\in\mathcal{U}_{u,i}}$, along with the messages sizes of $|\mathcal{U}_{u,i}|-1$ of the messages. We need only $|\mathcal{U}_{u,i}|-1$, since the sum of the sizes, $\sum_{k\in\mathcal{U}_{u,i}} D_k$, is already successfully received in the size packet described above. This requires $\sum_{k\in\mathcal{U}_i} D_k + (|\mathcal{U}_i|-1)\lceil \log_2 S \rceil$ information bits. These information bits are encoded with reliability $\epsilon_3$.
\end{enumerate}
In order to decode the packet destined to user $k$, it needs to decode four packets successfully. If one or more of these packets are not successfully decoded, the user can not decode the packet containing the message destined to that user. Thus, the reliabilities need to be chosen such that $(1-\epsilon_1)(1-\epsilon_2)(1-\epsilon_3)(1-\epsilon_4)$ is kept above or equal to $1-\epsilon$ to fulfill the reliability constraint in \eqref{eq:reliability}. If the $k$-th user is inactive, it needs only to decode the initial packet containing pointers to the UGFs and the bit field packet. It thereby achieves a reliability of $(1-\epsilon_4)(1-\epsilon_1)$.
We also point out that the described protocol reduces to a variant of the conventional protocol when $L = 1$.
We remark that the protocol specified above is one class among a large space of feasible protocols. For small $q$ is may be beneficial to use a different approach for conveying user activity. For example, one could encode the number of active users in an initial packet and encode an additional packet with the user identification numbers. Regarding the size packets, one can also encode all size packets jointly in each UGF jointly to enhance encoding efficiency at the expense of higher average power consumption at the users.
Assuming that $L_B$ divides $K$, we may sum up the blocklengths of all the packets, the average frame duration becomes
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{T_{L,L_B}}\nonumber\\
&=& \frac{K}{L_B}N\farg{L_B,\epsilon_1}+N\farg{\frac{K}{L_B} \mathrm{ptr}, \epsilon_4}\nonumber\\
&& {}+\frac{K}{L_B}\EBigg{\sum_{i=1}^{\lceil| \mathcal{U}_1|/L\rceil}\bigg(N\farg{ \lceil\log_2(L(S-1)+1) \rceil + \mathrm{ptr},\epsilon_2}\nonumber\\
&& \qquad\quad {}+ N\mathopen{}\bigg(\sum_{k\in\mathcal{U}_1} D_k + (|\mathcal{U}_1|-1)\lceil \log_2 S \rceil,\epsilon_3 \bigg) \bigg)}.
\end{IEEEeqnarray}
For the average power, we obtain
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{P_{L,L_B}}\nonumber\\
&=& N\farg{L_B,\epsilon_1}+N\farg{\frac{K}{L_B}\mathrm{ptr}, \epsilon_4}\nonumber\\
&&+\frac{1}{L_B}\EBigg{\sum_{i=1}^{\lceil| \mathcal{U}_1|/L\rceil}\bigg(N\farg{ \lceil \log_2(L(S-1)+1)\rceil + \mathrm{ptr},\epsilon_2}\nonumber\\
&& \qquad {}+ N\mathopen{}\bigg(\sum_{k\in\mathcal{U}_1} D_k + (|\mathcal{U}_1|-1)\lceil \log_2 S \rceil,\epsilon_3 \bigg) \bigg)}.
\end{IEEEeqnarray}
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ytick={1,5,10,15,20,25,30},
yticklabels={$1$,$5$,$10$,$15$,$20$,$25$,$30$},
ymin=1,
xmin=1,
ymax=25,
xmax=1.12,
xlabel={Average frame duration (normalized)},
ylabel={Average power (normalized)}
]
\addplot [
color=black,
solid,
mark=.,
mark options={fill=white},
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{broadcast_plot_K64_q0.50_sm1000_sb_1.dat};
\addplot [
color=red,
mark=*,
mark size=0.2pt,
only marks,
] table [x index = {0}, y index = {1}, col sep=comma]{broadcast_plot_points_K64_q0.50_sm1000_sb_1.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average transmission time and average power for the case $K=64$, $q=0.5$, $\alpha = 1000$, and $S=2$. Red dots are simulation points.}
\label{fig:simulation_points}
\end{figure}
\begin{figure}[!t]
\centering
\begin{tikzpicture}
\begin{axis}[
/pgf/number format/.cd, 1000 sep={},
grid=both,
tick align=center,
width=0.4\paperwidth,
height=0.35\paperwidth,
ymin=1,
ytick={1,5,10,15,20,25,30},
yticklabels={$1$,$5$,$10$,$15$,$20$,$25$,$30$},
xmin=1,
xmax=1.9,
ymax=30,
xlabel={Average frame duration (normalized)},
ylabel={Average power (normalized)},
legend entries={$\alpha=50$,$\alpha=100$,$\alpha=500$,$\alpha=1000$}
]
\addplot [
color=black,
solid,
mark=.,
mark options={fill=white},
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{broadcast_plot_K128_q0.50_sm50_sb_2.dat};
\addplot [
color=red,
solid,
mark=.,
mark options={fill=white},
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{broadcast_plot_K128_q0.50_sm100_sb_2.dat};
\addplot [
color=blue,
solid,
mark=.,
mark options={fill=white},
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{broadcast_plot_K128_q0.50_sm500_sb_2.dat};
\addplot [
color=green,
solid,
mark=.,
mark options={fill=white},
line width=1.0pt
] table [x index = {0}, y index = {1}, col sep=comma]{broadcast_plot_K128_q0.50_sm1000_sb_2.dat};
\end{axis}
\end{tikzpicture}
\caption{Trade-off between average frame duration and average power for the parameters are $N=128$, $q=0.5$ and $S=4$.}
\label{fig:res1}
\end{figure}
The specified protocol leaves the parameters $\epsilon_1$, $\epsilon_2$, $\epsilon_3$, $\epsilon_4$, $L$, and $L_B$ to be specified.
Now, we can trace the optimal trade-off between $T_{L,L_B}$ and $P_{L,L_B}$ by solving the optimization problem
\begin{IEEEeqnarray}{rCl}
\min_{\substack{L,L_B,\epsilon_1,\epsilon_2,\epsilon_3,\epsilon_4:\\ \prod_{j=1}^4 (1-\epsilon_j) \geq 1-\epsilon}} \E{T_{L,L_B}} + \beta \E{P_{L,L_B}}.\label{eq:opt_problem}
\end{IEEEeqnarray}
for a range of values of $\beta\geq 0$. The optimization problem is clearly not convex, and hence we find an approximate solution in the next section using a grid search for practical values of $K$, $\epsilon$, $q$, and $P_D$.
\section{Numerical results}
\label{sec:numerical}
In order to solve the optimization problem in \eqref{eq:opt_problem}, we compute $\E{T_{L,L_B}}$ and $\E{P_{L,L_B}}$ using $5000$ Monte Carlo simulations of the protocol for $L \in \{1,2,\cdots, K\}$ and $L_B$ equal to all powers of two between $1$ and $K$. We evaluate $\epsilon_1,\cdots,\epsilon_4$ over the four-dimensional grid $10\times 10\times 10\times 10$ grid. The average frame duration $\E{T_{L,L_B}}$ and average power $\E{P_{L,L_B}}$ are normalized according to the lower bounds in \eqref{eq:Tbound} and \eqref{eq:Pbound}, respectively. The normalization implies that any simulation point must be in the square $[1,\infty)\times [1,\infty)$. The trade-off between average total transmission and average power is computed as the lower convex envelope of the simulation points. This is depicted in Fig.~\ref{fig:simulation_points}, where the simulations points are shown as red dots and the lower convex envelope is the black curve. For the computation, we use $\mathrm{ptr}=16\ \mathrm{bits}$. Although the lower convex envelope is not directly achievable using our protocol, it can be achieved by time sharing between two sets of protocol parameters. Note that the lower-most point of the trade-off curve corresponds to the conventional extreme case where the messages of each user are encoded separately. The gap to $1$ is thus due overhead from control information.
Our results are depicted in Fig.~\ref{fig:res1} for the parameters $K=128$, $q=0.5$, $S = 4$, and $\alpha\in\{50,100,500,1000\}$. We observe that one can reduce the average frame duration by grouping users as proposed. Smaller values of $\alpha$ implies that messages are encoded less efficient, and hence grouping becomes an interesting option.
\section{Conclusions}
\label{sec:conclusions}
In this paper, we have addressed the problem of downlink transmission of short packets to $K$ users. Our main objective has been to highlight some of the challenges faced when the messages are small. Specifically, we used recent finite blocklength approximations to visualize the trade-offs between the average power of the each user and the average frame duration seen from the transmitter. To show this trade-off, we have designed a practical protocol that groups messages and thereby achieves more efficient coding rates. The key element in the protocol design is the encoding of control information.
\end{comment}
\appendices
\section{Proof of Lemma~\ref{lem:breve_phi}}\label{app:lem_proof}
By definition of the lower convex envelope, for every $\vect{x}\in\mathbb{R}_+^S$, there exists a vector $\vect{\nu}\in\mathbb{R}_+^S$ with $\vect{1}_S\tr \vect{\nu}=1$ and $I$ points $\vect{a}_i\in\mathbb{R}_+^S$, for $i\in\{1,\cdots,I\}$, such that
\begin{IEEEeqnarray}{rCl}
\breve{\phi}_{\beta}(\vect{x}) = \sum_{i=1}^{I} \nu_i \phi_\beta(\vect{a}_i)\label{eq:remark_convex_combination}
\end{IEEEeqnarray}
and such that
\begin{IEEEeqnarray}{rCl}
\vect{x} = \sum_{i=1}^I \nu_i \vect{a}_i.\label{eq:sum_to_x}
\end{IEEEeqnarray}
Since $N(n,\epsilon)$ is concave in $n$, we have that $\phi_\beta(\vect{x})$ is concave on the simplex $\mathcal{A}_{\kappa}\triangleq \{\vect{x}\in\mathbb{R}^S_+: \vect{1}_{S}\tr \vect{x} = \kappa\}$ for every $\kappa\in\mathbb{R}_+$. Consequently, for $i\in\{1,\cdots,I\}$, we have
\begin{IEEEeqnarray}{rCl}
\phi_\beta(\vect{a}_i) &=&\phi_\beta\farg{ \sum_{s=1}^S \frac{a_{i,s}}{\vect{1}_S\tr \vect{a}_i} \vect{\bar 0}^S_s\mathopen{}\left(\vect{1}_S\tr \vect{a}_i \right)} \label{eq:ineqm1}\\
&\geq& \sum_{s=1}^S \frac{a_{i,s}}{\vect{1}_S\tr \vect{a}_i}\phi_\beta\farg{ \vect{\bar 0}^S_s\mathopen{}\left(\vect{1}_S\tr \vect{a}_i\right)}\label{eq:ineq0}\\
&=& \sum_{s=1}^S \frac{a_{i,s}}{\vect{1}_S\tr \vect{a}_i}\phi^{(s)}_\beta\farg{\vect{1}_S\tr \vect{a}_i }.\label{eq:ineq1}
\end{IEEEeqnarray}
In \eqref{eq:ineqm1}, $\vect{0}_s^S(x)$ denotes an $S$-dimensional vector with $x$ in the $s$-th entry and zeroes in the rest, \eqref{eq:ineq0} follows by Jensen's inequality (concave) applied to $\phi_\beta(\vect{x})$ on the simplex $\mathcal{A}_{\vect{1}_S\tr \vect{a}_i}$, and \eqref{eq:ineq1} is by the definition of $\phi_{\beta}^{(s)}(\cdot)$ in \eqref{eq:phis_def}.
bWe can now lower-bound \eqref{eq:remark_convex_combination} as
\begin{IEEEeqnarray}{rCl}
\breve{\phi}_{\beta}(\vect{x}) &\geq& \sum_{s=1}^S \sum_{i=1}^{I} \frac{\nu_i a_{i,s}}{\vect{1}_S\tr \vect{a}_i} \phi^{(s)}_\beta\farg{\vect{1}_S\tr \vect{a}_i }\label{eq:ineq2}\\
&\geq& \sum_{s=1}^S \sum_{i=1}^{I} \frac{\nu_i a_{i,s}}{\vect{1}_S\tr \vect{a}_i} \breve{\phi}^{(s)}_\beta\farg{\vect{1}_S\tr \vect{a}_i }\label{eq:ineq4}\\
&\geq& \sum_{s=1}^S \left(\sum_{i=1}^{I} \frac{\nu_i a_{i,s}}{\vect{1}_S\tr \vect{a}_i} \right) \breve{\phi}^{(s)}_\beta\farg{ \frac{\sum_{i=1}^I\frac{\nu_i a_{i,s}}{\vect{1}_S\tr \vect{a}_i} \vect{1}_S\tr \vect{a}_i }{\sum_{i=1}^I \frac{\nu_i a_{i,s}}{\vect{1}_S\tr \vect{a}_i}} }\label{eq:ineq3}\\
&=& \sum_{s=1}^S \zeta_s \breve{\phi}^{(s)}_\beta\farg{x_s/\zeta_s}.\label{eq:ineq5}
\end{IEEEeqnarray}
Here, \eqref{eq:ineq2} is by \eqref{eq:remark_convex_combination} and \eqref{eq:ineq1}, \eqref{eq:ineq4} is by $\phi^{(s)}_\beta(x) \geq \breve{\phi}^{(s)}_\beta(x)$ for $x\geq 0$, \eqref{eq:ineq3} follows by Jensen's inequality (convex) applied to $\breve{\phi}^{(s)}_\beta(\cdot)$, and \eqref{eq:ineq5} follows by setting $\zeta_s \triangleq \sum_{i=1}^{I} \frac{\nu_i a_{i,s}}{\vect{1}_S\tr \vect{a}_i}$ and by using $\sum_{i=1}^I\nu_i a_{i,s} = x_s$ by \eqref{eq:sum_to_x}. Thus, we have shown that the LHS of \eqref{eq:breve_simple_comp_claim} is larger than or equal the RHS of \eqref{eq:breve_simple_comp_claim}.
Next, we establish the equality in \eqref{eq:breve_simple_comp_claim}. Suppose, on the contrary, that there exists a positive vector $\vect{\bar \zeta}\in\mathbb{R}^S$ such that $\vect{1}_S\tr \vect{\zeta} = 1$ and such that $\breve{\phi}_\beta(\vect{x}) > \sum_{s=1}^S \bar\zeta_s \breve{\phi}_\beta^{(s)}(x_s/\bar \zeta_s)$. This implies a contradiction:
\begin{IEEEeqnarray}{rCl}
\breve{\phi}_\beta(\vect{x}) &>& \sum_{s=1}^S \bar\zeta_s \breve{\phi}_\beta^{(s)}(x_s/\bar \zeta_s) \\
&=& \sum_{s=1}^S \bar\zeta_s \breve{\phi}_\beta(\vect{\bar 0}_s^S(x_s/\zeta_s)) \\
&\geq& \breve{\phi}_\beta(\vect{x}).\label{eq:contradiction}
\end{IEEEeqnarray}
Here, \eqref{eq:contradiction} follows by Jensen's inequality (convex) applied to $\breve{\phi}_\beta(\cdot)$. We conclude that \eqref{eq:breve_simple_comp_claim} must be satisfied with equality. Note that it is sufficient to write minimum instead of infimum in \eqref{eq:breve_simple_comp_claim} because we have shown the existence of a feasible point in \eqref{eq:breve_simple_comp_claim} that attains the minimum.
To show convexity of the optimization problem in \eqref{eq:breve_simple_comp_claim}, it is sufficient to show that the function $x \breve{\phi}^{(s)}(y/x)$ is convex in $x>0$ for a constant $y>0$ and $s\in\{1,\cdots,S\}$, i.e., for every $x_1>x_2>0$ and $\alpha \in[0,1]$, we need to show that
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\alpha x_1 \breve{\phi}^{(s)}(y/x_1) + (1-\alpha)x_2 \breve{\phi}^{(s)}(y/x_2)}\nonumber\\\quad
&\geq& (\alpha x_1 + (1-\alpha)x_2 )\breve{\phi}^{(s)}( y/(\alpha x_1 + (1-\alpha)x_2) ).\label{eq:convexity}
\end{IEEEeqnarray}
Fix, without loss of generality, arbitrary $x_1 > x_2 > 0$, $\alpha\in[0,1]$, and $s\in\{1,\cdots,S\}$.
Define the function
\begin{IEEEeqnarray}{rCl}
g(x) = \frac{\frac{y}{x_2} -x}{\frac{y}{x_2} - \frac{y}{x_1}} \breve{\phi}^{(s)}\farg{\frac{y}{x_1}} +\frac{x - \frac{y}{x_1}}{\frac{y}{x_2} - \frac{y}{x_1}} \breve{\phi}^{(s)}\farg{\frac{y}{x_2}}.\IEEEeqnarraynumspace
\end{IEEEeqnarray}
Note that $x g(y/x)$ and $g(x)$ are affine functions in $x>0$ and that $g(y/x_1) = \breve{\phi}^{(s)}(y/x_1)$ and $g(y/x_2) = \breve{\phi}^{(s)}(y/x_2)$. Thus, since $\breve{\phi}^{(s)}(\cdot)$ is convex, we have $g(x) \geq \breve{\phi}^{(s)}(x)$ for $x\in[y/x_1, y/x_2]$. To verify \eqref{eq:convexity}, we write
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\alpha x_1 \breve{\phi}^{(s)}(y/x_1) + (1-\alpha)x_2 \breve{\phi}^{(s)}(y/x_2)}\nonumber\\\quad &=& (\alpha x_1 + (1-\alpha)x_2)g( y/(\alpha x_1 + (1-\alpha)x_2) )\IEEEeqnarraynumspace\\
&\geq& (\alpha x_1 + (1-\alpha)x_2)\breve{\phi}^{(s)}( y/(\alpha x_1 + (1-\alpha)x_2) )\label{eq:convexity_argu}
\end{IEEEeqnarray}
This establishes the convexity of the optimization problem in \eqref{eq:breve_simple_comp_claim} because we can redo the above argument for all $x_2>x_1>0$ and $\alpha\in[0,1]$.
\bibliographystyle{IEEEtranTCOM}
|
\section{Introduction}
Investigations of ionosphere and magnetosphere with use of international
network of SuperDARN(Super Dual Auroral Radar Network) radars becomes
one of the basic techniques for analysis of the processes in the high-latitude
ionosphere and magnetosphere today \cite{Ruohoniemi_1988,Ruohoniemi_1989,Greenwald_1995,Chisham_2007,Ruohoniemi_2011}.
Expanding of the network to the mid-latitudes allows to investigate
expanding the polar effects to the mid-latitudes during geomagnetic
disturbances and storms\cite{Baker_2007}. In spite of wide use of
the data the inversion technique for this method - obtaining the parameters
of the ionosphere from the received signals is still under development.
Today taking into account the ionospheric refraction to the scattered
signal characteristics is important theoretical problem \cite{Nasyrov_1991,Uspensky_1994,Uspensky_1994a,Ponomarenko_2009,Gillies_2009,Ponomarenko_2010,Gillies_2011,Spaleta_2015}.
The basis for the backscattering technique, used by the SuperDARN
radars is the radar equation. It relates the shape of the correlation
function of the received signal with spectral density of irregularities
and characteristics of background ionosphere.
Currently the SuperDARN radar equation exists only in very simple
approximations \cite{Schiffler_1996}, which can lead to potential
problems in data interpretation. For example, correct taking into
account the refraction in the evaluation of the basic SuperDARN parameter
- irregularities velocity \cite{Gillies_2011}.
Traditionally, to interpret the SuperDARN measurements there is used
an analogy \cite{Schiffler_1996} with the radar equation in refraction-free
case \cite{Tatarsky_1967,Ishimaru_1999}. Initially, this radar equation
was developed for the case of scatterers that are small compared with
the Fresnel zone radius, and not always valid for irregularities with
the size compared with this radius. In \cite{Berngardt_1999,Berngardt_2000}
there was obtained the radar equation under assumption of smooth spatial changes
of spectral density of irregularities. This refraction-free radar
equation is valid for scattering by large-scale ionospheric irregularities
including irregularities elongated with the Earth magnetic field and
studied by SuperDARN radars. But correct taking into account the refraction
was not done yet.
In the paper we obtained the scalar radar equation for SuperDARN radars
within the first propagation hop, taking into account refraction and
single reflection from the ionosphere.
\section{Initial equations}
\subsection{Ray representation of Green's function}
Traditionally the problems of radiowave propagation and scattering
in smoothly inhomogeneous media are analyzed using Green's function
approach. Radiowave propagation in smoothly inhomogeneous media can
be described using geometrical optics (WKB approximation)\cite{Heading_2013,Kravtsov_1990}.
Geometrooptical approach allows to replace extended source (antenna)
by point source and distribution of the transmitted signal over the
ray exit directions (antenna pattern). The full geometrooptical Green's
function is matrix-valued (see, for example \cite{Kravtsov_1996,Liang_1997}):
\begin{equation}
\begin{array}{c}
\overrightarrow{E}(\overrightarrow{r},\omega)=\widehat{G}(\overrightarrow{R}_{0},\overrightarrow{r},\omega)\overrightarrow{a}(\omega)\\
\widehat{G}(\overrightarrow{R}_{0},\overrightarrow{r},\omega)=\frac{1}{F^{1/2}(\overrightarrow{R}_{0},\overrightarrow{r})}e^{i\frac{\omega}{c}\psi(\overrightarrow{R}_{0},\overrightarrow{r})}\hat{\Gamma}(\overrightarrow{R}_{0},\overrightarrow{r})
\end{array}\label{eq:GreensFunction}
\end{equation}
where $\hat{\Gamma}(\overrightarrow{R}_{0},\overrightarrow{r})$ -
polarization matrix, $F^{1/2}(\overrightarrow{R}_{0},\overrightarrow{r})$
- decrease of the signal with the range, $\psi(\overrightarrow{R}_{0},\overrightarrow{r})$
- geometrooptical eikonal; $\overrightarrow{R}_{0}$ is the position
of the transmitter.
The problem of the formula is impossibility to use it in the tasks
of multipath propagation, when signal has a number of different trajectories
to reach the same scattering point\cite{Kravcov_1980}. To generalize
(\ref{eq:GreensFunction}) for the problem of multipath propagation
lets use the ray coordinate system and analyze the field in a given
point as a result of interference of several geometrooptical rays
with different exit angles. In simpler scalar case (which does not
take into account the polarization of the electromagnetic wave) we
can make the equivalent ray representation for scalar Green's function
$G_{go}(\overrightarrow{R}_{0},\overrightarrow{\Lambda},\omega)$,
that after integrating over the ray coordinates results into standard
scalar Green's function:
\begin{equation}
G(\overrightarrow{R}_{0},\overrightarrow{r}',\omega)=\int G_{go}(\overrightarrow{R}_{0},\overrightarrow{\Lambda},\omega)\delta(\overrightarrow{r}'-\overrightarrow{R}(\overrightarrow{\Lambda}))F(\overrightarrow{\Lambda})d\overrightarrow{\Lambda}\label{eq:GreensFunctionRays}
\end{equation}
Here $G_{go}(\overrightarrow{R}_{0},\overrightarrow{\Lambda},\omega)$
is Green's function ray representation as a function of ray arguments
$\overrightarrow{\Lambda}$, antenna position $\overrightarrow{R}_{0}$
and frequency of the signal $\omega$; $\overrightarrow{R}(\overrightarrow{\Lambda})$
is ray trajectory for given ray coordinates $\overrightarrow{\Lambda}$;
\begin{equation}
F(\overrightarrow{\Lambda})=\left\Vert \frac{\partial R_{i}(\overrightarrow{\Lambda})}{\partial\Lambda_{j}}\right\Vert \label{eq:dR_dL}
\end{equation}
is Jacobian to convert from ray coordinates to spatial coordinates
(and at the same time the square of geometrical divergence);$\overrightarrow{\Lambda}=(\widehat{\Lambda};\Lambda_{\Vert})=(\widehat{\Lambda}_{\xi},\widehat{\Lambda}_{\eta};S)$
are the ray coordinates - two ray exit angles and trajectory group
length; $d\overrightarrow{\Lambda}=d\Lambda_{\Vert}d\widehat{\Lambda}_{\xi}d\widehat{\Lambda}_{\eta}$.
After taking into account antenna pattern $g(\widehat{\Lambda},\omega)$
of the transmitter, the signal in the position of the scatterer $\overrightarrow{r}$
can be calculated (with accuracy of constant multiplier) from the
integral representation (\ref{eq:GreensFunctionRays}), as a superposition
of the fields, propagating over the different trajectories:
\begin{equation}
u(\overrightarrow{r},\omega)=\int a(\omega)g(\widehat{\Lambda},\omega)G_{go}(\overrightarrow{R}_{0},\overrightarrow{\Lambda},\omega)\delta(\overrightarrow{r}-\overrightarrow{R}(\overrightarrow{\Lambda}))F(\overrightarrow{\Lambda})d\overrightarrow{\Lambda}\label{eq:signal_representation_rayGreen}
\end{equation}
where
\begin{equation}
G_{go}(\overrightarrow{R}_{0},\overrightarrow{\Lambda},\omega)=\frac{1}{F^{1/2}(\overrightarrow{\Lambda})}e^{i\frac{\omega}{c}\psi(\overrightarrow{\Lambda},\omega)}\label{eq:RayRepresentationFormula}
\end{equation}
$\psi(\overrightarrow{\Lambda},\omega)$ is eikonal in ray coordinates;
$\overrightarrow{R}(\overrightarrow{\Lambda})$ is ray trajectory;
$a(\omega)$ - transmitted signal spectrum for frequency $\omega$.
The given representation (\ref{eq:signal_representation_rayGreen})
has the simple physical sense: the signal that comes to the point
$\overrightarrow{r}$ is a superposition of geometrooptical rays,
transmitted from the antenna with different ray directions $\widehat{\Lambda}$
and amplitudes $g$. The correspondence of the representation (\ref{eq:signal_representation_rayGreen})
to the traditional geometrooptical representation for the signal can
be shown by integrating (\ref{eq:signal_representation_rayGreen})
over the $d\overrightarrow{\Lambda}$ and taking into account that
integrating of delta-function under integral produces the transformation
Jakobian (\ref{eq:dR_dL}) and summation over all propagation trajectories
$\nu$ to the given point $\overrightarrow{r}$:
\begin{equation}
\int g(\overrightarrow{\Lambda})\delta(\overrightarrow{r}-\overrightarrow{R}(\overrightarrow{\Lambda}))d\overrightarrow{\Lambda}=\sum_{\nu}\left\Vert \frac{\partial}{\partial\Lambda_{j}}R_{i}(\overrightarrow{\Lambda})\right\Vert _{\nu}^{-1}g(\overrightarrow{R}^{-1}(\overrightarrow{r}))\label{eq:delta-function-definition}
\end{equation}
After integration the signal (\ref{eq:signal_representation_rayGreen})
transforms into:
\begin{equation}
\left\{ \begin{array}{l}
u(\overrightarrow{r},\omega)=\sum\limits _{\nu}\frac{a(\omega)g(\widehat{\Lambda}_{\nu},\omega)}{F^{1/2}(\overrightarrow{\Lambda}_{\nu})}e^{i\frac{\omega}{c}\psi(\overrightarrow{R}(\overrightarrow{\Lambda}_{\nu}))}\\
\overrightarrow{r}=\overrightarrow{R}(\overrightarrow{\Lambda}_{\nu})
\end{array}\right.\label{eq:KravcovMultipathFormula}
\end{equation}
The trajectories $\nu$ are defined by second equation of the system
(\ref{eq:KravcovMultipathFormula}) for any given $\overrightarrow{r}$.
This equation is close to discussed in \cite{Kravcov_1980}.
\subsection{Ray equations: Hamilton representation}
The eikonal $\psi(\overrightarrow{\Lambda})$ is defined as the integral
over the curvilinear trajectory, definable as a propagation in background
ionosphere of the ray with given exit angle $\widehat{\Lambda}$.
The trajectory shape depends on the refractivity index distribution
$n(\overrightarrow{R}(\widehat{\Lambda},\sigma))$ in the ionosphere.
Traditionally, the calculations of rays for the decameter propagation
are made in hamiltonian optics approach \cite{Kravtsov_1990}. In
this approach, the ray trajectory is defined as a solution of the
Hamilton equations, and depends on the ray coordinates chosen. The
choice of ray coordinates depends on the convenience and on the structure
of Hamiltonian.
The eikonal $\psi$ in the equations (\ref{eq:GreensFunction},\ref{eq:RayRepresentationFormula},\ref{eq:KravcovMultipathFormula})
is also calculated from the Hamiltonian $H(\overrightarrow{P},\overrightarrow{R})$,
generalized momentum $\overrightarrow{P}$ and generalized coordinate
$\overrightarrow{R}$. One of the most used Hamiltonian $H(\overrightarrow{P},\overrightarrow{R})$
in geometrical optics is \cite{Kravtsov_1990}:
\begin{equation}
H(\overrightarrow{P},\overrightarrow{R})=\frac{1}{2}\left\{ P^{2}-n^{2}(\overrightarrow{R})\right\} =0\label{eq:ham_pn1}
\end{equation}
In this case the Hamilton equations, that allows to calculate the
ray trajectory and eikonal are:
\begin{equation}
\left\{ \begin{array}{c}
P=n(\overrightarrow{R})\\
\frac{\partial\overrightarrow{R}}{\partial S}=\overrightarrow{P}\\
\frac{\partial\overrightarrow{P}}{\partial S}=n\overrightarrow{\nabla}n(\overrightarrow{R})\\
\overrightarrow{\nabla}\psi=\overrightarrow{P}
\end{array}\right.\label{eq:gam_eq1}
\end{equation}
In this representation (\ref{eq:ham_pn1}) $S=\Lambda_{\Vert}$ is
ray coordinate, that is equivalent to the trajectory group length.
Eikonal, generalized coordinate and momentum are obtained by integrating
(\ref{eq:gam_eq1}) over the ray trajectory:
\begin{equation}
\left\{ \begin{array}{c}
\overrightarrow{R}=\overrightarrow{R}_{0}+\int\limits _{S_{0}}^{S}\overrightarrow{P}dS'\\
\overrightarrow{P}=\overrightarrow{P}_{0}+\int\limits _{S_{0}}^{S}n\overrightarrow{\nabla}n(\overrightarrow{R}(S'))dS'\\
\psi=\psi_{0}+\int\limits _{S_{0}}^{S}P^{2}dS'
\end{array}\right.\label{eq:gam_eq_solution}
\end{equation}
The hamiltonian optics representation is useful for numerical calculations
and is widely used in the problems of radiowave propagation (as characteristics
method)\cite{Kravtsov_1990,Settimi_2014}, including SuperDARN problems
\cite{Gauld_2002,Ponomarenko_2009,Ponomarenko_2010,Ponomarenko_2011,Berngardt_2015b}.
\subsection{Initial representation of the scattered signal}
One can see that ray representation of Green's function (\ref{eq:RayRepresentationFormula})
is the product of fast and slow oscillating functions. Spatial multiplier
$g(\widehat{\Lambda},\omega)$ is a smooth function of the frequency,
transmitted signal $a(\omega)$ is narrowband near carrier frequency
$\omega_{0}$. In this case we can represent fast oscillating phase
$\frac{\omega}{c}\psi(\overrightarrow{\Lambda},\omega)$ as a Taylor
series near $\omega_{0}$, with the small parameter $\left|\omega-\omega_{0}\right|\ll\omega_{0}$
and get the temporal representation of the propagating signal:
\begin{equation}
\begin{array}{c}
u(\overrightarrow{r},t)\approx e^{i\omega_{0}t}U(\overrightarrow{r},t)\\
U(\overrightarrow{r},t)\approx \int A(t-T_{r})g(\widehat{\Lambda},\omega_{0})F^{1/2}(\overrightarrow{\Lambda},\omega_{0})e^{ik_{0}\psi(\overrightarrow{\Lambda},\omega_{0})}\delta(\overrightarrow{r}-\overrightarrow{R}(\overrightarrow{\Lambda},\omega_{0}))d\overrightarrow{\Lambda}
\end{array}\label{eq:signal_propagation_narrowband}
\end{equation}
Here $k_{0}=\frac{\omega_{0}}{c}$ is wavenumber of transmitted signal;$U(\overrightarrow{r},t)$
is the complex envelope of propagating signal; $A(t-T_{r})$ is complex
envelope of the sounding signal;
\begin{equation}
T_{r,t}=\frac{\partial}{\partial\omega}\left(\frac{\omega}{c}\psi(\overrightarrow{\Lambda},\omega)\right)_{\omega=\omega_{0}}\label{eq:eikonal_delay}
\end{equation}
is the group delay of the propagating signal.
Necessary condition of narrowband signal (with spectral width $\Delta\omega$)
is defined as smallness of second differential of fast oscillating
phase in temporal representation of (\ref{eq:RayRepresentationFormula}):
\begin{equation}
(\Delta\omega)^{2}\frac{\partial^{2}}{\partial\omega^{2}}\left(\frac{\omega}{c}\psi(\overrightarrow{\Lambda},\omega)\right)_{\omega=\omega_{0}}\ll1\label{eq:narrowband_condition}
\end{equation}
So, the obtained equation (\ref{eq:signal_propagation_narrowband})
is valid in the regions, where antenna pattern, refractivity index
and geometrooptical divergence are changing slowly over the band $\Delta\omega$
of the sounding signal $a(\omega)$.
In Fig.\ref{fig:1} are shown the validity (\ref{eq:narrowband_condition})
regions calculated for reference ionosphere model, typical SuperDARN
frequency (10MHz) and typical radar pulse durations (100 and 300usec.).
The signal trajectories were calculated by the method of characteristics
for cold isotropic plasma given by IRI-2012 model \cite{Bilitza_2011}.
As one can see from Fig.\ref{fig:1}, the approximation of the narrowband
signal (\ref{eq:narrowband_condition}) is valid almost everywhere,
except for a small number of rays near the Pedersen ray after reflection
point. For longer pulses (300usec), this area is smaller than for
shorter ones (100usec). Qualitatively, this can be explained as follows:
near the Pedersen ray small changes of signal frequency will lead
to a big changes of the trajectory (high frequency components of the
signal propagates upward, and the low frequency components propagates
downward). Thus, when propagating over these trajectories the signal
envelope in the ionosphere is not the same as the transmitted signal
envelope, and one should whether do not consider any of these areas,
or take into account the signal distortions at these trajectories
correctly.
Below in the paper we consider only the trajectories at which the
narrowband condition (\ref{eq:narrowband_condition}) is valid, and
use in (\ref{eq:signal_propagation_narrowband}) the equality sign.
By the same way we can obtain the expression for the scattered signal
at the receiving antenna:
\begin{equation}
\begin{array}{c}
U(t)\approx k_{0}^{2}\int\varepsilon(\overrightarrow{r},t-T_{t})A(t-T_{r}-T_{t})D(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})e^{ik_{0}\Psi(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})}\cdot\\
\cdot\delta(\overrightarrow{r}-\overrightarrow{R}_{r}(\overrightarrow{\Lambda}_{r},\omega_{0}))\delta(\overrightarrow{r}-\overrightarrow{R}_{t}(\overrightarrow{\Lambda}_{t},\omega_{0}))d\overrightarrow{\Lambda}_{t}d\overrightarrow{\Lambda}_{r}d\overrightarrow{r}
\end{array}\label{eq:signal_propagation_narrowband-1}
\end{equation}
where
$\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t}$ - ray
coordinate systems relative to receiver and transmitter correspondingly;
$\psi_{r}(\overrightarrow{\Lambda}_{r},\omega_{0}),\psi_{t}(\overrightarrow{\Lambda}_{t},\omega_{0})$
are the trajectory phase lengths from the scatterer position $\overrightarrow{R}_{r}(\overrightarrow{\Lambda}_{r},\omega_{0})$
to the receiver and transmitter correspondingly;
$\Psi(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})=\psi_{r}(\overrightarrow{\Lambda}_{r},\omega_{0})+\psi_{t}(\overrightarrow{\Lambda}_{t},\omega_{0})$
is the full trajectory phase length (eikonal);
\begin{equation}
D(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})=g_{r}^{*}(\widehat{\Lambda}_{r},\omega_{0})g_{t}(\widehat{\Lambda}_{t},\omega_{0})F_{r}^{1/2}(\overrightarrow{\Lambda}_{r},\omega_{0})F_{t}^{1/2}(\overrightarrow{\Lambda}_{t},\omega_{0})\label{eq:D_def}
\end{equation}
is the spatial multiplier, defined by antenna patterns $g_{r}(\widehat{\Lambda}_{r},\omega_{0}),g_{t}(\widehat{\Lambda}_{t},\omega_{0})$
of receiver and transmitter correspondingly, and by geometrical divergence;
indexes $t,r$ mark trajectories of transmitting (trajectory from
transmitter to the scatterer) and receiving (trajectory from receiver
to scatterer) correspondingly; $\overrightarrow{R}_{t}(\overrightarrow{\Lambda_{t}}),\overrightarrow{R}_{r}(\overrightarrow{\Lambda_{r}})$
are the positions of the scatterer as a function of ray coordinates
(ray direction and radar range) in transmitter coordinate system and
receiver coordinate system correspondingly; $T_{r},T_{t}$ are group
delays (\ref{eq:eikonal_delay}) from scatterer to transmitter and
receiver correspondingly; delta-functions define the condition that
scatterer position corresponds to the coincidence of the ends of trajectories
from transmitter and receiver (and at the same time are keeping the
ability of several propagation trajectories to the scatterer); $\varepsilon(\overrightarrow{r},t)$
is dielectric permittivity variations.
Integrating over the $d\overrightarrow{r}$ allows to simplify the
expression (\ref{eq:signal_propagation_narrowband-1}) and represent
it as an integral over the ray coordinates of transmitting and receiving
rays:
\begin{equation}
\begin{array}{c}
U(t)=k_{0}^{2}\int\varepsilon(\overrightarrow{R}_{t}(\overrightarrow{\Lambda}_{t},\omega_{0}),t-T_{t})A(t-T_{r}-T_{t})D(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})\cdot\\
\cdot e^{ik_{0}\Psi(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})}\delta(\overrightarrow{R}_{r}(\overrightarrow{\Lambda}_{r},\omega_{0})-\overrightarrow{R}_{t}(\overrightarrow{\Lambda}_{t},\omega_{0}))d\overrightarrow{\Lambda}_{t}d\overrightarrow{\Lambda}_{r}
\end{array}\label{eq:signal_scattering_narrowband}
\end{equation}
Delta-function under the integral in the case of several propagation
trajectories to the scatterer will transformed into the cross-sum
over these trajectories after integrating over the $\overrightarrow{\Lambda}_{t},\overrightarrow{\Lambda}_{r}$.
It is important to note that this equation is valid for narrowband
sounding and received signals (\ref{eq:narrowband_condition}) (where
$\Delta\omega$ is the sum of the bands of sounding signal $a(\omega)$
and dielectric permittivity fluctuations $\varepsilon(\omega)$) and
in absence of caustics for given group delay $t$.
For obtaining SuperDARN radar equations we will consider this equation
in the cases of monostatic and bistatic sounding.
\section{Monostatic sounding: single trajectory case}
In monostatic case the position of the receiver and the transmitter
coincides, so the functions $\overrightarrow{R}_{t}(\overrightarrow{\Lambda_{t}}),\overrightarrow{R}_{r}(\overrightarrow{\Lambda_{r}})$
coincides too and can be represented as a same function $\overrightarrow{R}(\overrightarrow{\Lambda})$,
but from different ray arguments $\overrightarrow{\Lambda_{t}},\overrightarrow{\Lambda_{r}}$
. In this case the expression for the received signal (\ref{eq:signal_scattering_narrowband})
becomes:
\begin{equation}
\begin{array}{c}
U(t)=k_{0}^{2}\int\varepsilon(\overrightarrow{R}(\overrightarrow{\Lambda}_{t}),t-T_{t})A(t-T_{r}-T_{t})D(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})\cdot\\
\cdot e^{ik_{0}\Psi(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t})}\delta(\overrightarrow{R}(\overrightarrow{\Lambda}_{r})-\overrightarrow{R}(\overrightarrow{\Lambda}_{t}))d\overrightarrow{\Lambda}_{t}d\overrightarrow{\Lambda}_{r}
\end{array}\label{eq:signal_scattering_singlepos}
\end{equation}
Here we excluded all the evident arguments of the functions. The expression
(\ref{eq:signal_scattering_singlepos}) describes both single and
multiple trajectory cases for the case of monostatic sounding.
\subsection{Single trajectory case}
To illustrate the approach of obtaining radar equation lets analyze
the simplest case: there is only a single trajectory to reach scatterer
from the transmitter. This corresponds to the case when $\overrightarrow{R}(\overrightarrow{\Lambda}_{r})$
is bijective function. In this case the integral over the $d\overrightarrow{\Lambda}_{r}$
in (\ref{eq:signal_scattering_singlepos}) can be calculated. The
rest argument $\overrightarrow{\Lambda}_{t}$ can be renamed to $\overrightarrow{\Lambda}$,
additional multiplier $F_{r}^{-1}$ arises due to integration of delta-function,
and the expression for the scattered signal (\ref{eq:signal_scattering_singlepos})
becomes:
\begin{equation}
U(t)=\int\varepsilon(\overrightarrow{R}(\overrightarrow{\Lambda}),t-T(\overrightarrow{\Lambda}))A(t-2T(\overrightarrow{\Lambda}))D_{u}(\overrightarrow{\Lambda},\omega_{0})e^{i2k_{0}\Psi(\overrightarrow{\Lambda})}d\overrightarrow{\Lambda}\label{eq:signal4-1}
\end{equation}
where
\begin{equation}
D_{u}(\overrightarrow{\Lambda}_{r},\omega_{0})=g_{r}^{*}(\widehat{\Lambda},\omega_{0})g_{t}(\widehat{\Lambda},\omega_{0})\label{eq:D_def-1}
\end{equation}
Usually, when interpreting SuperDARN data, the average autocorrelation
function of the signal\cite{Greenwald_1995,Chisham_2007,Ribeiro_2013}
is analysed:
\begin{equation}
P(t,\Delta T)=\left\langle U(t)U^{*}(t+\Delta T)\right\rangle \label{eq:P_UU}
\end{equation}
To get the radar equation, by analogy with \cite{Berngardt_1999,Berngardt_2000},
lets use the following characteristics:
\begin{equation}
\Phi(\overrightarrow{r};\overrightarrow{\rho},\Delta T)=\left\langle \varepsilon(\overrightarrow{r},t)\varepsilon^{*}(\overrightarrow{r}+\overrightarrow{\rho},t+\Delta T)\right\rangle \label{eq:irreg-spec-dens-def}
\end{equation}
is the stationary spatio-temporal correlation function of the irregularities,
that defines statistical characteristics of the scatterers;
\begin{equation}
W\left(t,S,\Delta T,\Delta S\right)=A(t-S/c)A^{*}(t-S/c+\Delta T+\Delta S/c)\label{eq:weight-volume-def}
\end{equation}
is the weight volume, that defines spatial selection of the irregularities
over the radar range (defined by two-way distance $S$) and selection
of the irregularities over the correlation function lag (defined by
parameter $\Delta T$ ) ;
\begin{equation}
\begin{array}{l}
D_{\Sigma}(\overrightarrow{\Lambda},\overrightarrow{\Lambda}_{2})=D_{u}(\overrightarrow{\Lambda},\omega_{0})D_{u}^{*}(\overrightarrow{\Lambda}_{2},\omega_{0})=\\
=g_{r}^{*}(\widehat{\Lambda}_{r},\omega_{0})g_{t}(\widehat{\Lambda}_{t},\omega_{0})g_{r}^{*}(\widehat{\Lambda}_{r,2},\omega_{0})g_{t}(\widehat{\Lambda}_{t,2},\omega_{0})\\
\\
\end{array}\label{eq:d-sum-def}
\end{equation}
is spatial multiplier that defines the contribution of antenna patterns.
The average autocorrelation function of the received signal (\ref{eq:P_UU})
can be represented as following temporal and spectral radar equations:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\int\Phi(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{R}(\overrightarrow{\Lambda}')-\overrightarrow{R}(\overrightarrow{\Lambda}),\Delta T-T(\overrightarrow{\Lambda}')+T(\overrightarrow{\Lambda}))\cdot\\
\cdot W(t,T(\overrightarrow{\Lambda})c,\Delta T,\left\{ T(\overrightarrow{\Lambda}')-T(\overrightarrow{\Lambda})\right\} c)D_{\Sigma}(\overrightarrow{\Lambda},\overrightarrow{\Lambda}')e^{i2k_{0}\left\{ \psi(\overrightarrow{\Lambda})-\psi(\overrightarrow{\Lambda}')\right\} }d\overrightarrow{\Lambda}d\overrightarrow{\Lambda}'
\end{array}\label{eq:signal4-1-2-1-1}
\end{equation}
\begin{equation}
\begin{array}{c}
\widetilde{P}(t,\omega)=\int\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{k},\nu\right)\widetilde{W}\left(t,T(\overrightarrow{\Lambda})c,\omega-\nu,\left\{ T(\overrightarrow{\Lambda}')-T(\overrightarrow{\Lambda})\right\} c\right)\cdot\\
\cdot e^{i\nu\left\{ T(\overrightarrow{\Lambda})-T(\overrightarrow{\Lambda}')\right\} }D_{\Sigma}(\overrightarrow{\Lambda},\overrightarrow{\Lambda}')e^{-i\overrightarrow{k}\left(\overrightarrow{R}(\overrightarrow{\Lambda})-\overrightarrow{R}(\overrightarrow{\Lambda}')\right)}e^{i2k_{0}\left(\psi(\overrightarrow{\Lambda})-\psi(\overrightarrow{\Lambda}')\right)}d\overrightarrow{\Lambda}d\overrightarrow{\Lambda}'d\nu d\overrightarrow{k}
\end{array}\label{eq:ResSpectrum1}
\end{equation}
Here $\widetilde{A}$ marks Fourier transform of $A$ over one of
its arguments; $\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{k},\nu\right)$
is spectral density of the irregularities.
It is difficult to use these equations for interpreting the experimental
data because of the fast oscillating functions under integral. But,
some of the integrands in (\ref{eq:signal4-1-2-1-1},\ref{eq:ResSpectrum1})
do not depend on $\overrightarrow{\Lambda}'$, the other can be considered
slightly dependent on it. So we can successfully integrate the equations:
at first over the $d\overrightarrow{\Lambda}'$ , and than over the
$d\overrightarrow{\Lambda}$.
To obtain a convenient representation of (\ref{eq:signal4-1-2-1-1},\ref{eq:ResSpectrum1})
we need to integrate fast oscillating functions like:
\begin{equation}
I=\int Z(\overrightarrow{\Lambda})e^{i\left(-\overrightarrow{k}\overrightarrow{R}(\overrightarrow{\Lambda},\omega_{0})+2k_{0}\psi(\overrightarrow{\Lambda},\omega_{0})\right)}d\overrightarrow{\Lambda}\label{eq:integral2}
\end{equation}
where $Z(\overrightarrow{\Lambda})$ is a function that is smooth
in comparison with fast-oscillating exponent. By analogy with \cite{Berngardt_1999,Berngardt_2000}
lets use a combination of the stationary phase method \cite{Fedoryuk_1989}
and Fourier transform for this.
\subsection{Integrating fast oscillating functions}
\subsubsection{Convenient ray coordinates}
The initial equations for the scattered signal (\ref{eq:ResSpectrum1})
are given in their basic representation, without specifying the exact
ray coordinate system $\overrightarrow{\Lambda}$. Different ones
can be used for this \cite{Kravtsov_1990}. The fast oscillating
integrand (\ref{eq:integral2}) in hamiltonian representation for
case (\ref{eq:gam_eq_solution}) becomes:
\begin{equation}
I=\int Z(\overrightarrow{\Lambda})e^{i\left(\int\limits _{0}^{S}\left\{ -\overrightarrow{k}\overrightarrow{P}(\hat{\Lambda},S')+2k_{0}P^{2}(\hat{\Lambda},S')\right\} dS'\right)}d\widehat{\Lambda}_{\xi}d\widehat{\Lambda}_{\eta}dS\label{eq:integral4}
\end{equation}
The most convenient technique for the calculation of fast oscillating
integrals is the method of stationary phase (MSP)\cite{Fedoryuk_1989}.
In this case, the integral can be expressed as the sum of the contributions
of the stationary points (points at which the derivative of fast-changing
phase over the integration variable becomes zero), and the contribution
is inversely proportional to the square root of second derivative
of the fast oscillating phase at this point (characteristic scale).
The criterion for the applicability of this method is small changes
of the other integrands at the characteristic scale \cite{Fedoryuk_1989}.
To check convenience of the ray coordinate system lets calculate the
integral (\ref{eq:integral4}) by stationary phase method over the
ray exit angles $\hat{\Lambda}=(\widehat{\Lambda}_{\xi},\widehat{\Lambda}_{\eta})$
that is convenient for this problem \cite{Berngardt_1999,Berngardt_2000}.
The detailed calculation by stationary phase method showed, that it
is convenient to choose the coordinates system that provides
\begin{equation}
\left\{ \overrightarrow{P}\frac{\partial\overrightarrow{P}}{\partial\hat{\Lambda}}\right\} _{\Lambda_{\Vert}=const}=0\label{eq:P_dPdL}
\end{equation}
that is valid for refraction-free or homogeneous background ionosphere.
In this case obtaining and interpreting the radar equations will be
similar to the approach developed for VHF \cite{Berngardt_1999}.
Due to the generalized momentum $\overrightarrow{P}$ is perpendicular
to the wave front at any given moment \cite[p.24]{Babich_1972}, we
should use instead of the trajectory length coordinate $\Lambda_{\parallel}$
the trajectory group length $S$ (or group delay of the signal $dT=\frac{dL}{cn}$),
that provides (\ref{eq:P_dPdL}). In this case, one can use the system
of Hamiltonian equations, described above (\ref{eq:gam_eq1},\ref{eq:gam_eq_solution})
and analyze the integral (\ref{eq:integral4}).
Using (\ref{eq:gam_eq1},\ref{eq:gam_eq_solution}) it can be shown
that stationary point condition $\frac{\partial\varphi}{\partial\hat{\Lambda}}=0$
corresponds to:
\begin{equation}
\left\{ \frac{\partial}{\partial\hat{\Lambda}}\int\limits _{0}^{S}\left(\overrightarrow{k}\overrightarrow{P}\right)dS'\right\} _{\hat{\Lambda}=\hat{\Lambda}_{SF}}=\int\limits _{0}^{S}\left\{ \overrightarrow{k}\left(\frac{\partial\overrightarrow{P}}{\partial\hat{\Lambda}}\right)_{\bot\overrightarrow{P}}\right\} _{\hat{\Lambda}=\hat{\Lambda}_{SF}}dS'=0\label{eq:StatConditionScalar}
\end{equation}
This means that stationary point is the point where the projection
of the irregularities wave vector $\overrightarrow{k}$ to the perpendicular
to the generalized momentum $\overrightarrow{P}$ 'in average' (as
an integral along the trajectory) equals to zero. In other words,
the irregularities wave vector $\overrightarrow{k}$ is 'in average'
should be perpendicular to the wave front.
So the stationary point (\ref{eq:StatConditionScalar}) for coordinate
system $(\hat{\Lambda},S)$ can be represented as:
\begin{equation}
\int\limits _{0}^{S}\overrightarrow{k}\overrightarrow{P}dS'=-\int\limits _{0}^{S}kPdS'\label{eq:ST_integral}
\end{equation}
For locally-homogeneous trajectory, i.e. for trajectory, at which
the refraction coefficient is nearly constant at investigated part
of the trajectory ($P=const$), the stationary point can be represented
not as the 'average' integral condition (\ref{eq:StatConditionScalar}),
but as the local condition:
\begin{equation}
\overrightarrow{k}\overrightarrow{P}=-kP\label{eq:ST_loc}
\end{equation}
This condition is structurally coincides with the refraction-free
case, analyzed in \cite{Berngardt_1999}, and corresponds the the
well-known Wolf-Bragg condition.
\subsubsection{Integrating over the ray angles}
By integrating (\ref{eq:ResSpectrum1}) over the ray exit angles $\hat{\Lambda},\hat{\Lambda}'$
by MSP and using stationary phase condition in form (\ref{eq:ST_integral}),
we obtain:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\int\Phi\left(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{k},\Delta T+S/c-S'/c\right)W_{2,t}\left(\hat{\Lambda}_{SF},S,\Delta T,S-S'\right)\cdot\\
\cdot\left|\gamma_{2}\right|^{2}e^{-i\int\limits _{S'}^{S}P\left(k-2k_{0}P(\hat{\Lambda}_{SF},S_{1},\omega_{0})\right)dS_{1}}dSdS'd\overrightarrow{k}
\end{array}\label{eq:autocorfunction1-2-2-1-1}
\end{equation}
where
\begin{equation}
\begin{array}{c}
W_{2,t}\left(\hat{\Lambda}_{SF},S,\Delta T,S-S'\right)=W\left(t,2S,\Delta T,2\left\{ S-S'\right\} \right)V_{\Lambda}^{-1}(\hat{\Lambda}_{SF},S,\omega_{0})\cdot\\
\cdot V_{\Lambda}^{-1}(\hat{\Lambda}_{SF},S',\omega_{0})D_{\Sigma}(\hat{\Lambda}_{SF},S)
\end{array}\label{eq:W_2t}
\end{equation}
is weight volume that defines spatio-temporal resolution of the technique;
$V_{\Lambda}$ is effective scattering volume; $\hat{\Lambda}_{SF}$
is stationary point over $\hat{\Lambda}$;
\[
\gamma_{2}=\left(2\pi\right)^{n/2}e^{i\frac{\pi}{4}sgn(S'')}=2\pi i
\]
is additional multiplier from stationary phase method \cite{Fedoryuk_1989}.
From the initial equation (\ref{eq:ResSpectrum1}) it is obvious that
for slowly changing $\Phi\left(\overrightarrow{R}(\hat{\Lambda}_{SF},S);\right)W\left(t-2T(\hat{\Lambda}_{SF},S)\right)$
the stationary points over $\hat{\Lambda},\hat{\Lambda}'$ are similar
and marked as $\hat{\Lambda}_{SF}$ , and satisfy the Wolf-Bragg condition
(\ref{eq:StatConditionScalar}).
The basic element of the calculations by stationary phase method is
the second differential of fast oscillating integrand function $\psi$.
It defines the integral contribution of the stationary point to the
integral, and inverse angular size of the region, making this contribution
\cite{Fedoryuk_1989}:
\begin{equation}
V_{\Lambda}=k\left\Vert \frac{\partial^{2}\psi}{\partial\hat{\Lambda}_{i}\partial\hat{\Lambda}_{j}}\right\Vert ^{1/2}\label{eq:WR-def}
\end{equation}
In geometrical optics the Hessian $V_{\Lambda}$ can be defined through
the product of two main curvature radii $R_{GM},R_{Gm}$ \cite{Kravtsov_1990}:
\begin{equation}
V_{\Lambda}=kF^{1/2}=k\left\{ R_{GM}R_{Gm}\right\} ^{1/2}\label{eq:WR_RgRg}
\end{equation}
In spectral representation the radar equation (\ref{eq:autocorfunction1-2-2-1-1})
becomes:
\begin{equation}
\begin{array}{c}
\widetilde{P}(t,\omega)=\int\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{k},\nu\right)\widetilde{W}_{2,t}\left(\hat{\Lambda}_{SF},S,\omega-\nu,S-S'\right)\cdot\\
\cdot\left|\gamma_{2}\right|^{2}e^{i\nu\left\{ S/c-S'/c\right\} }e^{-i\int\limits _{S'}^{S}P\left(k-2k_{0}P(\hat{\Lambda}_{SF},S_{1},\omega_{0})\right)dS_{1}}dSdS'd\overrightarrow{k}d\nu
\end{array}\label{eq:RLU-Final-Integral}
\end{equation}
The criterion of applicability of the MSP is a small changes of integrands
$\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{k},\nu\right),\widetilde{W}_{2,t}\left(\hat{\Lambda}_{SF},S,\omega-\nu,S-S'\right)$
at angles about $V_{\Lambda}^{-1}$ (\ref{eq:WR_RgRg}), or roughly
equivalent, small changes of the integrands at distances of the order
Fresnel zone $\sqrt{\lambda R}$. This condition has equivalent in
the refraction-free case \cite{Berngardt_1999}. It should be noted
that the smoothness of the refraction index on the characteristic
scale of the Fresnel zone is one of the conditions of applicability
of geometrical optics \cite{Kravtsov_1990}. Thus, the resulting
representation of radar equation (\ref{eq:RLU-Final-Integral}) is
valid almost everywhere, where the original equation (\ref{eq:signal4-1})
is valid and, at the same time, where the average spectral density
$\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{k},\nu\right)$
varies smoothly over the spatial variable $\overrightarrow{R}(\overrightarrow{\Lambda})$.
\subsubsection{Integrating the radar equation over the range}
Lets calculate the integral from function (\ref{eq:RLU-Final-Integral})
that oscillates fast over the $\Delta S=S-S'$ with the phase $\varphi$:
\begin{equation}
\varphi=\nu\Delta S-\int\limits _{S-\Delta S}^{S}P\left(k-2k_{0}P(\hat{\Lambda}_{SF},S_{1},\omega_{0})\right)dS_{1}\label{eq:PhaseForRegularityCondition}
\end{equation}
As it will be shown, it is convenient to analyze the two extreme cases:
locally-homogeneous and locally-inhomogeneous one. The integration
domain over $\Delta S$ is limited by the weight volume $W_{2,t}\left(\hat{\Lambda}_{SF},S,\Delta T,\Delta S\right)$
carrier. For pulse sequences used at SuperDARN radars\cite{Chisham_2007}
this region is defined by elementary sounding pulse duration.
It can be shown that for SuperDARN radars the carrier of $W_{2,t}$
has a specific shape over $\Delta S$ - a peak near zero and a few
peaks at very large distances $\Delta S$. Generally, this volume
can be decomposed into the sum of a peaks. For large distances $\Delta S$
we integrate regions separated by thousands kilometers so the correlation
at large $\Delta S$ and integrals with large $\Delta S$ can be neglected.
Therefore lets consider the following two limiting cases: when it
is a linear function of $\Delta S$ (it is equivalent to the smallness
of second differential of the phase (\ref{eq:PhaseForRegularityCondition})
over $\Delta S$) and the integral can be calculated as Fourier transform,
and when (\ref{eq:PhaseForRegularityCondition}) is very nonlinear
function of $\Delta S$ (when the second differential of the phase
(\ref{eq:PhaseForRegularityCondition}) over $\Delta S$ is big enough)
and the integral can be calculated with the stationary phase method.
So the (\ref{eq:RLU-Final-Integral}) calculations can be differed
into two classes:
- the locally-homogeneous parts of trajectories (with nearly constant
refraction coefficient along the trajectory), at which the scattered
signal is accumulated coherently over $\Delta S$ near zero;
- the locally-inhomogeneous parts of trajectories, at which the main
contribution into the integral is made by relatively small area inside
the integration region $\Delta S$.
So the radar equation (\ref{eq:RLU-Final-Integral}) will be a sum
of two terms - the contributions from locally-homogeneous and from
locally-inhomogeneous trajectories.
\paragraph{Locally-homogeneous trajectory case}
In the case of locally-homogeneous trajectories the resulting signal
is defined by 2D stationary phase method over the ray exit directions
$\hat{\Lambda}$, and by the Fourier-transform over the radar range
$\Delta S$. The locally-homogeneous case is equivalent to the refraction-free
case \cite{Berngardt_1999}, and 2D stationary phase condition (\ref{eq:StatConditionScalar})
can be considered locally too (\ref{eq:ST_loc}).
It is obvious that background ionosphere can be considered as locally-homogeneous
for given sounding pulse only in the case of sufficiently short pulse.
However, it should satisfy the narrowband sounding requirement (\ref{eq:narrowband_condition}),
to initial formulas to be valid. When pulse becomes longer and longer
there comes a moment when the locally-homogeneous trajectory condition
stops working. In this case coherent accumulation of the signal stops
and scattered signal amplitude growing stops too. Qualitatively one
can suppose that maximal scattered power is reached when the pulse
duration and local-homogenety scale length becomes equal.
The locally-homogeneous trajectory condition (i.e. the maximal size
of the region, at which the phase $\varphi$ (\ref{eq:PhaseForRegularityCondition})
can be considered as linear function of $\Delta S$) in the first
approximation can be defined from smallness of the second differential
of phase (\ref{eq:PhaseForRegularityCondition}) as
\begin{equation}
\left(\frac{\Delta R_{A}}{J_{1}}\right)^{2},\left(\frac{\Delta T_{A}}{cJ_{1}}\right)^{2}<1\label{eq:LocallyRegularCondition}
\end{equation}
where $\Delta T_{A}$ is the elementary sounding pulse duration (defined
from the spatial resolution of sounding pulse $\Delta R_{A}$); $J_{1}$
- the characteristic scale of quadratic phase changes:
\begin{equation}
J_{1}=(2k_{0}P\frac{\partial P}{\partial S})^{-1/2}\label{eq:J1}
\end{equation}
If $\Delta L$ is a characteristic scale of changes of refractivity
index along the trajectory, then $J_{1}\sim\sqrt{\lambda_{0}\Delta L}$.
From the condition of locally-homogeneous trajectory (\ref{eq:LocallyRegularCondition})
one can see that the requirement is equivalent to small variations
of the background refraction coefficient along the trajectory and
to short sounding pulses. At these trajectories the radar equation
(\ref{eq:RLU-Final-Integral}) can be significantly simplified. The
integration over the $d(S-S')$ produces Fourier transform of $W_{2,t}$:
\begin{equation}
\begin{array}{c}
\widetilde{P}(t,\omega)=\int\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\overrightarrow{\Lambda});\overrightarrow{k},\nu\right)\cdot\\
\cdot\left|\gamma_{2}\right|^{2}\widetilde{\widetilde{W}}_{2,t}\left(\hat{\Lambda}_{SF},S,\omega-\nu,\left(kP(\hat{\Lambda}_{SF},S,\omega_{0})-2k_{0}P^{2}(\hat{\Lambda}_{SF},S,\omega_{0})\right)-\nu/c\right)dSd\overrightarrow{k}d\nu
\end{array}\label{eq:ResultRLU-quasihomogeneous}
\end{equation}
The narrowband of the sounding signal and spectral density of the
irregularities in this case determines the smallness of the wavenumbers
region involved into the scattering:
\begin{equation}
\left|kP(\hat{\Lambda}_{SF},S,\omega_{0})-2k_{0}P^{2}(\hat{\Lambda}_{SF},S,\omega_{0})\right|<\Delta\omega/c\label{eq:StatCondition-Local-mod}
\end{equation}
Taken together, these two conditions (\ref{eq:StatCondition-Local-mod},\ref{eq:ST_loc})
define space-phase matching for radiowaves and irregularities (Wolf-Bragg
condition): the main contribution to the scattering in a smoothly
irregular media is produced by irregularities spatial harmonics with
wavevectors perpendicular to the wave front of the propagating radio
wave, and with the wavenumber nearly equal to the doubled wave number
of radiowave propagating in the medium. It is necessary to note that
the closer $P$ (or $n$) to 0, the wider the integration region over
$dk$. The conditions (\ref{eq:StatCondition-Local-mod},\ref{eq:ST_loc})
are illustrated at Fig.\ref{fig:fig_dP}A.
To represent (\ref{eq:ResultRLU-quasihomogeneous}) as an integral
over the space $d\overrightarrow{r}$, one needs to calculate Jacobian
for the transformation $dSd\hat{k}\rightarrow d\overrightarrow{r}$.
To simplify the calculations, lets take into account Wolf-Bragg condition
($d\hat{k}=d\hat{P}$) and calculate the following two transformations:
$dSd\hat{P}\rightarrow dSd\hat{\Lambda}\rightarrow d\overrightarrow{r}$
.
Geometrical interpretation of $F$ is cross sectional area of the ray\cite{Kravtsov_1990}. From geometrical considerations (see Fig.\ref{fig:fig_dP}B) it is
obvious that
\begin{equation}
J_{\hat{P}\rightarrow\hat{\Lambda}}\approx\left\{ \frac{1}{2}\frac{\sigma}{F}\frac{\partial F}{\partial\sigma}\right\} ^{2}\label{eq:J_P_L}
\end{equation}
It is also clear that $J_{dSd\hat{\Lambda}\rightarrow d\overrightarrow{r}}=2n^{-1}\left|\frac{\partial\overrightarrow{r_{i}}}{\partial\overrightarrow{\Lambda_{j}}}\right|^{-1}=2n^{-1}F^{-1}$.
So
\begin{equation}
\begin{array}{c}
\widetilde{P}(t,\omega)=\int\widetilde{\widetilde{\Phi}}\left(\overrightarrow{r};\overrightarrow{k},\nu\right)\left|\gamma_{2}\right|^{2}k^{2}\left\{ \frac{2}{nF}J_{\hat{P}\rightarrow\hat{\Lambda}}\right\} \cdot\\
\cdot\widetilde{\widetilde{W}}_{2,t}\left(\hat{\Lambda}_{SF}(\overrightarrow{r}),S(\overrightarrow{r}),\omega-\nu,\left(kP(\hat{\Lambda}_{SF}(\overrightarrow{r}),S(\overrightarrow{r}),\omega_{0})-2k_{0}P^{2}(\hat{\Lambda}_{SF}(\overrightarrow{r}),S(\overrightarrow{r}),\omega_{0})\right)-\nu/c\right)d\overrightarrow{r}dkd\nu
\end{array}\label{eq:ResultRLU-quasihomogeneous-dR}
\end{equation}
where $W_{2,t}\left(\hat{\Lambda}_{SF},S,\Delta T,S-S'\right)$ is
defined by (\ref{eq:W_2t}).
The additional spatial multiplier $\left\{ \frac{2}{nF}J_{\hat{P}\rightarrow\hat{\Lambda}}\right\} $
is relatively smooth over the space (except the caustics regions,
where it becomes infinite). For refraction-free case it is equal to
$\frac{2}{r^{2}}$, and the radar equation (\ref{eq:ResultRLU-quasihomogeneous-dR})
becomes similar to the obtained in \cite{Berngardt_1999}.
In Fig.\ref{fig:2} are illustrated the locally-homogeneous areas
for model ionosphere. The figure shows that such basic areas are:
field caustics associated with reflection (170-270km); area corresponding
F2-peak (270km) and E-peak (120 km); and the region of linear propagation
out of the ionosphere (area 0-90km altitude).
\paragraph{Locally-inhomogeneous trajectory case}
With increase of the variations of the refractivity index along the
trajectory or/and increase of the sounding pulse duration the situation
changes gradually. In this case the phase (\ref{eq:PhaseForRegularityCondition})
can not be considered as a linear function of $\Delta S$, and Fourier
transform approximation can not be used, and one need to use other
approximate methods to calculate fast oscillating integral (\ref{eq:RLU-Final-Integral}).
In the case of locally-inhomogeneous trajectories integration of the
radar equation (\ref{eq:RLU-Final-Integral}) over $\Delta S$ can
be done by using stationary phase method. One also can not use the
local condition for stationary point (\ref{eq:ST_loc}) instead of
the integral one (\ref{eq:ST_integral}). This mean that the stationary
point will be defined from 3D structure of the integral.
Calculating the position of the stationary point $\hat{\Lambda}_{SF},S_{SF}$
over $S$ and $S'$ separately in (\ref{eq:RLU-Final-Integral}) gives
the following integral conditions for stationary phase:
\begin{equation}
\left\{ \begin{array}{c}
\int\limits _{S-\Delta S_{W}}^{S}\overrightarrow{k}\left(\frac{\partial\overrightarrow{P}}{\partial\Lambda}\right)_{\bot\overrightarrow{P}}dS_{1}=0\\
\overrightarrow{k}\overrightarrow{P}(\hat{\Lambda}_{SF},S_{SF})-2k_{0}P^{2}(\hat{\Lambda}_{SF},S_{SF})-\nu/c=0
\end{array}\right.\label{eq:StatConditionScalar-LoclaKriv}
\end{equation}
In the case when the stationary point $S_{SF}$ is a single one and
is within the carrier of $W(S)$ $[S-\Delta S_{W},S]$ for given $t$
it makes a big contribution into integral (\ref{eq:RLU-Final-Integral}).
In other cases the contribution is small. It should be noted that
multiplier outside the carrier of W(S) $\int\limits _{0}^{S-\Delta S_{W}}\overrightarrow{k}\left(\frac{\partial\overrightarrow{P}}{\partial\Lambda}\right)_{\bot}dS_{1}$
can be removed out the integral sign. So actually the MSP (\ref{eq:StatConditionScalar-LoclaKriv})
is calculated over the carrier of W(S).
In this case both stationary points (for $S$ and $S'$) coincide
and integral will lead to the $W(\Delta S=0)$, and the expression
(\ref{eq:RLU-Final-Integral}) becomes:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\int\widetilde{\Phi}\left(\overrightarrow{R}(\overrightarrow{\Lambda}_{SF});\overrightarrow{k},\Delta T\right)W\left(t,2T(\overrightarrow{\Lambda}_{SF})c,\Delta T,0\right)\cdot\\
\cdot\left|\gamma_{2}\gamma_{1}\right|^{2}J_{1}^{2}V_{\Lambda}^{-2}(\overrightarrow{\Lambda}_{SF},\omega_{0})D_{\Sigma}(...)d\overrightarrow{k}
\end{array}\label{eq:autocorfunction1-2-2-2}
\end{equation}
\begin{equation}
\begin{array}{c}
\widetilde{P}(t,\omega)=\int\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\overrightarrow{\Lambda}_{SF});\overrightarrow{k},\nu\right)\widetilde{W}\left(t,2T(\overrightarrow{\Lambda}_{SF})c,\omega-\nu,0\right)\cdot\\
\cdot\left|\gamma_{2}\gamma_{1}\right|^{2}J_{1}^{2}V_{\Lambda}^{-2}(\overrightarrow{\Lambda}_{SF},\omega_{0})D_{\Sigma}(...)d\overrightarrow{k}d\nu
\end{array}\label{eq:autocorfunction1-2-2-2-1}
\end{equation}
where
\[
\gamma_{1}=\left(2\pi\right)^{1/2}e^{i\frac{\pi}{4}}=\sqrt{2\pi i}
\]
is additional multiplier from stationary phase method \cite{Fedoryuk_1989};
$J_{1}$ is defined by (\ref{eq:J1}).
Taking into account the same considerations, as for analysis formula
(\ref{eq:ResultRLU-quasihomogeneous}), we can obtain:
\begin{equation}
\begin{array}{c}
P(t,\omega)=\int\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\hat{\Lambda}_{SF},S_{SF}(\hat{k} _{SF}k));\hat{k} _{SF} k,\nu\right)\widetilde{W}_{2,t}\left(\hat{\Lambda}_{SF},T(\hat{\Lambda}_{SF},S_{SF}(\hat{k} _{SF}k))c,\omega-\nu,0\right)\cdot\\
\cdot\left|\gamma_{2}\gamma_{1}\right|^{2}J_{1}^{2}(\hat{\Lambda}_{SF},S_{SF}(\hat{k} _{SF}k))D_{\Sigma}(...)J_{\hat{P}\rightarrow\hat{\Lambda}}k^{2}dkd\hat{\Lambda}d\nu
\end{array}\label{eq:RLU-LocInhom}
\end{equation}
To estimate the relation between contributions of locally-homogeneous
(\ref{eq:ResultRLU-quasihomogeneous}) and locally-inhomogeneous (\ref{eq:RLU-LocInhom}),
we can estimate $d\overrightarrow{r}\sim\left(\frac{S}{2}\right)^{2}\Delta R_{A}d\hat{\Lambda}$
in (\ref{eq:ResultRLU-quasihomogeneous}). So the relation can be
estimated as:
\[
\frac{P_{inhom}(t,\omega)}{P_{hom}(t,\omega)}\approx\frac{J_{1}^{2} J_{\hat{P}\rightarrow\hat{\Lambda}}\frac{\Delta\omega_{A}}{nc}\Delta\omega_{A}}{\left(\frac{S}{2}\right)^{2}\left\{ \frac{2}{nF}J_{\hat{P}\rightarrow\hat{\Lambda}}\right\} \Delta R_{A}^{2}\frac{\Delta\omega_{A}}{nc}\Delta\omega_{A}}\approx\frac{J_{1}^{2}}{\{\Delta R_{A}\}^{2}}
\]
So the energy of the scattered signal in locally-homogeneous case
approximately $\left(\Delta R_{A}/J_{1}\right)^{2}$ times higher
then in locally-inhomogeneous case.
In Fig.\ref{fig:3}A shown the value of $J_{1}^{2}/\left(\Delta R_{A}\right)^{2}$
for locally-homogeneous (=0dB, yellow areas), and for locally-inhomogeneous
trajectories (red and blue colors). The calculations for Fig.\ref{fig:3}A
are made with the reference model ionosphere (IRI-2012) and typical
frequency and duration of SuperDARN sounding signal.
In Fig.\ref{fig:3}A shown that the regions with locally-inhomogeneous
trajectories (red and blue) are located between the E- and F-maximum
and expected to produce scattering about 10-15dB less powerful than
the region of locally-homogeneous trajectories. Because of the field-aligned
scatterers (studied by SuperDARN radars) produce powerful aspect scattering
these areas look still important for future consideration.
\section{Bistatic sounding}
To analyze the bistatic sounding we assume, by analogy with \cite{Berngardt_2000}
that trajectories of the transmitted and received signal are spaced
far enough from each other. This means that the angle between the
incident and scattered ray bigger than the angular size of the area
of the stationary phase (\ref{eq:WR_RgRg}). The case of smaller angle
the problem degenerates into the monostatic case, already analyzed
(\ref{eq:autocorfunction1-2-2-2}). It should be noted that in presence
of refraction this condition in general is not connected with the
relative position of transmitter and receiver (in opposite to refraction-free
situation \cite{Berngardt_2000}), but is connected only with the
angle between the trajectories of transmitted and scattered signal
at the point of scattering.
\subsection{Single trajectory pair: initial equation}
Lets analyze the case of single trajectory pair: there is only one
propagation path from the transmitter to scattering point, and only
one path from receiver to the scattering point. And they differ.
We consider the scattered signal as the integral over the ray coordinates
$\overrightarrow{\Lambda}_{t}$ associated with the transmitter and
over the spatial coordinates $\overrightarrow{r}$ of the scatterer
position. In this case, the scattering signal (\ref{eq:signal_scattering_narrowband})
leads to the following correlation function $P(t,\Delta T)$:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\int\widetilde{\Phi}(\overrightarrow{r};\overrightarrow{k},\Delta T+T_{t}-T_{t}')\cdot\\
\cdot W\left(t,\left\{ T_{r}+T_{t}\right\} c,\Delta T,\left\{ \left(T_{r}+T_{t}\right)-\left(T_{r}'+T_{t}'\right)\right\} c\right)D(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})D(\overrightarrow{\Lambda}_{r}',\overrightarrow{\Lambda}_{t}',\omega_{0})\cdot\\
\cdot e^{i\overrightarrow{k}\left(\overrightarrow{r}-\overrightarrow{r}'\right)}e^{ik_{0}\Psi(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t})}e^{-ik_{0}\Psi(\overrightarrow{\Lambda}_{r}',\overrightarrow{\Lambda}_{t}')}\cdot\\
\cdot\delta(\overrightarrow{r}-\overrightarrow{R}(\overrightarrow{\Lambda}_{t}))\delta(\overrightarrow{r}'-\overrightarrow{R}(\overrightarrow{\Lambda}_{t}'))F^{-1}(\overrightarrow{\Lambda}_{r}(\overrightarrow{r}))F^{-1}(\overrightarrow{\Lambda}_{r}'(\overrightarrow{r}'))d\overrightarrow{\Lambda}_{t}d\overrightarrow{r}d\overrightarrow{\Lambda}_{t}'d\overrightarrow{r}'d\overrightarrow{k}
\end{array}\label{eq:RLU-result-3D-3}
\end{equation}
\subsection{Single trajectory pair: obtaining of the radar equation}
Integrating delta-functions over the $\overrightarrow{\Lambda}_{t},\overrightarrow{\Lambda}_{t}'$
in (\ref{eq:RLU-result-3D-3}) will result to $\overrightarrow{\Lambda}_{t}(\overrightarrow{r})$
and $\overrightarrow{\Lambda}_{t}'(\overrightarrow{r}')$ functions
under the integral and corresponding Jacobian $F^{-1}$:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\int\widetilde{\Phi}(\overrightarrow{r};\overrightarrow{k},\Delta T+T_{t}-T_{t}')\\
W\left(t,\left\{ T_{r}+T_{t}\right\} c,\Delta T,\left\{ \left(T_{r}+T_{t}\right)-\left(T_{r}'+T_{t}'\right)\right\} c\right)D(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})D(\overrightarrow{\Lambda}_{r}',\overrightarrow{\Lambda}_{t}',\omega_{0})\\
e^{i\overrightarrow{k}\left(\overrightarrow{r}-\overrightarrow{r}'\right)}e^{ik_{0}\Psi(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t})}e^{-ik_{0}\Psi(\overrightarrow{\Lambda}_{r}',\overrightarrow{\Lambda}_{t}')}\\
\left|\gamma_{3}\right|^{2}F_{4}^{-1}(\overrightarrow{r}(\overrightarrow{k}),\overrightarrow{r}'(\overrightarrow{k}))d\overrightarrow{r}d\overrightarrow{r}'d\overrightarrow{k}
\end{array}\label{eq:RLU-result-3D-3-1}
\end{equation}
where
\begin{equation}
F_{4}(\overrightarrow{r},\overrightarrow{r}')=F(\overrightarrow{\Lambda}_{r}(\overrightarrow{r}))F(\overrightarrow{\Lambda}_{r}'(\overrightarrow{r}'))F(\overrightarrow{\Lambda}_{t}(\overrightarrow{r}))F(\overrightarrow{\Lambda}_{t}'(\overrightarrow{r}'))\label{eq:F4}
\end{equation}
\[
\gamma_{3}=\left(2\pi\right)^{n/2}e^{i\frac{\pi}{4}sgn(S'')}=\left(2\pi i\right)^{3/2}
\]
is additional multiplier from MSP \cite{Fedoryuk_1989}.
We can integrate the expression (\ref{eq:RLU-result-3D-3-1}) over
the $d\overrightarrow{r}$ and $d\overrightarrow{r}'$ by MSP separately.
Integrating over $d\overrightarrow{r}$ will lead to the following
stationary phase condition:
\begin{equation}
\overrightarrow{k}+k_{0}\frac{\partial}{\partial\overrightarrow{r}}\psi_{t}(\overrightarrow{\Lambda}_{t})+k_{0}\frac{\partial}{\partial\overrightarrow{r}}\psi_{r}(\overrightarrow{\Lambda}_{r})=0
\end{equation}
Taking into account that $\overrightarrow{r}=\overrightarrow{R_{r}}(\overrightarrow{\Lambda}_{r})=\overrightarrow{R_{t}}(\overrightarrow{\Lambda}_{t})$
and taking into account the properties of hamiltonian trajectories
(\ref{eq:gam_eq_solution}), the stationary phase condition will take
the clear look that shows the selective properties of the scattering:
\begin{equation}
\overrightarrow{k}+k_{0}\overrightarrow{P}_{t}(\overrightarrow{\Lambda}_{t})+k_{0}\overrightarrow{P}_{r}(\overrightarrow{\Lambda}_{r})=0\label{eq:WD-conditon-2st}
\end{equation}
Doing the same over $d\overrightarrow{r}'$and assuming the stationary
point $\overrightarrow{r}^{0}$ to be only one for each $\overrightarrow{k}$
we can transform the radar equation (\ref{eq:RLU-result-3D-3-1})
to the form:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\int\widetilde{\Phi}(\overrightarrow{r};\overrightarrow{k},\Delta T)W\left(t,\left(T_{r}+T_{t}\right)\frac{c}{2},\Delta T,0\right)\\
V_{R}^{2}(\overrightarrow{r}_{SF})D_{\Sigma}(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})\left|\gamma_{3}\right|^{2}F_{4}^{-1/2}(\overrightarrow{r}(\overrightarrow{k}),\overrightarrow{r}(\overrightarrow{k}))d\overrightarrow{k}\\
\\
\end{array}\label{eq:RLU-result-3D-spaceK}
\end{equation}
From qualitative considerations, the effective scattering volume $V_{R}$
should be close to that obtained for refraction-free case \cite{Berngardt_2000},
but instead of product $R_{t}R_{s}$ there should be the curvature
radii (as previously shown, they will automatically appear in the
two-dimensional stationary phase method over the angles (\ref{eq:WR_RgRg})):
\begin{equation}
V_{R} \approx \left\{ R_{GM,t}R_{Gm,t}R_{GM,r}R_{Gm,r}\right\} ^{1/2}k_{0}^{-3/2}(\overrightarrow{P}_{t}\times\overrightarrow{P}_{r})^{-1}\left(R_{t}+R_{r}\right)^{-1/2}\label{eq:WR-3D}
\end{equation}
The effective scattering volume $V_{R}$ is determined as:
\begin{equation}
V_{R}=\left|\aleph\right|^{-1/2}\label{eq:WR_def}
\end{equation}
where
\begin{equation}
\aleph=\left\{ \frac{\partial}{\partial r_{j}}\left\{ k_{0}P_{t,i}+k_{0}P_{r,i}+k_{i}\right\} \right\} _{\overrightarrow{r}=\overrightarrow{r}_{0}}\label{eq:N_values}
\end{equation}
It is necessary to note that this condition corresponds to locally-homogeneous
trajectories. As can be seen from Fig.\ref{fig:2}, in the E-layer
(heights 100-130 km) both trajectories (transmitter-scatterer and
scatterer-receiver) can be considered as locally-homogeneous ones,
and correspondingly one can use a local formulation of the Wolf-Bragg
conditions (\ref{eq:WD-conditon-2st}). For the rest of the points,
in which the trajectory of locally-inhomogeneous, this approach is
not entirely correct and should be investigated separately. Expectedly,
that in this case the effective scattering volume will be less powerful,
and will be determined by the differential of the refraction index
along the propagation trajectory (similar to the case of the monostatic
scattering).
It is easy to show that in a refraction-free case the (\ref{eq:WR-3D})
transforms to the formulas obtained in \cite{Berngardt_2000}.
From the formulas (\ref{eq:WD-conditon-2st},\ref{eq:N_values}) it
is obvious that
\begin{equation}
\aleph=\frac{\partial k_{i}(\overrightarrow{r})}{\partial r_{j}}\label{eq:JacCorrespondency}
\end{equation}
are the Lam\'{e} coefficients for the transition to the coordinate system
associated with the wave vectors $\overrightarrow{k}$ from the coordinate
system associated with spatial vectors of stationary points $\overrightarrow{r}$
(in the case that this transformation is bijection).
This means that in case of bijection $\overrightarrow{k}\rightarrow\overrightarrow{r}$
the integral over stationary points in the wave vectors space (\ref{eq:RLU-result-3D-spaceK})
can be represented as an integral over space $\overrightarrow{r}$
, where the exact expression for $V_{R}$ (\ref{eq:WR-3D}) is not
important:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\int\widetilde{\Phi}(\overrightarrow{r},t-T_{t};\overrightarrow{k}(\overrightarrow{r}),\Delta T)W\left(t,\left(T_{r}+T_{t}\right)c,\Delta T,0\right)\\
\left|\gamma_{3}\right|^{2}D_{\Sigma}(\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\overrightarrow{\Lambda}_{r},\overrightarrow{\Lambda}_{t},\omega_{0})F_{4}^{-1/2}(\overrightarrow{r},\overrightarrow{r})d\overrightarrow{r}
\end{array}\label{eq:RLU-result-3D-spaceR}
\end{equation}
In this expression explicitly highlighted the dependence of the power
of the scattered signal from the scattering angle (\ref{eq:WD-conditon-2st})(in the refraction-free
case it is determined by the distance between the transmitter and
receiver) and from the Gaussian curvature of the beams $F_{4}^{1/2}$ (in the refraction-free
case it is determined by the distance to the investigated volume from
the receiver and transmitter, respectively). So, the effective weight
volume $V_{R}$ is defined by Jacobian for the transition from the wave vector
$\overrightarrow{k}$ space to usual space $\overrightarrow{r}$,
and is not important for spatial integral representation (\ref{eq:RLU-result-3D-spaceR}).
The shape of $V_{R}$ is important when one estimate the homogenety
scale for the media, necessary for the obtained formulas to be valid.
Also its exact shape should be important when the scattering is produced
by a single spatial harmonic (for example, during radioacoustical sounding
of the ionosphere \cite{Lataitis_1992}) and wavevector representation (\ref{eq:RLU-result-3D-spaceK})
may be more convenient.
\subsection{Multiple trajectories pairs}
If there are exist multiple propagation trajectories from the transmitter
to the scatterer ($M$ trajectories) and multiple paths from the scatterer
to the receiver ($N$ trajectories) the radar equation becomes more
complex than (\ref{eq:RLU-result-3D-spaceR}).
In this case integration over $\overrightarrow{\Lambda}_{t},\overrightarrow{\Lambda}_{t}'$
in (\ref{eq:RLU-result-3D-3-1}) will remove the delta function, but
produce some cross-sums over the trajectories, replacing the function
arguments with the corresponding functions $\overrightarrow{\Lambda}_{t,m}(\overrightarrow{r}),\overrightarrow{\Lambda}_{t,m'}'(\overrightarrow{r}'),\overrightarrow{\Lambda}_{r,n}(\overrightarrow{r}),\overrightarrow{\Lambda}_{r,n'}'(\overrightarrow{r}')$
(calculated based on the delta-function arguments), where indexes
$n,n',m,m'$ mark trajectories :
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\sum\limits _{n,n'\in N,m,m'\in M}\int\widetilde{\Phi}(\overrightarrow{r},t-T_{t,m};\overrightarrow{k},\Delta T+T_{t,m}-T_{t,m'}')\cdot\\
\cdot W\left(t,\left(T_{r,n}+T_{t,m}\right)c,\Delta T,\left(\left(T_{r,n}+T_{t,m}\right)-\left(T_{r,n'}'+T_{t,m'}'\right)\right)c\right)D_{\Sigma}(\overrightarrow{\Lambda}_{r,n},\overrightarrow{\Lambda}_{t,m},\overrightarrow{\Lambda}_{r,n'}',\overrightarrow{\Lambda}_{t,m'}',\omega_{0})\cdot\\
\cdot F_{4,n,n',m,m'}^{-1/2}(\overrightarrow{r},\overrightarrow{r'})\cdot e^{i\overrightarrow{k}\left(\overrightarrow{r}-\overrightarrow{r}'\right)}e^{ik_{0}\Psi_{n,m}(\overrightarrow{\Lambda}_{r,n},\overrightarrow{\Lambda}_{t,m})}e^{-ik_{0}\Psi_{n',m'}(\overrightarrow{\Lambda}_{r,n'}',\overrightarrow{\Lambda}_{t,m'}')}d\overrightarrow{r}d\overrightarrow{r}'d\overrightarrow{k}
\end{array}\label{eq:RLU-result-3D-3-1-2}
\end{equation}
where
\begin{equation}
F_{4,n,n',m,m'}(\overrightarrow{r},\overrightarrow{r'})=F(\overrightarrow{\Lambda}_{r,n}(\overrightarrow{r}))F(\overrightarrow{\Lambda}_{r,n'}'(\overrightarrow{r}'))F(\overrightarrow{\Lambda}_{t,m}(\overrightarrow{r}))F(\overrightarrow{\Lambda}_{t,m'}'(\overrightarrow{r}'))
\end{equation}
In this case the condition of the stationary phase for each pair of
trajectories will take an already discussed form (\ref{eq:WD-conditon-2st}):
\begin{equation}
\overrightarrow{k}+k_{0}\overrightarrow{P}_{t}(\overrightarrow{\Lambda}_{t,m})+k_{0}\overrightarrow{P}_{r}(\overrightarrow{\Lambda}_{r,m})=0\label{eq:Wolf-Bragg-equation-bistatic}
\end{equation}
Assuming the stationary point $\overrightarrow{r}_{n,m}^{0}$ to be
only one for each given $\overrightarrow{k}$ and for each selected
unordered pair of trajectories $(n,m)$ the radar equation will take
the form:
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\sum\limits _{n,n'\in N,m,m'\in M,}\int\widetilde{\Phi}(\overrightarrow{r},t-T_{t,m};\overrightarrow{k},\Delta T+T_{t,m}-T_{t,m'}')\cdot\\
\cdot W\left(t,\left(T_{r,n}+T_{t,m}\right)c,\Delta T,\left(\left(T_{r,n}+T_{t,m}\right)-\left(T_{r,n'}'+T_{t,m'}'\right)\right)c\right)D_{\Sigma}(\overrightarrow{\Lambda}_{r,n},\overrightarrow{\Lambda}_{t,m},\overrightarrow{\Lambda}_{r,n'}',\overrightarrow{\Lambda}_{t,m'}',\omega_{0})\cdot\\
\cdot\left|\gamma_{3}\right|^{2}F_{4,n,n',m,m'}^{-1}(\overrightarrow{r}(\overrightarrow{k}),\overrightarrow{r'}(\overrightarrow{k}))\\
\cdot e^{i\overrightarrow{k}\left(\overrightarrow{r}_{n,m}^{0}-\overrightarrow{r}_{n',m'}^{0}\right)}e^{ik_{0}\Psi(\overrightarrow{\Lambda}_{r,n,m},\overrightarrow{\Lambda}_{t,n,m})}e^{-ik_{0}\Psi(\overrightarrow{\Lambda}_{r,n',m'}',\overrightarrow{\Lambda}_{t,n',m'}')}V_{R,n,m}^{2}d\overrightarrow{k}
\end{array}\label{eq:RLU-result-3D-several-traj}
\end{equation}
When transforming to the space coordinate system the effective weight
volume \textbf{$V_{R,n,m}$} disappears, because it is equal to the
Jacobian of the transition from one coordinate system ($\overrightarrow{k}$)
to another ($\overrightarrow{r}$) (\ref{eq:WR_def},\ref{eq:JacCorrespondency}):
\begin{equation}
\begin{array}{c}
P(t,\Delta T)=\sum\limits _{n,n'\in N,m,m'\in M}\int\widetilde{\Phi}(\overrightarrow{r},t-T_{t,m};\overrightarrow{k},\Delta T+T_{t,m}-T_{t,m'}')\cdot\\
\cdot W\left(t,\left(T_{r,n}+T_{t,m}\right)c,\Delta T,\left(\left(T_{r,n}+T_{t,m}\right)-\left(T_{r,n'}'+T_{t,m'}'\right)\right)c\right)D_{\Sigma}(\overrightarrow{\Lambda}_{r,n},\overrightarrow{\Lambda}_{t,m},\overrightarrow{\Lambda}_{r,n'}',\overrightarrow{\Lambda}_{t,m'}',\omega_{0})\cdot\\
\cdot\left|\gamma_{3}\right|^{2}F_{4,n,n',m,m'}^{-1/2}(\overrightarrow{r},\overrightarrow{r})e^{i\overrightarrow{k}\left(\overrightarrow{r}_{n,m}^{0}-\overrightarrow{r}_{n',m'}^{0}\right)}e^{ik_{0}\Psi(\overrightarrow{\Lambda}_{r,n,m},\overrightarrow{\Lambda}_{t,n,m})}e^{-ik_{0}\Psi(\overrightarrow{\Lambda}_{r,n',m'}',\overrightarrow{\Lambda}_{t,n',m'}')}d\overrightarrow{r}\\
\\
\end{array}\label{eq:RLU-result-3D-several-traj-1}
\end{equation}
\section{Resulting SuperDARN radar equation}
In Fig.\ref{fig:3}B are shown the geometrooptical trajectories in
a model of spherically stratified ionosphere. From the figure it is
seen that within the region bounded by the caustics, there is a region
in which it is possible to reach the same point by multiple trajectories
(In Fig.\ref{fig:3}B the region is limited by the black triangle).
The upper limit for this region is the caustic associated with reflection
from the ionosphere by the lower ray, the left border is the caustic
associated with the boundary of the dead zone (i.e. ground backscatter
position), and the right border is the trajectory with the lowest
possible elevation angle.
Qualitative analysis has shown in the simple model case that the number
of trajectories to the each point, lying inside this region can be
two or three: the first is direct propagation trajectory, the second
is the trajectory reflected from the the ionosphere by lower ray,
and the last one is the trajectory reflected from the ionosphere by
the upper ray. Thus, in this area the case is possible when the signal
propagates towards and backwards the scatterer by substantially different
trajectories. The limit case for this effect is the area of caustics,
where to the same point at the same time comes a large number of rays,
that leads to significant gain of the signal.
In this paper we do not consider the caustic regions, but the area
inside the area bounded by the caustics can be considered.
The region, where all the 3 rays exist and intersect is marked by
blue triangle in Fig.\ref{fig:3}B. The region where at least two
trajectories exists is marked by black triangle. So the scattering
problem in the region can be considered as equivalent of bistatic
sounding problem.
In this case of several possible propagation trajectories from the
transmitter to the scatterer a situation arises that is equivalent
to the situation of bistatic sounding (\ref{eq:RLU-result-3D-several-traj},\ref{eq:RLU-result-3D-several-traj-1}).
This case does not arise in the refraction-free case, since in refraction-free
case there is only single trajectory exists between two points.
Lets consider the case of multiple trajectories in case of monostatic
sounding that is shown in Fig.\ref{fig:3}C. It can be seen that on
the trajectories, allowing the reflection from the ionosphere (i.e.,
at distances above 0.5 hop) there can be a few (at least two) signal
propagation trajectories to the scattering point. The trajectory marked
with 1 is the path of direct propagation (reflection-free trajectory).
The paths marked by 2,3 are the trajectories that includes reflection
from the ionosphere.
Lets obtain the radar equation in the case when all three trajectories
exists. When analyzing the signal autocorrelation function it is convenient
to analyze the simpliest case, valid for SuperDARN radars.
It is the case when selection over the delays is made with short enough
sounding pulses, spaced by large enough delay between each other.
The pulses are short enough to differ two trajectories by delay: $\left(T_{r,n}+T_{t,m}\right)-\left(T_{r,n'}'+T_{t,m'}'\right)>\Delta T_{A}$
where unordered pairs $(n,m)\neq(n',m')$. This condition also compensate
the additional caustic phase, that arise when wave reflects from the
ionosphere, and the caustic phase does not affect on resulting radar
equation.
In this case the short sounding pulse provides selection of correlating
trajectories when the difference between propagation delays is larger
than pulse duration. The final radar equation, after taking into account
symmetry considerations, becomes:
\begin{equation}
\begin{array}[t]{l}
\widetilde{P}(t,\omega)=\sum\limits _{n\in N}\int\limits _{|J_{1}/\Delta R_{A}|>1}\widetilde{\widetilde{\Phi}}\left(\overrightarrow{r};-\widehat{P}_{n}^{H}(\overrightarrow{r})k,\nu\right)A_{n}^{H}(\overrightarrow{r})\cdot\\
\,\,\cdot\widetilde{\widetilde{W}}_{2,t}\left(\hat{\Lambda}_{n}^{H}(\overrightarrow{r}),S_{n}(\overrightarrow{r}),\omega-\nu,\left(kP(\overrightarrow{r},\omega_{0})-2k_{0}P^{2}(\overrightarrow{r},\omega_{0})\right)-\nu/c\right)k^{2}dkd\nu d\overrightarrow{r}+\\
\\
+\int\limits _{|J_{1}/\Delta R_{A}|<1}\widetilde{\widetilde{\Phi}}\left(\overrightarrow{R}(\hat{\Lambda},S_{SF}(\hat{k} _{SF}^{I}k));\hat{k} _{SF}^{I}k,\nu\right)A^{I}(\overrightarrow{R}(\hat{\Lambda},S_{SF}(\hat{k} _{SF}^{I}k)))\cdot\\
\,\cdot\widetilde{W}_{2,t}\left(\hat{\Lambda},S_{SF}(\hat{k} _{SF}^{I}k),\omega-\nu,0\right)k^{2}dkd\nu d\hat{\Lambda}+\\
\\
+\sum\limits _{n,m\in N,n>m}\int\widetilde{\widetilde{\Phi}}(\overrightarrow{r};\overrightarrow{k}_{n,m}(\overrightarrow{r}),\nu)2(1+cos(\nu(T_{m}(\overrightarrow{r})-T_{n}(\overrightarrow{r}))))\cdot\\
\cdot A_{n,m}(\overrightarrow{r})\widetilde{W}\left(t,c(T_{n}(\overrightarrow{r})+T_{m}(\overrightarrow{r})),\omega-\nu,0\right)d\nu d\overrightarrow{r}
\end{array}\label{eq:final-SD-RLU3}
\end{equation}
where:
\begin{equation}
\begin{array}[t]{l}
A_{n}^{H}(\overrightarrow{r})=\left(2\pi\right)^{2} k_{0}^{4} \left\{ \frac{2}{PF}J_{\hat{P}\rightarrow\hat{\Lambda}}\right\} \\
A^{I}(\overrightarrow{r})=\left(2\pi\right)^{3} k_{0}^{4} \left|g_{r}^{*}(\widehat{\Lambda},\omega_{0})g_{t}(\widehat{\Lambda},\omega_{0})\right|^{2}J_{1}^{2}J_{\hat{P}\rightarrow\hat{\Lambda}}\\
A_{n,m}(\overrightarrow{r})=\frac{\left(2\pi\right)^{3} k_{0}^{4}}{F_{n}F_{m}}\left|g_{r}^{*}(\widehat{\Lambda}_{n}(\overrightarrow{r}),\omega_{0})g_{t}(\widehat{\Lambda}_{m}(\overrightarrow{r}),\omega_{0})\right|^{2}
\end{array}
\end{equation}
are the multipliers that strongly affected by background ionosphere
and define power changes due to geometrooptical and focusing effects.
Here all the obvious arguments are hidden, and $J_{1},J_{\hat{P}\rightarrow\hat{\Lambda}},\widetilde{\widetilde{W}}_{2,t},\widetilde{W},F$
and $V_{\Lambda}$ are defined by (\ref{eq:J1},\ref{eq:J_P_L},\ref{eq:W_2t},\ref{eq:weight-volume-def},\ref{eq:dR_dL},\ref{eq:WR_RgRg})
correspondingly. $N$ is the maximal number of trajectories to reach
the same point at the same time. In analyzed case $N=[1,2,3]$: reflection-free
ray{[}1{]}, lower ray{[}2{]} and Pedersen ray{[}3{]} (see Fig.\ref{fig:3}C).
The radar equation (\ref{eq:final-SD-RLU3}) does not contain any
oscillating terms and is convenient for qualiative analysis or numerical
simulations of scattering process. In most cases, $\Phi(k)$ can be considered
a smooth function of $k$ at scale $\Delta \omega / (cP)$ and radar equation (\ref{eq:final-SD-RLU3}) can be simplified even more.
The first two terms in (\ref{eq:final-SD-RLU3}) are associated with
the propagating of the signal over the identical trajectories towards
the scatterer and backwards. For any given scattering point whether
first or second term will be valid, because the integration of the
first term is made over the spatial area where $|J_{1}/\Delta R_{A}|>1$
(locally-homogeneous trajectory part), and second - over the spatial
area with $|J_{1}/\Delta R_{A}|<1$ (locally-inhomogeneous trajectory
part).
The first term in the radar equation (\ref{eq:final-SD-RLU3}) corresponds
to the stationary points located at locally-homogeneous trajectories
(which is traditionally investigated in SuperDARN tasks or in refraction-free
case by VHF radars), the stationary point is marked by $H$ upper
index and is defined by (\ref{eq:ST_loc}). It's usage is limited
by locally-homogeneous condition $|J_{1}/\Delta R_{A}|>1$, for SuperDARN
case it is valid for very limited regions, including ionospheric E-layer
and F-layer maximums (see Fig.\ref{fig:2}). Exit angle $\hat{\Lambda}_{n}^{H}(\overrightarrow{r})$,
group path $S_{n}(\overrightarrow{r})$ and wave direction $\widehat{P}_{n}^{H}(\overrightarrow{r})$,
as well as $A_{n}^{H}(\overrightarrow{r})$ value, that correspond
to investigated point $\overrightarrow{r}$ can be calculated using
ray-tracing technique (\ref{eq:gam_eq_solution}). In refraction-free
case this term transforms to the refraction-free radar equation \cite{Berngardt_1999}.
The second term in the radar equation (\ref{eq:final-SD-RLU3}) corresponds
to the stationary points located at locally-inhomogeneous trajectories,
the stationary point is marked by $I$ upper index and is defined
by (\ref{eq:StatConditionScalar-LoclaKriv}). It's usage is limited
by locally-inhomogeneous condition $|J_{1}/\Delta R_{A}|<1$. For
SuperDARN case it is valid for most part of the F-layer ionosphere
(see Fig.\ref{fig:2}). This case has no equivalent in refraction-free
situation. To calculate it one need to find the point with most intensive
scattering $S_{SF}(\hat{\Lambda},k)$ that is defined by system (\ref{eq:StatConditionScalar-LoclaKriv}).
It looks complex task, but the simplest way to estimate it, is to
calculate the wavevector direction $\hat{k}_{SF}^{I}$ that in average
is perpendicular to wavefront according to the first condition in
(\ref{eq:StatConditionScalar-LoclaKriv}) for given exit angle $\hat{\Lambda}$
and over weight volume $W(..)$ size: $\int\limits _{ct/2-\Delta R_{A}}^{ct/2+\Delta R_{A}}\hat{k}_{SF}^{I}\left(\frac{\partial\overrightarrow{P}}{\partial\Lambda}\right)_{\bot\overrightarrow{P}}dS_{1}=0$
. Then one can numerically find the point $S_{SF}$, at which the
Wolf-Bragg condition (second condition in (\ref{eq:StatConditionScalar-LoclaKriv}),
$k\hat{k}_{SF}^{I}\overrightarrow{P}(\hat{\Lambda},S_{SF})-2k_{0}P^{2}(\hat{\Lambda},S_{SF})-\nu/c=0$)
is satisfied for given wavenumber $k$, exit angle $\hat{\Lambda}$
and calculated average wavefront perpendicular $\hat{k}_{SF}^{I}$. It
is important to note, that $S_{SF}$ should be $\in[ct/2-\Delta R_{A},ct/2+\Delta R_{A}]$,
or integral becomes neglectable because the carrier of weight volume
$W(...)$. The $A^{I}(\overrightarrow{r})$ value and group delay
$T_{n}(\overrightarrow{r})$ to the investigated point can be calculted
during ray-tracing.
The third term (cross-sums) in the radar equation (\ref{eq:final-SD-RLU3})
is associated with propagation towards the scatterer over one trajectory
and backwards to the receiver over another trajectory, and is an equivalent
of bistatic scattering\cite{Berngardt_2000}. The stationary point
is defined by (\ref{eq:Wolf-Bragg-equation-bistatic}). Wavevector
$\overrightarrow{k}_{n,m}(\overrightarrow{r})$ can be found using
Wolf-Bragg condition (\ref{eq:Wolf-Bragg-equation-bistatic}) by ray-tracing
technique (\ref{eq:gam_eq_solution}) over two different trajectories
$n,m$ to the investigated point $\overrightarrow{r}$. $A_{n,m}(\overrightarrow{r})$
value, group delays $T_{n}(\overrightarrow{r}),T_{m}(\overrightarrow{r})$
and exit angles $\Lambda_{n}(\overrightarrow{r}),\Lambda_{m}(\overrightarrow{r})$
to the investigated point can be also calculated during ray-tracing.
This case does not arise in refraction-free situation. It should be
noted that existence of the third term was considered and discussed
for a static scatterer in \cite{Kravcov_1980}.
To take into account a weak absorption $\chi$ during radiowave propagation,
for example in D-layer \cite{Ginzburg_1970,Settimi_2014,Settimi_2015},
one should include the multiplier $e^{-2\int_{0}^{cT_{n}}\chi(\overrightarrow{R}(\hat{\Lambda}_{n},S',\omega_{0}))dS'}$
into each of $A_{n}^{H}(\overrightarrow{r}),A^{I}(\overrightarrow{r})$
and the multiplier $e^{-\int_{0}^{cT_{n}}\chi(\overrightarrow{R}(\hat{\Lambda}_{n},S',\omega_{0}))dS'-\int_{0}^{cT_{m}}\chi(\overrightarrow{R}(\hat{\Lambda}_{m},S',\omega_{0}))dS'}$
into $A_{n,m}(\overrightarrow{r})$.
\section{Discussion and conclusion}
Based on hamiltonian optics approach and ray representation of Green's
function we obtained the scalar radar equation (\ref{eq:final-SD-RLU3})
that is suitable for taking into account the refraction effects in
SuperDARN data (autocorrelation functions of the received signals)
obtained by these radars within the first hope distances (i.e. before
scattering of the signal from the ground). The main conclusions from
this radar equation are the following.
1. The radar equation shows that the integral kernel that relates
the measured correlation function of the signal with the correlation
function of the ionospheric irregularities, is the weight volume (or
ambiguity function) $W(...)$(\ref{eq:weight-volume-def}), the widely
studied for refraction-free case \cite{Lehtinen_1986,Berngardt_1999,Berngardt_2000}.
This weight volume is defined by the shape of a sounding signal and
by signal processing techniques.
Radar equation (\ref{eq:final-SD-RLU3}) is linear over $W(...)$, so making any linear transformation of the scattered signal correlation function or spectral power
leads just to equivalent transformation of the weight volume $W(...)$ \cite{Lehtinen_1986,Berngardt_2013} and does not affect on the structure
of the radar equation.
This allows, for example, to prove the validity
of using at SuperDARN radars the signal processing techniques and sounding sequences,
initially developed for refraction-free cases (\cite{Farley_1972,Berngardt_2015c},
etc.).
2. The radar equation shows that the spatial harmonics that produce
the main contribution to the scattering signal are determined by the
local wave vector and satisfy to the Wolf-Bragg condition (\ref{eq:ST_loc},\ref{eq:StatCondition-Local-mod}).
This is very close to refraction-free case \cite{Berngardt_1999,Berngardt_2000}.
This also confirms with the results of \cite{Gillies_2011}, and
allows to take it into account refraction effects in the velocities
of ionospheric irregularities measured by SuperDARN radars.
3. In contrast to the refraction-free case, the refraction significantly
affects the scattered signal amplitude. At first, this effect leads
to amplitude changes related to the focusing of the signal during
propagation (multipliers $F^{-1/2}$) in the region of the scattering.
The second effect is associated with attenuation of the scattered
signal due changing the wave length of sounding signal at intervals
of the length of the sounding pulse (parameter $J_{1}$ in the equation).
4. Obtained radar equation is not exactly correct for the Pedersen
rays after the point of theirs reflection. This is associated with
considerable frequency distortion of the propagating signal in this
area ( higher-frequency components are not reflected by the ionosphere)
and, as a consequence, the inapplicability of the undistorted signal
propagation model.
5. The highest amplitude of the scattered signal occurs in regions
close to the areas of focusing, and in the areas where the refractive
index varies slightly along the propagation trajectories. These areas
correspond to the caustic trajectories, the areas near the local maximum
of the electron density (Pedersen rays) and the areas where the ionospheric
refraction can be neglected (the areas below the E-layer). Outside
these areas, the amplitude of the scattered signal is reduced and
can be compensated only by a high aspect dependence of scattering
irregularities. The maximal scattering amplitude at Pedersen ray have
been detected, for example, as a result of interpretation of the first
Russian-Ukrainian experiment using EKB radar and UTR-2 radiotelescope
\cite{Berngardt_2015b}.
6. When signal propagates in an inhomogeneous ionosphere there are
areas in which the signal reaches the point of scattering by several
(two or more) different trajectories simultaneously. We modeled the
orthogonality conditions of wave vector to the magnetic field, important
for scattering by field-aligned irregularities, with taking into account
the structure of the Earth magnetic field (IGRF reference model used)
and model ionosphere (IRI-2012 reference model used). Modeling shows
that the bistatic mechanism should be important for polar radars looking
equatorward (see Fig.\ref{fig:6}). In spite in this case the angle
between propagation trajectory and magnetic field is far from orthogonality (about
50 degrees), the wavevector of the scattering irregularities $\overrightarrow{k}$
is orthogonal to magnetic field $\overrightarrow{B}$. This mechanism
allows to interpret strong field-aligned scattering even at very high
aspect angles between trajectory and magnetic field.
7. The first two mechanisms can be considered as a result of monostatic
scattering similar to the refraction-free case \cite{Tatarsky_1967,Ishimaru_1999,Berngardt_1999}.
The third mechanism should be considered as an equivalent of bistatic
sounding in refraction-free case \cite{Tatarsky_1967,Ishimaru_1999,Berngardt_2000}. This case
apparently has not been considered in SuperDARN tasks. The existence
of a third mechanism has been proposed and investigated for the case
of stationary irregularities for example in \cite{Kravcov_1980}.
It should be noted that in the last case, despite the possibility
of differencing the trajectories over the time of arrival, the measured
spectrum is distorted by $2(1+cos(\nu(T_{m}-T_{n})))$ multiplier
due to trajectory pairs with exactly the same group length. Conversely,
the existence of such distortions in the received signal spectral power allows to
detect the multiple propagation case.
\section{acknowledgments} The work of OB and AP has been done under
support of FSR program II.12.2.2. The work of KK has been done under
support of RFBR grant \#16-05-01006a. The numerical simulations have
been performed with High-performance computing cluster "Academician
V.M. Matrosov" (http://hpc.icc.ru)
|
\section{Introduction}
The 3-body problem is one of the most fundamental and well studied problems in Celestial Mechanics. A key result for this problem is that there exist only 5 relative equilibria, and that these exist for all levels of angular momentum \cite{euler,lagrange}. The properties of these special solutions have been deeply studied, and have motivated significant research in mechanics and dynamics. A hallmark of the classical problem is that the bodies are considered to be point masses, with no restrictions on how close they can come to each other. A recent variation of this problem has been posed that removes this one restriction \cite{scheeres_minE}, and supposes that these bodies can be rigid bodies with finite density, and hence have limits on their proximity. Such ``Full Body'' systems inherit the fundamental symmetries of the $N$-body problem \cite{scheeres_F2BP}, however they also demand that the rotational angular momentum, energy and dynamics of these rigid bodies be incorporated in the theory as well.
This paper studies the spherical 3-body problem under the assumption that the bodies are rigid and have finite density, and thus the separation between the bodies is constrained to be positive. This one change completely alters the character of the problem and, while the traditional Euler and Lagrange solutions still exist for large enough angular momentum values, a full 23 additional relative equilibria emerge from the analysis at all values of angular momentum, with a complex and rich bifurcation scheme.
The celestial mechanics of bodies with finite density and fixed shape can have dynamical evolution and relative equilibria that are quite distinct from that found in the classical Newtonian point mass $N$-body problem. These differences were previously explored in \cite{scheeres_minE} where several results were proven for the so-called ``Full Body Problem,'' in which the individual bodies are treated as rigid bodies with finite densities. Specifically, it was shown that, in opposition to the point mass $N$-body problem, the full body problem will always have a minimum energy configuration. Further, the number and variety of relative equilibria for that problem are greatly enhanced, and now include configurations where the bodies can rest on each other and configurations where different collections of resting bodies orbit each other, as well as the classical central configurations. One important aspect of this problem is that the existence and stability of configurations become a function of the total angular momentum of the system, a dependance that does not exist for the classical point mass $N$-body problem.
This paper studies the relative equilibria of one particular problem in the Full Body Problem (FBP) to completeness. Specifically, all relative equilibria of the planar spherical full 3-body problem, which consists of three spheres of equal density but arbitrary size, located in the plane perpendicular to the angular momentum.
The explicit methodology used was developed in \cite{scheeres_minE, scheeres_minE_chap}, and is fundamentally based on analysis of the amended potential as developed by Smale \cite{smaleI, smaleII} and motivated by observations from Arnold \cite{arnold1988mathematical}. The main theorem is stated and described at first, the problem is technically defined, then several results used to make the proof are listed, and finally all the detailed computations for the proof are given. Following the proof, a summary of the proof is provided, indicating how it establishes the theorem.
A main application of this result is to identify the stable states that can be physically achieved by a collection of self-gravitating bodies that can sustain contact. This situation happens in solar system dynamics when considering the physical nature of rubble pile asteroids \cite{fujiwara_science}, although there the number of individual grains can be quite large. Recent observations of comets, however, also show that they can be comprised of a few larger components that rest on each other and, given their changing spin rates due to outgassing, also mimic a system with changing angular momentum \cite{massironi2015two}.
A specific motivation from this current analysis would seek out natural situations that mimic the stable members of these configurations. As the existence of some of these stable configurations are somewhat unexpected, finding such configurations in nature would be especially interesting.
Additional avenues for exploration would be to expand the current analyses to study the dynamics of finite bodies as they interact with each other gravitationally and through impact. This sort of approach has been followed in the planetary sciences community in the study of rubble pile asteroids \cite{richardson_equilibrium, sanchez_icarus}. The current results can motivate the formation of stable shapes as a function of body morphology and total angular momentum.
In addition, the presence of finite densities and minimum energy configurations also enables the rigorous computation of energy limits for the Hill stability in the Full $N$-body problem \cite{IAU_namur, IAU_hawaii}, an avenue of further investigation for the current problem.
\section{Main Result}
\begin{theorem}
\label{theorem:1}
In the spherical full 3-body problem there exist a total of 28 distinct relative equilibrium configurations. No single class of relative equilibria exists or is energetically stable for all angular momentum, however at every value of angular momentum there exists at least 1 energetically stable relative equilibria. The pattern of relative equilibria can be fully represented in a bifurcation chart as the total angular momentum of the system varies from 0 to $\infty$.
\end{theorem}
The 28 different relative equilibria can be delineated in a few different ways. Figure \ref{fig:equilibria} shows these relative equilibria separated into 7 different classes, each with one to six distinct configurations. The figure shows 20 distinct configurations, with 8 of them having an alternate ordering not shown in the figure.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.4]{graphics/equilibria_2}
\caption{Examples of the 20 different equilibrium configurations are shown within different classes. Colors indicate whether or not the configuration can be stable for a range of angular momentum (green), for some special combinations of mass ratio and angular momentum (yellow), or if they are unstable for all angular momentum values and mass ratios (red). Stars indicate when a reordering of the masses provides another relative equilibrium. The detailed naming convention for the configurations is also introduced. }
\label{fig:equilibria}
\end{figure}
Figures \ref{fig:bif1}-\ref{fig:bif4} show the detailed sequences of bifurcations that occur as the total angular momentum of the system is increased. These diagrams are qualitative, but the sequence of bifurcations in the specific connected pathways are accurate and will be derived in the course of the proof. Note that the sequence of bifurcations keeps some of the configurations separated from each other. In other cases the sequence of bifurcations will change as the relative masses of the bodies are changed, in these cases multiple types of sequences are shown. Note in Fig.\ \ref{fig:bif2} that this particular sequence has at least one stable configuration for all values of angular momentum while none of the others has a stable configuration in the vicinity of $H= 0$. Also note that the sequences shown in Figs.\ \ref{fig:bif1}--\ref{fig:bif3} all have 2 stable configurations for large $H$.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.35]{graphics/bif_EA123}
\caption{Bifurcation diagram showing the possible branches relating to the ER123 configuration. Depending on the relative mass values two different bifurcation pathways exist. Within each pathway, the manner in which the EO configuration appears can shift between the two pathways shown. }
\label{fig:bif1}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.35]{graphics/bif_EA132}
\caption{Bifurcation diagram showing the possible branches relating to the ER132 configuration. ER132 links to the LR configuration, and always fissions into the same EA13-2 configuration. Depending on the relative mass values there can be a range of stable TR132 configurations. Within each pathway, the manner in which the EO configuration appears can also shift between the two pathways shown. The two different sequence need not have the same pattern of TR and EO bifurcations.}
\label{fig:bif2}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.35]{graphics/bif_EA213}
\caption{Bifurcation diagram showing the possible branches relating to the ER213 configuration. Depending on the relative mass values two different bifurcation pathways exist. Within each pathway, the manner in which the EO configuration appears can also shift between the two pathways shown. }
\label{fig:bif3}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.35]{graphics/bif_LO}
\caption{Bifurcation diagram showing the pathways to the LO configuration. }
\label{fig:bif4}
\end{figure}
\clearpage
\section{Problem Statement}
\subsection{The Full Body Problem}
A full body problem is defined as a set of $N$ rigid bodies that gravitationally attract each other and which have a finite density mass distribution, meaning that there are specific limits on how close they can come to each other \cite{scheeres_F2BP, scheeres_minE, scheeres_minE_chap}. The description of such a system can be directly incorporated into a Lagrangian framework where the coordinates ${\bf Q} = \left\{ Q_i; i = 1, 2, \ldots, 6N\right\}$ denote the absolute Cartesian coordinates of the bodies and the Euler angles that orient the bodies in space. The rigid body constraints place restrictions of the form $|Q_i - Q_j| \ge D_{ij}({\bf Q})$ on the system. The dynamics of the system can be described by a total Kinetic Energy and Gravitational Potential Energy and as it is an isolated system will conserve its total angular momentum, denoted as $\bfm{H}$, and can conserve its total energy, denoted as $E$, if internal dissipative forces are excluded.
The linear momentum can be removed to reduce the system to $6(N-1)$ coordinates that are purely relative to each other and an additional 3 degrees of freedom that orient the entire system with respect to inertial space. Denote the relative coordinates as ${\bf q} = \left\{ q_i; i = 1, 2, \ldots 6(N-1)\right\}$, noting that these can always be transformed to locally reformulate the constraints as $q_i \ge D_i({\bf q})$.
Thus, for the Full 3-Body Problem we have 12 twelve relative degrees of freedom between the three bodies. Of these only three are required to specify the relative positions of the bodies. The additional 9 correspond to each of the bodies having 3 degrees of freedom to orient themselves relative to the position configuration of the bodies. As we take the three bodies to be spheres, their relative orientation is not tracked, although we must still account for their rotational angular momentum and kinetic energy. Thus, for our purposes, our problem can be specified with only 3 degrees of freedom, plus the overall orientation of the system with respect to inertial space.
Before continuing we define the amended potential, which plays a fundamental role in the following.
\begin{definition}{\bf Amended Potential }
The Amended Potential is defined as the function ${\cal E} = \frac{H^2}{2 I_H} + \mathcal{U} $ where $H$ is the total angular momentum of the system, $I_H$ is the moment of inertia of the total system taken about a principal axis of the system, in general about the rotation axis $\hat{\bfm{H}}$ which points in the direction of the total angular momentum vector, and $ \mathcal{U} $ is the gravitational potential energy of the system. The terms $I_H$ and $ \mathcal{U} $ are functions only of the relative positions and attitudes of the bodies, and for $I_H$ their orientation relative to $\hat{\bfm{H}}$. The gradients of the Amended Potential with respect to the degrees of freedom equal the force exerted on that degree of freedom when at an equilibrium or resting configuration (\cite{arnold1988mathematical}, pp 66-67).
\end{definition}
\subsection{Spherical Full 3-Body Problem Statement}
Consider three bodies, ${\cal B}_i, i = 1,2,3$, each of which is a sphere of radius $R_i$ and, for convenience, assumed to have a common density $\rho$.
The positions of these bodies can be denoted in $\mathbb{R}^3$ by Cartesian position vectors $\bfm{r}_i$. The relative positions of these bodies are denoted as $\bfm{r}_{ij} = \bfm{r}_j - \bfm{r}_i$ and have the fundamental rigid body constraint $| \bfm{r}_{ij} | \ge (R_i + R_j)$ for $i\ne j$. This lower bound, due to the bodies having finite density, is what enables resting equilibria to occur. Each of the spheres can carry angular momentum in their spin rate, although due to their symmetry the specific orientation of these spheres are arbitrary in any frame. Thus, the internal relative configuration space of the system, $\bfm{q}$, is completely specified by only three quantities
\begin{eqnarray}
\bfm{q} & = & \left\{ r_{12}, r_{23}, r_{31} \ | \ r_{ij} \ge (R_i + R_j) \ \& \ |r_{ij} - r_{jk}| \le r_{ki} \le |r_{ij} + r_{jk}| \right\}
\end{eqnarray}
While the configuration of the system is uniquely defined by these distances, not all distances are allowable. This means that there are geometric constraints between some of the distances independent of the finite density assumption. Thus it is sometimes easier to define a unique configuration where the restriction is clearly obvious. One such is to specify the distances between two of the bodies and the angle between these two bodies centered on the third body (see Fig.\ \ref{fig:config}). Thus, denoting the bodies with the unique indexing $i, j, k$, the configuration can be specified as
\begin{eqnarray}
\bfm{q} & = & \left\{ r_{ij}, r_{jk}, \theta_{ki} \ | \ r_{lm} \ge (R_l + R_m) \right\}
\end{eqnarray}
where the final distance $r_{ki}$ can be explicitly computed from the cosine rule:
\begin{eqnarray}
r_{ki}^2 & = & r_{ij}^2 + r_{jk}^2 - 2 r_{ij} r_{jk} \cos\theta_{ki} \label{eq:d31}
\end{eqnarray}
Note that the angle $\theta_{ki}$ will also have constraints placed upon it, as the associated length must satisfy $r_{ki} \ge R_i + R_k$. These two expressions of the third degree of freedom, $\theta_{ki}$ or $r_{ki}$, will be used equivalently.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.4]{graphics/configuration}
\caption{Configuration of the system.}
\label{fig:config}
\end{figure}
There are additional degrees of freedom of the triad of bodies with respect to inertial space, which can be tracked by the unit vector of the angular momentum, $\hat{\bfm{H}}$, which are briefly discussed later.
The gravitational potential is equivalent to the 3-body point mass potential due to the symmetry of spherical mass distributions.
\begin{eqnarray}
\mathcal{U} & = & - {\cal G} \left[ \frac{M_1 M_2}{r_{12}} + \frac{M_2 M_3}{r_{23}} + \frac{M_3 M_1}{r_{31}} \right]
\end{eqnarray}
The moment of rotational inertia of each sphere is equal to $2 M_i R_i^2 / 5$. For a given placement of the three masses, the total inertia dyad of the system can be constructed as
\begin{eqnarray}
\bfm{I} & = & \frac{1}{M_1 + M_2 + M_3} \sum_{i=1}^2 \sum_{j=2}^3 M_i M_j \left[ r_{ij}^2 \bfm{U} - \bfm{r}_{ij} \bfm{r}_{ij} \right] + \sum_{i=1}^3 \frac{2 M_i R_i^2}{5} \bfm{U}
\end{eqnarray}
where $\bfm{U}$ is the identity dyad. Note that this form uses the Lagrange Identity and assumes that the center of mass is nominally at a zero point. The inertia matrix is orientable, but for the amended potential only its orientation relative to the constant angular momentum vector direction, $\hat{\bfm{H}}$, is needed. Dotting the dyad on both sides by this unit vector yields
\begin{eqnarray}
I_{H} & = & \frac{1}{M_1+M_2+M_3} \sum_{i=1}^{2}\sum_{j=2}^3 M_i M_j \left[ r_{ij}^2 - (\hat{\bfm{H}}\cdot\bfm{r}_{ij})^2 \right] + \sum_{i=1}^3 \frac{2 M_i R_i^2}{5}
\end{eqnarray}
The principal moments of inertia for a three point-mass system have the following relation: $I_{max} = I_{int} + I_{min}$. Furthermore the maximum moment of inertia of the point masses will always be perpendicular to the plane that the three bodies form \cite{greenwood_dynamics}. Thus, with our assumption that the body spins about its maximum moment of inertia, the quantities $(\hat{\bfm{H}}\cdot\bfm{r}_{ij}) = 0$ and the moment of inertia simplifies to
\begin{eqnarray}
I_{H} & = & M_1 M_2 r_{12}^2 + M_2 M_3 r_{23}^2 + M_3 M_1 r_{31}^2 + I_S \\
I_S & = & \frac{2}{5} M_1 R_1^2 + \frac{2}{5} M_2 R_2^2 + \frac{2}{5} M_3 R_3^2
\end{eqnarray}
For rotation about the intermediate and minimum moments of inertia, we note that $I_H$ will be strictly less than or equal to this above value, with equality between the intermediate and maximum only occurring when the minimum moment of inertia of the three particles (without the rigid sphere contributions) is 0. The maximum moment of inertia of the point masses can never be zero, due to the finite size of the particles.
\subsection{Normalization}
To simplify the discussion, normalize the system with a length and a mass scale. The length scale used is $R_T = R_1 + R_2 + R_3$, while the mass scale is $M_T = M_1 + M_2 + M_3$. Denote $m_i = M_i / M_T$, $r_i = R_i / R_T$, and $d_{ij} = r_{ij} / R_T$. In normalized coordinates the fundamental quantities take on the values
\begin{eqnarray}
\mathcal{U} & = & - \left[ \frac{m_1 m_2}{d_{12}} + \frac{m_2 m_3}{d_{23}} + \frac{m_3 m_1}{d_{31}} \right] \\
I_{H} & = & m_1 m_2 d_{12}^2 + m_2 m_3 d_{23}^2 + m_3 m_1 d_{31}^2 + I_S \label{eq:IH} \\
I_S & = & \frac{2}{5} m_1 r_1^2 + \frac{2}{5} m_2 r_2^2 + \frac{2}{5} m_3 r_3^2
\end{eqnarray}
with the angular momentum being normalized by the dividing factor $\sqrt{{\cal G} M_T^3 R_T}$ and the energy normalized by the dividing factor ${\cal G} M_T^2 / R_T $. For both $H$ and $E$ the same notational designation is kept for the normalized values.
The normalizations provide two identities:
\begin{eqnarray}
r_1 + r_2 + r_3 & = & 1 \\
m_1 + m_2 + m_3 & = & 1
\end{eqnarray}
There are also fundamental relationship between the $r_i$ and the $m_i$, assuming constant density.
\begin{eqnarray}
m_i & = & \frac{r_i^3}{r_1^3 + r_2^3 + r_3^3} \label{eq:massnorm} \\
r_i & = & \frac{m_i^{1/3}}{m_1^{1/3} + m_2^{1/3} + m_3^{1/3}}
\end{eqnarray}
\subsection{Parameterization of the Problem}
Any given variant of the F3BP can be identified with a point in a compact, 2-dimensional triangle, using either the masses or the radii. Plotting the radii $r_1, r_3$ or masses $m_1, m_3$ along two perpendicular axes each of them can only take values between 0 and 1, and that furthermore they will be bounded by a diagonal defined by $r_1 + r_3 \le 1$ or $m_1 + m_3 \le 1$. On the boundary of this equality $r_2 = m_2 = 0$. If, instead, a diagonal defined by $r_1+r_3 = r_{13} < 1$ or $m_1+m_3 = m_{13} < 1$ is drawn, then the value of the second radius or mass will equal $r_2 = 1 - r_{13}$ or $m_2 = 1 - m_{13}$. This also lends itself to a graphical description, shown in Fig.\ \ref{fig:triangle} for the masses. Every point within this triangle defines a unique F3BP in terms of its relative masses. In \cite{scheeres_minE} only the relative equilibria for the point $1/3, 1/3, 1/3$ was studied. This paper studies the bifurcation structure across the entire region, however due to the symmetry of the problem the study can be restricted to a specific region only. To that end, consider the restrictions
\begin{eqnarray}
0 \le m_3 \le m_2 \le m_1 \le 1 \\
0 \le r_3 \le r_2 \le r_1 \le 1
\end{eqnarray}
This region is shaded in Fig.\ \ref{fig:triangle}. There are 5 other equivalent triangles defined by reordering the different inequalities given above. The approach taken will be to exhaustively study all possible relative equilibria in the denoted region, the results of which can then be easily applied to all other regions.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.45]{graphics/triangle_shaded}
\caption{Triangle defined for the masses with the region of study shaded.}
\label{fig:triangle}
\end{figure}
With this convention, there are additional constraints for the masses and radii.
\begin{eqnarray}
\frac{1}{3} \le (r_1, m_1) \le 1 \\
0 \le (r_3, m_3) \le \frac{1}{3} \\
0 \le (r_3, m_3) \le (r_2, m_2) \le \frac{1}{2}
\end{eqnarray}
Previous research has exhaustively explored the bifurcation structure and properties for two general cases along the boundary of this triangle. One is at the point (1/3, 1/3, 1/3), when all masses and sizes are equal \cite{scheeres_minE}. In this case a more limited number of relative equilibria were found with a less complex bifurcation structure. The other case is for $m_3 = 0$, in essence just considering the two mass case with $0 \le m_2 \le 1/2 \le m_1 \le 1$, along the base of the triangle \cite{scheeres_minE}. In this region the number of relative equilibria are also much fewer and the bifurcation structure less complex.
For the problem we study, the spherical 3-body problem, we can easily just consider the planar motion of the system, with rotation occurring about a principal moment of inertia of the system. We note that the spheres contribute to the system's total angular momentum but have the same moment of inertia about any axis. In general we will assume that the system rotates about the maximum moment of inertia, but will justify this later.
\section{Background and Supporting Results}
A few definitions and supporting Lemmas are stated for use in this paper. Some of these are classical results while others have been considered more recently \cite{scheeres_minE, scheeres_minE_chap}, thus the proofs are only briefly reviewed to point out their salient features. Specific results for our current analysis are worked out in detail.
\begin{lemma}
\label{lemma:1}
The Total Energy of the system is conserved in the absence internal dissipation and equals $E = T_r + {\cal E}$, where $E$ is the total energy and $T_r$ is kinetic energy of the system components relative to each other, evaluated in the rotating frame with inertial angular velocity $\bfm{H} / I_H$.
\end{lemma}
\begin{proof}
For rotation about a principal axis of the system, ${\cal E}$ equals the amended potential as introduced by Smale \cite{smaleI, smaleII}, and specifically considered by Arnold for the 3-body problem in \cite{arnold1988mathematical}, pp. 66-67. For a system rotating about its principal axis the proofs in \cite{scheeres_minE_chap} apply, showing that the amended potential arises from a Routh reduction of the system. The Routhian is shown to have a Jacobi integral, which is identical to the total energy of the system.
\end{proof}
\begin{lemma}
\label{lemma:2}
The total energy of the system is strictly bounded from below by the amended potential: $\mathcal{ E} \le E$. If $E = \mathcal{ E}$, then $T_r = 0$. If the system is momentarily stationary ($T_r = 0$) and spins about a principal axis of inertia of the system, then $\mathcal{E} = { E}$. Thus the inequality is sharp and the lower bound can be achieved.
\end{lemma}
\begin{proof}
Lemma \ref{lemma:2} is proven in \cite{scheeres_minE_chap}. The proof establishes the inequality using the Cauchy Inequality applied to the angular momentum, and shows that it is sharp through direct construction.
\end{proof}
Another important feature of the system involves the existence of minimum energy configurations for full body systems. The following lemma establishes the existence of minimum energy states for all values of angular momentum. This result also provides the fundamental motivation for the current study.
\begin{lemma}
For a finite density distribution, the amended potential ${\cal E}$ has a global minimum for all values of angular momentum $H$.
\end{lemma}
\begin{proof}
The proof is given in \cite{scheeres_minE} and involves showing that ${\cal E}$ is compact and bounded over all possible values of the configuration space. This requires the finite density assumption, as this blocks individual point masses from coming arbitrarily close to each other. If body $i$ escapes to $\infty$ relative to bodies $j$ and $k$, then the amended potential takes on the value $\mathcal{E} = \mathcal{U}_{jk}$, and remains bounded in the interval $[- m_j m_k / (1-r_i), 0]$. If all three bodies escape to $\infty$ relative to each other then $\mathcal{E} = 0$.
\end{proof}
Given the definition of the amended potential and its properties relative to the total energy of the system, the relative equilibrium and energetic stability can be defined. Following this conditions under which these are satisfied are stated.
\begin{definition}{\bf Relative Equilibrium}
A given configuration $\bfm{q}^*$ is said to be a ``Relative Equilibrium'' if its internal kinetic energy is null ($T_r = 0$), meaning that ${\cal E} = E$ at an instant, and if it remains in this state over at least a finite interval of time.
\end{definition}
\begin{definition}{\bf Energetic Stability}
A given relative equilibrium $\bfm{q}^*$ is said to be ``Energetically Stable'' if any equi-energy deviation from that relative equilibrium requires a negative internal kinetic energy, $T_r < 0$, meaning that this motion is not allowed.
\end{definition}
\begin{lemma}
\label{lemma:3}
Consider a system with an amended potential ${\cal E}$ as defined above with $n$ degrees of freedom, $m$ of which are activated in such a way that only the variations $\delta q_j \ge 0$, $j = 1, 2, \ldots, m$ are allowed. The degrees of freedom $q_i$ for $m < i \le n$ are free.
The Necessary and Sufficient conditions for a system in a configuration $\bfm{q}^*$ to be in a relative equilibrium are that at this configuration:
\begin{enumerate}
\item
$T_r = 0 $
\item
${\cal E}_{q_j} \ge 0 \ \forall \ 1 \le j \le m$
\item
${\cal E}_{q_i} = 0 \ \forall \ m < i \le n $
\end{enumerate}
The Necessary and Sufficient conditions for a system in a relative equilibrium to be energetically stable are that:
\begin{enumerate}
\item
$ {\cal E}_{q_j} > 0 \ \forall \ 1 \le j \le m$
\item
$ \left[ \frac{\partial^2 {\cal E}}{\partial q_i \partial q_k} \right] > 0 \ \forall \ m < i,k \le n$
\end{enumerate}
\end{lemma}
\begin{proof}
The proof of Lemma \ref{lemma:3} is found in \cite{scheeres_minE_chap}. It relies on taking variations of the amended potential, asserting the principle of conservation of energy and using the general form of the Lagrange equations of motion for the full body system.
\end{proof}
Next a few results of relevance for the system are stated regarding bifurcation of relative equilibria and their stability.
Specific example bifurcations and their properties are shown in Fig.\ \ref{fig:bif_ex}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.35]{graphics/bif_ex}
\caption{Examples of bifurcations of interest in this problem, and their stability properties. }
\label{fig:bif_ex}
\end{figure}
\begin{definition}
{\bf Symmetric Bifurcation}
A bifurcation of two relative equilibria which follow a symmetric path relative to each other about a reflection line at changing values of angular momentum.
\end{definition}
\begin{definition}
{\bf H-Bifurcation}
An H-Bifurcation occurs when, under increasing angular momentum, a pair of relative equilibria appear in a degree of freedom $q$ that is not at a constraint. At its first appearance there must be a degeneracy of the form $\mathcal{E}_{qq} = 0$ that will generically disappear under increasing angular momentum.
\end{definition}
\begin{definition}
{\bf Fission} A collection of bodies in a relative equilibria with at least one active constraint is said to ``fission'' if, under an increase in angular momentum, the active constraint is released, meaning that a free relative equilibria intersects with it. Following fission the body may either transition into a new relative equilibrium without that active constraint or may no longer lie in any relative equilibrium associated with that configuration.
\end{definition}
\begin{definition}
{\bf Termination Fission}
A fission bifurcation where the relative equilibria disappear at higher values of angular momentum.
\end{definition}
\begin{definition}
{\bf Transition Fission}
A fission bifurcation where the relative equilibria continues with its constraint inactive at higher values of angular momentum.
\end{definition}
Now a particularly useful lemma is proven, which enables us to relate the stability of equilibrium points to how their coordinate changes as a function of angular momentum $H$.
\begin{lemma}
\label{theorem:X}
Assume a relative equilibria exists with a 1-1 relationship between a single degree of freedom $q$ and the angular momentum $H$, meaning that along the local family of relative equilibria as the angular momentum is changed only the degree of freedom $q$ changes. Then $\mathrm{sign}( \mathcal{E}_{qq} ) = \mathrm{sign}( \partial H/\partial q )$ at the relative equilibria. Thus, if $\partial H / \partial q < 0 $ the equilibrium point will be energetically unstable and if $\partial H / \partial q > 0 $ it could be stable, depending on the other degrees of freedom.
\end{lemma}
\begin{proof}
From the lemma statement it can be assumed that all other degrees of freedom lie in a relative equilibrium condition independent of the local value of $H$. Given this, assume that there exists a value $q^*$ such that the scalar equation $\left. \mathcal{E}_q\right|_* = - H^2 / (2 I_H^2) I_{Hq} + \mathcal{U}_q = 0$. This can be solved for $H^* = \left.\left(I_H \sqrt{ 2 I_{Hq} / \mathcal{U}_q}\right)\right|_*$, where the righthand side is a function of $q^*$ and by assumption is non-singular. Now consider a neighboring relative equilibrium at a different value of $H$ and hence $q$, with the values defined locally by the expansion
\begin{eqnarray*}
\mathcal{E}_q( H^* + \Delta H, q^* + \Delta q) & = & 0 + \left.\mathcal{E}_{qH}\right|_* \left.\frac{\partial H}{\partial q}\right|_* \Delta q + \left.\mathcal{E}_{qq}\right|_* \Delta q + \ldots
\end{eqnarray*}
Setting this to zero and solving for an arbitrary $\Delta q$ yields
\begin{eqnarray*}
\left.\mathcal{E}_{qq}\right|_* & = & - \left.\mathcal{E}_{qH}\right|_* \left.\frac{\partial H}{\partial q}\right|_*
\end{eqnarray*}
However, from the defining equation for $\mathcal{E}_q$ given above, it is seen that $\mathcal{E}_{qH} = - H / I_H^2 I_{Hq}$, where $I_{Hq} > 0$ by inspection of Eqn.\ \ref{eq:IH}. Thus, the sign of $\left.\mathcal{E}_{qq}\right|_*$ equals the sign of $\left.\frac{\partial H}{\partial q}\right|_*$.
\end{proof}
\begin{corollary}
\label{corrolary:X}
For an increasing angular momentum $H$, a free relative equilibria that ends in a Termination Fission is always unstable in the degree of freedom $q$. Conversely, a free relative equilibria that emanates from a Transition Fission is always stable in the degree of freedom $q$.
\end{corollary}
\begin{proof}
Assume the active constraint is defined to be $q=0$. By definition, a Termination Fission occurs when a relative equilibrium at $q^* > 0$ moves towards the general constraint $q = 0$ under increasing angular momentum. Thus $\partial H / \partial q < 0$ and from Lemma \ref{theorem:X} the relative equilibrium is unstable. Conversely, a Transition Fission occurs when a relative equilibrium at $q^* \ge 0$ moves away from the general constraint $q = 0$ under increasing angular momentum. Thus $\partial H / \partial q > 0$ and from Lemma \ref{theorem:X} the relative equilibrium is energetically stable in the degree of freedom $q$, although it may be unstable in other degrees of freedom.
\end{proof}
Finally, we end with a Lemma on the rotation axis that a stable configuration must have.
\begin{lemma}
Any relative equilibrium configuration not rotating about the maximum moment of inertia of the body will be energetically unstable.
\end{lemma}
\begin{proof}
If a body is in a relative equilibria it must rotate about a principal moment of inertia. It can then be treated as a rigid body, at least up to first order variations in its internal configuration and inertial orientation. If it is not rotating about its maximum moment of inertia, it must be rotating about its intermediate or minimum moment of inertia. The relevant total energy of the function system then equals $H^2 / (2 I_i)$ where $I_i$ is a principal moment of inertia (here ignoring internal variations). From the classical Poinsot construction the body will be at a saddle point of the energy function if rotating about the intermediate axis and will be at a local maximum of the energy function for rotating about the minimum axis. In either case, the rigid body rotation is not stable in the energetic sense as it can depart from this rotation axis while conserving energy with an increase in kinetic energy.
\end{proof}
Due to this result, we only consider rotation about the maximum moment of inertia, which will always lie in the plane containing the three bodies. In the degenerate case where the bodies are in a line, the system will rotate perpendicular to its line of symmetry.
With these Definitions, Lemmas and Corollaries stated, the relative equilibria and stability of the Full 3-body problem can be established.
\section{Existence, Stability and Bifurcation of Relative Equilibria}
In this section, having stated the theorem and developed the necessary background, the detailed proof of Theorem 1 is now given.
\begin{proof}
To systematically explore the existence, stability and bifurcation of the relative equilibria the systems with different numbers of degree of freedom constraints activated and conditions for these to be released are considered separately. The discussion starts with all three DOF constraints activated and progressing to fewer and fewer until all degrees of freedom are not constrained. The Appendix contains the detailed partial derivatives and variation conditions of the amended potential needed for the following discussions.
\subsection{Three Active Constraints: Lagrange Resting Configurations}
\paragraph{Existence:}
For the three constraints to be active requires that $d_{ij} = r_i + r_j = 1 - r_k$ for all of the indices. This configuration can only occur when the three bodies are mutually resting on each other.
The relative angle between adjacent grains are then defined by
\begin{eqnarray}
\cos\theta_{ki} & = & \frac{ (1-r_k)^2 + (1-r_i)^2 - (1-r_j)^2 }{ 2 (1-r_k) (1-r_i) } \\
\sin\theta_{ki} & = & \frac{\sqrt{ (1-r_k)^2(r_k-r_i r_j) + (1-r_i)^2(r_i-r_j r_k) + (1-r_j)^2(r_j-r_k r_i)} }{\sqrt{2} (1-r_k) (1-r_i)}
\end{eqnarray}
where $k, i$ take on all possible values.
The corresponding values of $I_H$ and $\mathcal{U}$ in this configuration are
\begin{eqnarray}
I_H & = & m_i m_j (1-r_k)^2 + m_j m_k (1-r_i)^2 + m_k m_i (1-r_j)^2 + I_S \\
\mathcal{U} & = & - \left[ \frac{m_i m_j}{1-r_k} + \frac{m_j m_k}{1-r_i} + \frac{m_k m_i}{1-r_j} \right]
\end{eqnarray}
There are two unique orderings of the resting configuration, mirroring the orbital Lagrange configuration, which results in two distinct relative equilibria. Due to this these configurations are called the Lagrange Resting (LR) configurations.
\paragraph{Stability:}
As this is the minimum distance for each of these bodies to achieve, this also implies that the potential energy will be minimized at this configuration. From this it can immediately be concluded that for $H=0$ this particular resting configuration is the minimum energy configuration of the system and hence is stable.
\paragraph{Bifurcation:}
As $H$ increases from zero this system should exist as a relative equilibrium for some range of $H$, but to discover the precise range when this holds requires that the transition from three to two active constraints be investigated. Thus, as angular momentum is increased, conditions for when one of these constraints is no longer enforced is sought, meaning that one of the degrees of freedom will have an allowable variation that decreases the energy. For this configuration each of the three distances can be tested in turn to see which will lose positivity first. For the condition tested, consider the angle variation $\delta\theta_{ki} \ge 0$, keeping the other two constraints $\delta d_{ij} = \delta d_{jk} = 0$. The condition for existence (and stability) of this configuration then becomes $\delta_{\theta_{ki}}\mathcal{E} \ge 0$. Evaluating this explicitly and substituting for the equal resting conditions yields
\begin{eqnarray}
\delta_{\theta_{ki}} {\cal E} & = & m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{(1-r_j)^3} \right] (1-r_k) (1-r_i) \sin\theta_{ki} \delta \theta_{ki} \label{eq:3bp3}
\end{eqnarray}
and substituting in for $\sin\theta_{ki}$ yields
\begin{eqnarray}
\delta_{\theta_{ki}} {\cal E} & = & \frac{m_k m_i}{\sqrt{2}} \left[ - \frac{H^2}{I_H^2} + \frac{1}{(1-r_j)^3} \right] \times \\
& & \sqrt{ (1-r_k)^2(r_k-r_i r_j) + (1-r_i)^2(r_i-r_j r_k) + (1-r_j)^2(r_j-r_k r_i)} \delta \theta_{ki}\nonumber
\end{eqnarray}
Changing $k,i$ for $i,j$ and $j,k$ only changes the items on the first line, and thus the controlling condition for the existence and stability of the Lagrange Resting configurations is
\begin{eqnarray}
\frac{1}{(1-r_j)^3} & > & \frac{H^2}{I_H^2}
\end{eqnarray}
which must hold for $j = 1,2,3$. Thus the minimum value of $r_j$ gives the minimum value of $H$ for the inequality to be violated. For the specified definitions this means that $j=3$ and the loss of stability occurs about the angle $\theta_{12}$, meaning that the Lagrange Resting configuration will undergo a Termination Fission by losing contact between its two largest bodies, pivoting about the smallest grain (see Fig. \ref{fig:bif2}).
\subsection{Two Active Constraints}
In this case two bodies rest on each other, but do not have the third contact active. A convenient way to express this is to have the two distances at their minimum value and leave the angle free, or $d_{ij} = 1 - r_k$, $d_{jk} = 1 - r_i$ with $\theta_{ki}$ only constrained by the resting limit, $d_{ki} \ge 1 - r_j$. For the moment assume that $\mathcal{E}_{d_{ij}} > 0$ and $\mathcal{E}_{d_{jk}} > 0$ (this will be checked later), and thus there is only one degree of freedom to be concerned with. Taking the first variation and substituting for the distances yields
\begin{eqnarray}
\delta_{\theta_{ki}} {\cal E} & = & m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] (1-r_k)(1-r_i) \sin\theta_{ki} \delta \theta_{ki}
\end{eqnarray}
which must now be identically equal to zero for the system to be in equilibrium. There are two possibilities, $\sin\theta_{ki} = 0$ or $- \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} = 0$. Both can occur and are discussed separately, the former is called the Euler Resting configuration and the latter the Transitional Resting configuration. No assumptions about the ordering of the bodies in terms of mass, unless specified.
Each case must be tested for when the configurations cease to exist, which will occur once one of the energy variations in the active distance constraints equals zero. These will be explicitly tested for each case to determine conditions at which these equilibria no longer exist.
\subsubsection{Euler Resting Configurations}
\paragraph{Existence:}
First consider the case when $\theta_{ki} = \pi$, noting that the angle cannot equal zero due to the finite radius constraints. Then the first variation is identically equal to zero and the bodies rest on a straight line with the ordering $i, j, k$, the system forming a relative equilibrium. These are notationally denoted as ER$ijk$, noting that configuration ER$kji$ is considered to be equivalent. Now the moment of inertia and potential energy take on the values
\begin{eqnarray}
I_H & = & m_i m_j (1-r_k)^2 + m_j m_k (1-r_i)^2 + m_k m_i (1+r_j)^2 + I_S \\
\mathcal{U} & = & - \left[ \frac{m_i m_j}{1-r_k} + \frac{m_j m_k}{1-r_i} + \frac{m_k m_i}{1+r_j} \right]
\end{eqnarray}
with the main difference from the Lagrange Resting (LR) configurations being that the distance $d_{ki} = 1+r_j$ due to the elongate geometry.
\paragraph{Stability:}
Under the assumption that the two distance variations are both positive (which is true for a low enough value of $H$), the stability of this relative equilibrium can be analyzed by computing the second order variation evaluated at the resting configuration.
\begin{eqnarray}
\delta_{\theta_{ki}\theta_{ki}} {\cal E} & = & - m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{(1+r_j)^3} \right] (1-r_k) (1-r_i) \delta \theta_{ki}^2 \label{eq:RE_thetasq}
\end{eqnarray}
Stability of this configuration occurs when $\delta_{\theta_{ki}\theta_{ki}} {\cal E} > 0$ which places a lower limit on the angular momentum for stability
\begin{eqnarray}
\frac{I_H^2}{(1+r_j)^3} & < & H^2
\end{eqnarray}
Note that the value of angular momentum is lower than the angular momentum at which the LR configurations cease to exist. Also, the stability transition occurs when the Transitional Resting configuration conditions are satisfied for the same configuration, indicating that a bifurcation occurs.
\paragraph{Bifurcation:}
For lower values of angular momentum the Euler Resting configuration exists, but is unstable and mimics an inverted pendulum. When the stability condition is satisfied, the system mimics a hanging pendulum and will remain stable until one of the energy distance variations becomes zero, indicating a transition from two active constraints to a single active constraint. To probe when this occurs, substitute the equilibrium condition into Eqn.\ \ref{eq:3bpij} to find
\begin{eqnarray}
\delta_{d_{ij}} {\cal E} & = & m_i \left\{ m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{(1-r_k)^3} \right] (1-r_k) \right. \nonumber \\
& & \left. + m_k \left[ - \frac{H^2}{I_H^2} + \frac{1}{(1+r_j)^3} \right] (1+r_j) \right\} \delta d_{ij}
\end{eqnarray}
Setting this to be greater than or equal to zero defines when the ER relative equilibrium configuration exists, and can be solved for as a condition on angular momentum
\begin{eqnarray}
\left[ m_j (1-r_k) + m_k (1+r_j) \right] \frac{H^2}{I_H^2} & \le & m_j \frac{1}{(1-r_k)^2} + m_k \frac{1}{(1+r_j)^2}
\end{eqnarray}
The precise value of $H$ when this is first violated is discussed in a later section. For the current analysis it suffices to note that this inequality is always satisfied when the ER configurations first become stable. Substituting $(H/I_H)^2 = 1 / (1+r_j)^3$ and simplifying yields
\begin{eqnarray}
0 & \le & r_j + r_k
\end{eqnarray}
which is trivially satisfied for any $j$ or $k$. It is also clear that a large enough $H$ will always be able to violate the existence condition. Generically, one of the two bodies $i$ or $k$ will separate from $j$, leaving the other body in contact and transitioning the configuration into the Euler Aligned configuration.
\subsubsection{Transitional Resting Configurations}
\paragraph{Existence:}
When the Euler Resting (ER) configurations becomes stable, a pair of solutions that satisfy the second equilibrium condition bifurcate from or into the resting configuration. The condition in general is $H^2 = I_H^2 / d_{ki}^3$, but now the moment of inertia $I_H$ becomes a function of $\theta_{ki}$ and must change with $H$ to maintain this condition. There are two branches, $\theta_{ki} > \pi$ and $\theta_{ki} < \pi$, and these give two different orderings of the configuration -- ultimately corresponding to the two different orientations of several of the equilibrium configurations.
The moment of inertia and potential energy now take on the more generalized form
\begin{eqnarray}
I_H & = & m_i m_j (1-r_k)^2 + m_j m_k (1-r_i)^2 + m_k m_i d_{ki}^2 + I_S \\
\mathcal{U} & = & - \left[ \frac{m_i m_j}{1-r_k} + \frac{m_j m_k}{1-r_i} + \frac{m_k m_i}{d_{ki}} \right] \\
d_{ki}^2 & = & (1-r_k)^2 + (1-r_i)^2 - 2 (1-r_k)(1-r_i)\cos\theta_{ki}
\end{eqnarray}
\paragraph{Stability:}
Evaluating the second variation of the energy with respect to $\theta_{ki}$ yields
\begin{eqnarray}
\delta^2_{\theta_{ki}}{\cal E} & = & m_k m_i \left[ 4 m_i m_k d_{ki}^2 - 3 I_H \right] \frac{(d_{ij} d_{jk} \sin\theta_{ki})^2}{I_H d_{ki}^5} \left(\delta\theta_{ki}\right)^2
\end{eqnarray}
Stability, when the configuration exists, then hinges on the sign of $4 m_i m_k d_{ki}^2 - 3 I_H$. Making the substitution from Eqn.\ \ref{eq:massnorm} the stability condition can be reduced to
\begin{eqnarray}
d_{ki}^2 & > & \frac{3}{r_i^3 r_k^3} \left[ \frac{2}{5}\left(r_i^3+r_j^3+r_k^3\right)\left(r_i^5+r_j^5+r_k^5\right) +
r_i^3 r_j^3 (1-r_k)^2 + r_j^3 r_k^3 (1-r_i)^2 \right]
\label{eq:TRstab}
\end{eqnarray}
where $1-r_j \le d_{ki} \le 1+r_j$.
Note that the equilibrium configuration does not necessarily exist across this entire range of mutual distances.
Specifically, the distance variation conditions must be verified for the configuration to exist. Substituting the equilibrium condition into Eqn.\ \ref{eq:3bpij} then yields the existence condition (after simplification)
\begin{eqnarray}
\frac{H^2}{I_H^2} & \le & \frac{1}{(1-r_k)^3}
\end{eqnarray}
where $k$ is the radius of either of the outer resting bodies. Note that the transitional resting configurations will always exist over some interval of angular momentum, as substituting the initial bifurcation conditions of $(H/I_H)^2 = 1 / (1+r_j)^3$ can be trivially shown to satisfy the above existence condition.
Again, note that $H$ can also always be chosen large enough for the existence condition to be violated.
There are three different possible situations to cover, investigated in detail below.
\paragraph{$i=1,j=2,k=3$} For this sequence the Transitional Resting equilibrium are unstable and migrate from the ER123 configuration (which they stabilize upon bifurcation from it) to the distance $d_{31} = 1-r_3$, at which point body 1 separates from the system. To determine instability evaluate Eqn.\ \ref{eq:TRstab} over the entire range of radius values and verify that it is never satisfied.
To see this consider the contact conditions from Eqns.\ \ref{eq:3bpij} and \ref{eq:3bpjk}. For these conditions to hold both must be greater than or equal to zero for a positive variation in the mutual distance $d_{ij}$ and $d_{jk}$. Substituting the equilibrium condition $(H/I_H)^2 = 1/d_{31}^3$ and simplifying, the condition for existence of the TR123 configuration is that both
\begin{eqnarray}
d_{31} & \ge & d_{12} \\
d_{31} & \ge & d_{23}
\end{eqnarray}
For the current configuration, $d_{12} = 1-r_3 > d_{23} = 1-r_1$. Thus the controlling condition is $d_{31} \ge 1 - r_3$. Now note that $d_{31} \ge 1-r_2$ and that $1-r_3 > 1-r_2$, thus this inequality is violated prior to the TR123 configuration reaching the LR configuration, and as noted occurs once $d_{31} = 1-r_3$. See Fig.\ \ref{fig:bif1} for the evolutionary path for this situation.
\paragraph{$i=3,j=1,k=2$} For this sequence the Transitional Resting equilibrium are unstable (determined as before) and migrate from the ER312 configuration (which they stabilize upon bifurcation from it) to the distance $d_{23} = 1-r_3$ when body 2 separates from the system. Similar to above, the condition for existence of the TR312 configuration is that
\begin{eqnarray}
d_{23} & \ge & d_{31} \\
d_{23} & \ge & d_{12}
\end{eqnarray}
For the current configuration, $d_{12} = 1-r_3 > d_{31} = 1-r_2$. Thus the controlling condition is $d_{23} \ge 1 - r_3$. Now note that $d_{23} \ge 1-r_1$ and that $1-r_3 > 1-r_1$, thus this inequality is violated prior to the TR312 configuration reaching the LR configuration, and as noted occurs once $d_{23} = 1-r_3$ again. See Fig.\ \ref{fig:bif3} for the evolutionary path for this situation.
\paragraph{$i=1,j=3,k=2$} For this sequence the Transitional Resting equilibria exist across the range of radius limits, going from ER132 to Lagrange Resting configurations. For this configuration there are ranges of parameters for which there are stable relative equilibria. To identify these regions compare the upper inequality limit to when the distance for stability is less than the maximum distance $1+r_j$. Plotting out this region delineates the small oval region in Fig.\ \ref{fig:TR132}. For parameter values within this region the evolution of the TR132 configuration becomes more complex. Specifically, the angular momentum profile in this region is such that there are two relative equilibria defined at a given level of angular momentum, one towards the LR configuration (which is always at a local maximum of the energy and thus is unstable) and one towards the ER132 configuration (which becomes a local minimum of the energy and thus is stable). Figure \ref{fig:bif2} shows the two different pathways that can occur.
\paragraph{Bifurcation:}
For the TR configurations which are always unstable, as the angular momentum is increased they first bifurcate into existence by stabilizing the ER configurations. Then as $H$ is increased they migrate towards more compact configurations. The TR123 and TR312 configurations then end with one of the bodies separating from the other two. The TR132 configuration migrates all the way to the LR configuration and destabilizes it, thus terminating both the LR and TR132 configurations.
For TR configurations that can be stable, indicated in Fig.\ \ref{fig:TR132}, the sequence is different. Here as $H$ is increased an $H$-Bifurcation occurs at an angle between the minimum (or maximum) constrained value of $\theta_{12}$ and $\pi$. To show this consider the angular momentum as a function of distance $d_{12}$, $H = I_H(d_{12}) / d_{12}^{3/2}$. Taking the partial of this with respect to $d_{12}$ shows that there is a zero in the interval $1-r_3 \le d_{12} \le 1+r_3$ (meaning that $H$ takes on an extreme value) whenever the stability condition in Eqn.\ \ref{eq:TRstab} is satisfied. Further, taking the second partial of $H$ and substituting the equilibrium condition shows that this is always positive, meaning that $H$ takes on a minimum value in the interval. Thus, as $H$ is increased the two equilibria exist on either side of the minimum, with no other equilibria emerging due to the definiteness of $H$ as a function of $d_{12}$. From Lemma \ref{theorem:X} the equilibria that moves down to the LR configuration must be unstable, and thus the equilibria that moves toward the ER configuration is stable (this also agrees with the condition as formulated in Eqn.\ \ref{eq:TRstab}).
Thus the unstable TR132 configuration continues down to the LR configuration and terminates it. The stable TR132 configuration moves toward the ER132 configuration and terminates there, stabilizing the ER132 configuration. The existence of these stable TR132 configurations is unexpected and breaks the symmetry otherwise seen in these configurations. The region where these occur correspond to grains with a nearly equal $r_1$ and $r_2$, with $r_3$ neither close to zero or to the size of the other grains.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.3]{graphics/TR132}
\caption{Region where stable TR132 configurations can lie.}
\label{fig:TR132}
\end{figure}
\subsection{One Active Constraint}
Now consider relative equilibria when there is a single active constraint. In this configuration two of the bodies rest on each other, say $i$ and $j$ and thus $d_{ij}=1-r_k$, and the third body is located by the distance $d_{jk}$ and by either $d_{ki}$ or the angle $\theta_{ki}$. There are two classes of relative equilibrium solutions in this class, with the two bodies in contact either being aligned with the third body, or with their line of contact being orthogonal to the third body. The former are called the Euler Aligned configurations and the latter the transverse, or Isoceles, configurations -- the terminology arising due to the structure that these make. For these structures there are two limiting cases, one where the final active constraint separates and the other where one of the free constraints becomes activated. The former occurs when the single active constraint configurations intersect with the orbital configurations. The latter occurs when it intersects with a double active configuration. These two classes of configurations are discussed in turn.
First it can be established that these are the only relative equilibrium configurations. Consider Eqns.\ \ref{eq:3bpjk} and \ref{eq:3bpki} in the Appendix, which both must equal zero. There are two possibilities for Eqn.\ \ref{eq:3bpki} to equal zero, either $\sin\theta_{ki} = 0$ or $H = I_H / d_{ki}^{3/2}$. Consider $\theta_{ki} = \pi$, as setting the angle to 0 is equivalent to a reordering of the bodies. Then for Eqn.\ \ref{eq:3bpjk} to equal zero the condition becomes $m_j \left( H^2/I_H^2 - 1 / d_{jk}^3\right) d_{jk} + m_i \left( H^2/I_H^2 - 1 / d_{ki}^3\right) d_{ki} = 0$. In this configuration $d_{ki} > d_{jk}$ and thus along this configuration it can never occur that $H^2 / I_H^2 = 1/d_{ki}^3$, meaning that this condition will not intersect with the $\theta_{ki} = \pi$ configuration. If $H$ is chosen such that $ I_H / d_{ki}^{3/2} < H < I_H / d_{jk}^{3/2}$ it is possible for the second condition to be satisfied, which is explored in more detail below.
The alternate condition to consider is $H^2 / I_H^2 = 1/d_{ki}^3$, with no immediate constraint on $\theta_{ki}$. Then, by substitution into Eqn.\ \ref{eq:3bpjk} yields the condition $H^2 / I_H^2 = 1/d_{jk}^3$, or $d_{ki} = d_{jk} = d$. From Eqn.\ \ref{eq:d31} and with $d_{ij} = 1 - r_k$ the condition on $\theta_{ki}$ becomes
\begin{eqnarray}
\cos\theta_{ki} & = & \frac{1}{2} \frac{1-r_k}{d}
\end{eqnarray}
Note that $d$ can always be chosen large enough for $\theta_{ki}$ to be well defined.
\subsubsection{Isosceles Configurations}
\label{sec:isosceles}
\paragraph{Existence:}
The Isosceles configurations are described by having two grains in contact, nominally $i$ and $j$, the third grain $k$ in non-contact with these grains, and with the line connecting the grains $i$ and $j$ being perpendicular to the line from grain $k$ to the center of mass of grains $i$ and $j$. These are referred to as IS$ij$-$k$, with the first two indices indicating the grains in contact and the separated third index the separated grain. In terms of Eqns.\ \ref{eq:3bpij}, \ref{eq:3bpjk} and \ref{eq:3bpki}, set $d_{ij} = 1-r_k$ and $d_{jk} = d_{ki} = d$, forming an Isosceles triangle. The equilibrium condition is then simply stated as
\begin{eqnarray}
\frac{H^2}{I_H^2} & = & \frac{1}{d^3} \\
I_H & = & m_i m_j (1-r_k)^2 + (m_j m_k + m_k m_i) d^2 + I_S
\end{eqnarray}
Making this substitution, see that $\delta_{d_{jk}}{\cal E} = \delta_{\theta_{ki}}{\cal E} = 0$. So long as $d_{jk} \ge 1-r_i$ and $d_{ki} \ge 1-r_j$, the remaining condition for this equilibrium to be satisfied is $ \delta_{d_{ij}}{\cal E} \ge 0$, which can be simplified to the condition
\begin{eqnarray}
d = d_{ki} = d_{kj} & \ge & 1-r_k
\end{eqnarray}
Now consider the existence of each of the possible combinations, in turn.
\paragraph{IS12-3}
Here the grains in contact are separated by a distance $1-r_3$ and the controlling distance of the equal legs of the triangle will be $d_{31} = 1-r_2$. Note that $1-r_3 > 1-r_2$, and thus the above existence condition will be violated when $d_{31} = d_{23} = 1-r_3$, and in fact the three grains will lie at the vertices of an equilateral triangle. This condition corresponds to the intersection of IS12-3 with the orbital Lagrange configuration, LO, and terminates the IS12-3 configuration, without having grain 3 contacting the other two grains. This sequence is isolated from the previous configurations as the three grains never come into contact and is shown in Fig.\ \ref{fig:bif4}.
\paragraph{IS23-1}
Now the grains in contact are separated by a distance $1-r_1$ and the controlling distance of the equal legs of the triangle will be $d_{12} = 1-r_3$. Now as $1-r_3 > 1-r_1$, grains 1 and 2 will touch prior to the grains 2 and 3 separating. Once grains 1 and 2 touch the configuration matches the end-state configuration of the TR123 configuration. Thus the TR123 and IS23-1 configurations terminates, as shown in Fig.\ \ref{fig:bif1}.
\paragraph{IS31-2}
Now the grains in contact are separated by a distance $1-r_2$ and the controlling distance of the equal legs of the triangle will be $d_{12} = 1-r_3$ again. Similar to before, $1-r_3 > 1-r_2$, so grains 1 and 2 will touch prior to the grains 1 and 3 separating. Once grains 1 and 2 touch the configuration matches the end-state configuration of the TR312 configuration. Thus the TR312 and IS31-2 configurations terminate, as shown in Fig.\ \ref{fig:bif3}.
\paragraph{Stability:}
Now consider the stability of the IS configurations. The condition $\delta_{d_{ij}}{\cal E} \ge 0$ is uniformly satisfied, except for the termination of the IS12-3 configuration noted above. Thus it is just needed to test whether the joint variations of $\delta d_{jk}$ and $\delta \theta_{ki}$ are positive definite or not. For this situation one must take the second partial of the energy with respect to both of these variations, evaluated at the relative equilibrium, and test the 2$\times$2 resulting matrix for whether it is positive definite.
\begin{eqnarray}
\delta^2{\cal E} & = & \left[ \begin{array}{cc} \delta d_{jk} & \delta\theta_{ki} \end{array} \right]
\left[ \begin{array}{cc} \frac{\partial^2{\cal E}}{\partial d_{jk} \partial d_{jk} } & \frac{\partial^2 {\cal E}}{\partial d_{jk} \partial\theta_{ki}} \\
\frac{\partial^2 {\cal E}}{\partial\theta_{ki}\partial d_{jk} } & \frac{\partial^2{\cal E}}{\partial\theta_{jk} \partial\theta_{jk} } \end{array} \right]
\left[ \begin{array}{c} \delta d_{jk} \\ \delta\theta_{ki}\end{array} \right]
\end{eqnarray}
A matrix is positive definite by Sylvester's Criterion if all of its leading principal minors are positive. A simpler, necessary condition, is that the diagonals of the matrix are all positive.
To that end, consider the term $\frac{\partial^2{\cal E}}{\partial\theta_{jk} \partial\theta_{jk} }$ evaluated at the equilibrium condition, which can be found to equal
\begin{eqnarray}
\frac{\partial^2{\cal E}}{\partial\theta_{jk} \partial\theta_{jk} } & = & m_k m_i \left[ m_k\left(m_i-3m_j\right)d_{ki}^2 - 3 I_S - 3 m_i m_j (1-r_k)^2 \right] \frac{\left(d_{ij}\sin\theta_{ki}\right)^2}{d_{ki}^3 I_H}
\end{eqnarray}
Note that the ordering of $i$ and $j$ does not matter, although the individual terms of the matrices may change. Thus, one can always choose to assign $i$ and $j$ such that $m_j > m_i$ to ensure that $m_i-3m_j < 0$, making the diagonal negative definite. Thus, any of the configurations will violate the necessary condition for the system to be positive definite, meaning that the Isosceles configurations are always unstable. Note that this instability mode is related to the angle $\theta_{ki}$ and not related to instability in the distance variation. Due to this, the IS family is always unstable even if it is formed from a Transition Fission.
\paragraph{Bifurcation:}
To end, note that in Figs.\ \ref{fig:bif1}, \ref{fig:bif3} and \ref{fig:bif4} a similar specific sequence for the evolution of all of the Isosceles configurations is shown, with them appearing as an $H$-Bifurcation with one branch continuing to $\infty$ and the other terminating at a TR configuration or ending at an LO configuration. The persistence of this structure can be proven, using a similar approach as used in discussing the bifurcation of the TR132. The relation between angular momentum and distance $d$ in these configurations is the simple expression $H = I_H(d) / d^{3/2}$. It can be shown that this function has a unique minimum positive value, and thus the Isosceles configurations bifurcate into existence when the angular momentum rises above this value. Further, it can be shown that the distance at which this bifurcation occurs is always greater than the associated contact distances for this configuration. This is shown by developing a specific inequality that must be satisfied, and then checking it by computing level sets across the domain of possible radii. Doing so reveals that the bifurcation at a non-zero value of $H$ always occurs away from any of the contact termination conditions. Thus the pattern of having one branch progress towards the TR configurations and the other branch extend to large distances can be inferred.
\subsubsection{Euler Aligned Relative Equilibria}
\label{sec:aligned}
\paragraph{Existence:}
The Euler Aligned relative equilibria are defined by having two grains in contact and the third at a distance along the centers of mass of the two grains in contact. Again, the grains in contact are $i$ and $j$ and grain $k$ is separated. The notation for these equilibria is EA$ij$-$k$, where the order is important. Specifically, note that EA$ij$-$k$ and EA$ji$-$k$ are different, with the grain $k$ rotated 180$^\circ$ relative to the other configuration. Thus, EA$ij$-$k$ can be organized from left to right and fits with the earlier notation. There are 6 different configurations that can be considered, EA12-3, EA21-3, EA13-2, EA31-2, EA23-1, EA32-1. In all these definitions the angle $\theta_{ki} = \pi$ and $\delta_{\theta_{ki}}{\cal E} = 0$. The two remaining conditions are then $\delta_{d_{ij}} {\cal E} \ge 0$ and $\delta_{d_{jk}} {\cal E} = 0$.
For existence, solve each of these conditions for the ratio $(H/I_H)^2$ to find
\begin{eqnarray}
\left(\frac{H}{I_H}\right)^2 & \le & \frac{1}{m_j(1-r_k)+m_kd_{ki}} \left[ \frac{m_j}{(1-r_k)^2} + \frac{m_k}{d_{ki}^2} \right] \\
\left(\frac{H}{I_H}\right)^2 & = & \frac{1}{m_i d_{ki} + m_j d_{jk}} \left[ \frac{m_i}{d_{ki}^2} + \frac{m_j}{d_{jk}^2} \right]
\label{eq:H_EA}
\end{eqnarray}
Note that $d_{ki} \ge 1 + r_j$ and $d_{jk} = d_{ki} - (1-r_k)$.
These conditions can be combined and rewritten into a standard form
\begin{eqnarray}
\frac{1}{d_{ki}^3} F( m_j/m_i, d_{jk}/d_{ki} ) & \le & \frac{1}{(1-r_k)^3} F( m_k/m_j, d_{ki}/(1-r_k))
\end{eqnarray}
where $F(\mu, x) = \left( 1 + \mu / x^2 \right) / \left( 1 + \mu x\right)$. Note the identity $x^3 F(\mu,x) = F(1/\mu, 1/x)$. When this inequality is violated grains $i$ and $j$ will separate and the configuration will cease to exist.
As a final step, define $r = d_{ki} / (1-r_k) > 1$, $\mu_{ij} = m_i/m_j$ and $d_{jk}/d_{ki} = 1 - 1/r$. Then the inequalities are written as
\begin{eqnarray}
F( \mu_{ji}, 1 - 1/r) & \le & F(\mu_{jk}, 1/r)
\end{eqnarray}
It can be shown (see Appendix) that $F(\mu,r)$ is monotonically decreasing in $r$ and is convex. From this it can be shown that $F( \mu, 1 - 1/r)$ is monotonically decreasing in $r$ and that $F( \mu, 1/r)$ is monotonically increasing. Thus the inequality can be crossed either 0 or 1 times and it is not needed to consider the possibility of multiple transitions in the existence of solutions. Given the well defined interval over which the parameter $r$ is defined, $r \in [(1+r_j)/(1-r_k), \infty)$, an explicit method for determining when these conditions exist can be developed.
First note that $\lim_{r\rightarrow\infty} F( \mu, 1 - 1/r) = 1$ and that $\lim_{r\rightarrow\infty} F(\mu, 1/r) \sim r^2/\mu + \ldots$. Thus the inequality is always satisfied when the distance between the grains in contact, $i$ and $j$, and the separated grain $k$, is large. This holds independent of the ordering of the indices, and means that all of the EA configurations exist when the $k$th grain is sufficiently distant from the two in contact. Thus, to ascertain whether the configuration exists across all possible values of $r$ it is only needed to check the condition at the minimum radius condition. Due to the topological properties of the two functions in the inequality, if the inequality is satisfied at the minimum value of $r$, then the given configuration exists across all distances in the interval. If it is violated at the minimum value of $r$, then there exists a distance at which the configuration ceases to exist.
Evaluating the inequality at the minimum distance $r = (1+r_j)/(1-r_k)$ yields
\begin{eqnarray}
F( \mu_{ji}, (1-r_i)/(1+r_j) ) & \le & F(\mu_{jk}, (1-r_k)/(1+r_j))
\end{eqnarray}
If this inequality is confirmed, then the configuration EA$ij$-$k$ exists across the whole domain and, by swapping indices $i$ and $k$, that then the configuration EA$kj$-$i$ does not exist by definition. Conversely, if the inequality for EA$ij$-$k$ is not confirmed at the lower limit, then the configuration EA$kj$-$i$ is trivially confirmed. The following discussions will assume that the configuration EA$ij$-$k$ exists all the way to contact, and thus that the configuration EA$kj$-$i$ does not and terminates at a finite distance from contact.
This means that whenever one EA configuration exists down to the Euler Resting configuration, that the alternate EA configuration does not, and terminates at a finite separation. The termination of the conjugate configuration occurs when that configuration intersects with the conditions for the orbital Euler configuration EO$ijk$, as by definition at termination $\delta_{d_{ji}}{\cal E} = 0$ by default and $\delta_{d_{kj}}{\cal E} = 0$ due to the contact constraint vanishing.
With these results in hand, the realms where the different Euler Aligned configurations exist can be evaluated. To do this, plot the level sets of the function
\begin{eqnarray}
F(\mu_{jk}, (1-r_k)/(1+r_j)) - F( \mu_{ji}, (1-r_i)/(1+r_j) ) & = & 0
\end{eqnarray}
As these functions are analytical and have no singularities, there are no computational issues with evaluating these level sets. The zero line delineates where a transition in the existence of these configurations occurs. In the region where the difference is positive, the EA$ij$-$k$ configuration exists down to contact, while in the region where the difference is negative, the EA$kj$-$i$ configuration exists down to contact. These distinctions are important as they control which grain will separate from an Euler resting configuration when angular momentum is increased. In the following the plots of these zero lines are displayed for the different possible configurations.
\paragraph{EA12-3 and EA32-1}
Figure \ref{fig:EA123} shows a plot of the level set of the inequality for the ordering 123, showing that there exists a region where the EA12-3 configuration exists down to the ER123 configuration, and where the EA32-1 configuration exists down to the ER123 configuration. The former exists in the region where the grains 2 and 3 are more similar sized, and the latter where the grains 1 and 2 are more similar sized. Which side of the line that configuration lies determines how the configuration will fission when it terminates. Figure \ref{fig:bif1} shows the different bifurcation pathways that occur.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.25]{graphics/EA123}
\caption{Fission chart for the ER123 configuration. For masses to the right of the line ER123 configuration fissions by having the smallest mass separate and transitioning into the EA$12$-$3$ configuration. For masses to the left of the line the ER123 configuration fissions by having the largest mass separate and transitioning into the EA32-1 configuration.}
\label{fig:EA123}
\end{figure}
\paragraph{EA13-2 and EA23-1}
Figure \ref{fig:EA132} shows a plot of the level set of the inequality for the ordering 132. Here, only the EA13-2 configuration exists down to the ER132 configuration, and thus the EA23-1 configuration always terminates at a finite distance. Not shown here explicitly is that at the left border, where $r_1 = r_2$, the two conditions are equivalent due to symmetry and both EA configurations extend down to the ER132 configuration. Figure \ref{fig:bif2} shows the different bifurcation pathways that occur.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.25]{graphics/EA132}
\caption{Fission chart for the ER132 configuration. For this configuration the intermediate mass always separates, transitioning into the EA13-2 configuration. }
\label{fig:EA132}
\end{figure}
\paragraph{EA31-2 and EA21-3}
Figure \ref{fig:EA312} shows a plot of the level set of the inequality for the ordering 312. There are two regimes again. When the grains are relatively equal in size the configuration EA21-3 continues down to ER312. Away from this geometric region, however, configuration EA31-2 continues down to ER312. Figure \ref{fig:bif3} shows the different bifurcation pathways that occur.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.25]{graphics/EA312}
\caption{Fission chart for the ER312 configuration. For masses to the right of the line the ER312 configuration fissions by having the intermediate mass separate and transitioning into the EA31-2 configuration. For masses to the left of the line the ER312 configuration fissions by having the smallest mass separate and transitioning into the EA21-3 configuration.}
\label{fig:EA312}
\end{figure}
\paragraph{Stability:}
For an EA$ij$-$k$ configuration to be stable requires $\delta_{d_{ij}}{\cal E} > 0$ and the second variations of ${\cal E}$ with respect to $d_{jk}$ and $\theta_{ki}$ be positive definite. The condition on $d_{ij}$ is automatically satisfied, except at specific transition points, once it is shown that a given configuration exists. Thus only the second order variation conditions need to be evaluated.
First, note that the cross variations $\delta^2_{d_{jk}\theta_{ki}}{\cal E}$ are identically zero, and only consider the second variations $\delta^2_{\theta_{ki}}{\cal E}$ and $\delta^2_{d_{jk}}{\cal E}$ separately. Computing the first of these and evaluating it at the nominal condition yields
\begin{eqnarray}
\delta^2_{\theta_{ki}}{\cal E} & = & - m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3}\right] (1-r_k) \left( d_{ki} - (1-r_k)\right) (\delta\theta_{ki})^2
\end{eqnarray}
Make the substitution $H^2/I_H^2 = \frac{1}{d_{ki}^3} F( m_j/m_i, 1 - (1-r_k)/d_{ki})$. Then, stability in this variation can be established by showing that
\begin{eqnarray}
F(m_j/m_i, 1 - (1-r_k)/d_{ki}) & > & 1
\end{eqnarray}
However, the function $F(\mu,1-1/r)$ was shown to be monotonically decreasing in $r$ with the limiting value of 1 as $r$ becomes arbitrarily large. Thus this is always satisfied and the EA configurations are always stable to variations in the angle $\theta_{ki}$.
All that is left is to consider when $\delta^2_{d_{jk}}{\cal E} > 0$. First, re-express Eqn.\ \ref{eq:3bpjk} as
\begin{eqnarray}
\delta_{d_{jk}} {\cal E} & = & \frac{ m_k}{I_H^2} \left\{ - H^2\left( m_j + m_i \right) + I_H^2 \left[ \frac{m_j}{d_{jk}^2} + m_i \frac{d_{jk} - d_{ij}\cos\theta_{ki}}{d_{ki}^3} \right] \right\} \delta d_{jk}
\end{eqnarray}
The term inside the brackets is identically zero at equilibrium, thus one does not need to take the variation of terms outside of the brackets. Taking the variation inside the brackets and simplifying yields
\begin{eqnarray}
\delta^2_{d_{jk}} {\cal E} & = & 2 m_k \left\{ \frac{2 m_k}{ I_H } \left( m_j d_{jk} + m_i d_{ki} \right) \left[ \frac{m_j}{d_{jk}^2} + \frac{m_i}{d_{ki}^2} \right] - \left[ \frac{m_j}{d_{jk}^3} + \frac{m_i}{d_{ki}^3} \right] \right\} \delta d_{jk} ^2
\end{eqnarray}
First consider the case when $d \sim d_{jk} \sim d_{ki} \gg 1$. The second variation then reduces to
\begin{eqnarray}
\delta^2_{d_{jk}} {\cal E} & = & 2 m_k \frac{m_j+m_i}{d^3} \delta d_{jk} ^2
\end{eqnarray}
which is always positive. Thus, all EA$ij$-$k$ configurations with large enough distances are stable.
Conversely, Corrolary \ref{corrolary:X} shows that the EA configurations are always unstable when they approach and touch the ER configurations. From this, consider the stability of the EA configurations as a function of separation. At their minimum separation, when terminating the ER configurations, they begin as unstable. As the distance is increased they eventually become stable, indicating that ${\cal E}_{d_{jk}d_{jk}}$ evaluated at the EA configuration must cross through zero at some specific equilibrium configuration, indicating the point where the $H$-Bifurcation occurs.
\paragraph{Bifurcation:}
Studying this aspect of the situation can provide qualitative insight into how the EA configurations bifurcate into existence and evolve as $H$ is increased. Consider Eqn.\ \ref{eq:H_EA}, which must be satisfied for an EA relative equilibrium configuration. As all of the terms on the right hand side are positive and bounded from below, there is an absolute minimum value such that if $H$ is below this value the equality cannot be satisfied and the EA relative configuration does not exist. At this value of angular momentum there will be a bifurcation from no relative equilibria to two relative equilibria, corresponding to the point identified above where ${\cal E}_{d_{jk}d_{jk}} = 0$. As $H$ increases further one branch of the EA relative equilibria must migrate towards the ER configuration and the other to larger distances, due to the uniqueness of this family. The branch that migrates to the larger separation will be stable while the branch that migrates to the contact configuration must be unstable, from Lemma \cite{theorem:X}. A similar bifurcation will occur for the configuration that intersects with the Euler Orbital family.
\subsection{No Active Constraint}
Finally consider the case where none of the constraints are active. Then the three conditions must all be identically zero. Due to the structure of the problem, it is well known that there are only 5 relative equilibria to this problem \cite{wintner}. These are divided into the Lagrange solutions, which lie at the vertices of an equilateral triangle \cite{lagrange}, and the Euler solutions, which lie in a single line and are appropriately spaced \cite{euler}.
\subsubsection{Lagrange Solutions}
For the Lagrange solutions, set $d = d_{12} = d_{23} = d_{31}$. From Eqn.\ \ref{eq:d31} and note that this requires $\cos\theta_{31} = 1/2$, meaning that $\theta_{31} = \pm 60^\circ$. Then the condition can be uniformly satisfied by choosing the distance $d$ such that $H^2 = \frac{I_H(d)^2}{d^3}$. Note that for all $d> \max(r_i+r_j) = r_1+r_2$ (given the assumed ordering) such a solution will always exist. However, for a given $H^2$ a solution to the non-contact case may not always exist. Indeed, since for this case $I_H = (m_1m_2 + m_2m_3 + m_3m_1) d^2 + I_S$, $H$ has a minimum value that is greater than zero, and thus will not exist for all values of angular momentum.
The point where the Lagrange Orbital configuration comes into existence can be explicitly probed. In general there are two possibilities. One is that it appears as a two branch family once the angular momentum goes above its minimum value. Then one branch will migrate inwards with increasing angular momentum and terminate by intersection with the IS12-3 family. Otherwise, if the minimum angular momentum point arises at a mutual distance less than $r_1+r_2$, then the inner IS12-3 family will transition directly into the LO family, as it is known that the equal mass case has this sort of a bifurcation \cite{scheeres_minE}, it is relevant to test for when this will occur. To do this just compute $\partial H/\partial d$ and solve for the zero to find
\begin{eqnarray}
d^2 & = & \frac{3 I_S}{m_1 m_2 + m_2 m_3 + m_3 m_1}
\end{eqnarray}
Which bifurcation structure ensues can be found by finding where this solution is greater or less than $(1-r_3)$. Figure \ref{fig:LOH} plots the region where the double branch occurs and where the single branch occurs. Figure \ref{fig:bif4} shows the two different bifurcation pathways.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.25]{graphics/LOH}
\caption{Transition line between the two different bifurcation modes of the Lagrange Orbital configuration. The two different patterns of bifurcation are shown in Fig.\ \ref{fig:bif4}. }
\label{fig:LOH}
\end{figure}
While it is well known that the classical 3-body problem is spectrally stable when the Routh Criterion is satisfied, it should be noted that the stability considered in this paper, energetic stability, is a stronger type of stability. An observation by Moeckel \cite{moeckel_central} shows that central configurations in the point-mass $N$-body problem never have a positive definite second variation of their energy, and thus it can be suspected that the same holds true for the Lagrange Orbital configuration in the Full body.
To test this, take the second order variation of $\mathcal{E}$, evaluated at the equilibrium, and determine if the resulting matrix of values is positive definite. Here it is simpler to take the 3 distances $d_{12}$, $d_{23}$ and $d_{31}$ as the degrees of freedom, with the general form, starting from Eqn.\ \ref{eq:dsym}, substituting the equilibrium condition and simplifying
\begin{eqnarray}
\frac{\partial^2\mathcal{E}}{\partial d_{ij}^2} & = & \frac{m_i m_j}{d I_H} \left[ \left(m_i m_j - 3 m_k(m_i+m_j)\right) d^2 - 3 I_S\right] \\
\frac{\partial^2\mathcal{E}}{\partial d_{ij}\partial d_{jk}} & = & \frac{4 m_i m_j^2 m_k}{d I_H}
\end{eqnarray}
where $d \ge 1-r_3$ for the specific case of interest.
For the full Hessian of $\mathcal{E}$, $[\partial^2\mathcal{E}/\partial d_{ij} d_{jk} ]$, to be positive definite utilize Sylvester's Theorem again, which states that a necessary and sufficient condition is that all of the principal minors of the Hessian matrix be positive. Thus, a necessary condition for being positive definite is that the diagonals all be positive. Should any of these be negative, then the matrix is not positive definite and hence the relative equilibrium configuration is not energetically stable. Consider the entry for $\frac{\partial^2\mathcal{E}}{\partial d_{23}^2}$. The controlling condition for stability is then that $\left[ m_2 m_3 - 3 m_1(m_2+m_3)\right] d^2 - 3 I_S$ be positive.
However, it is easy to show that the term $\left[ m_2 m_3 - 3 m_1(m_2+m_3)\right] < 0$, showing that the Lagrange Orbital configurations are always energetically unstable, consistent with Moeckel's result. First, restate the negative condition as $3 m_1 (m_2 + m_3) > m_2 m_3$, then note that $m_2 + m_3 > m_3$ and $m_1 > m_2$, establishing the inequality unequivocally.
\subsubsection{Euler Solutions}
For the Euler conditions consider $d_{ij} \ge r_i+r_j$, $d_{jk} \ge r_j+r_k$, and $\theta_{ki} = \pi$. This case also has $d_{ki} = d_{ij} + d_{jk}$. Both Eqns.\ \ref{eq:3bpij} and \ref{eq:3bpjk} must equal zero in this case, yielding the two conditions.
\begin{eqnarray}
0 & = & m_i m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ij}^3} \right] d_{ij} + m_i m_k \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] d_{ki} \\
0 & = & m_k m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{jk}^3} \right] d_{jk} + m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] d_{ki}
\end{eqnarray}
First, there is a more fundamental equality within these results
\begin{eqnarray}
m_i m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ij}^3} \right] d_{ij} & = m_k m_i \left[ \frac{H^2}{I_H^2} - \frac{1}{d_{ki}^3} \right] d_{ki} = & m_k m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{jk}^3} \right] d_{jk}
\end{eqnarray}
By inspection, with the knowledge that $d_{ki} \ge d_{jk}, d_{ij}$, note that
\begin{eqnarray}
\frac{1}{d_{ki}^3} & \le \frac{H^2}{I_H^2} \le & \frac{1}{\max{ (d_{jk}, d_{ij})}^3}
\end{eqnarray}
Alternately, this ratio can also be solved for the quantity $(H/I_H)^2$ to find
\begin{eqnarray}
\frac{H^2}{I_H^2} & = & \frac{1}{m_j d_{ij} + m_k d_{ki}} \left[ \frac{m_j}{d_{ij}^2} + \frac{m_k}{d_{ki}^2} \right] \\
\frac{H^2}{I_H^2} & = & \frac{1}{m_j d_{jk} + m_i d_{ki}} \left[ \frac{m_j}{d_{jk}^2} + \frac{m_i}{d_{ki}^2} \right]
\end{eqnarray}
which is the condition used to analyze how the ER configurations fissioned. Indeed, at the transition lines on Figs.\ \ref{fig:EA123} and \ref{fig:EA312} the resting configuration is in fact a central configuration, meaning that the relative attractions between the bodies will be balanced so long as their relative distances are preserved.
This can be generalized to identify the possible bifurcation pattern in the EO configurations. Assume, say, that bodies $i$ and $j$ are in contact and that as body $k$ is moved to a larger distance it reaches the point where the equality between the above conditions occurs, meaning that bodies $i$ and $j$ are now in a relative equilibrium condition and the entire system satisfies a central configuration conditions \cite{wintner}. At this point the relative distances between these bodies can be uniformly scaled with the ratio $H/I_H$ following along. The change in angular momentum with this scaling is not uniform, however, due to the $3I_S$ term in $I_H$. Of specific interest regarding the pattern of bifurcation is whether the angular momentum decreases or increases with this change in relative distance. In the following it can be shown that both conditions can occur in general.
Define the distance between bodies $k$ and $i$ where the EO conditions are satisfied (assuming $d_{ij} = 1-r_k$) as $d_{ki}^*$, and thus $d_{jk}^* = d_{ki}^* - (1-r_k)$, and define the ratio $H/I_H = \Omega^*$ at this point. Then, for increasing the distance the relative equilibria will all scale uniformly, meaning that
\begin{eqnarray}
\frac{H^2}{I_H^2} & = & \frac{\Omega^{*2}}{d^3}
\end{eqnarray}
where $d \ge 1$ and $d_{ij} = d(1-r_k)$, $d_{jk} = d d^*_{jk}$ and $d_{ki} = d d^*_{ki}$. With this structure, it is possible to compute the gradient
\begin{eqnarray}
\frac{\partial H}{\partial d} & = & \frac{\Omega^{*2}}{2 d^{5/2}} \left[ \left( m_i m_j (1-r_k)^2 + m_j m_k (d_{ki}^{*}-(1-r_k))^2 + m_k m_i d_{ki}^{*2}\right) d^2 - 3 I_S\right]
\end{eqnarray}
How the bifurcation occurs can be tested by plotting the level sets from $d_{ki}^* = (1+r_j)$ to large values. For the 1, 2, 3 and 2, 1, 3 orderings the gradient is positive towards the apex of the triangular region and can take on negative values near the base. Thus, the appearance of the EO orbits occur as a transition closer to the equal mass condition and as a bifurcation followed by a termination away from there. Precise limits could be computed, but would require root solving algorithms.
Finally, consider the stability of the Euler Orbital solutions. These are again suspected to be energetically unstable due to the instability of the point mass cases, however this should be checked given the changes in the current approach. First, note that the second order variation in $\theta_{ki}$ is uncoupled from the variations in distance, and evaluated at the equilibrium yields
\begin{eqnarray}
\delta^2_{\theta_{ki}} {\cal E} & = & - m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] d_{ij} d_{jk} (\delta \theta_{ki})^2
\end{eqnarray}
The quantity in the brackets is negative, as established above, and hence the angle variation is stable. For the distance variations the full $2\times 2$ Hessian matrix must be evaluated, however one can again just check the necessary conditions that the diagonals must all be positive. Taking the second order variations of both conditions from the diagonal of the Hessian matrix shows that both of the following conditions must be positive for stability
\begin{eqnarray}
m_i m_j - 3 (m_i+m_j)m_k & > & 0 \\
m_j m_k - 3 (m_j+m_k)m_i & > & 0
\end{eqnarray}
For the ordering $m_3 \le m_2 \le m_1$ it can be shown that for all combinations of $i$, $j$ and $k$ that at least one of these conditions will be violated, and hence the EO configurations are always unstable.
\end{proof}
\section{Summary}
To finish, the results are presented in light of the main theorem. First, consider the total number of relative equilibria found. For the no contact case recall the classical result of 5 distinct relative equilibria. When one contact is active the EA and IS relative equilibria were identified, which have 6 unique components each, raising the count to 17. For the two contact cases there are 3 ER and 6 TR configurations, resulting in a total of 26. Finally, for three contact cases there are the 2 LR configurations, leading to the total of 28.
Now consider the bifurcation patterns, which are focused on the transitions between the different contact cases, and the identification of when the $H$-Bifurcations can occur. The details of the transitions will be outlined, although a few observations can be given first. With regard to stability the system starts with only two stable LR configurations at low values of $H$ and eventually has six stable EA configurations for arbitrarily large values of $H$. Between these limits the number of stable configurations can vary, and to establish the precise sequence and number would require a more detailed investigation for a specific set of sizes. It is noted, however, that there always exist at least one stable configuration.
The bifurcation pattern seen in Fig.\ \ref{fig:bif2} is described first.
The LR configurations all end at a Termination Fission condition with the TR132 configurations. These configurations either arise from an $H$-Bifurcation (in a limited region of the parameter space) or more commonly emerge as a symmetric bifurcation as the unstable ER132 configuration stabilizes. Under increasing $H$ the ER132 configuration either ends with a Termination Fission with the unstable component of the EA13-2 configuration, or for a limited range of parameters ends with a Transition Fission into the EA13-2 configuration. This second occurrence is of great interest as it is the only occasion in which the Spherical Full 3-Body Problem will fission into a stable configuration. The EA13-2 configuration itself usually arises as an $H$-Bifurcation, with its unstable branch terminating as mentioned above and its stable branch existing for all $H$ with an increasing distance proportional to $H^2 $. The only exception is when it arises as a Transition Fission, as described above. The EA23-1 configuration has an evolution that is completely isolated from the rest of this chart. It arises through an $H$-Bifurcation, with its unstable component either having a Termination Fission with the EO132 configuration or a Transition Fission into an EO132 configuration. Its stable component continues for all larger values of $H$ with a similar asymptotic form as the EA13-2 configuration. The EO132 configuration can either arise as an $H$-Bifurcation or through a Transition Fission, however the EO components are always unstable.
It is important to note that at each stage of the system evolution with $H$, that there is at least one stable configuration, providing proof of that aspect of the theorem.
Now consider the bifurcation patterns for the ER123 and ER213 pathways, shown in Figs.\ \ref{fig:bif1} and \ref{fig:bif3}. These are similar, and distinct from the ER132 pathway. Each of these start out in unstable ER configurations and both stabilize by a symmetric bifurcation with the TR123 and TR213 configurations, respectively. The TR123 and TR213 configurations end with a Termination Fission with an IS23-1 and IS13-2 configuration, respectively. The IS configurations arise through an $H$-Bifurcation with the inner configuration ending with the Termination Fission mentioned above and the outer configuration extending for all $H$, ultimately with their size on the order of $H^2$, although the IS configurations are always unstable. The stable ER123 and ER213 configurations end with a Termination Fission into an EA12-3 or EA32-1 configuration for the ER123 case or an EA21-3 or EA31-2 configuration for the ER213 case. Limits where these transitions occur have been delineated in Figs.\ \ref{fig:EA123} and \ref{fig:EA312}. Note that at the transition between these fission pattern the ER123 and ER213 configurations are in a central configuration, a situation that does not happen for the ER132 configuration. When the EA inner configurations do not end with a Termination Fission with an ER configuration, they end with a Termination Fission or Transition Fission with the corresponding EO configuration. The outer EA configurations are all stable and have the same asymptotic structure for large $H$. The EO123 and EO213 configurations either arise as an $H$-Bifurcation or as a Transition Fission, with the details of these boundaries left for future investigation.
Finally consider the sequence involving the LO configuration, represented in Fig.\ \ref{fig:bif4}. This sequence is the least complex, with the IS12-3 configurations arising from an $H$-Bifurcation. The inner component ends with a Termination Fission or Transition Fission with the LO configuration. The LO configuration, in turn, either arises as an $H$-Bifurcation, with the inner component ending with a Termination Fission, or as a Transition Fission. All configurations in these sequences are unstable.
\section*{Appendix}
\subsection*{Partial Derivatives}
It is useful to state the relevant partial derivatives of the amended potential and its constituent terms, as a function of the distances and angles. In the following use the convention that the distances are denoted with indices $ij$ and $jk$ and the angle with indices $ki$.
If the third degree of freedom is the angle $\theta_{ki}$ then
\begin{eqnarray}
\frac{\partial {\cal E}}{\partial d_{ij}} & = & - \frac{H^2}{2 I_H^2} \frac{\partial I_H}{\partial d_{ij}} + \frac{\partial \mathcal{U} }{\partial d_{ij}} \\
\frac{\partial {\cal E}}{\partial \theta_{ki}} & = & - \frac{H^2}{2 I_H^2} \frac{\partial I_H}{\partial \theta_{ki}} + \frac{\partial \mathcal{U} }{\partial \theta_{ki}}
\end{eqnarray}
where
\begin{eqnarray}
\frac{\partial I_H}{\partial d_{ij}} & = & 2 m_i m_j d_{ij} + 2 m_i m_k \left( d_{ij} - d_{jk} \cos\theta_{ki} \right) \\
\frac{\partial I_H}{\partial \theta_{ki}} & = & 2 m_i m_k d_{ij} d_{jk} \sin\theta_{ki}
\end{eqnarray}
Similarly
\begin{eqnarray}
\frac{\partial \mathcal{U} }{\partial d_{ij}} & = & m_i m_j \frac{1}{d_{ij}^2} + m_i m_k \frac{\left( d_{ij} - d_{jk} \cos\theta_{ki} \right)}{d_{ki}^3} \\
\frac{\partial \mathcal{U} }{\partial \theta_{ki}} & = & m_i m_k \frac{ d_{ij} d_{jk} \sin\theta_{ki} }{d_{ij}^3}
\end{eqnarray}
If the third degree of freedom is the distance $d_{ki}$, and if not at a limiting constraint, then
\begin{eqnarray}
\frac{\partial {\cal E}}{\partial d_{ij}} & = & - \frac{H^2}{2 I_H^2} \frac{\partial I_H}{\partial d_{ij}} + \frac{\partial \mathcal{U} }{\partial d_{ij}}
\end{eqnarray}
where
\begin{eqnarray}
\frac{\partial I_H}{\partial d_{ij}} & = & 2 m_i m_j d_{ij}
\end{eqnarray}
and
\begin{eqnarray}
\frac{\partial \mathcal{U} }{\partial d_{ij}} & = & m_i m_j \frac{1}{d_{ij}^2}
\end{eqnarray}
\subsection*{Equilibrium Conditions}
For a relative equilibrium there are two different possibilities. Either $\delta{\cal E}=0$ or $\delta{\cal E} > 0$.
For either, the relevant statement of the variations is given in the following for the two different formulations of the third degree of freedom.
If the third degree of freedom is the angle $\theta_{ki}$ then the full set of variations are
\begin{eqnarray}
\delta_{d_{ij}} {\cal E} & = & m_i \left\{ m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ij}^3} \right] d_{ij} \right. \nonumber \\
& & \left. + m_k \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] \left[ d_{ij} - d_{jk}\cos\theta_{ki}\right] \right\} \delta d_{ij} \label{eq:3bpij} \\
\delta_{d_{jk}} {\cal E} & = & m_k \left\{ m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{jk}^3} \right] d_{jk} \right. \nonumber \\
& & + \left. m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] \left[ d_{jk} - d_{ij}\cos\theta_{ki}\right] \right\} \delta d_{jk} \label{eq:3bpjk}
\end{eqnarray}
\begin{eqnarray}
\delta_{\theta_{ki}} {\cal E} & = & m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] d_{ij} d_{jk} \sin\theta_{ki} \delta \theta_{ki} \label{eq:3bpki}
\end{eqnarray}
If the third degree of freedom is the distance $d_{ki}$ and the system is not at a constraint limit (i.e., $d_{ki} \ne |d_{ij} \pm d_{jk}|$), then the full set of variations are
\begin{eqnarray}
\delta_{d_{ij}} {\cal E} & = & m_i m_j \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ij}^3} \right] d_{ij} \delta d_{ij} \label{eq:dsym} \\
\delta_{d_{jk}} {\cal E} & = & m_j m_k \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{jk}^3} \right] d_{jk} \delta d_{jk} \\
\delta_{d_{ki}} {\cal E} & = & m_k m_i \left[ - \frac{H^2}{I_H^2} + \frac{1}{d_{ki}^3} \right] d_{ki} \delta d_{ki}
\end{eqnarray}
These expressions are used to develop the necessary and sufficient conditions for a configuration to be a relative equilibrium. If any two components are in contact, then the condition for the degree of freedom that is blocked should be $\delta{\cal E} > 0$ for all allowed variations, otherwise the condition should be $\delta{\cal E} = 0$.
\subsection*{Properties of the Function $F(\mu,x)$}
\begin{lemma}
The function
\begin{eqnarray}
F(\mu,x) = \frac{1+\frac{\mu}{x^2}}{1+\mu x}
\end{eqnarray}
is monotonically decreasing in $x$ and is convex in $x$ over the interval $x\in ( 0, \infty)$.
\end{lemma}
\begin{proof}
Consider the first derivative of the function with respect to $x$:
\begin{eqnarray}
F'(\mu,x) & = & \frac{- 2 \mu}{x^3} \frac{1}{1+\mu x} - \mu \frac{1+\frac{\mu}{x^2}}{(1+\mu x)^2}
\end{eqnarray}
By inspection it can be seen that all terms are negative and non-zero, and thus the function is monotonically decreasing in $x$.
Taking the second derivative with respect to $x$:
\begin{eqnarray}
F''(\mu,x) & = & \frac{6 \mu}{x^4} \frac{1}{1+\mu x} + \frac{4 \mu^2}{x^3} \frac{1}{(1+\mu x)^2}
+ 2 \mu^2 \frac{1+\frac{\mu}{x^2}}{(1+\mu x)^3}
\end{eqnarray}
By inspection again see that all terms are positive and non-zero and is thus convex.
\end{proof}
\section*{Acknowledgements}
The author acknowledges support from NASA grant NNX14AL16G from the Near Earth Objects Observation programs.
\newpage \bibliographystyle{plain}
|
\section{INTRODUCTION}
\label{sec:introduction}
%
%
%
%
%
%
%
%
Robotic setups often need fine-tuned controller parameters both at low- and task-levels. Finding an appropriate set of parameters through simplistic protocols, such as manual tuning or grid search, can be highly time-consuming. We seek to automate the process of fine tuning a nominal controller based on performance observed in experiments on the physical plant. We aim for information-efficient approaches, where only few experiments are needed to obtain improved performance.
Designing controllers for balancing systems such as in \cite{trimpe2012balancing} or \cite{mason2014full} are typical examples for such a scenario. Often, one can without much effort obtain a rough linear model of the system dynamics around an equilibrium configuration, for example, from first principles modeling. Given the linear model, it is then relatively straightforward to compute a stabilizing controller, for instance, using optimal control. When testing this nominal controller on the physical plant, however, one may find the balancing performance unsatisfactory, e.g. due to unmodeled dynamics, parametric uncertainties of the linear model, sensor noise, or imprecise actuation. Thus, fine-tuning the controller gains in experiments on the real system is desirable in order to partly mitigate these effects and obtain improved balancing performance.
We have a tuning scenario in mind, where a limited budget of experimental evaluations is allowed (e.g.\ due to limited experimental time on the plant, or costly experiments). The automatic tuning shall globally explore a given range of controllers
and return the best known controller after a fixed number of experiments.
During exploration, we assume that it is acceptable for the controller to fail, for example, because other safety mechanisms are in place \cite{akametalu2014reachability}, or it is uncritical to stop an experiment when reaching safety limits (as is the case in experiment considered herein).
For this scenario, we propose a controller tuning framework extending previous work \cite{trimpe2014self}.
Therein, a Linear Quadratic Regulator (LQR) is iteratively improved based on control performance observed in experiments. The controller parameters of the LQR design are adjusted using Simultaneous Perturbation Stochastic Approximation (SPSA) \cite{spall2003simultaneous} as optimizer of the experimental cost. It obtains a very rough estimate of the cost function gradient from few cost evaluations, and then updates the parameters in its negative direction. While control performance could be improved in experiments on a balancing platform in \cite{trimpe2014self}, this approach does not exploit the available data as much as could be done. Additionally, rather than exploring the space globally, it only finds local minima.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{./Pics/apollo_pole_balancing_brighter_low_q.eps}
\caption{The humanoid robot Apollo learns to balance poles of different lengths using the automatic controller tuning framework proposed herein.}
\label{fig:apollo}
\end{figure}
In contrast to \cite{trimpe2014self}, we propose the use of Entropy Search (ES) \cite{hennig2012entropy,villemonteix2009informational}, a recent algorithm for global Bayesian optimization, as the minimizer for the LQR tuning problem. ES employs a Gaussian process (GP) as a non-parametric model capturing the knowledge about the unknown cost function. At every iteration, the algorithm exploits all past data to infer the shape of the cost function. Furthermore, in the spirit of an active learning algorithm, it suggests the next evaluation in order to learn most about the location of the minimum. Thus, we expect ES to be more data-efficient than simple gradient-based approaches as in \cite{trimpe2014self}; that is, to yield better controllers with fewer experiments.
The main contribution of this paper is the development of an automatic controller tuning framework combining ES \cite{hennig2012entropy} with LQR tuning \cite{trimpe2014self}.
While ES has been applied to numerical optimization problems before, this work is the first to use it for controller tuning on a complex robotic platform. The effectiveness of the proposed auto-tuning method is demonstrated in experiments of a humanoid robot balancing a pole
%
(see Figure \ref{fig:apollo}). We present successful auto-tuning experiments for parameter spaces of different dimensions (2D and 4D), as well as for initialization with relatively good, but also poor initial controllers.
Preliminary results of this approach are presented in the workshop paper \cite{marco2015irosws}.
The presentation in this paper is more elaborate and new experimental results are included.
\textit{Related work:}
Automatic tuning of an LQR is also considered in
\cite{Gr84} and \cite{ClKaMo85}, for example. In these references, the tuning typically happens by first identifying model parameters from data, and then computing a controller from the updated model. In contrast, we tune the controller gain directly thus bypassing the model identification step. Albeit we exploit a nominal model in the LQR design, this model is not updated during tuning and merely serves to pre-structure the controller parameters.
Using Gaussian processes (GPs) for automatic controller tuning has recently also been proposed in \cite{schreiter20015,Berkenkamp2016SafeController,calandra2015bayesian,metzen2015ActiveBayesOpt}. In \cite{schreiter20015}, the space of controller parameters is explored by selecting next evaluation points of maximum uncertainty (i.e. maximum variance of the GP). In contrast, ES uses a more sophisticated selection criterion: it selects next evaluation points where the expected information gain is maximal in order to learn most about the global minimum. A particular focus of the method in \cite{schreiter20015} is on safe exploration. For this purpose, an additional GP distinguishing safe and unsafe regions (e.g. corresponding to unstable controllers) is learned.
Safe learning is also the focus in \cite{Berkenkamp2016SafeController}, where the Bayesian optimization algorithm for safe exploration from \cite{sui2015safe} is employed. This work restricts the exploration to controllers that incur a small cost with high probability. The method avoids unsafe controllers and finds the optimum within the safely reachable set of controllers. In contrast, ES explores globally and maximizes information gain in the entire parameter space, regardless of a potentially large costs incurred in an individual experiment.
The authors in \cite{calandra2015bayesian} use Bayesian optimization for learning gait parameters of a walking robot. The gait is achieved using a discrete event controller, and transitions are triggered based on sensor feedback and the learned parameters.
Same as herein, \cite{metzen2015ActiveBayesOpt} also uses Entropy Search for controller tuning, and extends this to contextual policies for different tasks. While \cite{schreiter20015} and \cite{metzen2015ActiveBayesOpt} present simulation studies (balancing an inverted pendulum and robot ball throwing, respectively), \cite{Berkenkamp2016SafeController} and \cite{calandra2015bayesian} demonstrate their algorithms in hardware experiments (quadrocopter and 4-DOF walking robot). To the authors' knowledge, \cite{Berkenkamp2016SafeController} and the work herein are the first to propose and experimentally demonstrate Bayesian optimization for direct tuning of continuous state-feedback controllers on a real robotic platform.
The task of learning a controller from experimental rewards (i.e.\ negative cost) is also considered in the rather large area of reinforcement learning (RL), see \cite{KoBaPe13} for a survey. However, the tools used here (GP-based optimization) differ from the classical methods in RL.
\textit{Outline of the paper:} The LQR tuning problem is described in Sec.~\ref{sec:problem}. The use of \ess for automating the tuning is outlined in Sec.~\ref{sec:es}. The experimental results are presented in Sec.~\ref{sec:results}, and the paper concludes with remarks in Sec.~\ref{sec:conc}.
\section{LQR TUNING PROBLEM}
\label{sec:problem}
In this section, we formulate the LQR tuning problem following the approach proposed in \cite{trimpe2014self}.
\subsection{Control design problem}
\label{ssec:control_des}
We consider a system that follows a discrete-time non-linear dynamic model
\begin{equation}
\bmd{x}_{k+1} = \bmd{f}(\bmd{x}_k,\bmd{u}_k,\bmd{w}_k)
\label{eq:non_lin_sys}
\end{equation}
with system states $\bmd{x}_k\in \mathbb{R}^{n_x}$, control input $\bmd{u}_k\in \mathbb{R}^{n_u}$, and zero-mean process noise $\bmd{w}_k$ at time instant $k$. We assume that (\ref{eq:non_lin_sys}) has an equilibrium at $\bmd{x}_k=\bmd{0}$, $\bmd{u}_k=\bmd{0}$ and $\bmd{w}_k=\bmd{0}$, which we want to keep the system at. We also assume that $\bmd{x}_k$ can be measured and, if not, an appropriate state estimator is used.
For regulation problems such as balancing about an equilibrium, a linear model is often sufficient for control design. Thus, we consider a scenario, where a linear model
\begin{equation}
\tilde{\bmd{x}}_{k+1}=\bmd{A}_{\text{n}}\tilde{\bmd{x}}_k + \bmd{B}_{\text{n}}\bmd{u}_k + \bmd{w}_k
\label{eq:nom_model}
\end{equation}
is given as an approximation of the dynamics (\ref{eq:non_lin_sys}) about the equilibrium at zero. We refer to (\ref{eq:nom_model}) as the \textit{nominal model}, while (\ref{eq:non_lin_sys}) are the true system dynamics, which are unknown.
A common way to measure the performance of a control system is through a quadratic cost function such as
\begin{equation}
J = \lim_{K\rightarrow\infty}\dfrac{1}{K}\mathbb{E}\left[ \sum_{k=0}^{K-1} \bmd{x}_k^T\bmd{Q}\bmd{x}_k + \bmd{u}_k^T\bmd{R}\bmd{u}_k \right]
\label{eq:cost_def}
\end{equation}
with positive-definite weighting matrices $\bmd{Q}$ and $\bmd{R}$, and $\mathbb{E}\left[ \cdot \right] $ the expected value. The cost (\ref{eq:cost_def}) captures a trade-off between control performance (keeping $\bmd{x}_k$ small) and control effort (keeping $\bmd{u}_k$ small).%
Ideally, we would like to obtain a state feedback controller for the non-linear plant (\ref{eq:non_lin_sys}) that minimized (\ref{eq:cost_def}). Yet, this non-linear control design problem is intractable in general. Instead, a straightforward approach that yields a locally optimal solution is to compute the optimal controller minimizing (\ref{eq:cost_def}) for the nominal model (\ref{eq:nom_model}). This controller is given by the well-known Linear Quadratic Regulator (LQR) \cite[Sec. 2.4]{anderson2007optimal}
\begin{equation}
\bmd{u}_k = \bmd{F}\bmd{x}_k
\label{eq:static_gain}
\end{equation}
whose static gain matrix $\bmd{F}$ can readily be computed by solving the discrete-time infinite-horizon LQR problem for the nominal model $(\bmd{A}_\text{n},\bmd{B}_\text{n})$ and the weights $(\bmd{Q},\bmd{R})$. For simplicity, we write
\begin{equation}
\bmd{F}=\text{lqr}(\bmd{A}_\text{n},\bmd{B}_\text{n},\bmd{Q},\bmd{R}).
\label{eq:controller_no_par}
\end{equation}
If (\ref{eq:nom_model}) perfectly captured the true system dynamics (\ref{eq:non_lin_sys}), then (\ref{eq:controller_no_par}) would be the optimal controller for the problem at hand. However, in practice, there can be several reasons why the controller (\ref{eq:controller_no_par}) is suboptimal: the true dynamics are non-linear, the nominal linear model (\ref{eq:nom_model}) involves parametric uncertainty, or the state is not perfectly measurable (e.g. noisy or incomplete state measurements). While still adhering to the controller structure (\ref{eq:static_gain}), it is thus beneficial to fine tune the nominal design (the gain $\bmd{F}$) based on experimental data to partly compensate for these effects. This is the goal of the automatic tuning approach, which is detailed next.
\subsection{LQR tuning problem}
\label{ssec:LQRpar}
Following the approach in \cite{trimpe2014self}, we parametrize the controller gains $\bmd{F}$ in (\ref{eq:static_gain}) as
\begin{equation}
\bmd{F}(\bmd{\theta})=\text{lqr}(\bmd{A}_\text{n},\bmd{B}_\text{n},\bmd{W}_x(\bmd{\theta}),\bmd{W}_u(\bmd{\theta}))
\label{eq:controller}
\end{equation}
where $\bmd{W}_x(\bmd{\theta})$ and $\bmd{W}_u(\bmd{\theta})$ are \textit{design weights} parametrized in $\bmd{\theta}\in \mathbb{R}^D$, which are to be varied in the automatic tuning procedure. For instance, $\bmd{W}_x(\bmd{\theta})$ and $\bmd{W}_u(\bmd{\theta})$ can be diagonal matrices with $\theta_j >0$, $j=1,\ldots,D$, as diagonal entries.
Parametrizing controllers in the LQR weights $\bmd{W}_x$ and $\bmd{W}_u$ as in \eqref{eq:controller}, instead of varying the controller gains $\bmd{F}$ directly, restricts the controller search space. This restriction is often desirable for practical reasons. First, we assume that the nominal model (albeit not perfect) represents the true dynamics reasonable well around an equilibrium. In this situation, one wants to avoid controllers that destabilize the nominal plant or have poor robustness properties, which is ensured by the LQR design\footnote{According to classical results in control theory \cite{kalman1964linear} and \cite{KoGoSe12}, any stabilizing feedback controller \eqref{eq:static_gain} that yields a return difference greater one (in magnitude) can be obtained for some $\bmd{W}_x$ and $\bmd{W}_u$ as the solution to the LQR problem. The return difference is relevant in the analysis of feedback loops \cite{anderson2007optimal}, and its magnitude exceeding one means favorable robustness properties. Therefore, the LQR parameterization \eqref{eq:controller} only discards controllers that are undesirable because they destabilize the nominal plant, or have poor robustness properties.}.
Second, further parametrizing $\bmd{W}_x$ and $\bmd{W}_u$ in $\bmd{\theta}$ can be helpful to focus on most relevant parameters or to ease the optimization problem. While, for example, a restriction to diagonal weights $\bmd{W}_x$ and $\bmd{W}_u$ is common practice in LQR design (i.e. $n_x + n_u$ parameters), it is not clear how one would reduce the dimensionality of the gain matrix $\bmd{F}$ ($n_x\times n_u$ entries) when tuning this directly. We expect this to be particularly relevant for high-dimensional problems, such as control of a full humanoid robot \cite{mason2014full}.
When varying $\bmd{\theta}$, different controller gains $\bmd{F}(\bmd{\theta})$ are obtained. These will affect the system performance through (\ref{eq:static_gain}), thus resulting in a different cost value from (\ref{eq:cost_def}) in each experiment. To make the parameter dependence of (\ref{eq:cost_def}) explicit, we write
\begin{equation}
J = J(\bmd{\theta}).
\end{equation}
The goal of the automatic LQR tuning is to vary the parameters $\bmd{\theta}$ such as to minimize the cost (\ref{eq:cost_def}).
\textit{Remark:} The weights $(\bmd{Q},\bmd{R})$ in (\ref{eq:cost_def}) are referred to as \textit{performance weights}. Note that, while the \textit{design weights} $\left( \bmd{W}_x(\bmd{\theta}),\bmd{W}_u(\bmd{\theta})\right) $ in (\ref{eq:controller}) change during the tuning procedure, the performance weights remain unchanged.
\subsection{Optimization problem}
\label{ssec:opti}
The above LQR tuning problem is summarized as the optimization problem
\begin{equation}
\arg\min J(\bmd{\theta}) \quad \text{s.t. } \bmd{\theta} \in \mathcal{D}
\label{eq:min_fun}
\end{equation}
where we restrict the search of parameters to a bounded domain $\mathcal{D}\subset \mathbb{R}^D$.
The domain $\mathcal{D}$ typically represents a region around the nominal design, where performance improvements are to be expected or exploration is considered to be safe.
The shape of the cost function in (\ref{eq:min_fun}) is unknown. Neither gradient information is available nor guarantees of convexity can be expected. Furthermore, (\ref{eq:cost_def}) cannot be computed from experimental data in practice as it represents an infinite-horizon problem. As is also done in \cite{trimpe2014self}, we thus consider the approximate cost
\begin{equation}
\hat{J} = \dfrac{1}{K}\left[ \sum_{k=0}^{K-1} \bmd{x}_k^T\bmd{Q}\bmd{x}_k + \bmd{u}_k^T\bmd{R}\bmd{u}_k \right]
\label{eq:cost_approx}
\end{equation}
with a finite, yet long enough horizon $K$. The cost (\ref{eq:cost_approx}) can be considered a noisy evaluation of (\ref{eq:cost_def}). Such an evaluation is expensive as it involves conducting an experiment, which lasts few minutes in the considered balancing application.
\section{LQR TUNING WITH ENTROPY SEARCH}
\label{sec:es}
In this section, we introduce Entropy Search (ES) \cite{hennig2012entropy} as the optimizer to address problem (\ref{eq:min_fun}). The key characteristics of \ess are explained in Sec.~\ref{ssec:gp_theory} to \ref{ssec:loss}, the resulting framework for automatic LQR tuning is summarized in Sec.~\ref{ssec:aLQR}, and Sec.~\ref{ssec:OtherMethodsThanES} briefly discusses related methods.
Here, we present only the high-level ideas of ES from a practical standpoint. The reader interested in the mathematical details, as well as further explanations, is referred to \cite{hennig2012entropy}.
\subsection{Underlying cost function as a Gaussian process}
\label{ssec:gp_theory}
ES is one of several popular formulations of Bayesian Optimization \cite{kushner1964new,jones1998efficient,srinivas2009gaussian,auer2003using,osborne2009gaussian}, a framework for global optimization in which uncertainty over the objective function $J$ is represented by a probability measure $p(J)$, typically a Gaussian process (GP) \cite{rasmussen2006gaussian}. The shape of the cost function (\ref{eq:cost_def}) is unknown; only noisy evaluations (\ref{eq:cost_approx}) are available.
A GP is a probability measure over a space of functions. It encodes the knowledge we have about the underlying cost function. Additional information about this cost, gathered through experiments (i.e. noisy evaluations of it), is incorporated by conditioning, which is an analytic operation if the evaluation noise is Gaussian; refer to \cite{rasmussen2006gaussian} for more details.
We model prior knowledge about $J$ as the GP
\begin{equation}
J(\bmd{\theta})\sim \mathcal{GP}\left( \mu(\bmd{\theta}),k(\bmd{\theta},\bmd{\theta}_*)\right)
\label{eq:gp}
\end{equation}
with mean function $\mu(\bmd{\theta})$ and covariance function $k(\bmd{\theta},\bmd{\theta}_*)$. Common choices are a zero mean function ($\mu(\bmd{\theta})=\bmd{0}$ for all $\bmd{\theta}$), and the squared exponential (SE) covariance function
\begin{equation}
k_\text{SE}(\bmd{\theta},\bmd{\theta}_*) = \sigma^2\exp\left[ -\dfrac{1}{2}(\bmd{\theta}-\bmd{\theta}_*)^\text{T}\bmd{S}(\bmd{\theta}-\bmd{\theta}_*) \right]
\label{eq:ker}
\end{equation}
which we also use herein. The covariance function $k(\bmd{\theta},\bmd{\theta}_*)$ captures the covariance between $J(\bmd{\theta})$ and $J(\bmd{\theta}_*)$. It can thus be used to encode assumptions about properties of $J$ such as smoothness, characteristic length-scales, and signal variance. In particular, the SE covariance function \eqref{eq:ker} models very smooth functions with signal variance $\sigma^2$ and length-scales $\bmd{S}=\text{diag}(\lambda_1,\lambda_2,\ldots,\lambda_D)$, $\lambda_j>0$.
We assume that the noisy evaluations (\ref{eq:cost_approx}) of (\ref{eq:cost_def}) can be modeled as
\begin{equation}
\hat{J}=J(\bmd{\theta})+\varepsilon
\label{eq:lik}
\end{equation}
with Gaussian noise $\varepsilon$ of variance $\sigma_\text{n}^2$, yielding the likelihood.
To simplify notation, we write $\bmd{\ev} = \{\hat{J}^i\}_{i=1}^N $ for $N$ evaluations at locations $\bmd{\Theta}=\{ \bmd{\theta}^i\}_{i=1}^N $. Conditioning the GP on the data $\left\lbrace \ev,\bmd{\Theta} \right\rbrace $ then yields another GP with posterior mean $\bar{\mu}(\bmd{\theta})$ and a posterior variance $\bar{k}(\bmd{\theta},\bmd{\theta}_*)$.
Figure \ref{fig:gp} provides an example for a one-dimensional cost function and three successive function evaluations.
As can be seen, the shape of the mean is adjusted to fit the data points, and the uncertainty (standard deviation) is reduced around the evaluations. In regions where no evaluations have been made, the uncertainty is still large.
We gather the hyperparameters of the GP in the set $\mathcal{H}=\left\lbrace \lambda_1,\lambda_2,\ldots,\lambda_D,\sigma,\sigma_\text{n}\right\rbrace$.
An initial choice of $\mathcal{H}$ is improved with every new data point $\hat{J}^i$ %
by maximizing the marginal likelihood, a popular approximation.
In addition, we use automatic relevance determination \cite[Sec. 5.1]{rasmussen2006gaussian} in the covariance function (\ref{eq:ker}), which removes those parameter dimensions with low influence on the cost as more data points become available.
\subsection{Probability measure over the location of the minimum}
\label{ssec:p_min}
A key idea of ES
is to explicitly represent the probability $p_\text{min}(\bmd{\theta})$
for the minimum location over the domain $\mathcal{D}$:
\begin{equation}
p_\text{min}(\bmd{\theta})\equiv p( \bmd{\theta} = \arg \min J(\bmd{\theta}) ), \quad \bmd{\theta} \in \mathcal{D} .
\label{eq:p_min_rough}
\end{equation}
The probability $p_\text{min}(\bmd{\theta})$ is induced by the GP for $J$: given a distribution of cost functions $J$ as described by the GP, one can in principle compute the probability for any $\bmd{\theta}$ of being the minimum of $J$. For the example GPs in Fig.~\ref{fig:gp}, $p_\text{min}(\bmd{\theta})$ is shown in green.
To obtain a tractable algorithm, ES approximates $p_\text{min}(\bmd{\theta})$ with finitely many points on a non-uniform grid that puts higher resolution in regions of greater influence.
\subsection{Information-efficient evaluation decision}
\begin{figure}[t!]
\centering
\subfigure[1 evaluation]{
\input{./Pics/GP1D/GP_post1.tex}
\includegraphics[width=1\columnwidth]{./Pics/GP1D/GP_post1.eps}}
\subfigure[2 evaluations]{
\input{./Pics/GP1D/GP_post2.tex}
\includegraphics[width=1\columnwidth]{./Pics/GP1D/GP_post2.eps}}
\subfigure[3 evaluations]{
\input{./Pics/GP1D/GP_post3.tex}
\includegraphics[width=1\columnwidth]{./Pics/GP1D/GP_post3.eps}}
\caption{Evolution of an example Gaussian process
for three successive function evaluations (orange dots),
reproduced with slight alterations from \cite{hennig2012entropy}.
The posterior mean $\bar{\mu}(\bmd{\theta})$ is shown in thick violet, two standard deviations $2\bar{\sigma}(\bmd{\theta})$ in thin violet, and the probability density as a gradient of color that decreases away from the mean. Two standard deviations of the likelihood noise 2$\sigma_\text{n}$ are represented as orange vertical bars at each evaluation.
Approximated probability distribution over the location of the minimum $p_\text{min}(\bmd{\theta})$ in green. This plot uses arbitrary scales for each object.}
\label{fig:gp}
\end{figure}
\label{ssec:loss}
The key feature of ES is the suggestion of new locations $\bmd{\theta}$, where \eqref{eq:cost_approx} should be evaluated to learn most about the location of the minimum. This is achieved by selecting the next evaluation point that maximizes the relative entropy
\begin{equation}
H = \int_{\mathcal{D}}{p_\text{min}({\bmd{\theta}})\log \dfrac{p_\text{min}({\bmd{\theta}})}{b({\bmd{\theta}})}}\text{d}\bmd{\theta}
\label{eq:entropy}
\end{equation}
between $p_\text{min}(\bmd{\theta})$ and the uniform distribution $b(\bmd{\theta})$ over the bounded domain $\mathcal{D}$. The rationale for this is that the uniform distribution essentially has no information about the location of the minimum, while a very ``peaked'' distribution would be desirable to obtain distinct potential minima. This can be achieved by maximization of the relative entropy \eqref{eq:entropy}.
For this, ES selects next evaluations where the first order expansion $\Delta H(\bmd{\theta})$ of the expected change in (\ref{eq:entropy}) is maximal. In this way, the algorithm efficiently explores the domain of the optimization problem in terms of information gain (cf. \cite[Sec. 2.5]{hennig2012entropy}). Conceptually, the choice of the locations $\bmd{\Theta}$ is made such that ``we evaluate where we expect to learn most about the minimum, rather than where we think the minimum is'' \cite[Sec. 1.1]{hennig2012entropy}.
In addition to suggesting the next evaluation, ES also returns its current \textit{best guess} of the minimum location; that is, the maximum of its approximation to $p_\text{min}(\bmd{\theta})$.
\subsection{Automatic LQR tuning}
\label{ssec:aLQR}
The proposed method for automatic LQR tuning is obtained by combining the LQR tuning framework from Section \ref{sec:problem} with \es; that is, using ES to solve \eqref{eq:min_fun}. At every iteration, ES suggests a new controller (through $\bmd{\theta}$ with \eqref{eq:controller}), which is then tested in an experiment to obtain a new cost evaluation (\ref{eq:cost_approx}). Through this iterative procedure, the framework is expected to explore relevant regions of the cost \eqref{eq:cost_def}, infer the shape of the cost function, and eventually yield the global minimum within $\mathcal{D}$. The automatic LQR tuning method is summarized in Algorithm \ref{alg:aLQR}.
The performance weights $(\bmd{Q},\bmd{R})$ encode the desired performance for the system (\ref{eq:non_lin_sys}). Thus, a reasonable initial choice of the parameters $\bmd{\theta}$ is such that the design weights $\left( \bmd{W}_x(\bmd{\theta}),\bmd{W}_u(\bmd{\theta})\right) $ equal $(\bmd{Q},\bmd{R})$. The obtained initial gain $\bmd{F}$ would be optimal if (\ref{eq:nom_model}) were the true dynamics. After $N$ evaluations,
ES aims to improve this initial choice, by selecting experiments which are expected to provide maximal information about a better parameter setting.
\begin{algorithm}[t!]
\caption{{\small Automatic LQR Tuning. As its inputs, \textsc{EntropySearch} takes the type of covariance function $k$, the likelihood $l$, a fixed number of evaluations $N$, and data points $\{\bmd{\Theta},\bmd{y}\}$. Alternative stopping criteria instead of stopping after $N$ iterations can be used.
}}
\begin{algorithmic}[1]
\State $\text{initialize } \bmd{\theta}^0 \text{; typically } \bmd{W}_x(\bmd{\theta}^0)=\bmd{Q}\text{, }\bmd{W}_u(\bmd{\theta}^0)=\bmd{R}$
\State $\hat{J}^{0}\gets$ \Call{CostEvaluation}{$\bmd{\theta}^{0}$} \Comment{Cost evaluation}
\State $\{\bmd{\Theta},\bmd{y}\} \gets \{\bmd{\theta}^0,\hat{J}^0\}$
\Procedure{EntropySearch}{$k$,$l$,$N$,$\{\bmd{\Theta},\bmd{y}\}$}
\For{$i=1\text{ to }N$}
\State $[ \bar{\mu},\bar{k}] \gets \text{GP}(k,l,\{\bmd{\Theta},\bmd{y}\})$ \Comment{GP posterior}
\State $p_\text{min} \gets \text{approx\_pmin}(\bar{\mu},\bar{k})$ \Comment{Approximate $p_\text{min}$}
\State $\bmd{\theta}^{i}\gets \arg\max\Delta H$ \Comment{Next location to evaluate at}
\State $\hat{J}^{i}\gets$ \Call{CostEvaluation}{$\bmd{\theta}^{i}$} \Comment{Cost evaluation}
\State $\{\bmd{\Theta},\bmd{y}\} \gets \{\bmd{\Theta},\bmd{y}\}\cup\{\bmd{\theta}^{i},\hat{J}^{i}\}$
\State $\bmd{\theta}^\text{BG}\gets \arg\max p_\text{min}$ \Comment{Update current ``best guess''}
\EndFor
\State \Return $\bmd{\theta}^\text{BG}$
\EndProcedure
\Statex
\Function{CostEvaluation}{$\bmd{\theta}$}
\State LQR design: $\bar{\bmd{F}}\gets \text{lqr}(\bmd{A}_\text{n},\bmd{B}_\text{n},\bmd{W}_x(\bmd{\theta}),\bmd{W}_u(\bmd{\theta}))$
\State update control law (\ref{eq:static_gain}) with $\bmd{F}=\bar{\bmd{F}}$
\State perform experiment and record $\{\bmd{x}_k\},\{\bmd{u}_k\}$
\State Evaluate cost: $\hat{J}\gets\frac{1}{K}\left[ \sum_{k=0}^{K-1} \bmd{x}_k^T\bmd{Q}\bmd{x}_k + \bmd{u}_k^T\bmd{R}\bmd{u}_k \right]$
\State \Return $\hat{J}$
\EndFunction
\end{algorithmic}
\label{alg:aLQR}
\end{algorithm}
\subsection{Relation to other GP-based optimizers}
\label{ssec:OtherMethodsThanES}
In addition to the novel ES, there exist a number of Bayesian optimization algorithms based on Gaussian process (GP) measures over the optimization objective. Most of these methods do not retain an explicit measure over the location of the optimum. While ES aims at collecting information about the minimum, these methods directly try to collect small function values (a concept known as \emph{minimizing regret}).
This strategy is encoded in several different heuristic evaluation utilities, including \emph{probability of improvement} (PI) \cite{kushner1964new}, \emph{expected improvement} (EI) \cite{jones1998efficient} and \emph{upper confidence bound} for GP bandits (GP-UCB) \cite{srinivas2009gaussian,auer2003using}. PI is the probability that an evaluation at a specific point lies below the current best guess, and EI is the expected value by which an evaluation might lie below the current best guess. GP-UCB captures the historically popular notion of ``optimism in the face of uncertainty'' and has the analytic appeal of coming with a theoretical worst-case performance guarantee.
The key difference between ES and these other methods is that they directly try to design experiments that yield increasingly low function values. This is the right strategy in settings where the performance of each individual experiment matters, e.g. the gait of a walking robot is improved online, but it is not allowed to fall. However, in a ``prototyping'' setting, where the sole use of experiments is to learn about a final good design, the numerical result of each experiment is less important than its information content.
The proposed automatic controller tuning framework relies on the ``prototyping'' setting, for which each experiment should be as informative as possible about the global minimum of the cost function.
One minor downside of ES is that it has higher computational cost than alternative methods, taking several seconds to decide on the next experiment. However, in our setting, where the physical experiments take significantly longer than this time, this is not a major drawback.
\section{EXPERIMENTAL RESULTS}
\label{sec:results}
In this section, we present auto-tuning experiments for learning to balance a pole as shown in Fig. \ref{fig:apollo}.
A video demonstration that illustrates the second experiment described in Sec.~\ref{ssec:results_2D} is available at \burl{https://am.is.tuebingen.mpg.de/publications/marco_icra_2016}.
\subsection{System description}
We consider a one-dimensional balancing problem: a pole linked to a handle through a rotatory joint with one degree of freedom (DOF) is kept upright by controlling the acceleration of the end-effector of a seven DOF robot arm (Kuka lightweight robot). Figure \ref{fig:apollo} shows the setup for two poles of different length.
The angle of the pole is tracked using an external motion capture system.
The continuous-time dynamics of the balancing problem (similar to \cite{schaal1997learning}) are described by:
\begin{align}
& mr^2\ddot{\psi}(t)-mgr\sin \psi(t)+mr\cos \psi(t)u(t)+\xi\dot{\psi}(t) = 0 \nonumber \\
& \ddot{s}(t) = u(t)
\label{eq:pole_dyn}
\end{align}
where $\psi(t)$ is the pole angle with respect to the gravity axis, $s(t)$ is the deviation of the end-effector from the zero position, and $u(t)$ is the end-effector acceleration.
Two poles with different lengths are used in the experiments.
The center of mass of the short pole lies at $r\simeq 0.33$ m from the axis of the rotatory joint, its mass is $m\simeq 0.27$ kg, the
friction coefficient is $\xi\simeq 0.012$ Nms, and the gravity constant is $g=9.81$ m/s$^\text{2}$.
For the long pole, we have $r\simeq 0.64$ m and $m\simeq 0.29$ kg.
A model (\ref{eq:nom_model}) of the system is obtained by linearization of (\ref{eq:pole_dyn}) about the equilibrium $\psi=0$, $s=0$ and discretization with a sampling time of 1 ms. Using the parameters of the short pole, we obtain its nominal model $(\bmd{A}_\text{n},\bmd{B}_\text{n})$.
The non-linear model (\ref{eq:pole_dyn}) assumes that we can command a discretized end-effector acceleration $u_k$ as control input to the system. In reality, this end-effector acceleration is realized through an appropriate tracking controller for the end-effector following a similar control structure as in \cite{righetti2014autonomous}.
The estimated end-effector position $s_k$ and velocity $\dot{s}_k$ are computed at a sampling rate of 1kHz from the robot's joint encoders using forward kinematics. The pole orientation is captured at 200 Hz by the motion capture system. From this data, we obtain estimates of pole angle $\psi_k$ and angular velocity $\dot{\psi}_k$ through numerical differentiation and low-pass filtering (2nd-order Butterworth, 10 Hz cutoff). With this scheme, no model is required to obtain estimates of all states (in contrast to the Kalman filter used in \cite{marco2015irosws}), and it can be used irrespective of which balancing pole is used.
The complete state vector of (\ref{eq:nom_model}) is given by $\bmd{x}_k = [ \psi_k, \dot{\psi}_k, s_k, \dot{s}_k]^\text{T}$.
When using a state-feedback controller (\ref{eq:static_gain}) for balancing, biases in the angle measurement
lead to a steady-state error in the end-effector position (cf. discussion in \cite[p. 67]{trimpe2012balancing} for a similar balancing problem). To compensate for such steady-state offsets, the state feedback controller (\ref{eq:static_gain}) is augmented with an integrator on the end-effector position, which is a standard way to achieve zero steady-state error (see e.g. \cite[Sec. 6.4]{astrom2010feedback}). That is, we implement the control law $u_k = \bmd{F}\bmd{x}_k + F_zz_k$ instead of (\ref{eq:static_gain}), where $z_k$ is the integrator state.
Although $F_z$ can readily be included in the LQR formulation (\ref{eq:controller}) and tuned alongside the other gains (as was done in \cite{marco2015irosws}), we fix $F_z=-0.3$ here for simplicity. Since the integrator is not a physical state (it is implemented in the controller) and merely affects the long-term behavior, we do not include it in the computation of the cost (\ref{eq:cost_approx}).
\subsection{Automatic LQR tuning: Implementation choices}
We choose the performance weights to be
\begin{equation}
\bmd{Q}=\text{diag}(1,100,10,200),\;\bmd{R}=10
\label{eq:per_weights}
\end{equation}
where $\text{diag}(\cdot)$ denotes the diagonal matrix with the arguments on the diagonal. We desire to have a quiet overall motion in the system. Therefore, we penalize the velocities $\dot{\psi}_k$ and $\dot{s}_k$ more than the position states.
We conducted two types of tuning experiments, one with two parameters and another one with four. The corresponding design weights are
\begin{itemize}
\item 2D tuning experiments:
\begin{equation}
\bmd{W}_x(\bmd{\theta})=\text{diag}(1,50\theta_1,10,50\theta_2),\; \bmd{W}_u(\bmd{\theta})=10
\label{eq:des_weights_ex_2D}
\end{equation}
where the parameters $\bmd{\theta} = \left[\theta_1,\theta_2 \right]$ can vary in $[0.01, 10]$, and $\bmd{\theta}^\text{0}=\left[2, 4\right] $ is chosen as initial value.
\item 4D tuning experiments:
\begin{equation}
\begin{array}{l}
\bmd{W}_x(\bmd{\theta})=\text{diag}(\theta_1,25\theta_2,10\theta_3,25\theta_4),\\
\bmd{W}_u(\bmd{\theta})=10
\end{array}
\label{eq:des_weights_ex_4D}
\end{equation}
with $\bmd{\theta} = \left[\theta_1,\theta_2,\theta_3,\theta_4 \right]$, $\theta_j \in [0.01, 10]$, and $\bmd{\theta}^\text{0}=\left[1, 4, 1, 8\right] $.
\end{itemize}
In both cases, the initial choice $\bmd{\theta}^\text{0}$ is such that the design weights equal the performance weights. That is, the first controller tested corresponds to the nominal LQR design \eqref{eq:controller_no_par}.
\begin{table}[b!]
\caption{Characterization of the gamma prior over $\mathcal{H}$}
\begin{center}
\begin{tabular}{rm{5ex}m{5ex}||m{5ex}m{5ex}@{}m{0pt}@{}}
& \multicolumn{2}{c||}{2D exploration} & \multicolumn{2}{c}{4D exploration} & \\
\cline{2-5}
& $\mathbb{E}\left[ \cdot \right] $ & $\text{Std}\left[ \cdot \right] $ & $\mathbb{E}\left[ \cdot \right] $ & $\text{Std}\left[ \cdot \right] $ & \\[1ex]
\hline
Lengthscale $\lambda_j$ & $2.5$ & $0.11$ & $2.00$ & $0.63$ & \\
Signal variance $\sigma$ & $0.2$ & $0.02$ & $0.75$ & $0.075$ & \\
Likelihood noise $\sigma_\text{n}$ & $0.033$ & $0.0033$ & $0.033$ & $0.010$ & \\
\end{tabular}
\end{center}
\label{tab:gamma_prior}
\end{table}
Balancing experiments were run for 2 minutes, i.e. a discrete time horizon of $K=1.2\cdot 10^5$ steps.
We start the experiments from roughly the same initial condition. To remove the effect of the transient and slightly varying initial conditions, we omit the first $30\, \text{s}$ from each experiment.
Because the nominal model does not capture the true dynamics, some LQR controllers obtained during the tuning procedure destabilized the system. This means
that the system exceeded either acceleration bounds or safety constraints on the end-effector position.
In these cases, the experiment was stopped and a fixed heuristic cost $J_\text{u}$ was assigned to the experiment.
Values for $J_\text{u}$ are typically chosen slightly larger than the performance of a stable but poor controller. We used $J_\text{u}=3.0$ and $J_\text{u}=5.0$ for the 2D and 4D experiments, respectively.
\begin{figure}[t!]
\centering
\subfigure[ES initialization]{
\includegraphics[width=1\columnwidth]{./Pics/results_2D/short_iter1_polished_shorter_ind.eps}
}
\subfigure[GP posterior after 20 iterations]{
\includegraphics[width=01\columnwidth]{./Pics/results_2D/short_iter20_polished1_no_next_eval_no_labels_indi.eps}}
\caption{GPs at (a) the start and (b) the end of the first tuning experiment. The GP mean is represented in violet and $\pm$ two standard deviations in grey. The red dot corresponds to the initial controller, computed at location $\bmd{\theta}^\text{0}=\left[2,4\right] $. The green dot represents the current best guess for the location of the minimum. The blue dot is the location suggested by ES to evaluate next,
and orange dots represent previous evaluations.
The best guess found after 20 iterations (green dot in (b)) has significantly lower cost than the initial controller (red dot).
}
\label{fig:short_pole_experiments}
\end{figure}
Before running ES, a few experiments were done
to acquire knowledge about the hyperparameters $\mathcal{H}$. A Gamma prior distribution was assumed over each hyperparameter with expected values and variances shown in Table \ref{tab:gamma_prior}. For the first iteration of ES, we use these expectations as initial set $\mathcal{H}$. After each iteration, $\mathcal{H}$ is updated as the result of maximizing the GP marginal likelihood.
\subsection{Results from 2D experiments}
\label{ssec:results_2D}
For the 2D experiments \eqref{eq:des_weights_ex_2D}, we first use a short pole (Fig.~\ref{fig:apollo}, right) and the best available linear model, showing that the framework is able to improve the initial controller. Secondly, we exchange the pole with one of double length (Fig.~\ref{fig:apollo}, left), but keep the same nominal model.
We show, for the latter case, that even with a $50\%$ underestimated model, the framework finds a stable controller with good performance. In both cases, we use the design weights (\ref{eq:des_weights_ex_2D}).
\subsubsection{Using an accurate nominal model}
ES was initialized with five evaluations, i.e. the initial controller $\bmd{\theta}^\text{0}$, and evaluations at the four corners of the domain $[0.01,10]^2$.
Figure \ref{fig:short_pole_experiments} (a) shows the 2D Gaussian process including the five initial data points.
The algorithm can also work without these initial evaluations; however, we found that they provide useful prestructuring of the GP and tend to speed up the learning. This way, the algorithm focuses on interesting regions more quickly.
Executing Algorithm \ref{alg:aLQR} for 20 iterations (i.e. 20 balancing experiments) resulted in the posterior GP shown in Figure \ref{fig:short_pole_experiments} (b).
The ``best guess'' $\bmd{\theta}^\text{BG}=\left[ 0.01,2.80 \right] $ (green dot) is what \ess suggests to be the location of the minimum of the underlying cost (\ref{eq:cost_def}).
In order to evaluate the result of the automatic LQR tuning, we computed the cost of the resulting controller (best guess after 20 iterations) in five separate balancing experiments. The average and standard deviation of these experiments are shown in Table \ref{tab:results_exp1_2} (left column, bottom), together with the average and standard deviation of the initial controller, computed in the same way before starting the exploration (left column, top). Even though the initial controller was obtained from the best linear model we had, the performance was still improved by $31.9\%$.
\subsubsection{Using a poor nominal model}
In this experiment, we take the same nominal model as in the previous case, but we use a longer pole in the experimental demonstrator (Fig.~\ref{fig:apollo}, left). The initial controller, computed with $\bmd{\theta}^0$, destabilizes the system, which can be explained by the nominal model significantly misrepresenting the true dynamics. As shown in Figure \ref{fig:long_pole_experiments}, after 20 iterations, ES suggested $\bmd{\theta}^\text{BG}=\left[ 3.25,0.01 \right]$ as the best controller. The results of evaluating this controller five times, in comparison to the initial controller, are shown in Table \ref{tab:results_exp1_2} (middle column). %
\begin{figure}[t!]
\centering
\includegraphics[width=1\columnwidth]{./Pics/results_2D/long_iter20_polished1_no_next_eval_no_labels_indi.eps}
\caption{Final GP posterior for the second tuning experiment using a wrong nominal model. The color scheme is the same as in Fig.~\ref{fig:short_pole_experiments}.
}
\label{fig:long_pole_experiments}
\end{figure}
\begin{table}[b!]
\caption{Cost values $\hat{J}$ for three tuning experiments} %
\begin{center}
\begin{tabular}{rm{4ex}m{4ex}||m{4ex}m{4ex}||m{4ex}m{3ex}}
& \multicolumn{2}{c||}{2D experiments} & \multicolumn{2}{c||}{2D experiments} & \multicolumn{2}{c}{4D experiments} \\
& \multicolumn{2}{c||}{Good model} & \multicolumn{2}{c||}{Poor model} & \multicolumn{2}{c}{Poor model} \\
\cline{2-7}
& mean & std & mean & std & mean & std \\
\hline
$\bmd{\theta}^\text{0}$ & $1.12$ & $0.11$ & $J_\text{u}$ & - & $J_\text{u}$ & - \\
$\bmd{\theta}^\text{BG}$ & $0.76$ & $0.058$ & $0.059$ & $0.012$ & $0.040$ & $0.0031$ \\
\end{tabular}
\end{center}
\label{tab:results_exp1_2}
\end{table}
\subsection{Results from 4D experiment}
The 4D tuning experiment, realized with the long pole, uses the same nominal model as in the previous experiments (i.e., a poor linear model for the real plant), and the design weights (\ref{eq:des_weights_ex_4D}). We show that the framework is able to improve the controller found during the 2D experiments with the long pole, but in a higher dimensional space.
The first controller $\bmd{\theta}^0$ destabilizes the system. After 46 iterations, ES suggests $\bmd{\theta}^\text{BG}=\left[ 4.21,7.47,0.43,0.01 \right]$, which in comparison with the 2D experiments with the long pole, performs about $31.7\%$ better (see Table \ref{tab:results_exp1_2}). We actually ran this experiment until iteration 50, however, the algorithm did not lead to further improvements.
Figure \ref{fig:cost_entropy_bg} shows the cost function evaluations over the course of the tuning experiment.
The fact that unstable controllers are obtained throughout the experiment
reflects how the global search tends to cover all areas.
Before starting the 2D experiments, we spent some effort selecting the method's parameters, such as hyperparameters and parameter ranges.
In contrast, we started the 4D experiments without such prior tuning. In particular, we kept the same performance weights, chose similar design weights, and started with the same values for the hyperparameters $\mathcal{H}$ and penalty $J_\text{u}$.
However, we had to restart the search twice in order to slightly adjust $\mathcal{H}$, and $J_\text{u}$.
In general, any method reasoning about functions on continuous domains from a finite number of data points relies on prior assumptions (see \cite[Sec.~1.1]{hennig2012entropy} for a discussion).
We were quite pleased with the outcome of the tuning experiments and, in particular, that not much had to be changed moving from the 2D to 4D experiment.
Nonetheless,
developing general rules for choosing the parameters of GP-based optimizers like ES (maybe specific for certain problems) seems important for future developments.
\begin{figure}[t!]
\centering
\input{./Pics/results_4D/cost.tex}
\includegraphics[width=1\columnwidth]{./Pics/results_4D/cost.eps}
\caption{Cost values obtained at each experiment distinguishing stable controllers (blue dots), and unstable controllers (red dots). %
}
\label{fig:cost_entropy_bg}
\end{figure}
\section{CONCLUDING REMARKS}
\label{sec:conc}
In this paper, we introduce Bayesian optimization for automatic controller tuning. We develop, and successfully demonstrate in experiments on a robotic platform, a framework based on LQR tuning \cite{trimpe2014self} and Entropy Search (\es) \cite{hennig2012entropy}. This work is the first to apply ES in experiments for automatic controller tuning.
The auto-tuning algorithm was demonstrated in a 2D and a 4D experiment, both when the method was initialized with an unstable and with a stable controller. While the 2D experiment could presumably also be handled by grid search or manual tuning, and thus mostly served as a proof of concept, the 4D tuning problem can already be considered difficult for a human. A key question for the development of truly automatic tuning methods is the amount of ``prior engineering'' that has to be spent to get the method to work. In particular, the 4D experiments were promising since not a lot of tuning, and only few restarts were necessary. However, questions pertaining to the prior choice or automatic adjustment of the method's parameters are relevant for future work.
Since the ES algorithm reasons about where to evaluate next in order to maximize the information gain of an experiment, we expect it to make better use of the available data and yield improved controllers more quickly than alternative approaches. Although ES has been shown to have superior performance on numerical problems \cite{hennig2012entropy},
investigating whether this claim holds true in practice
is future work.
A more challenging robotics scenario would be to use a consumer-grade vision system mounted on the head of the robot, instead of the current motion tracking system. This would provide observations of the pole state at a slower rate, larger measurement noise, and a potentially large delay. All these aspects may increase the cost function evaluation noise. It would be interesting to see how the system generalizes to these kinds of conditions, in future work.
\section*{ACKNOWLEDGMENT}
The authors thank Felix Grimminger and Heiko Ott for their support with the robot hardware and the pole design. We are grateful to Ludovic Righetti for his advice regarding the low-level tracking controllers, as well as Alexander Herzog for his help with the motion capture system.
|
\section{Introduction}
A metasurface is a composite material layer, designed
and optimized in order to control and transform
electromagnetic fields. The layer thickness is negligible
as compared to the wavelength in the surrounding space.
Conventional devices for wave transformations are either bulky and heavy (e.g., reflector antennas or lenses) or complicated and require active elements (transmitarray antennas, also called array lenses \cite{Zoya,transmit}). Therefore, it is quite tempting to become able to realize desired transformations (for example, focus or refract wave beams) using extremely thin passive layers.
Recently, there has been considerable interest and progress in creating metasurfaces for controlling reflected and transmitted waves, see recent reviews in \cite{new_review,phil, Caloz_review,Elefth_review, Alu_review}. Some limited manipulations of waves transmitted through a thin metasurface can be accomplished due to specifically designed phase gradient over the metasurface plane \cite{capasso,shalaev,grady,shalaev2}. The required phase gradient is achieved by precise adjustment of the phases of transmitted waves from each metasurface inclusion. Although this approach has enabled realizations of transmitarrays even at optical frequencies, it suffers from very low efficiency (less than 25\% of transmitted power) and cannot provide control of polarization of the transmitted waves (in fact, it suffers from uncontrollable polarization rotation by $90^\circ$). Subsequently, another approach based on generalized boundary conditions and the use of symmetric metasurfaces was proposed by several researchers \cite{Tailoring,elefth,alu,caloz}. It provided more efficient operation (more than 80\% of transmitted power) and manipulation of polarization \cite{polarization}. However, even this approach cannot ensure ideal performance \cite{passive,Elefth_review}, because these symmetric layers cannot be matched to impedances of two propagating waves (incident and transmitted) in different directions, and, therefore, they inevitably produce some reflections. Most recently, in Refs.~\cite{elefth2,ours} it has been shown that the use of metasurfaces with asymmetric response can open a possibility to realize metasurfaces for perfect refraction.
Known structures for manipulating reflection (both reflectarrays and metasurfaces) are able to control reflection phase at each point of the reflector surface and nearly fully reflect the incident power. Representative examples can be found in papers \cite{sun,bozh1,bozh2,mosall2,nanotechnology,prx4,veysi,li}. It has been believed that these properties can allow full control over reflected waves. However, it is not the case. As is shown in the submitted paper \cite{Alu} and in this paper, lossless fully reflecting metasurfaces designed to reflect a plane wave into another plane wave, always produce parasitic beams in undesired directions.
Without proper understanding of the physical properties of metasurfaces responsible for refraction and reflection phenomena it is not possible to create 100\% efficient metasurfaces with desired properties.
Here we address this problem by introducing a general approach to the design of metasurfaces for arbitrary manipulations of plane waves, both in transmission and reflection. We explain the main ideas of the proposed analytical approach to the synthesis of general functional metasurfaces using simple but enough general examples of metasurfaces for refraction or reflection of plane waves into arbitrary directions. In the first example, a metasurface between two isotropic half-spaces (generally different) is designed so that a plane wave incident from one space (the incidence angle $\theta_{\rm i}$) is fully refracted into a plane wave propagating in the second space (the refraction angle $\theta_{\rm t}$), without polarization change. We derive general conditions on the equivalent circuit parameters of the metasurface to ensure perfect refraction while the reflection coefficient is exactly zero (see Section~\ref{conditions}). Subsequently, we consider three different metasurface scenarios to satisfy these conditions (Sections~\ref{sec:teleportation}, \ref{sec:transmitarray}, and \ref{sc3}). The latter scenario was independently considered in \cite{elefth2}.
In the second example, we show how to design metasurfaces which fully reflect plane waves into an arbitrary direction (the reflection angle $\theta_{\rm r}$). In this example, there are two plane waves coexisting in the space in front of the reflecting metasurface. This issue complicates the study, but the solution allows us to approach the problem of synthesis of metasurfaces for the most general field transformations, where the main challenge is to account for interference between multiple plane waves. Indeed, any arbitrary field distribution can be represented as a series of plane waves that interfere on both sides of the metasurface.
In Section~\ref{sec:power1}, we examine conditions on the metasurface parameters for the perfect reflection regime. Similar conditions were obtained independently in \cite{Alu}. Next, in Sections~\ref{sc4}, \ref{OptDes}, and \ref{sc5}, we consider different scenarios for metasurface realizations.
We show that perfect control over both refraction and reflection using lossless metasurfaces requires careful engineering of spatial dispersion in the structure. To realize perfect refraction, we need only weak spatial dispersion in form of the artificial magnetism and bianisotropic omega coupling \cite{serd}. This effect is described by local relations between the exciting electric and magnetic fields and the induced polarizations in the unit cells. Perfect control over reflection using lossless metasurfaces appears to be possible only using strongly non-local metasurfaces: part of the power received in one area of the surface should be ``channelled'' and re-radiated at a different part of the surface. Lossless local-response metasurfaces (that is, conventional reflectarrays and earlier studied metamirrors) cannot create a perfect reflected plane wave in any direction except the specular and retro directions.
The results clarify the necessary physical properties of metasurfaces for ideal wave refraction and reflection and explain the limitations of earlier used design methods and earlier studied realizations in form of electric and magnetic sheets, inhomogeneous high impedance surfaces and reflectarrays. Possible routes towards realization of ideal and full control over refraction and reflection are identified and discussed.
\section{Control of transmission: Perfectly refracting metasurface}
As a first step we consider the problem of synthesis of metasurfaces for control of transmitted waves. We require that a given plane wave is fully refracted into another plane wave, without reflections or energy loss. The geometry of the problem is illustrated in Fig.~\ref{geom}.
\begin{figure}[h!]
\centering
\epsfig{file=fig1.eps, width=0.6\linewidth}
\caption{Illustration of the desired performance of an ideally refracting metasurface.}\label{geom}
\end{figure}
The metasurface is located in the $yz$-plane between two isotropic half-spaces with the characteristic impedances $\eta_1$ and $\eta_2$. We assume, without loss of generality of the approach, a transverse electric (TE, with respect to the normal to the surface) incident plane wave. Our approach can be used for waves of arbitrary polarizaitons, including arbitrary polarization transformations, by using the dyadic parameters instead of the scalar ones.
Let us assume that the metasurface is illuminated from medium 1 by a plane wave (with the wavenumber $k_1$ and the electric field vector $\_E_{\rm i}$) at an angle $\theta_{\rm i}$.
Requiring zero reflections, the tangential field components $\_E_{t1}$ and $\_H_{t1}$ on the illuminated side of the metasurface (at $x=0$) read
\begin{equation} \_E_{t1}=\_E_{\rm i}e^{-jk_1\sin\theta_{\rm i}z},\quad \_n\times \_H_{t1}=
\_E_{\rm i}{1\over \eta_1}\cos\theta_{\rm i} e^{-jk_1\sin\theta_{\rm i}z}, \l{plus} \end{equation}
where $z$ is the coordinate along the tangential component of the incident wavevector and the unit vector $\_n$ is orthogonal to the metasurface plane, pointing towards the source. The time-harmonic dependency in form $e^{j\omega t}$ is assumed.
We want to synthesize a metasurface which will transform this incident wave into a refracted wave
propagating in medium 2 (characterized by parameters $k_2$, $\eta_2$) in some other direction, specified by the angle $\theta_{\rm t}$, without any loss of power. Therefore, the required tangential fields behind the metasurface read
\begin{equation} \_E_{t2}=\_E_{\rm t}e^{-jk_2\sin\theta_{\rm t}z+j\phi_{\rm t}},\quad \_n\times \_H_{t2}=
\_E_{\rm t}{1\over \eta_2}\cos\theta_{\rm t} e^{-jk_2\sin\theta_{\rm t}z+j\phi_{\rm t}}. \l{minus}\end{equation} For generality, we assume that the refracted wave is phase-shifted by an arbitrary angle $\phi_{\rm t}$ with respect to the incident plane wave. With these notations, we can choose the origin of the $z$-axis so that both $\_E_{\rm i}$ and $\_E_{\rm t}$ will be real-valued vectors.
Obviously, the phase of the transmission coefficient
\begin{equation} \Phi_{\rm t}(z)=\angle (E_{t2}/E_{t1})=
-k_2\sin\theta_{\rm t}z+\phi_{\rm t}+
k_1\sin\theta_{\rm i}z \l{viki}\end{equation}
is not uniform over the surface, as long as $k_2 \sin\theta_{\rm t}\neq k_1 \sin\theta_{\rm i}$. Differentiating the above equation, one can find the
relation between the incidence and refraction angles in terms of the transmission coefficient phase gradient:
\begin{equation} k_1\sin\theta_{\rm i}-k_2\sin\theta_{\rm t}={d\Phi_{\rm t}(z)\over dz}.\l{Gen_Snell}\end{equation}
This result suggests the simplest approach to the realization of refractive surfaces: Designing a locally-periodical surface whose transmission coefficient is unity in the absolute value (lossless Huygens' sheet) and the phase of the transmission coefficient linearly changes in accordance with \r{Gen_Snell}. This method was used for a long time in antenna engineering (e.g. \cite{1979}) and more recently in designs of metasurfaces, both in microwaves (e.g. \cite{PRX}) and optics (e.g. \cite{capasso}). However, this approach does not lead to the desired perfect refraction \cite{passive}, and next we will explain how the desired performance can be realized exactly.
\subsection{Conditions on the equivalent circuit parameters}\label{conditions}
First, let us find the amplitude of the transmitted wave $\_E_{\rm t}$ which corresponds to full power transmission through the metasurface. Looking for possible realizations as metasurfaces with local response, we equate the normal (to the metasurface) components of the Poynting vector at each point of the metasurfaces, in the two media:
\begin{equation} {1\over 2}{\rm Re}(\_E_{t1}\times \_H_{t1}^*) = {1\over 2}{\rm Re}(\_E_{t2}\times \_H_{t2}^*), \end{equation}
and substitute the field values from \r{plus} and \r{minus}.
As a result, for metasurfaces with locally full power transmission we obtain
\begin{equation} \_E_{\rm t}=\_E_{\rm i}\sqrt{\cos\theta_{\rm i}\over \cos\theta_{\rm t}}\sqrt{\eta_2\over \eta_1}.\l{amp_tr}\end{equation}
Note that the amplitude of the transmitted wave can be larger or smaller than the amplitude of the incident plane wave, although the metasurface is lossless and the power is conserved in transmission. This result already tells about a limitation of the mentioned above simple design approach based only on engineering the transmission phase according to \r{Gen_Snell}.
Let us write the linear relations between the tangential fields at the two sides of the metasurface in form of an impedance matrix:
\begin{equation} \_E_{t1}=Z_{11} \_n\times \_H_{t1} +Z_{12} (-\_n\times \_H_{t2}), \l{11-12} \end{equation}
\begin{equation} \_E_{t2}=Z_{21} \_n\times \_H_{t1} +Z_{22} (-\_n\times \_H_{t2}),\l{21-22} \end{equation}
and find such values of the $Z$-parameters which correspond to this particular field transformation. Knowing the $Z$-parameters of a metasurface, we will be able to determine suitable topologies of constitutive elements (the unit-cell structures) which will realize the desired functionality. Furthermore, the use of the equivalent $T$-circuit (Fig.~\ref{T}) helps in understanding what physical properties the metasurface should have in order to provide the desired response.
\begin{figure}[h!]
\centering
\epsfig{file=fig2.eps, width=0.5\linewidth}
\caption{Equivalent $T$-circuit of a reciprocal metasurface for the considered case of one linear polarization (TE).}\label{T}
\end{figure}
Substituting the field values from \r{plus}, \r{minus}, and \r{amp_tr}, we get the following equations for the $Z$-parameters:
\begin{equation}
\begin{split} e^{-jk_1\sin\theta_{\rm i}z} & = Z_{11}\, {1\over \eta_1}\cos\theta_{\rm i} \, e^{-jk_1\sin\theta_{\rm i}z} \\
& -Z_{12}\, {1\over\sqrt{\eta_1 \eta_2}} \sqrt{\cos\theta_{\rm i} \cos\theta_{\rm t}} e^{-jk_2\sin\theta_{\rm t}z+j \phi_{\rm t}},\l{eq1}
\end{split}
\end{equation}
\begin{equation}
\begin{split} e^{-jk_2\sin\theta_{\rm t}z+j \phi_{\rm t}}& =Z_{21}\, {1\over\sqrt{\eta_1 \eta_2}} \sqrt{\cos\theta_{\rm i} \cos\theta_{\rm t}}\, e^{-jk_1\sin\theta_{\rm i}z}\\
& - Z_{22}\, {\cos\theta_{\rm t}\over \eta_2}\, e^{-jk_2\sin\theta_{\rm t}z+j \phi_{\rm t}}.\l{eq2a}\end{split}
\end{equation}
Obviously, there are infinitely many solutions for the unknown $Z$-parameters, because we have only two conditions imposed on four complex parameters. Note that solutions with complex values of impedance parameters mean that some of the components forming the metasurface are either lossy or active, but all these solutions still correspond to the {\it overall lossless} response of the metasurface, because the fields on the two sides of the metasurface form plane waves carrying the same power in both upper and lower half-spaces.
This observation suggests that we can impose some restrictions on the values of the
equivalent parameters of the metasurface for a specific transformation and achieve different realizations of metasurfaces which all perform the same operation on incident plane waves. The possibility of multiple realizations of arbitrary metasurfaces using the susceptibility model was discussed in Ref.~\cite{caloz}.
\subsection{Teleportation metasurface}
\label{sec:teleportation}
Considering equations \r{eq1} and \r{eq2a}, we observe that while the left-hand sides are single exponential functions (corresponding to either incident or transmitted wave), the right-hand sides are sums of two different exponential functions. This property indicates that in general the $Z$-parameters of the metasurface will depend on the coordinate $z$, that is, the metasurface is, in general, not uniform. However, there is an interesting and conceptually simple solution corresponding to a \emph{homogeneous} metasurface. If we assume that $Z_{12}=Z_{21}=0$, then both equations are satisfied with
\begin{equation} Z_{11}={\eta_1\over \cos\theta_{\rm i}},\qquad Z_{22}=-{\eta_2\over \cos\theta_{\rm t}}.\end{equation}
In this scenario, the metasurface is formed by a matched absorbing layer (the input resistance $Z_{11})$, a perfect electric conductor (PEC) sheet, and an active layer (an ``anti-absorber'' \cite{alutele,tele}) on the other side. The incident plane wave is totally absorbed in the matched absorber. The negative-resistance sheet (resistance $Z_{22}$) together with the wave impedance of medium 2 forms a self-oscillating system whose stable-generation regime corresponds to generation of a plane wave in the desired direction (the refraction angle $\theta_{\rm t}$). Indeed, the sum of the wave impedance of plane waves propagating at the angle $\theta_{\rm t}$ and the input impedance of the active layer is zero, and this is the necessary condition for stable generation. This structure is similar to the ``teleportation metasurface'' introduced in \cite{alutele,tele} for teleporting waves without changing the propagation direction. As shown in \cite{tele}, in that case if the reflector separating the resistive and active layers is made at least slightly imperfect (parameters $Z_{12}=Z_{21}$ are very small but not exactly zero), the amplitude and phase of the transmitted wave is synchronised with the incident field.
The teleportation metasurface is a theoretically perfect realization of the desired transformation of plane waves in transmission. In particular, when the incidence angle equals $\theta_{\rm i}$, the reflection coefficient is exactly zero. However, because the input resistance of the metasurface seen from medium 2 is negative, reflections of waves coming from this medium are very strong. Therefore, within the linear model of the negative resistance, the reflection coefficient tends to infinity for waves coming from the direction $\theta_{\rm t}$. Moreover, the amplitude of the field in medium 2 is established due to non-linear saturation of the negative resistance device. Therefore, it is probably practically impossible to ensure that the negative resistance saturates at exactly the desired amplitude of the generated wave.
Next, we consider an alternative realization, requiring perfect matching of the metasurface for waves coming from medium 2.
\subsection{Transmitarray}
\label{sec:transmitarray}
Let us consider alternative realizations demanding that the input impedance of the metasurface seen from medium 2 is matched to the wave impedance in medium 2, so that waves coming from the direction $\theta_{\rm t}$ will not produce any reflections. This requirement can be satisfied if we demand that
\begin{equation} Z_{22}={\eta_2\over \cos\theta_{\rm t}}.\end{equation}
Now we can find a realization of the metasurface as a nonreciprocal system where the ideal voltage source in the output branch is defined by
\begin{equation} Z_{21}={2\sqrt{\eta_1 \eta_2}\over \sqrt{\cos\theta_{\rm i} \cos\theta_{\rm t}}}\, e^{-j(k_2\sin\theta_{\rm t}-k_1\sin\theta_{\rm i})z+j \phi_{\rm t}},\end{equation}
as follows from \r{eq2a}.
If the desired response for illumination from medium 1 is the only requirement, we can set $Z_{12}=0$ and $Z_{11}={\eta_1\over \cos\theta_{\rm i}}$, so that for illuminations from medium 1 at the incidence angle $\theta_{\rm i}$ the metasurface is acting as a matched absorber (matched receiving antenna array).
This realization can be modeled by the corresponding nonreciprocal equivalent circuit, shown in Fig.~\ref{T_nonrec}.
\begin{figure}[h!]
\centering
\epsfig{file=fig3.eps, width=0.5\linewidth}
\caption{Equivalent $T$-circuit of a nonreciprocal transmitarray realization of refractive metasurfaces.}\label{T_nonrec}
\end{figure}
This realization reminds conventional transmitarrays \cite{Zoya}. The incident plane wave is received by a matched antenna array on one side of the surface and the wave is launched into medium 2 with a transmitting phase array antenna. In the ideal situation the transmitarray is overall lossless, as the resistance seen from the illuminated side is in fact the radiation resistance of the transmitting array (the two arrays need to be connected by matched cables).
The same model describes also the concept of field control and active cloaking using active Huygens' surfaces \cite{elefth,Elefth2}. In that scenario, there is no connection between the receiving side (realized as a matched absorber) and the active array. The incident field is assumed to be known and the amplitudes and phases of sources feeding the radiating array are set accordingly.
\subsection{Symmetrical double current sheets}
\label{sec:symmetrical}
Within the metasurface paradigm, the simplest approach to realization of refractive metasurfaces is to assume that the refraction is controlled by engineering surface densities of electric and magnetic current sheets, co-existing at the metasurface plane. It is obvious that sheets of \emph{only} electric or \emph{only} magnetic currents cannot offer the desired functionality because of the symmetry of the scattered fields in the forward and backward directions.
Because electric and magnetic surface current sheets are conveniently modeled by surface impedance relations, it appears reasonable to model refractive metasurfaces by two impedance relations which should hold both for the electric and magnetic surface current densities $\_J_{\rm e}$ and $\_J_{\rm m}$ \cite{Senior,Idemen,Kuester}:
\begin{equation} \_J_{\rm e}=\_n\times \_H_{t1}-\_n\times \_H_{t2}=Y_{\rm e}\_E_t=Y_{\rm e}{\_E_{t1}+\_E_{t2}\over 2},\l{Je}\end{equation}
\begin{equation} \_J_{\rm m}=-\_n\times(\_E_{t1}-\_E_{t2})=Y_{\rm m}\_H_t=Y_{\rm m}{\_H_{t1}+ \_H_{t2}\over 2}.\l{Jm} \end{equation}
Here $\_E_t$ and $\_H_t$ are the averaged tangential electric and magnetic fields at the metasurface plane. Forming sums and differences of \r{11-12} and \r{21-22}, it is easy to see that relations \r{Je} and \r{Jm} can hold only if the metasurface is symmetric and reciprocal, that is, when $Z_{11}=Z_{22}$ and $Z_{12}=Z_{21}$. Under these assumptions,
\begin{equation} Y_{\rm e}={2\over Z_{11}+Z_{12}}, \qquad Y_{\rm m}=2(Z_{11}-Z_{12}). \l{Yem}
\end{equation}
Since we have only two unknown complex parameters $Z_{11}$ and $Z_{12}$, the solution of \r{eq1} and \r{eq2a} becomes unique and it reads
\begin{equation} Z_{11}={\eta_1\over \cos\theta_{\rm i}} \frac{e^{-j\Phi_{\rm t}}+e^{j\Phi_{\rm t}}}
{e^{-j\Phi_{\rm t}}-{\eta_1 \cos\theta_{\rm t}\over \eta_2\cos\theta_{\rm i}} e^{j\Phi_{\rm t}} } ,\l{z_ap2}\end{equation}
\begin{equation} Z_{12}={\sqrt{\eta_1 \eta_2} \over \sqrt{\cos\theta_{\rm i} \cos\theta_{\rm t}} }\, {{\eta_1 \cos\theta_{\rm t}\over \eta_2\cos\theta_{\rm i}}+1 \over e^{-j\Phi_{\rm t}}- {\eta_1 \cos\theta_{\rm t}\over \eta_2\cos\theta_{\rm i}}e^{j\Phi_{\rm t}}},\l{z_ap1}\end{equation}
where $\Phi_{\rm t}$ is defined by \r{viki}.
We see that these parameters, as well as the electric sheet admittance and magnetic sheet impedance \r{Yem}, are complex numbers, which physically means that the surface is either lossy or active at different values of $z$. For a special case of refraction of a normally incident plane wave at $45^\circ$ such solution for sheet parameters has been published in \cite{Tailoring,circ} and later on discussed in e.g. \cite{Elefth_review}.
Inspecting \r{z_ap2} and \r{z_ap1}, we see that the metasurface parameters can be purely imaginary for all $z$, corresponding to passive lossless realizations, only if
\begin{equation} {\eta_1 \cos\theta_{\rm t}\over \eta_2\cos\theta_{\rm i}}=1,\l{Elefth_condition}\end{equation}
in which case
\begin{equation}
Z_{12}=j{\eta_1\over \cos\theta_{\rm i}}{1\over \sin \Phi_{\rm t}},\qquad
Z_{11}=j{\eta_1\over \cos\theta_{\rm i}}\cot\Phi_{\rm t}.\end{equation}
Corresponding surface admittances, given by \r{Yem}, are also purely imaginary and coincide with those derived in \cite{passive} in an alternative way.
Condition \r{Elefth_condition} physically means that the impedance of the incident plane wave at the top side of the metasurface (${\eta_1\over \cos\theta_{\rm i}}$) equals to the impedance of the refracted wave at the bottom side of the surface (${\eta_2\over \cos\theta_{\rm t}}$). It is, however, in contradiction with the desired field structure: Equations \r{plus} and \r{minus} imply that the ratio of the tangential field components (the wave impedance) must in general change if we require perfect refraction. Thus, lossless double current sheets modeled by impedance relations \r{Je} and \r{Jm} cannot realize perfectly refractive metasurfaces.
In paper \cite{passive} the requirement for equal impedances \r{Elefth_condition} was derived in a different way, demanding the absence of losses, and it was concluded that perfect refraction using lossless metasurfaces was not possible without reflections. Indeed, it is clear that adding some reflected field to \r{plus}, it is possible to make sure that the ratio of the tangential fields on top of the metasurface is the same as at the bottom. This approach is followed nearly in all current literature on lossless metasurfaces for refraction control: Nearly always only symmetric metasurfaces have been considered and used (see \cite{Caloz_review,Elefth_review}) and the realization is thought in form of a symmetric double-current sheet. This is the reason why earlier publications (see the review in \cite{Elefth_review}) state that there must be at least small reflections or there is a need to use active elements. The only known to us exception is the recent paper \cite{elefth2} where the problem is attacked using the generalized scattering matrix.
Next we show that perfect refraction at an angle which is not equal to the incidence angle is in fact possible using only lossless structures, but only if the surface is spatially dispersive, exhibiting bianisotropic omega coupling. This result has been independently obtained in \cite{Elefth_new_a}.
\subsection{Metasurface formed by lossless elements}\label{sc3}
In the above example realizations, metasurfaces contained both lossy and active elements, which may require complicated and expensive realizations. It is therefore of interest to consider if and how one can realize the same functionality using only reactive lossless components.
\subsubsection{Impedance matrix}
To answer this question, we again consider the main set of requirements on the $Z$-parameters of an ideal refractive metasurface
\r{eq1} and \r{eq2a} and look for a solution where all the $Z$-parameters are purely imaginary (i.e., $Z_{ij}=jX_{ij}$):
\begin{equation}
\begin{split} e^{-jk_1\sin\theta_{\rm i}z}&= jX_{11}\, {1\over \eta_1}\cos\theta_{\rm i} \, e^{-jk_1\sin\theta_{\rm i}z}\\
& -jX_{12}\, {1\over\sqrt{\eta_1 \eta_2}} \sqrt{\cos\theta_{\rm i} \cos\theta_{\rm t}} e^{-jk_2\sin\theta_{\rm t}z+j \phi_{\rm t}},
\end{split}
\end{equation}
\begin{equation}
\begin{split} e^{-jk_2\sin\theta_{\rm t}z+j \phi_{\rm t}}& =jX_{21}\, {1\over\sqrt{\eta_1 \eta_2}} \sqrt{\cos\theta_{\rm i} \cos\theta_{\rm t}}\, e^{-jk_1\sin\theta_{\rm i}z}\\
& - jX_{22}\, {\cos\theta_{\rm t}\over \eta_2}\, e^{-jk_2\sin\theta_{\rm t}z+j \phi_{\rm t}}.\end{split}
\end{equation}
This is a system of four real-valued equations for four real unknowns $X_{ij}$, which has a unique solution:
\begin{equation} X_{11}={\eta_1\over \cos\theta_{\rm i}}\cot \Phi_{\rm t} ,\l{x11}\end{equation}
\begin{equation} X_{22}={\eta_2\over \cos\theta_{\rm t}}\cot \Phi_{\rm t}, \l{x22}\end{equation}
\begin{equation} X_{12}=X_{21}={\sqrt{\eta_1 \eta_2}\over \sqrt{\cos\theta_{\rm i} \cos\theta_{\rm t}} }{1\over \sin\Phi_{\rm t}} .\l{x12} \end{equation}
For the case of zero phase shift ($\phi_{\rm t}=0$) formulas \r{x11}--\r{x22} agree with the result of \cite{elefth2}, obtained using the generalized scattering parameters approach.
The metasurfaces modeled by \r{x11}--\r{x12} are reciprocal ($X_{12}=X_{21}$). Indeed, the same solution follows from \r{eq1}--\r{eq2a} if we demand that a plane wave coming from the second medium (the incidence angle $\theta_{\rm t}$) is fully transmitted into the first medium in the direction $\theta_{\rm i}$.
The required physical properties of such metasurfaces
can be understood from the corresponding equivalent $T$-circuit (see Fig.~\ref{T}). The circuit is asymmetric, because $X_{11}\neq X_{22}$. This structure of the $Z$-matrix corresponds to bianisotropic omega layers, see a discussion in
\cite{Joni,AlbOmeg}.
Possible appropriate topologies include arrays of $\Omega$-shaped inclusions~\cite{PRL}, arrays of split rings, double arrays of patches (patches on the opposite sides of the substrate must be different to ensure proper magnetoelectric coupling) \cite{metamirror_AP,absorbers,abs_mohammad,abs_mohammad1}, etc. A more complicated set of three parallel reactive sheets was proposed in \cite{elefth2}.
Previuosly, probably only in paper \cite{elefth2} asymmetric metasurfaces were used for transmission management (equations \r{x11}--\r{x12} also appear in \cite{elefth2} for the case when $\phi_{\rm t}=0$). Note also that the role of the omega-type bianisotropy of metasurfaces has been discussed in the review paper \cite{phil}, and omega layers have been successfully used in single-layer metamirrors \cite{PRL}.
Comparing to the simple designs based on symmetrical metasurfaces (Section~\ref{sec:symmetrical}), we again see from \r{x11} and \r{x22} that lossless {\it symmetric} realizations with $X_{11}=X_{22}$ are possible only if ${\eta_1\over \cos\theta_{\rm i}}={\eta_2\over \cos\theta_{\rm t}}$, as we already saw from requirement \r{Elefth_condition}. If media $1$ and $2$ are the same, we can conclude that previously proposed symmetrical metasurfaces cannot provide perfect refraction (without parasitic reflections or energy loss).
\subsubsection{Unit-cell polarizabilities and appropriate topologies}\label{UnitCellPol}
Although the impedance matrix model provides a simple tool to design structures for desired wave transformations, it is not directly applicable for identifying appropriate topologies of the metasurface unit cells. Here we show how to determine what are the required properties of unit cells which realize ideally refractive metasurfaces. Knowing the polarizabilities of each unit cell, we can identify what polarization response should be generated in unit cells and what inclusions are needed to realize this response.
So-called collective polarizabilities \cite{modeboo} relate the tangential electric and magnetic dipole moments induced in the unit cell to the fields of the incident wave.
Knowing the $Z$-parameters of a metasurface is tantamount to knowing reflection and transmission coefficients. Writing them also in terms of the collective polarizabilities of unit cells, we can find the required polarizabilities which realize the desired response.
For the perfect refractive metasurfaces the collective polarizabilities of unit cells read (see \cite{suppl}
\begin{equation} \widehat{\alpha}_{\rm ee}^{yy}=\frac{S}{j\omega} \frac{\cos{\theta_{\rm i}} \cos{\theta_{\rm t}}}{\eta_1 \cos{\theta_{\rm t}} + \eta_2 \cos{\theta_{\rm i}}}
\left[ 2-\left(\sqrt{\frac{\eta_1 \cos{\theta_{\rm t}}}{\eta_2 \cos{\theta_{\rm i}}}} + \sqrt{\frac{\eta_2 \cos{\theta_{\rm i}}}{\eta_1 \cos{\theta_{\rm t}}}} \right) e^{\displaystyle j\Phi_{\rm t}(z)}
\right], \l{aee_tr}\end{equation}
\begin{equation} \widehat{\alpha}_{\rm mm}^{zz}=\frac{S}{j\omega} \frac{\eta_1 \eta_2}{\eta_1 \cos{\theta_{\rm t}} + \eta_2 \cos{\theta_{\rm i}}}
\left[ 2-\left(\sqrt{\frac{\eta_1 \cos{\theta_{\rm t}}}{\eta_2 \cos{\theta_{\rm i}}}} + \sqrt{\frac{\eta_2 \cos{\theta_{\rm i}}}{\eta_1 \cos{\theta_{\rm t}}}} \right) e^{\displaystyle j\Phi_{\rm t}(z)}
\right] ,\l{aeemm_tr}\end{equation}
\begin{equation} \widehat{\alpha}_{\rm em}^{yz}=-\widehat{\alpha}_{\rm me}^{zy}=\frac{S}{j\omega} \frac{\eta_2 \cos{\theta_{\rm i}} - \eta_1 \cos{\theta_{\rm t}}}{\eta_1 \cos{\theta_{\rm t}} + \eta_2 \cos{\theta_{\rm i}}},\l{aem_tr} \end{equation}
where $S$ is the unit-cell area and $\widehat{\alpha}_{\rm ee}^{yy}$, $\widehat{\alpha}_{\rm mm}^{zz}$, $\widehat{\alpha}_{\rm em}^{yz}$, $\widehat{\alpha}_{\rm me}^{zy}$ are, respectively, electric, magnetic, electromagnetic, and magnetoelectric polarizability components (coupling coefficients). The last two coefficients $\widehat{\alpha}_{\rm em}^{yz}$ and $\widehat{\alpha}_{\rm me}^{zy}$ imply so-called bianisotropic response in the unit cells which models the effect of weak spatial dispersion \cite{serd}. In other words, the incident electric (magnetic) field should induce also magnetic (electric) polarization in the unit cell. Here, $\widehat{\alpha}_{\rm em}^{yz}=-\widehat{\alpha}_{\rm me}^{zy}$, which is a typical characteristic of reciprocal \textit{omega} inclusions \cite{serd}.
As it can be expected, both the electric and magnetic polarizabilities in \r{aee_tr} and \r{aeemm_tr} depend on $z$, and this dependence is the same for both of them. This result reflects the requirement of zero reflection at any point of the metasurface, which demands the balance of the induced electric and magnetic surface currents at any point (Huygens' condition). On the other hand, the omega coupling coefficient in \r{aem_tr} is constant with respect to $z$ and depends only on the impedances and angles. This result reflects the fact that bianisotropic coupling of omega-type is necessary to ensure that the waves incident on both sides of the metasurface see the same surface impedance, so that reciprocal full transmission is realized. Since the impedances of the two waves depend only on the impedances of the media and on the two angles, the coupling coefficient also depends only on these parameters. As expected, we see that when the impedances of the incident and transmitted waves are the same, that is, ${\eta_1\over \cos\theta_{\rm i}}={\eta_2\over \cos\theta_{\rm t}}$, the required coupling coefficient vanishes.
Bianisotropic metasurfaces with the required properties defined by \r{aee_tr}--\r{aem_tr} can be realized as arrays of low-loss particles with the appropriate symmetry. As it was mentioned, for microwave applications, metallic canonical omega particles or double arrays of asymmetric patches can be used. Multilayered topologies were proposed in paper \cite{elefth2}. For optical applications, arrays of properly shaped dielectric particles were introduced as omega-type bianisotropic metasurfaces \cite{Alb1,Alb2}.
It is important to compare the polarizabilities \r{aee_tr}--\r{aem_tr} which are required for realizing perfect refraction with the polarizabilities found in earlier works on wave transformations in the transmitting regime (e.g., \cite{capasso,shalaev,grady,shalaev2,PRX}), where the design approach is based on the geometrical optics model and the ``generalized law of refraction'' \r{viki}. In that theory, the metasurface is assumed to be locally periodical, and the unit cells are designed so that the transmission coefficient has unit amplitude and the desired phase at every point.
These requirements are satisfied if the electric and magnetic polarizabilities read (taking the earlier considered special case of normal incidence and identical media at both sides \cite{PRX})
\begin{equation}
\widehat{\alpha}_{\rm ee}^{yy}= \frac{1}{\eta^2} \widehat{\alpha}_{\rm mm}^{zz}= {S\over j\omega\eta}\left(1-e^{j\Phi_{\rm t}(z)}\right),\l{old_alphas}\end{equation}
and the magnetoelectric coupling coefficient $\widehat{\alpha}_{\rm em}^{yz}$ is zero.
Periodical arrays formed by unit cells having these collective polarizabilities have unit transmissivity and the transmitted waves have the required phases $\Phi_{\rm t}(z)$, but when the cells are assembled into a non-uniform array, the performance becomes non-ideal. In other words, in order to ensure the desired response of the non-uniform metasurface, properties of periodical arrays formed by its unit cells must deviate from the simple geometrical-optics design recipe \r{viki}. This result is consistent with that in \cite{Alu_review}.
We can conclude that in order to ensure perfect refraction, it is not enough to make the metasuface bianisotropic (introducing asymmetry with respect to its two sides).
The electric and magnetic polarizabilities in the exact synthesis [see \r{aee_tr}--\r{aem_tr}] solution are also different as compared to the conventional synthesis solution \r{old_alphas}.
\section{Control of reflection: Perfectly reflecting metasurface}
In the previous case of refractive metasurfaces, there is only one single plane wave at every point of space. In order to be able to synthesize metasurfaces for general field transformations, we need to understand how to control several plane waves which propagate and interfere in the same space. This problem can be solved at an example of a perfectly reflective metasurface, which we consider next.
The geometry of the problem is shown in Fig.~\ref{geom_ref}.
\begin{figure}[h!]
\centering
\epsfig{file=fig4.eps, width=0.6\linewidth}
\caption{Illustration of the desired performance of an ideally reflecting metasurface. TE incidence is assumed and the metasurface is located in the $yz$-plane.}\label{geom_ref}
\end{figure}
The design goal is to fully reflect a plane wave coming from a given direction $\theta_{\rm i}$ into another plane wave propagating in a different and also arbitrary direction $\theta_{\rm r}$. Here, we consider the case when the polarization of the reflected wave is the same as that of the incident wave. Metasurfaces designed for full reflection were called \emph{metamirrors} in \cite{metamirror_AP,PRL}.
In this scenario, the desired field distribution at the surface of the metamirror is the superposition of two plane waves (the incident wave and the reflected wave):
\begin{equation}
\begin{array}{c}
\vspace*{.3cm}\displaystyle
\_E_{t1}=\_E_{\rm i}e^{-jk_1\sin\theta_{\rm i}z}+\_E_{\rm r}e^{-jk_1\sin\theta_{\rm r}z+j\phi_ {\rm r}},\\ \displaystyle
\quad \_n\times \_H_{t1}=\_E_{\rm i}{1\over \eta_1}\cos\theta_{\rm i} e^{-jk_1\sin\theta_{\rm i}z}-\_E_{\rm r}{1\over \eta_1}\cos\theta_{\rm r} e^{-jk_1\sin\theta_{\rm r}z+j\phi_ {\rm r}}.
\end{array}\label{eq:plusmirror}
\end{equation}
Here, $\_E_{t1}$ and $\_H_{t1}$ are the tangential (to the metamirror plane) components of the total electric and magnetic fields at the metamirror surface.
For generality, we assume that the reflected plane wave can have any desired phase shift $\phi_{\rm r}$ with respect to the incident wave. With these notations, we can choose the origin of the $z$-axis so that both $\_E_{\rm i}$ and $\_E_{\rm r}$ will be real-valued vectors.
Similarly to the refractive metasurface, we see that the phase of the reflection coefficient
\begin{equation} \Phi_{\rm r}(z)=-k_1\sin\theta_{\rm r}z+\phi_{\rm r}+
k_1\sin\theta_{\rm i}z\l{mohi}\end{equation}
depends on $z$, except the trivial case of specular reflection ($\theta_{\rm i}=\theta_{\rm r}$). Differentiating, we find the relation between the reflection and incidence angles in terms of the gradient of the reflection coefficient phase:
\begin{equation} k_1(\sin\theta_{\rm i}-\sin\theta_{\rm r})={d\Phi_{\rm r}(z)\over dz}.\l{Gen_refl}\end{equation}
Analogously with the transmitting regime, this result suggests a simple design approach: to realize a fully reflective surface (the amplitude of the reflection coefficient equals unity at each point) but with a linearly varying reflection phase, according to \r{Gen_refl}. Reflecting surfaces with engineered reflection phase are often called high impedance surfaces \cite{HIS} or reflectarrays \cite{encinar}. Such an approach has been used, for example, in \cite{sun,bozh1,bozh2,mosall2,nanotechnology,prx4,veysi,li,PRL} as well as in all known designs of reflectarrays.
However, similarly to refracting metasurfaces, in designing reflecting surfaces this simplistic method also does not allow us to exactly realize the desired performance.
Next, we present the theory of perfect reflecting surfaces and explore various reflection scenarios, with their advantages and limitations.
\subsection{Power flow into the metamirror}\label{sec:power1}
Applying the same method as in analysing metasurfaces for transmission control, we start from considering the power flow into the metamirror structure. The normal component of the Poynting vector at the reflector surface reads
\begin{equation} P_n= {1\over 2}{\rm Re}(\_E_{t1}\times \_H_{t1}^*) . \l{smet} \end{equation}
Substituting the required field distributions (\ref{eq:plusmirror}), we can write the normal component of the Poynting vector as
\begin{equation}
P_n={1\over 2\eta_1} \left[ - E_{\rm i}^2 \, \cos\theta_{\rm i} + E_{\rm i} E_{\rm r} \, ( \cos\theta_{\rm r} - \cos\theta_{\rm i} )\cos\Phi_{\rm r}(z) +E_{\rm r}^2 \, \cos\theta_{\rm r} \right],
\l{zero_power}
\end{equation}
where the reflection phase $\Phi_{\rm r}(z)$ is defined by \r{mohi}.
If this quantity is identically zero at all points along $z$, the metasurface locally (at every point) acts as a lossless reflector. Conventional realizations of non-uniform reflectors belong to this class of locally responding reflectors. Examining the above expression, we see that within this scenario, full transformation of an incident plane wave into a single reflected plane wave of the same polarization is impossible, except the cases of specular or retro-reflection, when $\theta_{\rm r}=\pm \theta_{\rm i}$ (this fact is proven also in \cite{Alu}). Indeed, the expression for the normal component of the Poynting vector \r{zero_power} contains an oscillating term, proportional to $\cos\Phi_{\rm r}(z)$, which can be zero only if $\cos\theta_{\rm r} = \cos\theta_{\rm i}$, that is, $\theta_{\rm r}=\pm \theta_{\rm i}$. Therefore, any local, passive and lossless non-uniform reflecting surface will create modulated reflected waves with spatial dependence of the fields different from the design target (\ref{eq:plusmirror}).
The same expression \r{zero_power} tells also that it is possible to reflect a plane wave into only one plane wave along a specified direction if we allow energy loss in the metasurface. To understand this conclusion, let us look for such \emph{constant} amplitude of the reflected wave $E_{\rm r}$ which ensures that $P_n\le 0$ for all $z$ (negative values of $P_n$ corresponds to flow of power into the surface, where it is absorbed). Obviously, this condition can be satisfied if $E_{\rm r}= E_{\rm i}$, since with this amplitude of the reflected field we have
\begin{equation}
P_n={E_{\rm i}^2\over 2\eta_1} ( \cos\theta_{\rm r} - \cos\theta_{\rm i} )\left[1+\cos\Phi_{\rm r}(z) \right].\l{alu_case}
\end{equation}
Since $1+\cos\Phi_{\rm r}(z)$ is non-negative, $P_n$ is negative or zero at all points of the metasurface $z$ if $\cos\theta_{\rm r} - \cos\theta_{\rm i}\le 0$. This realization scenario was introduced in \cite{Alu}. For instance, if the metamirror is excited by a normally incident plane wave ($\theta_{\rm i}=0$), it is possible to create a single reflected plane wave along any direction, because $\cos\theta_{\rm r}\le 1$ for any $\theta_{\rm r}$. However, as is seen from Eq.~\r{alu_case}, the amount of power which is lost in the metasurface increases with increasing difference between the incidence and reflection angles. In the limit of $\cos\theta_{\rm r} - \cos\theta_{\rm i}\rightarrow -1$, which corresponds to $\theta_{\rm i}\rightarrow 0$ and $\theta_{\rm r}\rightarrow \pi/2$, all incident power is completely absorbed. Figure~\ref{fig_efficomp} (dashed line) shows the efficiency of this scenario as a function of the reflection angle $\theta_{\rm r}$.
\begin{figure}[h]
\centering
\epsfig{file=fig5.eps, width=0.6\linewidth}
\caption{Comparison between the power efficiencies of the passive metamirror which reflects a single plane wave [surface impedance \r{Z11_alu_case}, dashed curve] and the optimized metamirror which minimizes reflections into non-desired directions [surface impedance \r{2waves}, solid curve] at normal incidence.}
\label{fig_efficomp}
\end{figure}
The efficiency $\zeta$ is defined as the ratio of the plane-wave power carried into the desired direction $P_{{\rm r}} = {|\_E_{\rm r}|^2\over {2 \eta_1}} \cos{\theta_{\rm r}}$ to the power of the incident plane wave $P_{{\rm i}} = {|\_E_{\rm i}|^2\over {2 \eta_1}} \cos{\theta_{\rm i}}$. As it is clear from \r{alu_case} and this figure, increasing the reflection angle results in decreasing the efficiency by a factor of $\cos\theta_{\rm r}/\cos\theta_{\rm i}$ (notice that $E_{\rm r}= E_{\rm i}$).
Actually, ideal reflection into a single plane wave without losing any power is possible, but only if we allow periodical flow of power into the metamirror structure and back into space. This conclusion is also evident from formula \r{zero_power}. Indeed, we see that if the amplitude of the reflected plane wave equals
\begin{equation}
E_{\rm r} =\frac{\sqrt{\cos\theta_{\rm i}}}{\sqrt{\cos\theta_{\rm r}}} \, E_{\rm i},
\l{nonmodulation}
\end{equation}
the normal component of the Poynting vector is a periodical function with zero average value:
\begin{equation}
P_n={E_{\rm i}^2\over 2\eta_1}\,\frac{\sqrt{\cos\theta_{\rm i}}}{\sqrt{\cos\theta_{\rm r}}} \,(\cos\theta_{\rm r}-\cos\theta_{\rm i})\cos \Phi_{\rm r}(z). \l{mod_power}\end{equation}
The metasurface performs the desired function perfectly, but the response must be \emph{strongly non-local}: the power which enters the metasurface structure in the areas where $P_n<0$ must be launched back from the areas where $P_n>0$. Alternatively, the perfect reflection can be achieved if the metasurface has active and lossy elements (being overall lossless in the average over the surface area).
We see again that there is no power flow into the metamirror at any point only if $\theta_{\rm r}=\pm \theta_{\rm i}$, in agreement with the previous conclusion.
\subsection{Required surface impedance}\label{sc4}
Following the introduced synthesis approach based on the impedance matrix, we write
the linear relation between the tangential fields at the metamirror surface. Assuming that the metamirror is a boundary and the fields behind it are zero ($\_E_{t2}=0$, $\_H_{t2}=0$), we need only one parameter of the $Z$-matrix \r{11-12}--\r{21-22}, the input impedance $Z_{11}$:
\begin{equation} \_E_{t1}=Z_{11}\, \_n\times \_H_{t1} .\l{BCref}\end{equation}
Substituting the desired field values from \r{plusmirror}, we get the following equation for the unknown input impedance $Z_{11}$:
\begin{equation}
\begin{split} &\_E_{\rm i}\,e^{-jk_1\sin\theta_{\rm i}z} + \_E_{\rm r}\,e^{-jk_1\sin\theta_{\rm r}z+j\phi_{\rm r}} \\
&= Z_{11}\, {1\over \eta_1}\left(\_E_{\rm i}\,\cos\theta_{\rm i} \, e^{-jk_1\sin\theta_{\rm i}z}-\_E_{\rm r}\,\cos\theta_{\rm r} \, e^{-jk_1\sin\theta_{\rm r}z+j\phi_{\rm r}}\right).\l{eq1_ref}
\end{split}
\end{equation}
For the ideally performing non-local metasurface, which produced the reflected wave with the amplitude given by \r{nonmodulation}, the corresponding input impedance reads
\begin{equation}
\begin{split} Z_{11}\, = \frac{\eta_1}{\displaystyle \sqrt{\cos\theta_{\rm i}\cos\theta_{\rm r}}} \, \frac{\sqrt{\cos\theta_{\rm r}} + \sqrt{\cos\theta_{\rm i}}\,e^{j\Phi_{\rm r}(z)} }{\displaystyle \sqrt{\cos\theta_{\rm i}} -\sqrt{\cos\theta_{\rm r}} \, e^{j\Phi_{\rm r}(z)}}.\l{Z_11_const_a}
\end{split}
\end{equation}
We see that the input impedance is a complex number, whose real part is a periodical function of $z$. Figure~\ref{Z11_a} presents the required input impedance for the case when $\theta_{\rm i}=0^\circ$ and $\theta_{\rm r}=70^\circ$.
\begin{figure}[h]
\centering
\includegraphics[width=.6\linewidth]{fig6.eps}
\caption{The required normalized input impedance $Z_{11}/\eta_1$ of the ideal metamirror for $\theta_{\rm i}=0^\circ$, $\theta_{\rm r}=70^\circ$, $\phi_{\rm r}=0^\circ$.}
\label{Z11_a}
\end{figure}
The real part of the input impedance periodically takes positive (loss) and negative (gain) values. The surface \textit{acts} as if it is lossy close to the regions where the reactive impedance is high (close to the regime of a perfect magnetic conductor, PMC) and active in the areas where the reactance is small (close to a perfect electric conductor, PEC). Importantly, this behaviour does not imply that the surface cannot be passive or lossless. We stress that, on the contrary, properly tuned metasurface with strongly non-local response can emulate such a metamirror: The power which passes through the input surface in the ``lossy'' regions is not absorbed but it is re-radiated from the ``active'' regions. Another possibility to realize the ideal performance dictated by impedance \r{Z_11_const_a} could be a metasurface with truly active and lossy elements where only the overall response is lossless.
\begin{figure}[h]
\centering
\includegraphics[width=.6\linewidth]{fig7.eps}
\caption{The required normalized input impedance $Z_{11}/\eta_1$ for passive metamirrors in the case when $\theta_{\rm i}=0^\circ$, $\theta_{\rm r}=70^\circ$, $\phi_{\rm r}=0^\circ$. One period of the metamirror along the $z$-coordinate is shown. The solid, dashed, and dotted lines correspond, respectively, to the lossy metamirrors [Eq.~\r{Z11_alu_case}], the lossless metamirrors creating two reflected plane waves [Eq.~\r{2waves}], and the conventional non-uniform reflectors [Eq.~\r{viki2}].}
\label{Z11_alu}
\end{figure}
As discussed in Section~\ref{sec:power1}, it is possible to eliminate the need to realize active input impedance (which increases the realization complexities), at the expense of losing some part of the incident power in the metamirror. The surface impedance of such a lossy metasurface, which creates a single plane wave in the desired direction, can be found from \r{eq1_ref} upon substitution of the reflected field amplitude $E_{\rm r}=E_{\rm i}$. The result reads
\begin{equation}
Z_{11}=\eta_1{e^{-jk_1\sin\theta_{\rm i}z} + e^{-jk_1\sin\theta_{\rm r}z+j\phi_{\rm r}} \over
\cos\theta_{\rm i} \, e^{-jk_1\sin\theta_{\rm i}z}-\cos\theta_{\rm r} \, e^{-jk_1\sin\theta_{\rm r}z+j\phi_{\rm r}}} \l{Z11_alu_case}.
\end{equation}
An example is plotted as a function of the coordinate in Fig.~\ref{Z11_alu} (the solid lines). As it is seen, the real part of the impedance is always non-negative, corresponding to the absorbed power given by \r{alu_case}.
So far we have demonstrated that a surface having the input impedance \r{Z11_alu_case} produces a single (non-modulated) reflected wave in the desired direction if the power loss in the metamirror is allowed [see Fig.~\ref{fig_efficomp}]. However, depending on the application requirements, it can be preferable to allow some modulation of the reflected wave but reduce the power loss.
In the next section, we present a scenario in which the metamirror is lossless at every point, and at the same time the reflections into non-desired directions are reduced.
\subsection{Optimizing reflections from lossless and local metamirrors}\label{OptDes}
It is possible to optimize the reactance ${\rm Im\mit}${$\{Z_{11}\}$} profile of a lossless metamirror in order to minimize reflections into non-desired directions based on particular optimization criteria. As one example, we notice that there is an interesting lossless design, where all the power which cannot be sent into the desired reflection direction $\theta_{\rm r}$ is reflected into the specular direction. To demonstrate this possibility, we consider the situation when the difference between the incidence angle $\theta_{\rm i}$ and $\theta_{\rm r}$ is large, so that only three propagating plane waves can exist in the Floquet spectrum of the propagating reflected field \cite{vardax}
\begin{equation}
{E}_{\rm r} = \sum_{n=-2}^{0}E_n e^{-jk_1 \left[ (n+1) \sin \theta_{\rm r} - n \sin \theta_{\rm i} \right]z}. \l{Floq}
\end{equation}
For the normal illumination ($\theta_{\rm i}=0$), this corresponds to $\theta_{\rm r}$ larger than $30^\circ$. It is easy to check that a set of three plane waves: the incident wave, the wave reflected into the desired direction ($n =0$), and the parasitic plane wave reflected into the specular direction ($n =-1$) exactly satisfy the boundary condition \r{BCref} with a purely reactive impedance
\begin{equation} Z_{11}(z)=j{\eta_1\over \cos\theta_{\rm r}}
\cot\left[\Phi_{\rm r}(z)/2 \right],\l{2waves}
\end{equation}
if the wave reflected in the desired direction $\theta_{\rm r}$ is given by
\begin{equation} E_{\rm 0}=E_{\rm i}{2\cos\theta_{\rm i}\over \cos\theta_{\rm i}+\cos\theta_{\rm r}},\end{equation}
and the wave reflected into the specular direction $\theta_{\rm i}$ is
\begin{equation} E_{\rm -1}=E_{\rm i}{\cos\theta_{\rm i}-\cos\theta_{\rm r}\over \cos\theta_{\rm i}+\cos\theta_{\rm r}}.\end{equation}
The amplitude of the Floquet harmonic $n=-2$ is equal to zero, and the evanescent part of the spectrum also vanishes. These amplitudes have been found by requiring that the normal component of the Poynting vector is identically zero at the surface. In this case, the metasurface is lossless and exhibits no strong spatial dispersion. Reciprocally, we can conclude that 100\% power reflection in the desired direction can be achieved by illuminating the metasurface by two plane waves at once, properly selecting their relative amplitudes, phases, and propagation directions.
It is interesting that the efficiency of transformation of the incident plane wave into the desired reflected plane wave is much better than for the passive lossy scenario (presented in Section~\ref{sec:power1}) where the parasitic reflections were absent. This conclusion is illustrated in Fig.~\ref{fig_efficomp} by comparing the efficiencies of these two cases.
The conventional approach for designing lossless non-uniform reflectors is based on the ``generalized reflection law'' \r{Gen_refl}, which corresponds to a linear phase variation along the metasurface. In that approach, the measurface is designed so that the \emph{local} reflection coefficient at every point has unit amplitude and the phase as dictated by \r{Gen_refl}. The local reflection coefficient is defined for an infinite uniform array, that is, the input impedance can be found from
\begin{equation}
\_E_{\rm i} + \_E_{\rm r}\,e^{j\Phi_{\rm r}(z)} = Z_{11}\, {1\over \eta_1}\left(\_E_{\rm i}\,\cos\theta_{\rm i}-\_E_{\rm r}\,\cos\theta_{\rm i} \, e^{j\Phi_{\rm r}(z)}\right),
\end{equation}
where $\_E_{\rm r}=\_E_{\rm i}$ and $\Phi_{\rm r}(z)$ is given by \r{mohi}.
The result reads
\begin{equation} Z_{11}=j{\eta_1\over \cos\theta_{\rm i}}\cot\left[\Phi_{\rm r}(z)/2 \right],\l{viki2}
\end{equation}
and an example is plotted in Fig.~\ref{Z11_alu}. One can see that the required surface impedance in the conventional reflectors is different from that of the lossless metamirror described by \r{2waves}. In the conventional reflectors, the reflected wave has a complex structure: Generally, several propagating plane waves in different directions and some evanescent fields localized close to the surface are excited. To study the field structure, one can use numerical simulations or the theoretical technique exploited for the case of refraction in \cite{bloch}.
\subsection{Unit-cell polarizabilities and appropriate topologies}\label{PolRef11}
Making use of the boundary conditions on the reflecting metasurface which tell that the tangential electric and magnetic fields are equal, correspondingly, to the surface magnetic and electric current densities, we can find relations between the surface impedance $Z_{11}$ and the collective polarizabilities of unit cells of the metamirror (see \cite{suppl}):
\begin{equation} {\eta_1\over \cos{\theta_{\rm i}}}\,{{\cos{\theta_{\rm i}}+\cos{\theta_{\rm r}}}\over {Z_{11}\cos{\theta_{\rm r}}+\eta_1}} = \frac{j\omega}{S} \left({\eta_1\over \cos{\theta_{\rm i}}}\,\widehat{\alpha}_{\rm ee}^{yy}+\widehat{\alpha}_{\rm em}^{yz} \right), \l{pol1}\end{equation}
\begin{equation} Z_{11}\,{{\cos{\theta_{\rm i}}+\cos{\theta_{\rm r}}}\over {Z_{11}\cos{\theta_{\rm r}}+\eta_1}} = \frac{j\omega}{S} \left({\cos{\theta_{\rm i}}\over \eta_1}\,\widehat{\alpha}_{\rm mm}^{zz}-\widehat{\alpha}_{\rm em}^{yz} \right). \l{pol2}\end{equation}
Here $S$ is the unit-cell area.
Obviously, these equations have infinitely many solutions for polarizabilities which realize the desired response. The metasurface can be either bianisotropic (omega coupling) or it can be a non-bianisotropic pair of electric and magnetic current sheets. For the non-bianisotropic realization we set
\begin{equation} \widehat{\alpha}_{\rm em}^{yz}=\widehat{\alpha}_{\rm me}^{zy}=0, \end{equation}
and find the unique solution
\begin{equation} \widehat{\alpha}_{\rm ee}^{yy}=\frac{S}{j\omega}\,{{\cos{\theta_{\rm i}}+\cos{\theta_{\rm r}}}\over {Z_{11}\cos{\theta_{\rm r}}+\eta_1}}, \end{equation}
\begin{equation} \widehat{\alpha}_{\rm mm}^{zz}=\frac{S}{j\omega}\, \frac{Z_{11}\,\eta_1}{\cos{\theta_{\rm i}}}\,{{\cos{\theta_{\rm i}}+\cos{\theta_{\rm r}}}\over {Z_{11}\cos{\theta_{\rm r}}+\eta_1}}.\end{equation}
We see that in the design of fully reflective metasurfaces, weak spatial dispersion effects are necessary at least in form of the artificial magnetism. If we demand that both magnetic polarizability and the bianisotropy coefficient are zero, the above equations have no solutions. The use of bianisotropy offers additional design flexibilities.
Knowing the collective polarizabilities required for the desired performance we can immediately see what are the appropriate topologies of unit cells. Since we need both electric and magnetic polarizations, the physical thickness of the reflecting layer must be different from zero, to allow formation of tangential magnetic moments in unit cells. For example, it is not possible to realize the desired performance by any patterning of a single, infinitesimally thin sheet of a perfect conductor. The non-bianisotropic realization scenario suggests the use of a single array of small particles which are polarizable both electrically and magnetically, such as small metal spirals as in \cite{PRX}. A typical realization based on the bianisotropic route is a high-impedance surface with a PEC ground plane (such as ``mushroom layers'' \cite{HIS}). An important advantage in using bianisotropic effects is the relaxed requirement on the strength of the magnetic response. Especially for optical applications, it is easier to realize strong bianisotropy (which is a first-order dispersion effect) as compared with the artificial magnetism (which is a weaker, second-order effect) \cite{serd}.
\subsection{Ideal metamirrors}\label{sec:power3}
We have seen that all local lossless non-uniform reflectors modulate the reflected waves, which reduces the power efficiency in the desired direction. The operation of conventional planar reflectors (such as high impedance surfaces \cite{HIS}, reflective diffraction gratings \cite{gratings}, and reflectarrays \cite{encinar}) are similar in this respect. Next we discuss the potentials of ideal metamirrors based on non-local and non-reciprocal surfaces.
As shown above, it is possible to synthesise an \textit{overall} lossless metamirror which would create an {\it unmodulated} reflected wave into any desired direction, satisfying the requirement
(\ref{eq:plusmirror}) exactly, with a constant value of the reflected plane wave amplitude $E_{\rm r}$. This goal can be achieved if we require that the normal component of the Poynting vector on the metasurface is zero only in the average over the metamirror period, and not necessarily is equal to zero at every point. In this case, the amplitude of the plane wave reflected into the desired direction is given by \r{nonmodulation}, and the normal component of the Poynting vector oscillates, according to Eq.~\r{mod_power}.
Realization of such metamirrors requires absorption of power in some areas of the surface and generation of power in some other areas or, alternatively, power channelling from one area to the other. Conceptually, this scenario of balanced loss and gain can be realized using the same two approaches which were found in the analysis of perfectly refractive metasurfaces: teleportation metasurface (Section~\ref{sec:teleportation}) and transmitarrays (Section~\ref{sec:transmitarray}).
In the former approach, one can envisage a realization in form of an array of small receiving antennas loaded by positive resistors in the areas where the energy should be partially absorbed, and by negative resistors where the energy should be launched back into space. This arrangement is similar to the teleportation metasurface described in Section~\ref{sec:teleportation}, where such arrays were positioned at the two opposite sides of a metal screen.
Alternatively, one can envisage a similar array of antennas, where the antennas of the absorbing areas are connected by cables to the antennas of the active areas. Thus, the power received at the absorbing areas is re-radiated by the active areas. It is important to note that both these devices should be non-reciprocal, as the ``active'' antennas should radiate power but not receive it back from space. Actual realization of both these concepts is a challenging task. As to the teleportation approach, one needs non-reciprocal antennas, which can be in principle realized using non-reciprocal materials like magnetized ferrites or using active components. There is also an interesting possibility to use parametric circuits for the same purpose \cite{non_rec_ant}. The non-reciprocal transmitarray approach in principle can be realized also in reflecting metasurfaces, using non-reciprocal circuits inside the metasurface, but it appears that the use of spatial modulation of the surface impedance by external forces (using unit cells equipped with varactors, for example), is more promising. Conceptually, the desired performance can be achieved by modulating (for example) varactors in all unit cells with the same amplitude but with different phases. Controlling the spatial distribution of the modulation phase, one can possibly realize parameteric amplification or absorption according to the design specifications. Initial work on space-time modulated metasurfaces \cite{space-time,s-t-Caloz} produced interesting and promising results, and we expect that developing this route may lead to realizations of theoretically perfectly operating lossless non-uniform metasurfaces.
\section{Perfectly reflecting polarizers}\label{sc5}
In the previous section, we considered metamirrors which reflect an incident plane wave into a desired direction. However, we encountered either active-lossy realizations of the metamirror or lossless reflection of modulated waves. In this section, we introduce a new solution for a lossless metamirror which ideally reflects the incident wave into the desired direction without any modulations. Since the main reason for modulations of reflected waves is interference between the incident and reflected fields, we construct a metamirror which reflects waves with the polarization orthogonal to that of the incident wave. As a simple canonical example, we consider the transformation of a transverse electric (TE) wave with the amplitude $E_{\rm i}^{y}$ into a transverse magnetic (TM) wave with the amplitude $E_{\rm r}^{z}=E_{\rm r} \cos{\theta_{\rm r}}$, propagating in the desired direction. Figure~\ref{geom3} shows the problem configuration.
\begin{figure}[h!]
\centering
\epsfig{file=fig8.eps, width=0.6\linewidth}
\caption{Illustration of the desired performance of an ideal metamirror which perfectly transforms a TE incident wave into a TM reflected wave.}\label{geom3}
\end{figure}
It is clear that in this case there is no interference between the incident and reflected waves. The desired tangential electric and magnetic fields at the metamirror surface read \begin{equation} \_E_{t1}=\hat{\_y} \, {E_{\rm i}^{y}}e^{-jk_1\sin\theta_{\rm i}z}+\hat{\_z} \, {E_{\rm r}^{z}}e^{-jk_1\sin\theta_{\rm r}z+\phi_{\rm r}} , \l{plusE} \end{equation} and \begin{equation} \_n\times \_H_{t1}= \hat{\_y} \, {\cos\theta_{\rm i}\over \eta_1} \, {E_{\rm i}^{y}}e^{-jk_1\sin\theta_{\rm i}z}-\hat{\_z} \, {1\over \eta_1 \cos\theta_{\rm r}} \,{E_{\rm r}^{z}}e^{-jk_1\sin\theta_{\rm r}z+\phi_{\rm r}}, \l{plusH} \end{equation}
respectively. Considering the metamirror as a boundary, the impedance relation between the tangential electric and magnetic fields \r{11-12}--\r{21-22} in this case reads~\cite{modeboo}
\begin{equation} \_E_{t1}=\overline{\overline{Z}}_{11} \cdot \_n\times \_H_{t1}, \l{Z_tens} \end{equation}
where the impedance has the matrix form
\begin{equation}
\overline{\overline{Z}}_{11} = \begin{bmatrix}
Z^{yy}_{11} & Z^{yz}_{11} \\
& \\
Z^{zy}_{11} & Z^{zz}_{11}
\end{bmatrix}.
\end{equation}
Notice, in contrast with the previous case, we should consider the full-rank impedance dyadics $\overline{\overline{Z}}$ in order to account for any possible polarization transformation. Substituting \r{plusE} and \r{plusH} into \r{Z_tens}, we obtain the following matrix equation \begin{equation}
\begin{bmatrix}
1\\
\\
\displaystyle R^{zy}e^{j\Phi_{\rm r}}
\end{bmatrix}=
\begin{bmatrix}
Z^{yy}_{11} & Z^{yz}_{11} \\
& \\
Z^{zy}_{11} & Z^{zz}_{11}
\end{bmatrix}
\begin{bmatrix}
\displaystyle {\cos\theta_{\rm i}\over \eta_1} \\
\\
\displaystyle {-1\over \eta_1 \cos\theta_{\rm r}} \,R^{zy}e^{j\Phi_{\rm r}}
\end{bmatrix},
\end{equation}
where $R^{zy}={E_{\rm r}^{z}}/{E_{\rm i}^{y}}$ and $\Phi_{\rm r}$ is defined in \r{mohi}.
The solution of the above equation for the lossless case (i.e., ${\rm Re\mit}\{ \overline{\overline{Z}}_{11} \} = 0$) is unique and reads as
\begin{equation}
\begin{bmatrix}
Z^{yy}_{11} & Z^{yz}_{11} \\
& \\
Z^{zy}_{11} & Z^{zz}_{11}
\end{bmatrix}
=j
\begin{bmatrix}
\displaystyle {\eta_1\over{\cos{\theta_{\rm i}}}}\cot{\Phi_{\rm r}} & \displaystyle {\eta_1\cos{\theta_{\rm r}}\over{R^{zy}}}{1\over{\sin{\Phi_{\rm r}}}} \\
& \\
\displaystyle {\eta_1\over{\cos{\theta_{\rm i}}}}R^{zy}{1\over{\sin{\Phi_{\rm r}}}} & \displaystyle {\eta_1{\cos{\theta_{\rm r}}}}\cot{\Phi_{\rm r}}
\end{bmatrix}.\l{Z_pol_rot}\end{equation}
Here, $Z^{yy}_{11}$ is the metamirror input impedance which is responsible for suppressing unwanted reflections in the specular direction. The proper values of the cross-components $Z^{yz}_{11}$ and $Z^{zy}_{11}$ ensure the polarization rotation, and, finally, $Z^{zz}_{11}$ is responsible for reflection with the orthogonal polarization in the desired direction.
Next, we apply the condition for power conservation (we demand that the normal component of the Poynting vector identically equals zero at the metasurface plane at each point to ensure local response) to find the required reflection coefficient $R^{zy}$. This condition reads
\begin{equation}
-\left({E_{\rm i}^{y}}\right)^2 \, \cos\theta_{\rm i} \, {1\over 2 \eta_1}+\left({E_{\rm r}^{z}}\right)^2 \, {1\over 2\eta_1\cos\theta_{\rm r}} = 0 ,
\l{zero_powerXpol}
\end{equation}
which defines the reflection coefficient for the perfect reflection regime:
\begin{equation} R^{zy}= \sqrt{\cos\theta_{\rm i}\cos\theta_{\rm r}}.\l{Rzy}\end{equation}
As it is clear from \r{Z_pol_rot}, realization of this scenario is possible with purely lossless metasurface elements. Moreover, since the reflected field does not interfere with the incident one, there is no field modulation. Therefore, the proposed metamirror provides an ideal and single reflecting wave.
\subsection{Unit-cell polarizabilities and appropriate topologies}\label{PolRef11_rot}
Following the procedure outlined in Sections~\ref{PolRef11}, we can find the relations for collective polarizabilities of unit cells of the proposed metamirror in the case of perfectly reflecting polarizers (see \cite{suppl}
for details):
\begin{equation} 1 = \frac{j\omega}{S} \left({\cos{\theta_{\rm i}}\over \eta_1}\,\widehat{\alpha}_{\rm mm}^{zz}+\widehat{\alpha}_{\rm me}^{zy} \right), \l{polrot1}\end{equation}
\begin{equation} R^{zy}e^{j\Phi_{\rm r}} = -\frac{j\omega}{S} \left({\cos{\theta_{\rm i}}\over \eta_1}\,\widehat{\alpha}_{\rm mm}^{yz}+\widehat{\alpha}_{\rm me}^{yy} \right), \l{polrot2}\end{equation}
\begin{equation} {\cos{\theta_{\rm i}}\over \eta_1} = \frac{j\omega}{S} \left({\cos{\theta_{\rm i}}\over \eta_1}\,\widehat{\alpha}_{\rm em}^{yz}+\widehat{\alpha}_{\rm ee}^{yy} \right), \l{polrot3}\end{equation}
\begin{equation} \frac{R^{zy}e^{j\Phi_{\rm r}}}{\eta_1\cos{\theta_{\rm r}}} =- \frac{j\omega}{S} \left({\cos{\theta_{\rm i}}\over \eta_1}\,\widehat{\alpha}_{\rm em}^{zz}+\widehat{\alpha}_{\rm ee}^{zy} \right). \l{polrot4}\end{equation}
Obviously, these equations have infinitely many solutions for polarizabilities which realize the desired response. Even restricting ourselves by reciprocal realizations, the metamirror can be either bianisotropic (both omega and chiral couplings) or it can be non-bianisotropic with anisotropic electric and magnetic responses. Here we show two simple design solutions. In the first design, the metamirror is modeled by anisotropic electric and magnetic polarizabilities. The non-zero polarizabilities read:
\begin{equation}
\widehat{\alpha}_{\rm mm}^{zz}= \frac{S}{j\omega} \, \frac{\eta_1}{\cos{\theta_{\rm i}}},
\l{ammzz}\end{equation}
\begin{equation}
\widehat{\alpha}_{\rm mm}^{yz}= - \frac{S}{j\omega} \, \frac{\eta_1}{\cos{\theta_{\rm i}}}\, R^{zy}e^{j\Phi_{\rm r}},
\l{ammyz}\end{equation}
\begin{equation}
\widehat{\alpha}_{\rm ee}^{yy}= \frac{S}{j\omega} \, \frac{\cos{\theta_{\rm i}}}{\eta_1},
\l{aeeyy}\end{equation}
\begin{equation}
\widehat{\alpha}_{\rm ee}^{zy}= - \frac{S}{j\omega} \, \frac{1}{\eta_1 \cos{\theta_{\rm r}}}\, R^{zy}e^{j\Phi_{\rm r}}.
\l{aeezy}\end{equation}
In this design, the bianisotropic properties are excluded, that is, $\widehat{\alpha}_{\rm me}^{zy}= \widehat{\alpha}_{\rm me}^{yy} =\widehat{\alpha}_{\rm em}^{yz} =\widehat{\alpha}_{\rm em}^{zz} =0$. Notice that there is no limitations on the selection of the polarizability components $\widehat{\alpha}_{\rm ee}^{yz}$, $\widehat{\alpha}_{\rm ee}^{zz}$, $\widehat{\alpha}_{\rm mm}^{zy}$, and $\widehat{\alpha}_{\rm mm}^{yy}$ (they can be chosen from considerations of reciprocity, for example).
Alternatively, another simple solution of system \r{polrot1}--\r{polrot4} can be found by suppressing the cross polarizability components (i.e., $\widehat{\alpha}_{\rm me}^{zy}=\widehat{\alpha}_{\rm mm}^{yz}=\widehat{\alpha}_{\rm em}^{yz}=\widehat{\alpha}_{\rm ee}^{zy}=0$ ). This implies that the metamirror possesses chiral bianisotropic response:
\begin{equation}
\widehat{\alpha}_{\rm mm}^{zz}= \frac{S}{j\omega} \, \frac{\eta_1}{\cos{\theta_{\rm i}}},
\l{ammzz2}\end{equation}
\begin{equation}
\widehat{\alpha}_{\rm me}^{yy}= -\frac{S}{j\omega} \, R^{zy}e^{j\Phi_{\rm r}},
\l{ameyy}\end{equation}
\begin{equation}
\widehat{\alpha}_{\rm ee}^{yy}= \frac{S}{j\omega} \, \frac{\cos{\theta_{\rm i}}}{\eta_1},
\l{aeeyy2}\end{equation}
\begin{equation}
\widehat{\alpha}_{\rm em}^{zz}= - \frac{S}{j\omega} \, \frac{e^{j\Phi_{\rm r}}}{R^{zy}}\, .
\l{aemzz}\end{equation}
while there is no limitation on $\widehat{\alpha}_{\rm ee}^{zz}$, $\widehat{\alpha}_{\rm em}^{yy}$, $\widehat{\alpha}_{\rm me}^{zz}$ and $\widehat{\alpha}_{\rm mm}^{yy}$ (they can be chosen from considerations of reciprocity). It can be shown that if we apply the reciprocity condition ($\widehat{\alpha}_{\rm em}^{yy}=-\widehat{\alpha}_{\rm me}^{yy}$ and $\widehat{\alpha}_{\rm me}^{zz}=-\widehat{\alpha}_{\rm em}^{zz}$ \cite{serd}) and choose
\begin{equation}
\widehat{\alpha}_{\rm ee}^{zz}= \frac{S}{j\omega} \, \frac{1}{\eta_1 \cos{\theta_{\rm i}}},
\l{ammzz222}\end{equation}
\begin{equation}
\widehat{\alpha}_{\rm mm}^{yy}= \frac{S}{j\omega} \, \eta_1 \cos{\theta_{\rm i}},
\l{aemzz222}\end{equation}
then the same metamirror dually operates both for TE and TM polarized incident waves.
One can note a similarity of the conditions on the polarizabilities \r{ammzz2}--\r{aemzz222} with those used earlier for realizing polarization transformers \cite{Niemi} and absorbers \cite{absor11,PRX}. Here we see that the amplitudes of the polarizabilities should be balanced (as shown in \cite{Niemi} for the normal incidence), and the ideal reflector operation is ensured by proper adjustments of the chirality parameter phase.
These solutions are only two possibilities, selected for their simplicity. Other solutions are possible considering \r{polrot1}--\r{polrot4}.
\section{Conclusions and discussions}
In this paper we have introduced a general approach to the synthesis of metasurfaces for arbitrary manipulations of plane waves. We have explained the main ideas of the method on two canonical examples: A metasurface which perfectly refracts plane waves incident at an arbitrary angle $\theta_{\rm i}$ into plane waves propagating in an arbitrary direction defined by the angle $\theta_{\rm t}$, and a metasurface which fully reflects a given plane wave into an arbitrary direction $\theta_{\rm r}$. The general synthesis approach shows a possibility for alternative physical realizations, and we have discussed different possible device realizations: self-oscillating teleportation metasurfaces, non-local metasurfaces, and metasurfaces formed by only lossless components. The crucial role of omega-type bianisotropy in the design of lossless-component realizations of perfectly refractive surfaces has been revealed.
The conventional approach to realization of refractive and reflecting metasurfaces as well as both transmitarray and reflectarray antennas is based on requiring full power transmission or reflection at each point of the surface and providing complete phase control over the transmitted and reflected waves. We have clarified the role of modifications in the required phase gradient for conventional planar refractive/reflective structures in gaining higher efficiencies. Moreover, we have revealed fundamental limitations of this classical technique and showed how the ideal performance can be realized. For full control over transmission, weak spatial dispersion in form of bianisotropic coupling is necessary, while ideal lossless reflectarray operation calls for the use of structures with a strongly non-local response to the incident fields or structures that transform polarization of reflected waves.
We think that the reason why the role of metasurface bianisotropy in controlling refraction has not been appreciated earlier is that in this field transformation the wave polarization should not change, and it appears natural to expect that bianisotropic effects, such as chirality, are not needed. However, as we have shown here, omega coupling effects, which do not change polarization, are crucial in engineering perfectly matched lossless refractive metasurfaces.
In contrast to perfectly refracting metasurfaces, creation of perfectly reflecting surfaces requires careful control over the interference of the incident and reflected waves. We have shown that ideal transformation of an incident plane wave into a reflected plane wave propagating at an angle different from what is dictated by the usual reflection law requires either active structures or passive lossless non-local metasurfaces. We have discussed the structure of reflected fields and proposed an optimal compromise realization using local and passive metasurfaces.
In the last part of the paper we have shown that the requirement of strong spatial dispersion or active inclusions for realization of perfect metamirrors can be lifted if the polarization of the reflected wave is orthogonal to that of the incident field. In this case there is no interference between the incident and reflected wave, and perfect reflection can be realized using only weak spatial dispersion effects (artificial magnetism and chirality), similarly to ideally refractive metasurfaces.
Since any exciting fields can be expressed in form of a plane-wave expansion, the developed approach can be generalized to metasurfaces for the most general field transformations. We hope that understanding of the physical requirements for perfect metasurface operation in both transmission and reflection regime as well as the developed synthesis method will open a way for design and realization of ultimately thin composite sheets for a broad range of applications, such as lenses, antennas, sensors, etc.
We would like to note that during the review process of this paper a related preprint \cite{last} has been published, which describes a conceptual realization of perfectly reflecting lossless metasurfaces in form of a set of three parallel reactive sheets. This structure exhibits the required non-local properties (``channeling'' energy in the transverse direction), according to the theory presented here.
\section*{Acknowledgment}
This work was supported in part by the Academy of Finland (project 287894).
|
\section*{Introduction}
Pressure ($P$) or temperature ($T$) induced modifications in crystal structures and associated effects on the lattice dynamics are commonly observed and reasonably well understood. On the contrary, transformations in amorphous systems between distinct aggregates having different local structure and density are more elusive. These polyamorphic (PA) transitions are often difficult to observe, since hampered by several concomitant factors. For instance, when the density is the order parameter, extreme thermodynamic conditions are required
to significantly alter this variable due to the low compressibility of amorphous, non-gaseous, systems. Furthermore, PA phenomena often happen in metastable thermodynamic regions, where they are overshadowed by competing effects, such as glass transition or crystal nucleation.
On a general ground, the best candidates to observe PA transitions are systems with an intrinsically open, often tetrahedral, local structure. In fact, the large free volume available in tetrahedral arrangements can in principle allow structural modifications even at moderate thermodynamic conditions. This was demonstrated to be the case in water \cite{Mishima}, liquid silicon \cite{Silicon}, germanium\cite{Germanium}, and phosphorus \cite{Phosphorus}, as well as in amorphous SiO$_2$ \cite{Silica} and GeO$_2$ \cite{Germania} (see Ref. [\onlinecite{Brazhkin}] for a review on the topic). In spite of a thorough experimental scrutiny, some general aspects of PA transitions are still obscure,
including the possible influence on the propagation of collective excitations. This can be particularly relevant at mesoscopic ($\sim$ nm) length-scales, where the dynamics is known to be strongly coupled with local atomic arrangements.
Inelastic neutron (INS) and x-ray (IXS) scattering are two classic experimental techniques commonly used to probe atomic and lattice motions; however severe technical difficulties hinder the observation of PA transitions. A major one relates to the fact that modifications in the local order involving, e.g. the coordination number, usually disappear when the sample is recovered to ambient conditions \cite{nota}. This imposes \emph{in situ} high-pressure experiments on very small samples for direct observation of PA transitions. This, in most practical situations, rules out the possibility of using INS and often causes problems of spectral background in IXS measurements. Although water \cite{water_IXS} and silica \cite{silica_IXS} are the two polyamorphic materials that have been most extensively investigated by inelastic spectroscopies, no signature of PA transitions has been reported in the THz spectrum in either case, mainly owing to two different reasons: in silica PA phenomena happen at pressures still prohibitively high for scattering measurements, while in water the PA transition is expected to take place in a deeply supercooled region, representing a sort of \emph{no man's land} in the thermodynamic plane \cite{Poole}.
Compared to SiO$_2$, its structural analogous GeO$_2$ has proven to be a better candidate for IXS investigations of PA phenomena due to both the larger tetrahedral cell, which shifts the onset of PA transitions to lower $P$'s, and the higher electronic number -- and consequently shorter x-ray absorption length -- which substantially enhances the IXS signal from the small-sized sample suited for the use of Diamond Anvil Cells (DAC) \cite{BenAnt_PRB}. Accordingly, we have studied the pressure-dependent spectrum of density fluctuation, $S(Q,\omega)$, of vitreous (v-)GeO$_2$ at ambient temperature by \emph{in situ} IXS measurements from ambient $P$ up to 26 GPa (see Methods for further details). This $P$ range has been chosen to well track the $P$-dependence below and above 9 GPa, pressure around which a sudden jump of the bond distance is reported and commonly ascribed to a transition from a tetrahedral to an octahedral local structure \cite{Germania,Smith}, or, in other terms, from an $\alpha$-quartz-like to a rutile-like local lattice organization. In a more recent x-ray diffraction work \cite{Drewitt} important structural changes have been observed to continuously occur for pressure spanning the 5-8.6 $GPa$ range, as later confirmed by oxygen K-edge IXS measurements \cite{Lelong}. Furthermore, previous studies based on classical molecular dynamics \cite{Peralta} predicted a main structural change from tetrahedra to octahedra arrangement at 3-7 GPa, slightly lower than observed here and reported in previous works.
In addition to diffraction and absorption measurements \cite{Micolaut}, signatures of a PA transition in v-GeO$_2$ have been sought for by investigating the vibrational behavior by Raman and infrared techniques \cite{Micolaut,Galeener}, which provided evidences of possible dynamic counterparts of the aforementioned PA transition. Combining experimental results with first-principle density functional theory (DFT) calculations, we aim at detecting and explaining the signature of the PA crossover in the THZ spectrum of v-Ge$O_2$.
\section*{Results}
Typical IXS spectra measured at ambient pressure and at $P =$ 26 GPa, i.e. respectively well below and above the PA transition, are shown in Fig. \ref{Fig1} along with the experimental resolution function. We recall here that PA transition of GeO$_2$ is here identified with the large transformation of the average Ge-O bond distance, $d_{O-Ge}$ in the $ 5 \lesssim P \lesssim 9 \,GPa$ range(see Ref. \cite{Germania} and \cite{Smith}). Although, as mentioned, such a crossover was also either observed as a sharp, yet continuous, P-increase \cite{Drewitt,Lelong} or to occur at slightly lower P values, it is unanimously found that it takes place for P $<$ 10 GPa.
A clear excess of scattering intensity with respect to the resolution profile can be observed in all spectra, indicating the presence of an appreciable contribution of collective excitations to the experimental spectra in all the probed ($Q$,$P$)-values. Such inelastic spectral wings turn into a ``double shoulder'' lineshape at higher $Q$'s for both $P$ values, hence suggesting the presence of two distinct collective modes. In order to determine the $Q$-dispersion relations of such modes and gain insights into their correlations to structural changes within the probed $P$-range, we performed a best-fit line-shape analysis to determine the characteristic frequencies of the high ($\Omega_\mathrm{HF}$) and low frequency ($\Omega_\mathrm{LF}$) excitations (see Methods for further details).
The dispersion curves obtained as the best-fit values of the $\Omega_\mathrm{LH}$ and $\Omega_\mathrm{HF}$ parameters in Eq. \ref{DHO_2} are displayed in Fig. \ref{Fig2}.
It may be noticed that at the lowest $Q$s some $\Omega_\mathrm{LH}$ are not reported in the plot; this owes to the corresponding vanishing intensity of the low frequency peak in the spectrum.\\
Even more striking is the transformation of the $P$-evolution of the generalized sound velocities $c_s$ (Fig. 3) extracted from the low Q ($\leq 10 nm^{-1}$) slope of the dispersions curves in Fig. \ref{Fig2}. The fact that IXS probes sound propagation at THZ frequencies and $nm$ distances makes IXS measurements more sensitive to local molecular arrangements and interactions than ultrasound or Brillouin light scattering (BLS) techniques. Derived velocities can thus differ from ultrasonic and hypersonic determinations (see for instance discussion in Ref. \cite{BenAnt_PRB}), bringing complementary information. To further stress the distinctive elastic behavior of the low-pressure and high-pressure polyamorph, the inset of Fig. \ref{Fig3} shows the value of the longitudinal modulus $ M = \rho c_s$ - with $\rho$ being the mass density, here derived from from Ref. \cite{Smith}.
Figs. \ref{Fig4} and \ref{Fig5} compare, respectively, the lowest $P$ ($P$ = 0 GPa) and highest $P$ ($P$ = 26 GPa) dispersions shown in Fig. \ref{Fig2} with the phonon dispersions computed for crystalline $\alpha$-quartz GeO$_2$ and rutile GeO$_2$ at the corresponding $P$s along high symmetry directions; here the thickness of the dispersion curves is proportional to the weight of the corresponding contributions to the $S(Q,\omega)$ (see Methods for further details). The green lines in Figs. \ref{Fig4} and \ref{Fig5} denote the $(\omega,Q)$-range dominated by the phonon modes of diamond. Within this region the very intense scattering contribution from the diamond anvils hamper the detection of the signal from the sample. Right panels in Figs. \ref{Fig4} and \ref{Fig5} show the computed total vibrational density of state (v-DOS). The v-DOS measured by Raman scattering at $P$ = 0 GPa \cite{Deshamps} and 32 GPa \cite{Durben} are also included for comparison \cite{note}
in Figs. \ref{Fig4} and \ref{Fig5}.
\section*{Discussion}
The quality of the data analysis can be readily judged from the overall good agreement between measured and best-fit line-shapes in Fig. \ref{Fig1}, where the spectral contributions of the low and high frequency modes to the total scattering intensity are also shown. By comparing inelastic modes in the spectra measured at $P$ = 0 GPa and $P$ = 26 GPa one immediately observes that both inelastic shift and relative intensity of the high frequency mode increase dramatically with pressure.
A substantial difference between low- and high-$P$ sound dispersions readily emerges from Fig. \ref{Fig2}: $\Omega_\mathrm{LF}$ and $\Omega_\mathrm{HF}$ at low $P$s (left panel) span over lower energy values and exhibit a moderate $Q$-dependence. In contrast, the high-$P$ dispersions (right panel) are featured by a strongly $Q$-dependent high-$\omega$ phonon branch and by a mildly $Q$-dependent low-$\omega$ branch. Such features become even more striking when considering that $P$ effects are relatively moderate within each subset of data. This result is in good agreement/compatible with the expected transition value of 9 GPa.\\
Further transformations of the acoustic properties of the sample across the PA crossover clearly emerge in Fig. \ref{Fig3} through the cusp-like $P$ dependence of $c_s$.\\
The drastic change is emphasized by the different slopes of the straight lines best fitting all reported $c_s$ data either below or above the crossover (dashed lines). These lines intercept at a P value ( $\approx$ 8.7 GPa) close to the known PA crossover pressure. The lower high $P$ slope clearly indicates a higher resistance of the sample to $P$-induced modifications of elastic properties. In a similar manner, in recent BLS
works \cite{Murakami,Huang} changes in the slope of the P-evolution of sound velocity and/or elastic moduli were interpreted as signatures of PA phenomena; more specifically, they have been correlated to changes in the coordination number. Here
the cusp-like behavior exhibited by $c_s$ data seems more pronounced. This is the likely consequence of the mentioned rough matching between probed distances and first neighboring molecules'.
As the crystalline $\alpha$-quartz and rutile structures are known to approximate first neighbor Ge-O bonding arrangements of v-GeO$_2$ for $P$ below and above the PA crossover \cite{Germania}, we use DFT results as an interpretative guidance for the observed $P$-evolution of the collective dynamics measured on glasses. Bearing in mind the different sample nature (crystalline vs glassy), experimental and computational results shown in Figs. \ref{Fig4} and \ref{Fig5} are overall consistent. In particular, it is worth noticing that both computed dispersions and v-DOS exhibit distinctive features, markedly different between the $\alpha$-quartz and rutile phases. Indeed, dispersion curves of $\alpha$-quartz clearly span over lower energies, and optical phonon modes exhibit a classical relatively mild $Q$ dependence. Accordingly, the v-DOS of this structure is characterized by the expected parabolic low-frequency trend, followed by van Hove singularities of comparable intensity and, finally, by a clear phonon gap in the 45-55 meV interval (see inset in Fig. \ref{Fig3}).
In contrast, for the the rutile lattice arrangement, the main contribution to $S(Q,\omega)$ comes from highly dispersive high-$\omega$ phonon branches. The v-DOS is dominated by a large van Hove singularity at about 45 meV and presents no evidence of phonon gaps below 60 meV. As a consequence, within the reported 0-60 meV range, the v-DOS of rutile spans comparatively higher $\omega$'s than that of $\alpha$-quartz. In view of these qualitative differences, the overall agreement of experimental data collected at pressures below 9 GPa with calculations for the $\alpha$-quartz structure (Fig. \ref{Fig4}) and of the experimental data collected at pressure above 9 GPa with the calculations for the rutile crystal structure (Fig. \ref{Fig5}) provide strong indications that the PA quartz-like to rutile-like transition in the local structure of v-GeO$_2$ \cite{Germania} has a well-defined counterpart in the phonon dispersion behavior.
A closer inspection of Fig. \ref{Fig5} reveals few additional interesting aspects. The IXS measurements unmistakably show that, at high $P$, the $S(Q,\omega)$ of v-GeO$_2$ is dominated by two modes with very different energy, one highly $Q$-dispersive, and a second weakly $Q$-dispersive. The presence of the latter was particularly evident in the Spring-8 data (P= 13 GPa) most likely because of the better contrast in the low-frequency spectral range (see Methods). The comparison with the computed dispersion curves of the rutile crystal allows us to tentatively associate these two branches to longitudinal (LA) and transverse acoustic (TA) modes, respectively.
In order to facilitate the comparison between $\Omega_{LF}$ values and the TA modes in the crystal, in Fig. \ref{Fig4} the lowest frequency TA branches of $\alpha$-quartz are also indicated as dashed lines in the three crystalline paths considered. However, the spectral contribution of these branches was found negligible in the crystal.
Whereas for the crystal phase TA contributions to $S(Q,\omega)$ are forbidden within the first Brillouin zone, they may become sizable in the glassy phase. This directly relates to the absence of long-rage translational symmetries, or equivalently well-defined Brillouin zones. "Pure" symmetric TA modes can be "contaminated" and acquire a mixed longitudinal and transverse character. Such mode-mixing is often referred to as the longitudinal-transverse (L-T) coupling \cite{Sampoli} and is the physical rationale behind the appearance of a shear mode in the $S(Q,\omega)$, which primarily couples with longitudinal movements only. It is worth noticing that a similar L-T coupling has been reported in various systems sharing a tetrahedral molecular arrangements, such as water \cite{water}, GeSe$_2$ \cite{GeSe2} and v-GeO$_2$ as well \cite{Bove}. It can thus be envisaged that the open and highly directional nature of such arrangement fosters the onset of an L-T coupling.
Furthermore, the comparison between left and right panels in Fig. \ref{Fig5} indicates that the value of $\Omega_\mathrm{LF}$ well corresponds to the low frequency excess of the v-DOS in the glass as compared to the crystal.
This intensity excess relates to almost universal feature of glasses essentially amounting in a peak in the reduced density of states v-GeO$_2/E^2$, customarily referred to as Boson peak (BP) \cite{nota2}.
The equivalence between the BP of a glass and the Van Hove singularity of the TA branch of the corresponding crystal has been recently demonstrated to be a sound hypothesis \cite{Chumakov}. We observe that the intensity ratio between TA and LA modes decreases upon increasing the pressure consistently with the observed pressure trend of the BP \cite{Niss}.
\section*{Conclusion}
In summary, we investigated the evolution of the THz spectrum of density fluctuations in v-GeO$_2$ as a function of pressure. We observed a clear transformation in the $Q$-dispersion behavior upon crossing the known polyamorphic transition occurring in the glassy phase at $\approx$ 9 GPa. Supported by DFT calculations, we interpreted this as the abrupt evolution from a quartz-like to a rutile-like behavior, concluding that the collective dynamics of v-GeO$_2$ in the THz range is strongly sensitive to the undergoing changes in the local structure. This is clearly indicated by a cusp-like behavior of the pressure dependence of generalized sound velocity and longitudinal modulus across the PA transition. Both these trends suggest that the high-P polyamorphic phase is characterized by a higher resistance to pressure induced modification of elastic properties, likely due to the more packed first neighbor arrangement.
Furthermore, presented data indicates that the inherent disorder characteristics of the glassy phase seems to foster the visibility of a low frequency transverse modes, which, especially at low pressures, is evident at high $Q$'s but it is still appreciable within the first pseudo-Brillouin zone. This mode in the glass is here found to be possibly related to the low frequency excess intensity in the vibrational density of state.
We finally remark that, when compared to more traditional structural measurements, investigations of the THz dynamics as the one presented in this work provide a complementary insight onto the PA transition linking it to transformations of elastic properties.
More in general, the results presented in this work pave the way toward future investigations of the dynamic aspects of polyamorphism phenomena by using THz probes of phonon-like modes.
\section*{Methods}
\subsection{IXS measurements and data analysis}
Two independent IXS experiments were carried out on two different IXS beamlines: beamline BL35XU \cite{BL35} at SPring-8 (Hyogo, Japan) and beamline ID28 \cite{ID28} at the European Synchrotron Radiation Facility (ESRF, Grenoble, France). These two triple-axis spectrometers have the same working principle, based upon high order reflections from nearly perfect Si crystals. The instruments were operated at the Si$(9,9,9)$ configuration for both monochromator and analyzer crystals, corresponding to an incident energy of 17.947 KeV and an overall resolution bandwidth of 3.0 meV (FWHM). The profile of the instrumental resolution function, $R(\omega)$ was estimated through the measurement of the spectral line-shape of an essentially elastic scatterer, specifically a cryogenically cooled sample of perspex at the $Q$ position of its first sharp diffraction peak (10 nm$^{-1}$). The spectrometer was rotated to probe within a single energy scan the 4.5-7.8 nm$^{-1}$ $Q$-range (BL35XU measurements) or the 5.4-13.8 nm$^{-1}$ $Q$ range (ID28 measurements). The focal spot of the beam at sample position was 35$\times$15 $\mu m^2$ and 30$\times$60 $\mu m^2$(horizontal x vertical FWHM) for measurements at BL35XU and ID28, respectively.
The v-GeO$_2$ sample was synthesized as described in a previous work~\cite{Mei_PRB_2010} (BL35XU experiments) or purchased (Sigma-Aldrich; ID28 experiments). In both cases, the samples were loaded in DAC without any pressure transmitting medium to maximize the scattering volume for high-pressure measurements. Samples used for BL35XU experiments were conditioned in Mao type DAC, equipped with tungsten gaskets, while samples used for ESRF experiments were conditioned in membrane driven Le Toullec type DAC, equipped with rhenium gasket. We stress that the metallic gasket, beside radially containing the sample, also acts as cleaning pinhole in the x-ray path. To further minimize, spurious quasi-elastic scattering contributions, DACs were placed in a specifically designed vacuum chamber equipped with motorized exit slits. We indifferently used 350 or 300 $\mu m$ flat culets diamond to cover the pressure range of interest.
A small chip of ruby placed in the sample chamber served as pressure gauge \cite{Ruby}. In order to measure the spectral lineshape with the needed statistical accuracy, we performed energy transfer scans typically lasting 18-24 hours. Prior to each spectral acquisition we measured the $Q$-dependent elastic scattering intensity $I(Q,\omega=0)$, systematically confirming the vitreous nature of the investigated phase.
The pressures probed in the SPing-8 experiment were 3.7, 13 and 20 GPa, while in the ESRF experiment we collected data at room pressure and 26 GPa, in both cases the sample was at ambient temperature. SPring-8 measurements were performed with a better spectral contrast in the low-frequency spectral region, likely due to the lower quasi-elastic scattering from the diamonds DAC windows, while in spectra collected at ESRF had substantially higher count-rate and, consequently, a better statistical accuracy.
The best-fit line-shape modeling of IXS spectra is based on a sum of two Damped Harmonic Oscillator (DHO) functions plus an essentially elastic central peak. Overall the used profile reads as
\begin{eqnarray}
\frac{S(Q,\omega)}{S(Q)}=f(\omega)\left[A\delta(\omega)+I_\mathrm{LF}\frac{\left(2\Omega_\mathrm{LF}\Gamma_\mathrm{LF}\right)^2}{(\omega^2-
\Omega_\mathrm{LF}^2)^2+
4\Gamma_\mathrm{LF}^2\omega^2}+I_\mathrm{HF}\frac{\left(2\Omega_\mathrm{HF}\Gamma_\mathrm{HF}\right)^2}
{(\omega^2-\Omega_\mathrm{HF}^2)^2+4\Gamma_\mathrm{HF}^2\omega^2}\right],
\label{DHO_2}
\end{eqnarray}
\noindent where $f(\omega)=\hbar\omega/k_BT[1-\exp(-\hbar\omega/k_B T)]$ is the detailed balance factor accounting for the frequency-dependent statistical population of the states, with $k_B$ being the Boltzmann constant and $\hbar=h/2\pi$ with $h$ being the Planck constant. The parameters $\Omega_\mathrm{LF} (\Omega_\mathrm{HF})$ and $\Gamma_\mathrm{LF} (\Gamma_\mathrm{HF})$ represent the shift and the width the low (high) frequency inelastic mode, respectively, while $A$, $I_\mathrm{LF}$ and $I_\mathrm{HF}$ are frequency independent scaling factors. A $Q$-dependence is assumed implicitly for all parameters. The double DHO profile of Eq. \ref{DHO_2} was used to fit the measured spectral line-shape, even though at the lowest $Q$s best-fit values of $I_\mathrm{LF}$ were negligible. Additional DHO terms were added to account for the inelastic scattering from the phonon modes of diamond. Due to the very high sound speed of diamond, such well resolved spectral features locate in the high-frequency side of the spectra and disperse out of the probed window at high $Q$'s (see Fig. \ref{Fig1}).
Ultimately, the best fit of measured spectra (see Fig. \ref{Fig1}) was achieved through the minimization a $\chi^2$ variable defined as the normalized distance between the experimental line shape and the following model profile:
\begin{equation}
I(Q,\omega) = K S(Q,\omega)\otimes R(\omega) + B,
\label{final}
\end{equation}
The symbol "$\otimes$" represents the convolution operator, while $K$ and $B$ are two frequency-independent constants representing, respectively, an overall intensity factor and a flat background, which also includes the electronic noise of the detectors ($<$ 1 mHz).
\subsection{Numerical computations}
To get further insights into the dynamic response of the sample, we complemented our experimental investigation with a numerical study. DFT calculations within the local density approximations were carried out using the ABINIT package~\cite{Gonze} and norm-conserving pseudopotentials in the Troullier-Martins scheme. The crystal structures of bulk GeO$_2$ were optimized using a kinetic energy cutoff of 58 Hatree, and the first Brillouin zone were sampled using a $4\times 4\times 4$ and $6\times 6\times 6$ $Q$-grid for the $\alpha$-quartz and rutile structures, respectively. Phonon dispersion curves were interpolated over a $4\times 4\times 4$ and $3\times 3\times 3$ $Q$-grid for $\alpha$-quartz and rutile structures of GeO$_2$, respectively, using the density-functional perturbation theory scheme described in ~\cite{Lee}. Nonanalytic corrections due to the long-ranged anisotropic dipole-dipole interactions were applied to the dynamic matrix. The computed zone-center phonon frequencies at $P=0$ GPa for $\alpha$-quartz and rutile structures are within 0.6 meV of the experimental values~\cite{Madon,Sharma} and previous DFT results~\cite{Hermet}. The weight of the individual contributions to $S(Q,\omega)$ at a discrete ${\bf Q}_i$ on the $j$-th phonon branch (represented in Figs. \ref{Fig3} and \ref{Fig4} as the thickness of the corresponding line after multiplying with $\omega({\bf Q}_i)$) is denoted as $S_{j}({{\bf Q_i}})$:
\begin{eqnarray}
S(Q,\omega)&=&\sum_{i,j}S_{j}({{\bf Q}_i})\delta(\omega-\omega_j({\bf Q}_i)),
\end{eqnarray}
with:
\begin{eqnarray}
S_{j}({{\bf Q}_i})&=&\frac{1+n_j({\bf Q}_i)}{2\omega_j({\bf
Q}_i)}\bigg|\sum_d\frac{f_d({\bf Q}_i)}{\sqrt{M_d}} {\bf Q}_i\cdot
\sigma_d^j({\bf Q}_i) \mathrm{exp}^{\imath{\bf Q}_i\cdot{\bf
d}}\bigg|^2 \nonumber\\
\end{eqnarray}
\noindent where $f_d({\bf Q})$ is the x-ray form factor and $\sigma_d^j({\bf Q})$ is the projection on to atom $d$ of eigenvector in the $j$-th branch. Effects due to the Debye-Waller factor were ignored since it induces a negligible $Q$ dependence within the probed momentum transfer interval.
|
\section{Introduction}
Graphical models and corresponding message-passing algorithms have attracted a great amount of attention due to their wide-spreading application in many fields, including signal processing, machine learning, channel and source coding, computer vision, decision making, and game theory ({e.g.}, see \cite{kschischang2001factor, loeliger2004introduction}).
Finding marginal and mode of a probability distribution are two basic problems encountered in the field of graphical models. Taking the rudimentary approach, the marginalization problem has exponentially growing complexity in alphabet size. However, using BP algorithm (firstly introduced in \cite{pearl1988probabilistic}) to solve this problem either exactly or approximately, we can reduce the computational complexity to a significant degree. It has been proved that applying BP on graphical models without cycles provides exact solution to the marginalization problem.
Furthermore, it has been observed that for general graphs, BP can find good approximations for marginalization (or finding mode) problems, \cite{kschischang2001factor, loeliger2004introduction}.
Although BP has many favourable properties, it suffers from some limiting drawbacks. First, in complex and densely interconnected graphs, BP may not be able to produce accurate results; and even worse, it may not converge at all. Second, since in many applications ({e.g.}, decoding of error-correcting codes) messages are of high dimensions, the computational complexity of BP algorithm will highly increase which leads to slow convergence rates.
To deal with the first drawback, some works have been done to propose alternative algorithms ({e.g.}, see \cite{yuille2002cccp, welling2001belief, heskes2002stable, pakzad_estimation_2005}).
Specifically, to improve the accuracy of estimated marginal distribution, a generalization algorithm to BP has been introduced by Yedidia et al.~\cite{yedidia_constructing_2005}, known as Generalized Belief Propagation (GBP) algorithm. In their proposed algorithm, local computation is performed by a group of nodes instead of a single node as in BP. According to many empirical observations, GBP outperforms BP in many situations; \cite{yedidia2001characterization, harel2003poset, welling2004choice, sibel2014application}.
However, although GBP algorithm provides accurate results in terms of marginal distribution, it suffers from high order of computation complexity, specially in case of large alphabet size.
To overcome the second aforementioned deficiency of BP, lots of research have been conducted to reduce BP complexity for different applications ({e.g.}, refer to \cite{felzenszwalb2006efficient,sudderth2010nonparametric,mcauley2011faster, isard2009continuously, kersting2009counting, coughlan2007dynamic, arulampalam2002tutorial, smith2013sequential}).
In a recent work by Noorshams~et al.~\cite{noorshams_stochastic_2013}, to tackle with the challenge of high complexity in the case of large alphabet size, they introduce an alternative stochastic version of BP algorithm with lower complexity. The main idea behind their work is that each node sends a randomly sampled message taken from a properly chosen probability distribution instead of computing the exact message update rule in each iteration.
Motivated by \cite{noorshams_stochastic_2013} and in order to mitigate the computational complexity of GBP, we extend GBP and propose stochastic GBP (SGBP) algorithm. SGBP has the advantage of reducing the complexity, while increasing the accuracy of estimation.
In contrast to SBP, SGBP algorithm can reduce the computational complexity only if certain topological conditions are met by the region graph (defined later) associated to a graphical model. However, the complexity gain can be larger than only one order of magnitude in alphabet size. In this work, we characterize these conditions and the amount of computational gain that we can obtain by performing SGBP instead of GBP. Determining these criteria, we hope that they provide some useful guidelines on how to choose the regions and construct the region graph in a way that results to a lower complexity algorithm with good accuracy.
The rest of the paper is organized as follows.
First, \S\ref{sec:ProblemStatement} introduces our problem statement. In \S\ref{sec:SGBP_Alg}, we present the proposed stochastic GBP and then derive the topological conditions that guarantee SGBP has lower complexity than GBP.
Moreover, theoretical convergence results have been provided as well. Finally, to validate our theoretical results, considering a specific graphical model, SGBP is simulated and the results are presented.
\section{Problem Statement}
\label{sec:ProblemStatement}
\subsection{Notation}
In the following, we introduce the notation that will be used in the paper. The random variables are represented by upper case letters and their values by lower case letters. Vectors and matrices are determined by bold letters. Sometimes, we use calligraphic letters to denote sets. When we have a set of random variables $X_1,\ldots,X_n$, we write $\vect{X}_\set{A}$ to denote $(X_i,i\in\set{A})$. An undirected graph $G = (\set{V},\set{E})$ is defined by a set of nodes $\set{V}=\{1,2,\ldots,n\}$ and a set of edges $\set{E} \subseteq \set{V}\times \set{V}$, where $(u, v) \in \set{E}$ if and only if nodes $u$ and $v$ are connected. Similarly, we can define a directed graph.
For every function $f(x_1,x_2,\ldots,x_n)$ where $f: {\mc{X}}^n \mapsto \mbb{R}$, we define the operator $\cal{L}$ as a map that turns this function to a vector ${\cal{L}}(f)\in\mbb{R}_{ {|\mc{X}|}^n\times 1}$ by evaluating $f$ at every input point.
For instance, considering ${\cal{X}}_{\{1,2\}}\in\{0,1\}$, for $f(x_1,x_2)$ we have
\[
{\cal{L}}(f) = \begin{bmatrix}
f(0,0)\\
f(0,1)\\
f(1,0)\\
f(1,1)\\
\end{bmatrix}.
\]
\subsection{Graphical Model}
Undirected graphical models, also known as Markov random fields (MRF), is a way to represent the probabilistic dependencies among a set of random variables having Markov properties using an undirected graph.
More precisely, we say that a set of random variables $X_1,\ldots,X_n$ form an MRF if there exists a graph $G=(\set{V},\set{E})$, where each $X_i$ is associated to the node $i \in \set{V}=\{1,\ldots,n\}$, and edges of the graph $G$ encode Markov properties of the random vector $\vect{X}=(X_1,\ldots,X_n)$. These Markov properties are equivalent to a factorization of the joint distribution of random vector $\vect{X}$ over the cliques of graph $G$ \cite{grimmett2010probability}.
In this paper, we focus on discrete random variables case where for all $j\in \set{V}$ we have $X_j\in{\mc{X}}\triangleq\{1,2,\ldots,d\}$. Moreover, we assume that the distribution of $\vect{X}$ is factorized according to
\[
p(\vect{x}) = \frac{1}{Z}\prod_{a\in \set{F}} \phi_a(\vect{x}_a)
\]
where $\set{F}$ is a collection of subsets of $\set{V}$ and $Z$ is a constant called the partition function. For the factor functions $\phi_a$, we have also $\phi_a\ge 0$. This factorization can be represented by using a bipartite graph $G_f=(\set{V},\set{F},\set{E}_f)$ called factor graph. In this representation, the variable nodes $\set{V}$ correspond to random variables $X_i$'s and factor nodes $\set{F}$ determine the factor functions $\phi_a$'s. Moreover, there exists an edge $(i,a)\in\set{E}_f$ between a variable node $i$ and a factor node $a$ if the variable $x_i$ appears in the factor $\phi_a$ (for more information on factor graphs refer to \cite{kschischang2001factor}).
\subsection{Region Graph}
In order to present the Yedidia's parent-to-child algorithm \cite{yedidia_constructing_2005} as well as introducing our stochastic GBP algorithm, we need to state some definitions as follows.
\begin{definition}[see \cite{yedidia_constructing_2005}]
A \emph{region graph} $G_r=(\set{R},\set{E}_r)$ defined over a factor graph $G_f=(\set{V},\set{F},\set{E}_f)$ is a directed graph in which for each vertex $v\in\set{R}$ (corresponding to a region) we have $v\subseteq \set{V}\cup\set{F}$. Each region $v$ has this property that if a factor node $a\in\set{F}$ belongs to $v$ then all of its neighbouring variable nodes have to also belong to $v$.
A directed edge $(v_p \rightarrow v_c)\in\set{E}_r$
may exist if $v_c\subset v_p$. If such an edge exists, $v_p$ is a \emph{parent} of $v_c$, or equivalently, $v_c$ is a \emph{child} of $v_p$.
If there exists a directed path from $v_a$ to $v_d$ on $G_r$, we say that $v_a$ is an \emph{ancestor} of $v_d$ and $v_d$ is a \emph{descendant} of $v_a$.
\end{definition}
Now, for each $R\in \set{R}$, we let $\set{P}(R)$ denotes for the set of all parents of $R$, $\set{A}(R)$ denotes for the set of all ancestors of $R$ and $\set{D}(R)$ denotes for the set of all descendants of $R$.
Moreover, we define $\set{E}(R) \triangleq R \cup \set{D}(R)$. Finally, for a region $R\in\set{R}$, we use $|R|$ to denote for the number of variable nodes in $R$.
\subsection{Parent-to-child GBP algorithm}
We may derive the BP message-passing equations using the fact that the belief at each variable node is the product of all the incoming messages received from its neighbouring factor nodes. Additionally, the beliefs over the set of variable nodes connecting to a factor node $a\in\set{F}$ is the product of the factor function $\phi_a$ multiplied by the incoming messages to the factor node $a$. Now marginalizing the second set of beliefs to find the belief over a variable node and equate it to the belief of that variable node which is found directly using the first equation, we can recover the BP update rules.
\begin{figure}[h]
\centering
\includegraphics[scale= 0.4]{Fig7}
\caption{Graph region of an arbitrary graph corresponding to the parents-to-child algorithm.}
\label{fig:ProblemSetup11}
\end{figure}
Yedidia et al., generalize this idea in \cite{yedidia_constructing_2005}, proposing an algorithm called parents-to-child GBP algorithm. As explained in \cite{yedidia_constructing_2005}, in the parent-to-child algorithm, we have only one kind of message $m_{P\rightarrow R}(\vect{x}_R)$ from a parent region $P$ to a child region $R$. Then for the belief of region $R\in\set{R}$ we have
\begin{align}\label{eq:GBP_RegionBeliefComputation}
b_R(\vect{x}_R) \propto \Phi_{R}(\vect{x}_R) \times \prod_{P\in {\cal{P}}(R)}m_{P\rightarrow R}(\vect{x}_{R})
\times \prod_{D\in {\cal{D}}(R)}\prod_{P'\in{\cal{P}}(D)\setminus \cal{E}(R)}m_{P'\rightarrow D}(\vect{x}_D)
\end{align}
where $\Phi_{R}(\vect{x}_R) \triangleq \prod_{a\in R} \phi_a(\vect{x}_a)$ (with an abuse of notation when we product over $a\in R$ we mean to product only over the factor indexes of $R$).
Then, the message update rule over each edge $(P, R)\in\set{E}_r$ follows by
\begin{align}
m_{P\rightarrow R}(\vect{x}_{R}) =& \frac{\sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\prod_{(I,J)\in N(P,R)}m_{I\rightarrow J}(\vect{x}_{J})}{\prod_{(I,J)\in D(P,R)}m_{I\rightarrow J}(\vect{x}_{J})} \label{eq:UpdateRule_FirstEquation}\\
=& \sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\hat{M}(\vect{x}_{T_{PR}})\label{eq:TGBP}
\end{align}
where $\Phi_{P\setminus R}(\vect{x}_{P'})\triangleq \frac{\Phi_{P}}{\Phi_{R}}(\vect{x}_{P'})$ and ${P'}$ is the set of all variables appear in $\frac{\Phi_{P}}{\Phi_{R}}(\vect{x}_{P'})$. In addition, we have also
\begin{equation}
N(P,R) \triangleq \Big\{(I,J) | (I, J)\in \set{E}_r, I\notin \set{E}(P), J\in \set{E}(P) \setminus \set{E}(R) \Big\} \label{eq:Def_Set_N(P,R)}
\end{equation}
and
\begin{equation}
D(P,R) \triangleq \Big\{(I,J) | (I, J)\in \set{E}_r, I\in \set{D}(P) \setminus \set{E}(R), J\in \set{E}(R) \Big\}. \label{eq:Def_Set_D(P,R)}
\end{equation}
Notice that the sets $N(P,R)$ and $D(P,R)$ can be calculated in advance. Moreover, $\hat{M}(\vect{x}_{T_{PR}})$ in \eqref{eq:TGBP} is defined as follows
\[
\hat{M}(\vect{x}_{T_{PR}})\triangleq\frac{\prod_{(I,J)\in N(P,R)}m_{I\rightarrow J}(\vect{x}_{J})}{\prod_{(I,J)\in D(P,R)}m_{I\rightarrow J}(\vect{x}_{J})},
\]
where ${T_{PR}}$ is the set of all variables that appear in the above ratio.
\begin{remark}
It can be easily observed that depending on the graph topology and the choice of regions, we may have either $P'\subset P$ or $P'=P$ in \eqref{eq:TGBP}.
For example, consider two pairwise Markov Random Fields presented in Figures \ref{fig:ProblemSetup-smallerp} and \ref{fig:biggerp}.
Considering $P=\{1,2,4,5,7,8\}$ and $R=\{2,5,8\}$ in Figure~\ref{fig:ProblemSetup-smallerp}, we have $\Phi_{(124578\setminus 258)}(\vect{x}_{P'}) = \phi_1\phi_4\phi_7\psi_{12}\psi_{14}\psi_{74}\psi_{78}$
which leads to $P'=\{1,2,4,7,8\} \subset P$.
On the other hand, choosing $P=\{1,2,4,5\}$ and $R=\{2,5\}$ in Figure~\ref{fig:biggerp}, we have $\Phi_{(1245\setminus 25)}(\vect{x}_{P'}) = \phi_1\phi_4\psi_{12}\psi_{14}\psi_{45}$. Hence, $P'=\{1,2,4,5\}=P$.
\hfill {\small $\blacksquare$}
\end{remark}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.8,very thick]
\node at (-1.7,5.5)(index1){};
\node at (-1.7,5) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node1l) {1};
\node at (-1.7,4) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node2l) {4};
\node at (-1.7,3) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node3l) {7};
\node at (-0.85,4)[circle,dashed,draw=green!200,minimum size=65mm] (node3x){};
\node at (+0.85,4)[circle,dashed,draw=green!200,minimum size=65mm] (node3x){};
\node at (0,5.5)(index2){};
\node at (0,5) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node1) {2};
\node at (0,4) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node2) {5};
\node at (0,3) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node3) {8};
\node at (1.7,5.5)(index3){};
\node at (1.7,5) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node1r) {3};
\node at (1.7,4) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node2r) {6};
\node at (1.7,3) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node3r) {9};
\draw[thick] (node3l)--(node3)--(node3r)--(node2r)--(node1r)--(node1)--(node1l)--(node2l);\draw[thick] (node2l)--(node3l);\draw[thick](node1)--(node2)--(node3);
\end{tikzpicture}
\caption{Graph region of a graph with $x_{P'}\subset x_{P}$.}
\label{fig:ProblemSetup-smallerp}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.2,very thick]
\node at (-2.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1l) {124578};
\node at (0,3) [rectangle,draw=black!50,fill=red!20!white,inner sep=10pt, minimum size=10mm] (node) {258};
\node at (2.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1r) {235689};
\draw[thick,->](node1l)--(node);\draw[thick,->](node1r)--(node);
\end{tikzpicture}
\caption{Graph region of a graph with $x_{P'}\subset x_{P}$.}
\label{fig:ProblemSetup-smallerreg}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=2,very thick]
\node at (-1.7,5.5)(index1){};
\node at (-1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1l) {1};
\node at (-1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2l) {4};
\node at (-1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3l) {7};
\node at (-0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (-0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (0,5.5)(index2){};
\node at (0,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1) {2};
\node at (0,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2) {5};
\node at (0,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3) {8};
\node at (1.7,5.5)(index3){};
\node at (1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1r) {3};
\node at (1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2r) {6};
\node at (1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3r) {9};
\draw[thick] (node3l)--(node3)--(node3r)--(node2r)--(node1r)--(node1)--(node1l)--(node2l)--(node2)--(node2r);\draw[thick] (node2l)--(node3l);\draw[thick](node1)--(node2)--(node3);
\end{tikzpicture}
\caption{Basic clusters in 9 nodes grid with $x_{P'}=x_{P}$}\label{fig:biggerp}
\end{center}
\end{figure}
\begin{remark}
In the parent-to-child algorithm, the message transmitted over each edge $(P, R) \in \set{E}_r$ can be considered as a vector by applying the operator $\mc{L}(\cdot)$.
Namely, by concatenating all possible messages, we define $\vect{m}_{P\rightarrow R} \triangleq \mc{L}(m_{P\rightarrow R})$ where $\vect{m}_{P\rightarrow R} \in \mbb{R}^{d^{|R|}}$.
Moreover, concatenating all the messages over all edges of the region graph, we define $\vect{m} \triangleq \{\vect{m}_{P\rightarrow R}\}_{ (P, R)\in \set{E}_r} \in \mbb{R}^\Delta$ where $\Delta=\sum_{(P,R)\in \set{E}_r} d^{|R|}$.
\hfill {\small $\blacksquare$}
\end{remark}
Now, we can state the complexity of the parent-to-child GBP algorithm as stated in Lemma~\ref{lem:GBP_Complexity}.
\begin{lemma}\label{lem:GBP_Complexity}
The computation complexity of the message update rule of the parent-to-child GBP algorithm associated with each edge, computed according to \eqref{eq:TGBP}, is ${\mc{O}}(d^{|P|})$.
\end{lemma}
\begin{proof}
For each fixed vector $\vect{x}_R$, the calculation of $m_{P\rightarrow R}(\vect{x}_{R}) =\sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\hat{M}(\vect{x}_{T_{PR}})$
needs $d^{|P\setminus R|}$ operations. Moreover, to find $m_{P\rightarrow R}(\cdot)$ completely, one needs to evaluate the above summation $\mc{O}(d^{|R|})$ times.
Consequently, the overall complexity of calculating $m_{P\longrightarrow R}(\vect{x}_{R}) $ is of the order ${\mc{O}}(d^{|R|}\times d^{|P\setminus R|})={\cal{O}}(d^{|P|})$.
\end{proof}
At each round of the parent-to-child algorithm, $t=1,2,\ldots$, every parent node $P$ of $R$ in the region graph calculates a message $m^{(t+1)}_{P\rightarrow R}$ and sends it to node $R$. Mathematically, this can be written as (see \cite{yedidia_constructing_2005})
\begin{align}\label{eq:UpdateRuleExpansion1}
m^{(t+1)}_{P\rightarrow R} (\vect{x}_R) =& \left[ \Upsilon_{P\rightarrow R}(m^{(t)}) \right] (\vect{x}_R) \nonumber\\
& \hspace{-2cm} = \sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\hat{M}^{(t)}(\vect{x}_{T_{PR}}) \nonumber\\
& \hspace{-2cm} = \sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} }} \sum_{\vect{x}_{(P\setminus R)\cap T_{PR} }} \Phi_{P\setminus R}(\vect{x}_{P'}) \hat{M}^{(t)}(\vect{x}_{ T_{PR}} ) \nonumber\\
& \hspace{-2cm} = k_{PR}^{(t)} \left( \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right) \hspace{-11pt} \sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} }} \sum_{\vect{x}_{(P\setminus R)\cap T_{PR} }} \Big[ \Phi_{P\setminus R}(\vect{x}_{P'}) \times Q^{(t)}(\vect{x}_{ T_{PR} \cap (P\setminus R)} | \vect{x}_{ T_{PR} \setminus (P\setminus R) }) \Big],
\end{align}
where
\begin{align}\label{eq:Dist_Q_Def}
Q^{(t)}(\vect{x}_{ T_{PR} \cap (P\setminus R)} | \vect{x}_{ T_{PR} \setminus (P\setminus R) }) \triangleq\frac{\hat{M}^{(t)} \left( \vect{x}_{ T_{PR} \cap (P\setminus R) } , \vect{x}_{ T_{PR} \setminus (P\setminus R)} \right) }{\sum_{\vect{x}'_{ T_{PR} \cap (P\setminus R) }} \hat{M}^{(t)} \left( \vect{x}'_{ T_{PR} \cap (P\setminus R) }, \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right) }
\end{align}
is a conditional distribution. Moreover,
\[
k_{PR}^{(t)} \left( \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right) \triangleq \hspace{-12pt} \sum_{\vect{x}'_{ T_{PR} \cap (P\setminus R)}} \hspace{-12pt} \hat{M}^{(t)} \left( \vect{x}'_{ T_{PR} \cap (P\setminus R)} , \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right).
\]
Hence, for the update rule we can write
\begin{align}
m^{(t+1)}_{P\rightarrow R} (\vect{x}_R) = k_{PR}^{(t)}
\hspace{-10pt} \sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} } } \hspace{-10pt} \mbb{E}_{[\vect{X}_{(P\setminus R)\cap T_{PR} } \sim Q^{(t)} ] } \big[\Phi_{P\setminus R}(\vect{X}_{P'})\big] \label{eq:UpdateRuleExpansion2}
\end{align}
Here and in the following, for brevity and clarity of notation, we will omit the dependence of $k_{PR}^{(t)}$ to the variables $\vect{x}_{ T_{PR} \setminus (P\setminus R) }$.
Now, notice that we can decompose the set $P'$ as follows
\[
P' = \left[ (P\setminus R)\cap T_{PR} \right] \cup \left[ (P\setminus R)\setminus T_{PR} \right] \cup \left[P' \setminus (P\setminus R) \right]
\]
because we always have $P\setminus R \subseteq P'$. By using this relation, we can rewrite \eqref{eq:UpdateRuleExpansion2} as follows
\begin{align} \label{eq:UpdateRuleExpression_Final}
m^{(t+1)}_{P\rightarrow R} (\vect{x}_R) = k_{PR}^{(t)}
\sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} } }
\mbb{E}_{ [\vect{X}_{(P\setminus R)\cap T_{PR} } \sim Q^{(t)} ] } \bigg[
\Phi_{P\setminus R} \Big(\vect{X}_{(P\setminus R)\cap T_{PR} }, \vect{x}_{(P\setminus R)\setminus T_{PR} }, \vect{x}_{P'\setminus (P\setminus R)} \Big) \bigg].
\end{align}
In \eqref{eq:UpdateRuleExpansion1}, $\Upsilon_{P\rightarrow R}:\mbb{R}^{\Delta} \mapsto \mbb{R}^{d^{|R|}}$ is the local update function of the directed edge $(P, R)\in \set{E}_r$. By concatenating all the local update functions over the edges of the region graph, we can define the global update function as
\begin{equation}\label{eq:GlobalUpdateFuncDef}
\Upsilon(\vect{m})=\Big[ \Upsilon_{P\rightarrow R}(\vect{m}) : (P,R)\in \set{E}_r \Big]
\end{equation}
where $\Upsilon: \mbb{R}^\Delta \mapsto \mbb{R}^\Delta$.
The goal of the (parent-to-child) GBP algorithm is to find a fixed point $\vect{m}^*$ that satisfies $\Upsilon(\vect{m}^*)=\vect{m}^*$. If a fixed point $\vect{m}^*$ is found, then the beliefs of random variables in a region $R$ is computed by applying \eqref{eq:GBP_RegionBeliefComputation}.
\section{Stochastic Generalized Belief Propagation Algorithm}
\label{sec:SGBP_Alg}
In this section, first we introduce our stochastic extension to the parent-to-child GBP algorithm, and then present a result on the criteria where this algorithm is able to mitigate the computation complexity of GBP.
Based on \eqref{eq:UpdateRuleExpression_Final}, we introduce our algorithm as stated in Algorithm~\ref{alg:SGBP_Alg}. The main idea of the algorithm is that under proper conditions (that will state in Theorem~\ref{thm:mainresult}), some parts of the message update rule \eqref{eq:TGBP} for each edge of the region graph can be written as an expectation as stated in \eqref{eq:UpdateRuleExpression_Final}.
\begin{algorithm}
\caption{Stochastic Generalized Belief Propagation (SGBP) algorithm.}
\label{alg:SGBP_Alg}
\begin{algorithmic}[1]
\STATE \textbf{Initialize the messages.}
\FOR{$t\in\{1,2,\ldots\}$ and each directed edge $(P, R)\in\set{E}_r$}
\STATE Choose a random vector $\vect{J}_{PR}^{(t+1)} \in \mc{X}^{| T_{PR} \cap (P\setminus R)|}$ according to the conditional distribution $Q^{(t)}(\vect{x}_{ T_{PR} \cap (P\setminus R)} | \vect{x}_{ T_{PR} \setminus (P\setminus R) })$ defined in \eqref{eq:Dist_Q_Def}.
\STATE Update the message $m_{P\longrightarrow R}^{(t+1)}$ with the appropriately tuned step size $\alpha^{(t)}={\mathcal{O}}(\frac{1}{t})$ according to
\begin{align}\label{eq:UpdateRule_Alg}
m_{P\rightarrow R}^{(t+1)} (\vect{x}_R) =(1-\alpha^{(t)}) m_{P\rightarrow R}^{(t)} (\vect{x}_R) +\alpha^{(t)} k_{PR}^{(t)} (\vect{x}_{T_{PR}\setminus (P\setminus R)}) \hspace{-10pt} \sum_{\vect{x}_{(P\setminus R)\setminus T_{PR} }} \hspace{-14pt} \Phi_{P\setminus R} \left(\vect{J}_{PR}^{(t+1)}, \vect{x}_{(P\setminus R)\setminus T_{PR} }, \vect{x}_{P'\setminus (P\setminus R)} \right)
\end{align}
\STATE $t = t + 1$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{remark}\label{remrk:DeterministicUpdateRule}
Note that when $(P\setminus R)\cap T_{PR}=\varnothing$, the update rule \eqref{eq:UpdateRule_Alg} becomes deterministic as stated in the following
\[
m_{P\rightarrow R}^{(t+1)}=(1-\alpha^{(t)})m_{P\rightarrow R}^{(t)} +\alpha^{(t)} k_{PR}^{(t)} (\vect{x}_{T_{PR}}) \Big[\sum_{\vect{x}_{(P\setminus R)}}\Phi_{P\setminus R} \left(\vect{x}_{(P\setminus R) }, \vect{x}_{P'\setminus (P\setminus R)} \right)\Big].
\]
Later, we will prove in Lemma~\ref{lem:Graph-region property} that this condition can only happen in update rules corresponding to the highest-level ancestors regions.
\hfill {\small $\blacksquare$}
\end{remark}
In contrast to SPB studied in \cite{noorshams_stochastic_2013}, the stochastic version of GBP does not always reduce the computational complexity in each iteration. Theorem~\ref{thm:mainresult} describes the topological and regional conditions for which the complexity of SGBP is less than GBP for a specific edge of the region graph.
\begin{theorem}\label{thm:mainresult}
Our proposed algorithm that runs over a region graph $G_r$ reduces the computation complexity of each message $m_{P\rightarrow R}$ (compared to GBP) if and only if the following conditions hold
\begin{itemize}
\item[(i)] $(P\setminus R)\cap T_{PR} \neq\varnothing$
\item[(ii)] $(P\setminus R) \nsubseteq T_{PR}$
\end{itemize}
\end{theorem}
\begin{proof}
The main idea of the proof lies in the fact that whether or not \eqref{eq:TGBP} can be written in the form of an expected value of \emph{potential functions} as stated in \eqref{eq:UpdateRuleExpression_Final}. If this happens, as presented in Algorithm~\ref{alg:SGBP_Alg}, the complexity of update rules can be reduced. Now, to be able to have an expectation operation in \eqref{eq:UpdateRuleExpression_Final}, we should have $(P\setminus R)\cap T_{PR} \neq\varnothing$.
Now, assuming condition (i) holds, we find the complexity of Algorithm~\ref{alg:SGBP_Alg}'s update rule over every edge $(P, R)\in\set{E}_r$ in each iteration. First, let us fix $\vect{x}_R$. To find the PMF of the random vector $\vect{J}$ which is given by \eqref{eq:Dist_Q_Def},
we need ${\mc{O}}(d^{|\{P\setminus R\}\cap T_{PR} |}\times d^{| T_{PR} \setminus \{P\setminus R\}|}) = {\mc{O}}(d^{| T_{PR} |})$ operations. Notice that since we have $ T_{PR} \setminus (P\setminus R) \subseteq R$ and $[(P\setminus R)\cap T_{PR} ] \cap [(P\setminus R)\setminus T_{PR} ]=\varnothing$, for every fixed $\vect{x}_R$, the PMF of $\vect{J}$ does not depend on the vector $\vect{x}_{(P\setminus R)\setminus T_{PR} }$.
This means that for a fixed $\vect{x}_R$, to find the summation in \eqref{eq:UpdateRule_Alg}, the PMF of $\vect{J}$ should only computed once.
Hence, the overall complexity of update rule \eqref{eq:UpdateRule_Alg} becomes
\[
\mc{O} \left( d^{| T_{PR} |}+d^{|R|}\big[d^{|(P\setminus R)\cap T_{PR} |} + d^{|(P\setminus R)\setminus T_{PR} |}+ d^{|(P\setminus R)\cap T_{PR} |}\big] \right)
\]
where the terms in the brackets count for a fixed $\vect{x}_R$ the computation complexity of $k(\vect{x}_{ T_{PR} \setminus {P\setminus R}})$, of the summation in \eqref{eq:UpdateRule_Alg}, and of taking a sample vector $\vect{J}$ from the above PMF, respectively. The above relation can be rewritten as follows
\[
\mc{O} \left( \max\big[d^{| T_{PR} |}, d^{|R|+|(P\setminus R)\setminus T_{PR} |},d^{|R|+|(P\setminus R)\cap T_{PR} |}\big] \right)
\]
Now, we can conclude that if $T_{PR} \neq \varnothing$ and $(P\setminus R)\not\subset T_{PR}$ then we have
\[
\mc{O} \left( \max\big[d^{| T_{PR} |}, d^{|R|+|(P\setminus R)\setminus T_{PR} |},d^{|R|+|(P\setminus R)\cap T_{PR} |}\big] \right) < {\cal{O}}(d^{|P|}),
\]
where the right hand side is the computation complexity of the parent-to-child GBP algorithm derived in Lemma~\ref{lem:GBP_Complexity}. This completes the proof of theorem.
\end{proof}
\begin{corollary}\label{cor:orderreduction}
Assuming that the conditions of Theorem~\ref{thm:mainresult} hold and denoting
\[
\eta_{PR}\triangleq\max{\Big[ | T_{PR} |,|R|+|(P\setminus R)\setminus T_{PR} |, |R|+|(P\setminus R)\cap T_{PR} | \Big]},
\]
Algorithm~\ref{alg:SGBP_Alg} reduces the computation complexity of message $m_{P\rightarrow R}$ of the order
$\mathcal{O}(d^{|P|-\eta_{PR}})=\mathcal{O}(d^{I_{PR}})$
where $I_{PR} \triangleq |P|-\eta_{PR}$. Notice that $I_{PR}$ can be larger than $1$.
\end{corollary}
\begin{example}\label{ex:RegionGraph_with_I=2}
In this example, we provide a graph, drawn in Figure~\ref{fig:I=2}, in which SGBP reduces the complexity of updating rule for the edge $123456\rightarrow 36$ (see Figure~\ref{fig:graph region of I=2}) with two order of magnitude. The updating rule for $m_{123456\rightarrow36}$ is as follows
\begin{align*}
m_{123456\rightarrow36}=& \sum_{x_1x_2x_4x_5} \frac{\Phi_{123456}(x_1,x_2,x_3,x_4,x_5,x_6)}{\Phi_{36}(x_3,x_6)} \times \frac{m_{2478\rightarrow 24}(x_2,x_4)}{1}\\
=&\sum_{x_2x_4}\tilde{\Phi}(x_2,x_3,x_4,x_6) m_{2478\rightarrow 24}(x_2,x_4)\\
=& k\:\mathbb{E}_{(X_2,X_4)\sim Q} \left[ \tilde{\Phi}(X_2,x_3,X_4,x_6) \right]
\end{align*}
where in this example by applying SGBP we will get $I_{123456\rightarrow36}=2$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.45]{Fig10}
\caption{A graph in which the SGBP reduces the complexity with two order of magnitude in alphabet size (see Example~\ref{ex:RegionGraph_with_I=2}).}
\label{fig:I=2}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.2,very thick]
\node at (-1.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1l) {2478};
\node at (4,3) [rectangle,draw=black!50,fill=red!20!white,inner sep=10pt, minimum size=10mm] (node1rr) {36};
\node at (0.5,3) [rectangle,draw=black!50,fill=red!20!white,inner sep=10pt, minimum size=10mm] (node) {24};
\node at (2.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1r) {123456};
\draw[thick,->](node1l)--(node);\draw[thick,->](node1r)--(node);
\draw[thick,->](node1r)--(node1rr);
\end{tikzpicture}
\caption{Associated graph region of a graphical model depicted in Figure~\ref{fig:I=2} with $I_{123456\rightarrow36}=2$ (see Example~\ref{ex:RegionGraph_with_I=2}).}
\label{fig:graph region of I=2}
\end{center}
\end{figure}
\hfill {\small $\blacksquare$}
\end{example}
\begin{corollary}
The complexity of the parent-to-child GBP algorithm is dominated by the computation complexity of message update rule of the highest level edges in the region graph $G_r$.
As a result, if the dominant message update rule that belongs to the highest-level ancestor regions with the largest size, satisfies the conditions of Theorem~\ref{thm:mainresult}, then no matter what are the complexity of other edges, Algorithm~\ref{alg:SGBP_Alg} will reduce the overall computation complexity of the parent-to-child GBP.
\end{corollary}
\subsection{Convergence Rate of SGBP Algorithm}
\label{sec:Convergence_SGBP}
In this section, we extend the convergence guarantees of \cite{noorshams_stochastic_2013} to SGBP. Our convergence theorem (Theorem~\ref{thm:Convergence_Results}) is based on imposing a sufficient condition similar to \cite{noorshams_stochastic_2013} that guarantees uniqueness and convergence of the parent-to-child GBP message updates. More precisely we assume that the global update function $\Upsilon(\cdot)$, defined in \eqref{eq:GlobalUpdateFuncDef}, is contractive, namely $\exists \nu, 0<\nu<2$ such that
\begin{align}
\|\Upsilon(\vect{m})-\Upsilon(\vect{m}')\|_2 \leq \left( 1-\frac{\nu}{2} \right) \| \vect{m}-\vect{m}' \|_2. \label{eq:ContractiveCond_Def}
\end{align}
Following similar proof technique to \cite{noorshams_stochastic_2013}, with some appropriate modifications, we can obtain the following results.
\begin{theorem}\label{thm:Convergence_Results}
Assume that, for a given region graph, the update function $\Upsilon$ is contractive with parameter $1-\frac{\nu}{2}$ as defined in \eqref{eq:ContractiveCond_Def}. Then, parent-to-child GBP has a unique fixed point $\vect{m}^*$ and the message sequence $\{\vect{m}^{(t)}_{P\rightarrow R}\}^{\infty}_{t=1}$ generated by the SGBP algorithm has the following properties:
\begin{itemize}
\item[i)] The result of SGBP is consistent with GBP, namely we have $\vect{m}^{(t)}\overset{\mathrm{a.s.}}{\longrightarrow}\vect{m}^*$ as $t\longrightarrow \infty$.
\item[ii)] Bounds on mean-squared error: Let us divide the fixed point message $\vect{m}^*$ into two parts, $\vect{m}^* = {(\vect{m}_{\mc{E}_{1}}^*,\vect{m}_{\mc{E}_{\sim 1}}^*)}$, where $\vect{m}^*_{\mc{E}_{r_1}}$ corresponds to those edges of the region graph that perform deterministic update rule (as stated in Algorithm~\ref{alg:SGBP_Alg} and Remark~\ref{remrk:DeterministicUpdateRule}), while $\vect{m}^*_{\mc{E}_{\sim 1}}$ corresponds to the edges that run the stochastic algorithm. In other words, $\mc{E}_{1}$ and $\mc{E}_{\sim 1}$ represents the edges in region graph in which the message updating rules are deterministic and stochastic respectively.
Choosing step size $\alpha^{(t)}=\frac{\alpha}{\nu(t+2)}$ for some fixed $1<\alpha<2$ and defining $\delta_i^{(t)}\triangleq \frac{\vect{m}_i^{(t)}-\vect{m}_i^*}{\|\vect{m}_i^*\|^2}$ for each $i\in\{1,\sim 1\}$, we have
\begin{align*}
\frac{\mathbb{E}[\|\delta^{(t)}\|^2_2]}{\|\vect{m}^*\|^2_2} &\leq \left(\frac{3^{\alpha}{\alpha}^2 \Lambda(\Phi',k_{lu})}{2^\alpha(\alpha-1)\nu^2}\right)\frac{1}{t}+\frac{\mathbb{E}[{\|{\delta}_{\mc{E}_{\sim 1}}^{(0)}\|}^2_2]}{\|\vect{m}_{\mc{E}_{\sim 1}}^*\|^2_2}\left(\frac{2}{t}\right)^\alpha \nonumber
\end{align*}
for all iteration $t=1,2,3,\ldots$ where $\Lambda(\Phi',k_{lu})$ is a constant which depends on some factor functions (through $\Phi'$) and some variable nodes (through $k_{lu}$). For more details refer to Appendix.
\item[iii)] High probability bounds on error: With step size $\alpha^{(t)}=\frac{1}{\nu(t+1)}$, for any $1>\epsilon>0$ and $\forall t=1,2,\ldots,$ we have
\begin{align*}
\delta^{(t+1)} \leq \frac{\Lambda(\Phi',k_{lu})}{\nu^2}\frac{1+\log(t+1)}{t+1}+\frac{4Q(\Phi',k_{lu})}{\nu^2\sqrt{\epsilon}}\frac{\sqrt{(1+\log(t+1))^2+4}}{t+1} \nonumber
\end{align*}
with at least probability $1-\epsilon$.
\end{itemize}
\end{theorem}
The proof of Theorem~\ref{thm:Convergence_Results} and more discussion about the overall complexity of SGBP versus GBP can be found in Appendix.
\subsection{The Overall Complexity of SGBP vs. GBP}
Note that the complexity of ordinary parent-to-child algorithm is dominated by the highest-level ancestor regions with largest number of variables. Let us assume that we have $N$ highest-level ancestor regions, shown by $A_i$ for $1\leq i\leq N$. Therefore, the complexity of parents-to-child algorithm is of order ${\mc{O}}(|d|^{A_{\max}})$ where $A_{\max}=\max \big[|A|_1,\ldots,|A|_N \big]$.
If the conditions of Theorem~\ref{thm:mainresult} hold between region $A_{\max}$ and all of its children (if there is more than one region with size $|A_{\max}|$ this statement should hold for all of them), then the SGBP algorithm (Algorithm~\ref{alg:SGBP_Alg}) will reduce the overall complexity, otherwise our algorithm will not affect the dominant computation complexity; though it may reduce the update rule complexity over some of the other edges which are not dominant in terms of complexity.
Now, under the contractivity assumption of the global update function, it can be inferred that parents-to-child algorithm associated with each edge $(P\rightarrow R)$ demands $t=\mathcal{O}(\log(\frac{1}{\epsilon}))$ iteration to achieve $\epsilon$ precision, while according to Theorem~\ref{thm:Convergence_Results}, to get the same precision $\epsilon$ in SGBP algorithm, $t=\mathcal{O}(\frac{1}{\epsilon})$ iteration is needed. Nonetheless, using the Corollary~\ref{cor:orderreduction}, the computation complexity of the dominant update rule of SGBP algorithm is of order $\mc{O}(d^{(|A_{\max}|-\eta)}\frac{1}{\epsilon})$ in comparison to $\mc{O}(d^{|A_{\max}|}\log(\frac{1}{\epsilon}))$ for GBP algorithm where $\eta=\max_{C\in \mc{R}: (A_{\max},C)\in \mc{E}_r} \eta_{A_{\max}C}$. In particular, if $d>\exp\left( \frac{\log(1/\epsilon)}{\eta} \right) = \left(\frac{1}{\epsilon} \right)^{(1/\eta)}$ then SGBP leads to lower complexity than GBP to achieve the same error $\epsilon$.
\subsection{Simulation Results}
Considering a pairwise MRF, in this section we present some simulation results to study the impact of our algorithm along verifying our theoretical results. We choose the so-called Potts model (which is a generalization to Ising model; see \cite{felzenszwalb2006efficient}) of size $3\times 3$ for our simulation purpose.
We have the following potentials assigned to each of the edges $(u,v)\in \mc{E}$
\[
\psi_{uv}(i,j)=\left\{ \begin{array}{ll}
1 & \mathrm{if} \; i=j,\\
\gamma & \mathrm{Otherwise}.
\end{array}\right.
\]
where $0<\gamma<1$. For the nodes' potential we have
\[
\phi_{u}(i)=\left\{ \begin{array}{ll}
1 & \mathrm{if} \; i=1,\\
\mu+\sigma Y & \mathrm{Otherwise}.
\end{array}\right.
\]
in which $\sigma$ and $\mu$ meet the conditions $0<\sigma\leq \mu$ and $\sigma+ \mu<1$ and $Y$ should have the uniform distribution in the span of $(-1,1)$ in addition to being independent from other parameters. We take the following steps to run our simulation. First, setting $\sigma=\mu=\gamma=0.1$, we run parent-to-child algorithm with region size of $4$
to get the asymptotic $m^*$. Second, with the same parameters and taking $\alpha^{(t)}=\frac{2}{(1+t)}$ for $d\in\{4,8,16,32\}$, we perform Algorithm~\ref{alg:SGBP_Alg} for the same region graph.
It is worth noting that to calculate $\frac{\mathbb{E}[\|\delta^{(t)}\|^2_2]}{\|\vect{m}^*\|^2_2}$, we run algorithm $20$ times and then average over error corresponding to each simulation.
As it illustrated in the simulation result of Figure~\ref{fig:FinalResult}, this result is in consistency with the Theorem~\ref{thm:Convergence_Results}. Moreover, you can also observe the running time comparison between SGBP and GBP algorithm in Figure \ref{fig:RunningtimeResult}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.65]{final}
\end{center}
\caption{The normalized mean-squared error of SGBP versus number of iterations for a Potts models of size $3\times 3$.}
\label{fig:FinalResult}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.65]{ratio}
\end{center}
\caption{The running time comparison between SGBP and GBP algorithm with $d=4$ for a Potts models of size $3\times 3$.}
\label{fig:RunningtimeResult}
\end{figure}
\section{Examples}
\subsection{A General Example}\label{ex:genex}
As mentioned before we would like to emphasize that under assumptions of the Theorem~\ref{thm:mainresult}, the complexity gain of our algorithm depends on the graph topology and choice of regions.
\begin{figure}[h]
\centering
\includegraphics[scale=0.50]{Fig3}
\caption{Graph region of an arbitrary graph with the assumption $\vect{x}_P=\vect{x}_{P'}$.}
\label{figgencase1}
\end{figure}
To further clarity this idea, let's discuss the following belief equation update formulas. Moreover, we show that how we can reformulate the parent-to-child algorithm in terms of expectation. Considering Figure~\ref{figgencase1}, we have the following update rules:
\begin{align*}
m_{B\rightarrow R}(\vect{x}_R) &= c\sum_{\vect{x}_{B\setminus R}} \Phi_{B\setminus R}(\vect{x}_{B})\frac{1}{m_{C\rightarrow H}(\vect{x}_H)m_{F\rightarrow H}(\vect{x}_H)}\\
&= c\sum_{\vect{x}_{B\setminus R}} \Phi_{B\setminus R}(\vect{x}_{B})\hat{M}(\vect{x}_{T=H})\\
&=k(\vect{x}_{H\setminus(B\setminus R)}) \sum_{\vect{x}_{(B\setminus R) \setminus H}}\mathbb{E}_{{\vect{x}}_{H\cap(B\setminus R)}}[\Phi_{B\setminus R}(\vect{x}_{B}),\vect{x}_{\{B\setminus R\}}],
\end{align*}
\begin{align*}
m_{A\rightarrow R}(\vect{x}_R)
&=c\sum_{\vect{x}_{A\setminus R}} \Phi_{A\setminus R}(\vect{x}_{A}),
\end{align*}
\begin{align*}
m_{B\rightarrow C}(\vect{x}_C)
&=c\sum_{\vect{x}_{B\setminus C}} \Phi_{B\setminus C}(\vect{x}_{B})\frac{m_{A\rightarrow R}(\vect{x}_{R})}{m_{R\rightarrow E}(\vect{x}_{E})\times m_{D\rightarrow G}(\vect{x}_{E})}\\
&=c\sum_{\vect{x}_{B\setminus C}} \Phi_{B\setminus C}(\vect{x}_{B})\hat{M}(\vect{x}_{T=R})\\
&=c\sum_{\vect{x}_{B\setminus C}} \Phi_{B\setminus C}(\vect{x}_{B})\hat{M}({\vect{x}}_{R\cap (B\setminus C)},{\vect{x}}_{R\setminus (B\setminus C)})\\
&=k(\vect{x}_{R\setminus (B\setminus C)})\sum_{\vect{x}_{(B\setminus C)\setminus R}}\mathbb{E}_{{\vect{x}}_{R\cap (B\setminus C)}} \left[\Phi_{B\setminus C}(\vect{x}_{(B\setminus C)\cap R},{\vect{x}_{{(B\setminus C)\setminus R}}},{\vect{x}}_{B\setminus (B\setminus C)}) \right],
\end{align*}
and
\begin{align*}
m_{D\rightarrow G}(x_G)
&=c\sum_{\vect{x}_{D\setminus G}} \Phi_{D\setminus G}(x_{D})\frac{m_{R\rightarrow D}(\vect{x}_{D})}{m_{D\rightarrow G}(\vect{x}_{G})}\\
&=k(\vect{x}_{D\cap G})\sum_{\vect{x}_{D\setminus (D\setminus G)}} \mathbb{E}_{\vect{x}_{D\setminus G}}\left[ \Phi_{D\setminus G}(\vect{x}_{D}),{\vect{x}_{D\setminus (D\setminus G)}} \right].
\end{align*}
Furthermore we have,
\begin{align*}
m_{E\rightarrow G} &= c\sum_{\vect{x}_{E\setminus G}(\vect{x}_{E})} \Phi_{E\setminus G}(\vect{x}_{E})\left[\frac{m_{R\rightarrow E}(\vect{x}_{E})\times m_{C\rightarrow E}(\vect{x}_{E})\times m_{C\rightarrow H}(\vect{x}_{H})\times m_{F\rightarrow H}(\vect{x}_{H})}{m_{D\rightarrow G}(\vect{x}_{G})} \right]\\
&=c\sum_{\vect{x}_{E\setminus G}(\vect{x}_{E})} \Phi_{E\setminus G}(\vect{x}_{E})\hat{M}(\vect{x}_{T=E})
\end{align*}
Note that the updating rule equation for $m_{A\rightarrow R}$ cannot be rewritten in the form of an expected value due to contradicting the first condition of Theorem~\ref{thm:mainresult}.
\begin{remark}
Considering conditions of Theorem~\ref{thm:mainresult} for Figure \ref{figgencase1}, it should be noticed that because these conditions are satisfied for $m_{B\rightarrow R}$ and $m_{B\rightarrow C}$, SGBP does help reducing complexity. However, since the first and the second condition does not hold for $m_{D\rightarrow G}$ and $m_{A\rightarrow R}$, respectively, the proposed algorithm does not improve the computation complexity of message updates over these edges.
\hfill {\small $\blacksquare$}
\end{remark}
\subsection{An Example in Smaller Regions}
In the following, we present some examples in which the impact of our algorithm in reduction of the complexity of ordinary GBP is shown. Furthermore, {we use some Matrix representation} for following examples to illustrate our idea clearly.
\begin{example}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=2,very thick]
\node at (-1.7,5.5)(index1){};
\node at (-1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1l) {1};
\node at (-1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2l) {4};
\node at (-1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3l) {7};
\node at (-0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (-0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (0,5.5)(index2){};
\node at (0,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1) {2};
\node at (0,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2) {5};
\node at (0,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3) {8};
\node at (1.7,5.5)(index3){};
\node at (1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1r) {3};
\node at (1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2r) {6};
\node at (1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3r) {9};
\draw[thick] (node3l)--(node3)--(node3r)--(node2r)--(node1r)--(node1)--(node1l)--(node2l)--(node2)--(node2r);\draw[thick] (node2l)--(node3l);\draw[thick](node1)--(node2)--(node3);
\end{tikzpicture}
\caption{Basic clusters in 9 nodes grid}\label{jamdecfig}
\label{fig:exageneral}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.5,very thick]
\node at (-4,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1ll) {1245};
\node at (-4,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2ll) {25};
\node at (-2,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1l) {2356};
\node at (-2,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2l) {45};
\node at (0,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1) {4578};
\node at (0,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2) {56};
\node at (-1,3) [rectangle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3) {5};
\node at (2,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1r) {5689};
\node at (2,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2r) {58};
\draw (node1ll)--(node2ll)--(node3)--(node2l)--(node1ll);
\draw (node1l)--(node2ll)--(node3);\draw (node1l)--(node2)--(node3); \draw (node1)--(node2r)--(node3);\draw (node1)--(node2l);\draw (node1r)--(node2);\draw (node1r)--(node2r);
\end{tikzpicture}
\caption{Graph region of parents to child GBP algorithm}
\label{fig13.5general}
\end{center}
\end{figure}
Let's move on from these definition to consider an example in Parent-to-child generalized belief propagation algorithm and the feasibility of having low complexity stochastic one. Consider the following figure with basic clusters of four nodes, in which the belief equations for the shown region graph are:
\begin{align}
b_{1245}&=k[\phi_1\phi_2\phi_4\phi_5\psi_{12}\psi_{14}\psi_{25}\psi_{45}][m_{36\rightarrow 25}m_{78\rightarrow 45}m_{6\rightarrow 5}m_{8\rightarrow 5}]\nonumber\\
b_{2356}&=k[\phi_2\phi_3\phi_5\phi_6\psi_{25}\psi_{56}\psi_{36}\psi_{23}][m_{14\rightarrow 25}m_{89\rightarrow 56}m_{8\rightarrow 5}m_{4\rightarrow 5}]\nonumber\\
b_{25}&=k[\phi_2\phi_5\psi_{25}][m_{14\rightarrow 25}m_{36\rightarrow 25}m_{4\rightarrow 5}m_{6\rightarrow 5}m_{8\rightarrow 5}]\nonumber\\
b_{45}&=k[\phi_4\phi_5\psi_{45}][m_{12\rightarrow 45}m_{78\rightarrow 45}m_{2\rightarrow 5}m_{6\rightarrow 5}m_{8\rightarrow 5}]\nonumber\\
b_{5}&=k[\phi_{5}][m_{2\rightarrow 5}m_{4\rightarrow 5}m_{6\rightarrow 5}m_{8\rightarrow 5}]
\end{align}
Then recalling our definitions in previous sections, we can derive message updating rules as follows:
\begin{align}
m_{12\rightarrow 45}&=c\sum_{x_1,x_2\in{{\cal{X}}}}\phi_1(x_1)\phi_2(x_2)\psi_{12}(x_1,x_2)\psi_{14}(x_1,x_4)\psi_{25}(x_2,x_5)\frac{m_{36\rightarrow 25}(x_2,x_5)}{m_{2\rightarrow 5}(x_5)}\nonumber\\
&=c\sum_{x_1,x_2\in{{\cal{X}}}}\Phi_{1245}(x_1,x_2,x_4,x_5)\hat{M}_{T=(2,5)}(x_2,x_5)\nonumber\\
&=k_1(x_5)\sum_{x_1}\mathbb{E}_{x_2\sim\hat{M}_{T}}\Phi_{1245}(x_1,J,x_4,x_5)\label{eq1}\\
m_{14\rightarrow 25}&=c\sum_{x_1,x_4\in{{\cal{X}}}}\phi_1(x_1)\phi_4(x_4)\psi_{14}(x_1,x_4)\psi_{12}(x_1,x_2)\psi_{45}(x_4,x_5)\frac{m_{78\rightarrow 45}(x_4,x_5)}{m_{4\rightarrow 5}(x_5)}\nonumber\\
&=c\sum_{x_1,x_4\in{{\cal{X}}}}\phi_1(x_1)\phi_4(x_4)\psi_{14}(x_1,x_4)\psi_{12}(x_1,x_2)\psi_{45}(x_4,x_5)\frac{m_{78\rightarrow 45}(x_4,x_5)}{m_{4\rightarrow 5}(x_5)}\nonumber\\
&=c\sum_{x_1,x_4\in{{\cal{X}}}}\Phi_{1425}(x_1,x_2,x_4,x_5)\hat{M}_{T=(4,5)}(x_4,x_5)\nonumber\\
&=k_2(x_5)\sum_{x_1}\mathbb{E}_{x_4\sim\hat{M}_{T}}\Phi_{1425}(x_1,x_2,J,x_5)\label{eq2}
\end{align}
where $k_1(x_5)=c\sum_{x_2}\hat{M}(x_2,x_5)$, $k_2(x_5)=c\sum_{x_4}\hat{M}(x_4,x_5)$, \begin{align*}\Phi_{1245}(x_1,x_2,x_4,x_5)=&\phi_1(x_1)\phi_2(x_2)\psi_{12}(x_1,x_2)\psi_{14}(x_1,x_4)\psi_{25}(x_2,x_5)\\
\Phi_{1425}(x_1,x_2,x_4,x_5)=&\phi_1(x_1)\phi_4(x_4)\psi_{14}(x_1,x_4)\psi_{12}(x_1,x_2)\psi_{45}(x_4,x_5)
\end{align*} and defining
\begin{align*}
\Phi_{25}(x_2,x_5)=&\phi_2(x_2)\psi_{25}(x_2,x_5)\\
\Phi_{45}(x_4,x_5)=&\phi_4(x_4)\psi_{45}(x_4,x_5)
\end{align*}
we have
\begin{align}
\label{eq3} m_{2\rightarrow 5}=c\sum_{x_2\in{\cal{X}}}[\phi_2(x_2)\psi_{25}(x_2,x_5)]\big[{m_{14\rightarrow 25}(x_2,x_5)m_{36\rightarrow 25}(x_2,x_5)}\big]\\
\label{eq4} m_{4\rightarrow 5}=c\sum_{x_4\in{\cal{X}}}[\phi_4(x_4)\psi_{45}(x_4,x_5)]\big[{m_{12\rightarrow 45}(x_4,x_5)m_{76\rightarrow 45}(x_4,x_5)}\big]\\\nonumber
\end{align}
Note that the equations \ref{eq3}, \ref{eq4} do not meet the requirements of theorem \ref{thm:mainresult}, so we use them without any alteration. However, we cam apply stochastic updating with equations \ref{eq1}, \ref{eq2}.
Therefore, the distribution of random index generation associated with equations \ref{eq1} and \ref{eq2} is \begin{align}p(j|i)=\frac{\hat{M}(j,i)}{\sum_{j'}{\hat{M}(j',i)}}\label{2node}\end{align}
so we can introduce our algorithm now as follows:
\begin{itemize}
\item For the message from four-node clusters to two-node-cluster, pick two indexes $i$ and $j$ with probabilities according to distribution \ref{2node}. Then, the updating rules for these nodes is:
\begin{align}
m_{12\rightarrow 45}^{(t+1)}&=(1-\alpha^{(t)})m_{12\rightarrow 45}^{(t)}+\alpha^{(t)} k_1(x_5)\sum_{x_1}\Phi(x_1,J,x_4,x_5)\\
m_{14\rightarrow 25}^{(t+1)}&=(1-\alpha^{(t)})m_{14\rightarrow 25}^{(t)}+\alpha^{(t)} k_2(x_5)\sum_{x_1}\Phi(x_1,x_2,J,x_5)
\end{align} where $\alpha^{(t)}=\mathcal{O}(\frac{1}{t})$ is some step size.
\item For the message from middle nodes to the node $5$, we do the same as following:
\begin{align}
m_{2\rightarrow 5}^{(t+1)}&=(1-\alpha^{(t)})m_{2\rightarrow 5}^{(t)}+\alpha^{(t)} c\:\Phi(J_{25},x_5)\\
m_{4\rightarrow 5}^{(t+1)}&=(1-\alpha^{(t)})m_{4\rightarrow 5}^{(t)}+\alpha^{(t)} c\:\Phi(J_{45},x_5)
\end{align}
\item So it can be seen that the complexity of calculating the distribution and the updating rule is of ${\cal{O}}(d^2)$ and ${\cal{O}}(d^3)$ respectively, which is less than ${\cal{O}}(d^4)$, the complexity of ordinary GBP.
\end{itemize}
\end{example}
\section{Introduction}
Graphical models and corresponding message-passing algorithms have attracted a great amount of attention due to their wide-spreading application in many fields, including signal processing, machine learning, channel and source coding, computer vision, decision making, and game theory ({e.g.}, see \cite{kschischang2001factor, loeliger2004introduction}).
Finding marginal and mode of a probability distribution are two basic problems encountered in the field of graphical models. Taking the rudimentary approach, the marginalization problem has exponentially growing complexity in alphabet size. However, using BP algorithm (firstly introduced in \cite{pearl1988probabilistic}) to solve this problem either exactly or approximately, we can reduce the computational complexity to a significant degree. It has been proved that applying BP on graphical models without cycles provides exact solution to the marginalization problem.
Furthermore, it has been observed that for general graphs, BP can find good approximations for marginalization (or finding mode) problems, \cite{kschischang2001factor, loeliger2004introduction}.
Although BP has many favourable properties, it suffers from some limiting drawbacks. First, in complex and densely interconnected graphs, BP may not be able to produce accurate results; and even worse, it may not converge at all. Second, since in many applications ({e.g.}, decoding of error-correcting codes) messages are of high dimensions, the computational complexity of BP algorithm will highly increase which leads to slow convergence rates.
To deal with the first drawback, some works have been done to propose alternative algorithms ({e.g.}, see \cite{yuille2002cccp, welling2001belief, heskes2002stable, pakzad_estimation_2005}).
Specifically, to improve the accuracy of estimated marginal distribution, a generalization algorithm to BP has been introduced by Yedidia et al.~\cite{yedidia_constructing_2005}, known as Generalized Belief Propagation (GBP) algorithm. In their proposed algorithm, local computation is performed by a group of nodes instead of a single node as in BP. According to many empirical observations, GBP outperforms BP in many situations; \cite{yedidia2001characterization, harel2003poset, welling2004choice, sibel2014application}.
However, although GBP algorithm provides accurate results in terms of marginal distribution, it suffers from high order of computation complexity, specially in case of large alphabet size.
To overcome the second aforementioned deficiency of BP, lots of research have been conducted to reduce BP complexity for different applications ({e.g.}, refer to \cite{felzenszwalb2006efficient,sudderth2010nonparametric,mcauley2011faster, isard2009continuously, kersting2009counting, coughlan2007dynamic, arulampalam2002tutorial, smith2013sequential}).
In a recent work by Noorshams~et al.~\cite{noorshams_stochastic_2013}, to tackle with the challenge of high complexity in the case of large alphabet size, they introduce an alternative stochastic version of BP algorithm with lower complexity. The main idea behind their work is that each node sends a randomly sampled message taken from a properly chosen probability distribution instead of computing the exact message update rule in each iteration.
Motivated by \cite{noorshams_stochastic_2013} and in order to mitigate the computational complexity of GBP, we extend GBP and propose stochastic GBP (SGBP) algorithm. SGBP has the advantage of reducing the complexity, while increasing the accuracy of estimation.
In contrast to SBP, SGBP algorithm can reduce the computational complexity only if certain topological conditions are met by the region graph (defined later) associated to a graphical model. However, the complexity gain can be larger than only one order of magnitude in alphabet size. In this work, we characterize these conditions and the amount of computational gain that we can obtain by performing SGBP instead of GBP. Determining these criteria, we hope that they provide some useful guidelines on how to choose the regions and construct the region graph in a way that results to a lower complexity algorithm with good accuracy.
The rest of the paper is organized as follows.
First, \S\ref{sec:ProblemStatement} introduces our problem statement. In \S\ref{sec:SGBP_Alg}, we present the proposed stochastic GBP and then derive the topological conditions that guarantee SGBP has lower complexity than GBP.
Moreover, theoretical convergence results have been provided as well. Finally, to validate our theoretical results, considering a specific graphical model, SGBP is simulated and the results are presented.
\section{Problem Statement}
\label{sec:ProblemStatement}
\subsection{Notation}
In the following, we introduce the notation that will be used in the paper. The random variables are represented by upper case letters and their values by lower case letters. Vectors and matrices are determined by bold letters. Sometimes, we use calligraphic letters to denote sets. When we have a set of random variables $X_1,\ldots,X_n$, we write $\vect{X}_\set{A}$ to denote $(X_i,i\in\set{A})$. An undirected graph $G = (\set{V},\set{E})$ is defined by a set of nodes $\set{V}=\{1,2,\ldots,n\}$ and a set of edges $\set{E} \subseteq \set{V}\times \set{V}$, where $(u, v) \in \set{E}$ if and only if nodes $u$ and $v$ are connected. Similarly, we can define a directed graph.
For every function $f(x_1,x_2,\ldots,x_n)$ where $f: {\mc{X}}^n \mapsto \mbb{R}$, we define the operator $\cal{L}$ as a map that turns this function to a vector ${\cal{L}}(f)\in\mbb{R}_{ {|\mc{X}|}^n\times 1}$ by evaluating $f$ at every input point.
For instance, considering ${\cal{X}}_{\{1,2\}}\in\{0,1\}$, for $f(x_1,x_2)$ we have
\[
{\cal{L}}(f) = \begin{bmatrix}
f(0,0)\\
f(0,1)\\
f(1,0)\\
f(1,1)\\
\end{bmatrix}.
\]
\subsection{Graphical Model}
Undirected graphical models, also known as Markov random fields (MRF), is a way to represent the probabilistic dependencies among a set of random variables having Markov properties using an undirected graph.
More precisely, we say that a set of random variables $X_1,\ldots,X_n$ form an MRF if there exists a graph $G=(\set{V},\set{E})$, where each $X_i$ is associated to the node $i \in \set{V}=\{1,\ldots,n\}$, and edges of the graph $G$ encode Markov properties of the random vector $\vect{X}=(X_1,\ldots,X_n)$. These Markov properties are equivalent to a factorization of the joint distribution of random vector $\vect{X}$ over the cliques of graph $G$ \cite{grimmett2010probability}.
In this paper, we focus on discrete random variables case where for all $j\in \set{V}$ we have $X_j\in{\mc{X}}\triangleq\{1,2,\ldots,d\}$. Moreover, we assume that the distribution of $\vect{X}$ is factorized according to
\[
p(\vect{x}) = \frac{1}{Z}\prod_{a\in \set{F}} \phi_a(\vect{x}_a)
\]
where $\set{F}$ is a collection of subsets of $\set{V}$ and $Z$ is a constant called the partition function. For the factor functions $\phi_a$, we have also $\phi_a\ge 0$. This factorization can be represented by using a bipartite graph $G_f=(\set{V},\set{F},\set{E}_f)$ called factor graph. In this representation, the variable nodes $\set{V}$ correspond to random variables $X_i$'s and factor nodes $\set{F}$ determine the factor functions $\phi_a$'s. Moreover, there exists an edge $(i,a)\in\set{E}_f$ between a variable node $i$ and a factor node $a$ if the variable $x_i$ appears in the factor $\phi_a$ (for more information on factor graphs refer to \cite{kschischang2001factor}).
\subsection{Region Graph}
In order to present the Yedidia's parent-to-child algorithm \cite{yedidia_constructing_2005} as well as introducing our stochastic GBP algorithm, we need to state some definitions as follows.
\begin{definition}[see \cite{yedidia_constructing_2005}]
A \emph{region graph} $G_r=(\set{R},\set{E}_r)$ defined over a factor graph $G_f=(\set{V},\set{F},\set{E}_f)$ is a directed graph in which for each vertex $v\in\set{R}$ (corresponding to a region) we have $v\subseteq \set{V}\cup\set{F}$. Each region $v$ has this property that if a factor node $a\in\set{F}$ belongs to $v$ then all of its neighbouring variable nodes have to also belong to $v$.
A directed edge $(v_p \rightarrow v_c)\in\set{E}_r$
may exist if $v_c\subset v_p$. If such an edge exists, $v_p$ is a \emph{parent} of $v_c$, or equivalently, $v_c$ is a \emph{child} of $v_p$.
If there exists a directed path from $v_a$ to $v_d$ on $G_r$, we say that $v_a$ is an \emph{ancestor} of $v_d$ and $v_d$ is a \emph{descendant} of $v_a$.
\end{definition}
Now, for each $R\in \set{R}$, we let $\set{P}(R)$ denotes for the set of all parents of $R$, $\set{A}(R)$ denotes for the set of all ancestors of $R$ and $\set{D}(R)$ denotes for the set of all descendants of $R$.
Moreover, we define $\set{E}(R) \triangleq R \cup \set{D}(R)$. Finally, for a region $R\in\set{R}$, we use $|R|$ to denote for the number of variable nodes in $R$.
\subsection{Parent-to-child GBP algorithm}
We may derive the BP message-passing equations using the fact that the belief at each variable node is the product of all the incoming messages received from its neighbouring factor nodes. Additionally, the beliefs over the set of variable nodes connecting to a factor node $a\in\set{F}$ is the product of the factor function $\phi_a$ multiplied by the incoming messages to the factor node $a$. Now marginalizing the second set of beliefs to find the belief over a variable node and equate it to the belief of that variable node which is found directly using the first equation, we can recover the BP update rules.
\begin{figure}[h]
\centering
\includegraphics[scale= 0.4]{Fig7}
\caption{Graph region of an arbitrary graph corresponding to the parents-to-child algorithm.}
\label{fig:ProblemSetup11}
\end{figure}
Yedidia et al., generalize this idea in \cite{yedidia_constructing_2005}, proposing an algorithm called parents-to-child GBP algorithm. As explained in \cite{yedidia_constructing_2005}, in the parent-to-child algorithm, we have only one kind of message $m_{P\rightarrow R}(\vect{x}_R)$ from a parent region $P$ to a child region $R$. Then for the belief of region $R\in\set{R}$ we have
\begin{align}\label{eq:GBP_RegionBeliefComputation}
b_R(\vect{x}_R) \propto \Phi_{R}(\vect{x}_R) \times \prod_{P\in {\cal{P}}(R)}m_{P\rightarrow R}(\vect{x}_{R})
\times \prod_{D\in {\cal{D}}(R)}\prod_{P'\in{\cal{P}}(D)\setminus \cal{E}(R)}m_{P'\rightarrow D}(\vect{x}_D)
\end{align}
where $\Phi_{R}(\vect{x}_R) \triangleq \prod_{a\in R} \phi_a(\vect{x}_a)$ (with an abuse of notation when we product over $a\in R$ we mean to product only over the factor indexes of $R$).
Then, the message update rule over each edge $(P, R)\in\set{E}_r$ follows by
\begin{align}
m_{P\rightarrow R}(\vect{x}_{R}) =& \frac{\sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\prod_{(I,J)\in N(P,R)}m_{I\rightarrow J}(\vect{x}_{J})}{\prod_{(I,J)\in D(P,R)}m_{I\rightarrow J}(\vect{x}_{J})} \label{eq:UpdateRule_FirstEquation}\\
=& \sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\hat{M}(\vect{x}_{T_{PR}})\label{eq:TGBP}
\end{align}
where $\Phi_{P\setminus R}(\vect{x}_{P'})\triangleq \frac{\Phi_{P}}{\Phi_{R}}(\vect{x}_{P'})$ and ${P'}$ is the set of all variables appear in $\frac{\Phi_{P}}{\Phi_{R}}(\vect{x}_{P'})$. In addition, we have also
\begin{equation}
N(P,R) \triangleq \Big\{(I,J) | (I, J)\in \set{E}_r, I\notin \set{E}(P), J\in \set{E}(P) \setminus \set{E}(R) \Big\} \label{eq:Def_Set_N(P,R)}
\end{equation}
and
\begin{equation}
D(P,R) \triangleq \Big\{(I,J) | (I, J)\in \set{E}_r, I\in \set{D}(P) \setminus \set{E}(R), J\in \set{E}(R) \Big\}. \label{eq:Def_Set_D(P,R)}
\end{equation}
Notice that the sets $N(P,R)$ and $D(P,R)$ can be calculated in advance. Moreover, $\hat{M}(\vect{x}_{T_{PR}})$ in \eqref{eq:TGBP} is defined as follows
\[
\hat{M}(\vect{x}_{T_{PR}})\triangleq\frac{\prod_{(I,J)\in N(P,R)}m_{I\rightarrow J}(\vect{x}_{J})}{\prod_{(I,J)\in D(P,R)}m_{I\rightarrow J}(\vect{x}_{J})},
\]
where ${T_{PR}}$ is the set of all variables that appear in the above ratio.
\begin{remark}
It can be easily observed that depending on the graph topology and the choice of regions, we may have either $P'\subset P$ or $P'=P$ in \eqref{eq:TGBP}.
For example, consider two pairwise Markov Random Fields presented in Figures \ref{fig:ProblemSetup-smallerp} and \ref{fig:biggerp}.
Considering $P=\{1,2,4,5,7,8\}$ and $R=\{2,5,8\}$ in Figure~\ref{fig:ProblemSetup-smallerp}, we have $\Phi_{(124578\setminus 258)}(\vect{x}_{P'}) = \phi_1\phi_4\phi_7\psi_{12}\psi_{14}\psi_{74}\psi_{78}$
which leads to $P'=\{1,2,4,7,8\} \subset P$.
On the other hand, choosing $P=\{1,2,4,5\}$ and $R=\{2,5\}$ in Figure~\ref{fig:biggerp}, we have $\Phi_{(1245\setminus 25)}(\vect{x}_{P'}) = \phi_1\phi_4\psi_{12}\psi_{14}\psi_{45}$. Hence, $P'=\{1,2,4,5\}=P$.
\hfill {\small $\blacksquare$}
\end{remark}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.8,very thick]
\node at (-1.7,5.5)(index1){};
\node at (-1.7,5) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node1l) {1};
\node at (-1.7,4) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node2l) {4};
\node at (-1.7,3) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node3l) {7};
\node at (-0.85,4)[circle,dashed,draw=green!200,minimum size=65mm] (node3x){};
\node at (+0.85,4)[circle,dashed,draw=green!200,minimum size=65mm] (node3x){};
\node at (0,5.5)(index2){};
\node at (0,5) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node1) {2};
\node at (0,4) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node2) {5};
\node at (0,3) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node3) {8};
\node at (1.7,5.5)(index3){};
\node at (1.7,5) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node1r) {3};
\node at (1.7,4) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node2r) {6};
\node at (1.7,3) [circle,draw=black!50,fill=red!20!white,inner sep=8pt, minimum size=8mm] (node3r) {9};
\draw[thick] (node3l)--(node3)--(node3r)--(node2r)--(node1r)--(node1)--(node1l)--(node2l);\draw[thick] (node2l)--(node3l);\draw[thick](node1)--(node2)--(node3);
\end{tikzpicture}
\caption{Graph region of a graph with $x_{P'}\subset x_{P}$.}
\label{fig:ProblemSetup-smallerp}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.2,very thick]
\node at (-2.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1l) {124578};
\node at (0,3) [rectangle,draw=black!50,fill=red!20!white,inner sep=10pt, minimum size=10mm] (node) {258};
\node at (2.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1r) {235689};
\draw[thick,->](node1l)--(node);\draw[thick,->](node1r)--(node);
\end{tikzpicture}
\caption{Graph region of a graph with $x_{P'}\subset x_{P}$.}
\label{fig:ProblemSetup-smallerreg}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=2,very thick]
\node at (-1.7,5.5)(index1){};
\node at (-1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1l) {1};
\node at (-1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2l) {4};
\node at (-1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3l) {7};
\node at (-0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (-0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (0,5.5)(index2){};
\node at (0,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1) {2};
\node at (0,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2) {5};
\node at (0,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3) {8};
\node at (1.7,5.5)(index3){};
\node at (1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1r) {3};
\node at (1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2r) {6};
\node at (1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3r) {9};
\draw[thick] (node3l)--(node3)--(node3r)--(node2r)--(node1r)--(node1)--(node1l)--(node2l)--(node2)--(node2r);\draw[thick] (node2l)--(node3l);\draw[thick](node1)--(node2)--(node3);
\end{tikzpicture}
\caption{Basic clusters in 9 nodes grid with $x_{P'}=x_{P}$}\label{fig:biggerp}
\end{center}
\end{figure}
\begin{remark}
In the parent-to-child algorithm, the message transmitted over each edge $(P, R) \in \set{E}_r$ can be considered as a vector by applying the operator $\mc{L}(\cdot)$.
Namely, by concatenating all possible messages, we define $\vect{m}_{P\rightarrow R} \triangleq \mc{L}(m_{P\rightarrow R})$ where $\vect{m}_{P\rightarrow R} \in \mbb{R}^{d^{|R|}}$.
Moreover, concatenating all the messages over all edges of the region graph, we define $\vect{m} \triangleq \{\vect{m}_{P\rightarrow R}\}_{ (P, R)\in \set{E}_r} \in \mbb{R}^\Delta$ where $\Delta=\sum_{(P,R)\in \set{E}_r} d^{|R|}$.
\hfill {\small $\blacksquare$}
\end{remark}
Now, we can state the complexity of the parent-to-child GBP algorithm as stated in Lemma~\ref{lem:GBP_Complexity}.
\begin{lemma}\label{lem:GBP_Complexity}
The computation complexity of the message update rule of the parent-to-child GBP algorithm associated with each edge, computed according to \eqref{eq:TGBP}, is ${\mc{O}}(d^{|P|})$.
\end{lemma}
\begin{proof}
For each fixed vector $\vect{x}_R$, the calculation of $m_{P\rightarrow R}(\vect{x}_{R}) =\sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\hat{M}(\vect{x}_{T_{PR}})$
needs $d^{|P\setminus R|}$ operations. Moreover, to find $m_{P\rightarrow R}(\cdot)$ completely, one needs to evaluate the above summation $\mc{O}(d^{|R|})$ times.
Consequently, the overall complexity of calculating $m_{P\longrightarrow R}(\vect{x}_{R}) $ is of the order ${\mc{O}}(d^{|R|}\times d^{|P\setminus R|})={\cal{O}}(d^{|P|})$.
\end{proof}
At each round of the parent-to-child algorithm, $t=1,2,\ldots$, every parent node $P$ of $R$ in the region graph calculates a message $m^{(t+1)}_{P\rightarrow R}$ and sends it to node $R$. Mathematically, this can be written as (see \cite{yedidia_constructing_2005})
\begin{align}\label{eq:UpdateRuleExpansion1}
m^{(t+1)}_{P\rightarrow R} (\vect{x}_R) =& \left[ \Upsilon_{P\rightarrow R}(m^{(t)}) \right] (\vect{x}_R) \nonumber\\
& \hspace{-2cm} = \sum_{\vect{x}_{P\setminus R}}\Phi_{P\setminus R}(\vect{x}_{P'})\hat{M}^{(t)}(\vect{x}_{T_{PR}}) \nonumber\\
& \hspace{-2cm} = \sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} }} \sum_{\vect{x}_{(P\setminus R)\cap T_{PR} }} \Phi_{P\setminus R}(\vect{x}_{P'}) \hat{M}^{(t)}(\vect{x}_{ T_{PR}} ) \nonumber\\
& \hspace{-2cm} = k_{PR}^{(t)} \left( \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right) \hspace{-11pt} \sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} }} \sum_{\vect{x}_{(P\setminus R)\cap T_{PR} }} \Big[ \Phi_{P\setminus R}(\vect{x}_{P'}) \times Q^{(t)}(\vect{x}_{ T_{PR} \cap (P\setminus R)} | \vect{x}_{ T_{PR} \setminus (P\setminus R) }) \Big],
\end{align}
where
\begin{align}\label{eq:Dist_Q_Def}
Q^{(t)}(\vect{x}_{ T_{PR} \cap (P\setminus R)} | \vect{x}_{ T_{PR} \setminus (P\setminus R) }) \triangleq\frac{\hat{M}^{(t)} \left( \vect{x}_{ T_{PR} \cap (P\setminus R) } , \vect{x}_{ T_{PR} \setminus (P\setminus R)} \right) }{\sum_{\vect{x}'_{ T_{PR} \cap (P\setminus R) }} \hat{M}^{(t)} \left( \vect{x}'_{ T_{PR} \cap (P\setminus R) }, \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right) }
\end{align}
is a conditional distribution. Moreover,
\[
k_{PR}^{(t)} \left( \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right) \triangleq \hspace{-12pt} \sum_{\vect{x}'_{ T_{PR} \cap (P\setminus R)}} \hspace{-12pt} \hat{M}^{(t)} \left( \vect{x}'_{ T_{PR} \cap (P\setminus R)} , \vect{x}_{ T_{PR} \setminus (P\setminus R) } \right).
\]
Hence, for the update rule we can write
\begin{align}
m^{(t+1)}_{P\rightarrow R} (\vect{x}_R) = k_{PR}^{(t)}
\hspace{-10pt} \sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} } } \hspace{-10pt} \mbb{E}_{[\vect{X}_{(P\setminus R)\cap T_{PR} } \sim Q^{(t)} ] } \big[\Phi_{P\setminus R}(\vect{X}_{P'})\big] \label{eq:UpdateRuleExpansion2}
\end{align}
Here and in the following, for brevity and clarity of notation, we will omit the dependence of $k_{PR}^{(t)}$ to the variables $\vect{x}_{ T_{PR} \setminus (P\setminus R) }$.
Now, notice that we can decompose the set $P'$ as follows
\[
P' = \left[ (P\setminus R)\cap T_{PR} \right] \cup \left[ (P\setminus R)\setminus T_{PR} \right] \cup \left[P' \setminus (P\setminus R) \right]
\]
because we always have $P\setminus R \subseteq P'$. By using this relation, we can rewrite \eqref{eq:UpdateRuleExpansion2} as follows
\begin{align} \label{eq:UpdateRuleExpression_Final}
m^{(t+1)}_{P\rightarrow R} (\vect{x}_R) = k_{PR}^{(t)}
\sum_{\vect{x}_{(P\setminus R) \setminus T_{PR} } }
\mbb{E}_{ [\vect{X}_{(P\setminus R)\cap T_{PR} } \sim Q^{(t)} ] } \bigg[
\Phi_{P\setminus R} \Big(\vect{X}_{(P\setminus R)\cap T_{PR} }, \vect{x}_{(P\setminus R)\setminus T_{PR} }, \vect{x}_{P'\setminus (P\setminus R)} \Big) \bigg].
\end{align}
In \eqref{eq:UpdateRuleExpansion1}, $\Upsilon_{P\rightarrow R}:\mbb{R}^{\Delta} \mapsto \mbb{R}^{d^{|R|}}$ is the local update function of the directed edge $(P, R)\in \set{E}_r$. By concatenating all the local update functions over the edges of the region graph, we can define the global update function as
\begin{equation}\label{eq:GlobalUpdateFuncDef}
\Upsilon(\vect{m})=\Big[ \Upsilon_{P\rightarrow R}(\vect{m}) : (P,R)\in \set{E}_r \Big]
\end{equation}
where $\Upsilon: \mbb{R}^\Delta \mapsto \mbb{R}^\Delta$.
The goal of the (parent-to-child) GBP algorithm is to find a fixed point $\vect{m}^*$ that satisfies $\Upsilon(\vect{m}^*)=\vect{m}^*$. If a fixed point $\vect{m}^*$ is found, then the beliefs of random variables in a region $R$ is computed by applying \eqref{eq:GBP_RegionBeliefComputation}.
\section{Stochastic Generalized Belief Propagation Algorithm}
\label{sec:SGBP_Alg}
In this section, first we introduce our stochastic extension to the parent-to-child GBP algorithm, and then present a result on the criteria where this algorithm is able to mitigate the computation complexity of GBP.
Based on \eqref{eq:UpdateRuleExpression_Final}, we introduce our algorithm as stated in Algorithm~\ref{alg:SGBP_Alg}. The main idea of the algorithm is that under proper conditions (that will state in Theorem~\ref{thm:mainresult}), some parts of the message update rule \eqref{eq:TGBP} for each edge of the region graph can be written as an expectation as stated in \eqref{eq:UpdateRuleExpression_Final}.
\begin{algorithm}
\caption{Stochastic Generalized Belief Propagation (SGBP) algorithm.}
\label{alg:SGBP_Alg}
\begin{algorithmic}[1]
\STATE \textbf{Initialize the messages.}
\FOR{$t\in\{1,2,\ldots\}$ and each directed edge $(P, R)\in\set{E}_r$}
\STATE Choose a random vector $\vect{J}_{PR}^{(t+1)} \in \mc{X}^{| T_{PR} \cap (P\setminus R)|}$ according to the conditional distribution $Q^{(t)}(\vect{x}_{ T_{PR} \cap (P\setminus R)} | \vect{x}_{ T_{PR} \setminus (P\setminus R) })$ defined in \eqref{eq:Dist_Q_Def}.
\STATE Update the message $m_{P\longrightarrow R}^{(t+1)}$ with the appropriately tuned step size $\alpha^{(t)}={\mathcal{O}}(\frac{1}{t})$ according to
\begin{align}\label{eq:UpdateRule_Alg}
m_{P\rightarrow R}^{(t+1)} (\vect{x}_R) =(1-\alpha^{(t)}) m_{P\rightarrow R}^{(t)} (\vect{x}_R) +\alpha^{(t)} k_{PR}^{(t)} (\vect{x}_{T_{PR}\setminus (P\setminus R)}) \hspace{-10pt} \sum_{\vect{x}_{(P\setminus R)\setminus T_{PR} }} \hspace{-14pt} \Phi_{P\setminus R} \left(\vect{J}_{PR}^{(t+1)}, \vect{x}_{(P\setminus R)\setminus T_{PR} }, \vect{x}_{P'\setminus (P\setminus R)} \right)
\end{align}
\STATE $t = t + 1$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{remark}\label{remrk:DeterministicUpdateRule}
Note that when $(P\setminus R)\cap T_{PR}=\varnothing$, the update rule \eqref{eq:UpdateRule_Alg} becomes deterministic as stated in the following
\[
m_{P\rightarrow R}^{(t+1)}=(1-\alpha^{(t)})m_{P\rightarrow R}^{(t)} +\alpha^{(t)} k_{PR}^{(t)} (\vect{x}_{T_{PR}}) \Big[\sum_{\vect{x}_{(P\setminus R)}}\Phi_{P\setminus R} \left(\vect{x}_{(P\setminus R) }, \vect{x}_{P'\setminus (P\setminus R)} \right)\Big].
\]
Later, we will prove in Lemma~\ref{lem:Graph-region property} that this condition can only happen in update rules corresponding to the highest-level ancestors regions.
\hfill {\small $\blacksquare$}
\end{remark}
In contrast to SPB studied in \cite{noorshams_stochastic_2013}, the stochastic version of GBP does not always reduce the computational complexity in each iteration. Theorem~\ref{thm:mainresult} describes the topological and regional conditions for which the complexity of SGBP is less than GBP for a specific edge of the region graph.
\begin{theorem}\label{thm:mainresult}
Our proposed algorithm that runs over a region graph $G_r$ reduces the computation complexity of each message $m_{P\rightarrow R}$ (compared to GBP) if and only if the following conditions hold
\begin{itemize}
\item[(i)] $(P\setminus R)\cap T_{PR} \neq\varnothing$
\item[(ii)] $(P\setminus R) \nsubseteq T_{PR}$
\end{itemize}
\end{theorem}
\begin{proof}
The main idea of the proof lies in the fact that whether or not \eqref{eq:TGBP} can be written in the form of an expected value of \emph{potential functions} as stated in \eqref{eq:UpdateRuleExpression_Final}. If this happens, as presented in Algorithm~\ref{alg:SGBP_Alg}, the complexity of update rules can be reduced. Now, to be able to have an expectation operation in \eqref{eq:UpdateRuleExpression_Final}, we should have $(P\setminus R)\cap T_{PR} \neq\varnothing$.
Now, assuming condition (i) holds, we find the complexity of Algorithm~\ref{alg:SGBP_Alg}'s update rule over every edge $(P, R)\in\set{E}_r$ in each iteration. First, let us fix $\vect{x}_R$. To find the PMF of the random vector $\vect{J}$ which is given by \eqref{eq:Dist_Q_Def},
we need ${\mc{O}}(d^{|\{P\setminus R\}\cap T_{PR} |}\times d^{| T_{PR} \setminus \{P\setminus R\}|}) = {\mc{O}}(d^{| T_{PR} |})$ operations. Notice that since we have $ T_{PR} \setminus (P\setminus R) \subseteq R$ and $[(P\setminus R)\cap T_{PR} ] \cap [(P\setminus R)\setminus T_{PR} ]=\varnothing$, for every fixed $\vect{x}_R$, the PMF of $\vect{J}$ does not depend on the vector $\vect{x}_{(P\setminus R)\setminus T_{PR} }$.
This means that for a fixed $\vect{x}_R$, to find the summation in \eqref{eq:UpdateRule_Alg}, the PMF of $\vect{J}$ should only computed once.
Hence, the overall complexity of update rule \eqref{eq:UpdateRule_Alg} becomes
\[
\mc{O} \left( d^{| T_{PR} |}+d^{|R|}\big[d^{|(P\setminus R)\cap T_{PR} |} + d^{|(P\setminus R)\setminus T_{PR} |}+ d^{|(P\setminus R)\cap T_{PR} |}\big] \right)
\]
where the terms in the brackets count for a fixed $\vect{x}_R$ the computation complexity of $k(\vect{x}_{ T_{PR} \setminus {P\setminus R}})$, of the summation in \eqref{eq:UpdateRule_Alg}, and of taking a sample vector $\vect{J}$ from the above PMF, respectively. The above relation can be rewritten as follows
\[
\mc{O} \left( \max\big[d^{| T_{PR} |}, d^{|R|+|(P\setminus R)\setminus T_{PR} |},d^{|R|+|(P\setminus R)\cap T_{PR} |}\big] \right)
\]
Now, we can conclude that if $T_{PR} \neq \varnothing$ and $(P\setminus R)\not\subset T_{PR}$ then we have
\[
\mc{O} \left( \max\big[d^{| T_{PR} |}, d^{|R|+|(P\setminus R)\setminus T_{PR} |},d^{|R|+|(P\setminus R)\cap T_{PR} |}\big] \right) < {\cal{O}}(d^{|P|}),
\]
where the right hand side is the computation complexity of the parent-to-child GBP algorithm derived in Lemma~\ref{lem:GBP_Complexity}. This completes the proof of theorem.
\end{proof}
\begin{corollary}\label{cor:orderreduction}
Assuming that the conditions of Theorem~\ref{thm:mainresult} hold and denoting
\[
\eta_{PR}\triangleq\max{\Big[ | T_{PR} |,|R|+|(P\setminus R)\setminus T_{PR} |, |R|+|(P\setminus R)\cap T_{PR} | \Big]},
\]
Algorithm~\ref{alg:SGBP_Alg} reduces the computation complexity of message $m_{P\rightarrow R}$ of the order
$\mathcal{O}(d^{|P|-\eta_{PR}})=\mathcal{O}(d^{I_{PR}})$
where $I_{PR} \triangleq |P|-\eta_{PR}$. Notice that $I_{PR}$ can be larger than $1$.
\end{corollary}
\begin{example}\label{ex:RegionGraph_with_I=2}
In this example, we provide a graph, drawn in Figure~\ref{fig:I=2}, in which SGBP reduces the complexity of updating rule for the edge $123456\rightarrow 36$ (see Figure~\ref{fig:graph region of I=2}) with two order of magnitude. The updating rule for $m_{123456\rightarrow36}$ is as follows
\begin{align*}
m_{123456\rightarrow36}=& \sum_{x_1x_2x_4x_5} \frac{\Phi_{123456}(x_1,x_2,x_3,x_4,x_5,x_6)}{\Phi_{36}(x_3,x_6)} \times \frac{m_{2478\rightarrow 24}(x_2,x_4)}{1}\\
=&\sum_{x_2x_4}\tilde{\Phi}(x_2,x_3,x_4,x_6) m_{2478\rightarrow 24}(x_2,x_4)\\
=& k\:\mathbb{E}_{(X_2,X_4)\sim Q} \left[ \tilde{\Phi}(X_2,x_3,X_4,x_6) \right]
\end{align*}
where in this example by applying SGBP we will get $I_{123456\rightarrow36}=2$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.45]{Fig10}
\caption{A graph in which the SGBP reduces the complexity with two order of magnitude in alphabet size (see Example~\ref{ex:RegionGraph_with_I=2}).}
\label{fig:I=2}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.2,very thick]
\node at (-1.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1l) {2478};
\node at (4,3) [rectangle,draw=black!50,fill=red!20!white,inner sep=10pt, minimum size=10mm] (node1rr) {36};
\node at (0.5,3) [rectangle,draw=black!50,fill=red!20!white,inner sep=10pt, minimum size=10mm] (node) {24};
\node at (2.5,5) [rectangle,draw=black!50,fill=green!20!white,inner sep=10pt, minimum size=10mm] (node1r) {123456};
\draw[thick,->](node1l)--(node);\draw[thick,->](node1r)--(node);
\draw[thick,->](node1r)--(node1rr);
\end{tikzpicture}
\caption{Associated graph region of a graphical model depicted in Figure~\ref{fig:I=2} with $I_{123456\rightarrow36}=2$ (see Example~\ref{ex:RegionGraph_with_I=2}).}
\label{fig:graph region of I=2}
\end{center}
\end{figure}
\hfill {\small $\blacksquare$}
\end{example}
\begin{corollary}
The complexity of the parent-to-child GBP algorithm is dominated by the computation complexity of message update rule of the highest level edges in the region graph $G_r$.
As a result, if the dominant message update rule that belongs to the highest-level ancestor regions with the largest size, satisfies the conditions of Theorem~\ref{thm:mainresult}, then no matter what are the complexity of other edges, Algorithm~\ref{alg:SGBP_Alg} will reduce the overall computation complexity of the parent-to-child GBP.
\end{corollary}
\subsection{Convergence Rate of SGBP Algorithm}
\label{sec:Convergence_SGBP}
In this section, we extend the convergence guarantees of \cite{noorshams_stochastic_2013} to SGBP. Our convergence theorem (Theorem~\ref{thm:Convergence_Results}) is based on imposing a sufficient condition similar to \cite{noorshams_stochastic_2013} that guarantees uniqueness and convergence of the parent-to-child GBP message updates. More precisely we assume that the global update function $\Upsilon(\cdot)$, defined in \eqref{eq:GlobalUpdateFuncDef}, is contractive, namely $\exists \nu, 0<\nu<2$ such that
\begin{align}
\|\Upsilon(\vect{m})-\Upsilon(\vect{m}')\|_2 \leq \left( 1-\frac{\nu}{2} \right) \| \vect{m}-\vect{m}' \|_2. \label{eq:ContractiveCond_Def}
\end{align}
Following similar proof technique to \cite{noorshams_stochastic_2013}, with some appropriate modifications, we can obtain the following results.
\begin{theorem}\label{thm:Convergence_Results}
Assume that, for a given region graph, the update function $\Upsilon$ is contractive with parameter $1-\frac{\nu}{2}$ as defined in \eqref{eq:ContractiveCond_Def}. Then, parent-to-child GBP has a unique fixed point $\vect{m}^*$ and the message sequence $\{\vect{m}^{(t)}_{P\rightarrow R}\}^{\infty}_{t=1}$ generated by the SGBP algorithm has the following properties:
\begin{itemize}
\item[i)] The result of SGBP is consistent with GBP, namely we have $\vect{m}^{(t)}\overset{\mathrm{a.s.}}{\longrightarrow}\vect{m}^*$ as $t\longrightarrow \infty$.
\item[ii)] Bounds on mean-squared error: Let us divide the fixed point message $\vect{m}^*$ into two parts, $\vect{m}^* = {(\vect{m}_{\mc{E}_{1}}^*,\vect{m}_{\mc{E}_{\sim 1}}^*)}$, where $\vect{m}^*_{\mc{E}_{r_1}}$ corresponds to those edges of the region graph that perform deterministic update rule (as stated in Algorithm~\ref{alg:SGBP_Alg} and Remark~\ref{remrk:DeterministicUpdateRule}), while $\vect{m}^*_{\mc{E}_{\sim 1}}$ corresponds to the edges that run the stochastic algorithm. In other words, $\mc{E}_{1}$ and $\mc{E}_{\sim 1}$ represents the edges in region graph in which the message updating rules are deterministic and stochastic respectively.
Choosing step size $\alpha^{(t)}=\frac{\alpha}{\nu(t+2)}$ for some fixed $1<\alpha<2$ and defining $\delta_i^{(t)}\triangleq \frac{\vect{m}_i^{(t)}-\vect{m}_i^*}{\|\vect{m}_i^*\|^2}$ for each $i\in\{1,\sim 1\}$, we have
\begin{align*}
\frac{\mathbb{E}[\|\delta^{(t)}\|^2_2]}{\|\vect{m}^*\|^2_2} &\leq \left(\frac{3^{\alpha}{\alpha}^2 \Lambda(\Phi',k_{lu})}{2^\alpha(\alpha-1)\nu^2}\right)\frac{1}{t}+\frac{\mathbb{E}[{\|{\delta}_{\mc{E}_{\sim 1}}^{(0)}\|}^2_2]}{\|\vect{m}_{\mc{E}_{\sim 1}}^*\|^2_2}\left(\frac{2}{t}\right)^\alpha \nonumber
\end{align*}
for all iteration $t=1,2,3,\ldots$ where $\Lambda(\Phi',k_{lu})$ is a constant which depends on some factor functions (through $\Phi'$) and some variable nodes (through $k_{lu}$). For more details refer to Appendix.
\item[iii)] High probability bounds on error: With step size $\alpha^{(t)}=\frac{1}{\nu(t+1)}$, for any $1>\epsilon>0$ and $\forall t=1,2,\ldots,$ we have
\begin{align*}
\delta^{(t+1)} \leq \frac{\Lambda(\Phi',k_{lu})}{\nu^2}\frac{1+\log(t+1)}{t+1}+\frac{4Q(\Phi',k_{lu})}{\nu^2\sqrt{\epsilon}}\frac{\sqrt{(1+\log(t+1))^2+4}}{t+1} \nonumber
\end{align*}
with at least probability $1-\epsilon$.
\end{itemize}
\end{theorem}
The proof of Theorem~\ref{thm:Convergence_Results} and more discussion about the overall complexity of SGBP versus GBP can be found in Appendix.
\subsection{The Overall Complexity of SGBP vs. GBP}
Note that the complexity of ordinary parent-to-child algorithm is dominated by the highest-level ancestor regions with largest number of variables. Let us assume that we have $N$ highest-level ancestor regions, shown by $A_i$ for $1\leq i\leq N$. Therefore, the complexity of parents-to-child algorithm is of order ${\mc{O}}(|d|^{A_{\max}})$ where $A_{\max}=\max \big[|A|_1,\ldots,|A|_N \big]$.
If the conditions of Theorem~\ref{thm:mainresult} hold between region $A_{\max}$ and all of its children (if there is more than one region with size $|A_{\max}|$ this statement should hold for all of them), then the SGBP algorithm (Algorithm~\ref{alg:SGBP_Alg}) will reduce the overall complexity, otherwise our algorithm will not affect the dominant computation complexity; though it may reduce the update rule complexity over some of the other edges which are not dominant in terms of complexity.
Now, under the contractivity assumption of the global update function, it can be inferred that parents-to-child algorithm associated with each edge $(P\rightarrow R)$ demands $t=\mathcal{O}(\log(\frac{1}{\epsilon}))$ iteration to achieve $\epsilon$ precision, while according to Theorem~\ref{thm:Convergence_Results}, to get the same precision $\epsilon$ in SGBP algorithm, $t=\mathcal{O}(\frac{1}{\epsilon})$ iteration is needed. Nonetheless, using the Corollary~\ref{cor:orderreduction}, the computation complexity of the dominant update rule of SGBP algorithm is of order $\mc{O}(d^{(|A_{\max}|-\eta)}\frac{1}{\epsilon})$ in comparison to $\mc{O}(d^{|A_{\max}|}\log(\frac{1}{\epsilon}))$ for GBP algorithm where $\eta=\max_{C\in \mc{R}: (A_{\max},C)\in \mc{E}_r} \eta_{A_{\max}C}$. In particular, if $d>\exp\left( \frac{\log(1/\epsilon)}{\eta} \right) = \left(\frac{1}{\epsilon} \right)^{(1/\eta)}$ then SGBP leads to lower complexity than GBP to achieve the same error $\epsilon$.
\subsection{Simulation Results}
Considering a pairwise MRF, in this section we present some simulation results to study the impact of our algorithm along verifying our theoretical results. We choose the so-called Potts model (which is a generalization to Ising model; see \cite{felzenszwalb2006efficient}) of size $3\times 3$ for our simulation purpose.
We have the following potentials assigned to each of the edges $(u,v)\in \mc{E}$
\[
\psi_{uv}(i,j)=\left\{ \begin{array}{ll}
1 & \mathrm{if} \; i=j,\\
\gamma & \mathrm{Otherwise}.
\end{array}\right.
\]
where $0<\gamma<1$. For the nodes' potential we have
\[
\phi_{u}(i)=\left\{ \begin{array}{ll}
1 & \mathrm{if} \; i=1,\\
\mu+\sigma Y & \mathrm{Otherwise}.
\end{array}\right.
\]
in which $\sigma$ and $\mu$ meet the conditions $0<\sigma\leq \mu$ and $\sigma+ \mu<1$ and $Y$ should have the uniform distribution in the span of $(-1,1)$ in addition to being independent from other parameters. We take the following steps to run our simulation. First, setting $\sigma=\mu=\gamma=0.1$, we run parent-to-child algorithm with region size of $4$
to get the asymptotic $m^*$. Second, with the same parameters and taking $\alpha^{(t)}=\frac{2}{(1+t)}$ for $d\in\{4,8,16,32\}$, we perform Algorithm~\ref{alg:SGBP_Alg} for the same region graph.
It is worth noting that to calculate $\frac{\mathbb{E}[\|\delta^{(t)}\|^2_2]}{\|\vect{m}^*\|^2_2}$, we run algorithm $20$ times and then average over error corresponding to each simulation.
As it illustrated in the simulation result of Figure~\ref{fig:FinalResult}, this result is in consistency with the Theorem~\ref{thm:Convergence_Results}. Moreover, you can also observe the running time comparison between SGBP and GBP algorithm in Figure \ref{fig:RunningtimeResult}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.65]{final}
\end{center}
\caption{The normalized mean-squared error of SGBP versus number of iterations for a Potts models of size $3\times 3$.}
\label{fig:FinalResult}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.65]{ratio}
\end{center}
\caption{The running time comparison between SGBP and GBP algorithm with $d=4$ for a Potts models of size $3\times 3$.}
\label{fig:RunningtimeResult}
\end{figure}
\section{Examples}
\subsection{A General Example}\label{ex:genex}
As mentioned before we would like to emphasize that under assumptions of the Theorem~\ref{thm:mainresult}, the complexity gain of our algorithm depends on the graph topology and choice of regions.
\begin{figure}[h]
\centering
\includegraphics[scale=0.50]{Fig3}
\caption{Graph region of an arbitrary graph with the assumption $\vect{x}_P=\vect{x}_{P'}$.}
\label{figgencase1}
\end{figure}
To further clarity this idea, let's discuss the following belief equation update formulas. Moreover, we show that how we can reformulate the parent-to-child algorithm in terms of expectation. Considering Figure~\ref{figgencase1}, we have the following update rules:
\begin{align*}
m_{B\rightarrow R}(\vect{x}_R) &= c\sum_{\vect{x}_{B\setminus R}} \Phi_{B\setminus R}(\vect{x}_{B})\frac{1}{m_{C\rightarrow H}(\vect{x}_H)m_{F\rightarrow H}(\vect{x}_H)}\\
&= c\sum_{\vect{x}_{B\setminus R}} \Phi_{B\setminus R}(\vect{x}_{B})\hat{M}(\vect{x}_{T=H})\\
&=k(\vect{x}_{H\setminus(B\setminus R)}) \sum_{\vect{x}_{(B\setminus R) \setminus H}}\mathbb{E}_{{\vect{x}}_{H\cap(B\setminus R)}}[\Phi_{B\setminus R}(\vect{x}_{B}),\vect{x}_{\{B\setminus R\}}],
\end{align*}
\begin{align*}
m_{A\rightarrow R}(\vect{x}_R)
&=c\sum_{\vect{x}_{A\setminus R}} \Phi_{A\setminus R}(\vect{x}_{A}),
\end{align*}
\begin{align*}
m_{B\rightarrow C}(\vect{x}_C)
&=c\sum_{\vect{x}_{B\setminus C}} \Phi_{B\setminus C}(\vect{x}_{B})\frac{m_{A\rightarrow R}(\vect{x}_{R})}{m_{R\rightarrow E}(\vect{x}_{E})\times m_{D\rightarrow G}(\vect{x}_{E})}\\
&=c\sum_{\vect{x}_{B\setminus C}} \Phi_{B\setminus C}(\vect{x}_{B})\hat{M}(\vect{x}_{T=R})\\
&=c\sum_{\vect{x}_{B\setminus C}} \Phi_{B\setminus C}(\vect{x}_{B})\hat{M}({\vect{x}}_{R\cap (B\setminus C)},{\vect{x}}_{R\setminus (B\setminus C)})\\
&=k(\vect{x}_{R\setminus (B\setminus C)})\sum_{\vect{x}_{(B\setminus C)\setminus R}}\mathbb{E}_{{\vect{x}}_{R\cap (B\setminus C)}} \left[\Phi_{B\setminus C}(\vect{x}_{(B\setminus C)\cap R},{\vect{x}_{{(B\setminus C)\setminus R}}},{\vect{x}}_{B\setminus (B\setminus C)}) \right],
\end{align*}
and
\begin{align*}
m_{D\rightarrow G}(x_G)
&=c\sum_{\vect{x}_{D\setminus G}} \Phi_{D\setminus G}(x_{D})\frac{m_{R\rightarrow D}(\vect{x}_{D})}{m_{D\rightarrow G}(\vect{x}_{G})}\\
&=k(\vect{x}_{D\cap G})\sum_{\vect{x}_{D\setminus (D\setminus G)}} \mathbb{E}_{\vect{x}_{D\setminus G}}\left[ \Phi_{D\setminus G}(\vect{x}_{D}),{\vect{x}_{D\setminus (D\setminus G)}} \right].
\end{align*}
Furthermore we have,
\begin{align*}
m_{E\rightarrow G} &= c\sum_{\vect{x}_{E\setminus G}(\vect{x}_{E})} \Phi_{E\setminus G}(\vect{x}_{E})\left[\frac{m_{R\rightarrow E}(\vect{x}_{E})\times m_{C\rightarrow E}(\vect{x}_{E})\times m_{C\rightarrow H}(\vect{x}_{H})\times m_{F\rightarrow H}(\vect{x}_{H})}{m_{D\rightarrow G}(\vect{x}_{G})} \right]\\
&=c\sum_{\vect{x}_{E\setminus G}(\vect{x}_{E})} \Phi_{E\setminus G}(\vect{x}_{E})\hat{M}(\vect{x}_{T=E})
\end{align*}
Note that the updating rule equation for $m_{A\rightarrow R}$ cannot be rewritten in the form of an expected value due to contradicting the first condition of Theorem~\ref{thm:mainresult}.
\begin{remark}
Considering conditions of Theorem~\ref{thm:mainresult} for Figure \ref{figgencase1}, it should be noticed that because these conditions are satisfied for $m_{B\rightarrow R}$ and $m_{B\rightarrow C}$, SGBP does help reducing complexity. However, since the first and the second condition does not hold for $m_{D\rightarrow G}$ and $m_{A\rightarrow R}$, respectively, the proposed algorithm does not improve the computation complexity of message updates over these edges.
\hfill {\small $\blacksquare$}
\end{remark}
\subsection{An Example in Smaller Regions}
In the following, we present some examples in which the impact of our algorithm in reduction of the complexity of ordinary GBP is shown. Furthermore, {we use some Matrix representation} for following examples to illustrate our idea clearly.
\begin{example}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=2,very thick]
\node at (-1.7,5.5)(index1){};
\node at (-1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1l) {1};
\node at (-1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2l) {4};
\node at (-1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3l) {7};
\node at (-0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (-0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,3.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (+0.85,4.5)[circle,dashed,draw=black!50,minimum size=55mm] (node3x){};
\node at (0,5.5)(index2){};
\node at (0,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1) {2};
\node at (0,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2) {5};
\node at (0,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3) {8};
\node at (1.7,5.5)(index3){};
\node at (1.7,5) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node1r) {3};
\node at (1.7,4) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node2r) {6};
\node at (1.7,3) [circle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3r) {9};
\draw[thick] (node3l)--(node3)--(node3r)--(node2r)--(node1r)--(node1)--(node1l)--(node2l)--(node2)--(node2r);\draw[thick] (node2l)--(node3l);\draw[thick](node1)--(node2)--(node3);
\end{tikzpicture}
\caption{Basic clusters in 9 nodes grid}\label{jamdecfig}
\label{fig:exageneral}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.5,very thick]
\node at (-4,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1ll) {1245};
\node at (-4,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2ll) {25};
\node at (-2,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1l) {2356};
\node at (-2,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2l) {45};
\node at (0,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1) {4578};
\node at (0,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2) {56};
\node at (-1,3) [rectangle,draw=black!50,fill=blue!20!white,inner sep=8pt, minimum size=8mm] (node3) {5};
\node at (2,5) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node1r) {5689};
\node at (2,4) [rectangle,draw=black!50,fill=blue!20!white,inner sep=10pt, minimum size=10mm] (node2r) {58};
\draw (node1ll)--(node2ll)--(node3)--(node2l)--(node1ll);
\draw (node1l)--(node2ll)--(node3);\draw (node1l)--(node2)--(node3); \draw (node1)--(node2r)--(node3);\draw (node1)--(node2l);\draw (node1r)--(node2);\draw (node1r)--(node2r);
\end{tikzpicture}
\caption{Graph region of parents to child GBP algorithm}
\label{fig13.5general}
\end{center}
\end{figure}
Let's move on from these definition to consider an example in Parent-to-child generalized belief propagation algorithm and the feasibility of having low complexity stochastic one. Consider the following figure with basic clusters of four nodes, in which the belief equations for the shown region graph are:
\begin{align}
b_{1245}&=k[\phi_1\phi_2\phi_4\phi_5\psi_{12}\psi_{14}\psi_{25}\psi_{45}][m_{36\rightarrow 25}m_{78\rightarrow 45}m_{6\rightarrow 5}m_{8\rightarrow 5}]\nonumber\\
b_{2356}&=k[\phi_2\phi_3\phi_5\phi_6\psi_{25}\psi_{56}\psi_{36}\psi_{23}][m_{14\rightarrow 25}m_{89\rightarrow 56}m_{8\rightarrow 5}m_{4\rightarrow 5}]\nonumber\\
b_{25}&=k[\phi_2\phi_5\psi_{25}][m_{14\rightarrow 25}m_{36\rightarrow 25}m_{4\rightarrow 5}m_{6\rightarrow 5}m_{8\rightarrow 5}]\nonumber\\
b_{45}&=k[\phi_4\phi_5\psi_{45}][m_{12\rightarrow 45}m_{78\rightarrow 45}m_{2\rightarrow 5}m_{6\rightarrow 5}m_{8\rightarrow 5}]\nonumber\\
b_{5}&=k[\phi_{5}][m_{2\rightarrow 5}m_{4\rightarrow 5}m_{6\rightarrow 5}m_{8\rightarrow 5}]
\end{align}
Then recalling our definitions in previous sections, we can derive message updating rules as follows:
\begin{align}
m_{12\rightarrow 45}&=c\sum_{x_1,x_2\in{{\cal{X}}}}\phi_1(x_1)\phi_2(x_2)\psi_{12}(x_1,x_2)\psi_{14}(x_1,x_4)\psi_{25}(x_2,x_5)\frac{m_{36\rightarrow 25}(x_2,x_5)}{m_{2\rightarrow 5}(x_5)}\nonumber\\
&=c\sum_{x_1,x_2\in{{\cal{X}}}}\Phi_{1245}(x_1,x_2,x_4,x_5)\hat{M}_{T=(2,5)}(x_2,x_5)\nonumber\\
&=k_1(x_5)\sum_{x_1}\mathbb{E}_{x_2\sim\hat{M}_{T}}\Phi_{1245}(x_1,J,x_4,x_5)\label{eq1}\\
m_{14\rightarrow 25}&=c\sum_{x_1,x_4\in{{\cal{X}}}}\phi_1(x_1)\phi_4(x_4)\psi_{14}(x_1,x_4)\psi_{12}(x_1,x_2)\psi_{45}(x_4,x_5)\frac{m_{78\rightarrow 45}(x_4,x_5)}{m_{4\rightarrow 5}(x_5)}\nonumber\\
&=c\sum_{x_1,x_4\in{{\cal{X}}}}\phi_1(x_1)\phi_4(x_4)\psi_{14}(x_1,x_4)\psi_{12}(x_1,x_2)\psi_{45}(x_4,x_5)\frac{m_{78\rightarrow 45}(x_4,x_5)}{m_{4\rightarrow 5}(x_5)}\nonumber\\
&=c\sum_{x_1,x_4\in{{\cal{X}}}}\Phi_{1425}(x_1,x_2,x_4,x_5)\hat{M}_{T=(4,5)}(x_4,x_5)\nonumber\\
&=k_2(x_5)\sum_{x_1}\mathbb{E}_{x_4\sim\hat{M}_{T}}\Phi_{1425}(x_1,x_2,J,x_5)\label{eq2}
\end{align}
where $k_1(x_5)=c\sum_{x_2}\hat{M}(x_2,x_5)$, $k_2(x_5)=c\sum_{x_4}\hat{M}(x_4,x_5)$, \begin{align*}\Phi_{1245}(x_1,x_2,x_4,x_5)=&\phi_1(x_1)\phi_2(x_2)\psi_{12}(x_1,x_2)\psi_{14}(x_1,x_4)\psi_{25}(x_2,x_5)\\
\Phi_{1425}(x_1,x_2,x_4,x_5)=&\phi_1(x_1)\phi_4(x_4)\psi_{14}(x_1,x_4)\psi_{12}(x_1,x_2)\psi_{45}(x_4,x_5)
\end{align*} and defining
\begin{align*}
\Phi_{25}(x_2,x_5)=&\phi_2(x_2)\psi_{25}(x_2,x_5)\\
\Phi_{45}(x_4,x_5)=&\phi_4(x_4)\psi_{45}(x_4,x_5)
\end{align*}
we have
\begin{align}
\label{eq3} m_{2\rightarrow 5}=c\sum_{x_2\in{\cal{X}}}[\phi_2(x_2)\psi_{25}(x_2,x_5)]\big[{m_{14\rightarrow 25}(x_2,x_5)m_{36\rightarrow 25}(x_2,x_5)}\big]\\
\label{eq4} m_{4\rightarrow 5}=c\sum_{x_4\in{\cal{X}}}[\phi_4(x_4)\psi_{45}(x_4,x_5)]\big[{m_{12\rightarrow 45}(x_4,x_5)m_{76\rightarrow 45}(x_4,x_5)}\big]\\\nonumber
\end{align}
Note that the equations \ref{eq3}, \ref{eq4} do not meet the requirements of theorem \ref{thm:mainresult}, so we use them without any alteration. However, we cam apply stochastic updating with equations \ref{eq1}, \ref{eq2}.
Therefore, the distribution of random index generation associated with equations \ref{eq1} and \ref{eq2} is \begin{align}p(j|i)=\frac{\hat{M}(j,i)}{\sum_{j'}{\hat{M}(j',i)}}\label{2node}\end{align}
so we can introduce our algorithm now as follows:
\begin{itemize}
\item For the message from four-node clusters to two-node-cluster, pick two indexes $i$ and $j$ with probabilities according to distribution \ref{2node}. Then, the updating rules for these nodes is:
\begin{align}
m_{12\rightarrow 45}^{(t+1)}&=(1-\alpha^{(t)})m_{12\rightarrow 45}^{(t)}+\alpha^{(t)} k_1(x_5)\sum_{x_1}\Phi(x_1,J,x_4,x_5)\\
m_{14\rightarrow 25}^{(t+1)}&=(1-\alpha^{(t)})m_{14\rightarrow 25}^{(t)}+\alpha^{(t)} k_2(x_5)\sum_{x_1}\Phi(x_1,x_2,J,x_5)
\end{align} where $\alpha^{(t)}=\mathcal{O}(\frac{1}{t})$ is some step size.
\item For the message from middle nodes to the node $5$, we do the same as following:
\begin{align}
m_{2\rightarrow 5}^{(t+1)}&=(1-\alpha^{(t)})m_{2\rightarrow 5}^{(t)}+\alpha^{(t)} c\:\Phi(J_{25},x_5)\\
m_{4\rightarrow 5}^{(t+1)}&=(1-\alpha^{(t)})m_{4\rightarrow 5}^{(t)}+\alpha^{(t)} c\:\Phi(J_{45},x_5)
\end{align}
\item So it can be seen that the complexity of calculating the distribution and the updating rule is of ${\cal{O}}(d^2)$ and ${\cal{O}}(d^3)$ respectively, which is less than ${\cal{O}}(d^4)$, the complexity of ordinary GBP.
\end{itemize}
\end{example}
|
\section{Introduction}\label{sec:intro}
Measurements at colliders such as the Tevatron and Large Hadron Collider (LHC) have a unique capacity to shed light upon the internal dynamics
of the proton and provide constraints upon proton PDFs~\cite{Rojo:2015acz}. However including large hadron collider datasets
in PDF fits can provide a significant challenge due to the large computational footprint of performing accurate
theoretical predictions over the many iterations required in a fitting procedure. In order to make the fullest use of current
and future LHC results, efficient strategies for the computation of these observables must therefore be employed.
The typical Monte Carlo software packages used to perform predictions for
hadron collider observables cannot be easily deployed in a PDF fit due the processing time
required to obtain accurate results (usually of the order of a few hours or more per data
point). To overcome such limitations, the typical strategy
adopted for fast cross section prediction relies on the precomputation
of the partonic hard cross sections in such a way that the standard numerical
convolution with any set of PDFs can be reliably approximated by means of
interpolation techniques.
Such interpolation strategies are implemented in the {\tt
APPLgrid}~\cite{Carli:2010rw} and {\tt
FastNLO}~\cite{Wobisch:2011ij} projects. For the computation of the
hard cross sections, these packages rely on external codes to which
they are interfaced by means of a suite of functions allowing for the
filling of PDF- and $\alpha_s$-independent look-up tables of
cross section weights. Monte Carlo programs such as {\tt MCFM}~\cite{Campbell:2010ff}
and {\tt NLOJet++}~\cite{Nagy:2003tz} have been interfaced directly to
{\tt APPLgrid}/{\tt FastNLO} and more recently de\-di\-ca\-ted
interfaces to automated general-purpose event
generators have been developed. The {\tt
aMCfast}~\cite{Bertone:2014zva} and {\tt
MCgrid}~\cite{DelDebbio:2013kxa} codes can generate interpolation grids in {\tt APPLgrid}/{\tt FastNLO} format
by extracting the relevant information from the {\tt
MadGraph5\_aMC@NLO}~\cite{Alwall:2014hca} and {\tt
SHERPA}~\cite{Gleisberg:2008ta} event generators respectively.
While these tools have proven to be invaluable in the extraction
of parton densities, the volume of experimental data made available by
LHC collaborations for use in PDF fits is already stretching the
capabilities of the typical fitting technology. A standard
global PDF fit may now include thousands of hadronic data points for which
predictions have to be computed thousands of times during the
minimisation process. As a consequence, performing these predictions
using the standard interpolating tools, $i.e.$
{\tt APPLgrid} and {\tt FastNLO}, starts to become prohibitively time-consuming.
For this reason a high-performance tool tailored specifically to the requirements of PDF analysis becomes increasingly
important.
The {\tt FastKernel} method was developed to
address this problem in the context of the NNPDF global analyses~\cite{Ball:2014uwa}. This method differs from the standard procedure
\`{a} la {\tt APPLgrid} or {\tt FastNLO} in that it maximises the
amount of information that is precomputed prior to fitting so as to
minimise the amount of operations required during the
fit. More specifically, the {\tt FastKernel} method relies on the
combination of precomputed hard cross sections with DGLAP evolution
kernels into a single look-up table, here called
a {\tt FastKernel} ({\tt FK}) table. In this way the prediction
for a given hadronic observable can be obtained by performing a simple matrix
product between the respective {\tt FK} table and PDFs evaluated directly at the fitting scale.
In this paper we present the {\tt APFELgrid} package, a public
implementation of the {\tt FastKernel} method in which the hard
partonic cross sections provided in an {\tt APPLgrid} look-up table
are combined with the DGLAP evolution kernels provided by the {\tt
APFEL}
package~\cite{Bertone:2013vaa}.
This paper proceeds as follows. In
Sect.~\ref{sec:FastKernel} we present the technical details of the
implementation of the {\tt FastKernel} method. This is followed in
Sect.~\ref{sec:benchmark} by a performance benchmark of the
{\tt APFELgrid} library and resulting {\tt FK} tables. Finally, in
Sect.~\ref{sec:conclusion} we summarise the results discussed in this
work.
\section{Interpolation tools for collider observables}\label{sec:FastKernel}
Hadron collider observables are typically computed in QCD by means of a double
convolution of parton densities with a hard scattering
cross section. Consider for example the calculation of a general
cross section $pp\to X$ with a set of PDFs $\{f\}$:
\begin{equation}\label{eq:hadconv}
\sigma_{pp\to X} =
\sum_{s}\sum_{p} \int dx_1\,dx_2\,
\hat{\sigma}^{(p)(s)}\,\alpha_s^{p+p_{\rm LO}}(Q^2) \,F^{(s)}(x_1,x_2, Q^2)\,,
\end{equation}
where $Q^2$ is the typical hard scale of the process, the index $s$
sums over the active partonic subprocesses in the calculation, $p$ sums over
the perturbative orders used in the expansion, $p_{\rm
LO}$ is the leading-order power of $\alpha_s$ for the process and
$\hat{\sigma}^{(p)(s)}$ is the N$^p$LO contribution to the
cross section for the partonic subprocess scattering $(s)\to
X$. $F^{(s)}$ represents the subprocess parton density:
\begin{equation}\label{eq:APPLsubproc}
F^{(s)}(x_1,x_2, Q^2) =\sum_{i,j} C^{(s)}_{ij}
f_i(x_{1},Q^2)f_j(x_{2},Q^2)\,,
\end{equation}
where the $C^{(s)}_{ij}$ matrix enumerates the combinations of PDFs
contributing to the $s$-th subprocess. The central observation of
tools such as {\tt APPLgrid} and {\tt FastNLO} is that the PDF and
$\alpha_S$ dependence may be factorised out of the convolution via
expansion over a set of interpolating functions, spanning $Q^2$ and
the two values of parton-$x$. For example one may represent the
subprocess PDFs and $\alpha_S$ in terms of Lagrange basis
polynomials $\mathcal{I}_\tau(Q^2)$, $\mathcal{I}_\alpha(x_1)$ and
$\mathcal{I}_\beta(x_2)$ as:
\begin{equation}\label{eq:interpolation}
\begin{array}{c}
\displaystyle \alpha_s^{p+p_{\rm LO}}(Q^2) \,F^{(s)}(x_1,x_2, Q^2)
=\\
\\
\displaystyle \sum_{\alpha,\beta,\tau} \alpha_s^{p+p_{\rm LO}}(Q_\tau^2)
\,F^{(s)}_{\alpha\beta , \tau}
\,\mathcal{I}_\tau(Q^2)\,\mathcal{I}_\alpha(x_1)
\,\mathcal{I}_\beta(x_2),
\end{array}
\end{equation}
where we use the shorthand $F^{(s)}_{\alpha\beta ,\tau} =
F^{(s)}(x_\alpha, x_\beta,Q_\tau^2)$. Using these expressions in the
double convolution of Eq.~(\ref{eq:hadconv}) one can finally obtain an
expression for the desired cross section which depends upon the
subprocess PDFs only through a simple product:
\begin{equation} \label{eq:applconv}
\sigma_{pp\to X} = \sum_p \sum_{s}\sum_{\alpha,\beta,\tau}
\alpha_s^{p+p_{\rm LO}}(Q^2_\tau)W_{\alpha\beta,\tau}^{(p)(s)} \, F_{\alpha\beta,\tau}^{(s)}\,,
\end{equation}
where
\begin{equation} \label{eq:applgrid}
W_{\alpha\beta,\tau}^{(p)(s)} = \mathcal{I}_\tau(Q^2)\int dx_1\,dx_2\,
\hat{\sigma}^{(p)(s)}\,\mathcal{I}_\alpha(x_1)
\,\mathcal{I}_\beta(x_2)\,,
\end{equation}
consists of the convolution of the hard cross section with the
interpolating polynomials. This information may be stored in a
precomputed look-up table. The final expression for the cross section
in Eq.~(\ref{eq:applconv}) is therefore a considerably simpler task to perform
inside a fit than the direct evaluation of the double convolution.
\subsection{The FastKernel method}
A number of tools ($e.g.$ {\tt
APFEL},
{\tt HOPPET}~\cite{Salam:2008qg} and {\tt QCDNUM}~\cite{Botje:2010ay})
are available which perform PDF evolution via an analogous
interpolation procedure. In such a way PDFs at a general scale $Q_\tau$ may be
expressed as a product of PDFs at some initial fitting scale $Q_0$ and an
\textit{evolution operator} obtained by the solution of the DGLAP equation.
\begin{equation}\label{eq:fastPDFfinal_recalled}
f_i(x_{\alpha},Q^2_\tau) = \sum_{k}
\sum_\beta A^\tau_{\alpha\beta, ik}\;
f_k(x_\beta,Q^2_0)\,,
\end{equation}
where latin indices run over PDF flavour, greek indices run over points in an initial-scale interpolating $x$-grid and
the evolution operator $A$ may be accessed directly in the {\tt APFEL} package. Given this operator, we may replace the
(general-scale) PDFs used in the subprocess parton density
Eq.~(\ref{eq:APPLsubproc}) with their equivalent expressions evaluated at the fitting scale as
\begin{equation}\label{eq:FK1}
\begin{array}{rcl}
F^{(s)}_{\alpha\beta,\tau} &=& \displaystyle \sum_{i,j} \sum_{k,l}
\sum_{\delta,\gamma} C^{(s)}_{ij}
\left[ A^\tau_{\alpha\delta ik}\;
f_k(x_\delta,Q^2_0) A^\tau_{\beta\gamma
jl}\; f_l(x_\gamma,Q^2_0) \right]\;\;\;
\\
\\
&=& \displaystyle \sum_{k,l}\sum_{\delta,\gamma}
\widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta}
f_k(x_\delta,Q^2_0) f_l(x_\gamma,Q^2_0)\,,
\end{array}
\end{equation}
with the object
\begin{equation}
\widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta} =
\sum_{i,j} C^{(s)}_{ij} A^\tau_{\alpha\delta ik}
A^\tau_{\beta\gamma jl}\,,
\end{equation}
combining the operations of subprocess density construction and PDF
evolution. Going further and using the expression for
subprocess parton densities in Eq.~(\ref{eq:FK1}) in the full
cross section calculation we obtain
\begin{equation}
\sigma_{pp\to X} = \sum_{k,l}\sum_{\delta,\gamma}\sum_p
\sum_{s} \sum_{\alpha,\beta}
\sum_{\tau}
\alpha_s^{p+p_{\rm LO}}(Q^2_\tau)W_{\alpha\beta,\tau}^{(p)(s)} \widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta}
f_k(x_\delta,Q^2_0) f_l(x_\gamma,Q^2_0)\,.
\end{equation}
Performing some further contractions it is possible to obtain an
extremely compact expression for the calculation of the cross section
in question, in terms of only the initial-scale PDFs and summing only
over the initial scale interpolating $x$-grid and the incoming parton
flavours:
\begin{equation}\label{eq:FK}
\sigma _{pp\to X} = \sum_{k,l}\sum_{\delta,\gamma}
\widetilde{W}_{kl,\delta\gamma} \,f_k(x_\delta,Q^2_0) f_l(x_\gamma,Q^2_0),
\end{equation}
where the object:
\begin{equation}\label{eq:FKTable}
\widetilde{W}_{kl,\delta\gamma} = \sum_p\sum_{s}\sum_{\alpha,\beta} \sum_{\tau}
\alpha_s^{p+p_{\rm LO}}(Q^2_\tau) W_{\alpha\beta,\tau}^{(p)(s)} \widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta}
\end{equation}
is referred to here as an {\tt FK} table, and combines the information
stored in {\tt APPLgrid}-style interpolated weight grids with
analogously interpolated DGLAP evolution operators. This combination
enables for a maximally efficient expression for the
calculation of observables at hadron colliders under PDF variation, without invoking
any additional approximation.
\subsection{Features and limitations of {\tt FK} tables}
The {\tt FK} product of Eq.~(\ref{eq:FK}) differs with respect to the
product in Eq.~(\ref{eq:applconv}) in several notable ways. Firstly
the typical {\tt APPLgrid} or {\tt FastNLO} products use as input PDFs at a
general scale, requiring that PDF evolution {\it e.g.}
Eq.~(\ref{eq:fastPDFfinal_recalled}) be performed for every variation of the PDFs
during the fit. In the {\tt FK} product
this evolution is pre-cached at the stage of {\tt FK} table
generation, requiring only initial-scale PDFs at the time of fitting.
This pre-caching of the evolution also removes the need to
sum over hard scale and perturbative order during the
fit, further reducing the number of operations required. As the {\tt FK}
product acts directly at the fitting scale, it benefits from the typically reduced number of active
partonic modes, with the sum over flavours in
Eq.~(\ref{eq:FK}) being limited to those directly parametrised in the
fit. Having reduced the calculation to such a simple form, it is also
straightforward to apply standard
computational tools such as multi-threading through e.g OpenMP or
Single Instruction Multiple Data (SIMD) operations such as SSE or AVX to
further reduce computational expense.
While these features provide significant performance enhancements, the
{\tt FK} table format is not suitable as a complete replacement for
tools such as {\tt APPLgrid}. The precomputation of the PDF evolution
necessarily means that all theory parameters such as perturbative
order, strong coupling and factorization/renormalization scales are
inextricably embedded in each {\tt FK} table. In order to perform PDF
fits including variations of these parameters, multiple {\tt FK}
tables must be computed, each with different theory settings. While
performing such a re-calculation directly from Monte Carlo codes would
be exceptionally time consuming, the data representation in {\tt
APPLgrid} files allows for an efficient (re)-combination with varying
theory parameters.
\section{Performance benchmarks}
\label{sec:benchmark}
We shall now examine the performance differences between
the two expressions for fast interpolated cross section prediction
Eq.~(\ref{eq:applconv}) ({\tt APPLgrid}) and Eq.~(\ref{eq:FK}) ({\tt
FK}).
In order to provide a comprehensive benchmark, we consider here a wide range of
processes including LHC and Tevatron electroweak vector boson
production
measurements~\cite{Aaij:2012mda,Aaij:2012vn,Chatrchyan:2013mza,Chatrchyan:2013uja,Chatrchyan:2012xt,Aad:2013iua,Aad:2011fp,Aad:2011dm,Aaltonen:2010zza},
$t\bar{t}$ total
cross sections~\cite{ATLAS:2012aa, Chatrchyan:2013faa,Chatrchyan:2012bra,Chatrchyan:2012ria},
double-differential Drell-Yan
cross sections~\cite{Chatrchyan:2013tia,CMS:2014jea} and inclusive jet
data~\cite{Chatrchyan:2012bja,Aad:2011fc,Aad:2013lpa,Abazov:2007jy}. Predictions
are performed over a wide range of kinematics, for a total of 52
source {\tt APPLgrid} files corresponding to the majority of available
LHC and Tevatron datasets applicable to PDF determination.
While the source {\tt APPLgrid} files have varying grid densities in
$x$ and $Q^2$, for the purposes of comparison the corresponding {\tt
FK} tables are produced consistently with 30 points in $x$, and at
an initial scale below the charm threshold, therefore with seven
active partonic species. These settings are chosen so as to provide a realistic
comparison, in a production environment the density and distribution of the
$x$-grid may be adjusted to match interpolation accuracy requirements. For these comparisons the {\tt FK}
table is stored as double-precision in memory for table generation and in
single-precision for the purposes of computing the {\tt FK} product.
In Fig.~\ref{fig:timings} we compare the average time taken per
datapoint for the {\tt FK} and {\tt APPLgrid} calculations, for all of
the 52 tables. We show timings for the {\tt FK} calculation in four
different configurations: AVX-OpenMP 2x (2 CPU cores), AVX, SSE3 and
the standard double precision product. Due to the inherent structural
differences between the {\tt FK} and {\tt APPLgrid} procedures,
results from the {\tt FK} calculation are systematically faster than
those from {\tt APPLgrid}. In particular, when comparing {\tt
FK} AVX-OpenMP 2x to {\tt APPLgrid} timings we obtain minimally a factor of
ten improvement in speed for electroweak vector boson production and a
maximum factor of 2000 improvement in predictions for inclusive jet
data. Across all processes and kinematic regions we observe
significant performance improvements from using the {\tt FK}
calculation even without the use of SIMD or multi-threading.
While sheer computational speed is typically the primary consideration
when com\-pu\-ting observables in a PDF fit, other factors such as
table size in the filesystem and memory, along with the computational
cost of pre-computing {\tt FK} tables must be considered. Indeed, the
computation of the {\tt FK} table in Eq.~(\ref{eq:FKTable}) requires a
great deal of operations which can be time consuming, particularly in
the case of source {\tt APPLgrid} files with very high interpolation
precision.
In Fig.~\ref{fig:performance} we examine the {\tt FK} table generation
time with {\tt APFELgrid}, {\tt FK} table file size and memory usage
of the {\tt FK} tables arising from the same source {\tt APPLgrid}
files as discussed in Fig.~\ref{fig:timings}. When examining the
table generation time per point, we observe timings from a few
milliseconds to 3.5 minutes per point, with differences arising from
the varying grid densities used in the input {\tt APPLgrid} files. In
terms of the grid size on disk, {\tt FK} tables are typically larger
than their corresponding {\tt APPLgrid} files, primarily as the {\tt
FK} file format is encoded in plain text for compatibility whereas
{\tt APPLgrid} files are expressed in binary as {\tt ROOT}
files. However when measuring the in-memory resident set size used by
the two procedures, the amount of system memory used by {\tt
FK} tables is systematically less than {\tt APPLgrid} files for all
processes considered here by at least 75\%. Note that this effect is
in part due to the differing precisions of the default representations.
\begin{figure}
\centering
\includegraphics[scale=0.6]{plots/t0a}
\caption{\small Performance comparisons between {\tt FK} with
AVX-OpenMP 2x, AVX, SSE3, double precision convolution and {\tt
APPLgrid} convolution time per point and process.}
\label{fig:timings}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.6]{plots/t0b}
\caption{\small {\tt APFELgrid} generation time per point and
comparison to {\tt APPLgrid} for grid size on disk and resident set size
(RSS).}
\label{fig:performance}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
In this work we have demonstrated that by employing the so-called {\tt
FastKernel} method, it is possible to convert an {\tt APPLgrid}
weight table into a derived format, referred to as an {\tt FK} table,
including the effects of PDF and $\alpha_s$ evolution. This procedure has been
implemented in the {\tt APFELgrid} package, supplied as a set of {\tt
C++} routines designed to supplement the PDF evolution library {\tt
APFEL} with {\tt FK} table generation capabilities.
The {\tt APFELgrid} package allows one to obtain a computationally
efficient expression for the calculation of hadronic cross sections,
in terms of only the initial-scale PDFs and summing only over the
initial scale interpolating $x$-grid and the incoming parton
flavours. The simple structure of the resulting
product makes {\tt FK} tables particularly suitable for the
efficient use of computational tools such as {\tt SIMD} and {\tt
OpenMP}.
We have shown that in several practical examples the numerical
evaluation of an {\tt FK} product is considerably faster than the
corresponding {\tt APPLgrid} product, even in the case where neither
SIMD or multi-threading are applied. {\tt FK} tables are supplied in a
simple plain-text format in order to simplify the construction of user
interfaces, and therefore are typically larger than
corresponding {\tt APPLgrids}. However we have shown that the
in-memory resident set sizes occupied by {\tt FK} tables are typically
smaller than those required by {\tt APPLgrids}, in our examples by at least 75\%.
The substantial speed improvement of {\tt FK} tables with respect to
{\tt APPLgrid} in association with the reduction in memory footprint
makes the {\tt APFELgrid} code a valuable tool for modern PDF fits
including large collider datasets.
The {\tt APFELgrid} package and associated documentation are publicly
available from the webpage:
\begin{center}
\url{https://github.com/nhartland/APFELgrid}
\end{center}
\section*{Acknowledgements}
The authors would like to thank members of the NNPDF Collaboration for their support and
motivation for this work; particularly Juan Rojo, Luigi del Debbio and Alberto Guffanti. We would also like
to thank Mark Sutton for helpful comments on the paper. V.~B. and N.~H. are
supported by an European Research Council Starting Grant ``PDF4BSM''.
S.~C. is supported by the HICCUP ERC Consolidator grant (614577).
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
Due to their usefulness in numerous areas of science~\cite{MR2662471,MR2059685, MR2778089}, engineering~\cite{5946867,MR2028016,MR2921716, MR2021601} and mathematics ~\cite{MR2149656, fickus2016tremain, MR1984549}, \textem{equiangular tight frames (ETFs)} are a class of frames that have received much attention in recent years ~\cite{MR3150919, MR2277977, MR2015832, MR2736147, MR2350682, MR3005267, Tremain_08, MR2235693, Zau, Fickus:2015aa}; however, ETFs rarely exist~\cite{szollHosi2014all}. Because a compactness argument shows that Grassmannian frames always exist ~\cite{MR1984549}, they often serve as ideal generalizations for ETFs in applications where low coherence is desired ~\cite{bgb15, MR1984549, bodmann2015achieving, MR2021601, MR3429342}.
A complex \textem{Grassmannian frame} is a set of $N$ unit vectors that spans $\mathbb C^M$ with the property that the maximal element of its \textem{angle set}, or set of pairwise absolute inner products, is minimal. A Grassmannian frame is \textem{$K$-angular} if the cardinality of its angle set is $K$ and it is \textem{tight} if it satisfies a scaled version of Parseval's identity.
Equiangular tight frames are $1$-angular Grassmannian frames characterized by achievement of the optimal lower bound of Welch~\cite{MR2059685, MR2711357}, but this is only possible if $N \leq M^2$ ~\cite{MR1984549, MR2021601}. When $N>M^2$, the orthoplex bound provides an alternative means for characterizing known examples of Grassmannian frames ~\cite{MR1418961, MR0074013, bodmann2015achieving}, but this bound can only be achieved if $N \leq 2(M^2-1)$ ~\cite{MR1418961, MR0074013, bodmann2015achieving}.
For the special case of unit-norm frames in $\mathbb C^2$, the isometric spherical embedding technique of Conway, Hardin, and Sloane~\cite{MR1418961} sends frame vectors to the unit sphere in $\mathbb R^3$. As has been previously shown in \cite{MR2882247}, leveraging T\'{o}th's spherical cap packing bound for the unit sphere in $\mathbb R^3$ ~\cite{fejes1943covering} yields a lower bound for optimal coherence that is stronger than the orthoplex bound whenever $N>6$ (Theorem~\ref{th_toth}), and this saturates for the case $N=12$, as exemplified by a tight Grassmannian frame consisting of $12$ vectors in $\mathbb C^2$ (Example~\ref{ex_12_2}). As with the cases of Grassmannian frames consisting of $N=4$ and $6$ vectors for $\mathbb C^2$, it is striking that this example also embeds perfectly into the vertices of a Platonic solid, an icosahedron in this case. Furthermore, we observe that this example generates a \textem{complex projective 5-design} and is thus relevant to combinatoral and quantum information literature.
Because Grassmannian frames that achieve the Welch bound are tight and have angle sets of minimal cardinality, it is natural to ask the following questions:
\begin{question}\label{q_gr_tight}
Is every Grassmannian frame tight?
\end{question}
\begin{question}\label{q_gr_angles}
If $\Phi$ is a Grassmannian frame not characterized by the Welch bound, can we infer anything about the cardinality of its angle set or its spectral properties?
\end{question}
In ~\cite{BK06}, the authors answered Question~\ref{q_gr_tight} in the negative for the real case by showing that Grassmannian frames consisting of $5$ vectors in $\mathbb R^3$ are always equiangular but never tight; furthermore, their result suggests that a plausible answer to Question~\ref{q_gr_angles} is that the cardinality of the angle set of a real Grassmannian frame should satisfy a minimality condition.
By considering two distinct examples of Grassmannian frames consisting of $5$ vectors in $\mathbb C^2$, we find the answers to these questions for the complex case to be more complicated than we anticipated. Strictly speaking, the answer to Question~\ref{q_gr_tight} for the complex case is also in the negative, because there exists a non-tight, $2$-angular Grassmannian frame with $5$ vectors for $\mathbb C^2$ (Example~\ref{ex_Gr_52_bi}) ; however, this question may be the wrong one to ask, because a tight, $3$-angular Grassmannian frame consisting of $5$ vectors over $\mathbb C^2$ also exists (Example~\ref{ex_Gr_52_tri}).
Nevertheless, we find that one still encounters a certain amount of rigidity if one stipulates tightness in the frame's design; in particular, we prove that every tight Grassmannian frame consisting of $5$ vectors in $\mathbb C^2$ must have an angle set with cardinality greater than $2$ (Theorem~\ref{th_5_2}).
The proof of this depends on basic properties about $2$-angular, tight frames (Proposition~\ref{prop_equi} and Theorem~\ref{th_btfs_odd}) and the observation that whenever the spherical embedding technique of ~\cite{MR1418961} is applied to a tight, unit-norm frame, then the embedded vectors in the higher-dimensional Euclidean sphere must sum to zero (Theorem~\ref{th_sph_emb_zero}).
In light of the coexistence of tight, $3$-angular and non-tight, $2$-angular Grassmannian frames in this scenario, we arrive at the following partial answer to the questions above. There exist cases where, for a given fixed number of vectors and fixed dimension, a Grassmannian frame may be constructed in multiple ways, where there is some trade-off between its spectral properties (tightness) and geometric properties (angle set).
The remainder of this paper is outlined as follows. In Section~\ref{sec:prelim}, we establish notation and terminology and collect a few basics facts about frame theory. In Section~\ref{sec:emb}, we recall the spherical embedding technique of ~\cite{MR1418961} and show that it embeds tight frames into zero-summing vectors. In Section~\ref{sec:bounds}, we recall the Welch bound and orthoplex bound, and we use the spherical embedding technique along with a result of T\'{o}th to improve upon these bounds for the case of $N>6$ vectors in $\mathbb C^2$. In this section, we also discuss several examples of Grassmannian frames that achieve these bounds. Finally, in Section~\ref{sec:gr_5_2}, we prove a few basic facts about $2$-angular tight frames and use these facts to prove that a tight Grassmannian frame consisting of $5$ vectors in $\mathbb C^2$ can never be $2$-angular.
\section{Preliminaries}\label{sec:prelim}
Let $\{e_j\}_{j=1}^M$ denote the canonical orthonormal basis for $\mathbb F^M$, where $\mathbb F = \mathbb R$ or $\mathbb C$ and let $I_M$ denote the $M \times M$ identity matrix. A set of vectors $\Phi = \{\phi_j\}_{j=1}^N \subset \mathbb F^M$ is a \textem{(finite) frame} if $\SPAN \{\phi_j\}_{j=1}^N = \mathbb F^M.$
It is often convenient to identify a frame $\Phi = \{\phi_j\}_{j=1}^N$ in terms of its \textem{synthesis matrix}
$$\Phi = \left[ \phi_1 \, \, \phi_2 \, \, ... \,\, \phi_N \right],$$
the $M \times N$ matrix with columns given by the \textem{frame vectors}. Just as we have written
$\Phi = \{\phi_j\}_{j=1}^N $
and
$\Phi = \left[ \phi_1 \, \, \phi_2 \, \, ... \,\, \phi_N \right]$ in the last sentence, we proceed with the tacit understanding that $\Phi$ is both a matrix and a set of vectors. Furthermore, we reserve the symbols $M$ and $N$ to refer to the dimension of the span of a frame and the cardinality of a frame, respectively.
A frame $\Phi = \{\phi_j\}_{j=1}^N$ is \textem{$A$-tight} if $\Phi \Phi^* = \sum_{j=1}^N \phi_j \phi_j^*= A I_M$ for some $A>0$ and it is \textem{unit-norm} if each frame vector has norm $\|\phi_j\|=1$. If $\Phi$ is unit-norm and $A$-tight, then $A=\frac{N}{M}$ because
$$N= \sum_{l=1}^M\sum_{j=1}^N |\langle e_l, \phi_j \rangle|^2 = \sum_{l=1}^M\sum_{j=1}^N \TR(\phi_j \phi_j^* e_l e_l^*) = A \sum_{l=1}^M \| e_l \|^2 = AM,$$
in which case also we have the identity
\begin{equation}\label{eq_tight_un}
\sum\limits_{l=1}^N | \langle \phi_j, \phi_l \rangle |^2 = \frac{N}{M},
\end{equation}
for every $j \in \{1,2,...,N \}$.
Given a unit-norm frame $\Phi = \{ \phi_j \}_{j=1}^N$, its \textem{frame angles} are the elements of the set
$$
A_\Phi : =\{ |\langle \phi_j, \phi_l \rangle | : j \neq l \},
$$ i.e. \textem{the angle set of $\Phi$}, and we say that $\Phi$ is \textem{$K$-angular} if $|A_\Phi| =K$ for some $K \in \mathbb N$. In the special cases that $K=1,2$ or $3$, we say that $\Phi$ is \textem{equiangular}, \textem{biangular} or \textem{triangular}, respectively.
If $\Phi = \{ \phi_j \}_{j=1}^N$ is $K$-angular with frame angles $c_1, c_2,...,c_K$, then we say that $\Phi$ is \textem{equidistributed} if there exist $m_1, m_2, ..., m_K \in \mathbb N$ such that
$$
\bigg| \Big\{ j' \in \{1,...,N\} : j' \neq j, |\langle \phi_j, \phi_{j'} \rangle | = c_k \Big\} \bigg| = m_k
$$
for every $j \in \{1,2,...,N\}$ and every $k \in \{1,2,...,K\}$. In this case, we call the positive integers $m_1,m_2,...,m_K$ the \textem{multiplicities} of $\Phi$ and remark that $\sum_{j=1}^K m_j= N-1$.
We let $\UNFs{F}$ denote the space of unit norm frames consisting of $N$ vectors in $\mathbb F^M$.
Given $\Phi = \{ \phi_j \}_{j=1}^N \in \UNFs{F}$, its \textem{coherence} is defined by
$$\mu(\Phi) = \max\limits_{j \neq l} |\langle \phi_j, \phi_l \rangle |$$
and we define the \textem{Grassmanian constant for the pair $(N,M)$} by
$$\Gnk{F} = \inf\limits_{{\Phi} \in \UNFs{F}} \max\limits_{j \neq l} |\langle \phi_j, \phi_l \rangle | .$$
We say that $\Phi \in \UNFs{F}$ is a \textem{Grassmannian frame} if
$$
\mu(\Phi) = \Gnk{F}.
$$
\section{Spherical Embedding}\label{sec:emb}
In ~\cite{MR1418961}, the authors observed that a unit-norm frame can be isometrically embedded into a sphere in some high dimensional real Hilbert space.
\begin{thm}\label{th_sph_emb}[Conway et al.,~\cite{MR1418961}]
Let $D=M^2-1$ if $\mathbb F = \mathbb C$ or let $D=\frac{(M+2)(M-1)}{2}$ if $\mathbb F = \mathbb R$. If $\Phi = \{\phi_j\}_{j=1}^N \in \UNFs{F}$, then the frame vectors can be isometrically embedded into the unit sphere in $\mathbb R^D$ via
$$\phi_j \mapsto y_j \in \mathbb R^{D}$$ such that, for all $j, l \in \{1,...,N\}$, we have
\begin{equation*}
| \langle \phi_j, \phi_l \rangle |^2 = \frac{1}{M} + \frac{M-1}{M} \langle y_j, y_l \rangle.
\end{equation*}
\end{thm}
\begin{proof}
Let $\Phi = \{\phi_j\}_{j=1}^N \in \UNFs{F}$. The frame vectors of $\Phi$ can be embedded into the "traceless" subspace of the $M \times M$ self-adjoint (or symmetric) matrices via the mapping
$$ \phi_j \mapsto \phi_j \phi_j^* - \frac{1}{M }I_M,$$
which is isomorphic to $\mathbb R^{D}$ by a dimension counting argument. In particular, these embedded vectors all lie on a sphere of radius $\sqrt{\frac{M-1}{M}}$, because the Hilbert Schmidt norm gives $\| \phi_j \phi_j^* - \frac{1}{M }I_M\|_{H.S.}^2 = \frac{M-1}{M}$ for every $j \in \{1,...,N\}$, and this embdedding is distance preserving because for $j \neq l$
\begin{align*}
\| \phi_j \phi_j^* - \phi_l \phi_l^*\|_{H.S.}^2
&= 2\left(1- tr( \phi_j \phi_j^* \phi_l \phi_l^* )\right) \\
&= 2\left(1- |\langle \phi_j, \phi_l \rangle|^2 )\right). \\
\end{align*}
By identifying $\phi_j \phi_j^* - \frac{1}{M }I_M$ and $\phi_j \phi_j^* - \frac{1}{M }I_M$ with vectors $\tilde y_j, \tilde y_l \in \mathbb R^{D}$ on a sphere of radius $\sqrt{\frac{M-1}{M}}$
and using that
$$
\|y_j - y_l\|^2 = 2\frac{M-1}{M} (1 - \langle \tilde y_j , \tilde y_l \rangle)
,
$$
we can rewrite this equation as
$$
|\langle \phi_ j, \phi_l \rangle|^2 = \frac{1}{M} + \frac{M-1}{M} \langle y_j , y_l \rangle,
$$
where $y_j$ and $y_l$ are the unit vectors in the direction of $\tilde y_j$ and $\tilde y_l$, respectively.
\end{proof}
We observe that whenever a unit-norm, tight frame is embedded into a higher dimensional sphere as above, then the embedded vectors are also zero-summing.
\begin{cor}\label{th_sph_emb_zero}
Let $D=M^2-1$ if $\mathbb F = \mathbb C$ or let $D=\frac{(M+2)(M-1)}{2}$ if $\mathbb F = \mathbb R$. If $\Phi = \{\phi_j\}_{j=1}^N \in \UNFs{F}$ is a tight frame, then the frame vectors can be isometrically embedded into the unit sphere in $\mathbb R^D$ via
$$\phi_j \mapsto y_j \in \mathbb R^{D}$$ such that, for all $j, l \in \{1,...,N\}$, we have
\begin{equation}\label{eq_sph_emb_angles}
| \langle \phi_j, \phi_l \rangle |^2 = \frac{1}{M} + \frac{M-1}{M} \langle y_j y_l \rangle
\end{equation}
and
\begin{equation}\label{eq_sph_emb_zero}
\sum\limits_{j=1}^N y_j = 0.
\end{equation}
\end{cor}
\begin{proof}
Let $\Phi = \{\phi_j\}_{j=1}^N \in \UNFs{F}$. As in the proof of Theorem~\ref{th_sph_emb}, we embed the frame vectors via
$$ \phi_j \mapsto \phi_j \phi_j^* - \frac{1}{M }I_M.$$
Summing over these matrices and using the tightness property, we have
$$
\sum\limits_{j=1}^N \left( \phi_j \phi_j^* - \frac{1}{M }I_M \right) = \frac{N}{M} I_M - \frac{N}{M} I_M= 0_{M},
$$
where $0_{M}$ denotes the $M \times M$ zero matrix. The claim then follows by mimicing the proof of Theorem~\ref{th_sph_emb}.
\end{proof}
\section{Lower Bounds for the Grassmannian constant}\label{sec:bounds}
The optimal lower bound for the Grassmannian constant is the Welch bound ~\cite{welch:bound},
$$
\Gnk{F} \ge \sqrt{\frac{N-M}{M(N-1)}}.
$$
A Grassmannian frame achieves this lower bound if and only if it is an equiangular, tight frame ~\cite{MR1984549, MR2711357}, but it is well-known that this cannot occur when $N>M^2$ if $\mathbb F = \mathbb C$ or $N>\frac{M(M+1)}{2}$ if $\mathbb F = \mathbb R$ ~\cite{MR1418961, MR2021601}.
By applying Theorem~\ref{th_sph_emb} to a sphere-packing result of Rankin ~\cite{MR0074013}, the authors of ~\cite{MR1984549} (see also ~\cite{henkel05, bodmann2015achieving}) extrapolated a lower bound for $\Gnk{F}$ that is stronger than the Welch bound whenever $N>M^2$ or $N> \frac{M(M+1)}{2}$ for $\mathbb F = \mathbb C$ or $\mathbb R$, respectively.
\begin{thm}\label{th_ortho} [Orthoplex bound, ~\cite{MR0074013,MR1984549,henkel05, bodmann2015achieving}]
Let $D=M^2-1$ if $\mathbb F = \mathbb C$ or let $D=\frac{(M+2)(M-1)}{2}$ if $\mathbb F = \mathbb R$.
If $N>D+1$, then
$$
\Gnk{F} \geq \frac{1}{\sqrt{M}} \, .
$$
\end{thm}
The following example comes from ~\cite{bodmann2015achieving}, where the authors constructed infinite families of tight, complex Grassmannian frames with coherence equal to the orthoplex bound, which they termed \textem{orthoplectic Grassmannian frames}. Of particular interest to us, it is a tight, triangular Grassmannian frame in $\Omega_{5,2} ( \mathbb C)$.
\begin{ex}\label{ex_Gr_52_tri}
Let $\omega = e^{2 \pi i /3}$ and let
$$
\Phi = \frac{1}{\sqrt{2}}
\left[
\begin{array}{ccccc}
\sqrt{2} & 0 & 1 & 1 & 1\\
0 & \sqrt{2} & 1 & \omega & \omega^2
\end{array}
\right].
$$
It is straightforward to check that $\Phi$ is tight and has frame angles $c_1 = \frac{1}{\sqrt{2}}, c_2 = \frac{1}{2}$ and $c_3 =0$, so it is triangular and it is a Grassmannian frame by Theorem~\ref{th_ortho}.
\end{ex}
The next example also achieves the orthoplex bound, so it is also a Grassmannian frame in
$\Omega_{5,2} ( \mathbb C)$; however, unlike the preceding example, this one is biangular but not tight.
\begin{ex}\label{ex_Gr_52_bi}
Let
$$
\Phi = \frac{1}{\sqrt{2}}
\left[
\begin{array}{ccccc}
\sqrt{2} & 1 & 1 & 1 & 1\\
0 & -1 & 1 & i & -i
\end{array}
\right].
$$
It is straightforward to check that $\Phi$ has frame angles $c_1 = \frac{1}{\sqrt{2}}$ and $c_2 = 0$, so it is biangular and it is a Grassmannian frame by Theorem~\ref{th_ortho}. However, $\Phi \Phi^* \neq \frac{5}{2} I_5$, so it is not tight.
\end{ex}
Just as the Welch bound cannot saturate when a frame's cardinality is too large, it also known that a Grassmannian frame's coherence must be greater than
the orthoplex bound when $N>2(M^2-1)$ in the complex case or $N>(M+2)(M-1)$ in the real case~\cite{MR0074013,MR1984549,henkel05, bodmann2015achieving}. Thus, the orthoplex bound can also be improved when there are too many vectors.
In the special case that $\mathbb F = \mathbb C$ and $M=2$, the embedding from Theorem~\ref{th_sph_emb} sends points from the sphere in $\mathbb C^2$ to points on the sphere in $\mathbb R^3$, so the classical spherical cap packing result of T\'{o}th ~\cite{fejes1943covering} leads to an improved lower bound for $\mu_{N,2} (\mathbb{C})$ when $N\geq 7$, as has been previously noted in \cite{MR2882247}.
\begin{thm}\label{th_toth}[T\'{o}th's Bound, ~\cite{fejes1943covering}]
Let $N\geq 3$. If $\{x_j\}_{j=1}^N$ is a set of unit vectors in $\mathbb R^3$, then
$$
\max\limits_{j \neq l} \langle x_j, x_l \rangle \geq \frac{1}{2} \csc^2\left(\frac{N \pi}{6(N-2)} \right) - 1
.
$$
\end{thm}
\begin{thm}\label{cor_toth}[\cite{MR2882247}]
If $N \geq 3$, then
$$
\mu_{N,2} (\mathbb C) \geq \frac{1}{2} \csc \left(\frac{N \pi}{6(N-2)} \right) .
$$
\end{thm}
\begin{proof}
Let $\Phi = \{ \phi_j \}_{j=1}^N \in \Omega_{N,2} (\mathbb C).$ By Theorem~\ref{th_sph_emb}, there exist points $\{y_j\}_{j=1}^N$ on the unit sphere in $\mathbb R^3$ such that
$$
| \langle \phi_j, \phi_l \rangle |^2 = \frac{1}{2} + \frac{1}{2} \langle y_j, y_l\rangle.
$$
The claim then follows from Theorem~\ref{th_toth} because
\begin{align*}
\max\limits_{j \neq l} | \langle \phi_j, \phi_l \rangle |^2
&= \frac{1}{2} + \frac{1}{2} \max\limits_{j \neq l} \langle y_j, y_l \rangle \\
&\geq \frac{1}{2} + \frac{1}{2} \left[ \frac{1}{2} \csc^2\left(\frac{N \pi}{6(N-2)} \right) - 1 \right]. \\
\end{align*}
\end{proof}
T\'{o}th's bound from Theorem~\ref{th_toth} saturates when $N=3,4,6,$ and $12$ ~\cite{fejes1943covering}. When $N=3$, the bound is obtained by the three vertices of an equilateral triangle centered at the origin in $\mathbb R^3$. For $N=4$, the bound is obtained by the vertices of a regular $3$-simplex (i.e. a tetrahedron) centered at the origin. For $N=6$, the bound is obtained by the vertices of an orthoplex (i.e. an $\ell^1$-ball or octahedron) centered at the origin and the the case $N=12$ corresponds to the twelve vertices of an icosahedron centered at the origin. Furthermore, for the cases $N=3,4,$ and $6$, it is known that there exist tight Grassmannian frames in $\Omega_{N,2} (\mathbb C)$ that not only achieve the lower bound in Theorem~\ref{cor_toth} but embed perfectly into the vertices of an equilateral triangle~\cite{MR1418961}, regular tetrahedron~\cite{MR2059685}, and regular octahedron~\cite{MR2337670}, respectively. Next, we exhibit an example of a tight Grassmannian frame in $\Omega_{12,2} (\mathbb C)$ that embeds perfectly into the vertices of a regular icosahedron.
\begin{ex}\label{ex_12_2} [Icosaplectic Grassmannian Frame]
Let $a=\sqrt{\frac{5+\sqrt{5} }{10}}$ and $b=\sqrt{1 -a^2}$ and let $\eta = e^{2 \pi i /5}$ and $\omega = e^{2 \pi i/10}$ be the primitive $5$th and $10$th roots of unity, respectively.
A straightforward computation shows that
$$
\Phi =
\left[
\begin{array}{cccccccccccc}
\scriptscriptstyle 1 & \scriptscriptstyle 0 & \scriptscriptstyle -b & \scriptscriptstyle -b & \scriptscriptstyle -b & \scriptscriptstyle -b & \scriptscriptstyle -b& \scriptscriptstyle a \scriptscriptstyle \omega & \scriptscriptstyle a \scriptscriptstyle \omega \scriptscriptstyle \eta & \scriptscriptstyle a \scriptscriptstyle \omega \scriptscriptstyle \eta^2 & \scriptscriptstyle a \scriptscriptstyle \omega \scriptscriptstyle \eta^3 &\scriptscriptstyle a \scriptscriptstyle \omega \scriptscriptstyle \eta^4 \\
\scriptscriptstyle 0 &\scriptscriptstyle 1 &
\scriptscriptstyle -a \scriptscriptstyle \omega & \scriptscriptstyle -a \scriptscriptstyle \omega \scriptscriptstyle \eta & \scriptscriptstyle -a \scriptscriptstyle \omega \scriptscriptstyle \eta^2 & \scriptscriptstyle -a \scriptscriptstyle\omega \scriptscriptstyle\eta^3 &\scriptscriptstyle -a \scriptscriptstyle\omega \scriptscriptstyle\eta^4 &
\scriptscriptstyle b &\scriptscriptstyle b &\scriptscriptstyle b &\scriptscriptstyle b &\scriptscriptstyle b
\end{array}
\right]
$$
is an equidistributed, triangular, tight frame in $\Omega_{12,2} (\mathbb C)$ with frame angles
$
c_1 = a > c_2 =b > c_3 =0 $
and corresponding multiplicities $m_1 =5 , m_2 =5 $ and $m_3 =1$. It follows from elementary trigonometry that the lower bound in Theorem~\ref{cor_toth} equals $c_1$, showing that $\Phi$ is a Grassmannian frame in $\Omega_{12,2} ( \mathbb C)$. If $Y = \{y_j\}_{j=1}^{12}$ denotes the unit vectors in $\mathbb R^3$ obtained via the embedding from Theorem~\ref{th_sph_emb}, then another computation using the identity
$$
| \langle \phi_j, \phi_l \rangle |^2 = \frac{1}{2} + \frac{1}{2} \langle y_j, y_l\rangle
$$
shows that the vectors of $Y$ must correspond to the vertices of a regular icosahedron.
\end{ex}
Finally, we remark that the frame from Example~\ref{ex_12_2} has some relevance for the combinatorial and quantum information literature. It is simple to check that
$$
\frac{1}{(12)^2}\sum\limits_{j,l=1}^N |\langle \phi_j, \phi_l \rangle |^{10} = \frac{1}{6},
$$
so $\Phi$ generates an (equally) \textem{weighted complex projective $5$-design} by Theorem~{2.3} of ~\cite{MR2337670}. Such objects are known to be optimal for linear quantum state determination with respect to a fixed number of measurements ~\cite{MR2269701, MR2337670}.
\section{Grassmannian Frames for $\mathbb C^2$ Consisting of $5$ vectors }\label{sec:gr_5_2}
In this section, we show that although $\mu_{5,2}(\mathbb C)$ can be achieved by both biangular and triangular frames, as in Examples~\ref{ex_Gr_52_bi} and \ref{ex_Gr_52_tri}, there is a necessaray trade-off between the cardinality of the angle set and tightness. In order to do this, we collect a few basic facts about \textem{biangular, tight frames (BTFs)}.
First, we show that every BTF is equidistributed. A specialized version of this result was shown for the case of {\it $2$-distant tight frames} in \cite{MR3325226}.
\begin{prop}\label{prop_equi}
If $\Phi=\{\phi_j \}_{j =1}^N$ is a biangular, tight frame for $\mathbb F^M$, then $\Phi$ is equidistributed.
\end{prop}
\begin{proof}
If $c_1, c_2$ are the frame angles of $\Phi$, then Equation~\ref{eq_tight_un} implies that for each $j \in \{1,2,...,N\}$, there exists a pair of positive integers $m_{1, j}$ and $m_{2,j}$ such that $m_{1, j} + m_{2,j} =N-1$ and
$$
m_{1,j} c_1^2 + m_{2,j} c_2^2
=
\sum\limits_{l =1, l \neq j}^N | \langle \phi_j, \phi_l \rangle |^2
=
\| \phi_j \|^2 \left( \frac{N}{M}- 1 \right)
=\frac{N-M}{M},
$$
where the last equality follows from the unit-norm property. Next, for $j,l \in \{1,2,...,N\}$, we compute:
\begin{align*}
(N-1)c_1^2+m_{2,j}(c_2^2-c_1^1)&=(m_{1,j}+m_{2,j})c_1^2
+ m_{2,j}(c_2^2-c_1^2)\\
&= m_{1,j}c_1^2+m_{2,j}c_2^2\\
&= m_{1,l}c_1^2+m_{2,l}c_2^2\\
&=(m_{1,l}+m_{2,l})c_1^2+m_{2,l}(c_2^2-c_2^1)\\
&= (N-1)c_1^2+m_{2,l}(c_2^2-c_1^2).
\end{align*}
Since $c_1\not= c_2$, it follows that $m_{2,j}=m_{2,l}$, which implies that
$m_{1,j}=m_{1,l}$.
\end{proof}
It is worth noting that, in general, K-angular tight frames are not equidistributed, so BTFs and ETFs are quite special in this regard. For instance, the frame from Example~\ref{ex_Gr_52_tri} is tight and $3$-angular, but it is not equidistributed. The second observation we need about BTFs concerns the multiplicities of their frame angles when $N$ is odd.
\begin{lemma}\label{lem_even}
Let $N \in \mathbb N$ be odd. If $\Phi=\{\phi_j \}_{j =1}^N$ is a biangular, equidistributed frame for $\mathbb F^M$ with frame angles $c_1, c_2$ and corresponding multiplicities $m_1, m_2$, then $m_1$ and $m_2$ are both even.
\end{lemma}
\begin{proof}
Let $M = \left( |\langle \phi_l, \phi_j \rangle | \right)_{j,l=1}^N$, the matrix obtained by taking the absolute values of the entries of the Gram matrix of $\Phi$.
If $m_1$ is odd, then $N m_1$ is odd, so $c_1$ occurs an odd number of times among the off-diagonal entries of $M$. However, $M$ is symmetric, so the number of occurrences of $c_1$ above the diagonal of $M$ equals the number of occurrences of $c_1$ below the diagonal,which implies that $N m_1$ is even, a contradiction. Therefore, $m_1$ is even, so the fact that $N-1 = m_1 + m_2$ is even implies that $m_2$ is also even.
\end{proof}
\begin{thm}\label{th_btfs_odd}
If $N$ is odd and $\Phi=\{\phi_j \}_{j =1}^N$ is a biangular, tight frame for $\mathbb F^M$, then $\Phi$ is equidistributed with even multiplicities.
\end{thm}
\begin{proof}
This follows directly from Proposition~\ref{prop_equi} and Lemma~\ref{lem_even}.
\end{proof}
Finally, we show that a tight, biangular Grassmannian frame can never exist in $\Omega_{5,2} ( \mathbb C)$.
\begin{thm}\label{th_5_2}
If $\Phi$ is a tight Grassmannian frame in $\Omega_{5,2} (\mathbb C)$, then $|A_\Phi| \geq 3$.
\end{thm}
\begin{proof}
First, note that Examples~\ref{ex_Gr_52_tri} and \ref{ex_Gr_52_bi} show that the lower bound in Theorem~\ref{th_ortho} is saturated in this setting. In particular, we know that $\mu_{5,2} (\mathbb C) = \frac{1}{\sqrt{2}}$.
By way of contradiction, suppose that $|A_\Phi|<3$. If $|A_\Phi|=1$, then $\Phi$ is an ETF, so it achieves the Welch bound, which means
$
\mu(\Phi) = \sqrt{\frac{3}{8}} < \frac{1}{\sqrt{2}},
$
a contradiction.
Therefore, it must be that $|A_\Phi| =2$, meaning $\Phi$ is a BTF.
Let $c_1,c_2$ denote the frame angles of $\Phi$. Because $\Phi$ is a Grassmannian frame, we may assume with no loss of generality that $c_1=\frac{1}{\sqrt{2}}.$ By Corollary~\ref{th_btfs_odd}, $\Phi$ is equidistributed with multiplicities $m_1=m_2=2$. Because $\Phi$ is a $\frac{5}{2}$-tight, unit-norm frame, Equation~\ref{eq_tight_un} becomes
$$
\frac{5}{2} = 1 + 2 c_1^2 + 2 c_2^2 = 1 + 2 \left(\frac{1}{\sqrt{2}} \right)^2 + 2 c_2^2,
$$
which implies that $c_2 = \frac{1}{2}$.
Next, let $Y = \{y_j\}_{j=1}^5$ be the zero-summing unit vectors obtained by embedding the vectors of $\Phi$ into $\mathbb R^3$, as in Corollary~\ref{th_sph_emb_zero}. Because $\Phi$ is equidistributed with multiplicities $m_1=m_2=2$, it follows from Equation~\ref{eq_sph_emb_angles} that the statement
\begin{equation*
\mathcal P_j :
\left| \{ \langle y_j, y_l \rangle = 0 : l \neq j \} \right| = \left| \{ \langle y_j, y_l \rangle= -1/2 = 0 : l \neq j \} \right| = 2
\end{equation*}
must be true for all $j \in \{1,2,3,4,5\}$. We will show that this contradicts the zero-summing property of $Y$.
After an appropriate choice of unitary rotation, there is no loss of generality in assuming that $y_1 = e_1$. The statement $\mathcal P_1$ implies that $y_1$ is orthogonal to two of the vectors of $Y$ and it has inner product $-\frac{1}{2}$ with the other two; therefore, we may further assume without loss of generality that, after an apropriate rotation, $y_2 = e_2$ and that $\langle y_1, y_3 \rangle =0$ and $\langle y_1, y_4 \rangle = \langle y_1, y_5 \rangle = -\frac{1}{2}$. Viewing $Y$ as a $3 \times 5$ matrix, these assumptions mean that its first row and it first two columns are completely determined.
$$
Y =
\left[
\begin{array}{ccccc}
1 & 0 & 0 & -\frac{1}{2} & -\frac{1}{2} \\
0 & 1 & * & * & * \\
0 & 0 & * & * & * \\
\end{array}
\right].
$$
We cannot have $\langle y_2, y_3 \rangle =0$, because then the statements $\mathcal P_1, \mathcal P_2$ and $\mathcal P_3$ force
$$
\langle y_1, y_4 \rangle = \langle y_2, y_4 \rangle = \langle y_3, y_4 \rangle,
$$
which in turn contradicts $\mathcal P_4$. Therefore, $\langle y_2, y_3 \rangle = -1/2$.
$$
Y =
\left[
\begin{array}{ccccc}
1 & 0 & 0 & -\frac{1}{2} & -\frac{1}{2} \\
0 & 1 & -\frac{1}{2} & * & * \\
0 & 0 & * & * & * \\
\end{array}
\right].
$$
The statement $\mathcal P_2$ then implies that either (i) $\langle y_2, y_4 \rangle =0$ and $\langle y_2, y_5 \rangle =-1/2$ or (ii) $\langle y_2, y_4 \rangle =-1/2$ and $\langle y_2, y_5 \rangle =0$. Since it is clear that these two cases are symmetric, we assume with no loss in generality that $\langle y_2, y_4 \rangle =0$ and $\langle y_2, y_5 \rangle =-1/2$.
$$
Y =
\left[
\begin{array}{ccccc}
1 & 0 & 0 & -\frac{1}{2} & -\frac{1}{2} \\
0 & 1 & -\frac{1}{2} & 0 & -\frac{1}{2} \\
0 & 0 & * & * & * \\
\end{array}
\right].
$$
Finally, the unit-norm condition means that the remaining entries of $Y$ must satisfy
$$
Y =
\left[
\begin{array}{ccccc}
1 & 0 & 0 & -\frac{1}{2} & -\frac{1}{2} \\
0 & 1 & -\frac{1}{2} & 0 & -\frac{1}{2} \\
0 & 0 & \pm \frac{\sqrt{3}}{2} & \pm \frac{\sqrt{3}}{2} & \pm \frac{\sqrt{2}}{2} \\
\end{array}
\right],
$$
but this contradicts the zero-summing condition, because there is no choice of signs for which
$$
\pm \frac{\sqrt{3}}{2} \pm \frac{\sqrt{3}}{2} \pm \frac{\sqrt{2}}{2} = 0.
$$
Therefore, $\Phi$ cannot exist, which completes the proof.
\end{proof}
\section*{Acknowledgments}
The authors were supported by
NSF DMS 1307685, NSF ATD 1321779, and ARO W911NF-16-1-0008.
\section*{References}
|
\section{Introduction}
Methods based on Markov chains play a critical role in statistical
inference, where they form the basis of Markov chain Monte Carlo
(MCMC) procedures for estimating intractable expectations~\cite[see,
e.g.,][]{Gel13BDA,Rob05MCMC}. In MCMC procedures, it is the
stationary distribution of the Markov chain that typically encodes the
information of interest. Thus, MCMC estimates are asymptotically
exact, but their accuracy at finite times is limited by the
convergence rate of the chain.
The usual measures of convergence rates of Markov chains---namely, the
total variation mixing time or the absolute spectral gap of the
transition matrix~\citep{Lev08Markov}---correspond to very strong
notions of convergence and depend on global properties of the
chain. Indeed, convergence of a Markov chain in total variation
corresponds to uniform convergence of the expectations of all
unit-bounded function to their equilibrium values. The resulting
uniform bounds on the accuracy of
expectations~\citep{Chu12Hoeffding,Gil98Chernoff,Jou10Curvature,Kon14Uniform,Leo04Hoeffding,
Lez01Berry,Pau12Conc,Sam00Concentration} may be overly pessimistic---not indicative of
the mixing times of specific expectations such as means and variances
that are likely to be of interest in an inferential setting.
Another limitation of the uniform bounds is that they typically assume
that the chain has arrived at the equilibrium distribution, at least
approximately. Consequently, applying such bounds requires either
assuming that the chain is started in equilibrium---impossible in
practical applications of MCMC---or that the burn-in period is
proportional to the mixing time of the chain, which is also unrealistic, if
not impossible, in practical settings.
Given that the goal of MCMC is often to estimate specific
expectations, as opposed to obtaining the stationary distribution, in
the current paper we develop a function-specific notion of convergence
with application to problems in Bayesian inference. We define a notion
of ``function-specific mixing time,'' and we develop function-specific
concentration bounds for Markov chains, as well as spectrum-based
bounds on function-specific mixing times. We demonstrate the utility
of both our overall framework and our particular concentration bounds
by applying them to examples of MCMC-based data analysis from the
literature and by using them to derive sharper confidence intervals
and faster sequential testing procedures for MCMC.
\subsection{Preliminaries}
We focus on discrete time Markov chains on $d$ states given by a $d
\times d$ transition matrix $P$ that satisfies the conditions of
irreducibility, aperiodicity, and reversibility. These conditions
guarantee the existence of a unique stationary distribution $\pi$.
The issue is then to understand how quickly empirical averages of
functions of the Markov chain, of the form $f: [d] \to [0,1]$,
approach the stationary average, denoted by
\begin{align*}
\mu & \mydefn \E_{X \sim \pi} [f(X)].
\end{align*}
The classical analysis of mixing defines convergence rate in terms of
the total variation distance:
\begin{align}
\label{eq:sup-TV}
d_{\TV} \big (p,~q \big) = \sup_{f \colon \Omega \rightarrow [0,~1]}
\Big| \E_{X\sim p}\big[f(X)\big] - \E_{Y \sim q} \big [ f(Y) \big]
\Big|,
\end{align}
where the supremum ranges over all unit-bounded functions. The mixing
time is then defined as the number of steps required to ensure that
the chain is within total-variation distance $\delta$ of the
stationary distribution---that is
\begin{align}
\label{EqnDefnClassicalMixing}
T(\delta) & \mydefn \min \Big \{ n \in \Nat \; \mid \; \max_{i \in
[d]} d_{\TV}\big(\pi_{n}^{(i)},~\pi\big) \leq \delta \Big \},
\end{align}
where $\Nat = \{1, 2, \ldots \}$ denotes the natural numbers,
and $\pi_{n}^{(i)}$ is the distribution of the chain state $X_{n}$
given the starting state $X_{0} = i$.
Total variation is a worst-case measure of distance, and the resulting
notion of mixing time can therefore be overly conservative when the
Markov chain is being used to approximate the expectation of a fixed
function, or expectations over some relatively limited class of
functions. Accordingly, it is of interest to consider the following
function-specific discrepancy measure:
\begin{defn}[$f$-discrepancy]
For a given function $f$, the $f$-discrepancy is
\begin{align}
d_{f} \big(p, ~q \big) = \big|\E_{X \sim p} \big [ f \big( X
\big) \big] - \E_{Y \sim q} \big [ f \big( Y \big) \big ]
\big |.
\end{align}
\end{defn}
\noindent The $f$-discrepancy leads naturally to a function-specific
notion of mixing time:
\begin{defn}[$f$-mixing time]
For a given function $f$, the $f$-mixing time is
\begin{align}
\Tf\big(\delta\big) = \min \Big \{ n \in \Nat \; \mid \; \max_{i \in
[d]} d_{f}\big(\pi_{n}^{(i)},~\pi\big) \leq \delta \Big \} .
\end{align}
\end{defn}
\noindent In the sequel, we also define function-specific notions of
the spectral gap of a Markov chain, which can be used to bound the
$f$-mixing time and to obtain function-specific concentration
inequalities.
\subsection{Related work}
Mixing times are a classical topic of study in Markov chain theory,
and there is a large collection of techniques for their
analysis~\citep[see,
e.g.,][]{Ald86Shuffling,Dia90StrongStat,Lev08Markov,Mey12Markov,Oll09Ricci,Sin92Multicomm}.
These tools and the results based on them, however, generally apply
only to worst-case mixing times. Outside of specific examples~\citep{Con06Riffle,Dia15Heisenberg},
relatively little is known about mixing with respect to individual functions or limited classes of
functions. Similar limitations exist in studies of concentration of
measure and studies of confidence intervals and other statistical
functionals that depend on tail probability bounds. Existing bounds
are generally uniform, or non-adaptive, and the rates that are
reported include a factor that encodes the global mixing properties of
the chain and does not adapt to the
function~\citep{Chu12Hoeffding,Gil98Chernoff,Jou10Curvature,Kon14Uniform,Leo04Hoeffding,Lez01Berry,Pau12Conc,Sam00Concentration}.
These factors, which do not appear in classic bounds for independent
random variables, are generally either some variant of the spectral
gap $\gamma$ of the transition matrix, or else a mixing time of the
chain $T\big(\delta_{0}\big)$ for some absolute constant $\delta_{0} >
0$. For example, the main theorem from \cite{Leo04Hoeffding} shows
that for a function $f \colon [d] \rightarrow [0,~1]$ and a sample
$X_{0} \sim \pi$ from the stationary distribution, we have
\begin{align}
\label{eq:unif-hoeffding}
\P \big( \big| \frac{1}{N}\sum_{n = 1}^{N} f\big(X_{n}\big) - \mu\big|
\geq \epsilon\big) \leq 2 \exp \Big \{ -\frac{\gamma_{0}}{2\big(2 -
\gamma_0\big)} \cdot \epsilon^2N \Big \},
\end{align}
where the eigenvalues of $P$ are given in decreasing order as \mbox{$1
> \lambda_2(P) \geq \cdots \geq \lambda_d(P)$,} and we denote the
spectral gap of $P$ by
\begin{align*}
\gamma_{0} \mydefn \min \big \{ 1 - \lambda_2(P),~1 \big \}.
\end{align*}
The requirement that the chain start in equilibrium can be relaxed by
adding a correction for the burn-in time~\citep{Pau12Conc}. Extensions
of this and related bounds, including bounded-differences-type
inequalities and generalizations to continuous Markov chains and
non-Markov mixing processes have also appeared in the literature
(e.g.,~\cite{Kon14Uniform,Sam00Concentration}).
The concentration result has an alternative formulation in terms of
the mixing time instead of the spectral gap~\citep{Chu12Hoeffding}.
This version and its variants are weaker, since the mixing time can be
lower bounded as
\begin{align}
\label{eq:Tmix-lbd-gamma}
T\big(\delta\big) \geq \big(\frac{1}{\gamma_{\ast}} - 1\big)
\log\big(\frac{1}{2\delta}\big) \geq \big(\frac{1}{\gamma_0} -
1\big)\log\big(\frac{1}{2\delta}\big),
\end{align}
where we denote the absolute spectral
gap~\citep{Lev08Markov} by
\[\gamma_{\ast} \mydefn \min\big(1 - \lambda_2 ,~1 -
\big|\lambda_{d}\big|\big) ~\leq~ \gamma_{0}.\]
In terms of the minimum probability $\pimin \mydefn \min_i \pi_i$, the
corresponding upper bound is an extra factor of
$\log\big(\frac{1}{\pimin}\big)$ larger, which potentially leads to a
significant gap between $\frac{1}{\gamma_{0}}$ and
$T\big(\delta_{0}\big)$, even for a moderate constant such as
$\delta_{0} = \frac{1}{8}$. Similar distinctions arise in our
analysis, and we elaborate on them at the appropriate junctures.
\subsection{Organization of the paper}
In the remainder of the paper, we elaborate on these ideas and apply
them to MCMC. In Section~\ref{sec:main-results}, we state some
concentration guarantees based on function-specific mixing times, as
well as some spectrum-based bounds on $f$-mixing times, and the
spectrum-based Hoeffding bounds they imply.
Section~\ref{sec:stat-apps} is devoted to further development of these
results in the context of several statistical models. More
specifically, in Section~\ref{subsec:confidence}, we show how our
concentration guarantees can be used to derive confidence intervals
that are superior to those based on uniform Hoeffding bounds and
CLT-type bounds, whereas in Section~\ref{subsec:testing}, we analyze the consequences
for sequential testing. In Section~\ref{subsec:practical}, we show that our mixing time and
concentration bounds improve over the non-adaptive bounds in real
examples of MCMC from the literature. Finally, the bulk of our proofs
are given in Section~\ref{sec:proofs}, with some more technical
aspects of the arguments deferred to the appendices.
\section{Main results}
\label{sec:main-results}
We now present our main technical contributions, starting with a set
of ``master'' Hoeffding bounds with exponents given in terms of
$f$-mixing times. As we explain in
Section~\ref{subsec:derived-hoeffding}, these mixing time bounds can
be converted to spectral bounds bounding the $f$-mixing time in terms
of the spectrum. (We give some techniques for the latter in
Section~\ref{subsec:mixing-bounds}).
Recall that we use \mbox{$\mu \mydefn \E_{\pi} [f]$} to denote the
mean. Moreover, we follow standard conventions in setting
\begin{align*}
\lambda_{\ast} \mydefn \max \big \{ \lambda_2(P),~ \big|\lambda_{d}(P)
\big| \big \} , \quad \mbox{and} \quad \lambda_{0} \mydefn \max \big
\{ \lambda_2(P),~ 0 \big \}.
\end{align*}
so that the absolute spectral gap and the
(truncated) spectral gap introduced earlier are given by
$\gamma_{\ast} \mydefn 1 - \lambda_{\ast}, \quad \mbox{and} \quad
\gamma_{0} \mydefn 1 - \lambda_{0}.$
In Section~\ref{subsec:mixing-bounds}, we define and analyze
corresponding function-specific quantities, which we introduce as
necessary.
\subsection{Master Hoeffding bound}
\label{subsec:master-hoeffding}
In this section, we present a master Hoeffding bound that provides
concentration rates that depend on the mixing properties of the chain
only through the $f$-mixing time $T_{f}$. The only hypotheses on
burn-in time needed for the bounds to hold are that the chain has been
run for at least $N \geq T_{f}$ steps---basically, so that thinning is
possible---and that the chain was started from a distribution
$\pi_{0}$ whose $f$-discrepancy distance from $\pi$ is small---so that
the expectation of each $f\big(X_n\big)$ iterate is close to
$\mu$---even if its total-variation discrepancy from $\pi$ is
large. Note that the latter requirement imposes only a very mild
restriction, since it can always be satisfied by first running the
chain for a burn-in period of $T_{f}$ steps and then beginning to
record samples.
\begin{thm}
\label{thm:hoeffding-eps2}
Given any fixed $\epsilon>0$ such that $\df \big( \pi_{0},~\pi \big)
\leq \frac{\epsilon}{2}$ and $N \geq \Tf \big(\frac{\epsilon
}{2}\big)$, we have
\begin{align}
\label{eq:hoeffding-eps2}
\P \Big[ \frac{1}{N}\sum_{n = 1}^{N} f\big(X_n\big) \geq \mu +
\epsilon \Big] \leq \exp \left \{ -\frac{\epsilon^2
N}{8\Tf\big(\frac{\epsilon}{2}\big)} \right \}.
\end{align}
\end{thm}
Compared to the bounds in earlier work~\cite[e.g.,][]{Leo04Hoeffding},
the bound~\eqref{eq:hoeffding-eps2} has several distinguishing
features. The primary difference is that the ``effective'' sample size
\begin{subequations}
\begin{align}
\label{EqnEffective}
N_{\eff} & \mydefn \frac{N}{\Tf(\epsilon /2)},
\end{align}
is a function of $f$, which can lead to significantly sharper bounds
on the deviations of the empirical means than the earlier uniform
bounds can deliver. Further, unlike the uniform results, we do not
require that the chain has reached equilibrium, or even approximate
equilibrium, in a total variation sense. Instead, the result applies
provided that the chain has equilibrated only approximately, and only
with respect to $f$.
The reader might note that if one actually has access to a
distribution $\pi_{0}$ that is $\epsilon/2$-close to $\pi$ in
$f$-discrepancy, then an estimator of $\mu$ with tail bounds similar
to those guaranteed by Theorem~\ref{thm:hoeffding-eps2} can be
obtained as follows: first, draw $N$ i.i.d. samples from $\pi_{0}$,
and second, apply the usual Hoeffding inequality for i.i.d. variables.
However, it is essential to realize that
Theorem~\ref{thm:hoeffding-eps2} does not require that such a
$\pi_{0}$ be available to the practitioner. Instead, the theorem
statement is meant to apply in the following way: suppose
that---starting from \emph{any} initial distribution---we run an
algorithm for $N \geq T_f(\epsilon/2)$ steps, and then use the last of
$N - T_f(\epsilon/2)$ samples to form an empirical average. Our
concentration result then holds with an effective sample size of
\begin{align}
\label{EqnEffectiveWithBurnIn}
N_{\eff}^{\mathrm{burnin}} & \mydefn \frac{N -
T_f(\epsilon/2)}{\Tf\left(\epsilon/2\right)} ~=~
\frac{N}{\Tf\left(\epsilon/2\right)} - 1.
\end{align}
In other words, the result can be applied with an arbitrary initial
$\pi_{0}$, and accounting for burn-in merely reduces the effective
sample size by one. By contrast, such an interpretation does not
actually hold for the original result of \cite{Leo04Hoeffding}: it
requires an initial sample $X_1 \sim \pi$, but such an exact sample is
not attainable after any finite burn-in period.
The appearance of the function-specific mixing time $\Tf$ in the
bounds comes with both advantages and disadvantages. A notable
disadvantage, shared with the mixing time versions of the uniform
bounds, is that spectrum-based bounds on the mixing time (including
our $f$-specific ones) introduce a $\log\big(\frac{1}{\pimin}\big)$
term that can be a significant source of looseness. On the other hand,
obtaining rates in terms of mixing times comes with the advantage that
any bound on the mixing time translates directly into a version of the
concentration bound (with the mixing time replaced by its upper
bound). Moreover, since the $\pimin^{-1}$ term is likely to be an
artifact of the spectrum-based approach, and possibly even just of the
proof method, it may be possible to turn the mixing time based bound
into a stronger spectrum-based bound with a more sophisticated
analysis. We go part of the way toward doing this, albeit without
completely removing the $\pimin^{-1}$ term.
An analysis based on mixing time also has the virtue of better
capturing the non-asymptotic behavior of the rate. Indeed, as a
consequence of the link~\eqref{eq:Tmix-lbd-gamma} between mixing and
spectral graph (as well as matching upper bounds~\citep{Lev08Markov}),
for any fixed function $f$, there exists a function-specific
spectral-gap $\gamma_{f} > 0$ such that
\begin{align}
\label{eq:Tf-asym-gamma}
\Tf \big(\frac{\epsilon }{2}\big) \approx \frac1{\gamma_{f}} \log
\Big( \frac{1}{\epsilon } \Big) + O \big( 1 \big), \quad \mbox{for}
\quad \epsilon \ll 1.
\end{align}
\end{subequations}
These asymptotics can be used to turn our aforementioned theorem into
a variant of the results of~\citet{Leo04Hoeffding}, in which
$\gamma_{0}$ is replaced by a value $\gamma_{f}$ that (under mild
conditions) is at least as large as $\gamma_{0}$. However, as we
explore in Section~\ref{subsec:practical}, such an asymptotic
spectrum-based view loses a great deal of information needed to deal
with practical cases, where often $\gamma_{f} =\gamma_{0}$ and yet
\mbox{$\Tf(\delta) \ll T(\delta)$} even for very small values of
$\delta > 0$. For this reason, part of our work is devoted to
deriving more fine-grained concentration inequalities that capture
this non-asymptotic behavior.
By combining our definition~\eqref{EqnEffective} of the effective
sample size $N_{\eff}$ with the asymptotic
expansion~\eqref{eq:Tf-asym-gamma}, we arrive at an intuitive
interpretation of Theorem~\ref{thm:hoeffding-eps2}: it dictates that
the effective sample size scales as $N_{\eff} ~\approx~
\frac{\gamma_{f} N }{ \log(1/\epsilon)}$ in terms of the
function-specific gap $\gamma_f$ and tolerance $\epsilon$. This
interpretation is backed by the Hoeffding bound derived in
Corollary~\ref{cor:hoeffding-derived-eps2} and it is useful as a
simple mental model of these bounds. On the other hand, interpreting
the theorem this way effectively plugs in the asymptotic behavior of
$T_{f}$ and does not account for the non-asymptotic properties of the
mixing time; the latter may actually be more favorable and lead to
substantially smaller effective sample sizes than the naive asymptotic
interpretation predicts. From this perspective, the master bound has
the advantage that any bound on $T_{f}$ that takes advantage of
favorable non-asymptotics translates directly into a stronger version
of the Hoeffding bound. We investigate these issues empirically in
Section~\ref{subsec:practical}.
Based on the worst-case Markov Hoeffding
bound~\eqref{eq:unif-hoeffding}, we might hope that the
\mbox{$\Tf(\frac{\epsilon}{2})$} term in
Theorem~\ref{thm:hoeffding-eps2} is spurious and removable using
improved techniques. Unfortunately, it is fundamental. This conclusion
becomes less surprising if one notes that even if we start the chain
in its stationary distribution and run it for \mbox{$N <
\Tf(\epsilon)$} steps, it may still be the case that there is a
large set $\Omega_{0}$ such that for $i \in \Omega_0$ and $1 \leq n
\leq N$,
\begin{align}
\label{eq:lower-bd-intuition}
\left|f(X_{n}) - \mu\right| \gg \epsilon~~ \text{a.s. if}~ X_{0} = i .
\end{align}
This behavior is made possible by the fact that large positive and
negative deviations associated with different values in $\Omega_{0}$
can cancel out to ensure that $\E\left[f\left(X_{n}\right)\right] =
\mu$ marginally. However, the lower
bound~\eqref{eq:lower-bd-intuition} guarantees that
\begin{align*}
\P\left(\frac{1}{N}\sum_{n = 1}^{N} f\left(X_{n}\right) \geq \mu +
\epsilon\right) & \geq \sum_{i \in \Omega_0} \pi_{i} \cdot
\P\left(\frac{1}{N}\sum_{n = 1}^{N} f\left(X_{n}\right) \geq \mu +
\epsilon~|~X_{0} = i\right) \\ & \geq \pi\left(\Omega_0\right),
\end{align*}
so that if $\pi\left(\Omega_0\right) \gg 0$, we have no hope of
controlling the large deviation probability unless $N \gtrsim
\Tf\left(\epsilon\right)$. We make this intuitive argument precise in
Section~\ref{sec:lower}.
\subsection{Bounds on $f$-mixing times}
\label{subsec:mixing-bounds}
We generally do not have direct access either to the mixing time
$T\big(\delta\big)$ or the $f$-mixing time
$\Tf\big(\delta\big)$. Fortunately, any bound on $\Tf$ translates
directly into a variant of the tail bound~\eqref{eq:hoeffding-eps2}.
Accordingly, this section is devoted to methods for bounding these
quantities. Since mixing time bounds are equivalent to bounds on
$d_{\TV}$ and $d_{f}$, we frame the results in terms of distances
rather than times. These results can then be inverted in order to
obtain mixing-time bounds in applications.
The simplest bound is simply a uniform bound on total variation
distance, which also yields a bound on the $f$-discrepancy. In
particular, if the chain is started with distribution $\pi_0$, then we
have
\begin{align}
\label{eq:mix-TV}
d_{\TV}\big(\pi_{n},~\pi\big) \leq \frac{1}{\sqrt{\pimin}} \cdot
\abslambda^{n} \cdot d_{\TV}\big(\pi_{0},~\pi\big).
\end{align}
In order to improve upon this bound, we need to develop
function-specific notions of spectrum and spectral gaps. The simplest
way to do this is simply to consider the (left) eigenvectors to which
the function is not orthogonal and define a spectral gap restricted
only to the corresponding eigenvectors.
\begin{defn}[$f$-eigenvalues and spectral gaps]
For a function \mbox{$f \colon [d] \rightarrow \R$,} we define
\begin{subequations}
\begin{align}
J_{f} & \mydefn \Big \{ j \in [d] \, \mid \, \lambda_{j} \neq 1
~\text{and}~ q_{j}^{T}f \neq 0 \Big \},
\end{align}
where $q_{j}$ denotes a left eigenvector associated with
$\lambda_{j}$. Similarly, we define
\begin{align}
\lambda_{f} = \max_{j \in J_{f}} \big|\lambda_{j}\big|, \quad
\mbox{and} \quad \gamma_{f} = 1 - \lambda_{f}.
\end{align}
\end{subequations}
\end{defn}
Using this notation, it is straightforward to show that if the chain
is started with the distribution $\pi_{0}$, then
\begin{align}
\label{eq:mix-gap}
d_{f} \big( \pi_{n},~\pi\big) \leq \sqrt{ \frac{\E_{\pi}\big[f^2
\big]}{\pimin}} \cdot \lambda_{f}^{n} \cdot
d_{f}\big(\pi_{0},~\pi\big) .
\end{align}
This bound, though useful in many cases, is also rather brittle: it
requires $f$ to be exactly orthogonal to the eigenfunctions of the
transition matrix. For example, a function $f_{0}$ with a good value
of $\lambda_{f}$ can be perturbed by an arbitrarily small amount in a
way that makes the resulting perturbed function $f_{1}$ have
$\lambda_{f} = \lambda_{\ast}$. More broadly, the bound is of little
value for functions with a small but nonzero inner product with the
eigenfunctions corresponding to large eigenvalues (which is likely to
occur in practice; cf.\ Section \ref{subsec:practical}), or in
scenarios where $f$ lacks symmetry (cf.\ the random function example
in Section~\ref{subsec:example-cycle}).
In order to address these issues, we now derive a more fine-grained
bound on $\df$. The basic idea is to split the lower $f$-spectrum
$J_{f}$ into a ``bad'' piece $J$, whose eigenvalues are close to $1$
but whose eigenvectors are approximately orthogonal to $f$, and a
``good'' piece $J_{f} \setminus J$, whose eigenvalues are far from $1$
and which therefore do not require control on the inner products of
their eigenvectors with $f$. More precisely, for a given set $J
\subset J_{f}$, let us define
\begin{align*}
\Delta_{J}^{\ast} \mydefn 2 \big |J \big| \times \max_{j \in J} \|
h_{j} \|_{\infty} \times \max_{j \in J} \big | q_{j}^{T} f \big|, &
\qquad \lambda_{J} \mydefn \max \Big \{ \big|\lambda_{j}\big| \, \mid
\, j \in J \Big \}, \quad \mbox{and} \\
\lambda_{-J} \mydefn \max \Big \{ \big|\lambda_{j}\big| \; \mid \; j
\in J_{f} \setminus J \Big \}.
\end{align*}
We obtain the following bound, expressed in terms of $\lambda_{-J}$
and $\lambda_{J}$, which we generally expect to obey the relation $1 -
\lambda_{-J} \ll 1 - \lambda_{J}$.
\begin{lem}[Sharper $f$-discrepancy bound]
Given $f \colon [d] \rightarrow [0,~1]$ and a subset $J \subset
J_{f}$, we have
\begin{align}
\df \big(\pi_{n},~\pi \big ) & \leq \Delta^{\ast}_J \: \lambda_{J}^{n}
\cdot d_{\TV}(\pi_0,~\pi) + \sqrt{\frac{\E_{\pi}\big [ f^2 \big]
}{\pimin}} \cdot \lambda_{-J}^{n} \, \df( \pi_{0}, ~\pi).
\end{align}
\label{lem:mix-gap-all}
\end{lem}
The above bound, while easy to apply and comparatively easy to
estimate, can be loose when the first term is a poor
estimate of the part of the discrepancy that comes from the $J$ part
of the spectrum. We can get a still sharper estimate by instead
making use of the following vector quantity that
more precisely summarizes the interactions between $f$ and $J$:
\begin{align*}
h_{J}\big(n\big) \mydefn \sum_{j \in J} \big(q_{j}^{T}f \cdot
\lambda_{j}^{n}\big) h_{j}.
\end{align*}
This quantity leads to what we refer to as an \emph{oracle adaptive
bound}, because it uses the exact value of the part of the
discrepancy coming from the $J$ eigenspaces, while using the same
bound as above for the part of the discrepancy coming from
$J_{f}\setminus J$.
\begin{lem}[Oracle $f$-discrepancy bound]
\label{lem:mix-gap-all-oracle}
Given \mbox{$f \colon [d] \rightarrow [0,~1]$} and a subset \mbox{$J
\subset J_{f}$,} we have
\begin{align}
\df \big(\pi_{n},~\pi \big) & \leq \big|\big(\pi_0 -\pi\big)^{T} h_{J}
\big (n \big ) \big| + \sqrt{\frac{\E_{\pi}\big [f^2 \big] }{\pimin}}
\cdot \lambda_{-J}^{n} \cdot \df \big( \pi_{0}, ~\pi \big).
\end{align}
\end{lem}
We emphasize that, although Lemma \ref{lem:mix-gap-all-oracle} is
stated in terms of the initial distribution $\pi_{0}$, when we apply
the bound in the real examples we consider, we replace all
quantities that depend on $\pi_0$ by their worst cases values, in
order to avoid dependence on initialization; this results in a $\|
h_{J} \big ( n \big) \|_{\infty}$ term instead of the dot product in
the lemma.
\subsection{Concentration bounds}
\label{subsec:derived-hoeffding}
The mixing time bounds from Section~\ref{subsec:mixing-bounds} allow
us to translate the master Hoeffding bound into a weaker but more
interpretable---and in some instances, more directly
applicable---concentration bound. The first result we prove along
these lines applies meaningfully only to functions $f$ whose absolute
$f$-spectral gap $\gamma_{f}$ is larger than the absolute spectral gap
$\gamma_{\ast}$. It is a direct consequence of the master Hoeffding
bound and the simple spectral mixing bound \eqref{eq:mix-gap}, and it
delivers the asymptotics in $N$ and $\epsilon$ promised in
Section~\ref{subsec:master-hoeffding}.
\begin{cor}
\label{cor:hoeffding-derived-eps2}
Given any $\epsilon > 0$ such that $d_{f} \big( \pi_{0},~\pi\big) \leq
\frac{\epsilon}{2}$ and $N \geq \Tf \big (\frac{\epsilon}{2}\big)$, we
have
\begin{align*}
\P \left[ \frac{1}{N} \sum_{n = 1}^{N} f(X_n) \geq \mu + \epsilon
\right] & \leq
\begin{cases}
\exp \left ( - \frac{\epsilon^2}{8} \: \frac{\gamma_{f} N}{ \log
\big(\frac{2}{\epsilon \sqrt{\pimin}} \big) } \right ) &
\text{ if } \epsilon \leq \frac{2\lambda_f}{\sqrt{\pimin}},\\
\exp \left ( - \frac{\epsilon^2 N}{8} \right ) & \text{ otherwise.}
\end{cases}
\end{align*}
\end{cor}
Deriving a Hoeffding bound using the sharper $f$-mixing bound given in
Lemma~\ref{lem:mix-gap-all} requires more care, both because of the
added complexity of managing two terms in the bound and because one of
those terms does not decay, meaning that the bound only holds for
sufficiently large deviations $\epsilon > 0$.
The following result represents one way of articulating the bound
implied by Lemma~\ref{lem:mix-gap-all}; it leads to improvements over
the previous two results when the contribution from the bad part of
the spectrum $J$---that is, the part of the spectrum that brings
$\gamma_{f}$ closer to $1$ than we would like---is negligible at the
scale of interest. Recall that Lemma~\ref{lem:mix-gap-all} expresses
the contribution of $J$ via the quantity $\Delta_{J}^{\ast}$.
\begin{cor}
\label{cor:hoeffding-derived-Jf}
Given a triple of positive numbers $(\Delta, \Delta_J,
\Delta_{J}^\ast)$ such that $\Delta_{J} \geq \Delta_{J}^{\ast}$ and $N
\geq \Tf \big(\Delta_{J} + \Delta\big)$, we have
\begin{align}
\P \left[ \frac{1}{N}\sum_{n = 1}^{N} f\big(X_n\big) \geq \mu +
2\left(\Delta_{J} + \Delta\right)\right] & \leq
\begin{cases}
\exp\big(- \frac{\big(\Delta_{J} + \Delta\big)^{2}}{2} \:
\frac{\big(1 - \lambda_{-J}\big)N}{ \log\big(\frac{1}{\Delta
\sqrt{\pimin}}\big) }\big) \text{ if } \Delta \leq
\frac{\lambda_{-J}}{\sqrt{\pimin}}, \\
\exp \big(-\frac{\big(\Delta_{J} + \Delta\big)^{2}N}{2}\big) \text{ if }
\Delta > \frac{\lambda_{-J}}{\sqrt{\pimin}}.
\end{cases}
\end{align}
\end{cor}
Similar arguments can be applied to combine the master Hoeffding
bounds with the oracle $f$-mixing bound
Lemma~\ref{lem:mix-gap-all-oracle}, but we omit the corresponding
result for the sake of brevity. The proofs for both aforementioned corollaries are in Section \ref{subsec:proofs-conc-derived}.
\subsection{Example: Lazy random walk on $C_{2d}$}
\label{subsec:example-cycle}
In order to illustrate the mixing time and Hoeffding bounds from
Section~\ref{subsec:mixing-bounds}, we analyze their predictions for
various classes of functions on the $2d$-cycle $C_{2d}$, identified
with the integers modulo $2d$. In particular, consider the Markov
chain corresponding to a lazy random walk on $C_{2d}$; it has
transition matrix
\begin{align}
\label{eq:P-C2d}
P_{uv} = \begin{cases} \frac{1}2 &~ \text{if}~ v = u, \\ \frac{1}{4}
&~ \text{if}~ v = u + 1 \mod{2d}, \\ \frac{1}{4} &~ \text{if}~ v = u
- 1 \mod{2d}, \\ 0 &~ \text{otherwise.}
\end{cases}
\end{align}
It is easy to see that the chain is irreducible, aperiodic, and
reversible, and its stationary distribution is uniform. It can be
shown~\citep{Lev08Markov} that its mixing time scales proportionally
\mbox{to $d^2$.} However, as we now show, several interesting classes
of functions mix much faster, and in fact, a ``typical'' function,
meaning a randomly chosen one, mixes much faster than the naive mixing
bound would predict.
\paragraph{Parity function.} The epitome of a rapidly mixing function is the parity function:
\begin{align}
\parity (u)) \mydefn \begin{cases}
1 &~ \text{if}~ u~ \text{is odd}, \\
0 &~ \text{otherwise.}
\end{cases}
\end{align}
It is easy to see that no matter what the choice of initial
distribution $\pi_0$ is, we have $\E\big[\parity(X_1)\big] = \frac{1}2
$, and thus $\parity$ mixes in a single step.
\paragraph{Periodic functions.} A more general class of examples arises from considering the
eigenfunctions of $P$, which are given by $g_{j}\big(u\big) =
\cos\big(\frac{\pi j u}{d}\big)$; \citep[see, e.g.,][]{Lev08Markov}.
We define a class of functions of varying regularity by setting
\begin{align*}
f_{j} = \frac{1 + g_{j}}2 , \quad \mbox{for each $j = 0, 1, \ldots,
d$.}
\end{align*}
Here we have limited $j$ to $0 \leq j \leq d$ because $f_{j}$ and $f_{2d -
j}$ behave analogously. Note that the parity function $\parity$
corresponds to $f_d$.
Intuitively, one might expect that some of these functions mix well
before $d^2$ steps have elapsed---both because the vectors $\{f_j, \;
j \neq 1 \}$ are orthogonal to the non-top eigenvectors with
eigenvalues close to $1$ and because as $j$ gets larger, the periods
of $f_{j}$ become smaller and smaller, meaning that their global
behavior can increasingly be well determined by looking at local
snapshots, which can be seen in few steps.
Our mixing bounds allow us to make this intuition precise, and our
Hoeffding bounds allow us to prove correspondingly improved
concentration bounds for the estimation of $\mu =
\E_{\pi}\big[f_{j}\big] = 1/2$. Indeed, we have
\begin{align}
\label{eq:gamma-trig}
\gamma_{f_j} = \frac{1 - \cos\big(\frac{\pi j}{d}\big)}2
\geq
\begin{cases} \frac{\pi^2 j^2 }{24d^2 } & \text{if}~ j \leq
\frac{d}2 , \\
\frac{1}2 & \text{if}~ \frac{d}2 < j \leq d.
\end{cases}
\end{align}
Consequently, equation~\eqref{eq:mix-gap} predicts that
\begin{align}
\label{eq:Tf-trig}
T_{f_j} \big(\delta\big) \leq \tilde{T}_{f_j}\big(\delta\big) =
\begin{cases}
\frac{24}{\pi^2 } \big[\frac{1}2 \log 2d+ \log \big(\frac{1}{\delta}
\big) \big] \cdot \frac{d^2 }{j^2 } &~ \text{if}~ j \leq
\frac{d}2 , \\
\log 2d + 2 \log\big
(\frac{1}{\delta}\big) &~ \text{if}~ \frac{d}2 < j \leq d,
\end{cases}
\end{align}
where we have used the trivial bound $\E_{\pi}\big[f^2 \big] \leq 1$
to simplify the inequalities. Note that this yields an improvement
over $\asymp d^2 $ for $j \gtrsim \log{d}$. Moreover, the
bound~\eqref{eq:Tf-trig} can itself be improved, since each $f_{j}$ is
orthogonal to all eigenfunctions other than $\ones$ and $g_{j}$, so
that the $\log{d}$ factors can all be removed by a more carefully
argued form of Lemma~\ref{lem:mix-gap-all}. It thus follows directly
from the bound~\eqref{eq:gamma-trig} that if we draw $N +
\tilde{T}_{f_j}\big(\frac{\epsilon}{2}\big)$ samples, we obtain the
tail bound
\begin{align}
\P \Big[ \frac{1}{N_0} \sum_{n = N_{\mathrm{b}}}^{N +
N_{\mathrm{b}}} f_{j}\big(X_n\big) \geq \frac{1}2 +
\epsilon \Big] & \leq
\begin{cases}
\exp \big(-\frac{3 d^2 }{\pi^2 j^2 } \cdot \frac{\epsilon^2 N}{
\log\big(2\sqrt{2d}/\epsilon\big)}\big) &~ \text{if}~ j \leq
\frac{d}2 , \\
\exp \big(-\frac{\epsilon^2 N}{16 \log \big(
2\sqrt{2d}/\epsilon \big)} \big) &~ \frac{d}2 < j
\leq d ,
\end{cases}
\end{align}
where the burn-in time is given by $N_{\mathrm{b}} =
\tilde{T}_{f_j}\big(\epsilon / 2\big)$. Note again that the
sharper analysis mentioned above would allow us to remove the
$\log{2d}$ factors.
\paragraph{Random functions.}
A more interesting example comes from considering a randomly chosen
function $f \colon C_{2d} \rightarrow [0,~1]$. Indeed, suppose that
the function values are sampled iid from some distribution $\nu$ on
$[0,~1]$ whose mean $\mu^{\ast}$ is $1/2$:
\begin{align}\label{eq:f-random}
\{ f(u), ~u \in C_{2d} \} ~ \overset{\text{iid}}{\sim} ~ \nu .
\end{align}
We can then show that for any fixed $\delta^{\ast} > 0$, with high
probability over the randomness of $f$, have
\begin{align}
\label{eq:Tf-random}
T_{f} (\delta) & \lesssim \frac{d \log{d} \big[\log{d} + \log \big(
\frac{1}{\delta} \big) \big]}{\delta^2 }, \qquad \mbox{for all
$\delta \in (0, \delta^\ast]$.}
\end{align}
For $\delta \gg \frac{\log{d}}{\sqrt{d}}$, this scaling is an
improvement over the global mixing time of order $d^2 \log(1/\delta)$.
The core idea behind the proof of equation~\eqref{eq:Tf-random} is to
apply Lemma~\ref{lem:mix-gap-all} with
\begin{align}
\label{eq:J-random}
J_{\delta} & \mydefn \left\{ j \in \Nat \cap [1, 2d-1] \; \mid \; j
\leq 4 \delta \sqrt{\frac{d}{\log{d}}}~~ \text{or} ~~ j \geq 2d - 4
\delta \sqrt{\frac{d}{\log{d}}} \right\} .
\end{align}
It can be shown that $\| h_{j}\|_{\infty} = 1$ for all $0 \leq j < 2d$
and that with high probability over $f$, $|q_{j}^{T}f| \lesssim
\sqrt{\frac{\log{d}}{d}}$ simultaneously for all $j \in J_{\delta}$,
which suffices to reduce the first part of the sharper $f$-discrepancy
bound to order $\delta$.
In order to estimate the rate of concentration, we proceed as
follows. Taking $\delta = c_{0} \epsilon$ for a suitably chosen
universal constant $c_{0} > 0$, we show that $\Delta_{J} \mydefn
\frac{\epsilon}{4} \geq \Delta_{J}^{\ast}$. We can then set $\Delta =
\frac{\epsilon}{4}$ and observe that with high probability over $f$,
the deviation in Corollary~\ref{cor:hoeffding-derived-Jf} satisfies
the bound $ 2\left(\Delta_{J} + \Delta\right) \leq \epsilon$. With
$\delta$ as above, we have $1 - \lambda_{-J} \geq
\frac{c_{1}\epsilon^{2}}{d\log{d}}$ for another universal constant
$c_{1} > 0$. Thus, if we are given $N + \Tf\big(\epsilon/2\big)$
samples for some $N \geq \Tf\big(\frac{\epsilon}{2}\big)$, then we
have
\begin{align}
\label{eq:conc-random}
\P \left[ \frac{1}{N} \sum_{n = \Tf(\epsilon/2)}^{N + \Tf (
\epsilon/2)} f(X_n) \geq \mu + \epsilon \right] \leq \exp
\left \{ -\frac{c_2 \epsilon^{4}N}{d \log{d }\big[ \log \big(
\frac{4}{\epsilon} \big) + \log{2d} \big ] } \right \},
\end{align}
for some $c_2 > 0$. Consequently, it suffices for the sample size to be
lower bounded by
\begin{align*}
N \gtrsim \frac{d \log{d}\big[\log\big(1/\epsilon\big) +
\log{d}\big]}{\epsilon^{4}},
\end{align*}
in order to achieve an estimation accuracy of $\epsilon$. Notice that
this requirement is an improvement over the $\frac{d^2 }{\epsilon^2 }$
from the uniform Hoeffding bound provided that $\epsilon \gg (\frac{
\log^2 {d}} {d})^{1/2}$. Proofs of all these claims can be found in
Appendix~\ref{app:proof-example-cycle}.
\subsection{Lower bounds}
\label{sec:lower}
Let us now make precise the intuitive argument set forth at the end of
Section~\ref{subsec:master-hoeffding}. The basic idea is to start
with an arbitrary candidate function $\delta : \left(0,~1\right)
\rightarrow \left(0,~1\right)$ such that
$\Tf\left(\frac{\epsilon}{2}\right)$ in the denominator of the
function-specific Hoeffding bound \eqref{eq:hoeffding-eps2} can be
replaced by $\Tf\left(\delta\left(\epsilon\right)\right)$ and show
that if $\delta\left(\epsilon\right) \geq \epsilon$, the replacement
is not actually possible. We prove this fact by constructing a Markov
chain (which is independent of $\epsilon$) and a function (which
depends on both $\epsilon$ and $\delta$) such that the Hoeffding bound
is violated for the Markov chain-function pair for some value of $N$
(which in general depends on the chain and $\epsilon$).
As the following precise result shows, our lower bound continues to
hold for an arbitrary constant in the exponent of the Hoeffding bound,
meaning that Theorem~\ref{thm:hoeffding-eps2} is optimal up to
constants. We give the proof in Section~\ref{subsec:proofs-lower}.
\begin{prop}
\label{prop:lower-bound}
For every constant $c_{1} > 0$ and $\epsilon \in (0,1)$, there exists
a Markov chain $P_{c_1}$, a number of steps $N = N(c_1,\epsilon)$ and
a function $f = f_{\epsilon}$ such that
\begin{align}
\P_{\pi}\left(\left|\frac{1}{N} \sum_{n = 1}^{N} f(X_{n}) -
\frac{1}{2}\right| \geq \epsilon\right) > 2 \cdot \exp \left(
-\frac{c_{1} N \epsilon^{2}}{\Tf \left (\delta
(\epsilon)\right)}\right).
\end{align}
\end{prop}
\section{Statistical applications}
\label{sec:stat-apps}
We now consider how our results apply to Markov chain Monte Carlo
(MCMC) in various statistical settings. Our investigation proceeds
along three connected avenues. We begin by showing, in
Section~\ref{subsec:confidence}, how our concentration bounds can be
used to provide confidence intervals for stationary expectations that
avoid the over-optimism of pure CLT predictions without incurring the
prohibitive penalty of the Berry-Esseen correction---or the global
mixing rate penalty associated with spectral-gap-based confidence
intervals. Then, in Section~\ref{subsec:testing}, we show how our results
allow us to improve on recent sequential hypothesis testing
methodologies for MCMC, again replacing the dependence on the spectral
gap by a dependence on the $f$-mixing time. Later, in
Section~\ref{subsec:practical}, we illustrate the practical
significance of function-specific mixing properties by using our
framework to analyze three real-world instances of MCMC, basing both
the models and datasets chosen on real examples from the literature.
\subsection{Confidence intervals for posterior expectations}
\label{subsec:confidence}
In many applications, a point estimate of $\E_{\pi}\big[f\big]$ does
not suffice; the uncertainty in the estimate must be quantified, for
instance by providing $(1-\alpha)$ confidence intervals for some pre-specified constant $\alpha$. In this section, we
discuss how improved concentration bounds can be used to obtain
sharper confidence intervals. In all cases, we assume the Markov chain
is started from some distribution $\pi_0$ that need not be the
stationary distribution, meaning that the confidence intervals must
account for the burn-in time required to get close to equilibrium.
We first consider a bound that is an immediate consequence of the
uniform Hoeffding bound given by~\cite{Leo04Hoeffding}. As one would
expect, it gives contraction at the usual Hoeffding rate but with an
effective sample size of $N_{\eff} \approx \gamma_{0} (N - T_{0})$,
where $T_{0}$ is the tuneable burn-in parameter. Note that this means
that no matter how small $T_{f}$ is compared to the global mixing time
$T$, the effective size incurs the penalty for a global burn-in and
the effective sample size is determined by the global spectral
parameter $\gamma_{0}$. In order to make this precise, for a fixed
burn-in level $\alpha_0 \in (0, \alpha)$, define
\begin{subequations}
\begin{align}
\epsilon_{N}( \alpha,~\alpha_{0}) & \mydefn \sqrt{2\big(2 -
\gamma_{0}\big)} \cdot \sqrt{\frac{\log\big(2/\big[\alpha -
\alpha_0\big]\big)}{\gamma_{0}\big[N -
T\big(\alpha_{0}\big)\big]}}.
\end{align}
Then the uniform Markov Hoeffding bound~\cite[Thm. 1]{Leo04Hoeffding}
implies that the set
\begin{align}
\label{eq:confidence-unif-hoeffding}
\UNIFINT \big( \alpha,~\alpha_{0} \big) = \left[ \frac{1}{N - T \big(
\alpha_0/2 \big)} \sum_{n = T \big( \alpha_{0}/ 2 \big) + 1}^{N} f
\big( X_{n} \big) \pm \epsilon_{N} \big( \alpha, ~\alpha_{0}\big)
\right]
\end{align}
\end{subequations}
is a $1-\alpha$ confidence interval. Full details of the proof are
given in Appendix~\ref{appsec:confidence-unif-hoeffding}.
Moreover, given that we have a family of confidence intervals---one
for each choice of $\alpha_0 \in (0, \alpha)$---we can obtain the
sharpest confidence interval by computing the infimum
$\epsilon_{N}^{\ast}\big(\alpha\big) \mydefn \inf \limits_{0 <
\alpha_{0} < \alpha} \epsilon_{N}\big(\alpha,~\alpha_{0}\big)$.
Equation~\eqref{eq:confidence-unif-hoeffding} then implies that
\begin{align*}
\UNIFINT\big(\alpha\big) = \Big[\frac{1}{N - T
\big(\alpha_0\big)}\sum_{n = T(\alpha_{0}/2) + 1}^{N} f (X_{n})
\pm \epsilon_{N}^{\ast}( \alpha) \Big]
\end{align*}
is a $1 - \alpha$ confidence interval for $\mu$.
We now consider one particular application of our Hoeffding bounds to
confidence intervals, and find that the resulting interval adapts to
the function, both in terms of burn-in time required, which now falls
from a global mixing time to an $f$-specific mixing time, and in terms
of rate, which falls from $\frac{1}{\gamma_{0}}$ to $T_{f}(\delta)$
for an appropriately chosen $\delta > 0$. We first note that the
one-sided tail bound of Theorem~\ref{thm:hoeffding-eps2} can be
written as $e^{-r_{N}(\epsilon)/8}$, where
\begin{align}
\label{eq:rN-def}
r_{N} (\epsilon) & \mydefn \epsilon^2 \left [ \frac{N}{T_{f} \big(
\frac{\epsilon}{2} \big)} - 1 \right ].
\end{align}
If we wish for each tail to have probability mass that is at most
$\alpha/2$, we need to choose $\epsilon > 0$ so that $r_{N} \big (
\epsilon \big) \geq 8 \log\frac2 {\alpha}$, and conversely any such
$\epsilon$ corresponds to a valid two-sided $\big(1 - \alpha\big)$
confidence interval. Let us summarize our conclusions:
\begin{thm}
\label{thm:confidence-adaptive-hoeffding}
For any width $\epsilon_{N} \in r_{N}^{-1}\big(\big[8
\log\big(2/\alpha\big),~\infty\big)\big)$, the set
\begin{align*}
\FUNCINT & \mydefn \left[ \frac{1}{N - T_{f}\big(\frac{\epsilon
}{2} \big)} \sum_{n = T_{f} \big( \frac{\epsilon}{2}
\big)}^{N} f\big(X_{n}\big) \pm \epsilon_{N} \right]
\end{align*}
is a $1 - \alpha$ confidence interval for the mean $\mu =
\E_{\pi}\big[f\big]$.
\end{thm}
In order to make the result more amenable to interpretation, first
note that for any $0 < \eta < 1$, we have
\begin{align}
\label{eq:rN-eta-def}
r_{N}\big(\epsilon\big) \geq \underbrace{\epsilon^2
\left[\frac{N}{T_{f}\big(\frac{\eta}{2}\big)} - 1\right]}_{
r_{N,\eta}(\epsilon)} \quad \mbox{valid for all $\epsilon \geq
\eta$.}
\end{align}
Consequently, whenever $r_{N,\eta}(\epsilon_{N}) \geq 8 \log\frac2
{\alpha}$ and $\epsilon_{N} \geq \eta$, we are guaranteed that a
symmetric interval of half-width $\epsilon_{N}$ is a valid $\big(1 -
\alpha\big)$-confidence interval. Summarizing more precisely, we
have:
\begin{cor}
\label{cor:confidence-adaptive-hoeffding-concrete}
Fix $\eta > 0$ and let
\begin{align*}
\epsilon_{N} = r_{N,\eta}^{-1}\big(8 \log \frac2 {\alpha}\big) & = 2\sqrt{2}
\sqrt{\frac{T_{f}\big(\frac{\eta}{2}\big) \cdot \log\big (2/
\alpha \big)}{N - T_{f} \big( \frac{\eta}{2} \big)}} .
\end{align*}
If $N \geq \Tf\big(\frac{\eta}{2}\big)$, then $\FUNCINT$ is a $1 -
\alpha$ confidence interval for $\mu = \E_{\pi}\big[f\big]$.
\end{cor}
Often, we do not have direct access to $T_{f}\big(\delta\big)$, but we
can often obtain an upper bound $\tilde{T}_{f}\big(\delta\big)$ that
is valid for all $\delta > 0$. In
Section~\ref{subsec:proofs-confidence}, therefore, which contains the
proofs for this section, we prove a strengthened form of Theorem
\ref{thm:confidence-adaptive-hoeffding} and its corollary in that
setting.
A popular alternative strategy for building confidence intervals using
MCMC depends on the Markov central limit theorem
(e.g.,~\citep{Fle08Trust,Jon01Honest,Gly09BatchMeans,Rob05MCMC}). If
the Markov CLT held exactly, it would lead to appealingly simple
confidence intervals of width
\begin{align*}
\tilde{\epsilon}_{N} = \sigfasym \: \sqrt{ \frac{\log (2/\alpha)}{N}},
\end{align*}
where $\sigfasym^2 \mydefn \lim_{N \rightarrow \infty} \frac{1}{N}
\Var_{X_{0} \sim \pi} \big [ \sum_{n = 1}^{N} f \big( X_{n} \big)
\big]$ is the asymptotic variance of $f$.
Unfortunately, the CLT does not hold exactly, even after the burn-in
period. The amount by which it fails to hold can be quantified using a
Berry-Esseen bound for Markov chains, as we now discuss. Let us adopt
the compact notation
$\tilde{S}_{N} = \sum_{n = 1}^{N} \big[ f \big( X_{n} \big) - \mu
\big].$
We then have the bound~\citep{Lez01Berry}
\begin{align}
\label{eq:markov-clt-be}
\big| \P \big( \frac{ \tilde{S}_{N}}{\sigfasym \sqrt{N}} \leq s \big)
- \Phi \big(s\big)\big| \leq \frac{e^{- \gamma_{0} N } }{3
\sqrt{\pimin}} + \frac{13}{\sigfasym \sqrt{\pimin}} \cdot
\frac{1}{\gamma_{0}\sqrt{N}},
\end{align}
where $\Phi$ is the standard normal CDF. Note that this bound
accounts for both the non-stationarity error and for non-normality
error at stationarity. The former decays rapidly at the rate $e^{-\gamma_{0}N}$, while the latter decays far more slowly, at the rate
$\frac{1}{\gamma_{0}\sqrt{N}}$.
While the bound~\eqref{eq:markov-clt-be} makes it possible to prove a
corrected CLT confidence interval, the resulting bound has two
significant drawbacks. The first is that it only holds for extremely
large sample sizes, on the order of $\frac{1}{\pimin\gamma_{0}^2 }$,
compared to the order $\frac{\log\big(1/\pimin\big)}{\gamma_{0}}$
required by the uniform Hoeffding bound. The second, shared by the
uniform Hoeffding bound, is that it is non-adaptive and therefore
bottlenecked by the global mixing properties of the chain. For
instance, if the sample size is bounded below as
\begin{align*}
N \geq \max\big(\frac{1}{\gamma_{0}}\log\big(\frac2 {\sqrt{\pimin}\alpha}\big)
~, ~\frac{1}{\gamma_{0}^2 }\frac{6084}{\sigfasym^2 \pimin \alpha^2 } \big),
\end{align*}
then both terms of equation~\eqref{eq:confidence-unif-hoeffding} are
bounded by $1/6$, and the confidence intervals take the form
\begin{align}
\label{eq:confidence-clt-be}
\BEINT = \left[\frac{1}{N}\sum_{n = 1}^{N} f\big(X_{n}\big) \pm
\sigfasym \sqrt{\frac{2 \log \big (6 / \alpha \big)} {N}} \right].
\end{align}
See Appendix~\ref{appsec:confidence-clt-be} for the justification of
this claim.
It is important to note that the width of this confidence interval
involves a hidden form of mixing penalty. Indeed, defining the
variance $\sigf^2 = \Var_{\pi}\big[f\big(X\big)\big]$ and $\rho_{f}
\mydefn \frac{\sigf^2 }{\sigfasym^2 }$, we can rewrite the width as
\begin{align*}
\epsilon_{N} & = \sigf \sqrt{ \frac{2 \log \big ( 6 /\alpha \big)}{
\rho_{f} N} }.
\end{align*}
Thus, for this bound, the quantity $\rho_{f}$ captures the penalty due
to non-independence, playing the role of $\gamma_{0}$ and $\gamma_{f}$
in the other bounds. In this sense, the CLT bound adapts to the
function $f$, but only when it applies, which is at a sample-size
scale dictated by the global mixing properties of the chain (i.e.,
$\gamma_{0}$).
\subsection{Sequential testing for MCMC}
\label{subsec:testing}
For some applications, full confidence intervals may be unnecessary;
instead, a practitioner may merely want to know whether $\mu =
\E_{\pi}[f]$ lies above or below some threshold $0 < r < 1$. In these
cases, we would like to develop a procedure for distinguishing between
the two possibilities, at a given tolerable level $0 < \alpha < 1$ of
combined Type I and II error. The simplest approach is, of course, to
choose $N$ so large that the $1 - \alpha$ confidence interval built
from $N$ MCMC samples lies entirely on one side of $r$, but it may be
possible to do better by using a sequential test. This latter idea was
recently investigated in \citet{Gyo15Test}, and we consider the same
problem settings that they did:
\begin{enumerate}[leftmargin=*]
\item[(a)] Testing with (known) indifference region, involving a
choice between
\begin{align*}
H_{0} & \colon \mu \geq r + \delta \\
H_{1} & \colon \mu \leq r - \delta;
\end{align*}
\item[(b)] Testing with no indifference region---that is, the same as
above but with $\delta = 0$.
\end{enumerate}
For the first setting (a), we always assume $0 < \delta < \nu \mydefn \min(\mu,1-\mu)$, and the
algorithm is evaluated on its ability to correctly choose between
$H_{0}$ and $H_{1}$ when one of them holds, but it incurs no penalty
for either choice when $\mu$ falls in the indifference region $\big(r
- \delta,~r + \delta\big)$. The error of a procedure $\alg$ can thus
be defined as
\begin{align*}
\err\big(\alg,~f\big) = \begin{cases}
\P\big(\alg\big(X_{1:\infty}\big) = H_{1}\big) & ~ \text{if}~ \mu
\in H_{0}, \\ \P\big(\alg\big(X_{1:\infty}\big) = H_{0}\big) & ~
\text{if}~ \mu \in H_{1}, \\ 0 &~ \text{otherwise.}
\end{cases}
\end{align*}
The rest of this subsection is organized as follows. For the first
setting (a), we analyze a procedure $\algfix$ that makes a decision
after a fixed number $N := N(\alpha)$ of samples. We also analyze a
sequential procedure $\algseq$ that chooses whether to reject
at a sequence $N_0,\ldots,N_k,\ldots$ of decision times. For the
second, more challenging, setting (b), we analyze $\algdiff$, which also
rejects at a sequence of decision times. For both $\algseq$ and
$\algdiff$, we calculate the expected stopping times of the
procedures.
As mentioned above, the simplest procedure $\algfix$ would choose a
fixed number $N$ of samples to be collected based on the target level
$\alpha$. After collecting $N$ samples, it forms the empirical average
$\hat{\mu}_{N} = \frac{1}{N}\sum_{n = 1}^{N} f\big(X_{n}\big)$ and
outputs $H_{0}$ if $\hat{\mu}_{N} \geq r + \delta$, $H_{1}$ if
$\hat{\mu}_{N} \leq r - \delta$, and outputs a special indifference
symbol, say $\mathrm{I}$, otherwise.
The sequential algorithm $\algseq$ makes decisions as to whether to
output one of the hypotheses or continue testing at a fixed sequence
of decision times, say $N_{k}$. These times are defined recursively
by
\begin{align}
N_{0} & = \big\lfloor M \cdot \min\big(\frac{1}{r},~\frac{1}{1 -
r}\big) \big\rfloor, \label{eq:seq-seq-N0} \\ N_{k} & = \big\lfloor
N_{0}\big(1 + \xi\big)^{k} \big\rfloor \label{eq:seq-seq-Nk},
\end{align}
where $M > 0$ and $0 < \xi < 2/5$ are parameters of the algorithm. At
each time $N_{k}$ for $k \geq 1$, the algorithm $\algseq$ checks if
\begin{align}
\label{eq:seq-indiff-stopping}
\hat{\mu}_{N_k} \in \big(r - \frac{M}{N_k}, r + \frac{M}{N_k} \big) .
\end{align}
If the empirical average lies in this interval, then the algorithm
continues sampling; otherwise, it outputs $H_0$ or $H_1$ accordingly
in the natural way.
For the sequential algorithm $\algdiff$, let $N_{0} > 0$ be chosen
arbitrarily,\footnote{In \citet{Gyo15Test}, the authors set $N_{0}
= \big\lfloor \frac{100}{\gamma_{0}}\big\rfloor$, but this is
inessential.} and let $N_{k}$ be defined in terms of $N_{0}$ as in
\eqref{eq:seq-seq-Nk}. It once again decides at each $N_{k}$ for $k
\geq 1$ whether to output an answer or to continue sampling, depending
on whether
\begin{align*}
\hat{\mu}_{N_k} \in \big(r - \epsilon_{k}\big(\alpha\big),~r +
\epsilon_{k}\big(\alpha\big)\big) .
\end{align*}
When this inclusion holds, the algorithm continues; when it doesn't
hold, the algorithm outputs $H_0$ or $H_1$ in the natural way.
The following result is restricted to the stationary case; later in
the section, we turn to the question of burn-in.
\begin{thm}
\label{thm:seq-err-all-adapt}
Assume that $\alpha \leq \frac2 {5}$. For $\algfix, \algseq,
\algdiff$ to all satisfy $ \err\big(\alg,~f\big) ~\leq~ \alpha $, it
suffices to (respectively) choose
\begin{align}
N &= \frac{2\Tf\big(\delta\big)
\log\big(\frac{1}{\alpha}\big)}{\delta^2 }, \\ M &=
\frac{8\Tf\big(\frac{\delta}2\big) \log\big(\frac2
{\sqrt{\alpha\xi}}\big)}{\delta},~ \text{and} \\
\label{eq:seq-seq-diff-adapt-eps2-epsk}
\epsilon_{k}\big(\alpha\big) & = \inf \Big \{ \epsilon > 0 \colon
\frac{\epsilon^2 }{8\Tf\big(\frac{\epsilon}{2}\big)} \geq
\frac{ \log \big(1 / \alpha \big) + 1 + 2 \log{k}}{N_{k}} \Big \},
\end{align}
where we let $\inf \emptyset = \infty$.
\end{thm}
Our results differ from those of~\cite{Gyo15Test} because the latter
implicitly control the worst-case error of the algorithm
\begin{align*}
\err\big(\alg\big) = \sup_{f \colon \Omega \rightarrow [0,~1]}
\err\big(\alg,~f\big),
\end{align*}
while our analysis controls $\err\big(\alg,~f\big)$ directly. The corresponding choices made in \cite{Gyo15Test} are
\begin{align*}
N = \frac{\log(1/\alpha)}{\gamma_{0}\delta^2 },
M = \frac{\log (\frac2
{\sqrt{\alpha\xi}})}{\gamma_{0}\delta} , \text{ and }
\epsilon_{k}(\alpha) = \sqrt{\frac{\log (1/\alpha) + 1 + 2 \log{k}}{\gamma_{0}N_{k}}}.
\end{align*}
Hence, the $\Tf$ parameter in
our bounds plays the same role that $\frac{1}{\gamma_0}$
plays in their uniform bounds.
As a result of this close correspondence, we easily see that our
results improve on the uniform result for a fixed function $f$
whenever it converges to its stationary expectation faster than the
chain itself converges--- i.e., whenever
$\Tf\big(\delta\big) \leq \frac{1}{2\gamma_{0}}$.
The value of the above tests depends substantially on their sample
size requirements. In setting (a), algorithm $\algseq$ is only
valuable if it reduces the number of samples needed compared to
$\algfix$. In setting (b), algorithm $\algdiff$ is valuable because of
its ability to test between hypotheses separated only by a point, but
its utility is limited if it takes too long to run. Therefore, we now
turn to the question of bounding expected stopping times.
In order to o carry out the stopping time analysis, we introduce the
true margin $\Delta = |r - \mu |$. First, let us introduce some
useful notation. Let $N(\alg)$ be the number of sampled collected by
$\alg$. Given a margin schedule $\big(\epsilon_{k}\big)$, let
\begin{align*}
k_{0}^{\ast}\big(\epsilon_{1:\infty}\big) \mydefn \min \big \{ k \geq 1
\colon \epsilon_{k} \leq \frac{\Delta}2 \big \}, \text{ and } N_{0}^{\ast}\big(\epsilon_{1:\infty}\big) \mydefn
N_{k_{0}^{\ast}\big(\epsilon_{1:\infty}\big)}.
\end{align*}
We can bound the expected stopping times of $\algseq,\algdiff$ in
terms of $\Delta$ as follows:
\begin{thm}
\label{thm:seq-stopping-time-all}
Assume either $H_{0}$ or $H_{1}$ holds. Then,
\begin{align}
\E\big[N(\algseq)\big] &\leq \big(1 + \xi\big)\big[\frac{M}{\Delta} +
\frac4 {\Delta}\sqrt{\frac{2\Tf\big(\delta/2\big)M}{\Delta} +
8\Tf\big(\delta/2\big)} + 1\big];\\ \E\big[N(\algdiff)\big]
&\leq \big(1 + \xi\big)\big(N_{0}^{\ast} + 1\big) + \frac{32\alpha
T_{f}\big(\Delta/4\big)}{\Delta^2 }.
\end{align}
\end{thm}
With minor modifications to the proofs in~\cite{Gyo15Test}, we can
bound the expected stopping times of their procedures as
\begin{align*}
\E [N(\algseq)] &\leq (1 + \xi) \Big \{ \frac{M}{\Delta} + \frac2
{\Delta}\sqrt{\frac{M}{\gamma_{0}\Delta} + \frac{4}{\gamma_{0}}} +
1 \Big \};\\
\E [ N (\algdiff)] & \leq (1 + \xi) \big(N_{0}^{\ast} + 1\big) +
\frac{4\alpha}{\gamma_{0}\Delta^2 }.
\end{align*}
In order to see how the uniform and adaptive bounds compare, it is
helpful to first note that, under either $H_{0}$ or $H_{1}$, we have
the lower bound $\Delta \geq \delta$. Thus, the dominant term in the
expectations in both cases is $(1 + \xi)M/\Delta$. Consequently, the
ratio between the expected stopping times is approximately equal to the
ratio between the $M$ values---viz.,
\begin{align}
\frac{M_{\mathrm{adapt}}}{M_{\mathrm{unif}}} \approx \gamma_{0}\Tf\big(\delta/2\big).
\end{align}
As a result, we should expect a significant improvement in terms of
number of samples when the relaxation time $\frac{1}{\gamma_{0}}$ is
significantly larger than the \mbox{$f$-mixing} time
$\Tf\big(\delta/2\big)$. Framed in absolute terms, we
can write
\begin{align*}
\bar{N}_{\mathrm{unif}} (\algseq) \approx
\frac{\log\big(2/\sqrt{\alpha\xi}\big)}{\gamma_{0}\delta\Delta} \quad
\mbox{and} \quad \bar{N}_{\mathrm{adapt}} (\algseq) \approx
\frac{\Tf\big ( \delta / 2 \big) \log \big( 2 / \sqrt{\alpha
\xi} \big)}{\delta \Delta}.
\end{align*}
Up to an additive term, the bound for $\algdiff$ is also qualitatively
similar to earlier ones, with $\frac{1}{\delta\Delta}$ replaced by
$\frac{1}{\Delta^2 }$.
\section{Analyzing mixing in practice}
\label{subsec:practical}
We analyze several examples of MCMC-based Bayesian analysis
from our theoretical perspective. These examples demonstrate that
convergence in discrepancy can in practice occur much faster than
suggested by naive mixing time bounds and that our bounds help narrow the gap between theoretical predictions and
observed behavior.
\begin{figure}[h]
\begin{center}
\begin{tabular}{cc}
\widgraph{0.5\linewidth}{figs/ORingMH-Spec-Fixed.pdf} &
\widgraph{0.5\linewidth}{figs/ClinicalTrial-p-Spec.pdf} \\
(a) & (b)
\end{tabular}
\begin{tabular}{c}
\widgraph{0.5\linewidth}{figs/SchizoMixture-CGibbs-Spec.pdf} \\
(c)
\end{tabular}
\caption{Spectra for three example chains: (a) Metropolis-Hastings for
Bayesian logistic regression; (b) collapsed Gibbs sampler for
missing data imputation; and (c) collapsed Gibbs sampler for a
mixture model.
\label{fig:example-spectra}}
\end{center}
\end{figure}
\subsection{Bayesian logistic regression}
Our first example is a Bayesian logistic regression problem
introduced by \citet{Rob05MCMC}. The data consists of $23$
observations of temperatures (in Fahrenheit, but normalized by
dividing by $100$) and a corresponding binary outcome---failure $(y =
1)$ or not $(y = 0)$ of a certain component; the aim is to fit a
logistic regressor, with parameters $\big(\alpha,~\beta\big) \in \R^2 $, to the
data, incorporating a prior and integrating over the model uncertainty
to obtain future predictions. More explicitly, following the analysis
in \citet{Gyo12Nonasym}, we consider the following model:
\begin{align*}
p \big( \alpha, ~\beta ~| ~b \big) & = \frac{1}{b} \cdot e^{\alpha}
\exp \big ( -e^{\alpha} / b \big) \\
p\big(y \, \mid \, \alpha,~\beta,~x\big) & \propto \exp
\big(\alpha + \beta x \big),
\end{align*}
which corresponds to an exponential prior on $e^{\alpha}$, an
improper uniform prior on $\beta$ and a logit link for prediction. As
in \citet{Gyo12Nonasym}, we target the posterior by running
a Metropolis-Hastings algorithm with a Gaussian proposal with covariance matrix
$\Sigma ~=~ \begin{pmatrix} 4 & 0 \\ 0 & 10
\end{pmatrix}.$
Unlike in their paper, however, we discretize the state space to
facilitate exact analysis of the transition matrix and to make our
theory directly applicable. The resulting state space is given by
\begin{align*}
\Omega = \Big \{ \big(\hat{\alpha} \pm i \cdot \Delta,~\hat{\beta} \pm
j \cdot \Delta\big) \, \mid \, 0 \leq i,~j \leq 8 \Big \},
\end{align*}
where $\Delta = 0.1$ and \mbox{$(\hat{\alpha},~\hat{\beta})$} is the
MLE. This space has $d = 17^2 = 289$ elements, resulting in a $289
\times 289$ transition matrix that can easily be diagonalized.
\citet{Rob05MCMC} analyze the probability of failure when the
temperature $x$ is $\SI{65}{\degree}\mathrm{F}$; it is specified by
the function
\begin{align*}
f_{65} \big( \alpha,~\beta \big) = \frac{\exp \big( \alpha + 0.65
\beta \big)}{1 + \exp \big( \alpha + 0.65 \beta \big)} .
\end{align*}
Note that this function fluctuates significantly under the posterior,
as shown in Figure~\ref{fig:oring-p65-hist}.
\begin{figure}[h!]
\centering
\widgraph{0.5\linewidth}{figs/ORingMH-P1_65-hist-Fixed.pdf}
\caption{Distribution of $f_{65}$ values under the posterior. Despite
the discretization and truncation to a square, it generally matches
the one displayed in Figure 1.2 in~\citet{Rob05MCMC}.
\label{fig:oring-p65-hist}}
\end{figure}
We find that this function also happens to exhibit rapid mixing. The discrepancy
$d_{f_{65}}$, before entering an asymptotic regime in which it decays
exponentially at a rate $1 - \gamma^{\ast} \approx 0.386$, first drops
from about $0.3$ to about $0.01$ in just $2$ iterations, compared to
the predicted $10$ iterations from the naive bound
$
d_{f} \big( n \big) \leq d_{\TV} \big( n \big) \leq
\frac{1}{\sqrt{\pimin}} \cdot \big(1 - \gamma^{\ast}\big)^{n} .
$
Figure~\ref{fig:oring-discrep-1} demonstrates this on a log scale,
comparing the naive bound to a version of the bound in
Lemmas~\ref{lem:mix-gap-all} and~\ref{lem:mix-gap-all-oracle}. Note
that the oracle $f$-discrepancy bound improves significantly over the
uniform baseline, even though the non-oracle version does not. In
this calculation, we took $J = \big \{ 2, \dots, 140 \big \}$ to
include the top half of the spectrum excluding $1$ and computed $\ \|
h_{j} \|_{\infty}$ directly from $P$ for $j \in J$ and likewise for
$q_{j}^{T} f_{65}$. The oracle bound is given by
Lemma~\ref{lem:mix-gap-all-oracle}. As shown in
panel (b) of Figure~\ref{fig:oring-discrep-1}, this decay is also faster than that
of the total variation distance.
\begin{figure}[h!]
\begin{tabular}{cc}
\widgraph{0.5\linewidth}{figs/ORingMH-Discrep-1-Fixed.pdf} &
\widgraph{0.5\linewidth}{figs/ORingMH-df-vs-TV.pdf}\\
(a) & (b)
\end{tabular}
\caption{(a) Discrepancies (plotted on log-scale) for $f_{65}$ as a
function of iteration number. The prediction of the naive bound is
highly pessimistic; the $f$-discrepancy bound goes part of the way toward
closing the gap and the oracle version of the $f$-discrepancy bound nearly
completely closes the gap in the limit and also gets much closer to
the right answer for small iteration
numbers. (b) Comparison of the function discrepancy $d_{f_{65}}$ and the
total variation discrepancy $d_{\TV}$. They both decay fairly
quickly due to the large spectral gap, but the function discrepancy
still falls much faster.
\label{fig:oring-discrep-1}}
\end{figure}
An important point is that the quality of the $f$-discrepancy bound depends
significantly on the choice of $J$. In the limiting case where $J$
includes the whole spectrum below the top eigenvalue, the oracle bound
becomes exact. Between that and $J = \emptyset$, the oracle bound
becomes tighter and tighter, with the rate of tightening depending on
how much power the function has in the higher versus lower
eigenspaces. Figure \ref{fig:oring-discrep-Js} illustrates this for a
few settings of $J$, showing that although for this function and this
chain, a comparatively large $J$ is needed to get a tight bound, the
oracle bound is substantially tighter than the uniform and non-oracle
$f$-discrepancy bounds even for small $J$.
\begin{figure}[h!]
\begin{center}
\begin{tabular}{cc}
\widgraph{0.5 \linewidth}{figs/ORingMH-Discrep-J50-Fixed.pdf} &
\widgraph{0.5 \linewidth}{figs/ORingMH-Discrep-J100-Fixed.pdf} \\
(a) & (b) \\
\widgraph{0.5 \linewidth}{figs/ORingMH-Discrep-J200-Fixed.pdf} &
\widgraph{0.5 \linewidth}{figs/ORingMH-Discrep-J287-Fixed.pdf} \\
(c) & (d)
\end{tabular}
\end{center}
\caption{Comparisons of the uniform, non-oracle function-specific, and oracle
function-specific bounds for various choices of $J$. In each case, $J = \{ 2,
\dots, \Jmax \}$, with $\Jmax = 50$ in panel (a), $\Jmax = 100$ in
panel (b), $\Jmax = 200$ in panel (c), and $\Jmax = 288$ in panel
(d). The oracle bound becomes tight in the limit as $\Jmax$ goes to
$d = 289$, but it offers an improvement over the uniform bound
across the board.}
\label{fig:oring-discrep-Js}
\end{figure}
\subsection{Bayesian analysis of clinical trials}
The problem of missing data often necessitates Bayesian analysis,
particularly in settings where uncertainty quantification is
important, as in clinical trials. We illustrate how our framework
would apply in this context by considering a clinical trials
dataset~\citep{Ber10Clinical,Gyo12Nonasym}.
The dataset consists of $n = 50$ patients, some of whom participated
in a trial for a drug and exhibited early indicators ($Y_{i}$) of
success/failure and final indicators ($X_{i}$) of success/failure. Among
the 50 patients, both indicator values are available for $n_{X} = 20$ patients;
early indicators are available for $n_{Y} = 20$ patients; and no indicators are
available for $n_{0} = 10$ patients. The analysis depends on the following
parameterization:
\begin{align*}
\P\big(X_{i} = 1~ \mid ~Y_{i} = 0\big) & = \gamma_{0}, \\ \P\big(X_{i}
= 1~ \mid ~Y_{i} = 1\big) & = \gamma_{1}, \\
\P\big(X_{i} = 1~ \mid ~Y_{i}~\text{missing}\big) & = p .
\end{align*}
Note that, in contrast to what one might expect, $p$ is to be
interpreted as the marginal probability that $X_{i} = 1$, so that in
actuality $p = \P\big(X_{i} = 1\big)$ unconditionally; we keep the
other notation, however, for the sake of consistency with past
work~\citep{Ber10Clinical,Gyo12Nonasym}. Conjugate uniform (i.e.,
$\Beta\big(1,~1\big)$) priors are placed on all the model parameters.
The unknown variables include the parameter triple
$\big(\gamma_{0},~\gamma_{1},~p\big)$ and the unobserved $X_{i}$
values for $n_{Y} + n_{0} = 30$ patients, and the full sample space is
therefore $\tilde{\Omega} = [0,~1]^{3} \times \big \{ 0,~1
\big \}^{30}$. We cannot estimate the transition matrix for this
chain, even with a discretization with as coarse a mesh as $\Delta =
0.1$, since the number of states would be $d = 10^{3} \times 2^{30}
\sim 10^{12}$. We therefore make two changes to the original MCMC
procedure. First, we collapse out the $X_{i}$ variables to bring the
state space down to $[0,~1]^{3}$; while analytically collapsing out
the discrete variables is impossible, we can estimate the transition
probabilities for the collapsed chain analytically by sampling the
$X_{i}$ variables conditional on the parameter values and forming a
Monte Carlo estimate of the collapsed transition
probabilities. Second, since the function of interest in the original
work---namely, $f\big(\gamma_0,~\gamma_1,~p\big) = \ones\big(p >
0.5\big)$---depends only on $p$, we fix $\gamma_{0}$ and $\gamma_{1}$
to their MLE values and sample only $p$, restricted to the unit
interval discretized with mesh $\Delta = 0.01$.
\begin{figure}[h!]
\centering
\widgraph{0.5\linewidth}{figs/ClinicalTrial-p-LogDiscDiffs.pdf}
\caption{Change in log discrepancy for the two functions $f(p) =
\ones\big(p \geq 0.5\big)$ and $f(p) = p$ considered above. Whereas
$f(p) = p$ always changes at the constant rate dictated by the
spectral gap, the indicator discrepancy decays more quickly in the
first few iterations.\label{fig:clinical-rates}}
\end{figure}
As Figure~\ref{fig:example-spectra} shows, eigenvalue decay occurs
rapidly for this sampler, with $\gamma^{\ast} \approx 0.86$. Mixing
thus occurs so quickly that none of the bounds---uniform or
function-specific---get close to the truth, due to the presence of the constant
terms (and specifically the large term $\frac{1}{\sqrt{\pimin}}
\approx 2.14 \times 10^{33}$). Nonetheless, this example still
illustrates how in actual fact, the choice of target function can make
a big difference in the number of iterations required for accurate
estimation; indeed, if we consider the two functions
\begin{align*}
f_1(p) \mydefn \ones\big(p > 0.5\big), \quad \mbox{and} \quad f_2(p)
\mydefn p,
\end{align*}
we see in Figure~\ref{fig:clinical-rates} that the mixing behavior
differs significantly between them: whereas the discrepancy for the
second decays at the asymptotic exponential rate from the
outset, the discrepancy for the first decreases faster
(by about an order of magnitude) for the first few
iterations, before reaching the asymptotic rate dictated by the spectral gap.
\subsection{Collapsed Gibbs sampling for mixture models}
Due to the ubiquity of clustering problems in applied statistics and machine
learning, Bayesian inference for mixture models (and their
generalizations) is a widespread application of MCMC
\citep{Gha05IBP,Gri04LDA,Jai07SplitMerge,Mim12SSVI,Nea00DP}. We
consider the mixture-of-Gaussians model, applying it to a
subset of the schizophrenic reaction time data analyzed in
\citet{Bel95Schizophrenia}. The subset of the data we consider
consists of $10$ measurements, with $5$ coming from healthy subjects
and $5$ from subjects diagnosed with schizophrenia. Since our
interest is in contexts where uncertainty is high, we chose the $5$
subjects from the healthy group whose reaction times were greatest
and the $5$ subjects from the schizophrenic group whose reaction
times were smallest. We considered a mixture with $K = 2$
components, viz.:
\begin{align*}
\mu_{b} & \sim \Norm\big(0,~\rho^2 \big),~ b = 0,~1, \\
\omega & \sim \Beta\big(\alpha_0,~\alpha_1\big) \\
Z_{i} ~|~ \omega & \sim \Bern\big(\omega\big) \\
X_{i} ~|~ Z_{i} = b,~\mu & \sim \Norm\big(\mu_{b},~\sigma^2 \big) .
\end{align*}
We chose relatively uninformative priors, setting $\alpha_0 = \alpha_1
= 1$ and $\rho = 237$. Increasing the value chosen in the original
analysis \citep{Bel95Schizophrenia}, we set $\sigma \approx 70$; we
found that this was necessary to prevent the posterior from being too
highly concentrated, which would be an unrealistic setting for
MCMC. We ran collapsed Gibbs on the indicator variables $Z_{i}$ by
analytically integrating out $\omega$ and $\mu_{0:1}$.
As Figure~\ref{fig:example-spectra} illustrates, the spectral gap for
this chain is small---namely, $\gamma_{\ast} \approx 3.83 \times
10^{-4}$---yet the eigenvalues fall off comparatively quickly after
$\lambda_2 $, opening up the possibility for improvement over the
uniform $\gamma_{\ast}$-based bounds. In more detail, define
\begin{align*}
z^{\ast}_{b} & \mydefn \begin{pmatrix} b & b & b & b & b & 1 - b & 1 -
b & 1 - b & 1 - b & 1 - b \end{pmatrix},
\end{align*}
corresponding to the cluster assignments in which the patient and
control groups are perfectly separated (with the control group being
assigned label $b$). We can then define the indicator for exact
recovery of the ground truth by
\begin{align*}
f(z) = \ones\big(z \in \big \{ z_{0}^{\ast},~z_{1}^{\ast}\big \}\big).
\end{align*}
As Figure~\ref{fig:mixture-DFvsTV} illustrates, convergence in terms
of $f$-discrepancy occurs much faster than convergence in total
variation, meaning that predictions of required burn-in times and
sample size based on global metrics of convergence drastically
overestimate the computational and statistical effort required to
estimate the expectation of $f$ accurately using the collapsed Gibbs
sampler. This behavior can be explained in terms of the interaction
between the function $f$ and the eigenspaces of $P$. Although the
pessimistic constants in the bounds from the uniform
bound~\eqref{eq:mix-TV} and the non-oracle function-specific bound
(Lemma~\ref{lem:mix-gap-all}) make their predictions overly
conservative, the oracle version of the function-specific bound
(Lemma~\ref{lem:mix-gap-all-oracle}) begins to make exact predictions
after just a hundred iterations when applied with $J = \big \{ 1,
\dots, 25 \big \}$; this corresponds to making exact predictions of
$T_{f}\big(\delta\big)$ for $\delta \leq \delta_{0} \approx 0.01$,
which is a realistic tolerance for estimation of $\mu$.
Panel (b) of Figure~\ref{fig:mixture-DFvsTV} documents this by
plotting the $f$-discrepancy oracle bound against the actual value of
$d_{f}$ on a log scale.
\begin{figure}[h!]
\begin{center}
\begin{tabular}{cc}
\widgraph{0.45\linewidth}{figs/SchizoMixture-DFvsTV.pdf} &
\widgraph{0.45\linewidth}{figs/SchizoMixture-ExactMatch-Disc.pdf}\\
(a) & (b)
\end{tabular}
\caption{(a) Comparison of the $f$-discrepancy $d_{f}$ and the total
variation discrepancy $d_{\TV}$ over the first $100$ iterations of
MCMC. Clearly the function mixes much faster than the overall
chain. (b) The predicted value of $\log{d_{f}}$ (according to the
$f$-discrepancy oracle bound---Lemma \ref{lem:mix-gap-all-oracle})
plotted against the true value. The predictions are close to sharp
throughout and become sharp at around $100$ iterations.}
\label{fig:mixture-DFvsTV}
\end{center}
\end{figure}
\begin{table}[h]
\begin{center}
\begin{tabular}{c|c|c}
Bound type & $T_{f}\big(0.01\big)$ & $T_{f}\big(10^{-6}\big)$
\\ \hline Uniform & 31,253 & 55,312 \\ Function-Specific & 25,374 & 49,434
\\ Function-Specific (Oracle) & 98 & 409 \\ Actual & 96 & 409
\end{tabular}
\caption{Comparison of bounds on $T_{f}\big(\delta\big)$ for different
values of $\delta$. The uniform bound corresponds to the bound
$T_{f}\big(\delta\big) \leq T\big(\delta\big)$, the latter of which
can be bounded by the total variation bound. The function-specific bounds
correspond to Lemmas~\ref{lem:mix-gap-all}
and~\ref{lem:mix-gap-all-oracle}, respectively. Whereas the uniform
and non-oracle $f$-discrepancy bounds make highly conservative predictions,
the oracle $f$-discrepancy bound is nearly sharp even for $\delta$ as large
as $0.01$.\label{tab:mixture-Tf-bds}}
\end{center}
\end{table}
The mixture setting also provides a good illustration of how the
function-specific Hoeffding bounds can substantially improve on the
uniform Hoeffding bound. In particular, let us compare the
$T_{f}$-based Hoeffding bound (Theorem~\ref{thm:hoeffding-eps2}) to
the uniform Hoeffding bound established by~\citet{Leo04Hoeffding}. At
equilibrium, the penalty for non-independence in our bounds is
$(2T_{f}(\epsilon/2))^{-1}$ compared to roughly
$\gamma^{-1}_{\ast}$ in the uniform bound. Importantly, however, our
concentration bound applies unchanged even when the chain has not
equilibrated, provided it has approximately equilibrated with respect
to $f$. As a consequence, our bound only requires a burn-in of
$T_{f}(\epsilon/2)$, whereas the uniform Hoeffding
bound does not directly apply for any finite burn-in. Table~\ref{tab:mixture-Tf-bds} illustrates the
size of these burn-in times in practice. This issue can
be addressed using the method of \citet{Pau12Conc}, but at the cost of
a burn-in dependent penalty $d_{\TV}(T_0) = \sup_{\pi_{0}}
d_{\TV}(\pi_{n},~\pi)$:
\begin{align}
\label{eq:uniform-hoeffding-burn-in}
\P \Big[ \frac{1}{N - T_{0}} \sum_{n = T_0}^{N} f(X_{n}) \geq
\mu + \epsilon \Big] \leq d_{\TV}\big(T_{0}\big) +
\exp \Big \{ -\frac{\gamma_{0}}{2 \big(1 - \gamma_{0} \big)} \cdot
\epsilon^2 \big[N - T_0 \big] \Big \},
\end{align}where we have let $T_{0}$ denote the burn-in time. Note that a matching
bound holds for the lower tail.
For our experiments, we computed the tightest version of the
bound~\eqref{eq:uniform-hoeffding-burn-in}, optimizing $T_0$ in the
range $\big[0,~10^{5}\big]$ for each value of the deviation
$\epsilon$. Even given this generosity toward the uniform bound, the
function-specific bound still outperforms it substantially, as
Figure~\ref{fig:mixture-conc-bds} shows.
\begin{figure}[h!]
\centering
\widgraph{0.5\linewidth}{figs/SchizoMixture-ExactMatch-Tails-N=1e6}
\caption{Comparison of the (log) tail probability bounds provided by
the uniform Hoeffding bound due to~\cite{Leo04Hoeffding} with one
version of our function-specific Hoeffding bound
(Theorem~\ref{thm:hoeffding-eps2}). Plots are based on $N = 10^{6}$
iterations, and choosing the optimal burn-in for the uniform bound
and a fixed burn-in of $409 \geq T_{f} \big( 10^{-6} \big)$
iterations for the function-specific bound. The function-specific
bound improves over the uniform bound by orders of
magnitude.\label{fig:mixture-conc-bds}}
\end{figure}
For the function-specific
bound, we used the function-specific oracle bound
(Lemma~\ref{lem:mix-gap-all-oracle}) to bound
$T_{f}\big(\frac{\epsilon}{2}\big)$; this nearly coincides with
the true value when $\epsilon \approx 0.01$ but deviates slightly for
larger values of $\epsilon$.
\section{Proofs of main results}
\label{sec:proofs}
This section is devoted to the proofs of the main results of this
paper.
\subsection{Proof of Theorem~\ref{thm:hoeffding-eps2}}
\label{subsec:proofs-conc-master}
We begin with the proof of the master Hoeffding bound from
Theorem~\ref{thm:hoeffding-eps2}. At the heart of the proof is the
following bound on the moment-generating function (MGF) for the sum of
an appropriately thinned subsequence of the function values
$\{f(X_n)\}_{n=1}^\infty$. In particular, let us introduce the
shorthand notation $\Xtil_{m,t} \mydefn X_{(m - 1) \Tf(\epsilon/2) +
t}$ and $N_0 \mydefn N / T_f(\frac{\epsilon}{2})$. With this
notation, we have the following auxiliary result:
\begin{lem}[Master MGF bound]
For any scalars $\newalpha \in \real$, $\epsilon \in (0,1)$, and
integer $t \in \big [0, \Tf(\frac{\epsilon}{2}) \big)$, we have
\begin{align}
\E \left [ \exp \Big( \newalpha \sum_{m = 1}^{N_0} f \big(\Xtil_{m,t}
\big) \Big) \right] & \leq \exp \left \{\big[
\frac{1}{2}\newalpha\epsilon + \newalpha \mu + \frac{1}2
\newalpha^2 \big] \cdot N_0 \right \}.
\end{align}
\label{lem:mgf-master}
\end{lem}
\noindent See Section~\ref{SecProofLemMGFMaster} for the proof of this
claim.
Recalling the definition of $\Xtil_{m,t}$, we have
\begin{align*}
\E \left[ e^{\alpha\sum_{n = 1}^N f(X_n) } \right] & = \E \left[ \exp
\Big \{ \alpha \sum_{t = 1}^{\Tf(\epsilon/2)} \sum_{m=1}^{N_0}
f(\Xtil_{m,t}) \Big \} \right] \\
& = \E \left [ \exp \Big \{ \alpha \Tf(\epsilon/2) \big[
\frac{1}{\Tf(\epsilon/2)} \sum_{t = 1}^{\Tf(\epsilon/2)} \sum_{m =
1}^{N_0} f(\Xtil_{m,t}) \big] \Big \} \right ] \\
& \leq \frac{1}{\Tf \big( \epsilon/2 \big)} \sum_{t = 0}^{ \Tf \big(\epsilon/2
\big) - 1} \E \left[ \exp \Big \{ \alpha \Tf \big(\epsilon/2\big)
\sum_{m = 1}^{N_0} f \big( \Xtil_{m,t} \big) \Big \} \right],
\end{align*}
where the last inequality follows from Jensen's inequality, as applied
to the exponential function.
Applying Lemma~\ref{lem:mgf-master} with $\newalpha = \alpha\Tf\big(\epsilon/2\big)$, we conclude
\begin{align*}
\E \big[e^{\alpha\sum_{n = 1}^{N} f(X_{n}) }\big] \leq \exp \Big \{
\big[\frac{1}{2}\alpha\Tf\big(\frac{\epsilon}{2}\big)\epsilon + \alpha\Tf\big(\frac{\epsilon}{2}\big)\mu + \frac{1}2
\alpha^2 \Tf^2\big(\frac{\epsilon}{2}\big) \big] \cdot N_0 \Big \},
\end{align*}
valid for $\alpha > 0$. By exponentiating and applying Markov's
inequality, it follows that
\begin{align*}
\P \left[ \frac{1}{N}\sum_{n = 1}^{N} f\big(X_n\big) \geq \mu +
\epsilon \right] & \leq e^{-\alpha(\mu + \epsilon)} \E
\big[e^{\alpha\sum_{n = 1}^{N} f(X_{n}) }\big] \\
& \leq \exp \left \{\frac{1}{2} \cdot \Big[ -\alpha \:
\Tf(\frac{\epsilon}{2}) \epsilon + \alpha^2 \:
\Tf^2(\frac{\epsilon}{2}) \Big] \: N_0(\frac{\epsilon}{2}) \right
\}. \nonumber
\end{align*}
The proof of Theorem~\ref{thm:hoeffding-eps2} follows by taking
$\alpha = \frac{\epsilon}{2\Tf\left(\frac{\epsilon}{2}\right)}$ since
\begin{align*}
\P \left[ \frac{1}{N}\sum_{n = 1}^{N} f\big(X_n\big) \geq \mu +
\epsilon \right] & \leq \exp \left \{ \frac{1}{2} \cdot \big[ -\frac{\epsilon^{2}}{2}
+ \frac{\epsilon^2}{4} \big] \cdot N_0 \right \} \\
& \leq \exp \left \{ -\frac{\epsilon^2 N_0}{8} \right \} \\
& = \exp \left \{ - \frac{\epsilon^2 N}{8 \Tf
\big(\frac{\epsilon}{2}\big)} \right \}.
\end{align*}
\subsubsection{Proof of Lemma~\ref{lem:mgf-master}}
\label{SecProofLemMGFMaster}
For the purposes of the proof, fix $t$ and let $W_{m} =
\tilde{X}_{m,t}$. For convenience, also define a dummy constant random
variable $W_{0} \mydefn 0$. Now, by assumption, we have
\begin{align*}
\left|\E\left[f\left(W_{1}\right)\right] - \mu\right| \leq
\frac{\epsilon}{2} ~~ \text{and} ~~ \left|\E\left[f\left(W_{m +
1}\right)~|~W_{m}\right] - \mu\right| \leq \frac{\epsilon}{2}.
\end{align*}
We therefore have the bound
\begin{align}
\label{eq:master-MGF-martingale}
\E \left[e^{\alpha\sum_{m} f\left(W_{m}\right)}\right] & \leq \E
\left[\prod_{m = 1}^{N_0} e^{\alpha\left[f\left(W_{m}\right) -
\E\left[f\left(W_{m}\right)~|~W_{m - 1}\right]\right]}\right]
\cdot e^{\alpha\mu N_{0} + \frac{\alpha\epsilon N_{0}}{2}}.
\end{align}
But now observe that the random variables $\Delta_{m} =
f\left(W_{m}\right) - \E\left[f\left(W_{m}\right)~|~W_{m - 1}\right]$
are deterministically bounded in $[-1,~1]$ and zero mean conditional
on $W_{m - 1}$. Moreover, by the Markovian property, this implies that
the same is true conditional on \mbox{$W_{< m} \mydefn W_{0:\left(m -
1\right)}$.} It follows by standard MGF bounds that
\begin{align*}
\E\left[e^{\alpha\Delta_{m}}~|~W_{< m}\right] \leq
e^{\frac{\alpha^{2}}{2}} .
\end{align*}
Combining this bound with inequality~\eqref{eq:master-MGF-martingale},
we conclude that
\begin{align*}
\E\left[e^{\alpha\sum_{m} f\left(W_{m}\right)}\right] & \leq
e^{\frac{\alpha^{2}}{2} \cdot N_{0}}\cdot e^{\alpha\mu N_{0} +
\frac{\alpha\epsilon N_{0}}{2}},
\end{align*}
as claimed.
\subsection{Proofs of
Corollaries~\ref{cor:hoeffding-derived-eps2}
and~\ref{cor:hoeffding-derived-Jf}}
\label{subsec:proofs-conc-derived}
In this section, we prove the derived Hoeffing bounds stated in
Corollaries~\ref{cor:hoeffding-derived-eps2}
and~\ref{cor:hoeffding-derived-Jf}.
\subsubsection{Proof of Corollary~\ref{cor:hoeffding-derived-eps2}}
\label{SecProofThmHoeffdingDerivedEps2}
The proof is a direct application of
Theorem~\ref{thm:hoeffding-eps2}. Indeed, it suffices to note that if
$\epsilon \leq \frac{2\lambda_{f}}{\sqrt{\pimin}}$, then
\begin{align*}
\Tf \big(\frac{\epsilon}{2}\big) \leq \frac{\log \big(
\frac{2}{\epsilon\sqrt{\pimin}}\big)}{\log \big(
\frac{1}{\lambda_f} \big)} = \frac{ \log \big( \frac{2}{\epsilon}
\big) + \frac{1}2 \log \big(
\frac{1}{\pimin} \big)}{\log \big(\frac{1}{\lambda_f}\big)},
\end{align*}
which yields the first bound. Turning to the second bound, note that
if $\epsilon > \frac{2\lambda_{f}}{\sqrt{\pimin}}$, then
equation~\eqref{eq:mix-gap} implies that
$\Tf\big(\frac{\epsilon}{2}\big) = 1$, which establishes the
claim.
\subsubsection{Proof of Corollary~\ref{cor:hoeffding-derived-Jf}}
The proof involves combining Theorem~\ref{thm:hoeffding-eps2} with
Lemma~\ref{lem:mix-gap-all}, using the setting \mbox{$\epsilon =
2\left(\Delta + \Delta_{J}\right)$.} We begin by combining the
bounds $\lambda_{J} \leq 1$, $d_{\TV}\left(\pi_{0},~\pi_{n}\right)
\leq 1$, $d_{f}\left(\pi_{0},~\pi_{n}\right) \leq 1$, and
$\E_{\pi}\left[f^{2}\right] \leq 1$ with the claim of
Lemma~\ref{lem:mix-gap-all} so as to find that
\begin{align*}
d_{f}\left(\pi_{0},~\pi_{n}\right) & \leq \Delta_{J}^{\ast} +
\frac{\lambda_{-J}^{n}}{\sqrt{\pimin}} \; \leq \; \Delta_{J} +
\frac{\lambda_{-J}^{n}}{\sqrt{\pimin}} .
\end{align*}
It follows that
\begin{align*}
\Tf (\frac{\epsilon}{2}) & = \Tf\left(\Delta_{J} + \Delta\right) \leq
\frac{\log \left(\frac{1}{\Delta}\right) + \frac{1}{2} \log
\left(\frac{1}{\pimin}\right)}{\log(\frac{1}{\lambda_{-J}})} \qquad
\mbox{whenever $\Delta \leq \frac{\lambda_{-J}}{\sqrt{\pimin}}$.}
\end{align*}
Plugging into Theorem~\ref{thm:hoeffding-eps2} now yields the first
part of the bound. On the other hand, if $\Delta >
\frac{\lambda_{-J}}{\sqrt{\pimin}}$, then Lemma~\ref{lem:mix-gap-all}
implies that $\Tf\left(\Delta_{J} + \Delta\right) = 1$, which proves
the bound in the second case.
\subsection{Proof of Proposition~\ref{prop:lower-bound}}
\label{subsec:proofs-lower}
In order to prove the lower bound in
Proposition~\ref{prop:lower-bound}, we first require an auxiliary
lemma:
\begin{lem}
\label{lem:lower-bound-helper}
Fix a function $\delta \colon \left(0,~1\right) \rightarrow
\left(0,~1\right)$ with $\delta\left(\epsilon\right) > \epsilon$. For
every constant $c_{0} \geq 1$, there exists a Markov chain $P_{c_0}$
and a function $f_{\epsilon}$ on it such that $\mu = \frac{1}{2}$,
yet, for $N = c_{0}\Tf\left(\delta\left(\epsilon\right)\right)$, and
starting the chain from stationarity,
\begin{align*}
\P_{\pi}\left(\left|\frac{1}{N}\sum_{n = 1}^{N}
f_{\epsilon}\left(X_{n}\right) - \frac{1}{2}\right| \geq
\epsilon\right) \geq \frac{1}{3} .
\end{align*}
\end{lem}
Using this lemma, let us now prove Proposition~\ref{prop:lower-bound}.
Suppose that we make the choices
\begin{align*}
c_{0} & \mydefn \left\lceil\frac{\log
7}{c_{1}\epsilon^{2}}\right\rceil \geq 1, \quad \mbox{and} \quad
N_{c_{1},\epsilon} \mydefn c_{0}\Tf\left(\epsilon\right),
\end{align*}
in Lemma~\ref{lem:lower-bound-helper}. Letting $P_{c_0}$ be the
corresponding Markov chain and $f_{\epsilon}$ the function guaranteed
by the lemma, we then have
\begin{align*}
\P_{\pi}\left(\left|\frac{1}{N_{c_1,\epsilon}}\sum_{n =
1}^{N_{c_1,~\epsilon}} f_{\epsilon}\left(X_{n}\right) -
\frac{1}{2}\right| \geq \epsilon \right) & \geq \frac{1}{3} \; > \;
\frac{2}{7} \; \geq \; 2 \cdot \exp\left(-\frac{c_{1}
N_{c_1,\epsilon}\epsilon^{2}}{\Tf\left(\delta(\epsilon)\right)}
\right).
\end{align*}
\paragraph{Proof of Lemma~\ref{lem:lower-bound-helper}:}
It only remains to prove Lemma~\ref{lem:lower-bound-helper}, which we
do by constructing pathological function on a chain graph, and letting
our Markov chain be the lazy random walk on this graph. For the
proof, fix $\epsilon > 0$, let $\delta = \delta\left(\epsilon\right)$
and let $\Tf = \Tf\left(\delta\right)$. Now choose an integer $d > 0$
such that $d > 2c_{0}$ and let the state space be $\Omega$ be the line
graph with $2d$ elements with the standard lazy random walk defining
$P$. We then set
\begin{align*}
f\left(i\right) & = \begin{cases}
\frac{1}{2} - \delta & ~ 1 \leq i \leq d, \\
\frac{1}{2} + \delta &~d + 1 \leq i \leq 2d .
\end{cases}
\end{align*}
It is then clear that $\Tf = 1$.
Define the bad event
\begin{align*}
\mathcal{E} = \left\lbrace X_{1} \in \left[0,~\frac{d}{2}\right] \cup
\left[\frac{3d}{2},~2d\right]\right\rbrace .
\end{align*}
When this occurs, we have
\begin{align*}
\left|\frac{1}{N'}\sum_{n = 1}^{N'} f\left(X_{n}\right) -
\frac{1}{2}\right| \geq \delta > \epsilon \qquad \mbox{with
probability one},
\end{align*}
for all $N' < \frac{d}{2}$. Since $N = c_{0} < \frac{d}{2}$, we can
set $N' = N$.
On the other hand, under $\pi$, the probability of $\mathcal{E}$ is
$\geq \frac{1}{3}$. (It is actually about $\frac{1}{2}$, but we want
to ignore edge cases.) The claim follows immediately.
\subsection{Proofs of confidence interval results}
\label{subsec:proofs-confidence}
Here we provide the proof of the confidence interval corresponding to
our bound (Theorem~\ref{thm:confidence-adaptive-hoeffding}). Proofs of
the claims~\eqref{eq:confidence-unif-hoeffding}
and~\eqref{eq:confidence-clt-be} can be found in
Appendix~\ref{app:proof-confidence}.
As discussed in Section~\ref{subsec:confidence}, we actually prove a
somewhat stronger form of
Theorem~\ref{thm:confidence-adaptive-hoeffding}, in order to guarantee
that the confidence interval can be straightforwardly built using an
upper bound $\tilde{T}_{f}$ on the $f$-mixing time rather than the
true value. Setting $\tilde{T}_{f} = \Tf$ recovers the original
theorem.
Specifically, suppose $\tilde{T}_{f} \colon \N \rightarrow \R_{+}$ is
an upper bound on $\Tf$ and note that the corresponding tail bound
becomes $e^{-\tilde{r}_{N}(\epsilon)/8}$, where
\begin{align*}
\tilde{r}_{N}(\epsilon) = \epsilon^2
\left[\frac{N}{\tilde{T}_{f}\big(\frac{\epsilon}{2}\big)} - 1 \right]
.
\end{align*}
This means that, just as before we wanted to make the rate $r_{N}$ in
equation~\eqref{eq:rN-def} at least as large as $8 \log\frac2
{\alpha}$, we now wish to do the same with $\tilde{r}_{N}$, which
means choosing $\epsilon_{N}$ with
$\tilde{r}_{N}\big(\epsilon_{N}\big) \geq 8 \log \frac2 {\alpha}$. We
therefore have the following result.
\begin{prop}
\label{prop:confidence-adaptive-hoeffding-bdd}
For any width $\epsilon_{N} \in \tilde{r}_{N}^{-1}\big(\big[8
\log\big(2/\alpha\big),~\infty\big)\big)$, the set
\begin{align*}
\FUNCINT = \left[ \frac{1}{N - T_{f}\big(\frac{\epsilon}{2} \big)}
\sum_{n = \tilde{T}_{f} \big( \frac{\epsilon}{2} \big)}^{N}
f\big(X_{n}\big) \pm \epsilon_{N} \right]
\end{align*}
is a $1 - \alpha$ confidence interval for $\mu = \E_{\pi}\big[f\big]$.
\end{prop}
\begin{proof}
For notational economy, let us introduce the shorthands
$\tau_{f}(\epsilon) = \Tf\big(\frac{\epsilon}{2}\big)$ and
$\tilde{\tau}_{f}\big(\epsilon\big) =
\tilde{T}_{f}\big(\frac{\epsilon}{2}\big)$. Theorem~\ref{thm:hoeffding-eps2}
then implies
\begin{align*}
\P \left[ \frac{1}{N - \tilde{\tau}_{f}}\sum_{n =
\tilde{\tau}_{f}}^{N} f\big(X_n\big) \geq \mu + \epsilon \right] &
\leq \exp\big(-\frac{N - \tau_{f}}{4\tau_{f}} \cdot \epsilon^2 \big)
\\
& \leq \exp\big(-\frac{N - \tilde{\tau}_{f}}{4\tilde{\tau}_{f}}
\cdot \epsilon^2 \big) \\ & =
\exp\big(-\frac{\tilde{r}_{N}\big(\epsilon\big)}{4}\big) .
\end{align*}
Setting $\epsilon = \epsilon_{N}$ yields
\begin{align*}
\P \left[ \frac{1}{N - \tilde{\tau}_{f}}\sum_{n =
\tilde{\tau}_{f}}^{N} f\big(X_n\big) \geq \mu + \epsilon_{N}
\right] & \leq \frac{\alpha}{2}.
\end{align*}
The corresponding lower bound leads to an analogous bound on the lower
tail.
\end{proof}
As we did with Corollary~\ref{cor:confidence-adaptive-hoeffding-concrete}, we can
derive a more concrete, though slightly weaker, form of this result
that is more amenable to interpretation. We derive the corollary from the
specialized bound by setting $\tilde{T}_{f} = T_{f}$.
To obtain this bound, define the
following lower bound, in parallel with
equation~\eqref{eq:rN-eta-def}:
\begin{align*}
\tilde{r}_{N}\big(\epsilon\big) \geq
\tilde{r}_{N,\eta}\big(\epsilon\big) \mydefn \epsilon^2
\big[\frac{N}{\tilde{T}_{f}\big(\frac{\eta}{2}\big)} - 1\big],~
\epsilon \geq \eta .
\end{align*}
Since this is a lower bound, we see that whenever $\epsilon_{N} \geq
\eta$ and $\tilde{r}_{N,\eta}\big(\epsilon_{N}\big) \geq 8 \log\frac2
{\alpha}$, $\epsilon_{N}$ is a valid half-width for a $\big(1 -
\alpha\big)$-confidence interval for the stationary mean centered
at the empirical mean. More formally, we have the following:
\begin{prop}
\label{prop:confidence-adaptive-hoeffding-concrete-bdd}
Fix $\eta > 0$ and let
\begin{align*}
\epsilon_{N} = \tilde{r}_{N,\eta}^{-1}\big(8 \log \frac2 {\alpha}\big)
& = 2\sqrt{2}\sqrt{\frac{\tilde{T}_{f}\big(\frac{\eta}{2}\big) \cdot
\log\big (2/ \alpha \big)}{N - \tilde{T}_{f} \big( \frac{
\eta}{2} \big)}} .
\end{align*}
If $N \geq \tilde{T}_{f}\big(\frac{\eta}{2}\big)$, then $\FUNCINT$
is a $1 - \alpha$ confidence interval for $\mu = \E_{\pi}\big[f\big]$.
\end{prop}
\begin{proof}
By assumption, we have
\begin{align*}
\eta \leq \epsilon_{N}\big(\eta\big) = 2
\sqrt{\frac{\tilde{T}_{f}\big(\frac{\eta}{2}\big) \cdot
\log\big(2/\alpha\big)}{N - \tilde{T}_{f}\big(\frac{\eta
}{2}\big)}}.
\end{align*}
This implies $\tilde{T}_{f}\big(\frac{\epsilon_{N}}{2}\big) \geq
\tilde{T}_{f}\big(\frac{\eta}{2}\big)$, which yields
\begin{align*}
\tilde{r}_{N} \big(\epsilon_{N}\big) = \epsilon_{N}^2
\big[\frac{N}{\tilde{T}_{f}\big(\frac{\epsilon_{N}}{2}\big)} -
1\big] \geq \epsilon_{N}^2
\big[\frac{N}{\tilde{T}_{f}\big(\frac{\eta}{2}\big)} - 1\big] =
8 \log\big(2/\alpha\big).
\end{align*}
But now Proposition~\ref{prop:confidence-adaptive-hoeffding-bdd} applies,
so that we are done.
\end{proof}
\subsection{Proofs of sequential testing results}
\label{subsec:proofs-testing}
In this section, we collect various proofs associated with our
analysis of the sequential testing problem.
\subsubsection{Proof of Theorem~\ref{thm:seq-err-all-adapt} for $\algfix$}
We provide a detailed proof when $H_1$ is true, in which case we have
$\mu \leq r - \delta$; the proof for the other case is analogous.
When $H_1$ is true, we need to control the probability
$\P\big(\algfix\big(X_{1:N}\big) = H_{0}\big)$. In order to do so,
note that Theorem~\ref{thm:hoeffding-eps2} implies that
\begin{align*}
\P\big(\algfix\big(X_{1:N}\big) = H_{0}\big) & =
\P\big(\frac{1}{N}\sum_{n = 1}^{N} f\big(X_{n}\big) \geq r +
\delta\big) \\ & \leq \P\big(\frac{1}{N}\sum_{n = 1}^{N}
f\big(X_{n}\big) \geq \mu + 2\delta\big) \\ & \leq
\exp\big(-\frac{\delta^2 N}{2\Tf\big(\delta\big)}\big).
\end{align*}
Setting $N = \frac{2T_{f}\left(\delta\right)
\log\big(\frac{1}{\alpha}\big)}{\delta^2 }$ yields the bound $\P
\big(\algfix\big(X_{1:N}\big) = H_{0}\big) \leq \alpha$, as claimed.
\subsubsection{Proof of Theorem \ref{thm:seq-err-all-adapt} for $\algseq$}
The proof is nearly identical to that given by~\cite{Gyo15Test}, with
$\Tf\big(\delta/2\big)$ replacing $\frac{1}{\gamma_{0}}$. We
again assume that $H_{1}$ holds, so $\mu \leq r - \delta$. In this
case, it is certainly true that
\begin{align*}
\err\big(\algseq,~f\big) & = \P\big(\exists k \colon
\algseq\big(X_{1:N_{k}}\big) = H_{0}\big) \\
& = \P\big(\exists k \colon \frac{1}{N_{k}}\sum_{n = 1}^{N_{k}}
f\big(X_{n}\big) \geq r + \frac{M}{N_{k}}\big) \\
& \leq \sum_{k = 1}^{\infty} \P\big(\frac{1}{N_{k}}\sum_{n = 1}^{N_k}
f\big(X_{n}\big) \geq r + \frac{M}{N_{k}}\big) .
\end{align*}
It follows by Theorem \ref{thm:hoeffding-eps2}, with $\epsilon_{k} =
\delta + \frac{M}{N_{k}}$, that
\begin{align*}
\P\big(\frac{1}{N_{k}}\sum_{n = 1}^{N_k} f\big(X_{n}\big) \geq r +
\frac{M}{N_{k}}\big) & \leq \P\big(\frac{1}{N_{k}}\sum_{n = 1}^{N_k}
f\big(X_{n}\big) \geq \mu + \delta + \frac{M}{N_{k}}\big) \\ & \leq
\exp\big(-\frac{\epsilon_{k}^2 N_{k}}{8\Tf\big(\frac{\epsilon_{k}
}{2}\big)}\big) \\ & \leq \exp\big(-\frac{\epsilon_{k}^2
N_{k}}{8\Tf\big(\frac{\delta}{2}\big)}\big).
\end{align*}
In order to simplify notation, for the remainder of the proof, we
define \mbox{$\tau \mydefn 4\Tf(\delta/2)$,} \mbox{$\beta \mydefn
\frac{\sqrt{\alpha\xi}}{2}$,} and \mbox{$\zeta_{k} \mydefn
\frac{\delta^2 N_{k}}{2 \tau \log(1/\beta)}$.} In terms of this
notation, we have $M = \frac{2\tau \log(1/\beta)}{\delta}$, and hence
that
\begin{align*}
\exp\big(-\frac{\epsilon_{k}^2 N_{k}}{2\tau}\big) & =
\exp\big(-\frac{1}{2\tau} \cdot \big(\delta^2 N_{k} + 2\delta M +
\frac{M^2 }{N_{k}}\big)\big) \\ & = \exp\big(-\big[\frac{\delta^2
N_{k}}{2\tau} + \log\big(1/\beta\big) + \frac{2\tau \log^2
\big(1/\beta\big)}{\delta^2 N_{k}}\big]\big) \\ & =
\exp\big(-\log\big(1/\beta\big)\big[1 + \zeta_{k} +
\zeta_{k}^{-1}\big]\big) \\ & = \beta \cdot
\exp\big(-\log\big(1/\beta\big)\big[\zeta_{k} +
\zeta_{k}^{-1}\big]\big) .
\end{align*}
It follows that the error probability is at most
\begin{align*}
\beta\sum_{k = 1}^{\infty}
\exp\big(-\log\big(1/\beta\big)\big[\zeta_{k} +
\zeta_{k}^{-1}\big]\big).
\end{align*}
We now finish the proof using two small technical lemmas, whose proofs
we defer to Appendix~\ref{app:proofs-proofs-testing}.
\begin{lem}
\label{lem:seq-seq-indiff-adapt-sum} In the above notation, we have
\begin{align*}
\sum_{k = 1}^{\infty} \exp \Big \{ -\log (1/\beta) \big[\zeta_{k} +
\zeta_{k}^{-1}\big] \Big \} & \leq 4\sum_{\ell = 0}^{\infty} \exp
\Big \{ -\log(1/\beta) \Big[\big(1 + \xi\big)^{\ell} + \big(1 +
\xi\big)^{-\ell} \Big] \Big \}.
\end{align*}
\end{lem}
\begin{lem}
\label{lem:seq-seq-indiff-adapt-group}
For any integer $c \geq 0$, we have
\begin{align*}
(1 + \xi)^{\ell} + (1 + \xi)^{-\ell} & \geq 2 ( c + 1 ) \quad
\mbox{for all $\ell \in \Big[\frac{9c}{5\xi}, \frac{9 (c + 1)}{5\xi}
\Big)$.}
\end{align*}
\end{lem}
Using this bound, and grouping together terms in blocks of size
$\frac{9}{5\xi}$, we find that the error is at most
\begin{align*}
4 \sum_{\ell = 0}^{\infty} \exp\big(-\log\big(1/\beta\big)\big[\big(1
+ \xi\big)^{\ell} + \big(1 + \xi\big)^{-\ell}\big]\big) & \leq
\frac{36}{5\xi} \cdot \sum_{c = 0}^{\infty} \beta^{2\big(c + 1\big)}.
\end{align*}
Since both $\alpha$ and $\xi$ are at most $\frac{2}{5}$, we have
$\beta = \frac{\sqrt{\alpha \xi}}{2} \leq \frac{1}{5}$, and hence the
error probability is bounded as
\begin{align*}
\frac{36\beta}{5\xi}\sum_{c = 0}^{\infty} \beta^{2\big(c + 1\big)}
\leq \frac{36\beta^{3}}{5\xi\big(1 - \beta^2 \big)} \leq
\frac{36\beta^2 }{25\xi\big(1 - \beta^2 \big)}\leq \frac{3\beta^2
}{2\xi} = \frac{3\alpha}{4} < \alpha .
\end{align*}
\subsubsection{Proof of Theorem~\ref{thm:seq-err-all-adapt} for $\algdiff$}
We may assume that $H_{1}$ holds, as the other case is
analogous. Under $H_{1}$, letting $k_{0}$ be the smallest $k$ such
that $\epsilon_{k} < \infty$, we have
\begin{align*}
\err (\algdiff,~f) & \leq \sum_{k = k_{0}}^{\infty}
\P\left(\hat{\mu}_{N_k} \geq r + \epsilon_{k}\right) \; \leq \;
\sum_{k = k_{0}}^{\infty} \P(\hat{\mu}_{N_k} \geq \mu +
2\epsilon_{k}).
\end{align*}
By Theorem~\ref{thm:hoeffding-eps2}, and the definition of
$\epsilon_{k}$, we thus have
\begin{align*}
\err \left(\algdiff,~f\right) \leq \sum_{k = k_0}^{\infty}
\exp\left(-\frac{N_{k}\epsilon_{k}^{2}}{8T_{f}\left(\frac{\epsilon_{k}}{2}\right)}\right)
& \leq \frac{\alpha}{2}\sum_{k = k_0}^{\infty} \frac{1}{k^2} \\ & =
\frac{\pi^{2}}{12} \, \alpha \; < \; \alpha,
\end{align*}
as claimed.
\subsubsection{Proof of Theorem~\ref{thm:seq-stopping-time-all} for $\algseq$}
We may assume $H_{1}$ holds; the other case is analogous. Note that
\begin{align*}
\E[N] & \leq N_{1} + \sum_{k = 1}^{\infty} \big(N_{k + 1} -
N_{k}\big)\P\big(N > N_{k}\big) \\
& \leq N_{1} + \sum_{k = 1}^{\infty} \big(N_{k + 1} -
N_{k}\big)\P\big(\frac{1}{N_{k}}\sum_{n = 1}^{N_{k}} f\big(X_{n}\big)
\in \big(r - \frac{M}{N_{k}},~r + \frac{M}{N_{k}}\big)\big) \\
& \leq N_{1} + \sum_{k = 1}^{\infty} \big(N_{k + 1} -
N_{k}\big)\P\big(\frac{1}{N_{k}}\sum_{n = 1}^{N_{k}} f\big(X_{n}\big)
> r - \frac{M}{N_{k}}\big) \\
& = N_{1} + \sum_{k = 1}^{\infty} \big(N_{k + 1} -
N_{k}\big)\P\big(\frac{1}{N_{k}}\sum_{n = 1}^{N_{k}} f\big(X_{n}\big)
> \mu + \Delta - \frac{M}{N_{k}}\big) \\
& \leq N_{1} + \sum_{k = 1}^{\infty} \big( N_{k + 1} - N_{k} \big)
\exp \left \{ -\frac{\big(\Delta N_{k} - M\big)_{+}^2
}{8\Tf\big(\delta/2\big)N_{k}} \right \} .
\end{align*}
Our proof depends on the following simple technical lemma, whose proof
we defer to Appendix~\ref{subapp:proof-seq-seq-stopping-time-sum}.
\begin{lem}
\label{lem:seq-seq-stopping-time-sum}
Under the conditions of Theorem~\ref{thm:seq-stopping-time-all}, we
have
\begin{align}
\sum_{k = 1}^{\infty} \big(N_{k + 1} - N_{k}\big) \exp \left \{
-\frac{\big(\Delta N_{k} - M\big)_{+}^2
}{8\Tf\big(\delta/2\big)N_{k}} \right \} & \leq \big ( 1 + \xi
\big) \big [ 1 + \int_{N_{1}}^{\infty} h(s) \der s\big],
\end{align}
where $h(s) \mydefn \exp \Big \{ -\frac {(\Delta s -
M)_{+}^2}{8\Tf(\delta/2) s} \Big \}$.
\end{lem}
Given this lemma, we then follow the argument of~\citet{Gyo15Test} in
order to bound the integral. We have
\begin{align*}
\int_{N_{1}}^{\infty} h\big(s\big) \der s \leq \frac4
{\Delta}\sqrt{\frac{2\Tf\big(\delta/2\big)M}{\Delta} +
8\Tf\big(\delta/2\big)}.
\end{align*}
To conclude, note that either $r \geq \Delta$ or $1 - r \geq \Delta$,
since $0 < \mu < 1$, so that $\min\big(\frac{1}{r},~\frac{1}{1 -
r}\big) \leq \frac{1}{\Delta}$. It follows that
\begin{align*}
N_{1} \leq \big(1 + \xi\big)N_{0} \leq \frac{(1 + \xi) M}{\Delta}.
\end{align*}
Combining the bounds yields the desired result.
\subsubsection{Proof of Theorem~\ref{thm:seq-stopping-time-all} for $\algdiff$}
For concreteness, we may assume $H_{1}$ holds, as the $H_{0}$ case is
symmetric. We now have that
\begin{align*}
\P \left[ N \geq N_{k} \right] & \leq \P \left[
\big|\frac{1}{N_{k}}\sum_{n = 1}^{N_k} f\big(X_{n}\big) - r\big|
\leq \epsilon_{k} \right] \; \leq \; \P \left[ \frac{1}{N_{k}}
\sum_{n = 1}^{N_k} f\big(X_{n}\big) \geq \mu + \Delta - \epsilon_{k}
\right].
\end{align*}
For convenience, let us introduce the shorthand
\begin{align*}
T_{f,k}^{+} & \mydefn \begin{cases} T_{f}\big(\frac{\Delta - \epsilon_{k}
}{2}\big) &~ \text{if}~ \epsilon_{k} \leq \Delta, \\ 1 &~
\text{otherwise} .
\end{cases}
\end{align*}
Applying the Hoeffding bound from Theorem~\ref{thm:hoeffding-eps2}, we
then find that
\begin{align*}
\P \left[ N \geq N_{k} \right] & \leq \exp \left \{
-\frac{N_{k}}{8T_{f,k}^{+}} \cdot \big(\Delta -
\epsilon_{k}\big)_{+}^2 \right \}.
\end{align*}
Observe further that
\begin{align*}
\E [N] & = N_{1} + \sum_{k = 1}^{\infty} \big(N_{k + 1} -
N_{k}\big)\P\big(N > N_{k}\big) \\
& \leq N_{k_{0}^{\ast} + 1} + \sum_{k = k_{0}^{\ast} + 1}^{\infty}
\big(N_{k + 1} - N_{k}\big)\P\big(N > N_{k}\big) \\
& \leq \big(1 + \xi \big) \big(N_{0}^{\ast} + 1 \big ) + \sum_{k =
k_{0}^{\ast} + 1}^{\infty} \big(N_{k + 1} - N_{k} \big) \P\big(N >
N_{k} \big).
\end{align*}
Combining the pieces yields
\begin{align}
\label{eq:ENstar-bd-1}
\E[N] & \leq \big(1 + \xi\big)\big(N_{0}^{\ast} + 1\big) + \sum_{k =
k_{0}^{\ast} + 1}^{\infty} \big(N_{k + 1} -
N_{k}\big)\exp\big(-\frac{N_{k}}{8T_{f,k}^{+}} \cdot \big(\Delta -
\epsilon_{k}\big)_{+}^2 \big).
\end{align}
The crux of the proof is a bound on the infinite sum, which we pull
out as a lemma for clarity.
\begin{lem}
\label{lem:seq-diff-stopping-time-sum}
The infinite sum~\eqref{eq:ENstar-bd-1} is upper bounded by
\begin{align*}
\sum_{k = k_{0}^{\ast} + 1}^{\infty} \big(N_{k + 1} -
N_{k}\big)\exp\big(-\frac{N_{k}}{8T_{f,k}^{+}} \cdot \big(\Delta -
\epsilon_{k}\big)_{+}^2 \big) \leq \alpha\cdot \sum_{m =
1}^{\infty}\exp\big(-m \cdot \frac{\Delta^2
}{32T_{f}\big(\frac{\Delta}{4}\big)}\big).
\end{align*}
\end{lem}
\noindent See Appendix~\ref{subapp:proof-seq-diff-stopping-time-sum}
for the proof of this claim.
Lemma~\ref{lem:seq-diff-stopping-time-sum} then implies that
\begin{align*}
\sum_{k = k_{0}^{\ast} + 1}^{\infty} \big(N_{k + 1} - N_{k}\big)
\exp\big(-\frac{N_{k}}{T_{f}\big(\frac{\Delta}{4}\big)} \cdot
\frac{\Delta^2}{32}\big) & \leq \alpha\cdot \sum_{m =
1}^{\infty}\exp\big(-m \cdot \frac{\Delta^2
}{32T_{f}\big(\frac{\Delta}{4}\big)}\big) \\
& = \frac{\alpha \exp\big(-\frac{\Delta^2 }{32T_{f}
\big(\frac{\Delta}{4}\big)}\big)}{1 -
\exp\big(\frac{\Delta^2}{32T_{f}\big(\frac{\Delta
}{4}\big)}\big)} \\ & \leq \frac{32\alpha
T_{f}\big(\frac{\Delta}{4}\big)}{\Delta^2} .
\end{align*}
The claim now follows from equation~\eqref{eq:ENstar-bd-1}.
\section{Discussion}
\label{sec:discuss}
A significant obstacle to successful application of statistical
procedures based on Markov chains---especially MCMC---is the
possibility of slow mixing. The usual formulation of mixing is in
terms of convergence in a distribution-level metric, such as the total
variation or Wasserstein distance. On the other hand, algorithms like
MCMC are often used to estimate equilibrium expectations over a
limited class of functions. For such uses, it is desirable to build a
theory of mixing times with respect to these limited classes of
functions and to provide convergence and concentration guarantees
analogous to those available in the classical setting, and our paper
has made some steps in this direction.
In particular, we introduced the $f$-mixing time of a function, and
showed that it can be characterized by the interaction between the
function and the eigenspaces of the transition operator. Using these
tools, we proved that the empirical averages of a function $f$
concentrate around their equilibrium values at a rate characterized by
the $f$-mixing time; in so doing, we replaced the worst-case
dependence on the spectral gap of the chain, characteristic of
previous Markov chain concentration bounds, by an adaptive dependence
on the properties of the actual target function. Our methodology
yields sharper confidence intervals, as well as better rates for
sequential hypothesis tests in MCMC, and we have provided evidence
that the predictions made by our theory are accurate in some real
examples of MCMC and thus potentially of significant practical
relevance.
Our investigation also suggests a number of further questions. Two important
ones concern the continuous and non-reversible cases. Both arise frequently in statistical
applications---for example, when sampling continuous parameters or when performing
Gibbs sampling with systematic scan---and are therefore of considerable interest. As uniform Hoeffding bounds
do exist for the continuous case and, more recently, have been established for the non-reversible case, we
believe many of our conclusions should carry over to these settings, although somewhat different methods of
analysis may be required.
Furthermore, in practical applications, it would be desirable to have methods for
estimating or bounding the $f$-mixing time based on samples. It would
also be interesting to study the $f$-mixing times of Markov chains that
arise in probability theory, statistical physics, and applied
statistics itself. While we have shown what can be done with spectral
methods, the classical theory provides a much larger arsenal of
techniques, some of which may generalize to yield sharper $f$-mixing
time bounds. We leave these and other problems to future work.
\subsection*{Acknowledgments}
The authors thank Alan Sly and Roberto Oliveira for helpful discussions about the lower bounds and
the sharp function-specific Hoeffding bounds (respectively).
This work was partially supported by NSF grant CIF-31712-23800,
ONR-MURI grant DOD 002888, and AFOSR grant FA9550-14-1-0016. In
addition, MR is supported by an NSF Graduate Research Fellowship and a
Fannie and John Hertz Foundation Google Fellowship.
\bibliographystyle{plainnat}
|
\section{Introduction}
\label{maru}
In this paper we are interested in some stochastic
properties of the so called Nelson model of an electrically charged spinless quantum particle coupled to a scalar boson field. These properties
will be formulated in terms of central limit theorem-type behaviours of functionals of the particle-field operators. While the
quantum field models discussed here are defined in terms of self-adjoint operators on a joint particle-field space of functions,
for our purposes a Feynman-Kac type approach will be more suitable. Then the related evolution semigroups can be represented in
terms of averages over the paths of suitable random processes, which has been much explored lately.
The Nelson model is defined by a self-adjoint operator of the form
\begin{equation}
\label{nelsonH}
H_{\rm N} = \hp \otimes \one + \one \otimes \hf + \hi,
\end{equation}
on a Hilbert space $\LR\otimes \fff$, where $\fff$ denotes the boson Fock space over $\LR$, and the components describe the Hamilton operators of the free particle,
free field, and particle-field interaction, respectively. The classical and the relativistic Nelson models differ by the choice
of the free particle operator $\hp$, which in the classical model is a Schr\"odinger operator and in the relativistic case a
relativistic Schr\"odinger operator, as given by the expressions \kak{schroedinger} and (\ref{relp}) below. On the functional
integral representation level this difference will appear in the fact that a Schr\"odinger operator generates a diffusion,
while a relativistic Schr\"odinger operator generates a jump process.
A functional CLT for the classical Nelson model has been first established by Betz and Spohn in \cite{BS}. They have shown that under the
Gibbs measure obtained from taking the marginal over the particle-generated component of the path measure in the functional integral
representation of $H_{\rm N}$, the process scaled by Brownian scaling converges in distribution to Brownian motion having
reduced diffusion coefficients. This means that the particle increases its effective mass due to the coupling to the boson field. The main
observation in this paper is that one can associate a martingale with functionals of the process, whose long time behaviour can be
predicted by using the martingale convergence theorem. The result for more general Markov processes is originally due to Kipnis
and Varadhan \cite[Theorem 1.8]{kv86}, and similar problems are studied also in \cite {bha82,CCG,ledoux11}.
Whenever in $\hp$ the external potential $V$ is chosen to be sufficiently regular, the operator semigroup $\{ e^{-t\hp} :
t \geq 0\}$ can be studied by a Feynman-Kac type representation, i.e., there exists a random process $\proo Z$ on a suitable
probability space such that
\eq{FK}
\left(e^{-t\hp }f\right)(x) = \mathbb{E}^x\left[e^{-\int_0^tV(Z_s) ds}f(Z_t) \right]
\en
holds for all Borel measurable $f$ on $\BR$, where the expectation is taken with respect to the path measure of $\proo Z$.
When $\hp $ is a classical Schr\"odinger operator, the process is a $d$-dimensional Brownian motion $\proo B$ and the path
measure is Wiener measure $\W$ on the space of continuous paths $C((0,\infty),\BR)$. When, however, ${H}_{\rm p}$
is a relativistic Schr\"odinger operator, the process changes to a $d$-dimensional L\'evy process with their corresponding
path measures, now on the Skorokhod space $D((0,\infty),\BR)$ of c\`adl\`ag paths (i.e., the paths are discontinuous,
but continuous from the left with right limits). Whenever the coupling between the particle and field is turned on, the boson
field will contribute by an infinite dimensional Ornstein-Uhlenbeck process so that in the path integral representation of the
evolution semigroup $\{ e^{-tH_{\rm N}}: t \geq 0\}$ a two-component random process will appear, as it will be explained below.
Due to the linear coupling between particle and field, one can integrate over the OU-component and the marginal distribution for
the particle will contain beside the given external potential also an effective pair-interaction potential resulting from the
interaction.
Assuming that the external potential is chosen in such a way that the bottom of the spectrum $E_{\rm p} = \inf \sigma(\hp)$
is an isolated eigenvalue, i.e., a ground state $\grp$ ($L^2$-normalized eigenfunction corresponding to the bottom of the spectrum)
of $\hp$ exists, by standard methods it can be shown that it is unique and has a strictly positive version, which we will choose
throughout. Using the ground state, we can define the unitary operator
$$
U: L^2(\BR,\grp^2dx) \to L^2(\BR,dx), \quad f \mapsto \grp f,
$$
and consider the self-adjoint operator $U^{-1}(\hp-E_{\rm p})U$. This operator generates a stationary Markov process, which we
denote by $\proo Y$ and call a {$P(\phi)_1$-process associated with $\hp$}. We then have with suitable test functions $f$ and $g$
the formula
\eq{ppp}
(f\grp, e^{-t(\hp-E_{\rm p}) } g\grp)_{L^2(\BR,dx)}=\int _\BR \Ebb^x[f(Y_0) g(Y_t)] \grp^2(x)dx.
\en
In what follows, we are interested in constructing a $P(\phi)_1$-process associated with the Nelson Hamiltonian. This will be then
obtained by a similar unitary operator, and will give rise to a two-component random process according to the separate contributions
of the particle and the fields operators. Our main aim is then to study a FCLT behaviour of this process in the sense of the
invariance principle due to \cite{Don}.
This paper is organized as follows. In Section 2 we discuss the functional integral representations of the classical and relativistic
Nelson models. In Section 3, we construct $P(\phi)_{1}$-process associated with the two models.
Section 4 is devoted to proving a functional central limit theorem for additive functionals associated with the Nelson models by using
the properties of the $P(\phi)_{1}$-process. We show some functionals of special interest for both cases and determine explicitly the
variance in the related FCLT. Finally, we make some remarks on extensions to related models in Section 5.
\section{Functional integral representations of the Nelson model}
\subsection{Functional integral representation of the free particle Hamiltonians}
We will consider the classical and relativistic Nelson models in parallel. In some aspects of the construction the relevant
property is the Markov property of the underlying processes, thus the expressions will appear similar with the difference that
the appropriate processes are applied, however, in some other aspects the path properties will become crucial and significant
differences appear.
Let $V: \BR \to \RR$ be a Borel-measurable function giving the potential. We denote the multiplication operator defined by $V$
by the same label. The energy of the free particle in the classical model is described by the Schr\"{o}dinger operator acting
on $\LR $ and formally written as
\eq{schroedinger}
\hp =-\half\Delta +V.
\en
The {relativistic} model is described by the relativistic Schr\"{o}dinger operator acting on $\LR $ and formally written as
\begin{equation}
\label{relp}
\tilde{H}_{\rm p} = \sqrt{-\Delta +m^2} - m + V,
\end{equation}
where the square-root operator is defined by Fourier transform in the standard way, and the parameter $m \geq 0$ is the rest mass
of the particle.
These Schr\"odinger operators can be defined in the sense of perturbation theory by choosing suitable conditions on $V$. However,
since we will use methods of functional integration, we are interested to choose $V$ in a way which allows a Feynman-Kac type
representation to hold. The natural choice is Kato-class, in each case given in terms of the related random processes.
In order to describe the classical case, consider the space $\X=C(\RR^+,\BR)$ of $\BR$-valued continuous functions on $\RR^+.$
Let $\proo B$ be $d$-dimensional Brownian motion defined on $(\X,\B(\X))$, where $\B(\X)$ is the $\s$-field generated by the cylinder
sets of $\X$, and denote by $\W^x$ the Wiener measure starting from $x$ at $t=0$. Also, consider $\D=D(\RR^+,\BR)$, the space of c\`adl\`ag
paths (i.e., continuous from right with left limits) with values in $\BR$, and $\B(\D)$ the $\s$-field generated by the cylinder sets of
$\D$. Let $\proo b$ denote a $d$-dimensional rotationally symmetric relativistic Cauchy process generated by (\ref{relp}) when $m > 0$,
and a rotationally symmetric Cauchy process when $m=0$. In each case we denote by $\mathscr P ^x$ the path measure of the process in question
starting from $x$ at time $t=0$. It is well-known that all these processes have the strong Markov property with respect to their natural
filtrations. When we do not need to specify the process, we will use the generic notation $\proo Z$, and it will be understood that
expectations are taken with respect to the own path measure of the process.
\begin{definition}\TTT{Kato-class}
\label{kato}
{\rm
We say that $V = V_+ - V_-$ is a{ Kato-class potential with respect to the random process $\proo Z$} whenever for its positive and
negative parts
$$
V_- \in {\cal K}^Z \quad \mbox{and} \quad V_+ 1_C \in {\cal K}^Z \;\, \mbox{for every compact set $C \subset \RR^d$},
$$
hold, where $f \in {\cal K}^Z$ means that
\begin{align}
\label{eq:Katoclass}
\lim_{t \rightarrow 0} \sup_{x \in \RR^d} \Ebb^x \left[\int_0^t |f(Z_s)| ds\right] = 0.
\end{align}
When $\proo Z = \proo B$, we call this space {Kato-class}, and when $\proo Z = \proo b$, we call it
{relativistic Kato-class}.
}
\end{definition}
By Khasminskii's lemma \cite[Lemma 3.37]{LHB} and its straightforward extension to relativistic Kato-class it follows that the
random variables $-\int_0^t V(Z_s) ds$ are exponentially integrable for all $t \geq 0$, and thus we can define the Feynman-Kac semigroup
\begin{equation}
T_t f(x) = \Ebb^x\left[e^{-\int_0^t V(Z_s) ds} f(Z_t)\right], \quad f \in L^2(\RR^d), \ t>0.
\label{FK}
\end{equation}
Using the Markov property and stochastic continuity of the process $\proo Z$, we can show that $\{T_t: t\geq 0\}$ is a strongly
continuous one-parameter semigroup of symmetric operators on $L^2(\RR^d)$. Then, by the Hille-Yoshida theorem, there exists a self-adjoint
operator $K$ bounded from below such that $e^{-tK} = T_t$. Using the generator $K$, we can give a definition to a classical and a
relativistic Schr\"odinger operator for Kato-class potentials.
\begin{definition}
\rm{
If $\proo Z = \proo B$, then we call the self-adjoint operator $K$ on $L^2(\BR)$ a
{Schr\"odinger operator with
Kato-class potential $V$}. If $\proo Z = \proo b$, we call $K$ a {relativistic Schr\"odinger operator with Kato-class potential $V$}.
}
\end{definition}
For simplicity, we keep using the notations \kak{schroedinger} and (\ref{relp}). In both the non-relativistic and relativistic cases we
have then the following Feynman-Kac formula.
\begin{proposition}[\textbf{Functional integral representation}]
Let $f,g\in L^2(\BR)$. If $V$ is of Kato-class, then for any of the operators $K$ defined above we have
\begin{align}
\lk f, e^{-tK}g\rk_\LR=\int_{\BR}\Ebb^x[\bar{f}(Z_0)e^{-\int_0^tV(Z_s)ds} g(Z_t)] dx.
\end{align}
In particular,
\begin{align}
(e^{-tK}g)(x)=\Ebb^x[e^{-\int_0^tV(Z_s)ds} g(Z_t)], \quad x \in \BR.
\end{align}
\end{proposition}
\begin{proof}
See \cite[Sections 3.3, 3.6]{LHB}, \cite[Section 4]{HIL12}.
\end{proof}
Below we will need two-sided processes $\pro b$, i.e., indexed by the time-line $\RR$ instead of usually by the
semi-axis $\RR^+$. These processes can be defined as follows. Consider the measurable space $({\Omega}, {\B}({\Omega}))$,
with c\`adl\`ag space ${\Omega}=D_{\rm r}(\RR;\RR^d)$, as well as $\widehat \Omega = D_{\rm r}(\RR^{+}, \RR^d) \times
D_{\rm l}(\RR^{+}, \RR^d)$ and $\widehat {\mathscr P }^x = {\mathscr P }^x\times{\mathscr P }^x$, where $D_{\rm l}(\RR^{+}, \RR^d)$
denotes c\`agl\`ad space (i.e., paths continuous from left with right limits). Let $\omega=(\omega_1,\omega_2)\in \widehat
\Omega$ and define
$$
\widehat b_t(\omega)=
\left\{
\begin{array}{ll}\omega_1(t), & t\geq0,\\
\omega_2(-t), & t<0.
\end{array} \right.
$$
Since $\widehat b_t(\omega)$ is c\`adl\`ag in $t \in \RR$ under $\widehat {\mathscr P }^x$, we define
$b: (\widehat \Omega,{\B}(\widehat \Omega)) \to (\Omega, {\B}(\Omega))$ by $b_t(\omega)= \widehat b_t(\omega)$. Then we
have that $b \in {\B}(\widehat \Omega)/{\B}({\Omega})$ since $b^{-1}(E)\in {\B}(\widehat \Omega)$, for all cylinder sets
$E\in {\B}({\Omega})$. Thus $b$ is an $\Omega$-valued random variable on $\widehat \Omega$. Denote again the image measure
of $\widehat {\mathscr P }^x$ on $({\Omega},{\B}({\Omega}))$ with respect to $b$ by ${\mathscr P }^x = \widehat {\mathscr P }^x
\circ b^{-1}$. The coordinate process denoted by the same symbol $b_t:\omega\in {\Omega} \mapsto \omega(t)\in \RR^d$ is a
Cauchy (respectively, relativistic Cauchy) process over $\RR$ on $({\Omega}, {\B}({\Omega}), {\mathscr P }^x)$. The properties
of the so obtained process can be summarized as follows.
\begin{proposition}
The following hold:
\begin{enumerate}
\item
${\mathscr P }^x(b_0=x)=1$
\item
the increments $(b_{t_i}-b_{t_{i-1}})_{1\leq i\leq n}$ are independent symmetric Cauchy (respectively, relativistic Cauchy)
random variables for any $0=t_0<t_1<\cdots<t_n$ with $b_t-b_s\stackrel{\rm d}{=}b_{t-s}$ for
$t>s$
\item
the increments $(b_{-t_{i-1}}-b_{-t_i})_{1\leq i\leq n}$ are independent symmetric Cauchy (respectively, relativistic Cauchy)
random variables for any $0=-t_0>-t_1>\cdots>-t_n$ with $b_{-t}- b_{-s}\stackrel{\rm d}{=}b_{s-t}$
for $-t>-s$
\item
the function $\RR \ni t\mapsto b_t(\omega)\in \RR$ is c\`adl\`ag for every $\omega \in \Omega$
\item
$b_t$ and $b_s$ are independent for $t>0$ and $s<0$.
\end{enumerate}
\end{proposition}
A completely similar construction can be made of two-sided Brownian motion, with simplifications due to path continuity.
\subsection{Nelson Hamiltonian in boson Fock space}
The Nelson Hamiltonian is defined on a Hilbert space in terms of a self-adjoint operator bounded from below.
Consider the boson Fock space $\fff$ over $\LR$ defined as
\begin{align*}
\fff =\bigoplus_{n=0}^{\infty} \fff^{(n)},
\end{align*}
where $\fff^{(n)}= \otimes_{\rm sym}^{n}\LR .$ The Fock space can be identified with the space of $l_2$-sequences
$(\psi^{(n)})_{n\in \mathbb{N}}$ such that $\psi^{(n)}\in \fff^{(n)}$ and
\begin{align}\label{normfok}
\|\psi\|_{\fff}^2=\sum_{n=0}^{\infty} \|(\psi^{(n)})\|^2_{\fff^{(n)}}<\infty.
\end{align}
We denote the ``smeared" annihilation and creation operators by $a(f)$ and $\add (f)$, $f: \mathbb C \to \BR$, $f\in \LR $,
respectively, satisfying the canonical commutation relations
\begin{align*}
[a(f),\add (g)]=(\bar{f},g)\one,\quad [a(f),a(g)]=0=[\add (f),\add (g)]
\end{align*}
on a dense domain of $\fff$. Using these operators, the field operator and its conjugate momentum on $\fff $ are defined,
respectively, by
\begin{align*}
\Phi(f)=\frac{1}{\sqrt{2}}(\add (\bar{f})+a(f)) \quad \mbox{and} \quad \Pi(f)=\frac{i}{\sqrt{2}}(\add (\bar{f})-a(f)).
\end{align*}
For real-valued $L^2$-functions $f, g$, the commutation relations become
\eq{ccr}
[\Phi(f),\Pi(g)]=i(f,g), \quad [\Pi(f),\Pi(g)]=[\Phi(f),\Phi(g)]=0.
\en
Denote by $d\Gamma(T):\fff\to\fff$ the second quantization of a self-adjoint operator $T:\LR\to\LR$, defined by
$$
d\Gamma(T)=0\oplus
\left[
\oplus_{n=1}^\infty \sum_{j=1}^n \underbrace{\one\otimes\cdots \stackrel{j \rm th}{T}\cdots\otimes \one}_{n-\rm fold}\right].
$$
The self-adjoint operator
\begin{align*}
\hf =d\Gamma(\omega),
\end{align*}
is the {free field Hamiltonian}, where
\eq{omega}
\omega(k)=\sqrt{|k|^{2}+\nu^{2}}
\en
is the dispersion relation, and $\nu\geq 0$ denotes the mass of a single boson. Formally, the free field Hamiltonian can be written
as
\begin{align}\label{hf1}
\hf =\int_\BR \omega(k) \add (k)a(k)dk.
\end{align}
Physically, it describes the total energy of the interaction-free boson field since $\add (k)a(k)$ gives the number of bosons carrying
momentum $k$ and $\omega(k)$ is the energy of a single boson. The commutation relations
\eq{ccr2}
[\hf , a(f)]=-a(\omega f),
\quad
[\hf , \add (f)]=\add (\omega f)
\en
hold for $f\in D(\omega)$ on a dense domain of $\fff$. Hence we deduce that
\eq{ou1}
[\hf ,\Phi(f) ]=-i\Pi(\omega f).
\en
Let $\varphi:\BR \rightarrow\RR $ be a function describing the charge distribution of the particle, denote by $\vp$
its Fourier transform, and write $\widetilde{\widehat {\varphi}}(k)=\widehat {\varphi}(-k)$. For every $x\in\BR$, define
\eq{hi}
\hi (x)=\frac{1}{\sqrt{2}}\left( \add ({\widehat {\varphi}e^{-ikx}}/{\sqrt{\omega}})+
a({\widetilde{\widehat {\varphi}}e^{ikx}}/{\sqrt{\omega}})\right).
\en
Let $\hhh=\LR\otimes\fff$. We define the {interaction Hamiltonian} $\hi : \hhh \rightarrow\hhh$ by the constant fiber direct integral
$(\hi \Psi)(x)=\hi (x)\Psi(x)$ for $\Psi\in \hhh $ such that $\Psi(x)\in D(\hi(x))$, for almost every $x\in\BR$. Here
we use the identification
$$
\hhh\cong \int_\BR^\oplus \fff dx = \Big\{F:\mathbb{R}^{d}\rightarrow \fff \, \Big| \,
\|F\|^2_{\hhh}= \int_{\BR}\| F(x)\| _{\fff}^{2} dx<\infty \Big \}.
$$
Formally, this is written as
\eq{e}
\hi (x)=\int_\BR\frac{1}{\sqrt{2\omega(k)}}(\widehat {\varphi}(k)e^{-ikx}\add (k)+{\widehat {\varphi}(-k)}e^{ikx}a(k))dk.
\en
The {Nelson Hamiltonian} describing the interacting particle-field system
is then defined by
\eq{nelson}
H_{\rm N}=\hp \otimes\one +\one \otimes \hf +\hi
\en
Similarly, the {relativistic Nelson Hamiltonian} is defined by the operator
\eq{relnelson}
\tilde{H}_{\rm N} = \tilde{H}_{\rm p} \otimes\one +\one \otimes \hf +\hi
\en
on the space $\hhh $.
We will use the following standing assumptions throughout this paper.
\begin{assumption}\label{assumption}
{\rm
The following conditions hold:
\begin{enumerate}
\item
\label{r1}$\overline{\widehat {\varphi}(k)}=\widehat {\varphi}(-k)$
and
$\widehat {\varphi}/\sqrt{\omega},
\widehat {\varphi}/\omega\in\LR$.
\item
\label{ir}
$\vp/\omega\sqrt\omega\in \LR$.
\item
The external potential $V=V_+-V_-$ is of Kato-class in the sense of Definition \ref{kato}.
\item
$\hp $ has a unique, strictly positive ground state $\grp \in D(\hp)$, with $\hp \grp =E_{\rm p}\grp $,
$\|\grp\|_{L^2({\BR})} = 1$, where $E_{\rm p}=\is({\hp})$. Similarly, $\thp $ has a unique, strictly positive ground state
$\tilde\grp \in D(\thp)$, with $\thp \tilde\grp =\tilde E_{\rm p}\tilde\grp $, $\|\tilde\grp\|_{L^2({\BR})} = 1$, where $\tilde
E_{\rm p}=\is({\tilde\hp})$.
\item
$H_{\rm N}$ has a unique, strictly positive ground state $\gr \in D(H_{\rm N})$, with $H_{\rm N}\gr=E\gr$,
$\|\gr\|_{\hhh} = 1$, where $E=\is(H_{\rm N})$. Similarly, $\tilde{H}_{\rm N}$ has a unique, strictly
positive ground state $\tilde \gr \in D(\tilde{H}_{\rm N})$ with $\tilde{H}_{\rm N}\tilde \gr=\tilde E\tilde \gr$,
$\|\tilde\gr\|_{\hhh } = 1$, where $\tilde E=\is(\tilde{H}_{\rm N})$.
\end{enumerate}
}
\end{assumption}
Denote the ``free" operators by
$$
H_0= \hp \otimes\one +\one \otimes \hf \quad \mbox{and} \quad \tilde H_0 = \tilde \hp \otimes\one +\one \otimes \hf.
$$
The spectrum of $H_0$ can be derived from the spectra of $\hp$ and $\hf$. We have $\s(\hp) = [0,\infty)$, $\s(\hf)=\{0\}
\cup[\nu,\infty)$, and in the case $\nu=0$ the bottom of the spectrum of $H_0$ is the edge of the continuous spectrum. In
general, it is not clear whether the bottom of the spectrum of $H_{\rm N}$ is in the point spectrum or not, however,
whenever it is, the eigenfunction associated with this eigenvalue is a ground state. The same considerations hold also
for the relativistic operators.
Using Assumption \ref{assumption} it follows that $\hi$ is symmetric, and thus $H_{\rm N}$, $\tilde H_{\rm N}$ are
self-adjoint operators.
\bp{selfadjoint}
$H_{\rm N}$ is a self-adjoint operator on $D(\hp\otimes\one)\cap D(\one\otimes \hf)$ and essentially self-adjoint on any core
of $H_0$. Similarly, $\tilde H_{\rm N}$ is self-adjoint on $D(\tilde\hp\otimes\one)\cap D(\one\otimes \hf)$ and essentially
self-adjoint on any core of $\tilde H_0$.
\ep
\proof
For any $F \in D(H_0)$ we have
$$
\|\hi F\|\leq 2 \|\vp/2\omega\| \|\hf ^\han F\|+ \|\vp/\sqrt{2\omega}\| \|F\|.
$$
Let $\varepsilon>0$ be arbitrary. We obtain that
$$
\|\hf ^\han F\|^2 \leq (F, (\hp+\hf)F)-E_{\rm p}\|F\|^2\leq \varepsilon\|H_0 F\|^2 +
\left(\frac{1}{4\varepsilon}+|E_{\rm p}|\right) \|F\|^2.
$$
Thus there exists $b_\varepsilon > 0$ such that
$$
\|\hi F\|\leq \varepsilon\| H_0 F\|+b_\varepsilon\|F\|,
$$
and the claim follows by the Kato-Rellich theorem, see \cite[Theorem 3.11]{LHB}. The second part of the statement
follows similarly.
\qed
\subsection{Nelson Hamiltonian in function space}
\subsubsection{$P(\phi)_1$-process for the free particle operators}
\begin{definition}\TTT{$P(\phi)_1$-process}
\label{matu}
\rm{
Let $(E, \ms F, P)$ be a probability space and $K$ be a self-adjoint operator in $L^2(E, dP)$, bounded from below.
We say that an $E$-valued stochastic process $\pro Z$ on a probability space $(\ms Y, \B , \Q^z)$ is a $P(\phi)_1$-process
associated with $((E, \ms F, P), K)$ if conditions 1-4 below are satisfied:
\begin{enumerate}
\item
$\Q^z(Z_0=z)=1$.
\item\TTT{Reflection symmetry}
$(Z_t)_{t\geq 0}$ and $(Z_t)_{t\leq0}$ are independent and ${Z}_t\stackrel{\rm d}{=}{Z}_{-t}$ for every $t\in\RR$.
\item\TTT{Markov property}
$({Z}_t)_{t\geq 0}$ and $({Z}_t)_{t\leq0}$ are Markov processes with respect to the fields $\s(Z_s,0\leq s\leq t)$ and
$\s(Z_s,t\leq s\leq 0)$, respectively.
\item
\TTT{Shift invariance}
Let $-\infty<t_0\leq t_1<...\leq t_n<\infty$, $f_j\in L^{\infty}(E,dP)$, $j=1,...,n-1$ and $f_0,f_n\in L^2(E,dP)$.
Then for every $s\in\RR$,
\begin{align}
\int_{E}\Ebb_{\Q^z}\bigg[ \prod_{j=0}^{n}f_{j}({Z}_{t_{j}})\bigg] dP(z) &=
\int_{E}\Ebb_{\Q^z}\bigg[ \prod_{j=0}^{n}f_{j}({Z}_{t_{j}+s})\bigg] dP(z)\non\\
&=(\one, f_0 e^{-(t_{1}-t_{0})K }f_1\cdots
f_{n-1}e^{-(t_{n}-t_{n-1})K}f_n).
\end{align}
\end{enumerate}
}
\end{definition}
Denote
\begin{equation}
\label{N}
d{\rm N}(y)=\grp^2(y) dy.
\end{equation}
Since by part (4) of Assumption (\ref{assumption}) the function $\grp$ is square integrable and $L^2$-normalized, $d{\rm N}$ is
a probability measure on $\BR$. Define the unitary operator $U_{\rm p}: L^{2}(\BR ,d{\rm N}) \rightarrow L^{2}(\BR,dy)$
by $U_{\rm p}: f \mapsto \grp f$. Using that $\grp$ is strictly positive, the image
\begin{align}
\label{Lp}
L_{\rm p}=U_{\rm p}^{-1}(\hp -E_{\rm p})U_{\rm p}=\frac{1}{\grp }(\hp -E_{\rm p})\grp,
\end{align}
of the Schr\"odinger operator \kak{schroedinger} under this map is well-defined and has the domain $D(\lp)=\{f\in L^2(\BR,d{\rm N})
\,|\, f\gr\in D(\hp)\}$. Since $e^{-t\lp} \one=\one$ for the identity function $\one\in L^2(\BR,d{\rm N})$, the operator $\lp$
is the generator of a Markov process.
\bp{pphi}
If $V$ is in Kato class, then
\begin{itemize}
\item[\rm 1.]
there exists a probability measure $\N^y$ on $(\X, \B(\X))$ such that the coordinate process $\pro Y$ on
$(\X, \B(\X),\N^y)$ is a $P(\phi)_1$-process starting from $y \in \BR$, associated with the pair
$((\BR,\B (\BR), d{\rm N}), \lp)$
\item[\rm 2.]
the function $t\mapsto Y_t$ is almost surely continuous.
\end{itemize}
\ep
\proof
See \cite[Theorem 3.106]{LHB}.
\qed
We can define a $P(\phi)_1$-process for the relativistic Schr\"odinger operator (\ref{relp}) in a similar way. As
above, denote
\begin{equation}
\label{tilN}
d\tilde {\rm N}=\tgrp^2({x}) dx,
\end{equation}
which is a probability measure on $\BR$ for similar reasons as for the non-relativistic operator. Taking now the unitary map
$\tilde U_{\rm p}:L^{2}(\BR, d\tilde {\rm N})\rightarrow L^{2}(\BR , dx)$, $f \mapsto \tgrp f$, we similarly obtain
\begin{align}
\label{tilLp}
\tilde{L}_{\rm p}=\tilde U_{\rm p}\f(\tilde{H}_{\rm p}-\tilde E_{\rm p}) \tilde U_{\rm p}=
\frac{1}{\tgrp}(\tilde{H}_{\rm p}-\tilde E_{\rm p})\tgrp,
\end{align}
which is again a Markov generator. Then we have
\begin{theorem}
If $V$ is in relativistic Kato class, then
\begin{enumerate}
\item[\rm 1.]
there exists a probability measure $\tilde{\N}^y$ on $(\D, \B(\D))$ such that the coordinate process
$\pro {\tilde Y}$ on $(\D, \B(\D),\tilde{\N}^y)$ is a $P(\phi)_1$-process starting from $y \in \BR$, associated with
$((\BR, \B (\BR), d\tilde{\rm N}),\tilde{L}_{\rm p})$
\item[\rm 2.]
the function $t\mapsto \tilde Y_t$ is almost surely c\`adl\`ag.
\end{enumerate}
\end{theorem}
\proof
See \cite[Theorem 5.1]{LO12}.
\qed
\medskip
\noindent
\subsubsection{Infinite dimensional Ornstein-Uhlenbeck process}
Let $\K $ be a Hilbert space over $\RR $, defined by the completion of $D(1/\sqrt\omega)\subset\LR$ with respect to the
norm determined by the scalar product
\begin{align}
(f,g)_{\K }=\int_{\BR }\overline{\widehat {f}(k)}\widehat {g}(k)\frac{1}{2\omega(k)}dk,
\end{align}
i.e.,
$$
\K=\overline{D(1/\sqrt\omega)}^{\|\cdot\|_\K}.
$$
Let $T:\K\to\K$ be a positive self-adjoint operator with Hilbert-Schmidt inverse such that $\sqrt{\omega}{T}^{-1}$ is
bounded. Define the space $C^{\infty}({T})=\displaystyle\cap_{n=1}^\infty D(T^{n})$, and write
\begin{align*}
\K _{n}=\overline{C^{\infty}({T})}^{\|{T}^{n/2}\cdot \|_{\K}}.
\end{align*}
We construct a triplet $\K _{+2}\subset \K\subset\K _{-2}$, where we identify $\K ^\ast _{+2}=\K _{-2}$. Write
$Q=\K _{-2}$, and endow $Q$ with its Borel $\s$-field $\B (Q)$, defining the measurable space $(Q, \B(Q))$.
Consider the set $\Y =C(\RR , Q)$ of continuous functions on $\RR$, with values in $Q$, and denote its Borel $\s$-field
by $\mathcal B(\Y)$. We define a {$Q$-valued Ornstein-Uhlenbeck process} $\pro \xi$,
$$
\RR \ni t\mapsto\xi_{t}\in Q
$$
on the probability space $(\Y ,\B (\Y), \G )$ with probability measure $\G $. Let $\xi_{t}(f)=\lk \!\lk \xi_{t}, f\rk \!\rk$
for $f\in \M _{+2}$, where $\lk \!\lk .,.\rk \!\rk $ denotes the pairing between $Q$ and $\M _{+2}$. Then for every $t \in \mathbb{R}$
and $f$ we have that $\xi_{t}(f)$ is a Gaussian random variable with mean zero and covariance
\begin{align}\label{eq1}
\Ebb_{\G }[\xi_{t}(f)\xi_{s}(g)]=\int_{\BR }\overline{\widehat {f}(k)}\widehat {g}(k)e^{-|t-s|\omega(k)}\frac{1}{2\omega(k)}dk.
\end{align}
Note that by \kak{eq1} every $\xi_{t}(f)$ can be uniquely extended to test functions $f\in\M $, which for simplicity we will
denote in the same way.
In what follows we will need conditional measures of this Gaussian measure. Since the conditional expectation
$\Ebb_{\G }[1_A|\s(\xi_0)]$ with respect to $\s(\xi_0)$ is trivially $\s(\xi_0)$-measurable, there exists a measurable
function $h:Q\to\RR$ such that $h\circ \xi_0(\omega)=\Ebb_{\G }[1_A|\s(\xi_0)](\omega)$. We will use the notation $h(\xi)=
\G(A|\xi_0=\xi)$, however, we remark that $\G (A|\xi_0=\xi)$ is well defined for $\xi\in Q\setminus N_A$ with a null set
$N_A$ only. Nevertheless, since $Q$ is a separable complete metric space, there exists a null set $N$ such that $\G (A|\xi_0=\xi)$
is well defined for all $A$ and $\xi\in Q\setminus N$. The notation $\G ^{\xi}(\cdot)=\G (\cdot \,| \xi_{0}=\xi)$ for the family
of conditional probability measures $\G (\cdot|\xi_0=\xi)$ on $\Y$ with $\xi\in Q\setminus N$ makes then sense, and it is seen
that $\G ^{\xi}$ is a regular conditional probability measure. Then we have $\Ebb_{\G }[...]=\int_{Q} \Ebb_{\G ^{\xi}}[...]
d{{\rm G}}(\xi)$, where ${{\rm G}}$ is the distribution of the random process $\pro \xi$ on the measurable space $(Q, \B (Q))$,
and it is the stationary measure of $\G $. Thus we are led to the probability space $(Q, \B(Q),{\rm G})$.
\medskip
\noindent
\subsubsection{Functional integral representation of the Nelson Hamiltonians}
The Wiener-It\^{o}-Segal isomorphism $\WIS :\fff \longrightarrow L^{2}(Q, d{{\rm G}})$, $\WIS \Phi(f)\WIS ^{-1}=\xi(f)$,
establishes a close connection between $\fff $ and $L^{2}(Q, d{{\rm G}})$. Using (\ref{N}) and the stationary measure obtained
above, define the product measure
\eq{pzero}
{\rm P}={\rm N}\otimes {{\rm G}},
\en
which is a probability measure on the product space $\hir$. The unitary map
$$
U_{\rm p}\otimes\WIS :\hhh \longrightarrow L^{2}(\hir , d{\rm P})
$$
establishes a unitary equivalence between $L^{2}(\hir , d{\rm P})$ and $\hhh $, and we make the identification
\eq{identification}
\hhh \cong \LR \otimes L^{2}(Q)\cong L^{2}(\hir ,d{\rm P}).
\en
For convenience, hereafter we write $ L^{2}(\hir , d{\rm P})$ simply as $L^{2}({\rm P})$, moreover $L^{2}({\rm N})$ and
$L^{2}( {{\rm G}})$ for $L^{2}(\BR , d{\rm N})$ and $L^{2}(Q, d{{\rm G}})$, respectively. The images of the free field
and interaction Hamiltonians on $L^2({\rm P})$ under this unitary map are given by
\eq{hf}
{\hf }^\WIS =\WIS \hf \WIS ^{-1}
\en
and
\eq{hi2}
{\hi }^\WIS (y)=\WIS \hi \WIS ^{-1}(y)=\xi(\tilde{\varphi}(\cdot-y)), \quad y\in\BR.
\en
Here $\tilde{\varphi}$ is the inverse Fourier transform of $\vp/\sqrt\omega$. To simplify the notations, we write again
$\hf $ for ${\hf }^\WIS $, and $\hi$ for ${\hi }^\WIS $. Then the classical Nelson Hamiltonian $H_{\rm N}$ is unitary equivalent
with
\begin{align}
H= L_{\rm p}\otimes\one +\one \otimes \hf +\hi
\end{align}
acting on $L^{2}({\rm P})$, where $L_{\rm p}$ is given by (\ref{Lp}).
Recall that
$\pro Y$ is the $P(\phi)_{1}$-process associated with the pair $((\X ,\B(\X), \N^y), \lp)$, and write
$$
d\N=d{\rm N}(y)d\N^y.
$$
The probability space for the joint system without the particle-field interaction is then the product space
$(\ms X\times\ms Y, \Sigma, \P_0)$, where $\Sigma=\B(\X)\otimes\B(\Y)$ and
$$
\P _{0}=\N \otimes\G.
$$
Define the shift operator $\tau_s: L^2(Q,d{{\rm G}})\mapsto L^2(Q,d{{\rm G}})$ by $\tau_s\xi(h)=\xi(h(\cdot-s))$. We have
then the following functional integral representation for the classical Nelson Hamiltonian $H$ in $L^2({\rm P})$.
\begin{proposition}[\textbf{Functional integral representation}]\label{fik}
Let $\Phi,\Psi\in L^{2}({\rm P})$ and suppose that $s\leq 0\leq t$. Then
\begin{align*}
(\Phi,e^{-(t-s)H}\Psi)_{L^{2}({\rm P})}= \Ebb_{\P _{0}}[\Phi(Y_{s},\xi_{s})e^{-\int_{s}^{t}
\tau_{Y_{r}}\xi_{r}(\vpp)dr}\Psi(Y_{t},\xi_{t})].
\end{align*}
\end{proposition}
See \cite[Theorem 6.2]{LHB}.
\qed
\begin{corollary}\label{c1}
Let $\Phi,\Psi\in L^{2}({\rm P})$,
$t_{0}\leq...\leq t_{n}$ and
$A_{0},...,A_{n}\in\B (\hir)$. Then,
\begin{align*}
&(\Phi,\one _{A_{0}}e^{-(t_{1}-t_{0})H}\one _{A_{1}}
e^{-(t_{2}-t_{1})H} \cdots \one _{A_{n-1}}e^{-(t_{n}-t_{n-1})H}\one _{A_{n}}
\Psi)_{L^2({\rm P})} \\
&= \Ebb_{\P _{0}}
\left [\left(\prod_{j=0}^{n}\one _{A_{j}}(Y_{t_{j}},\xi_{t_{j}})\right)
\ov{\Phi(Y_{t_{0}},\xi_{t_{0}})}e^{-\int_{t_{0}}^{t_{n}}\tau_{Y_{s}}\xi_{s}
(\vpp)ds}\Psi(Y_{t_{n}},\xi_{t_{n}})\right].
\end{align*}
\end{corollary}
\proof
We have
\begin{align*}
&(\Phi,\one _{A_{0}}e^{-sH}\one _{A_{1}}e^{-tH}
\one _{A_{2}}\Psi)\\
& = \Ebb_{\P _{0}}[\ov{\Phi(Y_{0},\xi_{0})}
\one _{A_{0}}(Y_{0},\xi_{0}) e^{-\int_{0}^{s}\tau_{Y_{r}}\xi_{r}
(\vpp)dr}
\one _{A_{1}}(Y_{s},\xi_{s}) e^{-\int_{s}^{s+t}\tau_{Y_{r}}\xi_{r}
(\vpp)dr}
\one _{A_{2}}(Y_{s+t},\xi_{s+t})\Psi(Y_{s+t},\xi_{s+t})]\\
& = \Ebb_{\P _{0}}[\ov{\Phi(Y_{0},\xi_{0})}\one _{A_{0}}(Y_{0},\xi_{0})\one _{A_{1}}(Y_{s},\xi_{s})
\one _{A_{2}}(Y_{s+t},\xi_{s+t})
e^{-\int_{0}^{s+t}\tau_{Y_{r}}\xi_{r}
(\vpp)dr}\Psi(Y_{s+t},\xi_{s+t})].
\end{align*}
By iterating this, we obtain
\begin{align*}
&(\Phi,\one _{A_{0}}e^{-(t_{1}-t_{0})H}\one _{A_{1}}e^{-(t_{2}-t_{1})H} \cdots\one _{A_{n-1}}e^{-(t_{n}-t_{n-1})H}\one _{A_{n}}\Psi) \\
&=\Ebb_{\P _{0}}\left[
\lk
\prod_{j=0}^{n}\one _{A_{j}}(Y_{t_{j}-t_{0}},\xi_{t_{j}-t_{0}})\rk \ov{\Phi(Y_{0},\xi_{0})}
e^{-\int_{0}^{t_{n}-t_{0}}\tau_{Y_{s}}\xi_s(\vpp)ds}\Psi(Y_{t_{n}-t_{0}},\xi_{t_{n}-t_{0}})\right].
\end{align*}
Since both $\N $ and $\G $ are invariant under time shift, we can replace $Y_{s}$ by $Y_{s+t_{0}}$ and $\xi_{s}$ by $\xi_{s+t_{0}}$,
to find that
\begin{align*}
&(\Phi,\one _{A_{0}}e^{-(t_{1}-t_{0})H}\one _{A_{1}}e^{-(t_{2}-t_{1})H} \cdots\one _{A_{n}}e^{-(t_{n}-t_{n-1})H}\Psi)\\
&=\Ebb_{\P _{0}}\left[
\lk \prod_{j=0}^{n}\one _{A_{j}}(Y_{t_{j}},\xi_{t_{j}})
\rk \ov{\Phi(Y_{t_{0}},\xi_{t_{0}})}
e^{-\int_{t_{0}}^{t_{n}}\tau_{Y_{s}}\xi_s(\vpp)ds}\Psi(Y_{t_{n}},\xi_{t_{n}})\right].
\end{align*}
\qed
Finally, for later use we quote the following representation formula using Wiener measure instead of the particle
$P(\phi)_1$-measure.
\begin{proposition}
\label{B}
Let $\Phi,\Psi\in L^{2}({\rm P})$ and $s\leq 0\leq t$.
Then
\begin{align*}
&(\Phi,e^{-(t-s)H}\Psi)_{L^{2}({\rm P})}\\
&=
\int_{\BR\times Q}\!\! \Ebb_{\W\otimes{\G}} ^{(y,\xi)}\left[\Phi(B_{s},\xi_{s})\grp(B_s)e^{-\int_{s}^{t}\tau_{B_{r}}\xi_r(\vpp)dr}
\Psi(B_{t},\xi_{t})\grp(B_t) e^{-\int_s^t V(B_r) dr}\right] dy\otimes d{{\rm G}}.
\end{align*}
\end{proposition}
\proof
See \cite[Theorem 6.3]{LHB}.
\qed
\noindent
For convenience, we write
$$
\P _{0}^{(y,\xi)}=\N^y \otimes\G^\xi, \quad (y,\xi)\in \hir,
$$
so that $\Ebb_{\P_0}[\cdots] = \int _\hir \Ebb_{\P_0}^{(y,\xi)}[\cdots] d{\rm P}(y,\xi)$.
The above transformations and constructions can be repeated for the relativistic model. Then we have
\begin{equation}
\label{tilP}
\tilde {\rm P}=\tilde {\rm N}\otimes {\rm G}
\end{equation}
and the relativistic Nelson Hamiltonian $\tilde H $ in $L^{2}(\tilde {\rm P})$ becomes
\eq{RN}
\tilde H =\tilde{ L}_{\rm p}\otimes\one +\one \otimes \hf +\hi,
\en
using (\ref{tilLp}). We write
$$
d\tilde{\N}=d\tilde {\rm N}(y)d\tilde{\N}^y.
$$
We consider the probability space $(\tilde{\X}\otimes\Y, \tilde{\Sigma},\tilde{\P}_0)$
where $\tilde{\Sigma}=\B(\tilde {\X})\otimes\B(\Y)$ and
$\tilde{\P}_0=\tilde{\N}\otimes\G.$
Then we have the following expression for the relativistic Nelson Hamiltonian in $L^2(\rm \tilde{P})$.
\begin{proposition}
Let $t_{0}\leq...\leq t_{n}$, $f_0,f_n\in L^2({\rm\tilde {P}})$
and $f_j\in L^\infty(\tilde{\rm P})$, for $j=1,\ldots,n-1$.
Then
\begin{align*}
(f_0, e^{-(t_{1}-t_{0})\tilde H }f_1
e^{-(t_{n}-t_{n-1})\tilde H }f_n)
= \Ebb_{\tilde{\P} _{0}}
\bigg [\ov{f_0(\tilde{Y}_{0},\xi_{0})}
\bigg(\prod_{j=1}^{n}f_j(\tilde{Y}_{t_{j}},\xi_{t_{j}})\bigg)
e^{-\int_{t_{0}}^{t_{n}}\tau_{\tilde{Y}_{s}}\xi_{s}
(\vpp)ds}
\bigg].
\end{align*}
\end{proposition}
\proof
The proof is analogous to the proof of Corollary \ref{c1} (see \cite[Theorem 6.2]{LHB}).
\qed
Now we can give a functional integral representation of $e^{-t\tilde{H}_{\rm N}}$ by making use of the L\'evy process $(\xx_t)_{t\in \RR}$
and the infinite dimensional OU-process
$\pro {\xi}$.
\begin{proposition}\label{fkl}
Let $\Phi,\Psi\in L^{2}(\tilde{\rm P})$ and $s\leq 0\leq t$.
Then
\begin{align*}
&(\Phi,e^{-(t-s)\tilde H }\Psi)_{L^{2}(\tilde{\rm P})}\\
&=
\int_{\BR\times Q}\!\! \Ebb_{\mathscr P\otimes\G}^{(y,\xi)}\left[
\Phi(\xx_{s},\xi_{s})\tgrp(\xx_s)
e^{-\int_{s}^{t}\tau_{\xx_{r}}\xi_r(\vpp)dr}
\Psi(\xx_{t},\xi_{t})\tgrp(\xx_t) e^{-\int_s^t V(\xx_r) dr}\right] dy\otimes d\rm G.
\end{align*}
\end{proposition}
\proof
The proof is similar to the proof of Proposition \ref{B} (see \cite[Theorem 6.3]{LHB}).
\qed
\noindent
In what follows we write $\tilde{\P}_0^{(y,\xi)}=\tilde{\N}^y\otimes\G^\xi$ for $(y,\xi)\in\hir$.
\section{$P(\phi)_{1}$-processes associated with the Nelson Hamiltonians}\label{sec}
\subsection{Classical Nelson model}
Since the ground state $\gr$ of $H$ is strictly positive and $L^2$-normalized,
we define the probability measure
$$
d{\rm M}=\gr ^{2}d{\rm P}
$$
on $\hir $. Also, we define the unitary operator $U_{\rm g}:L^{2}(\hir, d{\rm M})\rightarrow L^{2}(\hir , d{\rm P})$ by
$U_{\rm g}: \Phi \mapsto \gr \Phi$. We write $\K=L^{2}(\hir , d{\rm M})$ and define the self-adjoint operator
$$
\LN =\frac{1}{\gr }(H-E)\gr.
$$
Let
$\XXQ=C(\RR;\hir)$ be the set of continuous paths with values in $\hir$ and indexed by the real line $\RR$, and
$\B_Q$ the $\s$-field generated by the cylinder sets. The main result of this section is the following.
\bt{pp1}\TTT{$P(\phi)_1$-process for the classical Nelson Hamiltonian}
Let $(y,\xi)\in \hir$. Then the following hold.
\begin{enumerate}
\item[\rm 1.]
There exists a probability measure $\P^{(y,\xi)}$ on $(\XXQ, \B_Q)$ such that the coordinate process $\pro X$ on
$(\XXQ, \B_Q,\P^{(y,\xi)})$ is a $P(\phi)_1$-process associated with the pair\\
$\bigg((\hir, \B (\BR)\otimes \B (Q), d{\rm M}), \LN \bigg)$.
\item[\rm 2.]
The function $t\mapsto X_t$ is almost surely continuous.
\end{enumerate}
\et
In order to show this theorem we need a string of lemmas. The idea of proof is taken from \cite{LHB}.
Write $\Sigma'= \B (\BR)\otimes \B (Q)$ and define the family of set functions
$\{\M _{\Lambda} \, | \,\Lambda\subset[0,\infty), \#\Lambda <\infty\}$ on ${\Sigma'} ^{\#\Lambda}=\underbrace
{ \Sigma'\times\cdots\times \Sigma'}_{\#\Lambda-\mbox{\tiny{times}}}$ by
\begin{align*}
\M _{\Lambda}(A_{0}\times A_{1}\times...\times A_{n})
=\lk \one _{A_{0}},
e^{-(t_{1}-t_{0})L_{\rm N}}\one _{A_{1}}e^{-(t_{2}-t_{1})L_{\rm N}}\one _{A_{1}} \cdots\one _{A_{n-1}}e^{-(t_{n}-t_{n-1})L_{\rm N}}\one _{A_{n}} \rk_{\ms K}
\end{align*}
for $\Lambda=\{t_0,\ldots,t_n\}$, $n \in \mathbb{N}$. It is straightforward to show that the family of set functions $\M _{\Lambda}$
satisfies the Kolmogorov consistency relation
$$
\M _{\{t_{0},t_{1},...,t_{n+m}\}}((\times_{i=0}^{n}A_{i})\times (\times_{i=n+1}^{n+m}\hir ))=\M _{\{t_{0},t_{1},...,t_{n}\}}
(\times_{i=0}^{n}A_{i}).
$$
Define the projection $\pi_{\Lambda}: (\hir)^{[0,\infty)}\longrightarrow (\hir)^{\Lambda}$ by $w \longmapsto (w(t_{0}),...,w(t_{n}))$
for $\Lambda=\{t_{0},...,t_{n}\}$, $n \in \mathbb{N}$. Then
$$
\ms A = \{\pi_{\Lambda}\f(A) \,|\, A\in {\Sigma' }^{\#\Lambda},\, \#\Lambda<\infty \}
$$
is a finitely additive family of sets, and the Kolmogorov extension theorem \cite[Theorem 2.2]{KA} yields that there exists a unique probability measure $\M$
on $((\hir)^{[0,\infty)},\sigma(\ms A ))$ such that
$$
\M (\pi_{\Lambda}^{-1}(A_{1}\times...\times A_{n})=\M _{\Lambda}(A_{1}\times...\times A_{n}),
$$
for all $\Lambda\subset[0,\infty)$ with $\#\Lambda<\infty$ and $A_{j}\in \Sigma' $, and
\begin{align}
\label{p2}
\M _{\{t_0,\ldots,t_n\}}(A_0\times\cdots\times A_n)=\Ebb_\M \bigg[\prod_{j=0}^n \one_{A_{t_j}}(Z_{t_j})\bigg]
\end{align}
holds. Here $(Z_t)_{t\geq0}$ is the coordinate process defined by $Z_t(\omega)=\omega(t)$ for $\omega \in (\hir)^{[0,\infty)}$.
\begin{lemma}
\label{cont}
The random process $(Z_{t})_{t\geq 0 }$ on $((\hir)^{[0,\infty)}, \s(\ms A))$ has a continuous version.
\end{lemma}
\proof
We write $Z_t=(x_t,\xi_t)$, where $x_t\in\BR$ and $\xi_t\in Q$ are the coordinate processes $x_t(\omega)=\omega^1(t)$ and
$\xi_t(\omega)=\omega^2(t)$ for all $t\geqslant0$ and $\omega=(\omega^1,\omega^2)\in (\hir )^{[0,\infty)}$. Define
$\|Z_t\|_{\hir}=\sqrt{\|x_t\|^2_{\BR}+\|\xi_t\|_Q^2}$. Using the Kolmogorov-\v{C}entsov theorem \cite[Theorem 2.8]{KA}, the estimate
\eq{kc}
\Ebb_\M [\| Z_{t}-Z_{s}\| _{\hir }^{4}]\leq D|t-s|^{2}
\en
with some $D>0$ implies that $(Z_{t})_{t\geq 0 }$ has a continuous version. Since
$$
\|Z_t\|_{\hir}^4\leq 2(\|x_t\|^4_{\BR}+\|\xi_t\|_Q^4),
$$
it will suffice to prove the bounds
\begin{align}
&\label{33}
\Ebb_\M [\| x_{t}-x_{s}\| _{\BR}^{4}]\leq D_1|t-s|^{2},\\
&\label{34}
\Ebb_\M [\| \xi_{t}-\xi_{s}\| _{Q}^{4}]\leq D_2|t-s|^{2}.
\end{align}
To obtain \kak{33}, recall the moments formula $\Ebb[|B_t-B_s|^{2n}]= K_n |t-s|^n$, with a constant $K_n$ for $n\geq0$. Let
$x_t=(x_t^1,\ldots,x_t^d)$. By using the formula in Proposition \ref{B}, we have for all $1\leq i,j\leq d$ that
\begin{align*}
&\Ebb_\M [(x_t^i)^n (x_s^j)^m]\\
&=((x^i)^n\gr, e^{-(t-s) (H-E)} (x^j)^m\gr)_{L^2({\rm P})}\\
&=((x^i)^n\grp \gr, e^{-(t-s) (H_{\rm N}-E)} (x^j)^m\grp \gr)_{\LR\otimes L^2(Q)}\\
&=\int_\BR \!\! dx\Ebb_{\W\times \G}^x\left[
(B_{0}^i)^n (B_{t-s}^j)^m \grp(B_0)\gr(B_0,\xi_0)
\grp(B_{t-s})\gr(B_{t-s},\xi_{t-s})\right.\\
&\left.\hspace{3cm}\times
e^{-\int_0^{t-s} \tau_{B_r}\xi_r(\vpp ))dr}
e^{-\int_0^{t-s} V(B_r) dr}\right].
\end{align*}
Using the eigenvalue equations and the Feynman-Kac formula for the free particle and the full Nelson Hamiltonians, it follows
that
$$
\underset{x\in\BR} \sup|\grp(x)|=C_1<\infty \quad \mbox{and} \quad \underset{(x,\xi)\in \hir } \sup|\gr(x,\xi)|=
C_2<\infty.
$$
Thus we have
\begin{align*}
&\Ebb_\M [|x_t- x_s|^4]
=
\int_\BR \!\! dx\Ebb_{\W\times \G}^x\left[|B_{0}- B_{t-s}|^4 \grp(B_0)\gr(B_0,\xi_0)
\grp(B_{t-s})\gr(B_{t-s},\xi_{t-s}) \right. \\&
\qquad \left.\times
e^{-\int_0^{t-s} \tau_{B_r}\xi_r(\vpp )} e^{-\int_0^{t-s} V(B_r+x) dr} \right]\\
&\leq
C_2^2 \int_\BR \!\! dx \Ebb_{\W\times \G}\grp(x) \left[\grp(B_{t-s}+x)|B_{0}- B_{t-s}|^4 e^{-\int_0^{t-s}
\tau_{B_r+x}\xi_r(\vpp ))dr} e^{-\int_0^{t-s} V(B_r+x) dr} \right]\\
&\leq
C_2^2 \lk\int_\BR \!\! dx \Ebb_{\W\times \G}\left[|\grp(B_{t-s}+x)|^2 \right]\rk^\han\\
&\qquad \times
\lk \int_\BR \!\! dx |\grp(x)|^2 \Ebb_{\W}\left[ |B_{0}- B_{t-s}|^8 e^{-2\int_0^{t-s} V(B_r+x) dr}
\Ebb_{\G}\left[e^{-2 \int_0^{t-s} \tau_{B_r+x}\xi_r(\vpp ))dr}\right]\right] \right)^\han.
\end{align*}
Since
$$
\Ebb_{\G }[ e^{-\int_{0}^{t-s}\tau_{B_{r}}\xi_{r}(\vpp )dr}] \leq e^{({t-s}) \|\vp/\omega\|^2} =C,
$$
we see that
\begin{align*}
\Ebb_\M [|x_t- x_s|^4]
&\leq
C_2^2 C \|\grp\|^2\lk \Ebb_{\W}\left[|B_{0}- B_{t-s}|^{16} \right]\rk^{1/4}\lk
\Ebb_\W[e^{-4\int_0^{t-s} V(B_r+x) dr}]\rk^{1/4}\\
&\leq
C_2^2 C K_8^{1/4} \sup_{x\in\BR} \lk\Ebb_\W[e^{-4\int_0^{t-s} V(B_r+x) dr}]\rk^{1/4}|t-s|^2.
\end{align*}
Thus \kak{33} follows.
Next we prove \kak{34}.
Let $f\in \ms M_{+2}$.
In the same way as in the proof of \kak{33} we have
\begin{align*}
\Ebb_\M [\xi_t(f)^n \xi_s(f)^m]
&=(\xi(f)^n\gr, e^{-(t-s) (H-E)} \xi(f)^m\gr)_{L^2(d\rm P)}\\
&=(\xi(f)^n\grp \gr, e^{-(t-s) (H_{\rm N}-E)} \xi(f)^m\grp \gr)_{\LR\otimes L^2(Q)}\\
&=\int_\BR \!\! dx\Ebb_{\W\times \G}^x\left[
\xi_0(f)^n \xi_s(f)^m \grp(B_0)\gr(B_0,\xi_0)
\grp(B_{t-s})\gr(B_{t-s},\xi_{t-s})\right.\\
&\hspace{3cm}
\left.\times
e^{-\int_0^{t-s} \tau_{B_r}\xi_r(\vpp ))dr}
e^{-\int_0^{t-s} V(B_r) dr}
\right].
\end{align*}
Hence
\begin{align*}
&\Ebb_\M [|\xi_t(f)-\xi_s(f)|^4]\\
&=
\int_\BR \!\! dx\Ebb_{\W\times \G}^x
\left[
|\xi_0(f)-\xi_{t-s}(f)| ^4 \grp(B_0)\gr(B_0,\xi_0)
\grp(B_{t-s})\gr(B_{t-s},\xi_{t-s})\right.\\
&\hspace{3cm}
\left.\times
e^{-\int_0^{t-s} \tau_{B_r}\xi_r(\vpp ))dr}
e^{-\int_0^{t-s} V(B_r+x) dr}\right]\\
& \leq C_2^2
\int_\BR \!\! dx
\Ebb_{\W\times \G}
\left[\grp(x)
\grp(B_{t-s}+x)
|\xi_t(f)-\xi_{t-s}(f)| ^4 \right.\\
&\hspace{3cm}
\left.\times e^{-\int_0^{t-s} \tau_{B_r+x}\xi_r(\vpp ))dr}
e^{-\int_0^{t-s} V(B_r+x) dr}
\right]\\
&
\leq C_2^2
\int_\BR \!\! dx
\Ebb_{\W}
\left[\grp(x)
\grp(B_{t-s}+x)
e^{-\int_0^{t-s} V(B_r+x) dr}\right.\\
&\left.\hspace{3cm}\times
\Ebb_{\G}\left[
|\xi_t(f)-\xi_{t-s}(f)| ^4 e^{-\int_0^{t-s} \tau_{B_r+x}\xi_r(\vpp ))dr}\right]\right]
\\
&
\leq C_2^2
\int_\BR \!\! dx
\Ebb_{\W}
\left[\grp(x)
\grp(B_{t-s}+x)
e^{-\int_0^{t-s} V(B_r+x) dr}
\lk \Ebb_{\G}\left[
|\xi_t(f)-\xi_{t-s}(f)| ^8\right]\rk^\han\right.\\
&\hspace{3cm}\left.\times
\lk
\Ebb_{\G}\left[ e^{-2\int_0^{t-s} \tau_{B_r+x}\xi_r(\vpp ))dr}\right]\rk^\han
\right].
\end{align*}
By Lemma \ref{bound} below we have
\begin{align*}
&\Ebb_\M [|\xi_t(f)-\xi_s(f)|^4]\\
&
\leq C_2^2
|t-s|^2 \|f\|^2_{\ms M_{+2}}
C
\int_\BR \!\! dx
\Ebb_{\W}
\left[
\grp(x)\grp(B_{t-s}+x)
e^{-\int_0^{t-s} V(B_r+x) dr}
\right]\\
&\leq
C_2^2
|t-s|^2 \|f\|_{\ms M_{+2}}^2
C
\|\grp\|^2
\sup_{x\in\BR}
\lk
\Ebb_\W[
e^{-2\int_0^{t-s} V(B_r+x) dr}]\rk^\han.
\end{align*}
Hence
\eq{tg}
\Ebb_\M \left[\frac{|\xi_t(f)-\xi_s(f)|^4}{\|f\|^2_{\ms M_{+2}}}\right]
\leq D_2
|t-s|^2
\en
with a constant $D_2$.
Since $\|\xi_t-\xi_s\|_Q=\sup_{f\not=0}|\xi_t(f)-\xi_s(f)|/\|f\|_{\ms M_{+2}}$,
for every $\eps>0$ there exists $f_\eps\in \ms M_{+2}$ such that
$\|\xi_t-\xi_s\|_Q\leq |\xi_t(f_\eps)-\xi_s(f_\eps)|/\|f_\eps\|_{\ms M_{+2}}+\eps$. Thus
together with \kak{tg} we have
\eq{tg2}
\Ebb_\M \left[{|\xi_t-\xi_s|^4}\right]-\eps
\leq D_2 |t-s|^2,
\en
and thus \kak{34} follows.
\qed
\bl{bound}
We have
$\Ebb_\G[|\xi_t(f)-\xi_s(f)|^{2n}]\leq D_n|t-s|^n\|f\|_{\ms M_{+2}}^n$.
\el
\proof
Since $\xi_t(f)-\xi_s(f)$ is a Gaussian process and $\Ebb_\G[e^{iT(\xi_t(f)-\xi_s(f))}]=e^{-\half C T^2}$, where
$C=(\widehat f/\sqrt\omega, (1-e^{-|t-s|\omega})\widehat f/\sqrt\omega)$, by taking derivatives $2n$ times at $t=0$
on both sides we obtain
\begin{align*}
\Ebb_\G[|\xi_t(f)-\xi_s(f)|^{2n}]&\leq
D_n (\widehat f/\sqrt\omega, (1-e^{-|t-s|\omega}) \widehat f/\sqrt\omega)^{n}\\
&\leq
D_n |t-s|^n \|\widehat f\| ^n_{L^2}\\
&\leq D_n' |t-s|^n \|f\| ^n_{\ms M_{+2}}.
\end{align*}
Here we used that the embedding $i:\ms M_{+2}\to\LR$ is bounded \cite[p288]{LHB}.
\qed
We denote the continuous version of $(Z_t)_{t\geq 0}$ by $(\bar Z_t)_{t\geq 0}$, and the set of $\hir$-valued continuous
paths by $\XQ=C([0,\infty), \hir)$. Note that $(\bar Z_t)_{t\geq 0}$ is a stochastic process on the probability space
$((\hir)^{[0,\infty)}, \s(\A), \M )$, and the map
$$
\bar Z_\cdot:((\hir)^{[0,\infty)}, \s(\A), \M )\to (\XQ, \B^+ )
$$
is measurable, where $\B ^+$ denotes the $\s$-field generated by cylinder sets. This map induces the image measure $\P_+=\M \circ
\bar Z_\cdot \f$ on $(\XQ, \B ^+)$. Thus for the coordinate process $(X^+_t)_{t\geq0}$ on $(\XQ, \B^+ ,\P_+)$ we have
$\bar Z_t\stackrel{\rm d}{=}X^+_t$ for $t\geq0$. Let $(y,\xi)\in\hir $, and define the regular conditional probability measure
$$
\P_+^{(y,\xi)}(\cdot)=\P_+(\cdot \;| X^+_{0}=(y,\xi))
$$
on
$(\XQ, \B^+)$. Since the distribution of $X^+_{0}$ is $d{\rm M}$, we see that
$$
\P_+ (A)=\int_{\hir }\!\!\!\Ebb_{\P_+} ^{(y,\xi)}[\one _{A}] d{\rm M}(y,\xi).
$$
Using the measure $\P_+$ we have
\eq{pos}
(f, e^{-tL_{\rm N}}g)_{\K }=
\int _\hir \Ebb_{\P_+}^{(y,\xi)}[\bar f(X^+_0) g(X^+_t)]
d{\rm M}(y,\xi),
\en
which implies that
\eq{semigroup3}
\lk e^{-tL_{\rm N}}g\rk (y,\xi)=
\Ebb_{\P_+}^{(y,\xi)} [g(X^+_t)].
\en
\begin{lemma}
\label{markov}
The random process $(X^+_{t})_{t\geqslant0 }$ on $(\XQ,\B^+ ,\P_+^{(y,\xi)})$ has the Markov property with respect to
the natural filtration $\sigma(X^+_{t},\,0\leq s\leq t)$.
\end{lemma}
\proof
In this proof we set $z=(y,\xi), z_j=(y_j,\xi_j)\in \hir $ for notational simplicity.
Let
\begin{align}
p_{t}(z, A)=(e^{-t\LN }\one_A)(z), \quad A\in
\B(\BR)\times\B (Q).
\end{align}
Notice that
$$
p_{t}(z, A)=\Ebb_{\P_+}^{z}[\one _{A}(X^+_{t})]=\Ebb_{\P_+ }[\one _{A}(X^+_{t})| X^+_{0}=z].
$$
We show that $p_{t}(z, A)$ is a probability transition kernel, i.e.,
\begin{enumerate}
\item
$p_{t}(z, \cdot)$ is a probability measure on $\B (Q)$
\item
the function $z\mapsto p_{t}(z, A)$ is Borel measurable
\item
the Chapman-Kolmogorov identity
\eq{ck}
\int_\hir p_t(z,A) p_s(z_1,dz)=p_{s+t}(x_1,A)
\en
is satisfied.
\end{enumerate}
Note first that by \kak{pos} it is easy to see that $e^{-tL_{\rm N}}$ is positivity improving.
For every function $f\in\hhh $ such that $0\leq f\leq\one$, we have
$$
\lk e^{-t\LN }f\rk (z)=\Ebb_{\P_+}^{z}[f(X^+_{t})]\leq
\Ebb_{\P_+}^{z}[\one]=1.
$$
Then we deduce that $0\leq e^{t\LN }f\leq\one$ and $e^{t\LN }\one=\one$, and (1)-(2) follow.
We can also show that the finite dimensional distribution is given by
\begin{align*}
\Ebb_{\P_+}\bigg[\prod_{j=1}^{n}\one_{A_{j}}(X^+_{t_{j}})\bigg] =\int_{(\hir)^{n}}
\prod_{j=1}^{n}\one_{A_{j}}(z_j)\prod_{j=1}^{n}p_{t_{j}-t_{j-1}}(z_{j-1}, dz_j).
\end{align*}
Thus $(X^+_{t})_{t\geq0}$ is a Markov process by \cite[Proposition 2.17]{LHB}.
\qed
Now we extend $(X^+_t)_{t\geq 0}$ to a random process indexed by the full real line $\RR.$ Consider the product
probability space $(\widehat {\XQ },\widehat \B^+ ,\widehat \P_+^{(y,\xi)})$ with $\widehat {\XQ }=\XQ \times\XQ $,
$\widehat \B^+ =\B^+ \otimes \B^+ $ and $\widehat \P_+^{(y,\xi)}=\P_+^{(y,\xi)}\otimes\P_+^{(y,\xi)}$, and let
$\pro {\widehat X}$ be the stochastic process
$$
\widehat {X}_t(\omega)= \left\{\begin{array}{rl}
X_t^+(\omega_1) \quad t\geq 0,\\
X_{-t}^+(\omega_2) \quad t\leq0,
\end{array}
\right.\\
$$
on the product space, for $\omega=(\omega_1,\omega_2)\in\widehat {\XQ }$.
\begin{lemma}
\label{hiro}
It follows that
\begin{enumerate}
\item[\rm (1)]
${\widehat X_0}={(y,\xi)}$ a.s.
\item[\rm (2)]
${\widehat X_t},t\geq 0$ and ${\widehat X_s}, s<0$ are independent
\item[\rm (3)]
${\widehat X_t}\stackrel{\rm d}{=} {\widehat X_{-t}}$ for all $t\in\RR$,
\item[\rm (4)]
$({\widehat X}_t)_{t\geq 0}$ (resp. $({\widehat X}_t)_{t\leq0}$) is a Markov process with respect to
$\s( \widehat X_s,0\leq s\leq t)$ (resp. $\s( \widehat X_s,t\leq s\leq 0)$)
\item[\rm (5)]
for $f_0,...,f_n\in \K$ and $-t=t_{0}\leq t_{1}\leq\ldots\leq t_{n}=t$, we have
\eq{p3}
\Ebb_{\widehat {\P}_+}\bigg[ \prod_{j=0}^{n}f_{j}(\widehat X_{t_{j}})\bigg] =\lk f_{0},
e^{-(t_{1}+t)\LN }f_{1}...e^{-(t-t_{n-1})\LN }f_{n}\rk _{\K }.
\en
\end{enumerate}
\end{lemma}
\proof
(1)-(3) are straightforward.
(4) follows from Lemma \ref{markov}.
(5) follows from \kak{p2} and a simple limiting argument.
\qed
\medskip
\noindent
{\it Proof of Theorem \ref{pp1}.} We show that the stochastic process $\pro {\widehat X}$ defined on
$(\widehat {\XQ}, \widehat {\B }, \widehat \P_+^{(y,\xi)})$ is a $P(\phi)_1$-process associated with
$((\hir, \Sigma', {\rm M}), L_{\rm N})$. The Markov property, reflection symmetry and the shift invariance
property follow from Lemma \ref{hiro}. Continuity of $t\mapsto \widehat X_t$ has been shown in Lemma
\ref{cont}. Thus $\widehat X_\cdot: (\widehat {\XQ}, \widehat {\B }, \widehat \P_+^{(y,\xi)})\to
(\XXQ, \B_Q)$ is measurable and the image measure $\P^{(y,\xi)}=\widehat \P_+^{(y,\xi)}\circ \widehat X_\cdot\f$
defines a probability measure on $(\XXQ, \B_Q)$. Hence the coordinate process $\pro X$ on $(\XXQ, \B_Q,\P^{(y,\xi)})$
satisfies $X_t\stackrel{\rm d}{=}\widehat X_t$, and then $\pro X$ is a $P(\phi)_1$-process associated with
$((\hir, \Sigma', d{\rm M}_0), L_{\rm N})$.
\qed
\begin{lemma}
\label{rev1}
The random process $( X_{t})_{t\in\RR }$ is a reversible Markov process under $\P $, and its stationary measure
is ${\rm M}$, i.e., for every $n \geq 1$ we have that $( X_{t_{1}}, X_{t_{2}},..., X_{t_{n}})$ has the same
distribution as $( X_{\tau-t_{1}}, X_{\tau-t_{2}},..., X_{\tau-t_{n}})$ for all $t_{1},...,t_{n},\tau\in\mathbb{R}$.
\end{lemma}
\proof
Let $f,g\in\K $. Then
$\Ebb_{\P }[f( X_{t})g( X_{s})]=\lk f,e^{-|t-s|\LN }g\rk _{\K }$.
Thus the lemma follows.
\qed
\subsection{Relativistic Nelson model}
\label{bbb}
Define a probability measure on $\hir$ by
$$
d\tilde{\rm M}=\ggr ^{2}d\tilde {\rm P}.
$$
The unitary operator $ \widetilde{U}_{\rm g} :L^{2}(\hir, d\tilde{\rm M})\rightarrow L^{2}(\hir , d\tilde {\rm P})$
is defined by $ \Phi \mapsto \ggr \Phi$. Let $\tilde{\K} =L^{2}(\hir, d\tilde{\rm M})$. Define the operator
$$
\tilde{L}_{\rm N}=\frac{1}{\ggr }(\tilde H -\tilde E)\ggr.
$$
Let $\DD=D(\RR,\hir)$ be the space of c\`adl\`ag paths with values in $\hir$ on the whole real line, and $\tilde \B_Q$ the
$\sigma$-field generated by cylinder sets. Similarly to the classical Nelson model, we can construct a $P(\phi)_1$ process
for the relativistic Nelson Hamiltonian.
\bt{pp11}
\TTT{$P(\phi)_1$-process for the relativistic Nelson Hamiltonian}
Let $(y,\xi)\in \hir$.
Then ,
\begin{enumerate}
\item[\rm 1.]
There exists a probability measure
$\tilde{\P}^{(y,\xi)}$ on
$(\tilde {\X}_Q, \tilde{\B}_Q)$ such that
the coordinate process $\pro \oo$ on
$(\tilde {\X}_Q, \tilde{\B}_Q,\tilde{\P}^{(y,\xi)})$ is
$P(\phi)_1$-process
associated with the pair\\
$\bigg((\hir, \B (\BR)\otimes \B (Q), d\tilde{\rm M}), \tilde{L}_{\rm N}\bigg ).$
\item[\rm 2.]
The function $t\mapsto \oo_t$ is c\`adl\`ag a.s.
\end{enumerate}
\et
The proof of Theorem \ref{pp11} is parallel with that of Theorem \ref{pp1} except for path regularity. We only discuss this
part of the proof. Define the family of set functions $\{\tilde{\M} _{\Lambda} \,|\, \Lambda\subset\RR, \#\Lambda
<\infty\}$ on ${\Sigma'} ^{\#\Lambda}=\underbrace{\Sigma'\times\cdots\times \Sigma'}_{\#\Lambda}$ by
\begin{align*}
\tilde{\M}_{\Lambda}(A_{0}\times A_{1}\times...\times A_{n}) =\lk \one _{A_{0}}, e^{-(t_{1}-t_{0})\tilde {L}_N}
\one _{A_{1}}e^{-(t_{2}-t_{1})\tilde{L}_{\rm N}}...\one _{A_{n-1}}e^{-(t_{n}-t_{n-1})\tilde{L}_{\rm N}}\one _{A_{n}} \rk_{\tilde{\K} }
\end{align*}
for $\Lambda=\{t_0,\ldots,t_n\}$. By the same way as for the classical Nelson Hamiltonian, we define the projection
$\pi_{\Lambda}: (\hir)^{[0,\infty)}\longrightarrow (\hir)^{\Lambda}$ by $w \longmapsto (w(t_{0}),...,w(t_{n}))$
for $\Lambda=\{t_{0},...,t_{n}\}$, and $\ms A = \{\pi_{\Lambda}\f(A) \,|\, A\in {\Sigma' }^{\#\Lambda},\, \#\Lambda<\infty \}
$ is a finitely additive family of sets. Using the Kolmogorov extension theorem, there exists a unique probability measure
$\tilde{\M}$ on $((\hir)^{[0,\infty)}, \s(\ms A))$
such that
\begin{align}\label{measure}
\Ebb_{\tilde{\M}}\bigg[\prod_{j=0}^n \one_
{A_{t_j}}(\tilde{Z}_{t_j})\bigg]&=\tilde{\M}_{\{t_0,\cdots,t_n\}}(A_0\times\cdots\times A_n)\nonumber\\
&=
\lk \one _{A_{0}}, e^{-(t_{1}-t_{0})\tilde{L}_{\rm N}}\one _{A_{1}}e^{-(t_{2}-t_{1})\tilde{L}_{\rm N}}\one _{A_{1}}...
\one _{A_{n-1}}e^{-(t_{n}-t_{n-1})\tilde{L}_{\rm N}}\one _{A_{n}} \rk_{\tilde{\K} },
\end{align}
where $(\tilde {Z_t})_{t\geqslant0}$ is the coordinate process on $((\hir)^{[0,\infty)}, \s(\ms A),\tilde{\M})$.
The equality \kak{measure} leads to the following result:
\begin{lemma}\label{shift2}
The stochastic process $(\tilde{Z}_t)_{t\geqslant0}$ is shift invariant under $\tilde{\M}$, i.e, for $f_0,...,f_n
\in \tilde{\K} $ and $s\geqslant0$ it follows that
\begin{align}\label{shi2}
\Ebb_{\tilde{\M}}\bigg[ \prod_{j=0}^{n}f_{j}(\tilde{Z}_{t_{j}+s})\bigg] =\Ebb_{\tilde{\M}}
\bigg[ \prod_{j=0}^{n}f_{j}(\tilde{Z}_{t_{j}})\bigg] =\lk f_{0},e^{-(t_{1}+t)
\tilde{L}_{\rm N}}f_{1}\cdots e^{-(t-t_{n-1})\tilde{L}_{\rm N}}f_{n}\rk _{\tilde{\K} }.
\end{align}
\end{lemma}
Now we prove that $(\tilde{Z}_t)_{t\geqslant0}$ has c\`adl\`ag version under $\tilde{\M}$.
For this purpose, we need the following technical lemmas which make use of the ideas in \cite[pp 59-62]{Sato}.
Let $I\subset[0, \infty)$ and $\varepsilon>0$. We say that $\tilde{Z}_\cdot(\omega)$, with $\omega$ fixed,
has $\varepsilon$-oscillation $n$ times in $I$, if there exist $t_0,t_1,\ldots,t_n\in I $ such that $t_0<t_1<t_2\ldots<t_n$
and $\| \tilde{Z}_{t_{j}}-\tilde{Z}_{t_{j-1}}\|_{\hir}>\varepsilon$ for $j=1,\ldots,n$.
We say that $\tilde{Z}_\cdot(\omega)$ has $\varepsilon$-oscillation infinitely often in $I$, if, for every $n$,
$\tilde{Z}_\cdot(\omega)$ has $\varepsilon$-oscillation $n$ times in $I$. Let
\begin{align*}
&\Omega'=\lkk \omega \in \Omega \,|\, \lim_{s\in {\Bbb Q}, s \downarrow t} \tilde{Z}_s(\omega)
\mbox{ and } \lim_{s\in {\Bbb Q}, s \uparrow t} \tilde{Z}_s(\omega) \mbox{ exist in } \hir \mbox{ for all } t\geq 0 \rkk,\\
&
A_{N,k}=\lkk \omega \in\Omega \,|\, \tilde{Z}_t(\omega) \mbox{ does not have } 1/k-\mbox{oscillation infinitely often in }
[0, N] \cap {\Bbb Q} \rkk,\\
&\Omega''=\bigcap_{N=1}^{\infty}\bigcap_{k=1}^{\infty}A_{N,k}.
\end{align*}
Similarly as in \cite[Lemma 11.2]{Sato}, we can see that $\Omega''\subset \Omega'$. Define
$$
B(p,\varepsilon, I)=\lkk \omega\in \hirr \,|\, \tilde{Z}_t(\omega) \mbox{ has } \varepsilon-\mbox{oscillation }
p \mbox{ times in } I \rkk.
$$
\begin{lemma}\label{11.2}
For every $\varepsilon>0$ we have $\underset{|t-s|\rightarrow 0} \lim \tilde{\M} \lk \| \tilde{Z}_t- \tilde{ Z}_s \|_{\hir}
>\varepsilon \rk=0$ and $ \tilde{\M}(\Omega'')=~1$.
\end{lemma}
\proof
In this proof we set
$$
\KKK(y,\xi)=\tgrp(y)\tgr(y,\xi)
$$
and
$\Ebb[\cdots]=\int_{\hir }dy\otimes d{{\rm G}} \Ebb_{\mathscr P\times \G}^{(y,\xi)}[\cdots]$ for notational simplicity.
Consider
\begin{align*}
\TT&=
\underset{(y,\xi)\in\hir } \sup(\mathscr P^y\otimes \G^\xi)\lk \tilde{Z}_{t-s}\in B^c((y,\xi),\varepsilon)\rk \\
&=
(\mathscr P\otimes \G)\lk \tilde{Z}_{t-s}\in B^c(0,\varepsilon)\rk = (\mathscr P\otimes \G)\lk |\tilde{Z}_{t-s}|>\varepsilon\rk.
\end{align*}
By Proposition \ref{fkl} we have
$$
\tilde{\M}\lk \| \tilde{ Z}_t- \tilde {Z}_s\|_{\hir}>\varepsilon \rk =(\one, e^{-(t-s)\tilde{L}_{\rm N}}\one_{B^c(0,\eps)})_{\tilde{\K}},
$$
where $B^c((y',\xi'),\eps)=\{(y,\xi)\in\hir \,|\, \|(y,\xi)-(y',\xi')\|_{\hir} \geq\eps\}$. Also, we have
$$
\underset{(y,\xi)\in\hir } \sup|\tgr (y,\xi)|=\sqrt K_2<\infty \quad \mbox{and} \quad
\underset{y\in\BR } \sup \Ebb_{\mathscr P}^{y}\left[e^{-4\int_0^{t-s} V(\xx_r) dr} \right]\leq C_V.
$$
We have
\begin{align*}
& \tilde{\M}\lk \| \tilde{Z}_t- \tilde {Z}_s\|_{\hir}>\varepsilon \rk\\
&=\Ebb
\left[ \KKK(y,\xi) e^{-\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr} e^{-\int_0^{t-s} V(\xx_r) dr}
\KKK(\xx_{t-s},\xi_{t-s})
\one_{B^c((y,\xi),\eps)} (b_{t-s},\xi_{t-s}) \right] e^{(t-s)\tilde E}.
\end{align*}
The right-hand side can be evaluated as
\begin{align*}
& \tilde{\M}\lk \| \tilde{ Z}_t- \tilde{ Z}_s\|_{\hir}>\varepsilon \rk\\
&\leq K_2 \Ebb
\left[ \tgrp(y) e^{-\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr} e^{-\int_0^{t-s} V(\xx_r) dr}
\tgrp(\xx_{t-s})
\one_{B^c((y,\xi),\eps)} (b_{t-s},\xi_{t-s}) \right] e^{(t-s)\tilde E}.
\end{align*}
By using Schwarz inequality twice, we have
\begin{align*}
& \tilde{\M}\lk \| \tilde{Z}_t- \tilde{ Z}_s\|_{\hir}>\varepsilon \rk\\
&\leq K_2 \|\tgrp\|
\Ebb \left[ | \tgrp(y)|^2 e^{-2\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr} e^{-2\int_0^{t-s} V(\xx_r) dr}
\one_{B^c((y,\xi),\eps)} (b_{t-s},\xi_{t-s}) \right]^{\frac{1}{2}} e^{(t-s)\tilde E} \\
&\leq K_2 \|\tgrp\|
\Ebb\left[ | \tgrp(y)|^2\one_{B^c((y,\xi),\eps)} (b_{t-s},\xi_{t-s})\right]^{\frac{1}{4}} \\
&\hspace{3cm}\times
\Ebb\left[ | \tgrp(y)|^2 e^{-4\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr}
e^{-4\int_0^{t-s} V(\xx_r) dr}\right]^{\frac{1}{4}} e^{(t-s)\tilde E}.
\end{align*}
Since
$$\Ebb\left[ | \tgrp(y)|^2\one_{B^c((y,\xi),\eps)} (b_{t-s},\xi_{t-s})\right]^{\frac{1}{4}}\leq
J_\varepsilon^{1/4}\|\tgrp\|^{1/2}$$
and
\begin{align*}
&\Ebb\left[ | \tgrp(y)|^2 e^{-4\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr} e^{-4\int_0^{t-s} V(\xx_r) dr}\right] \\
&=
\int_{\hir} \Ebb_\mathscr P^y\left[|\tgrp(y)|^2 e^{-4\int_0^{t-s} V(\xx_r) dr}\right] \Ebb^\xi_{\G}\left[e^{-4\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr}
\right]dy\otimes d{\rm G}.
\end{align*}
By \cite[Section 6.5]{LHB} we have
\begin{align*}
\Ebb^\xi_{\G}\left[e^{-4\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr}\right]\leq e^{(t-s)\left(\int_\BR\frac{|\hat{\varphi}(k)|^2}
{\omega(k)^2}dk+\int_\BR\frac{|\hat{\varphi}(k)|^2}{\omega(k)^3}dk\right) }=C_{t-s}<\infty.
\end{align*}
We deduce that
\begin {align*}
&\Ebb\left[ | \tgrp(y)|^2 e^{-4\int_0^{t-s} \tau_{\xx_r}\xi_r(\vpp)dr} e^{-4\int_0^{t-s} V(\xx_r) dr}\right] \\
&\quad \leq
C_{t-s}\int_{\hir} |\tgrp(y)|^2 \Ebb_\mathscr P^y\left[e^{-4\int_0^{t-s} V(\xx_r) dr}\right] dy\otimes d{\rm G} \leq
C_{t-s}C_V \|\tgrp\|^2,
\end{align*}
and so we obtain
\begin{align}
\label{darui}
\tilde{\M}\lk \| \tilde{Z}_t- \tilde{Z}_s\|_{\hir}>\varepsilon \rk\leq K_2 C_V^{1/4} \|\tgrp\|^2 C_{t-s}
J_\varepsilon^{1/4}e^{(t-s)\tilde E}.
\end{align}
Next we show that $\underset{|t-s|\rightarrow 0}\lim\TT =0.$ In fact, we have
\begin{align*}
\TT
=(\mathscr P\otimes\G) \lk \{
\omega=(\omega^1,\omega^2)\in\hirr |
\| \xx_{t-s} (\omega^1)\| _\BR ^2+ \| \xi_{t-s}(\omega^2) \| _Q^2>{\varepsilon}^2 \} \rk .
\end{align*}
By the stochastic continuity of the L\'evy process $(\xx_t)_{t\geq 0}$ and the OU-process $(\xi_t)_{t\geq 0}$ we deduce that
$\underset {|t-s|\rightarrow 0}\lim \TT =0$.
Then
$ \underset{|t-s|\rightarrow 0} \lim \tilde{\M}\lk\| \tilde{ Z}_t- \tilde{Z}_s \| _{\hir }>\varepsilon \rk =0$ follows.
To see that $\tilde{\M}(\Omega'')$=1, it suffices to show that $\tilde{\M}(A_{N,k}^c)=0$ for any fixed $N$ and $k$. We have
\begin{align*}
&\tilde{\M}(A_{N,k}^c)
=\tilde{\M}\lk\lkk \tilde{Z}_t \mbox{ has }1/k-\mbox{oscillation infinitely often in } \big [0, N\big] \cap {\Bbb Q} \rkk \rk \\
& \leq \sum_{j=1}^{l} \tilde{\M} \lk \lkk \tilde{Z}_t\mbox{ has
}1/k-\mbox{oscillation infinitely often in }
\bigg [\frac{j-1}{l} N, \frac{j}{l}N\bigg] \cap {\Bbb Q} \rkk \rk \\
&=\sum_{j=1}^{l}\underset{p\rightarrow \infty}\lim \tilde{\M} \lk B
\lk p, \frac{1}{k},\bigg [\frac{j-1}{l} N, \frac{j}{l}N\bigg] \cap {\Bbb Q}\rk
\rk .
\end{align*}
We enumerate as $\{t_1,\ldots,t_n,\ldots\} = [\frac{j-1}{l} N, \frac{j}{l}N] \cap {\Bbb Q} $.
Thus
$$
\tilde{\M} \lk B
\lk p, \frac{1}{k},\bigg [\frac{j-1}{l} N, \frac{j}{l}N\bigg] \cap {\Bbb Q}\rk
\rk
=\lim_{n\to\infty}\tilde{\M} \lk B \lk p, \frac{1}{k}, \{t_1,\cdots,t_n\}\rk \rk .$$
Then by Proposition \ref{fkl} we obtain
\begin{align*}
&\tilde{\M} \lk B \lk p, 1/k, \{t_1,\ldots,t_n\}\rk \rk\\
&= e^{N\tilde E/l}
\Ebb \left[ \KKK (y,\xi) e^{-\int_0^{N/l} \tau_{\xx_r}\xi_r(\vpp)dr}
\KKK(\xx_{N/l},\xi_{N/l}) e^{-\int_0^{N/l} V(\xx_r) dr}
\one_{B(p, 1/k, \{t_1,\ldots,t_n\})}(\xx_{N/l},\xi_{N/l})\right].
\end{align*}
Hence in the same estimate preceding \kak{darui} we have
\begin{align}
\label{darui2}
&\tilde{\M}\lk B \lk p, \frac{1}{k}, \{t_1,\ldots,t_n\}\rk \rk\non\\
&\leq K_2
C_V^{1/4} \|\tgrp\|^2 e^{\half C_{N/l}}
\left(\underset{(y,\xi)\in\hir}\sup\mathscr P^y\otimes\G^\xi B\bigg(p, \frac{1}{k}, \{t_1,\ldots,t_n\}\bigg)\right)^{\frac{1}{4}}
e^{(N/l) \tilde E}.
\end{align}
By \cite[Lemma 11.4]{Sato}, furthermore we have
\eq{kore}
(\mathscr P^y\otimes\G^\xi)
\lk
B\bigg(p, \frac{1}{k}, \{t_1,\ldots,t_n\}\bigg)\rk
\leq
\lk
\sup_{s,t\in[0,N]\atop
t-s\in[0,N/l]}(\mathscr P\otimes\G)\left (|(b_s,\xi_s)-(b_t,\xi_t)|\geq\frac{1}{4k} \right)\rk ^p.
\en
Moreover, by stochastic continuity of $((\xx_t,\xi_t))_{t\geq 0}$,
we can prove uniform stochastic continuity, i.e.,
\eq{us}
\sup_{s,t\in[0,N]\atop
t-s\in[0,N/l]}(\mathscr P\otimes\G)\left (|(b_s,\xi_s)-(b_t,\xi_t)|\geq\frac{1}{4k} \right)\to 0
\en
as
$l\to\infty$ in Lemma \ref{satobook} below.
\qed
\begin{lemma}\label{satobook}
\kak{us} holds.
\end{lemma}
\proof
For notational simplicity we write $X'_s=(b_s,\xi_s)$.
Fix $a>0$. For any $t$ there exists $\delta_t>0$ such that
$(\mathscr P\otimes\G)(|X'_t-X'_s|\geq \varepsilon/2) \leq a/2$
for $|t-s|<\delta_t$ by stochastic continuity. Let $I_t=(t-\delta_t/2,t+\delta_t/2)$.
Since $I_t$ is compact, there exists a finite covering $I_{t_j}$, $j=1,\ldots,n$, such that
$\cup_{j=1}^n I_{t_j}\supset[0,N]$. Let $\delta=\min_{j=1,...,n}\delta_{t_j}$.
If $|s-t|<\delta$ and $s,t\in [0,N]$, then $t\in I_{t_j}$ for some $j$, hence $|s-t_j|<\delta_{t_j}$ and
$$
(\mathscr P\otimes\G) (|X'_t-X'_s|\geq \varepsilon)
\leq (\mathscr P\otimes\G) (|X'_t-X'_{t_j}|\geq \varepsilon)
+(\mathscr P\otimes\G)(|X'_{t_j}-X'_s|\geq \varepsilon) <a.
$$
Hence the lemma follows.
\qed
\begin{lemma}\label{cadlag}
The process $(\tilde{Z}_t)_{t\geq 0}$ has
a right continuous version with left limits (c\`adl\`ag)
with respect to $\tilde{\M}$.
\end{lemma}
\proof
Let $(\tilde{Z}')_{t\geqslant0}$ be a c\`adl\`ag process defined by
\begin{align}
\label{mod}
\tilde{Z}'_t(\omega)= \left\{\begin{array}{ll}
\underset {s\in {\Bbb Q}, s \downarrow t}\lim \tilde{ Z}_s(\omega) \quad& \omega\in\Omega'',\\
0 \quad& \omega\notin\Omega''.
\end{array}
\right.
\end{align}
By Lemmas \ref{11.2} the process $(\tilde{Z}_t)_{t\geq 0}$ is stochastically continuous, which implies that
there exists
a sequence $s_n$ such that
\begin{align}\label{mod2}
\underset {s_n\in {\Bbb Q}, s_n \downarrow t}\lim \tilde{Z}_{s_n}(\omega) =\tilde{Z}_t(\omega)
\end{align}
for $\omega \in \Omega '''=(\BR\times Q)^{[0,\infty)}\setminus N_t$ with some null set $N_t$.
We can also see by the definition of the process $(\tilde{Z}')_{t\geqslant0}$ that
\begin{align}\label{mod3}
\underset {s_n\in{\Bbb Q}, s_n \downarrow t}\lim \tilde{Z}_{s_n}(\omega) =\tilde{Z}'_t(\omega)
\end{align}
for $\omega \in \Omega ''$, and
$\tilde{\M}(\Omega'')=1$ by Lemma \ref{11.2}.
For each $t$ by \kak{mod2} and \kak{mod3} we can derive that
$\tilde{Z}_t(\omega) =\tilde{Z}'_t(\omega) $ for $\omega \in \Omega''\cap \Omega '''$, and
$\tilde {\M}(\Omega''\cap \Omega ''')=1$.
Then
$(\tilde{Z}')_{t\geqslant0}$ is a c\`adl\`ag version of
$(\tilde{Z})_{t\geqslant0}$.
\qed
We denote the c\`adl\`ag version of $(\tilde Z_t)_{t\geq 0}$ by $(\bar{\tilde {Z_t}})_{t\geq 0}$, and the set of $\hir$-valued c\`adl\`ag paths by
$\XQC =D([0,\infty), \hir)$. Note that $(\bar{ \tilde{Z_t}})_{t\geq 0}$ is a stochastic process on the probability space
$((\hir)^{[0,\infty)}, \s(\A), \tilde{\M} )$, and the map
$$
\bar{\tilde{ Z_\cdot}}:((\hir)^{[0,\infty)}, \s(\A), \tilde{\M} )\to (\XQC , \tilde{\B} ^+ )
$$
is measurable, where $\tilde{\B} ^+$ denotes the $\s$-field generated by cylinder sets. This map induces the image
measure $\PC _+=\tilde{\M} \circ {\bar{\tilde {Z_\cdot}}} \f$ on $(\XQC , \tilde{\B} ^+)$. Then the coordinate process
$({\tilde X^+}_t)_{t\geq0}$ on $(\XQC , \tilde{\B} ^+ ,\PC _+)$ satisfies that $\bar{ \tilde{Z_t}}\stackrel{\rm d}{=}
\tilde X^+_t$ for $t\geq0$. Let $(y,\xi)\in\hir $ and define the regular conditional probability measure on
$(\XQC , \tilde{\B} ^+)$ by $\PC _+^{(y,\xi)}(\cdot)=\PC _+(\cdot \,|\, \tilde X^+_{0}=(y,\xi))$.
\begin{lemma}
\label{markovc}
The process $(\tilde X^+_{t})_{t\geqslant0 }$ is a Markov process on $(\XQC ,\tilde{\B} ^+ ,\PC _+^{(y,\xi)})$ with
respect to the natural filtration $\sigma(\tilde X^+_{t},\,0\leq s\leq t)$.
\end{lemma}
\proof
The proof is the same as that of Lemma \ref{markov}.
\qed
We extend $(\tilde X^+_t)_{t\geq 0}$ to a Markov process to the whole real line $\RR$. This can be done in the same
way as the extension of $(X^+_t)_{t\geq 0}$ to a process on the whole real line as seen in the case of the classical
Nelson Hamiltonian in the previous section. Consider the product probability space $(\XQCH ,\widehat {\tilde{\B}} ^+ ,
\widehat \PC _+^{(y,\xi)})$ with $\XQCH =\XQC \times\XQC$, $\widehat {\tilde{\B}}^+ =
\tilde{\B} ^+ \otimes \tilde{\B} ^+ $, and $\widehat \PC _+^{(y,\xi)}=\PC _+^{(y,\xi)}\otimes\PC _+^{(y,\xi)}$. Let
$\pro {\widehat{\tilde X}}$ be a stochastic process on the product space, defined by $\widehat{\tilde X}_t(\omega)=
\tilde X_t^+(\omega_1)$ for $t\geq 0$, and $\widehat{\tilde X}_t(\omega)=\tilde X_{-t}^+(\omega_2)$ for $t\leq 0$, with $\omega=
(\omega_1,\omega_2)\in\XQCH $.
\begin{lemma}
\label{hiroc}
It follows that
\begin{enumerate}
\item[\rm 1.]
${\widehat{\tilde X}_0}={(y,\xi)}$ a.s.
\item[\rm 2.]
${\widehat{\tilde X}_t},t\geq 0$ and ${\widehat{\tilde X}_s}, s<0$ are independent
\item[\rm 3.]
${\widehat{\tilde X}_t}\stackrel{\rm d}{=}{\widehat{\tilde X}_{-t}}$ for all $t\in\RR$
\item[\rm 4.]
$({\widehat{\tilde X}}_t)_{t\geq 0}$ (resp. $({\widehat{\tilde X}}_t)_{t\leq0}$)
is a Markov process with respect to $\s( \widehat{\tilde X}_s,0\leq s\leq t)$ (resp.
$\s( \widehat{\tilde X}_s,t\leq s\leq 0)$)
\item[\rm 5.]
for $f_0,\ldots,f_n\in \tilde{\K}$ and $-t=t_{0}\leq t_{1}\leq\ldots\leq t_{n}=t$, we have
\eq{p33}
\Ebb_{\widehat {\PC }_+}\left[ \prod_{j=0}^{n}f_{j}(\widehat{\tilde X}_{t_{j}})\right] =\lk f_{0},
e^{-(t_{1}+t)\tilde{L}_{\rm N} }f_{1}...e^{-(t-t_{n-1})\tilde{L}_{\rm N} }f_{n}\rk _{\tilde{\K}}.
\en
\end{enumerate}
\end{lemma}
\proof
The proof is similar to the proof of Lemma \ref{hiro}
\qed
\medskip
\noindent
{\it Proof of Theorem \ref{pp11}:} We show that the random process $\pro {\widehat {\tilde X}}$ defined on
$({\XQCH}, \widehat {\tilde \B}^+, \widehat{\PC}_+^{(y,\xi)})$ is a $P(\phi)_1$-process associated with
$((\hir, \Sigma', \tilde {\rm M}), \tilde{L}_{\rm N})$. The Markov property, reflection symmetry and the shift invariance
follow from Lemma \ref{hiroc}.
The c\`adl\`ag property of the path $t\mapsto \widehat X_t$, $t\geq0$, (resp. c\`agl\`ad property of $t\mapsto \widehat X_t$,
$t\leq 0$) was shown in Lemma~\ref{cadlag}. Thus the map $\widehat{\tilde X}_\cdot: (\XQCH, \widehat {\tilde {\B }}^+,
\widehat {\PC}_+^{(y,\xi)})\to (\D, \tilde {\B}_Q)$ is measurable and the image measure $\PC^{(y,\xi)}=\widehat
{\PC}_+^{(y,\xi)}\circ \widehat{\tilde X}_\cdot\f$ defines a probability measure on $(\D, \tilde {\B}_Q)$. Hence
the coordinate process $\pro {\tilde X}$ on $(\D, \tilde{\B}_Q, \PC^{(y,\xi)})$ satisfies $\tilde X_t\stackrel{\rm d}{=}
\widehat {\tilde X}_t$, is a $P(\phi)_1$-process associated with the pair\\
$\bigg((\hir, \B(\BR)\otimes\B(Q), d\tilde {\rm M}_0), \tilde{L}_{\rm N}\bigg)$.
\qed
\section{Functional central limit theorems}
\label{shikaku}
\subsection{Classical Nelson model}
Next we discuss FCLT related to the classical and relativistic Nelson models, starting with the classical case. Let
\eq{mt}
M_{t}=f(X_{t})-f(X_{0})+\int_{0}^{t}\LN f(X_{s})ds,\quad {t\geqslant0},
\en
where $f\in D(\LN )\subset\K $.
\begin{lemma}\label{***}
$(M_t)_{t\geqslant0 }$ is a martingale with stationary increments under $\P$.
\end{lemma}
\proof
By Lemma \ref{markov} $(X_{t})_{t\geqslant0}$ is a Markov process with semigroup $T_{t}=e^{-t\LN }, t\geqslant0$. Using the Markov property,
we have
\begin{align} \label{m1}
\Ebb_{\P }[f(X_{t})|\mathcal{F}_{s}]=T_{t-s}f(X_{s}),\quad 0\leq s\leq t.
\end{align}
Since the function $t\rightarrow T_{t} $ is differentiable, we obtain
\begin{align}\label{ma2}
\frac{d}{dt}T_{t}f=-\LN T_{t} f=-T_{t} \LN f\quad\mbox{and}\quad T_{t}f-f=-\int_{0}^{t}\LN T_{s} f ds, \quad t\geq 0.
\end{align}
Hence
\begin{align}\label{ma1}
\Ebb_{\P }[M_{t}|\mathcal{F}_{s}]=M_{s}+\Ebb_{\P }\left[ f(X_{t})-f(X_{s})+\int_{s}^{t}\LN f(X_{r})dr\left| \mathcal{F}_{s}\right.\right] \quad\mbox{a.s.}
\end{align}
Using \kak{m1}-\kak{ma2} we show that the second term on the right hand side of \kak{ma1} is zero.
Indeed,
\begin{align*}
\lefteqn{
\Ebb_{\P }\left[ f(X_{t})-f(X_{s})+\int_{s}^{t}\LN f(X_{r})dr\left| \mathcal{F}_{s}\right.\right] } \\
&=
T_{t-s}f(X_{s})-f(X_{s})+\int_{s}^{t}\LN T_{r-s} f (X_{s})dr \\
&=
T_{t-s}f(X_{s})-f(X_{s})+\int_{0}^{t-s}\LN T_{r} f (X_{s})dr \\
&=
T_{t-s}f(X_{s})-f(X_{s})-T_{t-s}f(X_{s})+f(X_{s}) =0.
\end{align*}
By the shift invariance of the process $(X_{t})_{t\geqslant0}$, it then follows that $(M_t)_{t\geq0}$ is a martingale
under $\P$, and it has stationary increments.
\qed
We begin by proving a CLT for the process $(M_t)_{t\geqslant0 }$ under $\P$. The fundamental tool will be the
following martingale central limit theorem \cite[Section 5]{hel82}:
\bp{mart}
Let $(N_t)_{t\in\RR }$ be a martingale on a probability space $(\Omega, \F, P)$
with $$\alpha^{2}=\underset{t\rightarrow\infty}\lim\frac{1}{t}\Ebb_{P}[N_{t}^{2}]<\infty,$$
and assume that $(N_t)_{t\in\RR }$ has stationary increments. Then
$$\lim_{s\rightarrow \infty}\frac{1}{\sqrt{s}}N_{[st]}\stackrel {\rm d} = \alpha^{2}B_{t}.$$
\ep
\bl{s}\label{finie}
If $\Ebb_{\P }[f^{2}(X_{t})]<\infty$ and $\Ebb_{\P }[(\LN f)^{2}(X_{t})]<\infty$ for every $t \geq 0$ and $f\neq0$, then
\begin{align}
\label{a1}
&\underset{t\rightarrow\infty}\lim\frac{1}{t}\Ebb_{\P }[M_{t}^{2}]=2\lk f, \LN f \rk _{\K }.\\
\label{a2}
&\lk f, \LN f \rk _{\K }>0.
\end{align}
\el
\proof
We have
\begin{align}
\Ebb_{\P}[M_t^2]&=
\Ebb_{\P}[f^2(X_t)]
+
\Ebb_{\P}[f^{2}(X_0)]
-2 \Ebb_{\P} [f(X_0)f(X_t)]]
+2 \Ebb_{\P }\left[f(X_t) \int_{0}^{t}\!\!\!dr\LN f(X_{r})\right]\non \\
&\label{eqmart}-2 \Ebb_{\P }\left[f(X_0) \int_{0}^{t}\!\!\!dr\LN f(X_{r})\right]+
\Ebb_{\P }\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right].
\end{align}
Consider
$$
\lim_{t\rightarrow \infty} \frac{1}{t}\Ebb_{\P }\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right].
$$
Writing $T_{t}=e^{-t\LN }$, and using the shift invariance and Markov properties of $(X_{t})_{t\geqslant0}$, we obtain
\begin{align}
&
\Ebb_{\P }\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right] \nonumber\\
& = \Ebb_{\P }\left[\int_{0}^{t}\!\!\!ds\int_{0}^{t}\!\!\!dr
\LN f(X_{r})\LN f(X_{s})\right]
= \int_{0}^{t}\!\!\!ds\int_{0}^{t}\!\!\!dr\Ebb_{\P }[\LN f(X_{0})\LN f(X_{|r-s|})] \nonumber\\
&\label{pos1}
= \int_{0}^{t}\!\!\!ds\int_{0}^{t}\!\!\!dr
\Ebb_{\P }\left[\LN f(X_{0})\Ebb_{\P }[ \LN f(X_{|r-s|})\left | \mathcal{F}_{0}\right] \right] =
\int_{0}^{t}\!\!\!ds\int_{0}^{t}\!\!\!dr \lk T_{|s-r|}\LN f, \LN f \rk _{\K }.
\end{align}
Hence
\begin{align}
\label{ach1}
\lim_{t\rightarrow \infty} \frac{1}{t}\Ebb_{\P}
\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right]=
2\lk f, \LN f \rk _{\K}.
\end{align}
By Schwarz inequality, we have
\begin{align*}
\left| \Ebb_{\P }\left[f(X_t) \int_{0}^{t}\!\!\!dr\LN f(X_{r})\right]\right|&\leq
\lk\Ebb_{\P }\left[f^2(X_t)\right] \rk^{1/2}
\lk\Ebb_{\P }\left[\lk\int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right]\rk^{1/2} \\
&=
\lk \Ebb_{\P }\left[f^2(X_0)\right]\rk^{1/2} \lk \Ebb_{\P }\left[\lk\int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right] \rk^{1/2}.
\end{align*}
Thus we obtain
\begin{align}
\label{ach2}
\lim_{t\rightarrow \infty}\frac{1}{t} \Ebb_{\P }\left[f(X_t) \int_{0}^{t}\!\!\!dr\LN f(X_{r})\right]=0.
\end{align}
Moreover, by the same argument, we have
\begin{align}
\label{ach3}
\lim_{t\rightarrow \infty}\frac{1}{t} \Ebb_{\P }\left[f(X_0) \int_{0}^{t}\!\!\!dr\LN f(X_{r})\right]=0.
\end{align}
\iffalse
the Markov property of the process $(X_{t})_{t\in\mathbb{R}}$, we have
\begin{align*}
\frac{d}{dt}T_{t}f(X_r)=-\LN T_{t} f(X_r)=-(\LN f)(X_{t+r}).
\end{align*}
Then,
\begin{align}\label{cor}
f(X_t) \int_{0}^{t}\!\!\!dr\LN f(X_{r})=-f(X_t)\big(f(X_t)-f(X_0)\big)\quad\mbox{and}
\quad f(X_0) \int_{0}^{t}\!\!\!dr\LN f(X_{r})=-f(X_0)\big(f(X_t)-f(X_0)\big).
\end{align}
Then, by \kak{cor} and the shift invariance properties of the process $(X_{t})_{t\in\mathbb{R}}$, we have
\begin{align*}
\Ebb_{\P}[M_{t}^2]&=-\Ebb_{\P}[f^2(X_0)]
-\Ebb_{\P}[f^{2}(X_T)]+
2\Ebb_{\P}[f(X_{0})f(X_{t})]+\Ebb_{\P }\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right]\\
&=-2\Ebb_{\P }[f^{2}(X_0)]+2 \Ebb_{\P }[f(X_{0})f(X_{t})]+\Ebb_{\P }\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right]
\end{align*}
\fi
Furthermore, by Schwarz inequality again,
\begin{align}
\label{ach4}
|\Ebb_{\P }[f(X_{0})f(X_{t})]|&\leq \Ebb_{\P }[f^{2}(X_{t})]^{\frac{1}{2}}\Ebb_{\P }[f^{2}(X_{0})]^{\frac{1}{2}}
=\Ebb_{\P }[f^{2}(X_{0})].
\end{align}
Thus $\underset {t\rightarrow \infty}\lim\frac{1}{t}\Ebb_{\P }[f(X_{0})f(X_{t})]=0$. Then by \kak{eqmart} we conclude that
\begin{align}\label{var}
\lim_{t\rightarrow \infty}\frac{1}{t}\Ebb_{\P }[M_{t}^{2}]=\lim_{t\rightarrow\infty}\frac{1}{t}\Ebb_{\P }\left[
\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right]=2\lk f, \LN f \rk _{\K }.
\end{align}
Hence \kak{a1} follows. Next we prove \kak{a2}. By \kak{pos1} we can write
\begin{align*}
&\Ebb_{\P }\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right]=\int_{0}^{t}\!\!\!dr\int_{0}^{t}\!\!\!ds \lk
T_{|s-r|}\LN f, \LN f \rk _{\K }\\
&=2\int_{0\leq r\leq s\leq t} dr ds \lk T_{s-r}\LN f, \LN f \rk _{\K }
=2\int_{0\leq r\leq s\leq t} dr ds \lk T_{r}\LN f, \LN f \rk _{\K }\\
&=2\int_{0\leq r\leq s\leq t} dr ds \lk T_{\frac{r}{2}}\LN f, T_{\frac{r}{2}}^{\ast}\LN f \rk _{\K }
=2 \int_{0}^{t}\!\!\!dr(t-r) \lk T_{\frac{r}{2}}\LN f, T_{\frac{r}{2}}^{\ast}\LN f \rk _{\K }\\
&=4\int_{0}^{\frac{t}{2}}ds(t-2s) \lk T_{s}\LN f,T_{s}^*\LN f \rk_{\K}.
\end{align*}
We have $\LN=\LN^{\ast}$ and thus $T_t=T_t^{\ast}$ for all $t\geqslant0$, hence the invariant probability measure $\P$ is
reversible. Using now reversibility of $\P$, we obtain
\begin{align*}
2t \int_{0}^{\frac{t}{4}}ds \|T_s \LN f\|^2\leq \Ebb_{\P }\left[\lk \int_{0}^{t}\!\!\!dr\LN f(X_{r})\rk ^{2}\right]
\leq 4t \int_{0}^{\frac{t}{2}}ds \|T_s \LN f\|^2.
\end{align*}
This implies \kak{a2}.
\qed
\bt{martingale}
\TTT{Functional central limit theorem}
Let $\proo B$ be standard Brownian motion. Under the assumptions of Lemma \ref{finie} we have
\begin{align*}
\lim_{s\to\infty} \frac{1}{\sqrt{s}}M_{[st]} \stackrel {\rm d} = \s^2 B_{t},
\end{align*}
where
$\s^2=2\lk f\gr ,[H_{0},f]\gr \rk _{L^{2}({\rm P})}$.
\et
\proof
By Lemma \ref{***} the process $(M_t)_{t\geqslant0}$ is a martingale with stationary increments under $\P $.
Furthermore, by Lemma \ref{finie} we have that $\sigma^{2}$ is finite, hence by Proposition \ref{mart} the theorem follows.
To determine $\s^{2}$
note that $(H-E)f\gr=[H_0 ,f]\gr$. Thus $\sigma^{2}=2\lk f, \LN f \rk _{\K }=2\lk f\gr ,[H_{0},f]\gr \rk _{L^{2}({\rm P})}$.
\qed
For suitable $f$, define
\eq{clt}
L_t=\int_{0}^{t}\LN f(X_{s})ds,
\en
which is an additive functional of the reversible Markov process. We can obtain a central limit theorem for such additive
functionals by using Theorem \ref{martingale} and the fundamental result below, see \cite[Theorem 1.8]{kv86}.
\bp{kv}\TTT {Kipnis-Varadhan}
Let $(\Omega,\F, \proo \F, \mu)$ be a filtered probability space and $(A,\mu_0)$ a measurable space, where $\mu$ and $\mu_0$ denote probability measures on
$\Omega$ and $A$ respectively. Let $\proo Y$ be an $A$-valued Markov process with respect to $\proo \F$.
Assume that $\proo Y$ is a reversible and ergodic Markov process with respect to $\mu$.
Let $F:A\rightarrow\mathbb{R}$ be a $\mu_{0}$ square integrable function with $\int_A Fd\mu_{0}=0$. Suppose in addition that
$F$ is in the domain of $L^{-1/2}$, where $L$ is the generator of the process $\proo Y$. Let
$$
R_{t}=\int_{0}^{t}F(Y_{s})ds.
$$
Then there exists a square integrable martingale $\proo N$ with respect to $\proo {\mathcal {F}}$, with stationary increments,
such that
\begin{equation}\label{ac}
\lim_{t\rightarrow\infty}\frac{1}{\sqrt{t}}\sup_{0\leq s\leq t}|R_{s}-N_{s}|=0
\end{equation}
in probability with respect to $\mu$, where $R_{0}=N_{0}=0.$ Moreover,
\begin{align}\label{kv2}
\lim_{t\rightarrow\infty}\frac{1}{t}\mathbb{E}_{\mu}[|R_{t}-N_{t}|^{2}]=0.
\end{align}
\ep
Now we show a central limit theorem for the additive functional $L_t.$
\bt{main2}
\TTT{Functional central limit theorem}
Under the assumptions of Lemma~\ref{finie} the random process $\lk L_{t}\rk_{t\geq0}$ satisfies a functional central limit theorem
relative to $\P $, and the limit variance is given by $ \sigma^{2}=2\lk f\gr ,[H_{0},f]\gr \rk_{L^{2}({\rm P})}$.
\et
\proof
By Lemma \ref{rev1}, the process $(X_{t})_{t\geq0}$ is a reversible Markov process under $\P.$ On the other hand, we see by
Proposition \ref{fik} that the semigroup $(T_{t})_{t\geq0}$ associated to $(X_{t})_{t\geq0}$ is positive, i.e., the process is
ergodic. We have
\begin{align*}
\Ebb_{\P }[\LN f(X_{t})]=\lk \gr ,(H-E)f\gr \rk =\lk (H-E)\gr , f\gr \rk =0.
\end{align*}
Thus $\Ebb_{\P }[L_{t}]=0$, and the assumptions of Proposition \ref{kv} are satisfied. Then $(L_{t})_{t\geq 0}$ is also a martingale
up to a correction term that disappears in the scaling limit. In fact, by \eqref{ac} there exists an $(\F_t)$-martingale
$(N_{t})_{t\geq0}$ such that
\begin{align*}
\lim_{s\rightarrow\infty}\frac{1}{\sqrt{s}}\Ebb_{\P }[\sup_{0\leq t\leq s}|{N_{t}}-L_{t}|]=0,
\end{align*}
and by \eqref{mt} we have that $N_t=M_t,$ and hence
\begin{align}\label{vanish}
\lim_{s\rightarrow\infty}\frac{1}{\sqrt{s}}\Ebb_{\P }[\sup_{0\leq t\leq s}|{M_{t}}-L_{t}|]=0.
\end{align}
Moreover, by \eqref{var} we have
\begin{align}\label{varqu}
\lim_{t\rightarrow\infty}\frac{1}{t}\Ebb_{\P }[|M_{t}-L_t|^{2}]=0.
\end{align}
Finally, by \kak{vanish} the difference $M_t-L_t$ vanishes in the diffusive limit. By \kak{var} and since the martingale
$(M_{t})_{t\geq0}$ has stationary increments, we conclude by Theorem
\ref{martingale} that
$$
\lim_{s\rightarrow\infty}\frac{1}{\sqrt{s}}M_{st}\overset{\P}=\lim_{s\rightarrow\infty}\frac{1}{\sqrt{s}}L_{st}\overset{\P}=
\s^{2} B_{t}, \quad t\geqslant0.
$$
\qed
\subsection{Examples of the variance $\sigma^2$}
In this section we give some examples of direct interest of the functions $f\in D(L_{\rm N})$, $f:\hir\ni(x,\xi)\mapsto f(x,\xi)
\in \mathbb C$, in the FCLT, which allows to compute the variance $\sigma^{2}$ explicitly. In what follows we assume that
$h\in\LR$ is any test function and $\gamma\in \BR$ any real vector. Moreover, we will denote the vector in $L^2(Q)$
associated with the conjugate momentum $\Pi(h)$ in $\fff$, with the same symbol $\Pi(h)$, i.e., we have
\eq{CCR}
[\xi(h),\Pi(h')]=\half(h,h').
\en
\begin{example}
\rm{
Let $f(x,\xi)= \gamma\cdot x$ (a related example is given in \cite{BS}). We have
\begin{align*}
[H_{0},(\gamma \cdot x)]=[-\frac{1}{2}\Delta ,(\gamma \cdot x)]= -\gamma\cdot\nabla.
\end{align*}
Then
\begin{align*}
\sigma^{2}= 2\lk (\gamma \cdot x)\gr ,(-\gamma\cdot \nabla\gr )\rk
=-2\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk x_{j}\gr ,\nabla_{k}\gr \rk
\end{align*}
Denote $X_{jk}=\lk x_{j}\gr , \nabla_{k}\gr \rk $. For $j\neq k$ we have
\begin{align*}
X_{jk} =-\lk \nabla_{k}x_{j}\gr ,\gr \rk =-\lk x_{j}\nabla_{k}\gr ,\gr \rk =-\lk \nabla_{k}\gr , x_{j}\gr \rk =-\bar X_{jk},
\end{align*}
i.e., $\Ree X_{jk}=0$. For $j=k$ we have $\Ree X_{jk}=-\frac{1}{2}$ since
\begin{align*}
X_{jk}=-\lk \nabla_{k}x_{k}\gr , \gr \rk
=-\lk\gr , \gr \rk -\lk x_{k}\nabla_{k}\gr , \gr \rk
=-\|\gr \|^{2} - \lk \nabla_{k}\gr , x_{k}\gr \rk
=-1-\bar X_{jk}.
\end{align*}
Hence finally we get $\sigma^{2}=|\gamma|^{2}$, in particular,
\eq{BS}
|\gamma|^2-2(\gamma\cdot \nabla \gr, (H-E)^{-1} \gamma\cdot \nabla \gr)=0.
\en
}
\end{example}
\begin{example}
\rm{
Let $f(x,\xi)=\xi(h) $. By \kak{ou1}, we have
$$
[H_{0},f]=[\hf , \xi(h)]=-i\Pi(\omega h).
$$
With $X= 2 \lk \xi(h)\gr ,-i\Pi(\omega h)\gr \rk$, we obtain
\begin{align*}
X
&= -2i\lk \gr , \xi(h)\Pi(\omega h)\gr \rk \\
&= -2i\lk \gr , \Pi(\omega h)\xi(h)\gr \rk +2i\lk \gr , \Pi(\omega h)\xi(h)\gr \rk \\
&= -2i\lk \Pi^{\ast}(\omega h)\gr , \xi(h)\gr \rk + 2i\lk \gr , -i\|\sqrt \omega h\|^2 \gr \rk \\
&=-2i\lk -\Pi(\omega h)\gr , \xi(h)\gr \rk +2\|\gr \|_{L^{2}({\rm P})}^{2}\|\sqrt \omega h\|^2 =
2\|\sqrt \omega h\|^2-\overline{X},
\end{align*}
hence
$$
\sigma^{2}= \Ree X =\|\sqrt \omega h\|^2.
$$
}
\end{example}
\begin{example}
\rm{
Let $f(x,\xi)=(\gamma \cdot x)\xi(h)$. We have
\begin{align*}
[H_{0},(\gamma \cdot x)\xi(h)]&=[-\frac{1}{2}\Delta ,(\gamma \cdot x)]\xi(h)+(\gamma \cdot x)[\hf ,\xi(h)]\\
&= -\gamma\cdot\nabla\xi(h)-i(\gamma \cdot x)\Pi(\omega h).
\end{align*}
Then
\begin{align}\label{sig2}
\sigma^{2}&= 2\lk (\gamma \cdot x)\xi(h)\gr ,\xi(h)(-\gamma\cdot\nabla\gr )\rk +
2\lk (\gamma \cdot x)\xi(h)\gr ,-i(\gamma \cdot x)\Pi(\omega h)\gr \rk \\
&=-2\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk x_{j}\xi(h)\gr ,\xi(h)\nabla_{k}\gr \rk
+2\sum_{1\leq j,k\leq d} \gamma_{j}\gamma_{k}\lk x_{j}\xi(h)\gr ,-i\Pi(\omega h)x_{k}\gr \rk \nonumber.
\end{align}
Denote again $X_{jk}=\lk x_{j}\xi(h)\gr , \xi(h)\nabla_{k}\gr \rk $. For $j\neq k$
\begin{align*}
X_{jk} &=-\lk \xi(h)\nabla_{k}x_{j}\gr ,\xi(h)\gr \rk =-\lk \xi(h)x_{j}\nabla_{k}\gr ,\xi(h)\gr \rk \\
&=-\lk \xi(h)\nabla_{k}\gr , x_{j}\xi(h)\gr \rk =-\bar X_{jk}.
\end{align*}
For the diagonal part we have
\begin{align*}
X_{kk} &=-\lk \xi(h)\nabla_{k}x_{k}\gr , \xi(h)\gr \rk \\
&=-\lk \xi(h)\gr , \xi(h)\gr \rk -\lk \xi(h)x_{k}\nabla_{k}\gr , \xi(h)\gr \rk \\
&=-\lk \xi(h)\gr , \xi(h)\gr \rk - \lk \xi(h)\nabla_{k}\gr , x_{k}\xi(h)\gr \rk =-\|\xi(h)\gr \|^{2}-\bar X_{kk},
\end{align*}
i.e., $\Ree X_{kk} = -\frac{1}{2}\|\xi(h)\gr \|^{2}$. To determine the second term in \kak{sig2}, write now
$X_{jk} =\lk x_{j}\xi(h)\gr ,-i\Pi(\omega h)x_{k}\gr \rk $. We have
\begin{align*}
X_{jk} &=-i\lk x_{j}\gr ,x_{k}\xi(h)\Pi(\omega h)\gr \rk \\
&=-i\lk x_{j}\gr ,x_{k}\Pi(\omega h)\xi(h)\gr \rk +i\lk x_{j}\gr ,x_{k}\Pi(\omega h)\xi(h)\times\gr \rk \\
&=-i\lk x_{k}\Pi^{\ast}(\omega h)\gr ,x_{j}\xi(h)\gr \rk +\lk x_{j}\gr ,\|\sqrt \omega h\|^2x_{k}\gr \rk
=-\bar X_{jk} +\lk x_{j}\gr ,x_{k}\gr \rk \|\sqrt \omega h\|^2.
\end{align*}
Hence we finally obtain
$$
\sigma^{2}=|\gamma|^{2}\|\xi(h)\gr \|^2+2\|(\gamma \cdot x)\gr \|^2\|\sqrt \omega h\|^2.
$$
}
\end{example}
\begin{example}
\rm{
Let $f(x,\xi)=e^{i\xi(h)}$. We have
\begin{align*}
e^{-i\xi(h)} \hf e^{i\xi(h)}&=\hf +i[\hf , \xi(h)]+\frac{1}{2}[[\hf , i\xi(h)],i\xi(h)]\\
&=\hf +\Pi(\omega h)+\frac{1}{2}\|\sqrt \omega h\|^2.
\end{align*}
Thus
$$
\hf e^{i\xi(h)}=e^{i\xi(h)}\hf +e^{i\xi(h)}\Pi(\omega h)+\frac{1}{2}\|\sqrt \omega h\|^2e^{i\xi(h)}$$
and
$$
[H_{0},e^{i\xi(h)} ]=e^{i\xi(h)}\Pi(\omega h)+\frac{1}{2}\|\sqrt \omega h\|^2e^{i\xi(h)}$$
follow. This gives
\begin{align*}
\sigma^{2}&=2\lk e^{i\xi(h)}\gr ,[H_{0},e^{i\xi(h)} ]\gr\rk = 2\lk \gr ,\Pi(\omega h)\gr\rk +\|\sqrt \omega h\|^2.
\end{align*}
}
\end{example}
\begin{example}
\rm{
Let $f(x,\xi)=(\gamma \cdot x)e^{i\xi(h)}$. We have
\begin{align*}
[H_{0},(\gamma \cdot x)e^{i\xi(h)} ]&=[-\frac{1}{2}\Delta +\hf ,(\gamma \cdot x)e^{i\xi(h)} ]\\
&=[-\frac{1}{2}\Delta ,(\gamma \cdot x)]e^{i\xi(h)}+(\gamma \cdot x)[\hf , e^{i\xi(h)}]\\
&=-\gamma \cdot \nabla e^{i\xi(h)}+(\gamma \cdot x)e^{i\xi(h)}\lk \Pi(\omega h)+\frac{1}{2}\|\sqrt \omega h\|^2\rk.
\end{align*}
Then
\begin{align}\label{sig4}
\sigma^{2}&=2\lk (\gamma \cdot x)e^{i\xi(h)}\gr ,
[H_{0},(\gamma \cdot x)e^{i\xi(h)} ]\gr \rk \nonumber\\
&=2\lk (\gamma \cdot x)e^{i\xi(h)}\gr ,-\gamma \cdot \nabla\varphi _{g}e^{i\xi(h)}+(\gamma \cdot x)e^{i\xi(h)}
\lk \Pi(\omega h)+\frac{1}{2}\|\sqrt \omega h\|^2\rk \gr \rk\nonumber \\
&=2\lk (\gamma \cdot x)\gr ,-\gamma \cdot \nabla\gr \rk +
2\lk (\gamma \cdot x)\gr ,(\gamma \cdot x)\Pi(\omega h)\gr \rk +\|(\gamma \cdot x)\gr \|\|\sqrt \omega h\|^2\\
&=\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk x_{j}\gr ,\nabla_{k}\gr \rk +
2\lk (\gamma \cdot x)\gr ,(\gamma \cdot x)\Pi(\omega h)\gr
\rk+\|(\gamma \cdot x)\gr
\|^{2}\|\sqrt \omega h\|^2.\nonumber
\end{align}
To get the first term, denote $X_{jk}=-\lk x_{j}\gr , \nabla_{k}\gr \rk$. For the off-diagonal part we have
\begin{align*}
X_{jk} =\lk \nabla_{k}x_{j}\gr , \gr \rk =\lk x_{j}\nabla_{k}\gr , \gr \rk =\lk \nabla_{k}\gr , x_{j}\gr \rk =
-\bar X_{jk},
\end{align*}
and the diagonal part gives $X_{kk} =\lk \nabla_{k}x_{k}\gr , \gr \rk =1+\lk \nabla_{k}\gr , x_{k}\gr \rk$. Hence
in total
$$
\sigma^{2}=|\gamma|^{2}+\|(\gamma \cdot x)\gr \|^{2}\|\sqrt \omega h\|^2+
2\lk (\gamma \cdot x)\gr ,\Pi(\omega h)(\gamma \cdot x)\gr \rk.
$$
}
\end{example}
\begin{example}
\rm{
Let $f(x,\xi)=e^{i(\gamma \cdot x)+i\xi(h)}$. We have
\begin{align*}
&[H_{0},e^{i(\gamma \cdot x)+i\xi(h)} ]
= [-\frac{1}{2}\Delta +\hf , e^{i(\gamma \cdot x)+i\xi(h)}]
=[-\frac{1}{2}\Delta ,e^{i(\gamma \cdot x)}]e^{i\xi(h)}+[\hf , e^{i\xi(h)}]e^{i(\gamma \cdot x)}\\
&=e^{i(\gamma \cdot x)+i\xi(h)}(\frac{1}{2}|\gamma|^{2}-i\gamma\cdot \nabla)
+e^{i(\gamma \cdot x)+i\xi(h)}(\Pi(\omega h)+\frac{1}{2}\|\sqrt \omega h\|^2)
\end{align*}
Thus
\begin{align*}
\sigma^{2}&=2 \lk e^{i(\gamma \cdot x)+i\xi(h)}\gr , [H_{0},e^{i(\gamma \cdot x)+i\xi(h)} ]\gr \rk \\
&=|\gamma|^{2}-2\lk \gr ,i\gamma\cdot \nabla\gr \rk +2\lk \gr , \Pi(\omega h)\gr \rk +\|\sqrt \omega h\|^2.
\end{align*}
Let $X= \lk \gr ,i\gamma\cdot \nabla\gr \rk $.
Note that $\gr>0$.
Since $X\in \mathbb{R}$ and $-i X\in \RR$, we have $X=0$ and thus
$$
\sigma^{2}=|\gamma|^{2}+2\lk \gr , \Pi(\omega h)\gr \rk +\|\sqrt \omega h\|^2.
$$
}
\end{example}
\subsection{Relativistic Nelson model}
\label{ccc}
The previous constructions can be extended to the relativistic case. Let
\eq{mt1}
\tilde{M}_{t}=f(\oo_{t})-f(\oo_{0})+\int_{0}^{t}\tilde{L}_{\rm N} f(\oo_{s})ds,\quad t\geqslant0.
\en
\bt{martingale2}
$(\tilde{M}_t)_{t\geqslant0 }$ is a martingale with stationary increments under $\tilde{\P}$, and
\begin{align*}
\lim_{s\to\infty} \frac{1}{\sqrt{s}}\tilde M_{st} \stackrel {\rm d} =
\tilde \s^2
B_{t},
\end{align*}
where
$\tilde \s^2=2\lk f\gr ,[\tilde H_{0},f]\gr \rk $.
\et
\proof
The proof is an analogue of
Lemma \ref{***} and Theorem \ref{martingale}.
\qed
We conclude by some explicit cases of variances $\tilde{\sigma}^2$.
\begin{example}
\rm{
Let $g(x,\xi)= \gamma\cdot x$. We have
\begin{align*}
[\tilde{H}_{0},(\gamma \cdot x)]=[\sqrt{-\Delta+m^2} ,(\gamma \cdot x)]= \frac{-(\gamma\cdot\nabla)}{\sqrt{-\Delta+m^2}}.
\end{align*}
Then
\begin{align}\label{ge}
\tilde{\sigma}^2
&=2\lk (\gamma\cdot x)\ggr , \frac{-(\gamma\cdot\nabla)}{\sqrt{-\Delta+m^2}} \ggr \rk\nonumber\\
&=-2\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \frac{-\nabla_k}{\sqrt{-\Delta+m^2}} x_j \ggr , \ggr \rk\nonumber\\
&=-2\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , [x_j, \frac{\nabla_k}{\sqrt{-\Delta+m^2}} ] \ggr \rk-2\lk \frac{-\nabla_k}{\sqrt{-\Delta+m^2}} \ggr , x_j \ggr \rk\nonumber\\
&=-2\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , [x_j, \frac{\nabla_k}{\sqrt{-\Delta+m^2}}] \ggr \rk- \overline{\tilde{\sigma}^2}.
\end{align}
Moreover, we have
\begin{align*}
[x_j, \frac{\nabla_k}{\sqrt{-\Delta+m^2}} ] &=[x_j,\nabla_k] \frac{1}{\sqrt{-\Delta+m^2}}+\nabla_k [x_j,\frac{1}{\sqrt{-\Delta+m^2}}]\\
&=-\frac{\delta_{jk}}{\sqrt{-\Delta+m^2}}-\frac{\nabla_k\nabla_j}{({-\Delta+m^2})^{\frac{3}{2}}}.
\end{align*}
By \eqref{ge}, we obtain
\begin{align*}
\tilde{\sigma}^2&=2\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , \frac{1}{\sqrt{-\Delta+m^2}} \ggr \rk\delta_{jk}
+2\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , \frac{\nabla_k\nabla_j}{(-\Delta+m^2)^{\frac{3}{2}}} \ggr \rk- \overline{\tilde{\sigma}^2}.
\end{align*}
Hence we conclude that
\begin{align*}
\tilde{\sigma}^2&=
\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , \frac{1}{\sqrt{-\Delta+m^2}} \ggr \rk\delta_{jk}+
\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , \frac{\nabla_k\nabla_j}{(-\Delta+m^2)^{\frac{3}{2}}} \ggr \rk.
\end{align*}
}
\end{example}
\begin{example}
\rm{
Let $g(x,\xi)=( \gamma\cdot x) \xi(h)$. We have
\begin{align*}
[\tilde{H}_{0},(\gamma \cdot x)\xi(h)]
=[\sqrt{-\Delta+m^2},(\gamma \cdot x)] \xi(h)
+ (\gamma \cdot x)[\hf,\xi(h)].
\end{align*}
Thus
\begin{align*}
\tilde{\sigma}^2
&=2\lk (\gamma\cdot x)\xi(h)\ggr ,[\sqrt{-\Delta+m^2},(\gamma \cdot x)] \xi(h) \ggr \rk+2\lk (\gamma\cdot x)\xi(h)\ggr ,[\hf,\xi(h)] (\gamma \cdot x) \ggr \rk\\
&=\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \xi(h)\ggr , \frac{1}{\sqrt{-\Delta+m^2}}\xi(h) \ggr \rk\delta_{jk}+\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \xi(h)\ggr, \frac{\nabla_k\nabla_j}{(-\Delta+m^2)^{\frac{3}{2}}} \xi(h)\ggr \rk\\
&\hspace{3cm}+2\|(\gamma \cdot x)\ggr \|\|\sqrt \omega h\|^2.
\end{align*}
}
\end{example}
\begin{example}
\rm{
Let $g(x,\xi)=( \gamma\cdot x) e^{i\xi(h)}$. We have
\begin{align*}
[\tilde{H}_{0},(\gamma \cdot x)e^{i\xi(h)}]
=[\sqrt{-\Delta+m^2},(\gamma \cdot x)] e^{i\xi(h)}+ (\gamma \cdot x)[\hf,e^{i\xi(h)}].
\end{align*}
Similarly, we obtain
\begin{align*}
\tilde{\sigma}^2
&=2\lk (\gamma\cdot x)e^{i\xi(h)}\ggr ,[\sqrt{-\Delta+m^2},(\gamma \cdot x)] e^{i\xi(h)}\ggr \rk+2\lk (\gamma\cdot x)e^{i\xi(h)}\ggr ,[\hf,e^{i\xi(h)}] (\gamma \cdot x)\ggr \rk\\
&=\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , \frac{1}{\sqrt{-\Delta+m^2}} \ggr \rk\delta_{jk}+
\sum_{1\leq j,k\leq d}\gamma_{j}\gamma_{k}\lk \ggr , \frac{\nabla_k\nabla_j}{(-\Delta+m^2)^{\frac{3}{2}}} \ggr \rk\\
&\hspace{3cm}+\|(\gamma \cdot x)\ggr \|^2\|\sqrt \omega h\|^2+2\lk (\gamma \cdot x)\ggr ,\Pi(\omega h)(\gamma \cdot x)\ggr \rk .
\end{align*}
}
\end{example}
\begin{example}
\rm{
Let $g(x,\xi)=e^{i(\gamma \cdot x)+i\xi(h)}$. We have
\begin{align*}
[\tilde{H}_{0},e^{i\gamma \cdot x+i\xi(h)}]
=[\sqrt{-\Delta+m^2},e^{i\gamma \cdot x}] e^{i\xi(h)}+ e^{\gamma \cdot x}[\hf,e^{i\xi(h)}].
\end{align*}
Since
$
e^{-i\gamma \cdot x} \sqrt{-\Delta+m^2} e^{i\gamma \cdot x}=\sqrt{-(\nabla-i\gamma)^2+m^2}$,
we obtain
\begin{align*}
[\sqrt{-\Delta+m^2},e^{i\gamma \cdot x}]=e^{i\gamma \cdot x} \sqrt{-(\nabla-i\gamma)^2+m^2}- e^{i\gamma \cdot x} \sqrt{-\Delta+m^2}.
\end{align*}
Finally, we deduce that
\begin{align*}
\tilde{\sigma}^2
&=2\lk \ggr , \sqrt{-(\nabla-i\gamma)^2+m^2}\ggr \rk-2\lk \ggr , \sqrt{-\Delta+m^2} \ggr \rk+2 \lk \ggr , \Pi(\omega h)\ggr \rk
+\|\sqrt \omega h\|^2\\
&= 2\lk \ggr , \sqrt{(-i\nabla-\gamma)^2+m^2}\ggr \rk-2\lk \ggr , \sqrt{-\Delta+m^2} \ggr \rk+2 \lk \ggr , \Pi(\omega h)\ggr \rk
+\|\sqrt \omega h\|^2.
\end{align*}
}
\end{example}
\section{Concluding remarks}
Although in this paper we focused on the Nelson model, $P(\phi)_1$ processes and an FCLT can further be constructed also for related
models. We briefly mention two cases.
\medskip
\noindent
\emph{Nelson model with fixed total momentum $P$.} Let $V=0$. Then $H_{\rm N}$ is translation invariant, i.e.,
$[H_{\rm N}, T_{\rm tot}]=0$, where $T_{\rm tot}=p\otimes\one+\one\otimes\pf$ denotes the total momentum and
$\pf_{\mu}=d\Gamma (k_\mu)$. Thus $H_{\rm N}$ can be decomposed as
$$
H_{\rm N}=\int_\BR ^\oplus H_{\rm N}(P) dP,
$$ where
$$
H_{\rm N}(P)=\half (P-\pf)^2+\phi(0)+\hf
$$
is a self-adjoint operator in $\fff$, called Nelson Hamiltonian with total momentum $P\in\BR$. It is known that for sufficiently
small $|P|$ the operator $H_{\rm N}(P)$ has a ground state \cite{fro74}.
\medskip
\noindent
\emph{Pauli-Fierz model.} The Pauli-Fierz Hamiltonian in non-relativistic quantum electrodynamics is defined by
$$
H_{\rm PF}=\frac{1}{2m}(-i\nabla\otimes\one +\sqrt\alpha A)^2+V\otimes \one +\one\otimes \hf,
$$
where $A$ denotes the quantized radiation filed given by
$$A_\mu(x)=\frac{1}{\sqrt 2}\sum_{j=1,2}\int_\BR\left(\frac{\vp(k)}{\sqrt{\omega(k)}}e_\mu(k,j)e^{ikx}\add(k,j) +
\frac{\vp(-k)}{\sqrt{\omega(k)}}e_\mu(k,j)e^{-ikx} a(k,j)\right)dk.
$$
Here $e(k,1)$ and $e(k,2)$ denote polarisation vectors such that $k\cdot e(k,j)=0$ for $j=1,2$, and
$[a(k,j),\add(k',j')]=\delta_{jj'}\delta(k-k')$ is satisfied. Then $H_{\rm PF}$ is a self-adjoint operator in $L^2(\RR^3)
\otimes \fff(L^2(\RR^3\times\{1,2\}))$. The existence of the ground state is studied in \cite{bfs3, gll,hir00}.
\bigskip
\noindent
{\bf Acknowledgements:}
FH is financially supported by JSPS Open Partnership Joint Projects between Japan and Tunisia
``Non-commutative infinite dimensional harmonic analysis: A unified approach from representation theory and probability
theory", and is also supported by
Grant-in-Aid for Science Research (B)16H03942
from JSPS.
JL thanks IHES, Bures-sur-Yvette, for a visiting fellowship, where part of this paper has been written.
\bibliographystyle{amsplain}
|
\section{Introduction}
Is the tacit agreement (perhaps a kind of ``dogma'') of the directed completeness of semantic domains falling?
We might get this impression from the recent results of Dag Normann and Vladimir Sazonov
on the game model of PCF.
So we begin this introduction with a brief history of the models of PCF.
\subsection*{History}
PCF is a simply typed $\la$-calculus on integers with higher-order recursion.
The concept of PCF was formed by Dana Scott in 1969, see the historical document
\cite{Scott}.
It is used as a prototypical programming language to explore the relationship
between operational and denotational semantics, see the seminal paper of
Gordon Plotkin \cite{Plotkin}.
The first model of Scott was made of directed complete partial orders,
beginning with flat domains for integers and booleans and the full domain of continuous functionals
for higher-order types.
It was natural to demand directed completeness of the domains,
so that we get a definition of continuity that leads to closure under function spaces
and the existence of
all conceivable fixpoints for the semantics of recursive functions.
But the model contained ideal elements that were not realized in the language:
These were finite elements, like the ``parallel or'' function,
or infinite elements (as lubs of directed sets of finite elements).
The presence of
the finite unrealized elements (like ``parallel or'') already causes the model to be not
fully abstract (i.e.\ the denotational semantics does not match the operational one),
as was observed by Gordon Plotkin \cite{Plotkin}.
The question of a fully abstract model arose, where a model was supposed to consist of
directed complete partial orders,
following the established ``dogma''.
The first fully abstract cpo model was constructed by Robin Milner \cite{Milner} in 1977
from equivalence classes of finite combinator terms by an inverse limit of domains.
Later the same model was constructed by G\'erard Berry \cite{Berry}\ from equivalence classes
of proper PCF-terms by an ideal completion.
So this model was built on syntactic terms of the language, which was not considered satisfactory,
and the search for a purely mathematical fully abstract cpo model began.
The widely accepted solution was the game semantics after 1990
\cite{Abramsky/Jagadeesan,Hyland/Ong,Nickau}.
In game semantics a term of PCF is modeled by a strategy of a game,
i.e.\ by a process that performs a dialogue of questions and answers with the environment, the opponent.
These strategies are still intensional; the fully abstract model\ is formed by a quotient, the extensional
collapse.
The strategies can be identified with
PCF B\"ohm trees of a certain normal form,
see \cite[section 6.6]{Amadio/Curien}.
It was an open problem whether the model of game domains is isomorphic to Milner's fully abstract cpo-model,
i.e.\ whether its domains are cpos and so contain every element of the cpo-model.
This problem was solved by Dag Normann \cite{Normann}:
its domains are not cpos, i.e.\ there are directed sets that have no lub.
The example given by Normann is in type 3 and rather sophisticated.
Then Vladimir Sazonov made a first attempt to build a general theory for these
non-cpo domains \cite{Sazonov:models,Sazonov:natural}.
His main important insight was that functions are continuous only with respect to certain lubs of directed sets
that he calls ``natural lubs''; these are the hereditarily pointwise lubs in PCF.
He defines an abstract structure of ``natural domains'' \cite{Sazonov:natural} as a partial order
with an operator that designates certain lubs of (general, not only directed)
subsets as ``natural lubs'', fulfilling some axioms.
He shows that the category of natural domains and functions that are continuous w.r.t.\ the
directed natural lubs is a cartesian closed category (ccc).
He defines naturally finite elements and natural algebraicity w.r.t.\ the natural directed lubs,
and also shows that naturally algebraic, bounded complete natural domains form a ccc.
In a recent paper Normann and Sazonov \cite{Normann/Sazonov} show that in the game model of PCF,
the sequential functionals,
there is a Normann-example in a second-order type, there
are directed lubs that are not natural, and there are naturally finite elements
that are not finite in the classical sense.
The main results of the last paper are also covered in the recent textbook
of Longley and Normann: ``Higher-order computability'' \cite{Longley/Normann},
section 7.6.
Generally speaking, all these problems are due to a fundamental mismatch between
the two worlds that semantics is relating:
The world of syntax, of mechanism (in the form of programming languages and abstract machines),
of intension on one side,
and the more abstract world of domains and continuous functions, of extension
on the other side.
The problems are generally caused by restrictions on the syntactic side.
So the restriction to sequentiality caused the full abstraction problem for PCF.
Its solution, games, are constructions that stand somewhat in the middle between the two worlds.
Sazonov's natural domain theory accounts for another syntactic restriction:
that limits of ascending chains of finite elements can only be formed for those
chains that are ascending with the syntactic order, as a Boehm tree.
(This is not bound to sequentiality, as Sazonov shows in \cite{Sazonov:models}
a corresponding model for PCF with parallel conditional.)
Natural domains model this incompleteness,
but they miss an important property:
the existence of fixpoints of endofunctions (on domains with $\bot$).
(The domains of game semantics have this property,
a mechanism always has the fixpoints by construction.)
So a category of abstract incomplete domains (e.g.\ natural domains)
must be understood as a (cartesian closed) ``house'' in which several
more syntactic programming language models (with the fixpoint property) live together.
We pose as an open problem to find categories of abstract incomplete domains
with the existence of fixpoints.
We have seen that the game domain model, the model of sequential functionals,
results from a syntactic restriction to Boehm trees.
There are other syntactic restrictions conceivable,
the most extreme being a restriction just to the terms of PCF itself.
So in my paper \cite[section 3]{Mueller:berry} on Berry's conjectures
I have given the definition of a whole spectrum of
fully abstract models of PCF (``f-models'') as sets of ideals of equivalence classes of finite terms,
such that application is defined and every PCF-term has a denotation.
In this spectrum Milner's cpo model is the largest model,
the pure term model is the least,
and the game model is properly between the two.
\subsection*{Ideas of this paper}
In this paper we further explore the abstract domain theory of incomplete domains.
Our main objective is to find cartesian closed categories.
We begin high above the natural domains with the most general conceivable structure,
the directed-lub partial orders (dlubpo),
partial orders with designated directed lubs,
in the form of a relation $A\nlub_D a$,
meaning that the directed subset $A$ has the natural lub $a$.
The only axiom they must obey is the singleton axiom $\{a\}\nlub_D a$
(Sazonov's axiom 3).
We define lub-rule classes to classify axiom systems with a form like that
of natural domains.
A lub-rule on a partial order $D$ is a triple $(D,P\mathord{\leadsto} A)$
with $P$ a set of subsets of $|D|$ that each have a lub
and $A$ a subset of $|D|$ that has a lub.
This expresses the fact that in $D$ we can infer from the existence of lubs of the
elements of $P$ the existence of the lub of $A$.
This inference, this lub-rule, is ``valid'' if it is invariant under monotonic functions,
i.e.\ every monotonic function $f\typ D\fun E$ for some partial order $E$
that respects the lubs of the elements of $P$ respects also the lub of $A$.
We explore axiom systems (of dlubpos) that can be described by classes of valid lub-rules.
We explain the philosophical significance of our translation from an axiom system $S$ to
a lub-rule class as a ``partial extensionalization'' of (the intension of) $S$,
i.e.\ an extension that is between the pure syntax of $S$ and the full extension,
the class of dlubpos that fulfill $S$.
The validity of a lub-rule can be characterized by a closure operator $\operatorname{cl}_D$ on
subsets of dlubpos $D$ that infers from one element all elements below it,
and from a natural subset the natural lub of it.
This closure operator already appeared in the work of Bruno Courcelle and Jean-Claude Raoult
on completions of ordered magmas \cite{Courcelle/Raoult}.
A lub-rule class is complete if it encompasses all valid lub-rules.
The dlubpos generated by a complete lub-rule class are called closed dlubpos (cdlubpo).
They can be characterized by the closure operator $\operatorname{cl}_D$,
i.e.\ they fulfill the axiom S9 (closure):\\
If $A\sub |D|$ is directed with lub $a$ and $a\in \operatorname{cl}_D A$, then $A\nlub_D a$.\\
They form a ccc.
There are natural domains that are no cdlubpo.
In contrast to natural domains, cdlubpos have several characterizations
as canonical structures.
Cdlubpos are the dlubpos that are ``realized''
by ``restricted dcpos'' (rdcpo).
So complete domains are coming in again, and the ``dogma'' of completeness could be ``saved''.
The idea is to complete every cdlubpo with new improper elements
(the ``blind realizers'') to a dcpo that retains the data of the incomplete cdlubpo
as a subset, by an order embedding.
This idea of realization of a partial order by a dcpo goes back to Alex Simpson \cite{Simpson}:
In Simpson's approach every element of the partial order is realized by one or several realizers of the dcpo,
while every realizer of the dcpo realizes exactly one element of the partial order.
In our approach every element of the partial order is realized by exactly one realizer of the dcpo,
while every realizer of the dcpo realizes at most one element of the partial order.
(The two approaches could be combined, see the last section \ref{s:outlook} Outlook.)
In a sequel paper we will work out the connection between the categories
$\mathbf{Dlubpo}$ and $\mathbf{Rdcpo}$.
There is an adjunction between them,
which establishes an adjoint equivalence between the sub-cccs
$\mathbf{Cdlubpo}$ and $\mathbf{Crdcpo}$ (closed rdcpos).
This connection also makes it possible to transfer the theory of cccs of algebraic dcpos to the
realm of closed rdcpos resp.\ cdlubpos.
\subsection*{Outline of the paper}
\begin{enumerate}[1.]
\setcounter{enumi}{1}
\item Preliminaries and notation:\\
We repeat a concrete definition of cartesian closed categories
and basic definitions of dcpo theory.
Throughout the paper we will encounter closure and completion procedures
that all follow the same abstract scheme.
Here we extract this scheme as a fixpoint lemma on powersets and the definition
of ``rule systems'' with their deductions.
\item Sazonov's definition of natural domains revisited:\\
We repeat Sazonov's definition of natural domains,
and give a simpler equivalent axiom system.
Although we deviate from natural domains in the following sections,
we will refer to the new axioms that are inferred here.
We also show an example of a natural domain where natural lubs are
needed for general subsets, not only directed ones.
\item Directed-lub partial orders (dlubpo):\\
We define directed-lub partial orders as the most general structure we conceive.
In this category the exponents, if they exist, have a definition that
is generally different from the pointwise exponents of natural domains.
We give sufficient conditions on subcategories of $\mathbf{Dlubpo}$ to have
terminal, products and exponents like the normal ones.
The main theorems are that $\mathbf{Dlubpo}$ is no ccc,
but that the full subcategory of dlubpos that fulfill Sazonov's axiom S5
is already a ccc (with exponents of the general form).
This section contains the basic definitions of dlubpos,
but the results are mainly unrelated to the rest of the paper,
only few minor ones are used in the following sections.
For a first reading, I recommend to read only the basic definitions 4.1 to 4.4
and skip the rest of the section.
\item Lub-rule classes and closed dlubpos:\\
We introduce the (valid) lub-rules, lub-rule systems
and (complete, invariant) lub-rule classes we described above.
We define closed dlubpos (cdlubpo) fulfilling the closure axiom S9.
These are the dlubpos fulfilling all valid (directed) lub-rules.
Every cdlubpo is a natural domain.
The dlubpos generated by an invariant lub-rule class form a full reflective
subcategory of $\mathbf{Dlubpo}$.
\item The ccc of S10-dlubpos:\\
We introduce a new axiom S10 for dlubpos.
The category of these dlubpos is the largest full sub-ccc of $\mathbf{Dlubpo}$
that is generated by an invariant lub-rule class and has the
pointwise exponents.
Every natural domain and every cdlubpo is in this category.
\item Example of a natural domain that is no cdlubpo:\\
We first show by a simple finite example that the lub-rule class
corresponding to the axioms of natural domains is not complete.
Then we give an example of a natural domain that is no cdlubpo.
\item Algebraic dlubpos:\\
We show that every algebraic dlubpo that fulfills axiom S6 (cofinality) is a cdlubpo.
This means that algebraic natural domains and algebraic cdlubpos are the same.
\item Restricted partial orders and restricted dcpos:\\
We give a sufficient condition based on subcategory morphisms
for a dlubpo to be a cdlubpo.
We define restricted partial orders (rpo) and restricted dcpos (rdcpo)
and show that cdlubpos are exactly the dlubpos realized by rpos resp.\ rdcpos.
\item
Outlook
\end{enumerate}
\section{Preliminaries and notation}
Notation:
We mostly write function application without brackets, if possible.
Function application associates to the left.
If $f$ is a function that is defined on the elements of a set $A$,
then we write $\fset{f} A=\set{fa}{a\in A}$.
$\mathcal{P}(S)$ is the powerset of the set $S$.
We give some basic definitions of category theory and domain theory,
and then a fixpoint lemma on powersets with a definition of ``rule systems''.
\subsection{Category theory}
The main property of our categories of domains that interests us here is cartesian closedness.
We adopt a concrete definition from \cite[def. 4.2.5]{Amadio/Curien}:
\begin{defi}[cartesian closed category] Let $\mathbf{K}$ be a category.\hfill
\begin{enumerate}[\em(1)]
\item
A \defin{terminal object} in $\mathbf{K}$ is a $\top\in\mathbf{K}$ such that:\\
$\all{C\in\mathbf{K}}\exiun{h\typ C\fun \top}$
\item
A \defin{product} of $D,E\in \mathbf{K}$ is an object $D\pro E\in \mathbf{K}$
with \defin{projections} $\proi\typ D\pro E\fun D$
and $\prot\typ D\pro E\fun E$
such that:\\
$\all{C\in\mathbf{K}}\all{f\typ C\fun D}\all{g\typ C\fun E}\exiun{h\typ C\fun D\pro E}
(\proi\comp h=f \text{ and } \prot\comp h=g).$\\
The morphism $h$ is denoted by $\lpro f,g\rpro$, $\lpro\_,\_\rpro$
is called the \defin{pairing operator}.\\
For $f\typ D\fun D'$, $g\typ E\fun E'$:
$f\pro g\typ D\pro E\fun D'\pro E'$ is defined as $f\pro g=\lpro f\comp\proi, g\comp\prot\rpro$.
\item
Let $\mathbf{K}$ have all (binary) products.\\
An \defin{exponent} of $D,E\in\mathbf{K}$ is an object $D\mathbin{\Rightarrow} E\in \mathbf{K}$
with an \defin{evaluation morphism} $\eval\typ (D\mathbin{\Rightarrow} E)\pro D\fun E$ such that:\\
$\all{C\in\mathbf{K}}\all{f\typ C\pro D\fun E}\exiun{h\typ C\fun(D\mathbin{\Rightarrow} E)}
(\eval\comp(h\pro \id)=f).$\\
The morphism $h$ is denoted by $\operatorname{curry}(f)$,
$\operatorname{curry}$ is called the \defin{currying operator}.
\end{enumerate}
A \defin{cartesian closed category (ccc)} is a category that has a terminal object and
all (binary) products and all exponents.
\end{defi}
We will also see adjunctions and reflective subcategories,
we will mainly use the notations of \cite{Maclane}.
\subsection{Domain theory}
We take our definitions from the excellent textbook \cite{Amadio/Curien}.
A structure $D=(|D|,\lex_D)$ is a \defin{partial order (po)}
if $|D|$ is a set and $\lex_D$ is a binary relation on $|D|$
that is reflexive, transitive and antisymmetric.\\
$\lex_D$ is simply written $\lex$ if $D$ is clear from the context,
and this abbreviation also applies to other structures and indices.\\
A non-empty subset $A\sub|D|$ is \defin{directed}
if for all $a,b\in A$ there is some $c\in A$ with $a\lex c$ and $b\lex c$.\\
For a partial order $D$ (or some extension of a partial order with more data)
$\bar{\Pot}(D)$ is the set of subsets of $|D|$ that have a least upper bound (lub) in $D$;
and $\Pot^\uparrow(D)$ is the set of directed subsets of $|D|$ that have a lub in $D$.\\
For subsets $A,B\sub|D|$ it is written $A\sqsubseteq B$ if
for all $a\in A$ there is $b\in B$ with $a\lex b$;
and for $A\sub |D|$, $b\in |D|$: $A\lex b$ if for all $a\in A$ it is $a\lex b$.
$D$ is a \defin{directed complete partial order (dcpo)}
if every directed subset of $A\sub|D|$ has a least upper bound (lub), denoted $\Lubex A$.
If futhermore $D$ has a least element (written $\bot$),
then it is called a \defin{complete partial order (cpo)}.\\
For $D,E$ dcpos, a function $f\typ|D|\fun|E|$ is \defin{continuous}
if it
preserves directed lubs: for all directed $A\sub |D|$ it is $f(\Lubex A)=\Lubex(\fset{f} A)$.
(Then it is also monotonic,
i.e.\ $a\lex_D b \ifthen fa\lex_E fb$.)
$\mathbf{Dcpo}$ is the category of dcpos and continuous functions.
The continuity of $f$ is specified by $f\typ D\fun E$,
which always means that $f$ is a morphism in the category of $D$ and $E$.\\
$\mathbf{Dcpo}$ is a ccc:
The terminal object is the one point dcpo.
For dcpos $D,E$ the product is $D\pro E= (|D|\pro|E|, \lex_{D\pro E})$
where $\lex_{D\pro E}$ is pointwise.
The exponent is $D\mathbin{\Rightarrow} E = (\elem{D\mathbin{\Rightarrow} E},\lex_{D\mathbin{\Rightarrow} E})$
where $\elem{D\mathbin{\Rightarrow} E}$ is the set of continuous functions,
and $\lex_{D\mathbin{\Rightarrow} E}$ is the pointwise order.
Lubs of directed sets of continuous functions are taken pointwise.
\subsection{Fixpoint lemma and rule systems}
We will encounter many closure and completion procedures that all follow a certain abstract
scheme that is here extracted.
These results are certainly all well known,
but I did not find a reference with proofs in the literature that matches.
We will use ordinals, a short introduction to them can be found in \cite{Johnstone}.
\begin{defi}
Let $S$ be a set.\\
A function $f\typ \mathcal{P}(S)\fun \mathcal{P}(S)$ is a \defin{preclosure}
if it is \defin{increasing}: for all $A\sub S$ it is $fA\supseteq A$,
and it is \defin{monotonic}: for $A,B\sub S$, $A\sub B$ implies $fA\sub fB$.\\
An $A\sub S$ with $fA=A$ is called \defin{closed under} $f$.\\
For $A\sub S$ we define $f^0 A=A$, $f^{\alpha+1}A=f(f^{\alpha}A)$ for all ordinals $\alpha$,
$f^{\beta}A = \Join_{\alpha<\beta} f^{\alpha}A$ for all limit ordinals $\beta$.
\end{defi}
\begin{lem}[fixpoint lemma]\hfill\\
Let $S$ be a set, $f\typ \mathcal{P}(S)\fun\mathcal{P}(S)$ a preclosure\ and $A\sub S$.\\
Let $B$ be the intersection of all $A'\supseteq A$ with $fA'=A'$.\\
Then $B$ is the least set with $A\sub B\sub S$ and $fB=B$.
$B$ is called the \defin{closure} of $A$ under $f$.\\
The map from $A\sub S$ to the closure of $A$ under $f$ is called
the \defin{closure operator} for $f$.\\
It is $B=\Join_{\alpha\text{ ordinal}} f^{\alpha}A$.
Furthermore $B=f^{\gamma}A$ for some ordinal $\gamma$.
\end{lem}
\proof
It is $fB\supseteq B$ because $f$ is increasing.\\
It is $fB\sub fA'\sub A'$ for all $A'\supseteq A$ with $fA'=A'$, therefore $fB\sub B$.\\
Let $C=\Join_{\alpha \text{ ordinal}}f^{\alpha}A$.
We show $C\sub B$, i.e.\ $f^\alpha A\sub B$ for all ordinals $\alpha$, by induction on $\alpha$:\\
It is $f^0 A=A\sub B$.
$f^{\alpha+1}A=f(f^\alpha A)\sub fB=B$.\\
$f^{\beta}A= \Join_{\alpha<\beta}f^\alpha A\sub B$ for all limit ordinals $\beta$.\\
To prove $B\sub C$ we show that $fC=C$:\\
By Hartogs' lemma \cite{Hartogs}\cite[lemma 7.1]{Johnstone}, there is an ordinal $\gamma$
that cannot be mapped injectively into $S$.
Assume $f(f^\gamma A) \neq f^\gamma A$.\\
Then for every element $\alpha$ of $\gamma$, i.e.\ for every ordinal $\alpha<\gamma$,
we have $f(f^\alpha A)\neq f^\alpha A$.\\
If we map each $\alpha<\gamma$ to some new element of $f(f^\alpha A)$ that is not in $f^\alpha A$,
we get an injective map from $\gamma$ into $S$, contradiction.
So it must be $f(f^\gamma A)=f^\gamma A=C$.\\
As $B$ is the least set with $fB=B$ and $A\sub B$, we get $B\sub C$.
\qed
In all our applications of the fixpoint lemma,
the preclosure\ is generated by a rule system on the underlying set $S$:
\begin{defi}[rule system]\label{d:rulesystem}
A \defin{rule system} on a set $S$ is a relation $\mathord{\leadsto}\,\sub \mathcal{P}(S)\pro S$.\\
The elements of $\mathord{\leadsto}$ are called \defin{rules}.
The \defin{preclosure\ of} this rule system is\\
$f\typ \mathcal{P}(S)\fun\mathcal{P}(S)$ defined by
$fA= A\cup \set{a\in S}{\exi{B\sub A} B\mathord{\leadsto} a}$.\\
The closure operator for $f$ is called the \defin{closure operator} of the rule system,
corresponding to \defin{closures under} the rule system.\\
A \defin{deduction of} $a\in S$ \defin{from} $A\sub S$ in this rule system is some $(N,r,\operatorname{\mathit{lab}},\operatorname{\mathit{pre}})$
where $N$ is the set of \defin{nodes},
$r\in N$ is the \defin{root},
$\operatorname{\mathit{lab}}\typ N\fun S$ is the \defin{labelling function},\\
$\operatorname{\mathit{pre}}\typ \set{n\in N}{\operatorname{\mathit{lab}} n\notin A}\fun \mathcal{P}(N)$ is the \defin{predecessor function},\\
such that $\operatorname{\mathit{lab}} r=a$,\\
for all $n\in N$ with $\operatorname{\mathit{lab}} n \notin A$: $\fset{\operatorname{\mathit{lab}}}(\operatorname{\mathit{pre}} n)\mathord{\leadsto} \operatorname{\mathit{lab}} n$,\\
(we say that $\operatorname{\mathit{lab}} n$ is \defin{deduced} from $\fset{\operatorname{\mathit{lab}}}(\operatorname{\mathit{pre}} n)$),\\
for all $n\in N$ there is a unique finite path of length $i\geq 1$:\\
$r=n_1$, $n_2\in \operatorname{\mathit{pre}} n_1$, $n_3\in \operatorname{\mathit{pre}} n_2$,\ldots, $n_i=n$,\\
and for all $n\in N$ there is no infinite path $n=n_1$, $n_2\in \operatorname{\mathit{pre}} n_1$, $n_3\in \operatorname{\mathit{pre}} n_2$,\ldots.
(So such paths end in some $n_i\in N$ with $\operatorname{\mathit{lab}} n_i\in A$.)
If there is a deduction of $a$ from $A$,
then the pair $(A,a)$ is called a \defin{derived rule} of the rule system.
\end{defi}
So the deduction is a tree with root $r$ labelled $a$ and leaves labelled by elements of $A$,
such that each non-leaf node is deduced by a rule from its predecessors.
It has a well-founded structure and we can prove properties of its nodes by induction on this
structure.
\begin{lem}
Let $\mathord{\leadsto}$ be a rule system on a set $S$, and $f$ its preclosure.
Let $A\sub S$ and $a\in S$.\\
There is a deduction of $a$ from $A$ in $\mathord{\leadsto}\ \ifff a$ is in the closure of $A$ under $f$.
\end{lem}
\proof
$\ifthen$:
Let $(N,r,\operatorname{\mathit{lab}},\operatorname{\mathit{pre}})$ be the deduction of $a$ from $A$ in $\mathord{\leadsto}$.
We prove by induction on the deduction that for every node $n\in N$,
$\operatorname{\mathit{lab}} n$ is in the closure $B$ of $A$ under $f$.
This is clear as $A\sub B$ and $fB=B$.
$\thenif$:
We show by induction on $\alpha$ that for every ordinal $\alpha$
and every $a\in f^{\alpha} A$,
there is a deduction of $a$ from $A$.
It is clear for $\alpha=0$ and $\alpha$ a limit ordinal.
For $\alpha=\beta+1$,
$a$ is already in $f^{\beta}A$ or else there is some $C\sub f^\beta A$ with $C\mathord{\leadsto} a$.
In the second case, we build the deduction with root $r$ and $\operatorname{\mathit{lab}} r=a$
and connect $r$ to all the deductions of the elements of $C$ as predecessors.
\qed
\section{Sazonov's definition of natural domains revisited}\label{s:sazonov}
In this section we repeat Sazonov's definition of natural domains, and
give a simpler equivalent axiom system.
We also give an example that shows
that natural lubs are needed for general subsets,
not only directed ones.
Sazonov's natural domains \cite{Sazonov:natural} are a general extrapolation of
the domains of the game model of PCF.
They are just partial orders with a designation of some of the (general) subsets
with lubs as ``natural lubs''.
The natural lubs of directed sets must be respected by the naturally continuous functions.
The natural lubs have to obey certain closure conditions which are
taylored to ensure cartesian closedness of the resulting category under a certain
function space.
We translate the definition into a new notation:
instead of the partial natural-lub-operator $\biguplus$ we use a relation
symbol $\to$ for natural convergence.
\begin{defi}[Sazonov, def. 2.1 of \cite{Sazonov:natural}] $\phantom{X}$\\
A structure $D=(|D|,\lex_D,\nlub_D)$ is a \defin{natural domain} if\\
$|D|$ is a set of \defin{elements},\\
$\lex_D$ is a partial order on $|D|$,\\
$\nlub_D\; \sub \Pot^\uparrow(|D|)\pro |D|$ is the \defin{(natural) convergence relation} on directed sets,\\
($X\nlub_D x$ means: $X$ has the \defin{natural lub} $x$,
such an $X$ is called \defin{natural})\\
such that $\nlub_D$ can be extended by pairs $(X,x)$
with $X\sub|D|$ \emph{not} directed, $x\in |D|$, to a relation on all subsets,
which we also denote $\nlub_D$,
and the following axioms are fulfilled for the extended relation
(where axiom Sx corresponds to $\biguplus$x in Sazonov):\\
\textbf{(S1)} If $X\to x$, then $\Lubex X = x$.\\
\textbf{(S2)} If $X\sub Y \sub |D|$, $X\to x$ and $Y\lex x$, then $Y\to x$.\\
\textbf{(S3 singleton)} For all $x\in |D|$: $\{x\}\to x$.\\
\textbf{(S4(1))} If $\{\yij\}_{i\in I\!, j\in J}$ is a non-empty two-parametric family of elements in $|D|$,\\
and for every $i\in I$ it is $\{\yij\}_{j\in J} \to d_i$,\\
then $\{d_i\}_{i\in I} \to d \ifff \{\yij\}_{i\in I\!, j\in J}\to d$.\\
\textbf{(S4(2))} If $I$ is an index set with a directed partial order, and $\{\yij\}_{i,j\in I}$ is a two-parametric family
of elements in $|D|$ that is monotonic in each parameter $i$ and $j$,\\
then $\{\yij\}_{i,j\in I} \to y \ifff \{y_{ii}\}_{i\in I} \to y$.
\end{defi}
Note that we deviate from Sazonov's definition in that we take only
the part of the convergence relation on \emph{directed} subsets as the data of the structure,
and leave the existence of the extended relation as a side condition.
This is justified by the morphisms of the category,
the (naturally) continuous\ functions,
which must be continuous\ only w.r.t.\ the \emph{directed} natural lubs.
We get our natural domains by identifying Sazonov's natural domains
which have the same order and the same \emph{directed} natural lubs.
The result is an equivalent category.
We make this change because we define all other related structures of
incomplete domains by convergence of directed sets only.
The specification of the lub in the conclusions of the axioms is redundant,
it suffices to specify that the set in question is natural,
e.g.\ in axiom (S2): $Y$ is natural instead of $Y\to x$.
We do not count S1 as a proper axiom,
instead we treat it as a condition for the relation $\to$
that is always presupposed in the sequel.
The other proper axioms have a different character,
they deduce the naturalness of lubs from existing natural lubs.
We have translated the axiom system into a form
that clearly discerns its different parts.
It needs a clean-up and simplification.
First, we can further split axiom S4(1) into the two axioms
S4(1$\Rightarrow$) and S4(1$\Leftarrow$),
where the logical equivalence $\ifff$ is replaced by the indicated direction.
\begin{prop}
In the presence of axiom S3 (singleton):
axioms S2 and S4(1$\Leftarrow$) together are equivalent to the axiom:\\
\textup{\textbf{(S6 cofinality)}} If $X,Y\sub |D|$, $X\to x$ and $X\sqsubseteq Y\sqsubseteq x$, then $Y\to x$.
\end{prop}
\proof
S2 and S4(1$\Leftarrow$) $\ifthen$ S6:\\
Let $X\to x$ and $X\lex Y\lex x$.\\
Let $I=\set{(a,b)}{a\in X, b\in Y\text{ and }a\lex b}$ and $J=\{1,2\}$.\\
For $(a,b)\in I$ define $y_{(a,b)1}=a$ and $y_{(a,b)2}=b$ and $d_{(a,b)}=b$.\\
For every $(a,b)\in I$ it is $\{y_{(a,b)j}\}_{j\in J} \to d_{(a,b)}$, by axiom S3 and S2.\\
It is $\{y_{ij}\}_{i\in I,j\in J} \to x$ by axiom S2,
as $X\sub \{\yij\}_{i\in I\!, j\in J}$ by $X\sqsubseteq Y$.\\
We apply axiom S4(1$\Leftarrow$) and deduce $\{d_{(a,b)}\}_{(a,b)\in I} \to x$.\\
This set is a subset of $Y$, therefore by axiom S2 we get $Y\to x$.\\
The reverse direction is immediate.
\qed
\begin{prop}
Axiom S4(2) is redundant, it follows from axiom S6 (cofinality).
\qed
\end{prop}
Sazonov proposes an ``optional clause, which might be postulated as well'',
as a replacement for axiom S4(2):\\[\medskipamount]
\textbf{(S5)} If $X\sub Y\sub |D|$, $Y\to y$ and $Y\sqsubseteq X$, then $X\to y$.\\[\medskipamount]
This does not lead to a stronger axiom system:
\begin{prop}
Axioms S2 and S5 together are equivalent to axiom S6 (cofinality).
\end{prop}
\proof
S2 and S5 $\ifthen$ S6:\\
It is $X\sub X\cup Y\sub |D|$, $X\to x$, $X\cup Y\lex x$, so $X\cup Y\to x$ by axiom S2.\\
It is $Y\sub X\cup Y\sub|D|$, $X\cup Y\to x$, $X\cup Y\sqsubseteq Y$, so $Y\to x$ by axiom S5.\\
The reverse direction is immediate.
\qed
Next, we want to give axiom S4(1$\Rightarrow$) in a form with sets instead of two-parametric families,
to get rid of the index sets of unlimited size:
\begin{prop}
In the presence of axioms S3 (singleton) and S6 (cofinality), axiom S4(1$\Rightarrow$) is equivalent to the axiom:\\
\textup{\textbf{(S7 transitivity)}}
If $\overline{X}\sub \mathcal{P}(|D|)$ with $X\to d_X$ for all $X\in \overline{X}$,\\
then $\set{d_X}{X\in \overline{X}} \to d \ifthen \Join \overline{X} \to d$.
\end{prop}
\proof
S4(1$\Rightarrow$) follows immediately from S7.\\
For the reverse direction, we prove S7 from S4(1$\Rightarrow$).\\
First, we have to prove the two special cases (1) $\overline{X} = \emptyset$ and (2) $\overline{X}=\{\emptyset\}$.\\
(1) The conclusion of S7 reads $\emptyset\to d \ifthen \emptyset\to d$.\\
(2) It is $\emptyset\to d_{\emptyset}$ and the conclusion reads $\{d_{\emptyset}\}\to d \ifthen
\emptyset\to d$.\\
Now assume we do not have case (1) or (2).\\
Take $I=\overline{X}$ and $J=\Join\overline{X}$.\\
Both $I$ and $J$ are not empty, as is required by S4(1$\Rightarrow$). We define $y_{ij}$:\\
If $i\in I=\overline{X}$ is not empty,
then for all $j\in J$ take $y_{ij}=j$ if $j\in i$ and else some arbitrary element of $i$.\\
If $i\in I$ is empty, then for all $j\in J$ take $y_{ij} = \Lubex \emptyset$ (the least element),
this exists as $i=\emptyset \to d_{\emptyset} = \Lubex \emptyset$.\\
Then for all nonempty $i\in I$ it is $i\sub\{\yij\}_{j\in J}$ and $\{\yij\}_{j\in J}\sqsubseteq d_i$,\\
so by axiom S6 it is $\{\yij\}_{j\in J}\to d_i$.\\
If $i\in I$ is empty,
then it is $\{\yij\}_{j\in J} = \{\Lubex \emptyset\}\to \Lubex \emptyset$ by axiom S3.\\
So in every case it is $\{\yij\}_{j\in J}\to d_i$,
and we can draw the conclusion of S4(1$\Rightarrow$).\\
Furthermore, if not $\emptyset \in I$,
then $\{\yij\}_{i\in I\!, j\in J} = \Join\overline{X}$, and the conclusion of S7 follows.\\
If $\emptyset\in I$, then $\{\yij\}_{i\in I\!, j\in J}= \Join\overline{X}\cup\{\Lubex\emptyset\}$.\\
From $\{\yij\}_{i\in I\!, j\in J}\to d$ follows $\Join\overline{X}\to d$ by axiom S6.
This is the conclusion of S7.
\qed
In summary, we get a nice simpler axiom system as a base for the rest of the paper:
\begin{prop}\label{p:sazonov}
The axioms of the Sazonov natural domain are equivalent to
S1, S3 (singleton), S6 (cofinality) and S7 (transitivity).
\qed
\end{prop}
We can further combine axioms S6 and S7, for this we need a new definition:
\begin{defi}\label{d:under}
Let $D$ be the structure in the beginning of the definition of the Sazonov natural domain,
and $A,B\sub|D|$.\\
$A$ is \defin{under} $B$, $A\stackrel{\sqsubset}{\leftarrow}_D B$, if for all $a\in A$:\\
there is $b\in B$ with $a\lex_D b$ or
there is $B'\sub B$ with $B'\nlub_D a$.
\end{defi}
\begin{prop}\label{p:under}
In the presence of axiom S3 (singleton):
axioms S6 (cofinality) and S7 (transitivity) together are equivalent to the axiom:\\
\textup{\textbf{(S8 under)}}
If $A,B\sub|D|$, $A\to a$ and $A\stackrel{\sqsubset}{\leftarrow} B \sqsubseteq a$, then $B\to a$.
\end{prop}
\proof
S6 and S7 $\ifthen$ S8:\\
Let $\varphi\typ A\fun \mathcal{P}(|D|)$ be defined as:\\
$\varphi x =\{x\}$ if there is $y\in B$ with $x\lex y$,\\
otherwise take a choice of some $B'\sub B$ with $B'\to x$ and define $\varphi x = B'$.\\
Let $C=\Join(\fset{\varphi} A)$.
By the singleton axiom S3 and transitivity axiom S7 it is $C\to a$.\\
It is $C\sqsubseteq B\sqsubseteq a$, therefore $B\to a$ by the cofinality axiom S6.\\
The reverse direction is immediate.
\qed
So the axioms of the Sazonov natural domain are equivalent to S1, S3 and S8.\\
The proper axioms share a common form:
They are rules (in the sense of rule systems, definition \ref{d:rulesystem})
that deduce from the existence of the lubs of some sets
the existence of the lub of another set,
all in some configuration of a partial order.
One can get the impression that the axioms are ``complete'' in this respect,
that all ``necessary'' deductions can be made.
In section \ref{s:lubrule} we set up the general frame of lub-rule systems
where we can define what this completeness means.
It will turn out in section \ref{s:incomplete} that the axioms are in fact incomplete.
So we will have to replace them by a stronger axiom in order to get
a class of domains that can be realized by dcpos.
We now come to the morphisms of the category of natural domains.
\begin{defi}[Sazonov def. 2.3(a) in \cite{Sazonov:natural}] \hfill\\
Let $D,E$ be natural domains.\\
A function $f\typ |D|\fun|E|$ is \defin{(naturally) continuous}
if it is monotonic w.r.t.\ $\lex_D$ and $\lex_E$,
and if $X$ is directed and $X\nlub_D x$, then $\fset{f}X\nlub_E fx$.
(Monotonicity is here redundant.)\\
If $f$ is naturally continuous, we write $f\typ D\fun E$.
\end{defi}
In \cite{Sazonov:natural}\ it is shown that the natural domains with naturally continuous functions
form a ccc.\\
The terminal object is the one point domain $\{\{\bot\},\bot\lex \bot,\{\bot\}\to \bot\}$.\\
The product $D\pro E$ has $\elem{D\pro E}=|D|\pro|E|$,
$\lex_{D\pro E}$ is component-wise,
and $A\to_{D\pro E} a$ iff $\fset{\proi} A\nlub_D \proi a$ and $\fset{\prot} A\nlub_E \prot a$.\\
The exponent $D\mathbin{\Rightarrow} E$ has $\elem{D\mathbin{\Rightarrow} E}$ the set of naturally continuous functions,
$\lex_{D\mathbin{\Rightarrow} E}$ is pointwise,
and $F\to_{D\mathbin{\Rightarrow} E} f$ iff $Fx\nlub_E fx$ for all $x\in |D|$.
As we remarked above, we take only the part of the convergence relation on directed sets
as the data of a natural domain,
and leave the existence of the extended relation as a side condition.
The following example shows a case where natural lubs of general subsets are needed
as intermediary steps in the deduction of the natural lubs of directed subsets,
so that the extended relation is really needed.
\begin{figure}[htb]
\begin{pspicture}(-0.7,-0.2)(2.0,6.5)
\firsto{b11}{180}{b_{11}}
\slineo{b11}{b12}{90}{1}{180}{b_{12}}
\dlineon{b12}{b1}{90}{1}{180}{b_1}
\nexto{b1}{b21}{90}{1}{180}{b_{21}}
\slineo{b21}{b22}{90}{1}{180}{b_{22}}
\dlineon{b22}{b2}{90}{1}{180}{b_2}
\slineo{b1}{a1}{337.5}{1.5}{0}{a_1}
\slineo{b2}{a2}{337.5}{1.5}{0}{a_2}
\sline{a1}{a2}
\dlineon{a2}{a}{90}{1.5}{0}{a}
\sline{b1}{a}
\sline{b2}{a}
\end{pspicture}
\caption{Natural domain $E$ (partial)}\label{f:weakE}
\end{figure}
Figure \ref{f:weakE} shows only part of the natural domain $E$.
It first has
the elements $a$, $a_i$ and $b_i$ for $i\geq 1$.
The order is $a_i\lex a_j$ iff $i\leq j$,
$a_i\lex a$, $a_i\lex b_i$ and $b_i\lex a$ for $i\geq 1$.
And $\{a_i\}_{i\geq 1}\to a$ is natural.
It has further the elements of $B=\{b_{ij}\}_{i,j\geq 1}$ with ($b_{ij}\lex b_{ik}$
iff $j\leq k$) and $b_{ij}\lex b_i$ for all $i\geq 1$.
It is stipulated that the $\{b_{ij}\}_{j\geq 1}\to b_i$ are natural.
Moreover, we erect an ``artificial'' directed set over the elements of $B$,
whose elements are not depicted in the diagram:
every finite subset $X\sub B$ is an element of $E$,
the set of these elements is called $\overline{B}$.
For all $i,j$ it is defined $b_{ij}\lex \{b_{ij}\}$.
For $X,Y\in \overline{B}$ it is defined $X\lex_E a$ and ($X\lex_E Y$ iff $X\sub Y$).
$E$ is defined to be the natural domain that is completed by all conclusions of the
partial order axioms and the axioms of naturality.
$\overline{B}$ is a directed subset of $|E|$ with the lub $a$.
From $\{a_i\}_{i\geq 1}\to a$ we deduce $\{b_i\}_{i\geq 1}\to a$ by axiom S6,
then $B\to a$ by axiom S7,
then $\overline{B}\to a$ by axiom S6.
Here $\{b_i\}_{i\geq 1}\to a$ and $B\to a$ are intermediary steps that deduce the naturality of non-directed subsets.
There is no deduction of the directed natural lub $\overline{B}\to a$
from the given directed natural lubs that does not use a
non-directed intermediary step.
\section{Directed-lub partial orders (dlubpo)}
Here we explore categories larger than the category of natural domains.
We want to find a ``minimal part'' of the natural domain axioms
that already achieves cartesian closedness.
Less axioms means more chaos.
We always presuppose axiom S1, which we do not count as proper axiom.
When there is no other axiom valid,
we get the normal terminal object $\top$,
but we do not get all constant element morphisms $\la x.d\typ \top \fun D$, $d\in D$
some object.
So in order to constrain chaos,
we start with the ``most obvious'' axiom, the singleton axiom S3,
and get the category $\mathbf{Dlubpo}$ of directed-lub partial orders.
These are partial orders with a designation of some lubs of \emph{directed} subsets as natural.
This category has the expected terminal object, products
and the constant element morphisms.
We prove that the exponents do not always exist;
and if they exist, they are generally defined differently from those in the category of
natural domains.
(But the natural domains,
with their \emph{directed} natural subsets as data, are a full sub-ccc of $\mathbf{Dlubpo}$
such that their exponents coincide with the general exponents of $\mathbf{Dlubpo}$.)
We also give sufficient conditions on subcategories of $\mathbf{Dlubpo}$
to have products and exponents like the normal ones,
in analogy to lemma 5 of \cite{Smyth}.
We show that the dlubpos fulfilling axiom S5 (for directed sets) already form a ccc.
For a first understanding of the following sections only definitions 4.1 to 4.4
are necessary.
The rest of this section may be skipped,
although definition \ref{d:genfun},
propositions \ref{p:evalstar} and \ref{p:expo1}(1) will be used in the following sections.
\begin{defi}
A structure $D=(|D|,\lex_D,\nlub_D)$ is a \defin{directed-lub partial order (dlubpo)} if\\
$|D|$ is a set of \defin{elements} of $D$,\\
$\lex_D$ is a partial order on $|D|$, lubs are denoted by $\Lubex_D$,\\
$\nlub_D\; \sub \Pot^\uparrow(|D|,\lex_D)\pro|D|$ is the \defin{(natural) convergence relation},\\
if $A\nlub_D a$, then $A$ is a directed subset of $|D|$,
and $a\in |D|$ is its lub,\\
($a$ is called the \defin{natural lub} of $A$, $A$ is called a \defin{natural} (directed) subset),\\
$\nlub_D$ fulfills the \defin{singleton axiom} S3:
For every $d\in |D|$ it is $\{d\}\nlub_D d$.
\end{defi}
\begin{defi}
Let $D,E$ be dlubpos.\\
A function $f\typ|D|\to|E|$ is \defin{continuous} if it is monotonic w.r.t.\ $\lex_D$ and $\lex_E$
and if $X\nlub_D x$, then $\fset{f}X\nlub_E fx$.
If $f$ is continuous, we write $f\typ D\fun E$.\\
The dlubpos with continuous functions form a category $\mathbf{Dlubpo}$,
with normal function composition and identity functions.
\end{defi}
Note that Sazonov calls such a function ``naturally continuous'',
which we abbreviate to ``continuous'',
as it is clear from the context that $D,E$ are dlubpos.
\begin{defi}
The dlubpo $\top=(\{\bot\},\bot\lex\bot, \{\bot\}\to \bot)$
is called the \defin{terminal dlubpo}.\\
$\top$ is the categorical terminal object in $\mathbf{Dlubpo}$.
It need not be so in subcategories.
If the subcategory has $\top$ as terminal object,
we say that it has the normal terminal.
\end{defi}
\begin{defi}
Let $D,E$ be dlubpos.\\
We define the \defin{product dlubpo} $D\pro E = (|D\pro E|, \lex_{D\pro E}, \to_{D\pro E})$
as follows:\\
$|D\pro E| = |D| \pro |E|$,\\
$(a,b)\lex_{D\pro E} (a',b')$ iff $a\lex_D a'$ and $b\lex_E b'$,\\
if $A\sub|D\pro E|$ is directed and $(a,b)\in |D\pro E|$,
then $A\to_{D\pro E} (a,b)$ iff $\fset{\proi} A \nlub_D a$ and $\fset{\prot} A \nlub_E b$,
where we have the normal \defin{projections}
$\proi\typ |D\pro E|\fun |D|$ and
$\prot\typ |D\pro E|\fun |E|$.\\
(It is clear that $(a,b)$ is the lub of $A$.)\\
We also have the normal pairing functions:\\
For any dlubpo $C$, $f\typ C\fun D$ and $g\typ C\fun E$
it is $\lpro f,g\rpro\typ C\fun D\pro E$, $\lpro f,g\rpro x = (fx,gx)$.\\
All these functions are continuous,
and $D\pro E$ with $\proi,\prot$ is the categorical product in $\mathbf{Dlubpo}$.
It need not be so in subcategories.
If the subcategory has the $D\pro E$ as categorical products,
we say it has the normal products.
\end{defi}
\begin{defi}\label{d:genfun}
Let $\mathbf{K}$ be a subcategory of $\mathbf{Dlubpo}$, and $D,E \in \mathbf{K}$.\\
We define (polymorphically) the function
$\eval\typ \set{f}{f\typ D\fun E \text{ in } \mathbf{K}}\pro |D|\fun|E|$ by $\eval(f,d)=fd$.\\
We define the \defin{general function dlubpo} $D\expo_{\Kcat}^{\ast}E$ (relative to $\mathbf{K}$) by:\\
$|D\expo_{\Kcat}^{\ast}E| = \set{f}{f\typ D\fun E \text{ in } \mathbf{K}}$,\\
$f\lex_{D\expo_{\Kcat}^{\ast}E} g$ iff for all $x\in |D|$ it is $fx\lex_E gx$,\\
if $F\sub |D\expo_{\Kcat}^{\ast}E|$ is directed and $f\in |D\expo_{\Kcat}^{\ast}E|$, then
$F\to_{D\expo_{\Kcat}^{\ast}E} f$ iff
for all $A\sub|D\expo_{\Kcat}^{\ast}E|\pro |D|$ that are directed w.r.t.\ the component-wise order
with $\fset{\proi} A=F$ and $\fset{\prot} A \to d$ it is $\fset{\eval} A\nlub_E fd$.\\
For $\mathbf{K}= \mathbf{Dlubpo}$ we define $(D\expo^{\ast}E) = (D\mathbin{\Rightarrow}_{\mathbf{Dlubpo}}^{\ast} E)$.\\
We also have the curried functions:
For any dlubpo $C$, $f\typ C\pro D\fun E$ in $\mathbf{Dlubpo}$ we define
$\operatorname{curry}(f)\typ |C|\fun |D\expo^{\ast}E|$ by $\operatorname{curry}(f)c= \la d.f(c,d)$.
\end{defi}
We have to prove that in the definition $\Lubex F =f$:
for all $d\in|D|$, take $A=F\pro\{d\}$,
then $\fset{\eval} A\to fd$,
so $f$ is the pointwise lub of $F$.\\
$D\expo_{\Kcat}^{\ast}E$ fulfills the singleton axiom:
for $f\in |D\expo_{\Kcat}^{\ast}E|$ it is $\{f\}\to f$ because $f$ is continuous.
The function $\eval$ is continuous.
The trick of the definition is that not only the continuity in constant second arguments
(as in the case of natural domains),
but the whole continuity of $\eval$ is coded in the definition of $\to_{D\expo_{\Kcat}^{\ast}E}$.
We will see that if $\mathbf{Dlubpo}$ contains an exponent of $D$ and $E$,
then it is isomorphic to $D\expo^{\ast}E$.
\begin{prop}\label{p:evalstar}
$\eval\typ(D\expo_{\Kcat}^{\ast}E)\pro D\fun E$ is continuous.\\
The results of $\operatorname{curry}(f)$, $\operatorname{curry}(f)c$ for all $c\in |C|$, are continuous.\\
$\operatorname{curry}(f)$ is monotonic, but need not be continuous.
(We will see a counter-example below.)
\qed
\end{prop}
We will now give sufficient conditions for subcategories of $\mathbf{Dlubpo}$ to contain terminal objects,
products and exponents that coincide partially with the normal terminals, products
and the general function dlubpos.
These conditions are extracted from the corresponding results of \cite[lemma 5]{Smyth}
for full subcategories of the category of $\omega$-algebraic cpos.
Those results apply generally to partial order structures,
and they achieve full coincidence, i.e.\ isomorphism,
only for the order structure; while in our case the isomorphism does not extend
to the additional structure of natural convergence.
The natural convergence of a dlubpo can be chosen more arbitrarily.
The only way to achieve full isomorphism in our case seems to presuppose the normal structure
in the subcategory (or a set of structures that all together force the normal structure
on the categorical object in question),
which we will do afterwards.
Smyth's results are for full subcategories,
while we give them for general subcategories,
by extracting the needed morphisms from Smyth's proofs as conditions in the propositions.
The proofs are adaptations of Smyth's proofs.
We give the proofs mainly for those results that we need in the sequel.
\begin{prop}
Let $\mathbf{K}$ be a subcategory of $\mathbf{Dlubpo}$.
If $\mathbf{K}$ contains some object that is not empty,
and $\mathbf{K}$ contains a terminal object $D$ and all constant morphisms $\la x.a\typ D\fun D$
for $a\in |D|$,
then $D$ is isomorphic to the terminal dlubpo $\top=\{\{\bot\},\bot\lex\bot,\{\bot\}\to\bot\}$.
\end{prop}
\proof
$D$ cannot have two different elements,
since this would entail two different constant morphisms $D\fun D$.
$D$ cannot be empty, as there is a non-empty object.
Thus $D$ has just one element $a$,
and it must be $\{a\}\to a$ by the singleton axiom.
\qed
\begin{prop}[Smyth, lemma 5(ii) in \cite{Smyth}, also prop.\ 5.2.17(2) in \cite{Amadio/Curien}]\hfill\\
Let $\mathbf{K}$ be a subcategory of $\mathbf{Dlubpo}$ with the normal terminal object $\top$.\\
Let $D,E\in\mathbf{K}$ and $D\pro_\Kcat E$ be their categorical product in $\mathbf{K}$,
with the projections $\pi_{1\Kcat},\pi_{2\Kcat}$ and pairing functions $\lpro f,g \rpro_\mathbf{K}$.\\
Let $\mathbf{K}$ have all constant functions $\la x.a\typ \top\fun D$, for $a\in |D|$,
and $\la x.a\typ \top\fun E$, for $a\in |E|$.\\
(As $D,E$ fulfill the singleton axiom, all these functions are sure to be continuous,
so are morphisms in every \emph{full} subcategory $\mathbf{K}$.)
\begin{enumerate}[(1)]
\item The function $\varphi\typ D \pro_\Kcat E\fun D\pro E$, $\varphi=\lpro\pi_{1\Kcat},\pi_{2\Kcat}\rpro$,
is bijective.
(And of course it is continuous, so a morphism in $\mathbf{Dlubpo}$.)\\
The translation $\varphi$ respects projections and pairing functions:\\
$\proi\comp \varphi= \pi_{1\Kcat}$,
$\prot\comp \varphi= \pi_{2\Kcat}$,
$\varphi\comp\lpro f,g\rpro_{\mathbf{K}}=\lpro f,g\rpro$.
\item
Let, furthermore, some $C\in \mathbf{K}$ with some $c,c'\in|C|$, $c\lex c'$ and $c\neq c'$,
and for all $d\lex d'$ in $D$
the function $f\typ C\fun D$ be in $\mathbf{K}$ with $fx=d$ for $x\lex c$ and $fx=d'$ else,
and likewise for all $e\lex e'$ in $E$
the function $f\typ C\fun E$ with $fx=e$ for $x\lex c$ and $fx=e'$ else.\\
(As to the existence of these functions in \emph{full} subcategories, the remark above
applies again.)\\
Then $\varphi^{-1}$ is monotonic, so that $\varphi$ is an order-isomorphism.
$\varphi^{-1}$ is continuous in each component of its argument separately.
(But it need not be continuous.)
\end{enumerate}
\end{prop}
\proof is an adaptation of the proof of part (ii), lemma 5 of \cite{Smyth}.
\qed
\begin{prop}[Smyth, lemma 5(iii) in \cite{Smyth}, also prop.\ 5.2.17(3) in \cite{Amadio/Curien}]\label{p:expo1}
\hfill\\
Let $\mathbf{K}$ be a subcategory of $\mathbf{Dlubpo}$ with the normal terminal object $\top$
and all normal products $D\pro E$.\\
Let $D,E\in \mathbf{K}$ and $D\expo_{\Kcat}E$ be their categorical exponent in $\mathbf{K}$,
with the evaluation\\
$\eval_\Kcat \typ (D\expo_{\Kcat}E)\pro D\fun E$
and curried morphisms $\curry_\Kcat(f)$.\\
Let $\mathbf{K}$ have all constant functions $\la x.f\typ \top\fun(D\expo_{\Kcat}E)$,
for $f\in D\expo_{\Kcat}E$.\\
(As $D\expo_{\Kcat}E$ fulfills the singleton axiom,
all these functions are sure to be continuous, so are morphisms in every \emph{full} subcategory $\mathbf{K}$.)
\begin{enumerate}[(1)]
\item
The function $\varphi\typ (D\expo_{\Kcat}E)\fun (D\expo_{\Kcat}^{\ast}E)$, $\varphi=\operatorname{curry}(\eval_\Kcat)$
is bijective and monotonic.
(But need not be continuous.)\\
The translation $\varphi$ respects evaluation and curried functions:\\
$\eval(\varphi a,d)= \eval_\Kcat(a,d)$, $\varphi\comp\curry_\Kcat(f)=\operatorname{curry}(f)$.
\item
Let, furthermore, some $C\in \mathbf{K}$ with some $c,c'\in |C|$, $c\lex c'$ and $c\neq c'$,
and for all $f_1\lex f_2$ in $D\expo_{\Kcat}^{\ast}E$
the function $h\typ C\pro D\fun E$ be in $\mathbf{K}$ with
$h(u,x)=f_1 x$ for $u\lex c$ and $h(u,x)=f_2 x$ else.
Let $g_1,g_2\typ \top \fun C$ be in $\mathbf{K}$ with $g_1\bot = c$ and $g_2 \bot=c'$.\\
Then $\varphi^{-1}$ is monotonic, so that $\varphi$ is an order isomorphism.
(But need not be a $\mathbf{Dlubpo}$-isomorphism.)
\end{enumerate}
\end{prop}
\proof
\textbf{(1)}
The proof is a detailed version of the first part of the proof of part (iii), lemma 5 of \cite{Smyth}.
\begin{pspicture}(-2,-0.5)(7,3)
\firstn{DE}{D\expo_{\Kcat}E}
\nextn{DE}{xk}{0}{0.7}{\pro_\Kcat}
\nextn{xk}{D1}{0}{0.5}{D}
\nextn{DE}{T}{90}{2}{\top}
\nextn{xk}{x}{90}{2}{\pro}
\nextn{D1}{D2}{90}{2}{D}
\nextn{D2}{E}{0}{2}{E}
\ncline{->}{T}{DE}\Bput{$\curry_\Kcat(f')$}
\ncline{->}{D2}{D1}\Aput{$\id$}
\ncline{->}{D2}{E}\Aput{$f'$}
\ncline{->}{D1}{E}\Bput{$\eval_\Kcat$}
\end{pspicture}
$\varphi$ is \textbf{surjective}:\\
Let $f\in D\expo_{\Kcat}^{\ast}E$.\\
Let $f'\typ\top\pro D\fun E$, with $f'y=f(\prot y)$, $f'=f\comp \prot$ is in $\mathbf{K}$.\\
Let $h=(\curry_\Kcat(f'))\bot \in (D\expo_{\Kcat}E)$. We calculate:
\begin{align*}
(\varphi h)x &= \eval_\Kcat (h,x)\\
&= \eval_\Kcat((\curry_\Kcat(f'))\bot,x)\\
&= f'(\bot,x)\\
&= fx.
\end{align*}
So $\varphi h=f$.
$\varphi$ is \textbf{injective}:\\
Let $h'\in D\expo_{\Kcat}E$ with $\varphi h'=f$.
We have to show that $h'=h$ above.\\
Let $g\typ \top \fun (D\expo_{\Kcat}E)$ with $g\bot=h'$, $g$ is in $\mathbf{K}$ by hypothesis.\\
We show that $\eval_\Kcat\comp(g\pro \id)=f'$:
\begin{align*}
f'y &= f(\prot y)\\
&=(\varphi h')(\prot y)\\
&=\eval_\Kcat(h',\prot y)\\
&=(\eval_\Kcat\comp(g\pro \id))y.
\end{align*}
By the uniqueness of $\curry_\Kcat(f')$ in $\mathbf{K}$,
it must be $g=\curry_\Kcat(f')$.\\
Therefore $h'=g\bot=\curry_\Kcat(f')\bot =h$, as desired.\\
\textbf{(2)}
The proof is an adaptation of the second part of the proof of part (iii), lemma 5 of \cite{Smyth}.
\qed
We now come to situations, where a subcategory $\mathbf{K}$ of $\mathbf{Dlubpo}$ has a categorical
(terminal, product, exponent) object $D$ and also the normal (terminal, product, general function)
dlubpo $D'$ with a certain morphism $\varphi\typ D\fun D'$,
and where this implies that $\varphi$ is an isomorphism.
For the terminal and product we can give two propositions where we generalize
from $\mathbf{Dlubpo}$ to an arbitrary category $\mathbf{C}$.
\begin{prop}
Let $\mathbf{C}$ be a category and $\mathbf{K}$ a subcategory of $\mathbf{C}$.
\begin{enumerate}[(1)]
\item
If $\mathbf{C}$ has a terminal object $\top$, and $\mathbf{K}$ a terminal object $\top_{\mathbf{K}}$,
and $\mathbf{K}$ contains $\top$ and a morphism $\varphi\typ \top_{\mathbf{K}}\fun \top$,
then $\varphi$ and the unique (in $\mathbf{K}$) $\psi\typ \top \fun \top_{\mathbf{K}}$
are an isomorphism in $\mathbf{K}$.
\item
Let $D,E\in \mathbf{K}$.
If $\mathbf{C}$ has a product $D\pro E$, with $\proi,\prot,\lpro\rpro$,
and $\mathbf{K}$ a product $D\pro_\Kcat E$, with $\pi_{1\Kcat},\pi_{2\Kcat},\lpro\rpro_{\mathbf{K}}$,
and $\mathbf{K}$ contains $D\pro E$, $\proi,\prot$ and the morphism
$\varphi\typ D\pro_\Kcat E\fun D\pro E$, $\varphi=\lpro \pi_{1\Kcat},\pi_{2\Kcat}\rpro$,\\
then $\varphi$ and $\psi\typ D\pro E\fun D\pro_\Kcat E$, $\psi=\lpro\proi,\prot\rpro_{\mathbf{K}}$,
are an isomorphism in $\mathbf{K}$.\\
The translation $\varphi$ respects projections and pairing functions.
\end{enumerate}
\end{prop}
\proof
The proofs are easy and like the known proofs of the isomorphism of terminal objects
resp.\ products in a single category.
\qed
The situation for exponents is special and more complicated:
\begin{prop}\label{p:expo2}
Let $\mathbf{K}$ be a subcategory of $\mathbf{Dlubpo}$ with the normal terminal object $\top$
and all normal categorical products $\pro$. Let $D,E\in \mathbf{K}$.\\
Let $\mathbf{K}$ have a categorical exponent $D\expo_{\Kcat}E$ of $D,E$ with evaluation
$\eval_\Kcat$, curried morphisms $\curry_\Kcat(f)$
and all constant functions $\la x.f\typ \top\fun(D\expo_{\Kcat}E)$, for $f\in |D\expo_{\Kcat}E|$.\\
If $\mathbf{K}$ also contains the general function dlubpo (relative to $\mathbf{K}$) $D\expo_{\Kcat}^{\ast}E$
with the morphism $\eval$,
then\\
$\varphi\typ (D\expo_{\Kcat}E)\fun(D\expo_{\Kcat}^{\ast}E)$, $\varphi=\operatorname{curry}(\eval_\Kcat)$ and\\
$\psi\typ (D\expo_{\Kcat}^{\ast}E)\fun(D\expo_{\Kcat}E)$, $\psi=\curry_\Kcat(\eval)$ are an isomorphism in $\mathbf{Dlubpo}$.\\
$\psi$ is, of course, in $\mathbf{K}$.
If $\varphi$ is also in $\mathbf{K}$, then the isomorphism is also in $\mathbf{K}$.\\
The translation $\varphi$ respects evaluation and curried functions:\\
$\eval(\varphi a,d)= \eval_\Kcat(a,d)$, $\varphi\comp\curry_\Kcat(f)=\operatorname{curry}(f)$.
\end{prop}
\proof
By proposition \ref{p:expo1}(1), $\varphi$ is bijective and monotonic.
We have to prove that $\varphi$ is also continuous,
and that $\psi$ is the inverse of $\varphi$.
\begin{pspicture}(-2,-0.5)(7,3)
\firstn{DE}{D\expo_{\Kcat}^{\ast}E}
\nextn{DE}{xk}{0}{0.7}{\pro}
\nextn{xk}{D1}{0}{0.5}{D}
\nextn{DE}{T}{90}{1.9}{D\expo_{\Kcat}E}
\nextn{xk}{x}{90}{2}{\pro}
\nextn{D1}{D2}{90}{2}{D}
\nextn{D2}{E}{0}{2}{E}
\ncarc[arcangle=20]{<-}{T}{DE}\Aput{$\psi$}
\ncarc[arcangle=20]{<-}{DE}{T}\Aput{$\varphi$}
\ncline{<->}{D2}{D1}\Aput{$\id$}
\ncline{->}{D2}{E}\Aput{$\eval_\Kcat$}
\ncline{->}{D1}{E}\Bput{$\eval$}
\end{pspicture}
For all $f\in D\expo_{\Kcat}^{\ast}E$, $d\in|D|$,
it is
\begin{align*}
\eval_\Kcat((\psi f),d) &= \eval_\Kcat(\curry_\Kcat(\eval)f,d)\\
&= \eval(f,d).\\
\text{Then } \varphi(\psi f) &= \la d\in D. \eval_\Kcat(\psi f,d)\\
&= \la d\in D. \eval(f,d) = f.
\end{align*}
Therefore $\psi$ is the inverse of $\varphi$, as $\varphi$ is bijective.
To show that $\varphi$ is continuous, let $A\to a$ in $D\expo_{\Kcat}E$.
We have to show that $\fset{\varphi} A\to \varphi a$.\\
So let $B\sub (D\expo_{\Kcat}^{\ast}E)\pro|D|$ directed with $\fset{\proi}B=\fset{\varphi} A$
and $\fset{\prot}B\to d$.\\
We have to show that $\fset{\eval} B\to \eval(\varphi a,d)$.\\
Let $C=\set{(\psi f,b)}{(f,b)\in B}$. $C$ is directed, as $B$ is directed and $\psi$ is monotonic.\\
It is $\fset{\proi}C=A$ and $\fset{\prot}C=\fset{\prot} B\to d$. We calculate:
\begin{align*}
\fset{\eval} B &= \set{\eval_\Kcat(\psi f,b)}{(f,b)\in B}\\
&= \fset{\eval_\Kcat} C\\
&\to \eval_\Kcat(a,d)\\
&= \eval_\Kcat(\psi(\varphi a),d)\\
&= \eval(\varphi a,d).
\end{align*}
So $\varphi$ is an isomorphism in $\mathbf{Dlubpo}$.
\qed
\begin{thm}\label{t:dlubpo}
In $\mathbf{Dlubpo}$ $\top$ is the terminal object and $D\pro E$ is the categorical product with
$\proi,\prot$ and $\lpro f,g\rpro$.\\
If there is an exponent of $D,E$ in $\mathbf{Dlubpo}$,
then it is $D\expo^{\ast}E$ with evaluation $\eval$ and curried functions $\operatorname{curry}(f)$.\\
$\mathbf{Dlubpo}$ is not a ccc: there are dlubpos $D,E$ such that $D\expo^{\ast}E$ is not the exponent,
i.e.\ there is some $f\typ C\pro D\fun E$ such that $\operatorname{curry}(f)$ is not continuous.
\end{thm}
\proof
The first sentence follows from the definitions of $\top$ and $D\pro E$.
The second follows from proposition \ref{p:expo2}.\\
Here is the counter-example in figure \ref{f:DEC}.
\psset{unit=5mm}
\begin{figure}[htb]
\begin{pspicture}(-2.7,-1.3)(13,7.0)
\firsto{b1}{0}{b_1}
\slineo{b1}{b2}{90}{2}{0}{b_2}
\slineo{b2}{b3}{90}{2}{0}{b_3}
\dlineon{b3}{b}{90}{2}{0}{b}
\nextn{b}{D}{180}{1.2}{D=}
\nexto{b1}{a1}{180}{2}{180}{a_1}
\nexto{b2}{a2}{180}{2}{180}{a_2}
\sline{a1}{b2}
\sline{a2}{b3}
\nexto{b1}{a1n}{0}{3}{180}{a_1}
\nexto{a1n}{b1n}{0}{2}{0}{b_1}
\slineo{b1n}{b1nn}{90}{0.9}{0}{b_1'}
\slineo{b1nn}{b2n}{90}{1}{0}{b_2}
\slineo{b2n}{b2nn}{90}{0.9}{0}{b_2'}
\slineo{b2nn}{b3n}{90}{1}{0}{b_3}
\dlineon{b3n}{bn}{90}{2}{0}{b}
\nextn{bn}{E}{180}{1.2}{E=}
\nexto{b2n}{a2n}{180}{2}{180}{a_2}
\sline{a1n}{b2n}
\sline{a2n}{b3n}
\nexto{b1n}{bot}{225}{1}{\S}{\bot}
\sline{bot}{a1n}
\sline{bot}{b1n}
\sline{bot}{a2n}
\nexto{b1n}{c1}{0}{4}{180}{c_1}
\slineo{c1}{c2}{90}{2}{180}{c_2}
\slineo{c2}{c3}{90}{2}{180}{c_3}
\dlineon{c3}{c}{90}{2}{180}{c}
\nextn{c}{C}{180}{1.2}{C=}
\nexto{c1}{d1}{0}{2}{0}{d_1}
\nexto{c2}{d2}{0}{2}{0}{d_2}
\sline{d1}{c2}
\sline{d2}{c3}
\end{pspicture}
\caption{Example theorem \ref{t:dlubpo}, dlubpos $D,E,C$}\label{f:DEC}
\end{figure}
\psset{unit=1cm}
In the dlubpos $D,E,C$ the order is the reflexive, transitive closure of the relations
specified below.
They also contain all trivial natural lubs,
i.e.\ all natural lubs $A\to a$ where $a\in A$.
So also the subcategory of the dlubpos with all trivial natural lubs is not a ccc.
$D$ is the dlubpo with\\
$|D|=\{a_i\}_{i\geq 1}\cup\{b_i\}_{i\geq 1}\cup \{b\}$,\\
for all $i$: $a_i\lex b_{i+1}$, $b_i\lex b_{i+1}$,
$b$ greatest element,\\
$D'=\{a_i\}_{i\geq 1}\cup\{b_i\}_{i\geq 1}$,
$D'\to b$.
$E$ is the dlubpo with:\\
$|E|=\{a_i\}_{i\geq 1}\cup\{b_i\}_{i\geq 1}\cup \{b_i'\}_{i\geq 1}\cup\{b\}$,\\
for all $i$: $a_i\lex b_{i+1}$, $b_i\lex b_i'\lex b_{i+1}$,
$b$ greatest element, $\bot$ least element,\\
for all directed $S\sub|E|$ with the lub $b$ and some $b_i'\in S$ it is $S\to b$.
$C$ is the dlubpo with:\\
$|C|=\{d_i\}_{i\geq 1}\cup\{c_i\}_{i\geq 1}\cup\{c\}$, it is isomorphic to $D$,\\
for all $i$: $d_i\lex c_{i+1}$, $c_i\lex c_{i+1}$,
$c$ greatest element,\\
$C'=\{d_i\}_{i\geq 1}\cup\{c_i\}_{i\geq 1}$,
$C'\to c$.
We define a continuous\ function $f\typ C\pro D\fun E$ by defining $\operatorname{curry}(f)$.
In the following, we abbreviate $\operatorname{curry}(f)x=\overline{x}$ for all $x\in|C|$.
For all $n\geq 1$ we define:
\begin{align*}
\overline{c_n} a_1 &= \bot & \overline{d_n} a_1 &= \bot\\
\overline{c_n} a_i &= a_{i-1} \text{ for } 2\leq i\leq n & \overline{d_n} a_i &= a_{i-1} \text{ for } 2\leq i<n\\
\overline{c_n} a_i &= \bot \text{ for } i>n & \overline{d_n} a_i &= \bot \text{ for } i\geq n \\
\overline{c_n} b_i &= b_i' \text{ for } i<n & \overline{d_n} b_i &= b_i' \text{ for } i<n\\
\overline{c_n} b_i &= b_{n-1}' \text{ for } i\geq n & \overline{d_n} b_i &= b_n \text{ for } i\geq n\\
\overline{c_n} b &= b_{n-1}' & \overline{d_n} b &= b_n\\[3mm]\db
\overline{c} a_1 &= \bot\\
\overline{c} a_i &= a_{i-1} \text{ for } i\geq 2\\
\overline{c} b_i &= b_i' \text{ for } i\geq 1\\
\overline{c} b &= b
\end{align*}
For all $x\in|C|$, the function $\overline{x}$ is monotonic.
Furthermore, for all $n$ it is
$\overline{c_n}\lex \overline{c_{n+1}}$,
$\overline{d_n}\lex \overline{c_{n+1}}$,
$\overline{c_n}\lex \overline{c}$,
also $\overline{d_n}\lex \overline{d_{n+1}}$
and $\overline{c_n}\lex \overline{d_{n+1}}$,
so that we get the diagram figure \ref{f:barx}.
\begin{figure}[htb]
\begin{pspicture}(-0.7,-0.2)(2.5,3.5)
\firsto{c1}{180}{\overline{c_1}}
\slineo{c1}{c2}{90}{1}{180}{\overline{c_2}}
\slineo{c2}{c3}{90}{1}{180}{\overline{c_3}}
\dlineon{c3}{c}{90}{1}{180}{\overline{c}}
\nexto{c1}{d1}{0}{1}{0}{\overline{d_1}}
\slineo{d1}{d2}{90}{1}{0}{\overline{d_2}}
\slineo{d2}{d3}{90}{1}{0}{\overline{d_3}}
\sline{c1}{d2}
\sline{c2}{d3}
\sline{d1}{c2}
\sline{d2}{c3}
\end{pspicture}
\caption{Example theorem \ref{t:dlubpo}, functions $\bar{x}=\operatorname{curry}(f)x$}\label{f:barx}
\end{figure}
From all this follows that $f$ is monotonic.\\
$f$ is also continuous:
let $A\sub |C|\pro|D|$ directed with $\fset{\proi} A\to e_1$ and $\fset{\prot} A\to e_2$.
We have to show that $\fset{f}A\to f(e_1,e_2)$.
The proof is by going through the cases,
we leave out the cases where one (or two) of the natural lubs is a trivial natural lub.\\
The interesting case is $\fset{\proi} A=C'\to c$ and $\fset{\prot} A=D'\to b$:\\
For every $n$ there is some $(c_n,y)\in A$ and some $(x,b_n)\in A$.\\
As $A$ is directed, there must be some $(c_i,b_j)\in A$ with $i\geq n$ and $j\geq n$,
then it is $b_{n-1}'\lex f(c_i,b_j)$.
Therefore $b$ is the lub of $\fset{f} A$.
Also $f(c_i,b_j)=b_k'$ for some $k$.
Therefore $\fset{f}A\to b=f(c,b)$.
So $f$ is continuous.
We now prove that $\operatorname{curry}(f)$ is not continuous:\\
It is $C'\to c$ in $C$.
We define $\overline{C'}=\fset{(\operatorname{curry}(f))}C'$.
It is $\overline{c}=\operatorname{curry}(f)c$.
We prove that not $\overline{C'}\to \overline{c}$ in $D\expo^{\ast}E$.\\
Let $A\sub |D\expo^{\ast}E| \pro |D|$ be the set $A=\{(\overline{c_n},a_n)\}_{n\geq1} \cup \{(\overline{d_n},b_n)\}_{n\geq 1}$.\\
$A$ is directed, as for all $n$ it is
$(\overline{c_n},a_n)\lex (\overline{d_{n+1}},b_{n+1})$
and $(\overline{d_n},b_n)\lex (\overline{d_{n+1}},b_{n+1})$.\\
It is $\fset{\proi}A=\overline{C'}$ and $\fset{\prot}A=D'\to b$,
but $\fset{\eval}A = \{\bot\}\cup \{a_i\}_{i\geq 1}\cup\{b_i\}_{i\geq 1}$ does not have
$\eval(\overline{c},b)=b$ as natural lub,
as $\fset{\eval}A$ does not contain any $b_i'$.
\qed
We now come to the problem of finding a large sub-ccc of $\mathbf{Dlubpo}$
(that uses the general exponents $D\expo^{\ast}E$).
We conjecture that it is possible to find such a category that is the largest among those
sub-cccs that are generated by some kind of invariant lub-rule class,
see the following section \ref{s:lubrule},
analogous to the result of theorem \ref{t:S10} for subcategories that use the
(pointwise) exponent $D\expo E$.
But it seems that this comes at a price:
the needed axiom will be rather complicated,
it encodes the condition that curried functions are continuous.
And the definition of the notion of lub-rule class has to be augmented.
As we think that such a result will not be worth the effort,
we do not follow this,
but show here that the simple axiom S5 is sufficient for cartesian closedness,
though not in any sense maximal.
\begin{defi}
A dlubpo $D$ is an \defin{S5-dlubpo} if it fulfills axiom S5 for directed sets:
If $X\sub Y\sub|D|$, $X,Y$ directed, $Y\to y$ and $Y\sqsubseteq X$, then $X\to y$.
\end{defi}
Note that from $X\sub Y\sub|D|$, $Y$ directed and $Y\sqsubseteq X$
follows that $X$ is also directed.
Therefore by axiom S5 we can deduce from natural lubs of directed sets only
natural lubs of sets that are directed again.
So we get the same class of dlubpos when we demand axiom S5 for general sets in the definition,
contrary to the situation for natural domains.
\begin{prop}\label{p:S5}
Let $D,E,C$ be dlubpos.
\begin{enumerate}[(1)]
\item
The terminal dlubpo $\top$ fulfills S5.
\item
If $D,E$ fulfill S5, then $D\pro E$ fulfills S5.
\item
If $E$ fulfills S5, then $D\expo^{\ast}E$ fulfills S5.
\item
If $E$ fulfills S5, then for all $f\typ C\pro D\fun E$ it is $\operatorname{curry}(f)\typ C\fun (D\expo^{\ast}E)$ continuous.
\end{enumerate}
\end{prop}
\proof
\textbf{(1)} is clear.\\
\textbf{(2)}
Let $X\sub Y \sub |D\pro E|$ directed, $Y\to y$ and $Y\sqsubseteq X$.
Then $\fset{\proi}X\sub \fset{\proi}Y$ directed,
$\fset{\proi}Y\to \proi y$
and $\fset{\proi}Y\sqsubseteq \fset{\proi} X$,
so $\fset{\proi}X\to \proi y$.
Also $\fset{\prot}X\to \prot y$. Therefore $X\to y$.\\
\textbf{(3)}
Let $F\sub G\sub |D\expo^{\ast}E|$ directed, $G\to g$ and $G\sqsubseteq F$.
We have to show $F\to g$.\\
Let $A\sub |D\expo^{\ast}E|\pro |D|$ directed with $\fset{\proi}A=F$ and $\fset{\prot}A\to d$.
We have to show $\fset{\eval}A\to gd$.\\
Let $B=G\pro (\fset{\prot}A)$.
$B$ is directed, so $\fset{\eval}B\to gd$.\\
For every $(h,a)\in B$ there is some $(h',y)\in A$ with $h\lex h'$,
and some $(x,a)\in A$.\\
As $A$ is directed, there is some $(h'',a')\in A$ with $(h,a)\lex (h'',a')$.\\
So we get $B\sqsubseteq A$, therefore $\fset{\eval}B\sqsubseteq \fset{\eval}A$.\\
From $A\sub B$ follows $\fset{\eval}A\sub \fset{\eval}B$.\\
From all this follows $\fset{\eval}A\to gd$, as $E$ fulfills S5.\\
\textbf{(4)}
Let $C'\to c$ in $C$. Let $\overline{C'}=\fset{(\operatorname{curry}(f))}C'$ and $\overline{c}=\operatorname{curry}(f)c$.\\
We have to show $\overline{C'}\to \overline{c}$ in $D\expo^{\ast}E$.\\
Let $A\sub |D\expo^{\ast}E|\pro |D|$ directed with $\fset{\proi}A=\overline{C'}$
and $\fset{\prot}A\to d$.\\
We have to show $\fset{\eval}A\to \overline{c} d$.\\
Let $B\sub|C|\pro|D|$ with $B=C'\pro(\fset{\prot}A)$.
$B$ is directed and $B\to(c,d)$, so $\fset{f}B\to f(c,d)$.\\
For every $(b,a)\in B$ there is some $(\operatorname{curry}(f)b,y)\in A$ and some $(x,a)\in A$.\\
As $A$ is directed, there is some $(g,a')\in A$ with
$(\operatorname{curry}(f)b,y)\lex (g,a')$ and $(x,a)\lex (g,a')$,
so $f(b,a)\lex \eval(g,a')$.\\
Therefore we get $\fset{f}B\sqsubseteq \fset{\eval}A$.\\
Together with $\fset{\eval}A\sub \fset{f}B$ and $\fset{f}B\to \overline{c} d$
we get $\fset{\eval}A\to \overline{c} d$,
as $E$ fulfills S5.
\qed
It follows this theorem:
\begin{thm}
The full subcategory $\mathbf{S5dlubpo}$ of $\mathbf{Dlubpo}$, of all S5-dlubpos,
is cartesian closed,
with the normal terminal object $\top$,
the normal products $D\pro E$,
the general exponents $D\expo^{\ast}E$,
and the corresponding morphisms.
\qed
\end{thm}
\section{Lub-rule classes and closed dlubpos}\label{s:lubrule}
In this section we set up the general frame of lub-rule systems and classes,
in which we can describe axiom systems that are like those of
Sazonov natural domains and dlubpos.
In this frame we can give a precise definition of the completeness of the axiom system;
and in section \ref{s:incomplete} it will turn out that the axioms of
Sazonov natural domains are not complete.
Let us have a look at an example axiom from section \ref{s:sazonov}:\\
(S6 cofinality) If $X,Y\sub |D|$, $X\to x$ and $X\sqsubseteq Y\sqsubseteq x$, then $Y\to x$.\\
Here we first have as hypothesis of the axiom a subset with a natural lub, $X\to x$,
and a further subset $Y$ with an order-theoretic configuration, $X\sqsubseteq Y\sqsubseteq x$.
From this configuration follows $\Lubex Y=x$,
and $Y\to x$ is the conclusion of the axiom.
The important property of the configuration, $X\sqsubseteq Y\sqsubseteq x$,
is that it is invariant against monotonic functions $f$ into another
partial order $E$: it is $\fset{f} X\sqsubseteq \fset{f} Y\sqsubseteq fx$.
If $f$ also respects the lub $\Lubex X=x$, i.e.\ $\Lubex \fset{f} X=fx$,
then also in $E$ we can draw the conclusion $\Lubex \fset{f} Y=fx$.
We can say this property of invariance means that the conclusion $\Lubex Y=x$,
and $Y\to x$ in the axiom, is ``necessary''.
All the axioms we have seen so far are of this form.
(Except axiom S1 which we do not count as proper axiom.)
They are necessary deductions of a lub from some set of natural lubs in an
invariant order-theoretic configuration.
We call these axioms ``valid''.
Here is an axiom that is not valid:
If $X\sub |D|$ and $\Lubex X=x$, then $X\to x$.
It is not valid, because there are monotonic functions $f$ from $D$ that do not
respect $\Lubex X=x$.
The monotonic function $f$ needs to respect only the natural lubs of the
hypothesis of the axiom, and here there are none.
We now formalize these intuitions.
Our results should also be applicable to domains that use natural lubs of \emph{general}
subsets, like natural domains.
For this we define lub partial orders as general structures,
only for use in this section.
(Further below we will go over from lubpos to dlubpos.)
\begin{defi}
A structure $D=(|D|,\lex_D,\nlub_D)$ is a \defin{lub partial order (lubpo)} if\\
$|D|$ is a set of \defin{elements},\\
$\lex_D$ is a partial order on $|D|$,\\
$\nlub_D\;\sub \bar{\Pot}(D)\pro|D|$ is the \defin{(natural) convergence relation}
with $\Lubex A=a$ for $A\to a$.
(Remember that $\bar{\Pot}(D)$ is the set of subsets of $|D|$ that have a lub in $\lex_D$.)\\
If $(|D|,\lex_D)$ is a partial order and $P\sub\bar{\Pot}(D)$,
then $\nlub^P\;\sub \bar{\Pot}(D)\pro|D|$ is the relation with
$A\nlub^P(\Lubex A)$ for all $A\in P$,
so that $(|D|,\lex_D,\nlub^P)$ is the lubpo corresponding to the set $P$ of natural subsets.
\end{defi}
A lub-rule models an instance of the application of an axiom:
\begin{defi}
A \defin{lub-rule} on a partial order $D$ is a triple $(D,P\mathord{\leadsto} A)$
with $P\sub\bar{\Pot}(D)$ and $A\in \bar{\Pot}(D)$.
$P$ is called the \defin{pattern} of the lub-rule and $A$ its \defin{result}.
A lub-rule $(D,P\mathord{\leadsto} A)$ is \defin{valid} if for every partial order $E$
and every monotonic function $f\typ D\fun E$
that respects the lubs of the elements of $P$
(i.e.\ $\all{B\in P}f(\Lubex B)=\Lubex(\fset{f}B)$),
$f$ respects also the lub of $A$
(i.e.\ $f(\Lubex A)=\Lubex(\fset{f} A)$).
A lubpo $E$ \defin{fulfills} a lub-rule $(D,P\mathord{\leadsto} A)$ if:\\
If $(|E|,\lex_E)=D$ and all elements of $P$ are natural in $E$,
then $A$ is natural in $E$.
A \defin{lub-rule system} $\mathcal{R}$ on a partial order $D$ is a set of pairs (rules)
$(P\mathord{\leadsto} A)$ such that $(D,P\mathord{\leadsto} A)$ is a valid lub-rule on $D$.
This is a rule system in the sense of definition \ref{d:rulesystem},
so we get derived lub-rules from the lub-rule system.
(Proposition \ref{p:derived} below shows that the derived rules are valid.)
In this context we implicitely understand every rule $(P\mathord{\leadsto} A)$ of $\mathcal{R}$
as the lub-rule $(D,P\mathord{\leadsto} A)$.
\noindent This lub-rule system $\mathcal{R}$ is \defin{complete} if every valid lub-rule on $D$
is derivable from it.
\noindent A \defin{lub-rule class} is a class of valid lub-rules.
(They can be on different partial orders.)
\noindent A lub-rule class $\mathcal{R}$ \defin{generates} the class of lubpos that fulfill all lub-rules of $\mathcal{R}$.
A lub-rule class $\mathcal{R}$ \defin{induces} on every partial order $D$ the lub-rule system of
all $(P\mathord{\leadsto} A)$ with $(D,P\mathord{\leadsto} A)\in \mathcal{R}$.
$\mathcal{R}$ is \defin{complete} if all these induced lub-rule systems are complete.
A lub-rule is \defin{derived} from $\mathcal{R}$,
if it is derived in one of the induced lub-rule systems.
A lub-rule class $\mathcal{R}$ is \defin{invariant} if for every lub-rule $(D,P\mathord{\leadsto} A)\in \mathcal{R}$,
every partial order $E$ and every monotonic function $f\typ D\fun E$
that respects the lubs of the elements of $P$,
also the lub-rule $(E,\fset{(\fset{f})}P\mathord{\leadsto} \fset{f}A)$
is derivable in the lub-rule system induced by $\mathcal{R}$ on $E$.
(It is $\fset{(\fset{f})} P = \set{\fset{f} B}{B\in P}$.)
(The last is a valid lub-rule as the first is it.)
\end{defi}
\begin{prop}
If we have a lubpo $E$ and a lub-rule system $\mathcal{R}$ on $(|E|,\lex_E)$
such that $E$ fulfills every rule of $\mathcal{R}$,
then $E$ fulfills also every derived rule of $\mathcal{R}$.\\
Every complete lub-rule class is invariant.
\qed
\end{prop}
Let us demonstrate the translation from the axiom system to the lub-rule class for
the axiom S6 (cofinality).
It is translated to the lub-rule class of all lub-rules $(D,\{X\}\mathord{\leadsto} Y)$
such that $D$ is a partial order, $X\sub|D|$ with $\Lubex X=x$ for some $x$,
and $X\sqsubseteq Y\sqsubseteq x$.
All these are valid lub-rules.
The other axioms S3 and S7 of natural domains are likewise translated,
so that we get an invariant lub-rule class $\mathcal{S}$
for the axioms of Sazonov natural domains.
(The proof is immediate and boring.)
\subsection*{Philosophical significance: the quest for intension by partial extensionalization}
Let us reflect on what we have done so far.
The following is not mathematics, it is the thoughts that I had when discovering
this piece of mathematics.
Since the Logic of Port-Royal (1662),
philosophical logic makes the distinction between two kinds of meaning of a concept:
the \emph{intension} (or content),
i.e.\ the predicates the concept uses to describe the things which fall under it,
and the \emph{extension},
i.e.\ the class of all these things.
Mathematics, as based on set theory, is always extensional,
i.e.\ we use \emph{intensions} (like expressions, axioms),
to describe, define \emph{extensions} (like sets, classes).
We always talk in intensions about extensions.
But what can we do when we want to talk \emph{in mathematics} about an intension?
We have to \emph{extensionalize} an aspect of the intension,
we have to introduce new extensions.
This may occur at different levels, the intension covers a whole spectrum of them.
We are always sure to have the lowest level,
i.e.\ the purely syntactic text on which the intension is founded,
and the highest level,
the extension that the intension describes.
We call the last one the \emph{full extensionalization},
and a level properly between the lowest and the highest a \emph{partial extensionalization}.
In our concrete case we have as intension an axiom system for lubpos,
which defines as extension a class of lubpos.
We deliberately leave open the formal logic of the axiom system,
so that we describe the following translation informally.
The axiom system must be amenable to this translation to a lub-rule class,
which is our partial extensionalization.
A lub-rule abstracts a certain form of inference from the axiom, in extensional form.
It must be defined that the partial extensionalization describes the full one,
which is done by our definition of a lub-rule class generating a class of lubpos.
We have gained new objects,
the lub-rule classes, which can be used to analyse classes of lubpos.
There might be similar examples in the literature, but I know only the following one.
Sometimes the partial extensionalization works by forming equivalence classes
of the syntactic expressions of the intensions.
This is the case for the formalization of the notion of algorithm given by Noson
Yanofsky \cite{Yanofsky}.
Here the intensional objects are (primitive recursive) programs,
and the full extensions are the functions they describe.
The algorithms are defined as certain equivalence classes of programs.
\bigskip
Now we introduce the important cl-rule system on a lubpo and its closure operation,
and the lub-completion of a lubpo.
These procedures appear in \cite{Courcelle/Raoult}
and are used there for various completions of partial orders that respect lubs
that are already fixed.
In this section we use them to give a characterization of valid lub-rules;
this will be used in section \ref{s:incomplete} to prove the
incompleteness of the Sazonov axioms $\mathcal{S}$.
\begin{defi}[Courcelle/Raoult, proof of theorem 1 in \cite{Courcelle/Raoult}]\label{d:cl}
Let $D$ be a lubpo.
The \defin{cl-rule system} on $D$ is the rule system
(in the sense of def.\ \ref{d:rulesystem}) $\mathord{\leadsto}$ on $|D|$ given by the rules:\\
for all $a,b\in |D|$ with $b\lex a$: $\{a\}\mathord{\leadsto} b$,\\
for all $A\sub |D|$, $a\in |D|$ with $A\to a$: $A\mathord{\leadsto} a$.\\
The closure operator of the cl-rule system is called $\operatorname{cl}_D$.\\
An $A\sub|D|$ is \defin{closed} if it is closed under $\operatorname{cl}_D$.\\
(We remark that closed sets are a generalization of subsets of dcpos
that are closed w.r.t.\ the Scott topology.)\\
If we have a lubpo given this way: $D=(|D|,\lex_D,\nlub^P)$,
with its set of natural sets $P$,
then we abuse notation and write $\operatorname{cl}_P$ for $\operatorname{cl}_D$,
if $|D|$ and $\lex_D$ are clear from the context.\\
$\hat{D}$ is the set of all closed subsets of $D$.\\
The \defin{lub-completion} of $D$ is $\operatorname{lub-comp}(D)=(\hat{D},\lex)$,
with $\lex$ the inclusion $\sub$.\\
It is a complete lattice with $\Lubex \bar{A}= \operatorname{cl}_D(\Join \bar{A})$, for $\bar{A}\sub \hat{D}$.\\
There is an embedding $\operatorname{in}_D\typ |D|\fun\hat{D}$ defined by
$\operatorname{in}_D d = \operatorname{cl}_D\{d\}=\set{x\in D}{x\lex d}$.
\end{defi}
\begin{prop}[Courcelle/Raoult, proof of theorem 1 in \cite{Courcelle/Raoult}]\label{p:in}
The embedding $\operatorname{in}_D$ is an order embedding,
i.e.\ it is monotonic, injective, and the reverse is monotonic. It maps natural lubs in $D$ to lubs in $\hat{D}$,
i.e.\ for $A\to a$ in $D$ it is $\operatorname{in}_D a = \Lubex(\fset{\operatorname{in}_D} A)= \operatorname{cl}_D A$.
\end{prop}
\proof
$\operatorname{in}_D$ and its reverse are clearly monotonic.\\
Now let $A\to a$ in $D$.
It is $A\sub \Join(\fset{\operatorname{in}_D} A)$,
then $a\in \operatorname{cl}(\Join(\fset{\operatorname{in}_D} A)) = \Lubex(\fset{\operatorname{in}_D} A)$,\\
so $\operatorname{in}_D a \sub \Lubex(\fset{\operatorname{in}_D} A)$.\\
For the reverse, it is $\operatorname{in}_D b\sub\operatorname{in}_D a$ for all $b\in A$.
\qed
\begin{prop}\label{p:valid}
Let $(D,P\mathord{\leadsto} A)$ be a lub-rule and
$E=(|D|,\lex_D,\nlub^P)$ be the corresponding lubpo
with the set $P$ of natural subsets.
Let $\Lubex A=a$.\\
$(D,P\mathord{\leadsto} A)$ is valid $\ifff$ $a\in \operatorname{cl}_E(A)$, also written $a\in \operatorname{cl}_P(A)$.
\end{prop}
\proof
$\ifthen$:
We have the embedding $\operatorname{in}\typ (|E|,\lex)\fun \operatorname{lub-comp}(E)=(\hat{E},\lex)$.\\
By proposition \ref{p:in}, the function $\operatorname{in}$ respects the natural lubs of the elements of $P$.
Whence by validity it also respects the lub of $A$, so $\operatorname{in}(\Lubex A)=\Lubex(\fset{\operatorname{in}} A)$.\\
Further we get $\Lubex(\fset{\operatorname{in}} A) = \operatorname{cl}_E(\Join(\fset{\operatorname{in}} A))=\operatorname{cl}_E(A)$,
thus $a\in\operatorname{cl}_E(A)$.
\noindent $\thenif$:
Let $F=(|F|,\lex_F)$ be a partial order and
$f\typ D\fun F$ be a monotonic function that respects the lubs of the elements of $P$.
We have to show: $fa=\Lubex \fset{f}A$.\\
We prove:
For all $x\in \operatorname{cl}_E(A)$,
for all upper bounds $b$ of $\fset{f}A$, it is $fx\lex b$,
by induction on the deduction of $x\in \operatorname{cl}_E(A)$ in the cl-rule system on $E$.\\
(1) This is true for $x\in A$.\\
(2) Let $x$ be deduced from $y\in \operatorname{cl}_E(A)$ by $x\lex y$.
Then $fx\lex fy\lex b$.\\
(3) Let $x$ be deduced from $Y\sub\operatorname{cl}_E(A)$ by $Y\to x$.
Then $fx=\Lubex(\fset{f} Y)\lex b$.\\
It is clear that $fa$ is an upper bound of $\fset{f} A$.
The proved property for $x=a$ shows that it is the least upper bound.
\qed
\begin{prop}\label{p:derived}
Let $\mathcal{R}$ be a lub-rule system on a partial order $D$.
Every derived rule $P\mathord{\leadsto} A$ of $\mathcal{R}$ is valid.
\end{prop}
\proof
Let D1 be a deduction of $A$ from $P$ in the lub-rule system $\mathcal{R}$.\\
We prove by induction on D1 that for every node of D1 labelled with some deduced $A'$,
with $\Lubex A'=a'$,
it is $a'\in \operatorname{cl}_P(A')$.\\
(1) This is clear for $A'\in P$.\\
(2) Let $A'$ be deduced from some $P'$ at this node, $P'\mathord{\leadsto} A'$.\\
From proposition \ref{p:valid} follows that $a'\in \operatorname{cl}_{P'}(A')$.\\
Let D2 be a deduction of $a'$ from $A'$ in the cl-rule system on $(|D|,\lex_D,\to^{P'})$.\\
We prove by induction on D2 that for every node of D2 labelled with some deduced $x$,
it is $x\in\operatorname{cl}_P(A')$.\\
(2.1) This is clear for $x\in A'$.\\
(2.2) Let $x$ be deduced from $y$ by $x\lex y$.
It is $y\in\operatorname{cl}_P(A')$ by induction hypothesis,
so $x\in \operatorname{cl}_P(A')$.\\
(2.3) Let $x$ be deduced from $Y\in P'$ by $Y\to x$.\\
The elements of $P'$ were deduced in the deduction D1,
so by induction hypothesis of the outer induction on D1 it is $x\in \operatorname{cl}_P(A')$.
From the inner induction on D2 follows that $a'\in \operatorname{cl}_P(A')$.\\
From the outer induction on D1 follows that $\Lubex A\in\operatorname{cl}_P(A)$,
so $P\mathord{\leadsto} A$ is valid by proposition \ref{p:valid}.
\qed
The easiest way to get a complete lub-rule class is to take just all valid lub-rules,
which can be characterized by proposition \ref{p:valid}:
\begin{defi}
The \defin{canonical complete lub-rule class} $\mathcal{C}$ is the class of all valid lub-rules.
This is the set of all lub-rules $(D,P\mathord{\leadsto} A)$ with $\Lubex A\in \operatorname{cl}_P(A)$.
\end{defi}
\bigskip
We are mainly interested in the natural sets that are directed, as continuity is based on them,
and regard the undirected natural sets only as intermediary (may be necessary)
steps in the deduction of naturality of directed sets,
see the discussion in section \ref{s:sazonov}.
So we go over from lubpos to dlubpos.
\begin{defi}
A lubpo that fulfills the singleton axiom S3 is called a \defin{lubpo with singletons}.
For a lubpo with singletons we have the \defin{dlubpo for} $D$, $\delta(D)= (|D|,\lex_D,\to)$
with $\to$ the restriction of $\nlub_D$ to directed subsets.
A lub-rule class $\mathcal{R}$ is \defin{with singletons}
if $\mathcal{R}$ has for all partial orders $D$ and all $x\in |D|$ the (valid)
lub-rule $(D,\emptyset\mathord{\leadsto} \{x\})$.
A lub-rule $(D,P\mathord{\leadsto} A)$ is \defin{directed} if $A$ and the elements of $P$ are all directed.\\
For a lub-rule class $\mathcal{R}$ with singletons,
$\delta(\mathcal{R})$ is the lub-rule class of all derived lub-rules of $\mathcal{R}$ that are directed.\\
A lub-rule class $\mathcal{R}$ with singletons \defin{generates} the class of all
dlubpos $\delta(D)$ for $D$ a lubpo that fulfills the lub-rules of $\mathcal{R}$.
$\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ is the full subcategory of this class in $\mathbf{Dlubpo}$.
\end{defi}
\begin{prop}
A lub-rule class $\mathcal{R}$ with singletons generates the class of dlubpos
that fulfill all lub-rules of $\delta(\mathcal{R})$.
\end{prop}
\proof
We have to prove that very dlubpo $D$ that fulfills all lub-rules of $\delta(\mathcal{R})$
can be extended to a lubpo $D'=(|D|,\lex_D,\to')$ such that
$D'$ fulfills the lub-rules of $\mathcal{R}$ and $\delta(D')=D$.\\
Take the set $P$ of all (directed) natural sets of $D$ and build the closure $P'$ of $P$
under the lub-rule system induced by $\mathcal{R}$ on $(|D|,\lex_D)$.
Take $\to'\,=\,\to^{P'}$.
\qed
If we use all valid lub-rules,
the deduction of (directed) natural lubs from existing ones is always in one step.
Thus we can define the class of dlubpos that is generated by $\mathcal{C}$ by an axiom
on directed lubs that tests the condition of proposition \ref{p:valid}:
\begin{defi}
A dlubpo $D=(|D|,\lex_D,\nlub_D)$ is a \defin{closed dlubpo (cdlubpo)}
if it fulfills the axiom:\\
\textbf{(S9 closure)} If $A\sub|D|$ is directed with lub $a$
and $a\in \operatorname{cl}_D(A)$, then $A\nlub_D a$.
\end{defi}
Note that this axiom, as it stands, does not describe a complete lub-rule class.
But its extension is exactly the class of dlubpos generated by the complete lub-rule class $\mathcal{C}$.
\begin{prop}
A dlubpo $D$ is a cdlubpo $\ifff$ $D$ fulfills all valid directed lub-rules\\
$((|D|,\lex_D),P\mathord{\leadsto} A)$.\\
The category $\mathbf{Cdlubpo}$ of cdlubpos is the category $\mathcal{C}$-$\mathbf{cat}$.
\end{prop}
\proof
$\ifthen$:
Let $((|D|,\lex_D),P\mathord{\leadsto} A)$ be a valid lub-rule of directed sets such that all elements
of $P$ are natural in $D$, and $\Lubex A=a$.
Then $a\in \operatorname{cl}_P(A)$ by proposition \ref{p:valid},
therefore also $a\in \operatorname{cl}_D(A)$.
Thus $A\to a$.
$\thenif$:
Let $A\sub |D|$ be directed with lub $a$ and $a\in \operatorname{cl}_D(A)$.
Let $P$ be the set of all natural (directed) subsets of $|D|$.
Then $((|D|,\lex_D),P\mathord{\leadsto} A)$ is valid by proposition \ref{p:valid}. Thus $A\to a$.
\qed
As a preparation for the next sections, we now give some properties of the classes of dlubpos
that are generated by (invariant) lub-rule classes.
\begin{defi}
Let $D$ be a dlubpo with the set $P$ of natural directed subsets,
and $\mathcal{R}$ be a lub-rule class with singletons.
The $\mathcal{R}$-completion $\operatorname{\mathcal{R}-comp}(D)$ of $D$ is the dlubpo $(|D|,\lex_D,\to^{P'})$
with $P'$ the directed sets of the closure of $P$ under the lub-rule system
induced by $\mathcal{R}$.
\end{defi}
\begin{prop}\label{p:frcomp}
Let $D,E$ be dlubpos and $\mathcal{R}$ be a lub-rule class with singletons.
\begin{enumerate}[(1)]
\item
$D$ is in the class of dlubpos generated by $\mathcal{R}$ iff $\operatorname{\mathcal{R}-comp}(D)=D$.
\item
If $\mathcal{R}$ is invariant, then every continuous\ function $f\typ D\fun E$ is also
a continuous\ function $f\typ \operatorname{\mathcal{R}-comp}(D)\fun \operatorname{\mathcal{R}-comp}(E)$.
\end{enumerate}
\end{prop}
\proof
(1) is clear.\\
(2) Let $D'=\operatorname{\mathcal{R}-comp}(D)$ and $E'=\operatorname{\mathcal{R}-comp}(E)$.\\
We have to prove: For every $A\to a$ in $D'$ it is
$\fset{f} A\to fa$ in $E'$.\\
There is a deduction of $A$ from the set $P$ of natural directed sets of $D$
in the lub-rule system induced by $\mathcal{R}$.
We prove by induction on this deduction,
that in every node for the deduced $A'$ (with lub $a'$)
it is $\Lubex \fset{f} A'= fa'$ and $\fset{f} A'$ is in the closure of
$\fsset{f}P$ in the lub-rule system of $\mathcal{R}$ on $E$.\\
This is clear for $A'\in P$, as $f\typ D\fun E$ is continuous.\\
Now let $A'$ be deduced from the set of subsets $P'$
by the (valid) lub-rule $(D,P'\mathord{\leadsto} A')$.\\
By induction hypothesis $f$ respects the lubs of the elements of $P'$,
so also the lub of $A'$: $\Lubex \fset{f} A' = fa'$.\\
By induction hypothesis for all $B\in P'$, $\fset{f} B$ is in the closure of $\fsset{f} P$.
As $\mathcal{R}$ is invariant, $(E,\fsset{f}P'\mathord{\leadsto} \fset{f} A')$ is a derived lub-rule
in the lub-rule system of $\mathcal{R}$.
Therefore also $\fset{f} A'$ is in the closure of $\fsset{f}P$.
From the proved hypothesis follows that $\fset{f} A \to fa$.
\qed
\begin{thm}
Let $\mathcal{R}$ be an invariant lub-rule class with singletons.\\
Then $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ is a full reflective subcategory of $\mathbf{Dlubpo}$.\\
The adjunction is $\langle\operatorname{\mathcal{R}-comp},\operatorname{\mathcal{R}-inc},\operatorname{\mathcal{R}-\eta}\rangle \typ \mathbf{Dlubpo} \rightharpoonup \mbox{$\mathcal{R}$-$\mathbf{cat}$}$
with\\
$\operatorname{\mathcal{R}-comp}\typ \mathbf{Dlubpo}\fun \mbox{$\mathcal{R}$-$\mathbf{cat}$}$ the functor mapping each
$D\in \mathbf{Dlubpo}$ to $\operatorname{\mathcal{R}-comp}(D)$
and each $f\typ D\fun E$ to the same $f\typ\operatorname{\mathcal{R}-comp}(D)\fun\operatorname{\mathcal{R}-comp}(E)$,\\
$\operatorname{\mathcal{R}-inc}\typ \mbox{$\mathcal{R}$-$\mathbf{cat}$}\fun\mathbf{Dlubpo}$ the inclusion functor,\\
$\operatorname{\mathcal{R}-\eta}\typ \id_{\mathbf{Dlubpo}}\fun \operatorname{\mathcal{R}-inc}\comp\operatorname{\mathcal{R}-comp}$
the natural transformation with\\
$\operatorname{\mathcal{R}-\eta}_D\typ D\fun\operatorname{\mathcal{R}-inc}(\operatorname{\mathcal{R}-comp}(D))$, $\operatorname{\mathcal{R}-\eta}_D = \id$.
\end{thm}
\proof
$\operatorname{\mathcal{R}-comp}$ is a functor by proposition \ref{p:frcomp}.\\
$\operatorname{\mathcal{R}-\eta}$ clearly is a natural transformation.\\
We have to prove that for every $D\in \mathbf{Dlubpo}$, $E\in\mbox{$\mathcal{R}$-$\mathbf{cat}$}$,
$f\typ D\fun \operatorname{\mathcal{R}-inc}(E)$,
there is a unique $g\typ\operatorname{\mathcal{R}-comp}(D)\fun E$ with $f=\operatorname{\mathcal{R}-inc}(g)\comp \operatorname{\mathcal{R}-\eta}_D$.
We choose for $g$ the same function $f$,
as $f\typ \operatorname{\mathcal{R}-comp}(D)\fun\operatorname{\mathcal{R}-comp}(E)=E$ is continuous\ by proposition \ref{p:frcomp}.
\qed
\begin{defi}\label{d:function}
Let $D,E$ be dlubpos.\\
We define the \defin{(pointwise) function space dlubpo} $D\expo E$ by\\
$|D\expo E |$ the set of all continuous $f\typ D\fun E$,\\
$f\lex_{D\expo E}g$ iff for all $x\in |D|$ it is $fx\lex_E gx$,\\
if $F\sub|D\expo E|$ is directed and $f\in|D\expo E|$,
then $F\to_{D\expo E} f$ iff for all $d\in |D|$ it is
$Fd = \set{gd}{g\in F} \nlub_E fd$.
In this definition it is $\Lubex F=f$.
The singleton axiom is fulfilled.
\end{defi}
\begin{prop}\label{p:invariant}
Let $\mathcal{R}$ be an invariant lub-rule class with singletons and $D,E \in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$. Then
\begin{enumerate}[(1)]
\item
$\top\in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$ is the categorical terminal object,
\item
$D\pro E\in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$ is the categorical product,
\item
$D\expo E\in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$ (but need not be the categorical exponent).
\end{enumerate}
\end{prop}
\proof
\textbf{(1)}
On $(|\top|,\lex)$ the only valid lub-rules are
$S\mathord{\leadsto} \{\bot\}$ for $S\sub \{\emptyset,\{\bot\}\}$
and they are all fulfilled by $\top$.\\
\textbf{(2)}
Let $((|D\pro E|,\lex),P\mathord{\leadsto} A)$ be a derived directed lub-rule of $\mathcal{R}$,
such that the elements of $P$ are natural.
(We have to show that $A$ is natural.)\\
Then the elements of $\fsset{\proi}P$ and $\fsset{\prot}P$ are
also natural in $D$ resp.\ $E$.\\
$((|D|,\lex),\fsset{\proi}P\mathord{\leadsto} \fset{\proi}A)$ and
$((|E|,\lex),\fsset{\prot}P\mathord{\leadsto} \fset{\prot}A)$ are derived rules of $\mathcal{R}$,
by invariance of $\mathcal{R}$ w.r.t.\ the continuous\ functions $\proi$ resp.\ $\prot$.\\
Thus $\fset{\proi} A$ and $\fset{\prot}A$ are natural in $D$ resp.\ $E$.
Therefore $A$ is natural in $D\pro E$,
so $D\pro E$ fulfills the lub-rule $P\mathord{\leadsto} A$.
We get $D\pro E\in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$.\\
As $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ is a full subcategory of $\mathbf{Dlubpo}$,
$\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ has the projections and pairing functions,
so $D\pro E$ is the categorical product in $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$.\\
\textbf{(3)}
Let $((|D\expo E|,\lex),P\mathord{\leadsto} A)$ be a derived directed lub-rule of $\mathcal{R}$,
such that the elements of $P$ are natural.
(We have to show that $A$ is natural.)\\
Then the elements of $Px=\set{Bx}{B\in P}$, with $Bx=\set{fx}{f\in B}$,
are also natural in $E$.\\
$((|E|,\lex),Px\mathord{\leadsto} Ax)$ is a derived lub-rule of $\mathcal{R}$,
by invariance of $\mathcal{R}$ w.r.t.\ the function $\la f.fx$,
that respects the lubs of the $B\in P$.\\
Thus $Ax$ is natural in $E$ for every $x\in |D|$.
Therefore $A$ is natural in $D\expo E$,
so $D\expo E$ fulfills the lub rule $P\mathord{\leadsto} A$.
We get $D\expo E \in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$.
\qed
\section{The ccc of S10-dlubpos}
In this section we find a large cartesian closed full subcategory of $\mathbf{Dlubpo}$
whose exponents are the (pointwise) function spaces $D\expo E$ which are known
from natural domains and from definition \ref{d:function}.
The underlying structures are the S10-dlubpos,
which fulfill the singleton axiom and a new axiom S10.
All natural domains are also S10-dlubpos.
\begin{defi}
A dlubpo $D$ is an \defin{S10-dlubpo} if it fulfills the axiom:\\
\textbf{(S10)}
If $I$ is an index set with a directed partial order,\\
and $\{\yij\}_{i,j\in I}$ is a two-parametric family of elements of $|D|$
that is monotonic in each parameter $i$ and $j$,\\
and for all $i\in I$ it is $\{\yij\}_{j\in I}\nlub_D x_i$ for some $x_i$,
and $\{x_i\}_{i\in I}\nlub_D u$ for some $u$,\\
and for all $j\in I$ it is $\{\yij\}_{i\in I}\nlub_D z_j$ for some $z_j$,
and $\{z_j\}_{j\in I} \nlub_D u$,\\
then it is $\{y_{ii}\}_{i\in I}\nlub_D u$.\\
$\mathbf{S10dlubpo}$ is the full subcategory of $\mathbf{Dlubpo}$ of all S10-dlubpos.
\end{defi}
The axiom looks like a cross-over of axioms S4(1$\Rightarrow$) and S4(2) of natural domains.
But note that all the natural convergences in the axiom are directed,
by the monotonocity of the family.
\begin{prop}
The axiom S10 follows in a dlubpo from axioms (S4(1$\Rightarrow$) and S4(2)).
It also follows from axioms (S5 and S7(transitivity)).\\
So every natural domain
is an S10-dlubpo.
\qed
\end{prop}
\begin{prop}
The axioms of an S10-dlubpo (i.e.\ S3(singleton) and S10)
form valid lub-rules and an invariant lub-rule class.\\
So every cdlubpo is an S10-dlubpo.
\end{prop}
\proof
This is clear for the singleton axiom S3.\\
An instance of axiom S10 translates to a lub-rule $(D,P\mathord{\leadsto} A)$
for a partial order $D$ with an $I$ and $\{\yij\}_{i,j\in I}$ in $D$
and where $P$ consists of $\{\yij\}_{j\in I}$ for all $i\in I$ (with lub $x_i$),
$\{x_i\}_{i\in I}$ (with lub $u$),
$\{\yij\}_{i\in I}$ for all $j\in I$ (with lub $z_j$),
$\{z_j\}_{j\in I}$ (with lub $u$).
$A$ is $\{y_{ii}\}_{i\in I}$.
$u$ is the lub of $A$.
Let $E$ be another partial order and $f\typ D\fun E$ be monotonic and respecting the lubs of the
elements of $P$.
In the same way it can be seen that $fu$ is the lub of $\fset{f} A$,
so the lub-rule $(D,P\mathord{\leadsto} A)$ is valid.
It is also clear that $(E,\fsset{f} P\mathord{\leadsto} \fset{f} A)$ is a lub-rule for an instance
of axiom S10,
so that the lub-rule class is invariant.
\qed
\begin{defi}
For every directed partial order $(I,\lex)$,
that will be used as an index set,
we construct a dlubpo $\bar{I} =(|\bar{I}|,\lex_{\bar{I}},\to_{\bar{I}})$
as a completion of $I$ by a lub:\\
If $I$ contains an upper bound of itself,
then this will be called $t$ and $|\bar{I}|=I$ and $\lex_{\bar{I}}\; =\;\lex$.\\
Otherwise, $|\bar{I}|=I\cup \{t\}$ with a new top element $t$,
and $i\lex_{\bar{I}} j$ iff
($i,j\in I$ and $i\lex j$)
or ($i\in |\bar{I}|$ and $j=t$).\\
In every case, the convergence is $\{x\}\to_{\bar{I}} x$ for $x\in |\bar{I}|$,
and $I\to_{\bar{I}} t$.
\end{defi}
\begin{lem}\label{l:evalS10}
Let $\mathcal{R}$ be an invariant lub-rule class with singletons.\\
Let for every directed partial order $I$ be $I^{\ast}=\operatorname{\mathcal{R}-comp}(\bar{I})$.\\
Let $E$ be a dlubpo generated by $\mathcal{R}$.
Then the following are equivalent:
\begin{enumerate}[(1)]
\item
$E$ fulfills axiom S10.
\item
For every dlubpo $D$ it is $\eval\typ (D\expo E)\pro D\fun E$ continuous.
\item
For every directed partial order $I$ it is $\eval\typ (I^{\ast}\mathbin{\Rightarrow} E)\pro I^{\ast}\fun E$ continuous.
\item
For every dlubpo $D$ it is $(D\expo E) = (D\expo^{\ast}E)$.
\end{enumerate}
\end{lem}
\proof
\textbf{(1)$\Rightarrow$ (2)}
First, $\eval$ is monotonic.
Now let $A\sub (D\expo E) \pro D$ be directed with $A\to (f,d)$.
We have to show that $\fset{\eval}A\to fd$.
It is $\fset{\eval}A=\set{(\proi x)(\prot x)}{x\in A}$.\\
Let $I=A$ be the index set for axiom S10.
For $i,j\in I$ let $y_{ij}=(\proi i)(\prot j)$.\\
For $i\in I$ it is $\{\yij\}_{j\in I}\to(\proi i)d =: x_i$,
as $\{\prot j\}_{j\in I}\to d$ and as $\proi i$ is continuous.\\
It is $\{x_i\}_{i\in I}\to fd =: u$,
as $\{\proi i\}_{i\in I}\to f$ and therefore $\{\proi i\}_{i\in I}\, d\to fd$.\\
For $j\in I$ it is $\{\yij\}_{i\in I}\to f(\prot j) =: z_j$,
as $\{\proi i\}_{i\in I}\to f$.\\
It is $\{z_j\}_{j\in I}\to fd$, as $\{\prot j\}_{j\in I}\to d$.\\
Therefore the conditions of axiom S10 are fulfilled and $\{y_{ii}\}_{i\in I}\to fd$,
so $\fset{\eval} A\to fd$.
\textbf{(2)$\Rightarrow$ (4)}
For $F\sub |D\expo E|$ directed and $f\in |D\expo E|$ it is
$F\to_{D\expo E} f$ iff $F\to_{D\expo^{\ast}E} f$.
\textbf{(4)$\Rightarrow$ (3)}
Follows directly from proposition \ref{p:evalstar}.
\textbf{(3)$\Rightarrow$ (1)}
Let $I$ be an index set with a directed partial order,
and $\{\yij\}_{i,j\in I}$, $\{x_i\}_{i\in I}$, $\{z_j\}_{j\in I}$ and $u$ in $E$ as the axiom S10 describes.\\
For $i\in I$ we define functions $f_i\typ \bar{I}\fun E$ by
$f_ij=y_{ij}$ for $j\in I$,
and $f_it=x_i$.\\
$f_i$ is continuous, as $\{\yij\}_{j\in I}\to x_i$.\\
As $\mathcal{R}$ is invariant,
$f_i$ is also a continuous function $f_i\typ I^{\ast}\fun E$, by proposition \ref{p:frcomp}.\\
We define $f_t\typ \bar{I}\fun E$ by $f_tj=z_j$ for $j\in I$,
and $f_tt=u$.\\
$f_t$ is continuous, as $\{z_j\}_{j\in I}\to u$.
Again, $f_t$ is also a continuous function $f_t\typ I^{\ast}\fun E$.\\
Let $A\sub |I^{\ast}\mathbin{\Rightarrow} E|\pro|I^{\ast}|$ with $A=\set{(f_i,i)}{i\in I}$.
$A$ is directed.\\
We have $\{f_i\}_{i\in I}\to f_t$,
as (a) $\{f_ij\}_{i\in I}\to f_tj$ for $j\in I$, as $\{\yij\}_{i\in I}\to z_j$,\\
and as (b) $\{f_it\}_{i\in I}\to f_tt$, as $\{x_i\}_{i\in I}\to u$.\\
And we have $\{i\}_{i\in I}\to t$,
so $A\to (f_t,t)$.\\
As $\eval\typ (I^{\ast}\mathbin{\Rightarrow} E)\pro I^{\ast}\fun E$ is continuous,
we get $\fset{\eval}A\to \eval(f_t,t)=u$,
so $\{y_{ii}\}_{i\in I}\to u$.
\qed
As a difference to the general function space $D\expo^{\ast}E$,
the curried functions for $D\expo E$ are continuous:
\begin{prop}\label{p:currynormal}
Let $C,D,E$ be dlubpos.
For any $f\typ C\pro D\fun E$, $\operatorname{curry}(f)\typ C\fun(D\expo E)$,
with $\operatorname{curry}(f)c = \la d\in D. f(c,d)$, is continuous\ (and has continuous\ results).
\end{prop}
\proof
Let $A\sub |C|$ be directed with $A\to a$.\\
Then $\set{\la d\in D. f(x,d)}{x\in A}\to_{D\expo E} \la d\in D. f(a,d)$,
as $f$ is continuous\ in its first argument.
\qed
\begin{thm}\label{t:S10}\hfill
\begin{enumerate}[(1)]
\item
Let $\mathcal{R}$ be an invariant lub-rule class with singletons.\\
The subcategory $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ of $\mathbf{Dlubpo}$ is a ccc with pointwise exponents $D\expo E$\\
$\ifff$ every $E\in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$ fulfills axiom S10.\\
($\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ also has normal terminal and normal products.)
\item
From (1) follows that $\mathbf{S10dlubpo}$, and the category of natural domains,
and $\mathbf{Cdlubpo}$ are cccs with the pointwise exponent $D\expo E$.
\item
$\mathbf{S10dlubpo}$ is the largest full sub-ccc of $\mathbf{Dlubpo}$
which is generated by an invariant lub-rule class and has the pointwise exponents $D\expo E$.
\end{enumerate}
\end{thm}
\proof
(1) $\ifthen$:
By proposition \ref{p:expo1}(1) the evaluation function for the exponent in $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$
is the normal $\eval$.\\
For every directed partial order $I$ it is $I^{\ast}=\operatorname{\mathcal{R}-comp}(\bar{I})$ in $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$.\\
Let $E\in \mbox{$\mathcal{R}$-$\mathbf{cat}$}$.
It must be $\eval\typ (I^{\ast}\mathbin{\Rightarrow} E)\pro I^{\ast}\fun E$ continuous,
so by lemma \ref{l:evalS10}(3$\Rightarrow$ 1), $E$ fulfills S10.
(1) $\thenif$:
From proposition \ref{p:invariant} follows that $\top$ is in $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ the terminal object,
$D\pro E$ is in $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ the product, and $D\expo E$ is in $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$.\\
That $D\expo E$ is the exponent in $\mbox{$\mathcal{R}$-$\mathbf{cat}$}$ follows from lemma \ref{l:evalS10}(1$\Rightarrow$ 2)
and proposition \ref{p:currynormal}.
(2) and (3) follow immediately from (1).
\qed
\section{Example of a natural domain that is no cdlubpo}\label{s:incomplete}
We give the counter-examples of this section in two stages of increasing complexity.
The first example shows that the lub-rule class $\mathcal{S}$ corresponding
to the axioms of natural domains is not complete in the sense of Section \ref{s:lubrule}.
It is simpler, because it refers to general, non-directed natural subsets.
The second example shows a natural domain whose directed natural subsets do not form
a cdlubpo.
It is more complicated, because it has to refer to \emph{directed} natural subsets.
Of course, the second example alone would be enough for all.
\begin{thm}
The lub-rule class $\mathcal{S}$ corresponding to the axioms of natural domains
is not complete.
This means that there is a natural domain that does not fulfill axiom S9 (closure)
for general (not necessarily directed) subsets.
\end{thm}
\proof
Here is the example, a finite natural domain $D$:
\begin{pspicture}(-0.2,-0.7)(3.0,2.0)
\firsto{d}{\S}{d}
\slineon{d}{b}{45}{1}{180}{b}
\slineo{b}{e}{315}{1}{\S}{e}
\slineo{e}{c}{45}{1}{0}{c}
\slineon{b}{a}{45}{1}{90}{a}
\sline{a}{c}
\end{pspicture}
It has as (non-directed) natural lubs $\{b,c\}\to a$,
$\{d,e\}\to b$ and all natural lubs that can be deduced
from these by the axioms of natural domains:
$\{b,c,d\}\to a$, $\{b,c,e\}\to a$, $\{b,c,d,e\}\to a$
and all trivial natural lubs where the lub is an element of the natural subset.
But by a complete axiom system the further natural lub $\{d,c\}\to a$ could be deduced,
because $a\in \operatorname{cl}_{\{\{b,c\},\{d,e\}\}} \{d,c\}$.
\qed
It might still be that the natural domains
coincide with the cdlubpos,
the closed dlubpos with complete axiom system.
But we have the theorem:
\begin{thm}
There is a natural domain $D$ whose directed natural subsets do not form a cdlubpo.
\end{thm}
\proof
Here is the example,
an infinite natural domain $D$ that is no cdlubpo, see figure \ref{f:incomplete}.
\newcommand{\bcbc}[9]%
{\dlineo{#1}{\noden2}{\S}{1}{180}{#2}
\slineon{\noden2}{\noden4}{22.5}{1.5}{0}{#4}%
\dlineo{\noden4}{\noden5}{\S}{1}{0}{#5}%
\slineo{\noden5}{\noden6}{\S}{1}{0}{#6}%
\nexton{\noden6}{\noden7}{\S}{1}{0}{#7}%
\slineo{\noden7}{\noden3}{202.5}{1.5}{180}{#3}%
\sline{\noden2}{\noden3}%
\dlineo{\noden7}{\noden8}{\S}{1}{0}{#8}%
\slineo{\noden8}{\noden9}{\S}{1}{0}{#9}}
\begin{figure}[htb]
\begin{pspicture}(-0.7,-11.5)(10,4)
\firston{b1}{90}{b_1}
\dlineo{b1}{c12}{\S}{1}{180}{c_{12}}
\slineo{c12}{c11}{\S}{1}{180}{c_{11}}
\slineon{b1}{b2}{22.5}{2}{90}{b_2}
\newcommand{d}{d}
\bcbc{b2}{c_{22}}{c_{21}}{b_{22}}{c_{222}}{c_{221}}{b_{21}}{c_{212}}{c_{211}}
\slineon{b2}{b3}{22.5}{3.5}{90}{b_3}
\dlineo{b3}{c32}{\S}{1}{180}{c_{32}}
\slineon{c32}{b32}{22.5}{1.5}{0}{b_{32}}
\renewcommand{d}{e}
\bcbc{b32}{c_{322}}{c_{321}}{b_{322}}{c_{3222}}{c_{3221}}{b_{321}}{c_{3212}}{c_{3211}}
\slineo{c32}{c31}{\S}{7}{180}{c_{31}}
\slineon{c31}{b31}{22.5}{1.5}{0}{b_{31}}
\renewcommand{d}{f}
\bcbc{b31}{c_{312}}{c_{311}}{b_{312}}{c_{3122}}{c_{3121}}{b_{311}}{c_{3112}}{c_{3111}}
\dlineon{b3}{a}{22.5}{4}{0}{a \text{ top}}
\end{pspicture}
\caption{natural domain $D$ (partial)}\label{f:incomplete}
\end{figure}
The elements of $D$ are:\\
$a$ the top element (i.e.\ $x\lex a$ for all $x\in |D|$).\\
The elements of $B=\{b_i\}_{i\geq 1}$ with $b_i\lex b_j$ for $i\leq j$, it is $B\to a$.\\
For every $n\geq 1$ there are the elements $b_{nw}$,
where $w$ is a word of numbers $\geq 1$ with $1\leq \operatorname{length}(w)<n$,\\
and the elements $c_{nw}$, where $w$ is a word of numbers $\geq 1$ with $1\leq \operatorname{length}(w)\leq n$.\\
(Words of numbers will always be denoted by $w$.)\\
It is $c_{w1}\lex c_{w2}\lex \dots \lex b_w$
and $\{c_{w1}, c_{w2},\dots\}\to b_w$ and $c_w\lex b_w$ for all valid words $w$.\\
Let $C=\set{c_{nw}}{n\geq 1, \operatorname{length}(w)=n}$,
these are the $c_{nw}$ in the rightmost position under each $b_n$.
$C$ is not directed, it has lub $a$.\\
Like the example in section \ref{s:sazonov}, we erect an ``artificial'' directed set $\overline{C}$ on $C$,
which is not depicted in the diagram:\\
Let $\overline{C}$ be the set of all finite subsets of $C$.
All elements of $\overline{C}$ are also elements of $D$.\\
For all $A,B\in \overline{C}$: $A\lex_D B$ iff $A\sub B$, and of course $A\lex_D a$.\\
For all $x\in C$: $x\lex_D \{x\}\in \overline{C}$.
$\overline{C}$ is directed with lub $a$.\\
The natural domain $D$ is the reflexive, transitive closure of the order $\lex$ and the
closure of $\to$ under the axioms of natural domain.
The intuition of the example is this:
To deduce $C\to a$ (and $\overline{C}\to a$) one would start with $B\to a$ and stepwise
replace in $B$ elements $b_w$ by $\{c_{wi}\}_{i\geq 1}$
and elements $c_w$ by $b_w$,
working from left to right under each of the $b_i$,
until we reach the elements of $C$ and have replaced all other elements.
Under $b_i$ this takes at least $2i-1$ sequential steps,
so no deduction can do this for all $b_i$.
(There is no separate deduction of natural lubs $b_i$.
$X\to b_i$ for $X\sub \{c_{ij}\}_{j\geq 1}$ infinite are the only non-trivial natural lubs $b_i$.)
Therefore we have:
\begin{prop}
In $D$ it is not $C\to a$ and not $\overline{C}\to a$.
\end{prop}
\proof
In the following ``non-trivial natural lub'' means that the natural subset does not contain its own lub.
The only non-trivial natural lubs $\neq a$ are the $C'\to b_w$ with a valid word $w$ and
where $C'$ contains an infinite subset of $\{c_{wj}\}_{j\geq 1}$.
This is so because in such a natural set no $c_{wj}$
can be replaced by $b_{wj}$, as it is not $b_{wj}\lex b_w$.
\medskip
We will prove that every non-trivial $A\to a$ fulfills this \emph{condition cond}:\\
There is some $n\geq 0$, and an infinite set $N$ of numbers $\geq 1$,
and a map $d\typ N\fun A$,
such that for every $i\in N$ it is $d(i)=b_{iw}$ or $d(i)=c_{iw}$ for some word $w$ with $\operatorname{length}(w)\leq n$.
\medskip
Here is the proof:\\
Every $A\to a$ must be deduced in a chain
$B=A_1\stackrel{\sqsubset}{\leftarrow} A_2\stackrel{\sqsubset}{\leftarrow} \dots \stackrel{\sqsubset}{\leftarrow} A_m = A$,
where $A_i\to a$ is non-trivial.
See the definition \ref{d:under} of $\stackrel{\sqsubset}{\leftarrow}$ and proposition \ref{p:under}
and note that $a$ is the top element.\\
$B$ fulfills the condition cond.
We have to prove that if $A$ fulfills the condition cond, then also $A'$ with $A\stackrel{\sqsubset}{\leftarrow} A'$.
A step $A\stackrel{\sqsubset}{\leftarrow} A'$ means the following:
\begin{enumerate}[(1)]
\item
For every $b_i\in A$ (there is some $b_j\in A'$ with $j\geq i$)
or (there is $C'\sub A'$ with an infinite $C'\sub \{c_{ij}\}_{j\geq 1}$).\\
\item
For every $b_{iw}\in A$, with $\operatorname{length}(w)\geq 1$, (there is $b_{iw}\in A'$)
or (there is $C'\sub A'$ with an infinite $C'\sub \{c_{iwj}\}_{j\geq 1}$).\\
\item
For every $c_{ij}\in A$, (there is some $c_{ik}\in A'$ with $k\geq j$)
or (there is some $b_k\in A'$ with $k\geq i$)
or $b_{ij} \in A'$).\\
\item
For every $c_{iwj}\in A$, with $1\leq \operatorname{length}(w)\leq i-2$,
(there is some $c_{iwk}\in A'$ with $k\geq j$)
or ($b_{iw}\in A'$) or ($b_{iwj}\in A'$).\\
\item
For every $c_{iwj}\in A$, with $\operatorname{length}(w)=i-1$,
(there is some $c_{iwk}\in A'$ with $k\geq j$)
or ($b_{iw}\in A'$)
or (there is some $F\in A'$ with $F\in \overline{C}$ and $c_{iwj}\in F$).\\
\item
For every $F\in A$, with $F\in \overline{C}$,
there is $F'\in A$ with $F'\in \overline{C}$ and $F\lex F'$.
\end{enumerate}
We prove that $A'$ fulfills the condition cond. There are two cases:\\
(a) $A'\cap B$ is infinite:\\
Then $A'$ fulfills the condition cond with $n'=0$,
$N'=\set{i}{b_i\in A'\cap B}$, $d'(i)=b_i$.\\
(b) otherwise:\\
Then $A'$ fulfills the condition cond with the following data:\\
$n'=n+1$.\\
$N'$ is the set of all $i\in N$ with $i\geq n+1$ that fulfill the following two conditions:
\begin{itemize}
\item If $d(i)=b_i$, then $b_i\in A'$ or some $c_{ij}\in A'$.
\item If $d(i)=c_{ij}$, then (some $c_{ik}\in A'$ with $k\geq j$)
or ($b_{ij}\in A'$) or ($b_i\in A'$).
\end{itemize}
(These conditions rule out the cases of $i$ where $b_i$ or $c_{ij}$ is replaced in $A'$
only by a $b_k$ with $k>i$.
$N'$ is still infinite as the $i\in N$ with $i\geq n+1$ are infinitely many
and the ruled out cases of $i$ can only be finitely many,
otherwise we would have case (a).)
$d'(i)$ is defined for $i\in N'$ as follows:
\begin{itemize}
\item
If $d(i)=b_{iw}$:
if $b_{iw}\in A'$ then $d'(i)=b_{iw}$ else $d'(i)=c_{iwj}$ for some $j$ such that $c_{iwj}\in A'$.
\item
If $d(i)=c_{iwj}$:
if there is some $c_{iwk}\in A'$ then $d'(i)=c_{iwk}$\\
else (if $b_{iw}\in A'$ then $d'(i)=b_{iw}$ else $d'(i)=b_{iwj}$).\\
(Here it is $i\geq n+1$ and $\operatorname{length}(wj)\leq n$, so $\operatorname{length}(w)<i-1$.
Therefore $c_{iwj}$ is not replaced in $A'$ by some $F\in \overline{C}$.)
\end{itemize}
It is clear from the meaning of the step $A\stackrel{\sqsubset}{\leftarrow} A'$ above and the choice of $N'$
that $d'(i)$ is an element of $A'$.
The word length of the index of $d'(i)$ stays the same as for $d(i)$ or increases by $1$.
So we have proved that every non-trivial $A\to a$ fulfills the condition cond.
This proves that $C\to a$ and $\overline{C}\to a$ are not fulfilled in $D$.
\qed
\begin{prop}
From the specified natural lubs of $D$ it can be deduced by the closure axiom S9
that $\overline{C}\to a$.
\end{prop}
\proof
Let $C^{\ast}=\operatorname{cl}_P(\overline{C})$, where $P$ is the specified set of natural sets of $D$.
We have to show that $a\in C^{\ast}$.\\
First, for every $c\in C$ it is $c\inC^{\ast}$, as $c\lex_D \{c\}\in \overline{C}$.\\
For all $n\geq 1$, we now work under $b_n$:\\
We have that all $c_{nw}\in C^{\ast}$ with $\operatorname{length}(w)=n$, these are the elements from $C$.\\
If for some $1\leq k\leq n$, all $c_{nw}\in C^{\ast}$ with $\operatorname{length}(w)=k$,
then all $b_{nw}\in C^{\ast}$ with $\operatorname{length}(w)=k-1$.\\
If for some $1\leq k\leq n-1$, all $b_{nw}\in C^{\ast}$ with $\operatorname{length}(w)=k$,
then all $c_{nw}\inC^{\ast}$ with $\operatorname{length}(w)=k$.\\
As a result of this consecutive process,
we get $b_n\in C^{\ast}$ for all $n\geq 1$, so $a\in C^{\ast}$.
\qed
From the two propositions it is immediate that $D$ is not a cdlubpo.
\qed
\begin{rem}
Note that $D$ of the theorem is not a (naturally) algebraic natural domain
in the sense of definition 3.2 of \cite{Sazonov:natural}.
In the next section we show that every algebraic natural domain is a cdlubpo.
\end{rem}
\begin{rem}
Another question is that for an augmentation of the axioms of natural domains
so that they describe cdlubpos.
This could be a modification of axiom S8 (under)
so that the ``limit'' $L$ of an ascending $\stackrel{\sqsubset}{\leftarrow}$-chain $\lex a$
leads to a natural lub $L\to a$.
But this would need a complicated definition of ``limit'',
so we prefer our axiom S9 (closure) instead.
\end{rem}
\section{Algebraic dlubpos}
We adapt the definitions of (naturally) finite elements and (naturally) algebraic natural domains
from Sazonov \cite{Sazonov:natural}\ to the case of dlubpos.
We show that an algebraic dlubpo that fulfills axiom S6 (cofinality) is a cdlubpo.
This proves that (naturally) algebraic natural domains and algebraic cdlubpos are the same.
\begin{defi}[Sazonov, definitions 3.1, 3.2 of \cite{Sazonov:natural}]\label{d:alg}
Let $D$ be a dlubpo.\\
An element $a\in|D|$ is \defin{finite} if for every (directed) $B\to b$ with
$a\lex b$ there is $b'\in B$ with $a\lex b'$.
$D^0$ is the set of finite elements of $D$.\\
For $d\in|D|$ we define $\mathord{\down^0} d=\set{a\in D^0}{a\lex d}$.\\
$D$ is \defin{algebraic} if for every $d\in|D|$ it is $\mathord{\down^0} d\to d$.
(This includes that $\mathord{\down^0} d$ is directed.)
\end{defi}
Note that we omit Sazonov's prefix ``naturally'' (finite, algebraic)
as this is given by the fact that we are talking about dlubpos.
The next lemma shows that the finiteness property extends to closures.
\begin{lem}\label{l:fincl}
Let $D$ be a dlubpo.\\
Let $A\sub|D|$, $a\in D^0$ and $a\in \operatorname{cl}_D A$.
Then there is some $a'\in A$ with $a\lex a'$.
\end{lem}
\proof
Let $(N,r,\operatorname{\mathit{lab}},\operatorname{\mathit{pre}})$ be a deduction of $a$ from $A$ in the cl-rule system.
We construct a path
$r=n_1, n_2\in \operatorname{\mathit{pre}} n_1, n_3\in \operatorname{\mathit{pre}} n_2,\dots, n_k$
from the root $r$ to a leaf $n_k$ inductively as follows.
It is always $a\lex \operatorname{\mathit{lab}} n_i$.\\
If $\operatorname{\mathit{lab}} n_i=b$ and $\operatorname{\mathit{pre}} n_i=\{m\}$ with $b\lex \operatorname{\mathit{lab}} m$,
then we choose $n_{i+1}=m$.\\
If $\operatorname{\mathit{lab}} n_i=b$ and $\operatorname{\mathit{pre}} n_i=M$ with $\fset{\operatorname{\mathit{lab}}} M\to b$,
then there is some $m\in M$ with $a\lex \operatorname{\mathit{lab}} m$, as $a$ is finite.
We choose $n_{i+1}=m$.\\
By well-foundedness of the deduction, this process ends with a leaf $n_k$
with $\operatorname{\mathit{lab}} n_k=a'\in A$ and $a\lex a'$.
\qed
There is an interesting property of dlubpos connected to algebraicity:
\begin{defi}
Let $D$ be a dlubpo.
$D$ is \defin{finite-determined}
if there is a subset $F\sub |D|$ such that for every directed $A\sub|D|$ with lub $a$ it is:\\
$A\to a \ifff$ for all $b\in F$ with $b\lex a$ there is some $a'\in A$ with $b\lex a'$.
\end{defi}
The elements of $F$ are finite in the sense of definition \ref{d:alg}.
Every algebraic dlubpo $D$ fulfills the direction $\ifthen$ with $F=D^0$.
But a finite-determined $D$ need not be algebraic,
as it is not stipulated that every element $a$ of $D$ is a (natural) lub of the directed set
of finite elements below $a$.
On the other side our definition demands more,
namely the direction $\thenif$,
which is not entailed by algebraic dlubpos.
\begin{prop}
If $D$ is a finite-determined dlubpo and for every $a\in |D|$ it is
$a=\Lubex \mathord{\down^0} a$ directed,
then $D$ is algebraic.
\qed
\end{prop}
Remember that the axiom S6 (cofinality) for directed subsets is:\\
If $X,Y\sub|D|$ directed, $X\to x$ and $X\sqsubseteq Y\sqsubseteq x$, then $Y\to x$.
\begin{prop}[Reinhold Heckmann, personal communication]\label{p:algfin}
\hfill\\
If $D$ is an algebraic dlubpo with axiom S6 for directed subsets,
then $D$ is finite-determined with $F=D^0$.
\end{prop}
\proof
We must prove the direction $\thenif$ of the definition.\\
The right side means: $\mathord{\down^0} a\sqsubseteq A$.
It is $A\sqsubseteq a$ and $\mathord{\down^0} a\to a$.\\
By axiom S6 we get $A\to a$.
\qed
\begin{prop}\label{p:fincd}
If $D$ is a finite-determined dlubpo,
then $D$ is a cdlubpo.
\end{prop}
\proof
Let $D$ be finite-determined by the subset $F$.\\
Let $A\sub |D|$ be directed with lub $a$ and $a\in \operatorname{cl}_D A$.
We have to show $A\to a$.\\
Let $b\in F$ with $b\lex a$.
Then $b\in D^0$ and $b\in \operatorname{cl}_D A$.\\
By lemma \ref{l:fincl} there is some $a'\in A$ with $b\lex a'$.
\qed
\begin{thm}\hfill
\begin{enumerate}[(1)]
\item
If $D$ is an algebraic dlubpo with axiom S6 (cofinality) for directed subsets,
then $D$ is a cdlubpo.
\item
For any dlubpo $D$ the following are equivalent:\\
(a) $D$ is an algebraic dlubpo that fulfills axiom S6 for directed subsets.\\
(b) $D$ is an algebraic cdlubpo.\\
(c) $D$ is an algebraic natural domain.
\end{enumerate}
\end{thm}
\proof
(1) follows from propositions \ref{p:algfin} and \ref{p:fincd}.\\
(2) is clear.
\qed
\section{Restricted partial orders and restricted dcpos}
We want to establish cdlubpos as the most general canonical structures in which to build
non-complete programming language models,
corresponding to dcpos as the most general structures to build complete models.
(These structures still lack algebraicity/continuity.)
We have already seen that cdlubpos are just the dlubpos that are generated by a complete
lub-rule class.
In this section we see another characterization of cdlubpos that shows their canonicity:
they are just the dlubpos that are realized by ``restricted partial orders'' (rpos),
in which they are embedded.
This embedding can also be in a ``restricted dcpo'' (rdcpo),
which is a dcpo with a subset designated as the set of ``proper'' elements.
But before describing the rdcpos we give a property that is sufficient for
dlubpos to be cdlubpos.
\begin{prop}\label{p:detK}
Let $D$ be a dlubpo and $K$ be a subclass of the class of all
dlubpo morphisms with domain $D$.
(The morphisms of $K$ could be given by a subcategory in which $D$ lives.)\\
We say that \defin{``the natural lubs of $D$ are determined by $K$''}\\
if for all directed $A\sub |D|$ with lub $a$ it is:\\
if all $f\typ D\fun E$ in $K$ respect the lub of $A$ (i.e.\ $fa=\Lubex \fset{f} A$),
then $A\nlub_D a$.\\
In this case $D$ is a cdlubpo.
\end{prop}
\proof
Let $((|D|,\lex_D),P\mathord{\leadsto} A)$ be a valid directed lub-rule
and let all elements of $P$ be natural in $D$.
All $f\typ D\fun E$ in $K$ respect the lubs of the elements of $P$,
so also respect the lub of $A$, as the lub-rule is valid.\\
Therefore $A$ is natural in $D$ by the definition,
so $D$ fulfills the lub-rule.
Thus $D$ is a cdlubpo.
\qed
We now come to the realization of cdlubpos by restricted partial orders and restricted dcpos.
\begin{defi}
A structure $E=(|E|,\real{E},\lex_E)$ is a \defin{restricted partial order (rpo)} if\\
$|E|$ is a set of (proper) \defin{elements},\\
$\real{E}$ is a set of \defin{realizers}, with $|E| \sub \real{E}$,\\
$\lex_E$ is a partial order on $\real{E}$.\\
We define $\breal{E}=\real{E}\setminus|E|$, the set of \defin{blind realizers}.
\end{defi}
The idea of a partial order realized by a cpo appears in \cite{Simpson},
but there every realizer realizes exactly one proper element (there are no blind realizers),
and a proper element may be realized by several realizers.
\begin{defi}
An rpo $E$ \defin{realizes} a dlubpo $D$ \defin{by} an order isomorphism\\
$\varphi\typ (|D|,\lex_D)\fun (|E|,\lex_E)$
if for all directed $\Lubex A=a$ in $D$:
\[ A\nlub_D a \ifff \varphi a = \Lubex \fset{\varphi} A. \]
\end{defi}
\begin{prop}\label{p:rpo}\hfill
\begin{enumerate}[(1)]
\item
If an rpo $E$ realizes a dlubpo $D$, then $D$ is a cdlubpo.
\item
Every cdlubpo $D$ is realized by the rpo $(\fset{\operatorname{in}}|D|,\hat{D},\lex)$ by $\operatorname{in}\typ|D|\fun\fset{\operatorname{in}}|D|$,
see def.\ \ref{d:cl}.
\item
The cdlubpos are exactly the dlubpos realized by rpos.
\end{enumerate}
\end{prop}
\proof
\textbf{(1)}
This is essentially the proof of proposition \ref{p:detK},
when we construe $E$ as a dlubpo with all directed lubs natural.\\
Let $\varphi$ be the embedding of $D$ in $E$.
Let $((|D|,\lex),P\mathord{\leadsto} A)$ be a valid directed lub-rule such that the elements
of $P$ are natural in $D$.
$\varphi$ respects the lubs of the elements of $P$.
As the lub-rule is valid,
$\varphi$ respects also the lub of $A$,
so $\varphi(\Lubex A)=\Lubex\fset{\varphi} A$.\\
As $E$ realizes $D$ by $\varphi$, it is $A\to a$ in $D$.
So $D$ fulfills all valid directed lub-rules, it is a cdlubpo.\\
\textbf{(2)}
Let $A\to a$ in $D$.
Then $\operatorname{in}_D a = \Lubex(\fset{\operatorname{in}_D}A)=\operatorname{cl}_D A$ by proposition \ref{p:in}.\\
For the reverse:
It is $a\in\operatorname{cl}_D A$, therefore $A\to a$.
\qed
We want to realize our dlubpos in dcpos which can be constructed by a completion process.
We define the category of restricted dcpos:
\begin{defi}
A structure $E=(|E|,\real{E},\lex_E)$ is a \defin{restricted dcpo (rdcpo)}
if it is an rpo and $(\real{E},\lex_E)$ is a dcpo.\\
Let $D,E$ be rdcpos.
A function $f\typ\real{D}\fun\real{E}$ is \defin{continuous}
if it is continuous\ on the dcpos $(\real{D},\lex_D)$ and $(\real{E},\lex_E)$,
and $\fset{f}|D|\sub|E|$.
In this case we write $f\typ D\fun E$.\\
$\mathbf{Rdcpo}$ is the category of all rdcpos and continuous\ functions,
with normal function composition and identity functions.
\end{defi}
\begin{prop}
The cdlubpos are exactly the dlubpos realized by rdcpos.
\end{prop}
\proof
Follows from proposition \ref{p:rpo}, as $\hat{D}$ is a dcpo.
\qed
There is an adjunction between
the categories $\mathbf{Dlubpo}$ and $\mathbf{Rdcpo}$,
which establishes an adjoint equivalence between $\mathbf{Cdlubpo}$ and a full
subcategory $\mathbf{Crdcpo}$ (closed rdcpos) of $\mathbf{Rdcpo}$.
The details and further developments in this direction will be found in a sequel paper.
\section{Outlook}\label{s:outlook}
We have introduced lub-rule classes and closed dlubpos corresponding to
complete lub-rule classes as a canonical alternative to natural domains.
Closed dlubpos can also be characterized as the dlubpos realized by restricted dcpos.
We will explore this connection with rdcpos further in a sequel paper.
It will turn out that there is an adjunction between the categories
which permits the transfer of much of the theory of cccs of algebraic dcpos
to closed rdcpos resp.\ closed dlubpos.
The different approaches to realize partial orders by dcpos,
Simpson's \cite{Simpson} and ours,
call for a common generalization:
in the new concept a proper element may be realized by \emph{several} realizers (Simpson),
and a realizer may realize one or none element (as here).
This would be a category of dcpos that carry also some kind of \emph{partial} preorder,
the correct definition of it is not yet clear.
(Simpson has a preorder.)
We indicate further directions of research.
The question of the topology of non-complete domains is not yet fully answered,
Sazonov made some first observations on this in \cite{Sazonov:natural}.
In the introduction we posed as an open problem to find categories of abstract
incomplete domains with the existence of fixpoints of endofunctions.
Also interesting is the question to find such categories of concrete incomplete domains,
i.e.\ domains based on mechanisms like games.
\section*{Acknowledgement}
I thank Reinhold Heckmann for very carefully reading the drafts of this paper and many fruitful discussions
and hints. Proposition \ref{p:algfin} is due to him.\\
I thank Reinhard Wilhelm, Sebastian Hack and the members of their chair for their support,
esp.\ Roland Lei\ss a and Klaas Boesche for help with the computer.
|
\chapter{Influence Of Measure}\label{chap:measure_analysis}
In this chapter, we study the influence of measure of the set where $\Omega$-result holds, on its possible improvements. The
following proposition is an interesting application of the main theorem (Theorem~\ref{thm:omega_pm_measure}) of this chapter.
Let $\Delta(x)$ denotes the error term appearing in the assymptotic formula for average order of non-isomorphic abelian groups:
\begin{equation}\label{eq:non_isomorphic_ab_gr_delta}
\Delta(x)=\sum_{n\leq x}^*a_n - \sum_{k=1}^{6}\Big(\prod_{j \neq k}\zeta(j/k)\Big) x^{1/k},
\end{equation}
where $a_n$ denotes the number of non-isomophic abelian groups of order $n$.
One would expect that
\[\Delta(x)=O\left(x^{1/6+\epsilon}\right)\ \text{ for any } \epsilon>0\]
(see Section~\ref{sec:sub_abelian_group} for more details), so an analogus $\Omega_\pm$ bound for $\Delta(x)$ is also
expected. The proposition below gives a sufficient condition to obtain such an $\Omega_\pm$ bound.
\begin{prop}\label{prop:abelian_group}
Let $\delta$ be such that $0<\delta<1/42$, and $\Delta(x)$ be as in (\ref{eq:non_isomorphic_ab_gr_delta}). Then
either
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x=\Omega( T^{5/3+\delta} ),\]
or
\[\Delta(x)=\Omega_\pm(x^{1/6-\delta}).\]
\end{prop}
It may be conjectured that
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x=O( T^{5/3+\epsilon} )\]
for any $\epsilon>0$. By the above proposition, this conjecture implies
\[\Delta(x)=\Omega_\pm(x^{1/6-\epsilon}) \ \text{ for any } \ \epsilon>0.\]
We begin by assuming the conditions and notations given in Assumptions~\ref{as:for_continuation_mellintran}. Further we have the
following notations for this chapter.
\begin{notations}
For $i=0, 1, 2$, let $\alpha_i(T)$ denote positive monotonic functions such that
$\alpha_i(T)$ converges to a constant as $T\rightarrow \infty$. For example, in some cases
$\alpha_i(T)$ could be $1-1/\log(T)$, which tend to $1$ as $T\rightarrow \infty$.
For $i=0, 1$, let $h_i(T)$ be positive monotonically increasing functions such that $h_i(T)\rightarrow\infty$
as $T\rightarrow\infty$.
For a real valued and non-negative function $f$, we denote
\[\mathcal{A}(f(x)):=\{x\geq 1: |\Delta(x)|>f(x)\}.\]
\end{notations}
\section{Refining Omega Result from Measure}
Define an $\textbf{X}\text{-Set}$ as follows.
\begin{defi}
An infinite unbounded subset $\mathcal{S}$ of non-negative real numbers is called an $\textbf{X}\text{-Set}$ .
\end{defi}
Now we hypothesize a situation when there is a lower bound estimate for the second moment of the error term.
\begin{asump}\label{as:measure_to_omega}
Let $\mathcal{S}$ be an $\textbf{X}\text{-Set}$. Define a real valued positive bounded function $\alpha(T)$ on $\mathcal{S}$, such that
\[0\leq \alpha(T)<M<\infty\]
for some constant $M$. For a fixed
$T$, we write
\[\mathcal{A}_T:=[T/2, T]\cap\mathcal{A}(c_5 x^{\alpha(x)}) \ \text{ for } c_5>0.\]
For all $T\in \mathcal{S}$ and for constants $c_6, \ c_7 > 0$, assume the following bounds hold:
\begin{enumerate}
\item[(i)] \[\int_{\mathcal{A}_T}\frac{\Delta^2(x)}{x^{2\alpha+1}}\mathrm{d}x>c_6,\]
\item[(ii)]
\[\mu(\mathcal{A}_T)<c_7h_0(T),\quad \text{ and }\]
\item[(iii)] the function
\[x^{\alpha+1/2}h_0^{-1/2}(x)\]
is monotonically increasing for $x\in [T/2, T]$.
\end{enumerate}
\end{asump}
We note that the first assumption indicates an $\Omega$-estimate. The next two assumptions indicate that the measure of
the set on which the $\Omega$ estimate holds is not \lq too big\rq.
\begin{prop}\label{prop:refine_omega_from_measure}
Suppose there exists an $\textbf{X}\text{-Set}$ $\mathcal{S}$ for $\Delta(x)$, having properties as described in Assumptions~\ref{as:measure_to_omega}.
Let the constant $c_8$ be given by
\[c_8:=\sqrt{\frac{c_6}{2^{2M+1}c_7}}.\]
Then there exists a $T_0$ such that for all $T>T_0$ and $T\in \mathcal{S}$, we have
\[|\Delta(x)|>c_8 x^{\alpha+1/2}h_0^{-1/2}(x)\]
for some $x\in [T/2, T]$.
In particular
\[\Delta(x)=\Omega(x^{\alpha+1/2}h_0^{-1/2}(x)).\]
\end{prop}
\begin{proof}
If the statement of the above proposition is not true, then for all $x\in [T/2, T]$ we have
\[\Delta(x)\leq c_8 x^{\alpha + 1/2}h_0^{-1/2}(x).\]
From this, we may derive an upper bound for second moment of $\Delta(x)$:
\begin{align*}
\int_{\mathcal{A}_T}\frac{\Delta^2(x)}{x^{2\alpha+1}}\mathrm{d}x &\leq
\frac{c_8^2 T^{2\alpha+1}\mu(\mathcal{A}_T\cap[T/2, T])}{h_0(T)(T/2)^{2\alpha+1}} \\
&\leq c_8^2 2^{2M + 1} c_7 \leq c_6.
\end{align*}
This bound contradicts (i) of Assumptions~\ref{as:measure_to_omega}, which proves the proposition.
\end{proof}
The above proposition will be used in the next chapter to obtain a result on the error term appearing in the asymptotic formula
for $\sum_{n\leq x}^*|\tau(n, \theta)|^2$.
\section{Omega Plus-Minus Result from Measure}
In this section, we prove an $\Omega_\pm$ result for $\Delta(x)$ when $\mu(A_T)$ is big.
We formalize the conditions in the following assumptions.
\begin{asump}\label{as:measure_omega_plus_minus}
Suppose the conditions in Assumptions~\ref{as:for_continuation_mellintran} hold.
Let $l$ be an integer such that
\[l>\max(\sigma_2, 1),\]
and let $\alpha_1(T)$ be a monotonic function satisfying the inequality
\[0<\alpha_1(T)\leq \sigma_1.\]
Furthermore:
\begin{enumerate}
\item[(i)] the Dirichlet series $D(\sigma+it)$ has no pole when $\alpha_1(T)\le \sigma\le \sigma_1$;
\item[(ii)] if $|t|\leq T^{2l}$ and $\alpha_1(T)\le \sigma\le \sigma_1$, we have
\[|D(\sigma + it)|\leq c_9 (|t|+1)^{l-1}\]
for some constant $c_9>0$.
\end{enumerate}
\end{asump}
\begin{asump}\label{as:measure_omega_plus_minus_weak}
Suppose there exists $\epsilon>0$ such that the following holds:
\begin{flushleft}
if $D(\sigma+it)$ has no pole for $\alpha_1(T)-\epsilon< \sigma \le \sigma_1$ and $|t|\leq 2T^{2l}$, then
there exists a constant $c_{10}>0$ depending on $\epsilon$ such that
\[|D(\sigma + it)|\leq c_{10} (|t|+1)^{l-1}, \]
when $\alpha_1(T)\le \sigma \le \sigma_1$ and $|t|\leq T^{2l}$.
\end{flushleft}
\end{asump}
Assumptions~\ref{as:measure_omega_plus_minus_weak} says that if $D(s)$ does not have pole in
$\alpha_1(T)-\epsilon< \sigma \le \sigma_1,$ then it has polynomial growth in a certain region.
\begin{lem}\label{lem:perron_for_omegapm_measure}
Under the conditions in Assumptions~\ref{as:measure_omega_plus_minus}, we have
\[\Delta(x)
=\int_{\alpha_1-iT^{2l}}^{\alpha_1+iT^{2l}}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta + O(T^{-1}),\]
for all $x\in [T/2, 5T/2]$.
\end{lem}
\begin{proof}Follows from Perron summation formula.
\end{proof}
\begin{lem}[Balasubramanian and Ramachandra \cite{BaluRamachandra2}]\label{lem:ramachandra_trick}
Let $T\ge 1,$ $\delta_0>0$ and $f(x)$ be a real-valued integrable function such that
\[f(x)\geq0 \quad \text{ for } x\in [T-\delta_0T, \ 2T+ \delta_0T].\]
Then for $\delta>0$ and for a positive integer $l$ satisfying
$\delta l\leq \delta_0,$
we have
\[\int_T^{2T}f(x)\mathrm{d} x \leq \frac{1}{(\delta T)^l}\underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}f(x)\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l.\]
\end{lem}
\begin{proof}
For $0\le y_i \le \delta T$, $i=1,2,...,l$
\begin{align*}
\int_T^{2T}f(x)\mathrm{d} x \leq \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}f(x)\mathrm{d}{x},
\end{align*}
as $f(x)\ge 0$ in
\[\left[T-\sum_{1}^l y_i, 2T+\sum_{1}^l y_i\right]\subseteq [T-\delta_0T, 2T+ \delta_0T].\]
This gives
\begin{align*}
&\frac{1}{(\delta T)^l}\underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}f(x)\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l \\
& \geq \frac{1}{(\delta T)^l}\underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T}^{2T}f(x)\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l
= \int_{T}^{2T}f(x)\mathrm{d}{x}.
\end{align*}
\end{proof}
The next theorem shows that if $\Delta(x)$ does not change sign then the set on which $\Omega$-estimate holds can not be \lq too big\rq.
\begin{thm}\label{thm:upper_bound_measure}
Suppose the conditions in Assumptions~\ref{as:measure_omega_plus_minus} hold. Let $h_1(T)$ be a monotonically
increasing function such that $h_1(T)\rightarrow\infty$. Let $\alpha_2(T)$ be a bounded positive monotonic function, such that
\begin{align*}
& 0<\alpha_1(T)<\alpha_2(T)\leq \sigma_1, \text{ and} \\
& \frac{h_1(T)}{T^{\alpha_1}}\rightarrow \infty \text{ as } T\rightarrow \infty.
\end{align*}
If there exist a constant $x_0$ such that $\Delta(x)$ does not change sign on $\mathcal{A}(h_1(x))\cap[x_0, \infty)$, then
\[\mu(\mathcal{A}(x^{\alpha_2(x)})\cap [T, 2T])\leq 4h_1(5T/2)T^{1-\alpha_2(T)}+ O(1 + T^{1-\alpha_2(T) + \alpha_1(T)})\]
for $T\geq 2x_0$.
\end{thm}
\begin{proof}
Trivially we have
\[\mu(\mathcal{A}(x^{\alpha_2})\cap [T, 2T])\leq \int_T^{2T}\frac{|\Delta(x)|}{x^{\alpha_2}}\mathrm{d} x .\]
Using Lemma~\ref{lem:ramachandra_trick} on the above inequality, we get
\[\mu(\mathcal{A}(x^{\alpha_2})\cap [T, 2T])\leq
\frac{1}{(\delta T)^l}\underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}\frac{|\Delta(x)|}{x^{\alpha_2}}\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l,\]
where $\delta=\frac{1}{2l}.$
Let $\chi$ denote the characteristic function of the complement of $\mathcal{A}(h_1(x))$:
\[\chi(x)=\begin{cases}
1 \quad \mbox{ if } x \notin \mathcal{A}(h_1(x)),\\
0 \quad \mbox{ if } x \in \mathcal{A}(h_1(x)).
\end{cases}
\]
For $T\geq 2x_0$, $\Delta(x)$ does not change sign on
$$\left[T-\sum_{1}^l y_i, \ 2T+\sum_{1}^l y_i\right]\cap\mathcal{A}(h_1(x)),$$
as $0\le y_i\le \delta T$ for all $i=1,...,l$.
So we can write
the above inequality as
\begin{align} \label{eq:measure_omega_pm_first_bound}
\notag
\mu(\mathcal{A}(x^{\alpha_2})\cap [T, 2T])
&\leq \frac{2}{(\delta T)^l}
\underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}\frac{|\Delta(x)|}{x^{\alpha_2}}\chi(x)\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l\\
&+ \frac{1}{(\delta T)^l}
\left| \underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}\frac{\Delta(x)}{x^{\alpha_2}}\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l\right|.
\end{align}
Since $x\notin \mathcal{A}(h_1(x))$ implies $|\Delta(x)|\le h_1(x)$, we get
\begin{align}\label{eq:measure_omega_pm_trivial_part}
\notag
\frac{2}{(\delta T)^l}&
\underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}\frac{|\Delta(x)|}{x^{\alpha_2}}\chi(x)\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l \\
&\leq 4 h_1(5T/2) T^{1-\alpha_2}.
\end{align}
We use the integral expression for $\Delta(x)$ as given in Lemma~\ref{lem:perron_for_omegapm_measure}, and get
\begin{align}\label{eq:measure_omega_pm_perron_part}
\notag
& \frac{1}{(\delta T)^l}
\left| \underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}\frac{\Delta(x)}{x^{\alpha_2}}\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l\right|\\
\notag
&= \frac{1}{(\delta T)^l}\left| \underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i}
\int_{\alpha_1-iT^{2l}}^{\alpha_1+iT^{2l}}\frac{D(\eta)x^{\eta-\alpha_2}}{\eta}\mathrm{d}\eta~\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l\right| +O(1) \\
\notag
&\ll 1 + \frac{1}{(\delta T)^l}\left| \int_{\alpha_1-iT^{2l}}^{\alpha_1+iT^{2l}}\frac{D(\eta)}{\eta}
\underset{l~\mathrm{ times }\quad}{\int_0^{\delta T}\cdots\int_0^{\delta T}} \int_{T-\sum_{1}^l y_i}^{2T+\sum_{1}^l y_i} x^{\eta-\alpha_2}\mathrm{d}{x}~\mathrm{d} y_1 \ldots \mathrm{d} y_l~\mathrm{d}\eta \right| \\
\notag
&\ll 1 + \frac{1}{(\delta T)^l}\left| \int_{\alpha_1-iT^{2l}}^{\alpha_1+iT^{2l}}
\frac{D(\eta)(2T+l\delta T)^{\eta-\alpha_2 + l + 1}}{\eta\prod_{j=1}^{l+1}(\eta-\alpha_2+j)}\mathrm{d}\eta\right|\\ \notag
&\ll 1 + \frac{T^{\alpha_1-\alpha_2+l + 1}}{(\delta T)^l}
\int_{-T^{2l}}^{T^{2l}}\frac{(1+|t|)^{l-1}}{(1+ |t|)^{l+2}}\mathrm{d} t
\ll 1 + T^{1-\alpha_2+\alpha_1}.\\
\end{align}
The theorem follows from (\ref{eq:measure_omega_pm_first_bound}), (\ref{eq:measure_omega_pm_trivial_part})
and (\ref{eq:measure_omega_pm_perron_part}).
\end{proof}
\begin{thm}\label{thm:omega_pm_measure}
Consider $\alpha_1(T), \alpha_2(T), \sigma_1, h_1(T)$ as in Theorem~\ref{thm:upper_bound_measure},
and $\mathcal P$ as in Assumptions \ref{as:for_continuation_mellintran}.
Let $D(s)$ does not have a real pole in $[\alpha_1-\epsilon_0, \infty)-\mathcal{P}$, for some $\epsilon_0>0$.
Suppose there exists an $\textbf{X}\text{-Set}$ $\mathcal{S}$ such that for all $T\in \mathcal{S}$
\[\mu(\mathcal{A}(x^{\alpha_2})\cap [T, 2T])> 5h_1(5T/2)T^{1-\alpha_2}.\]
Then:
\begin{enumerate}
\item[(i)]under Assumptions~\ref{as:measure_omega_plus_minus}, we have
\[\Delta(x)=\Omega_\pm(h_1(x))\]
( In this case $\Delta(x)$ changes sign in $[T/2, 5T/2]\cap \mathcal{A}(h_1(x))$ for $T\in S$
and $T$ is sufficiently large.);
\item[(ii)]under Assumptions~\ref{as:measure_omega_plus_minus_weak}, we have
\[\Delta(x)=\Omega_\pm(x^{\alpha_1-\epsilon}), \quad \text{for any } \ \epsilon>0.\]
\end{enumerate}
\end{thm}
\begin{proof}
If the conditions in Assumptions~\ref{as:measure_omega_plus_minus} hold, then (i) follows from Theorem~\ref{thm:upper_bound_measure}.
To prove (ii), choose an $\epsilon$ such that $0<\epsilon<\epsilon_0$.
Now suppose
$\eta_0$ is a pole of $D$ for $\text{Re}(\eta)\geq\alpha_1(T)-\epsilon$ and $t\leq 2T^{2l}$, then by Theorem~\ref{thm:landu_omegapm}
\[\Delta(x)=\Omega_\pm(T^{\alpha_1-\epsilon}).\]
If there are no poles in the above described region of $\sigma + it$, then we are in the set-up of
Assumptions~\ref{as:measure_omega_plus_minus}, and get
\[\Delta(x)=\Omega_\pm(h_1(x)).\]
We have
\[T^{\alpha_1(T)}=o(h_1(T)),\]
which gives
\[\Delta(x)=\Omega_\pm(x^{\alpha_1-\epsilon}).\]
This completes the proof of (ii).
\end{proof}
\section{Applications}\vspace{0.05mm}
Now we shall see two examples demonstrating applications of Theorem~\ref{thm:omega_pm_measure}.
\subsection{Error term of the divisor function}
Let $d(n)$ denote the number of divisors of $n$:
\[d(n)=\sum_{d|n}1.\]
Dirichlet \cite[Theorem~320]{HardyWright} showed that
\[\sum_{n\leq x}^*\tau(n) = x\log(x) + (2\gamma -1)x + \Delta(x), \]
where $\gamma$ is the Euler constant and
\[\Delta(x)=O(\sqrt{x}).\]
Latest result on $\Delta(x)$ is due to Huxley \cite{HuxleyDivisorProblem}, which is
\[\Delta(x)=O(x^{131/416}).\]
On the other hand, Hardy \cite{HardyDirichletDivisor} showed that
\begin{align*}
\Delta(x)&=\Omega_+((x\log x)^{1/4}\log\log x),\\
&=\Omega_-(x^{1/4}).
\end{align*}
There are many improvements of Hardy's result. Some notable results are due to K. Corr{\'a}di and I. K{\'a}tai \cite{CorradiKatai},
J. L. Hafner \cite{Hafner}, and K. Sounderarajan \cite{Sound}. Below, we shall show that $\Delta(x)$ is $\Omega_\pm(x^{1/4})$ as a consequence of Theorem~\ref{thm:omega_pm_measure} and
results of Ivi{\'c} and Tsang ( see below ).
Moreover, we shall how that such fluctuations occur in $[T, 2T]$ for every sufficiently large $T$.
\noindent Ivi{\'c} \cite{Ivic_second_moment_divisor_problem} proved that for a positive constant $c_{11}$,
\[\int_{T}^{2T}\Delta^2(x)\mathrm{d} x \sim c_{11} T^{3/2}.\]
A similar result for fourth moment of $\Delta(x)$ was proved by Tsang \cite{Tsang}:
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x\sim c_{12} T^2,\]
for a positive constant $c_{12}$.
Let $\mathcal{A}$ denote the following set:
\[\mathcal{A}:=\left\{x:|\Delta(x)|>\frac{c_{11} x^{1/4}}{6}\right\}.\]
For sufficiently large $T$, using the result of Ivi{\'c} \cite{Ivic_second_moment_divisor_problem}, we get
\begin{align*}
\int_{[T, 2T]\cap \mathcal{A}} \frac{\Delta^2(x)}{x^{3/2}}\mathrm{d} x &=\int_T^{2T}\frac{\Delta(x)^2}{x^{3/2}}\mathrm{d} x
-\int_{[T, 2T]\cap\mathcal{A}^c}\frac{\Delta^2(x)}{x^{3/2}}\mathrm{d} x\\
& \geq \frac{1}{4T^{3/2}}\int_T^{2T}\Delta^2(x)\mathrm{d} x -\frac{c_{11}}{6} \\
& \geq \frac{c_{11}}{5} -\frac{c_{11}}{6} \geq \frac{c_{11}}{30}.
\end{align*}
Using Cauchy-Schwarz inequality and the result due to Tsang \cite{Tsang} we get
\begin{align*}
\int_{[T, 2T]\cap \mathcal{A}} \frac{\Delta^2(x)}{x^{3/2}}\mathrm{d} x &\leq \left(\int_{[T, 2T]\cap\mathcal{A}}\frac{\Delta^4(x)}{x^2}\mathrm{d} x\right)^{1/2}
\left(\int_{[T, 2T]\cap\mathcal{A}}\frac{1}{x}\mathrm{d} x\right)^{1/2}\\
& \leq \left(\frac{c_{12}\mu([T, 2T]\cap \mathcal{A})}{T}\right)^{1/2}.
\end{align*}
The above lower and upper bounds on second moment of $\Delta$ gives the following lower bound for measure of $\mathcal{A}$:
\begin{equation*}
\mu([T, 2T]\cap \mathcal{A})> \frac{c_{11}^2}{901 c_{12}}T,
\end{equation*}
for some $T\geq T_0$.
Now, Theorem ~\ref{thm:omega_pm_measure} applies with the following choices:
\[\alpha_1(T)=1/5, \quad \alpha_2(T)=1/4, \quad h_1(T)=\frac{c_{11}^2}{9000c_{12}}T^{1/4}.\]
Finally using Theorem~\ref{thm:omega_pm_measure}, we get that for all $T\geq T_0$ there exists $x_1, x_2 \in [T, 2T]$ such that
\begin{align*}
\Delta(x_1)> h_1(x_1) \ \text{ and } \ \Delta(x_2)< - h_1(x_2).
\end{align*}
In particular we get
\begin{align*}
\Delta(x)=\Omega_\pm(x^{1/4}).
\end{align*}
\subsection{Average order of Non-Isomorphic abelian Groups}\label{sec:sub_abelian_group}
Let $a_n$ denote the number of non-isomorphic abelian groups of order $n$.
The Dirichlet series $D(s)$ is given by
\[D(s)=\sum_{n=1}^{\infty}\frac{a_n}{n^s} =\prod_{k=1}^{\infty}\zeta(ks),\quad \text{Re}(s)>1.\]
The meromorphic continuation of $D(s)$ has poles at $1/k$, for all positive integer $k\geq 1$. Let the
main term $\mathcal{M}(x)$ be
\[\mathcal{M}(x)=\sum_{k=1}^{6}\Big( \prod_{j\neq k} \zeta(j/k) \Big)x^{1/k},\]
and the error term $\Delta(x)$ be
\[\sum_{n\leq x}^* a_n - \mathcal{M}(x).\]
Balasubramanian and Ramachandra \cite{BaluRamachandra2} proved that
\begin{equation*}
\int_T^{2T}\Delta^2(x)\mathrm{d} x=\Omega(T^{4/3}\log T),
\text{ and }
\Delta(x)=\Omega_{\pm}(x^{92/1221}).
\end{equation*}
Sankaranarayanan and Srinivas \cite{srini} improved the $\Omega_\pm$ bound to
\[ \Delta(x)=\Omega_{\pm}\left(x^{1/10}\exp\left(c\sqrt{\log x}\right)\right)\]
for some constant $c>0$.
An upper bound for the second moment of $\Delta(x)$ was first given by Ivi{\'c} \cite{Ivic_abelian_group},
and then improved
by Heath-Brown \cite{Brown} to
\[\int_T^{2T}\Delta^2(x)\mathrm{d} x\ll T^{4/3}(\log T)^{89}.\]
This bound of Heath-Brown is best possible in terms of power of $T$.
But for the fourth moment, the similar statement
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x\ll T^{5/3}(\log T)^C,\]
which is best possible in terms of power of $T$, is an open problem.
Another open problem is to show that
\[\Delta(x)=\Omega_\pm(x^{1/6-\delta})\ \text{ for any }\ \delta>0.\]
For $0<\delta<1/42$, we have stated in Proposition~\ref{prop:abelian_group} that either
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x=\Omega( T^{5/3+\delta} )\ \text{ or }\ \Delta(x)=\Omega_\pm(x^{1/6-\delta}).\]
Below, we present a proof of this proposition.
\begin{proof}[Proof of Proposition~\ref{prop:abelian_group}]
If the first statement is false, then we have
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x\leq c_{13} T^{5/3+\delta}, \]
for some constant $c_{13}$ depending on $\delta$ and for all $T\geq T_0$.
Let $\mathcal{A}$ be defined by:
\[\mathcal{A}=\{x: |\Delta(x)|>c_{14}x^{1/6}\}, \quad c_{14}>0.\]
By the result of Balasubramanian and Ramachandra \cite{BaluRamachandra2}, we have an $\textbf{X}\text{-Set}$ $\mathcal{S}$, such that
\[\int_{[T, 2T]\cap\mathcal{A}}\Delta^2(x)\mathrm{d} x \geq c_{15}T^{4/3}(\log T)\]
for $T\in S$. Using Cauchy-Schwartz inequality, we get
\begin{align*}
c_{15}T^{4/3}(\log T)&\leq \int_{[T, 2T]\cap\mathcal{A}}\Delta^2(x)\mathrm{d} x
\leq \left(\int_T^{2T}\Delta^4(x)\mathrm{d} x\right)^{1/2}(\mu(\mathcal{A}\cap[T, 2T]))^{1/2}\\
&\leq c_{13}^{1/2}T^{5/6+\delta/2}(\mu(\mathcal{A}\cap[T, 2T]))^{1/2}.
\end{align*}
This gives, for a suitable positive constant $c_{16}$,
\[\mu(\mathcal{A}\cap[T, 2T])\geq c_{16}T^{1-\delta}(\log T)^2.\]
Now we use Theorem~\ref{thm:omega_pm_measure}, (i), with
\[\alpha_2=\frac{1}{6}, \quad \alpha_1=\frac{13}{84}-\frac{\delta}{2}, \quad \mbox{ and } \quad h_1(T)=T^{1/6-\delta}.\]
So we get
\[\Delta(x)=\Omega_\pm(x^{1/6-\delta}).\]
This completes the proof.
\end{proof}
\chapter{Introduction}\label{chap:intro}
In 1896, Jacques Hadamard and
Charles Jean de la Vall\'{e}e-Poussin proved that
the number of primes upto $x$ is asymptotic to $x/\log x$. This result is well known as the Prime Number Theorem (PNT).
Below we state a version of this theorem (PNT*) in terms of the von-Mangoldt function.
\begin{defi}
For $n\in \mathbb{N}$,
the von-Mangoldt function $\Lambda(n)$ is defined as
\[\Lambda(n)=\begin{cases}
\log p \quad &\mbox{ if } \ n=p^r, \ r\in \mathbb{N}\text{ and } p \text{ prime,}\\
0 & \mbox{ otherwise .}
\end{cases}
\]
\end{defi}
\noindent
\begin{thmu}[PNT*]
For a constant $c_1>0$, we have
\[\sum_{n\leq x}^*\Lambda(n)=x + O\left( x\exp\left(-c_1(\log x)^{3/5}(\log\log x)^{-1/5}\right)\right),\]
where
\[\sum_{n\leq x}^*\Lambda(n)=\begin{cases}
\sum_{n\leq x}\Lambda(n) \quad & \mathrm{ if } \ x\notin \mathbb{N},\\
\sum_{n\leq x}\Lambda(n)-\Lambda(x)/2 \quad& \mbox{ otherwise .}
\end{cases}
\]
\end{thmu}
\noindent
For a proof of the above theorem see \cite[Theorem~12.2]{IvicBook}.
Proof of PNT* uses analytic continuation of the function
\[\zeta(s)=\sum_{n=1}^{\infty}\frac{1}{n^s},\]
defined for $\text{Re}(s)>1$.
The function $\zeta(s)$ is called the \lq Riemann zeta function\rq, named after the famous German mathematician
Bernhard Riemann. In 1859, Riemann showed that this
has a meromorphic continuation to the whole complex plane. He also showed
PNT by assuming that the meromorphic continuation of $\zeta(s)$
does not have zeros for $\text{Re}(s)>\frac{1}{2}$. This conjecture of Riemann is popularly known
as the \lq Riemann Hypothesis\rq \ (RH), and is an unsolved problem. Under RH, the upper bound for $\Delta(x)$ in PNT*
can be improved as in the following theorem:
\begin{thmu}[PNT**]
Let $\Delta(x)$ be defined as in PNT*. Further, if we assume RH, then
\[\Delta(x)=O\left(x^{\frac{1}{2}}\log^2 x\right).\]
\end{thmu}
\begin{proof}
See \cite{PNT_Under_RH}.
\end{proof}
\noindent
In fact, we shall see in Theorem~\ref{thm:landu_omegapm} that PNT** is equivalent to RH.
At this point, it is natural to ask the following questions:
\texttt{
\begin{itemize}
\item[-] Can we obtain a bound for $\Delta(x)$, better than the bound in PNT**?
\item[-] Is $\Delta(x)$ an increasing or a decreasing function?
\item[-] Can $\Delta(x)$ be both positive and negative depending on $x$?
\item[-] How large are positive and negative values of $\Delta(x)$?
\end{itemize}}
\noindent
We shall make an attempt to answer these question by obtaining $\Omega$ and $\Omega_\pm$ results.
The following result was obtained by
Hardy and Littlewood \cite{HardyLittlewoodPNTOmegapm} in the year 1916:
\begin{align}
\Delta(x)=\Omega_\pm\left(x^\frac{1}{2}\log\log\log x\right).\
\end{align}
The above $\Omega_\pm$ bound on $\Delta(x)$ gives some answer to our earlier questions.
It says that we can not have an upper bound for $\Delta(x)$ which is smaller than $x^\frac{1}{2}\log\log\log x$.
It also says that $\Delta(x)$ often takes both positive and negative values with magnitude of order $x^\frac{1}{2}\log\log\log x$.
This suggests, it is important to obtain $\Omega$ and $\Omega_\pm$ bounds for various other error terms. In this direction, Landau's
theorem \cite{Landau} (see Theorem~\ref{thm:landu_omegapm} below) gives an elegant tool to obtain $\Omega_\pm$ results. Applying this theorem, we have
\begin{align*}
\Delta(x)=\Omega_\pm\left(x^\frac{1}{2}\right).
\end{align*}
The advantage of Landau's method as compared to Hardy and Littlewood's method is in its
applicability to a wide class of error terms of various summatory functions. In Landau's method, the existence of a complex pole
with real part $\frac{1}{2}$ serves as a criterion for the existence of above limits. In this thesis, we shall investigate on a
quantitative version of Landau's result by obtaining the Lebesgue measure of the sets where $\Delta(x)>\lambda x^{1/2}$ and
$\Delta(x)<-\lambda x^{\frac{1}{2}}$, for some $\lambda>0$. We shall show that the large Lebesgue measure of the set where
$|\Delta(x)|>\lambda x^{\frac{1}{2}}$, for some $\lambda>0$ replaces the criterion of existence of a complex pole in Landau's method.
This approach has the advantage of getting $\Omega_\pm$ results even when no such complex pole exists. This is evident from some
applications which we discuss in this thesis.
\section{Framework}
In this thesis, we consider a sequence of real numbers $\{a_n\}_{n=1}^{\infty}$ having Dirichlet series
\[D(s)=\sum_{n=1}^{\infty}\frac{a_n}{n^s}\]
that converges in some half-plane.
The Perron summation formula (see Theorem~\ref{thm:perron_formula}) uses analytic properties of $D(s)$ to give
\[\sum^*_{n\leq x}a_n=\mathcal{M}(x)+\Delta(x),\]
where $\mathcal{M}(x)$ is the main term, $\Delta(x)$ is the error term ( which would be specified later ) and $\sum^*$ is defined as
\begin{equation*}
\sum^*_{n\leq x} a_n =
\begin{cases}
\sum_{n\leq x} a_n \ & \text{ if } x\notin \mathbb N \\
\sum_{n\leq x} a_n -\frac{1}{2} a_x \ &\text{if } x\in\mathbb N.
\end{cases}
\end{equation*}
In Chapter~\ref{chap:analytic_continuation}, we analyze the Mellin transform $A(s)$ of $\Delta(x)$, which is defined as:
\begin{defi}\label{def:mellin_transform}
For a complex variable $s$, the Mellin transform $A(s)$ of $\Delta(x)$ is defined by
\[A(s)=\int_{1}^{\infty}\frac{\Delta(x)}{x^{s+1}}\mathrm{d} x.\]
\end{defi}
\noindent
In general, $A(s)$ is holomorphic in some half plane.
We shall discuss a method to obtain a meromorphic continuation of $A(s)$ from the meromorphic continuation of $D(s)$.
In particular, we shall prove in Theorem~\ref{thm:analytic_continuation_mellin_transform} that under some natural assumptions
\[A(s)=\int_{\mathscr{C}}\frac{D(\eta)}{\eta(s-\eta)}\mathrm{d} \eta,\]
where the contour $\mathscr{C}$ is as in Definition \ref{def:contour} and $s$ lies
to the right of $\mathscr{C}$. Later, this result will complement Theorem~\ref{thm:omega_pm_main} and Theorem~\ref{thm:omega_pm_main_new} in their applications.
In Chapter~\ref{chap:landau_theorem}, we revisit Landau's method and obtain measure theoretic results.
Also we generalize a theorem of Kaczorowski and Szyd{\l}o \cite{KaczMeasure}, and a theorem of Bhowmik, Ramar{\'e} and Schlage-Puchta \cite{gautami}
in Theorem~\ref{thm:omega_pm_main_new}.
Let
\[\mathcal{A}(\alpha, T):=\{x: x\in[T, 2T], |\Delta(x)|>x^{\alpha}\},\]
and let $\mu$ denotes the Lebesgue measure on $\mathbb R$.
In Chapter~\ref{chap:measure_analysis}, we establish a connection between $\mu(\mathcal{A}(\alpha, T))$ and
fluctuations of $\Delta(x)$. In Proposition~\ref{prop:refine_omega_from_measure}, we see
that
\[\mu(\mathcal{A}(\alpha, T))\ll T^{1-\delta} \ \text{ implies } \ \Delta(x)=\Omega(x^{\alpha + \delta/2}).\]
However, Theorem~{\ref{thm:omega_pm_measure}}
gives that
\[ \mu(\mathcal{A}(\alpha, T))=\Omega(T^{1-\delta}) \ \text{ implies } \ \Delta(x)
=\Omega_\pm(x^{\alpha-\delta}),\]
provided $A(s)$ does not have a real pole for $\text{Re} (s) \geq \alpha-\delta$. In particular,
this says that either we can improve on the $\Omega$ result or we can obtain a tight $\Omega_\pm$ result for $\Delta(x)$.
In Chapter~\ref{chap:twisted_divisor} we study a twisted divisor function defined as follows:
\begin{equation}\label{eq:tau-n-theta_def}
\tau(n, \theta)=\sum_{d\mid n}d^{i\theta}\ \text{ for } \ \theta\neq 0.
\end{equation}
This function is used in \cite[Chapter 4]{DivisorsHallTenen} to measure the clustering of divisors.
We give a brief note on some applications of $\tau(n, \theta)$ in Section~\ref{chap:twisted_divisor}.\ref{sec:applications_tau_n_theta}.
In \cite[Theorem 33]{DivisorsHallTenen}, Hall and Tenenbaum proved that
\begin{equation}\label{eq:formmula_tau_ntheta}
\sum_{n\leq x}^*|\tau(n, \theta)|^2=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)
+\omega_3(\theta)x + \Delta(x),
\end{equation}
where $\omega_i(\theta)$s are explicit constants depending only on $\theta$. They also showed that
\begin{equation}\label{eq:upper_bound_delta}
\Delta(x)=O_\theta(x^{1/2}\log^6x).
\end{equation}
We give a proof of this formula in Theorem~\ref{thm:asymp_formula_tau_n_theta}. Also, we derive
$\Omega$ and $\Omega_\pm$ bounds for $\Delta(x)$ using techniques from previous chapters.
In Theorem~\ref{omega_integral}, we obtain an $\Omega$ bound for the second moment of $\Delta(x)$ by adopting
a technique due to Balasubramanian, Ramachandra and Subbarao \cite{BaluRamachandraSubbarao}.
The main theorems of this thesis, except Theorem~\ref{thm:omega_pm_main_new}, are from \cite{MeasureOmega},
which is a joint work of the author with A. Mukhopadhyay.
\section{Applications}
Now we conclude the introduction by mentioning few applications of the methods given in this thesis.
\subsection{Twisted Divisors}
Consider the twisted divisor function $\tau(n, \theta)$ defined in the previous section.
The Dirichlet series of $|\tau(n, \theta)|^2$ can be expressed in terms of the Riemann zeta function as:
\begin{equation}\label{eq:dirichlet_series_tauntheta}
D(s)=\sum_{n=1}^{\infty}\frac{|\tau(n, \theta)|^2}{n^s}=\frac{\zeta^2(s)\zeta(s+i\theta)\zeta(s-i\theta)}{\zeta(2s)}
\quad\quad \text{for}\quad \text{Re}(s)>1.
\end{equation}
In Theorem~\ref{thm:asymp_formula_tau_n_theta}, we shall show
\begin{equation*
\sum_{n\leq x}^*|\tau(n, \theta)|^2=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)
+\omega_3(\theta)x + \Delta(x),
\end{equation*}
where $\omega_i(\theta)$s are explicit constants depending only on $\theta$
and
\begin{equation*
\Delta(x)=O_\theta(x^{1/2}\log^6x).
\end{equation*}
The Dirichlet series $D(s)$ has poles at $s=1, 1\pm i\theta $ and at the zeros of $\zeta(2s)$.
Using a complex pole on the line $\text{Re}(s)=1/4$,
Landau's method gives
\[\Delta(x)=\Omega_{\pm}(x^{1/4}).\]
In order to apply the method of Bhowmik, Ramar{\'e} and Schlage-Puchta, we need
\[ \int_T^{2T} \Delta^2(x) \mathrm{d} x \ll T^{2\sigma_0+1+\epsilon}, \]
for any $\epsilon >0$ and $\sigma_0=1/4$; such an estimate is not possible due to
Corollary~\ref{coro:balu_ramachandra1}.
Generalization of this method in Theorem~\ref{thm:omega_pm_main} can be applied to get
\begin{align*}
\mu \left( \mathcal{A}_j\cap [T, 2T]\right)=\Omega\left(T^{1/2}(\log T)^{-12}\right) \quad \text{ for } j=1, 2,
\end{align*}
and here $\mathcal{A}_j$s' for $\Delta(x)$ are defined as
\begin{align*}
\mathcal{A}_1=\left\{x: \Delta(x)>(\lambda(\theta)-\epsilon)x^{1/4}\right\}
\ \text{and} \ \mathcal{A}_2=\left\{ x : \Delta(x)<(-\lambda(\theta)+\epsilon)x^{1/4}\right\},
\end{align*}
for any $\epsilon>0$ and $\lambda(\theta)>0$ as in (\ref{eqn:lambda_theta}). But under Riemann Hypothesis, we show in
(\ref{result:tau_n_theta_omegapm_underRH}) that the above $\Omega$ bounds can be improved to
\begin{align*}
\mu\left(\mathcal{A}_j\right) =\Omega\left(T^{3/4-\epsilon}\right),\ \text{ for } j=1, 2 \ \text{ and for any } \ \epsilon>0.
\end{align*}
Fix a constant $c_2>0$ and define
\[\alpha(T) =\frac{3}{8}-\frac{c_2}{(\log T)^{1/8}}.\]
In Corollary~\ref{coro:balu_ramachandra2}, we prove that
\[\Delta(T)=\Omega\left(T^{\alpha(T)}\right).\]
In Proposition~\ref{Balu-Ramachandra-measure}, we give an $\Omega$ estimate for the measure of the sets involved in the above bound:
\[ \mu(\mathcal{A}\cap [T,2T])=\Omega\left(T^{2\alpha(T)}\right),\]
where
\[\mathcal{A}=\{x: |\Delta(x)|\ge Mx^{\alpha}\}\]
for a positive constant $M>0$.
In Theorem~\ref{thm:tau_theta_omega_pm}, we show
that
\[ \text{either } \ \Delta(x)=\Omega\left(x^{ \alpha(x)+\delta/2}\right)
\ \text{ or } \
\Delta(x)=\Omega_{\pm}\left(x^{3/8-\delta'}\right), \]
for $0<\delta<\delta'<1/8$.
We may conjecture that
\[ \Delta(x)=O(x^{3/8+\epsilon}) \ \text{ for any } \epsilon>0.\]
Theorem~\ref{thm:tau_theta_omega_pm} and this conjecture imply that
\[\Delta(x)=\Omega_{\pm}(x^{3/8-\epsilon})\ \text{ for any } \epsilon>0.\]
\subsection{Square Free Divisors}
Let $\Delta(x)$ be the error term in the asymptotic formula for partial sums of the square free divisors:
\begin{align*}
\Delta(x)=\sum_{n\leq x}^* 2^{\omega(n)}-\frac{x\log x}{\zeta(2)}+\left(-\frac{2\zeta'(2)}{\zeta^2(2)} + \frac{2\gamma - 1}{\zeta(2)}\right)x,
\end{align*}
where $\omega(n)$ denotes the number of distinct primes divisors of $n$. It is known that $\Delta(x)\ll x^{1/2}$ (see \cite{holder}).
Let $\lambda_1>0$ and the sets $\mathcal{A}_j$ for $j=1, 2$ be defined as in Section~\ref{subsec:sqfree_divisors}:
\begin{align*}
\mathcal{A}_1=\left\{x: \Delta(x)>(\lambda_1-\epsilon)x^{1/4}\right\},\ \text{ and } \
\mathcal{A}_2=\left\{ x : \Delta(x)<(-\lambda_1+\epsilon)x^{1/4}\right\}.
\end{align*}
In (\ref{eq:sqfree_divisors_omegaset_uc}), we show that
\begin{equation*}
\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{1/2}\right) \text{ for } j=1, 2.
\end{equation*}
But under Riemann Hypothesis, we prove the following $\Omega$ bounds in (\ref{eq:sqfree_divisors_omegaset}):
\begin{equation*}
\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{1-\epsilon}\right), \text{ for } j=1, 2 \ \text{ and for any } \ \epsilon>0.
\end{equation*}
\subsection{Divisors}
Let $d(n)$ denotes the number of divisors of $n$:
\[d(n)=\sum_{d|n}1.\]
Dirichlet \cite[Theorem~320]{HardyWright} showed that
\[\sum_{n\leq x}^*d(n) = x\log x + (2\gamma -1)x + \Delta(x), \]
where $\gamma$ is the Euler constant and
\[\Delta(x)=O(\sqrt{x}).\]
Latest result on $\Delta(x)$ is due to Huxley \cite{HuxleyDivisorProblem}, which is
\[\Delta(x)=O(x^{131/416}).\]
On the other hand, Hardy \cite{HardyDirichletDivisor} showed that
\begin{align*}
\Delta(x)&=\Omega_+((x\log x)^{1/4}\log\log x),\\
&=\Omega_-(x^{1/4}).
\end{align*}
There are many improvements on Hardy's result due to K. Corr{\'a}di and I. K{\'a}tai \cite{CorradiKatai},
J. L. Hafner \cite{Hafner} and K. Sounderarajan \cite{Sound}. As a consequence of Theorem~\ref{thm:omega_pm_measure},
we shall show in Chapter~\ref{chap:measure_analysis} that
for all sufficiently large $T$ and for a constant $c_3>0$, there exist $x_1, x_2 \in [T, 2T]$ such that
\begin{align*}
\Delta(x_1)> c_3x_1 \ \text{ and } \ \Delta(x_2)< - c_3x_2.
\end{align*}
In particular, we get
\begin{align*}
\Delta(x)=\Omega_\pm(x^{1/4}).
\end{align*}
\subsection{Error Term in the Prime Number Theorem}
Let $\Delta(x)$ be the error term in the Prime Number Theorem:
\[\Delta(x)=\sum_{n\leq x}^*\Lambda(n)-x.\]
We know from Landau's theorem \cite{Landau} that
\[\Delta(x)=\Omega_\pm\left(x^{1/2}\right)\]
and from the theorem of Hardy and Littlewood \cite{HardyLittlewoodPNTOmegapm} that
\[\Delta(x)=\Omega_\pm\left(x^{1/2}\log\log x\right).\]
We define
\begin{align*}
\mathcal{A}_1=\left\{x: \Delta(x)>(\lambda_2-\epsilon)x^{1/2}\right\} \
\text{ and } \ \mathcal{A}_2=\left\{ x : \Delta(x)<(-\lambda_2+\epsilon)x^{1/2}\right\},\\
\end{align*}
where $\lambda_2>0$ be as in Section~\ref{subsec:pnt_error}.
If we assume Riemann Hypothesis, then the theorem of Kaczorowski and Szyd{\l}o ( see Theorem~\ref{thm:kaczorowski} below )
along with PNT** gives
\[ \mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(\frac{T}{\log^4 T}\right) \text{ for } j=1, 2.\]
However, as an application of Corollary~\ref{cor:measure_omega_pm_from_upper_bound}, we prove the following weaker bound unconditionally:
\[ \mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{1-\epsilon}\right), \text{ for } j=1, 2\ \text{ and for any }\epsilon>0.\]
\subsection{Non-isomorphic Abelian Groups}
Let $a_n$ be the number of non-isomorphic abelian groups of order $n$, and
the corresponding Dirichlet series is given by
\[ \sum_{n=1}^{\infty}\frac{a(n)}{n^s} = \prod_{k=1}^{\infty}\zeta(ks)\ \text{ for } \text{Re}(s)>1. \]
Let $\Delta(x)$ be defined as
\begin{equation*}
\Delta(x)=\sum_{n\leq x}^*a_n - \sum_{k=1}^{6}\Big(\prod_{j \neq k}\zeta(j/k)\Big) x^{1/k}.
\end{equation*}
It is an open problem to show that
\begin{equation}\label{conj1}
\Delta(x)\ll x^{1/6+\epsilon} \ \text{ for any }\epsilon>0.
\end{equation}
The best result on upper bound of $\Delta(x)$ is
due to O. Robert and P. Sargos \cite{sargos}, which gives
\[\Delta(x)\ll x^{1/4+\epsilon}\ \text{ for any }\epsilon>0.\]
Balasubramanian and Ramachandra \cite{BaluRamachandra2} proved that
\begin{equation*}
\int_T^{2T}\Delta^2(x)\mathrm{d} x=\Omega(T^{4/3}\log T).
\end{equation*}
Following the proof of Proposition \ref{Balu-Ramachandra-measure}, we get
\begin{equation*}
\mu\left( \{ T\le x\le 2T: |\Delta(x)|\ge \lambda_3 x^{1/6}(\log x)^{1/2}\} \right)
=\Omega(T^{5/6-\epsilon}),
\end{equation*}
for some $\lambda_2>0$ and for any $\epsilon>0$.
Sankaranarayanan and Srinivas \cite{srini} proved that
\[ \Delta(x)=\Omega_{\pm}\left(x^{1/10}\exp\left(c\sqrt{\log x}\right)\right)\]
for some constant $c>0$.
It has been conjectured that
\[\Delta(x)=\Omega_\pm(x^{1/6-\delta})\ \text{ for any } \delta>0.\]
In Proposition~\ref{prop:abelian_group}, we prove that either
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x=\Omega( T^{5/3+\delta} )\text{ or }\Delta(x)=\Omega_\pm(x^{1/6-\delta}),\]
for any $0<\delta< 1/42$.
The conjectured upper bound (\ref{conj1}) of $\Delta(x)$ gives
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x \ll T^{5/3+\delta}.\]
This along with Proposition~\ref{prop:abelian_group} implies that
\[\Delta(x)=\Omega_\pm(x^{1/6-\delta})\ \text{ for any }\ 0<\delta <1/42.\]
\chapter{Landau's Oscillation Theorem}\label{chap:landau_theorem}
In this chapter, we revisit a result due to Landau and obtain $\Omega_\pm$ results for $\Delta(x)$ using certain singularities of $D(s)$.
Also we shall measure the fluctuations of $\Delta(x)$ in terms of $\Omega$ bounds, which generalizes a result of Kaczorowski and Szyd{\l}o \cite{KaczMeasure},
and a result of Bhowmik, Schlage-Puchta and Ramar{\'e} \cite{gautami}.
\section{Landau's Criterion for Sign Change}
We begin with a result on real valued functions that do not change sign. This appears in a paper of Landau \cite{Landau}, attributed
to Phragm{\'e}n and stated without a proof. Here we present a proof of this result following \cite[II.1.3, Theorem~6]{TenenAnPr}.
\begin{thm}[Phragm{\'e}n-Landau]\label{thm:phragmen_landau}
Let $f(x)$ be a real valued piecewise continuous function defined for $x\geq 1$, and bounded on every compact intervals. Let $F(s)$ be
its Mellin transform:
\[F(s)=\int_1^\infty \frac{f(x)}{x^{s+1}}\mathrm{d} x,\]
converges absolutely in some complex right half plane. Also assume that $f(x)$ does not change sign for $x\geq x_0$, for some $x_0\geq 1$.
If $F(s)$ diverges for some real $s$, then there exist a real number $\sigma_0$ satisfying the following properties:
\begin{enumerate}
\item[(1)] the integral defining $F(s)$ is divergent for $s<\sigma_0$ and convergent for $s>\sigma_0$,
\item[(2)] $s=\sigma_0$ is a singularity of $F(s)$,
\item[(3)] and $F(s)$ is analytic for $\text{Re}(s)>\sigma_0$.
\end{enumerate}
\end{thm}
\begin{proof}
Let $\sigma_0$ be:
\[\sigma_0=\inf\{\sigma\in \mathbb R:F(\sigma)\ \text{ converges}\}.\]
We shall show that $\sigma_0$ satisfies the properties given in the theorem.
As $f(x)$ does not change
sign for $x\geq x_0$, convergence of $F(\sigma)$ implies the absolute convergence of $F(s)$ for $\text{Re}(s)\geq\sigma$. This proves (1) and (3).
To prove (2), we proceed by method of contradiction. Assume that $s=\sigma_0$ is not a singularity of $F(s)$. Then there exist
$\sigma_0'>\sigma_0$ and $r>\sigma_0'-\sigma_0$ such that $F(s)$ has the following Taylor series expansion:
\[F(s)=\sum_{k=0}^\infty\frac{1}{k!}F^{(k)}(\sigma_0')(s-\sigma_0')^k,\]
for all $s$ satisfying $|s-\sigma_0'|<r$.
\begin{claim}[1] For $\sigma_0'$ as above, we have
\[F(s)=\sum_{k=0}^{\infty}\frac{1}{k!}(s-\sigma_0')^k\int_1^{\infty}(-\log x)^k\frac{f(x)}{x^{\sigma_0'+1}}\mathrm{d} x.\]
\end{claim}
\textit{Proof of Claim (1).}
By Cauchy's integral formula, we can write
\[F^{(k)}(\sigma_0')= \frac{k!}{2\pi i}\int_{\mathcal C}\frac{F(z)}{(z-\sigma_0')^{k+1}}\mathrm{d} z,\]
where $\mathcal C$ is a circle with a small enough radius having its center at $\sigma_0'$.
So we have
\[F(s)=\sum_{k=0}^{\infty}\frac{(s-\sigma_0')^k}{2\pi i }\int_{\mathcal C}\frac{1}{(z-\sigma_0')^{k+1}}\int_1^{\infty}\frac{f(x)}{x^{z+1}}\mathrm{d} x~\mathrm{d} z.\]
Suppose we can exchange the integrals of $x$ and $z$, then
\begin{align*}
F(s)&=\sum_{k=0}^{\infty}\frac{(s-\sigma_0')^k}{k!}\int_1^{\infty}\frac{f(x)}{x}
\frac{k!}{2\pi i}\int_{\mathcal C}\frac{x^{-z}\mathrm{d} z}{(z-\sigma_0')^{k+1}}\mathrm{d} x\\
&=\sum_{k=0}^{\infty}\frac{1}{k!}(s-\sigma_0')^k\int_1^{\infty}(-\log x)^k\frac{f(x)}{x^{\sigma_0'+1}}\mathrm{d} x,
\end{align*}
which proves Claim~1 conditionally. The only thing remains is to show that we can exchange integrals of $x$ and $z$.
If we choose $\mathcal C$ with a small enough radius, then
\[\int_1^\infty\frac{f(x)}{x^{\text{Re}(z) + 1}}\mathrm{d} x\]
is absolutely convergent and so is the
double integral
\[\int_{\mathcal C}\frac{1}{(z-\sigma_0')^{k+1}}\int_1^{\infty}\frac{f(x)}{x^{z+1}}\mathrm{d} x~\mathrm{d} z.\]
By the theorem of
Fubinni and Tonelli \cite[Theorem~B.3.1,~(b)]{DeitmarHarmonic},
we can exchange these two iterated integrals. This completes the proof of Claim~1.
\begin{claim}[2]
For $|s-\sigma_0'|<r$, the integral
\[F(s)=\int_{1}^{\infty}\frac{f(x)}{x^{s+1}}\mathrm{d} x\]
converges.
\end{claim}
\textit{Proof of Claim (2).}
We shall simplify $F(s)$ using Claim~1. We write
\[ F(s)=\sum_{k=0}^{\infty}\frac{(\sigma_0'-s)^k}{k!}\int_{1}^{\infty}\frac{(\log x)^kf(x)}{x^{\sigma_0' +1}}\mathrm{d} x. \]
In the above identity, we can exchange the series and the integral as the series is absolutely convergent. So we have
\begin{align*}
&\int_{1}^{\infty}\frac{f(x)}{x^{\sigma_0'+1}}\left(\sum_{k=0}^{\infty}\frac{(\sigma_0'-s)^k}{k!}(\log x)^k\right)\mathrm{d} x\\
&=\int_{1}^{\infty}\frac{f(x)}{x^{\sigma_0'+1}}\exp((\sigma_0'-s)\log x)\mathrm{d} x
=\int_{1}^{\infty}\frac{f(x)}{x^{s+1}}\mathrm{d} x.
\end{align*}
This completes the proof of Claim~2.
But Claim~2 implies that we have a real number smaller than $\sigma_0$, say $\sigma_0''$, such that the integral of $F(\sigma_0'')$ converges. This
is a contradiction to the definition of $\sigma_0$. So $\sigma_0$ is a singularity of $F(s)$, which proves (2).
\end{proof}
The following theorem appears in \cite[Section~2]{AnderOsci} without a proof and is attributed to Landau. We shall
prove this theorem using Theorem~\ref{thm:phragmen_landau}.
\begin{thm}[Phragm{\'e}n-Landau-Anderson-Stark \label{thm:landau_representation_integral}]
Let $f(x)$ be a real valued piecewise continuous function defined on $[1, \infty)$,
bounded on every compact intervals, and does not change sign when $x>x_0$ for some $1<x_0<\infty$. Define
\[F(s):=\int_{1}^{\infty}\frac{f(x)}{x^{s+1}}\mathrm{d} x,\]
and assume that the above integral is absolutely convergent in some half plane.
Further, assume that we have an analytic continuation of $F(s)$ in a region containing a part of the real line
\[l(\sigma_0, \infty):=\{\sigma + i0: \sigma > \sigma_0\}.\]
Then the integral representing $F(s)$ is absolutely convergent for $\text{Re}(s)>\sigma_0$, and hence $F(s)$ is an analytic function in this region.
\end{thm}
\begin{proof}
By Theorem~\ref{thm:phragmen_landau}, if
\[\int_1^\infty \frac{f(x)}{x^{\sigma'+1}}\mathrm{d} x\]
diverges for some $\sigma'>\sigma_0,$
then there exist a real number $\sigma_0'\geq\sigma'>\sigma_0$ such that $F$ is not analytic at $\sigma_0'$. But this contradicts our assumption
that $F$ is analytic on $l(\sigma_0, \infty)$. So the integral
\[\int_1^\infty \frac{f(x)}{x^{\sigma'+1}}\mathrm{d} x \ \text{ converges } \ \forall \ \sigma'>\sigma_0,\]
and since $f$ does not change
sign for $x\geq x_0$, $F(s)$ converges absolutely for $\text{Re}(s)>\sigma_0$. This also gives that $F(s)$ is analytic for $\text{Re}(s)>\sigma_0$.
\end{proof}
The above two theorems give some criteria when a function does not change sign. In the next section we will use these results to
show the sign changes of $\Delta(x)$.
\section{$\Omega_\pm$ Results}
Consider the Mellin transform $A(s)$ of $\Delta(x)$. We need the following assumptions to apply Theorem~\ref{thm:landau_representation_integral}.
\begin{asump}\label{as:for_landau}
Suppose there exists a real number $\sigma_0$, $0<\sigma_0<\sigma_1$,
such that $A(s)$ has the following properties.
\begin{itemize}
\item[(i)] There exists $t_0\neq 0$ such that
\begin{equation*}
\lambda:=\limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)|A(\sigma+it_0)|>0 .
\end{equation*}
\item[(ii)] At $\sigma_0$ we have
\begin{align*}
l_s&:=\limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)A(\sigma) < \infty,\\
l_i&:= \liminf\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)A(\sigma) > -\infty.
\end{align*}
\item[(iii)]
The limits $l_i, l_s$ and $\lambda$ satisfy
\[l_i+\lambda>0\quad \text{ and } \quad l_s-\lambda<0.\]
\item[(iv)] We can analytically continue $A(s)$ in a region containing the real line
$l(\sigma_0, \infty)$.
\end{itemize}
\end{asump}
\begin{rmk}
From Assumptions~\ref{as:for_landau},(i), we see that $\sigma_0+it_0$ is a singularity of $A(s)$.
\end{rmk}
We construct the following sets for further use.
\begin{defi}\label{def:A1A2}
With $l_s, l_i$ and $\lambda$ as in Assumptions~\ref{as:for_landau} and for an $\epsilon$ such that $0<\epsilon<\min(l_i+\lambda, \lambda-l_s)$, define
\begin{align*}
&\mathcal{A}_1:=\{x:x \in [1, \infty), \Delta(x)>(l_i+\lambda-\epsilon)x^{\sigma_0}\},\\
\text{ and }\quad&\mathcal{A}_2:=\{x:x \in [1, \infty), \Delta(x)<(l_s-\lambda+\epsilon)x^{\sigma_0}\}.
\end{align*}
\end{defi}
Under Assumptions~\ref{as:for_landau} and using methods from \cite{KaczMeasure}, we can derive the following measure theoretic theorem.
\begin{thm}\label{thm:landu_omegapm}
Let the conditions in Assumptions~\ref{as:for_landau} hold. Then for any
real number $M>1$, we have
\begin{align*}
\mu(\mathcal{A}_1\cap[M, \infty])>0,\\
\text{ and }\quad \mu(\mathcal{A}_2\cap[M, \infty])>0.
\end{align*}
This implies
\begin{align*}
\Delta(x)=\Omega_\pm(x^{\sigma_0}).
\end{align*}
\end{thm}
\begin{proof}
We prove the Theorem only for $\mathcal{A}_1$ as the other part is similar.
Define
\begin{align*}
&g(x):=\Delta(x)-(l_i+\lambda-\epsilon)x^{\sigma_0}, \quad &&G(s):=\int_{1}^{\infty}\frac{g(x)}{x^{s+1}}{\mathrm{d} x};\\
&g^+(x):=\max(g(x), 0), \quad &&G^+(s):= \int_{1}^{\infty}\frac{g^+(x)}{x^{s+1}}{\mathrm{d} x};\\
&g^-(x):=\max(-g(x), 0), \quad &&G^-(s):= \int_{1}^{\infty}\frac{g^-(x)}{x^{s+1}}{\mathrm{d} x}.
\end{align*}
With the above notations, we have
\begin{align*}
&g(x)=g^+(x)-g^-(x), \\
\text{ and }\quad &G(s)=G^+(s)-G^-(s).
\end{align*}
Note that
\begin{align*}
G(s)&=A(s) - \int_1^{\infty}(l_i+\lambda-\epsilon)x^{\sigma_0-s-1}{\mathrm{d} x} \\
&=A(s) + \frac{l_i+\lambda-\epsilon}{\sigma_0-s},\quad \text{for } \text{Re}(s)> \sigma_0.
\end{align*}
So $G(s)$ is analytic wherever $A(s)$ is, except possibly for a pole at $\sigma_0$.
This gives
\begin{align}\label{eq:G_lim_sigma0t0}
\limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)|G(\sigma+it_0)|
=\limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)|A(\sigma+it_0)|=\lambda.
\end{align}
We shall use the above limit to prove our theorem.
We proceed by method of contradiction.
Assume that there exists an $M>1$ such that
\[\mu(\mathcal{A}_1\cap[M, \infty))=0.\]
This implies
\[G^+(\sigma)=\int_{1}^{\infty}\frac{g^+(x)}{x^{s+1}}{\mathrm{d} x} = \int_{1}^{M}\frac{g^+(x)}{x^{s+1}}{\mathrm{d} x} \]
is bounded for any $s$, and so is an entire function.
By Assumptions~\ref{as:for_landau},
$A(s)$ and $G(s)$ can be analytically continued on the line $l(\sigma_0, \infty)$.
As $G(s)$ and $G^+(s)$ are analytic on $l(\sigma_0, \infty)$, $G^-(s)$ is also analytic on $l(\sigma_0, \infty)$.
The integral for $G^-(s)$ is absolutely convergent for $\text{Re}(s)>\sigma_3+1$, and $g^-(x)$ is a piecewise continuous function bounded
on every compact sets. This suggests that we can apply Theorem~\ref{thm:landau_representation_integral} to
$G^-(s)$, and conclude that
\[G^-(s)=\int_{1}^{\infty}\frac{g^-(x)}{x^{s+1}}{\mathrm{d} x} \]
is absolutely convergent for $\text{Re}(s)>\sigma_0$.
From the above discussion, we summarize that the Mellin transforms of $g, g^+$ and $g^-$ converge absolutely
for $\text{Re}(s)>\sigma_0$. As a consequence, we see that $G(\sigma), G^+(\sigma)$ and $G^-(\sigma)$
are finite real numbers for $\sigma>\sigma_0$. For $\sigma>\sigma_0$, we compare $G^+(\sigma)$ and $G^-(\sigma)$ in the following two cases.
\begin{itemize}
\item [Case 1: ] $G^+(\sigma)<G^-(\sigma)$.
In this case,
\begin{align*}
(\sigma-\sigma_0)|G(\sigma+ it_0)|&\leq (\sigma-\sigma_0)|G(\sigma)|\\
&=-(\sigma-\sigma_0)G(\sigma)\\
&=-(\sigma-\sigma_0)A(\sigma) + l_i+\lambda-\epsilon.
\end{align*}
So we have
\begin{equation*}
\limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)|G(\sigma+it_0)|
\leq l_i+\lambda-\epsilon - \liminf\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)A(\sigma)
\leq \lambda-\epsilon.
\end{equation*}
This contradicts (\ref{eq:G_lim_sigma0t0}).
\item[Case 2: ] $G^+(\sigma)\geq G^-(\sigma)$.
We have,
\begin{align*}
(\sigma-\sigma_0)|G(\sigma+it_0)|&\leq (\sigma-\sigma_0)G^+(\sigma)\\
&=O(\sigma-\sigma_0) \quad (G^+(\sigma)\text{ being a bounded integral}).
\end{align*}
Thus
\[ \limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)|G(\sigma+it_0)|=0.\]
This contradicts (\ref{eq:G_lim_sigma0t0}) again.
\end{itemize}
Thus $\mu(\mathcal{A}_1\cap[M, \infty))>0$ for any $M>1$, which completes the proof.
\end{proof}
\section{Measure Theoretic $\Omega_\pm$ Results}
Now we know that $\mathcal{A}_1$ and $\mathcal{A}_2$
are unbounded. But we do not know how the size of these sets grow.
An answer to this question was given by Kaczorowski and Szyd{\l}o in \cite[Theorem~4]{KaczMeasure}.
\begin{thm}[Kaczorowski and Szyd\l o \cite{KaczMeasure}]\label{thm:kaczorowski_correct}
Let the conditions in Assumptions~\ref{as:for_landau} hold.
Also assume that for a non-decreasing positive continuous function $h$ satisfying
\[h(x)\ll x^{\epsilon},\]
we have
\begin{equation}\label{eq:normDelta}
\int_{T}^{2T}\Delta^2(x){\mathrm{d} x}\ll T^{2\sigma_0 + 1}h(T).
\end{equation}
Then as $T\rightarrow \infty$,
\[\mu\left( \mathcal{A}_j\cap[1, T]\right)=\Omega\Big(\frac{T}{h(T)}\Big)\quad \text{ for } j=1, 2.\]
\end{thm}
In \cite{KaczMeasure}, Kaczorowski and Szyd{\l}o applied this theorem to the error term appearing in the asymptotic formula for the fourth power moment
of Riemann zeta function. We write this error term as $E_2(x)$:
\[ \int_0^x \left|\zeta\left(\frac{1}{2}+it\right)\right|^4 \mathrm{d} t = x P(\log x) + E_2(x),\]
where $P$ is a polynomial of degree $4$. Motohashi \cite{Motohashi1} proved that
\[ E_2(x) \ll x^{2/3+\epsilon}, \]
and further in \cite{Motohashi2} he showed that
\[ E_2(x)=\Omega_{\pm}(\sqrt x). \]
Theorem of Kaczorowski and Szyd{\l}o ( Theorem~\ref{thm:kaczorowski} ) gives that there exist $\lambda_0, \nu>0$ such that
\[\mu\{1\le x\le T: E_2(x)>\lambda_0\sqrt x \} = \Omega(T/{(\log T)^{\nu}}) \]
and
\[\mu\{1\le x\le T: E_2(x)<-\lambda_0\sqrt x \} = \Omega(T/{(\log T)^{\nu}})\]
as $T\rightarrow \infty$.
These results not only prove $\Omega_{\pm}$-results, but also give quantitative estimates for the occurrences
of such fluctuations.
The above theorem of Kaczorowski and Szyd\l o has been generalized by Bhowmik, Ramar\'e and Schlage-Puchta by localizing the fluctuations
of $\Delta(x)$ to $[T, 2T]$.
Proof of this theorem follows from \cite[Theorem~2]{gautami} (also see Theorem~\ref{thm:gautami} below).
\begin{thm}[Bhowmik, Ramar\'e and Schlage-Puchta \cite{gautami}]\label{thm:kaczorowski}
Let the assumptions in Theorem~\ref{thm:kaczorowski_correct} hold.
Then as $T\rightarrow \infty$,
\[\mu\left( \mathcal{A}_j\cap[T, 2T]\right)=\Omega\Big(\frac{T}{h(T)}\Big)\quad \text{ for } j=1, 2.\]
\end{thm}
An application of the above theorem to Goldbach's problem is given in \cite{gautami}.
Let
\begin{align*}
&\sum_{n\leq x} G_k(n) = \frac{x^k}{k!} -k \sum_\rho \frac{x^{k-1+\rho}}{\rho(1+\rho)\cdots(k-1+\rho)} + \Delta_k(x),
\end{align*}
where the Goldbach numbers $G_k(n)$ are defined as
\[\quad G_k(n)=\sum_{\substack{n_1,\ldots n_k \\ n_1+\cdots+n_k=n}}\Lambda(n_1)\cdots\Lambda(n_k),\]
and $\rho$ runs over nontrivial zeros of the Riemann zeta function $\zeta(s)$.
Bhowmik, Ramar\'e and Schlage-Puchta proved that under Riemann Hypothesis
\begin{align*}
&&\mu\{T\le x\le 2T: \Delta_k(x)>(\mathfrak c_k + \mathfrak c_k')x^{k-1} \} = \Omega(T/{(\log T)^{6}})& \\
&\text{ and }&\mu\{T\le x\le 2T: \Delta_k(x)<(\mathfrak c_k-\mathfrak c_k') x^{k-1} \} = \Omega(T/{(\log T)^{6}})& \ \text{ as } T\rightarrow \infty,
\end{align*}
where $k\geq 2$ and $\mathfrak{c}_k, \mathfrak{c}_k'$ are well defined real number depending on $k$ with $\mathfrak{c}_k'>0$.
Note that Theorem~\ref{thm:kaczorowski_correct} implies Theorem~\ref{thm:kaczorowski}, but both the theorems are applicable to the same set of examples.
The main obstacle in applicability of these theorems is the condition (\ref{eq:normDelta}).
For example, if $\Delta(x)$ is the error term in approximating
$\sum_{n\leq x}|\tau(n, \theta)|^2$, we can not apply Theorem~\ref{thm:kaczorowski_correct} and Theorem~\ref{thm:kaczorowski}. However, the following theorem due to
the author and A. Mukhopadhyay \cite[Theorem~3]{MeasureOmega} overcomes this obstacle by replacing the condition (\ref{eq:normDelta}).
\begin{thm}\label{thm:omega_pm_main}
Let the conditions in Assumptions~\ref{as:for_landau} hold.
Assume that there is an
analytic continuation of $A(s)$ in a region containing the real line $l(\sigma_0, \infty)$.
Let $h_1$ and $h_2$ be two positive functions
such that
\begin{equation}\label{eq:second_moment_error}
\int_{[T, 2T]\cap \mathcal{A}_j}
\frac{\Delta^2(x)}{x^{2\sigma_0+1}}\mathrm{d} x\ll h_j(T)\quad\text{ for } j=1, 2.
\end{equation}
Then as $T \longrightarrow \infty$,
\begin{equation}\label{eq:omega_pm_measure}
\mu(\mathcal{A}_j\cap[T, 2T])=\Omega\Big(\frac{T}{h_j(T)}\Big)\quad\text{ for } j=1, 2.
\end{equation}
\end{thm}
Below we state an integral version of Theorem~\ref{thm:kaczorowski} as in \cite{gautami}.
\begin{thm}[Bhowmik, Ramar\'e and Schlage-Puchta \cite{gautami}]\label{thm:gautami}
Suppose the conditions in Assumptions~\ref{as:for_landau} hold, and let $h(x)$ be as in Theorem~\ref{thm:kaczorowski}. Then as $\delta\rightarrow 0^+$,
\[\int_1^\infty\frac{\mu(\mathcal{A}_j\cap[x, 2x])h(4x)}{x^{2+\delta}}\mathrm{d} x=\Omega\Big(\frac{1}{\delta}\Big),\ \text{ for } j=1, 2.\]
\end{thm}
In our next theorem, we generalize Theorem~\ref{thm:kaczorowski_correct}, \ref{thm:kaczorowski}, \ref{thm:omega_pm_main} and \ref{thm:gautami}.
\begin{thm}\label{thm:omega_pm_main_new}
Let the conditions in Theorem~\ref{thm:omega_pm_main} hold. Then as $\delta\rightarrow 0^+$,
\begin{equation}
\int_1^\infty\frac{\mu(\mathcal{A}_j\cap[x, 2x])h_j(x)}{x^{2+\delta}}\mathrm{d} x=\Omega\Big(\frac{1}{\delta}\Big)\quad\text{ for }j=1, 2.
\end{equation}
\end{thm}
\begin{proof}
We shall prove the theorem for $j=1$; the proof is similar for $j=2$.
We define $g, g^+, g^-, G, G^+$ and $G^-$, as in Theorem~\ref{thm:landu_omegapm}.
Let
\[m^\#(x):=h_1(x)\mu(\mathcal{A}_1\cap[x, 2x])x^{-1}.\]
First, we shall show:
\begin{claim}[1] As $\delta\rightarrow 0$,
\[ \sum_{k\geq 0}\frac{m^\#(2^k)}{2^{k\delta}}=\Omega\left(\frac{1}{\delta}\right).\]
\end{claim}
Assume that
\begin{equation}\label{eq:measure_contradic_A1}
\sum_{k\geq 0}\frac{m^\#(2^k)}{2^{k\delta}}=o\left(\frac{1}{\delta}\right).
\end{equation}
From the above assumption, we may obtain an upper bound for $G^+(\sigma)$ as follows:
\begin{align*}
& \int_{\mathcal{A}_1}\frac{g^+(x)\mathrm{d} x}{x^{\sigma+1}}
\leq \sum_{k\geq 0}\int_{\mathcal{A}_1\cap[2^k, 2^{k+1}]}\frac{\Delta(x)\mathrm{d} x}{x^{\sigma+1}}
\quad ( \text{as } \Delta(x)> g(x) \text{ on } \mathcal{A}_1 )\\
&\leq \sum_{k\geq 0}\left(\int_{\mathcal{A}_1\cap[2^k, 2^{k+1}]}\frac{\Delta^2(x)\mathrm{d} x}{x^{2\sigma_0+1}}\right)^{\frac{1}{2}}
\left(\frac{\mu(\mathcal{A}_1\cap[2^k, 2^{k+1}])}{2^{k(2\delta+1)}}\right)^{\frac{1}{2}}
\ (\text{where } \sigma-\sigma_0=\delta>0)\\
&\leq c_3 \sum_{k\geq 0}\left(\frac{h_1(2^k)\mu(\mathcal{A}_1\cap[2^k, 2^{k+1}])}{2^{k(2\delta+1)}}\right)^{\frac{1}{2}}
\leq c_3\sum_{k\geq 0}\left(\frac{m^\#(2^k)}{2^{2k\delta}}\right)^{\frac{1}{2}}.
\end{align*}
From the above inequality, we get
\begin{align}\label{eq:Gplus_upper_bound}
\delta G^+(\sigma)\ll \delta\left(\sum_{k\geq 0}\frac{1}{2^{k\delta}}\right)^\frac{1}{2}
\left(\sum_{k\geq 0}\frac{m^\#(2^k)}{2^{k(\sigma-\sigma_0)}}\right)^\frac{1}{2}
=o(1)
\end{align}
as $\delta\rightarrow 0^+$.
Therefore
\[G^+(s)=\int_1^{\infty}\frac{g^+(x)\mathrm{d} x}{x^{s+1}}\]
is absolutely convergent for $\text{Re}(s)>\sigma_0$, and so it is analytic in this region.
But
\[G^-(s)=G(s)-G^+(s),\]
and $G$ is analytic on $l(\sigma_0, \infty)$. So $G^-$ is also analytic on $l(\sigma_0, \infty)$.
Using Theorem~\ref{thm:landau_representation_integral}, we get
\[G^+(s)=\int_1^{\infty}\frac{g^+(x)\mathrm{d} x}{x^{s+1}}\]
is absolutely convergent for $\text{Re}(s)>\sigma_0$. Absolute convergence of the integrals of $G$ and $G^+$
implies that the Mellin transformation of $g^-(x)$, given by
\[ G^-(s)=\int_1^{\infty}\frac{g^-(x)\mathrm{d} x}{x^{s+1}},\]
is also absolutely convergent for $\text{Re}(s)>\sigma_0$. As a consequence, we get $G(\sigma), G^+(\sigma)$, and $G^-(\sigma)$
are finite non-negative real numbers for $\sigma>\sigma_0$.
As indicated in Case-1 of Theorem~\ref{thm:landu_omegapm}, we can not have
\[G^+(\sigma)<G^-(\sigma)\ \text{ when }\sigma > \sigma_0.\]
So we always have
\[G^+(\sigma)\geq G^-(\sigma).\]
Using (\ref{eq:Gplus_upper_bound}),
\[\limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)|G(\sigma+it_0)|\leq
\limsup\limits_{\sigma\searrow\sigma_0}(\sigma-\sigma_0)G(\sigma)=0.\]
This is a contradiction to (\ref{eq:G_lim_sigma0t0}), and so (\ref{eq:measure_contradic_A1}) is wrong. This proves our Claim.
Now we are ready to prove the theorem.
For $k\geq1$, observe that
\begin{align*}
\int_{k-1}^{k}\frac{m^\#(2^x)}{2^{\delta x}} \mathrm{d} x
&=\int_{k-1}^{k} \frac{h_1(2^x)\mu(\mathcal{A}_1\cap[2^x, 2^{x+1}])}{2^{x(\delta +1)}} \mathrm{d} x
=\int_{k-1}^{k}\int_{2^x}^{2^{x+1}}\frac{h_1(2^x)}{2^{\delta x +x}}\mathrm{d}\mathcal{A}_1(t)\mathrm{d} x\\
&(\text{where } \mathcal{A}_1(t) \ \text{ is the indicator function of }\ \mathcal{A}_1) \\
= &\int_{2^{k-1}}^{2^k}\int_{k-1}^{\frac{\log t}{\log2}}\frac{h_1(2^x)}{2^{x(1+\delta)}}\mathrm{d} x\mathrm{d}\mathcal{A}_1(t)
+\int_{2^k}^{2^{k+1}}\int_{\frac{\log t}{\log2}-1}^{k}\frac{h_1(2^x)}{2^{x(1+\delta)}}\mathrm{d} x\mathrm{d}\mathcal{A}_1(t)
\end{align*}
From the above identity, we have
\[\int_{k-1}^{k}\frac{m^\#(2^x)}{2^{\delta x}}\mathrm{d} x\geq \int_{2^k}^{2^{k+1}}\int_{\frac{\log t}{\log2}-1}^{k
\frac{h_1(2^x)}{2^{x(1+\delta)}}\mathrm{d} x\mathrm{d}\mathcal{A}_1(t)\]
and
\[\int_{k}^{k+1}\frac{m^\#(2^x)}{2^{\delta x}}\mathrm{d} x\geq
\int_{2^{k}}^{2^{k+1}}\int_{k}^{\frac{\log t}{\log2}}
\frac{h_1(2^x)}{2^{x(1+\delta)}}\mathrm{d} x\mathrm{d}\mathcal{A}_1(t).
\]
So we get
\begin{align}
\notag
\int_{k-1}^{k+1}\frac{m^\#(2^x)}{2^{\delta x}}\mathrm{d} x
&\geq \int_{2^k}^{2^{k+1}}\int_{\frac{\log t}{\log 2}-1}^{k}\frac{h_1(2^x)}{2^{x(1+\delta)}}\mathrm{d} x\mathrm{d}\mathcal{A}_1(t)
+\int_{2^k}^{2^{k+1}}\int_{k}^{\frac{\log t}{\log2}}\frac{h_1(2^x)}{2^{x(1+\delta)}}\mathrm{d} x\mathrm{d}\mathcal{A}_1(t)\\
\notag
&=\int_{2^k}^{2^{k+1}}\int_{\frac{\log t}{\log 2}-1}^{\frac{\log t}{\log 2}}\frac{h_1(2^x)}{2^{x(1+\delta)}}\mathrm{d} x \mathrm{d}\mathcal{A}_1(t).
\end{align}
Now, we may use the fact that $h_1$ is a monotonically increasing function having polynomial growth, and simplify the above calculation as follows:
\begin{align}
\notag
&\int_{k-1}^{k+1}\frac{m^\#(2^x)}{2^{\delta x}} \mathrm{d} x
\geq h_1(2^{k})\int_{2^k}^{2^{k+1}}\int_{\frac{\log t}{\log2}-1}^{\frac{\log t}{\log2}}
\frac{\mathrm{d} x}{2^{x(1+\delta)}}\mathrm{d}\mathcal{A}_1(t)\\
\notag
&= \frac{h_1(2^{k})}{\log 2}\int_{2^k}^{2^{k+1}}
\Big(2^{-\left(\frac{\log t}{\log2}-1\right)(1+\delta)}-2^{-\frac{\log t}{\log2}(1+\delta)}\Big)\mathrm{d}\mathcal{A}_1(t)\\
\label{simplification_m3}
&= \frac{h_1(2^{k})}{\log 2}\int_{2^k}^{2^{k+1}}\frac{2^{1+\delta}-1}{t^{1+\delta}}\mathrm{d}\mathcal{A}_1(t)
\geq \frac{h_1(2^{k})}{2^{(k+1)(\delta +1)}}\mu(\mathcal{A}_1\cap[2^k, 2^{k+1}])
\geq \frac{1}{4}\frac{m^\#(2^k)}{2^{k\delta}}.
\end{align}
Now using Claim~(1), we get
\[\int_0^\infty\frac{m^\#(2^x)}{2^{\delta x}} \mathrm{d} x\gg \sum_{k=1}^\infty\frac{m^\#(2^k)}{2^{k\delta}}=\Omega\left(\frac{1}{\delta}\right).\]
Changing the variable $x$ to $u=2^x$ in the above inequality gives
\begin{align*}
&&\frac{1}{\log2} \int_1^\infty\frac{m^\#(u)}{u^{1+\delta}}\mathrm{d} u =\Omega \left(\frac{1}{\delta}\right),\\
&\text{or }& \int_1^\infty\frac{\mu(\mathcal{A}_j\cap[u, 2u])h_j(u)}{u^{2+\delta}}\mathrm{d} u =\Omega \left(\frac{1}{\delta}\right).
\end{align*}
This proves the theorem.
\end{proof}
\begin{coro}\label{cor:measure_omega_pm_from_upper_bound}
Let the conditions given in Theorem~\ref{thm:omega_pm_main} hold. Suppose we have a monotonically increasing positive function $h$ such that
\begin{equation}
\Delta(x)=O(h(x)),
\end{equation}
then
\begin{equation}
\mu(\mathcal{A}_j\cap[T, 2T])=\Omega\left(\frac{T^{1+2\sigma_0}}{h^2(T)}\right)\quad \text{ for } j=1, 2.
\end{equation}
\end{coro}
\begin{coro}\label{coro:omega_pm_secondmoment}
Similar to Corollary~\ref{cor:measure_omega_pm_from_upper_bound}, we assume that the conditions in
Theorem~\ref{thm:omega_pm_main} hold. Then we have
\begin{equation}\label{eq:omega_pm_secondmoment}
\int_{[T, 2T]\cap \mathcal{A}_j}\Delta^2(x)\mathrm{d} x = \Omega(T^{2\sigma_0 + 1}) \quad \text{ for } j=1, 2.
\end{equation}
\end{coro}
\begin{proof}
This Corollary follows from the proof of Theorem \ref{thm:omega_pm_main_new}.
We shall prove this Corollary for $\mathcal{A}_1$, and the proof for $\mathcal{A}_2$ is similar.
Note that in the proof of Theorem~\ref{thm:omega_pm_main_new}, we showed that the integral for $G^+(s)$
is absolutely convergent for $\text{Re}(s)>\sigma_0$ by assuming (\ref{eq:measure_contradic_A1}). Then we got a contradiction which
proves Claim~(1) of Theorem~\ref{thm:omega_pm_main_new}. Now we proceed in a similar manner
by assuming (\ref{eq:omega_pm_secondmoment}) is false. So we have
\[\int_{[T, 2T]\cap \mathcal{A}_1}\Delta^2(x)\mathrm{d} x = o(T^{2\sigma_0 + 1}). \]
So for an arbitrarily small constant $\varepsilon$, we have
\begin{align*}
& |G^+(s)|\leq \int_{\mathcal{A}_1}\frac{g^+(x)\mathrm{d} x}{x^{\sigma+1}}
\leq \sum_{k\geq 0}\int_{\mathcal{A}_1\cap[2^k, 2^{k+1}]}\frac{\Delta(x)\mathrm{d} x}{x^{\sigma+1}}\\
&\leq \sum_{k\geq 0}\frac{1}{2^{k(\sigma-\sigma_0)}}\left(\int_{\mathcal{A}_1\cap[2^k, 2^{k+1}]}\frac{\Delta^2(x)\mathrm{d} x}{x^{2\sigma_0+1}}\right)^{1/2}\\
&\leq c_4(\varepsilon) + \varepsilon\sum_{k\geq k(\varepsilon)}\frac{1}{2^{k(\sigma-\sigma_0)}}, \\
\end{align*}
where $c_4(\varepsilon)$ is a positive constant depending on $\varepsilon$. From this we obtain that $G^+(s)$ is absolutely convergent
for $\text{Re}(s)>\sigma_0$. Now onwards the proof is same as that of Theorem~\ref{thm:omega_pm_main_new}.
\end{proof}
\section{Applications}
Now we demonstrate applications of our theorems in the previous section to error terms appearing in two well known asymptotic
formulas.
\subsection{Square Free Divisors}\label{subsec:sqfree_divisors}
Let $a_n=2^{\omega(n)}$, where $\omega(n)$ denotes the number of distinct prime factors of $n$; equivalently, $a_n$ denotes the number of square free
divisors of $n$. We write
\begin{align*}
\sum_{n\leq x}^* 2^{\omega(n)}=\mathcal{M}(x) + \Delta(x),
\end{align*}
where
\[\mathcal{M}(x)= \frac{x\log x}{\zeta(2)}+\left(-\frac{2\zeta'(2)}{\zeta^2(2)} + \frac{2\gamma - 1}{\zeta(2)}\right)x,\]
and by a theorem of H{\"o}lder \cite{holder}
\begin{equation}\label{eq:sqfree_divisors_ub_unconditional}
\Delta(x)\ll x^{1/2}.
\end{equation}
Under Riemann Hypothesis, Baker \cite{baker} has improved the above upper bound to
\[\Delta(x)\ll x^{4/11}.\]
It is easy to see that the Dirichlet series $D(s)$ has the following form:
\[D(s)=\sum_{n=1}^{\infty}\frac{2^{\omega(n)}}{n^s}=\frac{\zeta^2(s)}{\zeta(2s)}.\]
Let $A(s)$ be the Mellin transform of $\Delta(x)$ at $s$, and let $s_0$ be
the zero of $\zeta(2s)$ with least positive imaginary part:
\begin{equation}\label{first_zeta_zero}
2s_0=\frac{1}{2} + i 14.134\ldots.
\end{equation}
\begin{figure}
\begin{tikzpicture}[yscale=0.9]
\draw [<->][thin, gray] (-1, -4.4)--(-1, 4.4);
\node at (-1.3, 0.22) {$0$};
\draw [<->][thin, gray] (4, 0)--(-3, 0);
\fill (3, 0) circle[radius=2pt];
\node at (2.8, 0.3) {$1$};
\fill (0.5, 3) circle[radius=1.5pt];
\node at (0.3, 3.2) {\small{$s_0$}};
\draw [thick] (3.6,-0.1)--(3.6, 0.1);
\node at (3.35, 0.35) {$\frac{5}{4}$};
\draw [thick] (0.5,-0.1)--(0.5, 0.1);
\node at (0.3, 0.35) {$\frac{1}{4}$};
\draw [thick] (2,-0.1)--(2, 0.1);
\node at (1.7, 0.35) {$\frac{3}{4}$};
\draw [thick] (0,-0.1)--(0, 0.1);
\node at (-0.3, 0.35) {$\frac{1}{5}$};
\draw [thick] (-1.1,-0.7)--(-0.9, -0.7);
\node at (-1.5, -0.7) {\small{$-2$}};
\draw [thick] (-1.1,0.7)--(-0.9, 0.7);
\node at (-1.33, 0.7) {\small{$2$}};
\draw [thick] (-1.1,4)--(-0.9, 4);
\node at (-1.5, 4) {\small{$14$}};
\draw [dashed] [postaction={decorate, decoration={ markings,
mark= between positions 0.09 and 0.98 step 0.28 with {\arrow[line width=1pt]{>},}}}]
(3.6, -4.4)--(3.6, -0.7)--(2, -0.7)--(2, 0.7)--(3.6, 0.7)--(3.6, 4.4);
\draw [dotted, thick] [postaction={decorate, decoration={ markings,
mark= between positions 0.3 and 0.7 step 0.2 with {\arrow[line width=1pt]{>},}}}]
(3.6, -4.4)--(3.6, -0.7)--(0, -0.7)--(0, 4)--(1, 4)--(1, 0.7)--(3.6, 0.7)--(3.6, 4.4);
\end{tikzpicture}
\caption{Contours for square-free divisors.}\label{fg:sq_free_divisors}
\end{figure}
We define a contour $\mathscr{C}^{(1)}$ as union of the following five lines:
\begin{align*}
\mathscr{C}^{(1)}:=&\left(\frac{5}{4}-i\infty,\ \frac{5}{4}-i2\right]\cup\left[\frac{5}{4}-i2,\ \frac{3}{4}-i2\right]
\cup\left[\frac{3}{4}-i2,\ \frac{3}{4}+i2\right]\\
&\cup\left[\frac{3}{4}+i2,\ \frac{5}{4}+i2\right]\cup\left[\frac{5}{4}+i2,\ \frac{5}{4}+i\infty\right)
\end{align*}
The contour $\mathscr{C}^{(1)}$ is represented by \lq dashed\rq \ lines in Figure~\ref{fg:sq_free_divisors}.
By Theorem~\ref{thm:analytic_continuation_mellin_transform}, we have
\[A(s)=\int_1^\infty\frac{\Delta(x)}{x^{s+1}}\mathrm{d} x=\int_{\mathscr{C}^{(1)}}\frac{D(\eta)}{\eta(s-\eta)}\mathrm{d} \eta.\]
Now, we shift the contour $\mathscr{C}^{(1)}$ to form a new contour $\mathscr{C}^{(2)}$, so that
\[1, \ s_0,\ l\left(\frac{1}{4}, \infty\right)\]
lie to the right of $\mathscr{C}^{(2)}$. We have represented the contour $\mathscr{C}^{(2)}$ by dotted lines in Figure~\ref{fg:sq_free_divisors}.
Since $s_0$ is a pole of $D(s)$ and is on the right side of $\mathscr{C}^{(1)}$, we have
\begin{align*}
A(s)=\int_{\mathscr{C}^{(2)}}\frac{D(\eta)}{\eta(s-\eta)}\mathrm{d} \eta + \underset{\eta=s_0}{\text{Res}}\left(\frac{D(\eta)}{\eta(s-\eta)}\right).
\end{align*}
From the above formula, we may compute the following limits:
\begin{equation*}
\lambda_1:= \lim_{\sigma\searrow0}\sigma|A(\sigma+s_0)| = |s_0|^{-1} \left|\underset{\eta=s_0}{\text{Res}}D(\eta)\right|>0
\end{equation*}
and
\begin{equation*}
\lim_{\sigma\searrow0}\sigma A(\sigma+ 1/4)=0.
\end{equation*}
For a fixed $\epsilon_0>0$,
\begin{align*}
\mathcal{A}_1&=\left\{x: \Delta(x)>(\lambda_1-\epsilon_0)x^{1/4}\right\}\\
\text{and} \quad \mathcal{A}_2&=\left\{ x : \Delta(x)<(-\lambda_1+\epsilon_0)x^{1/4}\right\}.
\end{align*}
Using Corollary~\ref{cor:measure_omega_pm_from_upper_bound} and (\ref{eq:sqfree_divisors_ub_unconditional}), we get
\begin{equation}\label{eq:sqfree_divisors_omegaset_uc}
\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{1/2}\right) \text{ for } j=1, 2.
\end{equation}
Under Riemann Hypothesis
we may show (also see Proposition~\ref{prop:upper_bound_second_moment_twisted_divisor}),
\[\int_{T}^{2T}\Delta^2(x)\ll T^{3/2+\epsilon}\ \text{ for any } \epsilon>0.\]
The above upper bound and Theorem~\ref{thm:omega_pm_main} give
\begin{equation}\label{eq:sqfree_divisors_omegaset}
\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{1-\epsilon}\right), \text{ for } j=1, 2 \ \text{ and for any } \ \epsilon>0.
\end{equation}
\subsection{The Prime Number Theorem Error}\label{subsec:pnt_error}
Consider the error term in the Prime Number Theorem:
\[\Delta(x)=\sum_{n\leq x}^*\Lambda(n)-x.\]
Let
\[\lambda_2=|2s_0|^{-1}, \]
where $2s_0$ is the first nontrivial zero of $\zeta(s)$ as in (\ref{first_zeta_zero}).
We shall apply Corollary~\ref{cor:measure_omega_pm_from_upper_bound} to prove the following proposition.
\begin{prop}\label{prop:pnt}
We write
\begin{align*}
\mathcal{A}_1&=\left\{x: \Delta(x)>(\lambda_2-\epsilon_0)x^{1/2}\right\}\\
\text{and} \quad \mathcal{A}_2&=\left\{ x : \Delta(x)<(-\lambda_2+\epsilon_0)x^{1/2}\right\},
\end{align*}
for a fixed $\epsilon_0$ such that $0<\epsilon_0<\lambda_2$. Then
\[\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{1-\epsilon}\right), \text{ for } j=1, 2 \ \text{ and for any } \epsilon>0.\]
\end{prop}
\begin{proof}
Here we apply Corollary~\ref{cor:measure_omega_pm_from_upper_bound} in a similar way as in the previous application,
so we shall skip the details.
The Riemann Hypothesis, Theorem~\ref{thm:kaczorowski} and Theorem~PNT** give
\[\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(\frac{T}{\log^4 T}\right) \text{ for } j=1, 2; \]
this implies the proposition. But if the Riemann Hypothesis is false, then there exists a constant $\mathfrak{a}$, with $1/2<\mathfrak{a}\leq1$, such that
\[\mathfrak{a}=\sup\{\sigma:\zeta(\sigma+it)=0\}.\]
Using Perron summation formula, we may show that
\[\Delta(x)\ll x^{\mathfrak{a}+\epsilon},\]
for any $\epsilon>0$. Also for any arbitrarily small $\delta$, we have $\mathfrak{a}-\delta<\sigma'<\mathfrak{a}$ such that
$\zeta(\sigma'+it')=0$ for some real number $t'$. If $\lambda'':=|\sigma'+it'|^{-1}$, then by Corollary~\ref{cor:measure_omega_pm_from_upper_bound}
we get
\begin{align*}
\mu\left(\left\{x\in[T, 2T]:\Delta(x)>(\lambda''/2)x^{\sigma'}\right\}\right)&=\Omega\left(T^{1-2\delta-2\epsilon}\right)\\
\text{ and } \quad \mu\left(\left\{x\in[T, 2T]:\Delta(x)<-(\lambda''/2)x^{\sigma'}\right\}\right)&=\Omega\left(T^{1-2\delta-2\epsilon}\right).
\end{align*}
As $\delta$ and $\epsilon$ are arbitrarily small and $\sigma'>1/2$, the above $\Omega$ bounds imply the proposition.
\end{proof}
\begin{rmk}
Results similar to Proposition~\ref{prop:pnt} can be obtained for error terms in asymptotic formulas for partial sums of Mobius function and for
partial sums of the indicator
function of square-free numbers.
\end{rmk}
\begin{rmk}
In Section~\ref{subsec:sqfree_divisors}~and~\ref{subsec:pnt_error}, we saw that $\mu(\mathcal{A}_j)$ are large. Now suppose that
$\mu(\mathcal{A}_1\cup\mathcal{A}_2)$ is large, then what can we say about the individual sizes of $\mathcal{A}_j$? We may guess that $\mu(\mathcal{A}_1)$ and $\mu(\mathcal{A}_2)$
are both large and almost equal. But this may be very difficult to prove. In the next chapter, we shall show that if $\mu(\mathcal{A}_1\cup\mathcal{A}_2)$ is large,
then both $\mathcal{A}_1$ and $\mathcal{A}_2$ are nonempty.
\end{rmk}
\chapter{Analytic Continuation Of The Mellin Transform}\label{chap:analytic_continuation}
In this chapter, we express the error term $\Delta(x)$ as a contour integral using the Perron's formula. This allows us to obtain a
meromorphic continuation of $A(s)$ (see Definition~\ref{def:mellin_transform}) in terms of the meromorphic continuation of $D(s)$, which is the main
theorem of this chapter ( Theorem~\ref{thm:analytic_continuation_mellin_transform} ). This theorem will be used in the next chapter to obtain
$\Omega_\pm$ results for $\Delta(x)$.
\section{Perron's Formula}
Recall that we have a sequence of real numbers $\{a_n\}_{n=1}^{\infty}$, with its
Dirichlet series $D(s)$. The Perron summation formula approximates the partial sums of $a_n$ by expressing it as a contour integral
involving $D(s)$.
\begin{thm}[Perron's Formula, Theorem~II.2.1 \cite{TenenAnPr}]\label{thm:perron_formula}
Let $D(s)$ be absolutely convergent for $\text{Re}(s)>\sigma_c$, and let $\kappa>\max(0, \sigma_c)$. Then for $x\geq 1$, we have
\[\sum_{n\leq x}^*a_n=\int_{\kappa-i\infty}^{\kappa+i\infty}\frac{D(s)x^s}{s}\mathrm{d} s.\]
\end{thm}
\noindent
But in practice, we use the following effective version of the Perron's formula.
\begin{thm}[Effective Perron's Formula, Theorem~II.2.1 \cite{TenenAnPr}]\label{thm:effective_perron_formula}
Let $\{a_n\}_{n=1}^\infty, \ D(s)$ and $\kappa$ be defined as in Theorem~\ref{thm:perron_formula}. Then for $T\geq 1$ and $x\geq 1$, we have
\[\sum_{n\leq x}^*a_n=\int_{\kappa-iT}^{\kappa+iT}\frac{D(s)x^s}{s}\mathrm{d} s + O\left(x^\kappa\sum_{n=1}^\infty
\frac{|a_n|}{n^\kappa(1+T|\log(x/n)|)}\right).\]
\end{thm}
\noindent
The above formulas are used by shifting the line of integration, and thus by collecting the residues of $D(s)x^s/s$ at its poles lying to
the right of the shifted contour.
The residues contribute to the main term $\mathcal{M}(x)$, leaving an expression for $\Delta(x)$ as a contour integral.
So we write
\[\sum_{n\leq x}^*a_n=\mathcal{M}(x)+\Delta(x),\]
where $\mathcal{M}(x)$ is the main term and $\Delta(x)$ is the error term. We make the following natural assumptions on $D(s), \mathcal{M}(x)$ and $\Delta(x)$.
\begin{asump}\label{as:for_continuation_mellintran}
Suppose there exist real numbers $\sigma_1$ and $\sigma_2$ satisfying $0<\sigma_1<\sigma_2$,
such that
\begin{enumerate}
\item[(i)]
$D(s)$ is absolutely convergent for $\text{Re}(s)> \sigma_2$.
\item[(ii)]
$D(s)$ can be meromorphically continued to the half plane $\text{Re}(s)>\sigma_1$ with only finitely many poles
$\rho$ of $D(s)$ satisfying
\[ \sigma_1<\text{Re}(\rho)\leq\sigma_2. \]
We shall denote this set of poles by $\mathcal{P}$.
\item[(iii)]
The main term $\mathcal{M}(x)$ is sum of residues of $\frac{D(s)x^s}{s}$ at poles in $\mathcal P$:
\[\mathcal{M}(x)=\sum_{\rho\in \mathcal{P}}Res_{s=\rho}\left( \frac{D(s)x^s}{s}\right).\]
\end{enumerate}
\end{asump}
The above assumptions also imply:
\begin{note}\label{note:initial_assumption_consequences}
We may also observe:
\begin{enumerate}
\item[(i)] For any $\epsilon>0$, we have
\[|a_n|, |\mathcal{M}(x)|, |\Delta(x)|, \left|\sum_{n\leq x}a_n\right| \ll x^{\sigma_2+\epsilon}. \]
\item[(ii)] The main term $\mathcal{M}(x)$ is a polynomial in $x$, and $\log x$:
\[\mathcal{M}(x)=\sum_{j\in\mathscr{J}}\nu_{1, j}x^{\nu_{2, j}}(\log x)^{\nu_{3, j}},\]
where $\nu_{1, j}$ are complex numbers, $\nu_{2, j}$ are real numbers with $\sigma_1<\nu_{2, j}\leq \sigma_2$, $\nu_{3, j}$ are positive integers, and $\mathscr J$ is a finite index set.
\end{enumerate}
\end{note}
\noindent
To express $\Delta(x)$ in terms of a contour integration, we define the following contour.
\begin{defi}\label{def:contour}
Let $\sigma_1, \sigma_2$ be as defined in Assumptions~\ref{as:for_continuation_mellintran}.
Choose a positive real number $\sigma_3$ such that
$\sigma_3>\sigma_2.$
We define the contour $\mathscr{C}$, as in Figure~\ref{fg:contour_c0}, as the union of the following five
line segments:
\[\mathscr{C}=L_1\cup L_2\cup L_3 \cup L_4 \cup L_5 ,\]
where
\begin{align*}
L_1=&\{\sigma_3+iv: T_0 \leq v < \infty\},
&L_2=\{u+iT_0: \sigma_1\leq u\leq \sigma_3 \}, \\
L_3=&\{\sigma_1+iv: -T_0 \leq v \leq T_0\},
&L_4=\{u-iT_0: \sigma_1\leq u\leq \sigma_3 \}, \\
L_5=&\{\sigma_3+iv: -\infty< v \leq -T_0\}. \\
\end{align*}
\end{defi}
\begin{figure}
\begin{tikzpicture}[yscale=0.8]
\draw [<->][dotted] (0, -4.4)--(0, 4.4);
\node at (-0.5, 2) {$T_0$};
\draw [thick] (-0.1, 2 )--(0.1,2);
\node at (-0.3, 0.3) {$0$};
\node at (-0.5, -2) {$-T_0$};
\draw [thick] (-0.1,-2 )--(0.1, -2);
\draw [<->][dotted] (5, 0)--(-2, 0);
\draw [dotted] (2, -4.4)--(2, 4.4);
\node at (1.7, 0.3) {$\sigma_1$};
\draw [dotted] (3, -4.4)--(3, 4.4);
\node at (2.7, 0.3) {$\sigma_2$};
\draw [dotted] (4, -4.4)--(4, 4.4);
\node at (3.7, 0.3) {$\sigma_3$};
\draw [thick] [postaction={decorate, decoration={ markings,
mark= between positions 0.1 and 0.99 step 0.2 with {\arrow[line width=1.2pt]{>},}}}]
(4, -4.4)--(4, -2)--(2, -2)--(2, 2)--(4, 2)--(4, 4.4);
\node [below left] at (1.8, 1.6) {$\mathscr{C}$};
\end{tikzpicture}
\caption{Contour $\mathscr{C}$}\label{fg:contour_c0}
\end{figure}
\noindent
Now, we write $\Delta(x)$ as an integration over $\mathscr{C}$ in the following lemma.
\begin{lem}\label{lem:integral_exp_delta}
Under Assumptions~\ref{as:for_continuation_mellintran}, the error term $\Delta(x)$ can be expressed as:
\begin{equation*}
\Delta(x) = \int_{\mathscr{C}}\frac{D(\eta)x^{\eta}}{\eta}{\mathrm{d} \eta}.
\end{equation*}
\end{lem}
\begin{proof}
Follows from Theorem~\ref{thm:perron_formula}.
\end{proof}
\section{Analytic continuation of $A(s)$}
Now, we shall discuss a method to
obtain a meromorphic continuation of $A(s)$, which will serve as an important tool to obtain $\Omega_\pm$ results for $\Delta(x)$ in the
following chapter.
\noindent
Below we present the main theorem of this chapter.
\begin{thm}\label{thm:analytic_continuation_mellin_transform}
Under Assumptions-\ref{as:for_continuation_mellintran}, we have
\[A(s)=\int_{\mathscr{C}}\frac{D(\eta)}{\eta(s-\eta)}\mathrm{d}\eta,\]
when $s$ lies right to $\mathscr{C}$.
\end{thm}
\subsection{Preparatory Lemmas}
We shall need the following preparatory lemmas to prove the above theorem.
From Lemma~\ref{lem:integral_exp_delta}, we have:
\begin{equation}\label{eq:A_in_eta_x}
A(s)= \int_1^{\infty}\int_{\mathscr{C}}\frac{D(\eta)x^{\eta}}{\eta}{\mathrm{d}\eta}\frac{\mathrm{d} x}{x^{s+1}}.
\end{equation}
To justify Theorem~\ref{thm:analytic_continuation_mellin_transform}, we need to justify the interchange of
the integrals of $\eta$ and $x$ in (\ref{eq:A_in_eta_x}).
\begin{defi}
Define the following complex valued function $B(s)$:
\begin{align*}
B(s)&:=\int_{\mathscr{C}}\frac{D(\eta)}{\eta}\int_1^{\infty}\frac{\mathrm{d} x}{x^{s-\eta+1}}{\mathrm{d}\eta}\\
&=\int_{\mathscr{C}}\frac{D(\eta){\mathrm{d}\eta}}{(s-\eta)\eta}\quad \quad \text{ for } \text{Re}(s)>\text{Re}(\eta).
\end{align*}
\end{defi}
The integral defining $B(s)$ being absolutely convergent, we have $B(s)$ is
well defined and analytic.
\begin{defi}
For a positive integer $N$, define the contour $\mathscr{C}(N)$ as:
\[\mathscr{C}(N)=\{\eta \in \mathscr{C}: |\text{Im}(\eta)|\leq N\}.\]
\end{defi}
\begin{defi}
Integrating the integrals of $\eta$ and $x$, define $B_N(s)$ as:
\begin{align*}
B_N(s)&=\int_{\mathscr{C}(N)}\frac{D(\eta){\mathrm{d} \eta}}{\eta}\int_1^{\infty}\frac{\mathrm{d} x}{x^{s-\eta+1}}\\
&=\int_{\mathscr{C}(N)}\frac{D(\eta){\mathrm{d}\eta}}{(s-\eta)\eta}\quad \quad \text{ for } \text{Re}(s)>\text{Re}(\eta).
\end{align*}
\end{defi}
With above definitions we prove:
\begin{lem}\label{lem:FubiniForExchange}
The functions $B$ and $B_N$ satisfy the following identities:
\begin{align}
\label{eq:limitAn_prime}
B(s)&= \lim_{N\rightarrow \infty}B_N(s)\\
\label{eq:fubini_onAn_prime}
&=\lim_{N\rightarrow \infty}\int_1^{\infty}\int_{\mathscr{C}(N)}\frac{D(\eta)x^{\eta}}{\eta}{\mathrm{d}\eta}\frac{\mathrm{d} x}{x^{s+1}}.
\end{align}
\end{lem}
\begin{proof}
Assume $N>T_0$. To show (\ref{eq:limitAn_prime}), note:
\begin{align*}
|B(s)-B_N(s)|&\leq\left|\int_{\mathscr{C}-\mathscr{C}(n)} \frac{D(\eta){\mathrm{d}\eta}}{(s-\eta)\eta}\right| \\
&\ll \left| \int_{\sigma_3+iN}^{\sigma_3+i\infty}\frac{D(\eta){\mathrm{d}\eta}}{(s-\eta)\eta} + \int_{\sigma_3-i\infty}^{\sigma_3-iN}\frac{D(\eta){\mathrm{d}\eta}}{(s-\eta)\eta}\right|\\
& \ll \int_N^{\infty}\frac{\mathrm{d} v}{v^2} \ll \frac{1}{N}.\quad (\text{ substituting }\eta=\sigma_3+iv)
\end{align*}
This completes proof of (\ref{eq:limitAn_prime}).
We shall prove (\ref{eq:fubini_onAn_prime}) using a theorem of Fubini and Tonelli \cite[Theorem~B.3.1,~(b)]{DeitmarHarmonic}. To show that the integrals commute, we need to show that
one of the iterated integrals in (\ref{eq:fubini_onAn_prime}) converges absolutely. We note:
\begin{align*}
&\int_{\mathscr{C}(N)}\int_1^{\infty}\left|\frac{D(\eta)}{\eta x^{s-\eta+1}}\right| \mathrm{d} x|\mathrm{d}\eta|\\
&\ll \int_{\mathscr{C}(N)}\left|\frac{D(\eta)}{\eta (\text{Re}(s)-\text{Re}(\eta))}\right| |\mathrm{d}\eta|< \infty.
\end{align*}
This implies (\ref{eq:fubini_onAn_prime}).
\end{proof}
Let
\begin{equation}\label{eq:B_N_s_defi}
B'_N(s):= \int_1^{\infty}\int_{\mathscr{C}(N)}\frac{D(\eta)x^{\eta}}{\eta}{\mathrm{d}\eta}\frac{\mathrm{d} x}{x^{s+1}}.
\end{equation}
We re-write (\ref{eq:fubini_onAn_prime}) of Lemma~\ref{lem:FubiniForExchange} as:
\[\lim_{N\rightarrow\infty}B_N'(s)=B(s).\]
Observe that $A(s)=B(s)$, if
\[\lim_{N\rightarrow \infty} \int_1^{\infty}\int_{\mathscr{C}-\mathscr{C}(N)}
\frac{D(\eta)x^{\eta}}{\eta}{\mathrm{d}\eta}\frac{\mathrm{d} x}{x^{s+1}}=0;\]
can be shown by interchanging the integral of $x$ with the limit. For this,
we need the uniform convergence of the integrand, which we do not have. It is easy to see from Theorem~\ref{thm:effective_perron_formula}
that the problem arises when $x$ is an integer. To handle this problem, we shall divide the integral in two parts,
with one part having neighborhoods of integers.
\begin{defi}
For $\delta=\frac{1}{\sqrt N}$ ( where $N\geq 2$ ), we construct the following set as a neighborhood of integers:
\[\mathcal{S(\delta)}:=[1, 1+\delta]\cup(\cup_{m\geq 2}[m-\delta, m+\delta]).\]
\end{defi}
\noindent
Write
\begin{equation}\label{eq:A_min_BN}
A(s)-B'_N(s)=J_{1, N}(s) + J_{2, N}(s) - J_{3, N}(s),\\
\end{equation}
where
\begin{align*}
&J_{1, N}(s)= \int_{\mathcal{S}(\delta)^c}\int_{\mathscr{C}-\mathscr{C}_N}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta \frac{\mathrm{d} x}{x^{s+1}},&&\\
&J_{2, N}(s)= \int_{\mathcal{S}(\delta)}\int_{\sigma_3-i\infty}^{\sigma_3+i\infty}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta \frac{\mathrm{d} x}{x^{s+1}},&&\\
&J_{3, N}(s)= \int_{\mathcal{S}(\delta)}\int_{\sigma_3-iN}^{\sigma_3+iN}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta \frac{\mathrm{d} x}{x^{s+1}}.&&\\
\end{align*}
In the next three lemmas, we shall show that each of $J_{i, N}(s) \rightarrow 0$ as
$N \rightarrow \infty$.
\begin{lem}\label{lem:J1lim}
For $\text{Re}(s)=\sigma>\sigma_3+1$, we have the limit
\begin{equation*}
\lim_{N\rightarrow \infty} J_{1, N}(s)= 0.
\end{equation*}
\end{lem}
\begin{proof}
Using Theorem~\ref{thm:effective_perron_formula} with $x\in \mathcal{S}(\delta)^c$, we have
\begin{align*}
\left| \int_{\mathscr{C}-\mathscr{C}_N}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta \right|& \ll x^{\sigma_3}\sum_{n=1}^{\infty}
\frac{|a_n|}{n^{\sigma_3}(1+N|\log(x/n)|)} \\
&\ll \frac{x^{\sigma_3}}{N}\sum_{n=1}^{\infty}\frac{|a_n|}{n^{\sigma_3}}
+ \frac{1}{N}\sum_{x/2\leq n \leq 2x}\frac{x|a_n|}{|x-n|}\left(\frac{x}{n}\right)^{\sigma_3}\\
&\ll \frac{x^{\sigma_3}}{N} + \frac{x^{\sigma_3+1+\epsilon}}{\delta N}
\ll \frac{x^{\sigma_3+1+\epsilon}}{\sqrt N} \quad (\text{ as } \delta= N^{-\frac{1}{2}}).
\end{align*}
From the above calculation, we see that
\[|J_{1, N}|\ll \frac{1}{\sqrt N}\int_{1}^{\infty}x^{\sigma_3-\sigma+\epsilon}dx\ll \frac{1}{\sqrt N}\]
for $\sigma=\text{Re}(s)>\sigma_3+1+\epsilon$. This proves our required result.
\end{proof}
\begin{lem}\label{lem:J2lim}
For $\text{Re}(s)=\sigma>\sigma_3$,
\[\lim_{N\rightarrow \infty} J_{2, N}(s) = 0. \]
\end{lem}
\begin{proof}
Recall that
\begin{equation*}
\sum_{n\leq x}^*a_n
= \begin{cases}
\sum_{n<x} a_n + a_x/2 &\mbox{ if } x \in \mathbb{N}, \\
\sum_{n\leq x}a_n &\mbox{ if } x \notin \mathbb{N}.
\end{cases}
\end{equation*}
By Note~\ref{note:initial_assumption_consequences},
$$\sum_{n\leq x}^*a_n \ll x^{\sigma_3}.$$
Using this bound, we calculate an upper bound for $J_{2, N}$ as follows:
\begin{align*}
&\left|\int_{\mathcal{S}(\delta)}\int_{\sigma_3-i\infty}^{\sigma_3+i\infty}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta\frac{\mathrm{d} x}{x^{s+1}}\right|
\leq\int_{\mathcal{S}(\delta)}\frac{\left|\sum_{n\leq x}^* a_n\right|}{x^{\sigma+1}}\mathrm{d} x\\
&\ll \int_{\mathcal{S}(\delta)} x^{\sigma_3-\sigma-1}\mathrm{d} x
\ll \int_1^{1+\delta}x^{\sigma_3-\sigma-1} + \sum_{m=2}^{\infty} \int_{m-\delta}^{m+\delta}x^{\sigma_3-\sigma-1}\mathrm{d} x .
\end{align*}
This gives
\[|J_{2, N}(s)|\ll \delta + \sum_{m\geq 2}\left(\frac{1}{(m-\delta)^{\sigma-\sigma_3}}-\frac{1}{(m+\delta)^{\sigma-\sigma_3}}\right).\]
Using the mean value theorem, for all $m\geq 2$ there exists a real number $\overline{m}\in[m-\delta, m+\delta]$ such that
\begin{align*}
|J_{2, N}(s)| \ll \delta + \sum_{m\geq 2}\frac{\delta}{\overline{m}^{\sigma-\sigma_3+1}}
\ll \delta=\frac{1}{\sqrt N} \quad \text{ by choosing } \sigma> \sigma_3.
\end{align*}
This implies that $J_{2, N}\rightarrow 0$ as $N\rightarrow\infty$.
\end{proof}
\begin{lem}\label{lem:J3lim}
For $\sigma>\sigma_3$, we have
\[\lim_{N\rightarrow \infty} J_{3, N}(s) = 0.\]
\end{lem}
\begin{proof}
Consider
\begin{align*}
J_{3, N}(s) &= \int_{\mathcal{S}(\delta)} \int_{\sigma_3-iN}^{\sigma_3+iN}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta \frac{\mathrm{d} x}{x^{s+1}}.
\end{align*}
This double integral is absolutely convergent for $\text{Re}(s)>\sigma_3$. Using the Theorem of Fubini and Tonelli \cite[Theorem~B.3.1,~(b)]{DeitmarHarmonic},
we can interchange the integrals:
\begin{align*}
J_{3, N}(s)&=\int_{\sigma_3-iN}^{\sigma_3+iN}\frac{D(\eta)}{\eta}\int_{\mathcal{S}(\delta)}x^{\eta-s-1}{\mathrm{d} x}~{\mathrm{d}\eta}\\
&=\int_{\sigma_3-iN}^{\sigma_3+iN}
\frac{D(\eta)}{\eta}\left\{\int_1^{1+\delta}\frac{x^{\eta}}{x^{s+1}}\mathrm{d} x
+\sum_{m\geq 2}\int_{m-\delta}^{m+\delta}\frac{x^{\eta}}{x^{s+1}}\mathrm{d} x\right\}\mathrm{d}\eta.
\end{align*}
For any $\theta_1, \theta_2 $ such that $0<\theta_1<\theta_2<\infty$, we have
\begin{align*}
\int_{\theta_1}^{\theta_2}x^{\eta-s-1}\mathrm{d} x = \frac{1}{s-\eta}\left\{ \frac{1}{\theta_1^{s-\eta}}-\frac{1}{\theta_2^{s-\eta}} \right\}
= \frac{\theta_2-\theta_1}{\overline{\theta}^{s-\eta+1}},
\end{align*}
for some $\overline{\theta}\in [\theta_1,\theta_2]$.
Applying the above formula to $J_{3, N}(s)$, we get
\begin{align*}
J_{3, N}(s)&=\int_{\sigma_3-iN}^{\sigma_3+iN}
\frac{D(\eta)}{\eta}\sum_{m\geq 1}
\frac{2\delta}{\overline{m}^{s-\eta+1}}\mathrm{d}\eta
=2\delta\sum_{m\geq 1} \int_{\sigma_3-iN}^{\sigma_3+iN}\frac{D(\eta)}{\overline{m}^{s-\eta+1}\eta}\mathrm{d} \eta,
\end{align*}
where $\overline{1/2}\in [1, 1+\delta]$ and $\overline{m}\in [m-\delta, m+\delta]$ for all integers $m\geq2$.
In the above calculation, we can interchange the series and the integral as the series is absolutely convergent.
So we have
\begin{align*}
J_{3, N}(s)&\ll \delta \sum_{m\geq 1} \int_{-N}^{N}\frac{1}{(1+|v|)\overline{m}^{\sigma-\sigma_3+1}}\mathrm{d} v \quad
\text{( substituting }\eta=\sigma_3+iv\text{ )}\\
&\ll \delta \log N \sum_{m\geq 1} \frac{1}{\overline{m}^{\sigma-\sigma_3+1}} \ll \frac{\log N}{\sqrt N}.
\end{align*}
Here we used the fact that for $\sigma>\sigma_3$, the series
\[\sum_{m\geq 1}\frac{1}{\overline{m}^{s-\eta+1}}\]
is absolutely convergent.
This proves our required result.
\end{proof}
\subsection{Proof of Theorem \ref{thm:analytic_continuation_mellin_transform}}
\begin{proof}
From equation (\ref{eq:A_min_BN}) and Lemma \ref{lem:J1lim}, \ref{lem:J2lim} and \ref{lem:J3lim}, we get
\[A(s)=\lim_{N\rightarrow \infty}B'_N(s)\]
for $\text{Re}(s)>\sigma_3+1$, and where $B'_N(s)$ is defined by (\ref{eq:B_N_s_defi}).
From Lemma~\ref{lem:FubiniForExchange}, we have
\[B(s)=\lim_{N\rightarrow \infty}B'_N(s).\]
This gives $A(s)$ and $B(s)$ are equal for $\text{Re}(s)>\sigma_3+1$. By analytic continuation,
$A(s)$ and $B(s)$ are equal for any $s$ that lies right to $\mathscr{C}$.
\end{proof}
In this chapter, we shall use the meromorphic continuation of $A(s)$
derived in Theorem~\ref{thm:analytic_continuation_mellin_transform}
to obtain mesure theoretic $\Omega_\pm$ results for $\Delta(x)$.
\section{Alternative Approches}
Theorem~\ref{thm:analytic_continuation_mellin_transform} gives a way for meromorphic continuation of $A(s)$ by formulating
it as a contour integral. This theorem has its significance in terms of elegance and generality.
However, there are alternative and easier ways in many cases. Below we give an example.
Note that
\[\sum_{n=1}^{\infty}\frac{a_n}{n^s}=-\int_{1}^{\infty}\Big(\sum_{n\leq x}a_n\Big){\mathrm{d} x^{-s}}\quad\text{ for }\text{Re}(s)>\sigma_2.\]
This gives
\[\frac{D(s)}{s}=\int_{1}^{\infty}\Big(\sum_{n\leq x}a_n\Big)x^{-s-1}\mathrm{d} x\quad\text{ for }\text{Re}(s)>\sigma_2.\]
So we can express $A(s)$ as
\begin{equation}
A(s)=\frac{D(s)}{s}-\int_1^{\infty}M(x)x^{-s-1}\mathrm{d} x\quad\text{ for }\text{Re}(s)>\sigma_2.
\end{equation}
The above formula reduces the problem of meromorphically continuing $A(s)$ to that of
\[\int_1^{\infty}M(x)x^{-s-1}\mathrm{d} x.\]
To demonstrate this method, we consider the case when $D(\eta)$ has a pole at $\eta=1$ and
residue at this pole gives the main term $M(x)$, i.e $\mathcal P=\{1\}$. The following meromorphic functions may serve as
examples of $D(\eta)$ in this situation:
\[\frac{\zeta(s)}{\zeta(2s)}, \zeta^2(s), \frac{\zeta^2(s)}{\zeta(2s)}, -\frac{\zeta'(s)}{\zeta(s)}, \ldots .\]
For a small positive real number $r$, we can
write $M(x)$ as
\[M(x)=\frac{1}{2\pi i}\int_{|\eta-1|=r}\frac{D(\eta)x^\eta}{\eta}\mathrm{d}\eta.\]
Thus
\begin{align}
\notag
\int_{1}^{\infty}\frac{M(x)}{x^{s+1}}\mathrm{d} x
&=\int_1^{\infty}\frac{1}{2\pi i}\int_{|\eta-1|=r}\frac{D(\eta)x^\eta}{\eta}d\eta\frac{\mathrm{d} x}{x^{s+1}}\\
\notag
&=\frac{1}{2\pi i}\int_{|\eta-1|=r}\frac{D(\eta)}{\eta}\left(\int_1^{\infty}\frac{\mathrm{d} x}{x^{s-\eta+1}}\right) \mathrm{d}\eta\\
\notag
& \text{( using \cite[Theorem~B.3.1,~(b)]{DeitmarHarmonic} )}\\ \label{eq:analytic_continuation_main_term}
&=\frac{1}{2\pi i}\int_{|\eta-1|=r}\frac{D(\eta)}{\eta(s-\eta)}\mathrm{d}\eta.
\end{align}
Let the Laurent series expansion of $D(\eta)$ at $\eta=1$ be
\[\frac{D(\eta)}{\eta}=\sum_{n\leq N} \frac{b_n}{(\eta-1)^n}+H(\eta),\]
where $H(\eta)$ is holomorphic for $\text{Re}(\eta)>\sigma_1$. Plugging in this expression for $D(\eta)$ in (\ref{eq:analytic_continuation_main_term}),
we get
\begin{equation}\label{eq:analytic_continuation_main_term2}
\int_{1}^{\infty}\frac{M(x)}{x^{s+1}}\mathrm{d} x = \sum_{n\leq N} b_n \frac{1}{2\pi i}\int_{|\eta-1|=r}\frac{\mathrm{d}\eta}{(\eta-1)^n(s-\eta)}.
\end{equation}
Let $\text{Re}(s)\geq1+2r$, then
\[\frac{|\eta-1|}{|s-1|}\leq \frac{1}{2}\quad\text{ for }|\eta-1|=r.\]
This gives
\[\frac{1}{s-\eta}=\sum_{n=0}^{\infty}\frac{(\eta-1)^n}{(s-1)^{n+1}}\]
is an absolutely convergent series. Using the above expansion of $(s-\eta)^{-1}$
in (\ref{eq:analytic_continuation_main_term2}), we have
\begin{align*}
\int_{1}^{\infty}\frac{M(x)}{x^{s+1}}\mathrm{d} x &=
\sum_{n\leq N} b_n \frac{1}{2\pi i}\int_{|\eta-1|=r}\left\{\sum_{m=0}^{\infty}\frac{(\eta-1)^m}{(s-1)^{m+1}}\right\}
\frac{\mathrm{d}\eta}{(\eta-1)^n}\\
&=\sum_{n\leq N}\frac{b_n}{(s-1)^n} \quad (\text{ by \cite[Theorem~6.1]{ComplexAnKra}})\\
&=\frac{D(s)}{s}-H(s).
\end{align*}
Thus we got
\[A(s)=H(s)\ \text{ for } \ \text{Re}(s)\ge 1+2r.\]
But the right hand side is holomorphic for $\text{Re}(s)>\sigma_1$ hence the formula
gives analytic continuation of $A(s)$ in the half plane $Re(s)>\sigma_1$.
Similar calculations can be done when the main term $M(x)$ is more complicated.
\chapter*{Notations}
\addcontentsline{toc}{chapter}{Notations}
We denote the set of natural numbers by $\mathbb N$, the set of integers by $\mathbb Z$, the set of real numbers by $\mathbb R$, the set of
positive real numbers by $\mathbb R^+$,
and the set of complex numbers by $\mathbb C$.
\vspace{0.3cm}
\noindent
The notaion $i$ stands for $\sqrt{-1}$, the square root of $-1$ that belongs to the upper half plane in $\mathbb C$.
\vspace{0.3cm}
\noindent
We denote the Lebesgue mesure on the real line $\mathbb R$ by $\mu$.
\vspace{0.3cm}
\noindent
For $z=\sigma+it\in \mathbb C$, we denote $\sigma$ by $\text{Re}(z)$ and $t$ by $\text{Im}(z)$.
\vspace{0.3cm}
\noindent
Let $f(x)$ and $g(x)$ be a complex valued function on $\mathbb R^+$.
As $x\rightarrow\infty$, we write
\begin{itemize}
\item[] $f(x)=O(g(x))$, \ if \ $\lim_{x\rightarrow\infty}\left|\frac{f(x)}{g(x)}\right|>0$;
\item[] $f(x)=o(g(x))$, \ if \ $\lim_{x\rightarrow\infty}\left|\frac{f(x)}{g(x)}\right|=0$;
\item[] $f(x)\ll g(x)$, \ if \ $f(x)=O(g(x))$;
\item[] $f(x)\gg g(x)$, \ if \ $g(x)=O(f(x))$;
\item[] $f(x)\sim g(x)$, \ if \ $\lim_{x\rightarrow\infty}\frac{f(x)}{g(x)}=1$;
\item[] $f(x)\asymp g(x)$, \ if \ $0<\lim_{x\rightarrow\infty}\left| \frac{f(x)}{g(x)}\right| <\infty$.
\end{itemize}
\vspace{0.3cm}
\noindent
Let $f(x)$ be a complex valued function on $\mathbb R^+$, and let $g(x)$ be a positive monotonic function on $\mathbb R^+$.
As $x\rightarrow\infty$, we write
\begin{itemize}
\item[] $f(x)=\Omega(g(x))$, \ if \ $\limsup_{x\rightarrow \infty}\frac{|f(x)|}{g(x)}>0$;
\item[] $f(x)=\Omega_+(g(x))$, \ if \ $\limsup_{x\rightarrow \infty}\frac{f(x)}{g(x)}>0$;
\item[] $f(x)=\Omega_-(g(x))$, \ if \ $\liminf_{x\rightarrow \infty}\frac{f(x)}{g(x)}<0$;
\item[] $f(x)=\Omega_\pm(g(x))$, \ if \ $f(x)=\Omega_+(g(x))$ and $f(x)=\Omega_-(g(x))$.
\end{itemize}
\chapter*{Synopsis}\label{chap:synopsis}
\addcontentsline{toc}{chapter}{Synopsis}
This thesis studies fluctuation of error terms that appears in various asymptotic formulas and size of the sets where these fluctuations occur.
As a consequence, this approach replaces Landau's criterion on oscillation of error terms.
\begin{center}
\textbf{\large{General Theory}}
\end{center}
Consider a sequence of real numbers $\{a_n\}_{n=1}^{\infty}$ having Dirichlet series
\[D(s)=\sum_{n=1}^{\infty}\frac{a_n}{n^s},\]
which is convergent in some half-plane. As in Perron summation formula \cite[II.2.1]{TenenAnPr}, we write
\[\sum^*_{n\leq x}a_n=\mathcal{M}(x)+\Delta(x),\]
where $\mathcal{M}(x)$ is the main term, $\Delta(x)$ is the error term and $\sum^*$ is defined as
\begin{equation*}
\sum^*_{n\leq x} a_n =
\begin{cases}
\sum_{n\leq x} a_n \ & \text{if } x\notin \mathbb N, \\
\sum_{n< x} a_n +\frac{1}{2} a_x \ &\text{if } x\in\mathbb N.
\end{cases}
\end{equation*}
In this thesis, we obtain $\Omega$ and $\Omega_\pm$ estimates for $\Delta(x)$.
We shall use the Mellin transform of $\Delta(x)$ (defined below) to obtain such estimates.
\begin{defiu}
The Mellin transform of $\Delta(x)$ be $A(s)$, defined as
\[A(s)=\int_{1}^{\infty}\frac{\Delta(x)}{x^{s+1}}\mathrm{d} x.\]
\end{defiu}
In this direction, under some natural assumptions and for a suitably defined contour $\mathscr{C}$, we shall show that
\[A(s)=\int_{\mathscr{C}}\frac{D(\eta)}{\eta(s-\eta)}\mathrm{d} \eta.\]
In the above formula, the poles of $D(s)$ that lie left to $\mathscr{C}$ are all the poles that contributes to the main term $\mathcal{M}(x)$.
Landau \cite{Landau} used the meromorphic continuation of $A(s)$ to obtain $\Omega_\pm$ results for $\Delta(x)$. He proved that
if $A(s)$ has a pole at $\sigma_0+it_0$ for some $t_0\neq 0$ and has no real pole for $s\ge \sigma_0$, then
\[ \Delta(x)=\Omega_{\pm} (x^{\sigma_0}). \]
We shall show a quantitative version of Landau's theorem, which also generalizes a theorem of Gautami, Ramar\'e and Schlage-Puchta \cite{gautami}.
Below we state this theorem in a simplified way. We introduce the following notations to state these theorems.
\begin{defiu}
Let
\begin{align*}
\mathcal{A}^+_T(x^{\sigma_0})&:=\{T\leq x \leq 2T: \Delta(x)> \lambda x^{\sigma_0}\},\\
\mathcal{A}^-_T(x^{\sigma_0})&:=\{T\leq x \leq 2T: \Delta(x)< -\lambda x^{\sigma_0}\},\\
\mathcal{A}_T (x^{\sigma_0})&:= \mathcal{A}^+_T(x^{\sigma_0})\cup \mathcal{A}^-_T(x^{\sigma_0}),
\end{align*}
for some $\lambda, \sigma_0 >0$.
\end{defiu}
\begin{thmu
Let $\sigma_0>0$, and let the following conditions hold:
\begin{itemize}
\item[(1)] $A(s)$ has no real pole for $\text{Re}(s)\geq\sigma_0$,
\item[(2)] there is a complex pole $s_0=\sigma_0+it_0$, $t_0\neq 0$, of $A(s)$, and
\item[(3)] for positive functions $h^\pm(x)$ such that $h\pm(x)\rightarrow \infty$ as $x\rightarrow \infty$, we have
$$\int_{\mathcal{A}^\pm_T(x^{\sigma_0})}\frac{\Delta^2(x)}{x^{2\sigma_0+1}}\mathrm{d} x\ll h^\pm(T).$$
\end{itemize}
Then
\[\mu(\mathcal{A}^\pm_{T}(x^{\sigma_0})) = \Omega\left(\frac{T}{h^\pm(T)}\right),\]
where $\mu$ denotes the Lebesgue measure.
\end{thmu}
In the above theorem, Condition~2 is a very strong criterion. In the following theorem, we replace Condition~2 by an $\Omega$-bound of
$\mu(\mathcal{A}_T (x^{\sigma_0}))$ and obtain an $\Omega_{\pm}$-result from the given $\Omega$-bound.
\begin{thmu}
Let $\sigma_0>0$, and let the following conditions hold:
\begin{itemize}
\item[(1)] $A(s)$ has no real pole for $\text{Re}(s)\geq\sigma_0$, and
\item[(2)] $\mu( \mathcal{A}_T(x^{\sigma_0}))=\Omega(T^{1-\delta})$ for $0<\delta<\sigma_0$.
\end{itemize}
Then
\[\Delta(x)=\Omega_\pm(T^{\sigma_0-\delta'})\] for any $\delta'$ such that $0<\delta'<\delta$.
\end{thmu}
The above two theorems are applicable to a wide class of arithmetic functions.
Now we mention some results obtained by applying these theorems.
\begin{center}
\textbf{\large{A Twisted Divisor Function}}
\end{center}
Given $\theta\neq 0$, define
\[\tau(n, \theta)=\sum_{d|n}d^{i\theta}.\]
The Dirichlet series of $|\tau(n, \theta)|^2$ can be expressed in terms of Riemann zeta function as
\begin{equation*}
D(s)=\sum_{n=1}^{\infty}\frac{|\tau(n, \theta)|^2}{n^s}=\frac{\zeta^2(s)\zeta(s+i\theta)\zeta(s-i\theta)}{\zeta(2s)}
\quad \text{for}\quad \text{Re}(s)>1.
\end{equation*}
In \cite[Theorem 33]{DivisorsHallTenen}, Hall and Tenenbaum proved that
\begin{equation*}
\sum_{n\leq x}^*|\tau(n, \theta)|^2=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)
+\omega_3(\theta)x + \Delta(x),
\end{equation*}
where $\omega_i(\theta)$s are explicit constants depending only on $\theta$. They also showed that
\begin{equation*}
\Delta(x)=O_\theta(x^{1/2}\log^6x).
\end{equation*}
Here the main term comes from the residues of $D(s)$ at $s=1, 1\pm i\theta $.
All other poles of $D(s)$ come from zeros of $\zeta(2s)$. Using a pole on the line $\text{Re}(s)=1/4$,
Landau's method gives
\[\Delta(x)=\Omega_{\pm}(x^{1/4}).\]
We prove the following bounds for a computable $\lambda(\theta)>0$ and for any $\epsilon>0$:
\begin{align*}
&\mu\left(\{T \leq x \leq 2T: \Delta(x)>(\lambda(\theta)-\epsilon)x^{1/4}\}\right)
=\Omega\left(T^{1/2}(\log T)^{-12}\right),\\
&\mu\left(\{T \leq x \leq 2T: \Delta(x)<(-\lambda(\theta)+\epsilon)x^{1/4}\}\right)
=\Omega\left(T^{1/2}(\log T)^{-12}\right).
\end{align*}
For a constant $c>0$, define
\[\alpha(T) =\frac{3}{8}-\frac{c}{(\log T)^{1/8}}.\]
Applying a method due to Balasubramanian, Ramachandra and Subbarao \cite{BaluRamachandraSubbarao}, we prove
\[\Delta(T)=\Omega\left(T^{\alpha(T)}\right).\]
In fact, this method gives $\Omega$-estimate for the measure of the sets involved:
\[ \mu(\mathcal{A}\cap [T,2T])=\Omega\left(T^{2\alpha(T)}\right),\]
where
\[\mathcal{A}=\{x: |\Delta(x)|\ge Mx^{\alpha(x)} \}\]
and $M>0$ is a positive constant.
We also show that
\[ \text{either } \ \Delta(x)=\Omega\left(x^{ \alpha(x)+\delta/2}\right)
\ \text{ or } \
\Delta(x)=\Omega_{\pm}\left(x^{3/8-\delta'}\right),\]
for $0<\delta<\delta'<1/8$.
For any $\epsilon>0$, this result and the conjecture
\[ \Delta(x)=O(x^{3/8+\epsilon})\]
proves that
\[\Delta(x)=\Omega_{\pm}(x^{3/8-\epsilon}).\]
\begin{center}
\textbf{\large{Prime Number Theorem Error}}
\end{center}
Let $a_n$ be the von Mandoldt function $\Lambda(n)$:
\[\Lambda(n):=\begin{cases}
\log p \quad &\mbox{ if } n=p^r, \ r\geq 1, \ p \text{ prime },\\
0 & \mbox{ otherwise.}
\end{cases}.\]
Let
\[\sum_{n\le x}^* \Lambda_n = x + \Delta(x). \]
From the Vinogradov's zero free region for Riemann zeta function, one gets \cite[Theorem~12.2]{IvicBook}
\begin{equation*}
\Delta(x)=O\left( x\exp\left(-c(\log x)^{3/5}(\log\log x)^{-1/5}\right)\right)
\end{equation*}
for some constant $c>0$.
Hardy and Littlewood \cite{HardyLittlewoodPNTOmegapm} proved that
\[\Delta(x)=\Omega_\pm\left(x^{1/2}\log\log\log x\right).\]
But this result does not say about the measure of the sets, where the above $\Omega_\pm$ bounds are attained by $\Delta(x)$.
We obtain the following weaker result, but with an $\Omega$-estimates for the measure of the corresponding sets.\\
\noindent
Let $\lambda_1>0$ denotes a computable constant. For a fixed $\epsilon$, $0<\epsilon<\lambda_1$, we write
\begin{align*}
\mathcal{A}_1&:=\left\{x: \Delta(x)>(\lambda_1-\epsilon)x^{1/2}\right\},\\
\mathcal{A}_2&:=\left\{ x : \Delta(x)<(-\lambda_1+\epsilon)x^{1/2}\right\}.\vspace{5mm}~\\
\end{align*}
Then
\begin{align*}
\mu([T, 2T]\cap\mathcal{A}_j)
=\Omega\left(T^{1-\epsilon}\right), \text{ for } j=1, 2 \ \text{ and for any } \ \epsilon>0.
\end{align*}
Under Riemann Hypothesis, we have
\begin{align*}
\mu([T, 2T]\cap A_j)=\Omega\left(\frac{T}{(\log T)^4}\right) \text{ for } j= 1, 2.
\end{align*}
We also show the following unconditional
$\Omega$-bounds for the second moment of $\Delta$:
\begin{equation*}
\int_{[T, 2T]\cap \mathcal{A}_j}\Delta^2(x)\mathrm{d} x = \Omega(T^2) \quad \text{ for } j=1, 2.
\end{equation*}
\begin{center}
\textbf{\large{Non-isomorphic Abelian Groups}}
\end{center}
Let $a_n$ denote the number of non-isomorphic abelian groups of order $n$. We write
\begin{equation*}
\sum_{n\leq x}^*a_n = \sum_{k=1}^{6}b_kx^{1/k} + \Delta(x).
\end{equation*}
In the above formula, we define $b_k$ as
\[b_k:=\prod_{j=1, j\neq k}^\infty \zeta(j/k).\]
It is an open problem to show that
\begin{equation}\label{conj1s}
\Delta(x)\ll x^{1/6+\delta} \ \text{ for any } \ \delta>0.
\end{equation}
The best result on upper bound of $\Delta(x)$ is
due to O. Robert and P. Sargos \cite{sargos}, which gives
\[\Delta(x)\ll x^{1/4+\epsilon} \ \text{ for any } \ \epsilon>0.\]
Also Balasubramanian and Ramachandra \cite{BaluRamachandra2} proved that
\begin{equation*}
\Delta(x)=\Omega\left(x^{1/6}\sqrt{\log x}\right).
\end{equation*}
Following their method, we prove
\begin{equation*}
\mu\left( \{ T\le x\le 2T: |\Delta(x)|\ge \lambda_2 x^{1/6}(\log x)^{1/2}\} \right)
=\Omega(T^{5/6-\epsilon}),
\end{equation*}
for some $\lambda_2>0$ and for any $\epsilon>0$.
They also obtained
\[ \Delta(x)=\Omega_{\pm}(x^{92/1221}),\]
while it has been conjectured that
\[\Delta(x)=\Omega_\pm(x^{1/6-\delta}),\]
for any $\delta>0$.
We shall show that either
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x=\Omega( T^{5/3+\delta} )\text{ or }\Delta(x)=\Omega_\pm(x^{1/6-\delta}),\]
for any $0<\delta< 1/42$.
The conjectured upper bound (\ref{conj1s}) of $\Delta(x)$ gives
\[\int_T^{2T}\Delta^4(x)\mathrm{d} x \ll T^{5/3+\delta}.\]
This along with our result implies that
\[\Delta(x)=\Omega_\pm(x^{1/6-\delta})\ \text{ for any } \ 0<\delta <1/42.\]
\chapter{The Twisted Divisor Function}\label{chap:twisted_divisor}
Recall that in Chapter~\ref{chap:intro}, we have defined the twisted divisor function $\tau(n, \theta)$
as follows:
\[\tau(n, \theta)=\sum_{d | n} d^{i\theta}, \quad \text{for } \ \theta\in \mathbb R -\{0\},\ n\in \mathbb N.\]
We also have stated the following asymptotic formula:
\begin{equation*}
\sum_{n\leq x}^*|\tau(n, \theta)|^2=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)
+\omega_3(\theta)x + \Delta(x),
\end{equation*}
where $\omega_i(\theta)$s are explicit constants depending only on $\theta$ and
\[\Delta(x)=O_\theta(x^{1/2}\log^6x).\]
In this chapter, we give a proof of this formula (see Section~\ref{sec:asymptotic_formula}, Theorem~\ref{thm:asymp_formula_tau_n_theta}).
In Section~\ref{sec:twisted_divisor_application_landau}, we use Theorem~\ref{thm:omega_pm_main} to obtain some measure
theoretic $\Omega_\pm$ results. Further, we obtain an $\Omega$ bound for the second moment of $\Delta(x)$ in Section~\ref{sec:stronger_omega}
by adopting a technique due to Balasubramanian, Ramachandra and Subbarao \cite{BaluRamachandraSubbarao}. In the final section, we prove
that if the $\Omega$ bound obtained in the previous section can not be improved, then
\[\Delta(x)=\Omega(x^{3/8-\epsilon}) \ \text{ for any }\ \epsilon>0.\]
Now we motivate with a brief note on few applications of $\tau(n, \theta)$.
\section{Applications of $\tau(n, \theta)$}\label{sec:applications_tau_n_theta}
The function $\tau(n, \theta)$ can be used to study various properties related to the distribution of divisors of an integer:
\[\sum_{\substack{d|n \\ a\leq \log d \leq b}}^*1=\frac{1}{2\pi}\int_{-\infty}^{\infty}\tau(n, \theta)
\frac{e^{-ib\theta}-e^{-ia\theta}}{-i\theta};\]
here $\sum^*$ means that the corresponding contribution to the sum is $\frac{1}{2}$ if $e^a|n$ or $e^b|n$.
Below we present two applications.
\subsection{Clustering of Divisors}
The following function measures the clustering of divisors of an integer:
\[W(n, f):=\sum_{d, d'|n} f(\log(d/d')),\]
for some constant $c>0$ and for a function $f\in L^1(\mathbb R)$. We assume that $f$ has a Fourier transformation, say $\hat f$,
and $\hat f \in L^1(\mathbb{R})$.
\begin{prop}\label{prop:clustering}
With the above notations:
\[\sum_{n\le x} W(n, f) = \frac{1}{2\pi} \int_{-\infty}^{\infty} \hat{f}( \theta)\sum_{n\le x} |\tau(n,\theta)|^2 \mathrm{d}\theta.\]
\end{prop}
\begin{proof}
Note that by the Fourier inversion formula, we get
\begin{align*}
W(n, f)&= \sum_{d, d'|n} f(\log(d/d'))= \frac{1}{2\pi}\sum_{d, d'|n} \int_{-\infty}^{\infty} \hat{f}(\theta)\left(\frac{d}{d'}\right)^{i\theta}\mathrm{d} \theta\\
&= \frac{1}{2\pi}\int_{-\infty}^{\infty} \hat{f}(\theta) \left( \sum_{d, d'|n} \left(\frac{d}{d'}\right)^{i\theta} \right)\mathrm{d} \theta
= \frac{1}{2\pi}\int_{-\infty}^{\infty} \hat{f}(\theta)|\tau(n, \theta)|^2\mathrm{d} \theta.
\end{align*}
This implies the proposition.
\end{proof}
Using Proposition~\ref{prop:clustering} and the formula in (\ref{eq:formmula_tau_ntheta}), we may write
\begin{align*}
\sum_{n\le x} W(n, f) &= \frac{x\log x}{2\pi} \int_{-\infty}^{\infty} \hat{f}( \theta) \omega_1(\theta)\mathrm{d}\theta
+ \frac{x}{2\pi} \int_{-\infty}^{\infty} \hat{f}( \theta)\big(\omega_2(\theta)\cos(\theta\log x)+\omega_3(\theta)\big)\mathrm{d}\theta \\
&+ \frac{x}{2\pi} \int_{-\infty}^{\infty} \hat{f}( \theta)\Delta(x, \theta)\mathrm{d}\theta. \\
&\text{(In the above identity, we denoted $\Delta(x)$ by $\Delta(x, \theta)$.)}
\end{align*}
This gives that the function $\sum_{n\leq x}W(n, f)$ behaves like $x\log x$. Further, if we want to obtain more information on $\sum_{n\leq x}W(n, f)$,
we may analyzing other terms in the above formula. But now, we skip the details and refer to \cite[Chapter~4]{DivisorsHallTenen}.
\subsection{The Multiplication Table Problem}
The multiplication table problem asks for an estimate on the order of the growth of $|\text{Mul}(N)|$ as $N\rightarrow\infty$, where
\[\text{Mul}(N):=\{1\leq m \leq N^2: m= ab, \ a, b \in \mathbb{Z}\text{ and }1\leq a, b \leq N\}.\]
The initial attempts in this direction
are due to Erd{\H{o}}s \cite{ErdosMultTab}. He used a result of Hardy and Ramanujan \cite{HardyRamanujan} (also see \cite{MyExpo}) to show
\[|\text{Mul}(N)|\ll \frac{N^2}{(\log N)^{\nu_0}\sqrt{\log\log N}}\ \text{as} \ N\rightarrow\infty,\]
and here
\[\nu_0=1-\frac{1+\log\log 2}{\log2}.\]
Intuitively, the theorem of Hardy and Ramanujan says that most of the positive integers less than $x$ have around $\log\log x$ prime factors;
more precisely
\[\#\{n\leq x: |\omega(n)-\log\log n|<\sqrt{\log\log n}\}\sim x \ \text{as} \ x\rightarrow\infty.\]
This gives that most of the positive integers less than $N^2$ have around $\log\log N$ prime factors, whereas most of
the integers in the multiplication table
have around $2\log\log N$ prime factors, which implies $|\text{Mul}(N)|$ is $o(N^2)$. Erd{\H{o}}s used a refined version of this argument
to obtain the given upper bound for $|\text{Mul}(N)|$.
The best known bound on the asymptotic growth of $|\text{Mul}(n)|$ is due to Ford \cite{FordAnn}:
\[|\text{Mul}(N)|\asymp \frac{N^2}{(\log N)^{\nu_0}(\log\log N)^{3/2}},\ \text{ as } N\rightarrow\infty.\]
To obtain the expected lower bound for $|\text{Mul}(N)|$, Ford first proved that
\[|\text{Mul}(N)|\gg \frac{N^2}{(\log N)^2}\sum_{n\leq N^{1/8}}\frac{L(n)}{n},\
\text{ where } \ L(n):=\mu\left(\cup_{d|n}[\log(d/2), \log d]\right).\]
We may also observe that
\[\sum_{n\leq N^{1/8}}\frac{L(n)}{n}\geq\frac{\left(\sum_{n\leq N^{1/8}}\frac{d(n)}{n}\right)^2}{6\sum_{n\leq N^{1/8}}\frac{W(n)}{n}}.\]
Rest of the part in Ford's argument deals with the above sums involving the divisor function $d(n)$ and
$W(n):=W\left(n, 1_{[\frac{1}{2}, 2]}\right)$, where $1_{[\frac{1}{2}, 2]}$ is the indicator function of the interval $[\frac{1}{2}, 2]$.
We skip the details and refer to \cite{Fordy2y}.
\section{Asymptotic Formula for $\sum_{n\leq x}^*|\tau(n, \theta)|^2$}\label{sec:asymptotic_formula}
In this section, we shall prove the following asymptotic formula for $\sum_{n\leq x}^*|\tau(n, \theta)|^2$.
\begin{thm}[Theorem~33, \cite{DivisorsHallTenen}]\label{thm:asymp_formula_tau_n_theta}
Let $\theta\neq 0$ be a fixed real number. Then for $x\geq 1$, we have
\[\sum_{n\leq x}^*|\tau(n, \theta)|^2=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)
+\omega_3(\theta)x + O_\theta(x^{1/2}\log^6x)\,\]
where $\omega_i(\theta)$s are explicit constants depending only on $\theta$.
\end{thm}
\begin{proof}
Recall that the corresponding Dirichlet series $D(s)$ has the following meromorphic continuation:
\[D(s)=\sum_{1}^\infty\frac{|\tau(n, \theta)|^2}{n^s}=\frac{\zeta^2(s)\zeta(s+i\theta)\zeta(s-i\theta)}{\zeta(2s)},\ \text{ for } s>1.\]
For $x\geq 2$, we denote $\kappa=1+\frac{1}{\log x}$ and $T=x+|\theta|+1$. By Perron's formula
\[\sum_{n\leq x}^*|\tau(n, \theta)|^2=\frac{1}{2\pi i}\int_{\kappa-iT}^{\kappa+iT}D(s)x^s\frac{\mathrm{d} s}{s} + O(x^{\epsilon}).\]
After shifting the line of integration to $\text{Re}(s)=\frac{1}{2}$, we may estimate the contributions from horizontal lines as follows:
\begin{align*}
T^{-1}\int_{\frac{1}{2}}^{1}|D(\sigma \pm iT)|x^{\sigma}\mathrm{d} \sigma \ll T^{-1}\int_{\frac{1}{2}}^{1}T^{1-\sigma+\epsilon}x^{\sigma}\mathrm{d} \sigma \ll x^{\epsilon}.
\end{align*}
To obtain an asymptotic formula for $\sum_{n\leq x}^*|\tau(n, \theta)|^2$, we add up the residues from the poles $1, 1\pm i\theta$
after shifting the line of integration to $\text{Re}(s)=\frac{1}{2}$:
\begin{align*}
\sum_{n\leq x}^*|\tau(n, \theta)|^2 = \mathcal{M}(x) +
O\left(x^\epsilon + x^\frac{1}{2}\int_{-T}^{T}\left|\frac{\zeta^2(\frac{1}{2}+it)\zeta(\frac{1}{2}+i(t+\theta))\zeta(\frac{1}{2}+i(t-\theta))}{\zeta(1+2it)(\frac{1}{2} + it)}\right|\mathrm{d} t\right),
\end{align*}
where
\[\mathcal{M}(x)=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)+\omega_3(\theta)x.\]
If we write
\[\mathcal J(\mathfrak{a}, T):=\int_{-T}^{T}\frac{\zeta^4(\frac{1}{2}+i(\mathfrak{a}+t))}{\sqrt{t^2+\frac{1}{4}}}\mathrm{d} t \quad \text{ for } \ \mathfrak{a}, T
\in \mathbb R \ \text{ and } \ T\geq 1\ ,\]
then we have \cite[Theorem~5.1]{IvicBook}
\begin{equation}\label{eq:upper_bound_4thmoment_zeta}
\mathcal J(\mathfrak a, T)\ll_{\mathfrak a} \log^5 T.\
\end{equation}
To express $\Delta(x)$ in terms of $\mathcal J(\mathfrak a, T)$, observe that
\begin{align*}
\Delta(x)&=\sum_{n\leq x}^* |\tau(n, \theta)|^2 -\mathcal{M}(x)\\
&\ll x^\epsilon + x^\frac{1}{2}\int_{-T}^{T}\left|\frac{\zeta^2(\frac{1}{2}+it)\zeta(\frac{1}{2}+i(t+\theta))\zeta(\frac{1}{2}+i(t-\theta))}{\zeta(1+i2t)(\frac{1}{2} + it)}\right|\mathrm{d} t \\
&\ll x^\epsilon + x^\frac{1}{2}\log x \int_{-T}^{T}|\zeta^2(\frac{1}{2}+it)\zeta(\frac{1}{2}+i(t+\theta))\zeta(\frac{1}{2}+i(t-\theta))|\frac{\mathrm{d} t}{|\frac{1}{2} + it|} .\\
\end{align*}
From (\ref{eq:upper_bound_4thmoment_zeta}) and using the Cauchy-Schwartz inequality twice, we get
\begin{align*}
\Delta(x)\ll x^\epsilon + x^\frac{1}{2}\log x \mathcal J^\frac{1}{2}(0, x) \mathcal J^\frac{1}{4}(\theta, x) \mathcal J^\frac{1}{4}(-\theta, x)
\ll_{\theta} x^\frac{1}{2} \log ^6 x,
\end{align*}
which gives the required result.
\end{proof}
In the following sections, we shall obtain various $\Omega$ and $\Omega_\pm$ bounds for $\Delta(x)$.
\section{Oscillations of the Error Term}\label{sec:twisted_divisor_application_landau}
Here we shall apply results in Chapter~\ref{chap:landau_theorem} to $\Delta(x)$ and obtain some measure theoretic $\Omega_\pm$ results.
We begin by defining a contour $\mathscr{C}$ as given in Figure~\ref{fg:contour_c0_taun}:
\begin{align*}
\mathscr{C}=&\left(\frac{5}{4}-i\infty, \frac{5}{4}-i(\theta+1)\right]\cup \left[\frac{5}{4}-i(\theta+1), \frac{3}{4}-i(\theta+1)\right]\\
&\cup \left[\frac{3}{4}-i(\theta+1), \frac{3}{4}+i(\theta+1)\right]\cup \left[\frac{3}{4}+i(\theta+1), \frac{5}{4}+i(\theta+1)\right]\\
&\cup \left[\frac{5}{4}+i(\theta+1), \frac{5}{4}+i\infty \right).
\end{align*}
From Theorem~\ref{thm:perron_formula}, we have
\[\Delta(x)=\int_{\mathscr{C}}\frac{D(\eta)x^\eta}{\eta}\mathrm{d} \eta.\]
The above identity expresses the Mellin transform $A(s)$ of $\Delta(x)$ as a contour integral involving $D(s)$.
Using Theorem~\ref{thm:analytic_continuation_mellin_transform}, we write
\[A(s)=\int_{1}^{\infty}\frac{\Delta(x)}{x^{s+1}}\mathrm{d} x=\int_{\mathscr{C}}\frac{D(\eta)}{\eta(s-\eta)}\mathrm{d} \eta,\]
when $s$ lies right to the contour $\mathscr{C}$. Denote the first nontrivial zero of $\zeta(s)$ with least positive imaginary part by $2s_0$.
An approximate value of this point is
\[2s_0=\frac{1}{2}+i14.134\ldots .\]
Define the contour $\mathscr{C}(s_0)$, as in Figure~\ref{fg:contour_c0s0}, such that $s_0$ and any real number $s\geq1/4$ lie in the
right side of this contour. A meromorphic continuation of $A(s)$ to all $s$ that lies right side of $\mathscr{C}(s_0)$ is given by
\begin{equation}\label{eq:analytic_cont1/4}
A(s)=\int_{\mathscr{C}(s_0)}\frac{D(\eta)x^\eta}{\eta}{\mathrm{d} \eta} + \frac{\underset{\eta=s_0}{\text{Res}}D(\eta)}{s_0(s-s_0)}.
\end{equation}
\begin{center}
\begin{figure}
\begin{tikzpicture}[yscale=0.8]
\draw [<->][dotted] (-1, -4.4)--(-1, 4.4);
\node at (-1.2, 0.3) {$0$};
\draw [<->][dotted] (4, 0)--(-3, 0);
\draw [dotted] (3, -4.4)--(3, 4.4);
\fill (3, 0) circle[radius=2pt];
\node at (2.8, 0.3) {$1$};
\fill (3, 1.8) circle[radius=2pt];
\node at (2.48, 1.8) {\scriptsize{$1+i\theta$}};
\fill (3, -1.8) circle[radius=2pt];
\node at (2.48, -1.8) {\scriptsize{$1-i\theta$}};
\node at (1.8, 0.3) {$\frac{3}{4}$};
\draw [thick] [postaction={decorate, decoration={ markings,
mark= between positions 0.09 and 0.98 step 0.21 with {\arrow[line width=1pt]{>},}}}]
(3.5, -4.4)--(3.5, -2.2)--(2, -2.2)--(2, 2.2)--(3.5, 2.2)--(3.5, 4.4);
\node [below left] at (1.55, 2) {$\mathscr{C}$};
\end{tikzpicture}
\caption{Contour $\mathscr{C}$, for $D(s)=\sum_{n=1}^{\infty}\frac{|\tau(n, \theta)|^2}{n^s}$.}\label{fg:contour_c0_taun}
\end{figure}
\end{center}
\begin{center}
\begin{figure}
\begin{tikzpicture}[yscale=0.8]
\draw [<->][dotted] (-1, -4.4)--(-1, 4.4);
\node at (-1.2, 0.3) {$0$};
\draw [<->][dotted] (4, 0)--(-3, 0);
\draw [dotted] (3, -4.4)--(3, 4.4);
\fill (3, 0) circle[radius=2pt];
\node at (2.8, 0.3) {$1$};
\fill (3, 1.8) circle[radius=2pt];
\node at (2.48, 1.8) {\scriptsize{$1+i\theta$}};
\fill (3, -1.8) circle[radius=2pt];
\node at (2.48, -1.8) {\scriptsize{$1-i\theta$}};
\draw [dotted] (0, -4.4)--(0, 4.4);
\node at (-0.28, 0.3) {$\frac{1}{4}$};
\fill (0, 3) circle[radius=2pt];
\node at (-0.28, 3) {$s_0$};
\draw [thick] [postaction={decorate, decoration={ markings,
mark= between positions 0.03 and 0.99 step 0.147 with {\arrow[line width=1pt]{>},}}}]
(3.5, -4.4)--(3.5, -2.5)--(2, -2.5)--(2, -0.5)--(-0.6, -0.5)--(-0.6, 3.8)--(0.4, 3.8)--(0.4, 0.65)
--(2, 0.65)--(2, 2.5)--(3.5, 2.5)--(3.5, 4.4);
\node [below left] at (-0.55, 2) {$\mathscr{C}(s_0)$};
\end{tikzpicture}
\caption{Contour $\mathscr{C}(s_0)$}\label{fg:contour_c0s0}
\end{figure}
\end{center}
From (\ref{eq:analytic_cont1/4}), we calculate the following two limits:
\begin{equation}\label{eqn:lambda_theta}
\lambda(\theta):= \lim_{\sigma\searrow0}\sigma|A(\sigma+s_0)| = |s_0|^{-1} \left|\underset{\eta=s_0}{\text{Res}}D(\eta)\right|>0
\end{equation}
and
\begin{equation*}
\lim_{\sigma\searrow0}\sigma A(\sigma+ 1/4)=0.
\end{equation*}
For a fixed small enough $\epsilon>0$, define
\begin{align*}
\mathcal{A}_1&=\left\{x: \Delta(x)>(\lambda(\theta)-\epsilon)x^{1/4}\right\},\\
\mathcal{A}_2&=\left\{ x : \Delta(x)<(-\lambda(\theta)+\epsilon)x^{1/4}\right\}.
\end{align*}
Theorem~\ref{thm:asymp_formula_tau_n_theta} and Corollary~\ref{cor:measure_omega_pm_from_upper_bound} give
\begin{equation}\label{result:tau_n_theta_omegapm}
\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{1/2}(\log T)^{-12}\right) \text{ for } j=1, 2.
\end{equation}
Under Riemann Hypothesis, Theorem~\ref{thm:omega_pm_main} and Proposition~\ref{prop:upper_bound_second_moment_twisted_divisor} give
\begin{equation}\label{result:tau_n_theta_omegapm_underRH}
\mu\left(\mathcal{A}_j\cap[T, 2T]\right)=\Omega\left(T^{3/4-\epsilon}\right) \text{ for } j=1, 2.
\end{equation}
From Corollary~\ref{coro:omega_pm_secondmoment}, we get
\begin{equation}\label{result:tau_n_theta_secondmoment}
\int_{\mathcal{A}_j\cap[T, 2T]}\Delta^2(x)\mathrm{d} x = \Omega\left(T^{3/2}\right) \text{ for } j=1, 2.
\end{equation}
\section{An Omega Theorem }\label{sec:stronger_omega}
Recall that (see Theorem~\ref{thm:asymp_formula_tau_n_theta})
$$\sum_{n\le x} |\tau(n,\theta)|^2 =\mathcal{M}(x) + \Delta(x),$$
where the main term
\[\mathcal{M}(x)=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)+\omega_3(\theta)x\]
comes from the poles of $D(s)$ at $s=1, 1+i\theta$ and
$s=1-i\theta$.
We may observe from Corollary~\ref{coro:omega_pm_secondmoment} that if $D(s)$ has a complex pole at $s_0=\sigma_0+it_0$, other than $1+i\theta$
and $1-i\theta$, then
\[ \int_T^{2T}\Delta(x)\mathrm{d} x=\Omega(x^{2\sigma_0+1}).\]
By Riemann Hypothesis, the only positive value for $\sigma_0$ is $\frac{1}{4}$, which is same as (\ref{result:tau_n_theta_secondmoment}).
In this section,
we shall use a technique due to Balasubramanian, Ramachandra and Subbarao \cite{BaluRamachandraSubbarao} to improve this omega bound further.
Now we state the main theorem of this section.
\begin{thm}\label{omega_integral}
For any $c>0$ and for a sufficiently large $T$ depending on $c$, we get
\begin{equation}\label{lb-increasing}
\int_T^{\infty} \frac{|\Delta(x)|^2}{x^{2\alpha+1}}e^{-2x/y} \mathrm{d} u
\gg_c\exp\left( c(\log T)^{7/8} \right),
\end{equation}
where
\[ \alpha=\alpha(T)=\frac{3}{8} -\frac{c}{(\log T)^{1/8} }.\]
\end{thm}
\noindent
In particular, this implies
\[ \Delta(x)=\Omega (x^{3/8}\exp(-c(\log x)^{7/8}), \]
for some suitable $c>0$.
In order to prove the theorem, we need several lemmas, which form the content of this section.
We begin with a fixed $\delta_0 \in (0,1/16]$ for which we would choose a numerical value at the end of this section.
\begin{defi}
For $T>1$, let $Z(T)$ be the set of all $\gamma$ such that
\begin{enumerate}
\item $T\le \gamma \le 2T$,
\item either $\zeta(\beta_1+i\gamma)=0$ for some $\beta_1\ge \frac{1}{2}+\delta_0$ \\
or $\zeta(\beta_2+i 2\gamma)=0$ for some $\beta_2\ge \frac{1}{2} +\delta_0$.
\end{enumerate}
Let
\[ I_{\gamma,k} = \{ T\le t \le 2T: |t-\gamma| \le k\log^2 T \} \text{ for } k=1, 2.\]
We finally define
\[J_k(T)=[T,2T] \setminus \cup_{\gamma\in Z(T)} I_{\gamma,k}. \]
\end{defi}
\begin{lem}\label{size-J(T)}
With the above definition, we have for $k=1,2$
\[ \mu(J_k(T)) = T +O\left( T^{1-\delta_0/4} \log^3 T \right). \]
\end{lem}
\begin{proof}
We shall use an estimate on the function $N(\sigma, T)$, which is defined as
\[N(\sigma, T):=\left|\{\sigma'+it:\sigma'\ge\sigma,\ 0<t\leq T,\ \zeta(\sigma'+it)=0\}\right|.\]
Selberg \cite[Page~237]{Titchmarsh} proved that
$$N(\sigma, T) \ll T^{1-\frac{1}{4}(\sigma -\frac{1}{2})} \log T, \ \text{ for } \ \sigma>1/2.$$
Now the lemma follows from the above upper bound on $N(\sigma, t)$, and the observation that
$$|\cup_{\gamma\in Z(T)} I_{\gamma,k}| \ll N\left(\frac{1}{2}+ \delta_0, T\right)\log^2 T.$$
\end{proof}
The next lemma closely follows Theorem 14.2 of \cite{Titchmarsh}, but does not depend on Riemann Hypothesis.
\begin{lem}\label{estimate-on-J(T)}
For $t\in J_1(T)$ and $\sigma= 1/2+\delta$ with $\delta_0 < \delta < 1/4-{\delta_0}/2$, we have
$$|\zeta(\sigma+it)|^{\pm 1} \ll
\exp\left(\log\log t \left(\frac{\log t}{\delta_0}\right)^{\frac{1-2\delta}{1-2\delta_0}}\right)$$
and
\[ |\zeta(\sigma+2it)|^{\pm 1} \ll
\exp\left(\log\log t \left(\frac{\log t}{\delta_0}\right)^{\frac{1-2\delta}{1-2\delta_0}}\right).\]
\end{lem}
\begin{proof}
We provide a proof of the first statement, and the second statement can be similarly proved.
Let $1 <\sigma' \le \log t$. We consider two concentric circles centered at $\sigma'+it$,
with radius $\sigma'-1/2-\delta_0/2$ and $\sigma' -1/2- \delta_0$.
Since $t\in J_1(T)$ and the radius of the circle is $\ll \log t$, we conclude that
\[ \zeta(z)\neq 0 \ \text{ for } \ |z-\sigma'-it | \le \sigma' - \frac{1}{2} - \frac{\delta_0}{2} \]
and also $\zeta(z)$ has polynomial growth in this region.
Thus on the larger circle, $\log |\zeta(z)| \le c_{17}\log t$, for some constant $c_{17}>0$.
By Borel-Carath\'eodory theorem,
\[\ |z-\sigma'-it | \le \sigma' - \frac{1}{2} - \delta_0 \ \text{ implies } \
|\log \zeta(z)| \le \frac{c_{18}\sigma'}{\delta_0} \log t, \]
for some $c_{18}>0$.
Let $1/2+\delta_0< \sigma < 1$, and $\xi>0$ be such that $1+\xi< \sigma'$.
We consider three concentric circles centered at
$\sigma'+it$ with radius $r_1=\sigma'-1-\xi$, $r_2=\sigma'-\sigma$ and
$r_3=\sigma'-1/2-\delta_0$,
and call them $\mathcal C_1, \mathcal C_2$ and $\mathcal C_3$ respectively.
Let
$$M_i = \sup_{z\in \mathcal C_i} |\log \zeta(z)|.$$
From the above bound on $|\log\zeta(z)|$, we get
$$M_3 \le \frac{c_{18}\sigma'}{\delta_0} \log t.$$
Suitably enlarging $c_{18}$, we see that
\[ M_1 \le \frac{c_{18}}{\xi}.\]
Hence we can apply the Hadamard's three circle theorem to conclude that
\[ M_2 \le M_1^{1-\nu} M_3^\nu, \ \text{ for } \ \nu=\frac{\log(r_2/r_1)}{\log(r_3/r_1)}. \]
Thus
\[ M_2 \le \left( \frac{c_{18}}{\xi} \right)^{1-\nu}\left(\frac{c_{18}\sigma' \log t}{\delta_0}\right)^\nu. \]
It is easy to see that
\[ \nu=2-2\sigma + \frac{4\delta_0(1-\sigma)}{1+2\xi-2\delta_0}
+O(\xi) + O\left( \frac{1}{\sigma'}\right). \]
Now we put
\[ \xi=\frac{1}{\sigma'}=\frac{1}{\log\log t} .\]
Hence
\[ M_2 \le \frac{c_{18} \log^\nu t \log\log t}{\delta_0^\nu}
=\frac{c_{19} \log\log t}{\delta_0^\nu}(\log t)^{2-2\sigma+\frac{4\delta_0(1-\sigma)}{1+2\xi-2\delta_0} }, \]
for some $c_{19}>0$.
We observe that
\[ 2-2\sigma+\frac{4\delta_0(1-\sigma)}{1+2\xi-2\delta_0} <
2-2\sigma+\frac{4\delta_0(1-\sigma)}{1-2\delta_0} =\frac{1-2\delta}{1-2\delta_0}. \]
So we get
\[ |\log \zeta(\sigma +it) |
\le c_{19} \log\log t \left(\frac{\log t}{\delta_0}\right)^{\frac{1-2\delta}{1-2\delta_0}},\]
and hence the lemma.
\end{proof}
We put $y=T^{\mathfrak b}$, for a constant $\mathfrak b \ge 8$. Now suppose that
$$\int_T^{\infty} \frac{|\Delta(u)|^2}{u^{2\alpha+1}}e^{-u/y}du \ge \log^2 T,$$
for sufficiently large $T$. Then clearly
$$\Delta(u) =\Omega( u^{\alpha}) .$$
Our next result explores the situation when such an inequality does not hold.
\begin{prop}\label{main-prop}
Let $\delta_0<\delta<\frac{1}{4}-\frac{\delta_0}{2}$.
For $1/4+\delta/2 < \alpha <1/2$, suppose that
\begin{equation}\label{assumption}
\int_T^{\infty} \frac{|\Delta(u)|^2}{u^{2\alpha+1}}e^{-u/y}du \le \log^2 T
\end{equation}
for any sufficiently large $T$.
Then we have
$$\int_{\substack{Re(s)=\alpha \\ t\in J_2(T)}} \frac{|D(s)|^2}{|s|^2}
\ll
1 + \int_T^{\infty} \frac{|\Delta(u)|^2}{u^{2\alpha+1}}e^{-2u/y} du.$$
\end{prop}
Before embarking on a proof, we need the following technical lemmas.
\begin{lem}\label{gamma}
For $0\le \text{Re}(z) \le 1$ and $|Im(z)|\ge \log^2T$, we have
\begin{equation}\label{gamma1}
\int_T^{\infty} e^{-u/y}u^{-z} du =\frac{T^{1-z}}{1-z} + O(T^{-\mathfrak b'})
\end{equation}
and
\begin{equation}\label{gamma2}
\int_T^{\infty} e^{-u/y} u^{-z}\log u\ du =\frac{T^{1-z}}{1-z}\log T + O(T^{-\mathfrak b'}),
\end{equation}
where $\mathfrak b'>0$ depends only on $\mathfrak b$.
\end{lem}
\begin{proof}
By changing variable by $v= u/y$, we get
\[ \int_T^{\infty} \frac{e^{-u/y}}{u^z} du = y^{1-z} \int_{T/y}^{\infty} e^{-v} v^{-z} dv. \]
Integrating the right hand side by parts
\[ \int_{T/y}^{\infty} e^{-v} v^{-z} dv = \frac{e^{-T/y}}{1-z}\left( \frac{T}{y} \right)^{1-z}
+ \frac{1}{1-z} \int_{T/y}^{\infty} e^{-v} v^{1-z} dv \]
It is easy to see that
\[ \int_{T/y}^{\infty} e^{-v} v^{1-z} dv = \Gamma(2-z) + O\left( \left(\frac{T}{y} \right)^{2-Re(z)}\right). \]
Hence (\ref{gamma1}) follows using $e^{-T/y} =1+O(T/y)$ and Stirling's formula along with the assumption
that $ |Im(z)|\ge \log^2T$.
Proof of (\ref{gamma2}) proceeds in the same line and uses the fact that
\[ \int_{T/y}^{\infty} e^{-v} v^{1-z}\log v\ dv = \Gamma'(2-z)
+ O\left( \left(\frac{T}{y} \right)^{2-Re(z)}\log T\right). \]
Then we apply Stirling's formula for $\Gamma'(s)$ instead of $\Gamma(s)$.
\end{proof}
\begin{lem}\label{initial-estimates}
Under the assumption (\ref{assumption}), there exists $T_0$ with $T\le T_0 \le 2T$ such that
\begin{equation*}
\frac{\Delta(T_0)e^{-T_0/y}}{T_0^{\alpha}} \ll \log^2 T,
\end{equation*}
\begin{equation*}
\text{and}\quad\frac{1}{y}\int_{T_0}^{\infty}\frac{\Delta(u)e^{-u/y}}{u^{\alpha}} du \ll \log T.
\end{equation*}
\end{lem}
\begin{proof}
The assumption (\ref{assumption}) implies that
\begin{eqnarray*}
\log^2T &\ge & \int_T^{2T} \frac{|\Delta(u)|^2}{u^{2\alpha+1}}e^{-u/y}du \\
&=& \int_T^{2T} \frac{|\Delta(u)|^2}{u^{2\alpha}}e^{-2u/y}\frac{e^{u/y}}{u} du \\
&\ge & \min_{T\le u\le 2T}\left(\frac{|\Delta(u)|}{u^{\alpha}}e^{-u/y}\right)^2,
\end{eqnarray*}
which proves the first assertion.
To prove the second assertion, we use the previous assertion and Cauchy- Schwartz inequality along with assumption (\ref{assumption}) to get
\begin{eqnarray*}
\left( \int_{T_0}^{\infty}\frac{|\Delta(u)|^2}{u^{\alpha}}e^{-u/y}du \right)^2
&\le & \left( \int_{T_0}^{\infty}\frac{|\Delta(u)|^2}{u^{2\alpha+1}}e^{-u/y}du \right)
\left( \int_{T_0}^{\infty} u e^{-u/y}du \right) \\
&\ll & y^2 \log^2 T.
\end{eqnarray*}
This completes the proof of this lemma.
\end{proof}
We now recall a mean value theorem due to Montgomery and Vaughan \cite{MontgomeryVaughan}.
\begin{notation}
For a real number $\theta$, let $\|\theta\|:=\min_{n\in \mathbb Z}|\theta -n|.$
\end{notation}
\begin{thm}[Montgomery and Vaughan \cite{MontgomeryVaughan}]\label{mean-value}
Let $a_1,\cdots, a_N$ be arbitrary complex numbers, and let $\lambda_1,\cdots,\lambda_N$ be distinct real numbers such that
\[\delta = \min_{m,n}\| \lambda_m-\lambda_n\|>0.\]
Then
\[ \int_0^T \left| \sum_{n\le N} a_n \exp(i\lambda_n t) \right|^2 dt
=\left(T +O\left(\frac{1}{\delta}\right)\right)\sum_{n\le N} |a_n|^2.\]
\end{thm}
\begin{lem}\label{mean-value-estimate}
For $T\le T_0\le 2T$ and $\text{Re}(s)=\alpha$, we have
$$\int_T^{2T} \left| \sum_{n\le T_0}\frac{|\tau(n,\theta)|^2}{n^s}e^{-n/y}\right|^2 t^{-2} dt \ll 1.$$
\end{lem}
\begin{proof}
Using theorem \ref{mean-value}, we get
\begin{eqnarray*}
\int_T^{2T} \left| \sum_{n\le T_0}\frac{|\tau(n,\theta)|^2}{n^s}e^{-n/y}\right|^2 t^{-2} \mathrm{d} t
&\le& \frac{1}{T^2} \left( T \sum_{n\le T_0} |b(n)|^2
+ O\left( \sum_{n\le T_0} n|b(n)|^2\right)\right),\\
&&\text{where } \quad b(n)=\frac{|\tau(n,\theta)|^2}{n^{\alpha}}e^{-n/y}.
\end{eqnarray*}
Thus
\[ \sum_{n\le T_0} |b(n)|^2 \le \sum_{n\le T_0}\frac{d(n)^4}{n^{2\alpha}}
\ll T_0^{1-2\alpha+\epsilon}\
\text{ and }
\ \sum_{n\le T_0} n|b(n)|^2 \le \sum_{n\le T_0}\frac{d(n)^4}{n^{2\alpha-1}}
\ll T_0^{2-2\alpha + \epsilon}\]
for any $\epsilon>0$, since the divisor function $d(n)\ll n^\epsilon$. As we have $\alpha>0$, this completes the proof.
\end{proof}
\begin{lem}\label{mean-value-error}
For $\text{Re}(s)=\alpha$ and $T\le T_0 \le 2T$, we have
$$\int_T^{2T}
\left| \sum_{n\ge 0}\int_0^1 \frac{\Delta(n+x+T_0)
e^{-(n+x+T_0)/y}}{(n+x+T_0)^{s+1}} \mathrm{d} x \right|^2 \mathrm{d} t
\ll \int_T^{\infty} \frac{|\Delta(x)|^2}{x^{2\alpha+1}}e^{-2x/y} \mathrm{d} x.$$
\end{lem}
\begin{proof}
Using Cauchy- Schwarz inequality, we get
\begin{align*}
& \left| \sum_{n\ge 0}\int_0^1 \frac{\Delta(n+x+T_0)}{(n+x+T_0)^{s+1}}
e^{-(n+x+T_0)/y} \mathrm{d} x \right|^2 \\
\le& \int_0^1 \left| \sum_{n\ge 0} \frac{\Delta(n+x+T_0)}{(n+x+T_0)^{s+1}}e^{-(n+x+T_0)/y} \right|^2 \mathrm{d} x.
\end{align*}
Hence
\begin{align*}
&\int_T^{2T}
\left| \int_0^1 \sum_{n\ge 0}\frac{\Delta(n+x+T_0) e^{-(n+x+T_0)/y}}{(n+x+T_0)^{s+1}} \mathrm{d} x \right|^2 \mathrm{d} t \\
\le& \int_T^{2T}\int_0^1 \left| \sum_{n\ge 0} \frac{\Delta(n+x+T_0)}{(n+x+T_0)^{s+1}}
e^{-(n+x+T_0)/y} \right|^2 \mathrm{d} x \mathrm{d} t \\
=& \int_0^1 \int_T^{2T}\left| \sum_{n\ge 0} \frac{\Delta(n+x+T_0)}{(n+x+T_0)^{s+1}}
e^{-(n+x+T_0)/y} \right|^2 \mathrm{d} t \mathrm{d} x.
\end{align*}
From Theorem \ref{mean-value}, we can get
\begin{eqnarray*}
&&\int_T^{2T}\left| \sum_{n\ge 0} \frac{\Delta(n+x+T_0)}{(n+x+T_0)^{s+1}}e^{-(n+x+T_0)/y} \right|^2 \mathrm{d} t\\
&=& T\sum_{n\ge 0}\frac{ |\Delta(n+x+T_0)|^2}{(n+x+T_0)^{2\alpha+2}}
e^{-2(n+x+T_0)/y}
+ O\left( \sum_{n\ge 0} \frac{ |\Delta(n+x+T_0)|^2}{(n+x+T_0)^{2\alpha+1}}\right)\\
&\ll & \sum_{n\ge 0} \frac{ |\Delta(n+x+T_0)|^2}{(n+x+T_0)^{2\alpha+1}}e^{-2(n+x+T_0)/y}.
\end{eqnarray*}
Hence
\begin{eqnarray*}
&&\int_T^{2T}
\left| \sum_{n\ge 0}\int_0^1 \frac{\Delta(n+x+T_0) e^{-(n+x+T_0)/T}}{(n+x+T_0)^{s+1}} \mathrm{d} x \right|^2 \mathrm{d} t \\
&\ll & \int_0^1 \sum_{n\ge 0} \frac{ |\Delta(n+x+T_0)|^2}{(n+x+T_0)^{2\alpha+1}}e^{-2(n+x+T_0)/y}\mathrm{d} x
\ll \int_T^{\infty} \frac{|\Delta(x)|^2}{x^{2\alpha+1}}e^{-2x/y} \mathrm{d} x,
\end{eqnarray*}
completing the proof.
\end{proof}
\begin{proof}[\textbf{Proof of Proposition \ref{main-prop}.} ]
For $s=\alpha+it$ with $1/4 +\delta /2 < \alpha < 1/2$ and $t\in J_2(T)$, we have
\begin{eqnarray*}
\sum_{n=1}^\infty \frac{|\tau(n,\theta)|^2}{n^s} e^{-n/y}
&=& \frac{1}{2\pi i} \int_{2-i\infty}^{2+i\infty}
D(s+w) \Gamma(w) y^w \mathrm{d} w \\
&=& \frac{1}{2\pi i} \int_{2-i\log^2T}^{2+i\log^2T}
+O\left( y^2\int_{\log^2T}^{\infty} |D(s+w)||\Gamma(w)|\mathrm{d} w \right).
\end{eqnarray*}
The above error term is estimated to be $o(1)$. We move the integral to
$$\left[\frac{1}{4}+\frac{\delta}{2} -\alpha-i\log^2T,
\ \frac{1}{4}+\frac{\delta}{2} -\alpha+i\log^2 T\right].$$
Let $\delta'=1/4+\delta/2 -\alpha$.
In this region $\text{Re}(2s+2w)=1/2+\delta$ . So we can apply
Lemma~\ref{estimate-on-J(T)} to
conclude that
$D(s+w) =O(T^\kappa)$, for some constant $\kappa>0$. Thus the integrals along horizontal lines are
$o(1)$.
Since the only pole inside this contour is at $w=0$, we get
\begin{eqnarray*}
\sum_{n=1}^{\infty} \frac{|\tau(n,\theta)|^2}{n^s} e^{-n/y}
= D(s) + \frac{1}{2\pi i}
\int_{\delta'-i\log^2 T}^{\delta'+i\log^2 T}
D(s+w)\Gamma(w)y^w \mathrm{d} w +o(1).
\end{eqnarray*}
Since $\delta' <0$, the remaining integral can be shown to be $o(1)$ for $\mathfrak b\ge 8$.
Using $T_0$ as in Lemma~\ref{initial-estimates}, we now divide the sum into two parts:
$$D(s)= \sum_{n\le T_0} \frac{|\tau(n, \theta)|^2}{n^s} e^{-n/y}
+ \sum_{n > T_0} \frac{|\tau(n, \theta)|^2 }{n^s} e^{-n/y}+o(1).$$
To estimate the second sum, we write
\begin{eqnarray*}
\sum_{n > T_0} \frac{|\tau(n, \theta)|^2}{n^s} e^{-n/y}
&=& \int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^s} \mathrm{d} \left(\sum_{n\le x}|\tau(n, \theta)|^2 \right)\\
&=& \int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^s} \mathrm{d} (\mathcal{M}(x)+\Delta(x))\\
&=& \int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^s} \mathcal{M}'(x)\mathrm{d} x
+ \int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^s} \mathrm{d} (\Delta(x)).
\end{eqnarray*}
Recall that
\[ \mathcal{M}(x)=\omega_1(\theta)x\log x + \omega_2(\theta)x\cos(\theta\log x)
+\omega_3(\theta)x,\]
thus
\[ \mathcal{M}'(x)=\omega_1(\theta)\log x + \omega_2(\theta)\cos(\theta\log x)
-\theta\omega_2(\theta)\sin(\theta\log x)+\omega_1(\theta)+\omega_3(\theta).\]
Observe that
\[ \int_{T_0}^{\infty}\frac{ e^{-x/y}}{x^s}\cos(\theta\log x) \mathrm{d} x
=\frac{1}{2} \int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^{s+i\theta}} \mathrm{d} x
+\frac{1}{2} \int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^{s-i\theta}} \mathrm{d} x.\]
Applying Lemma~\ref{gamma}, we conclude that
\[ \int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^s} \mathcal{M}'(x)\mathrm{d} x = o(1).\]
Integrating the second integral by parts:
\begin{eqnarray*}
\int_{T_0}^{\infty} \frac{ e^{-x/y}}{x^s} \mathrm{d} (\Delta(x))
&=& \frac{e^{-T_0/y} \Delta(T_0)}{T_0^s} \\
&+& \frac{1}{y}\int_{T_0}^\infty \frac{ e^{-x/y}}{x^s}\Delta(x) \mathrm{d} x
-s\int_{T_0}^{\infty}\frac{ e^{-x/y}}{x^{s+1}}\Delta(x) \mathrm{d} x.
\end{eqnarray*}
Applying Lemma~\ref{initial-estimates}, we get
\begin{eqnarray*}
\sum_{n > T_0} \frac{|\tau(n, \theta)|^2}{n^s} e^{-n/y} &=&
s\int_{T_0}^{\infty}\frac{\Delta(x) e^{-x/y}}{x^{s+1}} \mathrm{d} x + O(\log T) \\
&=& s\sum_{n\ge 0} \int_{0}^{1}\frac{\Delta(n+x+T_0) e^{-(n+x+T_0)/y}}{(n+x+T_0)^{s+1}} \mathrm{d} x +O(\log T).
\end{eqnarray*}
Hence we have
$$D(s)= \sum_{n\le T_0} \frac{|\tau(n, \theta)|^2}{n^s} e^{-n/y}
+s\sum_{n\ge 0} \int_{0}^{1}\frac{\Delta(n+x+T_0) e^{-(n+x+T_0)/y}}{(n+x+T_0)^{s+1}} \mathrm{d} x +O(\log T) .$$
Squaring both sides, and then integrating on $J_2(T)$, we get
\begin{align*}
\int_{J_2(T)} \frac{|D(\alpha+it)|^2}{|\alpha+it|^2} \mathrm{d} t
& \ll \int_T^{2T} \left| \sum_{n\le T_0} \frac{|\tau(n, \theta)|^2}{n^s}
e^{-n/y} \right|^2 \frac{ \mathrm{d} t}{t^2} \\
& + \int_T^{2T}
\left| \sum_{n\ge 0} \int_{0}^{1}\frac{\Delta(n+x+T_0) e^{-(n+x+T_0)/y}}{(n+x+T_0)^{s+1}} \mathrm{d} x \right|^2 \mathrm{d} t.
\end{align*}
The proposition now follows using Lemma~\ref{mean-value-estimate} and Lemma~\ref{mean-value-error}.
\end{proof}
We are now ready to prove our main theorem of this section.
\begin{proof}[\textbf{Proof of Theorem \ref{omega_integral}}]
We prove by contradiction. Suppose that (\ref{lb-increasing}) does not hold.
Then, given any $N_0>1$, there exists $T>N_0$ such that
\[\int_T^{\infty} \frac{|\Delta(x)|^2}{x^{2\alpha+1}}e^{-2x/y} \mathrm{d} x
\ll \exp\left( c(\log T)^{7/8} \right),\]
for all $c>0$.
This gives
\[ \int_T^{\infty} \frac{|\Delta(x)|^2}{x^{2\beta+1}}e^{-2x/y} \mathrm{d} x \ll 1, \]
where
\[
\beta=\frac{3}{8}-\frac{c}{2(\log T)^{1/8}} .
\]
We apply Proposition~\ref{main-prop} to get
\begin{equation}\label{contra}
\int_{J_2(T)} \frac{|D(\beta+it)|^2}{|\beta+it|^2} \mathrm{d} t
\ll 1.
\end{equation}
Now we compute a lower bound for the last integral over $J_2(T)$.
Write the functional equation for $\zeta(s)$ as
$$\zeta(s) =\pi^{1/2-s}\frac{\Gamma((1-s)/2)}{\Gamma(s/2)}\zeta(1-s).$$
Using the Stirling's formula for $\Gamma$ function, we get
$$|\zeta(s)|=\pi^{1/2-\beta}t^{1/2-\beta}|\zeta(1-s)|\left(1+O\left(\frac{1}{T}\right)\right),$$
for $s=\beta+it$. This implies
$$|D(\beta+it)|=t^{2-4\beta}\frac{|\zeta(1-\beta+it)^2\zeta(1-\beta-it-i\theta)
\zeta(1-\beta-it+i\theta)|}{|\zeta(2\beta+i2t)|}.$$
Let $\delta_0=1/16$, and
\[\beta=\frac{3}{8} -\frac{c}{2(\log T)^{1/8} }=\frac{1}{2}-\delta \]
with
\[\delta=\frac{1}{8}+\frac{c}{2(\log T)^{1/8} }.\]
Then using Lemma~\ref{estimate-on-J(T)}, we get
\[ |\zeta(1-\beta+it)| = \left|\zeta\left(\frac{1}{2}+\delta+it\right)\right|
\gg \exp\left(\log\log t \left(\frac{\log t}{\delta_0}\right)^{\frac{1-2\delta}{1-2\delta_0}}\right).\]
For $t\in J_2(T)$ we observe that $t\pm\theta \in J_1(T)$, and so the same
bounds hold for $\zeta(1-\beta+it+i\theta)$ and $\zeta(1-\beta+it -i\theta)$.
Further
\[ |\zeta(2\beta+i2t)| = \left|\zeta\left(\frac{1}{2}+\left(\frac{1}{2}-2\delta\right)+i2t\right)\right|
\ll \exp\left(\log\log t \left(\frac{\log t}{\delta_0}\right)^{\frac{4\delta}{1-2\delta_0}}\right).\]
Combining these bounds, we get
\[ |D(\beta+it)| \gg t^{2-4\beta}
\exp\left(-5\log\log t \left(\frac{\log t}{\delta_0}\right)^{\frac{1-2\delta}{1-2\delta_0}}\right).\]
Therefore
\begin{eqnarray*}
\int_{J_2(T)} |D(\beta+it)|^2 \mathrm{d} t
&\gg & \int_{J_2(T)} |D(\beta+it)|^2 \mathrm{d} t
\gg \int_{J_2(T)} |D(\beta+it)|^2 \mathrm{d} t\\
&\gg & T^{4-8\beta}
\exp\left(-10\log\log T\left(\frac{\log T}{\delta_0}\right)^{\frac{1-2\delta}{1-2\delta_0}}\right)
\mu(J_2(T)) \\
&\gg & T^{5-8\beta}\exp\left(-10\log\log T \left(\frac{\log T}{\delta_0}\right)^{\frac{1-2\delta}{1-2\delta_0}}\right),
\end{eqnarray*}
where we use Lemma \ref{size-J(T)} to show that $\mu(J_2(T))\gg T$.
Now putting the values of $\delta$ and $\delta_0$ as chosen above, we get
$$\int_{J(T)} \frac{|D(\beta+it)|^2}{|\beta+it|^2} dt
\gg \exp\left(3c(\log T)^{7/8}\right),$$
since $\frac{1-2\delta}{1-2\delta_0}< 7/8$. This contradicts (\ref{contra}), and hence the theorem follows.
\end{proof}
The following two corollaries are immediate.
\begin{coro}\label{coro:balu_ramachandra1}
For any $c>0$ there exists an $\textbf{X}\text{-Set}$ $\mathcal{S}$, such that for sufficiently large $T$ depending on $c$
there exists an
\[ X \in \left[ T, \frac{T^{\mathfrak b}}{2}\log^2 T\right]\cap \mathcal{S}, \] for which we have
\[ \int_X^{2X} \frac{|\Delta(x)|^2}{x^{2\alpha+1}} dx \ge \exp\left( (c-\epsilon)(\log X)^{7/8}\right)\]
with $\alpha$ as in Theorem \ref{omega_integral} and for any $\epsilon>0$.
\end{coro}
\begin{coro}\label{coro:balu_ramachandra2}
For any $c>0$ there exists an $\textbf{X}\text{-Set}$ $\mathcal{S}$, such that for sufficiently large $T$ depending on $c$
there exists an
\[ x \in \left[ T, \frac{T^{\mathfrak b}}{2}\log^2 T\right]\cap \mathcal{S}, \] for which we have
\[ \Delta(x) \ge x^{3/8} \exp\left( - c(\log x)^{7/8}\right). \]
\end{coro}
\noindent We can now prove a "measure version" of the above result.
\begin{prop}\label{Balu-Ramachandra-measure}
For any $c>0$, let
\[\alpha(x)=\frac{3}{8}- \frac{c}{(\log x)^{1/8} }\] and
$\mathcal{A}=\{x: |\Delta(x)|\gg x^{\alpha(x)} \}$.
Then for every sufficiently large $X$ depending on $c$, we have
\[ \mu(\mathcal{A}\cap [X,2X])=\Omega(X^{2\alpha(X)}).\]
\end{prop}
\begin{proof}
Suppose that the conclusion does not hold, hence
\[ \mu(\mathcal{A}\cap [X,2X]) \ll X^{2\alpha(X)}.\]
Thus for every sufficiently large $X$, we get
\[ \int_{\mathcal{A}\cap [X,2X] }\frac{|\Delta(x)|^2}{x^{2\alpha+1}}dx
\ll X^{2\alpha}\frac{ M(X)}{X^{2\alpha+1}}=\frac{ M(X)}{X},\]
where $\alpha=\alpha(X)$ and $M(X)=\sup_{X\le x \le 2X} |\Delta(x)|^2$.
Using dyadic partition, we can prove
\[ \int_{\mathcal{A}\cap [T,y] }\frac{|\Delta(x)|^2}{x^{2\alpha+1}}dx \ll \frac{M_0(T)}{T}\log T, \
\text{ where }
\ M_0(T) =\sup_{T\le x\le y} |\Delta(x)|^2 \]
and $y=T^{\mathfrak b}$ for some $\mathfrak b>0$ and $T$ sufficiently large.
This gives
\[ \int_T^{\infty} \frac{|\Delta(x)|^2}{x^{2\alpha+1}}e^{-2x/y} dx
\ll \frac{M_0(T)}{T}\log T. \]
Along with (\ref{lb-increasing}), this implies
\[ M_0(T)\gg T\exp\left( \frac{c}{2}(\log T)^{7/8} \right).\]
Thus
\[|\Delta(x)| \gg x^{\frac{1}{2}} \exp\left(\frac{c}{4}(\log x)^{7/8}\right),\]
for some $x\in [T,y].$
This contradicts the fact that $|\Delta(x)| \ll x^{\frac{1}{2}} (\log x)^6.$
\end{proof}
\subsection{Optimality of the Omega Bound }
The following proposition shows the optimality of the omega bound in Corollary~\ref{coro:balu_ramachandra1}.
\begin{prop}\label{prop:upper_bound_second_moment_twisted_divisor}
Under Riemann Hypothesis (RH), we have
\[\int_{X}^{2X}\Delta^2(x)\mathrm{d} x\ll X^{7/4+\epsilon},\]
for any $\epsilon>0$.
\end{prop}
\begin{proof}
Theorem~\ref{thm:effective_perron_formula} (Perron's formula) gives
\begin{equation*}
\Delta(x)=\frac{1}{2\pi}\int_{-T}^{T}\frac{D(3/8+it)x^{3/8+it}}{3/8+it}\mathrm{d} t + O(x^\epsilon),
\end{equation*}
for any $\epsilon>0$ and for $T=X^2$ with $x\in[X, 2X]$. Using this expression for $\Delta(x)$, we write its second moment as
\begin{align*}
&\int_{X}^{2X}\Delta^2(x)\mathrm{d} x
= \int_{X}^{2X}\int_{-T}^{T}\int_{-T}^{T}\frac{D(3/8+ it_1)D(3/8+ it_2)}{(3/8+it_1)(3/8+ it_2)}x^{3/4+ i(t_1+t_2)}\mathrm{d} x \ \mathrm{d} t_1 \mathrm{d} t_2 \\
&\hspace{2.5 cm}
+ O\left(X^{1+\epsilon}(1+|\Delta(x)|)\right)\\
&\ll X^{7/4}\int_{-T}^{T}\int_{-T}^{T}\left|\frac{D(3/8+ it_1)D(3/8+ it_2)}{(3/8+it_1)(3/8+it_2)(7/4+ i(t_1+t_2))}\right|\mathrm{d} t_1 \mathrm{d} t_2
+ O(X^{3/2+\epsilon}).\\
\end{align*}
In the above calculation, we have used the fact that $\Delta(x)\ll x^{\frac{1}{2}+\epsilon}$ as in (\ref{eq:upper_bound_delta}). Also note that for
complex numbers $a, b$, we have $|ab|\leq \frac{1}{2}(|a|^2 + |b|^2)$. We use this inequality with
\[a=\frac{|D(3/8+it_1)|}{|3/8+it_1|\sqrt{|7/4+i(t_1+t_2)|}} \ \text{ and }
\ b=\frac{|D(3/8+it_2)|}{|3/8+it_2|\sqrt{|7/4+i(t_1+t_2)|}},\]
to get
\begin{align*}
\int_{X}^{2X}\Delta^2(x)\mathrm{d} x
&\ll X^{7/4}\int_{-T}^{T}\int_{-T}^{T}\left|\frac{D(3/8+ it_2)}{(3/8+it_2)}\right|^2\frac{\mathrm{d} t_1}{|7/4+ i(t_1+t_2)|} \mathrm{d} t_2 + O(X^{3/2+\epsilon})\\
&\ll X^{7/4}\log X\int_{-T}^{T}\left|\frac{D(3/8+ it_2)}{(3/8+it_2)}\right|^2 \mathrm{d} t_2 + O(X^{3/2+\epsilon}).
\end{align*}
Under RH, $|D(3/8+it_2)|\ll |t_2|^{\frac{1}{2}+\epsilon}$. So we have
\begin{align*}
\int_{X}^{2X}\Delta^2(x)\mathrm{d} x
\ll X^{7/4+\epsilon} \ \text{ for any } \ \epsilon>0.
\end{align*}
\end{proof}
\noindent
\textbf{Note.}
The method we have used in Theorem~\ref{omega_integral} has its origin from the Plancherel's formula in Fourier analysis.
For instance, we may observe from Theorem~\ref{thm:perron_formula} that under Riemann Hypothesis and other suitable conditions
\[\frac{\Delta(e^u)}{e^{u\sigma}}=\frac{1}{2\pi}\int_{-\infty}^{\infty}\frac{D(\sigma+it)e^{iut}}{\sigma + it}\mathrm{d} t \
\text{ for } \frac{1}{4}<\sigma\le\frac{1}{2}.\]
So $\frac{\Delta(e^u)}{e^{u\sigma}}$ is the Fourier transform of $\frac{D(\sigma+it)}{\sigma+it}$. By Plancherel's formula
\[\int_{-\infty}^{\infty}\frac{|\Delta(e^u)|}{e^{2u\sigma}}\mathrm{d} u
=\frac{1}{4\pi^2}\int_{-\infty}^{\infty}\left|\frac{D(\sigma+it)}{\sigma+it}\right|^2\mathrm{d} t.\]
Now we change the variable $u$ to $\log x$ and use the functional equation for $\zeta(s)$ to get
\begin{align*}
\int_1^\infty\frac{\Delta^2(x)}{x^{2\sigma+1}}\mathrm{d} x \asymp \int_1^{\infty}\left|\frac{D(\sigma+it)}{\sigma+it}\right|^2\mathrm{d} t
\gg \int_1^{\infty} t^{2-8\sigma-16\epsilon}
\end{align*}
for any $\epsilon>0$. We may choose $\sigma=\frac{3}{8}-\epsilon$, then the above integral on the left side is convergent. But if
$\Delta(x)\ll x^{\frac{3}{8}-\epsilon}$, then the integral in the right diverges. This gives
\[\Delta(x)=\Omega(x^{\frac{3}{8}-\epsilon}).\]
In \cite{BaluRamachandra1} and \cite{BaluRamachandra2}, Balasubramanian and Ramachandra used this insight to obtain
$\Omega$ bounds for
the error terms in asymptotic formulas for partial sums of
square-free divisors and counting function for non-isomorphic abelian groups.
This method requires the Riemann Hypothesis to be assumed in certain cases. Later Balasubramanian, Ramachandra and Subbarao \cite{BaluRamachandraSubbarao}
modified this technique to apply on error term in the asymptotic formula for the counting function of $k$-full numbers
without assuming Riemann Hypothesis. This method has been used by several authors including \cite{Nowak} and \cite{srini}.
\section{Influence of Measure on $\Omega_\pm$ Results}
In this section, we shall show that for any $\epsilon>0$,
$$\text{ if } \ \Delta(x)\ll x^{3/8+ \epsilon},
\ \text{ then } \
\Delta(x)=\Omega_\pm\left(x^{3/8-\epsilon}\right).$$
This improves our earlier result, which says that $\Delta(x)$ is $\Omega_\pm\left(x^{1/4}\right)$.
Now, we state the main theorem of this section.
\begin{thm}\label{thm:tau_theta_omega_pm}
Let $\Delta(x)$ be the error term of the summatory function of the twisted divisor function as in Theorem~\ref{thm:asymp_formula_tau_n_theta}.
For $c>0$, let
\[\alpha(x)=\frac{3}{8}-\frac{c}{(\log x)^{1/8}} .\]
Let $\delta$ and $\delta'$ be such that
\[0<\delta<\delta'<\frac{1}{8}.\]
Then either
\[\Delta(x)=\Omega\left(x^{ \alpha(x)+ \frac{\delta}{2}}\right) \ \text{ or } \
\Delta(x)=\Omega_{\pm}\left(x^{\frac{3}{8}-\delta'}\right).\]
\end{thm}
\noindent
To prove the above theorem, we estimate the growth of the Dirichlet series $D(\sigma+it)$ by assuming that it does not have poles
in a certain region.
\begin{lem}\label{lem:polynomial_growth_critical_strip_2}
Let $\delta$ and $\sigma$ be such that
\begin{align*}
0<\delta < \frac{1}{8},
\mbox{ and } \quad \frac{3}{8}-\delta \leq \sigma < \frac{1}{2}.
\end{align*}
If $D(\sigma+it)$ does not have a pole in the above mentioned range of $\sigma$, then for
\[\frac{3}{8} -\delta +\frac{\delta}{2(1 + \log\log (3 + |t|))}<\sigma < \frac{1}{2},\]
we have
\[D(\sigma + it)\ll_{\delta, \theta} |t|^{2-2\sigma+\epsilon}\]
for any positive constant $\epsilon$.
\end{lem}
\begin{proof}
Let $s=\sigma+ it$ with $3/8-\delta\leq\sigma<1/2$.
Recall that
\[D(s)=\frac{\zeta^2(s)\zeta(s+i\theta)\zeta(s-i\theta)}{\zeta(2s)}.\]
Using functional equation, we write
\begin{equation}\label{eq:functional_eq_tau_n_theta}
D(s)=\mathcal{X}(s)\frac{\zeta^2(1-s)\zeta(1-s-i\theta)\zeta(1-s+i\theta)}{\zeta(2s)},
\end{equation}
where $\mathcal{X}(s)$ is of order (can be obtained from Stirling's formula for $\Gamma$)
\begin{equation}\label{eq:upperbound_chi}
\mathcal{X}(\sigma+it)\asymp t^{2-4\sigma}.
\end{equation}
Using Stirling's formula and Phragm\'en-Lindel\"of principle, we get
\begin{equation*}
\zeta(1-s)|\ll |t|^{\sigma/2}\log t.
\end{equation*}
So we get
\begin{equation}\label{eq:upperbound_numerator_tauntheta}
|\zeta^2(1-s)\zeta(1-s-i\theta)\zeta(1-s+i\theta)|\ll t^{2\sigma}(\log t)^4.
\end{equation}
Now we shall compute an upper bound for $|\zeta(2s)|^{-1}$. This
can be obtained in a similar way as in Lemma~\ref{estimate-on-J(T)}.
We choose $t\geq 100$. Similar computation can be done when $t$ is negative.
Consider two concentric circles $\mathcal C_{1,1}$ and $\mathcal C_{1, 2}$, centered at $2+ it$ with radii
\[\frac{5}{4} + 2\delta \quad \text{ and }\quad\frac{5}{4}+2\delta - \frac{\delta}{1+ \log\log(|t| + 3)}.\]
The circle $\mathcal C_{1, 1}$ passes through $3/4-2\delta +i2t$ and $\mathcal C_{1, 2}$ passes through
$3/4-2\delta + \delta(1 + \log\log(|t| + 3))^{-1} + i2t$. By our assumption, $\zeta(z)$ does not have any zero for
$|z-2-it|\leq 5/4 + 2\delta$. This implies $\log\zeta(z)$ is a holomorphic function in this region. It is easy to see that
on the larger circle $\mathcal C_{1, 1}$, we have $\log|\zeta(z)|<\sigma'\log t$ for some positive constant $\sigma'$.
We apply Borel-Carath\'eodory theorem to get an upper bound for $\log\zeta(z)$ on $\mathcal C_{1, 2}$ :
\begin{align*}
|\log\zeta(z)|&\leq 3\delta^{-1}(1 + \log\log(t + 3))\left(\sigma'\log t + |\log\zeta(2+it)|\right)\\
& \leq 10\delta^{-1} \sigma' (\log\log t) \log t \quad \text{ for } z\in \mathcal C_{1, 2}.
\end{align*}
We may also note that if $\text{Re}(z-3/4-2\delta)>\delta(\log\log t )^{-1}$
and $\text{Im}(z)\leq t/2 $, then similar arguments give
\[|\log\zeta(z)|<\delta^{-1} \sigma'(\log\log t) \log t,\]
for some positive constant $\sigma'$ that has changed appropriately.
Now we consider three concentric circles $\mathcal C_{2, 1}, \mathcal C_{2, 2}, \mathcal C_{2, 3}$, centered at $\sigma'' + i2t$ and with radii
$r_1=\sigma''-1-\eta, r_2=\sigma''-2\sigma$ and $r_3=\sigma''-\delta_0$ respectively. Here
\[\delta_0=\frac{3}{4}-2\delta + \frac{\delta}{1+\log\log(t+3)}.\]
We shall choose $\sigma''=\eta^{-1}=\log\log t$.
Let $M_1, M_2, M_3$ denote the supremums of $|\log\zeta(z)|$ on $\mathcal C_{2, 1}, \mathcal C_{2, 2}, \mathcal C_{2, 3}$ respectively.
We have already calculated that
\[M_3\leq \delta^{-1}\sigma'(\log\log t)\log t.\]
It is easy to show that
\[M_1\leq \sigma'\log\log t,\]
where $\sigma'$ is again appropriately adjusted. Applying the three circle theorem
we conclude
\[M_2\leq \sigma'(\log\log t)\delta^{-a}\log^{a} t,\]
where
\begin{align*}
a=\frac{\log(r_2/r_1)}{\log(r_3/r_1)}&=\frac{1-2\sigma + \eta}{1-\delta_0 + \eta} + O\left(\frac{1}{\sigma''}\right)\\
&=\frac{4(1-2\sigma)}{1+8\delta} + O_\delta\left(\frac{1}{\log\log t}\right).
\end{align*}
This gives
\begin{equation}\label{eq:upperbound_denominator_tauntheta}
|\zeta(2s)|^{-1}\ll \exp\left(c(\log\log t) (\log t)^{\frac{4(1-2\sigma)}{1+8\delta}}\right),
\end{equation}
for a suitable constant $c>0$ depending on $\delta$.
The bound in the lemma follows from (\ref{eq:functional_eq_tau_n_theta}), (\ref{eq:upperbound_chi}), (\ref{eq:upperbound_numerator_tauntheta}) and
(\ref{eq:upperbound_denominator_tauntheta}).
\end{proof}
Now we complete the proof of Theorem \ref{thm:tau_theta_omega_pm}.
\begin{proof}[Proof of Theorem \ref{thm:tau_theta_omega_pm}]
Let $M$ be any large positive constant, and define
\[\mathcal{A}:=\mathcal{A}(Mx^{\alpha(x)}).\]
Then from Corollary~\ref{coro:balu_ramachandra1}, we have
\[\int_{[T, 2T]\cap \mathcal{A}}\frac{\Delta^2(x)}{x^{2\alpha(T)} +1} \mathrm{d} x \gg \exp\left(c(\log T)^{7/8}\right).\]
Assuming
\begin{equation}\label{eq:upper_bound_measure_asump}
\mu([T, 2T]\cap \mathcal{A})\leq T^{1-\delta} \quad \text{for} \ T>T_0,
\end{equation}
Proposition~\ref{prop:refine_omega_from_measure} gives
\[\Delta(x)=\Omega(x^{\alpha(x) +\delta/2})\]
as $h_0(T)=T^{1-\delta}$,
which is the first part of the theorem.
But if (\ref{eq:upper_bound_measure_asump}) does not hold,
then we have
\[\mu([T, 2T]\cap \mathcal{A})> T^{1-\delta} \]
for $T$ in an $\textbf{X}\text{-Set}$ .
We choose
\[h_1(T)=T^{\frac{3}{8}-\frac{2c}{(\log T)^{1/8}}-\delta}, \
\alpha_1(T)=\frac{3}{8}-\frac{3c}{(\log T)^{1/8}}-\delta, \ \alpha_2(T)=\alpha(T).\]
Let $\delta''$ be such that $\delta<\delta''<\delta'$. If $D(\sigma + it)$
does not have pole for $\sigma>3/8-\delta''$ then by Lemma~\ref{lem:polynomial_growth_critical_strip_2},
$D(\alpha_1(T) + it)$ has polynomial growth. So Assumptions~\ref{as:measure_omega_plus_minus_weak} is
valid.
Since
\[T^{1-\delta}>5h_1(5T/2)T^{1-\alpha_2(T)},\]
by case (ii) of Theorem~\ref{thm:omega_pm_measure} we have
\[\Delta(T)=\Omega_{\pm}\left(T^{\frac{3}{8}-\frac{3c}{(\log T)^{1/8}}-\delta''}\right).\]
The second part of the above theorem follows from the choice $\delta'>\delta''$.
\end{proof}
|
\chapter*{Acknowledgement}
\addcontentsline{toc}{chapter}{Acknowledgement}
I find no words to praise Allah, who is the sole owner of the treasures of knowledge. He, who inspired all the prophets and sent revelation to enable them to guide the humanity to the path of virtue, also inspired me, enlightened my mind, blessed me with sufficient wisdom and laid bare before me the hidden treasures of knowledge. He guided me through all the labyrinths of life and brought me to this stage that I have succeeded in doing something worthwhile. Whenever, I try to think that what I could do without His mercy, my faith in His benevolence, strengthens.
Infinite salutations be upon the Holy Prophet Muhammad (P.B.U.H) who brought Man out of the depth of ignorance and introduced him to the light of knowledge. In his sayings he laid great stress on acquiring knowledge and aroused deep love for it in his Ummah. His teachings encouraged me and proved to be a torch in the exploration of new realities.
Although on the cover page of this thesis only my name is mentioned but I acknowledge the services of so many people whose names may not all be enumerated. Their contributions and sincere motivations are gratefully acknowledged. I beg pardon to all those whose names I did not mention. But there are some names which I can never forget due to their relentless efforts.
Bundles of thanks to my reverend supervisor Sven Heinemeyer who allowed me to share his vast knowledge and get maximum experience from his lifelong experience. I could not have imagined having a better advisor and mentor for my Ph.D study. During the whole research he guided me beyond my expectations. Without his exemplary supervision, I could not do whatever I have done. He showed an attitude and substance of a genius; he continually and convincingly conveyed a spirit of adventure in regard to research. I do not remember a single moment during the whole course of my study, when I found him non-cooperative or reluctant to help. I wish that one day I would become as good an advisor to my students as he has been to me.
Special thanks to my co-supervisor Mario E Gomez, who provided me valuable information and sagacious suggestions during the entire course of my research. I am greatly indebted to him not only for his dynamic guidance and encouragement but also for his positive and constructive criticism. His lively remarks and encouragement refreshed my mind and gave me new energy to complete this cumbrous work. Without his persistent help this dissertation would not have been possible.
I am deeply thankful to Multidark mangement specially Susana Hernandez (Manager Multidark) and Carlos Mu\~{n}oz (Cooridinator Multidark) for their help and support during my Ph.D. I would like to pay my gratitude to Thomas Hahn, who provided me with every possible help whenever it was needed. This work would not have been possible without his help in the implimentation of new Feynman rules in {\tt FeynArts}, {\tt FormCalc}\ and {\tt FeynHiggs}. I also would like to thank Barbara Chazin for her help in writing the Spanish part of the thesis. I also cannot forget the earnest cooperation of my friends Federico van der Pahlen and Leo Galeta, who proved very helpful to me during my entire PhD.
Sincere thanks to my brother Muhammad Qamar, sisters, nephews and nieces whose love and friendly gossips lessened my worries, refreshed my mind and encouraged me to complete my work. Although I lived far from my relatives but communication with them provided emotional atmosphere and concentration that was needed for this cumbersome work.
I would like to pledge my gratitude and deepest emotion to my wife, Ayesha Rehman for her moral support and cheerful inspiration which did not let me down whenever I found my task tiresome and unaccomplished. I will never forget the time when she presented a cup of hot tea and tasty snacks in the middle of the night, keeping a lively smile on her face. I also would like to mention the names of my cute little daughters Ayerah Rehman and Ayezah Rehman whose innocent smile and lively antics provided a soothing and balmy effect on my burdened mind.
I bow my head in honor of my father Nazir Hussain (late) and my mother Hameeda Begum, who always loved me, blessed me with their noble advice, prayed for my success and bestowed upon me sufficient material resources. They spent many sleepless nights while praying for my success and worrying about my health. They always prayed to see the bud of their wishes bloom into a flower. It is all due to them whatever I have achieved today. It is heart wrenching for me to think that when I will return to my country my father will not be there to hug me and appreciate my efforts. After his death, I feel as if I have lost a source of great motivation and strong support.
I also express my deepest and immeasurable appreciations to my dear homeland i.e. Pakistan whose holy soil gave birth to so many talented people. Its serene environment and cheerful people prompted me to do something worthwhile that can add to its glory.
Last but not least I would like to acknowledge Spanish Government for providing me financial support through the Spanish MICINN's Consolider-Ingenio 2010 Programme under grant MultiDark CSD2009-00064. I also express my sincere gratitude to the Instituto de Fisica de Cantabria and Universidad de Cantabria for letting my dreams of being a student there be fulfilled. They generously provided me all sufficient means necessary for my research requirements.
Finally, I would like to say that, apart from me, this research will be immensely important for those who are interested to know about this subject. I hope they will find it valuable. I have tried heart and soul to gather all relevant documents regarding this subject. I do not know how far I am able to do that. Furthermore I do not claim that all the information included in this thesis is described perfectly. There may be shortcomings, factual errors and mistakes which I confess entirely to be mine. I humbly invite positive criticism for future guidance.
\chapter*{List of Publications}
\addcontentsline{toc}{chapter}{List of Publications}
This thesis is based on the following puplications.
\begin{itemize}
\item M.E.~G{\'o}mez, S.~Heinemeyer and M.~Rehman, {\em The Quark Flavor Violating Higgs Decay \boldmath{\ensuremath{h \to \bar b s + b \bar s}} in the MSSM},
{\tt arXiv:1511.04342 [hep-ph]}.
\item M.~G{\'o}mez, S.~Heinemeyer and M.~Rehman, {\em Effects of Sfermion Mixing induced by RGE Running in the Minimal Flavor Violating CMSSM},
{\em Eur. Phys. J.} {\bf C} (2015) 9, 434
{\tt arXiv:1501.02258 [hep-ph]}.
\item M.~G\'omez, T.~Hahn, S.~Heinemeyer, M.~Rehman, {\em Higgs masses and Electroweak Precision Observables in Lepton Flavor Violating MSSM},
{\em Phys.\ Rev.} {\bf D 90} (2014) 074016
{\tt arXiv:1408.0663 [hep-ph]}.
\end{itemize}
\chapter*{Introduction}
\addcontentsline{toc}{chapter}{Introduction}
The Standard Model (SM) of the fundamental interactions\cite{Glashow, Weinberg, Salam}, the results of immense experimental
and theoretical effort, elucidates the ingredients forming our universe and
how do they interact. SM asserts that our universe is made up of fermions
(spin 1/2 particles) interacting through fields of which they are the
sources. Among the fermions, there are six leptons and six quarks
categorized in three families and have their respective anti particles with
opposite quantum numbers. The particles associated with the interaction
fields are bosons (spin 1 particles) namely photon $(\gamma ),$ weak vector
boson ($W^{\pm }$, $Z$) and gluons ($g$) and a scalar particle Higgs ($H$). The gauge bosons act as
force carriers of electromagnetic, weak and strong interactions. Gravity is
not part of SM. Symmetries and invariance principles determine the form
of these forces. SM is based on the gauge group $SU(3)_{\rm C}\times SU(2)_{\rm L}\times U(1)_{\rm Y}$. The renormalizability and gauge invariance demands
the $SU(2)_{\rm L}\times U(1)_{\rm Y}$ symmetry to be spontaneously broken through
Higgs mechanism. All the predictions laid down by SM have been
experimentally confirmed. The discovery of the last missing piece namely the Higgs boson at large hadron collider (LHC) was
announced on 4th July 2012 at Conseil Europeen pour la Recherche Nucleaire (CERN) \cite{ATLAS:2013mma,CMS:yva}, proving SM\ the most accurate and elegant theory at present. In spite of all its successes SM is believed to be
a limiting case of a more general theory. There are well motivated theoretical
as well as experimental reasons which coerce us to go beyond the SM.
The first problem of the SM is related to the neutrino sector.
Neutrinos are strictly massless in the SM. Several key experiments with
solar, atmospheric, reactor and accelerator neutrinos observed the disappearance of electron or muon neutrinos, the
evidence enough for scientists to acquiesce neutrino oscillation\cite{Neutrino-Osc}.
This observation has confirmed that neutrinos have distinct
masses and that 3 neutrino flavors $\nu _{e},$ $\nu _{\mu }$ and $\nu _{\tau
}$ mix among themselves to form 3 mass eigenstates. The fact that neutrinos
are massive and mix implies non-conservation of lepton flavor, hence charged
lepton flavor violating processes (cLFV) are expected in lepton sector just
as quark flavor violating processes arise in quark sector.
The trivial extension to SM to accomodate neutrino masses is to introduce three
fermionic $ SU(3)_{\rm C}\times SU(2)_{\rm L}\times U(1)_{\rm Y}$ singlets (missing right handed neutrinos) and write down
the neutrino Yukawa couplings which generates neutrino masses via electroweak symmetry breaking (EWSB).
However this extension of SM requires extremely small Yukawa couplings and violate lepton number at low energy scale.
As right handed neutrinos and lepton number violation has not been observed at low energy, one should look
for a mechanism that can generate masses for left handed neutrinos at low energy and also respect
the non observation of right handed neutrinos and lepton number violation.
One of the solutions to overcome this problem is the so called ``seesaw mechanism'' \cite{seesaw:I} which can be used
to generate neutrino masses which not only allow large neutrino Yukawa couplings but also explain why left handed
neutrinos are lighter compared to other SM fermions. These mechanisms assume the nature of neutrinos to
be Majorana and existance of very massive particle that couple to the neutrinos in Yukawa analogue.
The neutrino masses are then generated by an effective dimension 5 operator.
This gives rise to neutrino physical states which are not flavor diagonal and consequently generate lepton flavor violation (LFV).
On the other hand SM also lack sufficient explaination in Higgs sector.
For example SM is renormalizable, yet it is believed that SM is valid only up to some
cut-off energy scale. This cut-off can be related to the scale where new physics appear, for example the Planck scale (10$^{19}$ GeV) where quantum gravity becomes important. One-loop corrections to the Higgs mass $M_H$ due to a fermion $f$ in the loop are given by
\begin{equation}
\delta M_{H}^{2}(f)=- \frac{\left\vert \lambda_f \right\vert^2}{8 \pi^2}[\Lambda^2+.....],
\end{equation}%
where $\lambda_f$ represents the fermion coupling to the Higgs field and $\Lambda$ is the ultraviolet cutoff used to regulate the loop integral. If one replace the cutoff by the Planck mass one obtains $\delta M_{H}^{2} \approx 10^{30} \,\, \mathrm{GeV}^2$. One could
cancel these large correction with a bare mass of the same order and opposite
sign. However, these two contributions should cancel with a precision of one
part in $10^{26}$ to provide the observed Higgs mass. This is the so-called ``hierarchy problem''.
Third and equally important issue is the Dark Matter (DM). First speculation about the DM was due to
astronomer Zwicky. In 1933, he observed that the mass from the luminous
matter coming from COMA cluster is much smaller than the total mass we can
derivate from the motion of the cluster member galaxies \cite{Zwicky}. There are now
several pieces of observational evidence for DM in our universe.
Gravitational lensing and the unexpected rotational curves of spiral
galaxies are among these observations that point to there being so-called
``missing mass'' throughout the universe. Recent results from WMAP\cite{WMAP} and PLANCK\cite{Planck}
give us our most accurate value for the total mass in the universe and how
it is divided among different types of matter and energy. There is no SM particle that can serve as a DM candidate.
Due to all these reasons one needs to find renormalizable theories that can
remove quadratic divergences in the Higgs boson mass, a theory that can provide us
with a DM candidate and can explain LFV. Supersymmetric extension of the SM\ namely
Minimal Supersymmetric Standard Model (MSSM)\cite{mssm}, is technically well equipped to
deal with above mentioned discrepancies. The MSSM predicts the existence of a
super-partner for each of the fundamental degree of freedom of the SM with spin
differing by half unit. The presence of super-partners called sparticles in
the loop cancels the quadratic divergences in the Higgs boson mass. Also
the lightest supersymmetric particle (LSP) can be a DM candidate. However, the MSSM, like SM, assume
neutrinos to be massless so simple version of MSSM has to be extended with a mechanism like the seesaw to make it consistent with experimental observation of neutrino masses and mixing\cite{Neutrino-Osc}.
Much of the effort has been devoted at the LHC to discover supersymmetry (SUSY). But as of yet no SUSY particle has been observed at the colliders\cite{Atlas:SUSY2015,CMS:SUSY2015}. Another approach to discover SUSY could be to study the indirect effects of the SUSY particles on other observables\cite{PomssmRep}. Flavor mixing offer a unique prospective in this regard since most of the indirect effects of the SUSY particles involve the flavor mixing observables.
First among these are the Flavor Changing Neutral Current (FCNC) processes in the quark sector. In SM, FCNC processes are absent at tree level and can only occur at one-loop level. The only source of FCNC's in the SM is the Cabibbi Kobayashi Maskawa (CKM) matrix. However these processes are highly supressed due to GIM cancellations. On the other hand, in the MSSM, possible misalignment between the quark and squark mass matrices is another source of flavor violation that can dominate the SM contribution by several orders of magnitude. Any possible experimental deviation from the SM prediction for FCNS's would be a clear evidence of new physics and possibly a hint for MSSM.
Similarly, SM predictions for cLFV are zero. Even seesaw extensions of the SM do not predict
sizable rates for these processes, the cLFV rates in SM seesaw models are almost 40 orders of magnitude
smaller than the present experimental bounds and consequently beyond the experimental reach. On the other hand seesaw extensions of MSSM are well capable of explaining the higher rates (touching the present experimental bounds) for these processes if observed.
Also after the discovery of the Higgs boson with mass $M_{h}\approx 125 \,\, \mathrm{GeV}$, one needs large radiative corrections. One obvious choice would be to choose scalar top mass heavier $\geq 1 \,\, \mathrm{TeV}$. However this could go into the direction of (re-)introducing tuning. One can avoid this problem by choosing large left-right or flavor mixing (instead of assuming heavy scalar top mass). Consequently the issue of flavor mixing in SUSY needs to be explored in detail, which precisely is the aim of the thesis in hand.
Within the MSSM, the possible presence of soft
SUSY-breaking (SSB) parameters in the squark and slepton sector, which are
off-diagonal in flavor space (mass parameters as well as trilinear
couplings) are the most general way to introduce flavor mixing within the
MSSM. For example in MSSM, the off-diagonality in the sfermion mass matrix reflects the
misalignment (in flavor space) between fermions and sfermions mass
matrices, that cannot be diagonalized simultaneously.
This misalignment can be produced from various
origins. For instance, off-diagonal sfermion
mass matrix entries can be generated by Renormalization Group Equations
(RGE) running. Going from a high energy scale, where no flavor violation is
assumed, down to the electroweak (EW) scale can generate such entries due to
presence of non diagonal Yukawa matrices in RGE's. This kind of approach in the
literature is known as the Minimal Flavor Violation (MFV)~\cite{MFV1,MFV2}, where flavor and ${\cal CP}$-violation in quark sector
is assumed to be entirely described by the CKM matrix, even in theories beyond the SM.
MFV scenerios are well motivated due to the fact that they do not introduce new sources of flavor and
${\cal CP}$-violation, which can potentially lead to large non-standard effects in
flavor processes, in conflict with experimental bounds particularly from the kaon and $B_d$
sectors which are well described by the SM upto an accuracy of the $\sim 10\%$
level~\cite{HFAgroup}. For the first and second generation squarks which are
sensitive to the data on $K^0-\bar{K}^0$ and $D^0-\bar{D}^0$ the constraints are very tight. However the third generation system is,
in principle, less constrained, since present data on $B^0-\bar{B}^0$ mixing still leaves
some room for new sources of flavor violation. This opens the prospect for the more general scenerios, namely the Non Minimal Flavor Violation (NMFV) scenerios, other then the MFV ones.
In this thesis we will present a systematic and simultanous study of the effects of flavor mixing on different observables in MFV as well as the NMFV scenerios. As a first step we will study squark and slepton mixing in the MSSM at low energy in Model-Independent (MI) way. For MI approach, we introduce flavor mixing parameters into the sfermion mass matrices by hand and do not consider the possible origin of these parameters. For the squark mixing we will be presenting the effects to electroweak precision observables (EWPO), B-Physics Observables (BPO) and quark flavor violating higgs decays (QFVHD). For slepton mixing we will study the effects to EWPO, higgs boson mass predictions and lepton flavor violating higgs decays (LFVHD).
In the second step we will extend our analysis to the source of flavor mixing and will analyze the flavor mixing induced by RGE running from GUT to EW scale. In this study we will work within the MFV hypothesis for squarks as well as sleptons. Consequently, we will investigate two models (more detailed definitions and citations will follow in the next chapters):
\begin{itemize}
\item[(i)]
the Constrained Minimal Supersymmetric Standard Model (CMSSM), where only flavor violation in the squark sector is present.
\item[(ii)]
the CMSSM augmented by the seesaw type~I mechanism\cite{seesaw:I},
called ``CMSSM-seesaw~I'' below.
\end{itemize}
In many analyses of the CMSSM, or extensions such as the NUHM1 or NUHM2
(see \citere{AbdusSalam:2011fc} and references therein),
the hypothesis of MFV has been used, and it has been assumed
that the contributions coming
from MFV are negligible not only for FCNC processes but for other observables like EWPO and Higgs masses as well, see, e.g.,
\citere{CMSSM-NUHM}.
We will analyze whether this assumption is justified, and
whether including these MFV effects could lead to additional constraints
on the CMSSM parameter space. In this respect we will evaluate in the CMSSM
and in the CMSSM-seesaw~I\ the following set of observables: BPO, in particular \ensuremath{{\rm BR}(B \to X_s \gamma)}, \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ and \ensuremath{\Delta M_{B_s}};
EWPO, in particular $M_W$ and the
effective weak leptonic mixing angle, $\sin^2\theta_{\mathrm{eff}}$; the masses of
the neutral and charged Higgs bosons in the MSSM, as well as cLFV and LFVHD.
The layout of the thesis is as follows. Chapter 1 contains the introduction to SM.
In chapter 2 we introduce MSSM and its seesaw extensions. Chapter 3 is devoted to
the calculational basis of the observables considered in this work.
In chapter 4 we will be presenting our results in the case of squark flavor mixing in MI approach and study the effects to the BPO, EWPO and QFVHD. In chapter 5 slepton mixing effects to EWPO, Higgs mass predictions and LFVHD in MI approach will be presented.
Chapter 6 will be focusing on our analysis in CMSSM and CMSSM-seesaw~I where we present the flavor mixing effects to EWPO, BPO, Higgs boson mass predictions, QFVHD, cLFV and LFVHD. Chapter 7 is devoted to the summary and conclusions.
\chapter{The Standard Model}
Symmetries play an important role in physics. Their presence in a particular
problem often simplifies the problem. Particle physicists, using the concept
of gauge symmetries, are able to build SM, which is a very successful model
to explain the fundamental particles and their interactions. The theory has
been formulated by writing the Lagrangian of the fundamental particles. The
Lagrangian has been written by using the concept of internal symmetries and
gauge invariance. All these aspects are discussed in detail hereafter and
subsequent discussion follows closely~\citeres{GL-Kane,Hollik:2010id}.
\section{Fundamental particles and forces}
Quarks and leptons (collectively called fermions, spin 1/2 particles) are (assumed to be)
elementary particles of nature. There are six types (flavors) of leptons and
quarks placed in three families. Fermions are chiral particles which
connotes that left and right handed fields transform differently. The left
handed components are placed in EW $SU(2)$ doublets and right
handed components are placed in EW singlets
\begin{eqnarray}
L &=&\binom{\nu _{e}}{e}_{L},\ \ \ \binom{\nu _{\mu }}{\mu }_{L},\ \ \
\binom{\nu _{\tau }}{\mu }_{L},\ \ \ \notag \\
&&e_{R},\ \ \ \mu _{R},\ \ \ \tau _{R},\ \label{lepton}
\end{eqnarray}
\begin{eqnarray}
Q &=&\binom{u}{d}_{L},\ \ \ \binom{c}{s}_{L},\ \ \ \binom{t}{b}_{L},\ \
\notag \\
&&\ u_{R},\ \ \ d_{R},\ \ c_{R},\ \ \ s_{R},\ \ \ t_{R},\ \ \ b_{R}
\label{quarks}
\end{eqnarray}
L on the left represents lepton and in the subscript on the right it means
left-handed. Neutrinos being left handed are absent in the EW
singlets. For the quarks another index is required to describe how the
quarks transform under $SU(3)$ transformation.
\begin{equation}
Q_{\alpha }=\binom{u_{\alpha }}{d_{\alpha }}_{L} \label{1.1}
\end{equation}
Quarks and leptons interact through unified EW and strong forces.
These forces are transmitted by the exchange of particles, called gauge
bosons $(\gamma $ , $W^{\pm }$ and $Z)$. \ These are the mediators of
the unified EW force and gluons are the mediators of strong force.
There is an additional particle called Higgs boson predicted by the SM,
which has implications with regard to the origin of mass. It was discovered
recently at LHC CERN \cite{ATLAS:2013mma,CMS:yva}.
\section{Gauge transformation and invariance}
All particles appear to have three kind of gauge invariances, $(U(1)$%
, $SU(2),$ $SU(3))$. The $U(1)$ is related to the electromagnetic charge,
the $SU(2)$ corresponds to the non-abelian weak isospin and $SU(3)$ is
associated with the non-abelian strong (color) charge. In 1961 Glashow \cite%
{Glashow} proposed $SU(2)_{\rm L}\times U(1)_{\rm Y}$ structure of the SM. Weinberg
and Salam \cite{Salam, Weinberg} extended his idea and employed the
hypothesis of spontaneous symmetry breaking in their gauge theory models to
generate masses of gauge boson and fermions. Later on the Glashow, Weinberg and
Salam model achieved the theoretical status when 't Hooft \cite{thooft}
demonstrated that the form of symmetry breaking would not spoil the
renormalizability possessed by the massless theory.
\section{The SM Lagrangian}
The complete Lagrangian for the SM can be written as
\begin{equation}
{\cal L} ={\cal L}_{\rm fermion}+{\cal L}_{\rm gauge}+{\cal L}_{\rm Higgs}
\label{fullLagr}
\end{equation}
${\cal L}_{\rm fermion}$ is given by the relation
\begin{equation}
{\cal L} _{\rm fermion}=\sum\limits_{f=L,Q}\bar{f}\iota \gamma ^{\mu }D_{\mu }f,
\label{Lfermi}
\end{equation}%
where L and Q are given in \refeq{lepton} and \refeq{quarks} and
$D_{\mu } $ is a covariant derivative given by
\begin{equation}
D_{\mu }=\ \partial _{\mu }-\iota g_{1}\frac{Y}{2}B_{\mu }-\iota
g_{2}\frac{\sigma^{i}}{2}W_{\mu }^{i}-\iota g_{3}\frac{\lambda ^{\alpha }}{2}%
G_{\mu }^{\alpha }.
\label{CovariantD}
\end{equation}
It is to be noted that whenever the terms in $ D_{\mu } $ act on a
fermionic state of different matrix form, they give zero, by definition. The
second term represents the $U(1)$ symmetry. $B_{\mu }$ is spin 1 field
needed to maintain gauge invariance and Y is the generator of $U(1)$
transformations, that is also called hypercharge. The $g_{1}$ is the $U(1)$
gauge coupling, the third and the fourth term represents $SU(2)$ and $SU(3)$
symmetries respectively, three W$_{\mu }^{i}$ for $SU(2)$ and eight $G_{\mu
}^{\alpha }$ for $SU(3)$, one for each generator ($\sigma ^{i},\lambda
^{\alpha })$ of transformation whereas $g_{2}$ and $g_{3}$ are the $SU(2)$
and $SU(3)$ gauge couplings respectively.
The Lagrangian for the $SU(2)_{\rm L}\times U(1)_{ \rm Y}$ gauge sector of the theory
is
\begin{equation}
{\cal L}_{\rm gauge}=-\frac{1}{4}W_{\mu \nu }^{i}W_{i}^{\mu \nu }-\frac{1}{4}%
F_{\mu \nu }F^{\mu \nu },
\label{Lgauge}
\end{equation}%
where $F_{\mu \nu }$ is the field strength tensor for $U(1)$ gauge boson $%
B_{\mu }$ and is given by%
\begin{equation}
F_{\mu \nu }=\partial _{\mu }B_{\nu }-\partial _{\nu }B_{\mu }
\label{FStrengthT}
\end{equation}%
and
\begin{equation}
W_{\mu \nu }^{i}=\partial _{\mu }W_{\nu }^{i}-\partial _{\nu }W_{\mu
}^{i}+g_{2}\varepsilon _{ijk}W_{\mu }^{j}W_{\nu }^{k}
\label{WStrengthT}
\end{equation}%
is the field strength tensor for the $SU(2)$ gauge boson, $\varepsilon
_{ijk}$ in the third term of \refeq{WStrengthT} is structure constant and
this term appears due to non-abelian nature of the $SU(2)$ group.
${\cal L}$ does not contain any mass term. In order to generate masses for
fermions and bosons, Higgs mechanism is introduced which will be discussed in \refse{Higgs-Mech}.
\section{Electroweak theory}
By using \refeq{Lfermi}, the $U(1)$ and $SU(2)$ terms for the
Lagrangian of the first generation of leptons can be written as
\begin{eqnarray}
{\cal L}_{\rm lepton} &=&\frac{g_{1}}{2}[Y_{L}(\bar{\nu}_{L}\gamma ^{\mu }\nu
_{L}+\bar{e}_{L}\gamma ^{\mu }e_{L})+Y_{R}\bar{e}_{R}\gamma ^{\mu
}e_{R})]B_{\mu } \notag \\
&&-\frac{g_{2}}{2}[\bar{\nu}_{L}\gamma ^{\mu }\nu _{L}W_{\mu }^{o}-\bar{e}%
_{L}\gamma ^{\mu }e_{L}W_{\mu }^{o}-\sqrt{2}\bar{\nu}_{L}\gamma ^{\mu
}e_{L}W_{\mu }^{+} \notag \\
&&-\sqrt{2}\bar{e}_{L}\gamma ^{\mu }\nu _{L}W_{\mu }^{-}],
\label{Llepton}
\end{eqnarray}%
as neutrinos do not have electromagnetic interactions, the terms of the form
$\frac{g_1}{2} Y_{L}\bar{\nu}_{L}\gamma^{\mu}\nu_{L}B_{\mu}$ must be
avoided. To do so the coefficient $Z_{\mu }\varpropto g_{1}Y_{L}B_{\mu
}-g_{2}W_{\mu }^{0}$ of the term $\bar{\nu}_{L}\gamma ^{\mu }\nu _{L}$ is
assumed to be orthogonal to the electromagnetic field $A_{\mu }$. After
diagonalization one gets
\begin{equation}
A_{\mu }=\frac{g_{2}B_{\mu }-g_{1}Y_{L}W_{\mu }^{0}}{\sqrt{%
g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}},
\label{Amuterm}
\end{equation}
\begin{equation}
Z_{\mu }=\frac{g_{1}Y_{L}B_{\mu }-g_{2}W_{\mu }^{0}}{\sqrt{%
g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}.
\label{Zmuterm}
\end{equation}%
Solving for $B_{\mu }$ and $W_{\mu }^{0}$, one gets
\begin{equation}
B_{\mu }=\frac{g_{2}A_{\mu }+g_{1}Y_{L}Z_{\mu }}{\sqrt{%
g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}},
\label{Bmuterm}
\end{equation}
\begin{equation}
W_{\mu }^{o}=\frac{g_{2}Z_{\mu }-g_{1}Y_{L}A_{\mu }}{\sqrt{%
g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}.
\label{Wmuterm}
\end{equation}
With these definitions the neutral current interactions of the electrons in
\refeq{Llepton} are modified as
\begin{eqnarray}
&&-A_{\mu }[\bar{e}_{L}\gamma ^{\mu }e_{L}\frac{g_{1}g_{2}Y_{L}}{\sqrt{%
g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}+\bar{e}_{R}\gamma ^{\mu }e_{R}\frac{%
g_{1}g_{2}Y_{R}}{2\sqrt{g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}], \notag \\
&&-Z_{\mu }[\bar{e}_{L}\gamma ^{\mu }e_{L}\frac{g_{1}^{2}Y_{L}^{2}-g_{2}^{2}%
}{\sqrt{g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}+\bar{e}_{R}\gamma ^{\mu }e_{R}\frac{%
g_{1}^{2}Y_{L}Y_{R}}{2\sqrt{g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}].
\label{AmuZmu}
\end{eqnarray}%
This gives
\begin{equation}
e=\frac{-g_{1}g_{2}Y_{L}}{\sqrt{g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}
\label{elecCharge1}
\end{equation}%
and
\begin{equation}
e=\frac{-g_{1}g_{2}Y_{R}}{2\sqrt{g_{2}^{2}+g_{1}^{2}Y_{L}^{2}}}
\label{elecCharge2}
\end{equation}%
From \refeq{elecCharge1} and \refeq{elecCharge2} it follows that
\begin{equation}
2Y_{L}=Y_{R}
\label{Yldef}
\end{equation}%
As $g_{1}$ can be redefined to absorb any change in Y$_{L},$ Y$_{L}$ has
been set to $-1$ and \refeq{elecCharge1} is modified as
\begin{equation}
e=\frac{g_{1}g_{2}}{\sqrt{g_{2}^{2}+g_{1}^{2}}}
\label{elecCharge3}
\end{equation}
\begin{equation}
e=g_{2}\sin \theta _{W}
\label{elecCharge4}
\end{equation}%
where $\theta_{W}$ is EW mixing angle with $\sin ^{2}\theta _{W}=(%
\frac{g_{1}}{\sqrt{g_{2}^{2}+g_{1}^{2}}})^{2}$.
\section{Spontaneous symmetry breaking}
\label{SSB-SM}
The Lagrangian in \refeq{Lfermi} does not contain any mass term and
mass terms can not be added explicitly by hand as it would break gauge
invariance. Mass terms are included in SM\ Lagrangian by the Higgs
mechanism, using the idea of spontaneous symmetry breaking. Consider the
Lagrangian for a scalar field $\phi$
\begin{equation}
{\cal L} =\frac{1}{2}\partial _{\mu }\phi \partial ^{\mu }\phi -(\frac{1}{2}%
\mu ^{2}\phi ^{2}+\frac{1}{4}\lambda \phi ^{4});\text{ \ \ \ \ \ \ \ \ \ \ }%
\lambda >0
\label{Lscalar}
\end{equation}%
Here
\begin{equation}
V=\frac{1}{2}\mu ^{2}\phi ^{2}+\frac{1}{4}\lambda \phi ^{4}.
\label{Potential}
\end{equation}%
If $\mu ^{2}>0$ then the vacuum corresponds to $\phi_{0} =0$ but if $\mu ^{2}<0$
then the minimum of the potential is
\begin{equation*}
\frac{\partial V}{\partial \phi }=0
\end{equation*}
\begin{equation}
\phi_0 (\mu ^{2}+\lambda \phi_0 ^{2})=0
\label{1.22}
\end{equation}
\begin{equation}
\phi_0 =\pm \sqrt{\frac{-\mu ^{2}}{\lambda }}=v
\label{1.23}
\end{equation}%
where $v$ is vacuum expectation value (VEV) of Higgs field $\phi $. To
determine the particle spectrum one must study the theory in the region of
the minimum by putting $\phi =v +\eta (x)$ and expanding around $\eta =0.$
Using $\phi =v +\eta (x)$ and \refeq{1.23} in \refeq{Lscalar} yields
\begin{equation}
{\cal L}=\frac{1}{2}\partial _{\mu } \eta \partial ^{\mu } \eta-(\lambda v \eta
^{2}+\lambda v \eta ^{3}+\frac{1}{4}\lambda \eta ^{4})+{\rm const.}
\label{1.24}
\end{equation}
The term in $\eta ^{2}$ has the correct sign so it can be interpreted as mass square and the vacuum does not
have the reflection symmetry of the original Lagrangian. This is called
spontaneous symmetry breaking.
\section{Higgs mechanism}
\label{Higgs-Mech}
The renormalizability and gauge invariance of the theory demands that the
symmetry $SU(2)_{\rm L}\times U(1)_{\rm Y}$ be spontaneously broken through Higgs
mechanism. For this purpose a complex weak doublet of Higgs scalar with hypercharge $Y=1$,
\begin{equation}
\Phi (x) =%
\begin{pmatrix}
\phi ^{+}(x) \\
\phi ^{0}(x) %
\end{pmatrix}
\label{Phi-Def1}
\end{equation}
is introduced which is coupled to the gauge fields through
\begin{equation}
{\cal L}_{\rm Higgs} = (D_{\mu } \Phi)^{\dagger}(D_{\mu } \Phi)-V (\Phi).
\label{LHiggs}
\end{equation}
In this case the covariant derivative is given by
\begin{equation}
D_{\mu }=\ \partial _{\mu }-\iota \frac {g_{1}}{2}B_{\mu }-\iota
g_{2}\frac{\sigma ^{i}}{2}W_{\mu }^{i}.
\label{CovariantD-Higgs}
\end{equation}
The Higgs field self-interaction enters through the Higgs potential with constants $\mu^2$ and $\lambda$,
\begin{equation}
V(\Phi)=-\mu^2 \Phi^{\dagger} \Phi+\frac{\lambda}{4}(\Phi^{\dagger} \Phi)^2.
\label{Higgs-Pot}
\end{equation}
In the ground state, the vacuum, the potential has a minimum. For $\mu^2$, $\lambda > 0 $, the minimum does not
occur for $\Phi = 0$; instead, V is minimized by all non-vanishing field configurations with $\Phi^{\dagger} \Phi =2 \mu^2\ \lambda$.
Selecting the one which is real and electrically neutral, one gets the VEV
\begin{equation}
\langle\Phi\rangle=\frac{1}{\sqrt 2}
\begin{pmatrix}
0 \\
v
\end{pmatrix}.
\label{VEV-SMHiggs}
\end{equation}
Although the Lagrangian is symmetric under gauge transformations of the full $SU(2) \times U(1)$ group,
the vacuum configuration $\langle\Phi\rangle$ does not have this symmetry: the symmetry has been spontaneously
broken. $\langle\Phi\rangle$ is still symmetric under transformations of the electromagnetic subgroup $U(1)_{\rm em}$, which
is generated by the charge Q, thus preserving the electromagnetic gauge symmetry.
The scalar feld in \refeq{Phi-Def1} can be written as
\begin{equation}
\Phi (x) =%
\begin{pmatrix}
\phi ^{+}(x) \\
\left( v+H(x)+\iota \chi(x)\right)/{\sqrt 2} %
\end{pmatrix},
\label{Phi-Def2}
\end{equation}
where the components $\phi^+$, $H$, $\chi$ have vacuum expectation values zero. Expanding the potential in \refeq{Higgs-Pot}
around the vacuum configuration in terms of the components yields a mass term for $H$, whereas $\phi^+$ and
$\chi$ are massless. Exploiting the invariance of the Lagrangian, the components $\phi^+$ and $\chi$ can be eliminated
by a suitable gauge transformation; this means that they are unphysical degrees of freedom (called Higgs
ghosts or would-be Goldstone bosons). Choosing this particular gauge where $\phi^+ = \chi = 0$, denoted as
the unitary gauge, the Higgs doublet field has the simple form
\begin{equation}
\Phi (x) =%
\frac{1}{{\sqrt 2}}\begin{pmatrix}
0 \\
v+H(x) %
\end{pmatrix}.
\label{Phi-Def3}
\end{equation}
The real field $H(x)$ thus describes physical neutral scalar particles, the Higgs bosons, with mass
\begin{equation}
M_{H}={\sqrt 2} \mu= {\sqrt \lambda v}.
\end{equation}
The gauge invariant Higgs–gauge field interaction in the kinetic part
of \refeq{LHiggs} gives rise to mass terms for the vector bosons in the non-diagonal form
\begin{equation}
\frac{1}{2}\left(\frac{g_2}{2}v\right)^2 (W^2_1+W^2_2)+\frac{1}{2}\left(\frac{v}{2}\right)^2
\begin{pmatrix}
W_\mu^3,B_\mu
\end{pmatrix}
\begin{pmatrix}
g_2^2 & g_1 g_2\\
g_1 g_2 & g_1^2
\end{pmatrix}
\begin{pmatrix}
W^{\mu,3}\\
B_\mu
\end{pmatrix}.
\label{LHiggsKinetic}
\end{equation}
The first term can be written as
\begin{equation}
\left(\frac{g_2}{2}v\right)^2 W^+_\mu W^{-\mu}.
\end{equation}
For the charged boson the expected mass term for the Lagrangian would be $m^2 W^+ W^-$, so we can conclude that the charged $W$ boson has indeed acquired a mass
\begin{equation}
M_{W}=\frac{1}{2}g_2 v.
\label{Wmass}
\end{equation}%
The second term in the \refeq{LHiggsKinetic} is not diagonal and we have to define new eigenvalues to find the particles with definite mass. In fact, we already have the answer in hand, because the combination of $B$ and $W^3$ appearing in \refeq{LHiggsKinetic} is just the combination we have called $Z_\mu$ (see \refeq{Zmuterm}).
From \refeq{LHiggsKinetic} and normalization of $Z$ in \refeq{Zmuterm}, we can conclude that the neutral gauge boson $Z$ acquires mass
\begin{equation}
M_{Z}=\frac{v }{2}\sqrt{g_{2}^{2}+g_{1}^{2}}=\frac{M_{W}}{\cos
\theta _{W}},
\label{1.27}
\end{equation}%
while the photon remains massless.
In SM, all quarks and charged fermions get their
masses through the Yukawa couplings with the Higgs field $\Phi$:
\begin{equation}
{\cal L}_{\rm Yukawa} = (Y^{u})_{ij}\bar{Q}_{L_{i}}\Phi^{*} u_{R_{j}}
+ (Y^{d})_{ij}\bar{Q}_{L_{i}}\Phi d_{R_{j}}
+ (Y^{e})_{ij}\bar{L}_{L_{i}}\Phi e_{R_{j}},
\label{Lyukawa}
\end{equation}
where $Y^{u}$, $Y^{d}$ and $Y^{e}$ are up-quark down-quark and charged leptons Yukawa coupling, $Q_{L}$ and $L_{L}$ are left handed quark and lepton doublets, $u_{R}$, $d_{R}$ and $e_{R}$ are $SU(2)_{L}$ -
singlet right-handed fields of up-type quarks, down-type quarks and charged leptons respectively and $i$, $j$ are the generation indices.
After the EW symmetry is broken by a nonzero VEV $v$
of the Higgs field, the Yukawa terms in \refeq{Lyukawa} yield the mass matrices of quarks and charged
leptons
\begin{eqnarray}
(m_{u})_{ij} = (Y^{u})_{ij} v ,\: (m_{d})_{ij} = (Y^{d})_{ij} v ,\: (m_{e})_{ij} = (Y^{e})_{ij} v
\label{yukawamassterms}
\end{eqnarray}
Neutrinos are massless in the SM. They cannot have Dirac masses
because there are no $SU(2)_{\rm L}$ - singlet (“sterile”) right-handed neutrinos $\nu_{R}$ in the SM.
\section{The CKM matrix}
\label{Sec:CKM}
The quark doublets introduced in \refeq{quarks} can have up-down
transitions of the form $u_{i}\rightarrow d_{i}$ mediated by the $W^{\pm },$
where $u_{i}$ can be any up type quark and $d_{i}$ represents any down type
quark. These kind of interactions are absent among lepton doublets due to
conservation of lepton flavor. The mixing among different generations
indicated by rare kaon decay led Cabibbo \cite{cabibbo} to introduce the
mixing angle $\theta _{c}$ called cabibbo angle so that the quark doublet
given in \refeq{quarks} is modified to
\begin{equation}
\binom{u}{d^{\prime }}=\binom{u}{d\cos \theta _{c}+s\sin \theta _{c}}.
\label{1.48}
\end{equation}
This means that the weak eigenstate $d^{\prime }$ is a linear combination
of real mass eigenstates $d$ and $s$. This concept was modified by S.L.
Glashow, J. Iliopoulos and L. Maiani \cite{GIM}. They were able to predict
the existence of charm quark even before its discovery. This completed the
two quark doublets. They explained the mixing with the help of $2\times 2$
unitary matrix. As the concept of quark mixing was indicated through the
rare kaon decays, there was also indication of ${\cal CP}$ violation in these decays.
So it was believed that the ${\cal CP}$ violation has its origin in quark mixing.
This idea was adopted by Kobayashi and Maskawa \cite{KM} to introduce the
third quark doublet, as ${\cal CP}$ violation cannot be accommodated by two
doublets, consequently, they proposed the 3$\times 3$ unitary matrix called
CKM matrix given by
\begin{equation}
V_{\rm CKM}=%
\begin{pmatrix}
V_{ud} & V_{us} & V_{ub} \\
V_{cd} & V_{cs} & V_{cb} \\
V_{td} & V_{ts} & V_{tb}%
\end{pmatrix}.
\label{CKM-Vud}
\end{equation}
Thus the rotation from the $SU(2)$ interaction eigenstate basis, $q^{\rm int}_{L,R}$, to the physical mass eigenstate basis,
$q_{L,R}^{\rm phys}$, is performed by the unitary transformations, $V^{u,d}_{L,R}$:
\begin{equation}
\left( \begin{array}{c} u^{\rm phys}_{L,R} \\ c^{\rm phys}_{L,R} \\ t^{\rm phys}_{L,R} \end{array} \right) =
V^u_{L,R} \left( \begin{array}{c} u^{\rm int}_{L,R} \\ c^{\rm int}_{L,R} \\ t^{\rm int}_{L,R} \end{array} \right)~,~~~~
\left( \begin{array}{c} d^{\rm phys}_{L,R} \\ s^{\rm phys}_{L,R} \\ b^{\rm phys}_{L,R} \end{array} \right) =
V^d_{L,R} \left( \begin{array}{c} d^{\rm int}_{L,R} \\ s^{\rm int}_{L,R} \\ b^{\rm int}_{L,R} \end{array} \right)~,
\end{equation}
such that the quark mass matrices in the physical basis are:
\begin{eqnarray}
\frac{v}{\sqrt 2}V^u_L Y^{u*}V^{u\dagger}_R&=&{\rm diag}\left(m_u,m_c,m_t\right),
\\
\frac{v}{\sqrt 2}V^d_L Y^{d*}V^{d\dagger}_R&=&{\rm diag}\left(m_d,m_s,m_b\right).
\end{eqnarray}
In short, the quark flavour mixing is encoded in the CKM matrix,
\begin{equation}
V_{\rm CKM}= V^u_L V^{d\dagger}_L.
\end{equation}
There are nine parameters in the CKM matrix as shown in \refeq{CKM-Vud} which can be reduced to four in the standard
parametrization \cite{ckmparameters}. The three Euler angles $\theta _{12},\
\theta _{13},\ \theta _{23}$ and one phase factor $\delta ,$ which accounts
for the ${\cal CP}$ violation. The CKM matrix in standard parameterization is given by
\begin{equation}
V_{\rm CKM}=%
\begin{pmatrix}
c_{12}c_{13} & s_{12}c_{13} & s_{13}e^{-\iota \delta } \\
-s_{12}c_{23}-c_{12}s_{13}e^{\iota \delta } &
c_{12}c_{23}-s_{12}s_{23}s_{13}e^{\iota \delta } & s_{23}c_{13} \\
s_{12}s_{23}-c_{12}c_{23}s_{13}e^{\iota \delta } &
-c_{12}s_{23}-s_{12}c_{23}s_{13}e^{\iota \delta } & c_{13}c_{23}%
\end{pmatrix}
\label{1.50}
\end{equation}
where $c_{ij}=\cos \theta _{ij}$ and $s_{ij}=\sin \theta _{ij}$ ($i,j=1,2,3)$. The elements of the CKM matrix exhibit a pronounced
hierarchy. While the diagonal elements are close to unity, the off-diagonal elements are small, such that e.g. $V_{ud}\gg V_{us}\gg V_{ub}$.
In terms of the angles $\theta_{ij}$ we have $s_{12}\gg s_{23}\gg s_{13}$. This fact is usually expressed in terms of the
Wolfenstein parameterization \cite{wolf-parameter}, which can be understood as an expansion in $\lambda=|V_{us}|$. This reads up to order $\lambda^3$
\begin{equation}
V_{\rm CKM}=%
\begin{pmatrix}
1-\frac{\lambda ^{2}}{2} & \lambda & A\lambda ^{3}\left( \rho -\iota \eta
\right) \\
-\lambda & 1-\frac{\lambda ^{2}}{2} & A\lambda ^{2} \\
A\lambda ^{3}\left( 1-\rho -\iota \eta \right) & A\lambda ^{2} & 1%
\end{pmatrix}
\label{1.51}
\end{equation}%
with parameters $A,\rho \ $\ and $\eta$ are assumed to be of order 1. The current values of the CKM elements, obtained from a global fit using all the available measurements and imposing the SM constraints, are collected in the following matrix \cite{Olive:CKM}:
\begin{equation}
V_{\rm CKM}=
\left(\begin{array}{ccc}
0.97427 \pm 0.00014 & 0.22536 \pm 0.00061 & 0.00355 \pm 0.00015 \\
0.22522 \pm 0.00061 & 0.97343 \pm 0.00015 & 0.0414 \pm 0.0012 \\
0.00886^{+0.00033}_{-0.00032} & 0.0405^{+0.0011}_{-0.0012} & 0.99914 \pm 0.00005
\end{array}\right).
\end{equation}
\chapter{Supersymmetry \& Its Seesaw Extention}
A SUSY transformation turns a bosonic state into a fermionic state, and vice versa. The
operator $Q$ that generates such transformations must be an anticommuting spinor, with
\begin{equation}
Q \left|{\rm Boson}\right\rangle=\left|{\rm Fermion}\right\rangle, \quad Q\left|{\rm Fermion}\right\rangle=\left|{\rm Boson}\right\rangle.
\label{SUSY-Trans1}
\end{equation}
Spinors are intrinsically complex objects, so $Q^{\dagger}$ (the hermitian conjugate of $Q$) is also a symmetry
generator. Because $Q$ and $Q^{\dagger}$ are fermionic operators, they carry spin angular momentum 1/2, so it is
clear that SUSY must be a spacetime symmetry.
The No-go theorem \cite{mandula} asserts that it is impossible to mix internal
and Lorentz space time symmetries (when described by the commutators only) in a non-trivial way. If one wants to
extend the space-time structure, one will be left with the only choice of
SUSY with graded Lie algebra. The simplest realization is given by
\begin{equation}
\{Q_{\alpha },Q_{\overset{.}{\alpha }}^{\dagger }\}=2\sigma _{\alpha \overset%
{.}{\alpha }}^{\mu }P_{\mu },
\label{1.35}
\end{equation}
\begin{equation}
\{Q_{\alpha },Q_{\beta }\}=\{Q_{\overset{.}{\alpha }},Q_{\overset{.}{\beta }%
}\}=0 ,
\label{1.36}
\end{equation}
\begin{equation}
\lbrack Q_{\alpha },P_{\mu }]=[Q_{\overset{.}{\alpha }},P_{\mu }]=0,
\label{1.37}
\end{equation}%
where $P^{\mu }$ is the
momentum generator of space-time translations and $\sigma ^{\mu }$ $%
=(1,\sigma ^{1},$ $\sigma ^{2},$ $\sigma ^{3})$. In the following sections, we will give some motivation for SUSY and review main aspacts of the SUSY, in particular the MSSM and its seesaw extension. The subsequent discussion follows closely~\citeres{Martin:1997ns,PomssmRep}.
\section{Motivation}
\label{motivation}
SUSY can successfully explain some of the major deficiencies of SM, as discussed in the introduction,
in a more natural way.
\begin{itemize}
\item \textbf {Hierarchy problem:}
The simplest form of SUSY can solve the hierarchy problem mentioned
in the introduction. Quadratic divergences appearing at one loop level in Higgs mass
vanish due to cancellation between bosons and fermions. Consider for example coupling of the Higgs field $H$ to a Dirac fermion $f$ with
a term in the Lagrangian $-\lambda_f H{\bar f}f$. The one-loop radiative corrections to the Higgs mass $M_H$ will be of the form
\begin{equation}
\delta M_{H}^{2}(f)=- \frac{\left\vert \lambda_f \right\vert^2}{8 \pi^2}[\Lambda^2-2m^2_f\ {\rm ln} \frac{\Lambda}{m_f}+.....]
\label{delMH-fermi}
\end{equation}%
where $m_f$ is the mass of the fermion in the loop. As can be seen from above equation Higgs boson mass is quadratically divergent. In the case of fermion (gauge boson), the chiral (gauge) symmetry constitutes the ``natural barrier'' preventing their masses to become arbitrarily
large. In the case of Higgs boson, there is no symmetry that protects the scalar mass and in the limit $M_H \rightarrow 0$, the symmetry of the model is not increased. SUSY constitutes
so far the most interesting answer to hierarchy problem. As we have mentioned in the introduction, SUSY associates a scalar particle
with every fermionic degree of freedom in the theory with, in principle, identical masses and gauge
quantum numbers. Therefore, in a supersymmetric theory we would have a new
contribution to the Higgs mass at one loop given by
\begin{equation}
\delta M_{H}^{2}({\tilde f})=- \frac{\lambda_{\tilde f}}{8 \pi^2}[\Lambda^2-2m^2_{\tilde f}\ {\rm ln} \frac{\Lambda}{m_{\tilde f}}+.....],
\label{delMH-Sfermi}
\end{equation}%
where $\lambda_{\tilde f}$ is the SUSY particle coupling to the Higgs field, $m_{\tilde f}$ is the mass of the SUSY particle in the loop. If we compare \refeq{delMH-fermi} and \refeq{delMH-Sfermi} we see that with $|\lambda_f|^2=-\lambda_{\tilde f}$ and $m_f = m_{\tilde f}$ we obtain a total correction $\delta M_H^2(f ) + \delta M_H^2({\tilde f} ) = 0 $, i.e. quadratic divergence cancels exactly. If SUSY was an exact symmetry of nature, particles and their
superpartners would have the same mass, and therefore the superpartners should have
been observed in collider experiments. However we have not found scalars exactly degenerate with the SM fermions. This means SUSY can not be an exact symmetry of nature, it must be a broken symmetry. By comparing \refeq{delMH-fermi} and \refeq{delMH-Sfermi}, we can see that we must
still require $|\lambda_f|^2=-\lambda_{\tilde f}$ if we want to ensure the cancellation of quadratic
divergences. SUSY can be broken only in couplings with positive mass dimension, as for instance the
masses. This is called ``soft SUSY-breaking'' \cite{Girardello-SSB}. Now if we take $m^2_{\tilde f}= m^2_f+ \delta^2 $ we
obtain a correction to the Higgs mass,
\begin{equation}
\delta M_H^2(f ) + \delta M_H^2({\tilde f} ) =\frac{\left\vert \lambda_f \right\vert^2}{8 \pi^2} \delta^2 {\rm ln} \frac{\Lambda}{m_{\tilde f}}+.....,
\label{delMH-SSB}
\end{equation}%
and this is only logarithmically divergent and proportional to mass difference between fermion and its scalar partner and is, therefor, under control.
\item \textbf {Gauge coupling unification:}
The idea of gauge unification gets simplified by the SUSY. The coupling
constants $\alpha _{1},$ $\alpha _{2}$ and $\alpha _{3}$ vary with energy
and the rate of the variation of these coupling constants depends on the particle
content of the theory. If these coupling constants are extrapolated to the
higher energies using the particle content of the SM, these do not meet at
the same point. However, when the same extrapolation is repeated using the
particle contents of the SUSY, the three coupling constants meet at the same
point \cite{Dimopoulos-GCU,Marciano-GCU,Amaldi-GCU}. The ``exact'' unification of the gauge couplings within the MSSM
may or may not be an accident. But it provides enough reasons to consider supersymmetric
models seriously as it links SUSY and grand
unification in an inseparable manner\cite{Mohapatra-GCU}.
\item \textbf {Dark Matter candidate:}
There is no particle in the SM that can serve as a DM candidate. However most of the SUSY models provide a particle which might explain missing mass in the universe. For example lightest neutralino could be a DM candidate in the MSSM (see details below).
\item \textbf{Supergravity:}
If SUSY is formulated as a local symmetry, a spin 2 particle corresponding to the graviton, the hypothetical particle
that mediates gravity, is introduced. Then the supersymmetric models of gravity called
supergravity have the elegant feature to link the SM fundamental interactions with gravity\cite{Nieuwenhuizen-SG}.
\end{itemize}
\section{Superpotential}
In this section we will describe the concept of superpotential. The aim is to arrive
at a recipe that will allow to write down the allowed interaction terms of a general
supersymmetric theory, so that later these results can be applied to the special case of the MSSM (see, e.g., the discussion in \citere{Martin:1997ns}).
The single-particle states of a supersymmetric theory fall into irreducible representations of the
SUSY algebra, called supermultiplets. Each supermultiplet contains both fermion and boson
states, which are commonly known as superpartners of each other.
Each supermultiplet contains an equal number of fermionic and bosonic degrees of freedom.
The minimum fermion content of any theory in four dimensions consists
of a single left-handed two-component Weyl fermion $\psi$. Since this is an
intrinsically complex object, it seems sensible to choose as its superpartner
a complex scalar field $\phi$. This combination of a two-component Weyl fermion and a complex scalar field
is called a chiral or matter or scalar supermultiplet.
The next-simplest possibility for a supermultiplet contains a spin-1 vector boson. If the theory is to
be renormalizable, this must be a gauge boson that is massless, at least before the gauge symmetry is
spontaneously broken. Its superpartner is therefore a massless spin-1/2 Weyl fermion, called gaugino.
Such a combination of spin-1/2 gauginos and spin-1
gauge bosons is called a gauge or vector supermultiplet.
The simplest action one can write down for chiral supermultiplet
just consists of kinetic energy terms for scalar and fermionic fields.
\begin{equation}
S=\int d^{4}x({\cal L}_{\rm scalar}+{\cal L}_{\rm fermion})
\label{1.38}
\end{equation}%
with
\begin{equation}
{\cal L}_{\rm scalar}=\partial ^{\mu }\phi ^{\ast }\partial _{\mu
}\phi ,\qquad {\cal L}_{\rm fermion}=\iota {\bar \psi} \bar{\sigma}^{\mu
}\partial _{\mu }\psi.
\end{equation}
This is called the massless, non-interacting Wess-Zumino model.
The number of fermionic degrees of freedom must be equal to the number of
bosonic degrees of freedom. But scalar field contains one degree of freedom
and one can add one more if a complex scalar field is introduced. However, a fermionic
field carries at least four components. Two of these degrees of freedom can be fixed by Dirac equation. It means, the algebra of SUSY only closes on-shell in this formulation. This can be fixed by a trick. One can invent a new complex scalar field F, which does not
have a kinetic term. Such fields are called auxiliary, and they are really just book-keeping devices that
allow the SUSY algebra to close off-shell. Thus the free part of the
Lagrangian is
\begin{equation}
{\cal L}_{\rm free}=\partial ^{\mu }\phi ^{\ast i}\partial _{\mu
}\phi_{i}+\iota {\bar \psi} ^{i}\bar{\sigma}^{\mu
}\partial _{\mu }\psi_i+ F^{*i}F_{i}~,
\end{equation}
where it is summed over repeated indices $i$ (not to be confused with the
suppressed spinor indices). Now the most general set of renormalizable interactions for
these fields that is consistent with SUSY must have
dynamical field content with mass dimension $\leq$ 4. So, the candidate terms that are also SUSY invariant
are:
\begin{equation}
{\cal L}_{\rm int}=-\frac{1}{2} W^{ij} \psi_{i}\psi_{j}+ W^{i}F_{i}+c.c.
\end{equation}
A very useful object $W$ called the superpotential is introduced.
\begin{equation}
W=\frac{1}{2} M^{ij} \phi_{i}\phi_{j}+ \frac{1}{6} y^{ijk} \phi_{i}\phi_{j}\phi_{k}
\end{equation}
where $M^{ij}$ is a symmetric mass matrix for the fermion fields, and $y^{ijk}$ is
the Yukawa coupling of the scalar $\phi_k$. One can write
\begin{equation}
W^{ij}=\frac{\partial^2}{\partial \phi_i \partial \phi_j} W, \quad W^{i}=\frac{\partial W}{\partial \phi_i}.
\end{equation}
The auxiliary fields $F_i$
and $F^{\ast i}$ can be eliminated using their classical equations of motion.
The part of ${\cal L}_{\rm free}$ + ${\cal L}_{\rm int}$ that contains the auxiliary fields is $F_iF^{\ast i}$ + $W^iF_i$ +
$W^{\ast}_{i} F^{\ast i}$, leading to the equations of motion
\begin{equation}
F_{i}=-W^*_i; \quad F^{*i}=-W^{i}.
\label{EQM-F}
\end{equation}
After making the replacement \refeq{EQM-F} in ${\cal L}_{\rm free}$ + ${\cal L}_{\rm int}$, one obtains the Lagrangian density
\begin{equation}
{\cal L}=\partial ^{\mu }\phi ^{\ast i}\partial _{\mu
}\phi_{i}+\iota {\bar \psi} ^{i}\bar{\sigma}^{\mu
}\partial _{\mu }\psi_i-\frac{1}{2} (W^{ij} \psi_{i}\psi_{j}+W^{*}_{ij} {\bar \psi}^{i}{\bar \psi}^{j})-W^{i} W^{*}_{i}.
\end{equation}
In short, the most general non-gauge interactions
for chiral supermultiplets are determined by a single analytic function of
the complex scalar fields, the superpotential $W$.
\section{The MSSM}
The supersymmetric version of SM is called the MSSM with $N=1$ generators, where N refer to
the number of distinct copies of $Q$ and $Q^{\dagger}$.
According to the MSSM each of the fundamental particle of SM has a superpartner
with spin differing by half unit. These particles are placed either in
chiral or gauge supermultiplet as shown in \refta{tab:chiral} and \refta{tab:gauge}.
\renewcommand{\arraystretch}{1.4}
\begin{table}[tb]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Superfields & spin 0 & spin 1/2 & $(SU(3)_C, \> SU(2)_L,\> U(1)_Y)$
\\
\hline
\hline
$\hat Q$ & $({\widetilde u}_L\>\>\>{\widetilde d}_L )$&
$(u_L\>\>\>d_L)$ & $(\>{\bf 3},\>{\bf 2}\>,\>{1\over 6})$
\\
$\hat U$
&${\widetilde u}^*_R$ & $u^\dagger_R$ &
$(\>{\bf \overline 3},\> {\bf 1},\> -{2\over 3})$
\\ $\hat D$ &${\widetilde d}^*_R$ & $d^\dagger_R$ &
$(\>{\bf \overline 3},\> {\bf 1},\> {1\over 3})$
\\ \hline
$\hat L$ &$({\widetilde \nu}\>\>{\widetilde e}_L )$&
$(\nu\>\>\>e_L)$ & $(\>{\bf 1},\>{\bf 2}\>,\>-{1\over 2})$
\\
$\hat E$
&${\widetilde e}^*_R$ & $e^\dagger_R$ & $(\>{\bf 1},\> {\bf 1},\>1)$
\\ \hline
$\hat H_2$ &$({\cal H}_2^+\>\>\>{\cal H}_2^0 )$&
$(\widetilde {\cal H}_2^+ \>\>\> \widetilde {\cal H}_2^0)$&
$(\>{\bf 1},\>{\bf 2}\>,\>+{1\over 2})$
\\ $\hat H_1$ & $({\cal H}_1^0 \>\>\> {\cal H}_1^-)$ & $(\widetilde {\cal H}_1^0 \>\>\> \widetilde {\cal H}_1^-)$&
$(\>{\bf 1},\>{\bf 2}\>,\>-{1\over 2})$
\\ \hline
\end{tabular}
\caption[Chiral supermultiplets in the MSSM]{Chiral supermultiplets in the MSSM, their field content, and their representations in the gauge groups.
Here $u=u,c,t$; $d=d,s,b$; $e=e,\mu,\tau$ and $\nu=\nu_{e},\nu_{\mu},\nu_{\tau}$.}
\vspace{-0.6cm}
\label{tab:chiral}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1.55}
\begin{table}[t]
\begin{center}
\vspace{1 cm}
\begin{tabular}{|c|c|c|c|c|}
\hline
Superfields & spin 1/2 & spin 1 & $(SU(3)_C, \> SU(2)_L,\> U(1)_Y)$\\
\hline
\hline
$\hat G^a$ &$ \widetilde g$& $g$ & $(\>{\bf 8},\>{\bf 1}\>,\> 0)$
\\
\hline
$\hat W^i$ & $ \widetilde W^\pm\>\>\> \widetilde W^0 $&
$W^\pm\>\>\> W^0$ & $(\>{\bf 1},\>{\bf 3}\>,\> 0)$
\\
\hline
$\hat B$ & $\widetilde B^0$&
$B^0$ & $(\>{\bf 1},\>{\bf 1}\>,\> 0)$
\\
\hline
\end{tabular}
\caption[Gauge supermultiplets in
the MSSM]{Gauge supermultiplets in
the MSSM, their field content, and their representations in the gauge groups.}
\vspace{-0.45cm}
\label{tab:gauge}
\end{center}
\end{table}
In order to keep anomaly cancellation, contrary to the SM a second
Higgs doublet is needed~\cite{glawei}. One Higgs doublet, ${\cal H}_{1}$,
gives mass to the $d$-type fermions (with weak isospin -1/2), the
other doublet, ${\cal H}_{2}$, gives mass to the $u$-type fermions (with weak
isospin +1/2).
All SM multiplets, including the two Higgs doublets, are extended to
supersymmetric multiplets, resulting in scalar partners for quarks and
leptons (``squarks'' and ``sleptons'') and fermionic partners for the
SM gauge boson and the Higgs bosons (``gauginos'' and ``gluinos'') as shown in
\refta{tab:chiral} and \refta{tab:gauge}.
The mass eigenstates of the gauginos are linear combinations of these
fields, denoted as ``neutralinos'' and ``charginos''. Also the left- and
right-handed squarks (and sleptons) can mix, yielding the mass
eigenstates (denoted by the indices
$1,2$ instead of $L,R$). The EW interaction eigenstates and mass eigenstates
of the MSSM particle spectrum are given in \refta{tab:MSSMparticles}. Here
mass eigenstates are written with the assumption of no flavor violation. If
flavor violation is assumed, all six up- and down-type squarks and all six charged sleptons mix separately
to give six mass eigenstates (see \refse{sec:sfermions}).
\begin{table}[htb]
\begin{center}
\renewcommand{\arraystretch}{1.8}
\begin{tabular}{|c|c|c|} \hline
\rule[-2ex]{0mm}{5ex}%
Particle & Electroweak eigenstate & Mass Eigenstate \\
\hline \hline
squarks & $\tilde u_L$,$\tilde u_R$,$\tilde d_L$,$\tilde d_R$ & $\tilde u_1$,$\tilde u_2$,$\tilde d_1$,$\tilde d_2$ \\
& $\tilde c_L$,$\tilde c_R$, $\tilde s_L$,$\tilde s_R$ & $\tilde c_1$,$\tilde c_2$,$\tilde s_1$,$\tilde s_2$ \\
& $\tilde t_L$,$\tilde t_R$,$\tilde b_L$,$\tilde b_R$ & $\tilde t_1$,$\tilde t_2$,$\tilde b_1$,$\tilde b_2$ \\
\hline
sleptons & $\tilde e_L$,$\tilde e_R$,$\tilde \nu_e$ & $\tilde e_1$,$\tilde e_2$,$\tilde \nu_e$ \\
& $\tilde \mu_L$,$\tilde \mu_R$,$\tilde \nu_{\mu}$ & $\tilde \mu_1$,$\tilde \mu_2$,$\tilde \nu_{\mu}$ \\
& $\tilde \tau_L$,$\tilde \tau_R$,$\tilde \nu_{\tau}$ & $\tilde \tau_1$,$\tilde \tau_2$,$\tilde \nu_{\tau}$ \\
\hline
neutralinos & $\tilde B$,$\tilde W$,$\tilde H_u^0$,$\tilde H_d^0$ & $\tilde \chi_1^0$,$\tilde \chi_2^0$,$\tilde \chi_3^0$,$\tilde \chi_4^0$ \\
\hline
charginos & $\tilde W^{\pm}$,$\tilde H_u^+$,$\tilde H_d^{-}$ & $\tilde \chi_1^{\pm}$,$\tilde \chi_2^{\pm}$\\
\hline
gauge boson & $B$,$W^1$,$W^2$,$W^3$ & $W^{\pm}$,$Z$,$\gamma$ \\
\hline
gluon and gluino & $g$,$\tilde g$ & $g$,$\tilde g$ \\
\hline
\end{tabular}
\end{center}
\caption[The EW interaction eigenstates and mass eigenstates of the MSSM.]
{The EW interaction eigenstates and mass eigenstates of the MSSM particles. No flavor mixing is assumed here.}
\label{tab:MSSMparticles}
\end{table}
At knowing the particle content of MSSM, one can write the most general $%
SU(3)_{\rm C}\times SU(2)_{\rm L}\times U(1)_{\rm Y}$ gauge invariant and renormalizable
superpotential as \cite{mssm}
\begin{eqnarray}
W_{\rm MSSM} =\epsilon _{ab}[Y_{ij}^{e}\hat{H}_{1}^{a}\hat{L}_{i}^{b}%
\hat{E}_{j}^{C}+Y_{ij}^{d}\hat{H}_{1}^{a}\hat{Q}_{i}^{b}\hat{%
D}_{j}^{C}+Y_{ij}^{u}\hat{H}_{2}^{a}\hat{Q}_{i}^{b}\hat{U}%
_{j}^{C}-\mu \hat{H}_{1}^{a}\hat{H}_{2}^{b}]
\label{superpotential}
\end{eqnarray}
where $\hat L$ represents the chiral multiplet of a $SU(2)_{\rm L}$ doublet
lepton, $\hat E$ a $SU(2)_{\rm L}$ singlet charged lepton, $\hat H_1$ and $\hat H_2$ two Higgs multiplets with opposite hypercharge.
Similarly $\hat Q$, $\hat U$ and $\hat D$ represent chiral multiplets of quarks of a
$SU(2)_{\rm L}$ doublet and two singlets with different $U(1)_{\rm Y}$ charges
whereas $i,j=1,2,3$ are family indices and a, b are $SU(2)$ indices. The symbol $\epsilon_{ab}$ is an anti-symmetric tensor with $\epsilon_{12}=1$.
As mentioned in \refse{motivation}, SUSY is not an exact symmetry of nature. It must be a broken symmetry.
The general set-up for the SSB
parameters is given by~\cite{mssm}
\begin{eqnarray}
\label{softbreaking}
-{\cal L}_{\rm soft}&=&(m_{\tilde Q}^2)_i^j {\tilde {\cal Q}}^{\dagger i}
{\tilde {\cal Q}}_{j}
+(m_{\tilde U}^2)^i_j {\tilde {\cal U}}_{i}^* {\tilde {\cal U}}^j
+(m_{\tilde D}^2)^i_j {\tilde {\cal D}}_{i}^* {\tilde {\cal D}}^j
\nonumber \\
& &+(m_{\tilde L}^2)_i^j {\tilde {\cal L}}^{\dagger i}{\tilde {\cal L}}_{j}
+(m_{\tilde E}^2)^i_j {\tilde {\cal E}}_{i}^* {\tilde {\cal E}}^j
\nonumber \\
& &+m^2_{H_1}{\cal H}_1^{\dagger} {\cal H}_1
+m^2_{H_2}{\cal H}_2^{\dagger} {\cal H}_2
+(B \mu {\cal H}_1 {\cal H}_2
+ {\rm h.c.})
\nonumber \\
& &+ ( ({\bar A}^d)_{ij}{\cal H}_1 {\tilde {\cal D}}_{i}^*{\tilde {\cal Q}}_{j}
+({\bar A}^u)_{ij}{\cal H}_2 {\tilde {\cal U}}_{i}^*{\tilde {\cal Q}}_{j}
+({\bar A}^e)_{ij}{\cal H}_1 {\tilde {\cal E}}_{i}^*{\tilde {\cal E}}_{j}
\nonumber \\
& & +\frac{1}{2}M_1 {\tilde B}_L^0 {\tilde B}_L^0
+\frac{1}{2}M_2 {\tilde W}_L^a {\tilde W}_L^a
+\frac{1}{2}M_3 {\tilde G}^a {\tilde G}^a + {\rm h.c.}).
\end{eqnarray}
Here we have used calligraphic capital letters for the sfermion fields in the
interaction basis with generation indices,
\begin{eqnarray}
\tilde {\cal U}_{1,2,3}&=&\tilde u_R,\tilde c_R,\tilde t_R ; \quad
\tilde {\cal D}_{1,2,3}=\tilde d_R,\tilde s_R,\tilde b_R ; \quad
\tilde {\cal Q}_{1,2,3}=(\tilde u_L \, \tilde d_L)^T, (\tilde c_L\, \tilde s_L)^T, (\tilde t_L \, \tilde b_L)^T \notag \\
\tilde {\cal E}_{1,2,3}&=&\tilde e_R,\tilde \mu_R,\tilde \tau_R ; \quad
\tilde {\cal L}_{1,2,3}=(\tilde \nu_{eL} \, \tilde e_L)^T, (\tilde \nu_{\mu L}\, \tilde \mu_L)^T, (\tilde \nu_{\tau L} \, \tilde \tau_L)^T
\end{eqnarray}
and all the gauge indices have been omitted.
Here $m_{\tilde Q}^2$ and
$m_{\tilde L}^2$ are $3 \times 3$
matrices in family space (with $i,j$ being the
generation indeces) for the soft masses of the
left handed squark ${\tilde {\cal Q}}$ and slepton ${\tilde {\cal L}}$
$SU(2)$ doublets, respectively. $m_{\tilde U}^2$, $m_{\tilde D}^2$ and
$m_{\tilde E}^2$ contain the soft masses for right handed up-type squark
${\tilde {\cal U}}$, down-type squarks ${\tilde {\cal D}}$ and charged
slepton ${\tilde {\cal E}}$ $SU(2)$ singlets, respectively. $\bar A^u$,
$\bar A^d$ and $\bar A^e$ are the $3 \times 3$ matrices for the trilinear
couplings for up-type squarks, down-type
squarks and charged slepton, respectively.
$m_{H_1}$ and $m_{H_2}$ contain the soft
masses of the Higgs sector. In the last line $M_1$, $M_2$ and $M_3$
define the bino, wino and gluino mass terms, respectively.
It is noteworthy that the terms in \refeq{superpotential} conserve lepton and baryon numbers,
which is neither required by gauge invariance nor by renormalization. One can add the terms of the form
\begin{eqnarray}
\epsilon _{ab}[\lambda _{ijk}%
\hat{L}_{i}^{a}\hat{L}_{j}^{b}\hat{E}_{k}^{c}+\lambda
_{ijk}^{\prime }\hat{L}_{i}^{a}\hat{Q}_{j}^{b}\hat{D}%
_{k}^{c}+\lambda _{ijk}^{\prime \prime }\hat{U}_{i}^{c}\hat{D}%
_{j}^{c}\hat{D}_{k}^{c}]
\end{eqnarray}
to \refeq{superpotential} where $\lambda _{ijk}$, $\lambda_{ijk}^{\prime }$ and $\lambda _{ijk}^{\prime \prime }$ are the R-parity violating couplings.
However these terms violate either lepton or baryon number by one unit, and presence of
these terms have dangerous impact on matter i.e. these terms
lead to fast proton decay, which is in contradiction to experimental observations.
So in order to avoid this situation we have to introduce ad-hoc symmetry,
known as R-parity, defined as \cite{fayet,Ferr}
\begin{equation}
R_{p}=(-1)^{3(B-L)+2s}
\label{1.46}
\end{equation}
where $B$ represents baryon number, $L$ the lepton number and $s$ the
intrinsic spin of the particle. Invariance of Lagrangian under R-parity
implies $-1$ phase for sparticle and $+1$ for SM particles. Under this condition the $L$ and $B$ number violating processes are
prohibited, this prevents the proton from decaying rapidly.
Though R-parity is introduced by hand just to safe proton decay, but it has
large impact on particle physics phenomenology. The
conservation of R-parity demands that the sparticles are always produced in
pairs. e.g. the LSP must be stable and
is assumed as the excellent candidate for DM. To detect a LSP, collider experiments
search for missing transverse energy that would arise if one of these particles were created
during a collision process and escaped undetected. For example, at the LHC, the major SUSY production processes are gluinos $\tilde g$ and
squarks $\tilde q$ e.g., $ p + p \rightarrow \tilde g + \tilde q$. These then decay into lighter SUSY particles. The final states involve two lightest neutralinos $\tilde \chi^{0}_{1}$ (giving rise to missing transverse energy $E_{T}^{\rm miss}$), quarks (jets) and leptons. The signal is
thus $E_{T}^{\rm miss}$+ jets +leptons, which should be observable at the LHC detectors.
\subsection{The scalar fermion sector}
\label{subsec:sfermions}
The squarks and charged sleptons mass term (sneutrinos being treated differently) of the MSSM Lagrangian is given by
\begin{equation}
{\cal L}_{m_{\tilde{f}}} = -\frac{1}{2}
\Big( \tilde{f}_L^{\dag},\tilde{f}_R^{\dag} \Big)\; {\bf M_{\tilde f}} \;
\left( \begin{array}{c} \tilde{f}_L \\[0.5ex] \tilde{f}_R \end{array} \right)~,
\label{squarksmassmatrix}
\end{equation}
where
\begin{equation}
\renewcommand{\arraystretch}{1.5}
{\bf M_{\tilde f}} = \left( \begin{array}{cc} M_{\tilde{f}}^2 + M_Z^2\cos 2\beta(I_3^f-Q_fs_\mathrm{w}^2) + m_f^2
& m_f X_f \\
m_f X_f
& M_{\tilde{f}'}^2 + M_Z^2 \cos 2\beta Q_fs_\mathrm{w}^2 + m_f^2 \end{array} \right) ,
\label{squarkmassenmatrix}
\end{equation}
with $X_f=A_f - \mu \{\cot \beta;\tan \beta\}$ and $\tan \beta = v_2/v_1$, the ratio of the VEV's of the two Higgs doublets, corresponds to $d$-type squarks and charged sleptons whereas $\cot \beta$ corresponds to $u$-type squarks.
The SSB term $M_{\tilde{f}'}$ represents right handed squarks and right handed charged sleptons.
Sneutrino mass term is given by
\begin{equation}
M_{\tilde{\nu}}^2
= m_{\tilde{L}}^2 + \left(\frac{1}{2} M_Z^2 \cos 2\beta \right)
\label{sneutrinomass}
\end{equation}
In order to diagonalize the sfermion mass matrix and to determine the physical mass
eigenstates the following rotation has to be performed:
\begin{equation}
\left( \begin{array}{c} \tilde{f}_1 \\ \tilde{f}_2 \end{array} \right) = \left( \begin{array}{cc} \cos\tsf & \sin\tsf \\ -\sin\tsf & \cos\tsf \end{array} \right)
\left( \begin{array}{c} \tilde{f}_L \\ \tilde{f}_R \end{array} \right) .
\label{squarkrotation}
\end{equation}
The mixing angle $\theta\kern-.20em_{\tilde{f}}$ is given for $\tan \beta > 1$ by:
\begin{eqnarray}
\cos\tsf &=& \frac{\sqrt {M_{\tilde f}^2 + M_Z^2\cos 2\beta(I_3^f-Q_fs_\mathrm{w}^2) + m_f^2 - m_{\tilde f_2}^2}}
{\sqrt{m_{\tilde f_1}^2-m_{\tilde f_2}^2 }} \\
\sin\tsf &=& \frac{m_f X_f}
{
\sqrt {M_{\tilde f}^2 + M_Z^2\cos 2\beta(I_3^f-Q_fs_\mathrm{w}^2) + m_f^2 - m_{\tilde f_2}^2} \sqrt{m_{\tilde f_1}^2-m_{\tilde f_2}^2 }
} ~.
\label{stt}
\end{eqnarray}
The masses are given by the eigenvalues of the mass matrix:
\begin{eqnarray}
\label{Squarkmasse}
m_{\tilde{f}_{1,2}}^2 &=& m_f^2+\frac{1}{2} \Big[M_{\tilde{f}}^2 + M_{\tilde{f}'}^2 + M_Z^2 \cos 2\beta I^f_3 \\
& & \mp \sqrt {[M_{\tilde f}^2-M_{\tilde{f}'}^2+ M_Z^2\cos 2\beta(I_3^f-Q_fs_\mathrm{w}^2)]^2+4m_f^2|X_f|^2}~\Big].
\end{eqnarray}
Since the non-diagonal entry of the mass matrix
\refeq{squarkmassenmatrix} is proportional to the fermion mass,
mixing becomes particularly important for $\tilde f = \tilde{t}$,
in the case of $\tan \beta \gg 1$ also for $\tilde f = \tilde{b}$.
\subsection{The Higgs sector of the MSSM}
\label{subsec:higgssector}
The two Higgs doublets form the Higgs potential~\cite{hhg}
\begin{eqnarray}
V &=& (m_1^2 + |\mu|^2) |{\cal H}_1|^2 + (m_2^2 + |\mu|^2) |{\cal H}_2|^2
- m_{12}^2 (\epsilon_{ab} {\cal H}_1^a{\cal H}_2^b + \mbox {h.c.}) \nonumber \\
& & + \frac{1}{8}({g_1}^2+{g_2}^2) \left[ |{\cal H}_1|^2 - |{\cal H}_2|^2 \right]^2
+ \frac{1}{2} {g_2}^2|{\cal H}_1^{\dag} {\cal H}_2|^2~,
\label{higgspot}
\end{eqnarray}
which contains $m_1, m_2, m_{12}$ as SSB parameters.
The doublet fields ${\cal H}_1$ and ${\cal H}_2$ are decomposed in the following way:
\begin{eqnarray}
{\cal H}_1 &=& \left( \begin{array}{c} {\cal H}_1^0 \\[0.5ex] {\cal H}_1^- \end{array} \right) \; = \; \left( \begin{array}{c} v_1
+ \frac{1}{\sqrt2}(\phi_1^0 - i\chi_1^0) \\[0.5ex] -\phi_1^- \end{array} \right)
\nonumber \\
{\cal H}_2 &=& \left( \begin{array}{c} {\cal H}_2^+ \\[0.5ex] {\cal H}_2^0 \end{array} \right) \; = \; \left( \begin{array}{c} \phi_2^+ \\[0.5ex]
v_2 + \frac{1}{\sqrt2}(\phi_2^0 + i\chi_2^0) \end{array} \right)~.
\label{higgsfeldunrot}
\end{eqnarray}
The potential (\ref{higgspot}) can be described with the help of two
independent parameters (besides $g_1$ and $g_2$):
$\tan \beta$ and $M_A^2 = -m_{12}^2(\tan \beta+\cot \beta)$,
where $M_A$ is the mass of the ${\cal CP}$-odd $A$ boson.
The diagonalization of the bilinear part of the Higgs potential,
i.e.\ the Higgs mass matrices, is performed via the orthogonal
transformations
\begin{eqnarray}
\label{hHdiag}
\left( \begin{array}{c} H^0 \\[0.5ex] h^0 \end{array} \right) &=& \left( \begin{array}{cc} \cos \alpha & \sin \alpha \\[0.5ex] -\sin \alpha & \cos \alpha \end{array} \right)
\left( \begin{array}{c} \phi_1^0 \\[0.5ex] \phi_2^0 \end{array} \right) \\
\label{AGdiag}
\left( \begin{array}{c} G^0 \\[0.5ex] A^0 \end{array} \right) &=& \left( \begin{array}{cc} \cos \beta & \sin \beta \\[0.5ex] -\sin \beta & \cos \beta \end{array} \right)
\left( \begin{array}{c} \chi_1^0 \\[0.5ex] \chi_2^0 \end{array} \right) \\
\label{Hpmdiag}
\left( \begin{array}{c} G^{\pm} \\[0.5ex] H^{\pm} \end{array} \right) &=& \left( \begin{array}{cc} \cos \beta & \sin \beta \\[0.5ex] -\sin \beta &
\cos \beta \end{array} \right) \left( \begin{array}{c} \phi_1^{\pm} \\[0.5ex] \phi_2^{\pm} \end{array} \right)~.
\end{eqnarray}
The mixing angle $\alpha$ is determined through
\begin{equation}
\tan 2\alpha = \tan 2\beta \; \frac{M_A^2 + M_Z^2}{M_A^2 - M_Z^2} ;
\qquad -\frac{\pi}{2} < \alpha < 0~.
\label{alphaborn}
\end{equation}
One gets the following Higgs spectrum:
\begin{eqnarray}
\mbox{2 neutral bosons},\, {\cal CP} = +1 &:& h^0, H^0 \nonumber \\
\mbox{1 neutral boson},\, {\cal CP} = -1 &:& A^0 \nonumber \\
\mbox{2 charged bosons} &:& H^+, H^- \nonumber \\
\mbox{3 unphysical Goldstone bosons} &:& G^0, G^+, G^- .
\end{eqnarray}
The masses of the gauge bosons are given in analogy to the SM:
\begin{equation}
M_W^2 = \frac{1}{2} g_2^2 (v_1^2+v_2^2) ;\qquad
M_Z^2 = \frac{1}{2}(g_1^2+g_2^2)(v_1^2+v_2^2) ;\qquad M_\gamma=0.
\end{equation}
\bigskip
At tree level the mass matrix of the neutral ${\cal CP}$-even Higgs bosons
is given in the $\phi_1$-$\phi_2$-basis
in terms of $M_Z$, $M_A$, and $\tan \beta$ by
\begin{eqnarray}
M_{\rm Higgs}^{2, {\rm tree}} &=& \left( \begin{array}{cc} m_{\Pe}^2 & m_{\PePz}^2 \\
m_{\PePz}^2 & m_{\Pz}^2 \end{array} \right) \nonumber\\
&=& \left( \begin{array}{cc} M_A^2 \sin^2\beta + M_Z^2 \cos^2\beta & -(M_A^2 + M_Z^2) \sin \beta \cos \beta \\
-(M_A^2 + M_Z^2) \sin \beta \cos \beta & M_A^2 \cos^2\beta + M_Z^2 \sin^2\beta \end{array} \right),
\label{higgsmassmatrixtree}
\end{eqnarray}
which by diagonalization according to \refeq{hHdiag} yields the
tree-level Higgs boson masses
\begin{equation}
M_{\rm Higgs}^{2, {\rm tree}}
\stackrel{\alpha}{\longrightarrow}
\left( \begin{array}{cc} m_{H,{\rm tree}}^2 & 0 \\ 0 & m_{h,{\rm tree}}^2 \end{array} \right)~.
\end{equation}
The mixing angle $\alpha$ satisfies
\begin{equation}
\tan 2\alpha = \tan 2\beta \frac{M_A^2 + M_Z^2}{M_A^2 - M_Z^2},
\quad - \frac{\pi}{2} < \alpha < 0 .
\label{alpha}
\end{equation}
Since we treat all MSSM parameters as real there is no mixing between
${\cal CP}$-even and ${\cal CP}$-odd Higgs bosons.
The tree-level results for the neutral
${\cal CP}$-even Higgs-boson masses of the MSSM read
\begin{equation}
m^2_{(H, h),{\rm tree}} =
\frac{1}{2} \left[ M_A^2 + M_Z^2
\pm \sqrt{(M_A^2 + M_Z^2)^2 - 4 M_Z^2 M_A^2 \cos^2 2\beta} \right]~.
\label{eq:mhtree}
\end{equation}
This implies an upper bound of $m_{h, {\rm tree}} \leq M_Z$ for the
light ${\cal CP}$-even Higgs-boson mass of the MSSM. The
direct prediction of an upper bound for the mass of the light ${\cal CP}$-even
Higgs-boson mass is one of the most striking phenomenological
predictions of the MSSM. The existence of such a bound, which does not
occur in the case of the SM Higgs boson, can be related to the fact that
the quartic term in the Higgs potential of the MSSM is given in terms of
the gauge couplings, while the quartic coupling is a free parameter in
the SM.
\subsection{Charginos}
\label{subsec:charginos}
The charginos $\tilde{\chi}_i^+\; (i=1,2)$ are four component Dirac
fermions. The mass eigenstates are obtained from the winos
$\tilde{W}^\pm$ and the charged higgsinos $\tilde{H}^-_1$,
$\tilde{H}^+_2$:
\begin{equation}
\tilde{W}^+ = \left( \begin{array}{c} -i \lambda^+ \\[0.5ex] i \bar{\lambda}^- \end{array} \right)
\quad;\quad
\tilde{W}^- = \left( \begin{array}{c} -i \lambda^- \\[0.5ex] i \bar{\lambda}^+ \end{array} \right)
\quad;\quad
\tilde{H}^+_2 = \left( \begin{array}{c} \psi^+_{H_2} \\[0.5ex] \bar{\psi}^-_{H_1} \end{array} \right)
\quad;\quad
\tilde{H}^-_1 = \left( \begin{array}{c} \psi^-_{H_1} \\[0.5ex] \bar{\psi}^+_{H_2} \end{array} \right)~.
\end{equation}
The chargino masses are defined as mass eigenvalues of the
diagonalized mass matrix,
\begin{equation}
{\cal L}_{\tilde{\chi}^+,{\rm mass}} = -\frac{1}{2}\,
\Big( \psi^+,\psi^- \Big) \left( \begin{array}{cc} 0 & {\bf X}^T \\ {\bf X} & 0 \end{array} \right)
\left( \begin{array}{c} \psi^+ \\ \psi^- \end{array} \right) + \mbox {h.c.}~,
\end{equation}
or given in terms of two-component fields
\begin{equation}
\left. \begin{array}{c}
\psi^+ = (-i\lambda^+, \psi^+_{H_2}) \\[1.5ex]
\psi^- = (-i\lambda^-, \psi^-_{H_1})
\end{array}
\right.~,
\end{equation}
where {\bf X} is given by
\begin{equation}
{\bf X} = \left( \begin{array}{cc} M_2 & \sqrt2\, M_W\, \sin \beta \\[1ex] \sqrt2\, M_W\,
\cos \beta & \mu \end{array} \right)~.
\end{equation}
The physical (two-component) mass eigenstates are obtained via
unitary $(2 \times 2)$~matrices {\bf U} and {\bf V}:
\begin{equation}
\left. \begin{array}{c}
\chi_i^+ = V_{ij}\, \psi_j^+ \\[1.5ex]
\chi_i^- = U_{ij}\, \psi_j^-
\end{array}
\right. \qquad i,j=1,2~.
\end{equation}
This results in a four-component Dirac spinor
\begin{equation}
\tilde{\chi}_i^+ = \left( \begin{array}{c} \chi_i^+ \\[0.5ex] \bar{\chi}_i^- \end{array} \right)
\qquad i=1,2~,
\end{equation}
where {\bf U} and {\bf V} are given by
\begin{equation}
{\bf U} = {\bf O}_- \qquad;\qquad
{\bf V} = \left\{\begin{array}{cc} {\bf O}_+ &
\quad\det {\bf X}>0 \\[1ex]
\sigma_3\, {\bf O}_+ & \quad\det {\bf X}<0 \end{array} \right.
\end{equation}
with
\begin{equation}
{\bf O}_\pm = \left( \begin{array}{cc} \cos\phi_\pm & \sin\phi_\pm \\[0.5ex]
-\sin\phi_\pm & \cos\phi_\pm \end{array} \right)~;
\end{equation}
$\cos\phi_\pm$ and $\sin\phi_\pm$ are given by
$(\epsilon = \mbox{sgn} [\det {\bf X}])$
\begin{eqnarray}
\tan\phi_+ = \frac{\sqrt{2}\, M_W(\sin \beta m_{\tilde{\chi}_1^+}
+ \epsilon\, \cos \beta m_{\tilde{\chi}_2^+})}
{(M_2\, m_{\tilde{\chi}_1^+}
+ \epsilon\, \mu\, m_{\tilde{\chi}_2^+} )} \nonumber \\
\tan\phi_- = \frac{-\mu\, m_{\tilde{\chi}_1^+}
- \epsilon\, M_2\, m_{\tilde{\chi}_2^+}}
{\sqrt{2}\, M_W (\sin \beta m_{\tilde{\chi}_1^+}
+ \epsilon\, \cos \beta m_{\tilde{\chi}_2^+})}~.
\end{eqnarray}
(If $\phi_+ < 0$ it has to be replaced by $\phi{_+} + \pi$.)
$m_{\tilde{\chi}_1^+}$ and $m_{\tilde{\chi}_2^+}$ are the eigenvalues
of the diagonalized matrix
\begin{eqnarray}
{\bf M}^2_{{\rm diag},\tilde{\chi}^+} &=&
{\bf V\, X^{\dagger}\, X\, V}^{-1} \; = \;
{\bf U^*\, X\, X^{\dagger}\, (U^*)}^{-1} \nonumber \\
{\bf M}_{{\rm diag},\tilde{\chi}^+} &=&
{\bf U^*\, X\, V}^{-1} \; =\;
\left( \begin{array}{cc} m_{\tilde{\chi}_1^+} & 0 \\[0.5ex] 0 & m_{\tilde{\chi}_2^+} \end{array} \right) .
\end{eqnarray}
They are given by
\begin{eqnarray}
m^2_{\tilde{\chi}_{1,2}^+} &=& \frac{1}{2}\, \bigg\{
M_2^2 + \mu^2 + 2M_W^2 \mp \Big[ (M_2^2-\mu^2)^2 \nonumber \\
& & +\; 4M_W^4\cos^2 2\beta + 4M_W^2(M_2^2+\mu^2+2\,\mu\, M_2\, \sin 2\beta)
\Big]^{\frac{1}{2}} \bigg\}~.
\label{Charmasse}
\end{eqnarray}
\subsection{Neutralinos}
\label{subsec:neutralinos}
Neutralinos $\tilde{\chi}_i^0\; (i=1,2,3,4)$ are four-component
Majorana fermions. They are the mass eigenstates of the
photino,~$\tilde{\gamma}$, the zino,~$\tilde Z$, and the neutral higgsinos,
$\tilde{H}^0_1$ and $\tilde{H}^0_2$, with
\begin{equation}
\tilde{\gamma} = \left( \begin{array}{c} -i \lambda_\gamma \\[0.5ex]
i \bar{\lambda}_\gamma \end{array} \right) \quad;\quad
\tilde{Z} = \left( \begin{array}{c} -i \lambda_Z \\[0.5ex] i \bar{\lambda}_Z \end{array} \right)
\quad;\quad
\tilde{H}^0_1 = \left( \begin{array}{c} \psi^0_{H_1} \\[0.5ex] \bar{\psi}^0_{H_1} \end{array} \right)
\quad;\quad
\tilde{H}^0_2 = \left( \begin{array}{c} \psi^0_{H_2} \\[0.5ex] \bar{\psi}^0_{H_2} \end{array} \right)~.
\end{equation}
Analogously to the SM, the photino and zino are mixed states from the
bino,~$\tilde B$, and the wino,~$\tilde W$,
\begin{equation}
\tilde{B} = \left( \begin{array}{c} -i \lambda^\prime \\[0.5ex] i \bar{\lambda}^\prime \end{array} \right)
\qquad;\qquad
\tilde{W}^3 = \left( \begin{array}{c} -i \lambda^3 \\[0.5ex] i \bar{\lambda}^3 \end{array} \right)~,
\end{equation}
with
\begin{eqnarray}
\tilde{\gamma} &=& \tilde{W}^3\, s_\mathrm{w} + \tilde{B}\, c_\mathrm{w} \nonumber \\
\tilde{Z} &=& \tilde{W}^3\, c_\mathrm{w} - \tilde{B}\, s_\mathrm{w}~.
\end{eqnarray}
The mass term in the Lagrange density is given by
\begin{equation}
{\cal L}_{\tilde{\chi}^0,{\rm mass}} = -\frac{1}{2}(\psi^0)^T\, {\bf Y}\,
\psi^0 + \mbox {h.c.}~,
\end{equation}
with the two-component fermion fields
\begin{equation}
(\psi^0)^T = (-i\lambda^\prime , -i\lambda^3 , \psi_{H_1}^0 ,
\psi_{H_2}^0)~.
\end{equation}
The mass matrix {\bf Y} is given by
\begin{equation}
\renewcommand{\arraystretch}{1.2}
{\bf Y} = \left( \begin{array}{cccc} M_1 & 0 & -M_Zs_\mathrm{w}\cos \beta & M_Zs_\mathrm{w}\sin \beta \\ 0 &
M_2 & M_Zc_\mathrm{w}\cos \beta & -M_Zc_\mathrm{w}\sin \beta \\
-M_Zs_\mathrm{w}\cos \beta & M_Zc_\mathrm{w}\cos \beta & 0 & -\mu \\ M_Zs_\mathrm{w}\sin \beta &
-M_Zc_\mathrm{w}\sin \beta & -\mu & 0 \end{array} \right)~.
\label{Y}
\end{equation}
The physical neutralino mass eigenstates are obtained with the
unitary transformation matrix~{\bf N}:
\begin{equation}
\chi_i^0 = N_{ij}\, \psi_j^0 \qquad i,j=1,\ldots,4,
\end{equation}
resulting in the four-component spinor (representing the mass
eigenstate)
\begin{equation}
\tilde{\chi}_i^0 = \left( \begin{array}{c} \chi_i^0 \\[0.5ex] \bar{\chi}_i^0 \end{array} \right)
\qquad i=1,\ldots,4~.
\end{equation}
The diagonal mass matrix is then given by
\begin{equation}
{\bf M}_{{\rm diag},\tilde{\chi}^0} = {\bf N^*\, Y\, N}^{-1}~.
\end{equation}
\subsection{Gluinos}
\label{subsec:gluinos}
The gluino, $\tilde{g}$, is the spin~1/2 superpartner (Majorana fermion) of
the gluon. According to the 8 generators of $SU(3)_C$ (colour octet),
there are 8 gluinos, all having the same Majorana mass
\begin{equation}
m_{\tilde{g}} = |M_3|~.
\end{equation}
In SUSY GUTs $M_1$, $M_2$ and $M_3$ are not independent but connected
via
\begin{equation}
m_{\tilde{g}} = M_3 = \frac{g_3^2}{g_2^2}\, M_2 \; = \;
\frac{\alpha_s}{\alpha_{\rm em}}\, s_\mathrm{w}^2\, M_2, \;\;
M_1 = \frac{5}{3} \frac{s_\mathrm{w}^2}{c_\mathrm{w}^2}\, M_2~.
\label{G-GUT}
\end{equation}
\subsection{Scalar fermion sector with flavor mixing}
\label{sec:sfermions}
In \refse{Sec:CKM} we saw how quarks are rotated from the EW interaction eigenstate basis to the mass eigenstate basis. Since squarks belong to the same supermultiplet, they need to be rotated parallel to the quarks. The rotation is performed via same matrix i.e. the CKM matrix and the relavent terms in the SSB Lagrangian given in \refeq{softbreaking} get rotated from the interaction eigenstate basis to what is known as the Super-CKM basis
\begin{eqnarray}
\label{eq:lsoft-superCKM}
\mathcal{L}_{\rm soft}
&=& - {\tilde U}_{Ri}^* m_{\tilde U_R ij }^2 \tilde {U}_{Rj}
- {\tilde D}_{Ri}^* m_{\tilde D_R ij}^2 {\tilde D}_{Rj}
- {\tilde U}_{Li}^* m_{\tilde U_L ij}^2 {\tilde U}_{Lj}
- {\tilde D}_{Li}^* m_{\tilde D_L ij}^2 {\tilde D}_{Lj} \nonumber \\
&&- {\tilde U}_{Li} {\cal A}^u_{ij} {\tilde U}^*_{Rj} {\cal H}_2^0
- {\tilde D}_{Li} (V_{\rm CKM})_{ki} {\cal A}^u_{kj} {\tilde U}^*_{Rj}
{\cal H}_2^+
- {\tilde U}_{Li} (V_{\rm CKM}^*)_{ik} {\cal A}^d_{kj} {\tilde D}^*_{Rj}
{\cal H}_1^- \nonumber \\
&&+ {\tilde D}_{Li} {\cal A}^d_{ij} {\tilde D}^*_{Rj} {\cal H}_1^0 + \text{h.c.},
\end{eqnarray}
where ${\tilde U}_{L,R}$ with $U = u, c, t$ represents up-type squarks, ${\tilde D}_{L,R}$ with $D = d, s, b$ represents down-type squarks in Super-CKM basis. The soft masses $m_{\tilde U_L}$, $m_{\tilde U_R}$, $m_{\tilde D_L}$, $m_{\tilde D_R}$ and trilinear couplings ${\cal A}^{q}$ with $q=u,d$ in Super-CKM basis are related to the EW interaction eigenstate basis by
\begin{eqnarray}
{\cal A}^{q} &=& V^{q}_L {\bar A}^q V^{q \dagger}_R, \quad
m_{\tilde U_R}^2 = V_R^u m_{\tilde U}^2 V_R^{u \dagger}, \nonumber \\
m_{\tilde D_R}^2 &=& V_R^d m_{\tilde D}^2 V_R^{d \dagger}, \quad
m_{\tilde U_L}^2 = V_L^u m_{\tilde Q}^2 V_L^{u \dagger}, \nonumber \\
m_{\tilde D_L}^2 &=& V_L^d m_{\tilde Q}^2 V_L^{d \dagger}.
\label{eq:su2}
\end{eqnarray}
In the Super-CKM basis, not only squarks with different flavor can mix among themselves but we will have left-right mixing also. This will results in $6 \times 6$
mass matrices for up-type and down-type squarks. The same arguments hold for the sleptons but in this case flavor mixing will be induced by the PMNS matrix of
the neutrino sector and transmitted by the (tiny) neutrino Yukawa couplings. Thus we will have $6 \times 6$ mass matrix for the charged sleptons in the so called Super-PMNS basis, however for the sneutrinos we have a $3 \times 3$ mass matrix, since within the MSSM even with type~I
seesaw (to be defined below), we have only three EW interaction eigenstates, ${\tilde \nu}_{L}$
with $\nu=\nu_e, \nu_\mu, \nu_\tau$ (right handed neutrinos decouple below their respective mass scale).
The non-diagonal entries in this $6 \times 6$ general matrix for sfermions
can be described in terms of a set of
dimensionless parameters $\delta^{FAB}_{ij}$ ($F=Q,U,D,L,E; A,B=L,R$; $i,j=1,2,3$,
$i \neq j$) where $F$ identifies the sfermion type, $L,R$ refer to the
``left-'' and ``right-handed'' SUSY partners of the corresponding
fermionic degrees of freedom, and $i,j$
indices run over the three generations.
(Non-zero values for the $\delta^{FAB}_{ij}$ are generated via the processes
discussed in the introduction.)
One usually writes the $6\times 6$ non-diagonal mass matrices,
${\mathcal M}_{\tilde u}^2$ and ${\mathcal M}_{\tilde d}^2$ being ordered respectively as $(\tilde{u}_L, \tilde{c}_L,
\tilde{t}_L, \tilde{u}_R, \tilde{c}_R, \tilde{t}_R)$, $(\tilde{d}_L, \tilde{s}_L, \tilde{b}_L, \tilde{d}_R,
\tilde{s}_R, \tilde{b}_R)$ in the Super-CKM basis, ${\mathcal M}_{\tilde l}^2$
being ordered as $(\tilde {e}_L, \tilde{\mu}_L, \tilde{\tau}_L, \tilde {e}_R, \tilde{\mu}_R,
\tilde{\tau}_R)$ in the Super-PMNS basis and write them in terms of left- and right-handed blocks
$M^2_{\tilde q \, AB}$, $M^2_{\tilde l \, AB}$ ($q=u,d$, $A,B=L,R$),
which are non-diagonal $3\times 3$ matrices,
\begin{equation}
{\cal M}_{\tilde q}^2 =\left( \begin{array}{cc}
M^2_{\tilde q \, LL} & M^2_{\tilde q \, LR} \\[.3em]
M_{\tilde q \, LR}^{2 \, \dagger} & M^2_{\tilde q \,RR}
\end{array} \right), \qquad \tilde q= \tilde u, \tilde d~,
\label{eq:blocks-matrix}
\end{equation}
where:
\begin{alignat}{5}
M_{\tilde u \, LL \, ij}^2
= & m_{\tilde U_L \, ij}^2 + \left( m_{u_i}^2
+ (T_3^u-Q_us_\mathrm{w}^2 ) M_Z^2 \cos 2\beta \right) \delta_{ij}, \notag\\
M^2_{\tilde u \, RR \, ij}
= & m_{\tilde U_R \, ij}^2 + \left( m_{u_i}^2
+ Q_us_\mathrm{w}^2 M_Z^2 \cos 2\beta \right) \delta_{ij} \notag, \\
M^2_{\tilde u \, LR \, ij}
= & \left< {\cal H}_2^0 \right> {\cal A}_{ij}^u- m_{u_{i}} \mu \cot \beta \, \delta_{ij},
\notag, \\
M_{\tilde d \, LL \, ij}^2
= & m_{\tilde D_L \, ij}^2 + \left( m_{d_i}^2
+ (T_3^d-Q_d s_\mathrm{w}^2 ) M_Z^2 \cos 2\beta \right) \delta_{ij}, \notag\\
M^2_{\tilde d \, RR \, ij}
= & m_{\tilde D_R \, ij}^2 + \left( m_{d_i}^2
+ Q_ds_\mathrm{w}^2 M_Z^2 \cos 2\beta \right) \delta_{ij} \notag, \\
M^2_{\tilde d \, LR \, ij}
= & \left< {\cal H}_1^0 \right> {\cal A}_{ij}^d- m_{d_{i}} \mu \tan \beta \, \delta_{ij}~,
\label{eq:SCKM-entries}
\end{alignat}
and
\begin{equation}
{\mathcal M}_{\tilde l}^2 =\left( \begin{array}{cc}
M^2_{\tilde l \, LL} & M^2_{\tilde l \, LR} \\[.3em]
M_{\tilde l \, LR}^{2 \, \dagger} & M^2_{\tilde l \,RR}
\end{array} \right),
\label{eq:slep-6x6}
\end{equation}
where:
\begin{alignat}{5}
M_{\tilde l \, LL \, ij}^2
= & m_{\tilde L \, ij}^2 + \left( m_{l_i}^2
+ (-\frac{1}{2} + s_\mathrm{w}^2 ) M_Z^2 \cos 2\beta \right) \delta_{ij}, \notag\\
M^2_{\tilde l \, RR \, ij}
= & m_{\tilde E \, ij}^2 + \left( m_{l_i}^2
-s_\mathrm{w}^2 M_Z^2 \cos 2\beta \right) \delta_{ij} \notag, \\
M^2_{\tilde l \, LR \, ij}
= & \left< {\cal H}_1^0 \right> {\cal A}_{ij}^e- m_{l_{i}} \mu \tan \beta \, \delta_{ij},
\label{eq:slep-matrix}
\end{alignat}
with, $i,j=1,2,3$, $Q_u=2/3$, $Q_d=-1/3$, $T_3^u=1/2$ and
$T_3^d=-1/2$. $(m_{u_1},m_{u_2}, m_{u_3})=(m_u,m_c,m_t)$, $(m_{d_1},m_{d_2},
m_{d_3})=(m_d,m_s,m_b)$ are the quark masses and $(m_{l_1},m_{l_2},
m_{l_3})=(m_e,m_\mu,m_\tau)$ are the lepton masses.
It should be noted that the non-diagonality in flavor comes
exclusively from the SSB parameters, that could be
non-vanishing for $i \neq j$, namely: the masses $m_{\tilde U_L \, ij}^2$,
$m_{\tilde U_R \, ij}^2$, $m_{\tilde D_L \, ij}^2$,
$m_{\tilde D_R \, ij}^2$, $m_{\tilde L \, ij}$, $m_{\tilde E \, ij}$
and the trilinear couplings, ${\cal A}_{ij}^f$.
In the sneutrino sector there is, correspondingly, a one-block $3\times
3$ mass matrix, that is referred to the $(\tilde \nu_{eL}, \tilde \nu_{\mu L},
\tilde \nu_{\tau L})$ Super-PMNS basis:
\begin{equation}
{\mathcal M}_{\tilde \nu}^2 =\left( \begin{array}{c}
M^2_{\tilde \nu \, LL}
\end{array} \right),
\label{eq:sneu-3x3}
\end{equation}
where:
\begin{equation}
M_{\tilde \nu \, LL \, ij}^2
= m_{\tilde L \, ij}^2 + \left(
\frac{1}{2} M_Z^2 \cos 2\beta \right) \delta_{ij},
\label{eq:sneu-matrix}
\end{equation}
It is important to note that due to $SU(2)_{\rm L}$ gauge invariance
the same soft masses $m_{\tilde Q \, ij}$ enter in both up-type and
down-type squarks mass matrices similarly $m_{\tilde L \, ij}$ enter in
both the slepton and sneutrino $LL$ mass matrices.
The SSB parameters for the up-type squarks differ from
corresponding ones for down-type squarks by a rotation with CKM
matrix. The same would hold for sleptons i.e.\ the soft SUSY-breaking
parameters of the sneutrinos would differ from the corresponding ones
for charged sleptons by a rotation with the PMNS matrix. However, taking
the neutrino masses and oscillations
into account in the SM leads to LFV effects that are extremely small.
For instance, in $\mu \to e \gamma$ they are of \order{10^{-47}} in case
of Dirac neutrinos with mass around 1~eV and maximal
mixing~\cite{Kuno:1999jp,DiracNu,MajoranaNu}, and of \order{10^{-40}} in case
of Majorana neutrinos~\cite{Kuno:1999jp,MajoranaNu}. Consequently we do not
expect large effects from the inclusion of neutrino mass effects here
and neglect a rotation with the PMNS matrix.
The sfermion mass matrices in terms of the $\delta^{FAB}_{ij}$ are given as
\begin{equation}
m^2_{\tilde U_L}= \left(\begin{array}{ccc}
m^2_{\tilde Q_{1}} & \delta_{12}^{QLL} m_{\tilde Q_{1}}m_{\tilde Q_{2}} &
\delta_{13}^{QLL} m_{\tilde Q_{1}}m_{\tilde Q_{3}} \\
\delta_{21}^{QLL} m_{\tilde Q_{2}}m_{\tilde Q_{1}} & m^2_{\tilde Q_{2}} &
\delta_{23}^{QLL} m_{\tilde Q_{2}}m_{\tilde Q_{3}}\\
\delta_{31}^{QLL} m_{\tilde Q_{3}}m_{\tilde Q_{1}} &
\delta_{32}^{QLL} m_{\tilde Q_{3}}m_{\tilde Q_{2}}& m^2_{\tilde Q_{3}}
\end{array}\right)~,
\label{mUL}
\end{equation}
\noindent
\begin{equation}
m^2_{\tilde D_L}= V_{\rm CKM}^\dagger \, m^2_{\tilde U_L} \, V_{\rm CKM}~,
\label{mDL}
\end{equation}
\noindent
\begin{equation}
m^2_{\tilde U_R}= \left(\begin{array}{ccc}
m^2_{\tilde U_{1}} & \delta_{12}^{URR} m_{\tilde U_{1}}m_{\tilde U_{2}} &
\delta_{13}^{URR} m_{\tilde U_{1}}m_{\tilde U_{3}}\\
\delta_{{21}}^{URR} m_{\tilde U_{2}}m_{\tilde U_{1}} & m^2_{\tilde U_{2}} &
\delta_{23}^{URR} m_{\tilde U_{2}}m_{\tilde U_{3}}\\
\delta_{{31}}^{URR} m_{\tilde U_{3}} m_{\tilde U_{1}}&
\delta_{{32}}^{URR} m_{\tilde U_{3}}m_{\tilde U_{2}}& m^2_{\tilde U_{3}}
\end{array}\right)~,
\end{equation}
\noindent
\begin{equation}
m^2_{\tilde D_R}= \left(\begin{array}{ccc}
m^2_{\tilde D_{1}} & \delta_{12}^{DRR} m_{\tilde D_{1}}m_{\tilde D_{2}} &
\delta_{13}^{DRR} m_{\tilde D_{1}}m_{\tilde D_{3}}\\
\delta_{{21}}^{DRR} m_{\tilde D_{2}}m_{\tilde D_{1}} & m^2_{\tilde D_{2}} &
\delta_{23}^{DRR} m_{\tilde D_{2}}m_{\tilde D_{3}}\\
\delta_{{31}}^{DRR} m_{\tilde D_{3}} m_{\tilde D_{1}}&
\delta_{{32}}^{DRR} m_{\tilde D_{3}}m_{\tilde D_{2}}& m^2_{\tilde D_{3}}
\end{array}\right)~,
\end{equation}
\noindent
\begin{equation}
v_2 {\cal A}^u =\left(\begin{array}{ccc}
m_u A_u & \delta_{12}^{ULR} m_{\tilde Q_{1}}m_{\tilde U_{2}} &
\delta_{13}^{ULR} m_{\tilde Q_{1}}m_{\tilde U_{3}}\\
\delta_{{21}}^{ULR} m_{\tilde Q_{2}}m_{\tilde U_{1}} &
m_c A_c & \delta_{23}^{ULR} m_{\tilde Q_{2}}m_{\tilde U_{3}}\\
\delta_{{31}}^{ULR} m_{\tilde Q_{3}}m_{\tilde U_{1}} &
\delta_{{32}}^{ULR} m_{\tilde Q_{3}} m_{\tilde U_{2}}& m_t A_t
\end{array}\right)~,
\label{v2Au}
\end{equation}
\noindent
\begin{equation}
v_1 {\cal A}^d =\left(\begin{array}{ccc}
m_d A_d & \delta_{12}^{DLR} m_{\tilde Q_{1}}m_{\tilde D_{2}} &
\delta_{13}^{DLR} m_{\tilde Q_{1}}m_{\tilde D_{3}}\\
\delta_{{21}}^{DLR} m_{\tilde Q_{2}}m_{\tilde D_{1}} & m_s A_s &
\delta_{23}^{DLR} m_{\tilde Q_{2}}m_{\tilde D_{3}}\\
\delta_{{31}}^{DLR} m_{\tilde Q_{3}}m_{\tilde D_{1}} &
\delta_{{32}}^{DLR} m_{\tilde Q_{3}} m_{\tilde D_{2}}& m_b A_b
\end{array}\right)~.
\label{v1Ad}
\end{equation}
\noindent \begin{equation}
m^2_{\tilde L}= \left(\begin{array}{ccc}
m^2_{\tilde L_{1}} & \delta_{12}^{LLL} m_{\tilde L_{1}}m_{\tilde L_{2}} & \delta_{13}^{LLL} m_{\tilde L_{1}}m_{\tilde L_{3}} \\
\delta_{21}^{LLL} m_{\tilde L_{2}}m_{\tilde L_{1}} & m^2_{\tilde L_{2}} & \delta_{23}^{LLL} m_{\tilde L_{2}}m_{\tilde L_{3}}\\
\delta_{31}^{LLL} m_{\tilde L_{3}}m_{\tilde L_{1}} & \delta_{32}^{LLL} m_{\tilde L_{3}}m_{\tilde L_{2}}& m^2_{\tilde L_{3}} \end{array}\right)\end{equation}
\noindent \begin{equation}
v_1 {\cal A}^e =\left(\begin{array}{ccc}
m_e A_e & \delta_{12}^{ELR} m_{\tilde L_{1}}m_{\tilde E_{2}} & \delta_{13}^{ELR} m_{\tilde L_{1}}m_{\tilde E_{3}}\\
\delta_{21}^{ELR} m_{\tilde L_{2}}m_{\tilde E_{1}} & m_\mu A_\mu & \delta_{23}^{ELR} m_{\tilde L_{2}}m_{\tilde E_{3}}\\
\delta_{31}^{ELR} m_{\tilde L_{3}}m_{\tilde E_{1}} & \delta_{32}^{ELR} m_{\tilde L_{3}} m_{\tilde E_{2}}& m_{\tau}A_{\tau}\end{array}\right)\label{v1Al}\end{equation}
\noindent \begin{equation}
m^2_{\tilde E}= \left(\begin{array}{ccc}
m^2_{\tilde E_{1}} & \delta_{12}^{ERR} m_{\tilde E_{1}}m_{\tilde E_{2}} & \delta_{13}^{ERR} m_{\tilde E_{1}}m_{\tilde E_{3}}\\
\delta_{21}^{ERR} m_{\tilde E_{2}}m_{\tilde E_{1}} & m^2_{\tilde E_{2}} & \delta_{23}^{ERR} m_{\tilde E_{2}}m_{\tilde E_{3}}\\
\delta_{31}^{ERR} m_{\tilde E_{3}} m_{\tilde E_{1}}& \delta_{32}^{ERR} m_{\tilde E_{3}}m_{\tilde E_{2}}& m^2_{\tilde E_{3}} \end{array}\right)\end{equation}
In this thesis, for simplicity, we are assuming that all $\delta^{FAB}_{ij}$
parameters are real, therefore, hermiticity of
${\mathcal M}_{\tilde q}^2$, ${\mathcal M}_{\tilde l}^2$ and
${\mathcal M}_{\tilde \nu}^2$ implies $\delta_{ij}^{FAB}= \delta_{ji}^{FBA}$.
The next step is to rotate the squark states from the Super-CKM basis,
${\tilde q}_{L,R}$, to the physical basis.
If we set the order in the Super-CKM basis as above,
$(\tilde{u}_L, \tilde{c}_L, \tilde{t}_L, \tilde{u}_R, \tilde{c}_R, \tilde{t}_R)$ and
$(\tilde{d}_L, \tilde{s}_L, \tilde{b}_L, \tilde{d}_R, \tilde{s}_R, \tilde{b}_R)$,
and in the physical basis as
${\tilde u}_{1,..6}$ and ${\tilde d}_{1,..6}$, respectively, these last
rotations are given by two $6 \times 6$ matrices, $R^{\tilde u}$ and
$R^{\tilde d}$,
\begin{equation}
\left( \begin{array}{c} \tilde u_{1} \\ \tilde u_{2} \\ \tilde u_{3} \\
\tilde u_{4} \\ \tilde u_{5} \\\tilde u_{6} \end{array} \right)
\; = \; R^{\tilde u} \left( \begin{array}{c} \tilde{u}_L \\ \tilde{c}_L \\\tilde{t}_L \\
\tilde{u}_R \\ \tilde{c}_R \\ \tilde{t}_R \end{array} \right) ~,~~~~
\left( \begin{array}{c} \tilde d_{1} \\ \tilde d_{2} \\ \tilde d_{3} \\
\tilde d_{4} \\ \tilde d_{5} \\ \tilde d_{6} \end{array} \right) \; = \; R^{\tilde d} \left( \begin{array}{c} \tilde{d}_L \\ \tilde{s}_L \\ \tilde{b}_L \\
\tilde{d}_R \\ \tilde{s}_R \\ \tilde{b}_R \end{array} \right) ~,
\label{newsquarks}
\end{equation}
yielding the diagonal mass-squared matrices for squarks as follows,
\begin{eqnarray}
{\rm diag}\{m_{\tilde u_1}^2, m_{\tilde u_2}^2,
m_{\tilde u_3}^2, m_{\tilde u_4}^2, m_{\tilde u_5}^2, m_{\tilde u_6}^2
\} & = &
R^{\tilde u} \; {\cal M}_{\tilde u}^2 \;
R^{\tilde u \dagger} ~,\\
{\rm diag}\{m_{\tilde d_1}^2, m_{\tilde d_2}^2,
m_{\tilde d_3}^2, m_{\tilde d_4}^2, m_{\tilde d_5}^2, m_{\tilde d_6}^2
\} & = &
R^{\tilde d} \; {\cal M}_{\tilde d}^2 \;
R^{\tilde d \dagger} ~.
\end{eqnarray}
Similarly we need to rotate the sleptons and sneutrinos from the Super-PMNS basis to the physical mass eigenstate basis,
\begin{equation}
\left( \begin{array}{c} \tilde l_{1} \\ \tilde l_{2} \\ \tilde l_{3} \\
\tilde l_{4} \\ \tilde l_{5} \\\tilde l_{6} \end{array} \right)
\; = \; R^{\tilde l} \left( \begin{array}{c} \tilde {e}_L \\ \tilde{\mu}_L \\\tilde{\tau}_L \\
\tilde {e}_R \\ \tilde{\mu}_R \\ \tilde{\tau}_R \end{array} \right) ~,~~~~
\left( \begin{array}{c} \tilde \nu_{1} \\ \tilde \nu_{2} \\ \tilde \nu_{3} \end{array} \right) \; = \; R^{\tilde \nu} \left( \begin{array}{c} \tilde \nu_{eL} \\ \tilde \nu_{\mu L} \\ \tilde \nu_{\tau L} \end{array} \right) ~,
\label{rotsquarks}
\end{equation}
with $R^{\tilde l}$ and $R^{\tilde \nu}$ being the respective $6\times 6$ and
$3\times 3$ unitary rotating matrices that yield the diagonal
mass-squared matrices as follows,
\begin{eqnarray}
{\rm diag}\{m_{\tilde l_1}^2, m_{\tilde l_2}^2,
m_{\tilde l_3}^2, m_{\tilde l_4}^2, m_{\tilde l_5}^2, m_{\tilde l_6}^2
\} & = &
R^{\tilde l} \; {\cal M}_{\tilde l}^2 \;
R^{\tilde l \dagger} ~,\\
{\rm diag}\{m_{\tilde \nu_1}^2, m_{\tilde \nu_2}^2,
m_{\tilde \nu_3}^2
\} & = &
R^{\tilde \nu} \; {\cal M}_{\tilde \nu}^2 \;
R^{\tilde \nu \dagger} ~.
\end{eqnarray}
\section{Minimal Flavor Violation}
The SM has been very successfully tested by
low-energy flavor observables both from the kaon and $B_d$
sectors. In particular, the two $B$~factories have established that
$B_d$ flavor and ${\cal CP}$-violating processes are well described
by the SM up to an accuracy of the $\sim 10\%$
level~\cite{HFAgroup}.
This immediately implies a tension between the solution
of the hierarchy problem, calling for a New Physics (NP) scale at or below
the TeV scale, and the explanation of the Flavor Physics data
require a multi-TeV NP scale, if the new flavor-violating
couplings are of generic size.
An elegant way to simultaneously solve the above problems
is provided by the MFV
hypothesis~\cite{MFV1,MFV2}, where flavor and ${\cal CP}$-violation in quark sector
is assumed to entirely originate from the CKM matrix, even
in theories beyond the SM. For example in the MSSM
the off-diagonality in the sfermion mass matrix reflects the
misalignment (in flavor space) between fermions and sfermions mass
matrices, that cannot be diagonalized simultaneously.
This misalignment can be produced from various
origins. For instance, off-diagonal sfermion
mass matrix entries can be generated by RGE running. Going from a high energy scale, where no flavor violation is
assumed, down to the EW scale, such entries can be generated due to
presence of non diagonal Yukawa matrices in RGE's.
For instance, in the CMSSM (see \citere{AbdusSalam:2011fc} and
references therein), the RGE effects on non-diagonal sfermion SSB
parameters are affected only by non-diagonal elements on
the Yukawa couplings and the trilinear terms which are taken as
proportional to the Yukawas at the GUT scale.
We choose the following form of the Yukawa matrices (working in the
Super-CKM basis~\cite{slha2}),
\begin{align}
Y^{d} = {\rm diag}(y_{d},y_{s},y_{b}), \quad
Y^{u} = V_{\rm CKM}^{\dag} {\rm diag}(y_{u},y_{c},y_{t})~.
\end{align}
Hence, all flavor violation in the quark and squark sector is generated
by the RGE's and controlled by the CKM matrix, i.e.\ the Yukawa
couplings have a strong impact on the size of the induced off-diagonal
entries in the squark mass matrices.
The situation is somewhat different in the slepton sector where
neutrinos are strictly massless (in the SM and the MSSM).
Consequently, there is no slepton mixing, which would induce
LFV in the charged sector, allowing not yet observed processes like $l_i \to l_j \gamma$
($i > j$; $l_{3,2,1} = \tau, \mu, e$)~\cite{bm}.
However in the neutral sector, we have strong experimental
evidence that shows that the neutrinos are massive and mix among
themselves~\cite{Neutrino-Osc}.
In order to incorporate this, a seesaw mechanism (to be defined below) is
used to generate neutrino masses, and the PMNS matrix plays the role of
the CKM matrix in the lepton sector. Extending the MFV hypothesis for
leptons~\cite{MFVinlepton} we can assume that the flavor mixing in the
lepton and slepton sector is induced and controlled by the seesaw mechanism.
\section{The Constrained MSSM}
Within the CMSSM the SSB
parameters are assumed to be universal at the Grand Unification scale
$M_{\rm GUT} \sim 2 \times 10^{16} \,\, \mathrm{GeV}$,
\begin{eqnarray}
\label{soft}
& (m_Q^2)_{i j} = (m_U^2)_{i j} = (m_D^2)_{i j} = (m_L^2)_{i j}
= (m_E^2)_{i j} = m_0^2\ \delta_{i j}, & \nonumber \\
& m_{H_1}^2 = m_{H_2}^2 = m_0^2, &\\
& m_{\tilde{g}}\ =\ m_{\tilde{W}}\ =\ m_{\tilde{B}}\ =\ m_{1/2}, &
\nonumber \\
& ({\bar A}^u)_{i j}= A_0 e^{i \phi_A} (Y^u)_{i j},\ \ \ ({\bar A}^d)_{i j}= A_0 e^{i \phi_A}
(Y^d)_{i j},\ \ \
({\bar A}^e)_{i j}= A_0e^{i \phi_A} (Y^e)_{i j}. & \nonumber
\end{eqnarray}
There is a common mass (square) for all the scalars, $m_0^2$, a single gaugino
mass, $m_{1/2}$, and all the trilinear SSB terms are directly
proportional to the corresponding Yukawa couplings in the superpotential
with a proportionality constant $A_0 e^{i \phi_A}$, containing a
potential non-trivial complex phase.
With the use of the RGE of the MSSM,
one can obtain the SUSY spectrum at the EW scale.
All the SUSY masses and mixings are then given as
a function of $m_0^2$, $m_{1/2}$, $A_0$, and
$\tan \beta$. We require radiative symmetry breaking to fix $|\mu|$ and
$|B \mu|$ \cite{rge,bertolini} with the tree--level Higgs potential.
By definition, this model fulfills the MFV hypothesis, since the only
flavor violating terms stem from the CKM matrix.
The important point is that, even in a model with universal SSB terms at some high energy scale as the CMSSM, some
off-diagonality in the squark mass matrices appears at the EW scale.
Working in the basis where the squarks are rotated parallel to the
quarks i.e. the Super-CKM basis, the squark mass
matrices are not flavor diagonal at the EW scale.
This is due to the fact that at $M_{\rm GUT}$ there exist two non-trivial
flavor structures, namely the two Yukawa matrices for the up and down quarks,
which are not simultaneously diagonalizable. This implies that
through RGE evolution some flavor mixing leaks into the sfermion mass matrices.
In a general SUSY model the presence of new flavor structures
in the SSB terms would generate large flavor mixing in
the sfermion mass matrices. However, in the CMSSM, the two Yukawa matrices are the only
source of flavor change. As always in the Super-CKM basis, any off-diagonal entry
in the sfermion mass matrices at the EW scale will be necessarily
proportional to a product of Yukawa couplings. This will play a crucial role in the analysis in chapter 6.
\section{Seesaw extensions of the MSSM}
As already mentioned in the introduction, the neutrino masses can be generated through dimension 5 operator.
There are many possible ways to form a dimension-5 gauge singlet term at low energy through the tree-level exchange
of a heavy particle at the high energy:
(i) each $L_L$-$\phi$ pair forms a fermion singlet,
(ii) each of the $L_L$-$L_L$ and $\phi$-$\phi$ pair forms a scalar triplet,
(iii) each $L_L$-$\phi$ pair forms a fermion triplet,
and (iv) each of the $L_L$-$L_L$ and $\phi$-$\phi$ pair forms a scalar singlet.
Case (i) can arise from the tree-level exchange of a right handed fermion singlet
and this corresponds to the Type-I seesaw mechanism
\cite{seesaw:I}.
Case (ii) arises when the heavy particle is a Higgs triplet giving rise to the
Type-II seesaw mechanism \cite{Magg:1980ut,Lazarides:1980nt}.
For case (iii) the exchanged particle
should be a right-handed fermion triplet, which corresponds to
the Type-III seesaw mechanism \cite{Foot-Type-III,Ma:2002pf}.
The last scenario gives terms only of the form $\overline{\nu_L^C} e_L$, which
cannot generate a neutrino mass. We describe Type-I seesaw mechanisms
in \refse{sec:mssmI} in detail.
\subsection{Supersymmetric Type-I seesaw model}
\label{sec:mssmI}
In order to provide an explanation for the (small) neutrino masses, the
MSSM can be extended by the type-I seesaw mechanism~\cite{seesaw:I}.
The superpotential for MSSM-seesaw I can be written as
\begin{eqnarray}
\label{superpotentialSeesaw1}
W_{\rm SI}&=&W_{\rm MSSM}+ Y_{\nu}^{ij}\epsilon_{\alpha \beta} {\hat H}_2^{\alpha} {\hat N}_i^c {\hat L}_j^{\beta}
+ \frac{1}{2} M_{N}^{ij} {\hat N}_i^c {\hat N}_j^c,
\end{eqnarray}
where $W_{\rm MSSM}$ is given in \refeq{superpotential} and ${\hat N}_i^c$
is the additional superfield that contains the three right-handed neutrinos,
$\nu_{Ri}$, and
their scalar partners, $\tilde \nu_{Ri}$. $M_N^{ij}$ denotes the $3\times3$ Majorana mass matrix for heavy right handed neutrino.
The full set of SSB terms is given by,
\begin{eqnarray}
\label{softbreakingSeesaw1}
-{\cal L}_{\rm soft,SI} &=& - {\cal L}_{\rm soft}
+(m_{\tilde \nu}^2)^i_j {\tilde \nu}_{Ri}^* {\tilde \nu}_{R}^j
+ (\frac{1}{2}B_{\nu}^{ij} M_{N}^{ij} {\tilde \nu}_{Ri}^* {\tilde \nu}_{Rj}^*
+A_{\nu}^{ij}h_2 {\tilde \nu}_{Ri}^* {\tilde l}_{Lj}+ {\rm h.c.}),
\end{eqnarray}
with ${\cal L}_{\rm soft}$ given by \refeq{softbreaking},
$(m_{\tilde \nu}^2)^i_j$, $A_{\nu}^{ij}$ and $B_{\nu}^{ij}$
are the new SSB parameters.
By the seesaw mechanism three of the neutral fields acquire heavy masses and
decouple at high energy scale that we will denote as $M_N$, below this scale the effective theory
contains the MSSM plus an operator that provides masses to the neutrinos.
\begin{equation}
W_{\rm EW,SI}=W_{\rm MSSM}+ \frac{1}{2}(Y_{\nu} L H_2)^{T} M_{N}^{-1} (Y_{\nu} L H_2),
\end{equation}
where $W_{\rm EW,SI}$ represent the MSSM seesaw I superpotential at EW scale. This framework naturally explains neutrino oscillations in agreement with
experimental data~\cite{Neutrino-Osc}. At the electroweak scale an
effective Majorana mass matrix for light neutrinos,
\begin{equation}
\label{meff}
m_{\rm eff}=-\frac{1}{2}v_u^2 Y_{\nu}\cdot M_{N}^{-1}\cdot Y^{ T}_{\nu},
\end{equation}
arises from Dirac neutrino Yukawa $Y_{\nu}$ (that can be assumed of
the same order as the charged-lepton and quark Yukawas), and heavy Majorana
masses $M_N$. The smallness of the neutrino masses implies that the scale
$M_N$ is very high, \order{10^{14} \,\, \mathrm{GeV}}.
From \refeqs{superpotentialSeesaw1} and (\ref{softbreakingSeesaw1})
we can observe that
one can choose a basis such that the Yukawa coupling matrix,
$Y^e_{ij}$, and the mass matrix of the right-handed neutrinos, $M_N^{ij}$, are
diagonalized as $Y^e_\delta$ and $M_R^\delta$, respectively. In this case
the neutrino Yukawa couplings $Y_{\nu}^{ij}$ are not generally diagonal,
giving rise to LFV. Here it is important to note that the lepton-flavor
conservation is not a consequence of the SM gauge symmetry, even in the absence
of the right-handed neutrinos.
Consequently, slepton mass terms can violate
the lepton-flavor conservation in a manner consistent with the gauge
symmetry. Thus the scale of LFV can be identified with the
EW scale, much lower than the right-handed neutrino scale $M_N$, leading
to potentially observable rates.
In the SM augmented by
right-handed neutrinos, the flavor violating processes such as
$\mu \to e \gamma$, $\tau \to \mu \gamma$ etc.,
whose rates are proportional to inverse powers of $M_R^\delta$, would be
highly suppressed with such a large $M_N$ scale, and hence are far
beyond current experimental bounds. However, in SUSY theories, the
neutrino Dirac couplings $Y_\nu$
enter in the RGE's of
the SSB sneutrino and slepton masses, generating LFV. In
the basis where the charged-lepton Yukawa couplings matrix $Y^{e}$ is diagonal, the soft
slepton-mass matrix acquires corrections that contain off-diagonal
contributions from the RGE running from $M_{\rm GUT}$ down to the
Majorana mass scale $M_N$,
of the following form (in the leading-log approximation)~\cite{LFVhisano}:
\begin{align}
(m_{\tilde L}^2)_{ij} &\sim \frac 1{16\pi^2} (6m^2_0 + 2A^2_0)
\left({Y_{\nu}}^{\dagger} Y_{\nu}\right)_{ij}
\log \left( \frac{M_{\rm GUT}}{M_N} \right) \, \nonumber\\
(m_{\tilde E}^2)_{ij} &\sim 0 \, \nonumber\\
({\bar A}^e)_{ij} &\sim \frac 3{8\pi^2} {A_0 Y^{e}}_i
\left({Y_{\nu}}^{\dagger} Y_{\nu}\right)_{ij}
\log \left( \frac{M_{\rm GUT}}{M_N} \right) \,
\label{offdiagonal}
\end{align}
Consequently, even if the soft scalar masses were universal at the
unification scale, quantum corrections between the GUT scale and
the seesaw scale $M_N$
would modify this structure via renormalization-group
running, which generates off-diagonal
contributions~\cite{Cannoni:2013gq,gllv,Mismatch,Antusch,EGL,casas-ibarra} at
$M_N$ in a basis such that $Y^{e}$ is diagonal. Below this
scale, the off-diagonal contributions remain almost unchanged.
Therefore the seesaw mechanism induces non trivial values for slepton
$\delta^{FAB}_{ij}$ resulting in a prediction for LFV decays
$l_i \to l_j \gamma$, $(i > j)$ that can be much larger than
the non-SUSY case. These rates depend on the structure of $Y_\nu$ at a seesaw
scale $M_N$ in a basis where $Y^e$ and $M_N$ are diagonal.
By using the approach of \citere{casas-ibarra} a general form of $Y_\nu$
containing all neutrino experimental information can be written as:
\begin{equation}
Y_\nu = \frac{\sqrt{2}} {v_2} \sqrt{M_R^\delta} R \sqrt{m_\nu^\delta} U^\dagger~,
\label{eq:casas}
\end{equation}
where $R$ is a general orthogonal matrix and $m_\nu^\delta$ denotes the
diagonalized neutrino mass matrix. In this basis the matrix~$U$ can be
identified with the $U_{\rm PMNS}$ matrix obtained as:
\begin{equation}
m_\nu^\delta=U^T m_{\rm eff} U~.
\end{equation}
In order to find values for the slepton generation mixing
parameters we need a
specific form of the product $Y_\nu^\dagger Y_\nu$ as shown in
\refeq{offdiagonal}. The
simple consideration of direct hierarchical neutrinos with a common
scale for right handed neutrinos provides a representative reference
value. In this case using \refeq{eq:casas} we find
\begin{equation}
Y_\nu^\dagger Y_\nu= \frac{2}{v_u^2}M_R U m_\nu^\delta U^\dagger~.
\label{eq:ynu2}
\end{equation}
Here $M_R$ is the common mass assigned to the $\nu_R$'s. In the conditions
considered here, LFV effects are independent of the matrix $R$.
For the forthcoming numerical analysis the values of the Yukawa couplings
etc.\ have to be set to yield values in agreement with the
experimental data for neutrino masses and mixings.
In our computation, by considering a normal hierarchy among the neutrino
masses, we fix
$m_{\nu_3} \sim \sqrt{\Delta m^2_{\text{atm}}} \sim 0.05 \,\, \mathrm{eV}$ and
require $m_{\nu_2}/m_{\nu_3}=0.17$,
$m_{\nu_2} \sim 100 \cdot m_{\nu_1}$ consistent with the measured values of
$\Delta m^2_{\text{sol}}$ and $\Delta m^2_{\text{atm}}$~\cite{neu-fits}.
The matrix $U$ is identified with $U_{\rm PMNS}$ with the ${\cal CP}$-phases set to
zero and neutrino mixing angles set to the center of their
experimental values.
One can observe that $m_{\rm eff}$ remains unchanged by consistent
changes on the scales of $M_N$ and $Y_\nu$. This is no longer correct
for the off-diagonal entries in the slepton mass matrices
(parameterized by slepton $\delta^{FAB}_{ij}$). These quantities have quadratic dependence
on $Y_\nu$ and logarithmic dependence on $M_N$, see \refeq{offdiagonal}.
Therefore larger values of $M_N$ imply larger LFV effects.
By setting $M_N=10^{14} \,\, \mathrm{GeV}$, the largest values of
$Y_\nu$ are of about 0.29, this
implies an important restriction on the parameters space arising from the
${\rm BR}(\mu\to e \gamma)$.
An example of models with almost degenerate $\nu_R$ can be found in
\cite{Cannoni:2013gq}. For our numerical analysis we tested several scenarios
and we found that the one defined here is the simplest and also the
one with larger LFV prediction.
\chapter{Precision Observables}
\label{precision-observables}
In this chapter we will present the calculational details and experimetal status of the various low energy observables considered in this thesis.
\section{Higher order corrections to EWPO}
\label{sec:EWPO-calc}
EWPO that are known with an accuracy at the per-mille level or better
have the potential to allow a discrimination between quantum effects of
the SM and SUSY models, see \citere{PomssmRep} for a review. Examples
are the $W$-boson mass $M_W$ and the $Z$-boson observables, such as the
effective leptonic weak missxing angle $\sin^2\theta_{\mathrm{eff}}$.
The $W$-boson mass can be evaluated from
\begin{equation}
M_W^2 \left( 1 - \frac{M_W^2}{M_Z^2} \right) =
\frac{\pi \alpha}{\sqrt{2} G_{\mu}} (1 + \Delta r)
\end{equation}
where $\alpha$ is the fine-structure constant and $G_{\mu}$ the Fermi
constant. This relation arises from comparing the prediction for muon
decay with the experimentally precisely known Fermi constant.
The one-loop contributions to $\Delta r$ can be written as
\begin{equation}
\Delta r = \Delta\alpha - \frac{c_\mathrm{w}^2}{s_\mathrm{w}^2}\Delta\rho + (\Delta r)_{\text{rem}},
\end{equation}
where $\Delta\alpha$ is the shift in the fine-structure constant due to the
light fermions of the SM, $\Delta\alpha \propto \log(M_Z/m_f)$, and $\Delta\rho$
is the leading contribution to the $\rho$ parameter~\cite{rho} from
(certain) fermion and sfermion loops (see below). The remainder part
$(\Delta r)_{\text{rem}}$ contains in particular the contributions from the
Higgs sector.
The effective leptonic weak mixing angle at the $Z$-boson resonance,
$\sin^2\theta_{\mathrm{eff}}$, is defined through the vector and axial-vector couplings
($g_{\text{V}}^\ell$ and $g_{\text{A}}^\ell$) of leptons ($\ell$) to the
$Z$~boson, measured at the $Z$-boson pole. If this vertex is written as
$i\bar \ell\gamma^\mu (g_{\text{V}}^\ell - g_{\text{A}}^\ell \gamma_5) \ell
Z_\mu$ then
\begin{equation}
\sin^2\theta_{\mathrm{eff}} = \frac 14 \left( 1 -
\mathop{\mathrm{Re}}\frac{g_{\text{V}}^\ell}{g_{\text{A}}^\ell}\right)\,.
\end{equation}
Loop corrections enter through higher-order contributions to
$g_{\text{V}}^\ell$ and $g_{\text{A}}^\ell$.
Both of these (pseudo-)observables are affected by shifts in the
quantity $\Delta\rho$ according to
\begin{equation}
\label{eq:precobs}
\DeM_W \approx \frac{M_W}{2}\frac{c_\mathrm{w}^2}{c_\mathrm{w}^2 - s_\mathrm{w}^2} \Delta\rho\,, \quad
\Delta\sin^2\theta_{\mathrm{eff}} \approx - \frac{c_\mathrm{w}^2 s_\mathrm{w}^2}{c_\mathrm{w}^2 - s_\mathrm{w}^2} \Delta\rho\,.
\end{equation}
The quantity $\Delta\rho$ is defined by the relation
\begin{equation}
\Delta\rho = \frac{\Sigma_Z^{\text{T}}(0)}{M_Z^2} -
\frac{\Sigma_W^{\text{T}}(0)}{M_W^2}
\label{eq:drho}
\end{equation}
with the unrenormalized transverse parts of the $Z$- and $W$-boson
self-energies at zero momentum, $\Sigma_{Z,W}^{\text{T}}(0)$. It
represents the leading universal corrections to the EWPO
induced by mass splitting between partners in
isospin doublets~\cite{rho}. Consequently, it is sensitive to the
mass-splitting effects induced by flavor mixing.
Within the SM the corrections to $\Delta\rho$ stem from the
splitting in one $SU(2)$ doublet. Due to the mixing of various scalar
fermion states the picture is slightly more involved in the MSSM.
In MSSM without flavor violation the well known results for the third
generation squark contribution to $\Delta\rho$ (without flavor mixing) can
be written as
\begin{eqnarray}
\Delta\rho = \frac{3G_{\mu}}{8 \sqrt{2}\pi^2} \big[
-\sin^2 \theta_{\tilde{t}} \cos^2 \theta_{\tilde{t}} F_{0}(m_{\tilde{t}_1}^2,m_{\tilde{t}_2}^2)
-\sin^2 \theta_{\tilde{b}} \cos^2 \theta_{\tilde{b}} F_{0}(m_{\tilde{b}_1}^2,m_{\tilde{b}_2}^2)
\nonumber \\
+\cos^2 \theta_{\tilde{t}} \cos^2 \theta_{\tilde{b}} F_{0}(m_{\tilde{t}_1}^2,m_{\tilde{b}_1}^2)
+\sin^2 \theta_{\tilde{b}} \cos^2 \theta_{\tilde{t}} F_{0}(m_{\tilde{t}_1}^2,m_{\tilde{b}_2}^2)
\nonumber \\
+\sin^2 \theta_{\tilde{t}} \cos^2 \theta_{\tilde{b}} F_{0}(m_{\tilde{t}_2}^2,m_{\tilde{b}_1}^2)
+\sin^2 \theta_{\tilde{t}} \sin^2 \theta_{\tilde{b}} F_{0}(m_{\tilde{t}_2}^2,m_{\tilde{b}_2}^2) \big]
\label{drhoMSSM}
\end{eqnarray}
with
\begin{align}
F_{0}(m^2_1, m^2_2) &=
m^2_1+m^2_2-\frac{2m^2_1 m^2_2}{m^2_1-m^2_2} \ln\left(\frac{m^2_1}{m^2_2}\right)~.
\label{Fun:f0}
\end{align}
In the absence of intergenerational mixing there are only $2 \times 2$
mixing matrices to be taken into account, here parametrized by $\theta_{\tilde{t}}$
($\theta_{\tilde{b}}$) in the scalar top (bottom) case.
Here one can see that squarks do not need to be the
$SU(2)$ partners to give contribution to $\Delta\rho$. In particular the
first two terms of \refeq{drhoMSSM} describe contributions from the
same type (up type or down type) of scalar quarks.
Going from this simple case to the one with generation mixing one
finds contribution from all three generations, including two $6 \times 6$
mixing matrices (which are difficult to analyze analytically).
The two gauge boson self-energies are then
given by (see also \citere{delrhoNMFV}),
\begin{align*}
\Sigma_{ZZ}(0) & =\frac{e^{2}}{288\pi^{2}s_\mathrm{w}^{2}c_\mathrm{w}^{2}}(-%
{\displaystyle\sum\limits_{s,t=1}^{6}}
{\displaystyle\sum\limits_{i,j=1}^{3}}
2[\frac{1}{8}F_{0}(m_{\tilde{u}_{s}}^{2},m_{\tilde{u}_{t}}^{2})+\frac{1}%
{4}(A_0^{\rm fin}(m_{\tilde{u}_{s}}^{2})+A_0^{\rm fin}(m_{\tilde{u}_{t}}^{2}))]\\
& \{3R_{t,j}^{\tilde{u}}R_{t,j}^{\tilde{u}^{\ast}}-4s_\mathrm{w}^{2}(R_{t,j}%
^{\tilde{u}}R_{t,j}^{\tilde{u}^{\ast}}+R_{t,3+j}^{\tilde{u}}R_{t,3+j}%
^{\tilde{u}^{\ast}})\}\\
& \{3R_{s,i}^{\tilde{u}}R_{s,i}^{\tilde{u}^{\ast}}-4s_\mathrm{w}^{2}(R_{s,i}%
^{\tilde{u}}R_{s,i}^{\tilde{u}^{\ast}}+R_{s,3+i}^{\tilde{u}}R_{s,3+i}%
^{\tilde{u}^{\ast}})\}\\
& -%
{\displaystyle\sum\limits_{s,t=1}^{6}}
{\displaystyle\sum\limits_{i,j=1}^{3}}
2[\frac{1}{8}F_{0}(m_{\tilde{d}_{s}}^{2},m_{\tilde{d}_{t}}^{2})+\frac{1}%
{4}(A_0^{\rm fin}(m_{\tilde{d}_{s}}^{2})+A_0^{\rm fin}(m_{\tilde{d}_{t}}^{2}))]\\
& \{3R_{t,j}^{\tilde{d}}R_{t,j}^{\tilde{d}^{\ast}}-2s_\mathrm{w}^{2}(R_{t,j}%
^{\tilde{d}}R_{t,j}^{\tilde{d}^{\ast}}+R_{t,3+j}^{\tilde{d}}R_{t,3+j}%
^{\tilde{d}^{\ast}})\}\\
& \{3R_{s,i}^{\tilde{u}}R_{s,i}^{\tilde{u}^{\ast}}-2s_\mathrm{w}^{2}(R_{s,i}%
^{\tilde{d}}R_{s,i}^{\tilde{d}^{\ast}}+R_{s,3+i}^{\tilde{d}}R_{s,3+i}%
^{\tilde{d}^{\ast}}\}\\
& +%
{\displaystyle\sum\limits_{s=1}^{6}}
{\displaystyle\sum\limits_{i=1}^{3}}
A_0^{\rm fin}(m_{\tilde{u}_{s}}^{2})[(3-4s_\mathrm{w}^{2})^{2}R_{s,i}^{\tilde{u}}%
R_{s,i}^{\tilde{u}^{\ast}}+16s_\mathrm{w}^{4}R_{s,3+i}^{\tilde{u}}R_{s,3+i}%
^{\tilde{u}^{\ast}}]\\
& +%
{\displaystyle\sum\limits_{s=1}^{6}}
{\displaystyle\sum\limits_{i=1}^{3}}
A_0^{\rm fin}(m_{\tilde{d}_{s}}^{2})[(3-2s_\mathrm{w}^{2})^{2}R_{s,i}^{\tilde{d}}%
R_{s,i}^{\tilde{d}^{\ast}}+4s_\mathrm{w}^{4}R_{s,3+i}^{\tilde{d}}R_{s,3+i}^{\tilde
{d}^{\ast}}])
\end{align*}
\begin{align*}
\Sigma_{WW}(0) & =\frac{e^{2}}{32\pi^{2}s_\mathrm{w}^{2}}(-%
{\displaystyle\sum\limits_{s,t=1}^{6}}
{\displaystyle\sum\limits_{i,j=1}^{3}}
4[\frac{1}{8}F_{0}(m_{\tilde{u}_{s}}^{2},m_{\tilde{d}_{t}}^{2})+\frac{1}%
{4}(A_0^{\rm fin}(m_{\tilde{u}_{s}}^{2})+A_0^{\rm fin}(m_{\tilde{d}_{t}}^{2}))]\\
& R_{s,i}^{\tilde{u}}R_{t,j}^{\tilde{d}}R_{s,j}^{\tilde{u}^{\ast}}%
R_{t,i}^{\tilde{d}^{\ast}}\\
& +%
{\displaystyle\sum\limits_{s=1}^{6}}
{\displaystyle\sum\limits_{i=1}^{3}}
A_0^{\rm fin}(m_{\tilde{u}_{s}}^{2})R_{s,i}^{\tilde{u}}R_{s,i}^{\tilde{u}^{\ast}}+%
{\displaystyle\sum\limits_{s=1}^{6}}
{\displaystyle\sum\limits_{i=1}^{3}}
A_0^{\rm fin}(m_{\tilde{d}_{s}}^{2})R_{s,i}^{\tilde{d}}R_{s,i}^{\tilde{d}^{\ast}}%
\end{align*}
Here
$R^{\tilde{u}}$ and $R^{\tilde{d}}$ are the $6\times6$ rotation matrices for
the up and down-type squarks respectively, see \refeq{newsquarks}.
The finite part of the one point integral function is given by
\begin{align}
A_{0}^{\rm fin}(m^{2})=m^{2}(1-\log\frac{m^{2}}{\mu^{2}})~.
\end{align}
Here it is important to note that
the corrections will come, as in \refeq{drhoMSSM}, from states
connected via $SU(2)$ as well as from ``same flavor'' contributions
stemming from the $Z$~boson self-energy, see \refeq{eq:drho}.
Larger splitting between ``same flavor'' states due to the
intergenerational mixing thus leads to the expectation of increasing
contributions to $\Delta\rho$ from flavor violation effects.
The effects from flavor violation in the squark
entering via $\Delta\rho$ were already evaluated in Ref.
\cite{delrhoNMFV} and included in {\tt FeynHiggs}.
We have calculated the effects of slepton flavor mixing to $\Delta\rho$ via {\tt FeynArts}/{\tt FormCalc}\ setup and added the results to {\tt FeynHiggs}\ for our numerical evaluation. The details about the changes made to {\tt FeynArts}, {\tt FormCalc}\ and {\tt FeynHiggs}\ will be discussed in \refse{sec:feynhiggs}.
In \reffi{FeynDiagZSelf} and \reffi{FeynDiagWSelf}, we show the generic Feynman diagrams for $Z$ and $W$ boson self energies that enter in the calculation of $\Delta\rho$.
\begin{figure}[htb!]
\begin{center}
\unitlength=1.0bp%
\begin{feynartspicture}(432,190)(4,2)
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/Sine}{0}
\FALabel(3.,8.93)[t]{$Z$}
\FAProp(20.,10.)(14.,10.)(0.,){/Sine}{0}
\FALabel(17.,11.07)[b]{$Z$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde u_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde u_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/Sine}{0}
\FALabel(3.,8.93)[t]{$Z$}
\FAProp(20.,10.)(14.,10.)(0.,){/Sine}{0}
\FALabel(17.,11.07)[b]{$Z$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde d_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde d_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/Sine}{0}
\FALabel(3.,8.93)[t]{$Z$}
\FAProp(20.,10.)(14.,10.)(0.,){/Sine}{0}
\FALabel(17.,11.07)[b]{$Z$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde l_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde l_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/Sine}{0}
\FALabel(3.,8.93)[t]{$Z$}
\FAProp(20.,10.)(14.,10.)(0.,){/Sine}{0}
\FALabel(17.,11.07)[b]{$Z$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde \nu_{j}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde \nu_{i}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$Z$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$Z$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde u_{s}$}
\FAVert(10.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$Z$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$Z$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde d_{s}$}
\FAVert(10.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$Z$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$Z$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde l_{s}$}
\FAVert(10.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$Z$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$Z$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde \nu_{i}$}
\FAVert(10.,10.){0}
\end{feynartspicture}
\end{center}
\caption[Generic Feynman diagrams for $Z$ boson self-energies]{Generic Feynman diagrams for $Z$~boson self-energies
containing squarks and sleptons in loops.
$\tilde u_{s,t}$,$\tilde d_{s,t}$ and $\tilde l_{s,t}$ denote the six mass eigenstates of up-type, down-type and charged sleptons respectively.
$\tilde \nu_{i,j}$ are the three sneutrinos states $\tilde \nu_{e}$,
$\tilde \nu_{\mu}$ and $\tilde \nu_{\tau}$.}
\label{FeynDiagZSelf}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\unitlength=1.0bp%
\begin{feynartspicture}(432,190)(4,2)
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/Sine}{0}
\FALabel(3.,8.93)[t]{$W$}
\FAProp(20.,10.)(14.,10.)(0.,){/Sine}{0}
\FALabel(17.,11.07)[b]{$W$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde u_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde d_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/Sine}{0}
\FALabel(3.,8.93)[t]{$W$}
\FAProp(20.,10.)(14.,10.)(0.,){/Sine}{0}
\FALabel(17.,11.07)[b]{$W$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde l_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde \nu_{i}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$W$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$W$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde u_{s}$}
\FAVert(10.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$W$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$W$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde d_{s}$}
\FAVert(10.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$W$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$W$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde l_{s}$}
\FAVert(10.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(5.,8.93)[t]{$W$}
\FAProp(20.,10.)(10.,10.)(0.,){/Sine}{0}
\FALabel(15.,8.93)[t]{$W$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde \nu_{i}$}
\FAVert(10.,10.){0}
\end{feynartspicture}
\end{center}
\caption[Generic Feynman diagrams for $W$ boson self-energies]{Generic Feynman diagrams for $W$ boson self-energies
containing squarks and sleptons in loops.
$\tilde u_{s,t}$,$\tilde d_{s,t}$ and $\tilde l_{s,t}$ denote the six mass eigenstates of up-type, down-type and charged sleptons respectively.
$\tilde \nu_{i,j}$ are the three sneutrinos states $\tilde \nu_{e}$,
$\tilde \nu_{\mu}$ and $\tilde \nu_{\tau}$.}
\label{FeynDiagWSelf}
\end{figure}
The present experimental uncertainties for the EWPO are \cite{LEPEWWG}
\begin{equation}
\label{EWPO-today}
\deM_W^{\text{exp,today}} \sim 15 \,\, \mathrm{MeV}, \quad
\delta\sin^2\theta_{\mathrm{eff}}^{\text{exp,today}} \sim 15 \times 10^{-5}\,,
\end{equation}
which will further be reduced~\cite{Baak:2013fwa} to
\begin{equation}
\label{EWPO-future}
\deM_W^{\text{exp,future}} \sim 4\,\, \mathrm{MeV}, \quad
\delta\sin^2\theta_{\mathrm{eff}}^{\text{exp,future}} \sim 1.3 \times 10^{-5}\,,
\end{equation}
at the ILC and at the GigaZ option of the ILC, respectively.
Even higher precision could be expected from the FCC-ee, see, e.g.,
\citere{fcc-ee-paris}.
The prediction of $M_W$ also suffers from various kinds of
theoretical uncertainties, parametric and intrinsic. Starting with the
parametric uncertainties, an
experimental error of $1 \,\, \mathrm{GeV}$ on $m_t$ yields a parametric uncertainty on
$M_W$ of about $6 \,\, \mathrm{MeV}$, while the parametric uncertainties induced by the
current experimental error of the hadronic contribution to the shift in
the fine-structure constant, $\Delta\alpha_{\rm had}$, and by the
experimental error of $M_Z$ amount to about $2 \,\, \mathrm{MeV}$ and $2.5 \,\, \mathrm{MeV}$,
respectively. The uncertainty of the $M_W$ prediction caused by the
experimental uncertainty of the Higgs mass
$\delta M_h^{\rm exp} \lsim 0.3 \,\, \mathrm{GeV}$ is
signifcantly smaller ($\approx 0.2 \,\, \mathrm{MeV}$). The intrinsic
uncertainties from unknown higher-order corrections in the case of no
flavor mixing have been estimated to be around (4.7-9.4)~MeV in the
MSSM \cite{Heinemeyer:2006px,Haestier:2005ja} depending on the SUSY mass
scale. For our forthcoming numerical analysis, we have added the parameteric uncertanities in quadrature and add
the result linearly to the uncertanity from
the unknown higher order corrections in the case of no flavor mixing.
We assume an additional 10\% uncertanity from the flavor mixing
contribution to $\Delta\rho^{\rm MSSM}$ and (via \refeq{eq:precobs})
add it linearly to the other uncertainties.
\section{Higher-order corrections in the Higgs sector}
\label{sec:higgs-ho}
In order to calculate one-loop corrections to the Higgs boson
masses, the renormalized Higgs boson
self-energies are needed. Here we follow the procedure used in
\citeres{mhcMSSMlong,mhiggsf1lC} (and references therein). The parameters appearing in the Higgs
potential, see \refeq{higgspot}, are renormalized as follows:
\begin{align}
\label{rMSSM:PhysParamRenorm}
M_Z^2 &\to M_Z^2 + \delta \MZ^2, & T_h &\to T_h +
\delta T_h, \\
M_W^2 &\to M_W^2 + \delta \MW^2, & T_H &\to T_H +
\delta T_H, \notag \\
M_{\rm Higgs}^2 &\to M_{\rm Higgs}^2 + \delta M_{\rm Higgs}^2, &
\tan \beta\, &\to \tan \beta\, (1+\delta\!\tan\!\beta\,). \notag
\end{align}
$M_{\rm Higgs}^2$ denotes the tree-level Higgs boson mass matrix given
in \refeq{higgsmassmatrixtree}. $T_h$ and $T_H$ are the tree-level
tadpoles, i.e.\ the terms linear in $h$ and $H$ in the Higgs potential.
The field renormalization matrices of both Higgs multiplets
can be set up symmetrically,
\begin{align}
\label{rMSSM:higgsfeldren}
\begin{pmatrix} h \\[.5em] H \end{pmatrix} \to
\begin{pmatrix}
1+\tfrac{1}{2} \dZ{hh} & \tfrac{1}{2} \dZ{hH} \\[.5em]
\tfrac{1}{2} \dZ{hH} & 1+\tfrac{1}{2} \dZ{HH}
\end{pmatrix} \cdot
\begin{pmatrix} h \\[.5em] H \end{pmatrix}~.
\end{align}
\noindent
For the mass counter term matrices we use the definitions
\begin{align}
\delta M_{\rm Higgs}^2 =
\begin{pmatrix}
\delta m_h^2 & \delta m_{hH}^2 \\[.5em]
\delta m_{hH}^2 & \delta m_H^2
\end{pmatrix}~.
\end{align}
The renormalized self-energies, $\hat{\Sigma}(p^2)$, can now be expressed
through the unrenormalized self-energies, $\Sigma(p^2)$, the field
renormalization constants and the mass counter terms.
This reads for the ${\cal CP}$-even part,
\begin{subequations}
\label{rMSSM:renses_higgssector}
\begin{align}
\ser{hh}(p^2) &= \se{hh}(p^2) + \dZ{hh} (p^2-m_{h,{\rm tree}}^2) - \delta m_h^2, \\
\ser{hH}(p^2) &= \se{hH}(p^2) + \dZ{hH}
(p^2-\tfrac{1}{2}(m_{h,{\rm tree}}^2+m_{H,{\rm tree}}^2)) - \delta m_{hH}^2, \\
\ser{HH}(p^2) &= \se{HH}(p^2) + \dZ{HH} (p^2-m_{H,{\rm tree}}^2) - \delta m_H^2~.
\end{align}
\end{subequations}
Inserting the renormalization transformation into the Higgs mass terms
leads to expressions for their counter terms which consequently depend
on the other counter terms introduced in~(\ref{rMSSM:PhysParamRenorm}).
For the ${\cal CP}$-even part of the Higgs sectors, these counter terms are:
\begin{subequations}
\label{rMSSM:HiggsMassenCTs}
\begin{align}
\delta m_h^2 &= \deM_A^2 \cos^2(\alpha-\beta) + \delta M_Z^2 \sin^2(\alpha+\beta) \\
&\quad + \tfrac{e}{2 M_Z s_\mathrm{w} c_\mathrm{w}} (\delta T_H \cos(\alpha-\beta)
\sin^2(\alpha-\beta) + \delta T_h \sin(\alpha-\beta)
(1+\cos^2(\alpha-\beta))) \notag \\
&\quad + \delta\!\tan\!\beta\, \sin \beta\, \cos \beta\, (M_A^2 \sin 2 (\alpha-\beta) + M_Z^2 \sin
2 (\alpha+\beta)), \notag \\
\delta m_{hH}^2 &= \tfrac{1}{2} (\deM_A^2 \sin 2(\alpha-\beta) - \delta \MZ^2 \sin
2(\alpha+\beta)) \\
&\quad + \tfrac{e}{2 M_Z s_\mathrm{w} c_\mathrm{w}} (\delta T_H \sin^3(\alpha-\beta) -
\delta T_h \cos^3(\alpha-\beta)) \notag \\
&\quad - \delta\!\tan\!\beta\, \sin \beta\, \cos \beta\, (M_A^2 \cos 2 (\alpha-\beta) + M_Z^2 \cos
2 (\alpha+\beta)), \notag \\
\delta m_H^2 &= \deM_A^2 \sin^2(\alpha-\beta) + \delta \MZ^2 \cos^2(\alpha+\beta) \\
&\quad - \tfrac{e}{2 M_Z s_\mathrm{w} c_\mathrm{w}} (\delta T_H \cos(\alpha-\beta)
(1+\sin^2(\alpha-\beta)) + \delta T_h \sin(\alpha-\beta)
\cos^2(\alpha-\beta)) \notag \\
&\quad - \delta\!\tan\!\beta\, \sin \beta\, \cos \beta\, (M_A^2 \sin 2 (\alpha-\beta) + M_Z^2 \sin
2 (\alpha+\beta))~. \notag
\end{align}
\end{subequations}
For the field renormalization we chose to give each Higgs doublet one
renormalization constant,
\begin{align}
\label{rMSSM:HiggsDublettFeldren}
{\cal H}_1 \to (1 + \tfrac{1}{2} \dZ{{\cal H}_1}) {\cal H}_1, \quad
{\cal H}_2 \to (1 + \tfrac{1}{2} \dZ{{\cal H}_2}) {\cal H}_2~.
\end{align}
This leads to the following expressions for the various field
renormalization constants in \refeq{rMSSM:higgsfeldren}:
\begin{subequations}
\label{rMSSM:FeldrenI_H1H2}
\begin{align}
\dZ{hh} &= \sin^2 \!\alpha\, \dZ{{\cal H}_1} + \cos^2 \!\alpha\, \dZ{{\cal H}_2}, \\[.2em]
\dZ{hH} &= \sin \alpha\, \cos \alpha\, (\dZ{{\cal H}_2} - \dZ{{\cal H}_1}), \\[.2em]
\dZ{HH} &= \cos^2 \!\alpha\, \dZ{{\cal H}_1} + \sin^2 \!\alpha\, \dZ{{\cal H}_2}~.
\end{align}
\end{subequations}
The counter term for $\tan \beta$ can be expressed in terms of the vacuum
expectation values as
\begin{equation}
\delta\tan \beta = \frac{1}{2} \left( \dZ{{\cal H}_2} - \dZ{{\cal H}_1} \right) +
\frac{\delta v_2}{v_2} - \frac{\delta v_1}{v_1}~,
\end{equation}
where the $\delta v_i$ are the renormalization constants of the $v_i$:
\begin{equation}
v_1 \to \left( 1 + \dZ{{\cal H}_1} \right) \left( v_1 + \delta v_1 \right), \quad
v_2 \to \left( 1 + \dZ{{\cal H}_2} \right) \left( v_2 + \delta v_2 \right)~.
\end{equation}
It can be shown that the divergent parts of $\delta v_1/v_1$ and
$\delta v_2/v_2$ are equal~\cite{mhiggsf1l,mhiggsf1lC}.
Consequently, one can set $\delta v_2/v_2 - \delta v_1/v_1$ to zero.
Similarly for the charged Higgs sector, the renormalized self-energy is expressed in terms of the unrenormalized one and the corresponding counter-terms as:
\noindent \begin{equation}
\hat{\Sigma}_{H^{-}H^{+}}\left(p^{2}\right)=\Sigma_{H^{-}H^{+}}\left(p^{2}\right)+\delta Z_{H^{-}H^{+}}\left(p^{2}-m^{2}_{H^{\pm},{\rm tree}} \right)-\delta m_{H^{\pm}}^{2},\end{equation}
where,
\noindent \begin{equation}
\delta m_{H^{\pm}}^{2}=\delta M_{A}^{2}+\delta M_{W}^{2}\end{equation}
and,
\noindent \begin{equation}
\delta Z_{H^{-}H^{+}}=\sin^{2}\beta \, \,\dZ{{\cal H}_1}
+\cos^{2}\beta \,\,\dZ{{\cal H}_2}. \end{equation}
The renormalization conditions are fixed by an appropriate
renormalization scheme. For the mass counter terms on-shell conditions
are used, resulting in:
\begin{align}
\label{rMSSM:mass_osdefinition}
\delta \MZ^2 = \mathop{\mathrm{Re}} \se{ZZ}(M_Z^2), \quad \delta \MW^2 = \mathop{\mathrm{Re}} \se{WW}(M_W^2),
\quad \deM_A^2 = \mathop{\mathrm{Re}} \se{AA}(M_A^2).
\end{align}
For the gauge bosons $\Sigma$ denotes the transverse part of the self-energy.
Since the tadpole coefficients are chosen to vanish in all orders,
their counter terms follow from $T_{\{h,H\}} + \delta T_{\{h,H\}} = 0$:
\begin{align}
\delta T_h = -{T_h}, \quad \delta T_H = -{T_H}~.
\end{align}
For the remaining renormalization constants for $\delta\tan \beta$, $\dZ{{\cal H}_1}$
and $\dZ{{\cal H}_2}$ the most convenient
choice is a $\overline{\rm{DR}}$\ renormalization of $\delta\tan \beta$, $\dZ{{\cal H}_1}$
and $\dZ{{\cal H}_2}$,
\begin{subequations}
\label{rMSSM:deltaZHiggsTB}
\begin{align}
\dZ{{\cal H}_1} &= \dZ{{\cal H}_1}^{\overline{\rm{DR}}}
\; = \; - \left[ \mathop{\mathrm{Re}} \Sigma'_{HH \; |\alpha = 0} \right]^{\rm div}, \\[.5em]
\dZ{{\cal H}_2} &= \dZ{{\cal H}_2}^{\overline{\rm{DR}}}
\; = \; - \left[ \mathop{\mathrm{Re}} \Sigma'_{hh \; |\alpha = 0} \right]^{\rm div}, \\[.5em]
\delta\!\tan\!\beta\, &= \frac{1}{2} (\dZ{{\cal H}_2} - \dZ{{\cal H}_1}) = \delta\!\tan\!\beta\,^{\overline{\rm{DR}}}~.
\end{align}
\end{subequations}
The corresponding renormalization scale, $\mu_{\DRbar}$, is set to
$\mu_{\DRbar} = m_t$ in all numerical evaluations.
Finally, in the
Feynman diagrammatic (FD) approach that we are following here, the higher-order corrected
${\cal CP}$-even Higgs boson masses are derived by finding the
poles of the $(h,H)$-propagator
matrix. The inverse of this matrix is given by
\begin{equation}
\left(\Delta_{\rm Higgs}\right)^{-1}
= - i \left( \begin{array}{cc} p^2 - m_{H,{\rm tree}}^2 + \hat{\Sigma}_{HH}(p^2) & \hat{\Sigma}_{hH}(p^2) \\
\hat{\Sigma}_{hH}(p^2) & p^2 - m_{h,{\rm tree}}^2 + \hat{\Sigma}_{hh}(p^2) \end{array} \right)~.
\label{higgsmassmatrixnondiag}
\end{equation}
Determining the poles of the matrix $\Delta_{\rm Higgs}$ in
\refeq{higgsmassmatrixnondiag} is equivalent to solving
the equation
\begin{equation}
\left[p^2 - m_{h,{\rm tree}}^2 + \hat{\Sigma}_{hh}(p^2) \right]
\left[p^2 - m_{H,{\rm tree}}^2 + \hat{\Sigma}_{HH}(p^2) \right] -
\left[\hat{\Sigma}_{hH}(p^2)\right]^2 = 0\,.
\label{eq:proppole}
\end{equation}
Similarly, in the case of the charged Higgs sector, the corrected Higgs mass is derived by the position of the pole in the charged Higgs propagator, which is defined by:
\noindent \begin{equation}
p^{2}-m^{2}_{H^{\pm},{\rm tree}} +
\hat{\Sigma}_{H^{-}H^{+}}\left(p^{2}\right)=0.
\label{eq:proppolech}
\end{equation}
The present experimental uncertanity at the LHC for the mass of
light neutral higgs $M_h$ is
$\leq 300 \,\, \mathrm{MeV}$~\cite{ATLAS:2013mma,CMS:yva}. This can possibly be reduced by
about 50\% at the LHC and below the level of $\sim 50 \,\, \mathrm{MeV}$ at the
ILC~\cite{dbd}.
Similarly, for the mass of heavy neutral higgs $M_H$
and charged higgs boson $M_{H^\pm}$ an uncertainity at the $1\%$ level could
be expected at the LHC~\cite{cmsHiggs}.
This sets the goal for the theoretical uncertainty, which should be
reduced to the same (or higher) level of accuracy.
Higher order corrections to the masses and mixing angles of the Higgs bosons in the MSSM have already been calculated in the literature. For the light and heavy ${\cal CP}$-even Higgs boson masses, complete one-loop contributions exist~\cite{ERZ,mhiggsf1lA,mhiggsf1lB,mhiggsf1lC}. Almost all the dominant contributions at two-loop level are also known~\cite{mhiggsletter,mhiggslong,mhiggslle,mhiggsFD2,bse,mhiggsEP0,mhiggsEP1,mhiggsEP1b,mhiggsEP2,mhiggsEP3,mhiggsEP3b,mhiggsEP4,mhiggsEP4b,mhiggsRG1,mhiggsRG1a}. For example, with the assumption of vanishing external momenta, the \order{\alpha_t\alpha_s} contributions have been calculated in the Feynman diagrammatic (FD) approach and effective potential (EP) approach and the \order{\alpha_t^2}, \order{\alpha_b\alpha_s}, \order{\alpha_t\alpha_b} and \order{\alpha_b^2} contributions are calculated in the EP approach. The momentum dependence at two-loop level was evaluated in \citeres{mhiggs2lp2, Mh-p2-BH4,Borowka:2015ura,Mh-p2-DDVS} and in ~\citere{mhiggsEP5}, a nearly full two-loop calculation in EP approach that also include the leading three-loop contributions has been presented. The code {\tt H3m}~\cite{mhiggsFD3l} adds the leading three-loop corrections of \order{\alpha_t\alpha_s^2} to the {\tt FeynHiggs}\ results. In the very recent work~\cite{Mh-logresum}, a combination of full one-loop results supplimented with leading and subleading two-loop contributions and a resummation of the leading and subleading logarithmic contributions from scalar-top sector is presented. This combination reduce the theoretical uncertainty from about 3 $\,\, \mathrm{GeV}$ to about 2 $\,\, \mathrm{GeV}$, for scalar-top masses at or below the $\,\, \mathrm{TeV}$ scale, for the light ${\cal CP}$-even Higgs boson mass.
Flavor violation effects for the case of squarks in MI approach were calculated in \cite{arana,arana-NMFV2}. We have calculated the effects of slepton mixing to the Higgs boson masses in {\tt FeynArts}/{\tt FormCalc}\ setup and added the result to the {\tt FeynHiggs}. The details about the changes in {\tt FeynArts}, {\tt FormCalc}\ and {\tt FeynHiggs}\ can be found in \refse{sec:feynhiggs}. We also calculate the effects of squark mixing in MFV CMSSM and MFV CMSSM-seesaw~I.
In \reffi{FeynDiagHSelf}, we show generic Feynman diagrams for Higg self energy while Feynman diagrams for tadpoles are shown in \reffi{FeynDiagHTad}.
\begin{figure}[htb!]
\begin{center}
\unitlength=1.0bp%
\begin{feynartspicture}(432,280)(4,3)
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/Straight}{-1}
\FALabel(10.,5.73)[t]{$u$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/Straight}{1}
\FALabel(10.,14.27)[b]{$u$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/Straight}{-1}
\FALabel(10.,5.73)[t]{$d$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/Straight}{1}
\FALabel(10.,14.27)[b]{$d$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/Straight}{-1}
\FALabel(10.,5.73)[t]{$u$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/Straight}{1}
\FALabel(10.,14.27)[b]{$d$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/Straight}{-1}
\FALabel(10.,5.73)[t]{$l$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/Straight}{1}
\FALabel(10.,14.27)[b]{$l$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde u_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde u_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde d_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde d_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde u_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde d_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde l_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde l_{s}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde l_{t}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde \nu_{i}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(6.,10.)(0.,){/ScalarDash}{0}
\FALabel(3.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(14.,10.)(0.,){/ScalarDash}{0}
\FALabel(17.,11.07)[b]{$\phi$}
\FAProp(6.,10.)(14.,10.)(0.8,){/ScalarDash}{-1}
\FALabel(10.,5.73)[t]{$\tilde \nu_{j}$}
\FAProp(6.,10.)(14.,10.)(-0.8,){/ScalarDash}{1}
\FALabel(10.,14.27)[b]{$\tilde \nu_{i}$}
\FAVert(6.,10.){0}
\FAVert(14.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(10.,10.)(0.,){/ScalarDash}{0}
\FALabel(15.,8.93)[t]{$\phi$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde u_{s}, \tilde d_{s}$}
\FAVert(10.,10.){0}
\FADiagram{}
\FAProp(0.,10.)(10.,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(20.,10.)(10.,10.)(0.,){/ScalarDash}{0}
\FALabel(15.,8.93)[t]{$\phi$}
\FAProp(10.,10.)(10.,10.)(10.,15.5){/ScalarDash}{-1}
\FALabel(10.,16.57)[b]{$\tilde l_{s}, \tilde \nu_{i}$}
\FAVert(10.,10.){0}
\end{feynartspicture}
\end{center}
\caption[Generic Feynman diagrams for the Higgs boson self-energies]{
Generic Feynman diagrams for the Higgs boson self-energies. $\phi$ denotes any of the Higgs bosons, $h$, $H$, $A$ or
$H^\pm$; $u$ stand for $u,c,t$; $d$ stand for $d,s,b$; $l$ stand for $e,\mu,\tau$; $\tilde u_{s,t}$, $\tilde d_{s,t}$ and $\tilde l_{s,t}$ are the six mass
eigenstates of up-type, down-type squarks and charged sleptons respectively and $\tilde \nu_{i,j}$ are the three
sneutrinos states $\tilde \nu_{e}$, $\tilde \nu_{\mu}$ and
$\tilde \nu_{\tau}$.}
\label{FeynDiagHSelf}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\unitlength=1.0bp%
\begin{feynartspicture}(432,190)(4,2)
\FADiagram{}
\FAProp(0.,10.)(7.5,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(7.5,10.)(7.5,10.)(14.,10.){/Straight}{-1}
\FALabel(11.,16.)[]{$u_{i}$}
\FAVert(7.5,10.){0}
\FADiagram{}
\FAProp(0.,10.)(7.5,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(7.5,10.)(7.5,10.)(14.,10.){/Straight}{-1}
\FALabel(11.,16.)[]{$d_{i}$}
\FAVert(7.5,10.){0}
\FADiagram{}
\FAProp(0.,10.)(7.5,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(7.5,10.)(7.5,10.)(14.,10.){/Straight}{-1}
\FALabel(11.,16.)[]{$l_{i}$}
\FAVert(7.5,10.){0}
\FADiagram{}
\FAProp(0.,10.)(7.5,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(7.5,10.)(7.5,10.)(14.,10.){/ScalarDash}{-1}
\FALabel(11.,16.)[]{$\tilde u_{s}$}
\FAVert(7.5,10.){0}
\FADiagram{}
\FAProp(0.,10.)(7.5,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(7.5,10.)(7.5,10.)(14.,10.){/ScalarDash}{-1}
\FALabel(11.,16.)[]{$\tilde d_{s}$}
\FAVert(7.5,10.){0}
\FADiagram{}
\FAProp(0.,10.)(7.5,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(7.5,10.)(7.5,10.)(14.,10.){/ScalarDash}{-1}
\FALabel(11.,16.)[]{$\tilde l_{s}$}
\FAVert(7.5,10.){0}
\FADiagram{}
\FAProp(0.,10.)(7.5,10.)(0.,){/ScalarDash}{0}
\FALabel(5.,8.93)[t]{$\phi$}
\FAProp(7.5,10.)(7.5,10.)(14.,10.){/ScalarDash}{-1}
\FALabel(11.,16.)[]{$\tilde \nu_{i}$}
\FAVert(7.5,10.){0}
\end{feynartspicture}
\end{center}
\caption[Generic Feynman diagrams for the Higgs boson tadpoles]{
Generic Feynman diagrams for the Higgs boson tadpoles.
$\phi$ denotes any of the Higgs bosons, $h$ or $H$; $u$ stand for $u,c,t$; $d$ stand for $d,s,b$; $l$ stand for $e,\mu,\tau$; $\tilde u_{s,t}$, $\tilde d_{s,t}$ and $\tilde l_{s,t}$ are the six mass
eigenstates of up-type, down-type squarks and charged sleptons respectively and $\tilde \nu_{i,j}$ are the three
sneutrinos states $\tilde \nu_{e}$, $\tilde \nu_{\mu}$ and
$\tilde \nu_{\tau}$.}
\label{FeynDiagHTad}
\end{figure}
\section{\boldmath{$B$}-physics observables}
\label{sec:bpo}
In this thesis, we also consider several $B$-physics observables (BPO):
\ensuremath{{\rm BR}(B \to X_s \gamma)}, \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ and \ensuremath{\Delta M_{B_s}}.
Concerning \ensuremath{{\rm BR}(B \to X_s \gamma)}\ included in the calculation are the most
relevant loop contributions to the Wilson coefficients:
(i)~loops with Higgs bosons (including the resummation of large $\tan \beta$
effects~\cite{Isidori:2002qe}),
(ii)~loops with charginos and
(iii)~loops with gluinos.
For \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ there are three types of relevant one-loop corrections
contributing to the relevant Wilson coefficients:
(i)~Box diagrams,
(ii)~$Z$-penguin diagrams and
(iii)~neutral Higgs boson $\phi$-penguin diagrams, where $\phi$ denotes the
three neutral MSSM Higgs bosons, $\phi = h, H, A$ (again large resummed
$\tan \beta$ effects have been taken into account).
In our numerical evaluation there are included
what are known to be the dominant contributions to
these three types of diagrams \cite{Chankowski:2000ng}: chargino
contributions to box and $Z$-penguin diagrams and chargino and gluino
contributions to $\phi$-penguin diagrams.
Concerning \ensuremath{\Delta M_{B_s}}, in the MSSM
there are in general three types of one-loop diagrams that contribute:
(i)~Box diagrams,
(ii)~$Z$-penguin diagrams and
(iii)~double Higgs-penguin diagrams (again including the resummation of
large $\tan \beta$ enhanced effects).
In our numerical evaluation there are included again what are known to be the
dominant contributions to these three types of diagrams in scenarios
with non-minimal flavor violation (for a review see, for instance,
\cite{Foster:2005wb}): gluino contributions to box
diagrams, chargino contributions to box and $Z$-penguin diagrams, and
chargino and gluino contributions to double $\phi$-penguin diagrams.
More details about the calculations employed can be
found in \citeres{arana,arana-NMFV2}.
We perform our numerical calculation with the {\tt BPHYSICS} subroutine
taken from the {\tt SuFla} code~\cite{sufla} (with some additions and
improvements as detailed in \citeres{arana,arana-NMFV2}), which has
been implemented as a subroutine into (a private version of) {\tt FeynHiggs}.
The experimental values used in the numerical analysis\footnote{
Using the most up-to-date value of
$\ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)} = 2.9\pm 0.7 \times 10^{-9}$~\cite{bmm-CMS-LHCb}
would have had a minor impact on our analysis.
}
and SM prediction of these observables
is given in the
\refta{tab:ExpStatus-BPO}~\cite{hfag:rad,Misiak:2009nr,Chatrchyan:2013bka,Aaij:2013aka,Buras:2012ru,hfag:pdg,Buras:1990fn,Golowich:2011cx}.
\begin{table}[htb!]
\begin{center}
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|}
\hline
Observable & Experimental Value & SM Prediction \\\hline
\ensuremath{{\rm BR}(B \to X_s \gamma)} & $3.43\pm 0.22 \times 10^{-4}$ & $3.15\pm 0.23 \times 10^{-4}$ \\
\ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)} & $(3.0)^{+1.0}_{-0.9} \times 10^{-9}$ & $3.23\pm 0.27 \times 10^{-9}$ \\
\ensuremath{\Delta M_{B_s}} & $116.4\pm 0.5 \times 10^{-10} \,\, \mathrm{MeV} $
& $(117.1)^{+17.2}_{-16.4} \times 10^{-10} \,\, \mathrm{MeV}$ \\
\hline
\end{tabular}
\caption[Experimental values of BPO with their SM prediction.]
{Experimental values (used in our numerical analysis) of $B$-physics observables with
their SM prediction.}
\label{tab:ExpStatus-BPO}
\renewcommand{\arraystretch}{1.0}
\end{center}
\end{table}
\section{\boldmath{\ensuremath{h \to \bar b s + b \bar s}}}
\label{sec:hbs-calc}
In SM the branching ratio \ensuremath{\br(\hbs)}\ can be
at most of \order{10^{-7}}~\cite{HdecNMFV}, too small to have a
chance of detection at the LHC. But because of the strong FCNC gluino
couplings and the $\tan \beta$-enhancement inherent to the MSSM Yukawa
couplings, we may expect several orders of magnitude increase of the
branching ratio as compared to the SM result, see
\citere{HdecNMFV, SUSY-QCD}~. This decay in the framework of the MSSM has been analyzed in the
literature: the SUSY-QCD contributions for this decay were calculated
in~\cite{HdecNMFV,SUSY-QCD}, and the SUSY-EW contributions using the mass
insertion approximation were calculated in~\cite{Demir}. Later
in~\cite{SUSY-EW} the SUSY-EW contributions and their interference
effects with the SUSY-QCD contribution were calculated using exact
diagonalization of the squark mass matrices. In all these
analysis, only LL mixing in the
squarks mass matrix was considered, and experimental constraints were
imposed only from \ensuremath{{\rm BR}(B \to X_s \gamma)}. Most recently in~\cite{SUSY-EW-RR} also RR
mixing has been included.
However mixing of the LR or RL elements
of the mass matrix and constraints from other BPO or potential other constraints were not taken into account
(except in the most recent analysis in~\cite{SUSY-EW-RR}).
We (re-)calculate full one-loop
contributions from SUSY-QCD as well as SUSY-EW loops with the help of
the {\tt FeynArts}~\cite{feynarts,famssm} and {\tt FormCalc}~\cite{formcalc} packages.
The lengthy analytical results are not shown here. We take into account the experimental constraints
not only from BPO but also from the EWPO. In the scalar quark sector we not only
consider the LL mixing, but also include the LR-RL and RR mixing for
our analysis of \ensuremath{\br(\hbs)}. For our numerical analysis we define
\begin{align}
\ensuremath{\br(\hbs)} = \frac{\Gamma(\ensuremath{h \to \bar b s + b \bar s})}{\Gamma_{h, {\rm tot}}^{\rm MSSM}}
\end{align}
where $\Gamma_{h, {\rm tot}}^{\rm MSSM}$ is the total decay width of the
light Higgs boson $h$ of the MSSM, as evaluated with
{\tt FeynHiggs}~\cite{feynhiggs,mhiggslong,mhiggsAEC,mhcMSSMlong,Mh-logresum}.
The contributing Feynman diagrams for the decay \ensuremath{h \to \bar b s + b \bar s}\ are shown in
\reffi{QFVHD-QCD}-\ref{QFVHD-Rem}.
Which BR might be detectable at the LHC or an $e^+e^-$ collider such as
the ILC can only be established by means of specific experimental
analyses, which, to our knowledge, do not exist yet. However, in the
literature it is expected to measure BR's at the level of $10^{-3}$ at
the LHC~\cite{HdecNMFV}.
In the clean ILC environment in general Higgs boson branching ratios
below the level of $10^{-4}$ can be observed, see
e.g.\ \citere{ILCreview} for a recent review. We will take this as a
rough guideline down to which level the decay \ensuremath{h \to \bar b s + b \bar s}\ could be observable.
Feynman diagram for SUSY-EW contributions to the decay $\ensuremath{h \to \bar b s + b \bar s}$ are shown in \reffi{QFVHD-Rem} and \reffi{QFVHD-EW}, and SUSY-QCD contributions are shown in \reffi{QFVHD-QCD}.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Diag/SUSYrem.eps}
\end{center}
\caption[Feynman diagrams for the decay $h\rightarrow b\bar{s}+\bar{b}s$]{Feynman diagrams showing SUSY-EW contributions (except neutralino-chargino) to the decay process $h\rightarrow b\bar{s}+\bar{b}s$.}
\label{QFVHD-Rem}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Diag/SUSYEW.eps}
\vspace{0.2cm}
\end{center}
\caption[Feynman diagrams for the decay $h\rightarrow b\bar{s}+\bar{b}s$]{Feynman diagrams showing neutralino-chargino contributions to the decay process $h\rightarrow b\bar{s}+\bar{b}s$.}
\label{QFVHD-EW}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Diag/SUSYQCD.eps}
\end{center}
\caption[Feynman diagrams for the decay $h\rightarrow b\bar{s}+\bar{b}s$]{Feynman diagrams showing SUSY-QCD contributions to the decay process $h\rightarrow b\bar{s}+\bar{b}s$.}
\label{QFVHD-QCD}
\end{figure}
\section{\boldmath{$l_i\rightarrow l_j \gamma$}}
Neutrino oscillation experiments\cite{Neutrino-Osc} have established the existence of lepton
flavor violation. So, as a natural consequence of neutrino
oscillations, one would expect flavour mixing in the charged lepton sector as well. This
mixing can be manifested in rare decay processes such as $\mu\to e\gamma$, $\tau\to e\gamma$, and
$\tau\to \mu\gamma$. However, if only the lepton Yukawa couplings carry this information on
flavour mixing, as in the SM with massive neutrinos, the expected
rates of these processes are extremely tiny\cite{Kuno:1999jp,DiracNu,MajoranaNu} being proportional to the ratio of
masses of neutrinos over the masses of the $W$ bosons. These values are very
far from the present experimental upper bounds \cite{Adam:2013mnn,Aubert:2009ag} that can be
read from \refta{tab:ExpStatus-LFV}.
The situation in the MSSM (extended by the seesaw mechanism) is completely different. Here lepton-slepton misallignment (generated by the presence of seesaw parameters in the RGE's) can dominate the SM contribution by several orders of magnitude. Thus making the study of rare LFV processes very attractive.
We analyze these processes in the framework of CMSSM (extended by Type I seesaw mechanism). MSSM contributions to these decays originate from lepton-slepton-neutralino and lepton-slepton-chargino couplings. The predictions for ${\rm BR}(l_i \to l_j \gamma)$
are obtained with {\tt SPheno 3.2.4}. We checked that the use of this code produces results similar to the ones obtained by our private codes used
in \citere{Cannoni:2013gq}.
\begin{table}[h!]
\begin{center}
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|}
\hline
Observable & Experimental value \\\hline
${\rm BR}(\mu \rightarrow e \gamma)$ & $ < 5.7 \times 10^{-13}$ \\
${\rm BR}(\tau \rightarrow e \gamma)$ & $ < 3.3 \times 10^{-8}$ \\
${\rm BR}(\tau \rightarrow \mu \gamma)$ & $ < 4.4 \times 10^{-8}$ \\
\hline
\end{tabular}
\caption{Present experimental status of LFV processes; their SM
prediction is zero.}
\label{tab:ExpStatus-LFV}
\renewcommand{\arraystretch}{1.0}
\end{center}
\end{table}
\clearpage
\section{\boldmath $h \rightarrow l_i^{\pm} l_j^{\mp}$}
\label{sec:LFVHD-calc}
Since the discovery of Higgs boson, special effort has been made to determine its properties. The motivation
for such an effort resides on understanding the mechanism for EWSB. At present, several
aspects of the Higgs boson are to some extent well known, in particular those related with some of its expected
“standard” decay modes, namely: $WW^{*}$, $ZZ^{*}$, $\gamma \gamma$, $b \bar{b}$ and $\tau \bar{\tau}$ . Currently, measurements of these decay modes have
shown compatibility with the SM expectations, although with large associated uncertainties \cite{CMS:2014ega}.
Indeed, it is due to these large uncertainties that there is still room for nonstandard decay properties, something
that has encouraged such searches at the LHC as well. Searches for invisible Higgs decays have been published in
\cite{CMSAad2014,CMSChatrchyan2014}. Recently CMS collaboration using the 2012 dataset taken at $\sqrt{s} = 8 \rm TeV$ with an integrated luminosity of 19.7 $\rm fb^{-1}$, has found a 2.5 $\sigma$ excess in the $h\rightarrow \mu \tau$ channel, which translates into ${\rm BR}(h\rightarrow \mu \tau) \approx 0.89^{+40}_{-37} \%$ \cite{CMSLFVHD}. However there is no statistically significant excess in the ATLAS results\cite{ATLAS-LFVHD}.
One needs to find the theoretical framework which can accomodate larger rates for LFVHD to explain CMS excess while still respecting the upper bounds on cLFV's. Efforts in such direction have been done in different contexts, with pioneer works in Refs. \cite{Pilaftsis1992,Diaz2000}. More recenty, Ref. \cite{Herrero2013} studied the problem in the MSSM, while \cite{Arhrib2013} in the R-parity violating MSSM. These decays have been considered as well in the inverse seesaw model in \cite{Herrero2014}. Possible effects due to vectorlike leptons have been investigated in \cite{Falkowski2014}. Extended scalar sectors involving several Higgs doublets and flavor symmetries (Yukawa textures) have been examined too \cite{Bhattacharyya2011,Bhattacharyya2012, Arroyo2013, Campos2014}. Finally, the Type-III Two Higgs Doublet Model has been considered in Refs. \cite{Davidson2010,Kopp2014}. Basically, the bottom line of these analyses is that unless one deals with extra Higgs doublets, LFVHD are below the LHC reach.
In this thesis we calculate the LFVHD in SUSY using FD approach.
We study the lepton-slepton misalignment effects to LFVHD, both in the MI approach and in MFV CMSSM-seesaw~I. We do not use mass
insertion approximation and exact diagonalization of the
slepton mass matrix is performed. Feynman diagrams entering our calculation are shown in \reffi{Diag_LFVHD} where first two rows correspond to the decay $h \rightarrow e^{\pm} \mu^{\mp}$, middle two rows correspond to $h \rightarrow e^{\pm} \tau^{\mp}$ and last two rows correspond to $h \rightarrow \mu^{\pm} \tau^{\mp}$.
For the analytical calculation we used {\tt FeynArts}/{\tt FormCalc}\ setup. For this purpose, we implimented LFV Feynman rules for the MSSM in these packages (see \refse{sec:feynhiggs} for details).
For numerical analysis we define the branching ratios of LFVHD as
\begin{equation}
{\rm BR}(h \rightarrow l_i^{\pm} l_j^{\mp})= \frac{\Gamma(h \rightarrow l_i^{\pm} l_j^{\mp})}{\Gamma(h \rightarrow l_i^{\pm} l_j^{\mp})+\Gamma_h^{\rm MSSM}}
\end{equation}
where $i,j=e, \mu, \tau$ and $\Gamma_h^{\rm MSSM}$ is total decay width of ${\cal CP}$-even light Higgs boson $h$.
\begin{figure}[htb!]
\begin{center}
\psfig{file=Diag/HEMUE.eps}
\vspace{1.0cm}
\psfig{file=Diag/HETAU.eps}
\vspace{0.7cm}
\psfig{file=Diag/HMUETAU.eps}
\end{center}
\caption{Feynman diagrams for LFV decays $h \rightarrow l_i^{\pm} l_j^{\mp}$.}
\label{Diag_LFVHD}
\end{figure}
\section{Changes in {\tt FeynArts}, {\tt FormCalc} and {\tt FeynHiggs}}
\label{sec:feynhiggs}
{\tt FeynArts}\cite{feynarts} and {\tt FormCalc}\cite{formcalc} provide
a high level of automation for perturbative calculations up to one loop. This is
particularly important for models with a large particle content such as
the MSSM \cite{famssm}. Here we briefly describe the recent extension
of the implementation of the MSSM in these packages to include LFV.
Details on the previous inclusion of NMFV can be found in
\citeres{feynarts,interplay}. This involves firstly the modification of
the slepton couplings in the existing {\tt FeynArts}\ model file for the MSSM
and secondly the corresponding initialization routines for the slepton
masses and mixings, i.e.\ the $6\times 6$ and $3\times3$ diagonalization
of the mass matrices in {\tt FormCalc}.
\subsection{{\tt FeynArts} Model File}
{\tt FeynArts}' add-on model file \texttt{FV.mod} applies algebraic
substitutions to the Feynman rules of \texttt{MSSM.mod} to upgrade
minimal to non-minimal flavor mixing in the sfermion sector. The
original version modified only the squark sector, i.e.\ NMFV, and needed
to be generalized to include LFV. We solved this by allowing the user
to choose which sfermion types to introduce non-minimal mixing for
through the variable \texttt{\$FV} (set before model initialization, of
course). For example,
\begin{verbatim}
$FV = {11, 12, 13, 14};
InsertFields[..., Model -> {MSSM, FV}]
\end{verbatim}
sets non-minimal mixing for all four sfermion types, with 11 = $\tilde \nu$,
12 = $\tilde l$, 13 = $\tilde u$, and 14 = $\tilde d$ as usual in \texttt{MSSM.mod}.
For compatibility with the old NMFV-only version, the default is
\verb|$FV = {13, 14}|.
\texttt{FV.mod} introduces the following new quantities:
\begin{center}
\begin{tabular}{ll}
\texttt{UASf[$s_1$,$s_2$,$t$]} &
the slepton mixing matrix $R$, where \\
& $s_1, s_2 = 1\dots 6$, \\
& $t = 1\,(\tilde \nu\/), 2\,(\tilde l\/), 3\,(\tilde u\/), 4\,(\tilde d\/)$, \\[1ex]
\texttt{MASf[$s$,$t$]} &
the slepton masses, where \\
& $s = 1\dots 6$, \\
& $t = 1\,(\tilde \nu\/), 2\,(\tilde l\/), 3\,(\tilde u\/), 4\,(\tilde d\/)$.
\end{tabular}
\end{center}
Entries $4\dots 6$ are unused for the sneutrino.
\subsection{Model initialization in {\tt FormCalc}}
The initialization of the generalized slepton-mixing parameters
\texttt{MASf} and \texttt{UASf} is already built into {\tt FormCalc}'s regular
MSSM model-initialization file \texttt{model\_mssm.F} but not turned on
by default. It must be enabled by adjusting the \texttt{FV}
preprocessor flag in \texttt{run.F}:
\begin{verbatim}
#define FV 2
\end{verbatim}
where 2 is the lowest sfermion type $t$ for which flavor violation is
enabled, i.e.\ $\tilde l$.
The flavor-violating parameters $\delta^{FAB}_{ij}$ are represented in {\tt FormCalc}\
by the \texttt{deltaSf} matrix:
\begin{center}
\begin{tabular}{ll}
\texttt{double complex deltaSf($s_1$,$s_2$,$t$)} &
the matrix $(\delta_t)_{s_1s_2}$, where \\
& $s_1, s_2 = 1\dots 6$ ($1\dots 3$ for $\tilde \nu$), \\
& $t = 2\,(\tilde l\/), 3\,(\tilde u\/), 4\,(\tilde d\/)$.
\end{tabular}
\end{center}
Since $\delta$ is an Hermitian matrix, only the entries above the
diagonal are considered. The $\delta^{FAB}_{ij}$ are located at the following
places in the matrix $\delta$:
$$
\left(\begin{array}{ccc|ccc}
\vspc~\cdot~ & \delta^{LLL}_{12} & \delta^{LLL}_{13} &
\vspc~\cdot~ & \delta^{ELR}_{12} & \delta^{ELR}_{13} \\
\vspc~\cdot~ & \vspc~\cdot~ & \delta^{LLL}_{23} &
\delta^{ERL*}_{12} & \vspc~\cdot~ & \delta^{ELR}_{23} \\
\vspc~\cdot~ & \vspc~\cdot~ & \vspc~\cdot~ &
\delta^{ERL*}_{13} & \delta^{ERL*}_{23} & \vspc~\cdot~ \\ \hline
\vspc~\cdot~ & \vspc~\cdot~ & \vspc~\cdot~ &
\vspc~\cdot~ & \delta^{ERR}_{12} & \delta^{ERR}_{13} \\
\vspc~\cdot~ & \vspc~\cdot~ & \vspc~\cdot~ &
\vspc~\cdot~ & \vspc~\cdot~ & \delta^{ERR}_{23} \\
\vspc~\cdot~ & \vspc~\cdot~ & \vspc~\cdot~ &
\vspc~\cdot~ & \vspc~\cdot~ & \vspc~\cdot~
\end{array}\right)
$$
The trilinear couplings $A_f$ acquire non-zero off-diagonal entries in
the presence of LFV through the relations
\begin{equation}
m_{f,i} (A_f)_{ij} = (M_{\tilde f,LR}^2)_{ij}\,,
\quad
i, j = 1\dots 3\, ,
\end{equation}
see \refeq{eq:slep-matrix}. These off-diagonal trilinear couplings (and
hence the $\delta$'s) appear directly in the Higgs--slepton--slepton
couplings, whereas all other effects are mediated through the masses
and mixings.
The described changes are contained in the {\tt FeynArts}\ 3.9 and {\tt FormCalc} 8.4
packages which are publicly available from {\tt www.feynarts.de.}
\subsection{Inclusion of LFV into {\tt FeynHiggs}}
As discussed above, the new corrections to the (renormalized)
Higgs-boson self-energies (and thus to the Higgs-boson masses), as well
as to $\Delta\rho$ (and thus to $M_W$ and $\sin^2\theta_{\mathrm{eff}}$) have been included in
{\tt FeynHiggs}~\cite{feynhiggs,mhiggslong,mhiggsAEC,mhcMSSMlong,Mh-logresum}.
The corrections are activated by setting one or more of the $\delta^{FAB}_{ij}$ to
non-zero values. All $\delta^{FAB}_{ij}$ that are not set are assumed to be zero.
The non-zero value can be set in three ways:
\begin{itemize}
\item
by including them in the input file, e.g.
\texttt{deltaLLL23~~~~0.1}
where the general format of the identifier is
\texttt{delta$F$$XY$$ij$, $F$ = L,E,Q,U,D, $XY$ = LL,LR,RL,RR, $ij$ = 12,23,13}
\item
by calling the subroutine \texttt{FHSetLFV(\ldots)} from your
Fortran/C/C++ code.
\item
by calling the routine \texttt{FHSetLFV[\ldots]} from your Mathematica
code.
\end{itemize}
The detailed invocation of \texttt{FHSetLFV} is given in the
corresponding man page included in the {\tt FeynHiggs}\ distribution. The LFV
corrections are included starting from {\tt FeynHiggs}\ version 2.10.2,
available from \texttt{feynhiggs.de}.
\chapter{Quark Flavor Mixing Effects in the Model Independent Approach}
MFV sceneraios put tight constraints on the possible value of the FCNC
couplings, especially for the first and second generation squarks which
are sensitive to the data on $K^0-\bar{K}^0$ and $D^0-\bar{D}^0$
mixing. However, the third generation is less constrained, since present
data on $B^0-\bar{B}^0$ mixing still leaves some room for FCNCs. This
allows some parameter space for the more general scenerios focusing on
the mixing between second and third generation (s)quarks.
One such example is the neutral higgs decay \ensuremath{h \to \bar b s + b \bar s}.
The SM contribution is highly suppressed
for this process but the SUSY-QCD quark-squark-gluino loop contribution
can enhance the MSSM contribuion by several orders of magnitude. Also the
SUSY-EW one loop contribution from quark-squark-chargino and
quark-squark-neutralino loop even though subdominent, can have sizable
effects on the \ensuremath{\br(\hbs)}, where in particular the interfrence effects of
SUSY-QCD and SUSY-EW loop corrections can be relevant.
This decay in the framework of the MSSM has been analyzed in the
literature: the SUSY-QCD contributions for this decay were calculated
in~\cite{HdecNMFV,SUSY-QCD}, and the SUSY-EW contributions using the mass
insertion approximation were calculated in~\cite{Demir}. Later
in~\cite{SUSY-EW} the SUSY-EW contributions and their interference
effects with the SUSY-QCD contribution were calculated using exact
diagonalization of the squark mass matrices. In all these
analysis, only LL mixing in the
squarks mass matrix was considered, and experimental constraints were
imposed only from \ensuremath{{\rm BR}(B \to X_s \gamma)}. Most recently in~\cite{SUSY-EW-RR} also RR
mixing has been included.
However mixing of the LR or RL elements
of the mass matrix and constraints from other BPO or potential other constraints were not taken into account
(except in the most recent analysis in~\cite{SUSY-EW-RR}).
In this chapter we will analyze the decay \ensuremath{h \to \bar b s + b \bar s}, evaluated at the full
one-loop level, by taking into account the experimental constraints
not only from BPO but also from the EWPO. In the scalar quark sector we will not only
consider the LL mixing, but also include the LR-RL and RR mixing for
our analysis of \ensuremath{\br(\hbs)}. We will analyze this
decay in the model independent approach where flavor mixing
parameters are put in by hand without any emphasis on the origin of this
mixing (but respecting the experimental bounds from BPO and EWPO). The results
presented in this chapter were published in \cite{EWPO-BPO-QFVHD}.
In the next section we enlist the input parameters for our MI analysis.
\section{Input parameters}
Regarding our choice of MSSM parameters for our forthcoming numerical
analysis, we have chosen the framework of \cite{Arana-Catania:2013nha},
This framework is well compatible with present data.
In this framework, six specific points in the MSSM
parameter space, have been selected. These points are allowed by
present data, including recent LHC searches and the measurements of the
muon anomalous magnetic moment. In \refta{tab:spectra} the
values of the various MSSM parameters as well as the values of the
predicted MSSM mass spectra are summarized. They were evaluated with
the program
{\tt FeynHiggs}~\cite{feynhiggs,mhiggslong,mhiggsAEC,mhcMSSMlong,Mh-logresum}. For
simplicity, and to reduce the number of
independent MSSM input parameters we have assumed equal soft masses for
the sleptons of the first and second generations (similarly for the
squarks), equal soft masses for the left and right slepton sectors
(similarly for the squarks, where $\tilde Q$ denotes the the
``left-handed'' squark sector, whereas $\tilde U$ and $\tilde D$ denote
the up- and down-type parts of the ``right-handed'' squark sector)
and also equal trilinear couplings for
the stop, $A_t$, and sbottom squarks, $A_b$. In the slepton sector we
just consider the stau trilinear coupling, $A_\tau$. The other trilinear
sfermion couplings are set to zero. Regarding the SSB parameters for the gaugino
masses, $M_i$ ($i=1,2,3$), we assume an approximate GUT relation. The
pseudoscalar Higgs mass $M_A$, and the $\mu$ parameter are also taken as
independent input parameters. In summary, the six points S1, \ldots, S6 are
defined in terms of the following subset of ten input MSSM parameters:
\begin{eqnarray}
m_{\tilde L_1} &=& m_{\tilde L_2} \; ; \; m_{\tilde L_3} \;
(\mbox{with~} m_{\tilde L_{i}} = m_{\tilde E_{i}}\,\,,\,\,i=1,2,3) \nonumber \\
m_{\tilde Q_1} &=& m_{\tilde Q_2} \; ; \; m_{\tilde Q_3} \;
(\mbox{with~} m_{\tilde Q_i} = m_{\tilde U_i} = m_{\tilde D_i}\,\,,\,\,i=1,2,3)
\nonumber \\
A_t&=&A_b\,\,;\,\,A_\tau \nonumber \\
M_2&=&2 M_1\, =\,M_3/4 \,\,;\,\,\mu \nonumber \\
M_A&\,\,;\,\, &\tan \beta
\end{eqnarray}
\begin{table}[h!]
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& S1 & S2 & S3 & S4 & S5 & S6 \\\hline
$m_{\tilde L_{1,2}}$& 500 & 750 & 1000 & 800 & 500 & 1500 \\
$m_{\tilde L_{3}}$ & 500 & 750 & 1000 & 500 & 500 & 1500 \\
$M_2$ & 500 & 500 & 500 & 500 & 750 & 300 \\
$A_\tau$ & 500 & 750 & 1000 & 500 & 0 & 1500 \\
$\mu$ & 400 & 400 & 400 & 400 & 800 & 300 \\
$\tan \beta$ & 20 & 30 & 50 & 40 & 10 & 40 \\
$M_A$ & 500 & 1000 & 1000 & 1000 & 1000 & 1500 \\
$m_{\tilde Q_{1,2}}$ & 2000 & 2000 & 2000 & 2000 & 2500 & 1500 \\
$m_{\tilde Q_{3}}$ & 2000 & 2000 & 2000 & 500 & 2500 & 1500 \\
$A_t$ & 2300 & 2300 & 2300 & 1000 & 2500 & 1500 \\\hline
$m_{\tilde l_{1}}-m_{\tilde l_{6}}$ & 489-515 & 738-765 & 984-1018 & 474-802 & 488-516 & 1494-1507 \\
$m_{\tilde \nu_{1}}-m_{\tilde \nu_{3}}$& 496 & 747 & 998 & 496-797 & 496 & 1499 \\
$m_{{\tilde \chi}_1^\pm}-m_{{\tilde \chi}_2^\pm}$ & 375-531 & 376-530 & 377-530 & 377-530 & 710-844 & 247-363 \\
$m_{{\tilde \chi}_1^0}-m_{{\tilde \chi}_4^0}$& 244-531 & 245-531 & 245-530 & 245-530 & 373-844 & 145-363 \\
$M_{h}$ & 126.6 & 127.0 & 127.3 & 123.1 & 123.8 & 125.1 \\
$M_{H}$ & 500 & 1000 & 999 & 1001 & 1000 & 1499 \\
$M_{A}$ & 500 & 1000 & 1000 & 1000 & 1000 & 1500 \\
$M_{H^\pm}$ & 507 & 1003 & 1003 & 1005 & 1003 & 1502 \\
$m_{\tilde u_{1}}-m_{\tilde u_{6}}$& 1909-2100 & 1909-2100 & 1908-2100 & 336-2000 & 2423-2585 & 1423-1589 \\
$m_{\tilde d_{1}}-m_{\tilde d_{6}}$ & 1997-2004 & 1994-2007 & 1990-2011 & 474-2001 & 2498-2503 & 1492-1509 \\
$m_{\tilde g}$ & 2000 & 2000 & 2000 & 2000 & 3000 & 1200 \\
\hline
\end{tabular}
\caption[Selected points in the MSSM parameter space.]{Selected points in the MSSM parameter space (upper part)
and their corresponding spectra (lower part).
All mass parameters and trilinear couplings are given in GeV.}
\label{tab:spectra}
\end{table}
The specific values of these ten MSSM parameters in \refta{tab:spectra},
to be used in the forthcoming analysis, are chosen to provide
different
patterns in the various sparticle masses, but all leading to rather
heavy spectra, thus they are naturally in agreement with the
absence of SUSY signals at LHC. In particular
all points lead to rather heavy squarks and gluinos above $1200\,\, \mathrm{GeV}$ and
heavy sleptons above $500\,\, \mathrm{GeV}$ (where the LHC limits would also
permit substantially lighter scalar leptons).
The values of $M_A$ within the interval
$(500,1500)\,\, \mathrm{GeV}$, $\tan \beta$ within the interval $(10,50)$ and a large
$A_t$ within $(1000,2500)\,\, \mathrm{GeV}$ are fixed such that a light Higgs boson
$h$ within the LHC-favoured range $(123,127)\,\, \mathrm{GeV}$ is obtained.
It should also be noted that the large chosen values of $M_A \ge 500$
GeV place the Higgs sector of our scenarios in the so called decoupling
regime\cite{Haber:1989xc},
where the couplings of $h$ to gauge bosons and fermions are close to
the SM Higgs couplings, and the heavy $H$ couples like the
pseudoscalar $A$, and all heavy Higgs bosons are close in mass.
Increasing $M_A$ the heavy
Higgs bosons tend to decouple from low energy physics and the light
$h$ behaves like $H_{\rm SM}$. This type of MSSM Higgs sector seems
to be in good agreement with recent LHC data\cite{LHCHiggs}.
We have checked with the code {\tt HiggsBounds}~\cite{higgsbounds} (but not yet taking into account the most recent update\cite{Higgsbounds-2015})
that the Higgs sector is in agreement with the LHC searches (where S3
is right ``at the border'').
Particularly, the so far absence of gluinos at
LHC, forbids too low $M_3$ and, therefore, given the assumed GUT
relation, forbids also a too low $M_2$. Consequently, the
values of $M_2$ and $\mu$ are fixed as to get gaugino masses compatible
with present LHC bounds.
Finally, we have also required that all our points lead to a prediction
of the anomalous magnetic moment of the muon in the MSSM that can fill
the present discrepancy between the SM prediction and the
experimental value (see \cite{Arana-Catania:2013nha} for more details).
\section{Experimental constraints on $\delta^{FAB}_{ij}$}
In this section we will present the present experimental constraints on
the squark mixing parameters $\delta^{FAB}_{ij}$ for the above mentioned MSSM
points S1\dots S6 defined in \refta{tab:spectra}. The experimental
constraints from BPO for the same set of parameters that we are using
were already calculated in \cite{arana-NMFV2} for one $\delta^{FAB}_{ij} \neq0$ ,
which we reproduce here for completeness in the \refta{tab:boundsS1S6}.
We now turn our attention to the constraints from $M_W$.
In \reffi{Fig:DMW} we show the
$M_W$ as a function of $\del{QLL}{23}$, $\del{ULR}{23}$ and
$\del{DLR}{23}$ in the scenarios S1 \ldots S6.
The area between the orange lines shows the allowed
value of $M_W$ with $3\sigma$ experimental uncertainty.
The corresponding constraints from $M_W$ on $\delta^{FAB}_{ij}$, also taking
into account the theoretical uncertainties as described at the end of
\refse{sec:EWPO-calc}, are shown in \refta{tab:EWPOboundsS1S6}.
No constraints can be found on the $\del{RR}{ij}$,
as their contribution to $M_W$ does not reach the MeV level, and consequently
we do not show them here. Furtheremore, the constraints for the
$\del{URL}{23}$ and $\del{DRL}{23}$ are similar to those for
$\del{ULR}{23}$ and $\del{DLR}{23}$, respectively, and not shown
here.
On the other hand, the constraints on $\del{QLL}{23}$ are modified by the
EWPO specially the region (-0.83:-0.78) for the point S5, which was
allowed by the BPO, is now excluded. The allowed intervals for the
points S1-S3 have also shrunk. However the point S4 was already excluded
by BPO, similarly the allowed interval for S6 do not get modified by
EWPO. The constraints on $\del{ULR}{23}$ and $\del{DLR}{23}$
are less restrictive then the ones from BPO except for the point S4
where the region (0.076:0.12) is excluded for $\del{DLR}{23}$ by
EWPO.
\begin{figure}[htb!]
\begin{center}
\psfig{file=Plots/MISQplots/DMWQLL23.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/MISQplots/DMWULR23.eps ,scale=0.52,angle=0,clip=}\\
\vspace{2.0cm}
\psfig{file=Plots/MISQplots/DMWDLR23.eps ,scale=0.52,angle=0,clip=}
\vspace{0.2cm}
\end{center}
\caption[$M_W$ as a function of squark $\delta^{FAB}_{ij}$]{$M_W$ as a function of $\del{QLL}{23}$ (upper left),
$\del{ULR}{23}$ (upper right) and $\del{DLR}{23}$ (lower). }
\label{Fig:DMW}
\end{figure}
\renewcommand{\arraystretch}{1.10}
\begin{table}[htb!]
\begin{center}
\resizebox{9.0cm}{!} {
\begin{tabular}{|c|c|c|} \hline
& & Total allowed intervals \\ \hline
$\delta^{QLL}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.27:0.28) \\ (-0.23:0.23) \\ (-0.12:0.06) (0.17:0.19) \\ excluded \\ (-0.83:-0.78) (-0.14:0.14) \\ (-0.076:0.14) \end{tabular} \\ \hline
$\delta^{ULR}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular}
& \begin{tabular}{c}
(-0.27:0.27) \\ (-0.27:0.27) \\ (-0.27:0.27) \\ excluded \\ (-0.22:0.22) \\ (-0.37:0.37) \end{tabular} \\ \hline
$\delta^{DLR}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.0069:0.014) (0.12:0.13) \\ (-0.0069:0.014) (0.11:0.13) \\ (-0.0069:0.014) (0.11:0.13) \\ (0.076:0.12) (0.26:0.30) \\ (-0.014:0.021) (0.17:0.19) \\ (0:0.0069) (0.069:0.076) \end{tabular} \\ \hline
$\delta^{URL}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.27:0.27) \\ (-0.27:0.27) \\ (-0.27:0.27) \\ excluded \\ (-0.22:0.22) \\ (-0.37:0.37) \end{tabular}
\\ \hline
$\delta^{DRL}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c} (-0.034:0.034) \\ (-0.034:0.034) \\ (-0.034:0.034) \\ excluded \\ (-0.062:0.062) \\ (-0.021:0.021) \end{tabular}
\\ \hline
$\delta^{URR}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} & \begin{tabular}{c}
(-0.99:0.99) \\ (-0.99:0.99) \\ (-0.98:0.97) \\ excluded \\ (-0.99:0.99) \\ (-0.96:0.94) \end{tabular} \\ \hline
$\delta^{DRR}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c} (-0.96:0.96) \\ (-0.96:0.96) \\ (-0.96:0.94) \\ excluded \\ (-0.97:0.97) \\ (-0.97:-0.94) (-0.63:0.64) (0.93:0.97)
\end{tabular} \\ \hline
\end{tabular}}
\end{center}
\vspace{-1em}
\caption[Present allowed (by BPO) intervals for the $\delta^{FAB}_{ij}$]
{Present allowed (by BPO) intervals for the $\delta^{FAB}_{ij}$ for the MSSM points defined in
\refta{tab:spectra}\cite{arana-NMFV2}.
}
\label{tab:boundsS1S6}
\vspace{-4em}
\end{table}
\renewcommand{\arraystretch}{1.55}
\renewcommand{\arraystretch}{1.1}
\begin{table}[htb!]
\begin{center}
\resizebox{9.0cm}{!} {
\begin{tabular}{|c|c|c|} \hline
& & Total allowed intervals \\ \hline
$\delta^{QLL}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.18:0.18) \\ (-0.18:0.18) \\ (-0.18:0.18) \\ (-0.53:-0.17)(0.10:0.45) \\ (-0.14:0.14) \\ (-0.23:0.23) \end{tabular} \\ \hline
$\delta^{ULR}_{23},\delta^{URL}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular}
& \begin{tabular}{c}
(-0.41:0.41) \\ (-0.41:0.41) \\ (-0.41:0.41) \\ (0.10:0.50) \\ (-0.39:0.39) \\ (-0.47:0.47) \end{tabular} \\ \hline
$\delta^{DLR}_{23},\delta^{DRL}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.43:0.43) \\ (-0.43:0.43) \\ (-0.43:0.43) \\ (0.16:0.99) \\ (-0.39:0.39) \\ (-0.49:0.49) \end{tabular} \\ \hline
\end{tabular}}
\end{center}
\caption[Present allowed (by $M_W$) intervals for the $\delta^{FAB}_{ij}$]
{Present allowed (by $M_W$) intervals for the squark mixing parameters
$\delta^{FAB}_{ij}$ for the selected S1-S6 MSSM points defined in
\refta{tab:spectra}.
}
\label{tab:EWPOboundsS1S6}
\end{table}
\renewcommand{\arraystretch}{1.55}
\section{\boldmath{\ensuremath{\br(\hbs)}}}
In order to illustrate the contributions from different diagrams we show
in \reffi{SUSY-CONT} the SUSY-EW, SUSY-QCD and total SUSY contribution
to $\Gamma(\ensuremath{h \to \bar b s + b \bar s})$ as a function of
$\del{QLL}{23}$ (upper left), $\del{DLR}{23}$ (upper right),
$\del{DRL}{23}$ (lower left) and $\del{DRR}{23}$ (lower
right). These four $\delta^{FAB}_{ij}$ are the only relevant ones, since we
are mainly concerned with the down-type sector, and mixing with the
first generation does not play a role.
In order to compare our results with the literature, we have used
the same set of input parameters as in \cite{SUSY-EW}:
\begin{align}
\mu &= 800 \,\, \mathrm{GeV},\; m_{\rm SUSY} = 800 \,\, \mathrm{GeV},\; A_f = 500 \,\, \mathrm{GeV}, \nonumber\\
M_A &= 400 \,\, \mathrm{GeV},\; M_2 = 300 \,\, \mathrm{GeV},\; \tan \beta = 35 \, ,
\end{align}
where we have chosen, for simplicity, $m_{\rm SUSY}$ as a common value for the
soft SUSY-breaking squark mass parameters,
$m_{\rm SUSY} = M_{\tilde Q} = M_{\tilde U} = M_{\tilde D}$,
and all the various trilinear parameters to be universal,
$A_f=A_t=A_b=A_c=A_s$. The value of the $\delta^{FAB}_{ij}$'s are varied from
-0.9~to~0.9, and GUT relations are used to calculate $M_1$ and
$M_3$.
In \citere{SUSY-EW}, only LL mixing was considered. In this limit we find
results in qualitative agreement with \citere{SUSY-EW}. This analysis has
been done just to illustrate the different contributions and we do not
take into account any experimental constraints. A detailed analysis for
realisitic SUSY scenerios (defined in \refta{tab:spectra}) constrained
by BPO and EWPO can be found below.
As can be seen in \reffi{SUSY-CONT}, for the decay width $\Gamma(\ensuremath{h \to \bar b s + b \bar s})$ the
SUSY-QCD contribution is dominant in all the cases. For LL mixing shown
in the upper left plot, the SUSY-QCD contribution reaches up
to \order{10^{-6}}, while the SUSY-EW contribution
reach up to \order{10^{-7}}, resulting in a total contribution ``in
between'', due to the negative interference between SUSY-EW and SUSY-QCD
contribution.
For LR and RL mixing, shown in the upper right and lower left plot,
respectively, the SUSY-QCD contribution reach up to the
maximum value of \order{10^{-2}}, while the SUSY-EW contribution reach
only up to \order{10^{-7}}. In this case total contriution is almost equal to
SUSY-QCD contribution as SUSY-EW contibution (and thus the interference)
is relatively neglible.
For RR mixing, shown in the lower right plot, the SUSY-EW
contribution of \order{10^{-10}} is again neglible compared to SUSY-QCD
contribution of \order{10^{-7}}.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/SUSYQLL23.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/MISQplots/SUSYDLR23.eps ,scale=0.52,angle=0,clip=}\\
\vspace{2.0cm}
\psfig{file=Plots/MISQplots/SUSYDRL23.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/MISQplots/SUSYDRR23.eps ,scale=0.52,angle=0,clip=}
\vspace{0.2cm}
\end{center}
\caption[$\Gamma(\ensuremath{h \to \bar b s + b \bar s})$ as a function of squark $\delta^{FAB}_{ij}$]{$\Gamma(\ensuremath{h \to \bar b s + b \bar s})$ as a function of $\del{QLL}{23}$ (upper left),
$\del{DLR}{23}$ (upper right), $\del{DRL}{23}$ (lower left) and
$\del{DRR}{23}$ (lower right).
}
\label{SUSY-CONT}
\end{figure}
Now we turn to realistic scenarios that are in agreement with
experimental data from BPO and EWPO. Starting point are the scenarios
S1\ldots S6 defined in \refta{tab:spectra}, where we vary the flavor
violating $\delta^{FAB}_{ij}$ within the experimentally allowed ranges following
the results given in \reftas{tab:boundsS1S6}, \ref{tab:EWPOboundsS1S6}.
We start with the scenarios in which we allow one of the $\delta^{FAB}_{ij}$ to
be varied, while the others are set to zero.
In \reffi{Fig:QFVHD} we show \ensuremath{\br(\hbs)}\ as a function of $\del{QLL}{23}$
(upper left), $\del{DLR}{23}$ (upper right), $\del{DRL}{23}$ (lower
left) and $\del{DRR}{23}$ (lower right), i.e.\ for the same set of
$\delta^{FAB}_{ij}$ that has been analyzed in \reffi{SUSY-CONT}.
It can be seen that allowing only one $\delta^{FAB}_{ij} \neq 0$ results in
rather small values of \ensuremath{\br(\hbs)}. LL (upper left) and RL (lower left plot)
mixing results in \order{10^{-7}} values for \ensuremath{\br(\hbs)}. One order of
magnitude can be gained in the RR mixing case (lower right). The largest
values of \ensuremath{\br(\hbs)}\ are obtained in the case of $\del{DLR}{23} \neq 0$
(upper right plot). Here in S4 and S5 values of
$\ensuremath{\br(\hbs)} \sim 2 \times 10^{-4}$ can be found, possibly in the reach of
future $e^+e^-$ colliders, see \refse{sec:hbs-calc}.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/QFVHDQLL23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MISQplots/QFVHDDLR23.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/QFVHDDRL23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MISQplots/QFVHDDRR23.eps ,scale=0.51,angle=0,clip=}\\
\end{center}
\caption[\ensuremath{\br(\hbs)}\ as a function of squark $\delta^{FAB}_{ij}$]{\ensuremath{\br(\hbs)}\ as a function of $\del{QLL}{23}$ (upper left),
$\del{DLR}{23}$ (upper right), $\del{DRL}{23}$ (lower left) and
$\del{DRR}{23}$ (lower right).}
\label{Fig:QFVHD}
\end{figure}
So far we have shown the effects of independent variations of
one $\delta^{FAB}_{ij}$. Obviously, a realistic model would include several
$\delta^{FAB}_{ij} \neq 0$ that may interfere, increasing or decreasing the results
obtained with just the addition of independent contributions.
GUT based MFV models that induce the flavor violation
via RGE running automatically generate several $\delta^{FAB}_{ij} \neq 0$ at the
EW scale.
In the following we will present results with two or three
$\delta^{FAB}_{ij} \neq 0$, where we combined the ones that showed the largest
effects.
In \reffis{Fig:S1S3QLLDLR23}-\ref{Fig:S4S6DLRDRR23}, in the left columns
we show the $3\,\sigma$ contours (with experimental and theory
uncertainties added linearly) of
\ensuremath{{\rm BR}(B \to X_s \gamma)}\ (Black), \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (Green), \ensuremath{\Delta M_{B_s}}\ (Blue) and $M_W$ (Red).
For non-visible contours the whole plane is allowed by that constraint.
The right columns show, for the same parameters, the results for \ensuremath{\br(\hbs)}.
In \reffis{Fig:S1S3QLLDLR23} and \ref{Fig:S4S6QLLDLR23} we present the
results for the plane ($\del{QLL}{23}$,$\del{DLR}{23}$) for S1\ldots
S3 and for S4\ldots S6, respectively. Similarly,
in \reffis{Fig:S1S3DLRDRR23} and \ref{Fig:S4S6DLRDRR23} we show the
($\del{DRR}{23}$, $\del{DLR}{23}$) plane.
The shaded area in the left columns indicates the area that is allowed
by all experimental constraints.
In the ($\del{QLL}{23}$, $\del{DLR}{23}$) planes one can see that the
large values for $\del{QLL}{23}$ are not allowed by $M_W$, on the
other hand, \ensuremath{{\rm BR}(B \to X_s \gamma)}\ mostly restricts the value of
$\del{DLR}{23}$.
The largest values for \ensuremath{\br(\hbs)}\ in each plane in the arrea allowed by the
BPO and the EWPO are summarized in the upper part of \refta{tab:brhbs-2d}.
One can see that
in most cases we find ${\rm BR}(\ensuremath{h \to \bar b s + b \bar s}) \sim \order{10^{-5}}$,
which would render the observation difficult at current and future
colliders. However, in the ($\del{QLL}{23},\del{DLR}{23}$) plane in the
scenarios S4 and S5 maximum values of \order{3 \times 10^{-4}} can be
observed, which could be detectable at future ILC measurements.
In the ($\del{DRR}{23}$, $\del{DLR}{23}$) plane for these two scenarios
even values of \order{10^{-3}} are reached, which would make a
measurement of the flavor violating Higgs decay relatively easy at the
ILC.
\renewcommand{\arraystretch}{1.2}
\begin{table}[htb!]
\begin{center}
\resizebox{11.0cm}{!} {
\begin{tabular}{|c|c|c|c|} \hline
Plane & MSSM point & Maximum possible value & Figure \\ \hline
($\del{QLL}{23},\del{DLR}{23}$) & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
$1.38 \times 10^{-5}$ \\ $1.39 \times 10^{-5}$ \\ $1.43 \times 10^{-5}$
\\ $3.34 \times 10^{-4}$ \\ $2.74 \times 10^{-4}$ \\ $1.36 \times 10^{-8}$
\end{tabular} &
\begin{tabular}{c}
\reffi{Fig:S1S3QLLDLR23} \\ \reffi{Fig:S1S3QLLDLR23} \\ \reffi{Fig:S1S3QLLDLR23} \\
\reffi{Fig:S4S6QLLDLR23} \\ \reffi{Fig:S4S6QLLDLR23} \\ \reffi{Fig:S4S6QLLDLR23}
\end{tabular}
\\ \hline
($\del{DRR}{23},\del{DLR}{23}$) & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular}
& \begin{tabular}{c}
$4.41 \times 10^{-6}$ \\ $3.32 \times 10^{-6}$ \\ $3.07 \times 10^{-5}$
\\ $1.66 \times 10^{-3}$ \\ $1.97 \times 10^{-3}$ \\ $6.03 \times 10^{-8}$
\end{tabular} &
\begin{tabular}{c}
\reffi{Fig:S1S3DLRDRR23} \\ \reffi{Fig:S1S3DLRDRR23} \\ \reffi{Fig:S1S3DLRDRR23} \\
\reffi{Fig:S4S6DLRDRR23} \\ \reffi{Fig:S4S6DLRDRR23} \\ \reffi{Fig:S4S6DLRDRR23}
\end{tabular}
\\ \hline\hline
\begin{tabular}{c}($\del{QLL}{23},\del{DLR}{23}$) \\ with $\del{DRR}{23}= 0.5$ \end{tabular} & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
$7.49 \times 10^{-6}$ \\ $7.33 \times 10^{-6}$ \\$3.50 \times 10^{-6}$
\\ Excluded \\ Excluded \\ Excluded \end{tabular} &
\begin{tabular}{c}
\reffi{Fig:S1S3QLLDLRDRR23} \\ \reffi{Fig:S1S3QLLDLRDRR23} \\ \reffi{Fig:S1S3QLLDLRDRR23} \\
\reffi{Fig:S4S6QLLDLRDRR23} \\ \reffi{Fig:S4S6QLLDLRDRR23} \\ \reffi{Fig:S4S6QLLDLRDRR23}
\end{tabular} \\ \hline
\end{tabular}}
\end{center}
\caption[Maximum $\ensuremath{\br(\hbs)}$ for two and three $\delta^{FAB}_{ij} \neq 0$ case]{Maximum possible value for $\ensuremath{\br(\hbs)}$ for two and three $\delta^{FAB}_{ij} \neq 0$ case for the selected S1-S6 MSSM points defined in
\refta{tab:spectra}.
}
\label{tab:brhbs-2d}
\end{table}
\renewcommand{\arraystretch}{1.55}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/S1QLLDLRBPhy.eps ,scale=0.70,angle=0,clip=}
\psfig{file=Plots/MISQplots/S1QLLDLRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S2QLLDLRBPhy.eps ,scale=0.70,angle=0,clip=}
\psfig{file=Plots/MISQplots/S2QLLDLRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S3QLLDLRBPhy.eps ,scale=0.70,angle=0,clip=}
\psfig{file=Plots/MISQplots/S3QLLDLRBRhbs.eps ,scale=0.52,angle=0,clip=}
\end{center}
\vspace{-2em}
\caption[Contours of EWPO, BPO and \ensuremath{\br(\hbs)}\ in ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane]{Left: Contours of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ (Black), \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (Green), \ensuremath{\Delta M_{B_s}} (Blue) and
$M_W$ (Red) in ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane for
points S1-S3. The shaded area shows the range of values allowed by all
constraints. Right: corresponding contours for \ensuremath{\br(\hbs)}.}
\label{Fig:S1S3QLLDLR23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/S4QLLDLRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S4QLLDLRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S5QLLDLRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S5QLLDLRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S6QLLDLRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S6QLLDLRBRhbs.eps ,scale=0.52,angle=0,clip=}
\vspace{-2em}
\end{center}
\caption[Contours of EWPO, BPO and \ensuremath{\br(\hbs)}\ in ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane]{Left: Contours of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ (Black), \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (Green), \ensuremath{\Delta M_{B_s}}\ (Blue) and
$M_W$ (Red) in ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane for
points S4-S6. The shaded area shows the range of values allowed by all
constraints. Right: corresponding contours for \ensuremath{\br(\hbs)}.}
\label{Fig:S4S6QLLDLR23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/S1DLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S1DLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S2DLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S2DLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S3DLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S3DLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}
\vspace{-2em}
\end{center}
\caption[Contours of EWPO, BPO and \ensuremath{\br(\hbs)}\ in ($\del{DRR}{23}$ , $\del{DLR}{23}$) plane]
{Left: Contours of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ (Black), \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (Green), \ensuremath{\Delta M_{B_s}}\ (Blue) and
$M_W$ (Red) in ($\del{DRR}{23}$ , $\del{DLR}{23}$) plane for points S1-S3.
The shaded area shows the range of values allowed by all
constraints. Right: corresponding contours for \ensuremath{\br(\hbs)}.}
\label{Fig:S1S3DLRDRR23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/S4DLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S4DLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S5DLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S5DLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S6DLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S6DLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}
\vspace{-2em}
\end{center}
\caption[Contours of EWPO, BPO and \ensuremath{\br(\hbs)}\ in ($\del{DRR}{23}$ , $\del{DLR}{23}$) plane]
{Left: Contours of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ (Black), \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (Green), \ensuremath{\Delta M_{B_s}}\ (Blue) and
$M_W$ (Red) in ($\del{DRR}{23}$ , $\del{DLR}{23}$) plane for points S4-S6.
The shaded area shows the range of values allowed by all
constraints. Right: corresponding contours for \ensuremath{\br(\hbs)}.}
\label{Fig:S4S6DLRDRR23}
\end{figure}
\medskip
As a last step in model independent analysis, we consider the case of
three $\delta^{FAB}_{ij} \neq 0 $ at a time. For this purpose we
scan the parameters in the ($\del{QLL}{23}$, $\del{DLR}{23}$)
plane and set $\del{DRR}{23} = 0.5$. For reasons of practicability
we choose {\em one} intermediate value for $\del{DRR}{23}$; a very small
value will have no additional effect, and a very large value of
$\del{DRR}{23}$ leads to large excluded areas in the
($\del{QLL}{23}$, $\del{DLR}{23}$) plane.
We show our results in \reffis{Fig:S1S3QLLDLRDRR23} and
\ref{Fig:S4S6QLLDLRDRR23} in the scenarios S1-S3 and S4-S6,
respectively. Colors and shadings are chosen as in the previous
analysis.
Here it should be noted that in S4 the whole plane is excluded by
$M_W$, and in S5 by \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (both contours are not visible). In S6 no
overlap between the four constraints is found, and again this scenario
is excluded. We have checked that also a smaller value of
$\del{DRR}{23} = 0.2$ does not qualitatively change the picture for
S4, S5 and S6.
The highest values that can be reached for \ensuremath{\br(\hbs)}\ in the three
remaining scenarios in the experimentally
allowed regions are shown in the lower part of \refta{tab:brhbs-2d}.
One can see only very small valus or \order{5 \times 10^{-6}} are
found, i.e.\ choosing $\del{DRR}{23} \neq 0$ did not lead to
observable values of \ensuremath{\br(\hbs)}.
To summarize, in our model independent analysis, allowing for more than
one $\delta^{FAB}_{ij} \neq 0$ we find
that the additional freedom resulted in somewhat larger values of
\ensuremath{\br(\hbs)}\ as compared to the case of only one non-zero $\delta^{FAB}_{ij}$.
In particular in the two scenarios S4 and S5 values of
$\ensuremath{\br(\hbs)} \sim 10^{-3} - 10^{-4}$ can be reached, allowing the
detection of the flavor violating Higgs decay at the ILC. The other
scenarios always yield values that are presumably too low for
current and future colliders.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/S1QLLDLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S1QLLDLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S2QLLDLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S2QLLDLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S3QLLDLRDRRBPhy.eps ,scale=0.69,angle=0,clip=}
\psfig{file=Plots/MISQplots/S3QLLDLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}
\vspace{-2em}
\end{center}
\caption[Contours of EWPO, BPO and \ensuremath{\br(\hbs)}\ in ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane]
{Left: Contours of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ (Black), \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (Green), \ensuremath{\Delta M_{B_s}}\ (Blue) and
$M_W$ (Red) in the ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane with
$\del{DRR}{23} = 0.5$ for points S1-S3. The shaded area shows the range
of values allowed by all
constraints. Right: corresponding contours
for \ensuremath{\br(\hbs)}.}
\label{Fig:S1S3QLLDLRDRR23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISQplots/S4QLLDLRDRRBPhy.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/MISQplots/S4QLLDLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S5QLLDLRDRRBPhy.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/MISQplots/S5QLLDLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MISQplots/S6QLLDLRDRRBPhy.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/MISQplots/S6QLLDLRDRRBRhbs.eps ,scale=0.52,angle=0,clip=}
\vspace{-2em}
\end{center}
\caption[Contours of EWPO, BPO and \ensuremath{\br(\hbs)}\ in ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane]
{Left: Contours of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ (Black), \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ (Green), \ensuremath{\Delta M_{B_s}}\ (Blue) and
$M_W$ (Red) in the ($\del{QLL}{23}$ , $\del{DLR}{23}$) plane with
$\del{DRR}{23} = 0.5$ for points S4-S6. The shaded area shows the range
of values allowed by all
constraints. Right: corresponding contours
for \ensuremath{\br(\hbs)}.}
\label{Fig:S4S6QLLDLRDRR23}
\end{figure}
\chapter{Lepton Flavor Mixing Effects in the Model Independent Approach}
In this chapter we analyse the lepton flavor mixing in MI approach. We use the same set of input parameter (\refta{tab:spectra}) that was used in the previous chater. As a first step, we have calculated the sensitivity of EWPO like the
$W$-boson mass or the effective weak leptonic mixing angle to the $\delta^{FAB}_{ij}$'s in slepton sector
entering in the $Z$ and $W$ boson self energies at one-loop
level through the $\rho$~parameter. Besides EWPO we also explore the effects of LFV on the MSSM Higgs
sector. We evaluate the effects of LFV on the predictions of the masses of the light and heavy
${\cal CP}$-even Higgs bosons, $M_h$ and $M_H$, as well as on the charged
Higgs-boson mass $M_{H^\pm}$. Here we do not calculate predictions for cLFV decays
in the MI approach as they were already explored in \cite{Arana-Catania:2013nha} for the same set of input parameters that we are using. They calculated the constraints on slepton $\delta^{FAB}_{ij}$'s from cLFV decays (mentioned in the following section). We have also calculated the predictions for LFVHD which will be presented in the last section. The results presented in this chapter were published in \cite{drhoLFV}.
\section{Constraints on \boldmath{$\delta^{FAB}_{ij}$} from cLFV decays}
\label{sec:limits}
We need to set the range of values for the explored
$\delta^{FAB}_{ij}$'s. We use the constraints (shown in \refta{boundsSpoints})
as taken from \citere{Arana-Catania:2013nha}, calculated from the following LFV processes.
\begin{itemize}
\item[1.-] Radiative LFV decays: $\mu \to e \gamma$, $\tau \to e \gamma$ and
$\tau \to \mu \gamma$. These are sensitive to the $\delta^{FAB}_{ij}$'s via the
$(l_il_j\gamma)_{\rm 1-loop}$ vertices with a real photon.
\item[2.-] Leptonic LFV decays: $\mu \to 3 e$, $\tau \to 3 e$ and $\tau
\to 3 \mu$. These are sensitive to the $\delta^{FAB}_{ij}$'s via the
$(l_il_j\gamma)_{\rm 1-loop}$ vertices with a virtual photon, via
the $(l_il_jZ)_{\rm 1-loop}$ vertices with a virtual $Z$, and via the
$(l_il_jh)_{\rm 1-loop}$, $(l_il_jH)_{\rm 1-loop}$ and
$(l_il_jA)_{\rm 1-loop}$ vertices with virtual Higgs bosons.
\item[3.-] Semileptonic LFV tau decays: $\tau \to \mu \eta$ and $\tau
\to e \eta$. These are sensitive to the $\delta^{FAB}_{ij}$'s via $(\tau \mu
A)_{\rm 1-loop}$ and $(\tau e A)_{\rm 1-loop}$ vertices,
respectively, with a virtual $A$, and via
$(\tau \mu Z)_{\rm 1-loop}$ and $(\tau e Z)_{\rm 1-loop}$ vertices,
respectively with a virtual $Z$.
\item[4.-] Conversion of $\mu$ into $e$ in heavy nuclei: These are
sensitive to the $\delta^{FAB}_{ij}$'s via the $(\mu e\gamma)_{\rm 1-loop}$
vertex with a virtual photon, via the $(\mu e Z)_{\rm 1-loop}$
vertex with a virtual $Z$, and via the $(\mu e h)_{\rm 1-loop}$ and
$(\mu e H)_{\rm 1-loop}$ vertices with a virtual $h$ and $H$
Higgs boson, respectively.
\end{itemize}
\renewcommand{\arraystretch}{1.1}
\begin{table}[htb!]
\begin{center}
\resizebox{15.0cm}{!} {
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& S1 & S2 & S3 & S4 & S5 & S6
\\ \hline
& & & & & & \\
$|\delta^{LLL}_{12}|_{\rm max}$ & $10 \times 10^{-5}$ & $7.5\times 10^{-5}$ & $5 \times 10^{-5}$& $6 \times 10^{-5}$ & $42\times 10^{-5}$ & $8\times 10^{-5}$ \\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{ELR}_{12}|_{\rm max}$ & $2\times 10^{-6}$ & $3\times 10^{-6}$ &
$4\times 10^{-6}$ & $3\times 10^{-6}$ & $2\times 10^{-6}$ & $1.2\times 10^{-5}$ \\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{ERR}_{12}|_{\rm max}$ & $1.5 \times 10^{-3}$& $1.2 \times 10^{-3}$ &
$1.1 \times 10^{-3}$ & $1 \times 10^{-3}$ & $2 \times 10^{-3}$ & $5.2 \times 10^{-3}$ \\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{LLL}_{13}|_{\rm max} $ & $5 \times 10^{-2}$ & $5 \times 10^{-2}$ &
$3 \times 10^{-2}$ & $3 \times 10^{-2}$& $23 \times 10^{-2}$ & $5 \times 10^{-2}$ \\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{ELR}_{13}|_{\rm max}$& $2\times 10^{-2}$ & $3\times 10^{-2}$ & $4\times 10^{-2}$ & $2.5\times 10^{-2}$ & $2\times 10^{-2}$ & $11\times 10^{-2}$ \\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{ERR}_{13}|_{\rm max}$ & $5.4\times 10^{-1}$ & $5\times 10^{-1}$ &
$4.8\times 10^{-1}$ &$5.3\times 10^{-1}$ & $7.7\times 10^{-1}$ & $7.7\times 10^{-1}$
\\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{LLL}_{23}|_{\rm max}$ & $6\times 10^{-2}$ & $6\times 10^{-2}$ &
$4\times 10^{-2}$& $4\times 10^{-2}$ & $27\times 10^{-2}$ & $6\times 10^{-2}$
\\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{ELR}_{23}|_{\rm max}$ & $2\times 10^{-2}$ & $3\times 10^{-2}$ &
$4\times 10^{-2}$ & $3\times 10^{-2}$ & $2\times 10^{-2}$ & $12\times 10^{-2}$
\\
& & & & & & \\
\hline
& & & & & & \\
$|\delta^{ERR}_{23}|_{\rm max}$ & $5.7\times 10^{-1}$ & $5.2\times 10^{-1}$ &
$5\times 10^{-1}$& $5.6\times 10^{-1}$ & $8.3\times 10^{-1}$ & $8\times 10^{-1}$
\\
& & & & & & \\
\hline
\end{tabular}}
\end{center}
\caption[Constraints on $|\delta^{FAB}_{ij}|$ from LFV decays.]{ Present upper bounds on the slepton mixing parameters $|\delta^{FAB}_{ij}|$ for the selected S1-S6 MSSM points defined in \refta{tab:spectra}. The bounds for $|\delta^{ERL}_{ij}|$ are similar
to those of $|\delta^{ELR}_{ij}|$.}
\label{boundsSpoints}
\end{table}
\section{Numerical results}
\label{sec:results}
We have implemented the full one-loop results for the $W$~and
$Z$~boson and the Higgs boson self-energies
in ${\tt FeynHiggs}$, including all LFV mixing terms (see \refse{sec:feynhiggs} for details). The analytical results are
lenghty and are not shown here.
They can, however, be found in the latest version of our code,
{\tt FeynHiggs}\,{\tt 2.10.2}.
For the numerical investigation we have analyzed all 12 slepton $\delta^{FAB}_{ij}$'s
for the MSSM scenarios defined in \refta{tab:spectra}.
In order to get a good understanding of the LFV effects to $\Delta\rho$ and
consequently $M_W$ and $\sin^2\theta_{\mathrm{eff}}$ we define
\begin{align}
\Delta\rho^{\rm LFV} &= \Delta\rho-\Delta\rho^{\rm MSSM}, \\
\deM_W^{\rm LFV} &= M_W-M_W^{\rm MSSM}, \\
\delta\sin^2\theta_{\mathrm{eff}}^{\rm LFV} &= \sin^2\theta_{\mathrm{eff}}-\sin^2\theta_{\mathrm{eff}}^{\rm MSSM},
\end{align}
where $\Delta\rho^{\rm MSSM}$, $M_W^{\rm MSSM}$ and $\sin^2\theta_{\mathrm{eff}}^{\rm MSSM}$ are the
values of the relevant observables with all $\delta^{FAB}_{ij} = 0$ (and the
latter two evaluated with the help of \refeq{eq:precobs}). Furthermore we define
\begin{align}
\Delta M_h^{\rm LFV} &= M_h - M_h^{\rm MSSM}, \\
\Delta M_H^{\rm LFV} &= M_H - M_H^{\rm MSSM}, \\
\Delta M_{H^\pm}^{\rm LFV} &= M_{H^\pm} - M_{H^\pm}^{\rm MSSM},
\end{align}
where $M_h^{\rm MSSM}$, $M_H^{\rm MSSM}$ and
$M_{H^\pm}^{\rm MSSM}$ corresponds to the
Higgs masses with all $\delta^{FAB}_{ij} = 0$.
The SM results for $M_W$ and $\sin^2\theta_{\mathrm{eff}}$ are $M_W=80.361 \,\, \mathrm{GeV}$ and
$\sin^2\theta_{\mathrm{eff}}=0.23152$ as evaluated with {\tt FeynHiggs}\ (using the approximation formulas
given in \citeres{MWSMapprox,sw2effSMapprox}).
The numerical values of $\Delta\rho$, $M_W$, $\sin^2\theta_{\mathrm{eff}}$,
$M_h$, $M_H$ and $M_{H^\pm}$ in the MSSM with all $\delta^{FAB}_{ij} = 0$ are
summarized in \refta{absolutevalues}.
\begin{table}[h!]
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& S1 & S2 & S3 & S4 & S5 & S6
\\ \hline
& & & & & & \\
$\Delta\rho$ & $2.66 \times 10^{-5}$ & $1.72\times 10^{-5}$ & $1.39 \times 10^{-5}$& $2.35 \times 10^{-4}$ & $2.36\times 10^{-5}$ & $2.14\times 10^{-5}$ \\
& & & & & & \\
\hline
& & & & & & \\
$M_W$ & $ 80.362$ & $ 80.362 $ & $80.361$ & $80.375$ & $80.364$ & $80.363$ \\
& & & & & & \\
\hline
& & & & & & \\
$\sin^2\theta_{\mathrm{eff}}$ & $0.23151$& $0.23152$ & $0.23152$ & $0.23143$ & $0.23150$ & $0.23151$ \\
& & & & & & \\
\hline
& & & & & & \\
$ M_{h} $ & $126.257$ & $126.629$ & $126.916$ & $123.205$& $123.220$ & $124.695$ \\
& & & & & & \\
\hline
& & & & & & \\
$M_{H} $ & $500.187$ & $999.580$ & $999.206$ & $1001.428$ & $1000.239$ & $1499.365$ \\
& & & & & & \\
\hline
& & & & & & \\
$M_{H^{\pm}} $ & $506.888$ & $1003.182$ & $1003.005$ &$1005.605$ & $1003.454$ & $1501.553$
\\
& & & & & & \\
\hline
\end{tabular}
\caption[The values of $\Delta\rho$, $M_W$, $\sin^2\theta_{\mathrm{eff}}$,
$M_h$, $M_H$ and $M_{H^\pm}$ with all $\delta^{AB}_{ij} = 0$.]{The values of $\Delta\rho$, $M_W$, $\sin^2\theta_{\mathrm{eff}}$,
$M_h$, $M_H$ and $M_{H^\pm}$ for the selected S1-S6 MSSM points defined
in \refta{tab:spectra} (i.e.\ with all $\delta^{FAB}_{ij} = 0$). Mass values are
in~GeV.}
\label{absolutevalues}
\end{table}
Our numerical results are shown in \reffi{figdLL13} to \reffi{figdRR23}.
The six plots in each figure are ordered as follows.
Upper left: $\Delta\rho^{\rm LFV}$,
upper right: $\deM_W^{\rm LFV}$,
middle left: $\delta\sin^2\theta_{\mathrm{eff}}^{\rm LFV}$,
middle right: $\Delta M_h^{\rm LFV}$,
lower left: $\Delta M_H^{\rm LFV}$,
and lower right: $\Delta M_{H^\pm}^{\rm LFV}$,
as a function of $\delta^{LL}_{13}$ (Fig.\ref{figdLL13}),
$\delta^{LLL}_{23}$ (Fig.\ref{figdLL23}),
$\delta^{ELR}_{13}$ (Fig.\ref{figdLR13}),
$\delta^{ELR}_{23}$ (Fig.\ref{figdLR23}),
$\delta^{ERL}_{13}$ (Fig.\ref{figdRL13}),
$\delta^{ERL}_{23}$ (Fig.\ref{figdRL23}),
$\delta^{ERR}_{13}$ (Fig.\ref{figdRR13})
and $\delta^{ERR}_{23}$ (Fig.\ref{figdRR23}).
The legends are shown only in the first plot of each figure.
We do not show results for LFV effects involving only the first and
second generation. While they are included for completeness in our
analytical results, they are expected to have a negligible effect on the
observables considered here. The latter is confirmed by the numerical
analysis presented in the next subsections.
Applying the most recent limits from the above listed LFV process yield
up-to-date limits on the $\delta^{FAB}_{ij}$~\cite{Arana-Catania:2013nha}.
Using the these upper bounds on $\delta^{FAB}_{ij}$, as given in the
\refta{boundsSpoints}, we calculate the corrections to the Higgs boson masses
and the EWPO. For each explored non-vanishing delta, $\delta^{FAB}_{ij}$, the
corresponding sfermion physical masses and the sfermion rotation matrices, as
well as the EWPO and Higgs masses were numerically
computed with {\tt FeynHiggs\,2.10.2}, where we have included the analytical
results of our calculations.
\subsection{EWPO}
\label{sec:ewpo}
We start with the investigation of the LFV effects on the EWPO.
The experimental bounds on $\delta^{FAB}_{12}$ where $A,B=L,R$ are very
strict (as discussed above, see \refta{boundsSpoints})
and does not yield sizable contribution. The bounds on the other
$\delta^{FAB}_{ij}$'s are relatively less strict but still in most cases we
do not get sizable contributions for EWPO (but now can quantify their
corresponding sizes). The only sizable contribution that we get comes from
$\delta^{LLL}_{23}$. The upper left plot in \reffi{figdLL23} shows our results for
$\Delta\rho$ as functions of $\delta^{LLL}_{23}$, under the presently allowed
experimental range given in \ref{boundsSpoints}, where, depending on the
choice of the scenario (S1 \ldots S6) values of up to \order{10^{-3}} can
be reached. The largest values are found in S5, where the largest values
of $\del{LLL}{23}$ of up to $\pm 0.3$ are permitted. For the same value
of $\del{LLL}{23}$ we find the largest contributions in S6, which
possesses the relatively largest values of SSB
parameters in the slepton sector. This indicates that in general large
contributions to the EWPO are possible as soon as heavy sleptons are
involved. Consequently, while such heavy sleptons are in general
difficult to detect directly at the LHC or the ILC, their presence could
be visible in case of large LFV contributions via a shift in the EWPO.
Turning to the (pseudo-)observables $M_W$ and $\sin^2\theta_{\mathrm{eff}}$, which are shown
in the upper right and middle left plot of \reffi{figdLL23},
respectively, we can
compare the size of the LFV contributions to the current and future
anticipated accuracies in these observables. The black line in both
plots indicates the result for $\del{LLL}{23} = 0$. The red line shows
the current level of accuracy, see \refeq{EWPO-today}, while the blue line
indicates the future ILC/GigaZ precision, see \refeq{EWPO-future}.
We refrain from putting the absolute values of these
observables, since their values strongly depend on the choice of the
stop/sbottom sector (see \citere{PomssmRep} and references therein),
which is independent on the slepton sector under investigation here.
While the current level of accurcay only has the potential to restrict
$\del{LLL}{23}$ in~S5 and~S6, the future accuracy, in particular for $\sin^2\theta_{\mathrm{eff}}$,
can set stringent bounds in all six scenarios.
The overall conclusion for the EWPO is that while $\del{LLL}{23}$ is most
difficult to restrict from ``conventional'' LFV observables, see
\refse{sec:limits}, it has (by far) the strongest impact on EWPO. Even with
the current precision, and even better with the (anticipated) future
accuracies, depending on the values of the scalar top/bottom sector new bounds
beyond the ``conventional'' LFV observables can be obtained.
\subsection{Higgs masses}
\label{sec:Mh}
We now turn to the effects of the LFV contribtions on the prediction of
the neutral ${\cal CP}$-even and the charged MSSM Higgs boson masses. As
discussed in \refse{sec:higgs-ho}, the theoretical accuracy should reach
a precision of $\sim 50 \,\, \mathrm{MeV}$ in the case of $M_h$ and about $\sim 1\%$
in the case of the heavy Higgs bosons. The calculation of $M_h$ in the
presence of NMFV in the scalar quark
sector, as obtained in \citere{arana}, indicated that from the colored
sector corrections of \order{10 \,\, \mathrm{GeV}} are possible (i.e.\ for NMFV
$\delta^{FAB}_{ij}$ in agreement with all other precision data). Similar or even
larger corrections where found for the heavy Higgs bosons, in particular
for the mass of the charged Higgs boson. Large corrections were
connected especially to non-zero values of $\del{ULR,URL}{23}$.
While the corrections from the scalar lepton sector are naturally much smaller
than from the scalar quark sector, it could be expected that the LFV
contributions can exceed future and possibly even current experimental
uncertainties. In the absence of the knowledge of the exact LFV
contributions a theoretical uncertainty had to be assigned at least at
the level of \order{100 \,\, \mathrm{MeV}} for $M_h$ and \order{10 \,\, \mathrm{GeV}} for $M_{H^\pm}$.
Both uncertainties are at the level (or exceeding) the future
anticipated accuracies for these Higgs-boson masses. Consequently, the
LFV have to be evaluated and analyzed in order to reach the required
level of precision.
As described above, the Higgs-boson masses are shown in the middle right
plot ($M_h$), the lower left ($M_H$) and the lower right plot ($M_{H^\pm}$)
in each figure. As expected from the NMFV analysis in the scalar quark
sector~\cite{arana}, the largest effects are found for
$\del{ELR,ERL}{23}$, but similarly for $\del{ELR,ERL}{13}$, indicating that
only the electroweak, but not the Yukawa couplings, play a relevant role
in these corrections. Contrary to the expectations, the corrections to
$M_h$ {\em always} stay below the level of a few~MeV. While this result
eliminates the above menioned uncertainty of \order{100 \,\, \mathrm{MeV}}, these
contributions are too small to yield a sizable numerical effect.
Turning to the heavy Higgs bosons, the contributions to $M_H$, most
sizable again for $\del{ELR,ERL}{23,13}$, do not exceed \order{100 \,\, \mathrm{MeV}}
and are thus effectively negligible. Substantially larger corrections
are found, in agreement with the expectations from \citere{arana} for
the charged Higgs-boson mass. They can reach the level of nearly
$-2 \,\, \mathrm{GeV}$, see \reffis{figdLR13} - \ref{figdRL23}. For the chosen values
of $M_A$ (or $M_{H^\pm}$) this stays below the level of~1\%. However, the
absolute size of the corrections is not connected to the value of $M_{H^\pm}$
in~S1-S6. Choosing starting values of $M_A$ somewhat smaller (requiering
a new evaluation of the corresponding bounds on the LFV $\delta^{FAB}_{ij}$),
could yield relative corrections to $M_{H^\pm}$ at the level
of~1\%. Furthremore, as in the case of the light Higgs-boson mass, the
explicit calculation of the LFV effects eliminates the theory
uncertainty associated to these effects, thus improving the theoretical
accuracy.
\subsection{{\boldmath ${\rm BR}(h \rightarrow l_i^{\pm} l_j^{\mp})$}}
As a last step in MI analysis, we present here the slepton mixing effects to the LFVHD. These decays were calculated using newly modified (see \refse{sec:feynhiggs}) {\tt FeynArts}/{\tt FormCalc}\ setup.
The constraints from cLFV decays on slepton $\delta^{FAB}_{ij}$'s are very tight and we do not expect large values for the BR's. In \reffi{fig:Hetau:Hmutau} we present our numerical results for BR($h \rightarrow e^{\pm} \tau^{\mp} $) and BR($h \rightarrow \mu^{\pm} \tau^{\mp} $) as a function of slepton mixing $\delta^{FAB}_{ij}$'s for the six points defined in the \refta{tab:spectra}. ${\rm BR}(h \rightarrow e^{\pm} \mu^{\mp})$ can only reach \order{10^{-17}} at maximum and we do not show them here. BR($h \rightarrow e^{\pm} \tau^{\mp} $) and BR($h \rightarrow \mu^{\pm} \tau^{\mp} $) can reach at most to \order{10^{-9}} for some parameter points, which is very small compared to the CMS excess \cite{CMSLFVHD}. The reason for such a small value in the experimentally allowed parameter range is the following. The same couplings namely chargino-lepton-slepton and neutralino-lepton-slepton are responsible for the cLFV decays and LFVHD, making it very difficult to find any larger values for LFVHD BR's. Our results show that if the excess shown in the CMS results\cite{CMSLFVHD} persists, we will need to find some other sources of LFV to explain CMS result. Lepton-slepton misalignment is not sufficient to explain this excess. On the other hand our results are in agreement with the ATLAS results \cite{ATLAS-LFVHD} which do not see any excess over SM background.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISLplots/droLL13.eps ,scale=0.53,angle=0,clip=}
\hspace{0.3cm}
\psfig{file=Plots/MISLplots/dmwLL13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinLL13.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MhLL13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHLL13.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MHpLL13.eps ,scale=0.53,angle=0,clip=}\\
\end{center}
\caption{EWPO and Higgs masses as a function of $\delta^{LLL}_{13}$.}
\label{figdLL13}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\hspace{0.3cm}
\psfig{file=Plots/MISLplots/droLL23.eps ,scale=0.53,angle=0,clip=}
\hspace{0.3cm}
\psfig{file=Plots/MISLplots/dmwLL23.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinLL23.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MhLL23.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHLL23.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MHpLL23.eps ,scale=0.53,angle=0,clip=}\\
\end{center}
\caption[EWPO and Higgs masses as a function of $\delta^{LLL}_{23}$.]{EWPO and Higgs masses as a function of $\delta^{LLL}_{23}$.
Solid red (blue) line shows the present (future) experimental uncertainty.}
\label{figdLL23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISLplots/droLR13.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/dmwLR13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinLR13.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MhLR13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHLR13.eps ,scale=0.53,angle=0,clip=}
\hspace{0.1cm}
\psfig{file=Plots/MISLplots/MHpLR13.eps ,scale=0.53,angle=0,clip=}\\
\end{center}
\caption{EWPO and Higgs masses as a function of $\delta^{ELR}_{13}$.}
\label{figdLR13}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISLplots/droLR23.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/dmwLR23.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinLR23.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MhLR23.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHLR23.eps ,scale=0.53,angle=0,clip=}
\hspace{0.1cm}
\psfig{file=Plots/MISLplots/MHpLR23.eps ,scale=0.53,angle=0,clip=}\\
\end{center}
\caption{EWPO and Higgs masses as a function of $\delta^{ELR}_{23}$.}
\label{figdLR23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISLplots/droRL13.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/dmwRL13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinRL13.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MhRL13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHRL13.eps ,scale=0.53,angle=0,clip=}
\hspace{0.1cm}
\psfig{file=Plots/MISLplots/MHpRL13.eps ,scale=0.53,angle=0,clip=}\\
\end{center}
\caption{EWPO and Higgs masses as a function of $\delta^{ERL}_{13}$.}
\label{figdRL13}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISLplots/droRL23.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/dmwRL23.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinRL23.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MhRL23.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHRL23.eps ,scale=0.53,angle=0,clip=}
\hspace{0.1cm}
\psfig{file=Plots/MISLplots/MHpRL23.eps ,scale=0.53,angle=0,clip=}\\
\end{center}
\caption{EWPO and Higgs masses as a function of $\delta^{ERL}_{23}$.}
\label{figdRL23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISLplots/droRR13.eps ,scale=0.53,angle=0,clip=}
\hspace{0.2cm}
\psfig{file=Plots/MISLplots/dmwRR13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinRR13.eps ,scale=0.53,angle=0,clip=}
\psfig{file=Plots/MISLplots/MhRR13.eps ,scale=0.53,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHRR13.eps ,scale=0.53,angle=0,clip=}
\hspace{0.2cm}
\psfig{file=Plots/MISLplots/MHpRR13.eps ,scale=0.53,angle=0,clip=}\\
\end{center}
\caption{EWPO and Higgs masses as a function of $\delta^{ERR}_{13}$.}
\label{figdRR13}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MISLplots/droRR23.eps ,scale=0.53}
\hspace{0.2cm}
\psfig{file=Plots/MISLplots/dmwRR23.eps ,scale=0.53}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/dsinRR23.eps ,scale=0.53}
\psfig{file=Plots/MISLplots/MhRR23.eps ,scale=0.53}\\
\vspace{0.5cm}
\psfig{file=Plots/MISLplots/MHHRR23.eps ,scale=0.53}
\hspace{0.3cm}
\psfig{file=Plots/MISLplots/MHpRR23.eps ,scale=0.53}\\
\end{center}
\caption{EWPO and Higgs masses as a function of $\delta^{ERR}_{23}$.}
\label{figdRR23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/LFVHD/LFVHDLLL13.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/LFVHD/LFVHDELR13.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.7cm}
\psfig{file=Plots/LFVHD/LFVHDERR13.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/LFVHD/LFVHDLLL23.eps ,scale=0.52,angle=0,clip=}\\
\vspace{0.7cm}
\psfig{file=Plots/LFVHD/LFVHDELR23.eps ,scale=0.52,angle=0,clip=}
\psfig{file=Plots/LFVHD/LFVHDERR23.eps ,scale=0.52,angle=0,clip=}\\
\end{center}
\caption[LFV decays $h \rightarrow e \tau $ and $h \rightarrow \mu \tau $ as a function of slepton mixing ]{Lepton flavor violating decays $h \rightarrow e \tau $ and $h \rightarrow \mu \tau $ as a function of slepton
mixing $\delta^{AB}_{ij}$ for the six points defined in the \refta{tab:spectra}.}
\label{fig:Hetau:Hmutau}
\end{figure}
\chapter{Flavor Mixing Effects in MFV CMSSM \& its Seesaw Extension}
After presenting the MI analysis in the previous chapters, here we will investigate the predictions for
off-diagonal sfermion SSB mass terms and flavor mixing effects in
the CMSSM and CMSSM-seesaw~I.
This work is motivated by the fact that in many analyses of the CMSSM, or extensions such as the NUHM1 or NUHM2
(see \citere{AbdusSalam:2011fc} and references therein),
the hypothesis of MFV has been used, and it has been assumed
that the contributions coming
from MFV are negligible not only for FCNC processes but for other observables like EWPO and Higgs masses as well, see, e.g.,
\citere{CMSSM-NUHM}.
In this chapter we will analyze whether this assumption is justified, and
whether including these MFV effects could lead to additional constraints
on the CMSSM parameter space. In this respect we evaluate in the CMSSM
and in the CMSSM-seesaw~I\ the following set of observables:
\begin{itemize}
\item BPO, in particular \ensuremath{{\rm BR}(B \to X_s \gamma)}, \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ and \ensuremath{\Delta M_{B_s}},
\item EWPO, in particular $M_W$ and the
effective weak leptonic mixing angle, $\sin^2\theta_{\mathrm{eff}}$,
\item the masses of
the neutral and charged Higgs bosons in the MSSM,
\item QFVHD in particular \ensuremath{h \to \bar b s + b \bar s},
\item cLFV decays in particular $\mu \to e \gamma$, $\tau \to e \gamma$, $\tau \to \mu \gamma$ as well as
\item LFVHD in particular $h \to e^{\pm} \mu^{\mp}$, $h \to e^{\pm} \tau^{\mp}$ and $h \to \mu^{\pm} \tau^{\mp}$.
\end{itemize}
In order to perform our calculations, we used {\tt SPheno}~\cite{Porod:2003um}
to generate the CMSSM (containing also the type~I seesaw) particle
spectrum by running RGE from the GUT down to the EW scale.
The particle spectrum was handed over in the form of an SLHA
file~\cite{SLHA} to
{\tt FeynHiggs}~\cite{feynhiggs,mhiggslong,mhiggsAEC,mhcMSSMlong,Mh-logresum} to
calculate EWPO and Higgs boson masses. The BPO were
calculated by the {\tt BPHYSICS} subroutine included in the SuFla
code~\cite{sufla} (see also \citeres{arana,arana-NMFV2} for the improved
version used here). QFVHD and LFVHD were calculated using {\tt FeynArts}/{\tt FormCalc}\ setup whereas cLFV decays were calculated with {\tt SPheno 3.2.4}. The following section describes the details of our computational setup.
The results presented in this chapter were published in \cite{MFV-CMSSM}.
\section{Computational setup}
\label{sec:GUTEW}
The SUSY spectra have been generated with the code
{\tt SPheno 3.2.4}~\cite{Porod:2003um} (for the CMSSM and the CMSSM-seesaw~I).
We defined the SLHA~\cite{SLHA} file at the GUT scale.
In a first step within {\tt SPheno}, gauge and
Yukawa couplings at $M_Z$ scale are calculated using tree-level
formulas. Fermion masses, the $Z$~boson pole mass, the fine structure constant
$\alpha$, the Fermi constant $G_F$ and the strong coupling constant
$\alpha_s(M_Z)$ are used as input parameters. The gauge and Yukawa
couplings, calculated at $M_Z$, are then used as
input for the one-loop
RGE's to obtain the corresponding values at the GUT scale
which is calculated from the requirement that $g_1 = g_2$.
The CMSSM boundary
conditions are then applied to the complete set of
two-loop RGE's and are evolved to the EW scale.
At this point the SM and SUSY radiative
corrections are applied to the gauge and Yukawa couplings, and the
two-loop RGE's are again evolved to GUT scale.
After applying the CMSSM boundary conditions again
the two-loop RGE's are run down to EW scale to get SUSY spectrum. This
procedure is iterated until the required precision is achieved.
The output is then written in the form of an SLHA, file which is used as
input to calculate low energy observables discussed below.
For our scans of the CMSSM-seesaw~I\ parameter space we use
{\tt SPheno 3.2.4}~\cite{Porod:2003um} with
the model ``see-saw type-I'' and apply a similar procedure to that in the CMSSM case. The neutrino
related input parameters are
included in the respective SLHA input blocks (see \citere{SLHA} for
details). The predictions for ${\rm BR}(l_i \to l_j \gamma)$
are also obtained with {\tt SPheno 3.2.4}, see the discussion in \refse{sec:Sl}.
We checked that the use of this code produces results similar to the ones obtained by our private codes used
in \citere{Cannoni:2013gq}.
\section{Input parameters}
In order to get an overview about the size of the effects in the CMSSM
parameter space, the relevant parameters $m_0$, $m_{1/2}$ have been
scanned as, or in case of $A_0$ and
$\tan \beta$ have been set to all combinations of
\begin{align}
m_0 &\eq 500 \,\, \mathrm{GeV} \ldots 5000 \,\, \mathrm{GeV}~, \\
m_{1/2} &\eq 1000 \,\, \mathrm{GeV} \ldots 3000 \,\, \mathrm{GeV}~, \\
A_0 &\eq -3000, -2000, -1000, 0 \,\, \mathrm{GeV}~, \\
\tan \beta &\eq 10, 20, 35, 45~,
\end{align}
with $\mu > 0$.
Primarily we are not interested in the absolute values for EWPO BPO and Higgs masses but the
effects that comes from flavor violation within the MFV framework,
i.e.\ the effect from the off-diagonal entries in the sfermion mass matrices.
We first calculate the low-energy observables by setting all $\delta^{FAB}_{ij}=0$
by hand. In a second step we evaluate the observables with the values of
$\delta^{FAB}_{ij}$ obtained through RGE running. We then evaluate the ``pure
MFV effects'',
\begin{align}
\ensuremath{\Delta {\rm BR}^{\rm MFV}(B \to X_s \gamma)} &\eq {\rm BR}(B \to X_s \gamma) - {\rm BR}^{\rm MSSM}(B \to X_s \gamma)~, \\
\ensuremath{\Delta {\rm BR}^{\rm MFV}(B_s \to \mu^+ \mu^-)} &\eq {\rm BR}(B_{s} \to \mu^+ \mu^-)
- {\rm BR}^{\rm MSSM}(B_{s} \to \mu^+ \mu^-)~, \\
\ensuremath{\Delta M_{B_s}^{\rm MFV}} &\eq \Delta M_{B_s} - \Delta M_{B_s}^{\rm MSSM}~,
\end{align}
where ${\rm BR}^{\rm MSSM}(B \to X_s \gamma)$,
${\rm BR}^{\rm MSSM}(B_s \to \mu^+ \mu^-)$ and $\Delta M_{B_{S}}^{\rm MSSM}$
corresponds to the values of relevant observables with all $\delta^{FAB}_{ij} =
0$. Furthermore we define
\begin{align}
\ensuremath{\Delta \Mh^{\rm MFV}} &\eq M_h - M_h^{\rm MSSM} \\
\ensuremath{\Delta \MH^{\rm MFV}} &\eq M_H - M_H^{\rm MSSM} \\
\ensuremath{\Delta \MHp^{\rm MFV}} &\eq M_{H^\pm} - M_{H^\pm}^{\rm MSSM}
\end{align}
where $M_h^{\rm MSSM}$, $M_H^{\rm MSSM}$ and
$M_{H^\pm}^{\rm MSSM}$ corresponds to the
Higgs masses with all $\delta^{FAB}_{ij} = 0$. Similarly we define for the EWPO
\begin{align}
\ensuremath{\Delta\rho^{\rm MFV}} &\eq \Delta\rho-\Delta\rho^{\rm MSSM} \\
\ensuremath{\Delta\MW^{\rm MFV}} &\eq M_W-M_W^{\rm MSSM} \\
\ensuremath{\Delta\sweff^{\rm MFV}} &\eq \sin^2\theta_{\mathrm{eff}}-\sin^2\theta_{\mathrm{eff}}^{\rm MSSM}
\end{align}
where $\Delta\rho^{\rm MSSM}$, $M_W^{\rm MSSM}$ and $\sin^2\theta_{\mathrm{eff}}^{\rm MSSM}$ are the
values of the relavant observables with all $\delta^{FAB}_{ij} = 0$.
\section{Effects of squark mixing in the CMSSM}
\label{sec:Sq}
In this section we analyze the effects from RGE induced flavor violating
mixing in the scalar quark sector in the CMSSM (i.e.\ with no mixing in
the slepton sector). The RGE running from the GUT scale to the EW has
been performed as described in \refse{sec:GUTEW}, with the subsequent
evaluation of the low-energy observables as discussed in
Chap:~\ref{precision-observables}.
In \reffis{fig:DelQLL13}-\ref{fig:Sq-MH-BPO} we show the results of
our CMSSM analysis in the $m_0$--$m_{1/2}$ plane for four
different combinations of $\tan \beta = 10, 45$ (left and right column)
and $A_0 = 0, -3000 \,\, \mathrm{GeV}$ (upper and lower row).
This set represents four ``extreme'' cases of the parameter space and give
an overview about the possible sizes of the effects and their
dependences on $\tan \beta$ and $A_0$ (which we verified with other, not
shown, combinations).
\subsection{Squark \boldmath{$\delta^{FAB}_{ij}$'s}}
We start with the three most relevant $\delta^{FAB}_{ij}$'s. In
\reffis{fig:DelQLL13}-\ref{fig:DelULR23} we show the results for
$\del{QLL}{13}$, $\del{QLL}{23}$ and $\del{ULR}{23}$, respectively,
which are expected to yield the largest results.
The values show the expected pattern of their size with
$\del{QLL}{23} \sim \order{10^{-2}}$ being the largest one, and
$\del{QLL}{13}$ and $\del{ULR}{23}$ about one or two orders of magnitude
smaller. All other $\delta^{FAB}_{ij}$ which are not shown reach only values of
\order{10^{-5}}.
One can observe an interesting pattern in these figures: the
values of $\delta^{FAB}_{ij}$ increase with larger values of either $\tan \beta$ or
$A_0$.
The values for $\del{QLL}{}$ increase with $m_0$, whereas the
$\del{ULR}{}$ and $\del{DLR}{}$ decrease with $m_0$.
This behavior can be understood for the RGE's of the non diagonal SUSY
breaking parameters (see, e.g., \citere{Martin:1993zk}), $\del{QLL}{}$'s are
defined as ratios of off-diagonal soft terms that grow with $m_0^2$ over
diagonal soft masses that also grow with $m_0$. However,
$\del{ULR}{}$'s and $\del{DLR}{}$'s arises from the ratio of the RGE generated
off-diagonal
trilinear terms which depend on the value of $A_0$, that is considered fixed
in our case, over $m_0$ growing diagonal soft masses.
As discussed above, these $\delta^{FAB}_{ij} \neq 0$ are often neglected
in phenomenological analyses of the CMSSM (see, e.g.,
\citere{CMSSM-NUHM}). We also emphasize that these effects are purely due to
the presenece of the CKM matrix on the RGE's, their contribution will vanish
when the mixing of the two first generation with the third generation is
neglected (as we have checked numerically).
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/TB10A0000/DelQLL13.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB10A3000/DelQLL13.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Squarks/TB45A0000/DelQLL13.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/DelQLL13.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of $\delta^{QLL}_{13}$ in the
$m_0$--$m_{1/2}$ plane.]{Contours of $\delta^{QLL}_{13}$ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM.}
\label{fig:DelQLL13}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/TB10A0000/DelQLL23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB10A3000/DelQLL23.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Squarks/TB45A0000/DelQLL23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/DelQLL23.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of $\delta^{QLL}_{23}$ in the
$m_0$--$m_{1/2}$ plane.]{Contours of $\delta^{QLL}_{23}$ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM.}
\label{fig:DelQLL23}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/TB10A0000/DelULR23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB10A3000/DelULR23.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Squarks/TB45A0000/DelULR23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/DelULR23.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of $\delta^{ULR}_{23}$ in the
$m_0$--$m_{1/2}$ plane.]{Contours of $\delta^{ULR}_{23}$ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM.}
\label{fig:DelULR23}
\end{figure}
\subsection{EWPO}
\label{EWPO-CMSSM-Res}
In \reffis{fig:SQ-delrho}-\ref{fig:SQ-delSW2} we analyze the effects of
the non-zero $\delta^{FAB}_{ij}$ on the EWPO \ensuremath{\Delta\rho^{\rm MFV}}, \ensuremath{\Delta\MW^{\rm MFV}}\ and \ensuremath{\Delta\sweff^{\rm MFV}},
respectively.
Here the same pattern is reflected for the EWPO, i.e.\ by increasing
the value of $\tan \beta$ or $A_0$, we find larger contributions to the
EWPO. In particular one can observe a non-decoupling effect for
large values of $m_0$. Larger soft SUSY-breaking
parameters with the non-zero values in particular of $\del{QLL}{23}$,
see above, lead to an enhanced splitting in masses belonging to an
$SU(2)$ doublet, and thus to an enhanced contribution to the
$\rho$-parameter. The corresponding effects on $M_W$ and $\sin^2\theta_{\mathrm{eff}}$,
for $m_0 \gsim 3 \,\, \mathrm{TeV}$, exhibit corrections that are several times
larger than the current experimental accuracy (whereas the SUSY
corrections with all $\delta^{FAB}_{ij} = 0$ decouple and go to
zero). Consequently, including the non-zero values of the $\delta^{FAB}_{ij}$
and correctly taking these corrections into account, would yield an
{\em upper} limit on $m_0$, which in the known analyses so far
is unconstrained from above~\cite{CMSSM-NUHM}.
A more detailed analysis within the CMSSM will be needed to determine
the real upper bound on $m_0$, which, however, is beyond the scope of
this thesis.
In order to gain more insight about the source of the large
corrections to $\Delta\rho$ (and thus to the EWPO), we show in
\reffi{fig:SQ-masses} several relative mass (square) differences,
$(m_2^2 - m_1^2)/(m_2^2 + m_1^2)$ in the
$m_0$--$m_{1/2}$ plane for fixed $A_0 = 0$ and $\tan \beta = 45$.
The left plot shows the mass difference for the two most stop-like
squarks (i.e.\ in the limit of zero inter-generational mixing they
coincide with the two scalar tops). The right plot shows the relative
mass difference for the lightest most stop-like and most sbottom-like
squark. (These results are simply the {\tt Spheno} output in our
scenario.)
In both cases one can see that the relative mass differences
increase (controlled by the non-zero $\delta^{FAB}_{ij}$ induced by the CKM matrix
in the RGE running) in a fashion similar as the $\del{QLL}{}$ discussed
above, i.e.\ in particular for $m_0 > m_{1/2} > 1 \,\, \mathrm{TeV}$. These
increasing mass differences lead (together with
contributions from the mixing matrices) to the observed increase of
$\Delta\rho$ as in \reffi{fig:SQ-delrho}.
Our findings can be briefly compared to the existing
literature. The EWPO in the context of flavor violation were evaluated
first in \citere{delrhoNMFV}, where correspondingly large corrections
were found for large $\del{QLL}{23}$ (in fact, that was the only
parameter dependence analyzed in that paper, and only the mixing between
the second and third generation of squarks was taken into
account). Subsequently, the EWPO were also evaluated for the full
three-generation mixing in \citere{Cao:2006xb}. The numerical analysis,
however, was restricted to a degenerate and fixed SUSY mass scale.
Correspondingly, no large effects with increasing SUSY mass scales were
analyzed and only relative small corrections were found. Due to the
different numerical set-up, however, there is no contradiction with our
results for $\Delta\rho$.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/TB10A0000/deltarho.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB10A3000/deltarho.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Squarks/TB45A0000/deltarho.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/deltarho.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of \ensuremath{\Delta\rho^{\rm MFV}}\ in the $m_0$--$m_{1/2}$ plane.]{Contours of \ensuremath{\Delta\rho^{\rm MFV}}\ in the $m_0$--$m_{1/2}$ plane for different
values of $\tan \beta$ and $A_0$ in the CMSSM.}
\label{fig:SQ-delrho}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/TB10A0000/MWMSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB10A3000/MWMSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Squarks/TB45A0000/MWMSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/MWMSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of \ensuremath{\Delta\MW^{\rm MFV}}\ in GeV in the
$m_0$--$m_{1/2}$ plane]{Contours of \ensuremath{\Delta\MW^{\rm MFV}}\ in GeV in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM.}
\label{fig:SQ-delMW}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/TB10A0000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB10A3000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Squarks/TB45A0000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.8cm}
\end{center}
\caption[Contours of \ensuremath{\Delta\sweff^{\rm MFV}}\ in the
$m_0$--$m_{1/2}$ plane.]{Contours of \ensuremath{\Delta\sweff^{\rm MFV}}\ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM.}
\label{fig:SQ-delSW2}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/DMSupup.eps,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Squarks/DMSupdown.eps,scale=0.51,angle=0,clip=}
\vspace{-0.2cm}
\end{center}
\caption[Contours of $(m_2^2 - m_1^2)/(m_2^2 + m_1^2)$ in the $m_0$--$m_{1/2}$
plane]{Contours of $(m_2^2 - m_1^2)/(m_2^2 + m_1^2)$ in the $m_0$--$m_{1/2}$
plane for fixed values of $A_0 = 0$ and $\tan \beta = 45$. Left: the two most
stop-like squarks (i.e.\ in the limit of zero inter-generational mixing they
coincide with the two scalar tops), right: the lightest most stop-like and most
sbottom-like squarks (see text).}
\label{fig:SQ-masses}
\end{figure}
\subsection{Higgs masses and the BPO}
In \reffi{fig:Sq-MH-BPO} we show the results of
our CMSSM analysis with the effects of
the non-zero $\delta^{FAB}_{ij}$ on the Higgs mass calculations and on the
BPO in the $m_0$--$m_{1/2}$ plane for $\tan \beta = 45$ and
$A_0 = -3000$.
We only show this ``extreme'' case, where smaller values of
$\tan \beta$ and $A_0$ would lead to smaller effects.
In the upper left, upper right and middle left plot we show \ensuremath{\Delta \Mh^{\rm MFV}},
\ensuremath{\Delta \MH^{\rm MFV}}\ and \ensuremath{\Delta \MHp^{\rm MFV}}, respectively. It can be seen that the effects on the
neutral Higgs boson masses are negligible w.r.t.\ the experimental
accuracy. The effects on $M_{H^\pm}$ can reach \order{100 \,\, \mathrm{MeV}},
where largest effects are found for both very small values
of $m_0$ and $m_{1/2}$ (dominated by $\del{ULR}{23}$) or very large values of
$m_0$ and $m_{1/2}$ (dominated by $\del{QLL}{13,23}$). Corrections of up to
$-300 \,\, \mathrm{MeV}$ are found, but still remaining
below the foreseeable future precision. Consequently,
also in the Higgs mass evaluation not taking into account the non-zero
values of the $\delta^{FAB}_{ij}$ is a good approximation.
In the middle right, lower left and lower right plot of
\reffi{fig:Sq-MH-BPO} we show the results for the BPO
\ensuremath{\Delta {\rm BR}^{\rm MFV}(B \to X_s \gamma)}, \ensuremath{\Delta {\rm BR}^{\rm MFV}(B_s \to \mu^+ \mu^-)}\ and \ensuremath{\Delta M_{B_s}^{\rm MFV}},
respectively. The effects in \ensuremath{\Delta {\rm BR}^{\rm MFV}(B \to X_s \gamma)}\ are of \order{-10^{-5}} and thus one
order of magnitude smaller than the experimenal accuracay. Similarly, we
find $\ensuremath{\Delta {\rm BR}^{\rm MFV}(B_s \to \mu^+ \mu^-)} \sim \order{10^{-10}}$ and
$\ensuremath{\Delta M_{B_s}^{\rm MFV}} \sim \order{10^{-15} \,\, \mathrm{GeV}}$, i.e.\ one or several orders of
magnitude below the experimental precision. This shows that for the BPO
neglecting the effects of non-zero $\delta^{FAB}_{ij}$ in the CMSSM is a good
approximation.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Squarks/TB45A3000/Mhh.eps,scale=0.50,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/MHH.eps,scale=0.50,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Squarks/TB45A3000/MHp.eps,scale=0.50,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/Brbsg.eps,scale=0.50,angle=0,clip=}\\
\vspace{0.5cm}
\psfig{file=Plots/Squarks/TB45A3000/Brxlplm.eps,scale=0.50,angle=0,clip=}
\psfig{file=Plots/Squarks/TB45A3000/Brdelmb.eps,scale=0.50,angle=0,clip=}\\
\end{center}
\vspace{-1em}
\caption[Contours of Higgs mass corrections and BPO
in the $m_0$--$m_{1/2}$ plane]{Contours of Higgs mass corrections (\ensuremath{\Delta \Mh^{\rm MFV}},
\ensuremath{\Delta \MH^{\rm MFV}}\ and \ensuremath{\Delta \MHp^{\rm MFV}}\ in GeV) and BPO (\ensuremath{\Delta {\rm BR}^{\rm MFV}(B \to X_s \gamma)}, \ensuremath{\Delta {\rm BR}^{\rm MFV}(B_s \to \mu^+ \mu^-)}\ and \ensuremath{\Delta M_{B_s}^{\rm MFV}})
in the $m_0$--$m_{1/2}$ plane for $\tan \beta = 45$ and \
$A_0 = -3000 \,\, \mathrm{GeV}$ in the CMSSM.}
\label{fig:Sq-MH-BPO}
\vspace{-3em}
\end{figure}
\subsection{\boldmath \ensuremath{\br(\hbs)}}
The results are shown in \reffi{fig:BRhbs}, where we display the
contours of \ensuremath{\br(\hbs)}\ in the ($m_0$, $m_{1/2}$) plane for $\tan \beta=10$,
$A_0=0$ (upper left), $\tan \beta=10$, $A_0=-3000 \,\, \mathrm{GeV}$ (upper right),
$\tan \beta=45$, $A_0=0$ (lower left) and $\tan \beta=45$, $A_0=-3000 \,\, \mathrm{GeV}$ (lower
right). By comparison with planes for other $\tan \beta$-$A_0$ combinations we
have varyfied that these four planes constitute a representative
example. The allowed parameter space can be deduced by comparing to the
results presented above and in \citeres{mc9}. While not all the planes are in
agreement with current constraints, large parts, in particular for
larger values of $m_0$ and $m_{1/2}$ are compatible with a combination
of direct searches, flavor and electroweak precision observables as well
as astrophysical data. Upper bounds on $m_0$ at the few~TeV level
could possibly be set by including the findings of \refse{EWPO-CMSSM-Res}
into a global CMSSM analysis.
In \reffi{fig:BRhbs} one can see that for most of parameter space values
of \order{10^{-7}} are found for \ensuremath{\br(\hbs)}, i.e.\ outside the reach of
current or future collider experiments. Even for the ``most
extreme'' set of parameters we have analyzed, $\tan \beta = 45$ and
$A_0 = -3000 \,\, \mathrm{GeV}$, no detectable rate has been found. Turning the
argument around, any observation of the decay \ensuremath{h \to \bar b s + b \bar s}\ at the (discussed)
future experiments would exclude the CMSSM as a possible model.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MFVBRHBS/TB10A0000.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MFVBRHBS/TB10A3000.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MFVBRHBS/TB45A0000.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MFVBRHBS/TB45A3000.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of ${\rm BR}(h \rightarrow b \bar{s}+\bar{b}s)$ in the
$m_0$--$m_{1/2}$ plane]{Contours of ${\rm BR}(h \rightarrow b \bar{s}+\bar{b}s)$ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM.}
\label{fig:BRhbs}
\end{figure}
\clearpage
\section{Effects of slepton mixing in CMSSM-seesaw~I.}
\label{sec:Sl}
In this section we analyze the effects of non-zero $\delta^{FAB}_{ij}$ values in
the CMSSM-seesaw~I.
In order to investigate the effects induced just by the mixings
in the slepton sector, such that we can compare their contribution from
the one produced by the mixings in the squak sector (and to discriminate
it from effects from mixings in the squark sector) we present here the
results with only $\delta^{FAB}_{ij}$ in
the slepton sector non-zero, i.e.\ after the RGE running with both CKM
and seesaw parameters non-zero, the $\delta^{FAB}_{ij}$ from the squark
sector are set to zero by hand at the EW scale. The effects of the
squark mixing in the CMSSM-seesaw~I\ are nearly
indistinguishable from the ones analyzed in the previous subsection.
As mentioned in \refse{sec:mssmI}, the calculations in this section are
done by using the values of $Y_\nu$ constructed from \refeq{eq:casas} with
degenerate $M_R$'s. The matrix $R$ is set to the identity since it does not
enter in \refeq{eq:ynu2} and therefore the slepton $\delta^{FAB}_{ij}$'s do not depend
on it. The matrix $m_\nu^\delta$ is a diagonal mass matrix adjusted to
reproduce neutrino masses at low energy compatible with the experimental
observations and with hierarchical neutrino masses. We performed our
computation by using the seesaw scale $M_N=10^{14} \,\, \mathrm{GeV}$. With this
choice the bound ${\rm BR}(\mu \to e \gamma) < 5.7 \times 10^{-13}$ ~\cite{Adam:2013mnn}
imposes severe restrictions on the $m_0$--$m_{1/2}$ plane, excluding
values of $m_0$ below 2--3~TeV (depending on $\tan \beta$ and $A_0$). The values of
the slepton $\delta^{FAB}_{ij}$ will increase as the scale $M_N$ increases but also
does the parameter space excluded by the ${\rm BR}(\mu \to e \gamma)$
bound. For example, by increasing $M_N$ by an order of magnitude, the largest
entries in the matrix $Y_\nu$ will become of \order{1} and the bound on
${\rm BR}(\mu \to e \gamma)$ will only be satisfied if $m_0\approx 5 \,\, \mathrm{TeV}$ (see more details below).
\subsection{Slepton \boldmath{$\delta^{FAB}_{ij}$'s}}
Our numerical results in the CMSSM-seesaw~I\ are shown in
\reffis{fig:DelLLL12} - \ref{SL-MH}. As in the CMSSM we present the
results in the $m_0$--$m_{1/2}$ plane for four combinations of
$\tan \beta = 10, 45$ (upper and lower row) and $A_0 = 0, -3000 \,\, \mathrm{GeV}$ (left
and right column), again capturing the ``extreme'' cases.
We start presenting the three most relevant $\delta^{FAB}_{ij}$.
\reffis{fig:DelLLL12}-\ref{fig:DelLLL23} show $\del{LLL}{12}$,
$\del{LLL}{13}$ and $\del{LLL}{23}$, respectively. As expected,
$\del{LLL}{23}$ turns out to be largest of \order{0.01}, while the
other two are about one order of magnitude smaller. The dependence on
$\tan \beta$ is not very prominent, but going from $A_0 = 0$ to $-3000 \,\, \mathrm{GeV}$
has a strong impact on the $\delta^{FAB}_{ij}$. For small $A_0$ the size of the
$\delta^{FAB}_{ij}$ is increasing with larger $m_0$ and $m_{1/2}$, for
$A_0 = -3000 \,\, \mathrm{GeV}$ the largest values are found for small $m_0$ and
$m_{1/2}$.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Sleptons/TB10A0000/DelLLL12.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB10A3000/DelLLL12.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Sleptons/TB45A0000/DelLLL12.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB45A3000/DelLLL12.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of $\delta^{LLL}_{12}$ in the
$m_0$--$m_{1/2}$ plane.]{Contours of $\delta^{LLL}_{12}$ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM-seesaw~I. }
\label{fig:DelLLL12}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Sleptons/TB10A0000/DelLLL13.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB10A3000/DelLLL13.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Sleptons/TB45A0000/DelLLL13.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB45A3000/DelLLL13.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of $\delta^{LLL}_{13}$ in the
$m_0$--$m_{1/2}$ plane.]{Contours of $\delta^{LLL}_{13}$ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM-seesaw~I. }
\label{fig:DelLLL13}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Sleptons/TB10A0000/DelLLL23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB10A3000/DelLLL23.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Sleptons/TB45A0000/DelLLL23.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB45A3000/DelLLL23.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of $\delta^{LLL}_{23}$ in the
$m_0$--$m_{1/2}$ plane.]{Contours of $\delta^{LLL}_{23}$ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM-seesaw~I.}
\label{fig:DelLLL23}
\end{figure}
\subsection{EWPO}
In \reffis{fig:SL-delrho}-\ref{fig:SL-delSW2} we show the results for
the EWPO. The same pattern and non-decoupling behavior for EWPO as
in the case of CMSSM (squark $\delta^{FAB}_{ij}$) can be observed. However, the
corrections induced by slepton flavor violation
are relatively small compared to squark case. For the most extreme
cases, i.e.\ the largest values of $m_0$, the corrections to $M_W$
turn out to be of the same order of the experimental uncertainty.
For those parts of the parameter space neglecting the effects of LFV
to the EWPO could turn out to be an insufficient approximation, in
particular in view of future improved experimental accuracies.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Sleptons/TB10A0000/deltarho.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB10A3000/deltarho.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Sleptons/TB45A0000/deltarho.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB45A3000/deltarho.eps ,scale=0.51,angle=0,clip=}\\
\end{center}
\caption[Contours of \ensuremath{\Delta\rho^{\rm MFV}}\ in the
$m_0$--$m_{1/2}$ plane.]{Contours of \ensuremath{\Delta\rho^{\rm MFV}}\ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM-seesaw~I.}
\label{fig:SL-delrho}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Sleptons/TB10A0000/MWMSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB10A3000/MWMSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Sleptons/TB45A0000/MWMSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB45A3000/MWMSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of \ensuremath{\Delta\MW^{\rm MFV}}\ in GeV in the
$m_0$--$m_{1/2}$ plane.]{Contours of \ensuremath{\Delta\MW^{\rm MFV}}\ in GeV in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM-seesaw~I.}
\label{fig:SL-delMW}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Sleptons/TB10A0000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB10A3000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/Sleptons/TB45A0000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB45A3000/SW2MSSM.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of \ensuremath{\Delta\sweff^{\rm MFV}}\ in the
$m_0$--$m_{1/2}$ plane.]{Contours of \ensuremath{\Delta\sweff^{\rm MFV}}\ in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and
$A_0$ in the CMSSM-seesaw~I. }
\label{fig:SL-delSW2}
\end{figure}
\subsection{Higgs masses}
Finally, in \reffi{SL-MH} we present the corrections to the Higgs
boson masses induced by slepton flavor violation.
Here we only show \ensuremath{\Delta \Mh^{\rm MFV}}\ (left) and \ensuremath{\Delta \MHp^{\rm MFV}} (right) for $\tan \beta = 10$ and
$A_0 = 0$. They turn out to be
negligibly small in both cases. Corrections to \ensuremath{\Delta \MH^{\rm MFV}}, which are not
shown, are even smaller. We have checked that these results hold also
for other combinations of $\tan \beta$ and $A_0$.
Consequently, within the Higgs sector the
approximation of neglecting the effects of the $\delta^{FAB}_{ij}$ is fully
justified.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/Sleptons/TB10A0000/Mhh.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/Sleptons/TB10A0000/MHp.eps ,scale=0.51,angle=0,clip=}
\end{center}
\caption[Contours of \ensuremath{\Delta \Mh^{\rm MFV}}\ (left) and \ensuremath{\Delta \MHp^{\rm MFV}}\ (right)
in the $m_0$--$m_{1/2}$ plane.]{Contours of \ensuremath{\Delta \Mh^{\rm MFV}}\ (left) and \ensuremath{\Delta \MHp^{\rm MFV}}\ (right)
in the $m_0$--$m_{1/2}$ plane for $\tan \beta = 10$ and $A_0 = 0$ in
the CMSSM-seesaw~I.}
\label{SL-MH}
\end{figure}
\subsection{\boldmath{${\rm BR}(l_i \rightarrow l_j \gamma)$}}
The experimental limit BR($\mu \to e \gamma)< 5.7 \times 10^{-13}$ put severe constraints on slepton $\delta^{FAB}_{ij}$'s as discussed before.
In \reffi{fig:BrmegSSI}, we show the predictions for BR($\mu \to e \gamma$)
in $m_0$--$m_{1/2}$ for different values of $A_0$ and $\tan \beta$ in CMSSM-seesaw~I.
The selected values of $Y_\nu$ result in a large prediction for, e.g.,
BR($\mu \to e \gamma$) that can eliminate some of the $m_0$--$m_{1/2}$
parameter plane, in particular combinations of low values of $m_0$ and
$m_{1/2}$. For $\tan \beta=10$ and $A_0=0$, BR($\mu \to e \gamma$) (upper left plot of \reffi{fig:BrmegSSI}) do not exclude any region in $m_0$--$m_{1/2}$ plane, whereas with $\tan \beta=10$ and $A_0=-3000$ lower left region below $m_0, m_{1/2}=2000 $ is excluded (see upper right plot of \reffi{fig:BrmegSSI}). For combinations like $\tan \beta=45,$ $A_0=0$ and $\tan \beta=45,$ $A_0=-3000$ even larger parts of the plane are excluded by BR($\mu \to e \gamma$).
In \reffi{fig:BrtegSSI} and \reffi{fig:BrtmgSSI}, we show the predictions for BR($\tau \to e \gamma$) and BR($\tau \to \mu \gamma$) respectively. It can be seen that these processes do not reach their respective experimental bounds ${\rm BR}(\tau \rightarrow e \gamma)< 3.3 \times 10^{-8}$, ${\rm BR}(\tau \rightarrow \mu \gamma)< 4.4 \times 10^{-8}$. Consequently they do not exclude any parameter space.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MSSMI/TB10A0000/Brmeg.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB10A3000/Brmeg.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MSSMI/TB45A0000/Brmeg.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB45A3000/Brmeg.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of BR($\mu \to e \gamma$) in the
$m_0$--$m_{1/2}$ plane]{Contours of BR($\mu \to e \gamma$) in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM-seesaw~I.}
\label{fig:BrmegSSI}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MSSMI/TB10A0000/Brteg.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB10A3000/Brteg.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MSSMI/TB45A0000/Brteg.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB45A3000/Brteg.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of BR($\tau \to e \gamma$) in the
$m_0$--$m_{1/2}$ plane]{Contours of BR($\tau \to e \gamma$) in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM-seesaw~I.}
\label{fig:BrtegSSI}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MSSMI/TB10A0000/Brtmg.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB10A3000/Brtmg.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MSSMI/TB45A0000/Brtmg.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB45A3000/Brtmg.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of BR($\tau \to \mu \gamma$) in the
$m_0$--$m_{1/2}$ plane]{Contours of BR($\tau \to \mu \gamma$) in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM-seesaw~I.}
\label{fig:BrtmgSSI}
\end{figure}
\subsection{\boldmath${\rm BR}(h \rightarrow l_i^{\pm} l_j^{\mp}$)}
\reffi{fig:HMueESSI} shows the results for BR($h \to e \mu$). The largest value is of the \order{10^{-16}} for low $m_0$ and $m_{1/2}$ values but is excluded from BR($\mu \to e \gamma$). In the allowed range they are typically \order{10^{-18}}. Similarly \reffi{fig:HTauESSI} and \reffi{fig:HTauMueSSI} shows the predictions for BR($h \to e \tau$) and BR($h \to \tau \mu$) respectively. Predictions of the \order{10^{-14}} and \order{10^{-12}} are possible for BR($h \to e \tau$) and BR($h \to \tau \mu$) in the lower left region of the $m_0$--$m_{1/2}$ plane respectively but are excluded from BR($\mu \to e \gamma$) bound. In the allowed region they are of the \order{10^{-16}} or less. These results are in a clear contradiction to the recently reported CMS excess\cite{CMSLFVHD}. If this excess seen in the CMS is confirmed in the future analysis, we will need models other than the CMSSM-seesaw~I\ to explain this excess. However our findings are in agreement with the ATLAS reports\cite{ATLAS-LFVHD}, where they do not see any significant excess over background. It remains to be seen how these results will develop with the LHC Run II.
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MSSMI/TB10A0000/HMueE.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB10A3000/HMueE.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MSSMI/TB45A0000/HMueE.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB45A3000/HMueE.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of BR($h \to e \mu$) in the
$m_0$--$m_{1/2}$ plane]{Contours of BR($h \to e \mu$) in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM-seesaw~I.}
\label{fig:HMueESSI}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MSSMI/TB10A0000/HTauE.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB10A3000/HTauE.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MSSMI/TB45A0000/HTauE.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB45A3000/HTauE.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of BR($h \to e \tau$) in the
$m_0$--$m_{1/2}$ plane]{Contours of BR($h \to e \tau$) in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM-seesaw~I.}
\label{fig:HTauESSI}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\psfig{file=Plots/MSSMI/TB10A0000/HTauMue.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB10A3000/HTauMue.eps ,scale=0.51,angle=0,clip=}\\
\vspace{0.2cm}
\psfig{file=Plots/MSSMI/TB45A0000/HTauMue.eps ,scale=0.51,angle=0,clip=}
\psfig{file=Plots/MSSMI/TB45A3000/HTauMue.eps ,scale=0.51,angle=0,clip=}\\
\vspace{-0.2cm}
\end{center}
\caption[Contours of BR($h \to \tau \mu$) in the
$m_0$--$m_{1/2}$ plane]{Contours of BR($h \to \tau \mu$) in the
$m_0$--$m_{1/2}$ plane for different values of $\tan \beta$ and $A_0$ in
the CMSSM-seesaw~I.}
\label{fig:HTauMueSSI}
\end{figure}
\chapter{Summary \& Conclusions}
\label{sec:conclusions}
SUSY proves to be a very powerful and technically well equiped theory as it successfully explains some of the major deficiencies
of the SM, but still lacks experimental endorsement. Direct searches for
sparticles at LHC did not succeed so far.
The other way around is to probe SUSY via virtual effects of additional
particles to the precision observables.
For example, in the MSSM, the fermion-sfermion misallignment can generate flavor changing effects that can dominate the SM effects by several orders of magnitude. Any possible experimental deviation from the SM results for the precision observales could be a hint of SUSY. Also, as this misallignment arises from the soft SUSY-breaking terms, this may provide guidlines for the SUSY model building. In this thesis, keeping the above mentioned points in mind, we studied the possible phenomenological consequences of flavor mixing to various observables.
The flavor mixing was parameterized in terms of a set of dimensionless parameters $\delta^{FAB}_{ij}$ ($F=Q,U,D,L,E; A,B =
L,R$; $i,j = 1,2,3$). In chapter 1, we reviewed some aspacts of the SM, similarly in chapter 2, a general introduction to MSSM and its seesaw extension was discussed. Calculational details for the considered observables were given in chapter 3 where we presented the higher order corrections to the electroweak precision observables (EWPO), higher order corrections to Higgs boson masses, calculational details of the $B$-physics observables (BPO), quark flavor violating Higgs decays (QFVHD) and lepton flavor violating Higgs decays (LFVHD). In order to calculate slepton mixing effects (squark mixing was already present), we prepared an add-on model file for {\tt FeynArts}\ to include lepton flavor violation in the already existing MSSM model file of the {\tt FeynArts}. {\tt FormCalc}\ drivers were also modified accordingly. The inclusion of lepton flavor violation (LFV) into {\tt FeynArts}/{\tt FormCalc}\ allowed us to
calculate the one-loop LFV effects on EWPO
(via the calculation of gauge-boson self-energies) as well on the
Higgs-boson masses of the MSSM (via the calculation of the Higgs-boson
self-energies). The corresponding results have been included in the
code {\tt FeynHiggs}\ and are publicly available from version 2.10.2 on.
We have (re-)caculated the decay \ensuremath{h \to \bar b s + b \bar s}\ in the
{\tt FeynArts}\ and {\tt FormCalc}\ setup. The BPO and EWPO constraints have been evalated
with the help of (a private version of) {\tt FeynHiggs}, taking into account the
full flavor violating one-loop corrections to $M_W$ and to the relevant
BPO (supplemented with further MSSM higher-order
corrections).
The effects of squarks mixing to EWPO, BPO and QFVHD such as $\ensuremath{h \to \bar b s + b \bar s}$ in the Model Independent (MI) approach were presented in chapter 4. This evaluation improved on existing analyses in various ways. We took into account the full set of SUSY QCD and SUSY EW corrections,
allowing for LL, RL, LR and RR mixing simultaneously. The parameter
space was restricted not only by BPO, but also by
EWPO, in particular the mass of the
$W$~boson. We have shown that $M_W$ can
yield non-trivial, additional restrictions on the parameter space of the
squark flavor violating $\delta^{FAB}_{ij}$.
In six representative scenarios, which are allowed by current
searches for SUSY particles and heavy Higgs bosons, we have evaluated
the allowed parameter space for the various $\delta^{FAB}_{ij}$ by applying BPO
and EWPO constraints. Within these allowed ranges we have then evaluated
\ensuremath{\br(\hbs)}. In the case of only one $\delta^{FAB}_{ij} \neq 0$ we have found that
only relatively large values of $\del{DLR}{23}$ could lead to rates of
$\ensuremath{\br(\hbs)} \sim 10^{-4}$, which could be in the detectable range of future
$e^+e^-$ colliders. Allowing two $\delta^{FAB}_{ij} \neq 0$
simultaneously lead to larger values up to $\ensuremath{\br(\hbs)} \sim 10^{-3}$,
which would make the observation at the ILC relatively easy. Allowing
for a third $\delta^{FAB}_{ij} \neq 0$, on the other hand, did not lead to
larger values of the flavor violating branching ratio.
The effects of slepton mixing to EWPO, Higgs boson masses and LFVHD in the MI approach were presented in chapter 5.
The numerical analysis was performed on the basis of same six benchmark
points as in the previous chapter. These benchmark
points represent different combinations of parameters in the sfermion
sector. The restrictions on the various slepton $\delta^{FAB}_{ij}$ in these six
scenarios, provided by experimental limits on LFV processes (such as
$\mu\to e\gamma$) were taken from \citere{Arana-Catania:2013nha}, and
the effects on EWPO and Higgs-boson masses were evaluated in the
experimentally allowed ranges. In this way we were able to provide a general
overview about the possible size of LFV effects and potential new
restrictions on the slepton $\delta^{FAB}_{ij}$ from EWPO and Higgs-boson masses.
The LFV effects in the EWPO turned out to be sizable for $\del{LLL}{23}$
but (at least in the scenarios under investigation) negligible for the
other $\delta^{FAB}_{ij}$. The effects of varying $\del{LLL}{23}$ in the
experimentally allowed ranges turned out to exceed the current
experimental uncertainties of $M_W$ and $\sin^2\theta_{\mathrm{eff}}$ in the case of heavy
sleptons. No new general bounds could be set on $\del{LLL}{23}$,
however, since the absolute values of $M_W$ and $\sin^2\theta_{\mathrm{eff}}$ strongly depend
on the choices in the stop/sbottom sector, which is disconnected from
the slepton sector presently under investigation. Such bounds could be
set on a point-by-point basis in the LFV MSSM parameter space, however.
Looking at the future anticipated accuracies, also lighter sleptons
yielded contributions exceeding that precision. It may therefore be
possible in the future to set bounds on $\del{LLL}{23}$ from EWPO that
are stronger than from direct LFV processes.
In the Higgs sector, based on evaluations for flavor violation in the
squark sector, non-negligible corrections to the light ${\cal CP}$-even Higgs
mass as well as to the charged Higgs-boson mass could be expected. The
associated theoretical uncertainties exceeded the anticipated future
precision for $M_h$ and $M_{H^\pm}$. Taking the existing limits on the
$\delta^{FAB}_{ij}$ from LFV processes into account, however, the corrections
mostly turned out to be small. For the light ${\cal CP}$-even Higgs mass they
stay at the few-MeV level. For the charged Higgs boson mass they can reach
\order{2 \,\, \mathrm{GeV}}, which, depending on the choice of the heavy Higgs-boson
mass scale, could be at the level of the future experimental precision.
More importantly, the theoretical uncertainty from LFV effects that
previously existed for the evaluation of the MSSM Higgs-boson masses,
has been reduced below the level of future experimental accuracy.
The predictions for the LFVHD in the MI approach were also presented in chapter 5. However due to very tight constraints on the slepton $\delta^{FAB}_{ij}$'s from cLFV decays, the BR's for these processes turned out to be very small.
Effects of squark mixing in the the CMSSM and slepton mixing in CMSSM extended by type I Seesaw under the
Minimal Flavor Violation (MFV) hypothesis were presented in chaptor 6. This work was motivated by the fact that in many
phenomenological analyses of the CMSSM the effects of intergenerational
mixing in the squark and/or slepton sector are neglected. However, such
mixings are naturally induced, assuming no flavor violation at the GUT
scale, by the RGE running from the GUT to the EW scale exactly due to the
presence of the CKM and/or the PMNS matrix.
In this sense the CMSSM and the CMSSM-seesaw~I\ represent two simple
``realistic'' GUT based models, in which flavor violation in induced solely by
RGE running.
The spectra of the CMSSM and CMSSM-seesaw~I\ have been numerically evaluated
with the help of the program {\tt SPheno}
by taking the GUT scale input run down via the appropriate
RGEs to the EW scale.
We have evaluated the predictions for BPO, MSSM
Higgs boson masses, EWPO in the CMSSM and
CMSSM-seesaw~I.
In order to numerically analyze the effects of neglecting
intergenerational mixing
these observables have been evaluated with the full spectrum at the EW
scale, as well as with the spectrum, but with all intergenerational
mixing set artificially to zero (as it has been done in
many phenomenological analyses).
The difference in the various observables
indicates the possible size of the effects neglected in those
analyses. In this way it can be checked whether neglecting those mixing
effects is a justified approximation.
Within the CMSSM we have taken a fixed grid of $A_0$ and $\tan \beta$,
while scanning the $m_0$--$m_{1/2}$ plane. We found that the value of
$\delta^{FAB}_{ij}$ increases with the increase of the $A_0$ or $\tan \beta$ values.
The Higgs boson masses receive corrections below current and future
experimental uncertainties,
where the shifts in $M_{H^\pm}$ were found largest at the level of
\order{100 \,\, \mathrm{MeV}}. Similarly for the BPO the induced
effects are at least one order of magnitude smaller than the current
experimental uncertainty. For those two groups of observables the
approximation of neglecting intergenerational mixing explicitly is a
viable option.
The picture changes for the EWPO.
We find that the masses of the squarks grow with $m_0$, and thus do the mixing terms,
inducing a splitting between masses in an $SU(2)$ doublet, leading
to a non-decoupling effect.
For $m_0 \gsim 3 \,\, \mathrm{TeV}$ the effects induced in $M_W$ and $\sin^2\theta_{\mathrm{eff}}$ are
found to be several times larger than the current experimental
uncertainties and could shift the CMSSM prediction outside the allowed
experimental range. In this way, taking the intergenerational mixing
into account could in principle set bounds
on $m_0$ that are not present in recent phenomenological analyses.
By investigating numerically squark mass differences, we
have shown that this behavior can be traced back to the non-decoupling
effects in the scalar quark mass matrices, provided by {\tt Spheno} when
taking into account the CKM matrix in the RGE running.
However, we would like to point out that this bound only holds
because of the particularly simple structure of the CMSSM and cannot
be extended easily to other, more complicated model frameworks.
In the final step of the numerical analysis within the CMSSM we have evaluated \ensuremath{\br(\hbs)}. Here we
have found that for most of parameter space values
of \order{10^{-7}} are found for \ensuremath{\br(\hbs)}, i.e.\ outside the reach of
current or future collider experiments.
Going to the CMSSM-seesaw~I\ the numerical results depend on the
concrete model definition. We have chosen a set of parameter that
reproduces correctly the observed neutrino data and simultaneously
induces large LFV effects and induces {\em relatively} large corrections
to the calculated observables. Consequently, parts of the parameter
space are excluded by the experimental bounds on ${\rm BR}(\mu \to e \gamma)$. However ${\rm BR}(\tau \to e \gamma)$ and ${\rm BR}(\tau \to \mu \gamma)$ do not reach to their respective experimental limits. Again predictions for the BR of LFVHD turned out very small in CMSSM-seesaw~I. We can conclude that we will need models other than the CMSSM-seesaw~I\ to explain the CMS excess (if it persists) for the channel ${\rm BR}(h \to \mu \tau)$.
Concerning the precision observables we find that
BPO are not affected, we also find that the
additional effects induced by slepton flavor violation on Higgs boson
masses are negligible. Again the EWPO are found to show the largest
impact, where for $M_W$ effects at the same level as the
current experimental accuracy have been observed for very large values
of $m_0$. As above, we would like to point out that these effects are due to
the relatively simple structure of the CMSSM-seesaw~I.
To summarize our MFV analysis: we have analyzed two ``realistic'' GUT based models in which
flavor violation is solely induced by the CKM matrix via RGE running (as
evaluated using the {\tt Spheno code}). We find that
artificially setting all flavor violating terms to zero in
the CMSSM and CMSSM-seesaw~I\ is an acceptable approximation for BPO, Higgs boson masses (evaluated using a private version of
{\tt FeynHiggs}). However, in the EWPO (also evaluated with {\tt FeynHiggs})
in our numerical analysis we find larger effects in the CMSSM and CMSSM-seesaw~I.
The numerical contributions are larger than the current
experimental accuracy in $M_W$ and $\sin^2\theta_{\mathrm{eff}}$. Taking those effects
correctly into account could in principle place new bounds on $m_0$ that
are not present in recent phenomenological analyses.
\chapter*{Introducci\'on}
\addcontentsline{toc}{chapter}{Introducci\'on}
El Modelo Est\'andar (ME) de la f\'{\i}sica de part\'{\i}culas \cite{Glashow, Weinberg, Salam}, fruto de un inmenso esfuerzo tanto te\'orico como experimental, muestra la naturaleza de los ingredientes que forman nuestro universo y c\'omo interact\'uan entre s\'{\i}. Seg\'un el ME, nuestro universo se compone de fermiones (part\'{\i}culas de spin 1/2), de los que seis son leptones y otros seis quarks, contenidos en tres familias. A cada fermi\'on le corresponde una anti-part\'{\i}cula con n\'umeros cu\'anticos opuestos. Las part\'{\i}culas asociadas con los campos de interacci\'on son bosones (part\'{\i}culas de spin 1); los fotones $(\gamma )$ y los bosones l ($W^{\pm }$ y $Z$) se asocian a la interacci\'on electrod\'ebil, los gluones ($g$) a la fuerte. La gravedad no es parte del ME. Las simetr\'{\i}as y los principios de invariancia determinan la forma de estas fuerzas, el ME se basa en el grupo gauge $SU(3)_{\rm C}\ \times SU(2)_{\rm L}\times U(1)_{\rm Y}$. La renormalizabilidad e invariancia gauge exigen que la simetr\'{\i}a $SU(2)_{\rm L}\times U(1)_{\rm Y}$ se rompa espont\'aneamente mediante el llamado mecanismo de Higgs. Todas las predicciones establecidas por ME se han confirmado experimentalmente. El descubrimiento de la \'ultima pieza que faltaba por conocer, el bos\'on de Higgs, se anunci\'o el 4 de julio de 2012 en el gran Colisionador de Hadrones (LHC) del Conseil Europ\'een pour la Recherche Nucleaire (CERN)\cite{ATLAS:2013mma,CMS:yva}. De este modo, el ME es la teor\'{\i}a m\'as precisa y elegante en la actualidad. Sin embargo, a pesar su \'exito, hay buenas razones tanto te\'oricas como experimentales que nos llevan m\'as all\'a del ME. De modo que puede pensarse que ME es un caso l\'{\i}mite de una teor\'{\i}a m\'as general.
El primer problema no explicado por el ME est\'a relacionado con el sector de los neutrinos, considerados sin masa en el ME. En varios experimentos se ha observado la desaparici\'on del neutrinos electr\'onico o mu\'onico. Esto ha aportado la evidencia suficiente para aceptar su oscilaci\'on de sabor \cite{Neutrino-Osc}. Esta observaci\'on ha confirmado que los neutrinos tienen masas diferentes y que los tres sabores de neutrinos $\nu _{e},$ $\nu _{\mu }$ y $\nu _{\tau}$ se mezclan entre a s\'{\i} para formar tres estados propios de masa. Esto implica la no conservaci\'on del sabor lept\'onico, por lo tanto, la predicci\'on de procesos de violacion de sabor de leptones cagados (cLFV), como ocurre en el sector de los quarks.
La extensi\'on más simple del ME para acomodar las masas de los neutrinos consiste en introducir tres singletes fermi\'onicos $ SU(3)_{\rm C}\times SU(2)_{\rm L}\times U(1)_{\rm Y}$ y acoplarlos a neutrinos mediante interacciones Yukawa, las masas de los neutrinos se generar\'{\i}an a trav\'es de ruptura de simetr\'{\i}a electrod\'ebil (EWSB). Sin embargo esta extensi\'on del ME requiere accoplamientos de Yukawa extremadamente peque\~nos y violar el n\'umero lept\'onico a baja energ\'{\i}a. Como ni los neutrinos dextr\'ogiros ni la violaci\'on del n\'umero lept\'onico se han observado a esa escala, es preciso buscar un mecanismo que explique las masas de los neutrinos respetando ambas evidencias. Una de las soluciones al problema es incluir en la teor\'{\i}a un mecanismo ``see-saw'' (mecanismo de balanc\'{\i}n) \cite{seesaw:I} para generar las masas de los neutrinos, el cual no s\'olo permite los acoplamientos de Yukawa, sino que adem\'as explica porque los neutrinos lev\'ogiros son m\'as ligeros que los otros fermiones de ME. Estos mecanismos asumen la existencia de neutrinos my pesados del tipo de Majorana, los cuales se acoplan a los del ME mediante interacciones Yukawa. Las masas de los neutrinos son generadas por un operador efectivo de dimension 5. Esto da lugar a estados f\'{\i}sicos de neutrinos que mezclan el sabor y en consecuencia predicen violaci\'on del sabor leptones (LFV).
Por otro lado ME tampoco explica suficientemente el sector de Higgs. Aunque el ME es renormalizable, se cree que es v\'alido s\'olo hasta cierta escala de energ\'{\i}a, la cual est\'a realcionada con la aparici\'on de f\'{\i}sica desconocida. Si esta escala se asocia con la integraci\'on de la gravedad en teor\'{\i}a, deber\'{\i}a estar en torno a la masa de Planck (10$^{19}$ GeV). De este modo, las correcciones a la masa del Higgs $M_H$ debidas a los fermiones ser\'{\i}an:
\begin{equation}
\delta M_{H}^{2}(f)=- \frac{\left\vert \lambda_f \right\vert^2}{8 \pi^2}[\Lambda^2+.....],
\end{equation}%
donde $\lambda_f$ representa el acoplamiento del fermi\'on $f$ al campo de Higgs y $\Lambda$ es el corte ultravioleta utilizado para regular el integral. Si \'este se reemplaza por la masa de Planck se obtiene $\delta M_{H}^{2} \approx 10^{30} \,\, \mathrm{GeV}^2$. Esta enorme correcci\'on se podr\'{\i}a cancelar con una masa original del mismo orden y signo opuesto. Sin embargo, estas dos contribuciones se deber\'{\i}an cancelar entre s\'{\i} con una precisi\'on de una parte en $10^{26}$ para explicar la masa del Higgs observada experimentalmete. Este es el llamado ``problema de la jerarqu\'{\i}a''.
El tercer problema que el ME no explica es el de la materia oscura (MO). Las primeras especulaciones sobre la existencia de MO se debieron al astr\'onomo Zwicky. En 1933, observ\'o que la masa total de la materia luminosa procedente del c\'umulo de Coma es mucho menor que la masa total podemos suponer por movimiento de las galaxias que lo integran \cite{Zwicky}. En la actualidad hay diversas muestras de la presencia de MO en nuestro universo. El efecto de lente gravitatoria y las curvas de rotaci\'on de las galaxias espirales son observaciones que apuntan a la existencia de la llamada ``masa perdida'' en el universo. Resultados recientes de los expeimentos WMAP \cite{WMAP} y Planck \cite{Planck} proporcionan un valor preciso de la masa del universo y la forma en que se divide entre los diferentes tipos de materia y energ\'{\i}a. No hay ninguna part\'{\i}cula del ME que pueda servir como candidato MO.
Por las razones anteriores hay que encontrar teor\'{\i}as renormalizables que puedan eliminar divergencias cuadr\'aticas en la masa del bos\'on de Higgs, proporcionar un candidato a MO y explicar el la violaci\'on del sabor lept\'onico. Una de \'estas, la extensi\'on supersim\'etrica del ME, cuya versi\'on m\'as simple es el {\it Minimal Supersymmetric Standard Model (MSSM)} \cite{mssm}, puede hacer frente a los problemas mencionados. El MSSM predice la existencia de una pareja para cada una de las part\'{\i}culas fundamentales de la ME a las que se atribuye un esp\'{\i}n que se diferencia en media unidad de sus compa\~neros del ME. La presencia de las super-part\'{\i}culas contribuye a cancelar las divergencias cuadr\'aticas en el bos\'on masa de Higgs. Tambi\'en la part\'{\i}cula supersim\'etrica m\'as ligera (LSP) puede ser un candidato a MO. Sin embargo, el MSSM, como el ME, asumen neutrinos sin masa, por lo que el MSSM tiene que ser ampliado para que sea consistente con las observaciones de las mezclas de sabor de \'estos \cite{Neutrino-Osc}.
Gran parte del esfuerzo del LHC se ha dedicado a descubrir la supersimetr\'{\i}a (SUSY), pero hasta el momento ninguna part\'{\i}cula SUSY se ha observado\cite{Atlas:SUSY2015,CMS:SUSY2015}. Otro enfoque para descubrir SUSY procede del estudio de los efectos indirectos de las part\'{\i}culas SUSY en otros observables \cite{PomssmRep}. Las mezclas de sabor ofrecen una perspectiva \'unica en este sentido, ya que la mayor parte de los efectos indirectos de las part\'{\i}culas SUSY proceden de observaciones en los que \'estas se producen. La primera de ellas, es el proceso de cambio de sabor en el sector de los quarks en corrientes neutras (FCNC). En ME, los procesos del tipo FCNC est\'an ausentes a nivel de \'arbol y s\'olo pueden ocurrir en nivel de un bucle. La \'unica fuente de FCNC de en el ME es la matriz de Cabibbo, Kobayashi y Maskawa (CKM), sin embargo no es significativa debido a la cancelación entre las diversas contribuciones (mecanismo GIM ). Por otro lado, en el MSSM, la posible desalineaci\'on entre las matrices de masa de los quarks y sus parejas supersim\'eticas (squarks) es otra fuente de violaci\'on de sabor, capaz de superar a la contribuci\'on ME en varios \'ordenes de magnitud. Cualquier posible desviaci\'on experimental de la predicci\'on de ME para la FCNS ser\'{\i}a una evidencia clara de nueva f\'{\i}sica y, posiblemente, un indicio del MSSM. Del mismo modo, las predicciones del MSSM para cLFV son cero. Incluso las extensiones del tipo ``see-saw'' del ME no predicen
tasas considerables para estos procesos. Las tasas cLFV en esta extensi\'on del ME son casi 40 \'ordenes de magnitud menor que las actuales l\'{\i}mites experimentales y, por consiguiente sin posibilidad de ser observadas. En cambio, la extensi\'on ``see-saw'' del MSSM, predice valores m\'as altos, cercanos a los l\'{\i}mites de observaci\'on actuales. Por otro lado, tras el descubrimiento del bos\'on de Higgs con una masa en torno a los 125 GeV, es preciso incorporar correcciones radiativas grandes para su explicaci\'on. Una masa superior a 1 TeV de la pareja supersim\'etrica del top, el s-top, podr\'{\i}a dar respuesta al problema, pero a costa de reintroducir un ajuste innatural de los par\'ametros. Sin embargo, esta inconveniencia puede evitarse con una mezcla fuerte entre las compontes quirales del stop o mediante una mezcla de sabor ente los s-quarks.
La forma m\'as general de introducir mezcla de sabor en el MSSM es a trav\'es de los par\'ametros que rompen la supersimetr\'{\i}a. Estos par\'ametros, dotan de masas moderamente grandes a las part\'{\i}culas supersim\'etricas. De este modo, no es posible con una \'unica rotaci\'on en el espacio del sabor diagonalizar simult\'aneamente las masas de los fermiones y las de sus correspondientes parejas supersim\'etricas.
Esta desalineaci\'on puede producirse por varias causas; un ejemplo son las ecuaciones de renormalizaci\'on (RGE): Aun partiendo de masas supersim\'etricas sin violaci\'on de sabor a una alta energ\'{\i}a, las RGE pueden inducirla debido a que contienen acoplamientos Yukawa no diagonales. Este tipo de enfoque es conocido en la literatura como la violaci\'on de sabor m\'{\i}nima (MFV) ~\cite{MFV1,MFV2}, donde se supone que el sabor y la violaci\'on de la simetr\'{\i}a CP en el sector de quarks se describe en su totalidad por la matriz CKM.
Los escenerios del tipo MFV est\'an bien motivados por el hecho de que no introducen nuevas fuentes de violaci\'on de sabor y de CP. Los cuales entrar\'{\i}an en conflicto con los l\'{\i}mites experimentales en los sectores de los kaones y $B_d$, descritos por el ME con una precisi\'on del\% ~ 10 \cite{HFAgroup}. Para la primera y segunda generaci\'on squarks, sensibles a los datos de $K^0 -\bar{K}^0$ y $D^0 -\bar{D}^0$, las restricciones son muy fuertes. Sin embargo, sistemas con la tercera generaci\'on est\'an menos limitados, ya que los datos de la mezcla $B^0 -\bar{B}^0$ a\'un dejan lugar para nuevas fuentes de violaci\'on sabor. Esto abre la posibilidad para scenerios m\'as generales como los de violaci\'on no m\'{\i}nima de sabor (NMFV), aparte de los de MFV.
En esta tesis, se presenta un estudio sistem\'atico y simult\'aneo de los efectos de la mezcla de sabor en diferentes observables utilizando el MFV y el NMFV. Como un primer paso estudiaremos las mezclas de squarks y sleptones en el MSSM a baja energ\'{\i}a, sin utilizar un modelo espec\'{\i}fico (MI). Para el enfoque MI, introducimos arbitrariamente los par\'ametros de mezcla de sabor en las matrices de masa de los sfermiones, sin tener en cuenta el origen de estos par\'ametros. Estudiamos los efectos de la mezcla de los squark en los observables de la interacci\'on electrod\'ebiles medidos con gran precisi\'on (EWPO), la f\'{\i}sica del quark $b$ (BPO) y las desintegraciones del bos\'on de Higgs que violan el sabor. Para la mezcla del sabor lept\'onico, estudiamos tambi\'en los efectos sobre EWPO, la masa del los bosones de Higgs y las desintegraciones de \'estos que violan el sabor lept\'onico (LFVHD). En segundo lugar, extendemos nuestro an\'alisis a la fuente de la mezcla de sabor. Para ello analizamos la mezcla de sabor inducida por las RGE en la evoulci\'on de los par\'ametros desde las escalas GUT y electrod\'ebil. En este estudio trabajamos con la hip\'otesis de MFV tanto para squarks como para sleptones. Por consiguiente, vamos a investigar dos modelos (en los siguientes cap\'{\i}tulos se introduciran m\'as definiciones y citas):
\begin{itemize}
\item[(i)]
El modelo supersim\'etrico m\'{\i}nimo con rotura de la supersimetr\'{\i}a mediante
par\'ame- tros universales (CMSSM). En este caso solo hay violaci\'on
de sabor en los squarks.
\item[(ii)]
El modelo CMSSM ampliado mediante un mecanismo ``see-saw'' de tipo I \cite{seesaw:I},
llamado ``CMSSM-seesaw~I''
\end{itemize}
En muchos an\'alisis del CMSSM o sus extensiones como el NUHM1 o NUHM2 (v\'ease la Ref. \cite{AbdusSalam:2011fc} y las referencias en \'el), se ha utilizado la hip\'otesis de MFV , asumiendo que contribuciones procedentes de MFV son insignificantes tanto para procesos FCNC como para otros observables como EWPO y la masa del bos\'on de Higgs masas ( ver por ejemplo \cite{CMSSM-NUHM}). En este trabajo vamos a analizar si esta suposici\'on est\'a justificada, y si la inclusi\'on de los efectos MFV podr\'{\i}an conducir a restricciones adicionales del espacio de par\'ametros del CMSSM. En este sentido, vamos a evaluar en el CMSSM y en el CMSSMI el siguiente conjunto de observables: BPO, en particular, \ensuremath{{\rm BR}(B \to X_s \gamma)}, \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ y \ensuremath{\Delta M_{B_s}}; EWPO, en particular, $M_W$ y el \'angulo de Weinberg efectivo, $\sin^2\theta_{\mathrm{eff}}$; la masas de los bosones de Higgs neutros y cargados en el MSSM, as\'{\i} como cLFV y LFVHD.
La disposici\'on de la tesis es la siguiente. El cap\'{\i}tulo 1 contiene una introducci\'on al ME. En el cap\'{\i}tulo 2 presentamos MSSM y sus extensiones ``see-saw''. El cap\'{\i}tulo 3 est\'a dedicado a la base de c\'alculo de los observables considerados en este trabajo. En el cap\'{\i}tulo 4 presentaremos los resultados en el caso de la mezcla de sabor squark en el enfoque MI y el estudio sus efectos para los observables BPO, EWPO y QFVHD. En el cap\'{\i}tulo 5, de estudian los efectos de la mezcla de los sleptones en EWPO, las correciones a las masas de los Higgs y LFVHD en el contexto MI. El cap\'{\i}tulo 6 se centrar\'a en el an\'alisis del CMSSM y CMSSMI, para los que presentamos los efectos de mezcla del sabor en los observabels EWPO, BPO, las predicciones de masas del bos\'on de Higgs, QFVHD, cLFV mezcla y LFVHD. El cap\'{\i}tulo 7 se reserva para las conclusiones.
\chapter*{Resumen y Conclusiones}
\addcontentsline{toc}{chapter}{Resumen y Conclusiones}
La teor\'{\i}a supersim\'etrica ha demostrado un enorme potencial para explicar algunas de los mayores problemas del Modelo Est\'ander (ME), aunque hasta la fecha no se haya encontrado ninguna evidencia experimental de sus predicciones. Por ejemplo, la b\'usqueda directa de part\'{\i}culas supersim\'etricas no ha tenido \'exito por el momento. Sin embargo, es posible detectar la presencia de las nuevas part\'{\i}culas en los cambios que \'estas producen en algunos par\'ametros medidos con gran precisi\'on. En particular, el modelo supersim\'etrico m\'{\i}nimo (MSSM) predice nuevas contribuciones al cambio de sabor (FC) de los fermiones debido a mezclas entre las masas de sus correspondientes parejas supersim\'etricas. Esta mezcla est\'a originada por los par\'ametros responsables de la rotura de la Supersimetr\'{\i}a, lo cual tiene un gran inter\'es desde el punto de vista del dise\~no de modelos supersim\'etricos concretos. El cambio de sabor derivado de la no alineaci\'on entre fermiones y sus parejas escalares no se manifiesta en la aproximaci\'on a nivel m\'as bajo (“tree level”) de los c\'alculos, pero s\'{\i} en el primer orden (“one-loop” level) cuya contribuci\'on puede ser importante para ciertos valores de los par\'ametros del MSSM.
En esta tesis, se ha estudiado la posible contribuci\'on de la mezcla de sabor fermi\'onico a varias observaciones. La mezcla de sabor se ha introducido por medio de un conjunto de par\'ametros adimensionales denominados $\delta^{FAB}_{ij}$ ($F=Q,U,D,L,E; A,B = L,R$; $i,j = 1,2,3$). En el cap\'{\i}tulo 1, se revisaron algunos aspectos del ME; en el cap\'{\i}tulo 2, se introdujo el MSSM y la extensi\'on de \'este que incluye un mecanismo del tipo “see-saw” para explicar las oscilaciones de sabor de los neutrinos. Los detalles de la contribuci\'on SUSY a algunos observables de inter\'es se presenta en el cap\'{\i}tulo 3, en concreto se consideran: observables de la teor\'{\i}a electro-d\'ebil medidos con gran precisi\'on (EWPO), correcciones a la masa del bos\'on de Higgs, detalles en el c\'omputo de la f\'{\i}sica relacionada con el quark b ($B$-physics observables (BPO)), desintegraciones del bos\'on de Higgs con violaci\'on de sabor de quark (QFVHD) y finalmente, desintegraciones del bos\'on de Higgs con violaci\'on de sabor lept\'onico (LFVHD). Para calcular los efectos de la mezcla del sector lept\'onico se elabor\'o un algoritmo adicional para {\tt FeynArts}, con \'el se incluye LFV en el modelo del MSSM que el paquete ya tiene definido. Con ello ampliamos la capacidad de los programas incluidos en {\tt FeynArts}/{\tt FormCalc}\ para computar el efecto del LFV en observables como EWPO (a partir del c\'omputo de las auto-energ\'{\i}as de los bosones gauge) y tambi\'en sobre la masa de los bosones de Higgs del MSSM. Los resultados correspondientes han sido incluidos en el el c\'odigo {\tt FeynHiggs}\ y est\'an disponibles para su libre distribuci\'on a partir de la versi\'on 2.10.2. Se revis\'o el c\'alculo de la desintegraci\'on $\ensuremath{h \to \bar b s + b \bar s}$ utilizando los c\'odigos actualizados de
{\tt FeynArts}\ y {\tt FormCalc}. Los c\'alculos para evaluar los observables BPO y EWPO se realizaron con la ayuda de {\tt FeynHiggs}\ (utilizando una versi\'on no p\'ublica), teniendo en cuenta la contribuci\'on a la de todos los t\'erminos que violan sabor, en el caso de $M_W$ y de los m\'as relevantes en el caso de BPO.
En el cap\'{\i}tulo 4 se han estudiado los efectos de la mezcla de sabor de los squarks en la observaci\'on de EWPO, BPO y QFVHD (por ejemplo en $\ensuremath{h \to \bar b s + b \bar s}$) de una manera independiente del modelo (MI) que produce la mezcla de sabor. Nuestro c\'alculo mejora otros previos en varios aspectos: se ha tenido en cuenta el total de las correcciones supersim\'etricas del tipo fuerte y electro-d\'ebil y, adem\'as, se permiti\'o la mezcla simult\'anea de contribuciones del tipo LL, RL, LR y RR. Tambi\'en se consider\'o la limitaci\'on del valor de los par\'ametros impuesta no solo por los BPO, sino tambi\'en por los EWPO, en particular la masa del bos\'on $M_W$. Se mostr\'o que la contribuci\'on a $M_W$ produce restricciones adicionales al espacio de los par\'ametros $\delta^{FAB}_{ij}$ que mezclan el sabor de los squarks.
En la evaluaci\'on de los posibles valores de los par\'ametros $\delta^{FAB}_{ij}$ se han teniendo en cuenta las limitaciones procedentes de los valores de los BPO y EWPO. Para ello, se consideraron seis escenarios representativos no excluidos ni por la b\'usqueda de part\'{\i}culas SUSY ni por el valor experimental de la masa del bos\'on de Higgs. Los valores de $\delta^{FAB}_{ij}$ obtenidos se usaron para calcular $\ensuremath{\br(\hbs)}$. En el caso de tomar s\'olo uno de los $\delta^{FAB}_{ij} \neq 0$ se encontr\'o que \'unicamente valores relativamente grandes de $\delta^{DLR}_{23}$ predicen valores de $\ensuremath{\br(\hbs)} \sim 10^{-4}$, detectables en futuros colisionadores $e^+e^-$ . Permitiendo dos $\delta^{FAB}_{ij} \neq 0$ simult\'aneamente se obtiene un valor $\ensuremath{\br(\hbs)} \sim 10^{-3}$, que podr\'{\i}a observarse en el ILC. En cambio, si se permite un tercer $\delta^{FAB}_{ij} \neq 0$, no incrementa m\'as el valor de esas predicciones.
En el cap\'{\i}tulo 5 se estudiaron los efectos de la mezcla de sabor de los sleptones en los EWPO, las masas de los bosones de Higgs de una manera independiente del modelo que origina la mezcla del sabor. Nuestro an\'alisis num\'erico toma como referencia seis modelos supersim\'etricos a los que se atribuyen determinados valores de los par\'ametros de modo que las propiedades de los fermiones sean diferentes en cada uno de los casos. Los valores de los $\delta^{FAB}_{ij}$ en los seis escenarios est\'an limitados por la no observaci\'on de la violaci\'on del sabor lept\'onico (LFV) en procesos como $\mu\to e \gamma$ \cite{Arana-Catania:2013nha}, lo que se ha tenido en cuenta en la evaluaci\'on del los EWPO y las masas de los bosones de Higgs. De este modo hemos podido computar, de una manera general, el posible impacto de las restricciones en el valor de los $\delta^{FAB}_{ij}$ debido a \'estos observables en la predicci\'on de procesos con LFV. Encontramos que éstas son considerables para $\del{LLL}{23}$ e insignificantes para el resto de los $\delta^{FAB}_{ij}$, al menos en los escenarios considerados en nuestra investigaci\'on. El efecto de variar los valores de $\del{LLL}{23}$ dentro de los intervalos permitidos experimentalmente implica contribuciones para $M_W$ y $\sin^2\theta_{\mathrm{eff}}$ que pueden exceder el margen de error con el que est\'an medidos. Sin embargo, esto no implica nuevas restricciones en los valores de $\del{LLL}{23}$, ya que los valores absolutos de $M_W$ y $\sin^2\theta_{\mathrm{eff}}$ dependen en gran medida del s-top y el s-bottom. Este sector est\'a desconectado del de los sleptones, objeto de nuestra investigaci\'on. Sin embargo, en algunos casos, los limites que resultan de los EWPO pueden ser m\'as restrictivos que los procedentes de la medida directa de procesos con LFV. En el sector de los bosones de Higgs, la introducci\'on de violaci\'on de sabor de los s-quarks implica contribuciones no triviales a las masas del bos\'on de Higgs neutro m\'as ligero $M_h$ y a la del cargado $M_{H^{\pm}}$. En ambos casos, la incertidumbre te\'orica en su determinaci\'on es superior a la experimental. Si consideramos los valores para $\delta^{FAB}_{ij}$ permitidos por los l\'{\i}mites de procesos con LFV , la contribuci\'on a ambas masas es peque\~na. Para $M_h$ es del orden de unos pocos MeVs mientras para $M_{H^{\pm}}$ puede llegar hasta 2 GeV. Esta \'ultima puede alcanzar el valor de la futura precisi\'on experimental, dependiendo de la masa del bos\'on de Higgs neutro m\'as pesado. Pero lo m\'as relevante, es el hecho de que la incertidumbre derivada de los efectos de LFV en la evaluaci\'on de las masas de los bosones de Higgs neutros se ha reducido hasta hacerse del mismo orden que la que se prev\'e alacanzar en experimentos futuros.
En el cap\'{\i}tulo 5 tambien se presentaron las predicciones para LFVHD siguiendo la t\'ecnica MI. Sin embargo, las severas restricciones impuestas por los procesos con cLFV hacen que las predicciones para esos procesos sean muy peque\~nas.
En el cap\'{\i}tulo 6 se estudiaron los efectos de la mezcla de squarks en el CMSSM y de sleptones en la extensi\'on de \'este con un mecanismo ``see-saw'' de tipo I. En ambos casos se utiliz\'o la hip\'otesis de violaci\'on de sabor m\'{\i}nima (MFV). Este trabajo fue motivado por el hecho de que muchos an\'alisis fenomenol\'ogicos del CMSSM no incluyen estos efectos. Sin embargo, aparecen de manera natural en la evoluci\'on de los par\'ametros del modelo entre las escalas de energ\'{\i}a de gran unficiación (GUT) y
electro-d\'ebil (EW) debido a la presencia de las matrices CKM y PMNS en las RGE´s. En este sentido, los modelos CMSSM y CMSSM-seesaw~I\ constituyen dos ejemplos sencillos de modelos con gran unificaci\'on en los que la violaci\'on de sabor procede \'unicamente de las RGE. El espectro de masas de las part\'{\i}culas supersim\'etricas en ambos casos se ha evaluado num\'ericamente mediante el programa {\tt Spheno}, a partir de los valores a la escala GUT.
Se calcularon las predicciones para BPO, y la masa de los bosones de Higgs en el CMSSM y el CMSSM-seesaw~I. Se evalu\'o el impacto de incluir la mezcla de sabor comparando el c\'omputo con el caso simple en el que \'esta se desprecia, como ocurre en otros an\'alisis previos al nuestro. Los resultados indican en qu\'e casos pueden ignorarse las mezclas de sabor.
En el caso del CMSSM se ha hecho un recorrido a trav\'es de una red de valores en el plano $m_0$--$m_{1/2}$ para valores fijos de $A_0$ y $\tan \beta$. Se encontr\'o que el valor de $\delta^{FAB}_{ij}$ aumenta al incrementar los valores de \'estos \'ultimos valores. Los valores de las correcciones a las masas de los bosones de Higgs son inferiores a la precisi\'on en su valor experimental (presente y futuro) . De manera an\'aloga el impacto sobre los BPO est\'a por debajo de la incertidumbre en su medida experimental. Por tanto, encontramos que para estos dos grupos de observables est\'a justificado el ignorar los efectos de la mezcla. La conclusi\'on es diferente en el caso de los EWPO, aqu\'{\i} encontramos que las masas de los squarks aumentan con $m_0$, y con ello los par\'ametros de mezcla, esto genera una diferencia entre las masas del doblete de $SU(2)$ que debe ser tenida en cuenta en el c\'omputo de estos observables. Para $m_0 \gsim 3 \,\, \mathrm{TeV}$ encontramos que induce a valores de $M_W$ y $\sin^2\theta_{\mathrm{eff}}$ que superan varias veces la incertidumbre experimental, hasta el punto llevar a las predicciones del CMSSM fuera de los l\'{\i}mites experimentales. De esta manera, nuestro an\'alisis permite establecer l\'{\i}mites para la masa del $m_0$ que no aparec\'{\i}an en trabajos anteriores. El origen de la diferencia en las masas de los squarks responsable de \'esta contribuci\'on fue corroborado num\'ericamente utilizando {\tt Spheno}, con el que se compararon los efectos de incluir o ignorar la matriz CKM en la integraci\'on de las RGE. Sin embargo, debemos se\~nalar que nuestras conclusiones son dif\'{\i}ciles de extrapolar a modelos m\'as complejos que los aqu\'{\i} utilizados.
El \'ultimo eslab\'on de nuestro an\'alisis con el CMSSM ha sido la evaluaci\'on del $\ensuremath{\br(\hbs)}$. En este caso, encontramos valores del orden de ${10^{-7}}$. Esto es, fuera del alcance de su detecci\'on en los experimentos proyectados para un futuro pr\'oximo. En el caso del CMSSM-seesaw~I\, los resultados num\'ericos dependen de c\'omo est\'e definido el modelo. Se eligi\'o un conjunto de par\'ametros que reproduce correctamente las observaciones referentes a los neutrinos y a su vez induce a contribuciones apreciables de LFV en los observables que estudiamos. En consecuencia, algunas regiones del espacio de par\'ametros est\'an exlu\'{\i}das por su predicci\'on a ${\rm BR}(\mu \to e \gamma)$. En cambio, las de los ${\rm BR}(\tau \to e \gamma)$ y ${\rm BR}(\tau \to \mu \gamma)$ no alcanzan sus respectivos l\'{\i}mites experimentales. Las predicciones para los BR de LFVHD son muy peque\~nas tambi\'en en el CMSSM-seesaw~I. Con ello concluimos que precisamos de modelos diferentes del CMSSM-seesaw~I\ para explicar la observaci\'on de ${\rm BR}(h \to \mu \tau)$ en el detector CMS del CERN. En lo tocante a los observables medidos con gran precisi\'on, encontramos que los BPO no est\'an afectados. Tampoco las predicciones de las masas de los bosones de Higgs. El mayor impacto aparece una vez m\'as en los EWPO, en el caso de la $M_W$ pueden ser del orden de la incertidumbre experimental.
En resumen, en nuestro estudio de MFV hemos utilizado dos modelos con gran unificaci\'on ``realistas'' en los que la violaci\'on de sabor es introducida al tener en cuenta la presencia de las matrices CKM y PMNS en las RGE's. Se encontr\'o que el desestimar los efectos de violaci\'on de sabor es adecuado para los BPO y las masas de los bosones de Higgs. Sin embargo, para los EWPO se encontraron efectos grandes. El valor de la contribuci\'on a $M_W$ y $\sin^2\theta_{\mathrm{eff}}$ es superior a la incertidumbre de su valor experimental. Por tanto, los efectos de violaci\'on de sabor, objeto de nuestro estudio imponen un nuevo l\'{\i}mite superior a $m_0$ que no se ha tenido en cuenta en otros an\'alisis fenomenol\'ogicos recientes.
|
\section{Introduction}
The essence of the black hole information paradox is that the symmetry
principle of General Relativity, general covariance, is incompatible
with a unitary quantum evolution, where the Hawking radiation \citep{hawking}
carries away the quantum information of the black hole. The internal
rigidity of quantum mechanics, and the success of holographic approaches
to string theory, such as AdS/CFT \citep{Maldacena:1997re} and the
BFSS matrix model \citep{Banks:1996vh}, lend weight to the hypothesis
that a unitary quantum description should be exact. One is then faced
with the problem of how to recover approximate general covariance
from such a description.
According to the principle of black hole complementarity, as introduced
in \citep{Susskind:1993if}, physics outside the stretched horizon
of a black hole is well described by a local effective field theory
but the local description does not extend inside the stretched horizon.
As far as outside observers are concerned, the black hole interior
is encoded into quantum mechanical degrees of freedom associated with
the stretched horizon and residing in a Hilbert space of finite dimension
given by the exponential of the Bekenstein-Hawking entropy of the
black hole. No information enters the black hole - from the outside
point of view infalling matter is absorbed, thermalized and re-emitted
by the stretched horizon - and the interior spacetime experienced
by a typical observer entering the black hole in free fall is postulated
to emerge from the stretched horizon degrees of freedom in a holographic
fashion. The interior hologram is constructed from a finite number
of degrees of freedom, so the interior bulk theory can at best be
approximately local. The question is whether a physical observer inside
a black hole, whose measurement resolution is limited both in space
and time by the finite size of the black hole, can detect a deviation
from local effective field theory in the time allotted before hitting
the curvature singularity.
A version of black hole complementarity, which addresses this question,
has been proposed by the authors and explored in some recent papers
\citep{Lowe:2014vfa,Lowe:2015eba}. The construction in \citep{Lowe:2014vfa}
applies to an observer falling into a black hole at a prescribed time
and is restricted to a limited time period before and after. For the
construction to work, the two complementary descriptions of the black
hole interior need to satisfy two key requirements. First of all,
from the point of view of the outside description, the minimal decoherence
time of a black hole must have a lower bound of order the scrambling
time $t_{scr}=4M\log(4M)$. This is the minimum time it takes outside
observers to extract quantum information from the black hole after
an infalling observer has been absorbed into the stretched horizon.
Second, from the viewpoint of the infalling observer, any quantum
information that entered the black hole more than a scrambling time
earlier must already have been erased from the interior bulk description
when the observer enters the black hole.
The second requirement was studied in \citep{Lowe:2015eba} using
a simple model for the bulk physics. It was argued that a sensible
holographic description with a finite $N$ (or more precisely the
bulk $\hbar$ non-zero), would correspond to a bulk theory with a
physical regulator\footnote{The need for a bulk regulator to reconcile
a finite black hole entropy with the number of states computed
semiclassically was noted already in \citep{tHooft:1984re}.}. A minimal
requirement is that such a theory should be capable of
describing events in a freely-falling frame outside the black hole
with a resolution down to a Planck length. Propagating that forward
in time, then leads to a lattice black hole model of the type studied
in \citep{Corley:1997ef} based on a Painlev\textipa{\'{e}}-Gullstrand
time-slicing. The time coordinate of the infalling frame can be matched
to the exterior timelike Killing vector at some finite distance outside
the black hole, allowing one to map time evolution in the holographic
model to time evolution in the black hole interior. By studying interior
propagation of massless fields in this lattice description, it was
found that the scrambling time emerges as the maximum coordinate time
a signal can propagate before hitting the curvature singularity. This
is in sharp contrast to the unregulated description, with exact general
covariance, where an infinite coordinate time might pass before collision
with the singularity.
In the present paper, we study a model for the holographic side of
this story. The predictions from the bulk side are that one should
see free propagation in the interior for a time at least of order
the scrambling time \citep{Lowe:2015eba} followed by the rapid onset
of large curvature effects with a timescale of order the Planck time.
Suppose we send in a small subsystem in a pure state into the black
hole. The subsystem can be viewed as a simple model for a freely falling
laboratory, where tests of spacetime locality are carried out. Eventually
the quantum information initially contained in the subsystem will
come out in the Hawking radiation. Since the Hawking radiation propagates
with respect to a local Hamiltonian in the exterior, any interactions
there will appear as effects that decohere the state from the interior
bulk viewpoint. Moreover, such interactions will appear highly non-local
in the infalling frame, and lead to apparent violations of quantum
mechanics for the infalling observer. Naively, one might predict that
measurements of the Hawking radiation might disrupt such a state
in an arbitrarily short time. However, if one insists that the detector
itself evolve according to the same Hamiltonian as the black hole
degrees of freedom, a finite minimal decoherence time emerges.
These statements will be quantified in a spin system we introduce
below as a simple model of the holographic stretched horizon. The
model we choose exhibits fast scrambling which is conjectured to be
a property of the stretched horizon degrees of freedom a wide class
of black holes \citep{Sekino:2008he}. The infalling Hamiltonian evolution
is mapped in the holographic model to a mean field Hamiltonian, dependent
on the initial state of the system. We compute the trace distance
between the states that are obtained by evolving the initial state
with respect to the exact Hamiltonian and with respect to the mean
field Hamiltonian. This trace distance provides a measure of the decoherence
of the infalling state. We find that the decoherence only becomes
significant after at least a scrambling time, matching precisely the
expectation from the regulated bulk theory. Moreover the timescale
for the rapid onset of decoherence also matches the bulk prediction.
The results support the version of black hole complementarity advocated
by the authors, where singularity approach is complementary to decoherence
of the infalling state, initially outlined in \citep{Lowe:2006xm}.
This represents an important step forward towards solving the black
hole information problem.
\section{Coherence/decoherence}
Let us begin by reviewing the basic ideas of decoherence, which involves
some system of interest $S$, interacting with some much larger system
$S^{C}$ which is often denoted the environment. We suppose the Hilbert
space factors as
\[
\mathcal{H}=\mathcal{H}_{S}\otimes\mathcal{H}_{S^{c}}
\]
Consider a pure state in $\mathcal{H}_{S}$
\[
|\Psi\rangle=\left(|\psi_{1}\rangle_{S}+|\psi_{2}\rangle_{S}\right)\otimes|\chi\rangle_{S^{C}}
\]
Under time evolution this becomes
\begin{equation}
|\Psi'\rangle=e^{-iHt}|\Psi\rangle=\sum_{i}c_{1i}|e_{i}\rangle\otimes|f_{1i}\rangle+c_{2i}|e_{i}\rangle\otimes|f_{2i}\rangle\label{eq:timeevo}
\end{equation}
where the $e_{i}$ are some basis of $\mathcal{H}_{S}$. If there
is decoherence, then it is a good approximation to assume $|f_{1i}\rangle$
is orthogonal to $|f_{2j}\rangle$ for \emph{any} $i$ and $j$. For
example, this will typically occur if the Hamiltonian is local in
position space and $|f_{1i}\rangle$ and $|f_{2i}\rangle$ are position
eigenstates. We will adopt the notation
\[
\Psi_{S}=\mathrm{Tr}_{S^{c}}|\Psi\rangle\langle\Psi|
\]
to denote the partial trace over the complement of $S$. If there
is decoherence, then to a good approximation
\begin{align}
\Psi'_{S} & \approx\sum_{i}\left(|c_{1i}|^{2}+|c_{2i}|^{2}\right)|e_{i}\rangle\langle e_{i}|\label{eq:diagden}
\end{align}
which means the probabilities add, without cross terms. The end result
for $\Psi'_{S}$ is then the same as if a measurement had collapsed
the wavefunction into the states $|e_{i}\rangle$. Note the probability
of each $|e_{i}\rangle$ is not necessarily equal, so $\Psi'_{S}$
need not be maximally mixed.
As it stands, this statement of decoherence is basis dependent. To
make a basis independent statement, one can instead quantify the purity
of the reduced density matrix $\Psi_{S}'$. One way to do this is to
compute
\[
P=\mathrm{Tr}_{S}(\Psi_{S}')^{2}
\]
which is known as the purity of a density matrix. $P=1$ for a pure
state since a pure state acts as a projector, $\left(\Psi_{S}'\right)^{2}=\Psi_{S}'$,
and by normalization $\mathrm{Tr}_{s}\Psi'_{S}=1$. For a mixed state
$0<P<1$.
Alternatively, one may use the von Neumann entropy
\[
\mathcal{S}=-\mathrm{Tr}_{s}\Psi'_{S}\log\Psi_{S}'
\]
to quantify the purity of the reduced density matrix. This vanishes
for a pure state. For a maximally mixed state $\Psi_{S}'=\mathbbm{1}/n$,
on the other hand, $\mathcal{S}=\log n$, with $n$ the dimension of the
Hilbert subspace $S$.
We can then formulate the decoherence time $t_{d}$ in the following
way. Assume at time $t=0$ $\Psi_{S}$ is in a pure state. Then define
the decoherence time $t_{d}$ as the time when
\begin{equation}
\mathcal{S}(\Psi_{S}(t_{d}))=\delta\log n\label{eq:decoentropy}
\end{equation}
for some choice of $\delta<1$. We are not aware of prior appearances
of this definition of decoherence time in the literature. This definition
should be useful in many other contexts.
In the following we will mostly be interested in studying finite dimensional
spin systems. In this class of models, we can reformulate the condition
\eqref{eq:decoentropy} as a condition on the trace distance using
the results of \citep{fannes1973}. We recall the definition
\begin{equation}
\left\Vert \Psi_{S}-\Phi_{S}\right\Vert _{1}=\mathrm{Tr}_{S}\sqrt{\left(\Psi_{S}-\Phi_{S}\right)^{\dagger}\left(\Psi_{S}-\Phi_{S}\right)}\label{eq:tracedis}
\end{equation}
In \citep{fannes1973} it is shown that
\begin{equation}
|\mathcal{S}(\Psi_{S})-\mathcal{S}(\Phi_{S})|\leq\left\Vert \Psi_{S}-\Phi_{S}\right\Vert _{1}\log n\label{eq:fannes}
\end{equation}
for two different density matrices in $\mathcal{H}_{S}$. Therefore
the definition of the decoherence time can be reformulated as
\begin{equation}
\left\Vert \Psi_{S}(t_{d})-\Phi_{S}(t_{d})\right\Vert _{1}=\delta\label{eq:tracedeco}
\end{equation}
for some fixed constant $\delta<1$ and some suitable choice for $\Phi_{S}$.
The state $\Phi_{S}(t)$ should be chosen to maintain purity under
time evolution for the subsystem of interest, but minimize the trace
distance as a function of time so the bound \eqref{eq:fannes} is
as useful as possible. For the models considered here, we will choose
$\Phi_{S}$ to evolve according to a local mean field Hamiltonian,
as we describe below.
In \citep{Lashkari:2011yi} the statement of fast scrambling was defined
in a similar way. The key distinction is that scrambling involves
a global mixing of the system, rather than only the mixing of a particular
subsystem of interest. The condition for scrambling would then require
that, \eqref{eq:tracedeco} should hold for all subsystems, suitably
defined, rather than some single small subsystem, as typically considered
in the decoherence problem.
\section{Toy Holographic Model\label{sec:Toy-Holographic-Model}}
While it is interesting to try to derive an effective holographic
model for the horizon degrees of freedom of a black hole from some
more fundamental description such as AdS/CFT or the Matrix Model,
our strategy will be to make some minimal assumptions about such a
description, and hope that it carries over to a more precise reconstruction.
The key assumption we will make of the model is that it exhibits fast
scrambling in the sense of \citep{Hayden:2007cs}, with a scrambling
time
\[
t\sim\beta\log S_{BH}
\]
with $\beta$ the inverse Hawking temperature of a black hole hole
with energy $E$ and $S_{BH}$ the Bekenstein-Hawking entropy of the
black hole. Later we will also be interested in carrying out computations
in the model for highly entangled states that will model the state
of an old black hole entangled with its Hawking radiation. As such,
we assume the model contains enough degrees of freedom to model the
interior of the black hole and its immediate vicinity. Thus we make
the identification that $S\sim N$ the number of sites in the model,
and $\beta$ will be scaled out of the problem. The near-horizon region
will not contain all the symmetries of the asymptotic region, so we
do not expect conformal symmetry (as in AdS/CFT) or supersymmetry
(as in the BFSS model) to be crucial in formulating this effective
model. At best the holographic model should contain a version of rotation/translation
symmetry, and time translation invariance.
A toy model that exhibits these features is discussed in \citep{Lashkari:2011yi}.
This is a spin model with a non-local pairwise interaction. There
are $N$ distinct sites with the Hilbert space of tensor product form
$\mathcal{H}=\mathcal{H}_{1}\otimes\cdots\otimes\mathcal{H}_{N}$.
The sites interact via a pairwise Hamiltonian
$H=\sum_{\left\langle x,y\right\rangle }H_{\left\langle x,y\right\rangle }$
summing over unordered pairs of sites. The Hamiltonian may therefore
be associated with a graph $G=(V,E)$ with $N$ vertices $V$, and
edges $E$ corresponding to the non-zero $H_{\left\langle x,y\right\rangle }$.
In order to have fast scrambling, the degree of the vertices $D$
should be of order the size of the system. We shall then set $D=N-1$.
To have a sensible limit for large $N$, we take the pairwise interactions
to be bounded $\left\Vert H_{\left\langle x,y\right\rangle }\right\Vert<c/D$, for some
constant $c$. Here the operator norm $\left\Vert {\cal O} \right\Vert$ may be defined as the absolute value of the maximum eigenvalue of the operator ${\cal O}$.
The Lieb-Robinson result \citep{lieb1972} places bounds on the norm
of the commutator of operators localized at different sites, as a
function of time. For local interactions, this is to be interpreted
as a proof of finite group velocity in nonrelativistic spin systems.
In the case at hand, where interactions are non-local, the same method
still yields a bound on the norm of the commutator for operators.
In particular, in \citep{Lashkari:2011yi} it is shown that
\begin{equation}
\left\Vert \left[O_{A}(t),O_{B}\right]\right\Vert
\leq\frac{4}{D}\left\Vert O_{A}\right\Vert \left\Vert O_{B}\right\Vert |A|e^{8ct}
\label{eq:liebrob}
\end{equation}
Here $O_{X}$ is a bounded norm operator acting in the Hilbert subspace
of the sites in the set $X$, and $B$ is chosen to be a single site.
The condition for scrambling is set up in \citep{Lashkari:2011yi} as
follows. Consider some Hilbert subspace $\mathcal{H}_{1}$ with dimension
of order 1, maximally entangled with some reference system $\mathcal{P}$,
which experiences no interactions. Here we set the system $S=\mathcal{H}_{1}$.
Under time evolution, the entanglement between $\mathcal{H}_{1}$
and $\mathcal{P}$ will decay, which may be quantified by the trace
distance
\begin{equation}
\left\Vert \Psi_{\mathcal{P}S}(t_{*})-\Psi_{\mathcal{P}}(0)
\otimes\Psi_{S}(t_{*})\right\Vert _{1}<\epsilon\mathrm{\,rank}\Psi_{\mathcal{P}}(0)
\label{eq:tracedist}
\end{equation}
for some constant $\epsilon\ll1$. This may in principle then be used
as a definition of scrambling time. A bound on the time $t_{*}$
can then be obtained by noting that it is bounded by the signaling
time from the space $S$ to its complement $S^{c}$, which may be
bounded using \eqref{eq:liebrob}.
First apply this to an initial state where $S$ is a single site,
and the complement subspace $S^{c}$ has dimension of order $N$. We assume
the initial state is of product form $|\Psi(0)\rangle=|\psi_{1}\rangle_{\mathcal{PH}_{1}}
\otimes|\psi_{2}\rangle_{\mathcal{H}_{2}}\otimes\cdots
\otimes|\psi_{N}\rangle_{\mathcal{H}_{N}}$.
Applying \eqref{eq:liebrob} with $B=S$ and $A=S^{c}$ one finds the timescale
$t_{*}$ is of order a constant.
For the black hole problem, the natural initial state to choose is
instead one where the black hole degrees of freedom are maximally
entangled with the exterior Hawking radiation. Now essentially the
roles of $S$ and $S^{c}$ are reversed. One takes the system $S$
to be of size of order $N$, with some small subsystem in a factor
pure state. The complement is then of size of order 1. To satisfy the
bound \eqref{eq:tracedist} one again requires signaling between $S$
and $S^{c}$, and this timescale is bounded by the Lieb-Robinson result.
This yields a timescale of order $\log N$ as expected for a fast
scrambling system.
\section{Mean field and bulk evolution\label{sec:Mean-field-and}}
At first sight, the results of the previous section are not encouraging
for the black hole complementarity scenario. While one can build holographic
models that exhibit fast scrambling with $t_{*}$ proportional to
$\log N$, it seems the decoherence time for some small Hilbert subspace
in such models will be very short. This is, however, not the right question
to ask in the black hole problem. Instead, what one should do is
build a model for a laboratory that one sends into the black hole, and then
ask whether that laboratory will have a decoherence time sufficiently
long that they will not be able to distinguish quantum mechanics failing
from their classical demise due to singularity approach.
The eventual failure of quantum mechanics in the infalling laboratory
can be traced to the existence of two distinct time evolutions for the
state in the lab subspace. One of these will be the exact Hamiltonian
evolution according to the holographic Hamiltonian $H$. The other
will be defined according to a mean field Hamiltonian $H_{MF}$,
that we describe in more detail shortly, and corresponds to the
usual notion of time evolution in the bulk spacetime.
It is important to note that not all states yield sensible mean field
evolutions. Moreover, as will be clear, the mean field Hamiltonian
depends on the state. We conjecture that states close to smooth bulk
spacetimes do have useful mean field descriptions, and that the mean
field evolution is dual to the usual time evolution with respect to
the bulk Hamiltonian. In the remainder of this section, we explore
these issues in the context of our holographic toy model.
The mean field approximation to the time evolution of a density matrix
is considered in some generality in \citep{muntean}. We begin by
briefly reviewing the standard mean field construction based on an
initial pure state of product form
\begin{equation}
|\Psi(0)\rangle=|\psi_{1}\rangle_{\mathcal{H}_{1}}
\otimes\cdots\otimes|\psi_{N}\rangle_{\mathcal{H}_{N}}
\label{eq:initial}
\end{equation}
and later on we adapt it to the case of a highly entangled state
corresponding to an old black hole.
Starting from \eqref{eq:initial} one builds a state dependent mean field
Hamiltonian
\begin{align}
H^{MF} & =\sum_{x}H_{x}^{MF}(t)\nonumber \\
H_{x}^{MF} & =\sum_{y}{\rm tr}_{y}
\left(H_{\left\langle x,y\right\rangle }\Psi_{y}^{MF}(t)\right)
\label{eq:meanham}
\end{align}
where $\Psi^{MF}$ evolves according to $H^{MF}$ from the
same initial state $|\Psi(0)\rangle$. A key point is that with these
definitions, and choice of initial state, the mean field Hamiltonian
never generates entanglement between different sites, and remains
in the same product form as the initial state. This mean field Hamiltonian
then has the expected properties of the holographic dual of the bulk
gravity Hamiltonian. As is well known, the bulk Hamiltonian generates
smooth time evolution all the way to the curvature singularity, with
minimal entanglement being generated. This feature is in fact the origin
of the information problem.\footnote{It should be noted the mean field
Hamiltonian depends on the choice
of initial state via \eqref{eq:meanham}. The state dependence of
the boundary to bulk map is emphasized in \citep{Papadodimas:2013jku}.
However in the present construction it is also important that the boundary
to bulk mapping is time dependent, which follows from $[H,H_{x}^{MF}]\neq0.$}
One then wishes to calculate the timescale for which the trace norm
distance between $\Psi_{x}(t)$ and $\Psi_{x}^{MF}(t)$ remains small.
This maps onto a problem solved in \citep{Lashkari:2011yi} for the spin model
considered above, via careful application of Lieb-Robinson bounds
applied to an expansion of the matrix element
\begin{equation}
\left\langle \Psi(t)|\mathbbm{1}-\Psi_{x}^{MF}(t)|\Psi(t)\right\rangle
=1-\left\langle \Psi_{x}^{MF}(t)|\Psi_{x}(t)|\Psi_{x}^{MF}(t)\right\rangle
\label{eq:matrixel}
\end{equation}
by making a Dyson series expansion in $H-H_{x}^{MF}$. This matrix
element in turn places a bound on the trace distance between the states
\eqref{eq:tracedis}. Using the result of \citep{fannes1973}, this
then places a bound on the von Neumann entropy $H(\Psi_{x}(t))$.
One finds
\begin{equation}
\left\langle \Psi_{x}^{MF}(t)|\Psi_{x}(t)|\Psi_{x}^{MF}(t)
\right\rangle \leq\frac{c'}{D}e^{c''t}
\label{eq:liebrob-lak}
\end{equation}
where $c'$and $c''$ are constants independent of $N$. For $D=N-1$
these quantities become of order 1 when $t\sim\log N$.
Making contact with black hole physics, the holographic description
should be useful both inside and outside the black hole horizon. An
initial state of the form \eqref{eq:initial} is relevant outside
the black hole horizon, or for a recently formed black hole prior
to scrambling. To make further progress we need to generalize the
mean field results of \citep{Lashkari:2011yi} to highly entangled
states.
Suppose we choose a maximally entangled initial state where we have
a pairwise entanglement between $\mathcal{H}_{2k}$ and $\mathcal{H}_{2k+1}$
for all $k\geq1$. Then we can almost map the problem into the one
just considered by coarse graining, and viewing
$\mathcal{H}_{2k}\otimes\mathcal{H}_{2k+1}$
as a pure state on a single coarse grained site.\footnote{We note that
the Schmidt decomposition \citep{Nielsen:2011:QCQ:1972505} implies this
special state is unitary equivalent to the generic maximally entangled state.
Converting from Schrodinger picture to Heisenberg picture, this may be
viewed as conjugation of the Hamiltonian by a constant unitary matrix.
For the pairwise interaction considered in this model, a general unitary
transformation will induce self-interactions, but preserve the pairwise form
of the Hamiltonian. The condition $\left\Vert H_{\langle x,y\rangle}\right\Vert<c/D$
is preserved by this unitary conjugation, so the above proof goes through.}
The new feature is that the coarse grained Hamiltonian now has a
self-interaction term. Such a term must be treated exactly in the
mean field approximation. For this initial state, we therefore define
\begin{align*}
H^{MF} & =\sum_{x}H_{x}^{MF}(t)\\
H_{x}^{MF} & =H_{\left\langle x,x\right\rangle }
+\sum_{y\backslash x}{\rm tr}_{y}
\left(H_{\left\langle x,y\right\rangle }\Psi_{y}^{MF}(t)\right)
\end{align*}
where the sums are over coarse grained sites $x=1,\cdots,N/2$. With
this Hamiltonian, we may then proceed as above to compute the trace
distance between the mean field state and the exact evolution, or
equivalently the von Neumann entropy of the exact reduced density
matrix, obtaining the same scaling with $N$ (though different constants
$c'$ and $c''$) via \eqref{eq:liebrob-lak}.
This is now a nice model for an old evaporating black hole after the
Page time $t\sim\beta S_{BH}$, where the interior degrees of freedom
are maximally entangled with the exterior Hawking radiation. The decoherence
time, defined according to the definition \eqref{eq:decoentropy}
is now of order $t\sim\log N$ matching the scrambling time. We also
note once a time of order the scrambling time has passed, the bound
\eqref{eq:liebrob-lak} increases with a rise time of order 1, matching
the bulk expectation of strong curvature effects with an onset of
order the Planck time.
The same kind of computations can be carried out for a variety of
initial states. For example, to mimic a an observer falling in and
carrying out quantum experiments, we can choose to separate the infalling
site $x$ into two sites $x_{1}$ and $x_{2}$ with
$\left\Vert \mathcal{H}_{1}\Vert\gg \Vert\mathcal{H}_{2}\right\Vert$.
Let us assume some strong coupling between these sites, with coupling
to the other sites bounded as above. Defining the decoherence time
using \eqref{eq:tracedeco} for the choice $S=x_{2}$ will lead to
a decoherence time of order 1, independent of $N$. This means the
decoherence time for measurements internal to the infalling lab can
be made rapidly, as expected. However it is still true that the combined
state on $\mathcal{H}_{1}\otimes\mathcal{H}_{2}$ will remain pure
for a time of order the scrambling time using the above construction.
Thus measurements of the Hawking radiation will not lead to rapid
decoherence of the state, as naively expected.
To test the idea that the minimal decoherence time due to measurement
of Hawking radiation really matches the scrambling time, one can try
to generalize the above discussion to any density matrix with a pure
state subfactor representing the infalling system. The above derivation
will generalize provided the mean field approximation holds for the
evolution of the subsystem of interest. It seems natural that states
representing smooth spacetime geometries will correspond to good mean
field states, however the converse need not be true. It would be very
interesting to see more directly how this class of states emerges
as a class of attractor states from large $N$ holographic theories.
\section{Comments}
In the above we have argued the minimal decoherence time of an infalling
state is to be identified with the scrambling time, subject to assumptions
about the form of the holographic model. In this context, we have
proven one of the key assumptions about the approach to black hole
complementarity described in \citep{Lowe:2014vfa,Lowe:2015eba}.
The other key assumption relied on details of the holographic reconstruction
of bulk spacetimes, namely that general covariance is softly broken
through the introduction of a Planck length regulator. This was explored
in a regulated model of the bulk in \citep{Lowe:2015eba}, where it
was found that it was sufficient for infalling degrees of freedom
to retain coherence for a scrambling time. That work describes the
details of the construction, including the important conditions placed
on bulk timeslices compatible with the regulator.
Since we are concluding the timescales match, this is an important
success for building interior degrees of freedom in a holographic
theory. Moreover, if the bound is saturated, one also predicts the
timescale of order 1 associated with the rapid rise expected from
strong curvature near the singularity. Away from this region the trace
distance will be of order $1/N$.
Since these corrections are to be essentially interpreted as violations
of quantum mechanics for the infalling observer, it is of great interest
to quantify to what extent these are tolerable. The trace distance
may be interpreted directly as the probability of an ideal experiment
detecting the difference between two states \citep{Nielsen:2011:QCQ:1972505}.
As an initial crude estimate, if we take $N\sim S_{universe}\approx10^{88}$
(assuming domination by CMB photons) and assume the nonlocal effect
produces a Planck energy particle with a probability $1/N$ per unit
Planck time per degree of freedom, we can apply it to the atoms in
the Earth's atmosphere to conclude one Planck energy ultra high energy
cosmic ray would appear about every $10^{7}$ years. This is conceivably
a detectable effect, but apparently rather harmless.
It is natural to conjecture some version of the same matching of scrambling
time with interior decoherence time holds in all holographic theories
of quantum gravity. It remains an important open problem to directly
derive holographic effective theories of the horizon degrees of freedom
from more fundamental descriptions such as AdS/CFT or Matrix Models,
and test this conjecture. It will also be very interesting to further
explore mean field approximations in such holographic descriptions.
Of course one expects the mean field (or master field formulation
of a large $N$ theory) will coincide with the bulk gravity description
at leading order. However having a formulation directly in the Hilbert
space of the underlying holographic description is needed to carry out
computations analogous to those of section~\ref{sec:Mean-field-and}.
We note that generic holographic states in more realistic models may
well contain singularities leading to a breakdown of the mean field approach. From the bulk
perspective, we would expect for sufficiently large $N$, a version of cosmic
censorship \citep{Penrose:1969pc} to hold. The additional singularities
will then be censored by their own horizons leaving a smooth geometry
outside where we expect mean field to remain accurate. It will be interesting to see what extent this may be derived in the holographic model considered here.
In the near term, it would be interesting to generalize the present
work beyond spin models to systems with an infinite dimensional Hilbert
space at each site. Some recent applications of Lieb-Robinson bounds
to such systems appear in \citep{Nachtergaele2009,Nachtergaele:2009jv,2014arXiv1410.8174N,2015arXiv151206319G}.
It appears promising that these results can be generalized to models
that exhibit fast scrambling.
|
\section{Introduction}
\label{intro}
The Standard Model of particle physics predicts the existence of three active
massless neutrino species: the electron ($\nu_{\rm e}$), muon ($\nu_\mu$) and tau ($\nu_\tau$)
neutrinos. However, the discovery of
lepton flavour oscillations has suggested that neutrinos are massive
particles, fixing the lower limit of the sum of neutrino masses to
\smnu$\equiv m_{\nu_e}+m_{\nu_\mu}+m_{\nu_\tau}\gtrsim 0.06$ eV\footnote{
More specifically, \smnu\ must be greater than approximately $0.06$ eV in the
normal hierarchy scenario and $0.1$ eV in the degenerate hierarchy.
}~\cite{lesgourgues06,lesgourgues12,lesgourgues14,
lesgourgues13}. This implies that, after becoming
non-relativistic, neutrino free-stream with large thermal
velocities that suppress the growth of neutrino densities
perturbations on scales smaller than the so-called ``free-streaming
length'', $\lambda_{\rm fs}(z,m_\nu) \simeq 8.1\,H_0\,(1+z)/H(z)\,(1\,\rm eV/m_\nu)\, \rm Mpc/h$,
where $m_\nu$ is the mass of the single neutrino species, $H(z)$ is
the so-called Hubble parameter, and $H_0$ the Hubble constant,
$H(z=0)$. As a consequence, due to gravitational backreaction effects,
also the evolution of cold dark matter (CDM) and
baryon~\cite{rossi14} densities is altered, and the
total matter power spectrum is largely suppressed at scales $\lambda<<\lambda_{\rm fs}$ .
Moreover, given the present mass constraints, neutrinos become non-relativistic after the epoch of
recombination, and, accordingly, modify the radiation density contribution.
The transition from the relativistic to the non-relativistic regimes postpones the matter radiation equality for a given value of
\omegam$\,h^2$
(where \omegam\ is the ratio, at $z=0$, between the matter density of the Universe and the critical
density, $\rho_{\rm c}$, and $h$ the Hubble constant $H_0$ in units of 100 km
s\mone Mpc\mone), and modifies the background evolution, slightly affecting
the properties of the primary cosmic microwave background (CMB)
anisotropies.
In addition, along their travel from the last scattering surface to the observer,
CMB photons undergo also secondary anisotropies, in particular they are red/blue-shifted as
they cross growing/decaying gravitational potential wells.
This effect is called the late Integrated Sachs-Wolfe effect (ISW), as it was first described by
Sachs and Wolfe in 1967~\cite{Sachs67}. During the matter
dominated era, the two effects of background expansion and gravitational attraction
compensate each other so that the total linear gravitational
potential, $\Phi$, produced by the Large Scale Structure (LSS) distribution in the
Universe, is constant in time, and the ISW effect, which depends on the
time derivative $\dot{\Phi}$, vanishes. In contrast, during \eg the Dark Energy (DE) dominated
era, the background expansion rate of the Universe increases and these
two effects do not compensate anymore, causing the
decaying of the gravitational potential perturbations. In this case,
$\dot{\Phi}$ is no longer vanishing: a CMB photon passing through an overdense
region gains more energy falling into the potential
well with respect to the energy lost while climbing out of it; a
CMB photon passing through an underdense region loses more energy climbing the potential hill than
the energy gained during its descent. Therefore, overdense
regions correspond to hotter spots in the CMB sky map, and underdense regions to colder ones.
In the linear regime, the total effect is represented by an increase
of the photon temperature power spectrum on very large scales, which
has actually been detected via full-sky CMB probes,
\eg\ Planck~\cite{Planck13,Planck15}, or via the cross-correlation of the CMB temperature with LSS data~\cite{2012MNRAS.426.2581G,2016arXiv160403939S,Granett2009}.
However, besides dark energy, also the cosmological background of
massive neutrinos produces a non-vanishing ISW effect. In fact, the neutrino free-streaming
makes the gravitational potential to evolve in time, producing a net
$\dot{\Phi} \neq 0$ even in the absence of a recent accelerated background expansion.
As shown in~\cite{Lesgourgues_etal_2008}, neutrino velocities generate an
excess of ISW effect at high redshifts $z$, due to their
impact on the linear growth factor. Unfortunately, this effect is not
directly observable, since its detection would require very precise data at large $l$ and high
redshifts, where the late ISW effect is masked by primordial
temperature anisotropies.
The non-linear growth of density perturbations modifies the previous picture,
producing additional temperature perturbations which give rise to the
so-called ``Rees-Sciama'' (RS) effect, directly related to the
momentum density in the non-linear regime ({\emph e.g.}
\cite{RS68,Cooray2002,Seljak96,Spergel98,Merkel_isw}, and references therein).
In fact, the accelerated non-linear growth of structure
increases the depth of the potential wells in overdense
regions, resulting in a reduction of the total CMB temperature, with
respect to the linear case. This partially cancels the ISW effect in
hotter regions. On the contrary, the RS effect increases the ISW
effect in underdense regions, since the saturation of the density contrast in
voids further suppresses the growth of the gravitational
perturbations. Also in this case, massive neutrinos have an impact,
since they alter the RS effect in a scale- and redshift-dependent way,
owing to the neutrino free-streaming scale, $\lambda_{\rm fs}$.
Besides the ISW and RS effects, CMB photons undergo also the gravitational
lensing~\cite{Lewis06} and the Sunyaev-Zeldovich (SZ)
effects~\cite{SZ:1969} generated by LSS. Massive neutrinos alter the
related auto-~\cite{Roncarelli_etal_2015} and
cross-correlation functions, with different impacts at
the linear and non-linear levels, mainly depending on their effect on
the total matter power spectrum and cluster number counts.
In this work, we present for the first time in the literature a full reconstruction of
the total (linear and non-linear) ISW-RS effect in the presence of
massive neutrinos, together with its cross-correlations with
CMB-lensing and weak-lensing signals.
Previous works \cite{Cai_isw,Jubilee_isw} have provided a similar
reconstruction in the standard $\Lambda$CDM massless case.
The present analyses make use of all-sky maps extracted via ray-tracing
across the ``Dark Energy and Massive Neutrino Universe'' (DEMNUni)
simulations, which are the largest N-body simulations to date with a
particle neutrino component. At present, as we explain in more details
in \S~\ref{demnuni} below, these simulations are characterised by a
baseline $\Lambda$CDM cosmology to which we add neutrinos with
different total masses, $M_\nu\equiv$\smnu. In the next future, we plan to extend
the DEMNUni set with the inclusion of an evolving dark energy
background, with different equations of state $w$, in order to study
the degeneracy between $M_\nu$ and $w$ at the non
linear level.
This paper is organised as follows. In \S~\ref{demnuni} we
present the DEMNUni simulations. In
\S~\ref{map-making}, we explain the map-making procedure, in
\S~\ref{auto-power} we present
the ISW-RS and lensing signals extracted
from the simulations, focusing the discussion on the angular auto
power spectra, and presenting the cross-correlation signals in \S~\ref{cross-power}. Finally,
in \S~\ref{conclu} we draw our main conclusions.
\section{The DEMNUni simulations}
\label{demnuni}
The DEMNUni simulations have been conceived for the analysis of different
probes, like galaxy surveys and CMB data,
and their cross-correlations, in the presence of massive neutrinos. In particular,
in order to investigate simultaneously the neutrino impact on
different CMB secondary anisotropies, \eg the ISW-RS and weak-lensing
effects, we
have produced a set of simulations with a volume big enough
to include the very large scale perturbation modes, and, at the same time, with a good mass resolution to
investigate the effects of small-scale non-linearities and neutrino
free-streaming. Moreover, for the accurate reconstruction of the
light-cone back to the starting redshift of the simulations, we have
assumed an output time-spacing small enough that possible systematic
errors, due to the interpolation between neighbouring redshifts along
the line of sight, result to be negligible.
The DEMNUni simulations have been performed using the tree particle
mesh-smoothed particle hydrodynamics (TreePM-SPH) code
\gadgetthree~\cite{springel05}, specifically modified by~\cite{viel10} to
account for the presence of massive neutrinos. This modified version of
\gadgetthree\ follows the evolution of CDM and neutrino particles,
treating them as two separated collisionless species. Given the
relatively high velocity dispersion, neutrinos have a characteristic clustering scale larger than
the CDM one, allowing to save computational time by neglecting the calculation of the
short-range gravitational force. This results in a different spatial resolution for the
two components, which for neutrinos is fixed by the PM grid (that we have
chosen to be eight times larger than the particle number), while for
CDM particles is about one order of magnitude higher.
The DEMNUni set of simulations has a starting redshift $z_{in}=99$,
and is characterised by a comoving volume of
(2 \hmone\ Gpc)\cub\, filled with 2048\cub\ dark matter
particles and, where present, 2048\cub\ neutrino particles. Given the large
amount of memory required by the simulations, baryon physics is not included.
The authors in Ref.~\cite{vanDaalen_etal_2011} found baryon effects to
be independent of cosmological parameters, suggesting that they
are also independent of the neutrino mass; therefore, our choice
should not affect the results presented in this work. This is
supported also by~\cite{Bird_etal2011}, where the authors show that
the neutrino induced suppression in the matter power spectrum is very much the
same also when neutrinos are considered in the presence of
baryons. Moreover, since we are seeking deviations (due to massive
neutrinos) from a fiducial reference $\Lambda$CDM model in terms of
$P(k)$ ratios, we expect that baryon feedback will cancel out in
this case, and that additional effects produced by the interplay of
neutrinos with baryon physics should be higher order in both (for
further details see~\cite{Castorina_etal_2015}).
We have produced a total of four different simulations, choosing the
cosmological parameters according to the
\planck\ results~\cite{planck14cp}, namely a flat \lcdm\ model
generalised to a \lcdmn\
framework by changing only the value of the sum of the three active neutrino masses $M_\nu=$\smnut,
respectively, and keeping fixed \omegam\ and the amplitude of primordial curvature
perturbations $A_{\rm s}$.
The simulations are characterised by a softening length
$\varepsilon=20$\hmone\ Kpc, and have been run on the Fermi IBM BG/Q
supercomputer at CINECA\footnote{http://www.cineca.it/},
Italy, employing about 1 Million cpu-hrs per
simulation (including the production of halo and sub-halo catalogues).
For each simulation we have produced 62 output logarithmically equispaced in the scale factor
$a=1/(1+z)$, in the redshift interval $z=0-99$, 49 of which lay
between $z=0$ and $z=10$. For each of the 62 output times, we have dumped
on-the-fly a particle snapshot composed by both CDM and neutrino
particles, a three-dimensional (3D) grid of the
gravitational potential, $\Phi$, with side size $L_{\rm
box}=2$\hmone\ Gpc and a mesh of 4096\cub\ cells, and a 3D grid of the time
derivative $\dot{\Phi}$, with same dimensions and resolution, for a
total of about 90 TB of data per simulation. Finally, in order to build halo catalogues,
we have post-processed each of the 62 particle snapshots with the
friends-of-friends (FoF) algorithm, included in
\gadgetthree~\cite{springel01,dolag09},
setting to 32 the minimum number of particles, thus fixing the halo
minimum mass to $M_{\rm FoF}\simeq2.5\times
10^{12}$\hmone\msun. Finally, the FoF catalogues have been processed via the
\subfind\ algorithm (also included in \gadgetthree\ ) so that the
initial FoF parent halos are split into multiple sub-halos, with the
result of an increase in the total number of identified objects and of
a lower minimum mass limit (for further details see~\cite{Castorina_etal_2015}).
In Fig.~\ref{pk} the power spectra extracted from the simulations
for two total neutrino masses are shown (upper panel), together with their ratios
with respect to the $\Lambda$CDM case (lower panel), for different neutrino
masses and at two different redshifts. Worth of note is the range of scales sampled by the
simulations, more than three order of magnitudes, thanks both to the
peculiar large volume and mass resolution characterising the DEMNUni
set. The lower panel of Fig.~\ref{pk} shows the well-known non-linear
damping caused by massive neutrinos on the total matter power
spectrum. Our findings recover previous results in the literature
~\cite{2014JCAP...03..011V, 2014JCAP...02..049C, 2009JCAP...05..002B,
2008JCAP...08..020B, 2010JCAP...01..021B, 2013MNRAS.428.3375A,
2013JCAP...03..019V, Massara2014,
Wagner2012, Upadhye2014}, in particular the excess of power
suppression with respect to the linear theoretical expectations. As
we will show in the next Sections, the non-linear behaviour of the
total matter power spectrum proportionally affects the CMB-lensing
and weak-lensing potentials from LSS, and the larger the total neutrino
mass $M_\nu$ is, the greater its impact on lensing quantities is. On
the contrary, we will show that, for the ISW-RS effect, lighter
neutrinos produce, at intermediate scales, a larger effect on
$\dot{\Phi}$, due to their higher thermal velocities, $v_{\rm
th}(z,m_\nu) \simeq 158\,(1+z)\,(1 {\rm eV}/m_\nu)\,{\rm km/s}$, and smaller
free-streaming comoving wave number in Fourier space, $k_{fs} \equiv 2\pi/\lambda_{fs}/(1+z)$.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[width=0.5\textwidth]{Pk_mnu0.17_062.eps}&
\includegraphics[width=0.5\textwidth]{Pk_mnu0.53_062.eps}\\
\includegraphics[width=0.5\textwidth]{Pknu_ratios_036.eps}&
\includegraphics[width=0.5\textwidth]{Pknu_ratios_048.eps}
\end{tabular}
\end{center}
\caption{Top panel: the total matter $P(k, z=0)$ for $M_\nu=0.17,
0.53$ eV, measured from the simulations (violet stars) compared to
the non-linear (NL) total matter $P(k)$ extracted from
\CAMB\ (solid blue line) combined with the \Halofit\ \cite{Smith_etal_2003} non-linear
corrections, including the neutrino contribution from \cite{Bird_etal2011}. Orange
triangles and red dashed lines represent the corresponding
neutrino $P(k)$ from simulations and \CAMB, respectively.
Bottom panel: ratios of the simulated total matter $P(k)$ for
$M_\nu=0.17, 0.30, 0.53$ eV (stars, triangles, diamonds,
respectively) wrt the simulated $P(k, M_\nu=0 \;{\rm eV})$ at $z=0.485,
1.458$. The dotted blue, dashed red, and dot-dashed orange lines
represent the linear (L) expectations from \CAMB.}
\label{pk}
\end{figure*}
\section{ISW-RS and weak-lensing potential reconstruction}
\label{map-making}
The temperature anisotropies induced by the total (linear and
non-linear) ISW-RS effect, in a direction ${\bf \hat{n}}$ on the sky, can be
computed as the integral of the time derivative of the physical
peculiar gravitational
potential, $\dot{\Phi}$, along the line of sight from the last
scattering (LS) surface to the present epoch at $t_0$~\cite{Sachs67}.
\begin{equation}
\label{eq1}
\Delta T ({\bf \hat{n}}) = \frac{2}{c^2}\bar{T_0}\int_{t_{\rm LS}}^{t_0}
\dot{\Phi}(t,{\bf \hat{n}}) \, dt,
\end{equation}
where $t$ is the cosmic time, $t_{\rm LS}$ the age of the Universe at
the LS surface, $\bar{T_0}=2.7255 {\rm K}$ the today CMB temperature,
and $c$ the speed of light. Equation~(\ref{eq1}) can be rewritten as
the integral over the radial comoving distance, $r$,
\begin{equation}
\label{eq1b}
\Delta T ({\bf \hat{n}})
= \frac{2}{c^3}\bar{T_0}\int_{0}^{r_{\rm LS}}
\dot{\Phi}(r{\bf \hat{n}})\, a \, dr,
\end{equation}
where $r_{\rm LS}$ is the radial comoving distance to the LS surface,
and $a$ is the scale factor of the Universe.
Analogously, the integral for the projected CMB lensing potential due to scalar perturbations with no
anisotropic stress reads
\begin{equation}
\label{eq2b}
\phi({\bf \hat{n}})\equiv
-2\int_0^{r_{\rm LS}} \frac{r_{\rm LS}-r}{r_{\rm LS}r}\,\frac{\Phi(r{\bf\hat{n}};c\eta_0-r )}{c^2}\,{\rm d}r\,,
\end{equation}
where $\eta_0$ is the present conformal time, and $\Phi$ is the physical
peculiar gravitational potential generated by density perturbations. For the purposes of this
work, the line-of-sight integration is made in the so-called
``Born-approximation'' along the \emph{undeflected} photon path,
which, for a given particle mass resolution, results to be accurate at
sub-percent level for weak-lensing and CMB-lensing calculations up to
$l\lesssim 3000$~\cite{Hilbert_etal_2007,Calabrese_etal_2015}.
The corresponding deflection-angle integral is
\begin{equation}
\label{deflection_angle}
\boldsymbol{\alpha}({\bf \hat{n}})\equiv
-2\int_0^{r_*}
\frac{r_{\rm LS}-r}{r_{\rm LS}r}\,\nabla_{\bf\hat{n}}\frac{\Phi(r{\bf\hat{n}};c\eta_0-r )}{c^2}\, {\rm d}r\,,
\end{equation}
where $[1/r]\nabla_{\hat{\bf n}}$ is
the two dimensional (2D) transverse derivative with respect to the
line-of-sight pointing in the direction ${\hat{\bf
n}}\equiv(\vartheta,\varphi)$.
We implement Eqs.~(\ref{eq1b})-(\ref{eq2b}) in our code
for CMB ray-tracing across the simulated $\Phi$ and $\dot{\Phi}$
distributions, in order to produce all-sky ISW-RS and weak-lensing
maps, as described in \S~\ref{map-macking} below.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[angle=90,origin=c,width=0.5\textwidth]{lcdm_isw_a2H_fac_totmap_projpot_4_48_Ho_minus_sign.eps}&
\includegraphics[angle=90,origin=c,width=0.5\textwidth]{nu0.17_isw_a2H_fac_totmap_projpot_4_47_Ho_minus_sign.eps}\\
\includegraphics[angle=90,origin=c,width=0.5\textwidth]{lcdm_cmbLens_totmap_projpot.eps}&
\includegraphics[angle=90,origin=c,width=0.5\textwidth]{nu0.53_cmbLens_totmap_projpot.eps}
\end{tabular}
\end{center}
\caption{Top panel: projected mono-dipole-subtracted full-sky maps of
the ISW-RS signal for $M_\nu=0$ eV (left) and $M_\nu=0.17$ eV
(right). They do not include primary
CMB anisotropies.
Bottom panel: projected mono-dipole-subtracted full-sky
maps of the lensing potential for $M_\nu=0$ eV (left) and
$M_\nu=0.53$ eV (right); bars indicate the modulus and orientation of the
deflection-angle field $\boldsymbol{\alpha}$.}
\label{fig_maps}
\end{figure*}
\subsection{Map-making procedure}
\label{map-macking}
As mentioned in \S~\ref{demnuni}, during the production stage of the
DEMNUni simulations, using a properly modified version of
\gadgetthree, we have dumped on-the-fly 62 cubic grids of
$\Phi$ and $\dot{\Phi}$, with a mesh of $4096^3$ cells, each of $\simeq 0.5h^{-1}$Mpc on
a side\footnote{The gravitational potential itself has been
calculated by first assigning the particles to the mesh with the
clouds-in-cells mass assignment scheme. The resulting density field has
then been Fourier transformed, multiplied with the Green's function of
the Poisson equation in Fourier space, and then transformed back to
obtain the potential. Also, a slight Gaussian smoothing on a scale $r_s$
equal to 1.25 times the mesh size has been applied in Fourier space in
order to eliminate residual anisotropies on the scale of the mesh, and a
deconvolution to filter out the clouds-in-cells mass assignment kernel
has been applied as well. The time derivative of the gravitational
potential has been computed via two-sided differentiation of two
potential grids dumped at two different step-times immediately
subsequent to the output times of each particle snapshot.}. Hence, the field
$\Phi$ corresponds to the density field of the simulations smoothed on
a scale of about $500\,h^{-1}{\rm kpc}$. This resolution is good enough to resolve scales $\gtrsim
10~h^{-1}$~Mpc, relevant for the effects analysed in this work.
As explained in \S~\ref{demnuni}, the resolution of the N-body
simulation (which contains structures down to the gravitational
softening length of $20\,h^{-1}{\rm kpc}$) is much greater, but not
necessary for the present study.
In order to build mock all-sky maps of the CMB temperature anisotropies $\Delta T$
described in Eq.~(\ref{eq1}), we employ the map-making procedure
developed by~\cite{Carbone_etal_2007}, adapted to the ISW-RS effect,
\ie CMB photons are ray-traced along the undeflected line of sight through
the 3D field $\dot{\Phi}$. We apply the same kind
of ray-tracing also to the 3D $\Phi$-grids, in
order to produce the same realisation of the Universe and compute the
cross-correlation signal between the ISW-RS temperature maps and the
CMB/weak-lensing potential maps.
To this aim, we stack the $\Phi$- and $\dot{\Phi}$-grids around the
observer, located at $z=0$, applying the replication and randomisation
procedure designed by~\cite{Carbone_etal_2007}.
This particular 3D tessellation scheme is required to avoid both the
repetition of the same structures along the line
of sight, and the generation of artifacts like
ripples in the simulated deflection-angle field, which can be avoided
only if the peculiar gravitational potential is continuous transversely to each
line of sight. With this procedure we produce a simulated volume
around the observer which is large
enough to carry out the integration over all the redshifts relevant
to this work. Finally, we select a pixelisation of the sky with a set of
directions ${\bf\hat{n}}\equiv(\vartheta,\varphi)$, following
the standard approach introduced by the
\textsf{HEALPix}\footnote{http://healpix.sourceforge.net} hierarchical
tessellation of the unit sphere~\cite{Healpix}.
In order to extract the impact of massive neutrinos on the ISW-RS
effect, mainly dominated by the free-streaming at high
redshifts~\cite{Lesgourgues_etal_2008}, we integrate $\dot{\Phi}$ along the line of sight up to $z
\simeq 21$ (for further details on the interpolation and integration
scheme see~\cite{Carbone_etal_2007}). For the simulated CMB lensing signal, previous
studies~\cite{Carbone_etal_2008,Carbone_etal_2012} indicate that an
integration up to $z \simeq 99$ is sufficient to recover mostly $\sim 99$\%
of the power. In the case of weak-lensing, we produce all-sky maps of
the lensing potential with different source redshifts $z_s$. In particular, we consider
that all the sources are placed on a spherical surface at redshifts
$z_s=2,5.5,8$, respectively. This is done for illustrative purposes,
\ie to quantify the impact of neutrino free-streaming at different
redshifts on the weak-lensing (WL) signal and its cross-correlation
with the total ISW-RS effect.
In the upper panel of Fig.~\ref{fig_maps}, we show projected
mono-dipole-subtracted full-sky maps of
the ISW-RS signal for the massless case (left top panel) and for $M_\nu=0.17$ eV (right top
panel). These maps represent the ISW-RS contribution alone to the CMB
temperature, Eq.~(\ref{eq1b}), and do not include primary CMB
anisotropies. In the lower panel, projected mono-dipole-subtracted full-sky
maps of the lensing potential
in the two cases, $M_\nu=0$ eV and $M_\nu=0.53$ eV, are
presented; bars here indicate the modulus and orientation of the
deflection angle field which represents the spatial gradient of the
lensing potential map, as defined in Eq.~(\ref{deflection_angle}).
These maps have been obtained via the
map-making technique described above, using a \textsf{HEALPix} pixelisation
parameter $N_{\rm side}=2048$, and have an angular resolution of $\simeq
1.72^\prime$, with 50331648 pixels in total.
In Fig.~\ref{fig_maps}, the correlation between the maps in the upper
and lower panels, \ie between the ISW-RS and lensing potential
realisations, is clearly visible, and shows how the origin of these two effects is due to
the same LSS distribution crossed by CMB photons.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[angle=90,origin=c,width=0.5\textwidth]{lcdm_nu0.53_diff_lensz1_totmap_projpot_synfast.eps}&
\includegraphics[angle=90,origin=c,width=0.5\textwidth]{lcdm_nu0.53_diff_cmbLens_totmap_projpot_synfast.eps}\\
\end{tabular}
\end{center}
\caption{Modulus of the deflection-angle vector difference between the massless case, $M_\nu=0$ eV, and the
massive neutrino case, $M_\nu=0.53$ eV. Left: all the
galaxy sources are supposed to be located at $z_s=1$. Right: the source is represented by CMB photons at
$z_s=1100$.}
\label{fig_diff_maps}
\end{figure*}
\section{The impact of massive neutrinos on auto power spectra}
\label{auto-power}
Massive neutrinos produce noticeable effects on CMB secondary
anisotropies, which in some cases, \eg for large neutrino masses,
can be even visually inspected. Let us first notice some
visible differences between the massless and massive neutrino cases:
in the lower panels of Fig.~\ref{fig_maps}, the neutrino
free-streaming suppresses, as expected, the lensing potential signal
with respect to the $\Lambda$CDM scenario (this can be especially observed by
an eye inspection of the filaments in the two maps, and from the bar
units of the modulus of the deflection angle); on the other
hand, in the upper panels, the effect is opposite, \ie in the
ISW-RS case the neutrino free-streaming produces a slight excess of
power with respect to the massless case, and this is caused by the
$\dot{\Phi}$ induced by hot neutrinos, mostly at high redshifts and
for lighter neutrino masses (for this reason the temperature range represented by the colour bar is
larger for the $M_\nu=0.17$ eV case). We
will explain this effect in more details in the next section, but let
us anticipate that smaller neutrino masses produce a
larger excess in the CMB temperature power spectrum.
In Fig.~\ref{fig_diff_maps} we show the modulus of the deflection
angle difference, between the massless case, $M_\nu=0$ eV, and the
massive neutrino case, $M_\nu=0.53$ eV. Here the plotted quantity is
$\sqrt{(\Delta \alpha_1)^2+(\Delta
\alpha_2)^2}$, where $\alpha_1$ and $\alpha_2$ are the two components of the
deflection angle $\boldsymbol{\alpha}$, and $\Delta \alpha_i$, with $i=1,2$, stands for the
difference between each component in the two cases $M_\nu=0$ eV and
$M_\nu=0.53$ eV.
In the left panel all the
galaxy sources are supposed to be located at $z_s=1$, while in the
right panel the source is represented by CMB photons at
$z_s=1100$. These maps visually show the suppression in structure
formation due to the presence of massive neutrinos, for two different cases of
weak-lensing. Given the limited spatial resolution of the potential grids,
smaller angular scales are captured via ray-tracing as
the source redshift increases, such that the right panel in
Fig.~\ref{fig_diff_maps} shows much more ``non-linear'' features with respect to the left panel.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_cambApril14_keta14000.eps}&
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_percentage_cambApril14_keta14000.eps}
\end{tabular}
\end{center}
\caption{Left: simulated angular power spectra of the total ISW-RS induced temperature
anisotropies, for $M_\nu=0,0.17,0.3,0.53$ eV (black solid, blue
dashed, red dot-dashed, and orange tri-dot-dashed lines,
respectively). Light-green symbols represent the linear contribution (ISW alone) from
\CAMB\ in the massless case, the vertical line at
$l\simeq 100$ corresponds roughly to the transition between the linear and
non-linear regimes.
Right: percent residuals, with respect to the
massless case, of the ISW-RS power spectra obtained via direct ray-tracing across the simulations.}
\label{fig_iswrs}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_0_1500Mpc.eps}&
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_percentage_0_1500Mpc.eps}\\
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_1500_2500Mpc.eps}&
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_percentage_1500_2500Mpc.eps}\\
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_2500_7000Mpc.eps}&
\includegraphics[width=0.5\textwidth]{iswTOT_a2H_fac_deconv_comparison_percentage_2500_7000Mpc.eps}
\end{tabular}
\end{center}
\caption{Left: simulated power spectra of the ISW-RS induced
anisotropies computed by integration in different redshift bins. Right: corresponding percent residuals with respect to the massless case.}
\label{fig_iswrs_z}
\end{figure*}
\subsection{The total ISW-RS effect}
\label{ISWRS}
The simulated power spectra of the total ISW-RS induced temperature
anisotropies, obtained via ray-tracing from $z=0$ to $z\simeq 21$, are
shown in the left panel of Fig.~\ref{fig_iswrs} for
four different total neutrino masses, $M_\nu=0,0.17,0.3,0.53$ eV. They
are represented by the black (solid), blue (dashed), red (dot-dashed),
and orange (tri-dot-dashed) lines, respectively. The light-green
symbols represent the linear contribution (ISW alone) from
\CAMB\footnote{http://camb.info} in the massless case, while the vertical line at
$l\simeq 100$ corresponds roughly to the transition between the linear and
non-linear regimes \cite{Scahefer2006,Scahefer2008}. Let us notice the good agreement between the
theoretical expectations from \CAMB\ and the simulation outputs at
multipoles $l>10$; at $l<10$ the simulated signals show a lack of power with
respect to the predictions, due to the finite size of the
simulation box; on the other hand we succeed in reproducing the non-linear
contribution for $l>100$, scales at which \CAMB\ predictions fail,
given that the RS contribution is not implemented in the code.
In the right panel of Fig.~\ref{fig_iswrs} it is possible to appreciate
the differences produced in the ISW-RS induced temperature
anisotropies by free-streaming neutrinos with different total
masses. In particular, here we show the
percent residuals with respect to the massless case; for
small masses, $M_\nu=0.17\,,0.3$ eV, we observe an excess of power of
about $10$\% and $5$\% respectively, at $50 \lesssim l \lesssim
150$, corresponding to the transition
from linear to non-linear regimes. This excess is indeed expected to
originate from the $\dot{\Phi}$ term induced by the slow decay of
gravitational and matter perturbations produced by hot neutrinos at
intermediate cosmological scales, happening both during the matter and
dark energy domination eras. On the other hand, for large neutrino masses, \eg
$M_\nu=0.53$ eV, which correspond to smaller neutrino thermal
velocities, the total effect consists in a power suppression, similar to what
happens for the lensing potential in the presence of massive
neutrinos.
Finally, on fully non-linear scales $l>200$, where the signal is
totally due to the RS effect, massive neutrinos decrease the total
temperature power with respect to the massless scenario, and this suppression increases
with the neutrino mass. This is reasonably due to the
non-linear excess of suppression of the total matter power spectra
with respect to linear expectations in the presence of massive
neutrinos (see the left panel of Fig.~\ref{pk} for $k\lesssim 10$ $h$/Mpc,
corresponding to the $\dot{\Phi}$-grid resolution), which for lower
thermal velocities may dominate the total effect.
In order to understand how the effect of free-streaming massive
neutrinos evolves with $z$, in the left panels of
Fig.~\ref{fig_iswrs_z} we show the power spectra of the ISW-RS induced
anisotropies computed in different redshift bins. We observe that for
low redshifts, $z\lesssim 1$, neutrinos produce a larger suppression in $C_l^{TT}$ at
$l>100$ as their mass increases; in this case the slow down of
structure formation in the presence of massive neutrinos has the
dominant role. On the other hand, for larger
redshifts, $z\gtrsim 1$, far from the epoch of recent acceleration, we
see an excess of power which peaks at the transition scale, $l\simeq
100$, and is larger for lighter neutrinos; as pointed out in
\cite{Lesgourgues_etal_2008}, this is due to the behaviour of the
linear growth rate in the presence of massive neutrinos, which
produces an evolving gravitational potential even in the absence of
dark energy.
More quantitatively, in the right panels of
Fig.~\ref{fig_iswrs_z}, we show the percent residuals with respect to
the $\Lambda$CDM case: in the upper panels, for $z\lesssim 1$, the
power suppression in the $C_l^{TT}$ has a trend similar to the matter
$P(k)$, increasing at large multipoles and somewhat proportional to
the neutrino mass ratios. In the lower panel, instead, at higher redshifts,
$1\lesssim z \lesssim 21$, the produced excess of power can even reach
$50$\% for small neutrino masses at $l\simeq 100$. The combination of the lack
of power at low redshifts and the excess at high redshifts produces
the total effect shown in the left panel of Fig.~\ref{fig_iswrs}.
Incidentally, let us notice how the impact of the limited simulation volume increases
with redshift, producing a larger lack of power at $l<10$ for larger $z$ values.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[width=0.5\textwidth]{cmb_lensing_comparison_cambApril14_keta14000.eps}&
\includegraphics[width=0.5\textwidth]{cmb_lensing_comparison_percentage_cambApril14_keta14000.eps}\\
\includegraphics[width=0.5\textwidth]{TT_lensed_comparison_4096_l30_z99_b2_seed2_ratios_noxlog.eps}&
\includegraphics[width=0.5\textwidth]{EE_lensed_comparison_4096_l30_z99_b2_seed2_ratios_noxlog.eps}\\
\includegraphics[width=0.5\textwidth]{BB_lensed_comparison_4096_l30_z99_b2_seed2.eps}&
\includegraphics[width=0.5\textwidth]{BB_lensed_comparison_4096_l30_z99_b2_seed2_ratios.eps}
\end{tabular}
\end{center}
\caption{Upper left panel: CMB lensing potential angular
power spectrum obtained by photon ray-tracing, from $z=0$ up to
$z=99$, across the total gravitational field of the DEMNUni
simulations. The dashed light-green curve represents the prediction
from \CAMB\ in the massless case; the dotted black, blue, red, and
orange lines represent the simulated lensing potential for
$M_\nu=0,0.17,0.3,0.53$ eV, respectively.
Top right panel: percent residuals
between the massive and massless neutrino cases. Here, tri-dot-dashed blue, red, and
orange lines are the simulated signals for $M_\nu=0,0.17,0.3,0.53$ eV, respectively.
Semi-analytical non-linear \CAMB\ predictions are represented by symbols, as
described in the legend.
Middle left panel: percent residuals, wrt to the massless case, of the simulated TT lensed power spectra for
$M_\nu= 0.17,0.3,0.53$ eV.
Middle right panel: percent residuals, wrt to the massless case, of the simulated EE lensed power spectra for
$M_\nu= 0.17,0.3,0.53$ eV.
Lower left panel: simulated lens-induced BB angular power
spectra for $M_\nu=0,0.17,0.3,0.53$ eV.
Lower right panel: percent residuals, wrt to the massless case, of the
simulated lens-induced BB power spectra for
$M_\nu= 0.17,0.3,0.53$ eV.}
\label{fig_cmblens}
\end{figure*}
\subsection{CMB and Weak lensing spectra}
Let us now focus the discussion on the effect of massive neutrinos on
lens-induced CMB secondary anisotropies. Since lensing traces
directly the total matter power spectrum, \ie the
gravitational potential $\Phi$ rather than the time derivative
$\dot{\Phi}$, the theoretical linear predictions from
\CAMB, combined with the \Halofit\ \cite{Smith_etal_2003} non-linear
corrections and including also the neutrino contribution from
\cite{Bird_etal2011}, succeed in reproducing mostly perfectly (as
compared to the simulated signal) the CMB-lensing
effects in the presence of massive neutrinos on all the scales of interest.
This result is presented in the top panels of Fig.~\ref{fig_cmblens};
in particular, the left panel shows the CMB lensing potential angular
power spectrum obtained by photon ray-tracing, from $z=0$ up to
$z=99$, across the total gravitational
field, generated both by CDM and massive neutrinos, from the DEMNUni
simulations. The dashed light-green curve represents the prediction
from \CAMB\ in the $\Lambda$CDM case; the dotted black, blue, red, and
orange lines represent the simulated lensing potential for
$M_\nu=0,0.17,0.3,0.53$ eV, respectively. Comparing the black and
light-green lines, \ie the $\Lambda$CDM curves, we observe
less power in the simulated spectrum with respect to
\CAMB\ predictions, at multipoles $l\gtrsim 1000$. This is due to the finite resolution of the
DEMNUni potential grids, about $\sim 0.5$ Mpc/$h$, much smaller
compared to the resolution of the dark matter simulations from
\cite{Takahashi_etal_2012} used for non-linear corrections in
\Halofit. On the other hand, looking at the top
right panel of Fig.~\ref{fig_cmblens}, \ie at the percent residuals
between the massive and massless neutrino cases, we do not observe
such lack of power with respect to the semi-analytical
\CAMB\ predictions (compare tri-dot-dashed lines against symbols),
implying that resolution effects cancel out
when focusing on the relative differences with respect to the
$\Lambda$CDM case. In particular, as expected, the integrated effect of
massive neutrinos produces a suppression of power in the
lensing potential
which increases with the neutrino mass, decreases with the
angular scale (low $l$), and at large multipoles tends to become
constant, approximately proportional to the ratios of the total
neutrino masses. We find an asymptotic power suppression of about $\Delta
C_l^{\phi\phi}\simeq 10\%, 19\%, 31\%$ for $M_\nu=0.17,\,0.30,\,0.53$ eV,
respectively. As mentioned above, this trend is directly related to
the behaviour of the total matter power spectra shown in
Fig.~\ref{pk}, as the lensing potential power spectrum can be written as the integral along the line
of sight of the matter power spectrum weighted by a geometrical
factor.
In the middle and lower panels of Fig.~\ref{fig_cmblens} we show the
effect of massive neutrinos on the lensed CMB temperature (TT), lensed E-mode
polarisation (EE), and lens-induced B-mode polarisation (BB) angular
power spectra. As in
Refs.~\cite{Carbone_etal_2008,Calabrese_etal_2015},
the lensed power
spectra have been obtained by modifying the
\LensPix\ code\footnote{http://cosmologist.info/lenspix/}\cite{Lewis2005},
in order to use directly as input the power spectra and phases of the
lensing potential
maps produced by ray-tracing across the DEMNUni
simulations. As expected, the percent residuals between the
massive and massless neutrino cases, presented in the middle panel of
Fig.~\ref{fig_cmblens}, show that free-streaming neutrinos alter the lensed TT
and EE power spectra. In particular, as the strength of the
gravitational potential decreases for larger neutrino masses, CMB
acoustic oscillations are less lensed, {\emph i.e.} they
are less smoothed and smeared out than in the massless case. This implies
that the larger the neutrino mass is, the larger the lensed TT and EE
power spectra are at large angular scales (small $l$), and the smaller
the lensing amplification of the so-called ``damping-tail'' at
small scales (high $l$) is. As a consequence, the production of the lens-induced B-mode
polarisation, via rotation of the E-mode pattern, is less enhanced in
the presence of massive neutrinos, and the amplitude of the BB power
spectrum decreases with increasing neutrino masses on all the
scales, as shown in the left lower panel of Fig.~\ref{fig_cmblens}.
As the right panel shows, the B-mode lack of
power due to the presence of massive neutrinos is mostly scale
independent (except at $l>1000$ where non-linear effects come
into play) \cite{Oyama2013, Oyama2016}. The reason is well known, {\emph i.e.} B-modes are due to
the transfer of E-mode power from small scales to large scales, and on
scales $1000<l<2000$ the effect of massive neutrinos on E-modes is
almost scale-independent, as the right middle panel of
Fig.~\ref{fig_cmblens} shows. Again, the B-mode power residuals
are somewhat proportional, for $l\rightarrow 0$, to the ratio of the neutrino masses
assumed in the simulations.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[width=0.5\textwidth]{lensing_pot_comparison_zs2.eps}&
\includegraphics[width=0.5\textwidth]{lensing_pot_comparison_percentage_zs2.eps}\\
\includegraphics[width=0.5\textwidth]{lensing_pot_comparison_zs5.5.eps}&
\includegraphics[width=0.5\textwidth]{lensing_pot_comparison_percentage_zs5.5.eps}\\
\includegraphics[width=0.5\textwidth]{lensing_pot_comparison_zs8.eps}&
\includegraphics[width=0.5\textwidth]{lensing_pot_comparison_percentage_zs8.eps}
\end{tabular}
\end{center}
\caption{Left: lensing potential angular power spectra for
sources all placed at redshifts $z=2$ (upper panel),
$z=5.5$ (middle panel), and $z=8$ (lower panel), respectively. The dotted black, blue, red and
orange lines represent the simulated signals for
$M_\nu=0,0.17,0.3,0.53$ eV, respectively.
The $\Lambda$CDM non-linear expectations from \CAMB\ are represented
by the dashed green line.
Right: corresponding percent residuals wrt the massless
case. Here, tri-dot-dashed blue, red, and
orange lines are the simulated signals for $M_\nu=0,0.17,0.3,0.53$ eV, respectively.
Semi-analytical non-linear \CAMB\ predictions are represented by symbols, as
described in the legend.}
\label{fig_weaklens}
\end{figure*}
Finally, in Fig.~\ref{fig_weaklens} we present the results for three
cases of weak-lensing, with sources all placed at redshifts $z=2$,
$z=5.5$, and $z=8$, respectively. As expected, in all the cases, we find the
lensing potential to behave in a way very similar to the CMB lensing potential. In
fact, its angular power spectrum is suppressed by
free-streaming massive neutrinos, and such suppression is scale
dependent, increasing with the multipoles $l$, as new scales $k$
exceed the free-streaming scale $k_{fs}$. We have modified \CAMB\ to
account for the computation of the weak-lensing signal for sources
placed all at the same redshift $z_s$, by substituting $z_{\rm LS}$ with
$z_s$ in the \CAMB\ routine ``equations.f90'' for the calculation of
the ``sources(3)'' variable; the corresponding analytical
expectations are represented by the dashed green line in the left
panels of Fig.~\ref{fig_weaklens}. As expected, given the finite
simulation volume, on large scales the simulated signal recovers
much better CAMB predictions as redshift decreases, while on small
scales we observe an increase of the lack of power due the limited
resolution, $\sim 0.5$ Mpc/$h$, of the gravitational potential grids used for
ray-tracing. Nonetheless, when looking at the relative differences
between the massive and massless cases, for different $M_\nu$ values,
the simulated signal recovers the expectations within
the $\sim 1$\% accuracy level (see the right panel of Fig.~\ref{fig_weaklens}).
This means that the effect of massive neutrinos on LSS formation
decouples from highly non-linear regime physics, as {\emph e.g.}
baryon effects, since, as well known, massive neutrinos escape the
potential field of small scale structures, causing such effects
to cancel out when considering relative differences.
\section{The impact of massive neutrinos on cross power spectra}
\label{cross-power}
Let us now consider the effects of massive neutrinos on
the cross angular power spectra between CMB/weak-lensing and the
ISW-RS signals. Since the ISW-RS effect is not directly observable,
the ISW-RS cross-correlation with weak-lensing
(together with its cross-correlation with galaxies, which we do not
discuss in this work)
allows to observe and measure the impact of
time-varying potentials on light travelling to us from the last
scattering surface.
It is worth noting that here, for the first time in the literature, we
present the non-linear behaviour of such cross-correlation signal, as extracted from
N-body simulations accounting for free-streaming massive
neutrinos. The linear counterpart can be computed using Boltzmann
codes as \CAMB\ or \CLASS\footnote{http://class-code.net/} \cite{CLASS}.
\subsection{ISW-RS--CMB-lensing cross-correlations}
\label{ISWRS--CMB-lensing cross}
The absolute values of the simulated cross power spectra between
CMB-lensing and the total ISW-RS induced temperature
anisotropies, obtained via ray-tracing from $z=0$ to $z\simeq 21$, are
shown in the left panel of Fig.~\ref{fig_cmbl_isw_cross} for
four different total neutrino masses, $M_\nu=0,0.17,0.3,0.53$ eV. They
are represented by the black (solid), blue (long-dashed), red (dashed),
and orange (dot-dashed) lines, respectively. The violet
tri-dot-dashed line represents the linear contribution from
\CAMB\footnote{As already mentioned in \ref{ISWRS}, at present Boltzmann codes are
not able to compute the non-linear contribution to $\dot{\Phi}$.} in
the massless case. Above all, let us observe that, using a finite box of
$2$ Gpc/$h$, we manage to recover the predicted linear signal starting
from very low multiples, $l\sim 10$. This is a confirmation of the
accuracy of the technique implemented to extract the ISW effect from the
DEMNUni simulations.
In addition, at the transition from
linear to non-linear scales, we recover also the expected sign inversion of the
cross-correlation spectra due to the negative
\emph{non-linear} correlation between the RS effect and matter density, which becomes dominant with respect to the net positive \emph{linear} correlation between
density and CMB temperature, produced by the decay of the
\emph{linear} gravitational potential, as the Universe expands in the
presence of dark energy \cite{Smith_etal_2009, Schaefer2011}.
This leads to a sign change of the cross-spectrum, which for a $\Lambda$CDM
model with the same cosmological parameters as in the
simulations, happens to be about at $l\sim 700$ \cite{Mangilli13, Lewis_RS}.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[width=0.5\textwidth]{iswCMBL_cross_a2H_fac_deconv_ABScomparison_choose_z_isw_late_keta14000_lewis.eps}&
\includegraphics[width=0.5\textwidth]{iswCMBL_cross_a2H_fac_deconv_comparison_percentage_choose_z_isw_late_keta14000.eps}\\
\includegraphics[width=0.5\textwidth]{iswCMBL_cross_a2H_fac_deconv_comparison_choose_z_isw_late_keta14000_lewis_new_sign_inversion.eps}&
\includegraphics[width=0.5\textwidth]{iswCMBL_cross_a2H_fac_deconv_comparison_choose_z_isw_late_keta14000_lewis_new_small.eps}
\end{tabular}
\end{center}
\caption{Left upper panel: absolute values of the simulated cross power spectra between
CMB-lensing and the total ISW-RS induced temperature
anisotropies, obtained via ray-tracing from $z=0$ to $z\simeq 21$,
for $M_\nu=0,0.17,0.3,0.53$ eV (solid black, long-dashed blue, dashed
red, and dot-dashed orange lines, respectively). The violet
tri-dot-dashed line represents the linear contribution from \CAMB\ in the
massless case.
Right upper panel: corresponding percent residuals wrt the massless
case on linear scales $l<400$. Here, blue, red, and
orange lines are the simulated signals for $M_\nu=0,0.17,0.3,0.53$ eV, respectively.
Semi-analytical linear \CAMB\ predictions are represented by symbols, as
described in the legend.
Bottom panels: simulated cross power
spectra for $l>400$. The shaded
grey area represents the cosmic variance associated to the signal.}
\label{fig_cmbl_isw_cross}
\end{figure*}
As the left top panel of Fig.~\ref{fig_cmbl_isw_cross} shows,
the major effect of massive neutrinos on the CMB-lensing/ISW-RS cross
spectrum consists of moving the sign inversion position
toward larger multipoles, producing a larger displacement as the
neutrino mass increases. This can be explained considering
that the larger the neutrino mass is, the larger the suppression of
structure formation is, and therefore cosmological perturbations tend to
stay in the linear regime on smaller scales than in the massless
case. This implies that free-streaming massive neutrinos not only
produce an excess of ISW-RS power, but also an excess of cross-correlation between CMB-lensing and the ISW-RS effect, and
this time such excess increases with larger neutrino masses on scales
$100<l<1000$, being a factor of $\sim 4$ for $M_\nu=0.3$ eV, at
$l\sim 600$. On multipoles $700<l<1000$, the shift of the sign inversion
is the the dominant feature, and, indeed, a future detection and
measurement of its position could be a further probe of the total
neutrino mass. On larger multipoles the
asympthotic suppression of the matter power spectra due to neutrino
free-streaming becomes dominant, and we recover
the usual trend, {\emph{i.e.}} larger neutrino masses produce a larger decrease of
the non-linear cross power spectrum.
In the right top panel of Fig.~\ref{fig_cmbl_isw_cross} we compare
our findings with \CAMB\ predictions on linear scales. On
multipoles $10<l<400$ the accuracy of the reconstructed signal is
very high, about $1-2$\%, and the residuals start to
increase only when non-linear effects come into play, {\emph{i.e}} at
$l>400$. On such scales, the behaviour of the simulated cross power
spectra is shown in the bottom panels of
Fig.~\ref{fig_cmbl_isw_cross}. In particular, the different curves in left panel
represent the sign inversion due to non-linearities, and the shaded
grey area the associated cosmic variance $\sqrt{[(C_l^{T\kappa})^2+C_l^{TT} C^{\kappa\kappa}_l]/(2l+1)}$, which unfortunately makes
measurements of this effect quite challenging, since the primary CMB temperature
anisotropies act as a foreground in this case.
\begin{figure*}[!ht]
\begin{center}
\setlength{\tabcolsep}{0.01pt
\begin{tabular}{c c}
\includegraphics[width=0.5\textwidth]{cl_isw_weakL_cross_deconv_comparison_abs_zs2.eps}&
\includegraphics[width=0.5\textwidth]{cl_isw_weakL_cross_VS_weakL_deconv_comparison_zs2.eps}\\
\includegraphics[width=0.5\textwidth]{cl_isw_weakL_cross_deconv_comparison_abs_zs5.eps}&
\includegraphics[width=0.5\textwidth]{cl_isw_weakL_cross_VS_weakL_deconv_comparison_zs5.5.eps}\\
\includegraphics[width=0.5\textwidth]{cl_isw_weakL_cross_deconv_comparison_abs_zs8.eps}&
\includegraphics[width=0.5\textwidth]{cl_isw_weakL_cross_VS_weakL_deconv_comparison_zs8.eps}
\end{tabular}
\end{center}
\caption{Left: absolute values of the WL/ISW-RS angular
cross spectra for sources placed at redshifts $z=2$ (upper panel),
$z=5.5$ (middle panel), and $z=8$ (lower panel), respectively. The
solid black, long-dashed blue, dashed red, and dot-dashed
orange lines represent the simulated signals for
$M_\nu=0,0.17,0.3,0.53$ eV, respectively. The $\Lambda$CDM linear expectations from \CAMB\ are represented
by the tri-dotted violet line. The light-green dotted line is the linear
ISW-RS/CMB-lensing cross spectrum from \CAMB, here shown for
comparison.
Right: percent residuals wrt the massless
case for $l<300$. Tri-dot-dashed blue, red, and
orange lines correspond to the simulated signals of the auto WL power
spectra for $M_\nu=0,0.17,0.3,0.53$ eV, respectively.
Solid blue, red, and orange lines are the residulars of the simulated
ISW-RS/WL cross spectra. The corresponding
semi-analytical linear \CAMB\ predictions are represented by symbols, as
described in the legend.}
\label{fig_weakl_isw_cross}
\end{figure*}
\subsection{ISW-RS--weak-lensing cross-correlations}
\label{ISWRS--WEAK-lensing cross}
Finally, we consider the cross-correlation between the ISW-RS and the
weak-lensing signals. The left panels of
Fig.~\ref{fig_weakl_isw_cross} show the cross spectra obtained with
lensing sources all placed at $z_s=2,\,5.5,\,8$ and for
$M_\nu=0,0.17,0.3,0.53$ eV, represented by the black (solid), blue (long-dashed), red (dashed),
and orange (dot-dashed) lines, respectively. The violet
tri-dot-dashed line represents the corresponding linear contribution from
\CAMB, while the light-green dotted line is the linear
ISW-RS/CMB-lensing cross spectrum from \CAMB, here shown for comparison. As in \S
\ref{ISWRS--CMB-lensing cross}, also in this case we find a mostly
perfect agreement between CAMB linear predictions and the simulated
signals, at scales $30 \lesssim l \lesssim 400$. On smaller multipoles,
$l\lesssim 30$ window effects take place producing a lack of power in
the simulated cross signal. As for the ISW-RS/CMB-lensing cross
spectrum, also in this case on larger multipoles, $l\gtrsim 400$,
non-linear effects produce a sign inversion, whose position stays
however mostly constant with increasing neutrino masses, even if we
notice that it is more shifted toward larger multipoles as $z_s$
increases (as expected from the theory of linear perturbation evolution).
In the right panels of Fig.~\ref{fig_weakl_isw_cross} we show the
residuals, with respect to the massless case, of the simulated
ISW-RS/weak-lensing cross spectra (solid lines), together with
linear predictions from \CAMB\ (symbols). For comparison, we also show
corresponding residuals for the simulated weak-lensing auto
spectra (tri-dot-dashed lines).
For $z_s=2$ we find an excellent agreement within $1$\% accuracy; at
larger $z_s$ the agreement is still good with an accuracy of about
$2-3$\%. This lower accuracy is probably due to percent differences
in the ISW-RS reconstruction between \CAMB\ and our ray-tracing
technique. Overall, at $30\lesssim l\lesssim 300$ we observe a scale
dependent suppression of the cross signal which increases with the
neutrino mass. At $l\gtrsim 300$, the trend starts to reverse, and for
lighter neutrino masses the cross spectra exceed their value in the
massless case.
It is worth to note that the suppression, due to massive neutrinos, of
the ISW-RS/weak-lensing cross spectra is smaller than for the
weak-lensing auto spectra (compare solid lines against tri-dot-dashed
lines). Moreover, while in the latter case this suppression is mostly
constant with $z_s$, in the former case it decreases with the
increase of the source redshift, and finally, for very high values
of $z_s$, neutrino free-streaming produces an excess of signal with
respect to the massless case, as we observe also in the CMB case to a
greater extent.
This is due to the larger excess of ISW-RS in the presence of massive neutrinos at
higher redshifts, far from dark energy domination at low $z$ (see
Fig.~\ref{fig_iswrs_z}).
This excess somehow balances the suppression in the
weak-lensing signal, and was first pointed out by
\cite{Lesgourgues_etal_2008} for the case of the ISW-galaxy
cross-correlation. Similar arguments explained in their \S~{\bf B}
hold also in our case. It has been also shown \cite{Schaefer2011} that
ideal cosmic variance limited experiments may detect, via
cross-correlation with lensing, the non-linear RS effect with a
significance of $\sim 3 \sigma$ integrating up to multipoles
$\ell=3\times10^3$. However, this significance is drastically reduced by the finite resolution
and noise of actual CMB experiments, so that the signal to noise ratio of the
non-linear RS effect results to be an order of magnitude smaller compared to that of
the linear ISW effect. This implies that it would be quite difficult to
measure it with present and near future CMB-LSS experiments.
Together with the results presented in \S~\ref{ISWRS--CMB-lensing
cross}, these represent the main findings of the present work. For the
first time in the literature, the cross-correlation between CMB/weak-lensing
and ISW-RS effects have been simulated, on a very large range of scales,
from the linear to the fully non-linear regimes, and in the presence of
massive neutrinos.
\section{Conclusions}
\label{conclu}
In this work we present full-sky maps, auto and cross angular spectra of the ISW-Rees-Sciama and
CMB/weak-lensing signals, from the linear to the fully non-linear
regimes, as extracted via direct ray-tracing across very large N-body
simulations including a massive neutrino component, the so-called
DEMNUni simulations. We assume a Planck-like baseline cosmology, and
add neutrinos with total masses $M_\nu=0,\,0.17,\,0.3,\,0.53$ eV,
fixing the normalisation of the matter power spectrum at CMB.
The analysis of these signals shows that
\begin{itemize}
\item
Free-streaming massive neutrinos induce a time variation, $\dot{\Phi}$,
of the gravitational potential, affecting mostly scales corresponding to the
transition from the linear to the non-linear regimes, $l\sim 100$, and producing
a non negligible contribution to the ISW effect (see Fig.~\ref{fig_iswrs}). The induced
$\Delta$T anisotropies are more important for light neutrino masses and
at high redshifts $z\gtrsim 1.5$, when dark energy is not the dominant component. At lower
redshifts the ISW effect becomes more suppressed as the neutrino mass
increases (Fig.~\ref{fig_iswrs_z}). Considering relative differences with respect to the massless
case, on linear scales we recover,
within $\sim 1-2$\%, accuracy, analytical expectations from \CAMB,
which, however, at the moment does not provide non-linear estimations of the ISW-RS signal.
\item
At non-linear scales, $l>200$, massive neutrinos decrease the angular
power spectrum corresponding to the Rees-Sciama effect. Such
suppression is larger for larger neutrino masses, and decreases with
increasing redshifts, as shown in Fig.~\ref{fig_iswrs_z}.
\item
Since lensing traces directly the matter power spectrum, which can be
largely suppressed for large neutrino masses, we recover a similar
suppression in the CMB- and weak-lensing signals (Figs.~\ref{fig_cmblens}-\ref{fig_weaklens}).
When looking at relative differences with respect to the massless
case, the agreement with \CAMB\ (non-linear neutrino
corrections included) is at $\sim 1\%$ level on the scales covered by the simulations.
The suppression of the lensing auto power spectra decreases with the
increase of the source redshift, $z_s$, with a maximum difference of
$\sim 10\%$ between $z_s=2$ and $z_s=1100$.
\item
Lensed TT, EE and BB spectra are consistently affected by massive
neutrinos (mid and lower panels of Fig.~\ref{fig_cmblens}). As
expected, since the strength of the
gravitational potential decreases for larger neutrino masses, CMB
acoustic oscillations are less smoothed and smeared out than in the
massless case. This implies that, as the neutrino mass increases, the lensed TT and EE power spectra are larger at small $l$, and the
so-called ``damping-tail'' is lower at high $l$.
Therefore the amplitude of the lens-induced B-mode power
spectrum decreases for larger neutrino masses on all the
scales, as shown in the left lower panel of Fig.~\ref{fig_cmblens}.
\item
Concerning the cross-correlation between ISW-RS and CMB-lensing
signals, at $l\lesssim 400$ we correctly recover the linear signal
from \CAMB, within $1-2 \%$ accuracy
(Fig.~\ref{fig_cmbl_isw_cross}). At the transition between the linear
and the non-linear regimes, $l\sim 700$, the simulated signal
correctly undergoes the sign-inversion expected by non-linear
semi-analytical calculations in the massless case \cite{Schaefer2011,Mangilli13, Lewis_RS}. This feature is
interestingly altered by the presence of massive neutrinos, as it moves
toward larger multipoles with increasing $M_\nu$, and the
displacement, with respect to the $M_\nu=0$ eV case, can be larger than a factor of
$2$ for $M_\nu=0.53$ eV. This is reasonably
expected, since massive neutrinos extend the linear regime to smaller
scales than in the massless case. Therefore, the cross power between
ISW-RS and CMB-lensing increases with larger neutrino masses at
$300\lesssim l \lesssim 1500$, and, {\emph e.g.}, at $l\sim 600$, we find an excess of cross
power of a factor of $\sim 4$ for $M_\nu=0.3$ eV.
At higher multipoles, in the fully non-linear regime, we find a
suppression of the signal as also occurs for the auto spectra.
\item
As shown in the left panel of Fig.~\ref{fig_weakl_isw_cross}, the cross-correlation
between ISW-RS and weak-lensing presents features similar to
the ISW-RS cross CMB-lensing signal, with the only difference given by a
lower total impact of massive neutrinos. Again, when looking at relative differences with respect to the massless
case, at $l\lesssim 300$ we correctly recover, within $1-2 \%$ accuracy, the linear
signal from \CAMB, which does not provide non-linear estimations at the moment.
The non-linear sign-inversion of the cross power is still present but less enhanced, while the
signal seems to increase with increasing $z_s$. In particular, the
excess of ISW-RS due to the presence of massive neutrinos makes the
cross power less suppressed with respect to the weak-lensing auto
power, and finally, for very high source redshifts, we observe a net
excess of power with respect to the massless case (right panel of Fig.~\ref{fig_weakl_isw_cross}).
\end{itemize}
The last two points represent the main findings of this work.
The cross-correlation between the ISW-RS and lensing signals enters the
computation of the lensed CMB temperature bispectrum~\cite{Lewis_RS}; therefore its
correct estimation at the non-linear level from N-body simulations,
and the knowledge of the neutrino impact on its amplitude may result to be
of extreme importance for the full evaluation of the CMB
temperature three-point function~\cite{Lewis2011}.
The latter probes the perturbation growth and expansion history of the
Universe, and hence can be
used to constrain dark energy and neutrino masses~\cite{Lewis06}. In addition, due to
non-linear structure evolution, on very small scales the lensing
potential is not a Gaussian field, and consequent additional
contributions
to the bispectrum may be evaluated directly from the simulated maps
obtained via the DEMNUni simulations. We reserve this for future work.
\acknowledgments
C.C. thanks Anna Mangilli, Julien Bel, Emiliano Sefusatti, Matteo Calabrese,
and Matteo Zennaro for very useful discussions.
C.C. thanks Matteo Viel for providing the N-GenIC code for
initial conditions, modified to take into account a massive neutrino
particle component. The DEMNUni simulations were carried out
at the Tier-0 IBM BG/Q machine, Fermi, of the Centro Interuniversitario del
Nord-Est per il Calcolo Elettronico (CINECA, Bologna, Italy), via the five
million cpu-hrs budget provided by the Italian SuperComputing Resource
Allocation (ISCRA) to the class--A proposal entitled ``The Dark Energy and
Massive-Neutrino Universe". C.C. acknowledges financial support from the INAF
Fellowships Programme 2010 and from the European Research Council through
the Darklight Advanced Research Grant (n. 291521).
K.D. and M.P. acknowledge support by the DFG Cluster of
Excellence ``Origin and Structure of the Universe'' and the
SFB-Tansregio TR33 ``The Dark Universe''.
\bibliographystyle{JHEPb}
|
\section{Introduction}
For hermitian symmetric domains $\mathcal H=G/K$ the complex structure on $\mathcal H$
corresponds to a decomposition $\mathop{Lie}(G) \otimes_{\mathbb R} {\mathbb C}$ into subspaces
$\mathop{Lie}(K) \oplus {\mathfrak p}_+ \oplus {\mathfrak p}_-$. The maximal compact subgroup $K$ of $G$
acts on the subspaces ${\mathfrak p}_{\pm}$, corresponding to the complexified holomorphic resp. antiholomorphic
tangent space of $\mathcal H$ at the origin. For arithmetic subgroups $\Gamma$ of the
group $G$ the Hilbert space $L^2(\Gamma\!\setminus\! G,dg)$
decomposes into the discrete spectrum $L^2_{\textrm{dis}}(\Gamma\!\setminus\! G,dg)$ and a continous
spectrum.
Part of this discrete spectrum is obtained as follows. By assumption the group $G$
is of hermitian type, hence there exist discrete series representations of $G$.
Among these are the representations $\pi$ of the holomorphic discrete series.
The isotypical subspace $L^2(\Gamma\!\setminus\! G,dg)_\pi$ of $L^2_{\textrm{dis}}(\Gamma\!\setminus\! G,dg)$
on which $G$ acts by one of these holomorphic discrete series representations $\pi$
is isomorphic to a finite direct sum of $\pi$ with multiplicity say $m(\pi)$.
It is
well known that these subspaces
$L^2(\Gamma\!\setminus\! G,dg)_\pi$ occur in the cuspidal part of the spectrum.
The projection operator
\begin{equation*}
P_\pi: L^2(\Gamma\!\setminus\! G,dg) \longrightarrow L^2(\Gamma\!\setminus\! G,dg)_\pi
\end{equation*}
can be studied by various techniques. Since a holomorphic discrete series representation
contains a unique lowest $K$-type $\tau$, it often suffices to study
the projection operator
$ P_\pi$ on the $\tau$-isotypic subspace $L^2(\Gamma\!\setminus\! G,dg)_\tau$
of the action of $K$ on $L^2(\Gamma\!\setminus\! G,dg)$.
Classically, the analysis of the projection operator $P_\pi$ is then often
achieved by passing
from functions $f$ in $L^2(\Gamma\!\setminus\! G,dg)_\tau$ to functions $h$
on $\mathcal H$, defined by $h(gK)= J_\tau(g) f(g)$ using an explicit cocycle factor $J_\tau$ whose definition
involves the lowest $K$-type of $\pi$. The functions $h$ on $\mathcal H$ so defined transform under the action of $\gamma\in \Gamma$ with
respect to the modular transformation property $h(\gamma Z)=J_\tau(\gamma,Z)h(Z)$. The functions $f$ in the subspace $L^2(\Gamma\!\setminus\! G,dg)_\pi$
correspond to the holomorphic functions on $\mathcal H$ with this transformation property. Putting $L^2$-conditions aside (for simplicity),
the projectors $P_\pi$ thus become holomorphic projectors.
Classically, they were studied in terms of Fourier expansions using the theory of Poincare series
in \cite{sturm1}, \cite{gross-zagier} for the classical case $G=\SL_2$ and in \cite{panchishkin} for the case of the symplectic groups
$G=\Sp_{m}$.
\medskip
For the case $G=\SL_2$ this becomes more concrete as follows. The holomorphic discrete
series $\pi=\pi_k$ of $\SL_2$ is parametrized by the integers $k\geq 2$, the weight of their lowest
$K$-type, and it is well known that $m(\pi_k)$ can be identified with the dimension
of the space of cuspidal modular forms $[\Gamma,k]_0$ of weight $k$ with respect to the arithmetic group $\Gamma$ on the complex upper half plane $\mathcal H$.
Suppose $\Gamma$ contains translations, so that the modular transformation property
$h(\gamma z)= j_k(\gamma,z)h(z)$ implies $h(z)=h(z+ n)$ for some integer $n$
that allows to expand $h(z)$ into a Fourier expansion
\begin{equation*}
h(z) \:=\: \sum_{t}\ a(t,y) \exp(2\pi i tz) \:=\: \sum_t \ b(t,y) \exp(2\pi i tx) \: .
\end{equation*}
For $z=x+iy$ the Fourier coefficients $b(t,y)$ are functions of the imaginary part.
In this special context it has been shown by Sturm \cite{sturm1} for $k>2$ and Gross-Zagier \cite{gross-zagier} for $k=2$
that, up to some explicit normalizing constant $c(k)$ depending only on $k$, the projector $P_\pi$ for $\pi=\pi_k$ from above corresponds to the following
holomorphic projector defined on the level of Fourier coefficients by
\begin{equation*}
b(t,y) \quad \mapsto \quad b_{\textrm{hol}}(t) \:=\: c(k) \cdot \int_0^\infty b(t,y) \exp(-2\pi ty) (ty)^k \frac{dx~dy}{y^2} \: .
\end{equation*}
So the holomorphic projection of $h(z)$ is given by the holomorphic modular
form $\sum_t b_{\textrm{hol}}(t) \exp(2\pi i tz)$.
\medskip
Concerning the higher dimensional cases studied in \cite{panchishkin} and \cite{holproj},
it turned out that for the holomorphic discrete series of scalar weight high enough again the holomorphic projectors $P_\pi$ are given
by analogous holomorphic projections on the level of Fourier coefficients generalizing
the Sturm projections from above. One could therefore believe, that this holds quite generally
for all holomorphic discrete series representations of the symplectic group $\Sp_m$.
However, in the higher rank case new phenomena occur. Although the Sturm
projection operators are defined, in general they do not coincide with the corresponding
projector operators in all cases. This may be the general situation for those holomorphic discrete series whose lowest $K$-type is small.
For the special case $m=2$ we analyze this in detail.
The holomorphic discrete series $\pi=\pi_k$ of the group $\Sp_2$ are indexed by their lowest $K$-types $k=(k_1,k_2)$ that are given by integers $k_1 \geq k_2 \geq 3$.
The \lq{smallest\rq}\ case is $k_1=k_2=3$, the case of scalar weight 3. We show
that in this case, as opposed to the cases of weight greater $3$,
the Sturm projection does not describe the projection operator $P_{\pi}$
for the holomorphic discrete series of scalar weight 3.
For this we use Poincar\'e series of exponential type of weight $(3,3)$.
Their meromorphic continuation is established by the methods of~\cite{holproj}.
That is, we use the resolvents of special Casimir operators whose pole behavior on $L^2(\Gamma\backslash \Sp_2({\mathbb R}))$ is described by representation theory.
By a careful inspection of the spectral components containing the $K$-type $(3,3)$ nontrivially, the continuation of the Poincar\'e series is seen to be
analytic in the critical point. More precisely, their spectral locus there consists of the expected holomorphic discrete series representation $\pi_{(3,3)}^-$ and in addition
the holomorphic representation $\pi_{(1,1)}^{\mathop {hol}}$ of weight one (Theorem~\ref{lem_diskretes_spek}).
Correspondingly, the Poincar\'e series written as modular forms on $\mathcal H$ have a decomposition
\begin{equation*}
p_T\:=\: f\:+\: \Delta_+^{[2]} (h_T),
\end{equation*}
where $f_T$ is a holomorphic cuspform of weight $(3,3)$ and $\Delta_+^{[2]} (h_T)$ is the nonholomorphic derivative of a holomorphic modular form of weight $(1,1)$
by the Maass operator $\Delta_+^{[2]}$. Both components are nontrivial in general (Theorem~\ref{satz_phantom_proj}). This reflects the property of Sturm's operator.
\begin{thm}
Let $h$ be a holomorphic cuspform on $\mathcal H$ of weight $(k,k)$. For $\kappa=k+2\geq 4$ Sturm's operator applied to the nonholomorphic image $\Delta_+^{[2]} (h)$ is zero, but for
$\kappa=3$ it does not vanish.
\end{thm}
So Sturm's operator will realize the
projection $P_{\pi}$ only up to some additional phantom projection $Q_\pi$ to
a space of phantom components
\begin{equation*}
L^2(\Gamma\!\setminus\! G,dg)_{\mathop{phan}(\pi)} \subset L^2(\Gamma\!\setminus\! G,dg)
\end{equation*}
defined by certain representations $\mathop{phan}(\pi)$ associated to $\pi$, which in the case above is $\mathop{phan}(\pi)=\pi_{(1,1)}^{\mathop{hol}}$.
\section{Notation and resolvents}
We follow the notation of \cite{holproj}.
Let $G=\Sp_m({\mathbb R})$ be the symplectic group of genus $m$.
Apart from section~\ref{diffoperators} we restrict to the case $m=2$.
Realize $G$ as the group of those $g\in M_{m,m}({\mathbb R})$ satisfying $g'Wg=W$ for
\begin{equation*}
W=\begin{pmatrix}0&-E_m\\E_m&0\end{pmatrix}.
\end{equation*}
We have the usual action of $G$ on the Siegel halfspace ${\mathcal H}$, for $g=\begin{pmatrix}a&b\\c&d\end{pmatrix}\in G$,
\begin{equation*}
g Z=(aZ+b)(cZ+d)^{-1} \:.
\end{equation*}
The stabilizer $K$ of $i=iE_m\in{\mathcal H}$ is a maximal compact subgroup of
$G$. It is isomorphic to the unitary group $U(m)$. We denote by
\begin{displaymath}
g\:\mapsto \:g i=:Z=X+iY
\end{displaymath}
the obvious isomorphism of $G/K$ to ${\mathcal H}$.
Let $\mathfrak g$ be the Lie algebra of $G$ realized as $\mathfrak g_{\mathbb C}\subset M_{2m,2m}(\mathbb C)$ consisting of those $g$ satisfying
$g'W+Wg=0$.
Then $\mathfrak g_{\mathbb C}=\mathfrak p_+\oplus\mathfrak p_-\oplus\mathfrak k_{\mathbb C}$,
where $\mathfrak k_{\mathbb C}$ is the Lie algebra of $K$ given by the matrices satisfying
\begin{equation*}
\begin{pmatrix}A&S\\-S&A\end{pmatrix}\:, \quad A'=-A\:, \quad S'=S\:,
\end{equation*}
and
\begin{equation*}
\mathfrak p_\pm=\left\{\begin{pmatrix}X&\pm iX\\\pm iX&-X\end{pmatrix},\quad X'=X\right\}.
\end{equation*}
Let $e_{kl}\in M_{m,m}(\mathbb C)$ be the elementary matrix having entries $(e_{kl})_{ij}=\delta_{ik}\delta_{jl}$
and let $X^{(kl)}=\frac{1}{2}(e_{kl}+e_{lk})$.
The elements $(E_\pm)_{kl}=(E_\pm)_{lk}$ of $\mathfrak p_\pm$ are defined to be those corresponding to $X=X^{(kl)}$,
$1\leq k,l\leq m$.
Then $(E_\pm)_{kl}$, $1\leq k\leq l\leq m$ form a basis of $\mathfrak p^\pm$.
A basis of $\mathfrak k_{\mathbb C}$ is given by $B_{kl}$, for $1\leq k,l\leq m$, where $B_{kl}$ corresponds
to $A_{kl}=\frac{1}{2}(e_{kl}-e_{lk})$ and $S_{kl}=\frac{i}{2}(e_{kl}+e_{lk})$.
Let $E_\pm$ be the matrix having entries $(E_\pm)_{kl}$.
Similarly, let $B=(B_{kl})_{kl}$ be the matrix with entries $B_{kl}$ and let $B^\ast$ be its transpose.
$E_+$, $E_-$, $B$ and $B^\ast$ are matrix valued matrices.
Formal traces of their formal products, e.g. $\trace(E_+E_-)$, are not invariant under cyclic permutations of their arguments.
The center ${\mathfrak z}_{\mathbb C}$ of the universal enveloping Lie algebra $\mathfrak U(\mathfrak g_{\mathbb C})$
is generated by $m$ elements.
The Casimir elements $C_1,C_2$ belong to ${\mathfrak z}_{\mathbb C}$,
\begin{equation*}
C_1=\frac{1}{2}(\trace(E_+E_-)+\trace(E_-E_+))+\trace(BB)\:,
\end{equation*}
\begin{eqnarray*}
C_2&=& \frac{1}{2}\bigl(\trace(E_+E_-E_+E_-)+\trace(E_-E_+E_-E_+)+\trace(B^4)+\trace((B^\ast)^4)\bigr)\\
&&+2\bigl(\trace(E_+E_-BB)+\trace(E_-E_+B^\ast B^\ast)\bigr)\\
&&-\sum_{i,j,k,l}\{(E_+)_{kl},(E_-)_{ij}\}B_{jk}B_{il}\\
&&+\frac{(m+1)^2}{2}(\trace(E_+E_-)+\trace(E_-E_+))\:.
\end{eqnarray*}
For a smooth representation $\pi$ of $G$ the actions of the Casimir elements restricted to scalar $K$-types $(\kappa,\dots,\kappa)$ are given by
\begin{equation*}
\pi(C_1)\:=\: \pi(\trace(E_+E_-))-\kappa m(m+1-\kappa)
\end{equation*}
and
\begin{eqnarray*}
\pi(C_2)&=&\pi(\trace(E_+E_-E_+E_-))+m\kappa^4\\
&&+((m+1)^2-2\kappa (m+1)+2\kappa^2)\bigl(\pi(\trace(E_+E_-))-\kappa m(m+1)\bigr)\:.
\end{eqnarray*}
For genus $m=2$ let
\begin{equation}\label{def_kompakte_CUA}
{\mathfrak h}_{\mathbb C} \:=\:{\mathbb C} B_{11}+{\mathbb C} B_{22} \:\subset\: {\mathfrak k}_{\mathbb C}
\end{equation}
be a common Cartan subalgebra for ${\mathfrak k}_{\mathbb C}$ and ${\mathfrak g}_{\mathbb C}$.
Let $\Delta^+$ be the set of positive roots for ${\mathfrak h}_{\mathbb C}$ such that their root spaces belong to ${\mathbb C} B_{12}+{\mathfrak p}^-$.
Writing $\Lambda=(\Lambda_1,\Lambda_2)$ for $\Lambda \in{\mathfrak h}_{\mathbb C}^\ast$, where $\Lambda_j=\Lambda(B_{jj})$,
these root spaces are
\begin{eqnarray*}
{\mathfrak g}_{(1,-1)}={\mathbb C} B_{12}, & {\mathfrak g}_{(2,0)}={\mathbb C} (E_-)_{11}\:,\\
{\mathfrak g}_{(1,1)}={\mathbb C} (E_-)_{12}, & {\mathfrak g}_{(0,2)}={\mathbb C} (E_-)_{22}\:.
\end{eqnarray*}
Half the sum of positive root is
\begin{equation*}
\delta\:=\:\delta_G\:=\:\frac{1}{2}\sum_{\Lambda\in\Delta^+}\Lambda=(2,1)\:,
\end{equation*}
while $\delta_K=\frac{1}{2}(1,-1)$ is half the sum of positive compact roots.
Applying the linear form $\Lambda$ to the images of $C_1$ and $C_2$ under the Harish-Chandra homomorphism we get
\begin{eqnarray*}
\Lambda(C_1)\:=\:\Lambda(\gamma(C_1))&=&\Lambda_{1}^2+\Lambda_{2}^2-5\:,\\
\Lambda(C_2)\:=\:\Lambda(\gamma(C_2))&=&\Lambda_{1}^4+\Lambda_{2}^4-17+3\Lambda(C_1)\:.
\end{eqnarray*}
The diagonal subalgebra $\mathfrak a_{\mathbb C}$ is another Cartan subalgebra, and by choosing Euclidean coordinates $\Lambda=(\Lambda_1,\Lambda_2)$ there
we get an isometric isomorphism to $\mathfrak h_{\mathbb C}$. So the above formulas retain valid.
Choosing the system of positive roots correspondingly,
\begin{equation*}
\Delta^+:=\{\alpha_1=(0,2),\alpha_2=(1,-1),\alpha_1+\alpha_2,\alpha_1+2\alpha_2\} \subset \mathfrak a_{\mathbb C}^\ast\:,
\end{equation*}
$\mathfrak a$ is the split component of the Borel subgroup
\begin{equation*}
B\:=\:\left\{\begin{pmatrix} T&X\\0&T'^{-1}\end{pmatrix}\mid T \textrm{ upper triangular }\right\}\:\subset\: G\:.
\end{equation*}
The Weyl group $W$ of $G$ acts on ${\mathfrak a}_{\mathbb C}^\ast$. It is generated by the simple reflections $s_{\alpha_1}$ and $s_{\alpha_2}$.
We also define $\mathfrak a_1=\ker(\alpha_1)$ to be the split component of the Klingen parabolic $P_1\supset B$, and
$\mathfrak a_2=\ker(\alpha_2)$ to be the split component of the Siegel parabolic $P_2\supset B$.
Let $u$ and $v$ be complex variables. In \cite[sec.~3]{holproj} there are chosen elements
\begin{eqnarray*}
D_+(u,\Lambda) &=& \prod_{\alpha\in\Delta\textrm{ long}} \bigl(\check{\alpha}(\Lambda)-u\bigr)\:,\\
D_-(v,\Lambda) &=& \prod_{\alpha\in\Delta\textrm{ short}} \bigl(\check{\alpha}(\Lambda)-v\bigr)\:,
\end{eqnarray*}
or equivalently,
\begin{eqnarray*}
D_+(u,\Lambda)&=& (\Lambda_1^2-u^2)(\Lambda_2^2-u^2)\:,\\
D_-(v,\Lambda)&=& \bigl((\Lambda_1+\Lambda_2)^2-v^2\bigr)\bigl((\Lambda_1-\Lambda_2)^2-v^2\bigr)\:.
\end{eqnarray*}
They are the images of the Casimir elements
\begin{eqnarray*}
D_+(u)&:=& \frac{1}{2}\bigl(C_1^2-C_2+11C_1-2(u^2-1)C_1+2(u^2-1)(u^2-4)\bigr)\:,\label{Dplus_in_koordinaten}\\
D_-(v)&:=& 2C_2-C_1^2-34C_1-2(v^2-9)C_1+(v^2-9)(v^2-1)\:\label{Dminus_in_koordinaten}
\end{eqnarray*}
under the Harish-Chandra homomorphism.
Let $\Gamma$ be any subgroup of finite index in the full modular group
$\Sp_2({\mathbb Z})$ containing
the group
\begin{equation*}
\Gamma_\infty\:=\:\{\begin{pmatrix}
\pm E_2&\ast\\0&\pm E_2
\end{pmatrix}\in\Sp_2({\mathbb Z})\}
\end{equation*}
of translations.
The space $L^2(\Gamma\backslash G)$ is a representation space for $G=\Sp_2({\mathbb R})$ by right translations. This $G$-action
comes along with an action of the universal enveloping algebra
$\mathfrak U({\mathfrak g}_{\mathbb C})$ on $\mathcal C^\infty$-vectors and action of the elements $D_+(u)$ and $D_-(v)$ allows extension to $L^2(\Gamma\backslash G)$.
Behavior and existence of the resolvents $R_\pm$ of the Casimir operators $D_\pm$ on $L^2(\Gamma\backslash G)$ are regulated by their spectrum.
We are interested in the scalar $K$-type $\kappa=(3,3)$.
\begin{prop}\label{prop_resolventen}
The resolvent $R_-(v)$ exists as a meromorphic function on $\re v >1$.
The resolvent $R_+(u)$ exists as a meromorphic function on $\re u>\frac{1}{2}$. Its spectral pole locus within $L^2(\Gamma\backslash G)_{(3,3)}$ at $u=1$
is given by the two discrete parameters $\Lambda=(2,1)$ and $\Lambda=(0,1)$.
\end{prop}
\begin{proof}[Proof of Proposition~\ref{prop_resolventen}]
The meromorphicity of the resolvents on the given domains is shown in~\cite[sec.~3]{holproj}.
The poles of $R_+(u)$ in $u=1$ are given by the $1$-dimensional continuous spectral component $K_{\alpha_1}(1)=(i{\mathbb R},1)$ and discrete components indexed by $\Lambda=(s,1)$.
By Proposition~\ref{lemma_kein_kont_spek} the first does not occur in $L^2(\Gamma\backslash G)_{(3,3)}$.
By Theorem~\ref{lem_diskretes_spek}, the remaining discrete parameters are the claimed.
\end{proof}
\section{On the spectrum of $L^2(\Gamma\backslash G)$}
We give results on the occurrence of the $K$-type $(3,3)$ in $L^2(\Gamma\backslash G)$.
\begin{prop}\label{lemma_kein_kont_spek}
In the $1$-dimensional continuous spectral component included in the parameter $K_{\alpha_1}(1)=(i{\mathbb R},1)$ the $K$-type $(3,3)$ is trivial.
\end{prop}
\begin{proof}[Proof of Proposition~\ref{lemma_kein_kont_spek}]
By \cite[Sec.~7.1]{konno} this spectral component for the general symplectic group $\tilde G=\mathop{GSp}_2(\mathbb A)$ is globally given by
\begin{equation*}
\int_{i{\mathbb R}} \mathop{ind}\nolimits_{P_{1}}^G(\omega_1\lvert\cdot\rvert^{it}\otimes \omega(\det))\lvert\det\rvert^{-it/2})~dt\:,
\end{equation*}
for unitary characters $\omega_1,\omega$ of $\mathbb G_m$, where for an element
\begin{equation*}
m(\lambda,g)\:=\:\begin{pmatrix}\lambda&&&\\
&a&&b\\
&&\nu/\lambda&\\
&c&&d
\end{pmatrix}
\end{equation*}
of $\tilde M_{Kl}$, where $\lambda\in\mathbb G_m$ and $g=\begin{pmatrix}
a&b\\
c&d
\end{pmatrix}\in\mathop{GL}_2$
with $\nu=\det g$, the character $ \omega_1\lvert\cdot\rvert^{it}\otimes \omega(\det))\lvert\det\rvert^{-it/2}$
is defined
as
\begin{equation*}
\omega_1\lvert\cdot\rvert^{it}\otimes \omega(\det))\lvert\det\rvert^{-it/2}(m(\lambda,g))\:=\:
\omega_1(\lambda)\lvert\lambda\rvert^{it} \omega(\nu)\lvert\nu\rvert^{-it/2}\:.
\end{equation*}
So at the real place the element
\begin{equation*}
m_0\:=\:m(1,\begin{pmatrix}
0&1\\-1&0
\end{pmatrix})\:\in\: K\cap M_{1}
\end{equation*}
always produces the value
\begin{equation*}
\omega_1\lvert\cdot\rvert^{it}\otimes \omega(\det))\lvert\det\rvert^{-it/2}(m_0)\:=\: 1\:,
\end{equation*}
while for the $K$-types $\pm(3,3)$ which equal $\det^{\pm3}$ on the unitary group $K\cong U(2)$ we must have (see~\ref{diffoperators})
\begin{equation*}
\det\nolimits^{\pm3}(\psi(m_0))\:=\: \pm i\:.
\end{equation*}
So this continuous spectral component does not contain $K$-type $(3,3)$.
\end{proof}
\begin{thm}\label{lem_diskretes_spek}
Let $\pi$ be an irreducible unitary representation of $G=\Sp_2({\mathbb R})$ with infinitesimal character in the Weyl group orbit of $\Lambda=(1,s)$ containing the $K$-type $(3,3)$ nontrivially.
Then either $\Lambda=(2,1)$ and $\pi$ is the holomorphic discrete series representation $\pi_{(3,3)}^-$ of minimal weight $(3,3)$,
or $\Lambda=(0,1)$ and $\pi$ is the holomorphic representation $\pi_{(1,1)}^{\mathop{hol}}$ of weight one (non-discrete series).
\end{thm}
Theorem~\ref{lem_diskretes_spek} is proven by the following series of lemmas.
\begin{lem}\label{lem_K-type_infchar}\cite[Theorem~1.1]{zhu}
If $\pi$ and $\pi'$ are irreducible representations of the same infinitesimal character containing the same scalar $K$-type nontrivially, then $\pi$ and $\pi'$ are equivalent.
\end{lem}
We include a simple proof of this lemma.
\begin{proof}[Proof of Lemma~\ref{lem_K-type_infchar}]
It follows from Casselman's subrepresentation theorem that $\pi$ and $\pi'$ are constituents of the same induced representation. By Peter Weyl's theorem a scalar $K$-type occurs with at most
multiplicity one in an induced representation. So $\pi$ equals $\pi'$ there.
\end{proof}
\begin{lem}\label{nzoukoudi-ausbeute}
Any irreducible unitary representation $\pi$ of $G$ with infinitesimal character $\Lambda$ Weyl conjugated to $(1,s)$ occurring in $L^2(\Gamma\backslash G)$
is either a discrete series representation or occurs in the following list of irreducible Langlands quotients $J'(P,\sigma,\nu)$ for parabolic subgroups $P\not=G$.
\begin{enumerate}
\item [(a)] For the Siegel parabolic subgroup $P=P_2$, $\sigma=\sigma_2^+$ is the discrete series representation of $M_2$ of minimal $(K\cap M_2)$-type $2$ and $\nu=0$.
Then $\pi$ has infinitesimal character $\Lambda=(1,-1)$.
\item [(b)] Let P=$P_{1}$ be the Klingen parabolic subgroup. Either $\sigma=(\sigma_1^\pm,\pm)$ is a discrete series representation of $M_{1}$ of minimal $(K\cap M_{1})$-type
$1$ and $\nu=e_1$.
Then $\pi$ has infinitesimal character $\Lambda=(1,0)$. Or $\sigma=(\sigma_2^\pm,-)$ is a discrete series representation of $M_{1}$ of minimal $(K\cap M_1)$-type
$2$ and $\nu=e_1$.
Then $\pi$ has infinitesimal character $\Lambda=(1,2)$.
\item [(c)] If $P=B$ is the Borel subgroup, $\sigma=1$ is the trivial representation of $M_B$ and $\nu=(1,0)$ or $\nu=(2,1)$.
Then $\pi$ has infinitesimal character equal to $\nu$.
\end{enumerate}
\end{lem}
\begin{proof}[Proof of Lemma~\ref{nzoukoudi-ausbeute}]
We follow Nzoukoudi's~\cite{nzoukoudi} classification of the irreducible unitary representations via Langlands quotients.
The parabolic subgroup $P=G$ produces the discrete series.
For the Siegel parabolic, the quotients $J'(P_2,\sigma,\nu)$ belong to limits discrete series representations $\sigma=\sigma_n^+$ of $M_2\cong\SL_2({\mathbb R})^\pm$ with
infintesimal character $(n-1)(e_1-e_2)$ and
characters $\nu=z(e_1-e_2)$ for complex $z$. For the infinitesimal character $\Lambda=(n+z,-n+z)$ of the quotient to belong to the Weyl orbit
$(1,s)^W$ we must have $z\in {\mathbb Z}$.
The unitary
constraint then is $0\leq z\leq 1$ and $n$ odd. The only possible choice is $\Lambda=(1,-1)$, $\nu=0$ and $\sigma=\sigma_2^+$.
For the Klingen parabolic, $\sigma=(\sigma_n^\pm,\pm)$ is a limit of discrete series of $M_{2}\cong\SL_2({\mathbb R})\times Z_2$ with infinitesimal character
$2(n-1)e_2$, and $\nu=2ze_1$, $z$ complex,
is the
character. So the quotient has infinitesimal character $\Lambda=(2z,2n)$, which belongs to the Weyl orbit $(1,s)^W$ only if $2z=\pm 1$. The unitary constraint then forces
$\sigma=(\sigma_n^\pm,-)$ for arbitrary $n\geq 0$ or $\sigma=(\sigma_1^\pm,+)$.
As $J'(P_2,\sigma,\nu)$ isn't discrete series,
we have the Eisenstein constraint $\lvert\!\lvert \Lambda\rvert\!\rvert^2\leq \lvert\!\lvert \delta\rvert\!\rvert^2= 5$ for the Langlands quotient to belong to the residual spectrum
of $L^2(\Gamma\backslash G)$. So $n=0,1$.
The subgroup $M_B$ of the Borel group is isomorphic to $Z_2\times Z_2$, and any representation $\sigma=\sigma^{\epsilon_1,\epsilon_2}$ of $M_B$ is described by two signs
$\epsilon_1,\epsilon_2$ on generators.
The infinitesimal character $\Lambda$ of the Langlands quotient equals $\nu=(z_1,z_2)$.
The unitary contraint implies that $\nu$ is either purely imaginary, or of the form $(x+iy,x-iy)$ with $0<x\leq \frac{1}{2}$ and $y\in {\mathbb R}$,
or of the form $(x,iy)$ with $0<x\leq 1$ and $y\in {\mathbb R}^\times$, or $z_1\geq z_2\geq 0$ are real and $z_1+z_2\leq 1$ or $(z_1,z_2)=(2,1)$ and $\sigma=1$.
So only in the last two cases it may belong to the Weyl orbit $(1,s)^W$.
In case $\Lambda=\nu=(1,iy)$ the infinitesimal character of the quotient is not real. So it doesn't appear in the residual spectrum of $L^2(\Gamma\backslash G)$.
The remaining possibilities are $\Lambda=(1,0)$ or $\Lambda=(1,2)$ with $\sigma=1$.
\end{proof}
\begin{lem}\label{discrete_series}
Among the discrete series of $G$ there is a unique one carrying the $K$-type $(3,3)$ nontrivially and having infinitesimal character in the Weyl orbit $(1,s)^W$.
This is the holomorphic discrete series representation $\pi_3^-$ of minimal $K$-type $(3,3)$.
\end{lem}
\begin{proof}[Proof of Lemma~\ref{discrete_series}]
For a semisimple real Lie group $G$ with $\mathop{rank} G=\mathop{rank} K$
the discrete series representations are parametrized by Harish-Candra parameters (infinitesimal characters) $\Lambda$
which belong the weight lattice and satisfy
$\langle \check\alpha ,\Lambda\rangle\not=0$ for all roots $\alpha\in\Delta$ and $\langle \check\alpha ,\Lambda\rangle>0$ for all positive compact roots $\alpha\in\Delta_c^+$.
There is a unique choice of positive roots $\Delta_\Lambda^+$ for which $\Lambda$ is dominant.
Then the Blattner weight for $\Lambda$ is given by
\begin{equation*}
\beta_\Lambda=\frac{1}{2}\sum_{\alpha\in\Delta_\Lambda^+}\alpha - \sum_{\alpha\in\Delta_c^+}\alpha\:.
\end{equation*}
The minimal $K$-type of $\pi_\Lambda$ is given by the Blattner paramter
\begin{equation*}
k\:=\: \Lambda +\beta_\Lambda\:,
\end{equation*}
and all other $K$-types of $\pi_\Lambda$ are of the form
\begin{equation*}
k+\sum_{\alpha\in\Delta_\Lambda^+} n_\alpha \alpha\:,
\end{equation*}
for nonnegative integers $n_\alpha$ (\cite[IX.7]{knapp}).
For $G=\Sp_2({\mathbb R})$ we choose the Cartan subalgebra $ \mathfrak h_{\mathbb C}$ of both $G$ and $K$ as in (\ref{def_kompakte_CUA})
Chooinge simple roots $\alpha_1=2\Lambda_2$ and $\alpha_2=\Lambda_1-\Lambda_2$ as before, the short root $\alpha_2$ is compact
(i.e. its root space belongs $\mathop{Lie}(K)_{\mathbb C}=\mathfrak k_{\mathbb C}$) and we choose
$\Delta_c^+=\{\alpha_2\}$. The root system of $\mathfrak g_{\mathbb C}$ with respect to $\mathfrak h_{\mathbb C}$ is
\begin{eqnarray*}
\Delta&=& \{\pm\alpha_1,\pm\alpha_2,\pm (\alpha_1+\alpha_2),\pm(\alpha_1+2\alpha_2)\}\\
&=& \{(0,\pm 2),(\pm1,\pm1),(\pm 2,0)\}\:.
\end{eqnarray*}
There are four sectors of weight vectors satisfying the above conditions for $\Delta_c^+$ corresponding to the dominant weights of the following four choices of positive roots.
For the general symplectic group $\GSp_2({\mathbb R})$ these reduce by equivalence to the choices
\begin{equation*}
\Delta_1^+\:=\:\{(0,2),(1,-1),(1,1),(2,0)\}\:,
\end{equation*}
where the dominant weights $\Lambda=(\Lambda_1,\Lambda_2)$ satisfy $\Lambda_1>\Lambda_2>0$,
and
\begin{equation*}
\Delta_2^+\:=\:\{(0,-2),(1,-1),(1,1),(2,0)\}\:,
\end{equation*}
for which dominant weights $\Lambda=(\Lambda_1,\Lambda_2)$ satisfy $\Lambda_1>-\Lambda_2>0$.
The holomorphic discrete series $\pi_k^-$ of $\GSp_2({\mathbb R})$ belong to $\Delta_1^+$. Here $\beta_1=(1,2)$ and the minimal $K$-type $k$ is given by $k=(\Lambda_1+1,3)$
for dominant infinitesimal character $(\Lambda_1,1)$. For the $K$-type $l$ to occur we must have that $l$ or $-l$ is contained in
$k+{\mathbb Z}_{\geq 0}(0,2)+{\mathbb Z}_{\geq 0}(1,-1)$. So the $K$-type $l=(3,3)$ only occurs in $\pi_k^-$ if $k=(3,3)$ with infinitesimal character $(2,1)$.
The nonholomorphic discrete series $\pi_k^+$ belong to $\Delta_2^+$. Then $\beta_2=(1,0)$, the dominant infinitesimal characters are $\Lambda=(\Lambda_1,\Lambda_2)$, where $\Lambda_1>-\Lambda_2>0$.
The Blattner parameter for $\Lambda=(\Lambda_1,-1)$ is $k=(\Lambda_1+1,-1)$.
The arising $K$-types $l$ are of the form $\pm l\in k+{\mathbb Z}_{\geq 0}(1,1)+{\mathbb Z}_{\geq 0}(0,-2)$. So $l=(3,3)$ can occur as a $K$-type in $\pi_k^+$ at most in case $k=(3,-1)$,
which has infinitesimal character $(2,-1)$ Weyl conjugated to that of $\pi_{(3,3)}^-$. But this is impossible by Lemma~\ref{lem_K-type_infchar}.
\end{proof}
\begin{lem}\label{lemma_ausschlussverfahren}
Concerning case (b) of Lemma~\ref{nzoukoudi-ausbeute}, the Langlands quotient belonging to $\sigma=(\sigma_1^-,+)$ with $\nu=e_1$ has nontrivial $K$-type $(3,3)$.
It is the holomorphic representation $\pi_{(1,1)}^{\mathop{hol}}$ of weight one.
In all other cases of Lemma~\ref{nzoukoudi-ausbeute} (a)-(c) the $K$-type $(3,3)$ is zero.
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lemma_ausschlussverfahren}]
The $(K\cap M_S)$-types of discrete series representation $\sigma_2^+$ in case (a) are included in $2+2{\mathbb Z}$. By Frobenius reciprocity,
the induced representation $\mathop{ind}_{P_S}^G(\sigma_2^+)$
does not contain the odd scalar $K$-type $(3,3)$, nor does its Langlands quotient.
Concerning the limits of discrete series in case (b), the holomorphic limit of discrete series $\sigma_1^-$ of $\SL_2({\mathbb R})$ contains the $K$-type $3$ nontrivially, as well does the
limit of discrete series $(\sigma_1^-,+)$ of $M_{Kl}\cong \SL_2({\mathbb R})\times Z_2$. Again by Frobenius reciprocity, the corresponding Langlands quotient contains the $K$-type $(3,3)$ nontrivially.
The irreducible Langlands quotients are pairwise inequivalent. The quotients left by cases (b) and (c) have infinitesimal characters conjugated to either $(2,1)$ or $(1,0)$.
If one of them contained a nontrivial $K$-type $(3,3)$, it was equivalent to either $\pi_{(3,3)}^-$ or $\pi_{(1,1)}^{\mathop{hol}}$ by Lemma~\ref{lem_K-type_infchar}.
\end{proof}
\section{Poincar\'e series for weight three}\label{section_konvergenz}
We define Poincar\'e series of weight $\kappa=3$.
For complex variables $u$ and $v$ and positive definite $(2,2)$-matrices $T$ with half-integral entries
let
\begin{equation*}
P_T(g,u,v) =\sum_{\gamma\in \Gamma_\infty\backslash \Gamma} H_T(\gamma g,s_1,s_2)\:\:,
\end{equation*}
where
\begin{equation*}
H_T(g,s_1,s_2)\:=\: \frac{\exp(2\pi i\trace(T Z))}{J_\kappa(g,i)}\trace(TY)^{s_1}
\det (Y)^{s_2}\:,
\end{equation*}
for $J_\kappa(g,Z)=\det(cZ+d)^\kappa$ for $g=\begin{pmatrix}\ast&\ast\\c&d\end{pmatrix}$,
and
\begin{equation*}
s_1\:=\:\frac{v-2u-1}{2}\quad\textrm{ and }\quad s_2\:=\:\frac{u-(\kappa-m)}{2}\:=\: \frac{u-1}{2}\:.
\end{equation*}
By~\cite[Cor. 4.4]{holproj} these Poincar\'e series belong to $L^2(\Gamma\backslash G)$ within their area of convergence
\begin{equation*}
A\:=\: \{(u,v)\in{\mathbb C}^2\mid \re u>2 \textrm{ and } \re v>5\}\:.
\end{equation*}
The function $H_T(g,s_1,s_2)$ is
nonholomorphic (in the variable $g$) apart from $(s_1,s_2)=(0,0)$. One expects the Poincar\'e series to have analytic continuation to the critical point $(s_1,s_2)=(0,0)$,
equivalently $(u,v)=(1,3)$, which is holomorphic with respect to $g$.
By the same method as for case $\kappa=4$ (\cite[Sec.~6]{holproj}) we show that indeed this analytic continuation exists along the line $s_1=0$, but that there is a nonholomorphic share.
\begin{thm}\label{analytische_fortsetzung}
The Poincar\'e series $P_T(\cdot,u,v)$ admit meromorphic continuation in $L^2(\Gamma\backslash G)_\kappa$ to the cone
\begin{equation*}
\{(u,v)\in {\mathbb C}^2\mid \re u>\frac{1}{2},\: \re v>1\}\:.
\end{equation*}
The poles are contained in a finite number of lines $u=\mathop{const.}$ and $v=\mathop{const.}$.
The limit
\begin{equation*}
P_T(\cdot,1,3)\::=\: \lim_{u\to 1} P_T(\cdot,u,2u+1)
\end{equation*}
exists as a function of $L^2(\Gamma\backslash G)_\kappa$. It has a $\mathcal C^\infty$-representative.
Its nonzero spectral components belong to the isotypical components of irreducible representations with infinitesimal characters $\Lambda=(2,1)$ and $\Lambda=(1,0)$.
\end{thm}
\begin{proof}[Proof of Theorem~\ref{analytische_fortsetzung}]
By abuse of notation, we omit the dependence on $T$ in our notations, i.e. $P(g,u,v):=P_T(g,u,v)$.
As in \cite{holproj} we use Casimir operators and their resolvents for the continuation.
Their actions depend on the weight $\kappa=3$. For the two Casimir operators $C_1$ and $C_2$ we calulate
\begin{eqnarray*}
C_1( P(g,u,v)) &=& 4(s_1^2+2s_1s_2+2s_2^2+2s_1+3s_2) P(g,u,v)\\%\label{Casimir1_angewandt}\\
&& -16\pi (s_1+s_2)P(g,u,v+2)\\
&& -8\det(T) s_1(s_1-1) P(g,u+2,v)\\
&& +32\pi\det(T) s_1P(g,u+2,v+2)\:
\end{eqnarray*}
and
\begin{eqnarray*}
&& C_2( P(g,u,v)) \:=\:\\
&&\hspace*{1cm}\bigl(
17u^4+2v^4-12uv^3+30u^2v^2-36u^3v+15u^2\\
&&\hspace*{6cm}+6v^2-18uv-32\bigr)P(g,u,v)\\%%
&&\hspace*{1cm}+256\pi^2(s_1+s_2)(s_1+s_2+1)P(g,u,v+4)\\%%
&&\hspace*{1cm}-128\pi (s_1+s_2)\bigl((s_1+s_2)^2+3(s_1+s_2)+\frac{23}{8}\bigr)P(g,u,v+2)\\
&&\hspace*{1cm}+32\det(T)^2s_1(s_1-1)(s_1-2)(s_1-3)P(g,u+4,v)\\%%
&&\hspace*{1cm}-256\pi\det(T)^2s_1(s_1-1)(s_1-2)P(g,u+4,v+2)\\%%
&&\hspace*{1cm}-16\det(T) s_1(s_1-1)(7u^2+3v^2-9uv-u+\frac{7}{2})P(g,u+2,v)\\%%
&&\hspace*{1cm}+512\pi^2\det(T)^2s_1(s_1-1)P(g,u+4,v+4)\\%%
&&\hspace*{1cm}-64\pi\det(T) s_1(4u^2-3uv-10u+9v-8)P(g,u+2,v+2)\\%%
&&\hspace*{1cm}-256\pi^2\det(T) (s_1+s_2)(4s_1+2s_2+1)P(g,u+2,v+4)\:
\end{eqnarray*}
We used the computer algebra system Magma to verify these results.
But for continuation we need to apply operators which produce Poincar\'e series of better convergence in either $u$ or $v$. These operators are the known $D_+(u)$ and $D_-(v)$, respectively:
\begin{equation*}
D_+(u)\::=\: \frac{1}{2}\bigl(C_1^2-C_2+11C_1-2(u^2-1)C_1+2(u^2-1)(u^2-4)\bigr)
\end{equation*}
and
\begin{equation*}
D_-(v)\::=\: 2C_2-C_1^2-34C_1-2(v^2-9)C_1+(v^2-9)(v^2-1)\:,
\end{equation*}
respectively. We get
\begin{eqnarray}
&& D_+(u) P(g,u,v)\:=\:\label{Omega_groesser_angewandt}\label{gleichung_fuer_D_+}\\
&&\quad\quad\quad+16\det(T)^2s_1(s_1-1)(s_1-2)(s_1-3)P(g,u+4,v)\nonumber\\
&&\quad\quad\quad-128\pi\det(T)^2 s_1(s_1-1)(s_1-2)P(g,u+4,v+2)\nonumber\\
&&\quad\quad\quad+256\pi^2\det(T)^2 s_1(s_1-1)P(g,u+4,v+4)\nonumber\\
&&\quad\quad\quad+8\det(T) s_1(s_1-1)(u+1)(v-2)P(g,u+2,v)\nonumber\\
&&\quad\quad\quad-32\pi\det(T) s_1(6s_1s_2+3s_1+8s_2^2-8)P(g,u+2,v+2)\nonumber\\
&&\quad\quad\quad+64\pi^2\det(T)(v-u-2)(u-2)P(g,u+2,v+4)\:,\nonumber
\end{eqnarray}
respectively,
\begin{eqnarray}
D_-(v)P(g,u,v)&=& +64\pi^2(v-u)(v-u-2)P(g,u,v+4)\label{Omega_kleiner_angewandt}\label{gleichung_fuer_D_-}\\
&&+32\pi (u-1)(v+1)(v-u-2)P(g,u,v+2)\nonumber\\
&&+128\pi\det(T) s_1(s_1-2)(v+1)P(g,u+2,v+2)\nonumber\\
&&-256\pi^2\det(T) (v-u-2)^2P(g,u+2,v+4)\:.\nonumber
\end{eqnarray}
{\it Step 1. Meromorphic continuation.}
In \cite[sec.~3]{holproj} the spectral poles of the resolvents $R_+(u)$ and $R_-(v)$ of $D_+(u)$ and $D_-(v)$, respectively, were studied.
They exist as meromorphic functions with the following properties.
\begin{prop}\cite[Prop.~3.1, 3.2, 3.3]{holproj}\label{prop_meromorphe_resolventen}
The resolvent $R_+(u)$ of $D_+(u)$ is meromorphic on $\re u>\frac{1}{2}$. On the $2$-dimensional continuous spectrum, which is included in the parameters $\re \Lambda=0$, it is holomorphic.
On the $1$-dimensional spectrum it is meromorphic with a finite number of simple poles $u=c$ including $u=1$, which arise from components $K_{\alpha_1}(x)$ or $K_{\alpha_1+2\alpha_2}(c)$.
On the discrete spectrum $R_+(u)$ is meromorphic with a finite number of poles corresponding to the roots of $(\Lambda_1^2-u^2)(\Lambda_2^2-u^2)$
for infinitesimal characters $\Lambda=(\Lambda_1,\Lambda_2)$.
The resolvent $R_-(v)$ of $D_-(v)$ is meromorphic on $\re v>1$. On the $2$-di\-men\-sional as well as on the $1$-dimensional continuous spectral components it is holomorphic.
On the discrete spectrum $R_+(u)$ is meromorphic with a finite number of poles corresponding to the roots of
$((\lambda_1+\Lambda_2)^2-v^2)((\lambda_1-\Lambda_2)^2-v^2)$ for infinitesimal characters $\Lambda=(\Lambda_1,\Lambda_2)$.
\end{prop}
Iterated application (see~\cite[sec.~4]{holproj}) of the resolvents $R_+(u)$ and $R_-(v)$ to the Poincar\'e series yields their meromorphic continuation as $L^2$-functions
to the largest area on which the resolvents exist, that is to the cone
\begin{equation*}
\{(u,v)\in {\mathbb C}^2\mid \re u>\frac{1}{2},\: \re v>1\}\:.
\end{equation*}
{\it Step 2. The $L^2$-limit $P(\cdot,1,3)\::=\: \lim_{u\to 1} P(\cdot,u,2u+1)$ exists.}
Let
\begin{equation*}
L^2(\Gamma\backslash G)\:=\: L^2_{\re\Lambda=0} (\Gamma\backslash G)\:\bigoplus_{\gamma,c}\:L^2_{\gamma,c}(\Gamma\backslash G)
\:\bigoplus_\Lambda \:L^2_\Lambda(\Gamma\backslash G)
\end{equation*}
be the spectral decomposition of $L^2(\Gamma\backslash G)$ and denote by
\begin{equation*}
P_\bullet(\cdot,u,v)
\end{equation*}
the according spectral components of the Poincar\'e series.
Notice that $(u,v)=(1,3)$ is an inner point of the area of meromorphicity.
We analyze the operator $D_+(u)$.
We choose $v=2u+1$, so the limit series has equation $s_1=0$.
Equation (\ref{Omega_groesser_angewandt}) simplifies on the intersection of $s_1=0$ with the cone of convergence to
\begin{equation*}
D_+(u) P(\cdot,u,2u+1)\:=\:64\pi^2\det(T)(u-1)(u-2)P(g,u+2,2u+5)\:,
\end{equation*}
where $P(\cdot, u+2,2u+5)$ actually is convergent in $(u,v)=(1,3)$.
As the meromorphic continuation is unique, this holds everywhere on $s_1=0$.
Equivalently, as $D_+(1)=D_+(u)-(u^2-1)(C_1-(u^2-4))$,
\begin{eqnarray*}
D_+(1) P(\cdot,u,2u+1) &=&
(u^2-1)\bigl(C_1-(u^2-4)\bigr) P(\cdot,u,2u+1)\\
&&+64\pi^2\det(T)(u-1)(u-2)P(\cdot,u+2,2u+5)\:.
\end{eqnarray*}
Thus,
\begin{eqnarray*}
D_+(1)^n P(\cdot,u,2u+1) &=&
(u^2-1)^n
\bigl(C_1-(u^2-4)\bigr)^n P(\cdot,u,2u+1)\\
&&+(u-1)\mathcal P(\cdot, u)\:,
\end{eqnarray*}
where $\mathcal P(\cdot, u)$ is a symbol for a ${\mathbb C}[u]$-linear combination of Poincar\'e series which actually converge in $(u,v)=(1,3)$.
Choosing $n$ greater than the pole order of $P(\cdot,u,v)$ in $(u,v)=(1,3)$, we have
\begin{equation*}
\lim_{u\to 1}\:\lvert\!\lvert (u^2-1)^n
\bigl(C_1-(u^2-4)\bigr)^n P(\cdot,u)\rvert\!\rvert\:=\:0
\end{equation*}
as well as
\begin{equation*}
\lim_{u\to 1}\:\lvert\!\lvert (u-1)\mathcal P(\cdot, u)\rvert\!\rvert\:=\:0\:.
\end{equation*}
Applying Schwarz' inequality we deduce
\begin{equation*}
\lim_{u\to 1}\:\lvert\!\lvert D_+(1)^n P(\cdot,u,2u+1)\rvert\!\rvert^2\:=\:0\:.
\end{equation*}
Written according to the spectral decomposition,
\begin{eqnarray*}
0&=& \sum_\Lambda \lvert D_+(1,\Lambda)\rvert^{2n}\lim_{u\to 1}\lvert\!\lvert P_\Lambda(\cdot,u,2u+1)\rvert\!\rvert^2\\
&& +\sum_{\gamma,c}\lim_{u\to 1}\lvert\!\lvert D_+(1)^nP_{\gamma,c}(\cdot,u,2u+1)\rvert\!\rvert^2\\
&&+\lim_{u\to 1}\lvert\!\lvert D_+(1)^nP_{\re\Lambda=0}(\cdot,u,2u+1)\rvert\!\rvert^2\:.
\end{eqnarray*}
So any single summand is zero:
The limit
$\lim_{u\to 1}\lvert\!\lvert P_\Lambda(\cdot,u,2u+1)\rvert\!\rvert$ exits for any discrete parameter $\Lambda$.
It is nonzero only if $ D_+(1,\Lambda)=(\Lambda_1^2-1)(\Lambda_2^2-1)$ is zero.
So the remaining nonzero discrete
components $P_{\Lambda}(\cdot,u,v)$ have parameters $\Lambda=(1,\Lambda_2)$ and are indeed analytically continued
in $(u,v)=(1,3)$.
On any continuous component apart from $K_\alpha(1)$,
the resolvent $R_+(1)$ exists, thus there we have
\begin{eqnarray*}
\lim_{u\to 1}\lvert\!\lvert P_{\bullet}(\cdot,u,2u+1)\rvert\!\rvert^2
&=& \lim_{u\to 1}\lvert\!\lvert R_+(1)^nD_+(1)^nP_{\bullet}(\cdot,u,2u+1)\rvert\!\rvert^2\\
&\leq & \lvert\!\lvert R_+(1)\rvert\!\rvert^{2n}_\bullet\cdot\lim_{u\to 1}
\lvert\!\lvert D_+(1)^n P_{\bullet}(\cdot,u,2u+1)\rvert\!\rvert^2
=0\:.
\end{eqnarray*}
On the component $K_\alpha(1)$ we have $D_+(u)=(u^2-1)M_+(u)$, where $M_+(u)$ can be parametrized by $(u^2+t^2)$ and is bounded from below by $u^2$.
So from
\begin{equation*}
0\:=\: \lim_{u\to 1}\lvert\!\lvert D_+(1)^nP_{\alpha,1}(\cdot,u,2u+1)\rvert\!\rvert \: =\:\lim_{u\to 1}(u^2-1)^n \lvert\!\lvert M_+(1)^nP_{\alpha,1}(\cdot,u,2u+1)\rvert\!\rvert
\end{equation*}
we deduce as before that $\lvert\!\lvert M_+(1)P_{\alpha,1}(\cdot,u,2u+1)\rvert\!\rvert$ exists.
As the resolvent $M_+^{-1}(1)$ is an operator bounded by $u^{-1}=1$, the limit
\begin{eqnarray*}
\lim_{u\to 1}\lvert\!\lvert P_{\alpha,1}(\cdot,u,2u+1)\rvert\!\rvert^2
&=& \lim_{u\to 1}\lvert\!\lvert M_+(1)^{-n}M_+(1)^nP_{\alpha,1}(\cdot,u,2u+1)\rvert\!\rvert^2\\
&\leq & \lvert\!\lvert M_+^{-1}(1)\rvert\!\rvert^{2n}_{\alpha,1}\cdot\lim_{u\to 1}
\lvert\!\lvert M_+(1)^n P_{\alpha,1}(\cdot,u,2u+1)\rvert\!\rvert^2
\:
\end{eqnarray*}
exists.
So the limit $P(\cdot,1,3):=\lim_{u\to 1}P(\cdot,u,2u+1)$ exists as an $L^2$-function.
\bigskip
{\it Step 3. The spectral locus of $P(\cdot,u,v)$ in $(1,3)$.}
We examine the possible spectral components left from Step~2.
Within the continuous spectrum, the only remaining component is indexed by $\Lambda=(1,i{\mathbb R})=K_{\alpha_1}(1)$. But by Proposition~\ref{lemma_kein_kont_spek}, in this component the $K$-type
$\kappa=(3,3)$ does not occur. As the $K$-type is passed on the continuation, the $P(\cdot,1,3)$ has weight $\kappa$. So its continuous spectral component is
identically zero.
The remaining discrete spectral components are indexed by $\Lambda=(\Lambda_1,1)$. By Proposition~\ref{lem_diskretes_spek}, only two components occur within $L^2(\Gamma\backslash G)_\kappa$:
The holomorphic discrete series representation $\pi_\kappa$ of minimal $K$-type $\kappa=(3,3)$ and of infinitesimal character $\Lambda=(2,1)$, and a non-discrete series representation
of infinitesimal character $\Lambda=(0,1)$.
{\it Step 4. There is a $C^\infty$-representative.}
The limit $P(\cdot,1,3)$ is the solution of an elliptic differential equation with $C^\infty$-coefficients. So itself is $C^\infty$ by regularity theory.
\end{proof}
\section{Phantom holomorphic projection}\label{holomorphe_projektion}
\subsection{Differential operators}\label{diffoperators}
Let $\rho$ be an irreducible unitary representation of $U(m)$ on a finite dimensional vector space $V_\rho$.
By the homomorphism $J:G\to\GL_m({\mathbb C})$, $J(g)=(ci+d)$, we get an isomorphism $\psi=J\mid_K:K\tilde\to U_m({\mathbb C})$. So $\rho(g):=\rho\circ\psi(g)$ is an irreducible unitary representation of $K$.
As $\rho$ is the restriction of an irreducible representation of $\GL_m({\mathbb C})$, the element $J_\rho(g)=\rho(ci+d)\in\mathop{Aut}(V_\rho)$ is well-defined for all $g\in G$.
In the following we make use of formulas developed in \cite[\S 3]{weissauersLN}. The notation there is according to the choice of the isomorphism $K\cong U(m)$
given by the complex conjugate $\bar\psi$ of $\psi$. This implies to work instead of $\rho$ with the representation $\rho(\bar\psi(g))=\rho(\bar g)$ on $V_\rho$, which
is isomorphic to the contragredient representation $\rho^\ast\circ\psi$ of $K$. So we must carfully replace $\rho^\ast$ by $\rho$ in some of the formulas in \cite[\S 3]{weissauersLN}.
For the Siegel upper halfspace $\mathcal H=G/K$ of genus $m$ let $C^\infty(\mathcal H, V_\rho)$ be the space of $C^\infty$-functions on $\mathcal H$ with values in the space $V_\rho$ of
the $K$-representation $\rho$. We have the isomorphism \cite[p. 30]{weissauersLN}
\begin{eqnarray*}
C^\infty(G,V_{\rho\circ\bar\psi})^K&\tilde\to& C^\infty(\mathcal H, V_\rho)\:,\\
f(g)&\mapsto& J_\rho(g)f(g)\:.
\end{eqnarray*}
Here $C^\infty(G,V_\rho)^K$ is the subspace of $K$-invariant functions in $C^\infty(G,V_\rho)=C^\infty(G)\otimes V_\rho$, on which $K$ acts by right translations $R_gf(x)=f(xg)$.
By Schur's lemma $C^\infty(G,V_\rho)^K$ is the $K$-isotypical component for the representation $(\rho\circ\bar\psi)^\ast$. But this is isomorphic to $\rho\circ\psi$ and
we can identify $C^\infty(G,V_\rho)^K$ with $C^\infty(G)_\rho$, the $\rho$-isotypical component of $C^\infty(G)$ on which $K$ acts by right translations. So we get isomorphisms
\begin{equation*}
\phi_\rho\::\:C^\infty(G)_\rho\:\tilde\to\: C^\infty(\mathcal H, V_\rho)\:.
\end{equation*}
The universal envelopping algebra $\mathfrak U({\mathfrak g}_{\mathbb C})$ of the complex Lie algebra ${\mathfrak g}_{\mathbb C}$ of $G$ acts from the right on $C^\infty(G)$ and this action commutes with the left action of $G$.
The abelian Lie algebra ${\mathfrak p}_+$ (respectively ${\mathfrak p}_-$) can be identified with the holomorphic (respectively antiholomorphic) tangent space of $\mathcal H$ in $Z=iE$.
So $\mathfrak U({\mathfrak p}_-)$ acts on $C^\infty(G)$ by leftinvariant differential operators.
\begin{lem}\label{pplus_hoechstgewichte}
For the adjoint representation of $K$ on $\mathfrak U(\mathfrak p_+)$ it holds
\begin{equation*}
\mathfrak U(\mathfrak p_+)\:\cong\:\Symm^\bullet\Symm^2({\mathbb C}^m) \:=\: \bigoplus_{\rho_k} V_{\rho_k}\:,
\end{equation*}
where $\rho_k$ runs through the irreducible repesentations of $K$ of highest weight $k=(k_1,\dots,k_m)$ for $k_i\in 2{\mathbb Z}$ and $k_1\geq k_2\geq\dots\geq k_m\geq 0$.
\end{lem}
\begin{proof}[Proof of Lemma~\ref{pplus_hoechstgewichte}]
This is \cite[Lemma~3]{weissauersLN} with respect to the change from $\bar\psi$ to $\psi$.
\end{proof}
The representation of $K$ on $ \mathfrak U(\mathfrak p_-)$ is dual to that on $\mathfrak U(\mathfrak p_+)$.
For any irreducible representation $\rho$ in $\mathfrak U(\mathfrak p_+)$ there are operators $E_+^\rho$ on the $\rho$-isotypical component of $\mathfrak U(\mathfrak p_+)$,
respectively $E_-^\rho$ in $\mathfrak U(\mathfrak p_-)$ on the $\rho^\ast$-isotypical component of $\mathfrak U(\mathfrak p_-)$ \cite[p.43f.]{weissauersLN}.
They map
a $K$-isotypical subspace $C^\infty(G)_\tau$ of $C^\infty (G)$ to the direct sum of $K$-isotypical subspaces $C^\infty(G)_{\tilde\tau}$, where
\begin{equation*}
\rho\otimes \tau\:=\:\bigoplus \tilde\tau\:,
\end{equation*}
respectively $\rho^\ast\otimes \tau$ in case $E_-^{\rho}$.
We get the Maass operators $E_+$ respectively $E_-$ by choosing $\rho=\rho_k$ of highest weight $k=(2,0,\dots,0)$.
More generally, we get Maass operators $E_+^{[\mu]}$ respectively $E_-^{[\mu]}$ by choosing
\begin{equation*}
k=(2,\dots,2,0,\dots,0)
\end{equation*}
where $\mu$ is the number of $2$s occurring.
By the above identifications $\phi_\tau$ and $\phi_{\rho\otimes\tau}=\bigoplus_{\tilde\tau}\phi_{\tilde\tau}$ we have the commutative diagramm
\begin{center}
\begin{tikzcd}
C^\infty(G)_\tau\arrow{d}{E_+^\rho} \arrow{r}{\phi_\tau} & C^\infty(\mathcal H,V_\tau) \arrow{d}{\Delta_+^\rho} \\
C^\infty(G)_{\rho\otimes\tau} \arrow{r}{\phi_{\rho\otimes\tau}} & C^\infty(\mathcal H,V_{\rho\otimes\tau})
\end{tikzcd},
\end{center}
respectively
\begin{center}
\begin{tikzcd}
C^\infty(G)_\tau\arrow{d}{E_-^\rho} \arrow{r}{\phi_\tau} & C^\infty(\mathcal H,V_\tau) \arrow{d}{\Delta_-^\rho} \\
C^\infty(G)_{\rho^\ast\otimes\tau} \arrow{r}{\phi_{\rho^\ast\otimes\tau}} & C^\infty(\mathcal H,V_{\rho^\ast\otimes\tau})
\end{tikzcd}.
\end{center}
An explicit description of the operators $\Delta_+^{[\mu]}$ respectively $\Delta_-^{[\mu]}$ is found in \cite[pp. 33, 44]{weissauersLN}.
$\Delta_+^{[\mu]}$ acts on $C^\infty(\mathcal H,V_\tau)=C^\infty(\mathcal H)\otimes V_\tau$ with values in $C^\infty(\mathcal H)\otimes V_{(2,\dots,2,0,\dots,0)}\otimes V_\tau$
For $h\in C^\infty(\mathcal H)$ and $v\in V_\tau$ it is defined up to a factor $2^\mu$ by
\begin{equation*}
\Delta_+^{[\mu]}\left(h(Z)\otimes v\right)\:=\:
(2i)^\mu(\tau\otimes \mathop{det}\nolimits^{\frac{1-\mu}{2}})(Y^{-1})\cdot \partial_Z^{[\mu]}\left((\tau\otimes\mathop{det}\nolimits^{\frac{1-\mu}{2}})(Y) h(Z)\otimes v\right)\:.
\end{equation*}
Here $\partial_Z=\frac{1}{2}(\partial_X-i\partial_Y)$ is the matrix valued operator with components $\frac{1}{2}(1+\delta_{ij})\frac{\partial}{\partial_{Z_{ij}}}$ for the symmetric
matrix $Z=X+iY\in\mathcal H$ with components $Z_{ij}$, and for a matrix $M$ let
$M^{[\mu]}=\bigwedge^\mu(M)$ be the matrix of the $\mu$-th exterior power of $M$, i.e. the $\binom{m}{\mu}\times\binom{m}{\mu}$-matrix of the minors of $M$ of size $\mu$
(see~\cite[p. 208ff]{freitag}).
Analogously, $\Delta_-^{[\mu]}$ acts on on $C^\infty(\mathcal H,V_\tau)=C^\infty(\mathcal H)\otimes V_\tau$ with values in $C^\infty(\mathcal H)\otimes V_{(0,\dots,0,-2,\dots,-2)}
\otimes V_\tau$ via
\begin{equation*}
\Delta_-^{[\mu]}\left(h(Z)\otimes v\right)\:=\: \Delta_-^{[\mu]}\left(h(Z)\right)\otimes v\:
\end{equation*}
for $h\in C^\infty(\mathcal H)$ and $v\in V_\tau$ and is defined up to a factor $2^\mu$ by
\begin{equation*}
\Delta_-^{[\mu]}\left(h(Z)\right)\:=\:
(-2i)^\mu Y^{[\mu]}(\mathop{det}\nolimits^{\frac{1-\mu}{2}})(Y^{-1})\cdot \bar\partial_{Z}^{[\mu]}\left((\mathop{det}\nolimits^{\frac{1-\mu}{2}})(Y)\cdot h(Z)\right)\cdot Y^{[\mu]}\:.
\end{equation*}
\subsection{Sturm's operator}
Let
$\mathcal Y=\{Y=Y'\in M_m({\mathbb R})\mid Y>0\}$ be the space of positive definite symmetric matrices and let
$\mathcal X=\{X=X'\in M_m({\mathbb R})\mid \lvert X_{jk}\rvert\leq \frac{1}{2}, 1\leq j,k\leq m\}$ be a stripe of width one.
Let the genus $m$ equal two.
Let $F\in \tilde{\mathcal M}_{\kappa}(\Gamma)$ be a (nonholomorphic modular) form of bounded growth (see \cite[2.4]{panchishkin})) of weight $\kappa$.
Let $f(g)=J_\kappa(g,i)^{-1}F(gi)$ be the preimage of $F$ under $\phi_\rho$ for $\rho=(\kappa,\kappa)$. Define correspondingly
$e_{T}(g)=J_\kappa(g,i)^{-1}\exp(2\pi i\trace(Tgi))$.
Sturm's operator for weight $\kappa$ for positive definite $T$ is
\begin{equation*}
F\:\mapsto\:a(T)\::=\:c(\kappa)^{-1}\det(T)^{\kappa-\frac{3}{2}}\int_{\Gamma_\infty \backslash G/K} f(g)\overline{e_T(g)}~dg\:.
\end{equation*}
Here $c(\kappa)=\sqrt \pi(4\pi)^{3-\kappa}\Gamma(\kappa-\frac{3}{2})\Gamma(\kappa-2)$ (\cite[p. 84]{panchishkin}) is chosen such that
Sturm's operator is the identity on Fourier coefficients of holomorphic forms.
Up to a constant Sturm's operator is the invariant pairing of $F$ with
$\exp(2\pi i\trace(TZ))$ of level $\kappa$,
\begin{equation*}
F\:\mapsto\: a(T)\:=\:c(\kappa)^{-1}\det(T)^{\kappa-\frac{3}{2}}\int_{\Gamma_\infty\backslash\mathcal H} F(Z)\exp(-2\pi i\trace(TZ))\det(Y)^{\kappa}dv_Z\:,
\end{equation*}
where $dv_Z=\frac{dX}{\det(Y)^{3/2}}\frac{dY}{\det(Y)^{3/2}}$ is the invariant measure on $\mathcal H$ such that $dg=dv_Zdk$, and $dY=\prod_{k\leq l}dY_{kl}$
as well as $dX=\prod_{k\leq l}dX_{kl}$. Replacing $F$ by its Fourier expansion
\begin{equation*}
F(Z)\:=\:\sum_{T=T'}A(T,Y)\exp({2\pi i\trace(T X)})
\end{equation*}
we get
\begin{equation*}
a(T)\:=\:c(\kappa)^{-1}\int_{\mathcal Y} A(T,Y)\exp(-2\pi\trace(TY))\det(TY)^{\kappa-\frac{3}{2}}\frac{dY}{\det(Y)^{\frac{3}{2}}}\:.
\end{equation*}
In case $\kappa\geq 4$ Sturm's operator
\begin{equation*}
S: F(Z)\:\mapsto\: \sum a(T)\exp(2\pi\trace(TZ))
\end{equation*}
realizes the orthogonal projection to the holomorphic part of $F$.
The Fourier expansion
\begin{equation*}
\tilde F(Z)\:=\:\sum_{T>0}a(T)e^{2\pi i\trace(T Z)}
\end{equation*}
gives rise to a holomorphic cuspform $\tilde F\in[\Gamma,\kappa]_0$ of weight $\kappa$, and for all $f\in[\Gamma,\kappa]_0$ it holds
\begin{equation*}
\langle F,f\rangle\: =\: \langle \tilde F,f\rangle\:,
\end{equation*}
where $\langle \cdot,\cdot\rangle$ is the scalar product of the Hilbert space $L^2_\kappa(\Gamma\backslash \mathcal H)$.
This is shown in \cite{holproj} (see also \cite{panchishkin} in case $\kappa\geq 5$) by using Poincar\'e series of weight $\kappa$ defined analogously
to ours which are holomorphic cuspforms for $\kappa\geq 4$.
We are interested in the action of Sturm's operator on images of $\Delta_+^{[\mu]}$ in the special case $\mu=2$. So let $k=(2,2)$ and $\rho_k=\mathop{det}^2$ then
\begin{equation*}
\Delta_+^{[2]}\::\:C^\infty(\mathcal H_2,V_\tau)\:\to\:C^\infty(\mathcal H_2,V_{\tau\otimes\mathop{det}\nolimits^2})\:,
\end{equation*}
and $\Delta_+^{[2]}\circ\phi_\tau=\phi_{\tau\otimes \mathop{det}\nolimits^2}\circ\det(E_+^{[2]})$, is explicitly given on $\mathcal H$ by
\begin{equation*}
\Delta_+^{[2]}(h)(Z)\:=\:(2i)^2(\tau\otimes \mathop{det}\nolimits^{-\frac{1}{2}})(Y^{-1})\mathop{det}(\partial_Z)\left((\tau\otimes\mathop{det}\nolimits^{-\frac{1}{2}})(Y)h(Z)\right)\:.
\end{equation*}
\begin{prop}\label{prop_sturm_operator}
Let $h\in[\Gamma,k]_0$ be a holomorphic cuspform on $\mathcal H$ of weight $(k,k)$, where $\kappa=k+2$.
Sturm's operator applied to the Fourier coefficients of the nonholomorphic image $\Delta_+^{[2]}(h)$ is zero in case $\kappa\geq 4$.
But for $\kappa=3$ it does not vanish.
\end{prop}
\begin{proof}[Proof of Proposition~\ref{prop_sturm_operator}]
We apply $\Delta_+^{[2]}$ to a holomorphic cuspform $h\in[\Gamma,k]_0$ on $\mathcal H_2$ of weight $(k,k)$. So $V_\tau={\mathbb C}$ and
\begin{equation*}
\tilde h(Z)\:=\:\Delta_+^{[2]}(h)(Z)\:=\: -4\det(Y^{-1})^{k-\frac{1}{2}}\mathop{det}(\partial_Z)\left(\det(Y)^{k-\frac{1}{2}}h(Z)\right)
\end{equation*}
is a function on $\mathcal H_2$.
(This formula for Maass' operator is also due to Shimura~\cite{shimura} and can be found in \cite[3.1]{panchishkin}.)
It has $K$-type $(k+2,k+2)$ and belongs to the automorphic representation generated by the holomorphic cuspform $h$.
Let
\begin{equation*}
h(Z)\:=\:\sum_{T=T'}a(T)\exp(2\pi i\trace(TZ))
\end{equation*}
be its Fourier expansion. The Fourier coefficients $b(T,Y)$ of the function
\begin{equation*}
\tilde h(Z)\:=\: b(T;Y)\exp(2\pi i\trace(TZ))
\end{equation*}
are
\begin{equation*}
-4a(T)\det(Y)^{\frac{1}{2}-k}\mathop{det}(\partial_Z)\left(\det(Y)^{k-\frac{1}{2}}\exp(2\pi i\trace(TZ))\right)\cdot \exp(-2\pi i\trace(TZ))\:.
\end{equation*}
Let
\begin{equation*}
A(T,Y)\:=\: b(T,Y)\exp(-2\pi \trace(TY))
\end{equation*}
be the coefficient in the Fourier expansion in $X$ only. For Sturm's formula we study whether the limit
\begin{equation}\label{sturm_limit}
\lim_{s\to 0}\int_{\mathcal Y}A(T,Y)\exp(-2\pi\trace(TY))\det(TY)^{k+2+s-\frac{3}{2}}\frac{dY}{\det(Y)^{\frac{3}{2}}}
\end{equation}
is zero.
We make use of the Lemmas~\ref{lemma_freitag_1}, \ref{lemma_freitag_2}, and \ref{lemma_gamma_level_2} below.
Applying Lemma~\ref{lemma_freitag_1} to the functions $f(Z)=\det(Y)^{k-\frac{1}{2}}$ and $g(Z)=\exp(2\pi i\trace(TZ))$ we get
\begin{eqnarray*}
\mathop{det}(\partial_Z)\left(\det(Y)^{k-\frac{1}{2}}\exp(2\pi i\trace(TZ))\right)&=&
-\frac{1}{4}C_2(k-\frac{1}{2})\det(Y)^{k-\frac{3}{2}}g\\
&&
-\frac{i}{2}C_1(k-\frac{1}{2})\det(Y)^{k-\frac{3}{2}}\trace(Y2\pi iT)g\\
&&
+\det(Y)^{k-\frac{1}{2}}(2\pi i)^2\det(T)g\:,
\end{eqnarray*}
where $C_2(k-\frac{1}{2})=(k-\frac{1}{2})k$ and $C_1(k-\frac{1}{2})=k-\frac{1}{2}$ by Lemma~\ref{lemma_freitag_2}.
So the Fourier coefficient $b(T,Y)$ is
\begin{equation}\label{derivative_fourier_coeff}
(4\pi)^2a(T)\det(T)\left(
\frac{k(k-\frac{1}{2})}{(4\pi)^2}\det(TY)^{-1}-\frac{(k-\frac{1}{2})}{4\pi}\det(TY)^{-1}\trace(YT)+1\right)\:.
\end{equation}
By a change of variables
\footnote{The integral is $\int_{\mathcal Y}b(T;Y)\exp(-4\pi\trace(TY))\det(TY)^{k-1/2+s}\frac{dY}{\det(Y)^{3/2}}$, which equals
$(4\pi)^2a(T)\det(T)\int_{\mathcal Y}\left(k(k-1/2)(4\pi)^{-2}\det(TY)^{-1}-(k-1/2)(4\pi)^{-1}\det(TY)^{-1}\trace(TY)+1\right)$
$\times \exp(-4\pi\trace(TY))\det(TY)^{k-1/2+s}\frac{dY}{\det(Y)^{3/2}}$.
For the change of variables $Y\mapsto 4\pi T^{\frac{1}{2}}Y T^{\frac{1}{2}}$ this equals
$(4\pi)^{-2(k+1/2+s-1)}a(T)\det(T)\sqrt{\pi}s(s-1/2)\Gamma(s+k-1/2)\Gamma(s+k-1)$}
$Y\mapsto 4\pi T^{\frac{1}{2}}Y T^{\frac{1}{2}}$, for the limit (\ref{sturm_limit}) we have to compute the integral
\begin{equation*}
\int_{\mathcal Y}\left(k(k-\frac{1}{2})\det(Y)^{-1}
-(k-\frac{1}{2})\det(Y)^{-1}\trace(Y)+1\right)\exp(-\trace(Y))\det(Y)^{k+\frac{1}{2}+s}\frac{dY}{\det(Y)^\frac{3}{2}}
\end{equation*}
up to the factor $(4\pi)^{-2(k-\frac{1}{2}+s)}a(T)\det(T)$,
which by Lemma~\ref{lemma_gamma_level_2} is given (up to a factor $\sqrt\pi$) by
\begin{eqnarray*}
&&k(k-\frac{1}{2})\Gamma(s+k-\frac{1}{2})\Gamma(s+k-1)\\
&&\hspace*{1cm}-2(k-\frac{1}{2})(s+k-\frac{1}{2})\Gamma(s+k-\frac{1}{2})\Gamma(s+k-1)+\Gamma(s+k+\frac{1}{2})\Gamma(s+k)\:.
\end{eqnarray*}
But this equals
\begin{equation*}
s(s-\frac{1}{2})\Gamma(s+k-\frac{1}{2})\Gamma(s+k-1)\:.
\end{equation*}
So for all $k>1$ respectively $\kappa=k+2>3$ the limit (\ref{sturm_limit}) is zero,
\begin{equation*}
\lim_{s\to 0}\int_{\mathcal Y}A(T,Y)\exp(-2\pi\trace(TY))\det(TY)^{k+2+s-\frac{3}{2}}\frac{dY}{\det(Y)^{\frac{3}{2}}}\:=\:0\:.
\end{equation*}
While in case $k=1$ (i.e. $\kappa=3$) it is a multiple of
\begin{equation*}
\lim_{s\to 0}s(s-\frac{1}{2})\Gamma(s+\frac{1}{2})\Gamma(s)\:=\:-\frac{1}{2}\Gamma(\frac{1}{2})\:\not=\:0\:.
\end{equation*}
Collecting constants and having in mind $c(3)=\frac{\pi}{2}$, the Sturm operator maps
\begin{equation*}
A(T,Y)\:\mapsto\: -\frac{1}{4\pi}a(T)\det(T)\:
\end{equation*}
in case $k=1$, while it is zero for $k>1$.
\end{proof}
\begin{remark}
In case of weights $(k,k)$ for $k=0, \frac{1}{2}$ the Fourier coefficients (\ref{derivative_fourier_coeff}) vanish.
For $k=\frac{1}{2}$ this follows from the theory of singular modular forms \cite{freitag_singular}, \cite{resnikoff}.
So for these weights Sturm's operator is supposed to establish the holomorphic projection, too.
\end{remark}
\begin{lem}\label{lemma_freitag_1}
For quadratic matrices $A,B$ define the symbol $2\cdot(A\cap B)$ by
\begin{equation*}
(A+B)^{[2]}\:=\:A^{[2]}+2\cdot (A\cap B)+B^{[2]}\:.
\end{equation*}
Then it holds
\begin{equation*}
\partial_Z^{[2]}(fg)\:=\: \partial_Z^{[2]}(f)g+2(\partial_Z(f)\cap\partial_Z(g))+f\partial_Z^{[2]}(g)\:.
\end{equation*}
Especially for $m=2$, it holds
\begin{equation*}
2\cdot\left(Y^{-1}\cap T\right)\:=\:\det(Y)^{-1}\trace(YT)\:.
\end{equation*}
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lemma_freitag_1}]
See~\cite[pp. 208, 211]{freitag}. In case of genus $m=2$ we have $A^{[2]}=\det(A)$, so $2(A\cap B)=A_{11}B_{22}+A_{22}B_{11}-A_{12}B_{21}-A_{21}B_{12}$ and the identity
$2(Y^{-1}\cap T)=\det(Y)^{-1}\trace(YT)$ follows.
\end{proof}
\begin{lem}\cite[p. 213]{freitag}\label{lemma_freitag_2}
It holds
\begin{equation*}
\partial_Y^{[h]}\det(Y)^\alpha\:=\:C_h(\alpha)\det(Y)^\alpha(Y^{-1})^{[h]}\:,
\end{equation*}
where $C_h(\alpha)=\alpha(\alpha+\frac{1}{2})\cdots(\alpha+\frac{(h-1)}{2})$.
\end{lem}
\begin{lem}\label{lemma_gamma_level_2}
It holds
\begin{equation}\label{gamma_level_2}
\int_{\mathcal Y} e^{-\trace(T Y)}\det(Y)^{s-3/2}~dY\:=\:\sqrt\pi\det(T)^{-s}\Gamma(s)\Gamma(s-\frac{1}{2})
\end{equation}
as well as
\begin{equation}\label{gamma_level_2_ableitung}
\int_{\mathcal Y} e^{-\trace(Y)}\trace(Y)\det(Y)^{s-1}~dY\:=\: 2\sqrt\pi(s+\frac{1}{2})\Gamma(s)\Gamma(s+\frac{1}{2})\:.
\end{equation}
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lemma_gamma_level_2}]
The identity~(\ref{gamma_level_2}) is well-known (e.g.~\cite[p. 467]{shimura}).
Differentiating by $\partial_T$ in $T=E_2$ we get
\begin{equation*}
\int_{\mathcal Y} e^{-\trace(Y)}Y\det(Y)^{s-3/2}~dY\:=\: sE_2\sqrt\pi\Gamma(s)\Gamma(s-\frac{1}{2})\:,
\end{equation*}
so especially
\begin{equation*}
\int_{\mathcal Y} e^{-\trace(Y)}\trace(Y)\det(Y)^{s-1}~dY\:=\: 2\sqrt\pi(s+\frac{1}{2})\Gamma(s+\frac{1}{2})\Gamma(s\frac{1}{2})\:.
\end{equation*}
\end{proof}
\subsection{Poincar\'e series}
Let $\rho$ be the irreducible representation of $K$ of minimal weight $(\kappa,\kappa)$ for $\kappa=3$. The images
\begin{equation*}
p_T(g i,s) \:=\: \phi_\rho \left(P_T(g,u,2u+1) \right)\:=\: J_\kappa(g,i) P_T(g,u,2u+1)\:,
\end{equation*}
of the Poincar\'e series $P(g,u,2u+1)$ under the isomorphism $\phi_\rho$ define Poincar\'e series on $\mathcal H$,
\begin{equation*}
p_T(Z,s)\:=\:\sum_{\gamma\in\Gamma_\infty\backslash \Gamma}e^{2\pi i\trace(T\gamma Z)}
\frac{\det(\im\gamma Z)^s}{J_\kappa(\gamma,Z)}\:.
\end{equation*}
Here $s$ and $u$ are related by $s=\frac{1}{2}(u-1)$. The Poincar\'e series $p_T$ inherit the analytic properties of their preimages $P_T$.
\begin{cor}
For $\re s+\frac{\kappa}{2}>2$ the series $p_T(z,s)$ converge absolutely and locally uniformly in $s$ and uniformly on the Siegel fundamental domain $\mathcal F$ for $\Gamma$.
They belong to $L^2_\kappa(\Gamma\backslash \mathcal H)$, the Hilbert space of functions on $\Gamma\backslash \mathcal H$ of weight
$\kappa$ with scalar product given by
\begin{equation*}
\langle f,g\rangle \:=\:\int_{\mathcal F} f(Z)\overline{g(Z)}\det(Y)^\kappa~dv_Z\:,
\end{equation*}
where $dv_Z=\det(Y)^{-(m+1)}dXdY$.
They have meromorphic continuations to $\re s>-\frac{1}{2}$ as functions in $L^2_\kappa(\Gamma\backslash \mathcal H)$, which are analytic in the critical point $s=0$. That is,
the limit
\begin{equation*}\label{poincare_auf_H}
p_T(\cdot)\::=\:\lim_{s\to 0}p_T(\cdot,s)\:\in\: L^2_\kappa(\Gamma\backslash \mathcal H)
\end{equation*}
exists in $L^2_\kappa(\Gamma\backslash \mathcal H)$ and is $C^\infty$.
\end{cor}
\begin{thm}\label{satz_phantom_proj}
The analytic continuations $p_T$ of the Poincar\'e series of weight $(3,3)$ to the critical point $s=0$ decompose into
\begin{equation*}
p_T\:=\: f_T\:+\:\Delta_+^{[2]}(h_T)\:,
\end{equation*}
where $f_T\in[\Gamma,3]_0$ is a holomorphic cuspform of weight $(3,3)$, and $h_T\in[\Gamma,1]$ is a holomorphic modular form of weight $(1,1)$.
In general the two forms $f_T$ and $h_T$ are nonzero.
The form $h_T$ is recovered from $p_T$ by application of the antiholomorphic Maass operator $\Delta_-^{[2]}$,
\begin{equation*}
\Delta_-^{[2]} (p_T)\:=\: \frac{3}{4}\cdot\: h_T\:.
\end{equation*}
\end{thm}
\begin{proof}[Proof of Theorem~\ref{satz_phantom_proj}]
The analytic continuations $p_T$ of the Poincar\'e series to $s=0$ have the claimed decomposition by Theorem~\ref{analytische_fortsetzung}.
Let
\begin{equation*}
F(Z)\:=\:\sum_{T=T'}A(T,Y)e^{2\pi i\trace(T X)}
\end{equation*}
be the Fourier expansion of a nonholomorphic modular form of bounded growth of weight $3$.
We use unfolding to
get for $\re s+\frac{3}{2}>2$
\begin{eqnarray*}
\langle F, p_T(\cdot,\bar s)\rangle &=&
\int_{\mathcal F}F(Z)\sum_{\gamma\in\Gamma_\infty\backslash \Gamma}e^{-2\pi i\trace(T\gamma\bar Z)}
\frac{\det(\im\gamma Z)^{s}}{\overline{j(\gamma,Z)}^3}\det(Y)^3~dv_Z\\
&=&
\int_{\mathcal F}\sum_{\gamma\in\Gamma_\infty\backslash \Gamma}F(\gamma Z)e^{-2\pi i\trace(T\gamma\bar Z)}
\det(\im\gamma Z)^{s+3}~dv_Z\\
&=&
\int_{\mathcal X}\int_{\mathcal Y} F(Z)e^{-2\pi i\trace(T\bar Z)}\det(Y)^{s+3}~dv_Z\:.
\end{eqnarray*}
But the last integral
\begin{eqnarray*}
&&\int_{\mathcal X}\int_{\mathcal Y} F(Z)e^{-2\pi i\trace(T \bar Z)}\det(Y)^{s+3}~dv_Z\\
&&\hspace*{1cm}=\int_{\mathcal X}\int_{\mathcal Y}\sum_{\tilde T}A(\tilde T,Y)e^{2\pi i\trace((\tilde T-T)X)}
e^{-2\pi\trace(TY)}\det(Y)^{s}~dX~dY\\
&&\hspace*{1cm}=\int_{\mathcal Y} A(T,Y)e^{-2\pi\trace(TY)}\det(Y)^{s}~dY\:
\end{eqnarray*}
exists for $\re s\geq 0$ (this indeed is the definition of bounded growth), and its
value at $s=0$ is up to a factor the image $a(T)$
of Sturm's operator applied to $F$.
As analytic continuation is unique, we have
\begin{equation*
\langle F,p_T\rangle\:=\:c(3)\det(T)^{-\frac{3}{2}}a(T)\:.
\end{equation*}
This especially applies to the nonholomorphic function $F=\Delta_+^{[2]}(h)$ for a holomorphic cuspform $h\in[\Gamma,1]_0$, where Sturm's operator is seen to be nonzero by
Proposition~\ref{prop_sturm_operator}. So the nonholomorphic component $\Delta_+^{[2]}(h_T)$ cannot be zero in general.
Similarly choosing $F\in[\Gamma,3]_0$, Sturm's operator is the identity on $F$. So the holomorphic component $f_T$ does not vanish in general.
The operator $\Delta_-^{[2]}$ is explicitly given by
\begin{equation*}
\Delta_-^{[2]} (f(Z))\:=\:(-4)\det(Y)^{5/2}\mathop{det}(\bar\partial_{ Z})\left(\det(Y)^{-1/2}f(Z)\right)\:.
\end{equation*}
Applied to the Poincar\'e series $p_T=f_T+\Delta_+^{[2]} (h_T)$ we have
\begin{equation*}
\Delta_-^{[2]} (p_T)\:=\: \Delta_-^{[2]}\circ\Delta_+^{[2]} (h_T)\:,
\end{equation*}
as the holomorphic cuspform $f_T$ is deleted by $\bar\partial_{ Z}$ and thus by $\Delta_-^{[2]}$. We first find for the weight $(1,1)$-function $h_T$
\begin{equation*}
\Delta_+^{[2]} (h_T)\:=\:\frac{1}{2}\det(Y)^{-1}h_T+2i\det(Y)^{-1}\trace(Y\partial_Z(h_T))-4\mathop{det}(\partial_{ Z})(h_T)\:.
\end{equation*}
Applying $\Delta_-^{[2]}$ to this sum, the second and third term delete each other, while for the first we have
\begin{eqnarray*}
\Delta_-^{[2]}\left(\frac{1}{2}\det(Y)^{-1}h_T\right)&=& -2\det(Y)^{\frac{5}{2}}\mathop{det}(\bar\partial_{ Z})\left(\det(Y)^{-\frac{3}{2}}h_T\right)\\
&=&\frac{1}{2}\det(Y)^{\frac{5}{2}} h_T\cdot C_2(-\frac{3}{2})\det(Y)^{-\frac{5}{2}} \:=\:\frac{3}{4}\cdot h_T\:,
\end{eqnarray*}
so
\begin{equation*}
\Delta_-^{[2]}\circ\Delta_+^{[2]} (h_T)\:=\: \frac{3}{4}\cdot h_T\:.
\end{equation*}
\end{proof}
|
\section{Introduction}
\label{Intro}
\IEEEPARstart{V}{isual} attention facilitates our ability to rapidly locate the most important information in a scene \cite{ Yarbus1967eye,neisser1967cognitive}. Such image regions are said to be salient since it is assumed that they attract greater attention by the visual system than other parts of the image. These sali-ent regions are expected to possess distinctive features when compared with others in the image. The study of saliency detection may reveal the attentional mechanisms of biological visual systems, as well as model their fixation selection behavior. On the other hand, as a component of low-level artificial vision processing, it facilitates subsequent processing such as object detection or recognition by reducing computational cost, which is a key consideration in real-time applications. For object detection, this would always be more efficient than dense sampling, provided one could ensure the accuracy of the attentional mechanism.
Visual saliency detection has received extensive attention by both psychologists and computer vision researchers \cite{Itti_etal98pami, Itti_Koch01nrn, tsotsos2008roles, NIPS2005_81, saliencypoggio, NIPS2006_897, le2006coherent,kienzle2007nonparametric, mahadevan2009spatiotemporal, NIPS2007_874, NIPS2007_1074, tsotsos2011computational, NIPS2008_0142, ittisurprise}, and many models have been proposed based on different assumptions. Generally speaking, there are two different processes that influence visual saliency; one is top-down and depends on the task at hand and the other is bottom-up, which is driven by the input image. The focus of the paper is bottom-up saliency for selecting attentional regions.
Many bottom-up computational models that simulate primate perceptual abilities have appeared in the literature. For example, in \cite{Itti_etal98pami, Itti_Koch01nrn, itti2000saliency} a center-surround mechanism is used to define saliency across scales, which is inspired by the putative neural mechanism. It has also been hypothesized that some visual inputs are intrinsically salient in certain background contexts and that these are actually task-independent \cite{Itti_etal98pami, Itti_Koch01nrn}. This model has established itself as the exemplar for saliency detection and consistently used for comparison in the literature. Similarly, there are also several proposed models which use other types of local information in different ways. In \cite{kadir2001saliency}, saliency is defined as the local complexity. Gao {\it et al.} \cite{gao2009discriminant,gao2007bottom,gao2008plausibility} proposed a bottom-up saliency model by using Kullback-Leibler (KL) divergence to measure the difference between a location and its surrounding area. In \cite{NIPS2005_81}, a model of overt attention with selection based on self-information is proposed, where the patches of an image are decomposed into a set of pre-learned bases and kernel density estimation is used to approximate the self-information.
Several models have been suggested to compute saliency using global information. In \cite{LCAV-CONF-2009-012}, the authors first transform the input color image into the $Lab$ color space and then define the saliency at each location as the difference between the $Lab$ pixel value and the mean $Lab$ value of the entire image. Harel {\it et al.} \cite{NIPS2006_897} proposed a graph-based solution that uses local computation to obtain a saliency map, which is everywhere dependent on global information. In \cite{avraham2009esaliency}, a saliency model called "extended saliency" was proposed, in which the "global exceptions" concept is used to replace the traditional preference for local contrast. Recently, a simple and fast algorithm, called the spectrum residual (SR), was proposed based on the Fourier Transform \cite{hou2007saliency}. The paper argues that the {\it spectrum residual} corresponds to image saliency. Following this, the Phase spectrum of the Fourier Transform (PFT) was introduced, which achieved nearly the same performance as the SR \cite{guo2008spatio}. Based on PFT, PQFT \cite{guo2008spatio} was also proposed by combining more features and using the quaternion Fourier Transform.
In this paper, we address the issue from three perspectives. Inspired by \cite{hou2007saliency}, we first consider saliency detection as a frequency domain problem. Unlike recent approaches which model saliency as a local phenomenon, we propose a new frequency domain para-digm, which permits the full use of global information. Second, instead of modeling saliency in an image, we define the concept of {\it non-saliency} using global information. Although in this paper we are solely concerned with determining saliency computationally, it is also interesting to consider the biological point of view. Research has suggested that objects viewed by the human visual system are thought to compete with each other to selectively focus our attention on a subset \cite{beck2005stimulus, yantis2005visual}. Objects that appear in the visual field will influence how they are viewed by suppressing each other. Consequently, many are inhibited, while those that are not, will ultimately predominate in the visual cortex to provide a focus of attention. In this paper, we model these inhibited regions as non-saliency. Compared with salient regions, which are very distinctive in the image, non-saliency can usually be modeled by common or uniform regions. These are then suppressed, thereby permitting salient objects to literally pop out. In this paper, non-saliency is modeled in the frequency domain. Third, we also address another issue, that of detecting salient regions of different sizes.
To date there is no consistent definition of saliency in the literature. The models of saliency are diverse. In several models, saliency detection mimics the fixation selection mechanism and tends to find small distinct regions or points, for example, SR \cite{hou2007saliency}, PFT \cite{guo2008spatio}, PQFT \cite{guo2010multiresolution} and AIM \cite{NIPS2005_81}. However, these may fail when detecting large saliency regions. Other papers tend to find large salient regions \cite{LCAV-CONF-2009-012, liu2010learning, goferman2010context, chengsaliency,yusaliencyregions, khan2009top}. Recently, scale-aware saliency \cite{Jacobson2010scale-aware} has been introduced to alleviate the problem of fixed scale in the spatial domain. We consider both small salient points as well as salient regions. For convenience, we will refer to both of these as salient regions, but of different size. We will show that the size of saliency regions is related to a scale parameter in the frequency domain.
We propose a new framework for saliency detection, which ostensibly, at first sight, seems to be similar to the {\it Convolution Theorem} but in fact we will show that it is not. We will demonstrate that the convolution of the amplitude spectrum with a Gaussian kernel of an appropriate scale is equivalent to a saliency detector. The proposed framework has the ability to both highlight small and large salient regions and to inhibit repeated distractors in cluttered images.
The contribution of this paper is threefold: 1) A frequency domain paradigm for saliency detection is proposed; 2) The detection of both small and large salient regions is treated as a whole in the proposed model; 3) We show that SR, PFT and the frequency-tuned model\cite{LCAV-CONF-2009-012} are, to some extent, special cases of the proposed model.
The paper is organized as follows: Section 2 is the description and review of related work. In section 3, we present the theoretical background of the proposed framework, called Spectrum Scale Space analysis (SSS). We present the saliency model (HFT) based on the Hypercomplex Fourier Transform in section 4. In section 5, we discuss experimental results. Concluding remarks and possible extensions are discussed in section 6.
\section{Related Work}
\label{relatedworks}
Recently, a simple and fast algorithm, called the {\it Spectrum Residual} (SR) was proposed in \cite{hou2007saliency}. This paper argues that the spectrum residual corresponds to image saliency. Thus given an image $f(x,y)$, it was first transformed into the frequency domain: $f(x,y)\xlongrightarrow{\mathcal F}{\mathcal F}(f)(u,v) $.
The amplitude ${\mathcal A}(u,v)=|{\mathcal F}(f)|$ and phase ${\mathcal P}(u,v) =angle({\mathcal F}(f))$ spectra are calculated, and then the log amplitude spectrum is obtained: ${\mathcal L}(u,v)=log({\mathcal A}(u,v))$. Given these definitions, the spectrum residual was defined as:
\begin{equation}
{\mathcal R}(u,v)={\mathcal L}(u,v)-h_{n}\star{\mathcal L}(u,v), \label {equ2}
\end {equation}
and the saliency map ${\mathcal S}(x,y)$ of the original image as:
\begin{equation}
{\mathcal S}(x,y)={\mathcal F}^{-1}[exp({\mathcal R}(u,v)+i\cdot{\mathcal P}(u,v))]. \label {equ3}
\end {equation}
In order to obtain a better visual display, the final saliency map was actually presented as\footnote{In this paper, $\vert\cdot\vert^2$ indicates computing the square of each element in the matrix.}:
\begin{equation}
{\mathcal S}(x,y)=g\star\vert{\mathcal F}^{-1}[exp({\mathcal R}(u,v)+i\cdot{\mathcal P}(u,v))]\vert^2, \label {equ4}
\end {equation}
where ${\mathcal F}$ and ${\mathcal F}^{-1}$ denote the Fourier and inverse Fourier Transforms, respectively; $h_{n}$ and $g$ are low-pass filters; $i$ is the imaginary function; ${\mathcal P}(u,v)$ denotes the phase spectrum of the image, {\it which is assumed to be preserved} when transforming back to the spatial domain. Equations (\ref{equ2}-\ref{equ4}) are from \cite{hou2007saliency}. The spectrum residual is the key idea of the SR, and the authors argued that it is this residual, combined with the original phase spectrum, that corresponds to the saliency in the image. However, in this paper:
1) We will show that the spectrum residual is of little significance;
2) For natural images, SR (or other similar models such as PFT \cite{guo2008spatio}) are, to some extent, equivalent to a gradient operator; and
3) Provide an explanation of why SR works in certain cases.
For convenience, we rewrite the standard {\it inverse Fourier Transform} as follows:
\begin{eqnarray}
&f(x,y)&={\mathcal F}^{-1}[exp(log {\mathcal A}(u,v)+i\cdot{\mathcal P}(u,v))], \label {equ5} \\
\Leftrightarrow&f(x,y)&={\mathcal F}^{-1}[{\mathcal A}(u,v)\cdot exp(i\cdot{\mathcal P}(u,v))], \label {equ6}\\
\Leftrightarrow&f(x,y)&={\mathcal F}^{-1}[{\mathcal F}(f)(u,v)]. \label {equ7}
\end{eqnarray}
Thus we can rewrite (\ref{equ3}) as follows:
\begin{equation}
{\mathcal S}(x,y)={\mathcal F}^{-1}[exp({\mathcal R}(u,v)\cdot exp(i\cdot{\mathcal P}(u,v))], \label {equ8}
\end {equation}
Defining $exp({\mathcal R}(u,v))$ as ${\mathcal A}_{SR}(u,v)$, (\ref{equ8}) is rewritten as:
\begin{equation}
{\mathcal S}(x,y)={\mathcal F}^{-1}[{\mathcal A}_{SR}(u,v)\cdot exp(i\cdot{\mathcal P}(u,v))]. \label {equ9}
\end {equation}
Comparing (\ref{equ6}) and (\ref{equ9}), we observe that if we replace the amplitude spectrum ${\mathcal A}(u,v)$ by the exponential of ${\mathcal R}(u,v)$, the saliency map is obtained\footnote {The phase spectra will no longer be plotted in the remaining figures in this paper, although, obviously they exist and are required for computing the transforms.}. (See the comparison in Fig. \ref{fig:sr}(a, b)). This is the key idea of SR.
\begin{figure}
\begin{center}
\includegraphics[width=.82\linewidth]{figures/456.pdf}
\end{center}
\caption{Spectrum residual given by SR contains little information corresponding to image saliency. (a) Obviously, the original image is reproduced by performing the inverse FT using the original amplitude and phase spectrum. (b) In SR, it is argued that saliency map can be obtained by replacing the $log({\mathcal A}(u,v))$ by the Spectrum Residual ${\mathcal R}(u,v)$. (c) If we replace the log amplitude spectrum $log{\mathcal A}(u,v)$ by random white noise(sic), we can obtain nearly the same saliency map.}
\label{fig:sr}
\end{figure}
In order to illustrate that the spectrum residual is of little significance, we generate a 2D white noise signal ${\mathcal W}(u,v)$, which has the same average value and maximum as the spectrum residual ${\mathcal R}(u,v)$. We then use ${\mathcal W}(u,v)$ to replace the spectrum residual and perform the inverse Fourier Transform as follows:
\begin{equation}
{\mathcal S}(x,y)={\mathcal F}^{-1}[exp({\mathcal W}(u,v)\cdot exp(i\cdot{\mathcal P}(u,v))]. \label {equ10}
\end{equation}
Fig.\ref{fig:sr}(c) shows this process. If we define $exp({\mathcal W}(u,v))$ as ${\mathcal A}_{W}(u,v)$, (\ref{equ10}) can be rewritten as follows:
\begin{equation}
{\mathcal S}(x,y)={\mathcal F}^{-1}[{\mathcal A}_{W}(u,v)\cdot exp(i\cdot{\mathcal P}(u,v))]. \label {equ11}
\end {equation}
Surprisingly, we can obtain nearly the same saliency map when we use white noise to replace the spectrum residual. This result clearly shows that the spectrum residual in \cite{hou2007saliency} contains little information corresponding to saliency. Why is this the case? Comparing (\ref{equ9}) and (\ref{equ11}), we find that the amplitude spectra used to perform the inverse Fourier Transform are $A_{SR}(u,v)$ and ${\mathcal A}_{W}(u,v)$. As shown in the third columns of Fig.{\ref{fig:sr}}(b, c), both ${\mathcal A}_{SR}(u,v)$ and ${\mathcal A}_{W}(u,v)$ are nearly horizontal planes compared (at the same scale) with ${\mathcal A}(u,v)$ shown in Fig.\ref{fig:sr}(a). That is to say, in both (\ref{equ9}) and (\ref{equ11}), the amplitude information is totally abandoned and only phase information plays a role.
Two questions arise:
(1) Why does SR yield a saliency map using only phase information?
(2) More important, is there any information corresponding to image saliency contained in the amplitude spectrum?
For the first question, our answer is that it only works for certain cases (detecting small salient regions in uncluttered scenes).
Also, consider \cite{guo2008spatio, guo2010multiresolution} where the authors propose a new saliency model called the Phase Fourier Transform (PFT). The saliency is computed using only phase information as follows:
\begin{eqnarray}
&{\mathcal S}(x,y)&={\mathcal F}^{-1}[exp(i\cdot{\mathcal P}(u,v))], \label {equ12} \\
\Leftrightarrow&{\mathcal S}(x,y)&={\mathcal F}^{-1}[1(u,v)\cdot exp(i\cdot{\mathcal P}(u,v))]. \label {equ14}
\end{eqnarray}
We observe that in PFT, the amplitude spectrum is (implicitly) also replaced by a horizontal plane. Therefore, we can deduce that, for natural images, both SR and PFT will produce nearly the same saliency map.
What does using the inverse Fourier Transform solely with phase information imply? We argue that for natural images, both SR and PFT are, to a certain degree, equivalent to a gradient operator combined with Gaussian post-processing (like the $g$ in (\ref{equ4})). This is because the amplitude spectrum of natural images always has higher values at low than at high frequencies \cite{ruderman1994statistics, srivastava2003advances}. Thus if the amplitude spectrum is replaced by a horizontal plane, all of the frequencies are being treated equally. That is to say, the lower frequencies are suppressed and the higher frequencies are enhanced. As is well-known, this implies a gradient enhancement operation. Based on the above discussion, we conclude that both SR and PFT will enhance the object boundaries and textured parts in an image. This indicates that they could work well only in detecting small salient regions where the center-surround contrast is very strong (see col. 1 and 3 of Fig. \ref{fig: SRvsGS}). However, they will have difficulty detecting large salient regions (col. 4) and those in a cluttered background (col. 5). To illustrate this point, we use a simple gradient operation combined with Gaussian post-filtering (as given by Algorithm \ref{alg:GM} below) and obtain nearly the same performance as the other two methods, as shown in Fig. \ref{fig: SRvsGS}.
Why is the performance of these models inadequate? The reason is that the information contained in the amplitude spectrum has been abandoned.
\begin{algorithm}[htb]
\caption{Procedure for computing the {\bf gradient and smoothing(G\&S)} }
\label{alg:GM}
\begin{algorithmic}[1]
\REQUIRE ~~\\
The resized image ${\mathcal I}$ with resolution $128\times128$
\ENSURE ~~\\
Saliency map ${\mathcal S}$ of ${\mathcal I}$ .
\vspace{0.15cm}
\STATE Convolve the input image with a Laplacian kernel, $ L$= [0 -1 0; -1 4 -1; 0 -1 0]. Obtain the gradient magnitude map $Gra$;
\STATE Convolve $Gra$ with a low-pass Gaussian filter kernel $g$, giving ${\mathcal S}=g{\star}|Gra|^2$;
\RETURN ${\mathcal S}$.
\end{algorithmic}
\end{algorithm}
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=7cm]{figures/24.pdf}
\end{overpic}
\caption{G\&S achieves nearly the same performance as SR and PFT.}
\label{fig: SRvsGS}
\end{center}
\end{figure}
In the next section, we will discuss the question regarding whether the amplitude spectrum contains any useful information about saliency. We will illustrate that the amplitude spectrum contains very important information and will develop a new framework for saliency detection in which we make full use of both the amplitude and phase.
\section{Convolution of the Amplitude Spectrum with a Low-Pass Gaussian Kernel Equals a Saliency Detector}
\label{SSS}
Many researchers have proposed models of saliency, which invariably then require the detection of salient {\it regions}. These regions are described as {\it distinctive }or {\it irregular} patterns, which possess a distinct feature distribution when compared with the rest of the image. In this paper, instead of searching for these irregular patterns, we model regular or so-called common patterns that do not attract much attention by our visual system. We refer to these patterns as being {\it non-salient}.
\subsection{Suppressing Repeated Patterns for Saliency Pop-Out}
In the proposed model, we assume that a natural image consists of several salient and many so-called regular regions. All of these entities (whether distinct or not) may be considered as visual stimuli that compete for attention in the visual cortex. In this regard, it has been shown that nearby neurons constituting receptive fields in the visual cortex mutually inhibit each other and interact competitively \cite{duncan1989visual}. As an example, in Fig. \ref{fig:repeatPatt}, if we divide the image into many patches (at a particular scale), we find that, some are distinctive, while others are quite similar to each other. The bottom part of Fig. \ref{fig:repeatPatt} shows the collection of patches from the last natural image above. We observe that several patterns appear many times (e.g., blue sky and grassy patches). We refer to these regular patches as {\it repeated patterns}, which correspond to {\it non-saliency}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.7\linewidth]{figures/26.pdf}
\end{center}
\caption{Regular(repeated) and anomalous patterns. Top: Four images; bottom: Collection of fragments from the last image above.}
\label{fig:repeatPatt}
\end{figure}
Clearly, the primate visual system is more sensitive to distinctive rather than repeated patterns in an image. Furthermore, the latter are very diverse. For example, consider the top row of Fig. \ref{fig:repeatPatt}. These exhibit several different examples of repeated patterns at different scales (including at the "scale" of 0 frequency for the uniform areas): grassy and sky patches (image 4), similar objects (image 1), road patches of the same color and texture (image 2), the 'L's (image 3), and so on. We model these repeated patterns and then suppress them, thereby producing the pop-out of the salient objects.
\subsection{Spikes in the Amplitude Spectrum Correspond to Repeated Patterns}
In this paper, we will illustrate that the amplitude spectrum contains important information corresponding to saliency and non-saliency. To be more precise, the spikes in the amplitude spectrum turn out to correspond to repeated patterns, which should be suppressed for saliency detection.
For convenience, we take a 1-D periodic signal $f(t)$ as an example. Suppose $f(t)$ can be represented by $f(t)=\sum_{n=-\infty}^{\infty} F(n)e^{jn\omega_1t}$, where $F_n=\frac{1}{T}\int_{-T/2}^{T/2}f(t)e^{-jn\omega_1t}dt$. Then the Fourier transform is given by:
\begin{equation}
{\mathcal F}(w)=2\pi\sum_{n=-\infty}^{\infty} F(n)\delta(\omega-n\omega_1). \label {equ13}
\end {equation}
From (\ref{equ13}), we can conclude that the spectrum of a periodic signal (repeated cycles) is a set of impulse functions (spikes). We note that this is based on the assumption that the signal is infinite. Therefore, given a more realistic finite length periodic signal, the shape of the spectrum will obviously be different but not degraded greatly.
Fig.\ref{fig:Spikes} provides an illustration of this point. Fig.\ref{fig:Spikes}(a) shows three signals with a different number of repeated patterns (cycles) while Fig.\ref{fig:Spikes}(b) shows their corresponding amplitude spectra. We observe that the larger the number of repeated cycles, the sharper the spikes in the spectrum. In order to quantify this notion, we define the {\it sharpness} of a spectrum $X$. We note that if we smooth the spikes by convolving the spectrum with a low-pass filter, the sharper the original spike, the more its peak height will be reduced. Therefore, the {\it sharpness} of $X$ can be defined as
${\mathcal \gamma}(X)= \parallel X-X{\star}h_{m}\parallel_\infty,$
where $h_{m}$ is a Gaussian kernel with fixed scale. The sharpness values of these three spectra in Fig. \ref{fig:Spikes} are 0.2320, 0.6091 and 1.3227 respectively.
Besides the sinusoid shown in the figure, other repeated signals also have this characteristic.
Next suppose there is one salient part that is embedded in a finite length periodic signal (row 1 of Fig. \ref{fig: ssf}). We will illustrate that this salient interval will not largely influence the spikes in the spectrum. That is to say, 1) The spikes will remain even though a salient part is embedded in the periodic signal; 2) The embedded salient part will not lead to very sharp spikes in the spectrum. The signal to be analyzed is defined as follows:
\begin{equation}
f(t)=g(t)+g_\tau(t)+s(t), \label {equ15}
\end {equation}
where $g(t)$ is a periodic signal with finite length $L$, equaling $p(t)$ inside the interval $(0, L)$ and 0 elsewhere;
$g_\tau(t)=-p(t)\cdot {r}(t)$;
$s(t)=p_s(t)\cdot {r}(t)$,
$s(t)$ is the salient part of $f(t)$, which for convenience is also defined as a portion of yet another periodic function $p_s(t) $; $p(t)$ and $p_s(t) $ are periodic functions with frequencies $\nu$ and $\nu_s$, respectively; ${r}(t)$ is a rectangular window function that equals 1 inside the interval $(t_0,t_0+\tau)$ and 0 elsewhere; we also suppose that $(t_0,t_0+\tau)\in (0,L)$ and $\tau \ll L$ (see row 1 of Fig. \ref{fig: ssf}). Thus the Fourier Transform of $f(t)$ can be represented as follows:
\begin{multline}
{\mathcal F}(f)(\omega)=\int_{-\infty}^{\infty}f(t)e^{-j\omega t}dt=
\int_{0}^{L}g(t)e^{-j\omega t}dt\\
+\int_{ t_0}^{ t_0+\tau}g_\tau(t)e^{-j\omega t}dt+\int_{ t_0}^{ t_0+\tau}s(t)e^{-j\omega t}dt\label{F(f)}.
\end{multline}
From (\ref{F(f)}), the spectrum of $f(t)$ consists of three terms. We assume that $\tau \ll L$. This implies that the first term has very sharp spikes in the amplitude spectrum as it contains many repeated patterns, while this is not true of the second and third terms. Consider $g_\tau(t)$ as an example. $g_\tau(t)$ is the point-wise product of signal $-p(t)$ and ${r}(t)$. According to the convolution theorem, ${\mathcal F}(g_\tau)(\omega)$ equals the convolution of $-{\mathcal F}(p)(\omega)$ with ${\mathcal F}(r)(\omega)$. Since ${\mathcal F}({r})(\omega)=\frac{2sin (\tau/2)}{\omega}e^{j\omega(t_0+\tau/2)}$ is a low-pass filter, the spikes in the amplitude spectrum of $-{\mathcal F}(p)(\omega)$ will be greatly suppressed. That is to say, there are no sharp spikes in the second term. This also occurs for the third term.
As discussed above, the sharpness of ${\mathcal F}(f)(\omega)$ is mainly determined by $g(t)$, while the latter two terms in (\ref{F(f)}) do not make a significant contribution to the spikes in the spectrum. In other words, since the first term corresponds to repeated patterns (non-salient) which lead to spikes, they can be suppressed by smoothing the spikes in the amplitude spectrum of ${\mathcal F}(f)(\omega)$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.9\linewidth]{figures/11.pdf}
\end{center}
\caption{Repeated patterns lead to sharp spikes: (a): Signals with different number of repeated cycles; (b): corresponding amplitude spectra.}
\label{fig:Spikes}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=7cm]{figures/12.pdf}
\put(-2,51) {{\small 1}}
\put(-2,43) {{\small 2}}
\put(-2,35) {{\small 3}}
\put(-2,27) {{\small 4}}
\put(-2,19) {{\small 5}}
\put(-2,12) {{\small 6}}
\put(-2,4) {{\small 7}}
\end{overpic}
\caption{Suppression of repeated patterns by using spectrum filtering. It is clear that the larger the repeated background, the sharper the spikes, leading to the suppression of the amplitude spectrum via filtering.}
\label{fig: ssf}
\end{center}
\end{figure}
\subsection{Suppressing Repeated Patterns Using Spectral Filtering}
A Gaussian kernel $h$ can be employed to suppress spikes in the amplitude spectrum $|{\mathcal F}\{f\}|$ of an image as follows\footnote {In the implementation of this equation, we found that suppressing spikes in the log amplitude spectrum rather than the amplitude spectrum yielded better results.}:
\begin{equation}
\mathcal{A_{S}}(u,v)=|{\mathcal F}\{f(x,y)\}|\star h.
\label{Eq:AmpSmth1}
\end{equation}
The resulting smoothed amplitude spectrum ${\mathcal{A}_{S}}$ and the {\it original} phase spectrum are combined to compute the inverse transform, which in turn, yields the saliency map:
\begin{equation}
{\mathcal S}={\mathcal F^{-1}}\{\mathcal{A_{S}}{(u,v)}e^{i\cdot\mathcal{P}{(u,v)}}\}.
\label{Eq:AmpSmth2}
\end{equation}
In order to improve the visual display of saliency, we define it hereafter as:
\begin{equation}
{\mathcal S}=g\star|{\mathcal F^{-1}}\{\mathcal{A_{S}}{(u,v)}e^{i\cdot\mathcal{P}{(u,v)}}\}|^{2}.
\label{Eq:AmpSmth3}
\end{equation}
Consider the very simple example shown in Fig.\ref{fig: ssf}. The input signal (row 1) is periodic, but there is a short segment for which a different frequency signal is apparent. The short segment is quite distinct from the background for human vision, so a saliency detector should be able to highlight it. Row 2 shows the amplitude spectrum: there are three very sharp spikes (labeled by solid boxes), one of which corresponds to the constant background (uniform part) at zero frequency and the other two correspond to the periodic background. In addition, there are two rounded maxima (labeled by dashed boxes) corresponding to the salient parts. The amplitude spectrum is then smoothed by a Gaussian kernel (row 3), and the signal is reconstructed using the smoothed amplitude and original phase spectrum (row 4). It is clear that both the periodic and the uniform background are largely suppressed while the salient segment is well preserved. Row 5 shows the saliency map after enhancing the signal shown in row 4 using post-processing. We can further analyze this in the frequency domain, as shown in row 6, which illustrates the components actually removed by the previous operations. Here the eliminated frequency components are mainly the low frequencies near zero frequency, as well as the periodic background. Row 7 presents these removed components in the spatial domain. We find that non-salient parts (including uniform parts) are well suppressed using amplitude filtering. This process suggests that convolution in the frequency domain of the amplitude spectrum with a Gaussian kernel is equivalent to an image saliency detector\footnote{One might mistakenly be confused to think that this convolution in the frequency domain is equivalent to multiplication in the spatial domain as in the convolution theory. Yet, this is not the case as we convolve only the amplitude and do not change the phase.}.
\subsection{Spectrum Scale-Space Analysis}
\label{sec:SSSA}
Repeated patterns (including uniform patterns) can be suppressed by smoothing the amplitude spectrum at an appropriate scale. However, which scale is the best in (\ref{Eq:AmpSmth1})? As shown in Fig. \ref{fig:Scale}, if the filter scale is too small, the repeated patterns cannot be suppressed sufficiently (row 2), while if the filter scale is too large, only the boundaries of the salient region are highlighted (row 4 and 5). Therefore it is important to select a proper scale for the Gaussian kernel. In fact, we will illustrate that different filter scales are required for different types of saliency. For example, a small-scale kernel is needed to detect large salient regions, while a large-scale kernel could be used to detect texture-rich or small salient regions (e.g. distant objects in the scene).
In this paper, we propose a Spectrum Scale-Space (SSS) for handling amplitude spectra at different scales, yielding a one-parameter family of smoothed spectra which is parameterized by the scale of the Gaussian kernel. Given an amplitude spectrum, $\mathcal{A}(u,v)$, of an image, the SSS is a family of derived signals $\Lambda(u,v;k)$ defined by the convolution of $\mathcal{A}$ with the series of Gaussian kernels:
\begin{equation}
g(u,v;k)=\frac{1}{\sqrt{2\pi}2^{k-1}t_0}e^{-(u^2+v^2)/(2^{2k-1}t_0^2)},
\label{Eq:GassianKernel}
\end{equation}
where $k$ is the scale parameter, $k=1,...,K$. $K$ is determined by the image size: $K=\lceil log_2{min\{H,W\}}\rceil+1$, where $H$, $W$ indicate the height and width of the image; $t_0=0.5$. Thus the scale space is defined as follows:
\begin{equation}
\Lambda(u,v;k)=(g(.,.;k)\star\mathcal{A})(u,v).
\label{Eq:SSS}
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[width=.9\linewidth]{figures/Figure_3.pdf}
\end{center}
\caption{The original 1D signal is shown in the first row of col. 1 with the 1D saliency maps below. The spectrum of the original signal is shown in the first row of col. 2, followed by the smoothed spectra associated with the paired saliency map in col. 1.}
\label{fig:Scale}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=7.5cm]{figures/f8.pdf}
\end{overpic}
\caption{Five 2-D examples are shown. The first column shows the original 2-D signals (images). The remaining images in each row present the set of saliency maps computed by smoothing the original image amplitude spectrum using different scales for the Gaussian kernels.}
\label{fig:Scale2}
\end{center}
\end{figure}
As an example, assume a 1-D signal. We first compute a series of filtered spectra according to the SSS model; then the saliency map is computed for each scale, as shown in Fig. \ref{fig:Scale}. The significance of the scale for saliency detection can easily be observed. In this example, smoothed spectrum 2 gives the best result. As the kernel scale goes to infinity, the spectrum tends to be a constant (horizontal plane in 2D), as shown in the last row of Fig. \ref{fig:Scale}. This is exactly the case proposed in \cite{hou2007saliency, guo2008spatio, guo2010multiresolution}.
Fig. \ref{fig:Scale2} shows 2D results obtained using different kernel scales, increasing from left to right. The best saliency map is labeled by a red square. We observe that broad regions pop out when smaller scale kernels are used, while distant objects or those with rich texture pop out when larger scale kernels are used. Thus given a natural image, a set of saliency maps is obtained from which one must be selected as the final saliency map. The criterion for achieving this will be discussed in section \ref{sM:HFT}.
Here, we suggest that the frequency-tuned model \cite{LCAV-CONF-2009-012} is, to some extent, a special case of the proposed model. In \cite{LCAV-CONF-2009-012}, the saliency map is defined as:
${\mathcal{S}}(x,y)=\Vert I_{\mu}-I_{\omega hc}(x,y)\Vert$,
where $I_{\mu}$ is the average $Lab$ vector of the entire image and $I_{\omega hc}(x,y)$ is a specific $Lab$ pixel vector from the Gaussian-filtered version of the original image. Authors compute a saliency map by removing the frequencies around the DC frequency (the {"}mean{"}). Previously, we have illustrated that there is always a very sharp spike around zero frequency, which corresponds to this {"}mean{"}. Hence, if we use a very small scale Gaussian kernel to smooth the spectrum, those components corresponding to the {"}mean{"} will be suppressed significantly.
\section{Saliency Using the Hypercomplex Fourier Transform (HFT)}
\label{sM:HFT}
In section \ref{SSS}, we discussed the saliency computation using only one feature map (that of intensity). However, in order to obtain better performance, more features are required, for example, color and motion information. Inspired by \cite{guo2008spatio, ell2002quaternion, ell2006hypercomplex}, we use the so-called hypercomplex matrix to combine multiple feature maps. Consequently, the {\it Hypercomplex Fourier Transform} (HFT) is employed to replace the {\it Fourier Transform} used in section \ref{SSS} for saliency computing.
\subsection{Hypercomplex Fourier Transform}
The input to the traditional Discrete Fourier Transform is a real matrix. Each image pixel is an element of the input matrix and is a real number. However, if we combine more than one feature into a hypercomplex matrix, each element is a vector and this hypercomplex matrix is a {\it vector field}. Thus, the traditional Fourier Transform becomes unsuitable for computational purposes.
The Hypercomplex Fourier Transform was proposed in \cite{ell2002quaternion}, in which the hypercomplex input was specified to be a quaternion\footnote{The quaternion is represented as $q=a+bi+cj+dk$, where a,b,c and d are real numbers and i,j,k satisfy $ i^{2}=j^{2}=k^{2}=ijk=-1$. A quaternion can also be represented as $q=S(q)+V(q)$, where $S(q)=a$ is the scalar and $V(q)=bi+cj+dk$ is the vector part. q is called a pure quaternion if $S(q)=0$.}. Given a hypercomplex matrix:
\begin{equation}
f(n,m)=a+bi+cj+dk,
\label{Eq:qft1}
\end{equation}
the discrete version of the HFT of (\ref{Eq:qft1}) is given by:
\begin{equation}
\mathcal {F_{H}}[u,v]=\frac{1}{\sqrt{MN}}\sum^{M-1}_{m=0}{\sum^{N-1}_{n=0}e^{-\mu 2\pi((\frac{mv}{M})+(\frac{nu}{N}))}}f(n,m),
\label{Eq:qft2}
\end{equation}
where $\mu$ is a unit pure quaternion and $\mu^{2}$ =-1. Note that $\mathcal {F_{H}}[u,v]$ is also a hypercomplex matrix. The inverse Hypercomplex Fourier Transform is given as:
\begin{equation}
f(n,m)=\frac{1}{\sqrt{MN}}\sum^{M-1}_{v=0}{\sum^{N-1}_{u=0}e^{\mu 2\pi((\frac{mv}{M})+(\frac{nu}{N}))}}\mathcal {F_{H}}[u,v].
\label{Eq:qft3}
\end{equation}
\subsection{Hypercomplex Representation of Multiple Feature Maps}
The Hypercomplex representation can be employed to combine multiple features (e.g., in \cite{guo2010multiresolution} the authors combine color, intensity and motion as the features). We define the input hypercomplex matrix as follows:
\begin{equation}
f(n,m)=w_{1}f_{1}+w_{2}f_{2}i+w_{3}f_{3}j+w_{4}f_{4}k,
\label{Eq:qft5}
\end{equation}
where ${w_{1}\text{-}w_{4}}$ are weights and ${f_{1}\text{-}f_{4}}$ are the feature maps (matrices). Similar to \cite{Itti_etal98pami}, we use three features to compute the saliency for the static input case:
\begin{equation}
f_{2}=\mathcal {I}_{s}=(r+g+b)/3,
\label{Eq:fm1}
\end{equation}
\begin{equation}
\label{Eq:fm2}
f_{3}=\mathcal {RG}=R-G,
\end{equation}
\begin{equation}
\label{Eq:fm3}
f_{4}=\mathcal {BY}=B-Y,
\end{equation}
where $r$, $g$, $b$ are the red, green and blue channels of an input color image and $R=r-(g+b)/2$, $G=g-(r+b)/2$, $B=b-(r+g)/2$, $Y=(r+g)/2-|r-g|/2-b$.
These three feature maps comprise the opponent color space representation of the input image (see part 1 of Fig. \ref{fig:pHFT}). Based on the work in \cite{guo2010multiresolution}, our approach has also been experimentally confirmed using videos by defining a motion feature $\mathcal M$ and setting $f_{1}=\mathcal M$ in (\ref{Eq:qft5}). In this paper, we consider only the static image case by employing just intensity and color information. We select the weights so that $w_{1}=0$, $w_{2}=0.5$, $w_{3}=w_{4}=0.25$.
\subsection{Computing the Saliency Map}
Given an image, the input is defined according to section 4.2. The Hypercomplex Fourier Transform, $\mathcal {F_{H}}[u,v]$, can be rewritten in polar form as follows:
\begin{equation}
\mathcal {F_{H}}[u,v]=\Vert\mathcal {F_{H}}[u,v]\Vert e^{\mu \Phi (u,v)},
\label{Eq:qft6}
\end{equation}
where $\Vert\cdot \Vert$ indicates the modulus for each element of a hypercomplex matrix;
$\mathcal {F_{H}}[u,v]$ can be considered as the frequency domain representation of $f(m,n)$. Its amplitude spectrum $\mathcal{A}(u,v)$, phase spectrum $\mathcal{P}(u,v)$ and the so-called eigenaxis spectrum $\mathcal{X}(u,v)$ are defined as:
$$\mathcal{A}(u,v)=\Vert\mathcal{F_{H}}(u,v)\Vert,$$
$$\mathcal{P}(u,v)={\Phi}{(u,v)}=tan^{-1}\frac{\Vert\mathcal{V}{(\mathcal{F}(u,v))}\Vert}{\mathcal{S}{(\mathcal{F}(u,v))}},$$
$$\mathcal{X}(u,v)=\mu (u,v)=\frac{\mathcal{V}({\mathcal{F}(u,v)})}{\Vert \mathcal{V}({\mathcal{F}(u,v)})\Vert},$$
where, $\mathcal{X}(u,v)$ is a pure quaternion matrix. These three spectra are shown in part 2 of Fig. \ref{fig:pHFT} \footnote{Here we use a monochrome image to represent the phase spectrum $\mathcal{P}(u,v)$ as it is a real matrix. This is different from \cite{ell2006hypercomplex}.}.
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{figures/27.pdf}
\end{center}
\caption{Procedure for computing saliency using the Hypercomplex Fourier Transform (HFT)}
\label{fig:pHFT}
\end{figure}
As discussed in Sec. \ref{SSS}, the amplitude spectrum contains important information about the scene. Similar to the discussion in sec. \ref{sec:SSSA}, we create the {\it Spectrum Scale Space} $\Lambda=\{\Lambda_{k}\}$ by smoothing $\mathcal{A}(u,v)$ with a series of Gaussian kernels according to (\ref{Eq:SSS}) (see Fig.\ref{fig:pHFT} (3)) while retaining unchanged the phase spectrum $\mathcal{P}(u,v)$ and eigenaxis spectrum $\mathcal{X}(u,v)$.
Observing the images in part 3 of Fig. \ref{fig:pHFT} reveals that when the scale $k$ is very small, the information contained in the amplitude plots is retained quite well, while when it becomes very large, the pertinent information basically is lost. Actually, the PQFT model is a special case of the proposed framework when the scale goes to infinity.
Thus, given a single (smoothed) amplitude spectrum $\Lambda_{k}$ (one layer in $\Lambda$) and the original phase and eigenaxis spectra, we can perform the inverse transform (\ref{Eq:qft3}) to give the saliency map at each scale:
\begin{equation}
{\mathcal{S}_{k}}=g\star\Vert\mathcal{F}_{\mathcal{H}}^{-1}\{\Lambda_{k}(u,v)e^{\mathcal{X} \mathcal{P} (u,v)}\}\Vert^{2},
\label{Eq:qft7}
\end{equation}
where $g$ is a Gaussian kernel at a fixed scale\footnote{For convenience, the scale parameter has been set to $0.05\cdot W$. Though this has been done to improve the visual display, it will nevertheless influence the ROC score when predicting human fixation \cite{ImageSignature2012}. We discuss this issue in detail in section \ref{Sec:Experiment}.}. Thus, we again obtain a series of saliency maps $\{\mathcal{S}_{k}\}$, shown in part 4 of Fig.\ref{fig:pHFT}. In the approach proposed in this paper, the final saliency map $\mathcal{S}$ is chosen from $\{\mathcal{S}_{k}\}$ by selecting the best scale $k_{p}$ according to criteria discussed in section \ref{findingscale}. The saliency model based on the Hypercomplex Fourier Transform is referred to as HFT in this paper.
The HFT model is summarized in Algorithm \ref{alg:HFT} \footnote{The input image is resized to $128 \times 128$ in the experiments.}.
\begin{algorithm}[htb]
\caption{{\bf HFT }saliency model}
\label{alg:HFT}
\begin{algorithmic}[1]
\REQUIRE ~~\\
The resized color image ${\mathcal{C}}$ with resolution $m\times n$
\ENSURE ~~\\
Saliency map ${\mathcal{S}}$ of ${\mathcal{C}}$
\vspace{0.15cm}
\STATE Compute the feature maps $\{\mathcal{I}$, $\mathcal{RG}$, $\mathcal{BY}\}$ of ${\mathcal{C}}$ according to (\ref{Eq:fm1})-(\ref{Eq:fm3});
\STATE Form the hypercomplex matrix $f(n,m)$ by combining these feature maps according to (\ref{Eq:qft5});
\STATE Perform the Hypercomplex Fourier Transform on $f(n,m)$ and compute the amplitude spectrum $\mathcal{A}$, phase spectrum, $\mathcal{P}$ and eigenaxis spectrum $\mathcal{X}$;
\STATE Smooth the amplitude spectrum with Gaussian kernels according to (\ref{Eq:GassianKernel}), thereby obtaining a spectrum scale space $\{\Lambda_{k}\}$;
\STATE Obtain a saliency map ${\mathcal{S}}_{k}$ according to (\ref{Eq:qft7}) for each $\Lambda_{k}$, thereby producing a sequence of saliency maps $\{{\mathcal{S}}_{k}\}$;
\STATE Find the best saliency map ${\mathcal{S}}$ from the set $\{{\mathcal{S}}_{k}\}$ and use it as the final saliency map according to the criterion introduced in (\ref{Eq:pscal5}) in section \ref{findingscale};
\RETURN ${\mathcal{S}}$.
\end{algorithmic}
\end{algorithm}
\subsection{Finding the Proper Scale}
\label{findingscale}
In section \ref{SSS}, we assumed that the best saliency map would appear at a specific scale in the sequence $\{{\mathcal{S}}_{k}\}$. Unlike the use of {\it entropy} in \cite{kadir2001saliency}, we employ it as the criterion for determining the optimal scale:
\begin{equation}
k_{p}=\underset{k}{\arg\min}\{\mathcal{H}\left(\mathcal{S}_k\right)\},
\label{Eq:pscal0}
\end{equation}
where $\mathcal{H}(x)=-\sum_{i=1}^{n}p_{i}\mathrm{log}p_{i}$ is the definition of entropy of $x$. The reason for using entropy is as follows. The saliency map can be considered as a probability map. In a desirable saliency map, the regions of interest would be assigned higher values and the rest of the map would be largely suppressed. Thus it is expected that the values in the saliency map histogram would cluster around certain values. The entropy of the saliency map would then be very small according to the definition of entropy.
Conventional entropy is based on the distribution of a variable $x$; if the histogram is given, the entropy of $x$ is determined. However, the spatial geometric information is ignored. As shown in Fig. \ref{fig:2dentropy}, images may possess the same histograms and therefore have the same entropy values, even though the spatial structure becomes more and more chaotic. Obviously, in saliency detection, we wish to avoid selecting a map with a high level of chaos.
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=7.3cm]{figures/29.pdf}
\put(5,12) {{\small 1}}
\put(19.8,12) {{\small 2}}
\put(34.6,12) {{\small 3}}
\put(49,12) {{\small 4}}
\put(63.3,12) {{\small 5}}
\put(78.2,12) {{\small 6}}
\put(92.4,12) {{\small 7}}
\end{overpic}
\caption{Binary images with the same histogram, but with different spatial structures}
\label{fig:2dentropy}
\end{center}
\end{figure}
Spatial geometric information needs to be considered in 2D signal analysis, and work related to this issue has been reported, such as the so-called 2D Entropy \cite{abutaleb1989automatic, chen1994fast}. Here we present a simple improved definition of entropy in order to make use of the spatial geometric information. We consider each pixel individually and require it to also depend on the values of its neighbors. We achieve this objective by employing a Gaussian kernel to filter the 2D signal, and then compute the conventional entropy on the smoothed 2D signal. Consequently, the new entropy is defined as:
$\mathcal{H}_{2D}(x)=\mathcal{H}\{g_{n}\star x\}$
where $g_{n}$ is a low-pass Gaussian kernel with a scale of $\varsigma$.
As shown in Fig. \ref{fig:2dentropy3}, if $\varsigma$ were too small, especially when $\varsigma$=0, Gaussian filtering would have a minor effect. If $\varsigma$ =1.2, the entropy value would increase as the image became more and more chaotic in the spatial domain. This is quite reasonable. However, if $\varsigma$ were too large, the entropy value would decrease. This is because the small structures in the 2D signal would be heavily destroyed by the Gaussian filter. Thus, on the one hand, we desire that $\varsigma$ should be as large as possible, because with larger $\varsigma$ the influence of a pixel could spread farther. On the other hand, we do not wish to destroy the small spatial structures. Therefore, $\varsigma$ should be related to the size of smallest region we expect to detect. Experiments indicate that $\varsigma=0.01\thicksim0.03\cdot{W}$ yields acceptable results.
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=7cm]{figures/31.pdf}
\end{overpic}
\caption{Computing 2D entropy of the 2D images in Fig. \ref{fig:2dentropy} using Gaussian kernels of different size.}
\label{fig:2dentropy3}
\end{center}
\end{figure}
Besides entropy, there is another issue to consider when choosing the proper scale $k$. In HFT, given $\{{\mathcal{S}}_{k}\}$, we avoid choosing saliency maps with a strong response at the border region by using a border avoidance strategy. Thus, a parameter $\lambda$ is defined for each candidate saliency map:
$\lambda_{k}=\sum\sum{\mathcal{K}{}(n,m)\cdot\mathcal{N}(\mathcal{S}_{k}(n,m))},$
where $\mathcal{K}{}$ is a 2D centered Gaussian mask of the same size as ${\mathcal{S}}$, $\sigma_{w}=W/4$, $\sigma_{h}=H/4$, and $\sum\sum\mathcal{K}{}(n,m)=1$. $\mathcal{N}(\cdot)$ is used to normalize ${\mathcal{S}}$, so that the summation of all the pixel values is 1. Note that $\lambda$ is not the same as the {\it center-bias} or {\it border cut} described in \cite{zhang2008sun}, since it is used only to {\it choose} a proper scale, but not to {\it modify} the saliency map as done in \cite{NIPS2006_897,judd2012pami}. {\it Center-bias} and {\it border cut} will be discussed in section \ref{Sec:Experiment}.
Thus, with this definition of 2D entropy and $\lambda$, $k_{p}$ in (\ref{Eq:pscal0}) is revised as follows:
\begin{equation}
k_{p}=\underset{k}{\arg\min}\{\lambda_{k}^{-1}\mathcal{H}_{2D}\left(\mathcal{S}_k\right)\}.
\label{Eq:pscal5}
\end{equation}
The model that uses the criterion in the above equation is HFT. In addition, we use entropy without the border-avoidance strategy as the criterion, and the results are labeled as HFT(e). Of course, it is possible that the performance of the proposed model might be improved with an even better criterion. For example, as shown in the row 2 of Fig.\ref{fig:Exp71}(F), HFT failed to highlight the two salient objects uniformly. However, we note that this was caused by the selection of an improper scale, notwithstanding the fact that the optimal scale did was present in the existing set, as shown in the third row of Fig.\ref{fig:Scale2}. In order to illustrate the potential power of the proposed model, we have also determined the optimal scale for each image visually by examining the ROC scores of the saliency maps. The results are reported in this paper and are labeled as HFT*.
\section{Experimental Results}
\label{Sec:Experiment}
Three experiments to evaluate the performance of the proposed HFT are discussed in this section: 1) Response to psychological patterns, 2) Predicting human fixations and 3) Predicting the object regions to which humans pay attention. Eight state-of-the-art methods were employed to perform the comparisons: Itti's model\cite{Itti_etal98pami}\footnote{This implementation comes in the GBVS package, see \url{http://www.klab.caltech.edu/~harel/share/gbvs.php}. The saliency toolbox STB (based on Itti's model) was used in the first experiment.}, DVA\cite{NIPS2008_0142}\footnote{The code is available at \url{http://www.its.caltech.edu/~xhou/}.}, GBVS\cite{NIPS2006_897}\footnote{ See \url{http://www.klab.caltech.edu/~harel/share/gbvs.php}.}, SR\cite{hou2007saliency}\footnote{The code is available at \url{http://www.its.caltech.edu/~xhou/}.}, PFT\cite{guo2008spatio}\footnote{PFT (Phase Fourier Transform) is an improved version of SR. Our implementation was done according to \cite{guo2008spatio}.}, PQFT\cite{guo2008spatio,guo2010multiresolution}\footnote{The code was provided by the first author in \cite{guo2010multiresolution}.}, AIM\cite{NIPS2005_81}\footnote{
See \url{http://www-sop.inria.fr/members/Neil.Bruce/AIM.zip}.} and SUN\cite{zhang2008sun}\footnote{The code is available at \url{http://cseweb.ucsd.edu/~l6zhang/}.}.
We evaluate the performance of saliency detection algorithms both qualitatively and by comparison to human observers. For the former, we essentially compare the saliency map to the original image by using a simple algorithm to determine an object map based on the saliency map. For the latter we require ground truth data. We use two kinds of ground truth in this paper, fixation data and salient regions labeled by human observers.
In section \ref{exp_fixation}, we have used freely available human fixation data \cite{NIPS2005_81} as ground truth to evaluate the algorithms listed earlier. ROC score (area under the ROC curve, AUC) is adopted to measure their performance.
In section \ref{exp_region}, we have also evaluated the algorithms using object regions labeled by humans (some examples of "labeled" results are shown in each second column of Fig.\ref{fig:Exp71}) as ground truth. In fact, the {\it available} eye tracking data only contain {\it positional} information \cite{hou2007saliency}. However, saliency detection algorithms in computer vision are assumed and expected to have the ability to detect salient object {\it regions} in a scene \cite{elazary2008salient}. For example, given a region such as a flower (see row 4 of Fig.\ref{fig:Exp71}(A) as an example), an algorithm should respond more or less uniformly within the whole region and not just along the boundary of the flower or at several points on the flower. Therefore, we use {\it salient region maps} labeled by humans as ground truth.
In this experiment, besides ROC, we also use the DSC (Dice Similarity Coefficient) as a measure to evaluate the overlap between the thresholded saliency map and the ground truth. The peak value of the DSC curve (PoDSC) is an important index of performance, as it corresponds to the optimal threshold and the best possible algorithm performance\cite{thomasevaluation2008}.
\subsection{How to Make Fair Quantitative Comparisons?}
There are two aspects which should be considered when making quantitative comparisons between two saliency models: scale and post-processing.
Certain models permit the usage of different image scales (input image size). Therefore, in these cases, we find the optimal scale by maximizing their performance, but for the other models it is necessary to use the default settings, as shown in Table \ref{alg_list}.
With regard to post-processing, most previous work has used the ROC directly without investigating any of the post-processing factors affecting the fairness of this approach. However, it is important to note that three factors dramatically influence the ROC score and PoDSC: {\bf 1)} {\bf Border cut (BC)} \cite{zhang2008sun}, {\bf 2)} {\bf Centre-bias setting (CB)} \cite{NIPS2006_897} and {\bf 3)} {\bf smoothing (SM)} \cite{ImageSignature2012,NIPS2006_897}.
In this paper, in order to make a fair comparison, the post-processing is calibrated. We first consider BC and CB by dividing the saliency models into three {\it subsets}: 1) models without any BC and CB; 2) models with BC and 3) models with CB, as shown, in Table \ref{alg_list}. In addition, the optimal smoothing parameter for each model is learnt in order to eliminate the influence of SM. We compare HFT class models (HFT, HFT(e) and HFT*) with each of the the three subsets.
1) When comparing HFT class models with models in subset 1, we compute the ROC and/or PoDSC directly;
2) When comparing HFT class models with subset 2, we set the border cuts for all of these models to be of equal size\footnote{In our experiments, we considered only the interior of the frame and the corresponding region in the ground truth when computing the ROC curve.}.
3) When comparing HFT class models with models in subset 3, we apply an {\it optimal} center-bias for each model individually, thereby ensuring that the ROC score for each model is maximized.
\begin{table}[htbp]
\caption{Three subsets of algorithms employed for comparison}
\label{alg_list}
\centering
\begin{tabular}{@{}clcccc@{}}
\toprule
&\multicolumn{1}{c}{}
&\multicolumn{1}{c}{}
&\multicolumn{3}{c}{post-processing effects}\\
\cmidrule(lr){4-6}
\scriptsize{} {Subset} & {Model} & {Image size}
& {SM} & {BC} & {CB}\\
\midrule
1 &HFT &$128\times128^{\ddag}$ & explicit & no & no\\
1 &SR/PFT &$64\times64^{\dag}$ & explicit & no & no\\
1 &SUN &$\frac{1}{8}$full size$^{\dag}$& implicit & no & no\\
\midrule
2 &AIM & $\frac{1}{2}$full size$^{\dag}$& implicit & yes & no\\
2 &DVA & $80\times120^{\ddag}$ & explicit & yes & no\\
2 &PQFT &$64\times64^{\dag}$& explicit & yes & no\\
2 &Itti & full size$^{\ddag}$ & explicit & yes & no\\
\midrule
3 &GBVS & full size$^{\ddag}$ & explicit & no & yes\\
\bottomrule
\end{tabular}\\[2pt]
\footnotesize
$^{\dag}$The optimal image size for this model. $^{\ddag}$The default image size.
\end{table}
\subsection{Response to Psychological Patterns}
\label{EXP:Psy}
Psychological patterns are employed to evaluate three aspects of the algorithms: 1) first we use them to evaluate basic detection ability; 2) then we evaluate their ability to detect salient regions of different sizes, and 3) we evaluate their tolerance to random noise.
Four kinds of psychological patterns are employed: salient orientation and colored patterns (part A in Fig. \ref{fig:Exp1}), salient shape patterns (part B), asymmetric patterns (part C) and patterns with missing items (part D). The first column in Fig. \ref{fig:Exp1} shows the original images and the second shows the saliency maps produced by HFT. The proto-objects given by HFT are superimposed on the original images in the first column. Our results are compared with SR, PFT, PQFT, STB and GBVS.
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=7.2cm]{figures/32.pdf}
\end{overpic}
\caption{Responses to so-called psychological patterns. The first column shows the boundary of the primary object computed by HFT, superimposed on the original image. For comparison, the remaining columns present the results obtained by the methods mentioned earlier.}
\label{fig:Exp1}
\end{center}
\end{figure}
As shown in part A of Fig. \ref{fig:Exp1}, the first pattern is salient due to the distinguishing color. Both HFT and PQFT obtain acceptable results, while SR, PFT and STB are unable to highlight the salient bar. The second image contains a salient bar having both different color and orientation. HFT, GBVS and PQFT succeed in highlighting the red bar, while the other methods fail. The salient bar in the third image has a distinguishing orientation, and only GBVS failed to locate it.
In part B of Fig. \ref{fig:Exp1}, HFT, SR, PFT and PQFT function well. However, in B3, although both SR and PFT are able to highlight the salient region, the so-called common regions are not suppressed correctly. In B2, STB highlights the wrong area and in B3, both STB and PQFT cannot detect the salient region.
All of the algorithms are able to find the asymmetric salient regions in C1-C2. However, the common regions are not suppressed sufficiently by the SR, PFT, GBVS and PQFT. All of the algorithms perform well for C3, although HFT achieves the best result. Finding the salient bar in C4 is apparently a more difficult task for humans and this also seems to be the case for SR, PFT and PQFT. In general, the results are not as good as those for C3, and STB and GBVS has even failed completely.
Sometimes a salient region is simply an empty area, as shown in D1 of Fig. \ref{fig:Exp1}. A good salient detector should be able to locate such a region as well. We find that HFT, PFT, GBVS and PQFT can detect the missing item successfully, although HFT does a better job.
Overall HFT performs the best for all the cases shown in Fig. \ref{fig:Exp1}. We also note that SR and PFT obtain more or less the same results in cases A1-C2. However, they produce different results in cases C3-D1\footnote{In section \ref{relatedworks}, we draw the conclusion that both SR and PFT will yield nearly the same results based on the assumption of a natural image input. However, for "unnatural" images, they will sometimes produce different results, as is also discussed in \cite{ guo2010multiresolution}.}. Since PQFT is an advanced version of PFT, its performance is an improvement over the latter, especially in the case of colored tokens. However, in the rest of the cases, PQFT achieves nearly the same performance as PFT and SR.
We expect that a saliency detector would highlight salient regions of difference sizes in an image that people would pay attention to \cite{einhauser2008objects,LCAV-CONF-2009-012}. In order to examine this issue, we created three patterns in which the size of the tokens increased progressively, as shown in Fig. \ref{fig:Exp2}. All of the algorithms responded well to the small regions (see row 1). However, as the size increased, the performance of PQFT, Itti and SR/PFT are decreased. We observe that both SR/PFT and PQFT only respond to the boundaries of the regions when the salient region is large, while both HFT and GBVS highlight the salient region uniformly.
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=6cm]{figures/34.pdf}
\end{overpic}
\caption{Responses to psychological patterns with salient regions of different size.}
\label{fig:Exp2}
\end{center}
\end{figure}
Finally, in order to evaluate the noise tolerance of each model, we added different amounts of Gaussian (row 1-3 of Fig. \ref{fig:Exp3}) and salt$\&$pepper (row 4) noise to the pattern. As shown in Fig. \ref{fig:Exp3}, the proposed HFT obtained the best overall performance, while GBVS also performed quite well. GBVS is an improved version of Itti's saliency model, and its anti-noise properties have also improved. We observe that SR/PFT and PQFT are quite sensitive to both Gaussian and salt and pepper noise.
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=6cm]{figures/35.pdf}
\end{overpic}
\caption{Responses to psychological patterns with noise.}
\label{fig:Exp3}
\end{center}
\end{figure}
\subsection{Predicting Human Attention Using Fixations}
\label{exp_fixation}
We have evaluated HFT and compared it with state-of-the-art methods using human fixation data. Bruce's database\cite{NIPS2005_81} was employed for this purpose. It includes 120 natural images as well as corresponding eye-tracking data. The {\it quantitative} results are shown in Fig. \ref{fig:Exp33}.
We first compare HFT class models with models in subset 1. There is no border cut and center-bias in these models, so we need only find the optimal smoothing scale to compare the models. Fig. \ref{fig:Exp33}(A) shows the ROC scores for each model with different smoothing scales; we observe that they achieve their maximum ROC scores at different smoothing levels. We use the peak ROC score to establish the performance of each model, so that the influence of smoothing is compensated. In Fig.\ref{fig:Exp33}(A)-(C), the peak performance of each algorithm is labeled by a triangle. As shown in Fig. \ref{fig:Exp33}(A), it is obvious that HFT obtains the best performance, while SR, PFT and SUN have about the same performance.
\begin{figure}[h]
\begin{center}
\begin{overpic}[width=7.5cm]{figures/45.pdf}
\end{overpic}
\caption{Performance (peak ROC score) comparison between the HFT class models and those in the three subsets. (A)Comparing HFT with models in subset 1; (B)Comparing HFT with models in subset 2; (B)Comparing HFT with models in subset 3.}
\label{fig:Exp33}
\end{center}
\end{figure}
When comparing HFT with models in subset 2, the border cut is set for both HFT class models and the models in the subset. In fact, none of the models employ the same border cut, and therefore are not immediately comparable. Without calibration, the models in subset 2 will have a mendacious ROC score. As shown in Fig. \ref{fig:Exp33}(B), both HFT class algorithms and Itti's model are the highest performing models, while AIM has a higher peak performance than DVA.
In the more recent literature, GBVS always yields a very high ROC score and outperforms other models. However,this is most likely because GBVS incorporates a global center-bias \cite{judd2012pami}. When comparing HFT class models with GBVS, we selected the optimal center-bias for both. We note from Fig.\ref{fig:Exp33}(C) that HFT class algorithms achieve quite a high performance level, all of them outperforming GBVS.
\subsection{Predicting Salient Regions That Humans Attend}
\label{exp_region}
Besides using fixation data, we also used object maps labeled by humans to evaluate the algorithms (Refer to \url{http://www.cim.mcgill.ca/~lijian} for details of how this database was obtained).
Obviously, different images present different levels of difficulty for any saliency detector. However, the existing saliency benchmarks in the literature are collections of images, with no attempt to categorize the difficulty of analysis required. In this paper, a database containing 235 images was collected using Google as well as the recent literature. The images in this database were divided into 6 categories: 1) 50 images with large salient regions; 2) 80 images with intermediate salient regions; 3) 60 images with small salient regions; 4) 15 images with cluttered backgrounds; 5) 15 images with repeating distractors; 6) 15 images with both large and small salient regions. In this section, we report both the overall performance of each model as well as the performance of each model for each category.
In this experiment, we report only the performance at the optimal smoothing level. Both the ROC score (AUC) and the peak value of the DSC curve (PoDSC) for each model were calculated as shown in Table \ref{Smooth}-\ref{GBVS}. Fig.\ref{fig:Exp71} shows some examples which permit a qualitative comparison for each category in the dataset. However, due to space limitations, we are unable to show all of the {\it qualitative} results. However, these are available at \url{http://www.cim.mcgill.ca/~lijian}.
Fig.\ref{fig:Exp71}(A) shows natural images with large salient regions, a situation that is challenging for many models. It is clear that HFT achieves the best performance. The AUC and PoDSC criteria also support this conclusion. We note that GBVS achieves reasonable results, but SR, PQFT and AIM only enhance the boundaries instead of highlighting the whole salient region uniformly.
In Fig.\ref{fig:Exp71}(B), there are five images with intermediate salient regions. For example, in the second one, there are five salient flowers in the scene. HFT and GBVS have detected these object regions correctly. However, all of the other models failed to highlight them uniformly.
The images in Fig.\ref{fig:Exp71}(C) contain distant objects and distractors (e.g., the skyline in row 1). We observe that most of the algorithms work well in detecting the small salient regions. However, sometimes Itti's method and GBVS fail to suppress distractors (see row 1). HFT only slightly outperforms the others for this category.
The backgrounds of the images in Fig.\ref{fig:Exp71}(D) are quite cluttered. This case is also difficult because many algorithms are quite sensitive to background noise. For example, consider the image in row 2. For SR, PQFT, Itti and AIM, the non-salient grassy surface of the ground is enhanced as much as the two insects. However, HFT and GBVS detect these two regions correctly. HFT achieves excellent performance in this category, which is also supported by the quantitative results.
Compared to salient objects, repeating distractors in the scene should not attract much attention from humans \cite{beck2005stimulus}. Fig.\ref{fig:Exp71}(E) shows images with salient objects among repeating objects. In both rows 2 and 3, HFT and GBVS suppress the repeating objects and enhance the salient object correctly. In row 4, there is a salient playing card among the five, and HFT, PQFT and SR highlight the salient one and suppress the other four. HFT achieved the best performance for this category as well.
If an image contains both large and small salient regions (see Fig.\ref{fig:Exp71}(F)), a detector should be able to detect both simultaneously. For example, in row 1, there are two flowers of different size but SR, PQFT, Itti and AIM only respond strongly on their boundaries. However, HFT and GBVS respond correctly. Nevertheless, as discussed earlier, HFT selects just one optimal scale to determine the final output. Hence, objects of different size are not all detected or enhanced uniformly in the saliency map (see row 5 of Fig.\ref{fig:Exp71}(F)).
Overall, the experimental results shown in Table \ref{Smooth}-\ref{GBVS} indicate that the HFT model achieves the best performance for all six categories. Moreover, HFT exhibits superior performance when detecting large salient regions and saliency in cluttered scenes.
As mentioned earlier, images in this database contain salient regions of different sizes. Interestingly, in \cite{hou2007saliency} it is suggested that in order to find objects at different scales it should be possible to use different resolutions of the input image. In order to investigate this issue, we created different resolutions of the input image and then fed them into SR and the other one-resolution models (see Table \ref{one-scale}). We used the criterion described in (\ref{Eq:pscal5}) to find the optimal saliency map as the final output. We note that the performance of these revised models has improved (as shown in Table \ref{one-scale}), although it is still lower than the HFT class models (See Table \ref{Smooth}-\ref{GBVS}.)
\begin{table}
\renewcommand{\arraystretch}{1.3}
\caption{Performance of the revised one-resolution models}
\label{one-scale}
\centering
\begin{tabular}{lcc}
\toprule
Model & \bfseries AUC (improvement) & \bfseries PoDSC (improvement)\\
\hline
SR & 0.8733 ($\uparrow$ 0.0210) & 0.4316 ($\uparrow$ 0.0387)\\
PFT & 0.8769 ($\uparrow$ 0.0243) & 0.4420 ($\uparrow$ 0.0456)\\
PQFT & 0.8951 ($\uparrow$ 0.0197) & 0.4963 ($\uparrow$ 0.0234)\\
SUN & 0.8470 ($\uparrow$ 0.0067) & 0.4139 ($\uparrow$ 0.0298)\\
AIM & 0.8858 ($\uparrow$ 0.0027) & 0.4992 ($\uparrow$ 0.0050)\\
\bottomrule
\end{tabular}
\end{table}
Although HFT has performed well in the experiments described in section \ref{EXP:Psy}
to \ref{exp_region}, it does fail in certain cases. HFT could not satisfactorily predict the correct human fixations for several of the "hard" images collected in \cite{judd2012pami}. HFT did predict the human fixations correctly in Fig.\ref{failcases} (a, b), although some incorrect responses did occur. However, HFT did a poor job for some images. For example, in (c) it incorrectly highlighted some parts of the clothes and failed to highlight the eyes, while in (g), some parts of the boundary of the face were wrongly highlighted. In both (e) and (f), people tended to pay attention to the text, but HFT locates regions with salient low-level features (e.g., the red flag and the clock). In (d), HFT totally failed to locate the salient heads.
Clearly prior knowledge and task information is not employed for bottom-up models. Therefore, these approaches focus on regions possessing distinct low-level features (color, intensity etc.) and sometimes may fail to highlight the regions that are known to interest people (e.g., humans, animals and other common objects). One way to solve this problem is to employ more complex features or invoke top-down cues.
Most of the bottom-up saliency models, such as Itti, Gao's model, AIM and so on, use local contrast or center surround paradigm. Similarly, models like SR can also be considered as pixel-level local contrast models (gradient operation). These work well for detecting small salient regions, but do not perform well in predicting large salient regions. There are two ways to alleviate this problem; one is to adopt a multi-scale strategy (as used in Itti's model), the other is to decrease the resolution of the input image and employ a large amount of blurring of the saliency maps (as used in SR, PFT).
Finally, perhaps it is unfair to characterize SR class models as being only pixel-level local contrast deterctors. As discussed earily, SR and PFT are special cases of the proposed HFT model when the scale goes to infinity in the frequency domain. Hence they have the ability to globally inhibit and suppress repeated patterns. However, other models based on local contrast will perform poorly in this case. Nevertheless, if there are no repeated patterns in the scene, the SR model will function as a gradient detector, and only enhance boundaries of objects.
\begin{figure}
\begin{center}
\includegraphics[width=.8\linewidth]{figures/failcases.pdf}
\end{center}
\caption{Hard image cases of HFT in predicting human fixations}
\label{failcases}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=.97\linewidth]{figures/7100.pdf}
\end{center}
\caption{Procedure for computing saliency using the Hypercomplex Fourier Transform (HFT)}
\label{fig:Exp71}
\end{figure*}
\begin{table*}[htbp]
\caption{Comparison between HFT class models and models in subset 1 (optimal smoothing parameters for each algorithm)}
\label{Smooth}
\centering
\begin{tabular}{@{}lp{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}@{}}
\toprule
&\multicolumn{2}{c}{Category 1}
&\multicolumn{2}{c}{Category 2}
&\multicolumn{2}{c}{Category 3}
&\multicolumn{2}{c}{Category 4}
&\multicolumn{2}{c}{Category 5}
&\multicolumn{2}{c}{Category 6}
&\multicolumn{2}{c}{Overall}\\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11} \cmidrule(lr){12-13} \cmidrule(lr){14-15}
Model & \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}\\
\midrule
{\bf HFT} &{\bf0.9424} &{\bf0.7252} &{\bf0.9146} &{\bf0.5481} &{\bf0.9351} &{\bf0.4563} &{\bf0.9448} &{\bf0.5856} &{\bf0.9193} &{\bf0.5778} &{\bf0.9535} &{\bf0.6976} &{\bf 0.9281} &{\bf 0.5417}\\
HFT(e)&0.9101 &0.6592 &0.9050 &0.5112 &0.9348 &0.4502 &0.9463 &0.6306 &0.8907 &0.5123 &0.9418 &0.6489 &0.9159 &0.5029\\
HFT* &0.9543 &0.7438 &0.9425 &0.6184 &0.9572 &0.5217 &0.9686 &0.6846 &0.9413 &0.6148 &0.9709 &0.7568 &0.9497 &0.5963\\
SR &0.8148 &0.5104 &0.8495 &0.4321 &0.9091 &0.3281 &0.7595 &0.2796 &0.7929 &0.3266 &0.8924 &0.5404 &0.8523 &0.3929\\
PFT &0.8064 & 0.5029 & 0.8426 & 0.4292 & 0.9269 & 0.3780 & 0.7294 & 0.2931 & 0.7724 & 0.3377 & 0.8967 & 0.5574 &0.8526 &0.3964\\
SUN &0.8218 &0.5393 &0.8457 &0.4522 &0.8838 &0.3026 &0.6994 &0.2452 &0.8067 &0.3773 &0.8778 &0.5555 &0.8403 &0.4018\\
\bottomrule
\end{tabular}
\end{table*}
\begin{table*}[htbp]
\caption{Comparison between HFT class models and models in subset 2 (optimal smoothing parameters and the same border cut for each model)}
\label{AIM}
\centering
\begin{tabular}{@{}lp{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}@{}}
\toprule
&\multicolumn{2}{c}{Category 1}
&\multicolumn{2}{c}{Category 2}
&\multicolumn{2}{c}{Category 3}
&\multicolumn{2}{c}{Category 4}
&\multicolumn{2}{c}{Category 5}
&\multicolumn{2}{c}{Category 6}
&\multicolumn{2}{c}{Overall}\\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11} \cmidrule(lr){12-13} \cmidrule(lr){14-15}
Model & \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}\\
\midrule
{\bf HFT} &{\bf0.9338} &{\bf0.7395} &{\bf0.9064} &{\bf0.5697} &{\bf0.9328} &{\bf0.4871} &{\bf0.9378} &{\bf0.6074} &{\bf0.9137} &{\bf0.5893} &{\bf0.9441} &{\bf0.7114} &{\bf0.9217} &{\bf0.5627}\\
HFT(e) &0.9010 &0.6867 &0.9004 &0.5402 &0.9312 &0.4621 &0.9471 &0.6568 &0.8660 &0.5286 &0.9340 &0.6743 &0.9102 &0.5289\\
HFT* &0.9478 &0.7584 &0.9387 &0.6434 &0.9578 &0.5503 &0.9660 &0.7012 &0.9374 &0.6330 &0.9644 &0.7658 &0.9470 &0.6226\\
AIM &0.8511 &0.6011 &0.8761 &0.5226 &0.9359 &0.4506 &0.8370 &0.3969 &0.8668 &0.4987 &0.9124 &0.6489 &0.8831 &0.4942\\
PQFT &0.8571 &0.6201 &0.8771 &0.5350 &0.9096 &0.3901 &0.8205 &0.3819 &0.8421 &0.4304 &0.9105 &0.6398 &0.8754 &0.4729\\
DVA &0.8075 &0.5736 &0.8565 &0.5095 &0.9038 &0.3957 &0.7618 &0.3639 &0.8250 &0.4553 &0.9048 &0.6262 &0.8510 &0.4642\\
Itti &0.8768 &0.6533 &0.8886 &0.5317 &0.9239 &0.3843 &0.8107 &0.3687 &0.8983 &0.5194 &0.9191 &0.6530 &0.8910 &0.4949\\
\bottomrule
\end{tabular}
\end{table*}
\begin{table*}[htbp]
\caption{Comparison between HFT class models and model in subset 3 (optimal smoothing parameters and center-bias for each model)}
\label{GBVS}
\centering
\begin{tabular}{@{}lp{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}p{.78cm}@{}}
\toprule
&\multicolumn{2}{c}{Category 1}
&\multicolumn{2}{c}{Category 2}
&\multicolumn{2}{c}{Category 3}
&\multicolumn{2}{c}{Category 4}
&\multicolumn{2}{c}{Category 5}
&\multicolumn{2}{c}{Category 6}
&\multicolumn{2}{c}{Overall}\\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11} \cmidrule(lr){12-13} \cmidrule(lr){14-15}
Model & \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}
& \scriptsize{AUC} & \scriptsize{PoDSC}\\
\midrule
{\bf HFT} &{\bf0.9565} &{\bf0.7548} &{\bf0.9296} &{\bf0.5688} &{\bf0.9504} &{\bf0.4755} &{\bf0.9381} &{\bf0.5799} &{\bf0.9523} &{\bf0.6318} &{\bf0.9578} &{\bf0.7020} &{\bf0.9414} &{\bf0.5665}\\
HFT(e) &0.9409 &0.7213 &0.9287 &0.5697 &0.9614 &0.5024 &0.9460 &0.6315 &0.9361 &0.6038 &0.9579 &0.7089 &0.9403 &0.5609\\
HFT* &0.9665 &0.7771 &0.9533 &0.6396 &0.9724 &0.5640 &0.9709 &0.7028 &0.9644 &0.6841 &0.9743 &0.7586 &0.9609 &0.6250\\
GBVS &0.9363 &0.6990 &0.9135 &0.5304 &0.9173 &0.3678 &0.9223 &0.5644 &0.9453 &0.6145 &0.9249 &0.6329 &0.9211 &0.5154\\
\bottomrule
\end{tabular}
\end{table*}
\section{Conclusions}
This paper proposes a new saliency detection framework for images, based on analyzing the spectrum scale-space. We show that the convolution of the image amplitude spectrum with a low-pass Gaussian kernel of an appropriate scale is equivalent to such an image saliency detector. The proposed approach is able to highlight both small and large salient regions and inhibit repeated patterns. We also illustrate that both SR and PFT are special cases of the proposed model when the scale parameter goes to infinity.
In order to fuse multi-dimensional feature maps, we employ the Hypercomplex Fourier Transform to replace the standard Fourier Transform for spectrum scale-space analysis.
To validate the proposed approach, we have performed saliency computations on both commonly used synthetic data as well as natural images, and then compared the results with state-of-the-art algorithms. In order to make a fair comparison when using the ROC and PoDSC as measures of performance, we have proposed an improved comparison procedure by considering the border cut, center-bias and smoothing effects. Experimental results indicate that the proposed model can predict human fixation data as well as the object regions labeled by humans. We also show that sometimes HFT may fail to predict human fixations. This is most likely because only low-level features are employed; Clearly, top-down or task-orientated cues are necessary for improving the performance of current saliency models in predicting human attention.
With regard to future work, firstly, it would be interesting to investigate other criteria for optimal scale selection. Entropy was employed in this paper to select the optimal scale automatically. However, we have observed that a much higher performance can be obtained by selecting the optimal scale manually. Secondly, in the proposed model, only one saliency map, corresponding to the optimal scale, is selected as the final one. However, we have noted that certain of the abandoned maps also contain meaningful saliency information. How to incorporate these in the determination of the saliency is left to future investigation. Thirdly, we intend to include top-down information to improve performance. The ultimate goal of our research is to develop a system for on-board pedestrian and vehicle detection, for which a considerable amount of top-down temporal data exists.
{\small
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Let $g(t)$ be a one parameter family of Riemannian metrics for $ t \in [0, T]$.
We say that the time-dependent $g_{ij}(x,t)$ is a generalized geometric flow if it is evolving by the following
equation
\begin{equation}\label{eq11}
\frac{\partial }{ \partial t} g_{ij} (x,t) = - 2 h_{ij}(x,t), \ \ \ \ (x, t) \in M \times [0,T]
\end{equation}
with $g_{ij}( x,0) = g_0(x)$, where $x \in M$, $0<T<T_\epsilon$ is the maximal time of existence, i.e., $T_\epsilon$ is the first time where the flow blows-up and $h_{ij}$ is a general time-dependent symmetric $(0, 2)$-tensor. Here $h_{ij}$ is smooth in both variables $t$ and $x$. This is obvious since $g_{ij}$ is smooth in both variables also.
The scaling factor $2$ in (\ref{eq11}) is insignificant while the negative sign may be important
in some specific applications for the purpose of keeping the flow either forward or backward in time. Two popular examples of geometric flows are the Ricci flow where $h_{ij} = R_{ij}$ (the Ricci curvature tensor) and the mean curvature flow, in which case $h_{ij} = H \Pi_{ij}$ (here $H$ is the mean curvature and $\Pi_{ij}$ is the second fundamental form on $M$).
One may impose boundedness condition on the tensor $h_{ij}$. In fact, such boundedness and sign assumptions on $h_{ij}$ are preserved as long as the flow exists, so it follows that the metrics are uniformly equivalent. Precisely, if $ -K_1 g \leq h \leq K_2g$, where $g(t), t \in [0, T]$ is the flow, then
\begin{equation}\label{eq2.3}
e^{-K_1 T}g(0) \leq g(t) \leq e^{K_2 T}g(0).
\end{equation}
To see the above bounds (\ref{eq2.3}) we consider the evolution of a vector form $|X|_g = g(X, X), X \in T_xM$. By the equation (\ref{eq11}) and by the boundedness of the tensor $h$ we have
$ | \partial_t g(X, X) | \leq K_2 g(X, X),$
which implies (by integrating from $t_1\ to\ t_2$)
$$ \Big| \log \frac{g(t_2)(X, X)}{g(t_1)(X, X)}\Big| \leq K_2 t \Big|_{t_1}^{t_2}.$$
Taking the exponential of this estimate with $t_1 =0$ and $t_2 =T$ yields $|g(t)| \leq e^{k_2 T}g(0)$ from which the uniform boundedness of the metric follows. Thus, if there holds boundedness assumption $ -K_1 g \leq h \leq K_2g$, the metric $g(t)$ are uniformly bounded below and above for all time $0 \leq t \leq T$ under the geometric flow (\ref{eq11}). Then, it does not matter what metric we use in the argument that follows.
In this paper, we
discuss the evolution, monotonicity and differentiability of the first nonzero eigenvalue of $p$-Laplace operator on the $n$-dimensional compact Riemannian manifold whose metric evolves by (\ref{eq11}). The $p$-Laplace operator is defined as
$$ \Delta_{p, g} f(x) := \ div (|\nabla f|^{p-2} \nabla f)(x)$$
for $ p \in [1, \infty)$, where $div$ is the divergence operator, the adjoint of gradient ($grad$) for the $L^2$-norm induced by $g$ on the space of differential forms. When $p=2$,
$$ \Delta_{2, g} f(x) = div \circ grad \ f(x) $$
is the usual Laplace-Beltrami operator. The eigenvalues and the corresponding eigenfunctions of $ \Delta_{p, g}$ satisfy the following nonlinear eigenvalue problem
\begin{equation}\label{eq12}
\Delta_{p, g} f = - \lambda |f|^{p-2} f , \hspace{1cm} f \neq 0.
\end{equation}
It is easily verifiable that the principal symbol of (\ref{eq12}) is nonnegative everywhere and strictly positive at the neighbourhood of the point where $\nabla f \neq 0$.
We know that (\ref{eq12}) has weak solutions with only partial regularity of class $C^{1, \alpha}, (0< \alpha < 1)$ in general. Interested readers can find the classical papers by L. Evans \cite{[Ev81]} and P. Tolksdorff \cite{[Tok84]}. Notice that the least eigenvalue of $\Delta_{p,g}$ on compact manifold without boundary is zero with the corresponding eigenfunction being a constant. Hence, we refer to the infimum of the positive eigenvalues as the first nonzero eigenvalue or simply the first eigenvalue. The first eigenvalue of $ \Delta_{p, g}$ is characterised by the min-max principle
\begin{equation}\label{eq13}
\lambda_{p,1} = \inf_{0\neq f \in W^{1, p}(M) }\Bigg\{ \frac{\int_M | \nabla f |_g^p \ d \mu_g}{\int_M | f|^p_g \ d\mu_g} \ \ \Big| \ \ f \neq 0, \ \ f \in W^{1, p}(M) \Bigg\},
\end{equation}
satisfying the following constraint $\int_M |f|_g^{p-2} f d \mu_g = 0$, where $d \mu_g$ is the volume measure on $(M,g)$. Obviously, the infimum does not change when one replaces $ W^{1, p}(M)$ by $C^\infty(M)$.
The corresponding eigenfunction is the energy minimizer of Rayleigh quotient (\ref{eq13}) and satisfies the following Euler-Lagrange equation
\begin{equation}
\int_M [ |\nabla f |^{p-2} \langle \nabla f, \nabla \phi \rangle - \lambda |f|^{p-2} \langle f, \phi \rangle ]d\mu_g= 0
\end{equation}
for $ \phi \in C^\infty_0(M)$ in the sense of distribution.
The problem of finding $ \lambda_{p,1}$ is related to the problem of finding the best constant $C(M)$ in the $L^p$-Sobolev inequality
$$ \| f \|_{L^{p^*}} \leq C(M) \| \nabla f \|_{L^p},\ \ \ p^* = np/(n-p),$$
which is obtained by continuous embedding of $W^{1,p}(M) \hookrightarrow L^p(M)$ under the sobolev norm
$$\|f\|_{W^{1,p}} = \Big(\int_M |f|^p d\mu + \int_M | \nabla f|^p d\mu \Big)^{\frac{1}{p}}.$$
It is well-known that $p$-Laplacian has discrete eigenvalues but still remains unknown whether it only has discrete eigenvalues for bounded connected domains. Another well-known results tell us that the first nonzero eigenvalue is simple and isolated \cite{[Linq1],[Linq2],[MaRos]}. Here the simplicity shows that any nontrivial eigenfunction corresponding to $\lambda_{p,1}$ does not change sign and that any two first eigenfunctions are constant multiple of each other.
In contrast to the spectrum of Laplace-Beltrami operator (the case $p=2$), the $p$-Laplacian is nonlinear in general. For instance, the case $p=1$ gives $div(\nabla u /|\nabla u|)$, the negative of mean curvature operator. Aside being of geometric interests, the $p$-Laplacian appears naturally in the study of non-newtonian fluids, nonlinear elasticity, heat radiation, porous media flow, rheology, petroleum extraction, Brownian motions, torsional creep problem among others. For details of physical applications of $p$-Laplacian, the reader is referred to \cite{[DiSe]}.
Moreover, it is not known if $\lambda_{p,1}$ or its corresponding eigenfunction is $C^1$-differentiable (or even locally Lipschitz) along any geometric flow of the form (\ref{eq11}). However, it has been pointed out that the differentiability for the case $p=2$ is a consequence of eigenvalue perturbation theory, see for instance \cite{[Kat84],[KL06]}. For this reason, the method of L. Ma \cite{[Ma06]}, which assumes differentiability of eigenvalues and eigenfunctions under the Ricci flow, can only be applied to find the monotonicity of first eigenvalue for the case $p=2$ along (\ref{eq11}). Now to avoid the differentiability assumption on the first eigenvalue and the corresponding eigenfunction in the case $p \neq 2$, we shall apply techniques of Cao \cite{[Ca07]} as used by Wu \cite{[Wu11]} and Wu, Wang and Zheng \cite{[WWZ]} under the Ricci flow to study the evolution and monotonicity of $\lambda_{p,1}(t) = \lambda_{p,1}(t, f(t))$, where $\lambda_{p,1}(t, f(t))$ and $f(t)$ are assumed to be smooth. The evolution and the monotonicity formulas for the first eigenvalue (in both cases $p=2$ and $p \neq 2$) derived here do not depend on the evolution of the eigenfunction. The eigenfunction only needs to satisfy certain normalization condition.
There are many results on the evolution and monotonicity of eigenvalues of the Laplace operator on evolving manifolds with or without curvature assumptions. One can find \cite{[Ca07],[Ca08],[Li07a],[Li07b]} under the Ricci flow, \cite{[Li10]} under Ricci-Harmonic map flow and \cite{[GPT]} along abstract geometric flow with entropy methods. The study of the properties of eigenvalues of the $p$-Laplacian on evolving manifold is still very young. The main aim of this paper is to investigate if those known properties of $\lambda_{p,1}$ on static metric and for the case $p=2$ on evolving metric can be extended to various geometric flow. We however intend to develop a unified algorithm that can be used for this purpose on time-dependent metrics. Many interesting results concerning the behaviour of $\lambda_{p,1}$ can be found in \cite{[KaFr],[Linq1],[Linq2],[Mao2],[MaRos]} for static metrics and \cite{[Ab15],[GLW16],[Mao1],[Wu11],[WWZ],[Zha12],[Zha13]} for evolving metrics along various geometric flow.
The rest of this paper is planned as follows. In Section \ref{sec2}, we discuss the main results of this paper. Firstly, we highlight some notations about differential geometry including metrics, gradient, divergence and integration by parts, which form the core of analytic tools used in the paper. We also prove some technical lemmas about evolutions of some geometric objects relating to the eigenvalues under the flow (\ref{eq11}). In Section \ref{sec3}, we study the first eigenvalue of Laplace-Beltrami operator ($p=2$) under this geometric flow. Here, we assume the first eigenvalue to be a function of time only and obtain its general evolution and monotonicity formula under certain condition. In Section \ref{sec4}, we use a different approach to derive $p$-eigenvalue's evolution and monotonicity without differentiability assumption on the eigenfunction. In fact, the differentiability of $p$-eigenvalue is a consequence of the monotonicity formula. In the last section, we list some examples of geometric flows where the approaches used in this paper are applicable. In fact, this section reveals that our generalised geometric flow is not a trivial generalisation.
\section{Preliminaries and main results}\label{sec2}
\subsection{Notation}
Throughout, $M$ will be taken to be a closed manifold (i.e., compact without boundary).
Most of our calculations are done in local coordinates, where $ \{ x^i \}$ is fixed in a neighbourhood of every point $ x \in M$. We shall adopt Einstein summation convention with repeated indices summed up.
The Riemannian metric $g(x)$ at any point $x \in M$ is a bilinear symmetric positive definite matrix denoted in local coordinates by
$$g_{ij}(x) = g_{ij} dx^i dx^j.$$
The Laplace-Beltrami operator acting on a smooth function $ f $ on $M$ is defined as divergence of gradient of $f$, written as
$$\Delta_g f := div \ grad \ f =\frac{1}{\sqrt{|g|}} \frac{\partial }{\partial x^i} \Big( \sqrt{|g|}g^{ij} \frac{\partial f}{\partial x^j} \Big), $$
where $|g| = \det (g_{ij})$ and $g^{ij} = (g_{ij})^{-1}$ are determinant and the inverse metric matrix respectively. By the above we note that
$$(grad \ f)^i = (\nabla f)^i = g^{ij} \frac{\partial f}{\partial x^j} \ \ \mbox{and} \ \ div X = \frac{1}{\sqrt{|g|}} \frac{\partial }{\partial x^i} ( \sqrt{|g|} X^i ), $$
where $X$ is a smooth vector field.
Also we have the metric norm
$$| \nabla f|^2_g = g^{ij} \nabla_i f \nabla_j f = \nabla^j f \nabla_j f.$$
The Riemann structure allows us to define Riemannian volume measure $d \mu_g$ on $M$ by
$$ d \mu_g = \sqrt{|g_{ij}(x)|}dx^i .$$
By the divergence theorem, we have the following integration by parts formulas:
Let $X$ be a vector field, $X = X^i \partial_i$ and $f \in C^\infty(M)$ be smooth function, then
$$ \int_M \langle - div X, f \rangle_g = \int_M \langle X, \nabla f \rangle_g = - \int_M \frac{1}{\sqrt {\det g}} f \partial_i ( X^i \sqrt{\det g }) \sqrt{\det g} \ d x^i .$$
Also for functions $f, h \in C^2(M)$
$$ \int_M f \Delta_g h \ d\mu = - \int_M \langle \nabla f, \nabla h \rangle_g d\mu = \int_M \Delta_g f \ h d\mu.$$
We write in local coordinates gradient, Hessian and covariant derivative as
$$\nabla f = f_i, \ \ \nabla \nabla f = \nabla_i \nabla_j f = f_{ij} \ \ and \ \ \frac{\partial}{\partial x^i} = \partial_i$$
respectively.
Also we write time derivative as $\frac{\partial}{\partial t} f = \partial_t f = f_t$.
\subsection{Evolution equations}
Interestingly, all the geometric quantities associated with the underlying manifold evolve as the Riemannian metric evolves along the geometric flow. This also serves as a motivation considering the behaviours of some other important geometric quantities such as eigenvalues of the manifold under the flow. The next two lemmas give us these evolutions.
\begin{lemma}\label{lem211}
Suppose a one-parameter family of smooth metrics $g(t)$ solves the geometric flow (\ref{eq11}), then, we have the following evolutions
\begin{align*}
\displaystyle & (1) \hspace{2cm} \frac{\partial}{\partial t } g^{ij} = 2 g^{ik} g^{jl} h_{kl} = 2 h^{ij}\\
\displaystyle & (2)\hspace{2cm} \frac{\partial}{\partial t } \Gamma^k_{ij} = - g^{kl} \Big( \partial_i h_{jl} + \partial_j h_{il} - \partial_l h_{ij} \Big) \\
\displaystyle & (3) \hspace{2cm} \frac{\partial}{\partial t} | \nabla f|^2 = 2 h_{ij} f_i f_j + 2 f_i f_{t,j} \\
\displaystyle & (4) \hspace{2cm} \frac{\partial}{\partial t} ( \Delta f ) = 2 h_{ij}f_{ij} + 2 \langle \ div \ h , \nabla f \rangle - \langle \nabla \mathcal{H} , \nabla f \rangle + \Delta f_t \\
\displaystyle & (5) \hspace{2cm} \frac{\partial}{\partial t} d \mu = - \mathcal{H} d\mu.
\end{align*}
\end{lemma}
Here $\mathcal{H} = g^{ij} h_{ij}$, the metric trace of a symmetric $2$-tensor $h_{ij}$ and $f$ is a smooth function defined on $M$.
\begin{proof}
Recall that both $g_{ij}$ and $\mathcal{S}_{ij}$ are symmetric tensors and $ g^{ij}g_{jl} = \delta^i_l$.
Note also that Levi-civita connection is not a tensor, but the time derivative of a connection is $(2, 1)$ tensor. Then $(2)$ holds as a tensor equation in any coordinates system and at any point. The proofs of $(1)$ and $(2)$ are the same as those of \cite[Lemmas 3.1 and 3.2]{[CK04]}. We give the computations in local coordinates which lead to the proofs of $(3)$ and $(4)$. In fact, we have
\begin{align*}
\partial_t (| \nabla f|^2) & = \partial_t(g^{ij} \partial_i f \partial_j f) \\
\displaystyle &= (\partial_t g^{ij}) \partial_i f \partial_j f + 2 g^{ij} \partial_i f \partial_j f_t \\
\displaystyle & = 2 h^{ij} \partial_i f \partial_j f + 2 g^{ij} \partial_i f \partial_j f_t \\
\displaystyle & = 2 h_{ij} f_i f_j + 2 f_i f_{t, j},
\end{align*}
which is exactly (3).
We prove (4) by using (1) and (2) as follows:
\begin{align*}
\frac{\partial}{\partial t} ( \Delta f) & = \frac{\partial}{\partial t} [ g^{ij} ( \partial_i \partial_j - \Gamma_{ij}^k \partial_k ) f ] \\
& = \frac{\partial}{\partial t} (g^{ij}) ( \partial_i \partial_j - \Gamma_{ij}^k \partial_k ) f + g^{ij} ( \partial_i \partial_j - \Gamma_{ij}^k \partial_k ) \frac{\partial}{\partial t} f - g^{ij} \Big( \frac{\partial}{\partial t} \Gamma_{ij}^k \Big) \partial_k f \\
& = 2 h^{ij} \nabla_i \nabla_j f + \Delta f_t + g^{ij} g^{kl} \Big( \nabla_i h_{jl} + \nabla_j h_{il} - \nabla_l h_{ij}\Big) \nabla_k f \\
&= \Delta f_t + 2 h_{ij} f_{ij} + 2 g^{kl} \Big( g^{ij} \nabla_i h_{jl} - \frac{1}{2} g^{ij} \nabla_l h_{ij} \Big) \nabla_k f \\
& = \Delta f_t + 2 h_{ij} f_{ij} +2 \langle \ div \ h , \nabla f \rangle - \mathcal{H} _i f_j.
\end{align*}
In local coordinates, the volume form is written as $ d \mu = \sqrt{|g| } dx^1 \wedge . . .\wedge dx^n$, then,
$$\frac{\partial}{\partial t} d\mu = \frac{\partial}{\partial t} \Big(\sqrt{| g| }dx^1 \wedge . . .\wedge dx^n \Big).$$
By chain rule of differentiation
$$\frac{\partial}{\partial t} \Big( \sqrt{| g| } \Big) = \frac{1}{2} \frac{1}{\sqrt{|g| }} \frac{\partial}{\partial t} | g| = \frac{1}{2} \frac{1}{\sqrt{|g| }} \frac{\partial | g|}{\partial g_{ij} } \frac{\partial g_{ij}}{\partial t }$$
$$\ \ \ \ \ = - \sqrt{|g |}g^{ij}h_{ij} = - \mathcal{H} \sqrt{|g| }.$$
Therefore
$$\frac{\partial}{\partial t} d \mu = - \mathcal{H} \ d\mu.$$
This completes the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem212}
Suppose a one-parameter family of smooth metrics $g(t)$ solves the geometric flow (\ref{eq11}), then, we have the following evolutions
\begin{align*}
\displaystyle & (1) \hspace{2cm} \frac{\partial}{\partial t} | \nabla f|^p = p | \nabla f|^{p-2} \Big\{ h_{ij} \nabla_i f \nabla_j f + g^{ij} \nabla_i f \nabla_j f_t \Big\}\\
\displaystyle & (2) \hspace{2cm} \frac{\partial}{\partial t} | \nabla f|^{p-2} = (p-2) | \nabla f|^{p-4} \Big\{ h_{ij} \nabla_i f \nabla_j f + g^{ij} \nabla_i f \nabla_j f_t \Big\}\\
\displaystyle & (3) \hspace{2cm} \frac{\partial}{\partial t} ( \Delta_{p, g} f ) = 2 h^{ij} \nabla_i ( Z \nabla_j f ) + g^{ij} \nabla_i ( Z_t \nabla_j f ) + g^{ij} \nabla_i (Z \nabla_j f_t )\\
\displaystyle & \hspace{5cm} + Z \Big\{2 \langle \ div \ h , \nabla f \rangle - \langle \nabla \mathcal{H} , \nabla f \rangle + \Delta f_t \Big\},
\end{align*}
where $Z:= | \nabla f|^{p-2}$ and $f$ is a smooth function defined on $M$. When $p=2$ we have (2) and (3) of Lemma \ref{lem211}.
\end{lemma}
The proof follows standard computation as in \cite[Lemma 2.2]{[Ab15]}. But we include it here for completeness.
\begin{proof}
By (3) of Lemma \ref{lem211} we can get
\begin{align*}
\frac{\partial}{\partial t} \Big(| \nabla f|^p \Big) & = \frac{\partial}{\partial t} \Big( | \nabla f|^2\Big)^{\frac{p}{2}}\\
\displaystyle &= \frac{p}{2} \Big( | \nabla f|^2\Big)^{\frac{p-2}{2}} \frac{\partial}{\partial t} \Big( | \nabla f|^2\Big)\\
\displaystyle &= \frac{p}{2} | \nabla f|^{p-2} \Big\{ 2 h_{ij} \nabla_i f \nabla_j f + 2 g^{ij} \nabla_i f \nabla_j f_t \Big\} \\
\displaystyle & = p | \nabla f|^{p-2} \Big\{ h_{ij} \nabla_i f \nabla_j f + g^{ij} \nabla_i f \nabla_j f_t \Big\},
\end{align*}
which is (1). (2) follows from the same calculation as the above. Let $Z:= | \nabla f|^{p-2}$, then
\begin{align*}
\frac{\partial}{\partial t} \Delta{p, g} f & = \frac{\partial}{\partial t} \Big(div ( | \nabla f|^{p-2} \nabla f )\Big) = \frac{\partial}{\partial t} \Big(g^{ij} \nabla_i (Z \nabla_j f )\Big) \\
\displaystyle &= \frac{\partial}{\partial t} \Big(g^{ij} \nabla_i Z \nabla_j f + Z g^{ij} \nabla_i \nabla_j f \Big) \\
\displaystyle &= \frac{\partial}{\partial t} \Big( g^{ij} \Big) \nabla_i Z \nabla_j f + g^{ij} \nabla_i Z_t \nabla_j f + g^{ij} \nabla_i Z \nabla_j f_t + Z_t \Delta f + Z (\Delta f)_t .
\end{align*}
By (1) and (3) of the Lemma, we have
\begin{align*}
\frac{\partial}{\partial t} \Delta{p, g} f & = 2 h^{ij} \nabla_i Z \nabla_j f + g^{ij} \nabla_i Z_t \nabla_j f + g^{ij} \nabla_i Z \nabla_j f_t + Z_t \Delta f \\
\displaystyle &\hspace{1cm} + Z \Big\{ 2 h_{ij} \nabla_i \nabla_j f + 2 \langle \ div \ h , \nabla f \rangle - \langle \nabla \mathcal{H} , \nabla f \rangle + \Delta f_t \Big\}\\
\displaystyle & = 2 h^{ij} \nabla_i Z \nabla_j f + 2 h_{ij} Z \nabla_i \nabla_j f + g^{ij} \nabla_i Z_t \nabla_j f + Z_t \Delta f + g^{ij} \nabla_i Z \nabla_j f_t\\
\displaystyle &\hspace{1cm} + Z \Delta f_t + Z \Big\{2 \langle \ div \ h , \nabla f \rangle - \langle \nabla \mathcal{H} , \nabla f \rangle \Big\}\\
\displaystyle & = 2 h^{ij} \nabla_i ( Z \nabla_j f ) + g^{ij} \nabla_i ( Z_t \nabla_j f ) + g^{ij} \nabla_i (Z \nabla_j f_t ) \\
\displaystyle &\hspace{1cm} + Z \Big\{2 \langle \ div \ h , \nabla f \rangle - \langle \nabla \mathcal{H} , \nabla f \rangle \Big\}.
\end{align*}
\end{proof}
\subsection{Main Results}
Recall that we already mentioned that sign assumption on the tensor $h_{ij}$ is preserved throughout the flow. To prove the monotonicity of $\lambda_{p,1}$, we will need the condition $(h_{ij} - \alpha \mathcal{H} g_{ij})(x,t) \geq 0$ for all $t \in [0, T]$. This condition is informed by the Hamilton's maximum principle for tensors. For clarity we state the principle without proof.
\begin{definition}({\bf Null-eigenvector assumption})
A quantity $Q(q,t) : Sym^2T^*M \times [0,T] \to Sym^2T^*M$ is said to satisfy the null eigenvector assumption if whenever $\omega_{ij}$ is a nonnegative symetric $(0,2)$-tensor at a point $q$ and if $X \in T_qM$ is such that $\omega_{ij} X^j =0$ and then
$$Q_{ij}(\omega,g)X^iX^j \geq 0$$
for any $t \in [0,T]$.
\end{definition}
The symetric tensor $\omega_{ij}$ is defined to be nonnegative if and only if $\omega_{ij} V^iV^j \geq 0$ for all vectors $V^i$ (i.e., the quadratic form induced by $\omega_{ij}$ is semi-positive definite) and we write $\omega_{ij} \geq 0$.
\begin{theorem}(\cite[Theorem 9.1]{[Ha82]},\cite[Theorem 4.6]{[CK04]})
Let $g(t)$ be a smooth one parameter family of Riemannian metrics satisfying (\ref{eq11}). Let $M_{ij}$ be a symmetric $(0,2)$-tensor satisfying
\begin{align}
\frac{\partial}{\partial t} M_{ij}(x,t) \geq \Delta M_{ij}(x,t) + \langle X, \nabla M_{ij}\rangle + Q(M_{ij},g(t)),
\end{align}
where $X$ is a time-dependent vector field and $Q(M_{ij},g(t))$ is a symmetric $(0,2)$-tensor which is locally Lipschitz in $x$, continuos in $t$ and satisfies the null-eigenvector assumption.
If $M_{ij}(p,0) \geq 0$ for all $p \in M$, then $M_{ij}(p,t) \geq 0$ for all $p \in M$ and $t \in [0,T_\epsilon)$
\end{theorem}
By the above maximum principle for tensors we can prove the following conclusion.
\begin{proposition}
Let $g(t)$ be a smooth one parameter family of Riemannian metrics satisfying (\ref{eq11}). If
$$(h_{ij} - \alpha \mathcal{H} g_{ij})(x,0) \geq 0,$$
then
\begin{align}
(h_{ij} - \alpha \mathcal{H} g_{ij})(x,t) \geq 0
\end{align}
for some $\alpha \in [0,\frac{1}{n}]$ and all $t \in [0, T].$
\end{proposition}
The first main results on evolution and monotonicity of $\lambda_{p,1}$ for the case $p=2$ are the following and are proved in Section \ref{sec3}.
\begin{theorem}\label{thm311}
Let $(M, g(t))$ be a closed manifold evolving by the geometric flow (\ref{eq11}) and $\lambda(t)$ be the first nonzero eigenvalue of the Laplacian $\Delta_{g(t)}$ corresponding to the eigenfunction $ \varphi(t, x)$, then $\lambda(t)$ evolves by
\begin{equation}
\frac{d}{d t} \lambda(t) \int_M \varphi^2 \ d\mu = \lambda(t) \int_M \mathcal{H} \varphi^2 \ d\mu - \int_M \mathcal{H} | \nabla \varphi |^2 \ d\mu + 2 \int_M \langle h, d \varphi \otimes d \varphi \rangle \ d\mu.
\end{equation}
\end{theorem}
By setting $ h_{ij} - \alpha \mathcal{H}g_{ij} \geq 0, \alpha \geq 1/2 $, along the flow we have the following monotonicity formula from the last theorem
\begin{equation}
\frac{d}{d t} \lambda(t) \geq \ \lambda(t) \frac{\int_M \mathcal{H} \varphi^2 \ d\mu }{ \int_M \varphi^2 \ d\mu} + ( 2 \alpha - 1) \frac{\int_M \mathcal{H} | \nabla \varphi |^2 \ d\mu }{\int_M \varphi^2 \ d\mu}.
\end{equation}
By the definition of $\lambda(t)$ or from (\ref{eq21}) we know that $\lambda(t) \int_M \varphi^2 \ d\mu = \int_M | \nabla \varphi |^2 \ d\mu$
and $\lambda(t) > 0$. Suppose further that $\mathcal{H}(x,0) \geq 0$, then,
\begin{equation}\label{eq25}
\frac{d}{d t} \lambda(t) \geq 2 \alpha \lambda(t) \frac{\int_M \mathcal{H} \varphi^2 \ d\mu }{ \int_M \varphi^2 \ d\mu} .
\end{equation}
The monotonicity of $\lambda(t)$ here depends on the sign of $\mathcal{H}$. Note that in applications the sign of $\mathcal{H}$ is usually preserved throughout the evolution. An interesting case is when the manifold is being evolved under the Ricci flow \cite{[Ha82]}, where the nonnegativity of scalar curvature is preserved.
Recall also that $\mathcal{H}$ evolves by
$$\frac{\partial}{ \partial t } \mathcal{H} = \beta + 2 |h_{ij}|^2$$
where $\beta := g^{ij} \partial_t h_{ij}$, in particular, under Ricci flow where $h_{ij} = R_{ij}$ and $\mathcal{H} = R$, we have $\beta = \Delta R$. Here we will assume that
\begin{equation}\label{eqB}
\beta - \Delta \mathcal{H} \geq 0.
\end{equation}
This is motivated by an error term appearing in a result of M\"uller \cite[Lemma 1.6]{[Mu10]}. For our case the error term reads; for any time-dependent vector field $X$ on $M$
\begin{equation}\label{eq224}
\mathcal{D}(X) := 2(R_{ij} -h_{ij})(X, X) + 2\langle \nabla \mathcal{H} - 2 div \ h , X\rangle + \partial_t \mathcal{H} - \Delta \mathcal{H} - 2|h_{ij}|^2,
\end{equation}
where $R_{ij}$ is the Ricci curvature tensor of $M$. Clearly the last three terms in (\ref{eq224}) above is the same as the quantity $ \beta - \Delta \mathcal{H}$. It does make sense to assume (\ref{eqB}) holds whenever $ \mathcal{D}(X)$ is nonnegative. An application of this is that we are on a steady or shrinking soliton (self-similar solution to the geometric flow) if the equality in (\ref{eqB}) holds.
Writing $|h_{ij}|^2 \geq \frac{1}{n} \mathcal{H}^2$ and using the condition that $\beta - \Delta \mathcal{H} \geq 0$, we have a governing differential inequality for the evolution of $\mathcal{H}$ as follows
\begin{equation}\label{eq318}
\frac{\partial}{\partial t} \mathcal{H} \geq \Delta \mathcal{H} + \frac{2}{n} \mathcal{H}^2.
\end{equation}
Suppose $\mathcal{H} \geq \mathcal{H}_{min},$ we can apply the maximum principle by comparing the solution of (\ref{eq318}) with that of the following ordinary differential inequality
\begin{equation}
\left \{ \begin{array}{l}
\displaystyle \frac{d \psi(t) }{d t} = \frac{2}{n} (\psi(t))^2 \\
\displaystyle \ \ \psi(0) = \mathcal{H}_{min}(0),
\end{array} \right.
\end{equation}
solving to
$$ \psi(t) \ \ \ = \ \ \frac{ \mathcal{H}_{min}(0)}{ 1 - \frac{2}{n} \mathcal{H}_{min}(0)t }.$$
Therefore
\begin{equation}\label{Risup}
\mathcal{H}_{g(t)} \ \ \geq \ \ \psi(t) \ \ = \ \ \frac{ \mathcal{H}_{min}(0)}{ 1 - \frac{2}{n} \mathcal{H}_{min}(0)t }
\end{equation}
for all $t \geq 0$ as long as the flow exists. Hence we write (\ref{eq25}) as follows
\begin{equation}\label{eq29}
\frac{d}{d t} \lambda(t) \geq 2 \alpha \psi(t) \lambda(t) .
\end{equation}
By this we can prove the following
\begin{theorem}
Let $(M, g(t))$ be a closed manifold evolving by the geometric flow (\ref{eq11}). Let $\lambda(t)$ be the first nonzero eigenvalue of the Laplacian $\Delta_{g(t)}$. Suppose
$$ \mathcal{H}_{g(t)} \ \ \geq \ \ \psi(t) \ \ = \ \ \frac{ \mathcal{H}_{min}(0)}{ 1 - \frac{2}{n} \mathcal{H}_{min}(0)t }$$
Then
\begin{equation}
\frac{d}{d t} \Big[ \lambda(t) \exp \Big( - 2 \alpha \int_0^T \psi(t) dt \Big) \Big] \geq 0,
\end{equation}
where $\alpha \geq \frac{1}{2}$ and
\begin{equation}
\lambda(t) \geq \lambda(0) e^{ 2 \alpha \int_0^T \psi(t) dt}
\end{equation}
for $0 \leq t \leq T$.
\end{theorem}
In the next we state our results concerning the general $p$. The proofs are discussed in Section \ref{sec4}.
\begin{theorem}\label{thm411}
Let $(M, g)$ be an $n$-dimensional closed Riemannian manifold evolving by the geometric flow (\ref{eq11}). Let $\lambda_{p,1}(t)$ be the first eigenvalue of the $p$-Laplacian on $M$ corresponding to the eigenfunction $u(t, x)$ at time $t \in [0, T]$. Then
\begin{equation}\label{eq48}
\left. \begin{array}{l}
\displaystyle\frac{d}{dt} \lambda_{p,1}(t) = \lambda_{p, 1}(t) \int_M \mathcal{H} | u|^p d\mu - \int_M \mathcal{H} | \nabla u|^p d\mu \\
\displaystyle \hspace{3cm}+ p \int_M | \nabla u|^{p-2} h^{ij} \nabla_i u \nabla_j u d\mu
\end{array} \right.
\end{equation}
for all time $ t \in[0, T]$.
Moreover, if it holds that
\begin{equation}\label{eq49}
h_{ij} - \alpha \mathcal{H} g_{ij} \geq 0 , \ \ \ \ \alpha \in [1/p, 1/n),
\end{equation}
then $ \lambda_{p,1}(t)$ is monotonically nondecreasing along the geometric flow and it is differentiable almost everywhere. Precisely,
\begin{equation}\label{eq410}
\frac{d}{dt} \lambda_{p,1}(t) \geq \lambda_{p, 1}(t) \int_M \mathcal{H} | u|^p d\mu +(\alpha p -1) \int_M \mathcal{H} | \nabla u|^p d\mu \geq 0
\end{equation}
provided $\mathcal{H}$ is nonnegative.
\end{theorem}
\begin{corollary}\label{cor4}
With the conditions of Theorem \ref{thm411}.
\begin{equation}\label{eq514}
\lambda_{p, 1}(t_2) \geq \lambda_{p, 1}(t_1) + \int_{t_1}^{t_2} \Theta(g(t), u(t) ) dt,
\end{equation}
where
$$ \Theta(g(t), u(t) ) = \lambda_{p, 1}(t) \int_M \mathcal{H} | u|^p d\mu +(\alpha p -1) \int_M \mathcal{H} | \nabla u|^p d\mu.$$
Furthermore, if $\mathcal{H} \geq \mathcal{H}_{min} > 0$ and satisfies the governing inequality (\ref{eq318}), it then holds for all time $t_1 < t_2$ that
\begin{equation}\label{eq414}
\lambda_{p, 1}(t_2) \geq \lambda_{p, 1}(t_1) \exp\Big\{ \alpha p \int_{t_1}^{t_2} \psi(t) dt \Big\}.
\end{equation}
\end{corollary}
Finally, we show that the following quantity
\begin{equation}
\lambda_{p,1}(t) \cdot \Big( \psi^{-1}_0 - \frac{2}{n}t \Big)^{\frac{\alpha np}{2}}
\end{equation}
is nondecreasing (see Theorem \ref{thm44}) and that $\lambda_{p,1}(t) $ is differentiable almost everywhere along the geometric flow.
\section{Eigenvalues of the Laplace-Beltrami operator}\label{sec3}
The eigenvalue problem involving Laplace-Beltrami operator on a closed manifold consists in finding all possible real $\lambda$ such that there exists non-trivial functions $u$ satisfyng
\begin{align}
\Delta_g u = - \lambda u.
\end{align}
It is well-known that $- \Delta_g$ has a discrete spectrum on a closed Riemannian manifold. This set consists of an infinite sequence
$$0=\lambda_0<\lambda_1 \leq \lambda_2 \leq ...\lambda_k \leq ... \to \infty \ \ \ \mbox{as} \ k \to \infty$$
and can be found that only constant functions correspond to $\lambda_0=0$. The eigenfunctions are $L^2(M,g)$ orthonormal basis $\{u_0, u_1, u_2,...,u_k,...\}$ of real $C^\infty(M,g)$ function such that
$$\Delta_g u_j = - \lambda_j u_j, \ \ \ j=1,2,...$$
while the eigenvalues are $L^2(M,g)$ orthogonal. By the min-max principle, the first non-zero eigenvalue $\lambda_1(M,g)$ can be characterised as follows
\begin{equation}
\lambda_1 = \inf_{0\neq u \in W_0^{1,2}(M) }\Bigg\{ \frac{\int_M | \nabla u |_g^2 \ d \mu_g}{\int_M | u|^2_g \ d\mu_g} \ \ \Big| \ \ u \neq 0, \ \ u \in W_0^{1, 2}(M,g) \Bigg\},
\end{equation}
satisfying $\int_M u d \mu_g = 0 $, where $ W_0^{1, 2}(M,g)$ is the completion of $C^\infty_0(M,g)$ with respect to the sobolev norm
$$\|u\|_{W^{1,2}} = \Big(\int_M |u|^2 d\mu + \int_M | \nabla u|^2 d\mu \Big)^{\frac{1}{2}}.$$
Note that the $p$-Laplacian $\Delta_{p,g}, \ (1\leq p \leq \infty)$ is a natural generalisation of the Laplace-Beltrami operator.
In this section, we consider the eigenvalues of the Laplace-Beltrami operator under the geometric flow, assuming the least eigenvalue $\lambda = \lambda(t)$ is a function of time only. Next we discuss the proof of Theorem \ref{thm311}.
Let $M$ be a closed Riemannian manifold and $g(t)$ evolve by the generalized geometric flow (\ref{eq11}) in the interval $0\leq t \leq T.$ Let $\varphi(t) = \varphi(t, x)$ be the corresponding eigenfunction to the first nonzero eigenfunction $\lambda(t) = \lambda(t, \varphi)$ of $ \Delta_{2, g(t)}= \Delta$, i.e,
\begin{equation}\label{eq21}
- \Delta \varphi(t, x) = \lambda(t) \varphi(t, x).
\end{equation}
Taking derivative with respect to time, we have
$$ - \Big(\frac{\partial}{\partial t} \Delta \Big) \varphi(t, x) - \Delta \frac{\partial}{\partial t} \varphi (t, x) = \Big(\frac{d}{d t} \lambda(t) \Big) \varphi(t, x) + \lambda(t) \frac{\partial}{\partial t} \varphi(t, x). $$
Multiplying the above by $\varphi(t, x)$ and integrate with respect to the volume measure on $M$, we have
$$ - \int_M \varphi \Big(\frac{\partial}{\partial t} \Delta \Big) \varphi \ d\mu - \int_M \varphi \Delta \frac{\partial}{\partial t} \varphi \ d\mu = \frac{d}{d t} \lambda(t) \int_M \varphi^2 \ d\mu + \lambda(t) \int_M \varphi \frac{\partial}{\partial t} \varphi \ d\mu.$$
Notice that by the application of integration by parts and (\ref{eq21})
$$ - \int_M \varphi \Delta \frac{\partial}{\partial t} \varphi \ d\mu = \lambda(t) \int_M \varphi \frac{\partial}{\partial t} \varphi \ d\mu,$$
then, we arrive at
\begin{equation}\label{eq22}
\frac{d}{d t} \lambda(t) \int_M \varphi^2 \ d\mu = -\int_M \varphi \Big(\frac{\partial}{\partial t} \Delta \Big) \varphi \ d\mu.
\end{equation}
Use the evolution of the Laplacian under the geometric flow, (i.e., (4) of Lemma (\ref{lem211})), so that we have
\begin{align*}
\frac{d}{d t} \lambda(t) \int_M \varphi^2 \ d\mu & = - 2 \int_M h_{ij} \nabla^i \nabla^j \varphi\ \varphi\ d\mu - 2 \int_M \langle div\ h , \nabla \varphi \rangle \varphi \ d\mu \\
\displaystyle & \hspace{2cm} + \int_M \langle \nabla \mathcal{H}, \nabla \varphi \rangle \varphi \ d\mu.
\end{align*}
We express the first and the last terms of the above as follows
\begin{align*}
- 2 \int_M \varphi\ h_{ij} \nabla^i \nabla^j \varphi\ d\mu & = \int_M \nabla^i ( 2 \varphi h_{ij} ) \nabla^j \varphi \ d\mu \\
\displaystyle & = 2 \int_M ( \nabla^i \varphi h_{ij} + \varphi \nabla_i h_{ij} ) \ \nabla^j \varphi \ d\mu \\
\displaystyle & = 2 \int_M h_{ij} \nabla^i \varphi \nabla^j \varphi \ d\mu + 2 \int_M \varphi \langle div\ h, \nabla \varphi \rangle \ d\mu
\end{align*}
and
\begin{align*}
\int_M \langle \nabla \mathcal{H}, \nabla \varphi \rangle \varphi \ d\mu & = - \int _M \mathcal{H} \nabla_i(\varphi \nabla_j \varphi) \ d\mu\\
\displaystyle & = - \int_M \mathcal{H} \nabla^i \varphi \nabla^j \varphi \ d\mu - \int_M \mathcal{H} \varphi \Delta \varphi \ d\mu \\
\displaystyle & = - \int_M \mathcal{H} | \nabla \varphi |^2 \ d\mu + \lambda(t) \int_M \mathcal{H} \varphi^2 \ d\mu.
\end{align*}
Putting these together we have
$$ \frac{d}{d t} \lambda(t) \int_M \varphi^2 \ d\mu = \lambda(t) \int_M \mathcal{H} \varphi^2 \ d\mu - \int_M \mathcal{H} | \nabla \varphi |^2 \ d\mu + 2 \int_M h_{ij} \nabla^i \varphi \nabla^j \varphi \ d\mu. $$
Hence, we have proved Theorem \ref{thm311} on the evolution of $\lambda_1$.
In some applications $\mathcal{H}$ may be required to be a constant or bounded by a constant. Thus when one takes $h_{ij}$ to be Ricci curvature tensor one is talking about a manifold with constant scalar curvature. In this situation we have the following as a corollary.
\begin{corollary}
Let $(M, g(t))$ be a closed manifold evolving by the geometric flow (\ref{eq11}). Let $\lambda(t)$ be the first nonzero eigenvalue of the Laplacian $\Delta_{g(t)}$. Then if $\mathcal{H} \geq C > 0$ in $M \times [t_0, t ]$ for some uniform constant $C$ along the flow we have
\begin{equation*}
\frac{d}{d t} \log \lambda(t) \geq C
\end{equation*}
and
\begin{equation*}
\lambda(t) \geq \lambda(t_0) e^{C(t - t_0)} \ for \ t > t_0.
\end{equation*}
\end{corollary}
\section{Nonlinear eigenvalue problem for $p$-Laplacian}\label{sec4}
In this section, we consider the nonlinear eigenvalue problem
\begin{equation}\label{eq41}
\Delta_{p, g} u = - \lambda |u|^{p-2} u, \hspace{1cm} u \neq 0 \ \ \ on \ M \times [0,T]
\end{equation}
with the normalization condition $\int_M |u|^p\ d\mu =1$.
We want to derive general evolution for the $p$-eigenvalues (eigenvalues of $\Delta_{p,g}$) and show that $\lambda_{p,1}$ is monotone on metrics evolving by the geometric flow. In order to do these we need to calculate time evolution for $\lambda_{p,1}$ and its corresponding eigenfunction. But unfortunately, we do not know whether $\lambda_{p,1}$ or its corresponding eigenfunction $(p\neq2)$ is $C^1$-differentiable or not along the flow. So a similar approach to the one in \cite{[Ca07]} (see also Wu \cite{[Wu11]}, Wu, Wang and Zheng \cite{[WWZ]} and Zhao \cite{[Zha13]}) will be used to avoid this difficulty. Precisely, let $(M, g_{ij}(t)), t \in [0, T]$ be a smooth compact Riemannian manifold evolving by the flow (\ref{eq11}).
Define a genral smooth function as follows
\begin{equation}
\lambda_{p,1}(u(t), t) : = - \int_M u(t) \Delta_p u(t) d\mu_{g(t)} = \int_M |\nabla u(t) |^pd\mu_{g(t)},
\end{equation}
where $u(t)$ is a smooth function satisfying the normalisation condition
\begin{equation}\label{eqdef}
\int_M|u(t)|^p d\mu_{g(t)} = 1, \ \ and \ \ \int_M |u(t)|^{p-2} u(t)d\mu_{g(t)} = 0.
\end{equation}
By this we claim that there exists a smooth function $u(s)$ at time $t=s \in [0,T]$ satisfying (\ref{eqdef}). To see this claim, we first assume that at time $t=s$, $u(s)$ is the eigenfunction corresponding to $\lambda_{p,1}(s)$ of $\Delta_{p.g(s)}$ which implies
\begin{equation*}
\int_M|u(s)|^p d\mu_{g(s)} = 1, \ \ and \ \ \int_M |u(s)|^{p-2} u(s)d\mu_{g(s)} = 0.
\end{equation*}
Then we consider the following smooth function
\begin{equation}\label{eqdef1}
h(t) = u(s) \Bigg( \frac{|g(s)|}{|g(t)|}\Bigg)^{\frac{1}{2(p-2)}}
\end{equation}
under the flow $g(t)$. We normalize this smooth function
\begin{equation}\label{eqdef2}
u(t) = \frac{h(t)}{\Big( \int_M |h(t)|^p d\mu_{g(t)} \Big)^{\frac{1}{p}}}
\end{equation}
under the flow $g(t)$.
By (\ref{eqdef2}) we can easily check that $u(t)$ satisfies (\ref{eqdef}).
Note that in general $\lambda_{p,1}(u,t)$ is not equal to $\lambda_{p,1}(t)$. But at time $t=s$, if $u(s)$ is the eigenfunction of the first eigenvalue $\lambda_{p,1}(s)$, then we conclude that
$$\lambda_{p,1}(u(s),s) = \lambda_{p,1}(s).$$
Notice that the normalisation condition implies
\begin{equation}
\frac{\partial} {\partial t}\Big( \int_M |u|^p d \mu \Big)\Big|_{t=s} = 0,
\end{equation}
which by direct computation (at $t=s$) yields the following
\begin{align*}
\frac{\partial} {\partial t}\Big( \int_M |u|^p d \mu \Big) = 0 &= \frac{\partial}
{\partial t}\Big( \int_M |u|^{p-1} u d \mu \Big)\\
\displaystyle& = \int_M(p-1) |u|^{p-2} u \frac{\partial u} {\partial t} d\mu +
\int_M |u|^{p-1} \frac{\partial } {\partial t} (u d\mu).
\end{align*}
By this it holds that
\begin{equation}\label{eq47}
\int_M |u|^{p-2} u \Big((p-1) \frac{\partial u} {\partial t} d\mu + \frac{\partial }
{\partial t} ( u d\mu) \Big) = 0.
\end{equation}
We now present a proposition
\begin{proposition}
Let $g(t)$ be a smooth solution of the flow (\ref{eq11}) on $M$. Let $\lambda_{p,1}(t)$ be the
first eigenvalue of the $p$-Laplacian under (\ref{eq11}). Assume $u(s)$ is the corresponding eigenfunction of $\lambda_{p,1}(t)$ at time $t=s \in [0, T]$, that is,
\begin{equation*}
\Delta_{p,g(s),p} u(s) = - \lambda_{p,1}(s) |u(s)|^{p-2} u(s).
\end{equation*}
Let $\lambda_{p,1}(u,t)$ be a smooth function defined by
\begin{equation}
\lambda_{p,1}(u(t), t) : = - \int_M u(t) \Delta_p u(t) d\mu_{g(t)}.
\end{equation}
Then we have
\begin{equation}\label{eq5.9}
\left. \begin{array}{l}
\displaystyle\frac{d}{dt} \lambda_{p,1}(u,t)\Big|_{t=s} = \lambda_{p, 1}(s) \int_M S_g | u(s)|^p
d\mu_{g(s)} - \int_M S_g | \nabla u(s)|^p d\mu_{g(s)} \\
\displaystyle \hspace{3cm}+ p \int_M | \nabla u(s)|^{p-2} \mathcal{S}_{ij} u_i u_j d\mu_{g(s)},
\end{array} \right.
\end{equation}
where $u(t)$ is any smooth function satisfying (\ref{eqdef}) such that at any time $ t = s \in[0, T]$, $u(s)$ is the eigenfunctin for $\lambda_{p,1}(s)$. Here $u_i$ denotes covariant derivative of $u$ with respect to the Levi-Civita connection of $g(t)$ at time $t=s$.
\end{proposition}
Since the evolution formula (\ref{eq5.9}) does not depend on the time derivative of the eigenfunction $u$, we have that
$$\frac{d}{dt} \lambda_{p,1}(u(t), t)=\frac{d}{dt}\lambda_{p,1}(t)$$
at some time $t=s$.
We are now set to prove a theorem about the evolution, monotonicity and differentiability (Theorem \ref{thm411}) of the first eigenvalue of the $p$-Laplacian under the geometric flow.
Clearly, we can now set
\begin{equation}
\lambda_{p,1}(t) = \lambda_{p,1}(u(t), t) = - \int_M u(t,x) \Delta_p u(t,x) d\mu_{g(t)}.
\end{equation}
The evolution of $\lambda_{p,1}$ then follows
\begin{equation}
\frac{d}{dt}\lambda_{p,1}(t) = \frac{d}{dt} \lambda_{p,1}(u(t), t) = - \frac{\partial} {\partial t} \int_M u(t,x)
\Delta_p u(t,x) d\mu_{g(t)},
\end{equation}
\subsection*{Proof of Theorem \ref{thm411}}
\begin{proof}
The proof also follows by direct computation using evolution of quantities in Lemmas \ref{lem211} and \ref{lem212}.
Denote $Z:= |\nabla u|^{p-2}$, then working in local orthonormal coordinates we have
\begin{align*}
\frac{\partial} {\partial t} \int_M u \Delta_p u d\mu &= \frac{\partial} {\partial t} \int_M g^{ij} \nabla_i[Z \nabla_j u ] u d\mu \\
\displaystyle& = \frac{\partial} {\partial t} \int_M \Big( g^{ij} \nabla_i Z \nabla_j u + Z \Delta u \Big) u d\mu \\
\displaystyle& = \int_M \frac{\partial} {\partial t} \Big( g^{ij} \nabla_i Z \nabla_j u + Z \Delta u \Big) u d\mu + \int_M \Delta_p u \frac{\partial} {\partial t} (u d\mu)\\
\displaystyle& =: I + II.
\end{align*}
By the evolution of $\Delta_{p,g}$ in Lemma \ref{lem212} we have
\begin{align*}
I & = \int_M 2 h^{ij} \nabla_i ( Z \nabla_j u ) + g^{ij} \nabla_i ( Z_t \nabla_j u ) + g^{ij} \nabla_i (Z \nabla_j u_t ) \\
\displaystyle &\hspace{1cm} + \int_M Z \Big\{2 \langle \ div \ h , \nabla u \rangle - \langle \nabla \mathcal{H} , \nabla u \rangle \Big\} u d\mu.
\end{align*}
Using integration by parts on the second and third terms of the last integral we have
\begin{align*}
I & = \int_M 2 h^{ij} \nabla_i ( Z \nabla_j u ) - \int_M g^{ij} Z_t \nabla_i u \nabla_j u - \int_M Z g^{ij} \nabla_i u \nabla_j u_t ) \\
\displaystyle &\hspace{1cm} + \int_M Z \Big\{2 \langle \ div \ h , \nabla u \rangle - \langle \nabla \mathcal{H} , \nabla u \rangle \Big\} u d\mu.
\end{align*}
Therefore we have after using the evolution $Z_t$ from Lemma \ref{lem212}
\begin{equation}\label{eq411}
\left. \begin{array}{l}
\displaystyle\frac{\partial} {\partial t} \int_M u \Delta_p u d\mu = \int_M 2 h^{ij} \nabla_i ( Z \nabla_j u ) ud\mu -(p-2) \int_M |\nabla u|^{p-2} h^{ij} \nabla_i u \nabla_j u d\mu\\
\displaystyle \hspace{3cm} -(p-1) \int_M |\nabla u|^{p-1} g^{ij} \nabla_i u \nabla_j u_t d\mu \\
\displaystyle \hspace{3cm} + \int_M Z \Big\{2 \langle \ div \ h , \nabla u \rangle - \langle \nabla \mathcal{H} , \nabla u \rangle \Big\} u d\mu \\
\displaystyle \hspace{3cm} + \int_M \Delta_p u \frac{\partial} {\partial t} (u d\mu).
\end{array} \right.
\end{equation}
Computing the first and the third terms on the right hand side (RHS for short) of (\ref{eq411})as follows
\begin{align*}
\int_M 2 h^{ij} \nabla_i ( Z \nabla_j u ) u d\mu &= - 2 \int_M \nabla_i ( h^{ij} u) Z \nabla_j u u d\mu \\
\displaystyle&=- 2 \int_M |\nabla u|^{p-2} h^{ij} \nabla_i u \nabla_j u d\mu - 2 \int_M Z \langle div h, \nabla u \rangle u d\mu.
\end{align*}
\begin{align*}
-(p-1) \int_M |\nabla u|^{p-1} g^{ij} \nabla_i u \nabla_j u_t d\mu &= (p-1) \int_M g^{ij}\nabla_i( |\nabla u|^{p-1} \nabla_i u ) u_t \\
\displaystyle&= (p-1) \int_M \Delta_p u u_t d\mu
\end{align*}
Putting these back into (\ref{eq411}) we have
\begin{align*}
\frac{\partial} {\partial t} \int_M u \Delta_p u d\mu &= -p \int_M |\nabla u|^{p-2} h^{ij} \nabla_i u \nabla_j u d\mu - \int_M Z \langle \nabla \mathcal{H} , \nabla u \rangle u d\mu \\
\displaystyle& \ \ \ + (p-1) \int_M \Delta_p u \ u_t d\mu + \int_M \Delta_p u \frac{\partial} {\partial t} (u d\mu).
\end{align*}
Using the integrability condition (\ref{eq47}) and the definition of $\Delta_p u$ in (\ref{eq41}), the last two terms on the RHS of the above equation vanish and we then arrive at
\begin{align}\label{eq412}
\frac{\partial} {\partial t} \int_M u \Delta_p u d\mu = -p \int_M |\nabla u|^{p-2} h^{ij} \nabla_i u \nabla_j u d\mu - \int_M Z \langle \nabla \mathcal{H} , \nabla u \rangle u d\mu.
\end{align}
The next is to compute the second term on the RHS of the last equality using integration by parts as follows
\begin{align*}
- \int_M Z \langle \nabla \mathcal{H} , \nabla u \rangle u d\mu & = \int_M \mathcal{H} \nabla_i(Z \nabla_j u \ u ) d\mu \\
\displaystyle& = \int_M \mathcal{H} \nabla_i(Z \nabla_j u ) u d\mu + \int_M \mathcal{H} Z \nabla_i u \nabla_j u d\mu \\
\displaystyle& = \int_M \mathcal{H} \Delta_p u\ u d\mu + \int_M \mathcal{H} |\nabla u|^{p-2} | \nabla u|^2 d\mu \\
\displaystyle& = - \lambda_{p, 1}(t) \int_M \mathcal{H} |u|^p d\mu + \int_M \mathcal{H} |\nabla u|^p d\mu.
\end{align*}
Putting this into (\ref{eq412}) we obtain (\ref{eq48}) at once. Hence, the first part of the theorem is proved. Using the condition (\ref{eq49}) in (\ref{eq48}) we have the monotonicity formula (\ref{eq49}) with the condition $\mathcal{H} \geq 0$.
\end{proof}
Clearly, when $p=2$ we have $\Delta_2 = \Delta_g$, the usual Laplace-Beltrami operator. Also $\lambda_{p, 1} = \lambda_1$, the first eigenvalue of $\Delta_g$ and the corresponding eigenfunction are smoothly differentiable. Then Theorem \ref{thm311} reduces to a corollary. This further explains that the $p$-Laplacian is a nonlinear generalisation of Laplace-Beltrami operator.
Integrating both sides of (\ref{eq410}) from $t_1$ to $t_2$ on a sufficiently small time interval $ t_1 \leq t \leq t_2$, $ t \in [0, T]$, we then obtain \ref{eq514}.
\begin{proof} of Corollary \ref{cor4}.
By the definition of $ \lambda_{p, 1}$ in (\ref{eq13}) and the normalization condition $\int_M |u|^p d\mu =1$ we know that
$$ \lambda_{p, 1} = \int_M |\nabla u|^p d\mu.$$
Then (\ref{eq410}) reduces to
\begin{equation}\label{eq415}
\frac{d}{dt} \lambda_{p, 1}(t) \geq \alpha p \psi(t) \lambda_{p, 1}(t)
\end{equation}
from which (\ref{eq414}) follows by integrating on the interval $[t_1, t_2]$ with $ t_1, t_2 \in [0, T].$
\end{proof}
Note that both $\lambda_1(t)$ and $\psi(t)$ are functions of time only.
Setting
$$\psi(0) = \mathcal{H}_{min}(0) = \psi_0,$$
we can evaluate
\begin{align*}
\int_{t_1}^{t_2} \psi(t) dt &= \int_{t_1}^{t_2} \Big( \frac{1}{\psi^{-1}_0 - \frac{2}{n}t} \Big) dt \\
\displaystyle & = - \frac{n}{2} \log ( \psi^{-1}_0 - \frac{2}{n}t) \Big|_{t_1}^{t_2} = \log \Bigg( \frac{ \psi^{-1}_0 - \frac{2}{n}t_1}{ \psi^{-1}_0 - \frac{2}{n}t_2} \Bigg)^{\frac{n}{2}}.
\end{align*}
Therefore integrating both sides of (\ref{eq415}) from $t_1$ to $t_2$ and together with the above equality, yields
\begin{equation}
\log \frac{\lambda_{p,1}(t_2)}{ \lambda_{p,1}(t_1)} \geq \log \Bigg( \frac{ \psi^{-1}_0 - \frac{2}{n}t_1}{ \psi^{-1}_0 - \frac{2}{n}t_2} \Bigg)^{\frac{\alpha np}{2}}
\end{equation}
for any time $t_1 < t_2$ sufficiently close to $t_2$. By this we have
\begin{equation*}
\lambda_{p,1}(t_2) \cdot \Big( \psi^{-1}_0 - \frac{2}{n}t_2 \Big)^{\frac{\alpha np}{2}} \geq \lambda_{p,1}(t_1) \cdot \Big(\psi^{-1}_0 - \frac{2}{n}t_1\Big)^{\frac{\alpha np}{2}}.
\end{equation*}
Then $\lambda_{p,1}(t) \cdot ( \psi^{-1}_0 - \frac{2}{n}t)^{\frac{\alpha np}{2}}$ is nondecreasing along the geometric flow. Hence we conclude this section with the following.
\begin{theorem}\label{thm44}
With the assumption of Theorem \ref{thm411}. The following quantity
\begin{equation*}
\lambda_{p,1}(t) \cdot \Big( \psi^{-1}_0 - \frac{2}{n}t \Big)^{\frac{\alpha np}{2}}
\end{equation*}
is nondecreasing and $\lambda_{p,1}(t) $ is differentiable almost everywhere along the geometric flow (\ref{eq11}).
\end{theorem}
The last theorem has been proved using a different method under the Ricci flow in \cite{[WWZ]}.
\subsection*{The differentiability of $\lambda_{p,1}(t)$}
Since $\lambda_{p,1}(t)$ is nondecreasing on the time interval $[0,T)$ (under curvature assuption of the theorem), by the classical Lebesgue's theorem [A. Mukherjea and K. Pothoven: Real and Functional Analyisis (Chap 4)], it is easy to see that $\lambda_{p,1}(t)$ is differentiable almost everywhere,
\begin{remark}
Our proofs of the first eigenvalue evolution and monotonicity do not use any differentiability of the first eigenvalue $\lambda_{p,1}(t)$ or its corresponding eigenfunction $u(t,x)$ of the $p$-Laplacian under $(RH)_\alpha$-flow. In fact, it is not known whether they are differentiable in advance.
It would be interesting to find out whether the corresponding first eigenfunction of the $p$-Laplacian is a $C^1$-differentiable function with respect to $t$-variable along $(RH)_\alpha$-flow.
\end{remark}
\section{Examples of geometric flows}
In this section, we give some examples of geometric flows where our results are valid. We remark that in these cases the error term $\mathcal{D}(X)$ and the quantity $\beta - \Delta \mathcal{H}$ are nonnegative. More examples can be found in \cite[Section 2]{[Mu10]}.
\subsection{Hamilton's Ricci flow \cite{[Ha82]}}
Let $(M, g(t))$ be a solution to the Hamilton's Ricci flow
\begin{equation}
\partial_t g_{ij}(t, x) = - 2 R_{ij}.
\end{equation}
This is the case where $h_{ij} =R_{ij}$ is the Ricci tensor and $\mathcal{H} =R$ is the scalar curvature on $M$. Here, the scalar curvature evolves by
$$\partial_t R = \Delta R + 2 |R_{ij}|^2.$$
By twice contracted second Bianchi identity $g^{ij} \nabla_i R_{jk} = \frac{1}{2} \nabla_k R$, which implies
$$ 2 \langle \ div \ h , \nabla f \rangle - \langle \nabla \mathcal{H} , \nabla f \rangle = 0, $$
the quantity $\mathcal{D}(X)$ vanishes identically and $\beta -\Delta R \equiv 0$. Note that the positivity of curvature is preserved along the Ricci flow \cite{[Ha82]}. The evolution equation and monotonicity formula for the first eigenvalue follow easily (see \cite{[Ca07]}-\cite{[GPT]} and \cite{[Li07a]}-\cite{[Ma06]}). A fundamental result here is Perelman's paper \cite{[Pe02]}, where he defines his $\mathcal{F}$-energy
\begin{equation}
\mathcal{F}(g_{ij}(t), u(t)) = \int_{M} (4 |\nabla u|^2 + R u^2) d \mu \ \ \ \ \ \ with \ \ \ \int_M u^2 d \mu = 1.
\end{equation}
and proves that it is monotonically nondecreasing. He also defines
\begin{equation}
\lambda_1 ( g_{ij}) = \inf \Big\{ \mathcal{F} (g_{ij}, f ) : f \in C_c^\infty (M) , \int_M e^{ - f} d \mu = 1 \Big\},
\end{equation}
with $ \lambda_1 ( g_{ij})$ (being the least eigenvalue of the geometric operator $- 4\Delta + R$) and its corresponding eigenfunction $u=e^{ - f} $ satisfying the eigenvalue problem
$$ - 4 \Delta u + R u = \lambda_1 ( g_{ij}) u.$$
He shows that monotonicity of $ \lambda_1 ( g_{ij})$ follows from that of $\mathcal{F}$.
\subsection{Ricci-harmonic map flow \cite{[Mu12]} }
Let $(M, g)$ and $(N, \xi)$ be compact (without boundary) Riemannian manifolds of dimensions $m$ and $n$ respectively.
Let a smooth map $ \varphi: M \rightarrow N$ be a critical point of the Dirichlet energy integral $E( \varphi) = \int_M | \nabla \varphi|^2 d \mu_g$, where $N$ is isometrically embedded in $\mathbb{R}^d, \ d \geq n,$ by the Nash embedding theorem. The configuration $(g(x ,t), \varphi(x, t)), t \in [0, T)$ of a one parameter family of Riemannian metrics $g(x, t)$ and a family of smooth maps $ \varphi(x, t)$ is defined to be Ricci-harmonic map flow if it satisfies the coupled system of nonlinear parabolic equations
\begin{equation}\label{eq514}
\left \{ \begin{array}{l}
\displaystyle \frac{\partial}{\partial t} g(x ,t) = - 2 Rc(x ,t) + 2 \alpha \nabla \varphi(x ,t) \otimes \nabla \varphi(x ,t)
\\ \ \\
\displaystyle \frac{\partial}{\partial t} \varphi(x ,t) = \tau_g \varphi(x ,t),
\end{array} \right.
\end{equation}
where $ Rc(x ,t)$ is the Ricci curvature tensor for the metric $g$, $\alpha(t) \equiv \alpha > 0$ is a time-dependent coupling constant, $\tau_g \varphi$ is the intrinsic Laplacian of $ \varphi$, which denotes the tension field of map $ \varphi$ and $\nabla \varphi \otimes \nabla \varphi = \varphi^* \xi$ is the pullback of the metric $\xi$ on $N$ via the map $\varphi$. See List \cite{[Li08]} when the target manifold is one dimensional.
Here $h_{ij} = R_{ij} - \alpha \partial_i \varphi \partial_j \varphi =:S_{ij}$, \ $\mathcal{H}= R - \alpha |\nabla \varphi|^2 =: S$ and
\begin{equation}
\partial_t S = \Delta S + 2|S_{ij}|^2 + 2 \alpha |\tau_g \varphi|^2 - 2 \dot{\alpha} |\nabla \varphi|^2.
\end{equation}
Using the twice contracted second Bianchi identity, we have
\begin{equation}
(g^{ij} \nabla_i S_{jk} - \frac{1}{2} \nabla_k S) X_j = - \alpha \tau_g \varphi \nabla_j\varphi X_j.
\end{equation}
A straightforward computation gives
\begin{equation}\label{eq44}
2 (Rc - \mathcal{S}) (\nabla u, \nabla u) = 2 \alpha | \nabla \varphi|^2 | \nabla u |^2 .
\end{equation}
Then, $ \mathcal{D}(X) = 2 \alpha | \tau_g \varphi - \nabla_X \varphi|^2 - 2 \dot{\alpha} |\nabla \varphi|^2 $ and $\beta - \Delta S = 2 \alpha | \tau_g \varphi|^2 - \dot{\alpha} |\nabla \varphi|^2 $ for all $X$ on $M$. Thus both $\mathcal{D}$ and $\beta - \Delta S$ are nonnegative as long as $\alpha(t)$ is nonincreasing in time. The first author has considered this in \cite{[Ab15]}. See also \cite{[Li10]} for the monotonicity of the first eigenvalue of Laplace-Betrami operator and versions of Perelman's entropy under the Ricci-harmonic map flow.
\subsection{Lorentzian mean curvature flow}
Let $M^n(t) \subset L^{n+1}$ be a family of space-like hypersurfaces in ambient
Lorentzian manifold evolving by Lorentzian mean curvature flow
$$\partial_t F(t, \cdot) = \Pi(t, \cdot) \nu(t, \cdot)$$
for $(t, \cdot) \in [0,T] \times M$, where $F(t, \cdot)$ is the position of $M^n$ in $L^{n+1}$ satisfying $F(0,\cdot)=F_0(\cdot)$. Here $\nu(t, \cdot)$ and $\Pi(t, \cdot)$
are respectively the outer normal vector and mean curvature at the point $F(t, \cdot).$ Then, the induced metric evolves by
$$\partial_tg_{ij} = 2 H \Pi_{ij}, $$
where $\Pi_{ij}$ denotes the components of the second fundamental form $\Pi$ on $M$ and $H = g^{ij} \Pi_{ij}$ denotes the mean curvature of $M$. In this case
$h_{ij} = - H \Pi_{ij}$ and $\mathcal{H} = - H^2$. Letting $ \widetilde{Rc}$ and $\widetilde{Rm}$ denote the Ricci and Riemman curvature tensor of $L^{n+1}$ respectively, we have the Gauss equation
$$ R_{ij} = \widetilde{R}_{ij} - H \Pi_{ij} + \Pi_{il} \Pi_{lj} + \widetilde{R}_{i0j0},$$
the Codazzzi equation
$$\nabla_i \Pi_{jk} - \nabla_k \Pi_{ij} = \widetilde{R}_{0jki},$$
the evolution equation
$$\partial_tH = \Delta H - H (|\Pi|^2 + \widetilde{Rc}(\nu, \nu)$$
and
$$\beta - \Delta H = 2 H^2 |\Pi|^2 + |\nabla H|^2 + 2 H \widetilde{Rc}(\nu, \nu).$$
See the explicit forms of the Gauss and the Codazzi equations
for $L^{n+1} = \mathbb{R}^{n+1}$ in \cite{[Hu84]}. Combining the above equation we obtain the quantity
\begin{equation}
\mathcal{D}(X) = 2|\nabla H - \Pi(X, \cdot)|^2 + 2 \widetilde{Rc}(H \nu - X, H \nu - X) + 2 \langle \widetilde{Rm}(X, \nu)\nu, X \rangle,
\end{equation}
where $\nu$ denotes future-oriented timelike unit normal vector on $M$. Obviously both $\mathcal{D}(X)$ and $\beta -\Delta H$ are nonnegative when assuming nonnegativity on sectional curvature of $L^{n+1}$. See \cite{[Zha13]} for the evolution and monotonicity of the first eigenvalue of $p$-Laplace operator under the $m^{th}$ powers of the mean curvature flow. See also Huisken \cite{[Hu84]}, the second author's paper \cite{[Mao1]} and \cite{[GLW16],[Zha12]} for related results.
\subsection{The Yamabe flow}
This is the case when $h_{ij} = \frac{1}{2}Rg_{ij}$, where $R$ is the scalar curvature of the metric. Yamabe flow is then the following evolution equation
\begin{equation}\label{eq518}
\left. \begin{array}{l}
\displaystyle \frac{\partial}{\partial t} g_{ij}(x ,t) = - R(x, t) g_{ij}(x ,t), \ \ (x, t) \in M \times [0, T]
\\ \ \\
g_{ij}(x ,0) = g_0(x)
\end{array} \right.
\end{equation}
as introduced by R. Hamilton who first establishes the existence of its unique solution for all time and shows that the metric $g(t)$ approaches constant as $t \rightarrow \infty$. His proof is done for volume preserving flow
\begin{equation}
\frac{\partial}{\partial t} g_{ij}(x ,t) = (r(t)- R(x, t)) g_{ij}(x ,t), \ \ (x, t) \in M \times [0, \infty),
\end{equation}
with $r(t) = Vol^{-1}(g(t))\int_M R d\mu$ is the average of scalar curvature for the metric in a conformal class. For more on the global existence and convergence of (\ref{eq518}) see Chow \cite{[Ch92]} and Ye \cite{[Ye94]}. Note that under the Yamabe flow the volume measure evolves as $\partial_t d\mu = n/2R d\mu$ and the normalization condition, $\partial_t(\int_M |u|^p d\mu) = 0$, implies
$$ \int_M p |u|^p u u_t d\mu - \frac{n}{2} \int_M R u d\mu = 0.$$
Here, the evolution of scalar curvature is given \cite{[Ch92]} as
\begin{equation}
\frac{\partial}{\partial t} R = (n-1) \Delta R + R^2
\end{equation}
and by the strong maximum principle
$$R(x, t) \geq \psi(t) = \frac{R_{min}(0)}{1 - R_{min}(0) t} $$
for all t.
We can also compute
\begin{align*}
\langle 2 div\ h - \nabla \mathcal{H}, \nabla f \rangle &= \langle 2 g^{ij} \nabla_i( \frac{1}{2} Rg_{ij}) - \frac{n}{2} \nabla_k R, \nabla f \rangle \\
\displaystyle & = \frac{2n-n}{2} \langle \nabla R, \nabla f \rangle
\end{align*}
and
$$\beta - \Delta \mathcal{H} = \frac{n(n-2)}{2} \Delta R.$$
These imply that the quantitty $\mathcal{D}(\nabla f)$ is nonnegative on the Einstein tensor
$$E_{ij} = R_{ij} - \frac{1}{2} R_{ij} \geq 0.$$
Hence our results hold. See \cite[Section 7]{[WWZ]}, where with assumption that $p \geq n$ and $R \geq 0$, they prove that $\lambda_{p, 1}$ is strictly increasing and differentiable almost everywhere along the Yamabe flow. Yamabe flow coincides with the Ricci flow on Riemann surfaces.
\subsection*{Acknowledgements}
The first author was partially supported by Nigeria Tetfund research grant through OSCOTECH, Esa-Oke. The second author was partially supported by the NSF of China (Grant No. 11401131). The first version of the paper was completed during the first author's research visit to African Institute for Mathematical Sciences (AIMS), Senegal in February 2016. He therefore thanks AIMS-Senegal for finacial suport during the visit. He also wishes to thank the Research chair, Prof. Moustapha Fall for the hospitality and useful discussions which helped in improving this paper.
|
\section{Introduction}
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{groJ1008-57_gamma_uncertainties}
\caption{Distribution of the relative uncertainties (90\% confidence level)
of the power-law photon indices, $\Gamma$, in our study of all
\textsl{RXTE}-observations of GRO~J1008$-$57 \citep{kuehnel2013a}. Fitting
all 43 observations separately results in the green histogram. As soon as we
perform a simultaneous fit with $\mathcal{P}=1$ global continuum parameter
the uncertainties decrease significantly as shown by the red histogram.
Finally, using $\mathcal{P}=2$ global parameters (blue histogram) results in
a median of ${\sim}6\%$ in the uncertainties (compare the arrows on top).}
\label{fig:1008gamma}
\end{figure}
Nowadays, the still increasing computation speed and available memory allows
us to analyze large datasets at the same time. Using X-ray spectra of
accreting neutron stars as an example, we have shown in a previous paper
\citep[][hereafter paper I]{kuehnel2015a} that loading and fitting the spectra
simultaneously has several advantages compared to the ``classical'' way of
X-ray data analysis, which is treating every observation individually. In
particular, instead of fixing parameters to a mean value one can determine
them by a joint fit to all datasets under consideration. Due to the
reduced number of degrees of freedom the remaining parameters can be better
constrained (see Fig.~\ref{fig:1008gamma} as an example). Furthermore,
parameters no longer need to be independent, but can be combined into
functions. For instance, the slope of the spectra might be described as a
function of flux with the coefficients of this function as fit-parameters.
The disadvantages of fitting many datasets simultaneously are, however, an
increased runtime and a complex handling because of the large number of
parameters. In paper I, we have introduced functions to facilitate this
handling, which have been implemented into the \texttt{Interactive Spectral
Interpretation System} (ISIS) \citep{houck2000a}. While these functions are
already available as part of the
\texttt{ISISscripts}\footnote{http://www.sternwarte.uni-erlangen.de/isis} they
are continuously updated and new features are implemented. One important
question, which we raised in paper I, is about the goodness of a
simultaneous fit as it is, e.g., calculated after the commonly used
$\chi^2$-statistics, particularly the case where some datasets are
not described well by the chosen model. Due to the potential large total
number of datasets, the information about failed fits can be
hidden in the applied fit-statistics. After we have given a reminder
about the terminology of simultaneous fits in Section~\ref{sec:reminder},
we describe the problem of detecting failed fits in more detail in
Section~\ref{sec:solution} and provide possible solutions. We will
conclude this paper by applying these solutions to examples in
Section~\ref{sec:examples}.
\subsection{Simultaneous Fits}
\label{sec:reminder}
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{terminology}
\caption{Terminology of simultaneous fits in ISIS, according to paper I. A
data-group consists of simultaneously taken datasets, here data-group A has $n$
and B has $m$ datasets. A model with $p+\mathcal{P}$ parameters is fitted
such that each data-group has its own set of $p_i$ parameters, called group
parameters. The common $\mathcal{P}$ parameters between the groups are the
so-called global parameters.}
\label{fig:terminology}
\end{figure}
As we have described in paper I, a \textit{data-group} contains all datasets
which have been taken simultaneously in time or, in general, represent the
same state of the observed object. In the example illustrated in
Fig.~\ref{fig:terminology} two data-groups, A and B, have been added to the
simultaneous fit, containing $n$ and $m$ datasets, respectively. Thus, a
dataset is labeled by the data-group it belongs to, e.g., B3 is the third
dataset in the second data-group. After a model with $p$ parameters has been
defined each of these data-groups is fitted by an individual set of parameters,
called \textit{group parameters}. Consequently, all datasets belonging to a
specific data-group are described by the same parameter values. A specific group
parameter can now be marked as a so-called \textit{global parameter}. The
value of the corresponding group parameters will now be tied to this global
parameter, i.e, this parameter has a common value among all data-groups.
Instead of tying group parameters together to a global value, a parameter
function may be defined, to which the group parameters are set instead. This
function takes, e.g., other parameters as input to calculate the value for
each group parameter. In this case, correlations between model parameters,
e.g, as predicted by theory can be implemented and fitted directly.
\section{Goodness of a simultaneous fit}
\label{sec:solution}
As an indicator for the goodness of a fit the analysis software used,
e.g ISIS or XSPEC \citep{arnaud1996a}, usually displays the fit-statistics
after the model has been fitted to the data. Here, we chose the
$\chi^2$-statistics since the developed functions for a simultaneous fit have
been first applied to accreting neutron stars. The high count rates satisfies
the Gaussian approxmation of the uncertainties, which are actually Poisson
distributed. In principle, however, the discussed issues and their solutions
can be generalized for any kind of fit-statistics.
For each datapoint $k$ the difference between the $\mathrm{data}$ and the
$\mathrm{model}$ is calculated and normalized by the measurement uncertainty,
$\mathrm{error}$, of the data. The sum over all $n$ datapoints is called the
$\chi$-square,
\begin{equation}\label{eq:chisqr}
\chi^2 = \sum_{k=1}^n \frac{(\mathrm{data}_k -
\mathrm{model}_k)^2}{\mathrm{error}_k^2}
\end{equation}
and is displayed after a fit. Additionally, the sum is normalized to the
total number of degrees of freedom, $n-p$ with the number of free
fit-parameters $p$, since the $\chi^2$ increases with $n$. This normalized sum,
called the reduced $\chi$-square,
\begin{equation}\label{eq:redchisqr}
\chi^2_\mathrm{red} = \frac{\chi^2}{n-p}
\end{equation}
is also displayed. For Gaussian distributed data the expected value
is $\chi^2_\mathrm{red}=1$ for a perfect fit of the chosen model. However,
once the probability distribution is changed, e.g., when a
spectrum has been rebinned, the expected value changes as well. Consequently,
a reliable measure for the goodness of the fit has to be defined with some
forethought.
The $\chi^2_\mathrm{red}$ threshold, for which a simultaneous fit is
acceptable, strongly depends on the considered case. In particular, a few
data-groups might not be described well by the chosen model, which would
result in an unacceptable $\chi^2_\mathrm{red}$ when fitted individually.
However, in case of a simultaneous fit, this information might be hidden in
the classical definition of the $\chi^2_\mathrm{red}$ (Eq.~\ref{eq:redchisqr}).
Let us consider $N$ data-groups and a model with $p$
group parameters and $\mathcal{P}$ global parameters. Then, the total
$\chi^2_\mathrm{red}$ is
\begin{equation}\label{eq:redchisqr:simfit}
\chi^2_\mathrm{red} = \frac{\sum_{i=1}^N \chi^2_i}{\sum_{i=1}^N (n_i -
p_i) - \mathcal{P}}
\end{equation}
with the number of degrees of freedom, $n_i - p_i$, and the $\chi^2_i$ for
each data-group $i$ after Eq.~\ref{eq:chisqr}. Now, we assume a failed fit
with $\chi^2_i \sim 2$ for a particular $i$ to be present, while for the
remaining data-groups $\chi^2_i \sim 1$. For $N \gtrsim 10$ the
$\chi^2_\mathrm{red}$ after Eq.~\ref{eq:redchisqr:simfit} is still near unity
and, thus, suggests a successful simultaneous fit. In the
following, we present three possibilities to investigate the
goodness of a simultaneous fit more carefully.
\subsection{Histogram of the goodness}
\label{sec:histogram}
A trivial but effective solution is to check the goodness of the fit for
each data-group individually. Here, in the chosen case of the
$\chi^2$-statistics, the $\chi^2_\mathrm{red,i}$ is calculated for each
data-group, $i$, after
\begin{equation}\label{eq:groupchi}
\chi^2_\mathrm{red,i} = \frac{\chi^2_i}{n_i - p_i}
\end{equation}
where $n_i$ are the number of datapoints in the data-group, and $p_i$ is the
number of free group parameters. Due to that the global parameters are
not taken into account here, the $\chi^2_\mathrm{red,i}$ is, however,
different to that performed by a single fit of the data-group.
In the case of a large number of data-groups, it is more convenient to sort the
$\chi^2_\mathrm{red,i}$ into a histogram to help investigating the goodness
of the fit to all data-groups. We have added such a histogram to the
simultaneous fit functions as part of the \texttt{ISISscripts}. After a fit
has been performed using the fit-functions \texttt{fit\_groups} or
\texttt{fit\_global} (see paper I) this histogram is added to the default
output of the fit-statistics. In this way, failed fits of specific data-groups
can be identified by the user at first glance.
\subsection{A combined goodness of the fit}
\label{sec:newstatistic}
Instead of a few failed fits to certain data-groups, one might ask if the
chosen model fails in the global context of a simultaneous fit. To
answer this question, a special goodness of the simultaneous fit is needed to
take its logical structure into account. As explained in
Section~\ref{sec:reminder}, a data-group represents a certain state of the
observed object, e.g., the datasets where taken at the same time. Thus, the
data-groups are statistically independent of each other. Calculating the
goodness of the fit in a traditional way, which is the $\chi^2_\mathrm{red}$
after Eq.~\ref{eq:redchisqr:simfit} in our case, does not, however, take
this aspect into account. As a solution we propose to define a
\textit{combined goodness of the fit} calculating the weighted mean of the
individual goodness of each data-group. In the case of $\chi^2$-statistics, a
combined reduced $\chi^2$ is calculated by
\begin{equation}\label{eq:combredchisqr}
\chi^2_\mathrm{red,comb.} = \frac{1}{N} \sum_{i=1}^N \frac{\chi_i^2}{n_i
- p_i - \mu_i \mathcal{P}}
\end{equation}
with $\chi_i^2$ computed after Eq.~\ref{eq:chisqr} for each data-group, $i$,
and a weighting factor, $\mu_i$, for the number of global parameters,
$\mathcal{P}$:
\begin{equation}\label{eq:mu}
\mu_i \approx (n_i - p_i) \times \sum_{j = 1}^N \frac{1}{n_j - p_j}
\end{equation}
Thus, $\mu_i$ normalizes the effect of data-group $i$ on the determination of
the global parameters, $\mathcal{P}$, by its number of degrees of freedom
relative to the total number of degrees of freedom of the simultaneous fit.
Equation~\ref{eq:mu} is, however, an approximation only. A data-group might not
be sensitive to a certain global parameter, e.g, if the spectra in this
data-group do not cover the energy range necessary to determine the parameter.
A failed fit to a specific data-group, for example with a high individual
$\chi^2_\mathrm{red,i}$, has a higher impact on the $\chi^2_\mathrm{red,comb.}$
(Eq.~\ref{eq:combredchisqr}) than on the traditional
$\chi^2_\mathrm{red}$ (Eq.~\ref{eq:redchisqr}). In general we expect
$\chi^2_\mathrm{red,comb.} \geq \chi^2_\mathrm{red}$, even if all data-groups
are fitted well. In the case of a good simultaneous fit (better than a certain
threshold), a weak feature in the data might still be unnoticed, if it is not
detected in any individual data-group. Such a feature can be investigated by
stacking the residuals, as outlined in the following section.
We note, however, that Eq.~\ref{eq:combredchisqr} is the result of an
empirical study. A more sophisticated goodness of a simultaneous fit
should be based on a different type of fit-statistics suitable for a
simultaneous analysis of many datasets, such as a Bayesian approach or a
joint likelihood formalism similar to \citet{anderson2015a}.
\subsection{Stacked residuals}
\label{sec:summedresiduals}
Once datasets can be technically stacked to achieve a higher
signal-to-noise ratio, e.g., when spectra have the same energy grid and
channel binning, further weak features might get visible. This is a common
technique in astrophysics \citep[see, e.g.,][]{ricci2001a,bulbul2014a}.
However, when stacked datasets are analyzed, differences in the individual
datasets, like source intrinsic variability, can no longer be revealed.
In case of a simultaneous fit, the residuals of all data-groups can be
stacked instead. The stacking dramatically increases the total exposure in
each channel bin. Thus, the stacked residuals of all data-groups,
$R(k)$, as a function of the energy bin, $k$, can be investigated to further
verify the goodness of the simultaneous fit
\begin{equation}\label{eq:sumresiduals}
R(k) = \sum_{i=1}^N \mathrm{data}_{i,k} - \mathrm{model}_{i,k}
\end{equation}
This task can be achieved using, e.g, the \texttt{plot\_data}
function\footnote{\url{http://space.mit.edu/home/mnowak/isis\_vs\_xspec/plots.html},
which is available through the \texttt{ISISscripts} as well.} written by
M.~A.~Nowak.
We can show that the combined reduced $\chi^2$ is effectively equal
to the goodness of a fit of the stacked data in the first place. Assuming the
same number of degrees of freedom, $n-p$, for each data-group,
Eq.~\ref{eq:combredchisqr} gives
\begin{equation}\label{eq:chisqrequal}
\chi^2_\mathrm{red,comb.} = \frac{1}{f} \sum_{i=1}^N
\sum_{k=1}^{n} \frac{(\mathrm{data}_{i,k} -
\mathrm{model}_{i,k})^2}{\mathrm{error}_{i,k}^2}
\end{equation}
with $f = N (n-p-\mu\mathcal{P})$ and having used Eq.~\ref{eq:chisqr}. Now,
the summand no longer depends on $i$ or $k$ explicitly. Thus, the order of the
sums in Eq.~\ref{eq:chisqrequal} may be switched. If we finally interpret $k$
as a spectral energy bin we end up with the goodness as a function of $k$:
\begin{align}
\begin{split}
\chi^2_\mathrm{red,comb.}(k) &\propto \sum_{i=1}^N
\frac{(\mathrm{data}_{i,k} -
\mathrm{model}_{i,k})^2}{\mathrm{error}_{i,k}^2} \\
&\propto \sum_{i=1}^N \mathrm{data}_{i,k}^2
\end{split}
\end{align}
This means that all datasets of the simultaneous fit are first summed up for
each energy bin in the combined reduced $\chi^2$. In contrast to
stacking the data in the first place, however, source variability can still
be taken into account during a simultaneous fit.
Note that once all data-groups have the same number of degrees of
freedom, the $\chi^2_\mathrm{red,comb.}$ (Eq.~\ref{eq:chisqrequal}) is equal
to the classical $\chi^2_\mathrm{red}$ (Eq.~\ref{eq:redchisqr}). To further
investigate the goodness of the simultaneous fit in such a case, the histogram
of the goodness of all data-groups (see Sec.~\ref{sec:histogram}) and, if
possible, the stacked residuals should be investigated.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{groJ1008-57_combined_xte_residuals}
\caption{The stacked spectra (a) and stacked residuals of all
individual data-groups (b) containing 43 \textsl{RXTE}-spectra of
GRO~J1008$-$57 (blue: PCA; red: HEXTE). Residual features are left in PCA,
which are caused by calibration uncertainties. These features are not
detected in that detail in the residuals of the single spectrum with the
highest signal (c).} \label{fig:1008}
\end{figure}
\section{Examples}
\label{sec:examples}
\subsection{GRO J1008$-$57}
\label{sec:example:gro1008}
The Be X-ray binary GRO J1008$-$57 was regularly monitored by \textsl{RXTE}
during outbursts in 2007 December, 2005 February, and 2011 April with a few
additional pointings by \textsl{Suzaku} and \textsl{Swift}. A detailed
analysis of the spectra has been published in \citet{kuehnel2013a} and in
paper I we have demonstrated, as an example, the advantages of a simultaneous
fit based on these data (see also Fig.~\ref{fig:1008gamma}). The
$\chi^2_\mathrm{red}$ of 1.10 with 3651 degrees of freedom (see Table~4
of K\"uhnel at al., 2013 \citet{kuehnel2013a}) calculated after Eq.~\ref{eq:redchisqr:simfit} indicates a
good fit of the underlying model to the data. Using the combined reduced
$\chi^2$ defined in Eq.~\ref{eq:combredchisqr} we find, however,
$\chi^2_\mathrm{red,comb.} = 1.68$. The reason for this significant worsening
of the goodness are calibration uncertainties in \textsl{RXTE}-PCA, which
are visible in the stacked residuals of all 43 data-groups as shown in
Fig.~\ref{fig:1008}: the strong residuals below 7\,keV are probably caused by
insufficient modeling of the Xe L-edges, the absorption feature at 10\,keV by
the Be/Cu collimator, and the sharp features around 30\,keV by the Xe K-edge
\citep[for a description of the PCA see][]{jahoda2006a}. These calibration
issues have been detected in a combined analysis of the Crab pulsar by
\citet{garcia2014a} as well. However, the calibration issues, which
are responsible for the high $\chi^2_\mathrm{red,comb.}$, do not affect the
continuum model of GRO~J1008$-$57 because of their low significance in the
individual data-groups. These calibrations features might have an influence,
however, in data with a much higher signal-to-noise ratio than the datasets
used here or once narrow features, such as emission lines, are studied.
\subsection{Vela X-1}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{velaX-1_XMM_iron_complex}
\caption{The iron line region of Vela X-1 can be nicely studied in this
stacked residuals of all 88 \textsl{XMM-Newton}-spectra (a). The model
includes the continuum shape only and, thus, does not take any fluorescent
line emission into account. The residuals of the single spectrum with the
highest signal show the K$\alpha$ emission line only (b). Note that
the residual flux in this line is $\sim$15 times lower compared to the
stacked residuals.}
\label{fig:vela}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{velaX-1_XMM_redchisqr}
\caption{Example for a histogram of the goodness of the fits. Here, the
distribution of the $\chi^2_\mathrm{red}$ of all individual data-groups
of the 88 \textsl{XMM-Newton}-spectra of Vela~X-1 are shown.}
\label{fig:velachi}
\end{figure}
Another excellent example for a simultaneous fit was performed by
\citet{martineznunez2014a}. These authors have analyzed 88 spectra recorded by
\textsl{XMM-Newton} during a giant flare of Vela~X-1. Although the continuum
parameters were changing dramatically within the ${\sim}100$\,ks observation a
single model consisting of three absorbed power-laws is able to describe the
data with a $\chi^2_\mathrm{red} = 1.43$ with 9765 degrees of freedom
\citep{martineznunez2014a}. Due to a global photon index for all power-laws
and data-groups the absorption column densities and iron line fluxes could be
constrained well.
Because every data-group is a single spectrum taken by the \textsl{XMM}-EPIC-PN
camera and a common energy grid was used, the $\chi^2_\mathrm{red,comb.}$
equals the $\chi^2_\mathrm{red}$ here. Thus, it is preferred to calculate the
stacked residuals of all data-groups according to Eq.~\ref{eq:sumresiduals}.
To demonstrate the advantage of this tool we have used the continuum model
only, i.e. without any fluorescence line taken into account, and evaluated
this model without any channel grouping to achieve the highest possible energy
resolution. The resulting stacked residuals in the iron line region
(5--9\,keV) are shown in Fig.~\ref{fig:vela}. The iron K$\alpha$ line at
${\sim}6.4$\,keV and the K$\beta$ line at ${\sim}7.1$\,keV are nicely
resolved. The tail following the K$\beta$ line is probably caused by a slight
mismatch of the continuum model with the data and requires a more detailed
analysis. Note that the flux of this mismatch is a few
$10^{-3}$\,photons\,s$^{-1}$\,cm$^{-2}$, which is detectable only in these
stacked residuals featuring 100\,ks exposure time and after the strong
continuum variability on ks-timescales has been subtracted.
As a further demonstrative example, the histogram of the goodness of the
fits of the 88 data-groups calculated after Eq.~\ref{eq:groupchi} is shown
in Fig.~\ref{fig:velachi}. The median $\chi^2_\mathrm{red}$ is around 1.3,
indicating that the model still could be improved slightly. Investigating
the three outliers with $\chi^2_\mathrm{red} > 2$ indeed proves that the
residuals around in the iron line region are left, which are responsible
for the high $\chi^2_\mathrm{red}$ and very similar to those shown in
Fig.~\ref{fig:vela}. There are no, however, extended residuals visible.
Thus, the continuum parameters presented in \citet{martineznunez2014a} are
still valid.
\subsection{XTE J1859+083}
\begin{figure}
\centering
\includegraphics[width=.985\linewidth]{xteJ1859p083_combined_xrt_residuals}
\caption{Seven stacked spectra of \textsl{Swift}-observations of
XTE~J1859+083 (a) and the residuals to the model (b). A weak
iron K$\alpha$ emission line at 6.4\,keV is visible, which is not detected
in the residuals of any individual spectrum (c).}
\label{fig:xte}
\end{figure}
The last example shown in this work is the outburst of the transient pulsar
XTE~1859+083 in 2015 April. This source was in quiescence since its bright
outburst in 1996/1997 \citep{corbet2009a}. During the recent outburst several
short observations by \textsl{Swift} were performed. A first analysis of these
data in combination with \textsl{INTEGRAL} spectra reports an absorbed
power-law shape of the source's X-ray continuum \citep{malyshev2015a}. We have
extracted and analyzed seven \textsl{Swift}-XRT spectra and can confirm these
findings. However, after examining the stacked residuals of all spectra an
iron K$\alpha$ emission line at 6.4\,keV shows up that has not been detected
before (see Fig.~\ref{fig:xte}). We define the equivalent width of this line
as a global parameter and find a value of $60 \pm 40$\,eV (uncertainty is at
the 90\% confidence level, $\chi^2_\mathrm{red,comb.} = 0.98$ with 585 degrees
of freedom).
\section{Summary}
We have continued developing functions to handle simultaneous fits in ISIS,
which we have introduced in paper~I. In particular, we have concentrated on
tools for checking the goodness of the fits to discover failed fits
of individual data-groups or global discrepancies of the model. We
propose to
\begin{itemize}
\item investigate the distribution of the goodness of fits to all
individual data-groups
\item calculate a combined goodness, here the
$\bf \chi^2_\mathrm{\bf red,comb.}$, which takes the individual nature of
each data-group into account
\item look at the stacked residuals of all data-groups to reveal weak
features
\end{itemize}
during a simultaneous fit in order to find the global best-fit. We
have demonstrated the tremendous benefit of analyzing the stacked
residuals by observations of three accreting neutron stars, in which we
could identify weak features that had not been detected before.
\begin{acknowledgements}
M.~K\"uhnel was supported by the Bundesministerium f\"ur Wirtschaft und
Technologie under Deutsches Zentrum f\"ur Luft- und Raumfahrt grants 50OR1113
and 50OR1207. The \texttt{SLxfig} module, developed by John E. Davis, was used
to produce all figures shown in this paper. We are thankful for the
constructive and critical comments by the reviewers, which were helpful to
significantly improve the quality of the paper.
\end{acknowledgements}
\renewcommand{\bibsep}{0pt}
|
\section{Introduction}
This article is devoted the study of the higher regularity properties of the free boundary of solutions to the \emph{thin obstacle} or \emph{Signorini problem}. To this end, we consider local minimizers to the functional
\begin{align*}
J(v):=\int\limits_{B_1^+}a^{ij}\p_iv\p_jv dx,\quad v\in \mathcal{K},
\end{align*}
with $\mathcal{K}:=\{u\in H^1(B_1^+)| \ u\geq 0 \mbox{ on } B_1'\times \{0\}\}$. Here $B_1^+ := \{x\in B_1 \subset \R^{n+1}| \ x_{n+1}\geq 0\}$ and $B_1':= \{x\in B_1 \subset \R^{n+1}| \ x_{n+1}=0\}$ denote the $(n+1)$-dimensional upper half ball and the co-dimension one ball, respectively. The tensor field $a^{ij}: B_1^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ is assumed to be uniformly elliptic, symmetric and $W^{1,p}(B_1^+)$ regular for some $p> n+1$. Here and in the sequel we use the summation convention.\\
Due to classical results on variational inequalities \cite{U87}, \cite{F10}, minimizers of this problem exist and are unique (under appropriate boundary conditions). Moreover, minimizers are $C^{1,\min\{1-\frac{n+1}{p},\frac{1}{2}\}}(B_{1/2}^+)$ regular (c.f. \cite{AC06}, \cite{KRSI}) and solve the following uniformly elliptic equation with \emph{complementary} or \emph{Signorini boundary conditions}
\begin{equation}
\label{eq:thin_obst}
\begin{split}
\p_i a^{ij} \p_j w &= 0 \mbox{ in } B_1^+,\\
w \geq 0, \ a^{n+1, j} \p_{j}w\leq 0,\ w (a^{n+1,j}\p_j w) &= 0 \mbox{ in } B_1' \times \{0\}.
\end{split}
\end{equation}
Here the bulk equation is to be interpreted weakly, while the boundary conditions hold pointwise. In particular, the constraint originating from the convex set $\mathcal{K}$ only acts on the boundary; in this sense the obstacle is \emph{thin}. The constraint on functions in $\mathcal{K}$ divides the boundary $B_1' \times \{0\}$ into three different regions: The \emph{contact set} $\Lambda_w := \{x\in B_{1}'\times \{0\}| \ w=0\}$, where the minimizer attains the obstacle, the \emph{non-coincidence set}, $\Omega_w:= \{x\in B_1' \times \{0\}| \ w>0\}$, where the minimizer lies strictly above the obstacle, and the \emph{free boundary}, $\Gamma_w := \partial \Omega_w$, which separates the contact set from the non-coincidence set. \\
As we seek to obtain a more detailed analysis of the (regular) free boundary under higher regularity assumptions on the metric tensor $a^{ij}$, we briefly recall the, for our purposes, most relevant known properties of the free boundary (c.f. \cite{CSS}, \cite{PSU}, \cite{GSVG14}, \cite{GPSVG15}, \cite{KRS14} \cite{KRSI}):
Considering metrics which need not be more regular than $W^{1,p}$ with $p\in(n+1,\infty]$ and carrying out a blow-up analysis of solutions, $w$, of (\ref{eq:thin_obst}) around free boundary points, it is possible to assign to each free boundary point $x_0\in \Gamma_w \cap B'_1$ the uniquely determined \emph{order of vanishing} $\kappa(x_0)$ of $w$ at this point (c.f. Proposition 4.2 in \cite{KRS14}\xspace):
\begin{align*}
\kappa(x_0):= \lim\limits_{r \rightarrow 0_+} \frac{\ln\left( r^{-\frac{n+1}{2}} \left\| w \right\|_{L^2(B_r^+(x_0))} \right)}{\ln(r)} .
\end{align*}
Since the order of vanishing satisfies the gap property that either $\kappa(x_0)= \frac{3}{2}$ or $\kappa(x_0)\geq 2$ (c.f. Corollary 4.2 in \cite{KRS14}\xspace), the free boundary can be decomposed as follows:
\begin{align*}
\Gamma_w \cap B_{1}' := \Gamma_{3/2}(w)\cup \bigcup\limits_{\kappa \geq 2} \Gamma_{\kappa}(w),
\end{align*}
where $\Gamma_{\kappa}(w):= \{x_0\in \Gamma_{w} \cap B_{1}'| \ \kappa(x_0)= \kappa\}$. Moreover, noting that the mapping $\Gamma_w \ni x_0\mapsto \kappa(x_0)$ is upper-semi-continuous (c.f. Proposition 4.3 in \cite{KRS14}\xspace), we obtain the set $\Gamma_{3/2}(w)$, which is called the \emph{regular free boundary}, is a relatively open subset of $\Gamma_w$. At each regular free boundary point $x_0\in \Gamma_{3/2}(w)$, there exists an $L^2$-normalized blow-up sequence $w_{x_0,r_j}$, which converges to a nontrivial global solution $w_{3/2}(Q(x_0)x)$ with flat free boundary. Here $w_{3/2}(x):=\Ree (x_n+ix_{n+1})^{3/2}$ is a model solution and $Q(x_0)\in SO(n+1)$ (c.f. Proposition 4.5 in \cite{KRS14}\xspace).
By a more detailed analysis the regular free boundary can be seen to be $C^{1,\alpha}$ regular (c.f. Theorem 2 in \cite{KRSI}\xspace) and a leading order expansion of solutions $w$ at the regular free boundary can be determined (c.f. Proposition 4.6 in \cite{KRSI}\xspace and Corollary 4.8 in \cite{KRSI}\xspace, c.f. also Proposition \ref{prop:asym2} in Section \ref{sec:asymp}).\\
In the sequel we will exclusively focus on the \emph{regular} free boundary. Due to its relative openness and by scaling, it is always possible to assume that the whole boundary in a given domain consists only of the regular free boundary. This convention will be used throughout the article; whenever referring to the ``free boundary'' without further details, we will mean the regular free boundary.
\subsection{Main results and ideas}
In this article our main objective is to prove \emph{higher} regularity of the (regular) free boundary if the metric $a ^{ij}$ is of higher (Hölder) regularity. In particular, we prove the analyticity of the free boundary for analytic coefficients:
\begin{thm}
\label{thm:higher_reg}
Let $a^{ij}:B_1^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ be a uniformly elliptic, symmetric, $W^{1,p}$ tensor field with $p\in(n+1,\infty]$. Suppose that $w:B_{1}^+ \rightarrow \R$ is a solution of the variable coefficient thin obstacle problem (\ref{eq:thin_obst}) with metric $a^{ij}$.
\begin{itemize}
\item[(i)] Then the regular free boundary $\Gamma_{3/2}(w)$ is locally a $C^{1,1-\frac{n+1}{p}}$ graph if $p<\infty$ and a $C^{1,1-}$ graph if $p=\infty$.
\item[(ii)] Assume further that $a^{ij}$ is $C^{k,\gamma}$ regular with $k\geq 1$ and $\gamma\in (0,1)$. Then the regular free boundary $\Gamma_{3/2}(w)$ is locally a $C^{k+1,\gamma}$ graph.
\item[(iii)] Assume in addition that $a^{ij}$ is real analytic. Then the regular free boundary $\Gamma_{3/2}(w)$ is locally real analytic.
\end{itemize}
\end{thm}
We note that these results are sharp on the Hölder scale. In deriving the sharp gain of a full derivative, the choice of our function spaces play a key role (c.f. the discussion below for details on the motivation of our function spaces and Remark \ref{rmk:optreg} in Section \ref{sec:IFT1} for the optimality on the Hölder scale and the role of our function spaces). \\
In addition to the previously stated results, we also deal with the regularity problem in the presence of inhomogeneities.
\begin{thm}
\label{thm:higher_reg_inhom}
Let $a^{ij}:B_1^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ be a $W^{1,p}$ tensor field with $p\in(2(n+1),\infty]$ and let $f:B_1^+ \rightarrow \R$ be an $L^p(B_1^+)$ function. Suppose that $w:B_{1}^+ \rightarrow \R$ is a solution of the variable coefficient thin obstacle problem with metric $a^{ij}$ and inhomogeneity $f$:
\begin{equation}
\label{eq:thin_obst_inhom}
\begin{split}
\p_i a^{ij} \p_j w &= f \mbox{ in } B_1^+,\\
w \geq 0, \ a^{n+1, j} \p_{j}w\leq 0,\ w (a^{n+1,j}\p_j w) &= 0 \mbox{ in } B_1' \times \{0\}.
\end{split}
\end{equation}
\begin{itemize}
\item[(i)] Then the regular free boundary $\Gamma_{3/2}(w)$ is locally a $C^{1,\frac{1}{2}-\frac{n+1}{p}}$ graph.
\item[(ii) ] Assume in addition that $a^{ij}$ is a $C^{k,\gamma}$ tensor field with $k\geq 1$ and $\gamma\in (0,1]$ and let $f:B_1^+ \rightarrow \R$ be a $C^{k -1,\gamma}$ function. Then the regular free boundary $\Gamma_{3/2}(w)$ is locally a $C^{k+[\gamma+ \frac{1}{2}], \gamma+\frac{1}{2}- [\gamma+ \frac{1}{2}]}$ graph.
\item[(iii)] Moreover, assume that $a^{ij}$ is a real analytic tensor field and let $f:B_1^+ \rightarrow \R$ be a real analytic function. Then the regular free boundary $\Gamma_{3/2}(w)$ is locally real analytic.
\end{itemize}
Here $[\cdot]$ denotes the floor function.
\end{thm}
We note that this in particular includes the set-up with non-zero obstacles with as low as $W^{2,p}$, $p\in(2(n+1),\infty]$, regularity (c.f. Section \ref{sec:nonzero}). To to best of our knowledge Theorems \ref{thm:higher_reg} and \ref{thm:higher_reg_inhom} are the first results on higher regularity for the thin obstacle problem with variable coefficients and inhomogeneities.\\
In order to obtain a better understanding for the gain of the free boundary regularity with respect to the regularity of the inhomogeneity, it is instructive to compare Theorem \ref{thm:higher_reg_inhom}, i.e. the situation of the variable coefficient \emph{thin} obstacle problem, with that of the variable coefficient \emph{classical} obstacle problem (c.f. \cite{KN77}, \cite{F10}): In the classical obstacle problem (for the Laplace operator) there is a gain of \emph{one} order of differentiability with respect to the inhomogeneity, i.e. if $f\in C^{k,\alpha}$, then the (regular) free boundary $\Gamma_w$ is $C^{k+1,\alpha}$ regular. This can be seen to be optimal by for instance considering an inhomogeneity which only depends on a single variable (the variable $x_n$ in whose direction the free boundary is a graph, i.e. $\Gamma_w=\{x\in B_1| \ x_n = g(x')\}$, with a choice of parametrization such that locally $|\nabla' g| \neq 0$), by using up to the boundary elliptic regularity estimates for all derivatives $\p_i w$ with respect to directions orthogonal to $e_n$ and by expressing the partial derivative $\p_{nn}w$ along the free boundary in terms of the parametrization $g$.\\
In contrast, in our situation of \emph{thin} obstacles, we gain \emph{three halves} of a derivative with respect to a general inhomogeneity. We conjecture that this is the optimal gain. As we are however dealing with a co-dimension two free boundary value problem, it seems harder to prove the optimality of this gain by similar means as for the classical obstacle problem. Yet, we remark that this gain of three-halves of a derivative also fits the scaling behavior (though not the regularity assumptions) of the inhomogeneities treated in \cite{DSS14}.\\
Let us explain the main ideas of deriving the regularity results of Theorems \ref{thm:higher_reg} and \ref{thm:higher_reg_inhom}:
In order to prove higher regularity properties of the free boundary, we rely on the partial Legendre-Hodograph transform (c.f. \cite{KN77}, \cite{KPS})
\begin{equation}\label{eq:L}
\begin{split}
T: B_1^+ &\rightarrow Q_+:=\{ y \in \R^{n+1}| \ y_n \geq 0, y_{n+1} \leq 0\},\\
y:=T(x)&:= (x'', \p_{n}w, \p_{n+1} w), \ v(y):= w(x) - x_n y_{n} - x_{n+1}y_{n+1},
\end{split}
\end{equation}
which allows us to fix the (regular) free boundary:
\begin{align*}
T(\Gamma_w) \subset \{y\in \R^{n+1}| \ y_n=y_{n+1}=0\}.
\end{align*}
The asymptotic expansion of $w$ around $\Gamma_w$ implies that the transformation $T$ is asymptotically a square root mapping. Similar arguments as in \cite{KPS} yield that $T$ is invertible with inverse given by
\begin{align*}
T^{-1}(y)= (y'', - \p_n v(y), -\p_{n+1}v(y)).
\end{align*}
Thus, the free boundary $\Gamma_w$ can be parametrized in terms of the Legendre function $v$ as
\begin{align*}
\Gamma_w \cap B_{1/2}' := \{x'\in B_{1/2}'| \ x_n = -\p_{n} v(x'',0,0)\}.
\end{align*}
Therefore, it suffices to study the regularity properties of the Legendre function $v$, in order to derive higher regularity properties of the free boundary $\Gamma_w$.\\
Pursuing this strategy and investigating the properties of the Legendre function $v$, we encounter several difficulties:\\
\emph{Nonlinearity and subellipticity of the transformed equation, function spaces.} In analogy to the observations in \cite{KPS} the Hodograh-Legendre transformation $T$ transforms the \emph{uniformly} elliptic equation for $w$ into a fully nonlinear, \emph{degenerate} (sub)elliptic equation for $v$ (c.f. Proposition \ref{prop:bulk_eq}). Moreover, studying the asymptotics of $v$ at the degenerate set of the nonlinear operator (which is the image of the free boundary under the transformation $T$), the linearized operator (at $v$) is identified a perturbation of the (subelliptic) \emph{Baouendi-Grushin} operator (c.f. Section~\ref{sec:grushin}).\\
In this context a central new ingredient and major contribution of the article enters: Seeking to deduce regularity by an application of the implicit function theorem instead of direct and tedious elliptic estimates (c.f. Section \ref{sec:IFT1}), we have to capture the relation between the linearized and nonlinear operators in terms of our \emph{function spaces} (c.f. Section \ref{sec:holder}). This leads to the challenge of finding function spaces which on the one hand mimic the asymptotics of the Legendre function. This is a key requirement, since the perturbative interpretation of the fully nonlinear operator (as a nonlinear Baouendi-Grushin type operator) crucially relies on the asymptotics close the the straightened free boundary. On the other hand, the spaces have to allow for good regularity estimates for the linearized equation which is of Baouendi-Grushin type. In this context, we note that Calderon-Zygmund estimates and Schauder estimates have natural analogues for subelliptic operators like the Baouendi-Grushin operator. The mismatch between the vector space structure (relevant for derivatives) and the subelliptic
geometry allows for nontrivial choices in the definition of Sobolev spaces and higher order H\"older spaces.\\
In order to deal with both of the described conditions, we introduce \emph{generalized Hölder} spaces which are on the one hand adapted to the Baouendi-Grushin operator (for instance by relying on the intrinsic geometry induced by this operator) and on the other hand measure the distance to an approximating polynomial with the ``correct'' asymptotics close to the straightened free boundary (c.f. Section \ref{sec:holder} for the definition and properties of our generalized Hölder spaces and the Appendix, Section \ref{sec:append} for the proofs of these results). These function spaces are reminiscent of Campanato type spaces (c.f. \cite{Ca64}, \cite{Mo09}) and also of the polynomial approximations used by De Silva and Savin \cite{DSS14}. While similar constructions are possible for elliptic equations they seem to be not relevant there. For our problem however they are crucial.\\
\emph{Partial regularity and the implicit function theorem.} Seeking to avoid lengthy and tedious higher order derivative estimates for the Legendre function $v$, we deviate from the previous strategies of proving higher regularity that are present in the literature on the thin obstacle problem. Instead we reason by the \emph{implicit function theorem} along the lines of an argument introduced by Angenent (c.f. \cite{AN90}, \cite{AN90a}, \cite{KL12}). In this context we pre-compose our Legendre function, $v$, with a one-parameter family, $\Phi_a$, of diffeomorphisms leading to a one-parameter family of ``Legendre functions'', $v_a$ (Section~\ref{subsec:IFT0}). Here the diffeomeorphisms are chosen such that the parameter dependence on $a$ is analytic and so that the diffeomorphisms are the identity outside of a fixed compact set, whereas at the free boundary infinitesimally they generate a family of translations in the tangential directions. The functions $v_a$ satisfy a similar fully nonlinear, degenerate elliptic equation as $v$. Invoking the analytic implicit function theorem, we then establish that solutions of this equation are necessarily analytic in the parameter $a$, which, due to the uniqueness of solutions, implies that the family $v_a$ depends on $a$ analytically. As the family of diffeomorphisms, $\Phi_a$, infinitesimally generates translations in the tangential directions, this immediately entails the partial analyticity of the original Legendre function $v$ in the tangential variables.\\
\emph{Corner domain, function spaces.}
Compared with the constant coefficient case in \cite{KPS}, the presence of variable coefficients leads to a completely new difficulty: By the definition of the Hodograph-Legendre transform \eqref{eq:L}, the transformation $T$ maps the upper half ball $B_1^+$ into the quarter space $Q_+$ (c.f. Section \ref{sec:Hodo}). In particular, the free boundary is mapped into the \emph{edge} of $Q_+$, which does not allow us to invoke standard interior regularity estimates there. \\
In contrast to the argument in \cite{KPS}, we cannot overcome this problem by reflecting the resulting solution so as to obtain a problem in which the free boundary is in the interior of the domain: Indeed, this would immediately lead to a loss of regularity of the coefficients $a^{ij}$ and hence would not allow us to prove higher regularity estimates up to the boundary.
Thus, instead, we have to work in the setting of an equation that is posed in the quarter space, where the singularity of the domain is centered at the straightened free boundary. This in particular necessitates regularity estimates in this (singular) domain which hold uniformly up to the boundary (c.f. Appendix, Section \ref{sec:quarter_Hoelder}). \\
In deducing these regularity estimates, we strongly rely on the form of our generalized Hölder spaces and on the interpretation of our fully nonlinear equation as a perturbation of the Baouendi-Grushin Laplacian in the quarter space which satisfies homogeneous Dirichlet data on $\{y_{n}=0\}$ and homogeneous Neumann data on $\{y_{n+1}=0\}$. As it is possible to classify and explicitly compute all the homogeneous solutions to this operator, an approximation argument in the spirit of \cite{Wa03} yields the desired regularity estimates in our generalized Hölder spaces (c.f. Appendix, Section \ref{sec:quarter_Hoelder}).\\
\emph{Low regularity metrics.}
In the case of only $W^{1,p}$ regular metrics with $p\in(n+1,\infty]$, and/or $L^p$ inhomogeneities, even \emph{away} from the free boundary a general solution $w$ is only $W^{2,p}$ regular. Thus, the previous arguments leading to the invertibility of the Hodograph-Legendre transform do not apply directly, as they rely on pointwise bounds for $D^2w$.
To resolve this issue, we use the \emph{splitting technique} from \cite{KRSI} and introduce a mechanism that exchanges \emph{decay} and \emph{regularity}: More precisely, we split a general solution $w$ into two components $w=\tilde{u}+u$. Here the first component $\tilde{u}$ deals with the low regularity of the coefficients and the inhomogeneity:
\begin{align*}
a^{ij}\p_{ij} \tilde{u} - \dist(x,\Gamma_w)^{-2}\tilde{u} = f - (\p_i a^{ij})\p_j w \mbox{ in } B_1 \setminus \Lambda_w, \ \tilde{u}=0 \mbox{ on } \Lambda_w.
\end{align*}
Due to the inclusion of the strongly coercive term $-\dist(x,\Gamma_w)^{-2}\tilde{u}$ in the equation, the solution $\tilde{u}$ has a \emph{strong decay} properties (compared to $w$) towards $\Gamma_w$. We hence interpret it as a controlled error.\\
The second contribution $u$ is now of better \emph{regularity} away from the free boundary $\Gamma_w$, as it solves the non-divergence form elliptic equation
\begin{align*}
a^{ij}\p_{ij} u = - \dist(x,\Gamma_w)^{-2}\tilde{u} \mbox{ in } B_1 \setminus \Lambda_w, \ \tilde{u}=0 \mbox{ on } \Lambda_w.
\end{align*}
Moreover, it captures the essential behavior of the original function $w$ (c.f. Lemma \ref{lem:lower1'} and Proposition \ref{prop:improved_reg1}). In particular, the free boundary $\Gamma_w$ is the same as the free boundary $\Gamma_u:=\partial_{B_1'}\{x\in B_1': u(x)>0\}$ of $u$. We then apply our previous arguments to $u$ and correspondingly obtain the regularity of the free boundary.
\subsection{Literature and related results}
The thin obstacle problem has been studied extensively beginning with the fundamental works of Caffarelli \cite{Ca79}, Uraltseva \cite{U85}, \cite{U87}, Kinderlehrer \cite{Ki81}, and the break through results of Athanasopoulos, Caffarelli \cite{AC06}, as well as Athanasopoulos, Caffarelli, Salsa \cite{ACS08} and Caffarelli, Silvestre, Salsa \cite{CSS}. While there is a quite good understanding of many aspects of the \emph{constant} coefficient problem, the \emph{variable} problem has only recently received a large amount of attention: Here, besides the early work of Uraltseva \cite{U87}, in particular the articles by Garofalo, Smit Vega Garcia \cite{GSVG14} and Garofalo, Petrosyan, Smit Vega Garcia \cite{GPSVG15} and the present authors \cite{KRS14}, \cite{KRSI} should be mentioned. While the methods differ -- the first two articles rely on a frequency function approach and an epiperimetric inequality, the second two articles build on a Carleman estimate as well as careful comparison arguments -- in both works the regularity of the regular free boundary is obtained for the variable coefficient problem under low regularity assumptions on the metric.\\
Hence, it is natural to ask whether the free boundary regularity can be improved if higher regularity assumptions are made on the coefficients and what the precise dependence on the regularity of the coefficients amounts to. In the constant coefficient setting, the higher regularity question has independently been addressed by De Silva, Savin \cite{DSS14}, who prove $C^{\infty}$ regularity of the free boundary by approximation arguments, and by Koch, Petrosyan, Shi \cite{KPS}, who prove analyticity of the free boundary. While the precise dependence on the coefficient regularity is well understood for the \emph{classical} obstacle with variable coefficients \cite{F10}, to the best of our knowledge this question has not yet been addressed in the framework of the \emph{variable} coefficient \emph{thin} obstacle problem.
\subsection{Outline of the article}
The remainder of the article is organized as follows: After briefly introducing the precise setting of our problem and fixing our notation in the following Section \ref{sec:prelim}, we recollect the asymptotic behavior of solutions of (\ref{eq:thin_obst}) in Section \ref{sec:asymp}. With this at hand, in Section \ref{sec:Hodo} we introduce the partial Hodograph-Legendre transformation in the case of $C^{k,\gamma}$ metrics with $k\geq 1$, obtain its invertibility (c.f. Proposition \ref{prop:invertibility}) and in Section \ref{sec:Legendre} derive the fully nonlinear, degenerate elliptic equation which is satisfied by the Legendre function $v$ (Proposition \ref{prop:bulk_eq}). Motivated by the linearization of this equation, we introduce our generalized Hölder spaces (c.f. Definitions \ref{defi:Hoelder}, \ref{defi:Hoelder1} and \ref{defi:spaces}) which are adapted to the geometry of the Baouendi-Grushin Laplacian (Section \ref{sec:holder}). Exploring the (self-improving) structure of the nonlinear equation for the Legendre function $v$ (c.f. Proposition \ref{prop:error_gain}), we deduce regularity properties of the Legendre function $v$ (c.f. Proposition \ref{prop:regasymp}) by an iterative bootstrap argument (c.f. Proposition \ref{prop:error_gain}) in Section \ref{sec:improve_reg}. In Section~\ref{sec:fb_reg} we build on this regularity result and proceed with the application of the implicit function theorem to prove the optimal regularity of the regular free boundary when the metrics $a^{ij}$ are $ C^{k,\gamma}$ Hölder regular for some $k\geq 1$ (c.f. Theorem \ref{prop:hoelder_reg_a}). Moreover, we also derive analyticity of the free boundary for analytic metrics (c.f. Theorem \ref{prop:analytic}). This provides the argument for the first two parts of Theorem \ref{thm:higher_reg}. Next, in Section~\ref{sec:W1p} we study the Hodograph-Legendre transformation for $W^{1,p}$ metrics with $p\in (n+1,\infty]$ and thus derive the optimal regularity result of Theorem \ref{thm:higher_reg} (i). Using similar ideas, we also discuss the necessary adaptations in proving regularity results in the presence of inhomogeneities and nonzero obstacles (c.f. Proposition \ref{prop:inhomo_2}). Finally, in the Appendix, Section \ref{sec:append}, we prove a characterization of our function spaces introduced in Section~\ref{sec:holder} and show an a priori estimate for the Baouendi-Grushin Laplacian in these function spaces (c.f. Section \ref{sec:quarter_Hoelder}). We also discuss auxiliary regularity and mapping properties which we use in the derivation of the asymptotics and in the application of the implicit function theorem (c.f. Sections \ref{sec:XY}, \ref{sec:kernel}).
\section{Preliminaries}
\label{sec:prelim}
\subsection{Conventions and normalizations}
\label{sec:conventions}
In the sequel we introduce a number of conventions which will be used throughout this paper.
Any tensor field $a^{ij}:B_1^+\rightarrow \R^{(n+1)\times (n+1)}_{sym}$ in this paper is uniformly elliptic, symmetric and at least $W^{1,p}$ regular for some $p\in (n+1,\infty]$. Furthermore, we assume that
\begin{itemize}
\item[(A1)]$a^{ij}(0)=\delta^{ij}$,
\item[(A2)] (Uniform ellipticity) $\frac{1}{2}|\xi|^2\leq a^{ij}(x)\xi_i\xi_j \leq 2|\xi|^2$ for each $x\in B_1^+$ and $\xi\in \R^{n+1}$,
\item[(A3)] (Off-diagonal) $a^{i,n+1}(x',0)=0$ for $i\in \{1,\dots,n\}$.
\end{itemize}
Here (A1)-(A2) follow from an affine transformation. The off-diagonal assumption (A3) is a consequence of a change of coordinates (c.f. for instance Section 2.1 in \cite{KRS14} and Uraltseva \cite{U85}), which allows to reduce \eqref{eq:thin_obst} to
\begin{equation}
\label{eq:varcoeff}
\begin{split}
\p_ia^{ij}\p_jw=0 &\text{ in } B_1^+,\\
w\geq 0,\quad \p_{n+1}w\leq 0, \ w(\p_{n+1}w)=0 &\text{ on } B'_1.
\end{split}
\end{equation}
Under the above assumptions (A1)-(A3), a solution $w$ to the thin obstacle problem is $C^{1,\min\{1-\frac{n+1}{p},\frac{1}{2}\}}_{loc}$ regular and of $\dist(x,\Gamma_w)^{3/2}$ growth at the free boundary $\Gamma_w$, i.e. for any $x_0\in \Gamma_w\cap B^+_{1/2}$,
\begin{equation}\label{eq:interior_est}
\sup _{ B_r(x_0)}|\nabla w|\leq C(n,p, \|a^{ij}\|_{W^{1,p}})\|w\|_{L^2(B_1^+)}r^{\frac{1}{2}}, \quad r\in (0,1/2).
\end{equation}
This regularity and growth behavior is optimal by the interior regularity and the growth behavior of the model solution $w_{3/2}(x)=\Ree(x_n+ix_{n+1})^{3/2}$. We refer to \cite{U85} for the $C^{1,\alpha}_{loc}$ regularity and to \cite{KRSI} for the optimal $C^{1,\min\{1-\frac{n+1}{p},\frac{1}{2}\}}_{loc}$ regularity as well as the growth result. In this paper we will always work with a solution $w\in C^{1,\min\{1-\frac{n+1}{p},\frac{1}{2}\}}_{loc}(B_1^+)$, for which \eqref{eq:interior_est} holds true.\\
In order to further simplify our set-up, we observe the following symmetry properties of our problem:\\
(Symmetry) Equation \eqref{eq:thin_obst} is invariant under scaling and multiplication. More precisely, if $w$ is a solution to \eqref{eq:thin_obst}, then for $x_0\in K\Subset B'_1$, for $c\geq 0$ and $\lambda>0$, the function
$$x\mapsto cw(x_0+\lambda x)$$
is a solution to \eqref{eq:thin_obst} (with coefficients $a^{ij}(x_0+\lambda\cdot)$) in $B_r^+$, $r\in (0, \lambda^{-1}(1-|x_0|)]$.\\
These symmetry properties are for instance crucial in carrying out rescalings around the (regular) free boundary:
Assuming that $x_0\in \Gamma_{3/2}(w)$ is a regular free boundary point and defining $w_{x_0,\lambda}(x):=w(x_0+\lambda x)/\lambda^{\frac{3}{2}}$, $\lambda\in (0,1/4)$, the asymptotic expansion around the regular free boundary (c.f. Proposition 4.6 in \cite{KRSI}\xspace) yields that
\begin{align*}
w_{x_0,\lambda}(x)\rightarrow a(x_0)w_{x_0}(x) \mbox{ in } C^{1,\beta}_{loc}(\R^{n+1}_+)
\end{align*}
for each $\beta\in (0,1/2)$ as $\lambda\rightarrow 0_+$. Here $w_{x_0}$ is a global solution with flat free boundary and $a(x_0)>0$ is a constant. \\
In this paper we are interested in the higher regularity of $\Gamma_{3/2}(w)$ under an appropriate higher regularity assumption on the metric $a^{ij}$. All the results given below are \emph{local} estimates around \emph{regular} free boundary points. Using the scaling and multiplication symmetries of the equation, we may hence without loss of generality suppose the following normalization assumptions (A4)-(A7):
\begin{itemize}
\item[(A4)] $0\in \Gamma_{3/2}(w)$,
\end{itemize}
that $w$ is sufficiently close to $w_{3/2}$ and that the metric is sufficiently flat in the following sense: For $\epsilon_0,c_\ast>0$ small
\begin{itemize}
\item[(A5)] $\|w-w_{3/2}\|_{C^1(B_1^+)}\leq \epsilon_0$,
\item[(A6)] $\|\nabla a^{ij}\|_{L^p(B_1^+)}\leq c_\ast$.
\end{itemize}
By \cite{KRSI}, if $\epsilon_0$, $c_\ast $ are sufficiently small depending on $n,p,\|w\|_{L^2(B_1^+)}$, then assumptions (A5)-(A6) imply that $\Gamma_w\cap B'_{1/2}\subset \Gamma_{3/2}(w)$ and that $\Gamma_w\cap B'_{1/2}$ is a $C^{1,\alpha}$ graph, i.e. after a rotation of coordinates $\Gamma_{w}\cap B_{1/2}' = \{x'=(x'',x_n,0)\cap B_{1/2}'| \ x_n = g(x'')\}$, for some $\alpha\in (0,1)$. Moreover, we have the following estimate for the (in-plane) outer unit normal $\nu_{x_0}$ of $\Lambda_w$ at $x_0$
\begin{equation}\label{eq:normal}
|\nu_{x_0}-\nu_{\tilde{x}_0}|\lesssim \max\{\epsilon_0,c_\ast\}|x_0-\tilde{x}_0|^\alpha, \text{ for any }x_0, \tilde{x}_0\in \Gamma_w\cap B'_{1/2},
\end{equation}
For notational simplicity, we also assume that
\begin{itemize}
\item[(A7)] $\nu_0=e_n$.
\end{itemize}
From now on, we will always work under the assumptions (A1)-(A7).
\subsection{Notation}
\label{sec:notation}
Similarly as in \cite{KRSI} we use the following notation:\\
\emph{Geometry.}
\begin{itemize}
\item $\R^{n+1}_+:=\{(x'',x_n,x_{n+1})\in \R^{n+1} | x_{n+1}\geq 0\}$.
\item $B_r(x_0):=\{x\in \R^{n+1}| |x-x_0|<r\}$, where $|\cdot|$ is the norm induced by the Euclidean metric, $B_r^+(x_0):=B_r(x_0)\cap \R^{n+1}_+$, $B'_r(x_0):=B_r(x_0)\cap \{x_{n+1}=0\}$. If $x_0$ is the origin, we simply write $B_r$, $B_r^+$ and $B'_r$.
\item Let $w$ be a solution of \eqref{eq:thin_obst}, then $\Lambda_w:=\{(x',0)\in B'_1| w(x',0)=0\}$ is the \emph{contact set}, $\Omega_w:=B'_1\setminus \Lambda_w$ is the \emph{positivity set}, $\Gamma_w:=\p_{B'_1}\Lambda_w\cap B'_1$ is the \emph{free boundary}, $\Gamma_{3/2}(w):=\{x\in \Gamma_w| \kappa_x=\frac{3}{2}\}$ is the \emph{regular set} of the free boundary, where $\kappa_x$ is the vanishing order at $x$.
\item For $x_0\in \Gamma_w$, we denote by $\mathcal{N}_{x_0}=\{x\in B^+_{1/4}(x_0)\big|\dist(x,\Gamma_w)\geq \frac{1}{2}|x-x_0|\}$ the non-tangential cone at $x_0$.
\item $\mathcal{C}'_\eta(e_n):=\mathcal{C}_\eta(e_n)\cap \{e_{n+1}=0\}$ is a tangential cone (with axis $e_n$ and opening angle $\eta$).
\item $Q_+:=\{(y'',y_n,y_{n+1})\in \R^{n+1} | y_n\geq 0, y_{n+1}\leq 0\}$.
\item $\tilde{\mathcal{B}}_r(y_0):=\{y\in \R^{n+1}| d_G(y,y_0)<r\}$, where $d_G(\cdot, \cdot)$ is the Baouendi-Grushin metric (c.f. Definition~\ref{defi:Grushinvf}). $\tilde{\mathcal{B}}_r^+(y_0):=\tilde{\mathcal{B}}_r(y_0)\cap Q_+$.
\item In the $Q_+$ with the Baouendi-Grushin metric $d_G(\cdot,\cdot)$, given $y_0\in P:=\{y_n=y_{n+1}=0\}$ we denote by $\mathcal{N}_G(y_0):=\{x\in \tilde{\mathcal{B}}_{1/4}^+(y_0)\big|\dist_G(y,P)\geq \frac{1}{2}d_G(y,y_0)\}$ the Baouendi-Grushin non-tangential cone at $y_0$.
\item We use the Baouendi-Grushin vector fields $Y_i$, $i\in\{1,\dots,n+1\}$, (c.f. Definition \ref{defi:Grushinvf}) and the modified Baouendi-Grushin vector fields $\tilde{Y}_i$, $i\in\{1,\dots,2n\}$ (c.f. Definition \ref{defi:Hoelder1}).
\item For $k\in \N$, we denote by $\mathcal{P}_k^{hom}$ the space of homogeneous polynomials (w.r.t. the Grushin scaling) of order $k$ (c.f. Definition \ref{defi:poly}), and by $\mathcal{P}_k$ the vector space of homogeneous polynomials of order less than or equal to $k$.
\end{itemize}
\emph{Functions and function spaces.}
\begin{itemize}
\item $w_{3/2}(x):= c_n \Ree(x_n + i x_{n+1})^{3/2}$, where $c_n>0$ is a normalization constant ensuring that $\| w_{3/2} \|_{L^2(B_{1}^+)}=1$.
\item $w_{1/2}(x):= c_n \Ree(x_n + i x_{n+1})^{1/2}$ and $\bar{w}_{1/2}(x):= - c_n \Imm(x_n + i x_{n+1})^{1/2}$, where $c_n>0$ denotes the same normalization constant as above.
\item We denote the \emph{asymptotic profile} at a point $x_0 \in \Gamma_{3/2}(w)\cap B_{1}'$ by $\mathcal{W}_{x_0}$. It is given by
\begin{align*}
\mathcal{W}_{x_0}(x)=a(x_0)w_{3/2}\left(\frac{(x-x_0)\cdot \nu_{x_0}}{(\nu_{x_0}\cdot A(x_0)\nu_{x_0})^{1/2}}, \frac{x_{n+1}}{(a^{n+1,n+1}(x_0))^{1/2}}\right).
\end{align*}
\item For a solution $w$ to (\ref{eq:varcoeff}) and a point $x_0\in \Gamma_{w}$ we define a \emph{blow-up sequence} $w_{x_0,\lambda}(x):=\frac{w(x_0 + \lambda x)}{\lambda^{3/2}}$ by rescaling with $\lambda \in (0,1)$. The asymptotic expansion from Proposition 4.6 in \cite{KRSI}\xspace implies that as $\lambda \rightarrow 0$ it converges to the \emph{blow-up profile} $\mathcal{W}_{x_0}(\cdot+ x_0)$. Here $\mathcal{W}_{x_0}$ denotes the asymptotic profile from above.
\item In the sequel, we use spaces adapted to the Baouendi-Grushin operator $\Delta_G$ and denote the corresponding Hölder spaces by $C^{k,\alpha}_{\ast}$ (c.f. Definitions \ref{defi:Hoelder}, \ref{defi:Hoelder1}). Moreover, relying on these, we construct our generalized Hölder spaces $X_{\alpha,\epsilon}, Y_{\alpha,\epsilon}$ which are appropriate for our corner domains (c.f. Definition \ref{defi:spaces}).
\item We use the notation $C_0(Q_+)$ to denote the space of all continuous functions vanishing at infinity.
\item Let $\R^{(n+1)\times (n+1)}_{sym}$ denote the space of symmetric matrices and let
$$G: \R^{(n+1)\times (n+1)}_{sym} \times \R^{n+1} \times \R^{n+1} \rightarrow \R, \ (M,P,y)\mapsto G(M,P,y),$$
with $M=(m_{k\ell})_{k \ell} \in \R^{(n+1) \times (n+1)}_{sym}$, $P = (p_1,\dots,p_{n+1})\in \R^{n+1}$ and $y=(y_1,\dots,y_{n+1})\in \R^{n+1}$. We denote the partial derivative with respect to the different components by
\begin{align*}
\p_{m_{k\ell}}G(M,P,y)&:= \frac{\p G(M,P,y)}{\partial m_{k \ell}},\\
\p_{p_{k}}G(M,p,y) &:=\frac{\p G(M,P,y)}{\partial p_{k}},\\
\p_{y_k}G(M,p,y)&:=\frac{\p G(M,P,y)}{\partial y_{k}}.
\end{align*}
\item $\D_G$ stands for the Baouendi-Grushin operator
\begin{align*}
\D_{G} v:= (y_n^2 + y_{n+1}^2)\D'' v + \p_{nn}v + \p_{n+1,n+1}v,
\end{align*}
where $\D''$ denotes the Laplacian in the tangential variables, i.e. in the $y''$-variables of $y=(y'',y_n,y_{n+1})$.
\end{itemize}
The notation $A\lesssim B$ means that $A\leq CB$ with $C$ depending only on dimension $n$.
\section{Hodograph-Legendre Transformation}
\label{sec:HLTrafo}
In this section we perform a partial Hodograph-Legendre transform of our problem (\ref{eq:varcoeff}). While fixing the free boundary, this comes at the price of transforming our uniformly elliptic equation in the upper half ball into a fully nonlinear, degenerate (sub)elliptic equation in the lower quarter ball (c.f. Sections \ref{sec:Hodo}, \ref{sec:Legendre}, Propositions \ref{prop:invertibility}, \ref{prop:bulk_eq}). In particular, in addition to the difficulties in \cite{KPS}, the domain in which our problem is posed now contains a corner. In spite of this additional problem, as in \cite{KPS} we identify the fully nonlinear equation as a perturbation of the Baouendi-Grushin operator with symmetry (i.e. with Dirichlet-Neumann data) by a careful analysis of the asymptotic behavior of the Legendre transform (c.f. Section \ref{sec:Legendre}, Example \ref{ex:linear} and Section \ref{sec:grushin}).
\subsection{Asymptotic behavior of the solution $w$}
\label{sec:asymp}
We begin by deriving and collecting asymptotic expansions for higher order derivatives of solutions to our equation (c.f. \cite{KRSI}). This will prove to be advantageous in the later sections (e.g. Sections \ref{sec:Hodo}, \ref{sec:Leg}).
\begin{prop}[\cite{KRSI}, Proposition 4.6]
Let $a^{ij}\in W^{1,p}(B_1^+, \R^{(n+1)\times (n+1)}_{sym})$ with $p\in (n+1,\infty]$ be a uniformly elliptic tensor.
Assume that $w:B_1^+ \rightarrow \R$ is a solution to the variable coefficient thin obstacle problem and that it satisfies the following conditions:
\label{prop:asym2}
There exist positive constants $\epsilon_0$ and $c_{\ast}$ such that
\begin{itemize}
\item[(i)] $\|w-w_{3/2}\|_{C^1(B_1^+)}\leq \epsilon_0$,
\item[(ii)] $\|\nabla a^{ij}\|_{L^p(B_1^+)}\leq c_\ast$.
\end{itemize}
Then if $\epsilon_0$ and $c_\ast$ are sufficiently small depending on $n,p$, there exists some $\alpha\in (0,1-\frac{n+1}{p}]$ such that $\Gamma_w\cap B_{1/2}^+$ is a $C^{1,\alpha}$ graph. Moreover, at each free boundary point $x_0\in \Gamma_w\cap B^+_{1/4}$, there exists an asymptotic profile, $\mathcal{W}_{x_0}(x)$,
\begin{align*}
\mathcal{W}_{x_0}(x)=a(x_0)w_{3/2}\left(\frac{(x-x_0)\cdot \nu_{x_0}}{(\nu_{x_0}\cdot A(x_0)\nu_{x_0})^{1/2}}, \frac{x_{n+1}}{(a^{n+1,n+1}(x_0))^{1/2}}\right),
\end{align*}
such that for any $x\in B_{1/4}^+(x_0)$
\begin{align*}
(i)\quad &\left|\p_i w(x)-\p_i \mathcal{W}_{x_0}(x)\right|\leq C_{n,p}\max\{\epsilon_0,c_{\ast}\}|x-x_0|^{\frac{1}{2}+\alpha}, \quad i\in\{1,\dots,n\},\\
(ii)\quad &\left| \p_{n+1} w(x)-\p_{n+1}\mathcal{W}_{x_0}(x)\right|\leq C_{n,p}\max\{\epsilon_0,c_{\ast}\} |x-x_0|^{\frac{1}{2}+\alpha},\\
(iii)\quad &\left|w(x)-\mathcal{W}_{x_0}(x)\right|\leq C_{n,p}\max\{\epsilon_0,c_{\ast}\} |x-x_0|^{\frac{3}{2}+\alpha}.
\end{align*}
Here $x_0\mapsto a(x_0)\in C^{0,\alpha}(\Gamma_w\cap B_{1/2}^+)$, $\nu_{x_0}$ is the (in-plane) outer unit normal of $\Lambda_w$ at $x_0$ and $A(x_0)=(a^{ij}(x_0))$. Furthermore, $w_{3/2}(x)= c_n \Ree(x_n+ i x_{n+1})^{3/2}$, where $c_n>0$ is a dimensional constant which is chosen such that $\| w_{3/2}\|_{L^{2}(B_1^+)}=1$.
\end{prop}
Assuming higher regularity of the metric allows us to use a scaling argument to deduce the asymptotics for higher order derivatives in non-tangential cones.
\begin{prop}
\label{prop:improved_reg}
Let $a^{ij}\in C^{k,\gamma}(B_1^+, \R^{(n+1)\times (n+1)}_{sym})$ with $k\geq 1$ and $\gamma\in (0,1)$ be uniformly elliptic. Let $\alpha>0$ be the Hölder exponent from Proposition \ref{prop:asym2}. There exist $\epsilon_0$ and $c_\ast$ sufficiently small depending on $n,p$ such that if
\begin{itemize}
\item[(i)] $ \|w-w_{3/2}\|_{C^1(B_1^+)}\leq \epsilon_0$,
\item[(ii)]$ [a^{ij}]_{\dot{C}^{k,\gamma}(B_1^+)}\leq c_\ast,$
\end{itemize}
then for each $x_0\in \Gamma_w\cap B^+_{1/4}$, an associated non-tangential cone $x\in \mathcal{N}_{x_0}:=\{x\in B^+_{1/4}(x_0)| \ \dist(x,\Gamma_w)\geq \frac{1}{2}|x-x_0|\}$ and for all multi-indeces $\beta$ with $|\beta|\leq k+1$ we have
\begin{align*}
\left|\p^\beta w(x)-\p^\beta \mathcal{W}_{x_0}(x)\right|& \leq C_{\beta,n,p} \max\{\epsilon_0,c_{\ast}\} |x-x_0|^{\frac{3}{2}+\alpha-|\beta|},\\
\left[\p^\beta w-\p^\beta \mathcal{W}_{x_0}\right]_{\dot{C}^{0,\gamma}(\mathcal{N}_{x_0}\cap (B_{3\lambda /4 }^+(x_0)\setminus B_{\lambda/2 }^+(x_0)))}& \leq C_{\beta,n,p}\max\{\epsilon_0,c_{\ast}\}\lambda^{\frac{3}{2}+\alpha-\gamma-|\beta|}.
\end{align*}
Here $\alpha$ is the same exponent as in Proposition~\ref{prop:asym2} and $\lambda \in (0,1)$.
\end{prop}
\begin{rmk}
\label{rmk:improved_reg}
It is possible to extend the above asymptotics for $x$ in the full neighborhood $B_{1/4}^+(x_0)$ as
\begin{align*}
&\left|\p^\beta w(x)-\p^\beta \mathcal{W}_{x_0}(x)\right|\leq C_{\beta,n,p} \max\{\epsilon_0,c_{\ast}\} |x-x_0|^{\frac{1}{2}+\alpha}\dist(x,\Gamma_w)^{-|\beta|+1}.
\end{align*}
Here it is necessary to introduce the distance to the free boundary instead of measuring it with a negative power of $|x-x_0|$.
\end{rmk}
Before coming to the proof of Proposition \ref{prop:improved_reg}, we state an immediate corollary, which will be important in the derivation of the asymptotics of the Legendre function in Proposition \ref{prop:holder_v} in Section \ref{sec:Leg}.
\begin{cor}
\label{cor:improved_reg}
Assume that the conditions of Proposition \ref{prop:improved_reg} hold. Let $$w_{x_0,\lambda}(x):=\frac{w(x_0+\lambda x)}{\lambda^{3/2}},\quad \lambda>0.$$ Then,
\begin{align*}
&[\p^\beta w_{x_0,\lambda}-\p^\beta \mathcal{W}_{x_0}(x_0+\cdot)]_{\dot{C}^{0,\gamma}(\mathcal{N}_{0}\cap (B_{3 /4}^+\setminus B_{1/2}^+))}\leq C_{n,p} \max\{\epsilon_0,c_{\ast}\}\lambda^{\alpha}.
\end{align*}
\end{cor}
\begin{proof}[Proof of Proposition \ref{prop:improved_reg}]
The proof of the proposition follows from elliptic estimates in Whitney cubes, which in turn are reduced to estimates on the scale one by scaling the problem.\\
We only prove the result for $k=1$ (i.e. in case of $|\beta|=2$) and restrict ourselves to the $L^{\infty}$ estimates. For $k> 1$ and for the second estimate the argument is similar. Moreover, we observe that the case $|\beta|=1$ is already covered in Proposition~\ref{prop:asym2}. We begin by considering the tangential derivatives of $w$: Let $\tilde{v}:=\p_\ell w$ with $\ell\in \{1,\dots, n\}$. Then $\tilde{v}$ satisfies
\begin{align*}
\p_i(a^{ij}\p_j\tilde{v})=\p_iF^i, \quad F^i=-(\p_\ell a^{ij})\p_jw,
\end{align*}
with the boundary conditions
\begin{align*}
\tilde{v}&=0 \text{ on } \Lambda_w, \quad \p_{n+1}\tilde{v}=0 \text{ on } B'_1\setminus \Lambda_w.
\end{align*}
Also, the derivative of the profile functions, $\p_\ell \mathcal{W}_{x_0}$, satisfies
\begin{align*}
\p_i(a^{ij}\p_j(\p_\ell \mathcal{W}_{x_0}))=g_1+g_2,
\end{align*}
with
\begin{align*}
g_1=(\p_{\ell} a^{ij}) \p_j\p_i \mathcal{W}_{x_0}, \ g_2=(a^{ij}(x)-a^{ij}(x_0))\p_{ij}\p_\ell\mathcal{W}_{x_0}.
\end{align*}
Seeking to combine the information on the functions $\tilde{v}$ and $\p_{\ell} \mathcal{W}_{x_0}$, we define
\begin{align*}
\tilde{u}(x):=\frac{w(x_0+\lambda x)-\mathcal{W}_{x_0}(x_0+\lambda x)}{\lambda^{\frac{3}{2}+\alpha}}, \ 0<\lambda<1/4.
\end{align*}
Due to the previous considerations,
$\p_\ell \tilde{u}$ satisfies the equation
\begin{align*}
\p_i(a^{ij}(x_0+\lambda \cdot)\p_j \p_\ell \tilde{u})=\p_i \tilde{F}^i - \tilde{g}_1 - \tilde{g}_2 \text{ in } B_1^{+}.
\end{align*}
Here
\begin{equation}
\begin{split}
\label{eq:tilde}
\tilde{F}^i(x)=\lambda^{\frac{1}{2}-\alpha}F^i(x_0+\lambda x),\\
\tilde{g}_1(x)=\lambda^{\frac{3}{2}-\alpha} g_1(x_0+\lambda x),\\
\tilde{g}_2(x)=\lambda^{\frac{3}{2}-\alpha} g_2(x_0+\lambda x).
\end{split}
\end{equation}
Moreover, by the asymptotics of $w$ at $x_0$ which were given in (iii) of Proposition~\ref{prop:asym2}, we obtain the following $L^{\infty}$ bound in the non-tangential cone $\mathcal{N}_0=\{x\in B^+_{1/4}|\dist(x,\Gamma_{w_{x_0,\lambda}})\geq \frac{1}{2}|x|\}$ for all $\ell \in \{1,\dots,n+1\}$:
\begin{align}\label{eq:max}
|\partial_{\ell} \tilde{u}|\lesssim C_{n,p}\max\{\epsilon_0,c_{\ast}\}.
\end{align}
Noting that
\begin{align*}
|F^i(x)|&\lesssim c_{\ast} \dist(x,\Gamma_w)^{1/2},\\
|g_1(x)|&\lesssim c_{\ast} \dist(x,\Gamma_w)^{-1/2},\\
|g_2(x)|&\lesssim c_{\ast} |x-x_0| \dist(x,\Gamma_w)^{-3/2},
\end{align*}
recalling that $\lambda \dist(x, \Gamma_{w_{x_0,\lambda}}) = \dist(\lambda (x-x_0), \Gamma_w)$ and using (\ref{eq:tilde}) yields
\begin{equation}
\label{eq:distresc}
\begin{split}
|\tilde{F}^i(x)|&\lesssim c_{\ast} \lambda^{1-\alpha}\dist(x,\Gamma_{w_{x_0,\lambda}})^{1/2},\\
|\tilde{g}_1(x)|&\lesssim c_{\ast} \lambda^{1-\alpha}\dist(x,\Gamma_{w_{x_0,\lambda}})^{-1/2},\\
|\tilde{g}_2(x)|&\lesssim c_{\ast} \lambda^{1-\alpha}\dist(x,\Gamma_{w_{x_0,\lambda}})^{-3/2}.
\end{split}
\end{equation}
By the definition of $\mathcal{N}_0$ the expressions involving the distance functions in (\ref{eq:distresc}) are uniformly (in $\lambda$) bounded in $B_1^+\setminus B_{1/4}^+$. Moreover, it is immediate to check that the semi-norms $[\tilde{F}^i]_{C^{0,\gamma}}$ are uniformly bounded.
For $\ell\in \{1,\dots, n\}$, we apply the $C^{1,\gamma}$ estimate to $\p_\ell\tilde{u}$, which holds up to the boundary, in $\mathcal{N}_0\cap (B^+_1\setminus B^+_{1/4})$ (note that with $\epsilon_0, c_\ast$ sufficiently small, $\mathcal{N}_0\cap (B_1^+\setminus B^+_{1/4})$ does not intersect the free boundaries $\Gamma_w$ or $\Gamma_{\mathcal{W}_{x_0}}$, thus $\tilde{u}$ satisfies either Dirichlet or Neumann conditions):
\begin{align}
\label{eq:tangential}
\|\p_\ell\tilde{u}\|_{C^{1,\gamma}(\mathcal{N}_0\cap (B^+_{3/4}\setminus B^+_{1/2}))} \lesssim \|\partial_{\ell}\tilde{u}\|_{L^\infty(\mathcal{N}_0\cap (B^+_1\setminus B^+_{1/4}))}+ c_{\ast}\lambda^{1-\alpha}.
\end{align}
In order to obtain a full second derivatives estimate, we now combine (\ref{eq:tangential}) with the equation for $\partial_{n+1}\tilde{u}$ to also obtain
\begin{align*}
\|\p_{n+1,n+1}\tilde{u}\|_{C^{0,\gamma}(\mathcal{N}_0\cap (B^+_{3/4}\setminus B^+_{1/2}))} \lesssim \sum\limits_{\ell=1}^{n}\|\partial_{\ell}\tilde{u}\|_{L^\infty(\mathcal{N}_0\cap (B^+_1\setminus B^+_{1/4}))}+ c_{\ast} \lambda^{1-\alpha}.
\end{align*}
Rescaling back and using \eqref{eq:max} consequently leads to
\begin{align*}
|\nabla \p_\ell w(x)-\nabla \p_\ell \mathcal{W}_{x_0}(x)|\lesssim \max\{\epsilon_0,c_{\ast}\}\lambda^{-\frac{1}{2}+\alpha} \text{ in } \mathcal{N}_{x_0}\cap (B^+_{3\lambda/4}(x_0)\setminus B^+_{\lambda/2}(x_0)),
\end{align*}
for $\ell\in \{1,\dots, n+1\}$.
Since this holds for any $\lambda\in (0,1/4)$, we conclude that
\begin{align*}
|\p^\beta w(x)-\p^\beta \mathcal{W}_{x_0}(x)|\lesssim \max\{\epsilon_0,c_{\ast}\} |x-x_0|^{-\frac{1}{2}+\alpha}, \quad x\in \mathcal{N}_{x_0},\ |\beta|=2.
\end{align*}
\end{proof}
\begin{rmk}
\label{rmk:normal}
In the next section we will strongly use the asymptotics of the first order derivatives $\p_\ell w$ with $\ell \in \{1,\dots,n+1\}$. Hence, for future reference we state them explicitly:
\begin{align*}
\p_e\mathcal{W}_{x_0}(x)&=b_e(x_0)w_{1/2}\left(\frac{(x-x_0)\cdot \nu_{x_0}}{(\nu_{x_0}\cdot A(x_0)\nu_{x_0})^{1/2}}, \frac{x_{n+1}}{(a^{n+1,n+1}(x_0))^{1/2}}\right),\\
\p_{n+1}\mathcal{W}_{x_0}(x)&=b_{n+1}(x_0)\bar w_{1/2}\left(\frac{(x-x_0)\cdot \nu_{x_0}}{(\nu_{x_0}\cdot A(x_0)\nu_{x_0})^{1/2}}, \frac{x_{n+1}}{(a^{n+1,n+1}(x_0))^{1/2}}\right),
\end{align*}
where
\begin{align*}
w_{1/2}(x)&= c_n \Ree(x_n + i x_{n+1})^{1/2},\quad \bar{w}_{1/2}(x)= - c_n \Imm(x_n + i x_{n+1})^{1/2},\\
b_e(x_0)&=\frac{3(e\cdot \nu_{x_0}) a(x_0)}{2(\nu_{x_0}\cdot A(x_0)\nu_{x_0})^{1/2}},\quad b_{n+1}(x_0)=\frac{3 a(x_0)}{2(a^{n+1,n+1}(x_0))^{1/2}},
\end{align*}
and $c_n>0$ is the same normalization constant as in Proposition \ref{prop:asym2}.
\end{rmk}
\begin{rmk}\label{rmk:convention}
For simplicity we can, and in the sequel will, further assume that $0\in \Gamma_w$ and that
\begin{align*}
\nu_0=e_n, \quad b_n(0)=b_{n+1}(0)=1\ (\text{which corresponds to } a(0)=2/3).
\end{align*}
Thus,
$$\nabla \mathcal{W}_0(x)=(0,w_{1/2}(x),\bar w_{1/2}(x)).$$
Moreover, under the assumptions of Proposition \ref{prop:asym2} we can also bound the $\dot{C}^{0,\alpha}$ semi-norm of $b_n, b_{n+1}$ by $\max\{\epsilon_0,c_{\ast}\}$.
\end{rmk}
Last but not least, we recall a sign condition on $\p_n w$ and $\p_{n+1}w$, which plays an important role in the determination of the image of the Hodograph-Legendre transform in (\ref{eq:mapT}) in Section \ref{sec:Hodo}. An extension of this to the set-up of $W^{1,p}$, $p\in (n+1,\infty]$, metrics is recalled in Section \ref{sec:ext}. As explained in \cite{KRSI} this requires an additional splitting step.
\begin{lem}[Positivity, \cite{KRSI}, Lemma 4.12.]
\label{lem:lower1}
Let $a^{ij}:B_1^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ be a tensor field that satisfies the conditions from Section \ref{sec:conventions} and in addition is $C^{1,\gamma}$ regular for some $\gamma \in (0,1)$. Let $w:B_1^+ \rightarrow \R$ be a solution of the thin obstacle problem with metric $a^{ij}$ and assume that it satisfies the normalizations from Section \ref{sec:conventions}. Then there exist positive constants $\eta= \eta(n)$ and $c=c(n)$ such that
\begin{align}
\label{eq:lower1}
\p_ew(x)\geq c\dist(x,\Lambda_w)\dist(x,\Gamma_w)^{-\frac{1}{2}}, \quad x\in B_{\frac{1}{2}}^+
\end{align}
for $e\in \mathcal{C}'_\eta(e_n):=\mathcal{C}_\eta(e_n)\cap \{e_{n+1}=0\}$, which is a tangential cone (with axis $e_n$ and opening angle $\eta$). Similarly,
\begin{align*}
\p_{n+1}w(x)\leq -c \dist(x,\Omega_w)\dist(x,\Gamma_w)^{-\frac{1}{2}}, \quad x\in B_{\frac{1}{2}}^+.
\end{align*}
\end{lem}
\subsection{Hodograph-Legendre transformation}
\label{sec:Hodo}
In this section we perform a partial Hodograph-Legendre transformation to show the higher regularity of the free boundary with zero obstacle. In the sequel, we assume that the metric satisfies $a^{ij}\in C^{1,\gamma}(B_1^+, \R^{(n+1)\times (n+1)}_{sym})$ with $\gamma\in (0,1)$.\\
We define the partial Hodograph-Legendre transformation associated with $w$ as
\begin{align}
\label{eq:def_Legendre}
T=T^w:B_1^+\rightarrow \R^{n+1}, \quad y=T(x)=(x'', \partial_{n} w(x), \partial_{n+1}w(x)).
\end{align}
The regularity of $w$ immediately implies that $T\in C^{0,1/2}(B_1^+)$. Moreover,
\begin{equation}
\label{eq:mapT}
\begin{split}
T(B_1^+\setminus B'_1)&\subset \{y_n>0, y_{n+1}<0\},\\
T(\Lambda_w)&\subset\{y_n=0, y_{n+1}\leq 0\}, \ T(B'_1\setminus \Lambda_w)\subset \{y_n>0, y_{n+1}=0\},\\
T(\Gamma_w)&\subset\{y_n=y_{n+1}=0\}.
\end{split}
\end{equation}
Here the first inclusion is a consequence of Lemma \ref{lem:lower1}.
Using the leading order asymptotic expansions from Section \ref{sec:asymp}, we prove the invertibility of the transformation:
\begin{prop}[Invertibility of $T$]\label{prop:invertibility}
Suppose that the assumptions of Proposition~\ref{prop:asym2} hold. Then, if $[\nabla a^{ij}]_{\dot{C}^{0,\gamma}(B_1^+)}\leq c_{\ast}$ and if $\epsilon_0$ and $c_\ast $ are sufficiently small, the map $T$ is a homeomorphism from $B_{1/2}^+$ to $T(B_{1/2}^+) \subset \{y\in \R^{n+1}| \ y_n\geq 0, y_{n+1}\leq 0\}$. Moreover, away from $\Gamma_w$, $T$ is a $C^1$ diffeomorphism.
\end{prop}
The proof of this result essentially relies on the facts that for each fixed $x''$, the transformation $T$ is asymptotically a square root mapping and the free boundary $\Gamma_w$ is sufficiently flat (i.e. it is a $C^{1,\alpha}$ graph with slow varying normals c.f. \eqref{eq:normal}). Hence, the main idea is to show the injectivity of $T$ on dyadic annuli around the free boundary. At these points the map $T$ is differentiable, which allows us to exploit the non-degeneracy of the derivative of $T$. To achieve this reduction to dyadic annuli we exploit the asymptotic structure of the functions $w$ (c.f. Propositions \ref{prop:asym2} and \ref{prop:improved_reg}).
\begin{proof}
\emph{Step 1: Homeomorphism.}\\
We begin with the injectivity of $T$ in $B_{1/2}^+$.
Since $T$ fixes the first $n-1$ variables, it is enough to show that for each $x_0\in \Gamma_w\cap B_{1/2}'$, $T$ is injective on the set $H_{x_0}:=\{(x''_0,x_n,x_{n+1})\}\cap B_{1/2}^+$. Moreover, as $\Gamma_w$ is given as a graph of a $C^{1,\alpha}$ function $g$, it suffices to prove that $T(x)\neq T(\tilde{x})$ for any two points $x,\tilde{x}\in H_{x_0}$ such that $x,\tilde{x}\notin \Gamma_w$.
In order to obtain this, we first prove that the mapping $T_1:=\psi\circ T$ is injective (and a homeomorphism) on $B_{1/2}^+$. Here $\psi:\R^{n+1}\rightarrow \R^{n+1}$ with $\psi(z)=(z'',z_n^2-z_{n+1}^2, -2z_nz_{n+1})$. Note that $T_1(x)=(x'',(\p_n w(x))^2 -(\p_{n+1} w(x))^2, - 2 \p_n w(x) \p_{n+1}w(x))$. We rely on the asymptotic expansion of $\nabla w$. In a second step, we then return to the mapping properties of $T$.\\
\emph{Step 1a: $T_1$ is a homeomorphism.}
We begin with the injectivity of $T_1$.
By Proposition~\ref{prop:improved_reg}, for $x\in B^+_{1/2}$
\begin{align*}
\p_n w(x)&= w_{1/2}(x)+\max\{\epsilon_0,c_{\ast}\}O(|x|^{\frac{1}{2}+\alpha}),\\
\p_{n+1}w(x)&=\bar{w}_{1/2}(x)+\max\{\epsilon_0,c_{\ast}\}O(|x|^{\frac{1}{2}+\alpha}).
\end{align*}
Hence, a direct computation gives that
\begin{align*}
&T_1(x)=x+E_0(x), \\
&\text{where } E_0:B_{1/2}^+\rightarrow \R^{n+1},\ |E_0(x)|=\max\{\epsilon_0,c_{\ast}\}O(|x|^{1+\alpha}).
\end{align*}
In general, by the explicit asymptotic expansions of $\p_nw$ and $\p_{n+1}w$ around $x_0\in \Gamma_w\cap B^+_{1/2}$ (c.f. Proposition~\ref{prop:improved_reg}) and by using the fact that $b_n,b_{n+1}\in C^{0,\alpha}(\Gamma_w\cap B^+_{1/2})$, $\nu(x_0)=\nu_{x_0}\in C^{0,\alpha}(\Gamma_w\cap B^+_{1/2})$ and $A(x_0)\in C^{0,\alpha}(\Gamma_w\cap B^+_{1/2})$, we have
\begin{equation}
\label{eq:identity}
\begin{split}
T_1(x)- T_1(x_0)&=(x-x_0) + E_{x_0}(x), \quad x\in B_{1/2}^+(x_0)\\
\text{where } |E_{x_0}(x)|&\lesssim \max\{\epsilon_0,c_{\ast}\} \left(|x-x_0||x_0|^\alpha+ |x-x_0|^{1+\alpha}\right).
\end{split}
\end{equation}
Here we recall that as indicated in Remark \ref{rmk:convention} we may assume that the Hölder constants of $b_n(x_0), b_{n+1}(x_0)$ are controlled by $\max\{\epsilon_0,c_{\ast}\}$.
From the identity (\ref{eq:identity}) we note that if $\epsilon_0, c_*$ are sufficiently small and if $x_0\in \Gamma_w\cap B_{1/2}^+$, then for $x\in B_{1/2}^+(x_0)$
\begin{align}
\label{eq:cont_boundary}
(1-\frac{1}{4})|T_1(x) - T_1(x_0)| \leq |x-x_0|\leq (1+\frac{1}{4})|T_1(x) - T_1(x_0)|.
\end{align}
Thus, if there are $x, \tilde{x}\in H_{x_0}$ with $x, \tilde{x}\notin \Gamma_w$ such that $T_1(x)=T_1(\tilde{x})$, then necessarily
\begin{align}\label{eq:quo}
\frac{1}{2}\leq \frac{|x-x_0|}{|\tilde{x}-x_0|}\leq 2.
\end{align}
Without loss of generality, we assume that $|x-x_0|\leq |\tilde{x}-x_0|$ and define $r:=|x-x_0|$. Then \eqref{eq:quo} implies that $x, \tilde{x}\in A_{r,2r}^+(x_0)\cap H_{x_0}$, where $A_{r,2r}^+(x_0)$ is the closed cylinder centered at $x_0$:
\begin{align*}
A^+_{r,2r}(x_0)&:=\{(x'',x_n,x_{n+1})\in B_1^+| \ |x''-x''_0|\leq r,\\
&\qquad r\leq \sqrt{(x_n-(x_0)_n)^2+(x_{n+1}-(x_0)_{n+1})^2}\leq 2r\}.
\end{align*}
Since $\Gamma_w$ is $C^{1,\alpha}$ with $|\nu_{x_0}-\nu_{\tilde{x}_0}|\lesssim \max\{\epsilon_0,c_\ast\}|x_0-\tilde{x}_0|^\alpha$, for any $x_0, \tilde{x}_0\in \Gamma_w\cap B'_{1/2}$ and $\nu_0=e_n$, we have that $\Gamma_w \cap A^+_{r,2r}(x_0)=\emptyset$ for a sufficiently small (but independent of $r$) choice of the constants $\epsilon_0, c_\ast$. Thus, $T_1$ is a $C^1$ mapping in $A^+_{r,2r}(x_0)\cap B_{1/2}^+$ (because $w$ is $C^{2,\gamma}$ away from $\Gamma_w$). We compute $DT_1$ in $A_{r,2r}^+(x_0)\cap B_{1/2}^+$. By using the asymptotics of $D w$ and $D^2w$ around $x_0$ (c.f. Propositions \ref{prop:asym2}, \ref{prop:improved_reg}), we obtain
\begin{align*}
|DT_1(x)-I|\lesssim \max\{\epsilon_0,c_{\ast}\} \left(|x_0|^{\alpha}+r^{2\alpha}\right), \quad x\in A_{r,2r}^+(x_0) \cap \mathcal{N}_{x_0}\cap B_{1/2}^+,
\end{align*}
where $I$ is the identity map.
Therefore, for sufficiently small, universal constants $\epsilon_0, c_\ast$, the map $T_1$ is injective in $A_{r,2r}^+(x_0)\cap \mathcal{N}_{x_0}\cap B_{1/2}^+$. This implies that $T_1(x)\neq T_1(\tilde{x})$.\\
\emph{Step 1b: $T_1:B_{1/2}^+ \rightarrow T_1(B_{1/2}^+)$ is a homeomorphism.}
By the continuity of $T_1$ and by the invariance of domain theorem, we infer that, as a mapping from $\inte(B_{1/2}^+)$ to $T_1(\inte(B_{1/2}^+))$, $T_1$ is a homeomorphism. We claim that this is also true for $T_1$ as a map from $B_{1/2}^+$ to $T_1(B_{1/2}^+)$. Indeed, due to our previous considerations in Step 1a, $T_1$ is injective (and hence invertible) on the whole of $B_{1/2}^+$ (as a map onto its image). Hence, it suffices to prove the continuity of the inverse. Here we distinguish three cases: Let $y\in T_1(B_{1/2}^+)$ and first assume that $y\in T_1(\Gamma_w\cap B_{1/2}')$. Then, (\ref{eq:cont_boundary}) immediately implies the continuity of $T_1^{-1}$ at $y$. Secondly, we assume that $y\in T_1(B_{1/2}'\setminus \Lambda_w)$. Let $x=T^{-1}_1(y)\in B_{1/2}'\setminus \Lambda_w$. Then we carry out an even reflection of $w$ about $x_{n+1}$ (and a corresponding partly even, partly odd reflection for $a^{ij}$) as described in Remark 3.8 in \cite{KRSI}. The resulting reflected function $\tilde{w}$ is still $C^{1,1/2}$ regular in a (sufficiently small) neighborhood $B_{\rho}(x) \subset B_{1/2}^+ \setminus \Lambda_w$ of $x$. Moreover, the $y_{n+1}$ -component of $T_1^{\tilde{w}}$ changes sign on passing from $x_{n+1}>0$ to $x_{n+1}<0$. Thus, the mapping $T_1^{\tilde{w}}$ is still injective as a mapping from $B_{\rho}(x)$ to $T^{\tilde{w}}_1(B_{\rho}(x))$. Since it is also continuous, the invariance of domain theorem implies that it is a homeomorphism from $B_{\rho}(x)$ to $T^{\tilde{w}}_1(B_{\rho}(x))$, which is an open subset in $\R^{n+1}$ containing $y$. In particular, this implies that our original mapping, $(T^{w}_1)^{-1}$, is continuous at $y\in T_1(B_{1/2}'\setminus \Lambda_w)$. Last but not least, for a point $y\in T_1(B_{1/2}' \cap \inte(\Lambda_w))$, we argue similarly. However, instead of using an even reflection, we carry out an odd reflection of $w$ about $x_{n+1}$. Again, we note that the associated map $T_1^{\tilde{w}}$ changes sign on passing from $x_{n+1}>0$ to $x_{n+1}<0$. Thus, arguing as in the second case, we again obtain the continuity of $(T_1^{w})^{-1}$ at $y$. Combining the results of the three cases therefore yields that $T_1$ is a homeomorphism as a map from $B_{1/2}^+$ to $T_1(B_{1/2}^+)$, which is relatively open in $\{y_n\geq 0, y_{n+1}\leq 0\}$.
\\
\emph{Step 1c: $T$ is a homeomorphism.}
By definition of $T_1$, we have that $T_1 = \psi\circ T$, where $\psi(x):=(x'', x_n^2 - x_{n+1}^2, -2 x_{n}x_{n+1})$. We show that the injectivity of $T$ follows immediately from the injectivity of $T_1$. As $T(B_1^+)\subset \{y\in \R^{n+1}| y_{n}\geq 0, y_{n+1}\leq 0\}$ and as $\psi$ is injective on this quadrant, we obtain $T(U)=\psi^{-1}\circ T_1(U)$ for any $U\subset B_1^+$. Since $T_1$ is open and $\psi$ is continuous, this implies that $T$ is open. Combining this with the continuity of $T$, we obtain that $T$ is homeomorphism from $B_{1/2}^+$ to $T(B_{1/2}^+)\subset \{y\in \R^{n+1}| y_n\geq 0, y_{n+1}\leq 0\}$.\\
\emph{Step 2: Differentiability.}
Recalling the regularity of the metric, $a^{ij}\in C^{1,\gamma}$ for $\gamma>0$, we observe that $w\in C^{2,\gamma}_{loc}(B_1^+\setminus \Gamma_w)$. Thus, $T$ is $C^1$ away from $\Gamma_w$. In order to show that $T$ is a $C^1$ diffeomorphism away from $\Gamma_w$, it suffices to compute its Jacobian. For $x\in B_{1/2}^+\setminus \Gamma_w$, let $x_0=(x'',g(x''),0)$ be the projection onto $\Gamma_w$. Then, by the asymptotics for $D^2w$ (Proposition~\ref{prop:improved_reg} applied in the non-tangential cone $\mathcal{N}_{x_0}$), we have
\begin{align*}
\det(DT(x))&=\p_{nn}w\p_{n+1,n+1}w-(\p_{n,n+1}w)^2 \\
&=\p_{nn}\mathcal{W}_{x_0}\p_{n+1,n+1}\mathcal{W}_{x_0}-(\p_{n,n+1}\mathcal{W}_{x_0})^2 + \max\{\epsilon_0,c_{\ast}\} O(|x-x_0|^{-1+\alpha}).
\end{align*}
A direct computation gives
\begin{align*}
&\p_{nn}\mathcal{W}_{x_0}\p_{n+1,n+1}\mathcal{W}_{x_0}-(\p_{n,n+1}\mathcal{W}_{x_0})^2\\
&=-\frac{9}{16}a(x_0)^2\frac{e_n\cdot \nu_{x_0}}{(\nu_{x_0}\cdot A(x_0)\nu_{x_0})(a^{n+1,n+1}(x_0))} \frac{1}{\tilde{r}},
\end{align*}
where
\begin{align*}
\tilde{r}=\left(\frac{((x-x_0)\cdot \nu_{x_0})^2}{(\nu_{x_0}\cdot A(x_0)\nu_{x_0})}+\frac{x_{n+1}^2}{a^{n+1,n+1}(x_0)}\right)^{1/2}.
\end{align*}
The $C^{0,\alpha}$ regularity of $\nu_{x_0}$ and the ellipticity of $A(x)=(a^{ij}(x))$ entail that
\begin{align*}
c|x-x_0|\leq \tilde{r}\leq C |x-x_0|, \text{ for some absolute constants }0<c<C<\infty.
\end{align*}
Thus,
\begin{equation}
\label{eq:jacobi}
\det(DT(x))=-c|x-x_0|^{-1}+\max\{\epsilon_0,c_{\ast}\} O(|x-x_0|^{-1+\alpha})<0.
\end{equation}
Therefore, after potentially choosing the constant $1/2=1/2(n,p,\alpha)>0$ even smaller, the implicit function theorem implies that $T$ and $T^{-1}$ are locally $C^{1}$. Due to the global invertibility, which we have proved above, the statement follows.
\end{proof}
\subsection{Legendre function and nonlinear PDE}
\label{sec:Legendre}
In this section we compute a partial Legendre transform of a solution $w$ of our problem (\ref{eq:varcoeff}). In this context it becomes convenient to view the equation (\ref{eq:varcoeff}) in non-divergence form and to regard the equation in the interior as a special case of the problem
\begin{align*}
a^{ij}\p_{ij} u = f(Du,u,y),
\end{align*}
for a suitable function $f$. In our case $f(Du,u,y)=-(\p_{i} a^{ij})\p_j u$.
Starting from this non-divergence form, we compute the equation which the Legendre function satisfies (c.f. Proposition \ref{prop:bulk_eq}). By considering the explicit example of the Legendre transform of $\mathcal{W}_{x_0}$ for $x_0=0$, we motivate that the fully nonlinear equation in the bulk is related to the Baouendi-Grushin operator (c.f. Example \ref{ex:linear}). \\
From now on we will work in the image domain $T(B_{1/2}^+)$, where $T$ is the partial Hodograph transformation defined in \eqref{eq:def_Legendre}. For simplicity, we set $U:=T(B_{1/2}^+)$ and denote the straightened free boundary by $P:=T(\Gamma_w\cap B_{1/2}')$. We recall that the Hodograph transform was seen to be invertible in $U$ (c.f. Proposition \ref{prop:invertibility}).
For $y\in U$, we define the partial Legendre transform of $w$ by the identity
\begin{equation}\label{eq:legendre}
v(y)=w(x)-x_{n}y_n-x_{n+1}y_{n+1}, \quad x=T^{-1}(y).
\end{equation}
A direct computation shows that
\begin{equation}\label{eq:dual}
\partial_{y_i}v=\partial_{x_i}w, \ i=1,\ldots, n-1,\quad \partial_{y_{n}}v=-x_n,\quad \partial_{y_{n+1}}v=-x_{n+1}.
\end{equation}
As a consequence of \eqref{eq:dual}, the free boundary $\Gamma_w\cap B_{1/2}'$ is parametrized by
\begin{align}
\label{eq:boundaryLH}
x_n=-\p_{y_n}v(y'',0,0).
\end{align}
As in \cite{KPS} the advantage of passing to the Legendre-Hodograph transform consists of fixing (the image of the) free boundary, i.e. by mapping it to the co-dimension two hyperplane $y=(y'',0,0)$. However, this comes at the expense of a more complicated, fully nonlinear, degenerate (sub)elliptic equation for $v$. We summarize this in the following:
\begin{prop}[Bulk equation]
\label{prop:bulk_eq}
Suppose that $a^{ij}\in C^{1,\gamma}(B_1^+, \R^{(n+1)\times (n+1)}_{sym})$ is uniformly elliptic.
Let $w:B_1^+ \rightarrow \R$ be a solution of the variable coefficient thin obstacle problem and let $v:U \rightarrow \R$ be its partial Legendre-Hodograph transform. Then $v\in C^{1}(U)$ and it satisfies the following fully nonlinear equation
\begin{equation}
\label{eq:nonlineq1}
\begin{split}
F(D^2v, D v, v,y)&=-\sum_{i,j=1}^{n-1}\tilde{a}^{ij}\det\begin{pmatrix}
\p_{ij}v& \p_{in}v & \p_{i,n+1}v\\
\p_{jn}v& \p_{nn}v & \p_{n,n+1}v\\
\p_{j,n+1}v & \p_{n,n+1}v &\p_{n+1,n+1}v
\end{pmatrix}\\
&+2\sum_{i=1}^{n-1}\tilde{a}^{i,n}\det\begin{pmatrix}
\p_{in}v & \p_{i,n+1}v\\
\p_{n,n+1}v & \p_{n+1,n+1}v
\end{pmatrix}\\
& \quad+2 \sum_{i=1}^{n-1}\tilde{a}^{i,n+1}\det\begin{pmatrix}
\p_{i,n+1}v & \p_{in}v\\
\p_{n,n+1}v & \p_{nn}v
\end{pmatrix}\\
&+\tilde{a}^{nn}\p_{n+1,n+1}v+\tilde{a}^{n+1,n+1}\p_{nn}v-2\tilde{a}^{n,n+1}\p_{n,n+1}v\\
&-\det\begin{pmatrix}
\p_{nn}v &\p_{n,n+1}v\\
\p_{n,n+1}v &\p_{n+1,n+1}v
\end{pmatrix}\left(\sum_{j=1}^{n-1}\tilde{b}^j\p_j v+\tilde{b}^ny_n+\tilde{b}^{n+1}y_{n+1}\right)=0,
\end{split}
\end{equation}
where
\begin{align*}
\tilde{a}^{ij}(y)&:=a^{ij}(x)\big|_{x=(y'',-\p_nv(y),-\p_{n+1}v(y))},\\
\tilde{b}^j(y)&:=\sum_{i=1}^{n+1}(\p_{x_i}a^{ij})(x)\big|_{x=(y'',-\p_nv(y),-\p_{n+1}v(y))}.
\end{align*}
Moreover, the following mixed Dirichlet-Neumann boundary conditions hold:
\begin{align*}
v=0\text{ on } U\cap \{y_n=0\}; \quad \p_{n+1}v=0 \text{ on } U\cap \{y_{n+1}=0\}.
\end{align*}
In particular, $\Gamma_w\cap B_{1/4}$ is parametrized by $x_n=-\p_{y_n}v(y'',0,0)$.
\end{prop}
\begin{rmk}
For convenience of notation, in the sequel we will also use the notation $F(v,y):=F(D^2v,Dv,v,y)$. We emphasize that the coefficients $\tilde{a}^{ij}(y)$ depend on $v$ nonlinearly.
\end{rmk}
The proof of Proposition \ref{prop:bulk_eq} follows by computing the corresponding changes of coordinates:
\begin{proof}
Due to the regularity of $T^{-1}$ and $w$ \eqref{eq:dual} directly entails that $v\in C^{ 1}(U)$. The condition $w=0$ on $\Gamma_w\cap B^+_{1/4}$ immediately translates into $v=0$ on $P$. Moreover, it is easy to check from \eqref{eq:dual} and the Signorini boundary condition of $w$, that $v=0$ on $U\cap \{y_n=0\}$ and $\p_{n+1}v=0$ on $U\cap \{y_{n+1}=0\}$.
Now we derive the equation for $v$. Recalling that
\begin{align*}
y=T(x)=(x', \partial_{x_n}w, \partial_{x_{n+1}} w), \quad x=T^{-1}(y)=(y', -\partial_{y_n}v, -\partial_{y_{n+1}}v),
\end{align*}
and using \eqref{eq:dual}, we have
\begin{align*}
DT=\begin{pmatrix}
I_{n-1} & 0\\
A(w)& H(w)
\end{pmatrix},\quad
DT^{-1}=\begin{pmatrix}
I_{n-1} & 0\\
A(v)& H(v)
\end{pmatrix} \mbox{ in } U\setminus P,
\end{align*}
where
\begin{align*}
A(w)=\begin{pmatrix}
\partial_{x_{n}x_1} w & \ldots & \partial_{x_{n}x_{n-1}} w\\
\partial_{x_{n+1}x_1} w & \ldots & \partial_{x_{n+1}x_{n-1}} w
\end{pmatrix},\quad
H(w)=\begin{pmatrix}
\partial_{x_{n}x_n} w & \partial_{x_{n}x_{n+1}} w\\
\partial_{x_{n+1}x_n} w & \partial_{x_{n+1}x_{n+1}}w
\end{pmatrix},\\
A(v)=-\begin{pmatrix}
\partial_{y_{n}y_1} v & \ldots & \partial_{y_{n}y_{n-1}} v\\
\partial_{y_{n+1}y_1} v & \ldots & \partial_{y_{n+1}y_{n-1}} v
\end{pmatrix},\quad
H(v)=-\begin{pmatrix}
\partial_{y_{n}y_n} v & \partial_{y_{n}y_{n+1}} v\\
\partial_{y_{n+1}y_n} v & \partial_{y_{n+1}y_{n+1}}v
\end{pmatrix}.
\end{align*}
Next we express $D^2w(x)$ in terms of $D^2v(y)$ if $y\in U\setminus P$. Since $(DT)^{-1}=DT^{-1}$, we immediately obtain
\begin{equation}\label{eq:relation}
H(w)=H(v)^{-1},\quad A(v)=-H(w)^{-1}A(w).
\end{equation}
Moreover, the identities \eqref{eq:relation} and \eqref{eq:dual} together with a direct calculation give
\begin{align}
(\partial_{y_iy_j}v)_{(n-1)\times (n-1)}&= (\partial_{x_ix_j}w)_{(n-1)\times (n-1)} - A(w)^t H(w)^{-1} A(w), \label{eq:hessianv}\\
(\partial_{x_ix_j}w)_{(n-1)\times (n-1)}&= (\partial_{y_iy_j}v)_{(n-1)\times (n-1)} - A(v)^t H(v)^{-1} A(v)\label{eq:hessianw}.
\end{align}
In order to compute the equation for $v$, we assume that $a^{ij}\in C^{1,\gamma}$ for some $\gamma>0$ and rewrite the equation for $w$ in non-divergence form
\begin{equation}\label{eq:nondivw}
a^{ij}\partial_{ij}w + (\partial_i a^{ij})\partial_j w=0.
\end{equation}
For convenience and abbreviation, we set
\begin{align*}
\tilde{a}^{ij}(y)&:=a^{ij}(x)\big|_{x=(y'',-\p_nv(y),-\p_{n+1}v(y))},\\
\tilde{b}^j(y)&:=\sum_{i=1}^{n+1}(\p_{x_i}a^{ij})(x)\big|_{x=(y'',-\p_nv(y),-\p_{n+1}v(y))}.
\end{align*}
Plugging \eqref{eq:relation}-\eqref{eq:hessianw} into (\ref{eq:nondivw}), and multiplying the resulting equation by
\begin{align*}
-J(v):=-\det\begin{pmatrix}
\p_{nn}v &\p_{n,n+1}v\\
\p_{n,n+1}v &\p_{n+1,n+1}v
\end{pmatrix},
\end{align*}
leads to the equation~(\ref{eq:nonlineq1}) for $v$.
\end{proof}
We conclude this section by computing the Legendre function of a 3/2-homogeneous blow-up of a solution to the variable coefficient thin obstacle problem.
\begin{lem}
\label{lem:asymp_profile}
Let $w:B_{1}^+ \rightarrow \R$ be a solution of the variable coefficient thin obstacle problem and let $x_0\in \Gamma_w\cap B_{1/2}$. Assume that $v$ is the Legendre function of $w$ under the Hodograph transformation $y=T^w(x)$. Then at $y_0=T^w(x_0)$, the Legendre function $v$ has the asymptotic expansion
$$v(y)= v_{y_0}(y)+ \max\{\epsilon_0,c_\ast\}O(|y-y_0|^{3+2\alpha}),$$
with the leading order profile
\begin{align*}
v_{y_0}(y) &=- \frac{4}{27 a^2(x_0)}\left( \left(\frac{\nu_{x_0}\cdot A(x_0)\nu_{x_0}}{(\nu_{x_0})_n} y_n \right)^3 \right.\\
& \quad \left. - 3 \left( \frac{\nu_{x_0}\cdot A(x_0)\nu_{x_0}}{(\nu_{x_0})_n} (a^{n+1,n+1}(x_0)) \right) y_ny_{n+1}^2 \right)\\
&\quad -g(y_0)y_n + y_{n}\frac{(y''-y_0)\cdot \nu_{x_0}''}{(\nu_{x_0})_n},
\end{align*}
where $\nu_{x_0}:= (\nu_{x_0}'', (\nu_{x_0})_n,0)=\frac{(-\nabla''g(x_0), 1, 0)}{\sqrt{1+|\nabla''g(x_0)|^2}}$ denotes the (in-plane) outer normal to $\Lambda_w$ at $x_0$.
\end{lem}
\begin{proof}
The claim follows from a straightforward calculation. Indeed, recall that $y(x)=(x'',\p_nw(x),\p_{n+1}w(x))$. From the asymptotics of $\p_nw, \p_{n+1}w$ around $x_0\in \Gamma_w$ in Proposition~\ref{prop:asym2}, we obtain the asymptotics of the inverse $x=x(y)$ around $y_0=T^w(x_0)$. Additionally, recalling that $v(y)=w(x(y))-x_n(y)y_n-x_{n+1}(y)y_{n+1}$, we obtain the claimed asymptotic expansion of $v$ around $y_0$.
\end{proof}
It turns out that the function $v_{y_0}(y)$ provides good intuition for the behavior of solutions to (\ref{eq:nonlineq1}).
In order to obtain a better idea about the structure of $F(v,y)$, we compute its linearization at $ v_0(y)$, which is the leading order expansion of $v$ at the origin. It is immediate from Lemma~\ref{lem:asymp_profile} (and using the normalization in Remark~\ref{rmk:normal} and Remark~\ref{rmk:convention}) that
$$v_0(y)=-\frac{1}{3}\left(y_n^3-3y_ny_{n+1}^2\right).$$
\begin{example}[Linearization at $v_0$]
\label{ex:linear}
Let $v_0$ be the Legendre function of the blow-up limit $\mathcal{W}_{0}$ at the origin, which itself is a global solution to the Signorini problem with constant metric $a^{ij}=\delta^{ij}$. Then, the Legendre function $v_0$ satisfies the nonlinear PDE
\begin{align*}
F(D^2v)=-\p_{nn}v-\p_{n+1,n+1}v+ \sum\limits_{i=1}^{n-1} \det
\begin{pmatrix}
\p_{ii}v & \p_{in}v & \p_{i,n+1}v\\
\p_{ni}v & \p_{nn}v & \p_{n,n+1}v\\
\p_{n+1,i}v & \p_{n+1,n}v & \p_{n+1,n+1}v
\end{pmatrix}=0.
\end{align*}
A direct computation leads to
\begin{align*}
\frac{\p F(M)}{\p m_{ij}}\big|_{M=D^2 v_0}=-
\begin{pmatrix}
4(y_n^2+y_{n+1}^2) & 0 & 0\\
0 & 1 &0\\
0& 0& 1
\end{pmatrix}.
\end{align*}
As a consequence, the linearization $L_{v_0}=D_vF \big|_{ v_0}=4(y_n^2+y_{n+1}^2)\Delta''+\p^2_{n,n}+\p^2_{n+1,n+1}$ is a \emph{constant coefficient Baouendi-Grushin operator}.
\end{example}
The previous example and the observation that around the origin $v$ is a perturbation of $v_0$ and $a^{ij}$ is a perturbation of the identity matrix, indicates that the linearization $D_v F$ (and hence $F$) can be viewed as a perturbation of the Baouendi-Grushin Laplacian.
Motivated by this, we introduce function spaces which are adapted to the Baouendi-Grushin operator in the next section.
\section{Function spaces}
\label{sec:holder}
In this section we introduce and discuss generalized H\"older spaces (c.f. Definition \ref{defi:spaces}, Proposition \ref{prop:decompI}) which are adapted to our equation (\ref{eq:nonlineq1}). These are the spaces in which we apply the implicit function theorem in Section~\ref{sec:fb_reg} to deduce the tangential regularity of the Legendre function $v$.
In order to define these spaces, we use the intrinsic geometry induced by the Baouendi-Grushin operator. In particular, we work with the intrinsic (or Carnot-Caratheodory) distance (c.f. Definition \ref{defi:Grushinvf}) associated with the Baouendi-Grushin operator and corresponding intrinsic Hölder spaces (c.f. Definitions \ref{defi:Hoelder}, \ref{defi:Hoelder1}).\\
Our function spaces are inspired by Campanato's characterization of the classical H\"older spaces \cite{Ca64} and are reminiscent of the function spaces used in \cite{DSS14}. They are constructed on the one hand to capture the asymptotics of the Legendre function and on the other hand to allow for elliptic estimates for the Baouendi-Grushin operator (c.f. Proposition \ref{prop:invert}).
\subsection{Intrinsic metric for Baouendi-Grushin Laplacian}
\label{sec:intrinsic}
In this section we define the geometry which is adapted to our equation (\ref{eq:nonlineq1}). This is motivated by viewing our nonlinear operator from (\ref{eq:nonlineq1}) as a variable coefficient perturbation of the constant coefficient \emph{Baouendi-Grushin} operator (c.f. Example \ref{ex:linear})
\begin{align*}
\D_G:= (y_n^2 + y_{n+1}^2)\D'' + \p_n^2 + \p_{n+1}^2.
\end{align*}
The Baouendi-Grushin operator is naturally associated with the \emph{Baouendi-Grushin vector fields} and an \emph{intrinsic metric}:
\begin{defi}
\label{defi:Grushinvf}
Let $Y_i:=\sqrt{y_n^2+y_{n+1}^2}\p_i$, $i\in\{1,\dots, n-1\}$, $Y_n:=\p_n$, $Y_{n+1}:=\p_{n+1}$ denote the \emph{Baouendi-Grushin vector fields}. The metric associated with the vector fields $Y_i$ is
\begin{align}
\label{eq:metr}
ds^2=\sum_{j=1}^{n-1}\frac{dy_j^2}{y_n^2+y_{n+1}^2}+dy_n^2+dy_{n+1}^2.
\end{align}
More precisely it is defined by the following scalar product in the tangent space:
\begin{align*}
g_{y}(v,w):= (y_n^2 + y_{n+1}^2)^{-1}\left(\sum\limits_{j=1}^{n-1}v_j w_j \right) + v_n w_n + v_{n+1} w_{n+1},
\end{align*}
for all $y \in \R^{n+1}$, $v,w \in \spa\{Y_i(y)| \ i\in \{ 1,\dots, n+1\}\}$.
Let $d_G$ be the distance function associated with this sub-Riemannian metric (or the associated \emph{Carnot-Caratheodory metric}):
\begin{multline*}
d_G(x,y) := \inf \{ \ell(\gamma)| \ \gamma: [a,b] \subset \R \rightarrow \R^{n+1} \mbox{ joins } x \mbox{ and } y,\\
\dot{\gamma}(t)\in \spa\{Y_i(\gamma(t))| \ i\in \{1,\dots, n+1\}\}\},
\end{multline*}
where
\begin{align*}
\ell(\gamma) := \int_{a}^{b} \sqrt{g_{\gamma(t)}(\dot{\gamma}(t), \dot{\gamma}(t))}dt.
\end{align*}
\end{defi}
\begin{rmk}\label{rmk:equi_dist}
We remark that for the family of dilations $\delta_\lambda(\cdot)$ which is defined by $\delta_\lambda(y'',y_n,y_{n+1}):=(\lambda^2 y'',\lambda y_{n},\lambda y_{n+1})$, we have $d_G(\delta_\lambda( p), \delta_\lambda (q))=|\lambda|d_G(p,q)$ for $p,q\in \R^{n+1}$. Moreover, from \eqref{eq:metr} for $\sqrt{y_n^2+y_{n+1}^2}\sim 1$ we have $ds^2\sim dy_1^2+\dots+dy_{n+1}^2$. Using these, it is possible to directly verify that $d_G$ is equivalent to the following quasi-metric
\begin{align*}
d(x,y)=|x_n-y_n|+|x_{n+1}-y_{n+1}|+\frac{|x''-y''|}{|x_n|+|x_{n+1}|+|y_n|+|y_{n+1}|+|x''-y''|^{1/2}}.
\end{align*}
\end{rmk}
\begin{rmk}
\label{rmk:original_variables}
In order to elucidate our choice of metric, we derive its form in our original $x$-coordinates. To this end, we consider the case of the flat model solution $w(x)=\mathcal{W}_{0}(x)$. Denoting the Euclidean inner product on $\R^{n+1}$ by $g_0$ and defining $g_{\mathcal{W}_{0}}$ as the Baouendi-Grushin inner product from Definition~\ref{defi:Grushinvf}, (\ref{eq:metr}) (up to constants), we obtain that $g_{\mathcal{W}_{0}}=(x_n^2+x_{n+1}^2)^{-\frac{1}{2}}T_\ast g_{0}$, where $T$ is the Legendre transformation associated with $\mathcal{W}_0$.
\end{rmk}
The previously defined intrinsic metric induces a geometry on our space. In particular, it defines associated Baouendi-Grushin cylinders/balls:
\begin{defi}
\label{defi:Grushincylinder}
Let $0<r\leq 1$. We set
$$\mathcal{B}_r:= \{y\in \R^{n+1}| \ |y''|\leq r^2, \ y_{n}^2 + y_{n+1}^2 \leq r^2 \}$$
to denote the closed \emph{non-isotropic Baouendi-Grushin cylinders}. For
$$y_0\in P:=\{(y'',y_n,y_{n+1})| y_n=y_{n+1}=0\}$$ we further define $\mathcal{B}_r(y_0):=y_0+\mathcal{B}_r$.
In the quarter space, we restrict the cylinders to the corresponding intersection
$$\mathcal{B}_{r}^+(y_0):=\mathcal{B}_{r}(y_0)\cap Q_{+}, \quad \text{where }Q_+:= \{y\in \R^{n+1}| \ y_{n}\geq 0, y_{n+1}\leq 0\}.$$
\end{defi}
\begin{rmk}
Due to Remark~\ref{rmk:equi_dist}, there are constants $c,C>0$ such that for any $y_0\in P$
\begin{align*}
\mathcal{\tilde{B}}_{cr}(y_0)\subseteq \mathcal{B}_r(y_0)\subseteq \mathcal{\tilde{B}}_{Cr}(y_0),\quad \text{where }\mathcal{\tilde{B}}_r(y_0)=\{y| d_G(y,y_0)< r\}.
\end{align*}
In the sequel, with slight abuse of notation, for $y_0\in P$ we will not distinguish between $\mathcal{\tilde{B}}_r(y_0)$ and $\mathcal{B}_r(y_0)$ for convenience of notation.
\end{rmk}
\subsection{Function spaces}
\label{sec:functions}
In the sequel, we consider the intrinsic H\"older spaces which are associated with the geometry introduced in Section \ref{sec:intrinsic}:
\begin{defi}
\label{defi:Hoelder}
Let $\Omega$ be a subset in $\R^{n+1}$ and let $\alpha\in (0,1]$. Then
\begin{align*}
C^{0,\alpha}_\ast(\overline{\Omega}):=\left\{u:\overline{\Omega}\rightarrow \R| \ \sup_{x,y\in \overline{\Omega}}\frac{|u(x)-u(y)|}{d_G(x,y)^\alpha}<\infty\right\}.
\end{align*}
Let
$$[u]_{\dot{C}^{0,\alpha}_\ast(\overline{\Omega})}:=\sup_{x,y\in \overline{\Omega}}\frac{|u(x)-u(y)|}{d_G(x,y)^\alpha}.$$
For $u\in C^{0,\alpha}_\ast(\overline{\Omega})$ we define
$$\|u\|_{C^{0,\alpha}_\ast(\overline{\Omega})}:=\|u\|_{L^\infty(\Omega)}+[u]_{C^{0,\alpha}_\ast(\overline{\Omega})}.$$
\end{defi}
\begin{rmk}
The mapping $\| \cdot \|_{C_{\ast}^{0,\alpha}}: C_{\ast}^{0,\alpha} \rightarrow [0,\infty)$ is a norm. By Remark \ref{rmk:equi_dist}
\begin{align*}
C^{0,\alpha}_{\ast}(\overline{\Omega}) \hookrightarrow C^{0,\frac{\alpha}{2}}(\overline{\Omega}).
\end{align*}
Hence, the pair $(C^{0,\alpha}_\ast(\overline{\Omega}), \| \cdot \|_{C^{0,\alpha}_{\ast}(\bar{\Omega})})$ is a Banach space.
\end{rmk}
Based on the spaces from Definition \ref{defi:Hoelder}, we can further define higher order H\"older spaces:
\begin{defi}
\label{defi:Hoelder1}
Let
$$\tilde{Y}_1=y_n\p_1, \quad \tilde{Y}_2=y_{n+1}\p_1, \quad \dots, \quad \tilde{Y}_{2n-1}=\p_n,\quad \tilde{Y}_{2n}=\p_{n+1}.$$
For $k\in \mathbb{N}$, $k\geq 1$, we say that $u\in C^{k,\alpha}_\ast(\overline{\Omega})$, if for all $\sigma_i\in \{1,\ldots, 2n\}$, $1\leq i\leq k$, the functions $u, \tilde{Y}_{\sigma_1}\cdots \tilde{Y}_{\sigma_i}u$ are continuous and $\tilde{Y}_{\sigma_1}\cdots\tilde{Y}_{\sigma_k}u \in C^{0,\alpha}_\ast(\overline{\Omega})$.
We define
\begin{equation}
\label{eq:norm1}
\begin{split}
\|u\|_{C^{k,\alpha}_\ast(\overline{\Omega})}& =\|u\|_{L^{\infty}(\overline{\Omega})} \\
&+ \sum_{j=1}^{k-1}\sum_{\sigma_1,\ldots,\sigma_j\in \{1,\dots, 2n\}}\|\tilde{Y}_{\sigma_1}\cdots \tilde{Y}_{\sigma_j}u\|_{L^\infty(\overline{\Omega})} \\
& +\sum_{\sigma_1,\ldots,\sigma_k\in \{1,\dots,2n\}}\|\tilde{Y}_{\sigma_1}\cdots \tilde{Y}_{\sigma_k}u\|_{C^{0,\alpha}_\ast(\overline{\Omega})}.
\end{split}
\end{equation}
\end{defi}
\begin{rmk}
The space $C^{k,\alpha}_{\ast}(\overline{\Omega})$ equipped with $\| \cdot \|_{C^{k,\alpha}_{\ast}(\overline{\Omega})}$ is a Banach space.\\
\end{rmk}
Building on the previously introduced Hölder spaces, we proceed to define the function spaces which we use to prove the higher regularity of the Legendre function $v$. These spaces, their building blocks and their role in our argument are reminiscent of the higher regularity approach of De Silva and Savin \cite{DSS14}. In contrast to the approach of De Silva and Savin we however use them in the \emph{linear} set-up in the sense that the (regular) free boundary has been fixed by the Legendre-Hodograph transform (at the expense of working with a degenerate (sub)elliptic, fully nonlinear equation). In this situation the approximation approach of De Silva and Savin simply becomes a Taylor expansion of our solution at the straightened free boundary.
Moreover, we do not carry out the expansion up to arbitrary order, but only up to order less than five. Beyond this we work with the implicit function theorem (c.f. Theorem \ref{prop:hoelder_reg_a} in Section \ref{sec:IFT1}), which is more suitable to the variable coefficients set-up. In particular, this restriction to an essentially leading order expansion with respect to the non-tangential variables allows us to avoid dealing with \emph{regularity issues in the non-tangential directions}. Working in a conical domain and with metrics and inhomogeneities which are not necessarily symmetric with respect to the non-tangential directions, we thus ignore potential higher-order singularities in the non-tangential variables. This has the advantage of deducing the desired partial regularity result in the tangential directions, which then entails the free boundary regularity, without having to deal with potentially arising non-tangential singularities.\\
Roughly speaking, our spaces interpolate between the regularity of the function at
$P=\{y_n=y_{n+1}=0\}$ (at which the Baouendi-Grushin operator is only degenerate elliptic)
and at $\{\frac{1}{2}<y_n^2+y_{n+1}^2< 2\}$ (in which the Baouendi-Grushin operator is uniformly elliptic region). In order to make this rigorous, we need the notion of an \emph{homogeneous polynomial}:
\begin{defi}[Homogeneous polynomials]
\label{defi:poly}
Let $k\in \N$. We define the \emph{space of homogeneous polynomials of degree less than or equal to $k$} as
\begin{align*}
\mathcal{P}_k=&\{p_k(y)| \ p_k(y)=\sum_{|\beta|\leq k}a_\beta y^{\beta},\\
&\text{ such that }a_\beta=0 \text{ whenever }\sum_{i=1}^{n-1}2\beta_i+\beta_n+\beta_{n+1}>k\}.
\end{align*}
Moreover, we define the \emph{space of homogeneous polynomials of degree exactly $k$} as
\begin{align*}
\mathcal{P}_k^{hom}=&\{p_k(y)| \ p_k(y)=\sum_{|\beta|\leq k}a_\beta y^{\beta},\\
&\text{ such that }a_\beta=0 \text{ whenever }\sum_{i=1}^{n-1}2\beta_i+\beta_n+\beta_{n+1}\neq k\}.
\end{align*}
\end{defi}
The definition of the homogeneous polynomials is motivated by the scaling properties of our operator $\Delta_G$. More precisely, we note the following dilation invariance property: if $u$ solves $\D_G u = f$, then the function $v(y):= u(\delta_\lambda(y))$, where $\delta_\lambda(y)=(\lambda^2y'',\lambda y_n,\lambda y_{n+1})$, solves
\begin{align*}
\D_G v = \lambda^2 f_\lambda,
\end{align*}
where $f_{\lambda}(y)= f(\delta_\lambda(y))$.
This motivates to count the order of the tangential variables $y''$ and the normal variables $y_n$, $y_{n+1}$ differently and define the homogeneous polynomials with respect to the Grushin scaling: $p_k(\delta_\lambda(y))=\lambda^k p_k(y)$ for $p_k\in \mathcal{P}_k^{hom}$.
\begin{rmk}
We observe that for instance $P\in \mathcal{P}_3$ is of the form
\begin{align*}
P(y)&=c_0+\sum_{i=1}^{n+1}a_iy_i+\sum_{k\in \{1,\dots,n-1\},\ell\in \{n,n+1\}}a_{k\ell}y_ky_\ell\\
&+ \left(c_1y_n^3+c_2y_n^2y_{n+1}+c_3y_ny_{n+1}^2+c_4y_{n+1}^3\right).
\end{align*}
\end{rmk}
Using the notion of homogeneous polynomials, we further define an adapted notion of differentiability at the co-dimension two hypersurface $P$:
\begin{defi}\label{defi:diff}
Let $k\in \N$ and $\alpha \in (0,1]$. Given a function $f$, we say that \emph{$f$ is $C^{k,\alpha}_\ast$ at $P$}, if at each $y_0\in P$ there exists an approximating polynomial $P_{y_0}(y)=\sum a_\beta(y_0)(y-y_0)^{\beta}\in \mathcal{P}_k$ such that
\begin{align*}
f(y)=P_{y_0}(y)+O(d_G(y,y_0)^{k+2\alpha}), \quad \text{as } y\rightarrow y_0.
\end{align*}
\end{defi}
\begin{rmk}
We note that for a multi-index $\beta$ satisfying $\sum_{i=1}^{n-1}2\beta_i+\beta_n+\beta_{n+1}\leq k$, the evaluation $\partial^{\beta}P_{y_0}(y_0) = \beta! a_{\beta}(y_0)$ corresponds to the (classical) $\beta$ derivative of $f$ at $y_0$, i.e. $\p^\beta f(y_0)=\beta! a_{\beta}(y_0)$.
\end{rmk}
With this preparation, we can finally give the definition of our function spaces:
\begin{defi}[Function spaces]
\label{defi:spaces}
Let $\epsilon, \alpha\in (0,1]$. Then,
\begin{align*}
X_{\alpha,\epsilon}:=&\{v\in C^{2,\epsilon}_\ast (Q_+) \cap C_0(Q_+)| \ \supp(\Delta_G v)\subset \mathcal{B}_1^+,\ v\text{ is } C^{3,\alpha}_\ast \text{ at } P, \\
&
v=0 \text{ on } \{y_n=0\},\ \p_{n+1}v=0\text{ on } \{y_{n+1}=0\}, \ \p_{nn}v=0 \text{ on } P,\\
& \text{and }\| v \|_{X_{\alpha, \epsilon}}<\infty\},\\
Y_{\alpha,\epsilon}:=&\{f\in C^{0,\epsilon}_\ast(Q_+)| \ \supp(f) \subset \mathcal{B}_1^+, \ f \text{ is } C^{1,\alpha}_\ast \text{ at }P, \ f=\p_{n+1}f=0 \text{ on }P,\\
& \text{and }\|f\|_{Y_{\alpha,\epsilon}}<\infty\}.
\end{align*}
The corresponding norms are defined as
\begin{align*}
\|f\|_{Y_{\alpha, \epsilon}}& : =\sup_{\bar y\in P}[d_G(\cdot,\bar y)^{-(1+2\alpha-\epsilon)}(f-P_{\bar y})]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+(\bar y))} ,\\
\text{ where } & P_{\bar y}(y)=y_n \p_nf(\bar y);\\
\| v \|_{X_{\alpha,\epsilon}} &:=\sup_{ \bar{y}\in P} \left(\|d_G(\cdot, \bar y)^{-(3+2\alpha)}(v-P_{\bar y})\|_{L^{\infty}(\mathcal{B}_3^+(\bar y))} \right.\\
& \left. +\sum\limits_{i,j=1}^{n+1}[d_G(\cdot, \bar y)^{-(1+2\alpha- \epsilon)}Y_i Y_j (v-P_{\bar y})]_{\dot{C}^{0,\epsilon}_{\ast}(\mathcal{B}_3^+(\bar y))}
+ [v]_{C^{2,\epsilon}_{\ast}(Q_+ \setminus \mathcal{B}_3^+(\bar y))}\right),\\
\text{ where } & P_{\bar y}(y)=\p_nv(\bar y)y_n + \sum_{i=1}^{n-1}\p_{in}v(\bar y)(y_i-\bar y_i) y_n +\frac{1}{6}\p_{nnn}v(\bar y)y_n^3\\
& \qquad +\frac{1}{2}\p_{n,n+1,n+1}v(\bar y)y_ny_{n+1}^2.
\end{align*}
\end{defi}
Let us discuss these function spaces $X_{\alpha,\epsilon}$: They are subspaces of the Baouendi-Grushin H\"older spaces $C^{2,\epsilon}_\ast(Q_+)$, with the additional properties that these functions are $C^{3,\alpha}_\ast$ along the edge $P$ and that they satisfy the symmetry conditions $v=0$ on $\{y_n=0\}$ and $\p_{n+1}v=0$ on $\{y_{n+1}=0\}$. The condition $\p_{nn}v=0$ on $P$ is a necessary compatibility condition which ensures that $\Delta_G$ maps $X_{\alpha,\epsilon}$ to $Y_{\alpha,\epsilon}$. The boundary conditions together with the $C^{3,\alpha}_\ast$ regularity allow us to conclude that any function in $X_{\alpha,\epsilon}$ has the same type of asymptotic expansion at $P$ as the Legendre function $v$. The support condition on $\Delta_G v$ together with the decay condition at infinity ($v\in C_0(Q_+)$ is a continuous function in $Q_+$ vanishing at infinity) is to ensure that $(X_{\alpha,\epsilon},\|\cdot\|_{X_{\alpha,\epsilon}})$ is a Banach space. \\
The spaces are the ones in which we apply the Banach implicit function theorem later in Section~\ref{sec:IFT1}. They are constructed in such a way as to
\begin{itemize}
\item[(i)] mimic the asymptotics behavior of our Legendre functions (which are defined in (\ref{eq:legendre})) around the straightened regular free boundary $P$. In particular, the Legendre functions $v$ associated with solutions $w$ of (\ref{eq:varcoeff}) are contained in the spaces $X_{\alpha,\epsilon}$ for a suitable range of $\alpha,\epsilon$ (c.f. Proposition \ref{prop:error_gain2}).
\item[(ii)] The spaces are compatible with the mapping properties of the fully nonlinear, degenerate, (sub)elliptic operator $F$ from (\ref{eq:nonlineq1}) (c.f. Proposition \ref{prop:nonlin_map}).
\item[(iii)] They are compatible with the linearization of the operator $F$ (c.f. Proposition \ref{prop:linear}). In particular they allow for ``Schauder type'' estimates for the Baouendi-Grushin Laplacian.
\end{itemize}
\begin{rmk}
\label{rmk:homo}
\begin{itemize}
\item[(i)] We note that by our support assumptions
\begin{align*}
& \|d_G(\cdot,\bar y)^{-(1+2\alpha)}(f-P_{\bar y})\|_{L^{\infty}(\mathcal{B}_3^+(\bar y))} \\
& \quad \leq C [d_G(\cdot,\bar y)^{-(1+2\alpha-\epsilon)}(f-P_{\bar y})]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+(\bar y))} .
\end{align*}
Similarly, by interpolation, we control all intermediate Hölder norms of $v-P_{\bar{y}}$ by $\| v\|_{X_{\alpha,\epsilon}}$.
\item[(ii)] We remark that the norms of $X_{\alpha,\epsilon}$ and $Y_{\alpha,\epsilon}$ only contain homogeneous contributions and do not include the lower order contributions which would involve the norms of the approximating polynomials $p(y):= \sum\limits_{|\alpha|\leq k} a_{\alpha}y^{\alpha}\in \mathcal{P}_k^{hom}$:
$$|p|_{k}:= \sum\limits_{\beta}|a_{\beta}| .$$ Yet, this results in Banach spaces as additional support conditions are imposed on $f, \D_G v$. The Banach space property is shown in Lemma~\ref{lem:Banach} in the Appendix.
\end{itemize}
\end{rmk}
For locally defined functions we use the following spaces:
\begin{defi}[Local function spaces]
\label{defi:spaces_loc}
Given $\alpha,\epsilon\in (0,1]$ and $R>0$.
\begin{align*}
X_{\alpha,\epsilon}(\mathcal{B}_R^+):=&\{v\in C^{2,\epsilon}_\ast(\mathcal{B}_R^+)| v\text{ is } C^{3,\alpha}_\ast \text{ at } P\cap \mathcal{B}_R, \\
&v=0 \text{ on } \{y_n=0\}\cap \mathcal{B}_R,\ \p_{n+1}v=0\text{ on }\{y_{n+1}=0\}\cap \mathcal{B}_R, \\
&\p_{nn}v=0\text{ on } P\cap \mathcal{B}_R \text{ and } \|v\|_{X_{\alpha,\epsilon}(\mathcal{B}_R^+)}<\infty\},
\end{align*}
where
\begin{align*}
\|v\|_{X_{\alpha,\epsilon}(\mathcal{B}_R^+)}:=\sup_{\bar y\in P\cap \mathcal{B}_R}\left(\sum_{i,j=1}^{n+1}[d_G(\cdot, \bar y)^{-(1+2\alpha-\epsilon)}Y_iY_j(v-P_{\bar y})]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+(\bar y)\cap \mathcal{B}_R^+)}\right.\\
\left.+\|d_G(\cdot, \bar y)^{-(3+2\alpha)}(v-P_{\bar y})\|_{L^{\infty}(\mathcal{B}_3^+(\bar y)\cap \mathcal{B}_R^+)}
+ |P_{\bar y}|_{3}\right),
\end{align*}
with $P_{\bar y}$ being as in Definition~\ref{defi:spaces}.\\
Similarly,
\begin{align*}
Y_{\alpha,\epsilon}(\mathcal{B}_R^+):=&\{f\in C^{0,\epsilon}_\ast(\mathcal{B}_R^+)| f \text{ is } C^{1,\alpha}_\ast \text{ at } P\cap \mathcal{B}_R,\\
&f=\p_{n+1}f=0\text{ on }P\cap \mathcal{B}_R \text{ and } \|f\|_{Y_{\alpha,\epsilon}(\mathcal{B}_R^+)}<\infty\},
\end{align*}
where
\begin{align*}
\|f\|_{Y_{\alpha,\epsilon}(\mathcal{B}_R^+)}:=\sup_{\bar y\in P\cap \mathcal{B}_R}\left(\| d_G(\cdot, \bar y)^{-(1+2\alpha)}(f-P_{\bar y})\|_{L^{\infty}(\mathcal{B}_3^+(\bar y)\cap \mathcal{B}_R^+)} \right.\\
\left. + [d_G(\cdot, \bar y)^{-(1+2\alpha)}(f-P_{\bar y})]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+(\bar y)\cap \mathcal{B}_R^+)} +|P_{\bar y}|_1 \right),
\end{align*}
with $P_{\bar y}$ being as in Definition~\ref{defi:spaces}.
\end{defi}
For the functions in $X_{\alpha,\epsilon}$ and $Y_{\alpha,\epsilon}$, the following characterization will be useful. We postpone the proof to the Appendix, Section \ref{sec:decomp}.
\begin{prop}[Characterization of $X_{\alpha,\epsilon}$ and $Y_{\alpha,\epsilon}$]
\label{prop:decompI}
Let $v\in X_{\alpha,\epsilon}$ and $f\in Y_{\alpha,\epsilon}$ and $2\alpha>\epsilon$. Let $r=r(y):=\sqrt{y_n^2+y_{n+1}^2}$ denote the distance from $y$ to $P$. Let $y'':=(y'',0,0)\in P$.
\begin{itemize}
\item[(i)] Then $\p_n f(y'')\in C^{0,\alpha}(P)$. Moreover, there exists $f_1(y)\in C^{0,\epsilon}_{\ast}(Q_+)$ vanishing on $P$, such that for $y\in \mathcal{B}_3^+$
\begin{align*}
f(y)=\p_n f(y'')y_n+ r^{1+2\alpha-\epsilon}f_1(y).
\end{align*}
\item[(ii)] Then $\p_nv(y'')\in C^{1,\alpha}(P)$, $\p_{nnn}v(y''), \p_{n,n+1,n+1}v(y'') \in C^{0,\alpha}(P)$. Moreover, there exist functions $C_1,V_i, C_{ij}\in C^{0,\epsilon}_{\ast}(Q_+)$, $i,j\in\{1,\dots,n+1\}$, vanishing on $P$, such that for $y\in \mathcal{B}_3^+$
\begin{align*}
v(y) &= \p_n v(y'') y_n + \frac{\p_{nnn}v(y'')}{6}y_n^3 +\frac{\p_{n,n+1,n+1}v(y'')}{2}y_ny_{n+1}^2+ r^{3+2\alpha-\epsilon}C_1(y), \\
\p_{i}v(y)&= \p_{in}v(y'') y_n + r^{1+2\alpha-\epsilon}V_i(y),\quad i\in \{1,\dots, n-1\},\\
\p_{n}v(y)& = \p_{n}v(y'') + \frac{\p_{nnn}v(y'')}{2}y_n^2 + \frac{\p_{n,n+1,n+1}v(y'')}{2}y_{n+1}^2 + r^{2+2\alpha-\epsilon}V_n(y),\\
\p_{n+1}v(y) & = \p_{n,n+1,n+1}v(y'') y_n y_{n+1} + r^{2+2\alpha-\epsilon}V_{n+1}(y),\\
\p_{ij}v(y)&=r^{-1+2\alpha-\epsilon}C_{ij}(y),\\
\p_{in}v(y)&= \p_{in}v(y'') + r^{2\alpha-\epsilon}C_{in}(y),\\
\p_{i,n+1}v(y)&=r^{2\alpha-\epsilon}C_{i,n+1}(y),\\
\p_{n,n}v(y)&= \p_{nnn}v(y'') y_n + r^{1+2\alpha-\epsilon}C_{n,n}(y),\\
\p_{n,n+1}v(y)&=\p_{n,n+1,n+1}v(y'')y_{n+1}+r^{1+2\alpha-\epsilon}C_{n,n+1}(y),\\
\p_{n+1,n+1}v(y)&=\p_{n,n+1,n+1}v(y'')y_{n} + r^{1+2\alpha-\epsilon}C_{n+1,n+1}(y).
\end{align*}
\end{itemize}
Moreover, $C_1(y)=0=V_{n+1}(y)$ on $\{y_n=0\}$.
For the decompositions in (i) and (ii) we have
\begin{align*}
&[\p_{n}f]_{\dot{C}^{0,\alpha}(P\cap \mathcal{B}_3^+)}+ [f_1]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)}\leq C \|f\|_{Y_{\alpha,\epsilon}},\\
&[\p_{in}v]_{\dot{C}^{0,\alpha}(P\cap \mathcal{B}_3^+)} +[\p_{nnn}v]_{\dot{C}^{0,\alpha}(P\cap \mathcal{B}_3^+)}+[\p_{n,n+1,n+1}v]_{\dot{C}^{0,\alpha}(P\cap \mathcal{B}_3^+)}\\
& \quad + \sum\limits_{i,j=1}^{n+1}[C_{ij}]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)}
+ \sum\limits_{j=1}^{n+1}[V_{j}]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)}
+ [C_1]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)} \leq C \|v\|_{X_{\alpha,\epsilon}}.
\end{align*}
\end{prop}
\begin{rmk}
\label{rmk:characterize}
It is immediate that any $f\in C^{0,\epsilon}_\ast(Q_+)$ with $\supp(f)\subset \mathcal{B}_3^+$ which satisfies the decomposition in (i) is in $Y_{\alpha,\epsilon}$. Moreover,
$$\|f\|_{Y_{\alpha,\epsilon}}\leq C\left([\p_{n}f]_{\dot{C}^{0,\alpha}(P)}+ [f_1]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)}\right).$$
Similarly, it is not hard to show that functions $v$ satisfying the decomposition in (ii) with $\supp(\Delta_Gv)\subset \mathcal{B}_3^+$ are in $X_{\alpha,\epsilon}$. In particular, this implies that Proposition \ref{prop:decompI} gives an equivalent characterization of the spaces $X_{\alpha,\epsilon}$ and $Y_{\alpha,\epsilon}$.\\
Motivated by the decomposition of Proposition \ref{prop:decompI}, we sometimes also write
\begin{align*}
Y_{\alpha,\epsilon} = y_n C^{0,\alpha} + r^{1+2\alpha- \epsilon} C_{\ast}^{0,\epsilon}.
\end{align*}
\end{rmk}
At the end of this section, we state the following a priori estimate (which should be viewed as a Schauder type estimate for the Baouendi-Grushin operator):
\begin{prop}
\label{prop:invert}
Let $\alpha \in (0,1)$, $\epsilon \in (0,1)$, $v\in X_{\alpha,\epsilon}$, $f\in Y_{\alpha,\epsilon}$ and
\begin{align*}
\D_G v = f.
\end{align*}
Then we have
\begin{align*}
\| v\|_{X_{\alpha,\epsilon}} \leq C \|f\|_{Y_{\alpha,\epsilon}}.
\end{align*}
\end{prop}
\begin{rmk}
We note that the a priori estimates exemplify a scaling behavior which depends on the support of the respective function. More precisely, let $=\D_G v$ be supported in $\mathcal{B}_{\mu}^+$ for some $\mu>0$. Then,
\begin{align*}
\|v\|_{X_{\alpha,\epsilon}} \leq C(1+|\mu|^{1+2\alpha-\epsilon}) \|\D_G v\|_{Y_{\alpha,\epsilon}}.
\end{align*}
\end{rmk}
The proof of Proposition \ref{prop:invert} follows by exploiting the scaling properties of our operator and polynomial approximations. This method is in analogy to the Campanato approach (c.f. \cite{Ca64}) to prove Schauder estimates for the Laplacian (c.f. also \cite{Gia}, \cite{Wang92} and \cite{Wa03} for generalizations to elliptic systems, fully nonlinear elliptic and parabolic equations and certain subelliptic equations). Our generalization of these spaces is adapted to our thin free boundary problem. We postpone the proof of Proposition \ref{prop:invert} to the Appendix, Section \ref{sec:quarter_Hoelder}.
\section[Regularity of the Legendre Function]{Regularity of the Legendre Function for $C^{k,\gamma}$ Metrics}
\label{sec:improve_reg}
In this Section we return to the investigation of the Legendre function. We recall that in Section~\ref{sec:Legendre} we transformed the free boundary problem (\ref{eq:varcoeff}) into a fully nonlinear Baouendi-Grushin type equation for the Legendre function $v$. In this section, we will study the regularity of the Legendre function $v$ in terms of the function spaces $X_{\alpha,\epsilon}$ from Section~\ref{sec:holder}.\\
In the whole section we assume that the metrics $a^{ij}$ are $C^{1,\gamma}$ Hölder regular for some $\gamma\in (0,1)$. We start by showing that the Legendre function $v$ (associated with a solution $w$ to \eqref{eq:thin_obst}) is in the space $X_{\alpha,\gamma}$ introduced in Section~\ref{sec:holder}. This is a consequence of transferring the asymptotics of $w$ (which were derived in Propositions \ref{prop:asym2}, \ref{prop:improved_reg} in Section~\ref{sec:asymp}) to $v$. Here $\alpha$ is the (a priori potentially very small) H\"older exponent of the free boundary from Section~\ref{sec:asymp}.\\
In Section~\ref{subsec:improvement} (c.f. Propositions \ref{prop:error_gain}, \ref{prop:error_gain2}), we exploit the structure of our nonlinear equation to improve the regularity of $v$ from $X_{\alpha,\epsilon}$ to $X_{\delta,\epsilon}$ for any $\delta\in (0,1)$ and $\epsilon \in (0,\gamma]$. In particular this implies that the free boundary $\Gamma_w$ is $C^{1,\delta}$ regular for any $\delta\in (0,1)$.
\subsection{Asymptotics of the Legendre function}
\label{sec:Leg}
Throughout this section we assume that $w$ solves the thin obstacle problem \eqref{eq:varcoeff} with coefficients $a^{ij}\in C^{k,\gamma}$ for $k\geq 1$ and $\gamma\in (0,1]$. Moreover, we always assume that the conditions (A1)-(A7) are satisfied. We further recall from Section~\ref{sec:Legendre} that the Legendre function $v$ is originally defined on $U=T(B_{1/2}^+)$. However, after a rescaling procedure we may assume that $v$ is defined in $\mathcal{B}_2^+$. \\
We first rewrite the asymptotics of $w$ (stated in Corollary~\ref{cor:improved_reg}) in terms of the corresponding Legendre function $v$.
\begin{prop}\label{prop:holder_v}
Let $a^{ij}\in C^{k,\gamma}$ with $k\geq 1$ and $\gamma\in (0,1)$. Let $w$ be a solution to the variable coefficient thin obstacle problem and let $v$ be its Legendre function as defined in \eqref{eq:legendre}. Assume that $w$ satisfies the asymptotic expansion in Proposition~\ref{prop:improved_reg} with some $\alpha\in (0,1)$.
Then for any $\hat y\in \mathcal{B}_1^+$ with $\sqrt{\hat y_n^2+\hat y_{n+1}^2}=:\lambda \in (0,1)$, and any multi-index $\beta$ with $|\beta|\leq k+1$
\begin{align*}
\left[D^{\beta}v-D^{\beta} v_{\hat y''}\right]_{\dot{C}^{0,\gamma}_\ast(\mathcal{B}_{\lambda/4}^+(\hat y))} \leq C(|\beta|,n,p) \max\{\epsilon_0, c_*\} \lambda^{3+2\alpha-\gamma-2|\beta''|-|\beta_n|-|\beta_{n+1}|},
\end{align*}
where $v_{\hat y''}$ is the leading order expansion of $v$ at $\hat y''$ as defined in Lemma~\ref{lem:asymp_profile}.
\end{prop}
\begin{proof}
We argue in three steps in which we successively simplify the problem:\\
\emph{Step 1: First reduction -- Scaling.} Given $\hat y\in \mathcal{B}_1^+$ with $\sqrt{\hat y_n^2+\hat y_{n+1}^2}=\lambda>0$, we project $\hat y$ onto $P=\{y_n=y_{n+1}=0\}$ and let $\hat y''=(\hat y'',0,0)$ denote the projection point.
Let
$$\hat v_{\hat y'',\lambda}(\zeta):=\frac{v(\hat y''+\delta_\lambda( \zeta))-\p_nv(\hat y'')(\lambda \zeta_n)}{\lambda^3}, $$
with $\delta_\lambda (\zeta)=(\lambda^2\zeta'',\lambda \zeta_n,\lambda \zeta_{n+1})$.
We note that
$\hat v_{\hat y'',\lambda}$ is the Legendre function for $w_{x_0,\lambda^2}(\xi):=w(x_0+\lambda^2\xi)/\lambda^3$ with $x_0=T^{-1}(\hat y'')$ . We set
$$\hat \zeta=\left(0,\frac{\hat y_n}{\lambda}, \frac{\hat y_{n+1}}{\lambda}\right)\in B''_1\times \mathcal{S}^1.$$
With this rescaling, it suffices to show that
\begin{align}\label{eq:deri_v}
\left[ D^{\beta} \hat v_{\hat y'',\lambda}- D^{\beta} \hat v_{\hat y''} \right]_{\dot{C}^{0,\gamma}_\ast(\mathcal{B}_{1/4}^+(\hat \zeta))}\leq C(|\beta|,n,p) \max\{\epsilon_0, c_*\} \lambda^{2\alpha},
\end{align}
where $\hat v_{\hat{y}''}$ denotes the Legendre function for $w_{x_0}(\xi)=\lim_{\lambda\rightarrow 0_+} w_{x_0,\lambda^2}(\xi)$.
Indeed, the conclusion of Proposition \ref{prop:holder_v} then follows by undoing the rescaling in \eqref{eq:deri_v}.\\
\emph{Step 2: Second reduction.} Given any multi-index $\beta$ with $|\beta |\leq k+1$, it is possible to express $D^{\beta}\hat v_{\hat y'',\lambda}$ as a function of $w_{x_{0},\lambda^2}$ and its derivatives:
$$D^\beta \hat v_{\hat y'',\lambda}(y)= F_\beta(D^{\tilde{\alpha}} w_{x_{0},\lambda^2}(x))\big|_{x=(T^{w_{x_0,\lambda^2}})^{-1}(y)},\quad |\tilde{\alpha}|\leq |\beta|.$$
Here $F_\beta$ is an analytic function on the open set $\{J(w_{x_{0},\lambda^2})\neq 0\}$. Let $\hat \xi=(T^{w_{x_{0},\lambda^2}})^{-1}(\hat \zeta)$. We note that a sufficiently small choice of $\lambda_0\in(0,1)$ implies that our change of coordinates is close to the square root mapping (c.f. Proposition \ref{prop:asym2}). Combining this with the observation that on scales of order one (i.e. when $\zeta_n^2+\zeta_{n+1}^2\sim 1$) the Baouendi-Grushin metric is equivalent to the Euclidean metric, results in the inclusions $(T^{w_{x_{0},\lambda^2}})^{-1}(\mathcal{B}_{1/4}(\hat \zeta)), (T^{w_{x_{0}}})^{-1}(\mathcal{B}_{1/4}(\hat \zeta))\subset B_{1/2}(\hat \xi)$ and $B_{3/4}(\hat \xi)\cap \Gamma_{w_{x_{0},\lambda^2}}=\emptyset$, $B_{3/4}(\hat \xi)\cap \Gamma_{w_{x_{0}}}=\emptyset$. Thus, to show \eqref{eq:deri_v}, it then suffices to prove
\begin{align}\label{eq:multi_deri}
\left[F_\beta(D^{\tilde{\alpha}} w_{x_{0},\lambda^2} )- F_\beta(D^{\tilde{\alpha}} \mathcal{W}_{x_0}(x_0 +\cdot))\right]_{\dot{C}^{0,\gamma}(B_{1/2}^+(\hat \xi))} \leq C(|\beta|,n,p) \max\{\epsilon_0, c_*\} \lambda^{2\alpha},
\end{align}
and
\begin{align}\label{eq:inverse_T}
\|(T^{w_{x_0,\lambda^2}})^{-1}-(T^{w_{x_0}})^{-1}\|_{C^{1}(\mathcal{B}^+_{3/4})}\leq C \max\{\epsilon_0, c_*\} \lambda^{2\alpha}
\end{align}
for $\lambda \in (0,\lambda_0)$.
Indeed, once we have obtained \eqref{eq:multi_deri}-\eqref{eq:inverse_T}, \eqref{eq:deri_v} follows by the equivalence of the Euclidean and Baouendi-Grushin geometries at scales of order one and a triangle inequality.\\
\emph{Step 3: Proof of (\ref{eq:multi_deri}) and conclusion.}
To show \eqref{eq:multi_deri}, we use a Taylor expansion and write
$$F_\beta(D^{\tilde{\alpha}} w_{x_{0},\lambda^2})-F_\beta(D^{\tilde{\alpha}} \mathcal{W}_{x_0}(x_0 + \cdot))=R_{\tilde{\alpha}} (w_{x_{0},\lambda^2})(D^{\tilde{\alpha}} w_{x_{0},\lambda^2}-D^{\tilde{\alpha}} \mathcal{W}_{x_0}(x_0 + \cdot)),$$
where
$$R_{\tilde{\alpha}}(w_{x_{0},\lambda^2})(x)=\int_0^1\p_{m_{\tilde{\alpha}}}F_\beta(tD^{\tilde{\alpha}}w_{x_{0},\lambda^2}(x)+(1-t)D^{\tilde{\alpha}}\mathcal{W}_{x_0}(x_0 + x))dt.$$
Since $-C\leq J(tw_{x_{0},\lambda^2}+(1-t)\mathcal{W}_{x_0})\leq -c$ in $B_{1/2}^+(\hat \xi)$ and since $w_{x_{0},\lambda^2}, \mathcal{W}_{x_0}(x_0 + \cdot) \in C^{k+1,\gamma}(B_{1/2}^+(\hat \xi))$, we have that $R_{\tilde{\alpha}}(w_{x_{0},\lambda^2})\in C^{0, \gamma}(B_{1/2}^+(\hat \xi))$ with uniform bounds in $\lambda$.
Next we recall that by Corollary \ref{cor:improved_reg}
$$\left[D^{\tilde{\alpha}}w_{x_{0},\lambda^2}-D^{\tilde{\alpha}}\mathcal{W}_{x_0}(x_0 + \cdot) \right]_{\dot{C}^{0,\gamma}(B_{1/4}^+(\hat{\xi}))} \leq C(\beta)\max\{ \epsilon_0, c_{\ast}\} \lambda^{2\alpha}.$$ Combining this with the fact that $R_{\tilde{\alpha}}(w_\tau)(\xi)\in C^{0, \gamma}(B_{1/4}^+(\hat \xi))$, yields \eqref{eq:multi_deri}.
To show \eqref{eq:inverse_T}, we first observe that $\|T^{w_{x_0,\lambda^2}}-T^{w_{x_0}}\|_{C^1(B_{3/4}(\hat \xi))}\leq C \max\{\epsilon_0, c_*\}\lambda^{2\alpha}$ by Corollary~\ref{cor:improved_reg} and the definition of $T$. Then using the uniform boundedness of $\|D (T^{w_{x_0,\lambda^2}})^{-1}\|_{L^\infty(\mathcal{B}^+_{1/2}(\hat \xi))}$ and $\|D (T^{w_{x_0}})^{-1}\|_{L^\infty(\mathcal{B}^+_{1/2}(\hat \xi))}$ we obtain the desired estimate.
\end{proof}
Using the spaces from Definition \ref{defi:spaces}, we apply Proposition \ref{prop:holder_v} to quantify the regularity and asymptotics of our Legendre function:
\begin{prop}
\label{prop:regasymp}
Under the assumptions of Proposition~\ref{prop:holder_v} we have $v\in X_{\alpha, \mu}(\mathcal{B}_{1}^+)$
for all $\mu\in(0,\gamma]$. In particular, for $y_0\in P\cap \mathcal{B}_{1}$ there exist functions $C_{k\ell}\in C^{0,\gamma}_{\ast}(\mathcal{B}_{1}^+(y_0))$ with $C_{k l}(y'',0,0)=0$ for all $k,l \in \{1,\dots,n+1\} $ such that the following asymptotics are valid:
\begin{align*}
\p_{ij}v(y)& = (y_n^2+y_{n+1}^2)^{-1}d_G(y,y_0)^{1+2\alpha-\gamma}C_{ij}(y),\\
\p_{in}v(y)&=\frac{(e_i\cdot \nu_{T^{-1}(y_0)})}{(e_n\cdot \nu_{T^{-1}(y_0)})}+d_G(y,y_0)^{2\alpha-\gamma}C_{in}(y),\\
\p_{i,n+1}v(y)&= d_G(y,y_0)^{2\alpha-\gamma}C_{i,n+1}(y),\\
\begin{pmatrix}
\partial_{nn} v(y) & \partial_{n,n+1} v(y)\\
\partial_{n+1,n} v(y) & \partial_{n+1,n+1}v(y)
\end{pmatrix}
&=\begin{pmatrix} a_0(y_0) y_n & a_1(y_0)y_{n+1}\\
a_1(y_0)y_{n+1}& a_1(y_0)y_n
\end{pmatrix} \\
& \quad + d_G(y,y_0)^{1+2\alpha- \gamma}\begin{pmatrix} C_{nn}(y)& C_{n,n+1}(y)\\
C_{n,n+1}(y)& C_{n+1,n+1}(y)
\end{pmatrix} .
\end{align*}
Here $i,j\in\{1,\dots,n-1\}$.
\end{prop}
\begin{rmk}
We emphasize that in the expression for $\p_{ij}v$ it is not possible to replace $(y_n^2 + y_{n+1}^2)^{-1}$ by $d_G(y,y_0)^{-2}$. This is in analogy with Remark \ref{rmk:improved_reg}.
\end{rmk}
\begin{proof}
For each $y_0\in P$ the proof of the asymptotics in the non-tangential cone $\mathcal{N}_G(y_0)$ follows directly from Proposition~\ref{prop:holder_v}, Lemma \ref{lem:asymp_profile} (which yields the explicit expressions of the leading order asymptotic expansions) and a chain of balls argument. In order to obtain the asymptotic expansion in the whole of $\mathcal{B}_{1}^+(y_0)$, we use the regularity of the coefficient functions and the triangle inequality. We only present the argument for $\p_{in}v$, since the reasoning for the other partial derivatives is analogous. Hence, let $y\in \mathcal{B}_{1}^+(y_0)$ but $y\notin \mathcal{N}_G(y_0)$. Let $\bar{y}\in P\cap \mathcal{B}_1$ denote the projection of $y$ onto $P$. Then, by the triangle inequality, we may assume that $d_G(y,\bar{y}), d_G(\bar{y},y_0)\leq C d_G(y,y_0)$. Thus, by virtue of the regularity of $\nu_{T^{-1}(y_0)}$, we have that
\begin{align*}
\left|\p_{in}v(y)- \frac{(e_i\cdot \nu_{T^{-1}(y_0)})}{(e_n\cdot \nu_{T^{-1}(y_0)})} \right| &\leq \left| \p_{in}v(y) -\frac{(e_i\cdot \nu_{T^{-1}(\bar{y})})}{(e_n\cdot \nu_{T^{-1}(\bar{y})})}\right| \\
& \quad + \left| \frac{(e_i\cdot \nu_{T^{-1}(\bar{y})})}{(e_n\cdot \nu_{T^{-1}(\bar{y})})} -\frac{(e_i\cdot \nu_{T^{-1}({y_0})})}{(e_n\cdot \nu_{T^{-1}({y_0})})} \right|\\
& \leq C d_G(y,\bar{y})^{2\alpha} +C|\bar{y}- y_0|^{\alpha} \leq C d_G(y,y_0)^{2\alpha}.
\end{align*}
The Hölder estimates are analogous.
\end{proof}
\begin{rmk}
\label{rmk:close}
For later reference we conclude this section by noting that the closeness condition (A5) (and the asymptotics from Proposition \ref{prop:holder_v}) implies that
$$\left\| v- v_{0}\right\|_{X_{\alpha,\epsilon}(\mathcal{B}_{1}^+)} \leq C\max\{\epsilon_0,c_\ast\},$$
where $v_0(y)= -\frac{1}{3}(y_n^3-3y_ny_{n+1}^2)$ is the leading order expansion of $v$ at the origin.
This will be used in the perturbation argument in Section \ref{sec:grushin} and in the application of the implicit function theorem in Section \ref{sec:IFT1}.
\end{rmk}
\subsection{Improvement of regularity}
\label{subsec:improvement}
In this section we present a bootstrap argument to infer higher regularity of the Legendre function. By virtue of the previous section, we have that $v\in X_{\alpha,\epsilon}(\mathcal{B}_1^+)$ for some potentially very small value of $\alpha\in(0,\gamma]$. In this section we improve this regularity modulus further by showing that $\alpha$ can be chosen arbitrarily close to one. To this end we argue in two steps: By an expansion, we first identify the structure of $F$ in terms of a leading order linear operator and additional higher order controlled contributions (c.f. Proposition \ref{prop:error_gain}). Then in a second step, we use this to bootstrap regularity (c.f. Proposition \ref{prop:error_gain2}). \\
In the sequel we use the following abbreviations:
\begin{align*}
G^{ij}(v)&:=-\det\begin{pmatrix}
\p_{ij}v& \p_{in}v & \p_{i,n+1}v\\
\p_{jn}v& \p_{nn}v & \p_{n,n+1}v\\
\p_{j,n+1}v & \p_{n,n+1}v &\p_{n+1,n+1}v
\end{pmatrix}, \ i,j\in\{1,\dots,n-1\},\\
G^{i,n}(v)&:=2\det\begin{pmatrix}
\p_{in}v & \p_{i,n+1}v\\
\p_{n,n+1}v & \p_{n+1,n+1}v
\end{pmatrix}, \ i\in\{1,\dots,n-1\},\\
G^{i,n+1}(v)&:=2\det\begin{pmatrix}
\p_{i,n+1}v & \p_{in}v\\
\p_{n,n+1}v & \p_{nn}v
\end{pmatrix}, \ i\in\{1,\dots,n-1\},\\
G^{n,n}(v)&:=\p_{n+1,n+1}v,\\
G^{n+1,n+1}(v)&:=\p_{nn}v,\\
G^{n,n+1}(v)&:=-\p_{n,n+1} v,\\
J(v)&:=\det\begin{pmatrix}
\p_{nn}v &\p_{n,n+1}v\\
\p_{n,n+1}v &\p_{n+1,n+1}v
\end{pmatrix}.
\end{align*}
With slight abuse of notation, we thus interpret $G^{ij}$ and $J$ as functions from the symmetric matrices $\R^{(n+1)\times(n+1)}_{sym}$ to $\R$ and recall the notation for partial derivatives of $G^{ij}$ with respect to the components $m_{k\ell}$ from Section \ref{sec:notation}:
\begin{align*}
\p_{m_{k\ell}}G^{ij}(M)=\frac{\p G^{ij}(M)}{\p m_{k\ell}}, \ M=(m_{k\ell})\in \R^{(n+1)\times (n+1)}_{sym}.
\end{align*}
With these conventions, the nonlinear equation \eqref{eq:nonlineq1} from Section~\ref{sec:Legendre}, which is satisfied by $v$, turns into
\begin{align}\label{eq:nonlin_2}
F(v,y):=\sum_{i,j=1}^{n+1}\tilde{a}^{ij}(y)G^{ij}(v)-J(v)\left(\sum_{j=1}^{n-1}\tilde{b}^j\p_j v+\tilde{b}^ny_n+\tilde{b}^{n+1}y_{n+1}\right)=0.
\end{align}
Relying on this structure, we derive a (self-improving) linearization. More precisely,
for each $y_0\in P\cap \mathcal{B}_{1}^+$, we will linearize the equation at $v_{y_0}$, where
\begin{equation}\label{eq:v0}
\begin{split}
v_{y_0}(y)&=\p_nv(y_0)y_n+\sum_{i=1}^{n-1}\p_{in}v(y_0)(y_i-(y_0)_i)y_n\\
&+\frac{\p_{nnn}v(y_0)}{6}y_n^3+\frac{\p_{n,n+1,n+1}v(y_0)}{2}y_ny_{n+1}^2,
\end{split}
\end{equation}
is the up to order three asymptotic expansion of $v$ at $y_0$.
The linearization then leads to the following self-improving structure:
\begin{prop}\label{prop:error_gain}
Let $a^{ij}(x)\in C^{1,\gamma}$ for some $\gamma\in (0,1]$.
Assume that $v\in X_{\alpha,\epsilon}(\mathcal{B}_{1}^+)$ with $\alpha\in (0,1]$ solves $F(v,y)=0$. Then at each point $y_0\in P\cap \mathcal{B}_{1/2}^+$, we have the following expansion in $\mathcal{B}_r^+(y_0)$ with $0<r<1/2$:
\begin{align*}
F(v,y)=L_{y_0}v + P_{y_0}(y)+E_{y_0}(y).
\end{align*}
Here
$$L_{y_0}v=\tilde{a}^{ij}(y_0)\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}v=D_vF\big|_{(v,y)=(v_{y_0},y_0)} v,$$
$$P_{y_0}(y)=\tilde{a}^{ij}(y_0)\left(G^{ij}(v_{y_0})-\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}v_{y_0}\right)\in \mathcal{P}_{1}^{hom},$$
$P_{y_0}(y)$ is of the form $c_0(y_0)y_n$ and $E_{y_0}(y)$ is an error term satisfying
\begin{equation*}
\begin{split}
&\left\|d_G(\cdot,y_0)^{-\eta_0}E_{y_0}\right\|_{L^\infty(\mathcal{B}_{1/2}^+(y_0))}+\left[d_G(\cdot,y_0)^{-(\eta_0-\epsilon)}E_{y_0}\right]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_{1/2}^+(y_0))}\leq C,
\end{split}
\end{equation*}
for $\eta_0=\min\{1+4\alpha,3\}$.
\end{prop}
\begin{rmk}[The role of $\alpha$, $2\alpha$ and $4\alpha$]
As already seen in Proposition~\ref{prop:decompI}, the parameter $\alpha$ in the space $Y_{\alpha,\epsilon}$ refers to the (tangential) H\"older regularity (w.r.t. the Euclidean metric) of the quotient $\frac{f}{y_n}\big|_P$ for any function $f\in Y_{\alpha,\epsilon}$. The parameter $2\alpha$ originates from the different scalings of the Euclidean metrics and Baouendi-Grushin metrics (c.f. Remark~\ref{rmk:equi_dist}). More precisely, for any $y''_1,y''_2\in P$, $|y''_1-y''_2|\sim d_G(y''_1,y''_2)^{2}$ by Remark~\ref{rmk:equi_dist}, which accounts for the $2\alpha$ in the definition of the norm $\|f\|_{Y_{\alpha,\epsilon}}=\sup_{\bar y\in P}[d_G(\cdot, \bar y)^{-(1+2\alpha-\epsilon)}(f-P_{\bar y})]_{\dot{C}^{0,\epsilon}_\ast}$. The parameter $4\alpha$ in Proposition~\ref{prop:error_gain} indicates an \emph{improvement} of the tangential regularity of $L_{y_0}v/y_n$ at $y_0$ from $C^{0,\alpha}$ to $C^{0,2\alpha}$.
\end{rmk}
\begin{rmk}\label{rmk:error_gain3}
By using the explicit expression of $v_{y_0}$ and $G^{ij}(v)$, it is possible to compute the form of the leading order operator:
\begin{align*}
L_{y_0}&:=\sum_{i,j=1}^{n+1}\tilde{a}^{ij}(y_0)\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}\\
&=\sum_{i,j=1}^{n-1}\tilde{a}^{ij}\left((A_1)^2y_{n+1}^2-A_0A_1y_n^2\right)\p_{ij}\\
&\quad +2\sum_{i,j=1}^{n-1}\tilde{a}^{ij}\left(B_jA_1y_n\p_{in}-B_jA_1y_{n+1}\p_{i,n+1}\right)+\sum_{i,j=1}^{n-1}\tilde{a}^{ij}B_iB_j\p_{n+1,n+1}\\
&\quad +2\sum_{i=1}^{n-1}\tilde{a}^{in}\left(A_1y_n\p_{in}-A_1y_{n+1}\p_{i,n+1}+B_i\p_{n+1,n+1}\right)\\
& \quad +\tilde{a}^{nn}\p_{n+1,n+1}+\tilde{a}^{n+1,n+1}\p_{n,n}.
\end{align*}
Here the coefficients $\tilde{a}^{ij}$ are evaluated at $y_0$ and $A_0,A_1,B_j$ are constants depending on $y_0$:
\begin{align*}
A_0&:=\p_{nnn}v(y_0),\quad A_1:=\p_{n,n+1,n+1}v(y_0),\\
B_j&:=\p_{jn}v(y_0),\quad j\in \{1,\dots, n-1\}.
\end{align*}
To obtain this, we have used the off-diagonal assumption (A3) for the metric $a^{ij}$, i.e. $a^{i,n+1}(x',0)=0$ for $i\in\{1,\dots, n\}$.\\
We note that the operator $L_{y_0}$ is a self-adjoint, constant coefficient Baouendi-Grushin type operator. It is hypoelliptic as an operator on $\R^{n+1}$ after an odd reflection in the $y_n$ variable and an even reflection in the $y_{n+1}$ variable (c.f. \cite{JSC87}).
\end{rmk}
\begin{proof}[Proof of Proposition~\ref{prop:error_gain}]
The proof of this result relies on a successive expansion of the coefficients and the nonlinearities.
Thus, we first expand $\tilde{a}^{ij}(y)$, $G^{ij}(v)$ and the lower order term $J(v)\left(\cdots\right)$ in \eqref{eq:nonlin_2} separately and then combine the results to derive the desired overall expansion. \\
\emph{Step 1: Expansion of the leading term.}
\emph{Step 1 a: Expansion of the coefficients.} For the coefficients $\tilde{a}^{ij}$ we have
\begin{align*}
\tilde{a}^{ij}(y)=\tilde{a}^{ij}(y_0)+E_2^{y_0,ij}(y),
\end{align*}
where
the error term $E^{y_0,ij}_2(y)$ satisfies
\begin{equation}
\label{eq:Holderweight0}
\begin{split}
\left\|d_G(\cdot,y_0)^{-2}E^{y_0,ij}_2\right\|_{L^\infty(\mathcal{B}_{1}^+(y_0))}
+\left[d_G(\cdot,y_0)^{-(2-\epsilon)}E^{y_0,ij}_2\right]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_{1}^+(y_0))}\leq C.
\end{split}
\end{equation}
\emph{Proof of Step 1a:}
The claim follows from the differentiability of $a^{ij}(x)$ and the asymptotics of $\nabla v$. Using the abbreviations $\xi(y):=(y'',-\p_nv(y),-\p_{n+1}v(y))$ and $\xi_0:=(y_0'',-\p_{n}v(y_0),-\p_{n+1}v(y_0))=(y_0'',-\p_nv(y_0),0)$, we obtain
\begin{equation}
\label{eq:expansion_1}
\begin{split}
\tilde{a}^{ij}(y)=a^{ij}(\xi(y))&=a^{ij}(\xi_0)+ (a^{ij}(\xi)-a^{ij}(\xi_0))\\
&=a^{ij}(\xi_0) + \int\limits_{0}^{1}\nabla_x a^{ij}((1-t)\xi_0 + t \xi(y))dt \cdot (\xi_0 - \xi(y)) .
\end{split}
\end{equation}
Hence, (\ref{eq:expansion_1}) turns into
\begin{equation}
\label{eq:expansion_2}
\begin{split}
\tilde{a}^{ij}&=\tilde{a}^{ij}(y_0) +E_2^{y_0,ij}(y),
\end{split}
\end{equation}
where
\begin{align*}
E_2^{y_0,ij}(y)&:=\int\limits_{0}^{1}\nabla_x'' a^{ij}((1-t)\xi_0 + t \xi(y)) dt \cdot (y'' - y_0'')\\
& \quad + \sum\limits_{k=n}^{n+1}\int\limits_{0}^{1}\nabla_{x_k} a^{ij}((1-t)\xi_0 + t \xi(y)) dt \cdot (\p_k v(y) - \p_k v(y_0)).
\end{align*}
Recalling the asymptotics of $v$ from Definition~\ref{defi:spaces} for $v$, we infer that for all $y\in \mathcal{B}_{1/2}^+(y_0)$
\begin{align*}
|E_2^{y_0,ij}(y)|&\leq C \left|(y''-y''_0,-\p_nv(y) + \p_{n}v(y_0),-\p_{n+1}v(y))\right|\leq Cd_G(y,y_0)^{2}.
\end{align*}
Thus, we have shown that $d_G(y,y_0)^{-2}E_2^{y_0,ij}(y)\in L^\infty(\mathcal{B}_{1/2}^+(y_0))$.
Similarly, we also obtain $[d_G(y,y_0)^{-(2-\epsilon)}E_2^{y_0, ij}]_{\dot{C}^{0,\epsilon}(\mathcal{B}_{1/2}^+(y_0))}\leq C$. This, together with Definition~\ref{defi:spaces}, yields the second estimate in (\ref{eq:Holderweight0}).\\
\emph{Step 1 b: Expansion of the functions $G^{ij}$.}
For the (nonlinear) functions $G^{ij}(v)$ we have for all $i,j\in\{1,\dots,n+1\}$
\begin{align*}
G^{ij}(v) = G^{ij}(v_{y_0})+\p_{m_{k \ell}}G^{ij}(v_{y_0}) \p_{k \ell} (v-v_{y_0}) + E^{y_0,ij}_1(y),
\end{align*}
where
the error $E^{y_0,ij}_{1}(y)$ satisfies the bounds
\begin{equation}
\label{eq:Holderweight}
\begin{split}
\left\|d_G(\cdot,y_0)^{-(1+4\alpha)}E^{y_0,ij}_1\right\|_{L^\infty(\mathcal{B}_{1}^+(y_0))}
+\left[d_G(\cdot,y_0)^{-(1+4 \alpha -\epsilon)}E^{y_0,ij}_1\right]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_1^+(y_0))}\leq C.
\end{split}
\end{equation}
\emph{Proof of Step 1b:}
To show the claim, we first expand
\begin{equation}
\label{eq:expand_v}
\begin{split}
G^{ij}(v)&=G^{ij}(v_{y_0})+\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}(v -v_{y_0})\\
&\quad +\frac{1}{2}\p^2_{m_{k\ell}m_{\xi\eta}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0})\p_{\xi\eta}(v-v_{y_0})\\
&\quad +\frac{1}{6}\p^3_{m_{k\ell}m_{\xi\eta}m_{hs}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0})\p_{\xi\eta}(v-v_{y_0})\p_{hs}(v-v_{y_0})\\
&=G^{ij}(v_{y_0})+\p_{m_{k \ell}}G^{ij}(v_{y_0}) \p_{k \ell} (v-v_{y_0})+ E^{y_0,ij}_1(y).
\end{split}
\end{equation}
Here the error term is given by
\begin{align*}
E^{y_0,ij}_1(y)&=\frac{1}{2}\p^2_{m_{k\ell}m_{\xi\eta}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0})\p_{\xi\eta}(v-v_{y_0})\\
&+\frac{1}{6}\p^3_{m_{k\ell}m_{\xi\eta}m_{hs}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0})\p_{\xi\eta}(v-v_{y_0})\p_{hs}(v-v_{y_0}).
\end{align*}
Hence, it remains to prove the error estimate (\ref{eq:Holderweight}). To this end we estimate each term from the expression for $E_1^{y_0,ij}$ separately. We begin by observing that
\begin{equation}
\label{eq:det}
\begin{split}
e^{y_0,ij}(y):&=\p_{m_{k\ell}m_{\xi\eta}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0})\p_{\xi\eta}(v-v_{y_0})\\
&= \det \begin{pmatrix} \p_{ij}v_{y_0} & \p_{in}(v-v_{y_0}) & \p_{i,n+1}(v-v_{y_0}) \\
\p_{jn}v_{y_0} & \p_{nn}(v-v_{y_0}) & \p_{n,n+1}(v-v_{y_0}) \\
\p_{j,n+1}v_{y_0} & \p_{n,n+1}(v-v_{y_0}) & \p_{n+1,n+1}(v-v_{y_0})
\end{pmatrix}\\
& \quad + \det \begin{pmatrix} \p_{ij}(v-v_{y_0}) & \p_{in}v_{y_0} & \p_{i,n+1}(v-v_{y_0}) \\
\p_{jn}(v-v_{y_0}) & \p_{nn}v_{y_0} & \p_{n,n+1}(v-v_{y_0}) \\
\p_{j,n+1}(v-v_{y_0}) & \p_{n,n+1}v_{y_0} & \p_{n+1,n+1}(v-v_{y_0})
\end{pmatrix}\\
&\quad +\det \begin{pmatrix} \p_{ij}(v-v_{y_0}) & \p_{in}(v-v_{y_0}) & \p_{i,n+1}v_{y_0} \\
\p_{jn}(v-v_{y_0}) & \p_{nn}(v-v_{y_0}) & \p_{n,n+1}v_{y_0} \\
\p_{j,n+1}(v-v_{y_0}) & \p_{n,n+1}(v-v_{y_0}) & \p_{n+1,n+1}v_{y_0}
\end{pmatrix},
\end{split}
\end{equation}
and
\begin{align*}
\tilde{e}^{y_0,ij}(y)&:=\p^3_{m_{k\ell}m_{\xi\eta}m_{hs}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0})\p_{\xi\eta}(v-v_{y_0})\p_{hs}(v-v_{y_0})\\
&=\left\{
\begin{array}{ll}
3G^{ij}(v-v_{y_0})\text{ if } i,j\in\{1,\dots, n-1\},\\
0 \quad \text{ if } i \text{ or }j\in\{n,n+1\}.
\end{array}
\right.
\end{align*}
For simplicity we only present the estimate for $e^{y_0,ij}(y)$ in detail.
To estimate the difference $\p_{k\ell}(v-v_{y_0})$ for $\ell,k\in\{1,\dots, n+1\}$ in $\mathcal{B}_{1/2}^+(y_0)$, we use the definition of our function space $X_{\alpha,\epsilon}$ (in the form of Definition~\ref{defi:spaces} or in the form of the decomposition from Proposition~\ref{prop:decompI}) to obtain that for $y\in \mathcal{B}_{1/2}^+(y_0)$
\begin{align*}
|\p_{ij}(v-v_{y_0})(y)|&\leq C d_G(y,P)^{-1+2\alpha},\\
|\p_{in}(v-v_{y_0})(y)|&\leq C d_G(y,y_0)^{2\alpha},\\
|\p_{nn}(v-v_{y_0})(y)|&\leq C d_G(y,y_0)^{1+2\alpha}.
\end{align*}
Using this and plugging the explicit expression for $\p_{ij}v_{y_0}$ (c.f. \eqref{eq:v0}) into \eqref{eq:det} gives
$
|e^{y_0,ij}(y)| \leq Cd_G(y,y_0)^{1+4\alpha}.
$
Similarly, $|\tilde{e}^{y_0,ij}(y)|\leq Cd_G(y,y_0)^{1+6\alpha}$.
Hence, we obtain
\begin{align*}
|E^{y_0,ij}_1(y)|\leq C|e^{y_0,ij}(y)|+C|\tilde{e}^{y_0,ij}(y)|\leq d_G(y,y_0)^{1+4\alpha}.
\end{align*}
Moreover, it is not hard to deduce that $d_G(y,y_0)^{-(1+4\alpha-\epsilon)}E_1^{y_0,ij}(y)\in \dot{C}^{0,\epsilon}_\ast(\mathcal{B}_{1/2}^{+}(y_0))$. This concludes the proof of Step 1b.\\
\emph{Step 1c: Concatenation.}
We show that the leading order term $\tilde{a}^{ij}G^{ij}(v)$ has the expansion
\begin{align*}
\tilde{a}^{ij}G^{ij}(v)=\tilde{a}^{ij}(y_0)\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}v + P_{y_0}(y)+E^{y_0}_3(y),
\end{align*}
where
\begin{align}
\label{eq:P3}
P_{y_0}(y)=\tilde{a}^{ij}(y_0)\left(G^{ij}(v_{y_0})-\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}v_{y_0}\right)\in \mathcal{P}_1^{hom},
\end{align}
is a polynomial of the form $c(y_0)y_n$, and $E^{y_0}_3(y)$ satisfies for $\eta_0=\min\{3,1+4\alpha\}$
\begin{equation}
\label{eq:err_3}
\|d_G(\cdot,y_0)^{-\eta_0}E^{y_0}_3\|_{L^\infty(\mathcal{B}_{1/2}^+(y_0))}+\left[d_G(\cdot,y_0)^{-(\eta_0-\epsilon)}E^{y_0}_3\right]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_{1/2}^+(y_0))}\leq C.
\end{equation}
\emph{Proof of Step 1c:} Using the expansions for $\tilde{a}^{ij}$ and for $G^{ij}(v)$ from Steps 1a and 1b, we obtain
\begin{align*}
\tilde{a}^{ij}(y)G^{ij}(v)&=\tilde{a}^{ij}(y_0)G^{ij}(v)+E_{2}^{y_0,ij}(y)G^{ij}(v)\\
&=\tilde{a}^{ij}(y_0)\left(G^{ij}(v_{y_0})+\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0}) +E_1^{y_0,ij}(y)\right)\\
&\quad +E_2^{y_0,ij}(y)G^{ij}(v)\\
&=\tilde{a}^{ij}(y_0)\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}v +P_{y_0}(y) +E_3^{y_0}(y),
\end{align*}
where
\begin{align*}
E_3^{y_0}(y) := \tilde{a}^{ij}(y_0)E_1^{y_0,ij}(y)+E^{y_0,ij}_2(y)G^{ij}(v).
\end{align*}
Recalling the error bounds from Steps 1a, 1b and further observing
\begin{align*}
\left| G^{ij}(v) \right| \leq Cd_G(y,y_0),
\end{align*}
entails (\ref{eq:err_3}).\\
\emph{Step 2: Expansion of the lower order contributions.}
For the lower order contribution the asymptotics of $v$ immediately yield
\begin{align*}
\left|J(v)(y)\left( \sum\limits_{j=1}^{n-1} \tilde{b}^j(y) \p_j v(y) + \tilde{b}^n(y)y_n + \tilde{b}^{n+1}(y)y_{n+1}\right)\right|\leq Cc_{\ast} d_G(y,P)^3.
\end{align*}
Here we used that $\|\nabla a^{ij}\|_{L^{\infty}}\leq C c_{\ast}$. Hence, this error is small compared with the error term $E^{y_0}_3(y)$ from the leading order expansion (c.f. \eqref{eq:err_3}).
\end{proof}
The previous proposition allows us to apply an iterative bootstrap argument to obtain higher regularity for $v$.
\begin{prop}
\label{prop:error_gain2}
Assume that $v\in X_{\alpha,\epsilon}(\mathcal{B}_{1}^+)$ for some $\alpha\in (0,1]$, $\epsilon \in (0,\gamma]$, and that it satisfies $F(v,y)=0$ with $a^{ij}(x)\in C^{1,\gamma}$ for some $\gamma\in (0,1]$. Then
$v\in X_{\delta,\epsilon}(\mathcal{B}_{1/2}^+)$ for any $\delta\in (0,1)$.
\end{prop}
\begin{proof}
If $1+4\alpha<3$, i.e. $0\leq \alpha<1/2$, Proposition~\ref{prop:error_gain} and Remark~\ref{rmk:error_gain3} yield that for each fixed $y_0\in P\cap \mathcal{B}_{1/2}$ the Legendre function $v$ solves
\begin{align}
\label{eq:eq}
L_{y_0}v=L_{y_0}v_{y_0}+\tilde{f} \text{ in } \mathcal{B}_{1/2}^+(y_0),
\end{align}
where $L_{y_0}$ is the ``constant coefficient" Baouendi-Grushin type operator from Remark~\ref{rmk:error_gain3}, $L_{y_0}v_{y_0}=c(y_0)y_n$ and the function $\tilde{f}(y)$ is such that
$$\|d_G(\cdot,y_0)^{-(1+4\alpha)}\tilde{f}\|_{L^\infty(\mathcal{B}_{1/2}^+(y_0))}+[d_G(\cdot,y_0)^{-(1+4\alpha-\epsilon)}\tilde{f}]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_{1/2}^+(y_0))}\leq C,$$
where $C$ depends on $\|v\|_{X_{\alpha,\epsilon}}$ and $[D a^{ij}]_{\dot{C}^{0,\gamma}}$ and is in particularly independent of $y_0$. We apply the compactness argument from the Appendix (c.f. the proof of Proposition~\ref{prop:Hoelder0}) at each point $y_0\in P\cap \mathcal{B}_{1/2}$. This is possible as $L_{y_0}$ is a self-adjoint, constant coefficient subelliptic operator of Baoundi-Grushin type which is hypoelliptic after suitable reflections (c.f. Remark \ref{rmk:error_gain3} and \cite{JSC87}). We note that as in the case of the Grushin operator there are no fourth order homogeneous polynomials with symmetry (even about $y_{n+1}$ and odd about $y_n$) which are solutions to the equation $L_{y_0}v=0$. Combining the above approximation result along $P\cap \mathcal{B}_{1/2}$ with the $C^{2,\epsilon}_\ast$, $\epsilon\leq \gamma$, estimate in the corresponding non-tangential region (with respect to $P$) leads to $v\in X_{2\alpha,\epsilon}(\mathcal{B}_{1/4}^+)$.
We repeat the above procedure until after finitely many, say, $k$, steps, $1+4k\alpha>3$. This results in $v\in X_{\delta,\epsilon}(\mathcal{B}_{1/2^{k}}^+)$ for every $\delta\in (0,1)$ (where we used the nonexistence of homogeneous fourth order approximating polynomials). Repeating this procedure in $\mathcal{B}^+_{1/2}(\bar y)$ for $\bar y\in P\cap \mathcal{B}_{1/2}^+$ and by a covering argument, we obtain that $v\in X_{\delta,\epsilon}(\mathcal{B}_{1/2}^+)$ for every $\delta\in (0,1)$.
\end{proof}
\section[Free Boundary Regularity]{Free Boundary Regularity for $C^{k,\gamma}$ Metrics, $k\geq 1$}
\label{sec:fb_reg}
In this section we apply the implicit function theorem to show that the regular free boundary is locally in $C^{k+1,\gamma}$, if $a^{ij}\in C^{k,\gamma}$ with $k\geq 1$ and $\gamma\in (0,1)$. Moreover, we also argue that the regular free boundary is locally real analytic, if $a^{ij}$ is real analytic. \\
In order to invoke the implicit function theorem, we discuss the mapping properties of the nonlinear function $F$ in the next two sections. More precisely, we prove that
\begin{itemize}
\item the nonlinearity $F$ maps $X_{\delta,\epsilon}(\mathcal{B}_1^+)$ to $Y_{\delta,\epsilon}(\mathcal{B}_1^+)$ for any $\delta\in (0,1)$ and $\epsilon$ sufficiently small (c.f. Section \ref{sec:nonlinmap}),
\item and that its linearization in a neighborhood of $v_0$ is a perturbation of the Baouendi-Grushin Laplacian and is hence invertible (c.f. Section \ref{sec:grushin}).
\end{itemize}
Then in Section \ref{subsec:IFT0} we introduce an one-parameter family of diffeomorphisms which will form the basis of our application of the implicit function theorem in Section \ref{sec:IFTAppl}. In Section \ref{sec:IFTAppl} we apply the implicit function theorem argument to show the regularity of the free boundary in $C^{k,\gamma}$, $k\geq 1$ metrics and analytic metrics, which yields the desired proof of Theorem \ref{thm:higher_reg}.
\subsection{Mapping properties of $F$}
\label{sec:nonlinmap}
As a consequence of the representation of $F$ which was derived in Proposition \ref{prop:error_gain} we obtain the following mapping properties for our nonlinear function $F$.
\begin{prop}
\label{prop:nonlin_map}
Let $a^{ij}:B_1^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ be a $C^{k,\gamma}$ tensor field with $k\geq 1$ and $\gamma \in (0,1]$. Assume that the nonlinear function $F$ is as in (\ref{eq:nonlin_2}). Then for any $\delta\in (0,1]$ and $\epsilon\in (0,\gamma)$, we have that
\begin{align*}
F:X_{\delta,\epsilon}(\mathcal{B}_1^+) \rightarrow Y_{\delta,\epsilon}(\mathcal{B}_1^+).
\end{align*}
\end{prop}
These properties will be used in Propositions \ref{prop:reg_a} and \ref{prop:invertible} to establish the mapping properties of the nonlinear function to which we apply the implicit function theorem in Section \ref{sec:IFT1}
\begin{proof}
The mapping properties of $F$ are an immediate consequence of the representation for $F$ that was obtained in Proposition \ref{prop:error_gain}. Indeed, given any $u\in X_{\delta,\epsilon}$, by Proposition \ref{prop:error_gain}, we have
\begin{align*}
F(u, y) = \sum\limits_{i,j=1}^{n+1} \tilde{a}^{ij}(y_0) \p_{m_{k \ell}} G^{ij}(u_{y_0}) \p_{k \ell } u + P_{y_0}(y) + E_{y_0}(y).
\end{align*}
Due to Proposition~\ref{prop:decompI}, Proposition~\ref{prop:error_gain} and Remark~\ref{rmk:error_gain3} we infer that $P_{y_0}(y) + E_{y_0}(y)\in Y_{\delta,\epsilon}(\mathcal{B}_1^+)$. For the remaining linear term we note that similarly as in the proof of Proposition \ref{prop:error_gain}
\begin{align*}
\tilde{a}^{ij}(y_0) \p_{m_{k \ell}} G^{ij}(u_{y_0}) \p_{k \ell } u & = \tilde{a}^{ij}(y_0) \p_{m_{k \ell}} G^{ij}(u_{y_0}) \p_{k \ell } u_{y_0} \\
& \quad + \tilde{a}^{ij}(y_0) \p_{m_{k \ell}} G^{ij}(u_{y_0}) \p_{k \ell } (u-u_{y_0}) \\
& = c(y'')y_n + r^{1+2\alpha-\epsilon}f(y),
\end{align*}
with $c(y'')\in C^{0,\delta}(\R^{n-1}\cap \mathcal{B}_1^+)$ and $f(y)\in C^{0,\epsilon}_{\ast}(\mathcal{B}_1^+)$.
This implies the result.
\end{proof}
\subsection{Linearization and the Baouendi-Grushin operator}
\label{sec:grushin}
In this section we compute the linearization of $F$ and show that it can be interpreted as a perturbation of the Baouendi-Grushin Laplacian. We treat the cases $a^{ij}\in C^{k,\gamma}$ with $k=1$ and $k\geq 2$ simultaneously as only small modifications are needed in the argument.
If $a^{ij}\in C^{k,\gamma}$ with $k\geq 2$, by using the notation in \eqref{eq:nonlin_2},
\begin{align*}
F(D^2v, Dv,y):=&\sum_{i,j=1}^{n+1}a^{ij}(y'',-\p_nv,-\p_{n+1}v)G^{ij}(v)-J(v)\sum_{j=1}^{n-1}b^j(y'',-\p_nv,-\p_{n+1}v)\p_j v\\
&-J(v)b^n(y'',-\p_nv,-\p_{n+1}v)y_n-J(v)b^{n+1}(y'',-\p_nv,-\p_{n+1}v)y_{n+1},
\end{align*}
In the case of $a^{ij}\in C^{1,\gamma}$ we view $F$ as
\begin{align*}
F(D^2v, Dv,y)&=\sum_{i,j}^{n+1}a^{ij}(y'',-\p_nv,-\p_{n+1}v)G^{ij}(v)+ f(y),\\
&\text{where } f(y)=-J(v(y))\left(\sum_{j=1}^{n-1}\tilde{b}^j(y)\p_j v(y)+\tilde{b}^n(y)y_n+\tilde{b}^{n+1}(y)y_{n+1}\right).
\end{align*}
Here we view $F$ as a mapping $F:\R^{(n+1)\times(n+1)}_{sym}\times \R^{n+1}\times U\rightarrow \R$ and introduce the abbreviations
\begin{align*}
F_{k\ell}(M,P,y)&:=\frac{\partial F(M,P,y)}{\partial m_{k\ell}}, \quad M=(m_{k\ell})\in \R^{(n+1)\times(n+1)}_{sym},\ P\in \R^{n+1},\ y\in \R^{n+1},\\
F_k(M,P,y)&:=\frac{\partial F(M,P,y)}{\partial p_{k}}.
\end{align*}
For notational convenience, we use the conventions $$F_{k\ell}(v,y):=F_{k\ell}(D^2v, Dv, y), \quad F_k(v,y):=F_{k}(D^2v, Dv,y), \quad F(v,y):=F(D^2v, Dv,y).$$
The linearization of $F$ at $v$ is
\begin{align}
\label{eq:op_lin}
L_v:=F_{k\ell}(v,y) \p_{k \ell} + F_k (v,y) \p_k.
\end{align}
In the case $a^{ij}\in C^{k,\gamma}$ with $k\geq 2$, we have
\begin{align*}
F_{k\ell}(v,y) &= \sum\limits_{i,j=1}^{n+1} \tilde{a}^{ij} \p_{m_{k \ell}} G^{ij}(v) - \p_{m_{k \ell}} J(v) \left(\sum_{j=1}^{n-1}\tilde{b}^j\partial_jv+\tilde{b}^ny_n+\tilde{b}^{n+1}y_{n+1}\right),\\
F_k(v,y) &= -J(v)\sum_{j=1}^{n-1}\tilde{b}^j, \mbox{ for } k\in\{1,\dots,n-1\},\\
F_{k}(v,y) &= \sum\limits_{i,j=1}^{n+1} b^{ij}_k G^{ij}(v)- J(v)\left(\sum_{j=1}^{n-1}b^j_k\partial_jv+b^n_k y_n+b^{n+1}_ky_{n+1}\right), \mbox{ for } k\in\{n,n+1\}.
\end{align*}
Here
\begin{align*}
\tilde{a}^{ij}&:= a^{ij}|_{(y'',-\p_n v(y), - \p_{n+1} v(y))}, \quad \tilde{b}^j=\sum_{i=1}^{n+1}\p_{x_i}a^{ij}|_{(y'',-\p_nv(y),-\p_{n+1}v(y))}\\
b^{ij}_k &:= \p_{x_k} a^{ij}|_{(y'',-\p_n v(y), - \p_{n+1} v(y))},\\
b^{j}_k &:= \sum_{i=1}^{n+1}\p_{x_k x_i} a^{ij}|_{(y'',-\p_n v(y), - \p_{n+1} v(y))}.
\end{align*}
In particular, we note that the linearization of $F$ already involves second order derivatives of our metric $a^{ij}$. \\
In the case $a^{ij}\in C^{1,\gamma}$, we have
\begin{align*}
F_{k\ell}(v,y)&=\sum_{i,j=1}^{n+1}\tilde{a}^{ij}\p_{m_{k\ell}}G^{ij}(v), \quad k,\ell\in\{1,\dots,n+1\},\\
F_k(v,y)&=0,\ k\in\{1,\dots, n-1\}; \quad F_k(v,y)=\sum\limits_{i,j=1}^{n+1} b^{ij}_k G^{ij}(v), \quad k\in\{n,n+1\}.
\end{align*}
Let
$v_0(y):=-\frac{1}{6}\left(y_n^3-3y_ny_{n+1}^2\right)
$ be the (scaled) blow-up of $v$ at $0$.
A direct computation shows that $F_{k\ell}(v_0,0)\p_{k\ell}=\Delta_G$, where $\Delta_G$ is the standard Baouendi-Grushin Laplacian in Section~\ref{sec:intrinsic}. Thus, we write
\begin{align*}
L_v&=\Delta_G+\left(F_{k \ell}(v,y)-F_{k \ell}(v_0,0)\right)\p_{k \ell} +F_k(v,y)\p_k\\
&=:\Delta_G+\mathcal{P}_v.
\end{align*}
With this at hand, we can prove the following mapping properties for $L_v$:
\begin{prop}
\label{prop:linear}
Let $L_v, \mathcal{P}_v$ be as above. Assume furthermore in the case of $a^{ij}\in C^{k,\gamma}$ with $k\geq 2$ that $[D^2_xa^{ij}]_{\dot{C}^{0,\gamma}}\leq c_\ast$. Given $\delta\in (0,1]$ and $\epsilon\in (0,\min\{\delta,\gamma\})$, let $X_{\delta,\epsilon}(\mathcal{B}_1^+)$ and $Y_{\delta,\epsilon}(\mathcal{B}_1^+)$ be the spaces from Definition~\ref{defi:spaces_loc}. Then
\begin{align*}
L_v : X_{\delta,\epsilon}(\mathcal{B}_1^+) \rightarrow Y_{\delta,\epsilon}(\mathcal{B}_1^+).
\end{align*}
Moreover, if $\|v-v_0\|_{X_{\delta,\epsilon}(\mathcal{B}_1^+}\leq \delta_0$, then
$$\|\mathcal{P}_v w\|_{Y_{\delta,\epsilon}(\mathcal{B}_1^+)}\lesssim \max\{\delta_0,c_\ast\}\|w\|_{X_{\delta,\epsilon}(\mathcal{B}_1^+)}, \quad \text{for all } w\in X_{\delta,\epsilon}(\mathcal{B}_1^+).$$
\end{prop}
\begin{proof}
We first show the claims of the proposition in the case of $a^{ij}\in C^{k,\gamma}$ with $k\geq 2$. We begin by arguing that $L_v w \in Y_{\delta,\epsilon}(\mathcal{B}_1^+)$ if $w\in X_{\delta, \epsilon}(\mathcal{B}_1^+)$.
\begin{itemize}
\item[(a)] To show this, we observe that
\begin{align*}
&\sum_{k,\ell}\p_{m_{k\ell}}G^{ij}(v)\p_{k\ell}w
=\det\begin{pmatrix}
\p_{ij}w & \p_{in}w & \p_{i,n+1}w\\
\p_{in}v & \p_{nn}v & \p_{n,n+1}v\\
\p_{i,n+1}v & \p_{n+1,n}v & \p_{n+1,n+1}v
\end{pmatrix} \\
&+ \det\begin{pmatrix}
\p_{ij}v & \p_{in}v & \p_{i,n+1}v\\
\p_{in}w & \p_{nn}w & \p_{n,n+1}w\\
\p_{i,n+1}v & \p_{n+1,n}v & \p_{n+1,n+1}v
\end{pmatrix}
+\det\begin{pmatrix}
\p_{ij}v & \p_{in}v & \p_{i,n+1}v\\
\p_{in}v & \p_{nn}v & \p_{n,n+1}v\\
\p_{i,n+1}w & \p_{n+1,n}w & \p_{n+1,n+1}w
\end{pmatrix},
\end{align*}
if $i,j\in\{1,\dots, n-1\}$ and for the remaining indices $(i,j)$ the expression $\sum_{k,\ell}G^{ij}(v)\p_{k\ell}w$ is similar. Thus, the mapping property of $\sum_{i,j}\sum_{k,\ell}\tilde{a}^{ij}\p_{m_{k\ell}}G^{ij}(v)\p_{k\ell}w$ follows along the lines of the proof of Proposition~\ref{prop:nonlin_map} and Proposition~\ref{prop:error_gain}.
\item[(b)] We discuss the term $\sum\limits_{i,j=1}^{n+1}\sum\limits_{\ell=n}^{n+1}b^{ij}_\ell\p_\ell w G^{ij}(v)$. As $b^{ij}_\ell$ is $C^{1,\gamma}$, it satisfies the same decomposition as $\tilde{a}^{ij}$ in \eqref{eq:Holderweight0}. Using the characterization for $\p_nw$, $\p_{n+1}w$ from Proposition~\ref{prop:decompI} (ii), we have that around $y_0\in P\cap \mathcal{B}_{1/2}$
\begin{align*}
\sum_{\ell=n,n+1}b^{ij}_\ell\p_\ell w = \tilde{b}^{ij}_n(y_0)\p_nw(y_0)+ E_{y_0}^{ij},
\end{align*}
where $E_{y_0}^{ij}$ satisfies the same error bounds as (\ref{eq:Holderweight0}). Moreover, the functions $\tilde{b}^{ij}_n$ inherit the off-diagonal condition of $a^{ij}$, i.e. $\tilde{b}^{i,n+1}_n=0$ on $\{y_{n+1}=0\}$ for $i\in \{1,\dots, n\}$. Thus, using exactly the same estimate as for $\tilde{a}^{ij}G^{ij}(v)$ in Step 1 of Proposition~\ref{prop:error_gain}, we obtain that $\sum\limits_{i,j=1}^{n+1}\sum\limits_{\ell=n}^{n+1}b_\ell^{ij}\p_\ell w G^{ij}(v)\in Y_{\delta, \epsilon}(\mathcal{B}_1^+)$.
\item[(c)] The term
\begin{align*}
-J(v)\left( \sum\limits_{j=1}^{n-1}b^j_k \p_j v + b^n_k y_n + b^{n+1}_k y_{n+1} \right)\p_k w,\quad k\in \{n,n+1\}
\end{align*}
is a lower order term contained in $Y_{\delta,\epsilon}(\mathcal{B}_1^+)$ as it is bounded by $r(y)^3=\dist(y,P)^3$ and satisfies the right Hölder bounds.
\item[(d)] Similarly, the contribution
\begin{align*}
-\p_{m_{k\ell}}J(v)\left( \sum\limits_{j=1}^{n-1}\tilde{b}^j \p_j v + \tilde{b}^n y_n + \tilde{b}^{n+1} y_{n+1} \right)\p_{k\ell} w,\quad k,\ell\in \{n,n+1\}
\end{align*}
is a lower order term contained in $Y_{\delta,\epsilon}(\mathcal{B}_1^+)$ as it is also bounded by $\dist(y,P)^3$ and satisfies the right Hölder bounds.
\end{itemize}
We continue by proving the bounds for $\mathcal{P}_v$. To this end, we again consider the individual terms in the linearization separately:
\begin{itemize}
\item Estimate of $\left(F_{k\ell}(v,y)-F_{k\ell}(v_0,0)\right)\p_{k\ell}w$. We will only present the details of the estimate for $k,\ell\in \{1,\dots,n-1\}$. The estimates for the remaining terms are similar. \\
For $k,\ell\in \{1,\dots, n-1\}$, from \eqref{eq:op_lin} we have
\begin{align*}
F_{k\ell}(v,y)=-\tilde{a}^{k\ell}(y)J(v).
\end{align*}
By assumption $v\in X_{\delta,\epsilon}(\mathcal{B}_1^+)$ is in a $\delta_0$ neighborhood of $v_0$ and by Proposition~\ref{prop:decompI} we have the decomposition ($r=r(y)=(y_n^2+y_{n+1}^2)^{1/2}$)
\begin{align*}
\p_{ij}v(y)&=r^{-1+2\delta-\epsilon}C_{ij}(y),\\
\p_{in}v(y)&=B_i(y'')+r^{2\delta-\epsilon}C_{in}(y),\\
\p_{i,n+1}v(y)&=r^{2\delta-\epsilon}C_{i,n+1}(y),\\
\p_{nn}v(y)&=A_0(y'')y_n+r^{1+2\delta-\epsilon}C_{n,n}(y),\\
\p_{n,n+1}v(y)&=A_1(y'')y_{n+1}+r^{1+2\delta-\epsilon}C_{n,n+1}(y),\\
\p_{n+1,n+1}v(y)&=A_1(y'')y_{n}+r^{1+2\delta-\epsilon}C_{n+1,n+1}(y),
\end{align*}
where $B_i, A_0, A_1\in C^{0,\delta}(P\cap \mathcal{B}_1)$ with
\begin{equation}\label{eq:smallness}
[B_i]_{\dot{C}^{0,\delta}}+\|A_0-1\|_{L^\infty}+\|A_1-(-1)\|_{L^\infty}\lesssim \delta_0,
\end{equation}
and $C_{ij}\in C^{0,\epsilon}_\ast(\mathcal{B}_1^+)$ with $[C_{ij}]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_1^+)}\lesssim \delta_0$.
Thus, for $k,\ell\in \{1,\dots, n-1\}$
\begin{align*}
&\quad F_{k\ell}(v,y)-F_{k\ell}(v_{0},0)\\
&=-\left(\tilde{a}^{k\ell}(y)-\tilde{a}^{k\ell}(0)\right)J(v_{0})-\tilde{a}^{k\ell}(y)\left(J(v)-J(v_{0})\right)\\
&\lesssim c_\ast(y_n^2+y_{n+1}^2)+ \delta_0 (y_n^2+y_{n+1}^2).
\end{align*}
Consequently,
\begin{align*}
\left|\left(F_{k\ell}(v,y)-F_{k\ell}(v_0,0)\right)\p_{k\ell}w(y)\right|\lesssim \max\{c_\ast,\delta_0\}r^{1+2\delta}|r^{-(1-2\delta)}\p_{k\ell}w(y)|.
\end{align*}
Moreover, it is not hard to check that
\begin{align*}
\left[r(\cdot)^{-(1+2\delta-\epsilon)}(F_{k\ell}(v,\cdot)-F_{k\ell}(v_0,0))\p_{k\ell}w\right]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_{1}^+)}\lesssim \max\{c_\ast,\delta_0\}\|w\|_{X_{\delta,\epsilon}(\mathcal{B}_1^+)}.
\end{align*}
Hence, for $k,\ell\in \{1,\dots, n-1\}$
$$\left\|(F_{k\ell}(v,y)-F_{k\ell}(v_0,0))\p_{k\ell}w\right\|_{Y_{\delta,\epsilon}(\mathcal{B}_1^+)}\lesssim \max\{c_\ast,\delta_0\}\|w\|_{X_{\delta,\epsilon}(\mathcal{B}_1^+)}.$$
For the remaining second order terms in the linearization we argue similarly.
\item Estimate for $F_k(v,y)\p_k$. To estimate the lower order terms $F_k(v,y)$, we need to assume that $a^{ij}\in C^{2,\gamma}$ for some $\gamma>0$. Furthermore, we assume $[D^2_xa^{ij}]_{\dot{C}^{0,\gamma}}\leq c_\ast$ (this assumption on the second derivatives becomes necessary as the term $F_k(v,y)$ with $k\in \{n,n+1\}$ involves $D^2a$, c.f. \eqref{eq:op_lin}). Note that combining this with the assumption (A6) and using an interpolation estimate we have $\|D^2_xa^{ij}\|_{L^\infty}\leq Cc_\ast$.\\
We begin with the terms with $k\in\{1,\dots, n-1\}$, i.e. with
\begin{align*}
F_k(v,y)&=-J(v)\sum_{j=1}^{n-1}\tilde{b}^j, \quad \tilde{b}^j=\sum_{i=1}^{n+1}\p_{x_i}a^{ij}|_{(y'',-\p_nv,-\p_{n+1}v)}.
\end{align*}
As by our assumption (A6) $\|\tilde{b}^i\|_{L^\infty}\lesssim c_\ast$, the asymptotics of $J(v)$ immediately yield that $\|F_k(v,y)\p_k w\|_{Y_{\delta,\epsilon}(\mathcal{B}_1^+)}\lesssim c_\ast\|w\|_{X_{\delta,\epsilon}(\mathcal{B}_1^+)}$ as long as $\epsilon\leq \gamma$.\\
For the contributions with $k\in\{n,n+1\}$, $F_k(v,y)$ is of the same structural form as the original nonlinear function $F(v,y)$, however with coefficients which contain an additional derivative, i.e. $a^{ij}$ is replaced by $b^{ij}_k$ and $\tilde{b}^j$ is replaced by $b^j_k$ (c.f. \eqref{eq:op_lin}).
Thus, by the argument in Section \ref{sec:nonlinmap} on the mapping properties of the nonlinear function $F(v,y)$, we infer that $F_k(v,y)\p_k w \in Y_{\alpha,\epsilon}(\mathcal{B}_1^+)$, $k\in\{n,n+1\}$ (for $\epsilon \leq \gamma$) for any $w\in X_{\alpha,\epsilon}(\mathcal{B}_1^+)$. Moreover, it satisfies $\|F_{k}(v,y)\p_k w\|_{Y_{\delta,\epsilon}(\mathcal{B}_1^+)}\lesssim C [D^2 a^{ij}]_{\dot{C}^{0,\gamma}}\|w\|_{X_{\delta,\epsilon}(\mathcal{B}_1^+)}$.
\end{itemize}
Combining the previous observations concludes the proof of Proposition \ref{prop:linear} in the case that $k\geq 2$.\\
For the case $k=1$, we notice that if $a^{ij}\in C^{1,\gamma}$ the linearization is simply given by $L_{v}=F_{k\ell}(v,y)\p_{k\ell}+\sum\limits_{i,j=1}^{n+1} b^{ij}_k G^{ij}(v)\p_k$. Thus, a similar proof as for the case $C^{k,\gamma}$ with $k\geq 2$ applies.
\end{proof}
\subsection{Hölder Regularity and Analyticity}
\label{sec:IFT1}
In this section we apply the implicit function theorem to show that if $a^{ij}\in C^{k,\gamma}$ with $k\geq 1$ and $\gamma\in (0,1)$, then the regular free boundary is locally in $C^{k+1,\gamma}$. \\
To this end, we first define a one-parameter family of diffeomorphisms which we compose with our Legendre function to create an ``artificially parameter-dependent problem''. Due to the regularity properties of $F$, this is then exploited to deduce the existence of a solution to the parameter-dependent problem, which enjoys good regularity properties in the artificially introduced parameter (c.f. Proposition \ref{prop:reg_a}). Finally, this regularity is transfered from the parameter variable to the original variables yielding the desired regularity properties of our Legendre function (c.f. Theorems \ref{prop:hoelder_reg_a}, \ref{prop:analytic}). This then proves the claims of Theorem \ref{thm:higher_reg}.\\
In the sequel, we will always assume that $v$ is a Legendre function (c.f. (\ref{eq:legendre})) which is associated with a solution of the variable coefficient thin obstacle problem (\ref{eq:varcoeff}), which satisfies the normalizations (A4) and (A5). The coefficient metric $a^{ij} \in C^{k,\gamma}$, $k\geq 1$, is assumed to obey the conditions (A1)-(A3) as well as (A6). We also suppose that $[D^2_xa^{ij}]_{\dot{C}^{0,\gamma}}\leq c_\ast$ if $k\geq 2$. By rescaling we assume that $v$ is well defined in $\mathcal{B}_2^+$. We have shown in Proposition~\ref{prop:error_gain2} that $v\in X_{\delta,\epsilon}(\mathcal{B}_1^+)$ for any $\delta\in (0,1)$ and $\epsilon\in (0,\gamma]$. Furthermore, we recall that $\|v-v_0\|_{X_{\delta,\epsilon}(\mathcal{B}_1^+)}\leq C\max\{\epsilon_0,c_\ast\}$ (c.f. Remark~\ref{rmk:close}), where $v_0(y)=-\frac{1}{6}(y_n^3-3y_ny_{n+1}^2)$ is the model solution, which is the (rescaled) blow-up of $v$ at $0$.
\subsubsection{An infinitesimal translation.}
\label{subsec:IFT0}
For $y\in \R^{n+1}$ and $a\in \R^{n-1}$ fixed, we consider the following ODE
\begin{equation}
\label{eq:ODE}
\begin{split}
\phi'(t)&=a ((3 /4 )^2-|\phi(t)|^2)^5_+ \eta(y_n,y_{n+1}),\\
\phi(0)&=y''.
\end{split}
\end{equation}
Here $\eta$ is an in the $y_n, y_{n+1}$ variables radially symmetric smooth cut-off function supported in $\{(y_n,y_{n+1})| \ y_n^2+y_{n+1}^2<1/2\}$, which is equal to one in $\{y_n^2+y_{n+1}^2\leq 1/4\}$. We denote the unique solution to the above ODE by $\phi_{a,y}(t)$ and let $$\Phi_a(y):= (\phi_{a,y}(1),y_n,y_{n+1}).$$
Due to the $C^5$ regularity of the right hand side of (\ref{eq:ODE}), we obtain that $\phi_{a,y}(1)$ is $C^5$ in $y$.
Moreover, an application of a fixed point argument yields that $\phi_{a,y}(1)$ is analytic in the parameter $a$. We summarize these properties as:
\begin{lem}
For each $a\in \R^{n-1}$, $\Phi_a:\R^{n+1}\rightarrow \R^{n+1}$ is a $C^5$ diffeomorphism. The mapping $\R^{n-1}\ni a\mapsto \Phi_a\in C^5(\R^{n+1})$ is analytic.
\end{lem}
Moreover, we note that $\Phi_a$ enjoys further useful properties:
\begin{lem}
\label{prop:psi}
\begin{itemize}
\item[(i)] For each $a\in \R^{n-1}$, $\Phi_a(\{y_n=0\})\subset\{y_n=0\}$ and $\Phi_a(\{y_{n+1}=0\})\subset\{y_{n+1}=0\}$. Moreover, $\Phi_a(y)=y$ if $y\notin \{y\in \R^{n+1}| \ |y''|<\frac{3 }{4}, \ y_n^2+y_{n+1}^2<\frac{1}{2}\}$.
\item[(ii)] For each $a\in \R^{n-1}$ and $y\in \{y\in\R^{n+1}| \ y_n^2+y_{n+1}^2<\frac{1}{4}\}$, we have $\Phi_a(y)=(\phi_{a,y''}(1),y_n,y_{n+1})$, i.e. $\Phi_a$ only acts on the tangential variables.
\item[(iii)] For each $a\in \R^{n-1}$, $\p_{n+1}\Phi_a(y'',y_n,0)=(0,\dots, 0,1)$.
\item[(iv)] At $a=0$, $\Phi_0(y)=y$ for all $y\in \R^{n+1}$.
\end{itemize}
\end{lem}
\begin{proof}
This follows directly from the definition of $\Phi_a$ and a short calculation.
\end{proof}
Let $v$ be a Legendre function as described at the beginning of Section~\ref{sec:IFT1}. We use the family of diffeomorphisms $\Phi_a$ to define a one-parameter family of functions:
$$v_a(y):=v(\Phi_a(y)).$$
We first observe that the space $X_{\delta,\epsilon}$ (c.f. Definition \ref{defi:spaces}) is stable under the diffeomorphism $\Phi_a$:
\begin{lem} \label{prop:psi2}
If $v\in X_{\delta,\epsilon}(\mathcal{B}_{1}^+)$, then $v_a=v\circ \Phi_a\in X_{\delta,\epsilon}(\mathcal{B}_{1}^+)$ as well.
\end{lem}
\begin{proof}
We first check that $v_a$ satisfies the Dirichlet-Neumann boundary condition. Indeed, by (i) in Lemma~\ref{prop:psi}, if $v=0$ on $\{y_n=0\}$, then $v_a=0$ on $\{y_n=0\}$ as well. To verify the Neumann boundary condition, we compute
\begin{align*}
\p_{n+1}v_a(y)=\sum_{k=1}^{n+1}\p_{k}v(z)\big|_{z=\Phi_a(y)}\p_{n+1} \Phi_a^k(y).
\end{align*}
Thus by (i) and (iii) of Lemma~\ref{prop:psi}, $\p_{n+1}v_a=0$ on $\{y_{n+1}=0\}$.
Next by property (ii) of Lemma~\ref{prop:psi}, for $y\in \{y\in \R^{n+1}| y_n^2+y_{n+1}^2<\frac{1}{4}\}$ and $i,j\in\{1,\dots,n-1\}$,
\begin{align*}
\p_iv_a(y)&=\sum_{k=1}^{n-1}\p_k v(z)\big|_{z=\Phi_a(y)}\p_i\Phi^k_a(y''), \\
\p_{ij}v_a(y)&=\sum_{k,\ell=1}^{n-1}\p_{k\ell}v(z)\big|_{z=\Phi_a(y)}\p_i\Phi^k_a(y'') \p_j\Phi^\ell_a(y'') + \sum_{k=1}^{n-1}\p_kv(z)\big|_{z=\Phi_a(y)}\p_{ij}\Phi_a^k(y''),\\
\p_{in}v_a(y)&=\sum_{k=1}^{n-1}\p_{kn}v(z)\big|_{z=\Phi_a(y)}\p_i\Phi_a^{k}(y'').
\end{align*}
Thus, combining these calculations with the fact that $\Phi_a$ fixes the $(y_n,y_{n+1})$ variables ((ii) of Lemma~\ref{prop:psi}), it is not hard to check that $v_a$ satisfies the decomposition in Proposition~\ref{prop:decompI} in the region $\{y\in \R^{n+1}| y_n^2+y_{n+1}^2<\frac{1}{4}\}$. The regularity of $\Phi_a$ and of $v$, entails that $v_a\in C^{2,\epsilon}_\ast$ outside of the region $\{y\in \R^{n+1}| y_n^2+y_{n+1}^2<\frac{1}{4}\}$. Thus, $v_a\in X_{\delta,\epsilon}(\mathcal{B}_1^+)$.
\end{proof}
Since $v$ satisfies $F(v,y)=0$, the function $v_a(y)=v(\Phi_a(y))$ solves a new equation $F_a(u,y)=0$. Here
\begin{align}\label{eq:Fa}
F_a(u,y)=F(u(\Phi^{-1}_a(z)),z)\big|_{z=\Phi_a(y)}.
\end{align}
For this equation we note the following properties:
\begin{prop}\label{prop:reg_a}
Let $a^{ij}\in C^{k,\gamma}$ with $k\geq 1$, $\gamma\in (0,1]$. Then for each $a\in \R^{n-1}$, $F_a$ maps $X_{\delta,\epsilon}(\mathcal{B}_{1}^+)$ into $Y_{\delta,\epsilon}(\mathcal{B}_{1}^+)$. Moreover,
\begin{itemize}
\item[(i)] for each $a\in \R^{n-1}$, the mapping
\begin{align*}
F_a(\cdot, y): X_{\delta,\epsilon}(\mathcal{B}_1^+)\rightarrow Y_{\delta,\epsilon}(\mathcal{B}_1^+), \quad u\mapsto F_a(u,y),
\end{align*}
is $C^{k-1,\gamma-\epsilon}$ in $u$.
\item[(ii)] For each $u\in X_{\alpha,\epsilon}(\mathcal{B}_1^+)$, the mapping
\begin{align*}
F_{\cdot}(u,y):\R^{n-1} \rightarrow Y_{\delta,\epsilon}(\mathcal{B}_1^+), \ a \mapsto F_a(u,y),
\end{align*}
is $C^{k-1,\gamma-\epsilon}$ in $a$. If $a^{ij}$ is real analytic in $B_1^+$, then $F_a$ is real analytic in $a$.
\end{itemize}
\end{prop}
\begin{proof}
We first check the mapping property of $F_a$. Let $\Psi_a(z):=\Phi_a^{-1}(z)$ and let $\tilde{u}_a(z):=u(\Phi_a^{-1}(z))$. A direct computation shows that for $ i,j\in\{1,\dots,n-1\}$, $\eta,\xi\in\{n,n+1\}$ and $y\in\{y_n^2+ y_{n+1}^2 \leq \frac{{1}}{4}\}$
\begin{align*}
\p_{ij}\tilde{u}_a(z)&=\sum_{k,\ell=1}^{n-1}\p_{k\ell}u(\Psi_a(z))\p_i \Psi_a^{k}(z) \p_j\Psi_a^\ell(z)+\sum_{k=1}^{n-1}\p_k u(\Psi_a(z))\p_{ij}\Psi_a^k(z),\\
\p_{i\xi}\tilde{u}_a(z)&=\sum_{k=1}^{n-1}\p_{k\xi}u(\Psi_a(z))\p_i\Psi^{k}_a(z),\\
\p_{\xi}\tilde{u}_a(z)&=\p_{\xi}u(\Psi_a(z)),\\
\p_{\eta\xi}\tilde{u}_a(z)&=\p_{\eta\xi}u(\Psi_a(z)).
\end{align*}
By property (ii) of Lemma~\ref{prop:psi} and a similar argument as in Lemma~\ref{prop:psi2} we have that $\tilde{u}_a=u\circ \Psi_a \in X_{\delta,\epsilon}(\mathcal{B}_{{1}/4}^+)$, if $u\in X_{\delta,\epsilon}(\mathcal{B}_{1}^+)$. Thus, by Proposition \ref{prop:nonlin_map}, $F(\tilde{u}_a,z)\in Y_{\delta,\epsilon}(\mathcal{B}_{1/4}^+)$. By (ii) in Lemma~\ref{prop:psi}, $F(\tilde{u}_a,z)\big|_{z=\Phi_a(y)}\in Y_{\delta,\epsilon}(\mathcal{B}_{1/4}^+)$ as well. Outside of $\{y_{n}^2 + y_{n+1}^2 \leq \frac{1}{4}\}$, the statement follows without difficulties.\\
Next we show the regularity of $F_a(u,y)$ in $u$ and in the parameter $a$ which were claimed in the statements (i) and (ii). We first show that when $a=0$, $u\mapsto F(u,y)$ is $C^{k-1,\gamma-\epsilon}$. Indeed, we recall the expression of $F(v,y)$ from the beginning of Section~\ref{sec:grushin}. By a similar estimate as in Proposition~\ref{prop:linear} we have that $u\mapsto \sum_{i,j}a^{ij}(z'',-\p_nu,-\p_{n+1}u)G^{ij}(u)$ is $C^{k-1,\gamma-\epsilon}$ regular. To estimate the contribution $J(u)(b^j(u)\p_ju+b^n(u)y_n+b^{n+1}(u)y_{n+1})$ (in the case $k\geq 2$), we use the bound
\begin{align*}
\left|b(u_1,x)-b(u_2,y)\right|\lesssim \|b\|_{\dot{C}^{0,\gamma}}\|u_1-u_2\|^{\gamma-\epsilon}_{X_{\alpha,\epsilon}(\mathcal{B}_1^+)}|x-y|^\epsilon.
\end{align*}
Here we used the decomposition property of $u\in X_{\delta,\epsilon}(\mathcal{B}_1^+)$ from Proposition~\ref{prop:decompI}, that $b$ is $C^{0,\gamma}$ as a function of its arguments and the definition $b(u,y):=b(y'',-\p_nu,-\p_{n+1}u)$. Combining this we infer that
$$\|(D^{k-1}_{u_1}F-D^{k-1}_{u_2}F)(h^{k-1})\|_{Y_{\alpha,\epsilon}(\mathcal{B}_1^+)}\lesssim_k\|u_1-u_2\|_{X_{\alpha,\epsilon}(\mathcal{B}_1^+)}^{\gamma-\epsilon}\|h\|_{X_{\alpha,\epsilon}(\mathcal{B}_1^+)}^{k-1}.$$
To show the regularity of $u\mapsto F_a(u,y)$ for nonzero $a$, we use the definition of $F_a$ in \eqref{eq:Fa} and the computation for $D^2\tilde{u}_a$ from above. The argument is the same as for $a=0$.\\
Now we show the regularity of $F_a(u,y)$ in $a$ for fixed $u$. We only show the case when $k=1$. The remaining cases follow analogously. We recall that
\begin{align*}
F(u,z)&=\sum_{i,j}^{n+1}a^{ij}(z'',-\p_{n}u,-\p_{n+1}u)G^{ij}(u)+ f(z),\\
&\text{where } f(z)=-J(v(z))\left(\sum_{j=1}^{n-1}\tilde{b}^j(z)\p_j v(z)+\tilde{b}^n(z)z_n+\tilde{b}^{n+1}(z)z_{n+1}\right),
\end{align*}
and that $F_a(u,y)=F(u(\Psi_a(z)),z)\big|_{z=\Phi_a(y)}$ from \eqref{eq:Fa}.
Since $a\mapsto \Psi_a$ and $a\mapsto \Phi_a$ are real analytic and since $F(v,z):X_{\delta, \epsilon}(\mathcal{B}_{1}^+)\rightarrow Y_{\delta,\epsilon}(\mathcal{B}_{1}^+)$ is $C^{k-1,\gamma-\epsilon}$ regular in $v$, it suffices to note the regularity of the mappings
\begin{align*}
a &\mapsto a^{ij}((\Phi_a(y))'',-\p_n u,-\p_{n+1}u)G_a^{ij}(u)
\end{align*}
in $a$ as functions from $\R^{n-1}$ to $Y_{\delta,\epsilon}(\mathcal{B}_1^+)$. For the term $f(z)|_{z=\Phi_a(y)}$, since $f(z)$ has the form $f(z)=r(z)^{3-\epsilon}\tilde{f}(z)$ with $r(z)=(z_n^2+z_{n+1}^2)^{1/2}$ and since $\tilde{f}(z)$ is $C^{0,\gamma}$ in its tangential variables (due to the regularity of $\tilde{b}^j$ and the fact that $f(z)$ involves only lower order derivatives of $v$), the map
$$\R^{n-1}\ni a\mapsto \tilde{f}_a(y):=\tilde{f}(\Phi_a(y))\in C^{0,\epsilon}_\ast$$
is $C^{0,\gamma-\epsilon}$ regular.
\end{proof}
\subsubsection{Application of the implicit function theorem and regularity}
\label{sec:IFTAppl}
With this preparation we are now ready to invoke the implicit function theorem. We seek to apply the implicit function theorem in the spaces $X_{\delta,\epsilon}$ and $Y_{\delta,\epsilon}$ from Definition \ref{defi:spaces}. However, the Legendre function $v$ is only defined in $\mathcal{B}_1^+$. Thus, we extend it into the whole quarter space $Q_+$. In order to avoid difficulties at (artificially created) boundaries, we base our argument not on $v_a$ but instead consider $w_a:= v_a - v$, where $v$ is the original Legendre function. For this function we first note that $\supp(w_a) \Subset \mathcal{B}_{3 /4}^+$, which follows from the definition of the diffeomorphism $\Phi_a$. Moreover, $w_a$ solves the following fully nonlinear, degenerate elliptic equation:
\begin{align*}
\tilde{F}_a(w_a,y) := F_a(w_a+v,y)=0 \mbox{ in } \mathcal{B}_{1}^+.
\end{align*}
We extend $w_a$ to the whole quarter space $Q_+$ by setting $w_a=0$ in $Q_+\setminus \mathcal{B}_{1}^+$. Using $w_a=0$ in $Q_+\setminus \mathcal{B}_{3/4}^+$, the function $w_a$ solves the equation
\begin{align*}
G_a(w_a, y):= \eta(d_G(y,0)) \tilde{F}_a(w_a + v,y) + (1-\eta(d_G(y,0))) \D_G w_a= 0,
\end{align*}
in $Q_+$. Here $\eta: [0,\infty) \rightarrow \R$ is a smooth cut-off function with $\eta(s)=1$ for $s\leq \frac{3 }{4}$ and $\eta(s)=0$ for $s\geq 1$.
This extension is chosen such that the operator is of ``Baouendi-Grushin type'' around the degenerate set $P=\{y_n=y_{n+1}=0\}$ and the Baouendi-Grushin type estimates from Proposition \ref{prop:invert} and from the Appendix, Section \ref{sec:quarter_Hoelder}, can be applied in a neighborhood of $P$. The function $G_a$ satisfies the following mapping properties:
\begin{prop}\label{prop:invertible}
Assume that $a^{ij}\in C^{k,\gamma}$ with $k\geq 1$ and $\gamma\in (0,1]$. Given $a\in \R^{n-1}$, $G_a$ maps from $X_{\delta,\epsilon}$ into $Y_{\delta,\epsilon}$ for each $\delta\in (0,1)$ and $\epsilon\in (0,\gamma)$. Let $L:=D_w G_a\big|_{(a,w)=(0,0)}$ be the linearization of $G_a$ at $w=0$ and $a=0$. Then $L:X_{\delta,\epsilon}\rightarrow Y_{\delta,\epsilon}$ is invertible.
\end{prop}
\begin{proof}
Let $\bar{\eta}(y):= \eta(d_G(y,0))$. By Proposition~\ref{prop:error_gain} the Legendre function $v\in X_{\delta,\epsilon}(\mathcal{B}_{1}^+)$ for any $\delta\in (0,1)$. Thus by Proposition~\ref{prop:reg_a}, $\tilde{F}_a(w,y)=F_a(w+v,y)\in Y_{\delta,\epsilon}(\mathcal{B}_{1}^+)$ for any $w\in X_{\delta,\epsilon}$ and $y\in \mathcal{B}_{1}^+$. Since the Baouendi-Grushin Laplacian also has this mapping property, i.e. $\Delta_G:X_{\delta,\epsilon}\rightarrow Y_{\delta,\epsilon}$, and using the support assumption of $\eta$, we further observe that $G_a=\bar \eta F_a+(1-\bar\eta)\Delta_G$ maps $X_{\delta,\epsilon}$ into $Y_{\delta,\epsilon}$.
By (iv) in Proposition~\ref{prop:psi}, it is not hard to check that the linearization of $G_a$ at $(0,0)$ is given by
\begin{equation}
\begin{split}
\label{eq:lin2}
L=(D_w G_a)|_{(0,0)}= \bar{\eta} \left(F_{k\ell}(v,y)\p_{k\ell}+ F_{k}(v,y)\p_k\right) + (1-\bar{\eta})\D_G.
\end{split}
\end{equation}
Firstly, by Proposition \ref{prop:linear}, $L$ maps $X_{\delta,\epsilon}$ into $Y_{\delta,\epsilon}$. Moreover, $L$ can be written as $L=\Delta_G+\bar\eta\mathcal{P}_v$. Since $\|v-v_0\|_{X_{\alpha,\epsilon}(\mathcal{B}_1^+)}\lesssim \max\{\epsilon_0,c_\ast\}$ by Remark~\ref{rmk:close}, Proposition~\ref{prop:linear} implies that $\|\bar\eta \mathcal{P}_v(w)\|_{Y_{\delta,\epsilon}}\lesssim \max\{\epsilon_0,c_\ast\}\|w\|_{X_{\delta,\epsilon}}$ for $\epsilon\in (0,\gamma)$. Thus if $\epsilon_0$ and $c_\ast$ are sufficiently small, $L:X_{\delta,\epsilon}\rightarrow Y_{\delta,\epsilon}$ is invertible, as $\Delta_G:X_{\delta,\epsilon}\rightarrow Y_{\delta,\epsilon}$ is invertible (c.f. Lemma~\ref{lem:inverse}).
\end{proof}
\begin{thm}[H\"older regularity]
\label{prop:hoelder_reg_a}
Let $a^{ij}\in C^{k,\gamma}(B_1^+,\R^{(n+1)\times(n+1)}_{sym})$ with $k\geq 1$ and $\gamma\in (0,1)$. Let $w:B_1^+\rightarrow \R$ be a solution of the variable coefficient thin obstacle problem with metric $a^{ij}$. Then locally $\Gamma_{3/2}(w)$ is a $C^{k+1,\gamma}$ graph.
\end{thm}
\begin{proof}
\emph{Step 1: Almost optimal regularity.}
We apply the implicit function theorem to $G_a: X_{\delta,\epsilon}\rightarrow Y_{\delta,\epsilon}$ with $\delta$ and $\epsilon$ chosen such that $\epsilon\in (0,\gamma/2)$, $\delta=1-\epsilon \in (0,1)$ (as explained above, for $k=1$ we here interpret the lower order term as a function of $y$ in the linearization).
We note that as a consequence of Proposition~\ref{prop:reg_a}, for $v\in X_{\delta,\epsilon}$, $G_a(v)$ (interpreted as the function $G_{\cdot}(v): \R^{n-1}\ni a \mapsto G_a(v)\in Y_{\delta,\epsilon}$) is $C^{k-1,\gamma-\epsilon}$ in $a$. Thus, the implicit function theorem yields a unique solution $\tilde{w}_a$ in a neighborhood $B''_{\epsilon_0}(0)\times \mathcal{U}$ of $(0,0)\in \R^{n-1}\times X_{\delta,\epsilon}$ (c.f. Proposition \ref{prop:reg_a}). Moreover, the map $\R^{n-1}\ni a\mapsto \tilde{w}_a\in X_{\delta,\epsilon}$ is $C^{k-1,\gamma-\epsilon}$.
Hence, for all multi-indices $\beta=(\beta'',0,0)$ with $|\beta|=k-1$,
\begin{align*}
\left\| \frac{\partial^{\beta}_a \tilde{w}_{a_1}- \partial^{\beta}_{a} \tilde{w}_{a_2}}{|a_1-a_2|^{\gamma-\epsilon}}\right\|_{X_{\delta,\epsilon}}\leq C \left\| \partial_a^{\beta} \frac{G_{a_1}(\tilde{w}_{a_1})-G_{a_2}(\tilde{w}_{a_1})}{|a_1-a_2|^{\gamma-\epsilon}} \right\|_{Y_{\delta,\epsilon}} <\infty.
\end{align*}
In particular,
\begin{align}
\label{eq:differences}
\left[ \frac{\partial^{\beta}_a \p_{in}\tilde{w}_{a_1}- \partial^{\beta}_{a} \p_{in}\tilde{w}_{a_2}}{|a_1-a_2|^{\gamma-\epsilon}}\right]_{\dot{C}^{0,\delta}(P)}\leq C \left\| \partial_a^{\beta} \frac{G_{a_1}(\tilde{w}_{a_1})-G_{a_2}(\tilde{w}_{a_1})}{|a_1-a_2|^{\gamma-\epsilon}} \right\|_{Y_{\delta,\epsilon}} <\infty.
\end{align}
Since by Lemma~\ref{prop:psi2} $w_a=v_a-v\in \mathcal{U}$ if $a\in B''_{\epsilon_1}(0)$ for some sufficiently small radius $\epsilon_1$, the local uniqueness of the solution implies that $\tilde{w}_a=w_a$ for $a\in B''_{\epsilon_1}(0)$. Thus, $v_a=v+w_a=v+\tilde{w}_a$ is $C^{k-1,\gamma-\epsilon}$ in $a$. Combined with (\ref{eq:differences}) this in particular implies that for any multi-index $\beta=(\beta'',0,0)$ with $|\beta|= k-1$
\begin{align}
\label{eq:reg_tan}
\left[\frac{\partial^{\beta}_a \p_{in}v_{a_1} - \partial^{\beta}_a \p_{in}v_{a_2}}{|a_1 - a_2|^{\gamma - \epsilon}} \right]_{\dot{C}^{0,\delta}(P)}\leq C <\infty.
\end{align}
Recalling that the $a$-derivative corresponds to a tangential derivative in $\mathcal{B}_{1/2}$ and the fact that Hölder and Hölder-Zygmund spaces agree for non-integer values (c.f. \cite{Triebel}), this implies that for any multi-index $\beta=(\beta'',0,0)$ with $|\beta|\leq k-1$, $\p^\beta\p_{in} v\in C^{1,\gamma-2\epsilon}(P\cap \mathcal{B}_{1/2})$. By the characterization of the free boundary as $\Gamma_w = \{x \in B_{1}'| x_n = -\p_{n}v(x'',0,0)\}$ this implies that $\Gamma_w$ is a $C^{k+1,\gamma-2\epsilon}$ graph for any $\epsilon\in (0,\gamma/2)$. As $\epsilon>0$ can be chosen arbitrarily small, this completes the proof of the \emph{almost} optimal regularity result.\\
\emph{Step 2: Optimal regularity.}
In order to infer the \emph{optimal} regularity result, we argue by scaling and our previous estimates. More precisely, we have that
\begin{equation*}
\label{eq:est}
\begin{split}
[\Delta_{a}^{\gamma-\epsilon}\p_{in}\partial_a^{\beta} \tilde{w}_a ]_{\dot{C}^{0,\delta}}
&\leq \|\Delta_{a}^{\gamma-\epsilon}\p_{in}\partial_a^{\beta} \tilde{w}_a\|_{X_{\delta,\epsilon}} \leq C \| \Delta_{a}^{\gamma-\epsilon}\partial_a^{\beta} G_a(\tilde{w}_{a_1}) \|_{Y_{\delta,\epsilon}}\\
& \leq C \left( \| \Delta_{a}^{\gamma-\epsilon}\partial_a^{\beta} G_a^1(\tilde{w}_{a_1}) \|_{Y_{\delta,\epsilon}} + \| \Delta_{a}^{\gamma-\epsilon}\partial_a^{\beta} G_a^2(\tilde{w}_{a_1}) \|_{Y_{\delta,\epsilon}}\right).
\end{split}
\end{equation*}
Here $G^{1}_a(\cdot)$ is the term that originates from $ F^1(v,y)= \sum\limits_{i,j=1}^{n+1}\tilde{a}^{ij}(y)G^{ij}(v)$ and $G^{2}_a(\cdot)$ is the contribution that originates from the lower order contribution
$$ F^2(v,y)= -J(v(y))\left(\sum_{j=1}^{n-1}\tilde{b}^j(y)\p_j v(y)+\tilde{b}^n(y)y_n+\tilde{b}^{n+1}(y)y_{n+1}\right).$$
The notation $\D_a^{\gamma-\epsilon}$ denotes the difference quotient in $a$ with exponent $\gamma-\epsilon$.
We now consider the norms on the right hand side of (\ref{eq:est}) more precisely and consider their rescalings. A typical contribution of $\| \Delta_{a}^{\gamma-\epsilon}\partial_a^{\beta} G_a^2(\tilde{w}_{a_1}) \|_{Y_{\delta,\epsilon}}$ for instance is
\begin{align*}
[ r^{-(1+2\delta-\epsilon)}\Delta_{a}^{\gamma-\epsilon} \partial^{\beta}_a \tilde{b}^{j}_a J(v_{a_1})\p_j v_{a_1} ]_{C^{0,\epsilon}_{\ast}(\mathcal{B}_{2}^+)}.
\end{align*}
We focus on this contribution and on the case $k=1$. The other terms can be estimated by using similar ideas. We consider the rescaled function $v_{\lambda,a}(y)$, where $v_{\lambda}(y):=\frac{v(\delta_{\lambda}(y))}{\lambda^{3}}$ (with $\delta_\lambda(y)=(\lambda^2y'',\lambda y_n,\lambda y_{n+1})$) and $v_{\lambda,a}(y):= v_{\lambda}(\Phi_a(y))$. The function $w_{\lambda, a}(y):=v_{\lambda,a}(y)-v_{\lambda}(y)$ is defined as its analogue from above. It is compactly supported in $\mathcal{B}_{3/4}^+$ (by definition of $\Phi_a$) and the functions $v_{\lambda}$ and $w_{\lambda,a}$ satisfy similar equations as $v, w_a$. Thus, we may apply estimate (\ref{eq:differences}) to $w_{\lambda,a}$. Inserting $\delta =1 -\epsilon$, using the support condition for $w_{\lambda,a}$ yields (with slight abuse of notation, as there are additional right hand side contributions, which however by the compact support assumption on $w_{\lambda,a}$ have the same or better scaling)
\begin{align*}
[\p_{ijn}v_{\lambda}]_{C^{0,\gamma-2\epsilon}(\mathcal{B}_{1}^+\cap P)} &\leq C \lambda^{-1}[r^{-3+3\epsilon}J(v)|_{\delta_{\lambda}(y)}\p_j v|_{\delta_{\lambda}(y)}]_{C^{0,\epsilon}_{\ast}}[\tilde{b}^{j}|_{\delta_{\lambda}(y)}]_{C^{0,\gamma}(\mathcal{B}_{2}^+)}\\
& \leq C \lambda^{-1}[r^{-3+\epsilon}J(v)|_{\delta_{\lambda}(y)}\p_j v|_{\delta_{\lambda}(y)}]_{C^{0,\epsilon}_{\ast}(\mathcal{B}_{2}^+)}[\tilde{b}^{j}|_{\delta_{\lambda}}]_{C^{0,\gamma}(\mathcal{B}_{2}^+)}.
\end{align*}
Comparing this to the left hand side of the estimate and rescaling both sides of the inequality therefore amounts to
\begin{align*}
\lambda^{2+2\gamma-4\epsilon} [\p_{ijn}v]_{C^{0,\gamma-2\epsilon}(\mathcal{B}_{ \lambda}^+\cap P)} & \leq C \lambda^{2+2\gamma}[r^{-3+\epsilon}J(v)\p_j v]_{C^{0,\epsilon}_{\ast}(\mathcal{B}_{2 \lambda}^+)}[\tilde{b}^{j}]_{C^{0,\gamma}(\mathcal{B}_{2 \lambda}^+}),
\end{align*}
which yields
\begin{align*}
[\p_{ijn}v]_{C^{0,\gamma-2\epsilon}(\mathcal{B}_{ \lambda}^+\cap P)} & \leq C \lambda^{4\epsilon}[r^{-3+\epsilon}J(v)\p_j v]_{C^{0,\epsilon}_{\ast}(\mathcal{B}_{2\lambda}^+)}[\tilde{b}^{j}]_{C^{0,\gamma}(\mathcal{B}_{2 \lambda}^+)}.
\end{align*}
As a result considering two points $x,y\in P$ with $|x-y|= \lambda^2$, yields
\begin{align*}
\frac{|\p_{ijn}v(x)-\p_{ijn}v(y)|}{|x-y|^{\gamma-2\epsilon}} \leq C \lambda^{4\epsilon} = C |x-y|^{2\epsilon}.
\end{align*}
Thus,
\begin{align*}
[\p_{ijn}v]_{C^{0,\gamma}(\mathcal{B}_{ \lambda}^+\cap P)} \leq C,
\end{align*}
which proves the optimal regularity result.
\end{proof}
\begin{rmk}[$\gamma=1$]
\label{rmk:gamma=1}
As expected from elliptic regularity, we can only deduce the full $C^{k+1,\gamma}$ regularity of the free boundary in the presence of $C^{k,\gamma}$ metrics, if $\gamma=1$. This is essentially a consequence of the elliptic estimates of Proposition \ref{prop:error_gain2}. On a technical level this is exemplified in the fact that in Step 2 of the previous proof, we for instance also have to deal with the term $(\D_a^{\gamma-\epsilon} a^{n,n})\p_{n+1,n+1}v_{\lambda}$ with the expansion $\p_{n+1,n+1}v(y) = a_1(y'')y_n + r^{1+2\delta-\epsilon}C_{n+1,n+1}(y)$ with $\delta\in (0,1)$. As the coefficients $a_1(y'')$ are in general not better than $C^{0,\delta}$, we do not have the full gain of $\lambda^{4\epsilon}$ if $\gamma=1$.
\end{rmk}
\begin{rmk}[Optimal regularity]
\label{rmk:optreg}
Let us comment on the optimality of the gain of the free boundary regularity with respect to the regularity of the metric $a^{ij}$:
Proposition \ref{prop:bulk_eq} in combination with our linearization results (c.f. Example \ref{ex:linear} and Section \ref{sec:grushin}) illustrates that $F$ can be viewed as a nonlinear perturbation of the degenerate, elliptic (second order) Baouendi-Grushin operator with metric $a^{ij}$. As such, we can not hope for a gain of more than \emph{two orders} of regularity for $v$ compared to the regularity of the metric $a^{ij}$ (by interior regularity in appropriate Hölder spaces, c.f. Section \ref{sec:holder}). Hence, for the regular free boundary we can in general hope for a gain of at most \emph{one order} of regularity with respect to the regularity of the metric. This explains our expectation that the regularity results from Theorem \ref{thm:higher_reg} are sharp higher order regularity results.\\
By a simple transformation it is possible to construct an example to the sharpness of this claim:
In $\R^3$ the function $w(x_1,x_2,x_3)=\Ree((x_2-x_1)/\sqrt{2}+ix_3)^{3/2}$ is a solution to the thin obstacle problem $\Delta w=0$ in $\R^3_+$ with the free boundary $\Gamma_w=\{(x_1,x_2,0)\in B'_1: x_2=x_1\}$. Applying a transformation of the form $$y(x):=(x_1,h(x_2),x_3),$$
with $h$ being a $W^{k+1,p}$ diffeomorphism from $(-1,1)$ to $(-1,1)$ yields that $\tilde{w}(y):=\Ree((h^{-1}(y_2)-y_1)/\sqrt{2}+iy_3)^{3/2}$ solves the variable coefficient thin obstacle problem
\begin{align*}
\p_{11}\tilde{w}+ \p_2(h'(x_2)\p_2 \tilde{w})+\p_{33}w=0 \mbox{ in } B'_1,
\end{align*}
with Signorini conditions on $B'_1$. We note that the free boundary of $\tilde{w}$ is given by the graph $\Gamma_{\tilde{w}}=\{(y_1,y_2,0)\in B'_1: y_2=h(y_1)\}$. If $h$ is not better than $W^{k+1,p}$ regular, the coefficients in the bulk equation are no more than $W^{k,p}$ regular. The free boundary is $W^{k+1,p}$ regular. Since it is a graph, it does not admit a more regular parametrization.\\
We further note that our choice of function spaces was crucial in deducing the full gain of regularity for the free boundary with respect to the metric. Indeed, considering the equation (\ref{eq:nonlineq1}), we note that also second order derivatives of the metric are involved. Yet, in order to deduce regularity of the free boundary (which corresponds to \emph{partial} regularity of the Legendre function $v$) this loss of regularity does not play a role as it is a ``lower order bulk term'' (this is similar in spirit to the gain of regularity obtained in boundary Harnack inequalities).
\end{rmk}
Finally, we give the argument for the analyticity of the free boundary in the case that the coefficients $a^{ij}$ are analytic:
\begin{thm}[Analyticity]
\label{prop:analytic}
Let $a^{ij}:B_{1}^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ be an analytic tensor field. Let $w:B_{1}^+ \rightarrow \R$ be a solution of the variable coefficient thin obstacle problem with metric $a^{ij}$. Then locally $\Gamma_{3/2}(w)$ is an analytic graph.
\end{thm}
\begin{proof}
This follows from the analytic implicit function theorem (c.f. \cite{Dei10}). Indeed, due to Proposition \ref{prop:reg_a}, $F_a$ is a real analytic function in $a$ and hence also $G_a$ is a real analytic function in $a$. Applying the analytic implicit function theorem similarly as in Step 1 of the previous proof, we obtain an in $a$ analytic function $\tilde{w}_a$. As before, this coincides with our function $w_a$. Therefore $w_a$ depends analytically on $a$. As differentiation with respect to $a$ however directly corresponds to differentiation with respect to the tangential directions $y''$, $w$ (and hence $v$) is an analytic function in the tangential variables.
\end{proof}
\begin{rmk}[Regularity in the normal directions]
In Theorems \ref{prop:hoelder_reg_a} and \ref{prop:analytic} we proved partial analyticity for the Legendre function $v$: We showed that in the \emph{tangential} directions, the regularity of $v$ in a quantitative way matches that of the metric (i.e. a $C^{k,\gamma}$ metric yields $C^{k+1,\gamma}$ regularity for $\p_n v(y'',0,0)$). Although this suffices for the purposes of proving regularity of the (regular) free boundary, a natural question is whether it is also possible to obtain corresponding higher regularity for $v$ in the \emph{normal} directions $y_n, y_{n+1}$. Intuitively, an obstruction for this stems from working in the corner domain $Q_+$. That this set-up of a corner domain really imposes restrictions on the normal regularity can be seen by checking a compatibility condition: As we are considering an expansion close to the regular free boundary point, we know that the Legendre function asymptotically behaves like a multiple of the function $v_0(y)=-(y_1^3 - 3 y_1 y_2^2)$. If additional regularity were true in the normal directions, we could expand the Legendre function $v$ further, for instance into a fifth order polynomial (with symmetry obeying the mixed Dirichlet-Neumann boundary conditions), which has $v_0$ as its leading order expansion. Hence, working in the two-dimensional corner domain $Q_+:=\{y_1\geq 0, y_2\leq 0\}$, we make the ansatz that
\begin{equation}
\label{eq:ansatz}
v(y) = -(y_1^3 - 3y_1 y_2^2) + c_1 y_1^4 + c_2 y_1^2 y_2^2 + c_3 y_1 y_2^3 + c_4 y_1^5 + c_5 y_1^3 y_1^2 + c_6 y_1^2 y_2^3 + c_7 y_1 y_2^4 + h.o.t,
\end{equation}
where $h.o.t$ abbreviates terms of higher order. We seek to find conditions on the metric $a^{ij}$ which ensure that such an expansion for $v$ up to fifth order exists. Without loss of generality we may further assume that
\begin{align*}
a^{ij}(0) = \delta^{ij}, \ a^{12}(x_1,0)=a^{21}(x_1,0)=0,
\end{align*}
which corresponds to a normalization at zero and the off-diagonal condition on the plane $\{x_2=0\}$. Transforming the equation
\begin{align*}
\p_i a^{ij} \p_j w = 0 \mbox{ in } \R^2_+,
\end{align*}
into the Legendre-Hodograph setting with the associated Legendre function $v$ yields
\begin{align*}
&a^{11}(-\p_1 v, -\p_2 v) \p_{22}v + a^{22}(-\p_1 v, -\p_2 v) \p_{11}v - 2 a^{12}(-\p_1 v, - \p_2 v) \p_{12}v\\
& - J(v)[\p_1 a^{11}(-\p_1 v, -\p_2 v) +\p_2 a^{12}(-\p_1 v, - \p_2 v) ] y_1 \\
&- J(v)[\p_1 a^{12}(-\p_1 v, - \p_2 v) +\p_2 a^{22}(-\p_1 v, -\p_2 v) ] y_2 =0 \mbox{ in } Q_+:= \{y_1,y_1 \geq 0\}.
\end{align*}
Here $J(v) = \det\begin{pmatrix} \p_{11}v & \p_{12}v\\
\p_{21}v & \p_{22}v \end{pmatrix}$.
Carrying out a Taylor expansion of the metric thus gives
\begin{align*}
&\Delta v -2\left(\p_2a^{12}(0)(-\p_2 v)\right)\p_{12}v\\
&+\left((-\p_1 v)\p_1a^{11}(0)+(-\p_2 v)\p_2a^{11}(0)\right)\p_{22}v+\left((-v_1)\p_1a^{22}(0)+(-v_2)\p_2a^{22}(0)\right)\p_{11}v\\
&-\det\begin{pmatrix}
\p_{11}v & \p_{12}v\\
\p_{21}v & \p_{22}v
\end{pmatrix}\left((\p_1a^{11}(0)+\p_2a^{21}(0))y_1+\p_2a^{22}(0)y_2\right)+h.o.t.=0.
\end{align*}
Inserting the ansatz (\ref{eq:ansatz}) into this equation, matching all terms of order up to three and using the off-diagonal condition, eventually yields the compatibility condition
\begin{align*}
\p_2(a^{11} + a^{22})(0)=0.
\end{align*}
Due to our normalization this necessary condition for having a polynomial expansion up to degree five can thus be formulated as
\begin{align*}
(\p_2 \det(a^{ij}))(0)=0.
\end{align*}
In particular this shows that on the transformed side, i.e. in the Legendre-Hodograph variables, one cannot expect arbitrary high regularity for $v$ in the normal directions $y_n, y_{n+1}$ in general. Compatibility conditions involving the metric $a^{ij}$ have to be satisfied to ensure this.
\end{rmk}
\section{$W^{1,p}$ Metrics and Nonzero Obstacles}
\label{sec:W1p}
In this section we consider the previous set-up in the presence of inhomogeneities $f\in L^p$ and possibly only Sobolev regular metrics. More precisely, in this section we assume that $a^{ij}:B_1^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ is a uniformly elliptic $W^{1,p}$, $p\in (n+1,\infty]$, metric and consider a solution $w$ of the variable coefficient thin obstacle problem with this metric:
\begin{equation}
\label{eq:inhom}
\begin{split}
\p_{i} a^{ij} \p_j w & = f \mbox{ in } B_1^+,\\
w \geq 0,\ \p_{n+1}w \leq 0, \ w\p_{n+1}w&=0 \mbox{ on } B_1'.
\end{split}
\end{equation}
We will discuss two cases:
\begin{itemize}
\item[(1)] $f=0$, $a^{ij}\in W^{1,p}$ with $p\in (n+1,\infty]$,
\item[(2)] $a^{ij}\in W^{1,p}$, $f\in L^p$ with $p\in (2(n+1),\infty]$.
\end{itemize}
In both cases all the normalization conditions (A1)-(A7) from Section \ref{sec:conventions} as well as the asymptotic expansions (c.f. Proposition~\ref{prop:asym2}) hold. We observe that case (2) in particular contains the setting with non-flat obstacles.
\subsection{Hodograph-Legendre transformation for $W^{1,p}$ metrics}
\label{sec:ext}
In the sequel, we discuss how the results from Sections \ref{sec:asymp}- \ref{sec:Legendre} generalize to the less regular setting of $W^{1,p}$, $p\in (n+1,\infty]$, metrics. We note that in this case the solution $w$ is only $W^{2,p}_{loc}(B_1^+\setminus\Gamma_w)$ regular away from the free boundary $\Gamma_w$. Thus, our Hodograph-Legendre transformation method from the previous sections does not apply directly (as it relies on the pointwise estimates of $D^2v$, and hence $D^2w$). Thus, a key ingredient in our discussion of this set-up will be the splitting result, Proposition 3.9, from \cite{KRSI}. In order to apply it, we extend $w$ and the metric $a^{ij}$ from $B_1^+$ to $B_1$ by an even reflection as in \cite{KRSI}. We now split our solution into two components, $w=u+\tilde{u}$, where $\tilde{u}$ solves
\begin{align}
\label{eq:split1}
a^{ij}\p_{ij}\tilde{u}-\dist(x,\Gamma_w)^{-2}\tilde{u}=f - (\p_i a^{ij})\p_j w \text{ in } B_1\setminus \Lambda_w, \quad \tilde{u}=0\text{ on }\Lambda_w,
\end{align}
and the function $u$ solves
\begin{align}
\label{eq:split2}
a^{ij}\p_{ij}u=-\dist(x,\Gamma_w)^{-2}\tilde{u}\text{ in } B_1\setminus \Lambda_w, \quad u=0\text{ on } \Lambda_w.
\end{align}
As in \cite{KRSI} the intuition is that $\tilde{u}$ is a ``controlled error'' and that $u$ captures the essential behavior of $w$. Moreover, as we will see later, $u$ will be $C^{2,1-\frac{n+1}{p}}_{loc}$ regular away from $\Gamma_w$ and that $\Gamma_w = \Gamma_u$ (c.f. the discussion below Lemma \ref{lem:lower1'}). Thus, in the sequel, we will apply the Hodograph-Legendre transformation to the function $u$.\\
In order to support this intuition, we recall the positivity of $\p_e u$ as well as the fact that $u$ inherits the complementary boundary conditions from $w$.
\begin{lem}[\cite{KRSI}, Lemma 4.11]
\label{lem:lower1'}
Let $a^{ij}\in W^{1,p}(B_1^+, \R^{(n+1)\times(n+1)}_{sym})$ and let $f\in L^p(B_1^+)$. Suppose that either
\begin{itemize}
\item[(1)] $p\in (n+1,\infty]$ and $f=0$ or,
\item[(2)] $p\in (2(n+1),\infty]$.
\end{itemize}
Let $w:B_1^+ \rightarrow \R$ be a solution to the thin obstacle problem with inhomogeneity $f$, and let $u$ be defined as at the beginning of this section. Then we have that $u\in C^{2,1-\frac{n+1}{p}}_{loc}(B_1^+\setminus \Gamma_w)\cap C^{1,\min\{\frac{1}{2},1-\frac{n+1}{p}\}}_{loc}(B_1^+)$. Moreover, there exist constants $c, \eta>0$ such that for $e\in \mathcal{C}'_\eta(e_n)$, $\p_eu $ satisfies the lower bound
\begin{align*}
\p_{e}u (x) \geq c\dist(x,\Lambda_w)\dist(x,\Gamma_w)^{-\frac{1}{2}}.
\end{align*}
A similar statement holds for $\p_{n+1}u$ if $\Lambda_w$ is replaced by $\Omega_w$.
\end{lem}
We remark that the lower bound in Lemma 3.13 does not necessarily hold for $\p_ew$. This is due to the insufficient decay properties of $\p_e \tilde{u}$ in the decomposition $\p_ew = \p_e\tilde{u}+\p_eu$. More precisely, the decay of $\p_e\tilde{u} $ to $\Lambda_w$ is in general only of the order $\dist(x,\Lambda_w)^{1-\frac{n+1}{p}}$, which cannot be controlled by $\dist(x,\Lambda_w)$.\\
We further note that the symmetry of $u$ about $x_{n+1}$ and the regularity of $u$ imply that $\p_{n+1}u=0$ in $B'_1\setminus \Lambda_w$. In particular, this yields the complementary boundary conditions:
\begin{align*}
u \p_{n+1}u = 0 \mbox{ on } B_1'.
\end{align*}
Most importantly, Lemma \ref{lem:lower1'} combined with the previous observations on the behavior of $\nabla u$ on $B_1'$ implies that $\Gamma_w = \Gamma_u$. Hence, seeking to investigate $\Gamma_w$, it suffices to study $u$ and its boundary behavior. In this context, Lemma \ref{lem:lower1'} plays a central role as it allows us to deduce the sign conditions for $\p_e u$ and $\p_{n+1}u$ which are crucial in determining the image of the Legendre-Hodograph transform which we will associate with $u$.\\
In accordance with our intuition that $\tilde{u}$ is a ``controlled error", the function $u$ inherits the asymptotics of the solution $w$ around $\Gamma_w$. As in Proposition \ref{prop:invertibility} in Section \ref{sec:Hodo}, this is of great importance in proving the invertibility of the Legendre-Hodograph transform which we will associate with $u$. We formulate the asymptotic expansions in the following proposition:
\begin{prop}
\label{prop:improved_reg1}
Let $a^{ij}\in W^{1,p}(B_1^+, \R^{(n+1)\times(n+1)}_{sym})$ and let $f\in L^p(B_1^+)$.
Suppose that either
\begin{itemize}
\item[(1)] $p\in (n+1,\infty]$ and $f=0$ or,
\item[(2)] $p\in (2(n+1),\infty]$.
\end{itemize}
Let $w:B_1^+ \rightarrow \R$ be a solution to the thin obstacle problem with inhomogeneity $f$, and let $u$ be defined as at the beginning of this section. There exist small constants $\epsilon_0>0$ and $c_\ast>0$ depending on $n,p$ such that if
\begin{itemize}
\item[(i)] $ \|w-w_{3/2}\|_{C^1(B_1^+)}\leq \epsilon_0$,
\item[(ii)]$ \|\nabla a^{ij}\|_{L^p(B_1^+)}+\|f\|_{L^p(B_1^+)}\leq c_\ast,$
\end{itemize}
then the asymptotics (i)-(iii) in Proposition~\ref{prop:asym2} hold for $\p_eu$, $\p_{n+1}u$ and $u$. The exponent $\alpha$ in the error term satisfies $\alpha\in (0,1-\frac{n+1}{p}]$ in case (1) and $\alpha\in (0,\frac{1}{2}-\frac{n+1}{p}]$ in case (2).
\end{prop}
\begin{proof}
By the growth estimate of Remark 3.11 in \cite{KRSI} we have that
\begin{equation}
\label{eq:auxv}
\begin{split}
|\tilde{u}(x)|& \lesssim c_\ast \dist(x,\Lambda_w)\dist(x,\Gamma_w)^{\frac{3}{2}-\frac{n+1}{p}} \text{ in case (1);}\\
|\tilde{u}(x)|& \lesssim c_\ast \dist(x,\Lambda_w)\dist(x,\Gamma_w)^{1-\frac{n+1}{p}} \text{ in case (2).}
\end{split}
\end{equation}
In particular this implies that
\begin{align*}
&|\tilde{u}(x)|\lesssim c_\ast \dist(x,\Gamma_w)^{\frac{3}{2}+\delta_0},\quad
|\nabla \tilde{u}(x)|\lesssim c_\ast \dist(x,\Gamma_w)^{\frac{1}{2}+\delta_0},\\
&\text{where }\delta_0=\left\{\begin{array}{ll}
1-\frac{n+1}{p} &\text{ if } p\in (n+1,2(n+1)],\\
\frac{1}{2}-\frac{n+1}{p} &\text{ if } p\in (2(n+1),\infty].
\end{array}
\right.
\end{align*}
Since $\delta_0>0$, in both cases the functions $\tilde{u}$ and $\nabla \tilde{u}$ are of higher vanishing order at $\Gamma_w$ compared to the leading term in the corresponding asymptotics of $w$ and $\nabla w$ (which are of order $\dist(x,\Gamma_w)^{3/2}$ and $\dist(x,\Gamma_w)^{1/2}$).
\end{proof}
In addition to these results the second order asymptotics for $u$ (not for the whole function $w$) remain valid under the conditions of Proposition \ref{prop:improved_reg1}. More precisely we have the following result:
\begin{prop}
\label{prop:improved_reg'}
Under the same assumptions as in Proposition~\ref{prop:improved_reg1}, we have the following:
For each $x_0\in \Gamma_w\cap B^+_{1/2}$, for all $x$ in an associated non-tangential cone $\mathcal{N}_{x_0}$ and for all multi-indeces $\beta$ with $|\beta|\leq 2$,
\begin{align*}
\left|\p^\beta u(x)-\p^\beta \mathcal{W}_{x_0}(x)\right|& \leq C_{n,p,\beta} \max\{\epsilon_0, c_*\} |x-x_0|^{\frac{3}{2}+\alpha-|\beta|},\\
\left[\p^\beta u-\p^\beta \mathcal{W}_{x_0}\right]_{\dot{C}^{0,\gamma}(\mathcal{N}_{x_0}\cap (B_{3\lambda/4}(x_0)\setminus B_{\lambda/2}(x_0)))}& \leq C_{n,p,\beta} \max\{\epsilon_0, c_*\} \lambda^{\frac{3}{2}+\alpha-\gamma-|\beta|}.
\end{align*}
Here $\gamma=1-\frac{n+1}{p}$ and $\lambda \in (0,1)$.
\end{prop}
\begin{proof}
We only prove the case of $|\beta|=2$, the other cases are already contained in Proposition \ref{prop:improved_reg1}. Since the arguments for case (1) and (2) are similar we only prove case (2), i.e. $p\in (2(n+1),\infty]$. As in Proposition \ref{prop:improved_reg} the result follows from scaling.
We consider the function
\begin{align*}
\bar{u}(x):= \frac{u(x_0+ \lambda x)- \mathcal{W}_{x_0}(x_0+\lambda x)}{\lambda^{3/2 + \alpha}},
\end{align*}
and note that it satisfies
\begin{align*}
a^{ij}(x_0 + \lambda \cdot) \p_{ij }\bar{u} = \tilde{G} + \tilde{g}_1, \quad \ell\in\{1,\dots,n\},
\end{align*}
for $x\in \mathcal{N}_{0}\cap (B_{1}\setminus B_{1/4})$ and $\mathcal{N}_0:= \{x\in B_{1/4}^+| \ \dist(x,\Gamma_{w_{x_0,\lambda}}) > \frac{1}{2}|x|\}$. Here
\begin{align*}
\tilde{G}(x) & := -\lambda^{1/2 - \alpha}\dist(x_0+\lambda x,\Gamma_w)^{-2}\tilde{u}(x_0+\lambda x),\\
&=-\lambda^{-\frac{3}{2}-\alpha}\dist(x,\Gamma_{w_{x_0,\lambda}})^{-2}\tilde{u}(x_0+\lambda x),\\
\tilde{g}_1(x) & := \lambda^{1/2-\alpha} (a^{ij}(x_0 + \lambda x)-a^{ij}(x_0)) \partial_{ij} \mathcal{W}_{x_0}(x_0 +\lambda x).
\end{align*}
In the definition of $\tilde{g}_1$ we have used that $a^{ij}(x_0)\p_{ij}\mathcal{W}_{x_0}=0$ in $ \mathcal{N}_{x_0}\cap (B_{\lambda}(x_0)\setminus B_{\lambda/4}(x_0))$. Using \eqref{eq:auxv} and the regularity of $\tilde{u}$ (and abbreviating $\gamma=1-\frac{n+1}{p}$) yields
\begin{align*}
\|\tilde{G}\|_{C^{0,\gamma}(\mathcal{N}_{0}\cap (B_{1}\setminus B_{1/4}))}\leq C \lambda^{-\frac{3}{2}-\alpha} \lambda^{2-\frac{n+1}{p}}=C\lambda^{\frac{1}{2}-\alpha-\frac{n+1}{p}}.
\end{align*}
Recalling the $C^{0,1-\frac{n+1}{p}}$ regularity of $a^{ij}$ and the explicit expression of $\mathcal{W}_{x_0}$, we estimate
\begin{align*}
\|\tilde{g}_1\|_{C^{0,\gamma}(\mathcal{N}_{0}\cap (B_{1}\setminus B_{1/4}))}\leq C\lambda^{\frac{1}{2}-\alpha}\lambda^{1-\frac{n+1}{p}}\lambda^{-\frac{1}{2}}=C\lambda^{1-\alpha-\frac{n+1}{p}}.
\end{align*}
Hence, applying the interior Schauder estimate to $\bar u$ we obtain
\begin{align*}
\|\bar u\|_{C^{2,\gamma}(\mathcal{N}_{0}\cap (B_{3/4}\setminus B_{1/2}))}&\leq C\left(\|\tilde{ G}\|_{C^{0,\gamma}(\mathcal{N}_0\cap (B_{1}\setminus B_{1/4}))}+\|\tilde{g}_1\|_{C^{0,\gamma}(\mathcal{N}_0\cap (B_{1}\setminus B_{1/4}))}\right.\\
&\quad \left.+\|\bar u\|_{L^\infty(\mathcal{N}_0\cap (B_{1}\setminus B_{1/4}))}\right)\\
&\leq C\left(\lambda^{\frac{1}{2}-\alpha-\frac{n+1}{p}}+1 \right)\leq C \quad (\text{since }\alpha\in (0,\frac{1}{2}-\frac{n+1}{p}]).
\end{align*}
Scaling back, the error estimates become
\begin{align*}
\|\p_e u -\p_e \mathcal{W}_{x_0}\|_{L^\infty(\mathcal{N}_{x_0}\cap (B_{3\lambda/4}(x_0)\setminus B_{\lambda/2}(x_0)))}&\leq C\lambda^{\frac{1}{2}-\alpha},\\
\|\p_{ee'}u-\p_{ee'}\mathcal{W}_{x_0}\|_{L^\infty(\mathcal{N}_{x_0}\cap (B_{3\lambda/4}(x_0)\setminus B_{\lambda/2}(x_0)))}
&\leq C\lambda^{-\frac{1}{2}-\alpha},\\
[\p_{ee'}u-\p_{ee'}\mathcal{W}_{x_0}]_{\dot{C}^{0,\gamma}(\mathcal{N}_{x_0}\cap (B_{3\lambda/4}(x_0)\setminus B_{\lambda/2}(x_0)))}
&\leq C\lambda^{-\frac{1}{2}-\alpha-\gamma}.
\end{align*}
Since this holds for every $\lambda\in (0,1)$, we obtain the asymptotic expansion for $\p_e w $ and $\p_{e e'}w$. The asymptotics for $\p_{ij}w$ with $i$ or $j=n+1$ are derived analogously.
\end{proof}
Due to the above discussion, the associated Hodograph transform with respect to $u$,
$$
T(x):=(x'',\p_n u(x), \p_{n+1}u(x)),$$
still enjoys all the properties stated in Section \ref{sec:Hodo}. In particular, it is possible to define the associated Legendre function
\begin{align}
\label{eq:Leg_split}
v(y):= u(x)-x_n y_n - x_{n+1} y_{n+1},
\end{align}
for $x= T^{-1}(y)$. This function satisfies an analogous nonlinear PDE as the one from Section \ref{sec:Legendre}:
\begin{align*}
\tilde{F}(v,y) = g(y).
\end{align*}
Here,
\begin{equation}\label{eq:w1p}
\begin{split}
\tilde{F}(v,y)&=-\sum_{i,j=1}^{n-1}\tilde{a}^{ij}\det\begin{pmatrix}
\p_{ij}v& \p_{in}v & \p_{i,n+1}v\\
\p_{jn}v& \p_{nn}v & \p_{n,n+1}v\\
\p_{j,n+1}v & \p_{n,n+1}v &\p_{n+1,n+1}v
\end{pmatrix}\\
&+2\sum_{i=1}^{n-1}\tilde{a}^{i,n}\det\begin{pmatrix}
\p_{in}v & \p_{i,n+1}v\\
\p_{n,n+1}v & \p_{n+1,n+1}v
\end{pmatrix}+2 \sum_{i=1}^{n-1}\tilde{a}^{i,n+1}\det\begin{pmatrix}
\p_{i,n+1}v & \p_{in}v\\
\p_{n,n+1}v & \p_{nn}v
\end{pmatrix}\\
&+\tilde{a}^{nn}\p_{n+1,n+1}v+\tilde{a}^{n+1,n+1}\p_{nn}v-2\tilde{a}^{n,n+1}\p_{n,n+1}v,\\
\tilde{a}^{ij}(y)&:=a^{ij}(x)\big|_{x=(y'',-\p_nv(y),-\p_{n+1}v(y))},\\
J(v)& := \p_{nn} v(y) \p_{n+1,n+1}v(y) - (\p_{n,n+1}v(y))^2,\\
g(y) &:= -J(v(y)) \dist(T^{-1}(y), T^{-1}(P))^{-2}\tilde{u}(T^{-1}(y)),\quad P:=\{y_n=y_{n+1}=0\}.
\end{split}
\end{equation}
From the asymptotics of $J(v)$ and \eqref{eq:auxv} for $\tilde{u}$ we have
\begin{align}
\label{eq:w1p_g}
|g(y)|\leq \left\{ \begin{array}{ll}\
C(y_n^2+y_{n+1}^2)^{\frac{3}{2}-\frac{n+1}{p}} &\text{ if } p\in (n+1,\infty] \text{ and } f=0,\\
C(y_n^2+y_{n+1}^2)^{1-\frac{n+1}{p}} &\text{ if }p\in (2(n+1),\infty].
\end{array}\right.
\end{align}
The result of Proposition \ref{prop:improved_reg'} in combination with an argument as in the proof of Proposition \ref{prop:holder_v} also yields that $v\in X_{\alpha,\epsilon}$ for a potentially very small $\alpha>0$.\\
We summarize all this in the following Proposition:
\begin{prop}
\label{eq:bulk_new}
Let $a^{ij}\in W^{1,p}$ and let $f\in L^p$.
Suppose that either
\begin{itemize}
\item[(1)] $p\in (n+1,\infty]$ and $f=0$ or,
\item[(2)] $p\in (2(n+1),\infty]$.
\end{itemize}
Let $v:T(B_{1/2}^+) \rightarrow \R$ be the Legendre function associated with $u$ defined in (\ref{eq:Leg_split}). Then $v\in C^1(T(B_{1/2}^+) )\cap X_{\alpha,\epsilon}(\mathcal{B}_{r_0}^+)$ for some $\alpha\in (0,1)$ (which is the same as in Proposition \ref{prop:improved_reg'}) and it satisfies the fully nonlinear equation
\begin{align*}
\tilde{F}(v,y) = g(y).
\end{align*}
Here $\tilde{F}, g$ are as in (\ref{eq:w1p}) and $g$ satisfies the decay estimate (\ref{eq:w1p_g}).
\end{prop}
\begin{rmk}
\label{rmk:decay}
We note that the leading contribution in the decay estimate for $g$ originates from the decay behavior of $\tilde{u}$ in \eqref{eq:auxv}. Therefore, the decay of $g$ is influenced by $-(\p_{i}a^{ij})\p_j w$ and by the inhomogeneity $f$ from \eqref{eq:split1}.
\end{rmk}
\subsection{Regularity of the free boundary}
\label{sec:free_boundary_reg_1}
In this section we discuss the implications of the results from Section \ref{sec:ext} on the free boundary regularity. In order to understand the different ingredients to the regularity results, we treat two different scenarios: First we address the setting of $W^{1,p}$ metrics with $p\in (n+1,\infty]$, and zero obstacles, i.e. with respect to Sections \ref{sec:HLTrafo} - \ref{sec:fb_reg} we present a result under even weaker regularity assumptions of the metric (c.f. Section \ref{subsec:w1p_zero}). Secondly, in Section \ref{sec:nonzero} we address the set-up with inhomogeneities. This in particular includes the case of non-zero obstacles. We treat this in the $W^{1,p}$ and the $C^{k,\gamma}$ framework.
\subsubsection{$W^{1,p}$ metrics without inhomogeneity}
\label{subsec:w1p_zero}
We now specialize to the setting in which $a^{ij}\in W^{1,p}$ with $p\in (n+1,\infty]$ and $f=0$ in (\ref{eq:inhom}). In this framework, we prove the following quantitative regularity result for the free boundary:
\begin{prop}[$C^{1,1-\frac{n+1}{p}}$ regularity]
\label{prop:W1p}
Let $a^{ij}\in W^{1,p}$ with $p\in (n+1,\infty]$ and $f=0$. Let $w$ be a solution of (\ref{eq:inhom}) and assume that the normalizations (A1)-(A7) from Section \ref{sec:conventions} hold. Then, if $p<\infty$, $\Gamma_w$ is a $C^{1,1-\frac{n+1}{p}}(B_{1/2}')$ graph and if $p=\infty$, it is a $C^{1,1-}(B_{1/2}')$ graph.
\end{prop}
\begin{proof}
We prove this result similarly as in the case of $C^{1,\gamma}$ metrics but instead of working with the original solution $w$, we work with the modified function $u$ from Section~\ref{sec:ext}.\\
We begin by splitting $w=u+\tilde{u}$ as in Section \ref{sec:ext}. Moreover, we recall that by Lemma~\ref{lem:lower1'} (and the discussion following it) $\Gamma_w = \Gamma_{u}$. The Legendre function $v$ with respect to $u$ (c.f. (\ref{eq:Leg_split})) satisfies the nonlinear equation $\tilde{F}(v,y)=g(y)$ (c.f. \eqref{eq:w1p}), which in the notation in Section~\ref{subsec:improvement}, can be written as
$$\tilde{F}(v,y)=\sum_{i,j=1}^{n+1}\tilde{a}^{ij}(y)G^{ij}(v)=g(y).$$
Furthermore, $g$ satisfies the decay condition (\ref{eq:w1p_g}).
Keeping this in the back of our minds, we begin by proving analogues of Propositions \ref{prop:error_gain}, \ref{prop:error_gain2}.
To this end, we use a Taylor expansion to obtain that
$$\tilde{a}^{ij}(y)=\tilde{a}^{ij}(y_0)+ \hat E^{y_0,ij}(y), \quad y\in \mathcal{B}_{1/2}^+(y_0),$$
for each $y_0\in P\cap \mathcal{B}_{1/2}$.
Due to the $C^{0,1-\frac{n+1}{p}}$ Hölder regularity of $a^{ij}$, for $\epsilon \in (0,1-\frac{n+1}{p})$ the function $\hat E^{y_0,ij}(y)$ satisfies
\begin{equation}\label{eq:w1p_metric}
\begin{split}
\left\|d_G(\cdot,y_0)^{-2(1-\frac{n+1}{p})}\hat E^{y_0,ij}\right\|_{L^\infty(\mathcal{B}_{1/2}^+(y_0))}\\
+\left[d_G(\cdot,y_0)^{-2(1-\frac{n+1}{p}-\epsilon) }\hat E^{y_0,ij}\right]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_{1/2}^+(y_0))}\leq C.
\end{split}
\end{equation}
Recalling \eqref{eq:expand_v}, we expand the nonlinear function $G^{ij}(v)$ as
$$G^{ij}(v)=G^{ij}(v_{y_0})+\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}(v-v_{y_0})+\tilde{E}^{y_0,ij}_1(y),$$
where $v_{y_0}$ is the asymptotic profile of $v$ at $y_0$ and $\tilde{E}^{y_0,ij}_1(y)$ denote the same functions as in \eqref{eq:expand_v}.
Due to Proposition~\ref{prop:improved_reg1} the error term $\tilde{E}_1^{y_0,ij}$ satisfies the estimate (\ref{eq:Holderweight}). Hence, as in Step 1c of Proposition~\ref{prop:error_gain}, we can rewrite our nonlinear equation $\tilde{F}(v,y)=g(y)$ as
\begin{align*}
L_{y_0}v= L_{y_0}v_{y_0}+\tilde{f}
\end{align*}
with $L_{y_0}=\tilde{a}^{ij}(y_0)\p_{m_{k\ell}}G^{ij}(v_{y_0})\p_{k\ell}$ being the same leading term as in Remark~\ref{rmk:error_gain3} and
\begin{align*}
\tilde{f}(y)&=-\tilde{a}^{ij}(y_0)\tilde{E}^{y_0,ij}_1(y)-\hat E^{y_0,ij}(y)G^{ij}(v_{y_0})+g(y).
\end{align*}
Due to the error bounds for $\tilde{E}_1^{y_0,ij}$ and $\hat E^{y_0,ij}(y)$, the linear estimate for $G^{ij}(v)$ and the estimate \eqref{eq:w1p_g} for $g$, we infer that
\begin{align*}
|\tilde{f}(y)|\leq C d_G(y,y_0)^{\eta_0}, \quad \eta_0=\min\left\{1+4\alpha, 3-\frac{2(n+1)}{p}\right\}.
\end{align*}
Hence, as long as $1+4\alpha<3-\frac{2(n+1)}{p}$ we bootstrap regularity as in Proposition~\ref{prop:error_gain2}, in order to obtain an increasingly higher modulus of regularity for $v$ at $P$. In particular, by the compactness argument in the Appendix, c.f. Section \ref{sec:quarter_Hoelder}, this allows us to conclude that the Legendre function $v$ is in $X_{\delta,\epsilon}(\mathcal{B}_{1/2}^+)$ for all $\delta \in (0,1-\frac{n+1}{p}]$ if $p<\infty$ and in $X_{\delta,\epsilon}(\mathcal{B}_{1/2}^+)$ for all $\delta \in (0,1-\frac{n+1}{p})$ if $p= \infty$. This shows the desired regularity of $v$ and hence of $\Gamma_u$.
\end{proof}
\subsubsection{Regularity results in the presence of inhomogeneities and obstacles}
\label{sec:nonzero}
In this section we consider the regularity of the free boundary in the presence of non-vanishing inhomogeneities $f$. In particular, this includes the presence of obstacles (c.f. Remark \ref{rmk:obstacles_1}).
In this set-up we show the following results:
\begin{prop}[Inhomogeneities]
\label{prop:inhomo_2}
Let $w$ be a solution of the thin obstacle problem with metric $a^{ij}$ satisfying the assumptions (A1)-(A7) from Section \ref{sec:conventions}.
\begin{itemize}
\item[(i)] Assume further that $a^{ij}\in W^{1,p}(B_1^+, \R^{(n+1)\times(n+1)}_{sym})$ and $f\in L^p$ for some $p\in (2(n+1),\infty]$. Then $\Gamma_w$ is locally a $C^{1,\frac{1}{2}-\frac{n+1}{p}}$ graph.
\item[(ii)] Assume further that $a^{ij}\in C^{k,\gamma}(B_1^+, \R^{(n+1)\times(n+1)}_{sym})$ and that $f\in C^{k-1,\gamma}$ with $k\geq 1$, $\gamma \in (0,1)$. Then we have that $\Gamma_w$ is locally a $C^{k+[\frac{1}{2}+\gamma], (\frac{1}{2}+\gamma - [\frac{1}{2}+\gamma])}$ graph.
\end{itemize}
\end{prop}
We point out that compared with the result without inhomogeneities we lose half a derivative. This is due to the worse decay of the inhomogeneity in (\ref{eq:w1p_g}).\\
Similarly as for the zero obstacle case the proofs for Proposition \ref{prop:inhomo_2} rely on the Hodograph-Legendre transformation. In case (i) of Proposition~\ref{prop:inhomo_2}, we consider the Legendre transformation with respect to the modified solution $u$ after applying the splitting method. This is similar as in Section~\ref{subsec:w1p_zero}, where we dealt with $W^{1,p}$ metrics with zero right hand side. In case (ii) of Proposition~\ref{prop:inhomo_2}, we consider the Legendre transformation with respect to the original solution $w$. We remark that the presence of the inhomogeneity changes neither the leading order asymptotic expansion of $\nabla w$ around the free boundary, nor of the second derivatives $D^2w$ in the corresponding non-tangential cones (assuming $\|f\|_{L^\infty}$ is sufficiently small, which can always be achieved by scaling). In particular, in this case the Hodograph-Legendre transformation is well defined, and the asymptotic expansion of the Legendre function (c.f. Section~\ref{sec:Leg}) remains true.
\begin{proof}
We prove the result of Proposition \ref{prop:inhomo_2} in three steps. First we consider the set-up of (i). Then we divide the setting of (ii) into the cases $k=1$ and $k\geq 2$.
\begin{itemize}
\item In the case of $W^{1,p}$ metrics and $W^{2,p}$ obstacles, we proceed similarly as in Section~\ref{subsec:w1p_zero} by using the splitting method from above. The only changes occur when we estimate the inhomogeneity $g(y)$, where $g(y)$ is as in (\ref{eq:w1p}). Indeed, in the case of $f\neq 0$ we can in general only use the decay estimate \eqref{eq:w1p_g} for $g(y)$. This yields
\begin{align*}
|g(y)|\leq Cd_G(y,y_0)^{\eta_0},\quad \eta_0=\min\left\{1+4\alpha, 2-\frac{2(n+1)}{p}\right\}.
\end{align*}
Thus, we obtain that $v\in X_{\delta,\epsilon}(\mathcal{B}_{1/2}^+)$ for all $\delta \in (0,\frac{1}{2}-\frac{n+1}{p}]$. In particular, this entails that $\p_{in}v\in C^{0,\frac{1}{2}-\frac{n+1}{p}}(P\cap \mathcal{B}_{\frac{1}{2}})$. Hence, the regular free boundary $\Gamma_{\frac{3}{2}}(w)$ is locally a $C^{1,\frac{1}{2}-\frac{n+1}{p}}$ submanifold.
\item In the case of a $C^{1,\gamma}$ metric $a^{ij}$ and a $C^{0,\gamma}$ inhomogeneity $f$, we carry out an analogous expansion as in Proposition \ref{prop:error_gain} and estimate the right hand side of the equation by $d_G(y,y_0)^{2}$. Hence, an application of the bootstrap argument from Proposition \ref{prop:error_gain2} implies that $v\in X_{\delta,\epsilon}(\mathcal{B}_{1/2}^+)$ for all $\delta \in (0,\frac{1}{2}]$. Combining this with the application of the implicit function theorem as in Section \ref{sec:IFT1} hence yields that $\p_{in}v \in C^{[1/2+\gamma], (1/2+\gamma - [1/2+\gamma])}$. This implies the desired regularity.
\item
In the case of $C^{k,\gamma}$, $k\geq 2$ metrics we first apply the implicit function theorem (note that in our set-up the functional $\tilde{F}_a(w_a,y)=F_a(w_a+v,y)-g_a(y)$ is still $C^{k-1,\gamma-\epsilon}$ regular in the parameter $a$). In contrast to the argument in Section \ref{sec:IFT1} we can however now only apply the implicit function theorem in the spaces $X_{\delta,\epsilon}$ with $\delta \in (0,1/2]$. Thus, by the implicit function theorem argument (Step 1 in Theorem~\ref{prop:hoelder_reg_a} in Section \ref{sec:IFT1}) we infer that $\p_{in}v\in C^{k+[1/2+\gamma], (1/2+\gamma - [1/2+\gamma])}$.
\end{itemize}
This concludes the proof of Proposition \ref{prop:inhomo_2}.
\end{proof}
Finally, we comment on the relation of our regularity results with inhomogeneities and the presence of non-zero obstacles.
\begin{rmk}
\label{rmk:obstacles_1}
We note that the set-up of the present Section \ref{sec:W1p} (c.f. \eqref{eq:inhom}) in particular includes the set-up on non-zero obstacles: Indeed, let $a^{ij}:B_1^+ \rightarrow \R^{(n+1)\times (n+1)}_{sym}$ be a uniformly elliptic $W^{1,p}$, $p\in (2(n+1),\infty]$, metric satisfying (A1)-(A3) and let $\phi:B'_1 \rightarrow \R$ be a $W^{2,p}$ function. Suppose that $\tilde{w}$ is a solution to the thin obstacle problem with metric $a^{ij}$ and obstacle $\phi$. Then $w:=\tilde{w}-\phi$ is a solution of the thin obstacle problem
\begin{align*}
\p_{i} a^{ij} \p_j w & = f \mbox{ in } B_1^+,\\
\p_{n+1}w \leq 0, \ w \geq 0, \ w\p_{n+1}w&=0 \mbox{ in } B_1'.
\end{align*}
Hence, the inhomogeneity now reads $f=-\p_ia^{ij}\p_j\phi$ and is in $L^p$. In particular, Proposition \ref{prop:inhomo_2} is applicable and yields the $C^{1,\frac{1}{2}-\frac{n+1}{p}}$ regularity of the free boundary. Analogous reductions hold for more regular metrics and non-vanishing obstacles.
\end{rmk}
\section{Appendix}
\label{sec:append}
Last but not least, we provide proofs of the estimates which we used in the application of the implicit function theorem. This in particular concerns the spaces $X_{\delta,\epsilon}, Y_{\delta,\epsilon}$ and the mapping properties of $\D_G$ in these: After giving the proof of the characterization of the spaces $X_{\delta,\epsilon}$, $Y_{\delta,\epsilon}$ in terms of decompositions into Hölder functions (c.f. Proposition \ref{prop:decompI}) in Section \ref{sec:decomp}, we present the proof of the (local) $X_{\delta,\epsilon}$ estimates for solutions of the Baouendi-Grushin operator with mixed homogeneous Dirichlet-Neumann data (c.f. Proposition \ref{prop:invert}) in Section \ref{sec:quarter_Hoelder}. Here we argue by an iterative approximation argument, which exploits the scaling properties of the Baouendi-Grushin operator similarly as in \cite{Wa03}. Finally in Sections \ref{sec:XY} and \ref{sec:kernel}, we use this to show the necessary mapping properties of $\D_G$ in the spaces $X_{\delta,\epsilon},Y_{\delta,\epsilon}$.
\subsection{Proof of Proposition \ref{prop:decompI}}
\label{sec:decomp}
In this section we present the proof of the characterization of the spaces $X_{\alpha,\epsilon}$, $Y_{\alpha,\epsilon}$ in terms of decompositions into Hölder functions (c.f. Proposition \ref{prop:decompI} in Section \ref{sec:functions}).
\begin{proof}
We argue in two steps and first discuss the decomposition of functions in $Y_{\alpha,\epsilon}$ and then the corresponding property of functions in $X_{\alpha,\epsilon}$:\\
(i) Given $f\in Y_{\alpha,\epsilon}$, we denote
$$f_0(y''):=\p_nf(y''), \quad f_1(y):=r(y)^{-(1+2\alpha-\epsilon)}(f(y)-f_0(y'')y_n).$$
In particular, this yields
$f(y)=f_0(y'')y_n+r^{1+2\alpha-\epsilon}f_1(y)$. Moreover, we note that $f_1$ is well-defined on $P$, where it vanishes as a consequence of the boundedness of the $Y_{\alpha,\epsilon}$ norm and of Remark \ref{rmk:homo}. Hence, it suffices to prove the Hölder regularity of $f_0$ and $f_1$. \\
To show that $f_0 \in C^{0,\alpha}(P)$ (in the classical sense), we consider points $y_0,y_1\in P$, $y_0\neq y_1$ and a point $y=(y'',y_n,y_{n+1})\notin P$ with the property that $r(y)=|y_0-y_1|^{1/2}$ and $y_0,y_1\in \mathcal{B}_{2r(y)}(y)$. Then, by the boundedness of the norm and by recalling the estimates in Remark \ref{rmk:homo} (i), we have
\begin{align*}
|f(y)-f_0(y_0)y_n|&\leq C r^{1+2\alpha},\quad |f(y)-f_0(y_1)y_n|\leq C r^{1+2\alpha}.
\end{align*}
Thus, by the triangle inequality,
$$|f_0(y_0)y_n-f_0(y_1)y_n|\leq Cr^{1+2\alpha}.$$
Choosing $y$ with $y_{n+1}=0$, $|y_n|=r(y)>0$ and dividing by $|y_{n}|$ yields
\begin{align*}
|f_0(y_0)-f_0(y_1)|\leq Cr^{2\alpha}=|y_0-y_1|^\alpha.
\end{align*}
This shows the $C^{0,\alpha}$ regularity of $f_0$ (if $\alpha \in (0,1]$).\\
We proceed with the $C^{0,\epsilon}_\ast(Q_+)$ regularity of $f_1$. First we observe that since $|f(y)-f_0(y'')y_n|\leq C r(y)^{1+2\alpha}$ (which follows from Remark \ref{rmk:homo} (i)), we immediately infer that $|f_1(y)|\leq Cr(y)^{\epsilon}$. Thus, if $y_1, y_2 \in Q_+$ are such that
$d_G(y_1,y_2)\geq \frac{1}{10}\max\{r(y_1), r(y_2) \}$, we have
\begin{align*}
|f_1(y_1)-f_1(y_2)| \leq C r(y_1)^{\epsilon}+Cr(y_2)^\epsilon\leq C d_G(y_1,y_2)^{\epsilon}.
\end{align*}
If $y_1,y_2\in Q_+$ are such that $d_G(y_1,y_2)<\frac{1}{10}\max\{r(y_1),r(y_2)\}$, then there is a point $\bar y\in P$ such that $y_1,y_2\in \mathcal{B}_1^+(\bar y)$ (for example assuming $r(y_1)\geq r(y_2)$ we can let $\bar y=(y''_1,0,0)$). Then the H\"older regularity follows from the $C^{0,\epsilon}_\ast(\mathcal{B}_1^+(\bar y))$ regularity of $d(\cdot, \bar y)^{-(1+2\alpha-\epsilon)}(f-P_{\bar y})$ and the $C^{0,\alpha}(P)$ regularity of $f_0$.
More precisely,
\begin{equation}
\label{eq:f1}
\begin{split}
&\quad \left|f_1(y_1)-f_1(y_2)\right|\\
&=\left|r(y_1)^{-1-2\alpha+\epsilon}\left(f(y_1)-f_0(y''_1)(y_1)_n\right)-r(y_2)^{-1-2\alpha+\epsilon}\left(f(y_2)-f_0(y_2'')(y_2)_n\right)\right|\\
&\leq r(y_1)^{-1-2\alpha+\epsilon}\left|\left(f(y_1)-f_0(y''_1)(y_1)_n\right)-\left(f(y_2)-f_0(y''_1)(y_2)_n\right)\right|\\
& \quad + r(y_1)^{-1-2\alpha+\epsilon}\left|f_0(y''_1)(y_2)_n-f_0(y_2'')(y_2)_n\right|\\
&\quad + |r(y_1)^{-1-2\alpha+\epsilon}-r(y_2)^{-1-2\alpha+\epsilon}||f(y_2)-f_0(y_2'')(y_2)_n|.
\end{split}
\end{equation}
By the definition of the norm of $Y_{\alpha,\epsilon}$, we have
\begin{align*}
&\left|\left(f(y_1)-f_0(y''_1)(y_1)_n\right)-\left(f(y_2)-f_0(y''_1)(y_2)_n\right)\right|\\
&=\left|(f(y_1)-P_{y''_1}(y_1))-(f(y_2)-P_{y''_1}(y_2))\right|\lesssim r(y_1)^{1+2\alpha-\epsilon}d_G(y_1,y_2)^{\epsilon}.
\end{align*}
Moreover, the $C^{0,\alpha}$ regularity of $f_0$ as well as $|(y_2)_n|\sim r$ yields
\begin{align*}
r(y_1)^{-1-2\alpha+\epsilon}\left|f_0(y''_1)(y_2)_n-f_0(y_2'')(y_2)_n\right|&\lesssim r(y_1)^{-1-2\alpha+\epsilon}r(y_1)d_G(y_1,y_2)^{2\alpha}\\
& \lesssim d_G(y_1,y_2)^{\epsilon}.
\end{align*}
Here we have used that $2\alpha \geq \epsilon$ and that w.l.o.g. $0\leq r(y_2)\leq r(y_1)$.
Finally, the last term in (\ref{eq:f1}) is estimated by the $C^{0,\epsilon}_\ast$ regularity of $r(y_1)^{\epsilon}$ and by recalling the definition of the norm on $Y_{\alpha,\epsilon}$ in combination with Remark \ref{rmk:homo} once more. Combining all the previous observations, we have
\begin{align*}
|f_1(y_1)-f_1(y_2)|\lesssim d_G(y_1,y_2)^{\epsilon}.
\end{align*}
This completes the proof of (i).\\
(ii) The proof for the decomposition of $v$ is similar. Given any $y\in Q_+\setminus P$, we denote by $y_0:=(y'',0,0)\in P$ the projection of $y$ onto $P$. Since $v$ is $C^{3,\alpha}_\ast$ at $y_0$, there exists a Taylor polynomial
\begin{align*}
P_{y_0}(z)=\p_n v(y_0)z_n+\p_{in}v(y_0)(z_i-y_i)z_n + \frac{1}{6}\p_{nnn}v(y_0)z_n^3+\frac{1}{2}\p_{n,n+1,n+1}(y_0)z_nz_{n+1}^2,
\end{align*}
such that $|v(z)-P_{y_0}(z)|\leq Cd_G(z,y_0)^{3+2\alpha}$ for each $z\in \mathcal{B}_{1}^+(y_0)$. Due to the regularity of $v$ at $P$, the coefficients have the desired regularity properties: $\p_{n}v(y'')\in C^{1,\alpha}(P\cap \mathcal{B}_{1/2})$, $\p_{in}v(y''), \p_{nnn}v(y''), \p_{n,n+1,n+1}v(y'')\in C^{0,\alpha}(P\cap \mathcal{B}_{1/2})$. Moreover, their Hölder semi-norms are bounded from above by $C\|v\|_{X_{\alpha,\epsilon}}$. \\
In order to show the $C^{0,\epsilon}_\ast$ estimates of $C_1$, $V_i$ and $C_{ij}$, we argue similarly as in (i) for $f\in Y_{\alpha,\epsilon}$. For simplicity we only present the argument for
$$V_n(y):=r^{-(2+2\alpha-\epsilon)}\p_n(v-P_{y''})(y), \quad y\in \mathcal{B}_1^+\setminus P.$$
The others are analogous.
First, the boundedness of the first two terms in the norm $\|v\|_{X_{\alpha,\epsilon}}$ (c.f. Definition~\ref{defi:spaces}) and an interpolation estimate imply that for each $y\in \mathcal{B}_1^+\setminus P$ fixed, $d_G(z,y'')^{-(2+2\alpha-\epsilon)}\p_n(v-P_{y''})(z)$ as a function of $z$ is in $C^{0,\epsilon}_\ast(\mathcal{B}_{r(y)/2}(y))$, with norm bounded by $C\|v\|_{X_{\alpha,\epsilon}}$.
Next, for any points $z_1$ and $z_2$ in the non-tangential ball $\mathcal{B}_{r(y)/2}(y)$ we have
\begin{align*}
&\quad |V_n (z_1)-V_n(z_2)|\\
&=\left|d(z_1,z''_1)^{-(2+2\alpha-\epsilon)}\p_{n}(v-P_{z''_1})(z_1)-d(z_2,z''_2)^{-(2+2\alpha-\epsilon)}\p_{n}(v-P_{z''_2})(z_2)\right|\\
&\leq \left|d(z_1,z''_1)^{-(2+2\alpha-\epsilon)}\p_{n}(v-P_{z''_1})(z_1)-d(z_2,z''_1)^{-(2+2\alpha-\epsilon)}\p_{n}(v-P_{z''_1})(z_2)\right|\\
& \quad + \left|\left(d(z_2,z''_1)^{-(2+2\alpha-\epsilon)}-d(z_2,z''_2)^{-(2+2\alpha-\epsilon)}\right)\p_{n}(v-P_{z''_1})(z_2)\right|\\
& \quad +\left|d(z_2,z''_2)^{-(2+2\alpha-\epsilon)}\p_{n}(P_{z''_2}-P_{z''_1})(z_2)\right|:=I+II+III.
\end{align*}
By the definition of the $X_{\alpha,\epsilon}$-norm and by interpolation, $I\leq C\|v\|_{X_{\alpha,\epsilon}}d_G(z_1, z_2)^{\epsilon}$. Using the fact that $|\p_{n}(v-P_{z''_1})(z_2)|\leq C\|v\|_{X_{\alpha,\epsilon}} d_G(z_2,z''_1)^{2+2\alpha}$ and that $d_G(z''_2,z''_1)\leq Cd_G(z_2,z_1)\leq C\min \{d_G(z_2,z''_2),d_G(z_1,z''_1)\}$ for $z_1,z_2$ in the non-tangential ball $\mathcal{B}_{r(y)/2}(y)$, we also have that $II\leq C\|v\|_{X_{\alpha,\epsilon}}d_G(z_2,z_1)^{\epsilon}$. To estimate $III$ we notice that
\begin{align*}
\p_n P_{z''_1}(z)&=\p_nv(z''_1)+\sum_{i=1}^{n-1}\p_{in}v(z''_1)(z_i-(z_1)_i)\\
& \quad +\frac{1}{2}\p_{nnn}v(z''_1)z_n^2+\frac{1}{2}\p_{n,n+1,n+1}v(z''_1)z_{n+1}^2.
\end{align*}
Recalling that $\p_{n}v(y'')\in C^{1,\alpha}(P)$ and using a Taylor expansion of $\p_nv(y'')$ at $z''_1$, we infer that
\begin{align*}
\left|\p_nv(z''_2)-\left(\p_nv(z''_1)+\sum_{i=1}^{n-1}\p_{in}v(z''_1)((z_2)_i-(z_1)_i)\right)\right|\leq C\|v\|_{X_{\alpha,\epsilon}}|z''_2-z''_1|^{1+\alpha}.
\end{align*}
Thus, recalling the $C^{0,\alpha}$ regularity of $\p_{nnn}v(z'')$ and $\p_{n,n+1,n+1}v(z'')$, we obtain
\begin{align*}
\p_n( P_{z''_1}-P_{z''_2}) (z_2)&= \left(\p_n v(z''_1)+\sum_{i=1}^{n-1}\p_{in}v(z''_1)( (z_2)_i-(z_1)_i) - \p_n v(z''_2)\right)\\
&+\frac{1}{2}\left(\p_{nnn}v(z''_1)-\p_{nnn}v(z''_2)\right)(z_2)_n^2 \\
&+\frac{1}{2}\left(\p_{n,n+1,n+1}v(z''_1)-\p_{n,n+1,n+1}v(z''_2)\right)(z_2)_{n+1}^2\\
&\leq C\|v\|_{X_{\alpha,\epsilon}}d_G(z''_2,z''_1)^{2(1+\alpha)}+ C\|v\|_{X_{\alpha,\epsilon}}d_G(z''_1,z''_2)^{2\alpha}d_G(z_2,z''_2)^2.
\end{align*}
Due to the same reason as for $f$, this implies the estimate for $III$.
\end{proof}
\subsection{Proof of Proposition~\ref{prop:invert}}
\label{sec:quarter_Hoelder}
In this section, we present the proof of the (local) $X_{\delta,\epsilon}$ estimates for the Baouendi-Grushin operator (c.f. Proposition \ref{prop:invert} in Section \ref{sec:functions}).
We begin by recalling the natural energy spaces associated with the Baouendi-Grushin operator:
\begin{defi}
\label{defi:GrushinLp}
Let $\Omega\subset \R^{n+1}$ be an open subset.
The Baouendi-Grushin operator is naturally associated with the following Sobolev spaces (recall Definition~\ref{defi:Hoelder1} for the vector fields $\tilde{Y}_j$):
\begin{align*}
M^{1}(\Omega)&:= \{u\in L^2(\Omega)| \tilde{Y}_ju \in L^2(\Omega) \mbox{ for } j\in \{1,\dots,2n\}\},\\
M^{2}(\Omega)&:= \{u\in L^2(\Omega)|\tilde{Y}_ju, \tilde{Y}_{k}\tilde{Y}_\ell u\in L^2(\Omega)\text{ for } j,k,\ell\in \{1,\ldots, 2n\}\}.
\end{align*}
\end{defi}
We prove Proposition~\ref{prop:invert} in two steps. Firstly, we obtain a polynomial approximation (in the spirit of Campanato spaces) near the points at which the ellipticity of the operator degenerates, $P:=\{y_n=y_{n+1}=0\}$ (c.f. Proposition \ref{prop:Hoelder0}). Then we interpolate these estimates with the uniformly elliptic estimates which hold away from the degenerate points. Here we follow a compactness argument which was first outlined in this form by Wang, \cite{Wa03}. It proceeds via approximation and iteration steps.\\
In the sequel, we deduce a first regularity estimate in the energy space. This serves as a compactness result for the following approximation lemmata:
\begin{prop}
\label{prop:Sobolevreg}
Let $0<r\leq R<\infty$. Let $f:\mathcal{B}_{R}^+(0) \rightarrow \R$ be an $L^{2}$ function and let $u:\mathcal{B}_{R}^+\rightarrow \R$ be a solution of
\begin{equation}
\label{eq:Grushin}
\begin{split}
\Delta_G u& =f \text{ in } \mathcal{B}_R^+,\\
u&=0 \text{ on } \mathcal{B}_R^+\cap \{y_n=0\},\\
\p_{n+1}u&=0 \text{ on } \mathcal{B}_R^+\cap\{y_{n+1}=0\}.
\end{split}
\end{equation}
Then
\begin{align}
\label{eq:Lpapriori}
\left\| u \right\|_{M^{2}(\mathcal{B}_r^+)} \leq C(n,r,R)\left( \| f\|_{L^{2}(\mathcal{B}_{R}^+)} + \| u\|_{L^{\infty}(\mathcal{B}_{R}^+)} \right).
\end{align}
\end{prop}
\begin{proof}
The result is obtained by an even and odd reflection from the whole space result (in particular by the kernel estimate, see e.g. Lemma \ref{lem:ker} in Section \ref{sec:kernel}).
\end{proof}
With Proposition \ref{prop:Sobolevreg} at hand, we prove our first approximation result: We approximate solutions of the \emph{inhomogeneous} Baouendi-Grushin equation by solutions of the \emph{homogeneous} equation, provided the inhomogeneity is sufficiently small.
\begin{lem}
\label{lem:compactness}
Assume that $u:\mathcal{B}_{1}^+ \rightarrow \R$ is a solution of (\ref{eq:Grushin}) which satisfies
\begin{align*}
\frac{1}{|\mathcal{B}_{1}^+(0)|}\int\limits_{\mathcal{B}_1^+(0)} u^2 dx \leq 1.
\end{align*}
For any $\epsilon>0$ there exists a constant $\delta=\delta(\epsilon)>0$ such that if
\begin{align*}
\frac{1}{|\mathcal{B}_{1}^+(0)|} \int\limits_{\mathcal{B}_1^{+}(0)}f^2 dx \leq \delta^2,
\end{align*}
then there is a solution $h$ of the homogeneous Baouendi-Grushin equation with mixed Dirichlet-Neumann data, i.e.
\begin{equation}
\label{eq:hGrushin}
\begin{split}
\D_G h & = 0 \mbox{ on } \mathcal{B}_R^+(0),\\
h&=0 \mbox{ on } \{y_{n}=0\}\cap \mathcal{B}_R^+(0),\\
\p_{n+1} h&=0 \mbox{ on } \{y_{n+1}=0\}\cap \mathcal{B}_R^+(0),
\end{split}
\end{equation}
such that
\begin{align*}
\frac{1}{|\mathcal{B}_{1/2}^+(0)|} \int\limits_{\mathcal{B}_{1/2}^+(0)}|u-h|^2 dx \leq \epsilon^2.
\end{align*}
\end{lem}
\begin{proof}
We argue by contradiction and compactness. Assume that the statement were wrong. Then there existed $\epsilon>0$ and sequences, $\{u_m\}_{m}$, $\{f_m\}_m$, such that on the one hand
\begin{equation}
\label{eq:contra}
\begin{split}
\D_G u_m &= f_m \mbox{ in } \mathcal{B}_1^+(0),\\
u_m & = 0 \mbox{ on } \{y_n = 0\}\cap \mathcal{B}_{1}^+(0),\\
\p_{n+1} u_m &= 0 \mbox{ on } \{y_{n+1}=0\} \cap \mathcal{B}_{1}^+(0).
\end{split}
\end{equation}
and
\begin{align*}
&\frac{1}{|\mathcal{B}_{1}^+(0)|}\int\limits_{\mathcal{B}_1^+(0)} u_m^2 dx \leq 1, \ \frac{1}{|\mathcal{B}_{1}^+(0)|} \int\limits_{\mathcal{B}_1^{+}(0)}f_m^2 dx \leq \frac{1}{m}.
\end{align*}
On the other hand
\begin{align*}
\frac{1}{|\mathcal{B}_{1}^+(0)|}\int\limits_{\mathcal{B}_1^+(0)} |u_m - h|^2 dx \geq \epsilon^2,
\end{align*}
for all $h$ which satisfy the homogeneous equation (\ref{eq:hGrushin}). By (\ref{eq:Lpapriori}), we however have compactness for $u_m$ in $M^{1}$:
\begin{align*}
u_m \rightarrow u_0 \mbox{ in } M^{1}(\mathcal{B}_{3/4}^+(0)).
\end{align*}
Testing the weak form of (\ref{eq:contra}) with a $C_{0}^{\infty}(\mathcal{B}_{1/2}^+(0))$ function, we can pass to the limit and infer that $u_0$ is a weak solution of the homogeneous bulk equation from (\ref{eq:hGrushin}).
Finally, by the boundedness of $u_m\in M^{2}$ and the corresponding trace inequalities or a reflection argument, we obtain that $u_0$ satisfies the mixed Dirichlet-Neumann conditions from (\ref{eq:hGrushin}). This yields the desired contradiction.
\end{proof}
We now prove a further approximation result for solutions of the homogeneous Baouendi-Grushin equation in the quarter space. More precisely, we now seek to approximate solutions of the homogeneous equation (\ref{eq:contra}) by associated (eigen-) polynomials.
To this end, we recall the notion of \emph{homogeneous polynomials} in Section~\ref{sec:holder}.
\begin{rmk}
We note that all homogeneous polynomial solutions (e.g. the ones up to degree five) of $\Delta_Gv=0$ which satisfy the Dirichlet-Neumann boundary conditions can be computed explicitly. For instance, the degree less than five polynomial solutions are given by the linear combination of
\begin{align*}
y_n, \ y_jy_n, \ j\in\{1,\dots,n-1\},\ y_{n}^3 - 3y_n y_{n+1}^2.
\end{align*}
\end{rmk}
Using the notion of homogeneous polynomials, we proceed to our second approximation lemma:
\begin{lem}
\label{lem:approx}
Let $u:\mathcal{B}_{1}^+(0) \rightarrow \R$ be a solution of (\ref{eq:hGrushin}) with $\| u \|_{L^2(\mathcal{B}_{1}^+(0))}\leq \bar{c}$. Then there exists a polynomial $p$ of homogeneous degree less than or equal to three which solves (\ref{eq:hGrushin}), i.e.
\begin{align*}
p(y)=y_n \left(a_0+\sum_{i=1}^{n-1}a_iy_i+b(y_n^2-3y_{n+1}^2)\right),
\end{align*}
such that for all $0<r\leq \frac{1}{2}$
\begin{align}
\label{eq:approx}
\frac{1}{|\mathcal{B}_{r}^+(0)|}\int\limits_{\mathcal{B}_r^+(0)}|u-p|^2 dy \leq C(\bar{c}) r^{10},
\end{align}
and
\begin{align*}
\sum_{i=0}^{n-1}|a_i|+|b|\leq C\bar c,
\end{align*}
where $C$ is a universal constant.
\end{lem}
\begin{proof}
After a conformal change of variables, the Baouendi-Grushin operator can be rewritten as
\begin{align*}
\D_G = (\dt^2 + \D_{\Sigma}),
\end{align*}
where $\Sigma:= \{(y'',y_n,y_{n+1})| \ |y''|^4 + y_n^2 + y_{n+1}^2 = 1\}$ denotes the Baouendi-Grushin sphere. In our setting this is augmented with the Dirichlet and Neumann conditions from (\ref{eq:hGrushin}). The eigenfunctions of $\D_{\Sigma}$ can be extended in the radial direction to yield homogeneous solutions of the homogeneous Baouendi-Grushin equation. As the Baouendi-Grushin operator is hypoelliptic, these solutions are polynomials (this remains true in the cone $\Sigma \cap \mathcal{B}_1^+$, as the eigenfunctions on the Baouendi-Grushin quarter sphere can be identified as a subset of the eigenfunctions on the whole sphere by appropriate (even and odd) reflections). Moreover, the eigenfunctions on $\Sigma$ are orthogonal and as a consequence, the same is true for the correspondingly associated polynomials. The Baouendi-Grushin polynomials hence form an orthogonal basis into which a solution of the homogeneous Baouendi-Grushin problem can be decomposed.
We denote these polynomials by $p_k(y)$ and normalize them with respect to $\mathcal{B}^{+}_{1}(0)$. Since $u$ is bounded in $L^2(\mathcal{B}_{1}^+(0))$, we have
\begin{align}
\label{eq:poly_decomp}
u(y) = \sum\limits_{k=0}^{\infty} \alpha_k p_k(y) \mbox{ with } \sum\limits_{k=0}^{\infty}|\alpha_k|^2 \leq \bar{c}^2.
\end{align}
The previous decomposition can also be seen ``by hand'': Making the ansatz that a homogeneous solution of the Baouendi-Grushin problem is of the form
\begin{align*}
u(t,\theta) = \sum\limits_{k\in \Z} \alpha_k(t) u_{k}(\theta),
\end{align*}
where $u_k(\theta)$ denotes the spherical eigenfunctions, we obtain that
\begin{align*}
0& =(u_k, \D_G u)_{L^{2}(\Sigma)} = \alpha_k''(t) + (u_k, \D_{\Sigma} u)_{L^2(\Sigma)} \\
& = \alpha_k''(t) -\lambda_k^2 \alpha_k(t)
+ \int\limits_{\partial \Sigma} u_k (\nu\cdot \nabla_{\Sigma} u) d\mathcal{H}^{n-1} - \int\limits_{\partial \Sigma} u (\nu\cdot \nabla_{\Sigma} u_k) d\mathcal{H}^{n-1}.
\end{align*}
Here $\lambda_k^2$ is the eigenvalue associated with $u_k(\theta)$ and $\nu:\partial \Sigma \rightarrow \R^{n-1}$ is the outer unit normal field.
As both $u$ and $u_k$ satisfy the mixed Dirichlet-Neumann boundary conditions, this yields that
\begin{align*}
\alpha''_k(t) - \lambda_k^2 \alpha_k(t)=0.
\end{align*}
As the Dirichlet data imply that $\alpha_k(-\infty)=0$, this results in $\alpha_k(t) = \alpha_k(0) e^{|\lambda_k| t} $. By hypoellipticity, $\lambda_k \in \Z$, so that we obtain a decomposition into polynomials, after undoing the conformal change of coordinates.
After an appropriate normalization, we again infer (\ref{eq:poly_decomp}).\\
We define $p(y):= \sum\limits_{k=0}^{3} \alpha_k p_k(y)$.
Thus, recalling that there are no eigenpolynomials of (homogeneous) degree four which satisfy our mixed Dirichlet-Neumann conditions and computing the difference $u-p$, we arrive at
\begin{align*}
\left\| u- p \right\|_{L^2(\mathcal{B}_r^+)}^2 = \sum\limits_{k=5}^{\infty}|a_k|^2 \| p_k\|_{L^2(\mathcal{B}_r^+)}^2
&\leq \sum\limits_{k=5}^{\infty}|a_k|^2 r^{10}|\mathcal{B}_r^+|\| p_k\|_{L^2(\mathcal{B}_1^+)}^2\\
&\leq r^{10+ 2n}C(\bar{c}),
\end{align*}
where we used the scaling of the Baouendi-Grushin cylinders from Definition \ref{defi:Grushincylinder} and the boundedness of $u$ (c.f. (\ref{eq:poly_decomp})).
This yields the desired result.
\end{proof}
\begin{rmk}
\label{rmk:approx}
We stress that the approximation from Lemma \ref{lem:approx} is not restricted to third order polynomials. It can be extended to polynomials of arbitrary (homogeneous) degree.
\end{rmk}
Combining the previous results, we obtain the key building block for the iteration which yields regularity at the hyperplane $\{y_n=y_{n+1}=0\}$ at which the ellipticity of the Baouendi-Grushin operator degenerates.
\begin{lem}[Iteration]
\label{prop:iteration}
Let $\alpha\in(0,1)$.
Assume that $u:\mathcal{B}_{1}^+ \rightarrow \R$ is a solution of (\ref{eq:Grushin}) which satisfies
\begin{align*}
\frac{1}{|\mathcal{B}_{1}^+(0)|}\int\limits_{\mathcal{B}_1^+(0)} u^2 dx \leq 1.
\end{align*}
There exist a radius $r_0\in (0,1)$, a universal constant $C>0$ and a constant $\epsilon>0$ such that if
\begin{align*}
\frac{1}{|\mathcal{B}_{1}^+(0)|} \int\limits_{\mathcal{B}_1^{+}(0)}f^2 dx \leq \epsilon^2,
\end{align*}
then there exists a polynomial $p$ of order less than or equal to three
satisfying (\ref{eq:hGrushin}), i.e.
\begin{align*}
p(y)=y_n \left(a_0+\sum_{i=1}^{n-1}a_iy_i+b(y_n^2-3y_{n+1}^2)\right),
\end{align*}
such that
\begin{align*}
\frac{1}{|\mathcal{B}_{r_0}^+(0)|}\int\limits_{\mathcal{B}_{r_0}^+(0)} |u-p|^2 dy \leq r^{2(3+2\alpha)}_0,
\end{align*}
and
\begin{align*}
\sum_{i=0}^{n-1}|a_i|+|b|\leq C.
\end{align*}
\end{lem}
\begin{proof}
By our first approximation result, Lemma \ref{lem:compactness}, there exists a function $h$ which solves (\ref{eq:hGrushin}) and satisfies
\begin{align}
\label{eq:approx1}
\int\limits_{\mathcal{B}_{1/2}^+(0)}|u-h|^2 dy \leq \delta^2.
\end{align}
In particular, $\| h \|_{L^2(\mathcal{B}_{1/2}^+(0))}\leq C$. Hence, by our second approximation result, Lemma \ref{lem:approx}, there exists a (homogeneous) third order Baouendi-Grushin polynomial satisfying the Dirichlet-Neumann condition such that
\begin{align*}
\frac{1}{|\mathcal{B}_{r}^+(0)|} \int\limits_{\mathcal{B}_{r}^+(0)} |h-p|^2 dy \leq C r^{10}, \quad \mbox{ for all } 0<r<1/2.
\end{align*}
Consequently, by rescaling, we obtain for each $0<r<1/2$,
\begin{align*}
\frac{1}{|\mathcal{B}_r^{+}(0)|} \int\limits_{\mathcal{B}_r^+(0)}|u-p|^2 dy & \leq 2 \frac{1}{|\mathcal{B}_r^{+}(0)|} \int\limits_{\mathcal{B}_r^+(0)}|u-h|^2 dy
+ 2 \frac{1}{|\mathcal{B}_r^{+}(0)|} \int\limits_{\mathcal{B}_r^+(0)}|h-p|^2 dy\\
& \leq 2 r^{-2n} \int\limits_{\mathcal{B}_{1/2}^+(0)}|u-h|^2 dy + 2C r^{10}\\
& \leq 2 r^{-2n}\delta^2 + 2C r^{10}, \
\end{align*}
where we used (\ref{eq:approx1}) to estimate the first term. First choosing $0<r_0<1$ universal, but so small such that $2C r_0^{10}\leq \frac{1}{2}r_0^{2(3+2\alpha)}$, and then choosing $\delta>0$ universal such that $2 r_0^{-2n}\delta^2 \leq\frac{1}{2}r_0^{2(3+2\alpha)}$, yields the desired result.
\end{proof}
As a corollary of Lemma \ref{prop:iteration}, we can iterate in increasingly finer radii.
\begin{cor}
\label{cor:iteration}
Let $\alpha\in(0,1)$.
Assume that $u:\mathcal{B}_{1}^+ \rightarrow \R$ is a solution of (\ref{eq:Grushin}) which satisfies
\begin{align*}
\frac{1}{|\mathcal{B}_{1}^+(0)|}\int\limits_{\mathcal{B}_1^+(0)} u^2 dx \leq 1,
\end{align*}
and that for each $k\in \mathbb{N}_+$
\begin{align*}
\frac{1}{|\mathcal{B}_{r_0^{k-1}}^+(0)|} \int\limits_{\mathcal{B}_{r_0^{k-1}}^{+}(0)}f^2 dx \leq \epsilon^2 r_0^{2(k-1)(1+2\alpha)}.
\end{align*}
Then there exists a polynomial $p_k$ of (homogeneous) degree (less than or equal to) three solving (\ref{eq:hGrushin}) such that
\begin{align*}
\frac{1}{|\mathcal{B}_{r^{k}_0}^{+}(0)|} \int\limits_{\mathcal{B}_{r^{k}_0}^+(0)}|u-p_k|^2 dx \leq r^{2k (3+ 2\alpha)}_0.
\end{align*}
Moreover, it is of the form
\begin{align*}
p_k(y) = y_n\left(a^0_k + \sum\limits_{j=1}^{n-1}a_k^j y_j\right) + b_k (y_n^3 - 3y_n y_{n+1}^2),
\end{align*}
and we have
\begin{equation}
\label{eq:coefficients}
\begin{split}
|a^0_k - a^0_{k-1}| &\leq C r_0^{k(2+{2\alpha})},\\
|a_k^j - a_{k-1}^j| &\leq C r_0^{{2k\alpha}},\quad j=1,\dots, n-1,\\
|b_k - b_{k-1}| & \leq C r_0^{{2k\alpha}}.
\end{split}
\end{equation}
\end{cor}
\begin{proof}
We argue by induction on $k$ and take $p_0 = 0$ and $p_1$ as the polynomial from Proposition \ref{prop:iteration}. We assume that the statement is true for $k$ and show it for $k+1$. For that purpose, we consider the rescaled and dilated functions
\begin{align*}
u_k(y):= \frac{(u-p_k)(r_0^{2k} y'', r_0^k y_n, r_0^k y_{n+1})}{r_0^{k(3+{2\alpha})}}.
\end{align*}
Hence,
\begin{align*}
\D_G u_k = \frac{r_0^{2 k}f_{r_0}}{r_0^{(3+{2\alpha})k}} = r_0^{-k (1+{2\alpha})} f_{r_0},
\end{align*}
where $f_{r_0}(y'',y_n,y_{n+1})= f(r_0^{2} y'', r_0 y_n, r_0 y_{n+1})$.
Using the smallness assumption on $f$, we obtain that
\begin{align*}
\frac{1}{|\mathcal{B}_{1}^{+}(0)|} \int\limits_{\mathcal{B}_{1}^+(0)}|r_0^{-k(1+{2\alpha})} f_{r_0}|^2 dx \leq r_0^{-2k (1+{2\alpha})} \frac{1}{|\mathcal{B}_{r_0^{k}}^{+}(0)|} \int\limits_{\mathcal{B}_{r_0^k}^+(0)}f^2 dx \leq \epsilon^2.
\end{align*}
Hence by Proposition \ref{prop:iteration}, we obtain a (homogeneous) polynomial, $q$, of degree less than or equal to three, which satisfies (\ref{eq:hGrushin}) and is of the form
\begin{align*}
q(y'',y_n, y_{n+1}) = y_n\left(a_0+ \sum_{j=1}^{n-1}a_j y_j\right) + b (y_n^3 - 3y_n y_{n+1}^2),
\end{align*}
and
$$\sum_{j=0}^{n-1}|a_j|+|b|\leq C,$$
such that
\begin{align*}
\frac{1}{|\mathcal{B}_{r_0}^{+}(0)|} \int\limits_{\mathcal{B}_{r_0}^+(0)}|u_k - q|^2 dx \leq r_0^{2(3+{2\alpha})}.
\end{align*}
Rescaling therefore gives us that the polynomial
\begin{align*}
p_{k+1}(y'',y_n,y_{n+1}) = p_k(y'',y_{n},y_{n+1}) + r_0^{(3+{2\alpha})k}q\left(\frac{y''} {r_0^{2k}}, \frac{y_n}{r_0^k}, \frac{y_{n+1}}{r_0^k} \right),
\end{align*}
satisfies the claim of the corollary.
\end{proof}
We summarize the previous compactness and iteration arguments in the following intermediate result:
\begin{prop}
\label{prop:Hoelder0}
Let $\alpha\in(0,1)$.
Assume that $u:\mathcal{B}_{1}^+ \rightarrow \R$ is a solution of (\ref{eq:Grushin}). Suppose that the inhomogeneity $f:\mathcal{B}_{1}^+(0)\rightarrow \R$ is $C^{1,\alpha}_{\ast}$ at $y=0$ in the sense of Definition \ref{defi:diff}, i.e.
\begin{align*}
|f - f(0) - \p_n f(0) y_n | \leq F_0 r^{1+{2\alpha}}
\end{align*}
for any $0<r<1$.
Then there is a polynomial $p$ with \\
$\|p\|_{C^{3}(\mathcal{B}_{1}^+)}\leq C \left(\| u\|_{L^{2}(\mathcal{B}_1^+)} + |f(0)|+|\p_nf(0)|\right)$ of (homogeneous) degree less than or equal to three such that
\begin{align*}
\frac{1}{|\mathcal{B}_{r}^{+}(0)|} \int\limits_{\mathcal{B}_{r}^+(0)}|u-p|^2 dx \leq C(\| u\|_{L^2(\mathcal{B}_1^+(0))}^2 + F_0^2)r^{2(3+{2\alpha})}.
\end{align*}
\end{prop}
\begin{proof}
Without loss of generality we may assume that $f(0)=\p_n f(0)=0$. Indeed, this follows by considering the function $v(y'',y_n,y_{n+1}):= u(y) - q(y)$, where $q(y)$ is a homogeneous polynomial of homogeneous degree less than or equal to three such that $\Delta_G q=f(0)+\p_{n}f(0)y_n$ and $q=0$ on $\{y_n=0\}$, $\p_{n+1}q=0$ on $\{y_{n+1}=0\}$ (for example, one can consider $q(y)=\frac{1}{2}f(0) y_n^2+ cy_ny_{n+1}^2 +dy_n^3$ with $2c+6d=\p_nf(0)$).
Considering $\tilde{v}:=\frac{\epsilon}{F_0}v$, then also gives the smallness assumptions of Corollary \ref{cor:iteration}. Thus, for each $k\in \N_+$ there exists a Baouendi-Grushin polynomial $p_k$ such that
\begin{align*}
\frac{1}{|\mathcal{B}_{r_0^k}^+(0)|} \int\limits_{\mathcal{B}_{r_0^k}^+(0)}|\tilde{v}-p_k|^2 dy \leq r_0^{2k(3+{2\alpha})}.
\end{align*}
Due to the estimates (\ref{eq:coefficients}) on the coefficients of $p_k$, which were derived in Corollary \ref{cor:iteration}, $p_k \rightarrow p_{\infty}$, where $p_{\infty}$ is a polynomial of (homogeneous) degree at most three and which satisfies
\begin{align*}
\frac{1}{|\mathcal{B}_{r_0^k}^+(0)|} \int\limits_{\mathcal{B}_{r_0^k}^+(0)}|p_{\infty}-p_k|^2 dy \leq C r_0^{2k(3+{2\alpha})}.
\end{align*}
Consequently, by the triangle inequality
\begin{align*}
\frac{1}{|\mathcal{B}_{r_0^k}^+(0)|} \int\limits_{\mathcal{B}_{r_0^k}^+(0)}|\tilde{v} -p_{\infty}|^2 dy \leq C r_0^{2k(3+{2\alpha})}
\end{align*}
for $k\in \N$. Rescaling then yields the desired result.
\end{proof}
\begin{rmk}
\begin{itemize}
\item The previous result yields the ``H\"older regularity at the point'' $y=0$. For other points $y_0=(y_0'',0,0)$ an analogous result holds by translation invariance of the equation and the boundary conditions in the $y''$ directions (c.f. (\ref{eq:Grushin})). In this translated case, the conditions on the inhomogeneity $f:\mathcal{B}_{1}^+(y_0)\rightarrow \R$ read
\begin{align*}
|f - f(y_0) - \p_{n}f(y_0) y_n| \leq F_0 r^{2(1+{2\alpha})}.
\end{align*}
\item Instead of imposing the $C^{1,\alpha}_{\ast}$ condition in the sense of Definition \ref{defi:diff}, it would have sufficed to assume the weaker condition
\begin{align*}
\frac{1}{|\mathcal{B}_{r}^{+}(y_0)|} \int\limits_{\mathcal{B}_{r}^+(y_0)}|f - f(y_0) - \p_{n}f(y_0) y_n|^2 dy \leq F_0 r^{2(1+{2\alpha})}.
\end{align*}
\item In order to argue as we have outlined above, we have to require the compatibility condition $\p_{n+1}f(y'',0,0)=0$ (c.f. Definition \ref{defi:spaces}). However, apart from the described $C^{1,\alpha}_{\ast}$ regularity, we do not have to pose further restrictions on $f$.
\end{itemize}
\end{rmk}
Building on the precise description of the regularity of solutions close to the hyperplane $\{y_n=y_{n+1}=0\}$, we can now derive the full regularity result of Proposition \ref{prop:invert} by additionally invoking the uniform ellipticity which holds at a sufficiently far distance from $P$. This then concludes the argument for Proposition \ref{prop:invert}.
\begin{proof}[Proof of Proposition~\ref{prop:invert}]
It suffices to prove the corresponding regularity result in $X_{\alpha,\epsilon}(\mathcal{B}_3^+)$ (c.f. Definition \ref{defi:spaces_loc}).
Indeed, the Hölder estimate,
\begin{align*}
\|v\|_{C^{2,\epsilon}_{\ast}(Q_+)} \lesssim \|\D_G v\|_{C^{0,\epsilon}_{\ast}(Q_+)},
\end{align*}
follows similarly. As a consequence of the support assumption on $\Delta_G v$, this then yields the bound
\begin{align*}
\|v\|_{C^{2,\epsilon}_{\ast}(Q_+)} \lesssim \|\D_G v\|_{Y_{\alpha,\epsilon}},
\end{align*}
which together with the local estimate in $X_{\alpha,\epsilon}(\mathcal{B}_3^+)$ provides the full bound from Proposition \ref{prop:invert}.\\
\emph{Step 1. Polynomial approximation at $P=\{y_n=y_{n+1}=0\}$.} We note that for $f\in Y_{\alpha,\epsilon}$ and $y_0\in P$ , there exists a first order polynomial $p_{y_0}(y)$ which is of the form $p_{y_0}(y) =f_0(y_0)y_n$ such that
\begin{align*}
\frac{1}{|\mathcal{B}_s^+(y_0)|}\int_{\mathcal{B}_s^+(y_0)}|f(y)-f_0(y_0)y_n|^2\leq C s^{2(1+2\alpha)}, \quad \forall s\in(0,1).
\end{align*}
By considering $v(y)-f_0(y_0)y_n^3/6$ and by still denoting the resulting function by $v$, we may assume that $f_0(y_0)=0$. The same arguments as before lead to the existence of a third order (in the homogeneous sense) polynomial $P_{y_0}$, where $$P_{y_0}(y)=a_0\left(y_n^3-3y_ny_{n+1}^2\right)+\sum_{i=1}^{n-1}b_iy_ny_i+c_0y_n,$$ for some constants $a_0,b_i,c_0$ depending on $y_0$, such that
\begin{align}
\label{eq:approx_a}
\frac{1}{|\mathcal{B}_s^+(y_0)|}\int_{\mathcal{B}_s^+(y_0)}|v-P_{y_0}|^2\leq C\left(\|v\|_{L^2(\mathcal{B}_1^+)}^2+\|f\|_{Y_{\alpha,\epsilon}}^2\right)s^{2(3+2\alpha)}
\end{align}
for any $0<s<1/2$. \\
\emph{Step 2. Interpolation.} For $y\notin P$ with $\sqrt{y_n^2+y_{n+1}^2}=\lambda>0$, let
\begin{align*}
\tilde{v}_\lambda(\xi):=\frac{(v-P_{y_0})(y_0+\lambda ^2 \xi'',\lambda \xi_n, \lambda \xi_{n+1})}{\lambda ^{3+2\alpha}},
\end{align*}
where $y_0$ is the projection of $y$ on $P$. Let $\xi_0$ be the image point of $y$ under this rescaling. By Step 1, $\tilde{v}_\lambda(\xi)\in L^2(\mathcal{B}_{1/2}(\xi_0))$ with
\begin{align}
\label{eq:rhs_est}
\|\tilde{v}_\lambda\|_{L^2(\mathcal{B}_1(\xi_0))}\leq C\left(\|v\|_{L^2}+\|f\|_{Y_{\alpha,\epsilon}}\right).
\end{align}
Moreover,
\begin{align*}
\Delta_G \tilde{v}_\lambda(\xi)= f_{\lambda}(\xi),
\end{align*}
where $f_{\lambda}(\xi):=\frac{1}{\lambda^{\epsilon}}f(y_0+\lambda^2\xi'',\lambda\xi_n,\lambda \xi_{n+1})$. We note that by the definition of $Y_{\alpha,\epsilon}$ and by $f_0(y''_0)=0$,
\begin{align*}
\| f_{\lambda}\|_{C^{0,\epsilon}(\mathcal{B}_{1/2}(\xi_0))} \leq \| f\|_{Y_{\alpha,\epsilon}}.
\end{align*}
In $\mathcal{B}_{1/2}(\xi_0)$, $\Delta_G$ is uniformly elliptic. Thus, by the classical $C^{2,\epsilon}$ Schauder estimates
\begin{align*}
\|\tilde{v}_\lambda\|_{C^{2,\epsilon}(\mathcal{B}_{1/4}(\xi_0))} \leq C\left( \|\tilde{v}_\lambda\|_{L^2(\mathcal{B}_{1/2}(\xi_0))}+\|f_\lambda\|_{C^{0,\epsilon}(\mathcal{B}_{1/2}(\xi_0))}\right).
\end{align*}
Rescaling back and letting $\tilde{v}:=v-P_{y_0}$, we in particular infer
\begin{align*}
\lambda^{-1-2\alpha+\epsilon}\sum\limits_{i,j=1}^{n+1}[Y_i Y_j \tilde{v}]_{C^{0,\epsilon}(\mathcal{B}_{\lambda/4}(y))}
\leq C\left(\|v\|_{L^2(\mathcal{B}_1^+)}+\|f\|_{Y_{\alpha,\epsilon}}\right).
\end{align*}
Here we used (\ref{eq:rhs_est}) to estimate the right hand side contribution.
Recalling the $L^{\infty}$ estimate $\| v\|_{L^{\infty}(Q_+)} \leq C \| \D_G v \|_{L^{\infty}}$ (c.f. the kernel bounds in Lemma \ref{lem:ker} in Section \ref{sec:kernel}) and the support conditions for $\D_G v$ and for $f$ allows us to further bound
\begin{align*}
\|v\|_{L^2(\mathcal{B}_1^+)}+\|f\|_{Y_{\alpha,\epsilon}} \leq C \| f\|_{Y_{\alpha,\epsilon}}.
\end{align*}
This implies
\begin{align}\label{eq:err_est}
\lambda^{-1-2\alpha+\epsilon}\sum\limits_{i,j=1}^{n+1}[Y_i Y_j \tilde{v}]_{C^{0,\epsilon}_\ast(\mathcal{B}_{\lambda/4}(y))}
+ \lambda^{-1-2\alpha}\sum\limits_{i,j=1}^{n+1}\|Y_i Y_j \tilde{v}\|_{L^{\infty}(\mathcal{B}_{\lambda/4}(y))}
\leq C\|f\|_{Y_{\alpha,\epsilon}}.
\end{align}
Passing through a chain of non-tangential balls, we infer that \eqref{eq:err_est} holds in a non-tangential cone at $y_0$:
\begin{align*}
&\sum\limits_{i,j=1}^{n+1}[d_G(y,y_0)^{-1-2\alpha+\epsilon}Y_i Y_j \tilde{v}]_{C^{0,\epsilon}_\ast(\mathcal{N}_G(y_0))} \\
&+ \sum\limits_{i,j=1}^{n+1}\|d_G(y,y_0)^{-1-2\alpha}Y_i Y_j \tilde{v}\|_{L^{\infty}(\mathcal{N}_G(y_0))}
\leq C\|f\|_{Y_{\alpha,\epsilon}}.
\end{align*}
We note that it is possible to derive \eqref{eq:err_est} for $v-P_{\bar y}$ at each $\bar y\in P$ and that hence $v$ is $C^{3,\alpha}_{\ast}(P)$ in the sense of Definition \ref{defi:diff}. As by Proposition~\ref{prop:decompI} the map $\bar y\mapsto P_{\bar y}$ is $C^{0,\alpha}(P)$ regular, a triangle inequality and a covering argument yield the estimate in the full neighborhood of $y_0$
\begin{align*}
&\sum\limits_{i,j=1}^{n+1}[d_G(y,y_0)^{-1-2\alpha+\epsilon}Y_i Y_j \tilde{v}]_{C^{0,\epsilon}_\ast(\mathcal{B}_1^+(y_0))} \\
&+ \sum\limits_{i,j=1}^{n+1}\|d_G(y,y_0)^{-1-2\alpha}Y_i Y_j \tilde{v}\|_{L^{\infty}(\mathcal{B}_1^+(y_0))}
\leq C\|f\|_{Y_{\alpha,\epsilon}}.
\end{align*}
This concludes the local estimate and hence concludes the proof of Proposition \ref{prop:invert}.
\end{proof}
\subsection{Invertibility of the Baouendi-Grushin Laplacian in $X_{\alpha,\epsilon} $, $Y_{\alpha,\epsilon}$}
\label{sec:XY}
We provide the proofs of the completeness of the spaces $X_{\alpha,\epsilon},Y_{\alpha,\epsilon}$ (c.f. Definition~\ref{defi:spaces}) and the desired invertibility of the Baouendi-Grushin Laplacian as an operator from $X_{\alpha,\epsilon}$ to $Y_{\alpha,\epsilon}$ (c.f. Lemma \ref{lem:inverse}).
\begin{lem}
\label{lem:Banach}
Let $X_{\alpha,\epsilon},Y_{\alpha,\epsilon}$ be as in Definition~\ref{defi:spaces}. Then $(X_{\alpha,\epsilon},\| \cdot \|_{X_{\alpha,\epsilon}}), (Y_{\alpha,\epsilon}, \| \cdot \|_{Y_{\alpha,\epsilon}})$ are Banach spaces.
\end{lem}
\begin{proof}
(i) We first note that by the definition of $Y_{\alpha,\epsilon}$, $\supp(f)\subset \mathcal{B}_3^+$ . Hence, it suffices to consider the behavior of functions on $\bar{\mathcal{B}_3^+}$. By Proposition~\ref{prop:decompI} and Remark~\ref{rmk:characterize} a function $f\in Y_{\alpha,\epsilon}$ can be decomposed as
\begin{align*}
f(y)=f_0(y'')y_n+ r(y)^{1+2\alpha-\epsilon}f_1(y),\quad r(y)=\sqrt{y_n^2+y_{n+1}^2},
\end{align*}
with $f_0\in C^{0,\alpha}(P)$ and $f_1\in C^{0,\epsilon}_\ast(Q_+)$ and $f_0$, $f_1$ are obtained by Taylor approximation of $f$ (c.f. the proof of Proposition \ref{prop:decompI} in Section \ref{sec:decomp}).
Moreover, $[f_0]_{\dot{C}^{0,\alpha}(P\cap \mathcal{B}_3)} +[f_1]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)}$ is equivalent to $\|f\|_{Y_{\alpha,\epsilon}}$. Thus, in order to obtain the desired Banach property, it suffices to show the equivalence of the homogeneous Hölder norms and their inhomogeneous counterparts for $y\in \mathcal{B}_3^+$. \\
We start by making the following observation: For any $f\in Y_{\alpha,\epsilon}$, $\supp(f)\subset \mathcal{B}_3^+$ (in combination with the definition of $f_0, f_1$) implies that
\begin{align*}
f_0(y'')=0 \mbox{ and } f_1(y)=0 \mbox{ for } y=(y'',y_n, y_{n+1}) \mbox{ such that } (y'',0,0) \in P\setminus \mathcal{B}_3^+.
\end{align*}
Thus,
\begin{align*}
\|f_0 \|_{L^\infty(P\cap \mathcal{B}_3)} \leq C [f_0]_{\dot{C}^{0,\alpha}(P \cap \mathcal{B}_3)}, \quad \|f_1\|_{L^\infty(\mathcal{B}_{3}^+)}\leq C[f_1]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)}.
\end{align*}
In particular, this immediately entails that $\|f_0\|_{C^{0,\alpha}(P\cap \mathcal{B}_3)}\leq C[f_0]_{\dot{C}^{0,\alpha}(P\cap \mathcal{B})}$ and $\|f_1\|_{C^{0,\epsilon}_\ast(\mathcal{B}_3^+)}\leq C[f_1]_{\dot{C}^{0,\epsilon}_\ast(\mathcal{B}_3^+)}$. Therefore, $Y_{\alpha,\epsilon}$ is a Banach space. \\
(ii) Let $v\in X_{\alpha,\epsilon}$. Since $\supp(\Delta_G v)\subset \mathcal{B}_3^+$, we infer that
\begin{align}\label{eq:compact_supp2}
\|\Delta_G v\|_{C^{0,\epsilon}_\ast(Q_+)}\leq C [\Delta_G v]_{\dot C^{0,\epsilon}_\ast(Q_+)}\leq C\|v\|_{X_{\alpha,\epsilon}}.
\end{align}
Moreover, the Dirichlet-Neumann boundary conditions allow us to extend $v$ and $\Delta_Gv$ evenly about $y_{n+1}$ and oddly about $y_n$. After the extension, the assumption that $v\in C_0(Q_+)$ yields the representation
$$v(x)=\int\limits_{\R^{n+1}} K(x,y) \Delta_G v(y)dy,$$
where $K$ is the fundamental solution of $\Delta_G$ in $\R^{n+1}$ (c.f. Lemma \ref{lem:ker} in Section \ref{sec:kernel}). We remark that a priori $v$ deviates from $\int K(x,y) \Delta_G v(y)dy$ by (at most) a third order polynomial as we only control the semi-norm $[v]_{C^{2,\epsilon}_{\ast}(Q_+\setminus \mathcal{B}_1^+)}$ in the bulk and the deviation of $Y_{i}Y_{j}v$ at the boundary of $\mathcal{B}_3^+$.
However, the decay property at infinity forces $v$ to coincide with $\int K(x,y) \Delta_Gv(y) dy$. By the kernel estimates for the fundamental solution (c.f. Lemma \ref{lem:ker} in Section \ref{sec:kernel}) and by the support assumption (\ref{eq:compact_supp2})
\begin{align*}
\| v \|_{L^{\infty}(Q_+)} \leq C \| \D_G v\|_{L^\infty(Q_+)} \leq C \|v\|_{X_{\alpha,\epsilon}}.
\end{align*}
Thus, we are able to control the $L^\infty $ norm of the coefficients of the approximating polynomial $P_{\bar y}$ at each point $\bar y\in P$.
\end{proof}
Last but not least, we show the invertibility of the Baouendi-Grushin Laplacian as an operator on these spaces:
\begin{lem}
\label{lem:inverse}
Let $X_{\alpha,\epsilon},Y_{\alpha,\epsilon}$ be as in Definition~\ref{defi:spaces}. Then, $\D_G:X_{\alpha,\epsilon} \rightarrow Y_{\alpha,\epsilon}$ is an invertible operator.
\end{lem}
\begin{proof}
We show that for each $f\in Y_{\alpha,\epsilon}$, there exists a unique $u\in X_{\alpha,\epsilon}$ such that $\Delta_Gu=f$. Moreover, by Section \ref{sec:quarter_Hoelder}
\begin{align}\label{eq:apriori}
\|u\|_{X_{\alpha,\epsilon}}\leq C\|f\|_{Y_{\alpha,\epsilon}}.
\end{align}
Indeed, given $f\in Y_{\alpha,\epsilon}$, we extend $f$ oddly about $y_{n}$ and evenly about $y_{n+1}$ and (with slight
abuse of notation) still denote the extended function by $f$. Let $u(x)= \int K(x,y) f(y)dy$, where $K$ is the kernel from Section \ref{sec:kernel}. In particular, the decay estimates for $K$ (c.f. Lemma \ref{lem:ker} in Section \ref{sec:kernel}) imply that $u\in C_0(\R^{n+1})$. Since $f\in L^\infty(\R^{n+1})$ and $\supp(f)\subset \mathcal{B}_3$, we obtain that $u\in M^{2,p}(\R^{n+1})$ for any $1<p<\infty$ (c.f. the Calderon-Zygmund estimates in Section \ref{sec:kernel}). Moreover, by the symmetry of the extension, $u$ is odd in
$y_n$ and even in $y_{n+1}$, which implies that $u=0$ on $\{y_n=0\}$ and $\p_{n+1}u=0$ on
$\{y_{n+1}=0\}$. We restrict $u$ to $Q_+$ and still denote it by $u$. By the interior estimates from
the previous Section \ref{sec:quarter_Hoelder} and a scaling argument, we further obtain that
$u\in X_{\alpha,\epsilon}$ and that it satisfies \eqref{eq:apriori}.
It is immediate that $\supp(\Delta_Gu)=\supp(f)\subset \mathcal{B}_3^+$. Moreover, by using the equation, $\p_{nn}u=f-(y_n^2+y_{n+1}^2)\sum_{i=1}^{n-1}\p_{ii}u-\p_{n+1,n+1}u=0 $ on $\{y_{n}=y_{n+1}=0\}$. This shows
the existence of $u\in X_{\alpha,\epsilon}$ which satisfies $\Delta_Gu=f$. Due to \eqref{eq:apriori} such a function $u$ is unique in $X_{\alpha,\epsilon}$.
\end{proof}
\subsection{Kernel estimates for the Baouendi-Grushin Laplacian}
\label{sec:kernel}
Last but not least, we provide the arguments for the mapping properties of the Baouendi-Grushin Laplacian in the whole space setting. This in particular yields the kernel bounds, which are used in the previous subsection.\\
Our main result in this section are the following Calderon-Zygmund estimates:
\begin{prop}[Calderon-Zygmund estimates]
\label{prop:CZ}
Let $Y:=(Y_1,\dots,Y_{n+1})$ with $Y_i$ denoting the vector fields from Definition \ref{defi:Grushinvf} and let $F=(F^1,\dots,F^{n+1})\in L^p(\R^{n+1},\R^{n+1})$, $f\in L^p(\R^{n+1})$.
Suppose that
\[ \Delta_G u = Y_i F^{i}. \]
Then, there exists a constant $c_n= c(n)>0$ such that
\[ \sum\limits_{i=1}^{n+1} \Vert Y_i u \Vert_{L^p(\R^{n+1})} \le c_{n} \frac{p^2}{p-1} \Vert F \Vert_{L^p(\R^{n+1})}. \]
If
\[ \Delta_ G u = f ,\]
then there exists a constant $c_n= c(n)>0$ such that
\[ \sum\limits_{i,j=1}^{n+1} \Vert Y_{i}Y_j u \Vert_{L^p(\R^{n+1})} \le c_{n} \frac{p^2}{p-1} \Vert f \Vert_{L^p(\R^{n+1})}. \]
If $0<s<1$ and $F\in \dot C^s(\R^{n+1}, \R^{n+1})$, then there exists a constant $c_n= c(n)>0$ such that
\[ \sum\limits_{i=1}^{n+1} \Vert Y_i u \Vert_{\dot C^s(\R^{n+1})} \le c_{n} \frac1{s(1-s)} \Vert F \Vert_{\dot C^s(\R^{n+1})}. \]
Moreover, if $F$ is supported on ball of radius one,
\[ \sum\limits_{i=1}^{n+1} \Vert Y_i u (x)\Vert_{L^\infty(\R^{n+1})}
\le c \Vert F \Vert_{C^{s}(B_1)}. \]
\end{prop}
The key auxiliary result to infer the regularity estimates of Proposition \ref{prop:CZ} is the following existence and regularity result for a kernel to our problem:
\begin{lem}
\label{lem:ker}
Let $u:\R^{n+1}\rightarrow \R$ be a solution of $\D_G u = f$. Then there exists a kernel $k(z,w): \R^{(n+1)\times (n+1)}\rightarrow \R$ such that
\begin{align*}
u(x)= \int\limits_{\R^{n+1}} k(x,y)f(y)dy.
\end{align*}
Let $\tilde{Y}^{\alpha}$ denote the composition of the vector fields $\tilde{Y}_{\alpha_{1}}\dots \tilde{Y}_{\alpha_{|\alpha|}}$ where $\tilde{Y}_i$, $i\in\{1,\dots,2n\}$, denote the modified vector fields from Definition \ref{defi:Hoelder1}. Then for all multi-indeces $\alpha, \beta$ the following estimates hold
\[ \left|\tilde{Y}_{z}^\alpha \tilde{Y}_{w}^\beta k(z,w)\right|
\le c_{\alpha, \beta} d_G(z,w)^{2-|\alpha|-|\beta|} (\vol(B_{d_G(z,w)}(z)))^{-1} .
\]
Here the subscript $z,w$ in the vector fields $\tilde{Y}_z^{\alpha}, \tilde{Y}_w^{\beta}$ indicates the variable the vector fields are acting on.
\end{lem}
Relying on this representation, we can proceed to the proof of Proposition \ref{prop:CZ}:
\begin{proof}[Proof of Proposition \ref{prop:CZ}]
Let $K_{ij}(z,w)= Y_{i,z} Y_{j,w} k(z,w)$ for any pair $i,j\in\{1,\dots,n+1\}$, where the indeces $z,w$ refer to the variables which the vector field are acting on and $k(z,w)$ denotes the kernel from Lemma \ref{lem:ker}. The function $K_{ij}(z,w)$ is related to the
obvious Calderon-Zygmund operator $T$ which maps $L^p$ to $L^p$. This proves the desired $L^p$ bounds. Hence, it remains to prove the Hölder estimates. Formally, it maps constants to zero. Thus,
\[ Tf(x) = T(f-f(x))(x)= \int\limits_{\R^{n+1}} K_{ij}(x,y)(f(y)-f(x)) dy .\]
Now let $d_G(z,w) = 3$. We choose a smooth cutoff function $\phi$ which is equal to $1$ in $\mathcal{B}_1(0)$ and equal to $0$ outside $\mathcal{B}_{3/2}(0)$ and set
$f(x) = \phi(x-z) f(x) + \phi(x-w)f(x) + (1-\phi(x-z)-\phi(x-w)) f(x). $
We claim that $|T(f)(w)-T(f)(z)| \le c $. This follows from the kernel estimates of Lemma \ref{lem:ker}.
\end{proof}
Finally, to conclude our discussion of the mapping properties of the Baouendi-Grushin operator, we present the proof of Lemma \ref{lem:ker}:
\begin{proof}[Proof of Lemma \ref{lem:ker}]
We begin by considering the equation
\[ \Delta_G u = \sum\limits_{i=1}^{n+1}Y_i F \mbox{ in } \R^{n+1}. \]
For this we have the energy estimate
\[ \sum\limits_{i=1}^{n+1} \Vert Y_i u \Vert_{L^2(\R^{n+1})}^2 \le \sum\limits_{i=1}^{n+1} \Vert F^i \Vert_{L^2(\R^{n+1})}^2 . \]
Also, the Sobolev embedding
\[ \Vert u \Vert_{L^p(\R^{n+1})} \le c \sum\limits_{i=1}^{n+1} \Vert Y_i u \Vert_{L^q(\R^{n+1})} , \]
holds with
\[ \frac1q - \frac1{2n-2} = \frac1p. \]
By duality, if
\[ \frac12 - \frac1{2n-2} = \frac1p, \]
we have the embedding
\[ \Vert f \Vert_{\dot M^{-1}(\R^{n+1})} \leq c \Vert f \Vert_{L^{p'}(\R^{n+1})} .\]
Here $\dot{M}^{-1}(\R^{n+1})$ denotes the dual space of $\dot{M}^1(\R^{n+1})$ (and $\dot{M}^{1}(\R^{n+1})$ is the homogeneous version of the space introduced in Definition \ref{defi:GrushinLp} in Section \ref{sec:quarter_Hoelder}).
As discussed in Section \ref{sec:holder} the symbol of $\D_G$ defines the sub-Riemannian metric
\[ g_{y}(v,w) = (y_n^2 + y_{n+1}^2)^{-1} \sum\limits_{i=1}^{n-1} v_{i}w_{i} + v_{n}w_{n} + v_{{n+1}}w_{{n+1}}, \]
which itself correspondingly defines a metric $d_G$ on $\R^{n+1}$.
The operator $\D_G$ satisfies the Hörmander condition with the vector fields
\[ \tilde{Y}_i, \ i\in\{1,\dots,2n\}. \]
Hence, it is hypoelliptic and any local distributional solution is smooth.
More precisely, if $u \in L^1(B_1(y))$ satisfies $\Delta_G u =0$, then for any multi-index $\alpha$
\[ \Vert \tilde{Y}^{\alpha} u \Vert_{L^\infty(B_{1/2}(y))}
\le c_{\alpha} \Vert u \Vert_{L^1(B_1(y))}. \]
Let $\frac1p+\frac1{2n-2}= \frac12$ and let $p'$ the Hölder conjugate
exponent. Then, by the embeddings, if $f \in L^{p'}$ and
\[ \Delta_G u = f \mbox{ in } \R^{n+1}, \]
then $\Vert u \Vert_{L^p(\R^{n+1})} \le c \Vert f \Vert_{L^{p'}(\R^{n+1})}$.
By the Schwartz kernel theorem there is a kernel $k(z,w)$ so that
\[ u(z) = \int\limits_{\R^{n+1}} k(z,w) f(w) dw. \]
More precisely, if $f \in \dot M^{-1}$, then $u \in \dot M^{-1}$.
In particular, if $f$ is supported in ball $B_1(w)$ then $u$ is a solution to
the homogeneous problem outside. In particular if $ d_G(z,w) \ge 3$ then $u$ is bounded together with all derivatives in $B_1(z)$. We fix $z$. Then,
\[ M^{-1}(B_1(w)) \ni f \to u(z), \]
is a linear continuous map, which is represented by $\tilde w \to k(z,\tilde w) \in M^1(B_1(w))$. Since $\Delta_G$ is self-adjoint,$k(z,w) = k(w,z)$. Repeating previous arguments we see that
\[ \tilde{Y}^\alpha_w k(z,\tilde w) \]
is bounded in $B_{1/2}(w)$. Repeating the arguments and dualizing once more we obtain that
\[ |\tilde{Y}_z^\alpha \tilde{Y}^\beta_w k(z,w)| \le c, \]
provided $d_G(z,w)=1$. Hence rescaling leads to the desired kernel estimates.
\end{proof}
|
\section{Introduction}\label{sec:intro}
Toussaint \cite{T88} introduced the sphere-of-influence graph of a finite set of points in Euclidean space for applications in pattern analysis and image processing (see \cite{T2014} for a recent survey).
This notion was later generalized to so-called closed sphere-of-influence graphs \cite{HJLM93} and to $k$-th closed sphere-of-influence graphs \cite{KZ2004}.
Our setting will be a $d$-dimensional normed space $\mathcal N$ with norm $\norm{\cdot}$.
We denote the ball with center $c\in\mathcal N$ and radius $r$ by $B(c,r)$.
\begin{defn}
Let $k\in\mathbb N$ and let $\setbuilder{c_i}{i=1,\dots,m}$ be a family of points in the
$d$-dimensional normed space $\mathcal N$.
For each $i=\{1,\dots,m\}$, let $r_i^{(k)}$ be the smallest $r$ such that
\[\setbuilder{j\in\mathbb N}{j\neq i, \norm{c_i-c_j}\leq r}\] has at least $k$ elements.
Define the \emph{$k$-th closed sphere-of-influence graph on
$V=\setbuilder{c_i}{i=1,\dots,m}$} by joining $c_i$ and $c_j$ whenever
$B(c_i,r_i^{(k)})\cap B(c_j,r_j^{(k)})\neq\emptyset$.
\end{defn}
F\"uredi and Loeb \cite{FL94} gave an upper bound for the minimum degree of any closed sphere-of-influence graph in $\mathcal N$ in terms of a certain packing quantity of the space (see also \cites{MQ94, S94}.)
\begin{defn}
Let $\vartheta(\mathcal N)$ denote the largest number of points in the ball
$B(o,2)$ of the normed space $\mathcal N$ such that any two points are at distance
at least $1$, and one of the points is the origin~$o$.
\end{defn}
F\"uredi and Loeb \cite{FL94} showed that any closed sphere-of-influence graph in $\mathcal N$ has a vertex of degree smaller than $\vartheta(\mathcal N)\leq 5^d$.
(It is clear that $\vartheta(\mathcal N)$ is bounded above by the number of balls of radius $1/2$ that can be packed into a ball of radius $5/2$, which is at most $5^d$ by volume considerations.)
Guibas, Pach and Sharir \cite{GPS94} showed that any $k$-th closed sphere-of-influence graph in $d$-dimensional Euclidean space has a vertex of degree at most $c^dk$.
In this note we show the following more precise result, valid for all norms, and generalizing the result of F\"uredi and Loeb \cite{FL94} mentioned above.
\begin{thm}\label{thm:kcsig}
Every $k$-th sphere-of-influence graph on at least two points in a normed space $\mathcal N$ has at least two vertices of degree smaller than
$\vartheta(\mathcal N)k\leq 5^d k$.
\end{thm}
\begin{cor}
A $k$-th sphere-of-influence graph on $n$ points in $\mathcal N$ has at most $(\vartheta(\mathcal N)k-1)n\leq (5^d k -1)n$ edges.
\end{cor}
\begin{proof}[Proof of Theorem~\ref{thm:kcsig}]
Let $V=\{c_1,c_2,\dots,c_m\}$.
Relabel the vertices $c_1,c_2,\dots,c_m$ such that $r_1^{(k)}\leq r_2^{(k)}\leq\dots\leq r_m^{(k)}$.
We define an auxiliary graph $H$ on $V$ by joining $c_i$ and $c_j$ whenever
$\norm{c_i-c_j}<\max\{r_i^{(k)},r_j^{(k)}\}$.
Thus, if $\setbuilder{c_i}{i\in I}$ is an independent set in $H$, then no ball in $\setbuilder{B(c_i,r_i^{(k)})}{i\in I}$ contains the center of another in its interior.
We next bound the chromatic number of $H$.
\begin{lem}\label{lem:color}
The chromatic number of $H$ does not exceed $k$.
\end{lem}
\begin{proof}
Note that for each $i\in\{1,\dots,m\}$, the set \[\setbuilder{j<i}{c_ic_j\in E(H)} =
\setbuilder{j<i}{\norm{c_i-c_j}<r_i^{(k)}}\] has less than $k$ elements.
Therefore, we can greedily color $H$ in the order $c_1,
c_2,\dots, c_m$ by $k$ colors.
\end{proof}
We next show that the degrees of $c_1$ and $c_2$ (corresponding to the two smallest $r_i^{(k)}$) are both at most $\vartheta(\mathcal N)k$, which will complete the proof of Theorem~\ref{thm:kcsig}.
We first need the so-called ``bow-and-arrow'' inequality of \cite{FL94}.
\begin{lem}[F\"uredi--Loeb \cite{FL94}]\label{cor:bowarrow}
For any two non-zero elements $a$ and $b$ of a normed space,
\[ \norm{\un{a}-\un{b}}\geq\frac{\norm{a-b}-\abs{\norm{a}-\norm{b}}}{\norm{b}}.\]
\end{lem}
\begin{proof}
Without loss of generality, $\norm{a}\geq\norm{b}>0$.
Then
\begin{align*}
\norm{a-b} &=\norm{\norm{a}\un{a}-\norm{b}\un{b}}\\
&=\norm{\norm{b}(\un{a}-\un{b})+(\norm{a}-\norm{b})\un{a}}\\
&\leq \norm{b}\norm{\un{a}-\un{b}}+\norm{a}-\norm{b}. \qedhere
\end{align*}
\end{proof}
The next lemma is abstracted with minimal hypotheses from \cite{MQ94}*{Proof of Theorem~6} (see also
\cite{FL94}*{Proof of Theorem~2.1}).
\begin{lem}\label{lem:satellite}
Consider the balls $B(v_1,\lambda_1)$ and $B(v_2,\lambda_2)$ in the normed space $\mathcal N$, such that
$\max\{\lambda_1,\lambda_2\}\geq 1$, $v_1\notin \operatorname{int}(B(v_2,\lambda_2))$, $v_2\notin \operatorname{int}(B(v_1,\lambda_1))$ and $B(v_i,\lambda_i)\cap B(o,1)\neq\emptyset$ \textup{(}$i=1,2$\textup{)}.
Define $\pi\colon\mathcal N\to B(o,2)$ by \[\pi(x)=\begin{cases} x &\text{if $\norm{x}\leq 2$,}\\ \frac{2}{\norm{x}}x & \text{if $\norm{x}\geq 2$.} \end{cases}\]
Then $\norm{\pi(v_1)-\pi(v_2)}\geq 1$.
\end{lem}
\begin{proof}
In terms of the norm, we are given that $\norm{v_1-v_2}\geq\max\{\lambda_1,\lambda_2\}\geq 1$, $\norm{v_1}\leq\lambda_1+1$, and $\norm{v_2}\leq\lambda_2+1$.
Without loss of generality, $\norm{v_2}\leq\norm{v_1}$.
If $v_1,v_2\in 2K$ then $\norm{\pi(v_1)-\pi(v_2)}=\norm{v_1-v_2}\geq 1$.
If $v_1\notin 2K$ and $v_2\in 2K$, then
\begin{align*}
\norm{\pi(v_1)-\pi(v_2)} &=\norm{2\un{v_1}-v_2} \geq \norm{v_1-v_2}-\norm{v_1-2\un{v_1}}\\
&= \norm{v_1-v_2}-(\norm{v_1}-2)\geq\lambda_1 - (\lambda_1+1) + 2 =1.
\end{align*}
If $v_1,v_2\notin 2K$, then
\begin{align*}
\norm{\pi(v_1)-\pi(v_2)} &=\norm{2\un{v_1}-2\un{v_2}} \geq 2\frac{\norm{v_1-v_2}-\norm{v_1}+\norm{v_2}}{\norm{v_2}} \quad\text{by Lemma~\ref{cor:bowarrow}}\\
&\geq 2\left(\frac{\lambda_1-(\lambda_1 + 1)}{\norm{v_2}} + 1\right) = \frac{-2}{\norm{v_2}} + 2 \geq -1+2=1.\qedhere
\end{align*}
\end{proof}
We can now finish the proof of Theorem~\ref{thm:kcsig}.
Let $c\in\{c_1,c_2\}$ be the point with smallest or second-smallest $r_i^{(k)}$.
By Lemma~\ref{lem:color} we can partition the set of neighbors of $c$ in the $k$-th closed sphere-of-influence graph on $V$ into $k$ classes $N_1,\dots,N_k$ so that each $N_i$ is an independent set in $H$.
We may assume that the radius $r_i^{(k)}$ corresponding to $c$ is 1.
Then each ball in $\setbuilder{B(c_j,r_j^{(k)})}{c_j\in N_i}$ intersects $B(c,1)$, and the center of no ball is in the interior of another ball.
By Lemma~\ref{lem:satellite}, $\setbuilder{\pi(p-c)}{p\in N_i}$ is a set of points contained in $B(o,2)$ with a distance of at least $1$ between any two.
That is, $\card{N_i\setminus\operatorname{int}(B(c,1))} \leq\vartheta(\mathcal N)-1$ for each $i=1,\dots,k$.
Since there are at most $k-1$ points in $V\cap \operatorname{int}(B(c,1))\setminus\{c\}$,
it follows that the degree of $c$ is at most $\sum_{i=1}^k \card{N_i\setminus\operatorname{int}(B(c,1))} + k-1 \leq(\vartheta(\mathcal N)-1)k + k-1 = \vartheta(\mathcal N)k-1$.
\end{proof}
\bibliographystyle{amsalpha}
|
\section{Introduction}\label{sec:intro}
The $k$-disjoint paths problem and related routing problems are among
the central problems in combinatorial optimisation. In the most basic
variant of the $k$-disjoint paths problem, a graph $G$ is given with
$k$ pairs $(s_1, t_1)$, $\dots$, $(s_k, t_k)$ of vertices and the
task is to find $k$ pairwise vertex-disjoint paths linking each
$s_i$ to its corresponding target $t_i$.
The problem is well known to be NP-complete \cite{MR57:11691}. On
undirected graphs with a fixed number $k$ of source/terminal pairs,
Robertson and Seymour proved in their monumental graph minor series
\cite{GM-series} that the problem is polynomial-time
solvable. In fact, they showed that it is fixed-parameter
tractable with parameter $k$: it can be solved in cubic time for every
fixed value of $k$.
For directed graphs, the problem is computationally much harder. Fortune et
al.~\cite{FortuneHW80} proved that it is already NP-complete
for only $k=2$ source/terminal pairs. In particular, this also implies
that it is not fixed-parameter tractable on directed graphs.
Following this result a lot of work has gone into establishing more
efficient algorithms on restricted classes of digraphs.
Fortune et al.~\cite{FortuneHW80} showed that the problem can be
solved in time $n^{O(k)}$ on acyclic digraphs, that is, it is
polynomial-time for every fixed $k$. However, as proved by Slivkins
\cite{Slivkins10}, the problem is $W[1]$-hard on acyclic digraphs, and
therefore unlikely to be fixed-parameter tractable. On the other hand,
Cygan et al. \cite{CyganMPP13} proved that the problem is
fixed-parameter tractable with parmeter $k$ when restricted to planar
digraphs. Related to this, Amiri et al.~\cite{AmiriGolKreSie14} proved
that the problem remains NP-complete even in upward planar graphs, but
admits a single exponential fixed-parameter algorithm.
Disjoint paths problems have also been studied intensively in the area
of approximation algorithms, both on directed and undirected graphs (see,
e.g.,~\cite{ChekuriKS06,KolliopoulosS04,AndrewsCGKTZ10,ChekuriKS04,ChekuriKS05,Chuzhoy12,ChekuriE13,ChuzhoyLi12,ChekuriE14}).
The goal is, given an input graph $G$ and demands
$(s_1, t_1), \dots, (s_k, t_k)$ to \emph{route} as many pairs as
possible in polynomial time. There are many variations what it means
for a pair to be routable. In particular, a problem studied
intensively in the
approximation literature is a relaxed version of disjoint paths where
the paths are no longer required to be fully disjoint. Instead, they
may intersect but every vertex of the graph is allowed to be contained
in at most $c$ paths, for some fixed constant $c$. This is called
\emph{congestion $c$ routing}. In particular, the well-linked decomposition
framework developed in \cite{ChuzhoyLi12} for undirected
graphs and later generalised to digraphs in \cite{ChekuriE14} has
proved to be very valuable for obtaining good approximation
algorithms for disjoint paths problems on planar graphs and digraphs.
In this paper, we are interested in exact solutions for high congestion
routing on acyclic digraphs. More precisely, we study the following problem.
\begin{definition}
\begin{enumerate}
\item Let $G$ be a digraph and let
$I := \{ (s_1, t_1), \dots, (s_k, t_k) \}$ be a set of pairs of
vertices. Let $c\geq 1$. A \emph{$c$-routing} of $I$ is a set
$\{ P_1, \dots, P_k\}$ of paths such that, for all
$1\leq i \leq k$, path $P_i$ links $s_i$ to $t_i$ and no vertex
$v\in V(G)$ appears in more than $c$ paths from
$\{P_1, \dots, P_k\}$.
\item Let $k, c\geq 1$. In the \kcrouting{k}{c} problem, a digraph $G$ is given in the input together with a
set $I := \{(s_1, t_1), \dots, (s_k, t_k)\}$ of $k$ pairs of
vertices (the demands); the task is to decide whether there is a
$c$-routing of $I$ in $G$.
\end{enumerate}
\end{definition}
We consider \kcrouting{k}{c} on acyclic digraphs. First, it is not
very difficult to show that, for every fixed $c\ge 1$, we can
generalize the $n^{O(1)}$ time algorithm of
Fortune~et~al.~\cite{FortuneHW80} to \kcrouting{k}{c}. By revisiting
the W[1]-hardness proof of Slivkins~\cite{Slivkins10} and making appropriate
modifications, we can establish that the problem remains W[1]-hard for
every fixed congestion $c\ge 1$. Moreover, by doing the proof in a
more modern way (reducing from general subgraph isomorphism instead of
maximum clique and invoking a lower bound of Marx
\cite{marx-toc-treewidth}), we can show that the $n^{O(k)}$ time
algorithm is essentially best possible with respect to the exponent of
$n$. This lower bound is under the Exponential-Time Hypothesis (ETH),
which can be informally stated as $n$-variable \textsc{3Sat} cannot be
solved in time $2^{o(n)}$ (see
\cite{MR1894519,DBLP:journals/eatcs/LokshtanovMS11,DBLP:books/sp/CyganFKLMPPS15}
for more background).
\begin{theorem}\label{thm:hardness}
For any fixed integer $c\ge 1$, \kcrouting{k}{c} is \textup{W[1]}-hard parameterized by $k$ and, assuming ETH, cannot be solved in time $f(k)n^{o(k/\log k)}$ for any computable function $f$.
\end{theorem}
While Theorem~\ref{thm:hardness} shows that the problem does get
easier for any larger fixed congestion $c\ge 1$, intuitively one
expects the problem to get simpler at some point: after all, the
problem is trivial if $c\ge k$. Therefore, we study the complexity of
the problem in settings close to this extreme case.
The main algorithmic result of this paper is to show that for any fixed value
of $d\ge 1$, the $\kcrouting{k}{k-d}$ problem can be solved in time
$n^{O(d)}$. That is, the exponent of the polynomial bounding the
running time of the algorithm only depends on $d$ but not
on the number $k$.
\begin{theorem}\label{thm:main-algo}
For every fixed $d\ge 1$, the $\kcrouting{k}{k-d}$ problem on
acyclic digraphs can be solved in time $n^{O(d)}$.
\end{theorem}
A simple corollary of Theorem~\ref{thm:hardness} shows that
\kcrouting{k}{k-d} is unlikely to be fixed-parameter tractable and the
running time of Theorem~\ref{thm:main-algo} essentially cannot be improved
(assuming ETH). Observe that if we set $d := k-1$, then
\kcrouting{k}{k-d} is simply the standard $k$-disjoint path problem,
thus any algorithmic result for \kcrouting{k}{k-d} parameterized by
$d$ would imply the essentially same algorithmic result for the fully
disjoint version parameterized by $k$.
\begin{corollary}\label{cor:hard2}
\kcrouting{k}{k-d} is \textup{W[1]}-hard parameterized by $d$ (if $k$
is part of the input) and, assuming ETH, cannot be solved in time
$f(k)n^{o(d/\log d)}$ for any computable function $f$.
\end{corollary}
\medskip\noindent\textbf{Organisation. } The paper is organised as follows. In
Section~\ref{sec:polytime} we fix some notation and prove our main
algorithmic result. The corresponding lower bound in then proved in
Section~\ref{sec:lowerbounds}
\section{Preliminaries}
We review basic notation and concepts of graph theory needed in the
paper. We refer to \cite{Diestel10,BangJensenG10} for background.
Let $G$ be a digraph. We write $V(G)$ and $E(G)$ for its set of
vertices and edges, respectively. We assume that there is no edge with
the same head and tail, i.e.~there are no loops in the digraphs we
consider in this paper. If $(u,v)\in E(G)$ is an edge, then $u$ is its
\emph{tail} and $v$ is its \emph{head}. $G$ is \emph{simple} if there
are no two distinct edges which have the same tail and the same
head. Otherwise we call $G$ a \emph{multi digraph}.
A \emph{path} $P$ in a digraph $G$ is determined by a sequence $(v_1,
\dots, v_{\ell})$ of vertices such that $v_i\not=v_j$ for all $1 \leq i < j \leq \ell$
and $(v_i, v_{i+1})\in E(G)$ for all $1\leq i < \ell$. We write $E(P)$
for the set $\{ (v_i, v_{i+1}) \st 1\leq i \leq \ell-1\}$ of edges
appearing in $P$ and $V(P)$ for the set $\{ v_1, \dots, v_\ell\}$ of
vertices. We say that $P$ \emph{links} $v_1$ to $v_{\ell}$.
Two paths $P_1$
and $P_2$ are \emph{edge disjoint} if $E(P_1)\cap E(P_2)=\emptyset$.
\section{A polynomial-time algorithm on acyclic digraphs}
\label{sec:polytime}
In this section we prove the first main result of this paper,
Theorem~\ref{thm:main-algo}, which we repeat here for convenience.
\medskip
\begin{TheoremNo}{\ref{thm:main-algo}}
For every fixed $d\ge 1$, the $\kcrouting{k}{k-d}$ problem on
acyclic digraphs can be solved in time $n^{O(d)}$.
\end{TheoremNo}
We first need some additional notation and prove some auxiliary lemmas.
\begin{definition}
Let $G$ be a digraph and let $\mathcal{L}$ be a set of paths in $G$. For
every $v\in V(G)$ we define the \emph{congestion} of $v$ with
respect to $\mathcal{L}$ as the number of paths in $\mathcal{L}$ containing $v$.
\end{definition}
The following lemma provides a simple extension of the algorithm
from~\cite{FortuneHW80} for disjoint paths in acyclic digraphs.
\begin{lemma}\label{lem:constant-number-of-paths}
On acyclic digraphs $G$ the $\kcrouting{k}{c}$ probem can be
solved in time $n^{O(k)}$, where $n:= |G|$.
\end{lemma}
\begin{proof}
In~\cite{FortuneHW80}, Fortune et al. proved that the $k$-disjoint
paths problem can be solved in time $n^{O(k)}$ on any $n$-vertex acyclic
digraph $G$.
Let $G$, $(s_1, t_1), \dots, (s_k, t_k)$ and $c$ be given.
We construct a new digraph $H$ with $V(H) := V(G)\times \{ 1, \dots,
c\}$ and $E(H) := \{ \big( (u,i), (v, j)\big) \st (u,v)\in
E(G)\}$.
Then $H$ contains $k$ pairwise vertex disjoint paths $P_1,
\dots, P_k$ such that $P_i$ links $(s_i, 1)$ to $(t_i, 1)$ if, and
only if, there is a positive solution to the $(k, c)$-Congestion
Routing Problem on~$G$. By the algorithm in~\cite{FortuneHW80} we
can decide whether the paths $P_1, \dots, P_k$ exist in $H$ in time
$|V(H)|^{O(k)}$ and hence in time $(c\cdot n)^{O(k)} = n^{O(k)}$ as $c\leq n$.
\end{proof}
We will use this lemma in the form given in the next corollary.
\begin{corollary}\label{cor:constant-number-of-paths}
For $c, k\geq 0$ such that $k\in O(c)$, the $\kcrouting{k}{c}$
problem can be solved on any acyclic $n$-vertex digraph $G$ in time $n^{O(c)}$.
\end{corollary}
The next lemma provides the main reduction argument for proving Theorem~\ref{thm:main-algo}.
\begin{lemma}\label{lem:reducing-number-of-paths}
Let $G$ be an acyclic directed graph and let $d \geq 1$ and $k>3d$. Let
$ I := \{ (s_1,t_1), \dots, (s_k,t_k)\}\subseteq V(G)\times V(G)$ be a set of
source/terminal pairs. There exists a $(k{-}d)$-routing of $I$ if,
and only if, for
every pair $(s, t)\in I$ there is a path in $G$ from $s$ to $t$ and
there is a subset $I'\subsetneq I$ of order
$|I'| = k-1$ such that there is a $(k-d-1)$-routing of $I'$.
\end{lemma}
\begin{proof}
The if direction is easy to see. Let $\SSS' := \{P_1, \dots,
P_{k-1}\}$ be a $(k-d-1)$-routing of a set $I'\subseteq I$ of order
$k-1$. Let $s, t$ be such that $I = I' \cup \{ (s, t)\}$. By
assumption there is a simple path $P$ from $s$ to $t$ in
$G$. Then $\SSS := \SSS' \cup \{ P\}$ is a $(k-d)$-routing of $I$.
For the reverse direction let $I := \{ (s_1, t_1), \dots, (s_k,
t_k)\}$ and let $\hat\SSS := \{ \hat P_1, \dots, \hat P_k\}$ be
a $(k-d)$-routing of $I$ such that $\hat P_i$ links $s_i$ to $t_i$, for
all $1\leq i \leq k$.
We define a multi digraph $G'$ on the same vertex set $V(G)$ as $G$
as follows. For every pair $u, v\in V(G')$ such that $e = (u,v)\in
E(G)$ and every $1\leq i \leq k$, if $e$ occurs on the path $\hat P_i \in
\SSS$, then we add a new edge $e^i = (u, v)$ to $G'$. Hence, if any
edge $e\in E(G)$ is used by $\ell$ different paths in $\hat\SSS$, then
$G'$ contains $\ell$ parallel edges between the endpoints of $e$. In
the rest of the proof we will work on the multi digraph $G'$. We can
now take a set $\SSS := \{ P_1, \dots, P_k\}$ of pairwise edge
disjoint paths, where $P_i$ is the
path from $s_i$ to $t_i$ induced by the edge set $\{ e^i \st e\in
E(\hat{P}_i)\}$. That is, by using the parallel edges, we can turn
the routing $\hat{\SSS}$ into a
$(k{-}d)$-routing $\SSS$ of $I$ where the
paths are mutually edge disjoint.
In the remainder of the proof we will construct a subset
$I'\subsetneq I$ of order $k-1$ and a $(k-d-1)$-routing of
$I'$ in $G'$ which is pairwise edge disjoint. This naturally induces
a $(k-d-1)$-routing of $I'$ in $G$. Note that in $G'$, if
$\mathcal{L}$ is any set of pairwise edge disjoint paths, then the
congestion of any vertex with respect to $\mathcal{L}$ is at most the
congestion of the vertex with respect to $\SSS$ (and thus $\hat
\SSS$) in $G'$ (and $G$, respectively).
Let $\sqsubseteq$ be a topological ordering of $G'$ and let
$A := \{ a_1, \dots, a_{\ell}\}$ be the set of vertices of congestion
$k-d$ with respect to $\SSS$ such that $a_i \sqsubseteq a_j$ whenever
$i<j$. As $k>3d$, for all $1\leq i < {\ell}$ there is a path in $G$ from
$a_i$ to $a_{i+1}$.
For $1\leq i \leq k$, an \emph{atomic subpath} of $P_i$ (with respect to
$\SSS$) is a subpath of $P_i$ that starts and ends in a vertex of
$A\cup \{ s_i, t_i\}$ and is internally vertex disjoint from $A$.
Hence, every path $P_i\in \SSS$ consists of the concatenation $P^1_1
\cdot \dots \cdot P_i^{{\ell}_i}$ of its atomic subpaths where we identify the
last vertex of $P^j_i$ with the first vertex of $P^{j+1}_i$ for all $1\le j <
\ell_i$. Note that any
two atomic subpaths of paths $P_i, P_j$ in $\SSS$ are pairwise edge
disjoint.
Let $I' \subset I$ be a subset of order $k-1$. A routing $\SSS' :=
\{ P'_1, \dots, P'_{k-1}\}$ of
$I'$ is \emph{conservative with respect to $\SSS$} if it consists of
pairwise edge disjoint paths and every path in
$\SSS'$ consists of a concatenation of atomic subpaths of paths in
$\SSS$. In the sequel,
whenever we speak of a conservative $I'$-routing we implicitly mean
that it is conservative with respect to $\SSS$.
If $\SSS'$ is a conservative $I'$-routing with respect to
$\SSS$, then it consists of pairwise edge disjoint paths and hence
for every $v\in V(G)$ the congestion of $v$ with respect to $\SSS'$
is at most the congestion of $v$ with respect to $\SSS$.
Let $1\leq i_1 < i_2\leq {\ell}$ and let $1\leq j \leq k$. Let $\SSS'$ be
a conservative $I'$-routing. An \emph{$(i_1, i_2)$-jump of colour $j$} is a subpath $P'$ of
$P_j$ from $a_{i_1}$ to $a_{i_2}$ such that for all $i$ with $i_1 < i < i_2$
the vertex $a_i$ is not on $P_j$. Note that any jump is an atomic
subpath. We call the jump $P'$
\emph{free with respect to $\SSS'$} if $P'$ is not used by any path
in $\SSS'$.
We are now ready to complete the proof of the lemma. Note first
that, as $k>3d$,
for any three vertices $b_1, b_2, b_3\in A$ there is a path $P\in
\SSS$ that contains $b_1, b_2, b_3$. Hence, we can choose an $h \in \{
1, \dots, k\}$ such that $a_1, a_{\ell} \in V(P_h)$ and there is a
vertex $a_r$ with $1<r<\ell$ such that $a_r\in V(P_h)$.
Let $I' := I\setminus \{ (s_h, t_h)\}$.
If $A\subseteq V(P_h)$, then $\SSS \setminus \{ P_h\}$ is a $(k-d-1)$-routing
of $I'$ and we are done. Otherwise, for every
vertex $a_r\in A$ which has congestion $k-d$ with respect to $\SSS
\setminus \{ P_h\}$ there are $i, j$ with $i < r < j$ and an $(i,j)$-jump $P$
of colour~$h$. This follows as $a_1, a_{\ell}\in V(P_h)$. Note also that $a_1$
and $a_{\ell}$ have congestion $k-d-1$ in $\SSS\setminus \{ P_h\}$.
Note that this jump $P$ is free with respect to $\SSS\setminus \{
P_h\}$.
Thus, it is easily seen that $\SSS\setminus\{P_h\}$ satisfies the
following two properties:
\begin{enumerate}
\item For every vertex $a_r$ of congestion $k-d$ with
respect to $\SSS\setminus \{P_h\}$ there are indices $i<r<j$ such
that there is a free $(i,j)$-jump $P$ with respect to
$\SSS\setminus \{P_h\}$.
\item For any three vertices $b_1, b_2, b_3$ of congestion $k-d$
with respect to $\SSS\setminus \{P_h\}$ there is a path $Q\in \SSS\setminus \{P_h\}$ with $\{b_1,
b_2, b_3\} \subseteq V(Q)$.
\end{enumerate}
Now let $\SSS'$ be a routing of $I'$ which satisfies Condition
$1$ and $2$ (with respect to $\SSS'$ instead of
$\SSS\setminus\{P_h\}$) and, subject to this, the number of vertices of
congestion $k-d$ with respect to $\SSS'$ is minimal.
We claim that $\SSS'$ is a $(k-d-1)$-routing of $I'$.
Let $\SSS' := \{ Q_1, \dots, Q_{k-1}\}$. Towards a
contradiction, suppose there is a vertex $a_r$ of congestion $k-d$
with respect to $\SSS'$. As $\SSS'$ is conservative, we have $a_r\in A$. Hence, by
assumption, there are $i<r<j$ and a free $(i,j)$-jump $P$ with
respect to $\SSS'$.
Let $Q_h$ be a path in $\SSS'$ that contains $a_i, a_r$ and $a_j$, which
exists by Condition $2$.
Let $Q_h := Q_h^1\cup Q_h^2\cup Q_h^3$ where
\begin{itemize}
\item $Q_h^1$ is the initial subpath of $Q_h$ from its first vertex
to $a_i$,
\item $Q_h^2$ is the subpath starting at $a_i$ and ending in
$a_j$ and
\item $Q_h^3$ is the subpath starting in $a_j$ and ending at
the end of $Q_h$.
\end{itemize}
We define $Q'_h := Q_h^1 \cup P \cup Q_h^2$, i.e.~$Q'_h$ is the path obtained
from $Q_h$ by replacing the inner subpath $Q_h^2$ by the
$(i,j)$-jump $P$.
Let $\mathcal{L} := (\SSS'\setminus \{ Q_h \}) \cup \{ Q'_h \}$. Then
$\mathcal{L}$ is a routing of $I'$. It is also conservative as we have only
rerouted a single path along a free jump.
We need to show that for all $b_1, b_2, b_3$ of congestion $k-d$
with respect to $\mathcal{L}$ there is a path $Q\in \mathcal{L}$ containing $b_1,
b_2, b_3$. By assumption, such
a path $Q'$ exists in $\SSS'$. If $Q' \not= Q_h$, then we are
done. So suppose $Q_h = Q'$. But then this implies that $b_s\not\in \{ a_{i+1}, \dots, a_{j-1}\}$ for
all $1\leq s \leq 3$ as otherwise the congestion of $b_s$ would have
dropped to $k-d-1$ in $\mathcal{L}$. But then $b_1, b_2, b_3\in V(Q'_h)$.
It remains to show that for every vertex $a_s$ of congestion $k-d$
with respect to $\mathcal{L}$ there is a free $(i, j)$-jump for some
$i<s<j$.
As before, by assumption, there are $s_1 < s < s_2$ and a free $(s_1,
s_2)$-jump with respect to $\SSS'$. If this jump is not $P$, then it
still exists with respect to $\mathcal{L}$ and we are done. So suppose this
jump is $P$, which implies that $i<s<j$. Furthermore, $a_s\not\in
Q_h$ as otherwise the congestion of $a_s$ in $\mathcal{L}$ would be
$k-d-1$. But then, there must be indices $i_1, i_2$ with $i\leq i_1
< s < i_2\leq j$ such that $a_{i_1}, a_{i_2} \in V(Q_h)$ and
$a_{s'}\not\in V(Q_h)$ for all $i_1 < s' < i_2$. Hence, the atomic
subpath $Q''$ of $Q_h$ from $a_{i_1}$ to $a_{i_2}$ is an $(i_1, i_2)$-jump
as required. As $Q''\subseteq Q_h^2$, this jump is now free.
Finally, the vertex $a_r$ now has congestion $k-d-1$ with respect to
$\mathcal{L}$ as $a_r$ is not contained in $Q'_h$. Hence, $\mathcal{L}$ has fewer
vertices of congestion $k-d$ than $\SSS'$, contradicting the
choice of $\SSS'$. Thus, $\SSS'$ must have been a $(k-d-1)$-routing
of $I'$ as required. This completes the proof of the lemma.
\end{proof}
The previous lemma has the following consequene which essentially
implies Theorem~\ref{thm:main-algo}.
\begin{corollary}\label{cor:reducing-number-of-paths}
Let $G$ be an acyclic digraph, let $d\geq 0$ and let $I := \{ (s_1,
t_1), \dots, (s_k, t_k)\}$ be a set of pairs of vertices such that
for all $1\leq i \leq k$ there is a path in $G$ linking $s_i$ to $t_i$. $G$
contains a $(k-d)$-routing of $I$ if, and only if, there is a subset
$I'\subseteq I$ of order $|I'| \leq 3d$ such that $G$ contains a
$(k{-}dc)$-routing of $I'$. If $k\geq 3d$, then $I'$ can be chosen of
size exactly $3d$.
\end{corollary}
We are now ready to prove Theorem~\ref{thm:main-algo}.
\begin{proof}[Proof of Theorem~\ref{thm:main-algo}]
Let $G, k, d$ and $I := \{(s_1, t_1), \dots, (s_k, t_k)\}$ be
given. Let $n := |G|$. If for some $1\leq i \leq k$ there is no path
in $G$ from $s_i$ to $t_i$, then the answer is no and we are
done. If $k\leq 3d$, then we can apply
Corollary~\ref{cor:constant-number-of-paths} to compute the answer in
time $n^{O(d)}$ as required.
Otherwise, by Corollary~\ref{cor:reducing-number-of-paths},
there is a $(k-d)$-routing for $I$ in $G$ if, and only if, there is
a subset $I'\subsetneq I$ of order $3d$ such that $I'$ has a
$(3d-d)$-routing. There are ${k \choose 3d} \leq k^{3d} \leq n^{3d}$
subsets $I'$ of order $3d$. By
Corollary~\ref{cor:constant-number-of-paths}, we can decide for any
such $I'$ of order $3d$ in time $n^{O(d)}$ whether a
$(3d-d)$-routing of $I'$ exists. Hence, by iterating through all
possible subsets $I'$, we can decide in time $n^{O(d)}$ whether
there is a $(k-d)$-routing of $I$ in $G$.
\end{proof}
\section{Lower Bounds}\label{sec:lowerbounds}
\newcommand{\textsc{Partitioned Subgraph Isomorphism}\xspace}{\textsc{Partitioned Subgraph Isomorphism}\xspace}
In this section, we prove Theorem~\ref{thm:hardness} by a reduction from \textsc{Partitioned Subgraph Isomorphism}\xspace.
The input of the \textsc{Partitioned Subgraph Isomorphism}\xspace problem consists of a graph $H$ with vertex set
$\{u_1,\dots,u_k\}$ and a graph $G$ whose vertex set is partitioned
into $k$ classes $V_1$, $\dots$, $V_k$. The task is to find a mapping
$\mu:V(H)\to V(G)$ such that $\mu(u_i)\in V_i$ for every $1\le i \le
k$ and $\mu$ is a subgraph embedding, that is, if $u_i$ and $u_j$ are
adjacent in $H$, then $\mu(u_i)$ and $\mu(u_j)$ are adjacent in
$G$.
\begin{theorem}[\cite{marx-toc-treewidth}]\label{thm:subgraph}
Assuming ETH, \textsc{Partitioned Subgraph Isomorphism}\xspace cannot be solved in time $f(k)n^{o(k/\log
k)}$ (where $k=|V(H)|$) for any computable function $f$, even when $H$ is assumed to be 3-regular and bipartite.
\end{theorem}
To prove Theorem~\ref{thm:hardness}, we need a reduction from \textsc{Partitioned Subgraph Isomorphism}\xspace (for 3-regular bipartite graphs) to \kcrouting{k}{c}, where the number $k$ of demands is linear in the number of vertics of $H$.
\newcommand{\overline{Q}}{\overline{Q}}
\newcommand{\underline{Q}}{\underline{Q}}
\newcommand{\overline{q}}{\overline{q}}
\newcommand{\underline{q}}{\underline{q}}
\begin{proof}[Proof (of Theorem~\ref{thm:hardness})]
We prove the theorem by a reduction from \textsc{Partitioned Subgraph Isomorphism}\xspace. Let $H$ and $G$
be two graphs, let $V(H)=\{u_1,\dots, u_{k}\}$, and let
$(V_1,\dots, V_{k})$ be a partition of $V(G)$. By copying vertices
if necessary, we may assume that every $V_i$ has the same size $n$;
let us denote by $\{v_{i,1},\dots,v_{i,n}\}$ the vertices in
$V_i$. By Theorem~\ref{thm:subgraph}, we may assume that $H$ is
3-regular and bipartite. This means that $H$ has exactly $h=3k/2$
edges and both partite classes contain $k/2$ vertices. Without loss
of generality, we can assume that $U_1=\{u_1,\dots,u_{k/2}\}$ and
$U_2=\{u_{k/2+1},\dots,u_k\}$ are the two partite classes. Let us
fix an arbitrary ordering $e_1$, $\dots$, $e_{h}$ of the edges of
$H$.
\textbf{Construction.} We construct an instance of $(k,c)$-Congestion Routing the following
way. We construct a directed graph $D$ that contains, for every
$1\le i \le k$, two directed paths $\overline{Q}_i$ and $\underline{Q}_i$ (see Figure~\ref{fig:hard}). Path
$\overline{Q}_i$ has $n(h+1)+1$ vertices: it contains the vertices $\overline{q}_{i,0}$,
$\dots$, $\overline{q}_{i,n}$ in this order and additionally, for every
$1\le j \le n$, the vertices $\overline{q}_{i,j,1}$, $\dots$, $\overline{q}_{i,j,h}$
are inserted between $\overline{q}_{i,j-1}$ and $\overline{q}_{i,j}$. The path $\underline{Q}_i$
is defined the same way, with vertices $\underline{q}$ instead of $\overline{q}$. For
every $1\le \ell \le h$, we introduce two vertices $s_\ell$ and
$t_\ell$. Then we complete the construction of the graph $D$ by
introducing further edges as follows.
\begin{figure}[t]
\begin{center}
{\small \svg{\linewidth}{hard}}
\end{center}
\caption{Part of the directed graph $D$ constructed in the proof of Theorem~\ref{thm:hardness} with $k=4$, $h=6$, and $n=5$. For clarity, we consider only one edge $e_4$ of $H$, which connects $u_1$ and $u_3$, and assume that the only edge between $V_1$ and $V_3$ is between $v_{1,3}$ and $v_{3,5}$. The highlighted red paths show the paths $P^v_1$, $P^v_3$, and $P^e_4$ of the solution.}\label{fig:hard}
\end{figure}
\begin{itemize}
\item For every $1\le i \le k$ and $1\le j \le n$, we introduce the edge $(\overline{q}_{i,j-1},\underline{q}_{i,j})$ (the curved bypass edges in Figure~\ref{fig:hard}).
\item For every $1\le i \le k$, $1\le j \le n$, and $1\le s \le h$, we introduce the edge $(\overline{q}_{i,j,s},\underline{q}_{i,j,s})$ (the vertical edges in Figure~\ref{fig:hard}).
\item For every $1\le \ell \le h$, we do the following. Suppose that edge
$e_\ell$ of $H$ connects $u_{i_a}$ and $u_{i_b}$ for some $1\le i_a \le k/2$ and
$k/2+1\le i_b \le k$. Then for every pair of vertices $v_{i_a,j_a}\in V_{i_a}$ and $v_{i_b,j_b}\in V_{i_b}$ that are adjacent in $G$, we introduce the following three edges into $D$: $(s_\ell,\overline{q}_{i_a,j_a,\ell})$, $(\underline{q}_{i_a,j_a,\ell},\overline{q}_{i_b,j_b,\ell})$, and $(\underline{q}_{i_b,j_b,\ell},t_\ell)$.
\end{itemize}
To complete the construction of the $(k,c)$-Congestion Routing instance, we define the following set of $k+2k(c-1)+h$ demands:
\begin{itemize}
\item For every $1\le i \le k$, we introduce the demand $(\overline{q}_{i,0},\underline{q}_{i,n})$ (vertex demands).
\item For every $1\le i \le k$, we introduce $c-1$ copies of the demand $(\overline{q}_{i,0},\overline{q}_{i,n})$ (blocking demands).
\item For every $1\le i \le k$, we introduce $c-1$ copies of the demand $(\underline{q}_{i,0},\underline{q}_{i,n})$ (blocking demands).
\item For every $1 \le \ell \le h$, we introduce the demand $(s_\ell,t_\ell)$ (edge demands).
\end{itemize}
Note that, for every fixed $c\ge 1$, the number of demands is $O(k)$. In the rest of the proof, we show that a routing with congestion $c$ exists if and only if the \textsc{Partitioned Subgraph Isomorphism}\xspace instance has a solution. Then the \textup{W[1]}-hardness and lower bound stated in Theorem~\ref{thm:subgraph} implies the same hardness results for the routing problem.
\textbf{Subgraph embedding $\Rightarrow$ routing.}
Suppose first that vertices $v_{1,z_1}\in V_1$, $\dots$, $v_{k,z_k}\in V_k$ form a solution to the \textsc{Partitioned Subgraph Isomorphism}\xspace instance. We construct a routing that contains the following paths, satisfying the demands defined above:
\begin{itemize}
\item For every $1\le i \le k$, the vertex demand $(\overline{q}_{i,0},\underline{q}_{i,n})$ is satisfied by
a path $P^v_i$ that goes from $\overline{q}_{i,0}$ to $\overline{q}_{i,z_i-1}$ on $\overline{Q}_i$, uses the edge $(\overline{q}_{i,z_i-1},\underline{q}_{i,z_i})$, and the goes from $\underline{q}_{i,z_i}$ to $\underline{q}_{i,n}$ on $\underline{Q}_i$.
\item For every $1\le i \le k$, each of the $c-1$ copies of the blocking demand $(\overline{q}_{i,0},\overline{q}_{i,n})$ is satisfied by a path going on $\overline{Q}_i$.
\item For every $1\le i \le k$, each of the $c-1$ copies of the blocking demand $(\underline{q}_{i,0},\underline{q}_{i,n})$ is satisfied by a path going on $\underline{Q}_i$.
\item For every $1 \le \ell \le h$, the edge demand $(s_\ell,t_\ell)$ is satisfied by a 5-edge path $P^e_\ell=(s_\ell,\overline{q}_{i_a,z_{i_a},\ell},$ $\underline{q}_{i_a,z_{i_b},\ell},\overline{q}_{i_b,z_{i_b},\ell},\underline{q}_{i_b,z_{i_b},\ell},t_\ell)$.
\end{itemize}
It is easy to verify that these are indeed paths: all the required
edges exist. We claim that each vertex of $D$ is used by at most $c$
of these paths. It is easy to see that two paths $P^v_{i'}$ and $P^v_{i''}$ with $i\neq i''$ satisfying vertex
demands do intersect, and this is also true for any two paths $P^e_{\ell'}$ and $P^e_{\ell''}$ with $\ell'\neq \ell''$
satisfying edge demands (note that each vertex of the path $P^e_\ell$ has $\ell$ in its index). The crucial
observation is that the path $P^v_i$ does intersect the path $P^e_\ell$ for any $\ell$. The only way this could possibly happen is
if edge $e_\ell$ of $H$ connects $u_{i_a}$ with $u_{i_b}$, and $i$ is
equal to $i_a$ or $i_b$. But the path $P^e_\ell$
uses only vertex $\overline{q}_{i_a,z_{i_a},\ell}$ from $\overline{Q}_{i_a}$ and vertex
$\underline{q}_{i_a,z_{i_b},\ell}$ from $\underline{Q}_{i_b}$, while the path $P^v_i$ does not use these vertices, as it jumps from
$\overline{q}_{i,z_i-1}$ to $\underline{q}_{i,z_i}$. Thus each vertex is used by at most $c-1$ paths satisfying a blocking demand and at most one additional path satisfying a vertex or edge demand. We can conclude that each vertex is used by at most $c$ of the paths, what we had to show.
\textbf{Routing $\Rightarrow$ subgraph embedding.} Next we show that given a routing with congestion $c$, it is possible to construct the required subgraph embedding from $H$ to $G$. It is clear that the path satisfying the blocking demand $(\overline{q}_{i,0},\overline{q}_{i,n})$ is exactly $\overline{Q}_i$: after leaving $\overline{Q}_i$, there is no way to return back to it. Similarly, the solution must use path $\underline{Q}_i$ to satisfying the blocking demand $(\underline{q}_{i,0},\underline{q}_{i,n})$. It is also clear that the path $P^v_i$ satisfying the vertex demand $(\overline{q}_{i,0},\underline{q}_{i,n})$ has to be contained in the union of $\overline{Q}_i$ and $\underline{Q}_i$. Let $1\le z_i \le n$ be the smallest value such that $\underline{q}_{i,z_i}$ is on path $P^v_i$ (note that this value is positive, as vertex $\underline{q}_{i,0}$ cannot be reached from $\overline{q}_{i,0}$). Observe that path $P^v_i$ uses every vertex of $\underline{Q}_i$ from $\underline{q}_{i,z_i}$ to $\underline{q}_{i,n}$ (as it cannot leave $\underline{Q}_i$). Moreover, since $P^v_i$ does not use the part of $\underline{Q}_i$ from $\underline{q}_{i,0}$ to $\underline{q}_{i,z_i-1}$ by definition, it has to use the part of $\overline{Q}_i$ from $\overline{q}_{i,0}$ to $\overline{q}_{i,z_i-1}$.
We claim that mapping vertex $u_i$ of $H$ to vertex $v_{i,z_i}$ of
$G$ is a correct subgraph embedding of $H$ into $G$. To show this,
suppose that edge $e_i$ of $H$ connects $u_{i_a}$ and $u_{i_b}$
with $1\le i_a \le k/2$ and $k/2+1\le i_b\le k$; we need to show
that $v_{i_a,z_{i_a}}\in V_{i_a}$ and $v_{i_b,z_{i_b}}\in V_{i_b}$ are adjacent. Consider
the path $P^e_\ell$ satisfying edge demand $(s_\ell,t_\ell)$. By
construction, the vertex of $P^e_\ell$ after $s_\ell$ has to be on
the path $\overline{Q}_{i_a}$ and the vertex of $P^e_\ell$ before $t_\ell$
has to be on $\underline{Q}_{i_b}$. The only way to go from $\overline{Q}_{i_a}$ to
$\underline{Q}_{i_b}$ is to use an edge of the form
$(\underline{q}_{i_a,j_a,\ell},\overline{q}_{i_b,j_b,\ell})$: the only way we can leave
the union of $\overline{Q}_{i_a}$ and $\underline{Q}_{i_a}$ is to enter some $\overline{Q}_i$
with $k/2+1 \le i \le k$, and there is no edge connecting
$\overline{Q}_{i_b}$ or $\underline{Q}_{i_b}$ with any $\overline{Q}_i$ with $k/2+1 \le i \le k$ and $i\neq i_b$
(this is the part of the proof where we use that $H$ is bipartite). We claim that $j_a=z_{i_a}$.
If $j>z_{i_a}$, then $\underline{q}_{i_a,j_a,\ell}$ is also used by
the $c-1$ paths satisfying the blocking demand
$(\underline{q}_{i_a,0},\underline{q}_{i_a,n})$ and (as we have seen) the path
$P^v_{i_a}$, contradicting the assumption that the routing has
congestion $c$. If $j<z_{i_a}$, then there is no way for the path
$P^e_\ell$ to reach $\underline{q}_{i_a,j_a,\ell}$ from $s_\ell$: each vertex of the path $\overline{Q}_{i_a}$ from $\overline{q}_{i_a,0}$ to $\overline{q}_{i_a,j_a}$
is used by $c-1$ paths satisfying the blocking demand
$(\underline{q}_{i_a,0},\underline{q}_{i_a,n})$ and (as shown above) by the path
$P^v_{i_a}$. This shows $j_z=z_{i_a}$ and a similar argument shows $j_b=z_{i_b}$. Now the existence of the edge
$(\underline{q}_{i_a,z_a,\ell},\overline{q}_{i_b,z_b,\ell})$ means, by construction, that $G$ contains an edge between $v_{i_a,z_a}\in V_{i_a}$ and $v_{i_b,z_b}\in V_{i_b}$, what we had to show.
\end{proof}
\section{Conclusion}
In this paper we have studied the $\kcrouting{k}{c}$ problem on
acyclic digraphs. It is easy to see that the $n^{O(k)}$ algorithm in \cite{FortuneHW80} for solving the
disjoint paths problem on acyclic digraphs can
be extended to an $n^{O(k)}$ algorithm for $\kcrouting{k}{c}$. As we
proved in Theorem~\ref{thm:hardness}, the $n^{O(k)}$ time
algorithm is essentially best possible with respect to the exponent of
$n$, under the Exponential-Time Hypothesis (ETH). We therefore studied
the extreme cases of relatively high congestion $k-d$ for some fixed
value of $d$. In Theorem~\ref{thm:main-algo} we showed that in this
case we can obtain an $n^{O(d)}$ algorithm on acyclic digraphs,
i.e.~the algorithm only depends on the offset $d$ in
$\kcrouting{k}{k-d}$ but not on the number $k$ of demand pairs. The
proof relied on a reduction argument that shows that as long as $k$ is big
enough compared to $d$, then a demand pair can be eliminated without
changing the answer.
It will be interesting to see whether our result can be extended to
larger classes of digraphs. In particular classes of digraphs of
bounded directed tree width would be a natural target. On such
classes, the $k$-disjoint paths problem can be solved in time
$n^{O(k + w)}$, where $w$ is the directed tree width of the input
digraph (see \cite{JohnsonRST01}). It is conceivable that our results
extend to bounded directed tree width classes and we leave this for
future research.
\bibliographystyle{plainurl}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.